[llvm] [X86] Fold VPERMV(MASK,CONCAT(LO,HI)) -> VPERMV3(WIDEN(LO),MASK',WIDEN(HI)) (PR #129708)

Simon Pilgrim via llvm-commits llvm-commits at lists.llvm.org
Wed Mar 5 05:20:45 PST 2025


https://github.com/RKSimon updated https://github.com/llvm/llvm-project/pull/129708

>From c464a85b95928bd688ad90058f83902dc79b0b9d Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Tue, 4 Mar 2025 13:59:38 +0000
Subject: [PATCH] [X86] Fold VPERMV(MASK,CONCAT(LO,HI)) ->
 VPERMV3(WIDEN(LO),MASK',WIDEN(HI))

If the VPERMV node is shuffling a source that is concatenated from separate subvectors, attempt to shuffle from the separate subvectors directly using an equivalent VPERMV3 node
---
 llvm/lib/Target/X86/X86ISelLowering.cpp       |   37 +
 .../any_extend_vector_inreg_of_broadcast.ll   |  172 +-
 .../vector-interleaved-store-i16-stride-3.ll  | 1488 +++----
 .../vector-interleaved-store-i16-stride-4.ll  |  244 +-
 .../vector-interleaved-store-i16-stride-6.ll  | 2396 ++++++------
 .../vector-interleaved-store-i16-stride-7.ll  | 3436 +++++++++--------
 .../vector-interleaved-store-i16-stride-8.ll  | 2010 ++++------
 .../vector-interleaved-store-i32-stride-8.ll  |   64 +-
 .../vector-interleaved-store-i8-stride-5.ll   | 1044 +++--
 .../vector-interleaved-store-i8-stride-6.ll   | 1012 +++--
 .../vector-interleaved-store-i8-stride-7.ll   | 2878 +++++++-------
 .../vector-interleaved-store-i8-stride-8.ll   |   74 +-
 llvm/test/CodeGen/X86/vector-reduce-ctpop.ll  |   38 +-
 13 files changed, 7308 insertions(+), 7585 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 882833f15b432..4f249da09095d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -42606,6 +42606,43 @@ static SDValue combineTargetShuffle(SDValue N, const SDLoc &DL,
 
     return SDValue();
   }
+  case X86ISD::VPERMV: {
+    // Combine VPERMV to VPERMV3 if the source operand can be freely split.
+    SmallVector<int, 32> Mask;
+    SmallVector<SDValue, 2> SrcOps, SubOps;
+    SDValue Src = peekThroughBitcasts(N.getOperand(1));
+    if ((Subtarget.hasVLX() ||
+         (VT.is512BitVector() && Subtarget.hasAVX512())) &&
+        getTargetShuffleMask(N, /*AllowSentinelZero=*/false, SrcOps, Mask) &&
+        collectConcatOps(Src.getNode(), SubOps, DAG)) {
+      assert(Mask.size() == NumElts && "Unexpected shuffle mask size");
+      assert(SrcOps.size() == 1 && "Unexpected shuffle ops");
+      assert((SubOps.size() == 2 || SubOps.size() == 4) &&
+             "Unexpected split ops");
+      // Bail if we were permuting a widened vector.
+      if (SubOps[SubOps.size() - 1].isUndef())
+        return SDValue();
+      // Bail if any subops would have folded into the concat.
+      if (any_of(SubOps, [](SDValue Op) { return isShuffleFoldableLoad(Op); }))
+        return SDValue();
+      // Concat 4x128 back to 2x256.
+      if (SubOps.size() == 4) {
+        SubOps[0] = concatSubVectors(SubOps[0], SubOps[1], DAG, DL);
+        SubOps[1] = concatSubVectors(SubOps[2], SubOps[3], DAG, DL);
+      }
+      // Convert mask to 2 operand shuffle.
+      int HalfElts = NumElts / 2;
+      for (int &M : Mask)
+        M += M >= HalfElts ? HalfElts : 0;
+      SDValue Lo = widenSubVector(SubOps[0], false, Subtarget, DAG, DL,
+                                  VT.getSizeInBits());
+      SDValue Hi = widenSubVector(SubOps[1], false, Subtarget, DAG, DL,
+                                  VT.getSizeInBits());
+      return lowerShuffleWithPERMV(DL, VT, Mask, DAG.getBitcast(VT, Lo),
+                                   DAG.getBitcast(VT, Hi), Subtarget, DAG);
+    }
+    return SDValue();
+  }
   case X86ISD::VPERMV3: {
     // Combine VPERMV3 to widened VPERMV if the two source operands can be
     // freely concatenated.
diff --git a/llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast.ll b/llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast.ll
index 951a2b4cafa26..70c71ea6c8983 100644
--- a/llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast.ll
+++ b/llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast.ll
@@ -4242,16 +4242,15 @@ define void @vec384_i32_widen_to_i64_factor2_broadcast_to_v6i64_factor6(ptr %in.
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512F-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,13,0,15]
-; AVX512F-NEXT:    vpermd %zmm0, %zmm1, %zmm0
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rdx), %ymm1, %ymm1
-; AVX512F-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa %ymm0, (%rcx)
-; AVX512F-NEXT:    vmovdqa %ymm1, 32(%rcx)
+; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512F-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,21,0,23]
+; AVX512F-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm2, %ymm0
+; AVX512F-NEXT:    vpaddb 32(%rdx), %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb (%rdx), %ymm2, %ymm1
+; AVX512F-NEXT:    vmovdqa %ymm1, (%rcx)
+; AVX512F-NEXT:    vmovdqa %ymm0, 32(%rcx)
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -4259,16 +4258,15 @@ define void @vec384_i32_widen_to_i64_factor2_broadcast_to_v6i64_factor6(ptr %in.
 ; AVX512DQ:       # %bb.0:
 ; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512DQ-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,13,0,15]
-; AVX512DQ-NEXT:    vpermd %zmm0, %zmm1, %zmm0
-; AVX512DQ-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rdx), %ymm1, %ymm1
-; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vmovdqa %ymm0, (%rcx)
-; AVX512DQ-NEXT:    vmovdqa %ymm1, 32(%rcx)
+; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,21,0,23]
+; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-NEXT:    vextracti64x4 $1, %zmm2, %ymm0
+; AVX512DQ-NEXT:    vpaddb 32(%rdx), %ymm0, %ymm0
+; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm2, %ymm1
+; AVX512DQ-NEXT:    vmovdqa %ymm1, (%rcx)
+; AVX512DQ-NEXT:    vmovdqa %ymm0, 32(%rcx)
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
@@ -4371,16 +4369,15 @@ define void @vec384_i32_widen_to_i96_factor3_broadcast_to_v4i96_factor4(ptr %in.
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512F-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,13,14,0]
-; AVX512F-NEXT:    vpermd %zmm0, %zmm1, %zmm0
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rdx), %ymm1, %ymm1
-; AVX512F-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa %ymm0, (%rcx)
-; AVX512F-NEXT:    vmovdqa %ymm1, 32(%rcx)
+; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512F-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,21,22,0]
+; AVX512F-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm2, %ymm0
+; AVX512F-NEXT:    vpaddb 32(%rdx), %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb (%rdx), %ymm2, %ymm1
+; AVX512F-NEXT:    vmovdqa %ymm1, (%rcx)
+; AVX512F-NEXT:    vmovdqa %ymm0, 32(%rcx)
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -4388,16 +4385,15 @@ define void @vec384_i32_widen_to_i96_factor3_broadcast_to_v4i96_factor4(ptr %in.
 ; AVX512DQ:       # %bb.0:
 ; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512DQ-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,13,14,0]
-; AVX512DQ-NEXT:    vpermd %zmm0, %zmm1, %zmm0
-; AVX512DQ-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rdx), %ymm1, %ymm1
-; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vmovdqa %ymm0, (%rcx)
-; AVX512DQ-NEXT:    vmovdqa %ymm1, 32(%rcx)
+; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,21,22,0]
+; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-NEXT:    vextracti64x4 $1, %zmm2, %ymm0
+; AVX512DQ-NEXT:    vpaddb 32(%rdx), %ymm0, %ymm0
+; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm2, %ymm1
+; AVX512DQ-NEXT:    vmovdqa %ymm1, (%rcx)
+; AVX512DQ-NEXT:    vmovdqa %ymm0, 32(%rcx)
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
@@ -4508,16 +4504,15 @@ define void @vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3(ptr %i
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512F-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,13,14,15]
-; AVX512F-NEXT:    vpermd %zmm0, %zmm1, %zmm0
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rdx), %ymm1, %ymm1
-; AVX512F-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa %ymm0, (%rcx)
-; AVX512F-NEXT:    vmovdqa %ymm1, 32(%rcx)
+; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512F-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,21,22,23]
+; AVX512F-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm2, %ymm0
+; AVX512F-NEXT:    vpaddb 32(%rdx), %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb (%rdx), %ymm2, %ymm1
+; AVX512F-NEXT:    vmovdqa %ymm1, (%rcx)
+; AVX512F-NEXT:    vmovdqa %ymm0, 32(%rcx)
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -4525,16 +4520,15 @@ define void @vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3(ptr %i
 ; AVX512DQ:       # %bb.0:
 ; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512DQ-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,13,14,15]
-; AVX512DQ-NEXT:    vpermd %zmm0, %zmm1, %zmm0
-; AVX512DQ-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rdx), %ymm1, %ymm1
-; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vmovdqa %ymm0, (%rcx)
-; AVX512DQ-NEXT:    vmovdqa %ymm1, 32(%rcx)
+; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,21,22,23]
+; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-NEXT:    vextracti64x4 $1, %zmm2, %ymm0
+; AVX512DQ-NEXT:    vpaddb 32(%rdx), %ymm0, %ymm0
+; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm2, %ymm1
+; AVX512DQ-NEXT:    vmovdqa %ymm1, (%rcx)
+; AVX512DQ-NEXT:    vmovdqa %ymm0, 32(%rcx)
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
@@ -4636,12 +4630,11 @@ define void @vec384_i32_widen_to_i192_factor6_broadcast_to_v2i192_factor2(ptr %i
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512F-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,13,14,15]
-; AVX512F-NEXT:    vpermd %zmm0, %zmm1, %zmm0
-; AVX512F-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512F-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,21,22,23]
+; AVX512F-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpaddb (%rdx), %ymm2, %ymm0
 ; AVX512F-NEXT:    vmovdqa %ymm0, (%rcx)
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
@@ -4650,12 +4643,11 @@ define void @vec384_i32_widen_to_i192_factor6_broadcast_to_v2i192_factor2(ptr %i
 ; AVX512DQ:       # %bb.0:
 ; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512DQ-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,13,14,15]
-; AVX512DQ-NEXT:    vpermd %zmm0, %zmm1, %zmm0
-; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
+; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,21,22,23]
+; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm2, %ymm0
 ; AVX512DQ-NEXT:    vmovdqa %ymm0, (%rcx)
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
@@ -4769,16 +4761,15 @@ define void @vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3(ptr %i
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512F-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpmovsxbq {{.*#+}} xmm1 = [0,7]
-; AVX512F-NEXT:    vpermq %zmm0, %zmm1, %zmm0
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rdx), %ymm1, %ymm1
-; AVX512F-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa %ymm0, (%rcx)
-; AVX512F-NEXT:    vmovdqa %ymm1, 32(%rcx)
+; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512F-NEXT:    vpmovsxbq {{.*#+}} xmm2 = [0,11]
+; AVX512F-NEXT:    vpermi2q %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm2, %ymm0
+; AVX512F-NEXT:    vpaddb 32(%rdx), %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb (%rdx), %ymm2, %ymm1
+; AVX512F-NEXT:    vmovdqa %ymm1, (%rcx)
+; AVX512F-NEXT:    vmovdqa %ymm0, 32(%rcx)
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -4786,16 +4777,15 @@ define void @vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3(ptr %i
 ; AVX512DQ:       # %bb.0:
 ; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512DQ-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm1 = [0,7]
-; AVX512DQ-NEXT:    vpermq %zmm0, %zmm1, %zmm0
-; AVX512DQ-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rdx), %ymm1, %ymm1
-; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vmovdqa %ymm0, (%rcx)
-; AVX512DQ-NEXT:    vmovdqa %ymm1, 32(%rcx)
+; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm2 = [0,11]
+; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm0, %zmm2
+; AVX512DQ-NEXT:    vextracti64x4 $1, %zmm2, %ymm0
+; AVX512DQ-NEXT:    vpaddb 32(%rdx), %ymm0, %ymm0
+; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm2, %ymm1
+; AVX512DQ-NEXT:    vmovdqa %ymm1, (%rcx)
+; AVX512DQ-NEXT:    vmovdqa %ymm0, 32(%rcx)
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
@@ -4898,12 +4888,11 @@ define void @vec384_i64_widen_to_i192_factor3_broadcast_to_v2i192_factor2(ptr %i
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512F-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpmovsxbq {{.*#+}} xmm1 = [0,7]
-; AVX512F-NEXT:    vpermq %zmm0, %zmm1, %zmm0
-; AVX512F-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512F-NEXT:    vpmovsxbq {{.*#+}} xmm2 = [0,11]
+; AVX512F-NEXT:    vpermi2q %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpaddb (%rdx), %ymm2, %ymm0
 ; AVX512F-NEXT:    vmovdqa %ymm0, (%rcx)
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
@@ -4912,12 +4901,11 @@ define void @vec384_i64_widen_to_i192_factor3_broadcast_to_v2i192_factor2(ptr %i
 ; AVX512DQ:       # %bb.0:
 ; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
 ; AVX512DQ-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm1 = [0,7]
-; AVX512DQ-NEXT:    vpermq %zmm0, %zmm1, %zmm0
-; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm0, %ymm0
+; AVX512DQ-NEXT:    vpaddb 32(%rsi), %ymm1, %ymm1
+; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm2 = [0,11]
+; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm0, %zmm2
+; AVX512DQ-NEXT:    vpaddb (%rdx), %ymm2, %ymm0
 ; AVX512DQ-NEXT:    vmovdqa %ymm0, (%rcx)
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll
index 4a2e7d55d3e88..114b9f2c858bc 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll
@@ -1674,58 +1674,59 @@ define void @store_i16_stride3_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
 ; AVX512-NEXT:    vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4],xmm3[5],xmm5[6,7]
 ; AVX512-NEXT:    vinserti128 $1, %xmm6, %ymm3, %ymm3
-; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm4[4,5,6,7]
-; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [5,5,0,6,6,0,7,7,0,8,8,0,9,9,0,10]
-; AVX512-NEXT:    vpermd (%rdx), %zmm4, %zmm5
-; AVX512-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm3))
-; AVX512-NEXT:    vmovdqa 32(%rdi), %ymm3
-; AVX512-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX512-NEXT:    vmovdqa 32(%rsi), %ymm8
-; AVX512-NEXT:    vmovdqa {{.*#+}} ymm9 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
-; AVX512-NEXT:    vpshufb %ymm9, %ymm8, %ymm8
-; AVX512-NEXT:    vpor %ymm3, %ymm8, %ymm3
-; AVX512-NEXT:    vmovdqa 48(%rdi), %xmm8
-; AVX512-NEXT:    vmovdqa 48(%rsi), %xmm10
-; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
-; AVX512-NEXT:    vpshufb %xmm7, %xmm11, %xmm7
-; AVX512-NEXT:    vprold $16, %xmm10, %xmm10
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm10[2],xmm8[3,4],xmm10[5],xmm8[6,7]
-; AVX512-NEXT:    vinserti128 $1, %xmm7, %ymm8, %ymm7
+; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm3[0,1,2,3],zmm4[4,5,6,7]
+; AVX512-NEXT:    vmovdqa (%rdx), %ymm3
+; AVX512-NEXT:    vmovdqa 32(%rdx), %ymm5
+; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [21,21,0,22,22,0,23,23,0,0,0,0,1,1,0,2]
+; AVX512-NEXT:    vpermi2d (%rdx), %zmm5, %zmm6
+; AVX512-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm4))
+; AVX512-NEXT:    vmovdqa 32(%rdi), %ymm4
+; AVX512-NEXT:    vmovdqa {{.*#+}} ymm8 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512-NEXT:    vmovdqa 32(%rsi), %ymm9
+; AVX512-NEXT:    vmovdqa {{.*#+}} ymm10 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
+; AVX512-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
+; AVX512-NEXT:    vpor %ymm4, %ymm9, %ymm4
+; AVX512-NEXT:    vmovdqa 48(%rdi), %xmm9
+; AVX512-NEXT:    vmovdqa 48(%rsi), %xmm11
+; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
+; AVX512-NEXT:    vpshufb %xmm7, %xmm12, %xmm7
+; AVX512-NEXT:    vprold $16, %xmm11, %xmm11
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm11[2],xmm9[3,4],xmm11[5],xmm9[6,7]
+; AVX512-NEXT:    vinserti128 $1, %xmm7, %ymm9, %ymm7
 ; AVX512-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm7[4,5,6,7]
-; AVX512-NEXT:    vmovdqa (%rdx), %ymm7
-; AVX512-NEXT:    vmovdqa 32(%rdx), %ymm8
-; AVX512-NEXT:    vmovdqa {{.*#+}} ymm10 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512-NEXT:    vpshufb %ymm10, %ymm8, %ymm11
-; AVX512-NEXT:    vpermd %ymm8, %ymm4, %ymm4
-; AVX512-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm4
-; AVX512-NEXT:    vinserti64x4 $1, %ymm4, %zmm11, %zmm4
-; AVX512-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm3 & mem)
-; AVX512-NEXT:    vmovdqa (%rdi), %ymm3
-; AVX512-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX512-NEXT:    vmovdqa (%rsi), %ymm6
-; AVX512-NEXT:    vpshufb %ymm9, %ymm6, %ymm6
-; AVX512-NEXT:    vpor %ymm3, %ymm6, %ymm3
-; AVX512-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm3
-; AVX512-NEXT:    vprold $16, %xmm0, %xmm6
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm8 = xmm1[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm6 = xmm8[0,1],xmm6[2],xmm8[3,4],xmm6[5],xmm8[6,7]
+; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm7[4,5,6,7]
+; AVX512-NEXT:    vmovdqa {{.*#+}} ymm7 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512-NEXT:    vpshufb %ymm7, %ymm5, %ymm9
+; AVX512-NEXT:    vpmovsxbd {{.*#+}} ymm11 = [5,5,0,6,6,0,7,7]
+; AVX512-NEXT:    vpermd %ymm5, %ymm11, %ymm5
+; AVX512-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm5
+; AVX512-NEXT:    vinserti64x4 $1, %ymm5, %zmm9, %zmm5
+; AVX512-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm4 & mem)
+; AVX512-NEXT:    vmovdqa (%rdi), %ymm4
+; AVX512-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512-NEXT:    vmovdqa (%rsi), %ymm8
+; AVX512-NEXT:    vpshufb %ymm10, %ymm8, %ymm8
+; AVX512-NEXT:    vpor %ymm4, %ymm8, %ymm4
+; AVX512-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm4
+; AVX512-NEXT:    vprold $16, %xmm0, %xmm8
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm9 = xmm1[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
 ; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
 ; AVX512-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
-; AVX512-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
-; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[4,5,6,7]
-; AVX512-NEXT:    vpshufb %ymm10, %ymm7, %ymm1
+; AVX512-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm0
+; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
+; AVX512-NEXT:    vpshufb %ymm7, %ymm3, %ymm1
 ; AVX512-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [0,0,0,0,1,1,0,2]
-; AVX512-NEXT:    vpermd %ymm7, %ymm2, %ymm2
+; AVX512-NEXT:    vpermd %ymm3, %ymm2, %ymm2
 ; AVX512-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
 ; AVX512-NEXT:    vpandn %ymm2, %ymm3, %ymm2
 ; AVX512-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm0 & zmm3)
 ; AVX512-NEXT:    vmovdqa64 %zmm1, (%rcx)
-; AVX512-NEXT:    vmovdqa64 %zmm4, 128(%rcx)
-; AVX512-NEXT:    vmovdqa64 %zmm5, 64(%rcx)
+; AVX512-NEXT:    vmovdqa64 %zmm5, 128(%rcx)
+; AVX512-NEXT:    vmovdqa64 %zmm6, 64(%rcx)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
@@ -1752,58 +1753,59 @@ define void @store_i16_stride3_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
 ; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4],xmm3[5],xmm5[6,7]
 ; AVX512-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm3, %ymm3
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm4[4,5,6,7]
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [5,5,0,6,6,0,7,7,0,8,8,0,9,9,0,10]
-; AVX512-FCP-NEXT:    vpermd (%rdx), %zmm4, %zmm5
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm3))
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm3
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm8
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm8, %ymm8
-; AVX512-FCP-NEXT:    vpor %ymm3, %ymm8, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa 48(%rdi), %xmm8
-; AVX512-FCP-NEXT:    vmovdqa 48(%rsi), %xmm10
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
-; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm11, %xmm7
-; AVX512-FCP-NEXT:    vprold $16, %xmm10, %xmm10
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm10[2],xmm8[3,4],xmm10[5],xmm8[6,7]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm7, %ymm8, %ymm7
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm3[0,1,2,3],zmm4[4,5,6,7]
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm3
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm5
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [21,21,0,22,22,0,23,23,0,0,0,0,1,1,0,2]
+; AVX512-FCP-NEXT:    vpermi2d (%rdx), %zmm5, %zmm6
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm4))
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm4
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm8 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm9
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
+; AVX512-FCP-NEXT:    vpor %ymm4, %ymm9, %ymm4
+; AVX512-FCP-NEXT:    vmovdqa 48(%rdi), %xmm9
+; AVX512-FCP-NEXT:    vmovdqa 48(%rsi), %xmm11
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
+; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm12, %xmm7
+; AVX512-FCP-NEXT:    vprold $16, %xmm11, %xmm11
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm11[2],xmm9[3,4],xmm11[5],xmm9[6,7]
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm7, %ymm9, %ymm7
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm7[4,5,6,7]
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm7
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm8
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm8, %ymm11
-; AVX512-FCP-NEXT:    vpermd %ymm8, %ymm4, %ymm4
-; AVX512-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm4
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm11, %zmm4
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm3 & mem)
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm3
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm6
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm6, %ymm6
-; AVX512-FCP-NEXT:    vpor %ymm3, %ymm6, %ymm3
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm3
-; AVX512-FCP-NEXT:    vprold $16, %xmm0, %xmm6
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm1[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm6 = xmm8[0,1],xmm6[2],xmm8[3,4],xmm6[5],xmm8[6,7]
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm7[4,5,6,7]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm5, %ymm9
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm11 = [5,5,0,6,6,0,7,7]
+; AVX512-FCP-NEXT:    vpermd %ymm5, %ymm11, %ymm5
+; AVX512-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm5
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm9, %zmm5
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm4 & mem)
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm4
+; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm8
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm8, %ymm8
+; AVX512-FCP-NEXT:    vpor %ymm4, %ymm8, %ymm4
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm4
+; AVX512-FCP-NEXT:    vprold $16, %xmm0, %xmm8
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm1[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
 ; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[4,5,6,7]
-; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm7, %ymm1
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm0
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm3, %ymm1
 ; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [0,0,0,0,1,1,0,2]
-; AVX512-FCP-NEXT:    vpermd %ymm7, %ymm2, %ymm2
+; AVX512-FCP-NEXT:    vpermd %ymm3, %ymm2, %ymm2
 ; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
 ; AVX512-FCP-NEXT:    vpandn %ymm2, %ymm3, %ymm2
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm0 & zmm3)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, (%rcx)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rcx)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rcx)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 128(%rcx)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rcx)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -1830,58 +1832,59 @@ define void @store_i16_stride3_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
 ; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4],xmm3[5],xmm5[6,7]
 ; AVX512DQ-NEXT:    vinserti128 $1, %xmm6, %ymm3, %ymm3
-; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm4[4,5,6,7]
-; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [5,5,0,6,6,0,7,7,0,8,8,0,9,9,0,10]
-; AVX512DQ-NEXT:    vpermd (%rdx), %zmm4, %zmm5
-; AVX512DQ-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm3))
-; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm3
-; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512DQ-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX512DQ-NEXT:    vmovdqa 32(%rsi), %ymm8
-; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm9 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
-; AVX512DQ-NEXT:    vpshufb %ymm9, %ymm8, %ymm8
-; AVX512DQ-NEXT:    vpor %ymm3, %ymm8, %ymm3
-; AVX512DQ-NEXT:    vmovdqa 48(%rdi), %xmm8
-; AVX512DQ-NEXT:    vmovdqa 48(%rsi), %xmm10
-; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
-; AVX512DQ-NEXT:    vpshufb %xmm7, %xmm11, %xmm7
-; AVX512DQ-NEXT:    vprold $16, %xmm10, %xmm10
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm10[2],xmm8[3,4],xmm10[5],xmm8[6,7]
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm7, %ymm8, %ymm7
+; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm3[0,1,2,3],zmm4[4,5,6,7]
+; AVX512DQ-NEXT:    vmovdqa (%rdx), %ymm3
+; AVX512DQ-NEXT:    vmovdqa 32(%rdx), %ymm5
+; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [21,21,0,22,22,0,23,23,0,0,0,0,1,1,0,2]
+; AVX512DQ-NEXT:    vpermi2d (%rdx), %zmm5, %zmm6
+; AVX512DQ-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm4))
+; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm4
+; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm8 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512DQ-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512DQ-NEXT:    vmovdqa 32(%rsi), %ymm9
+; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm10 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
+; AVX512DQ-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
+; AVX512DQ-NEXT:    vpor %ymm4, %ymm9, %ymm4
+; AVX512DQ-NEXT:    vmovdqa 48(%rdi), %xmm9
+; AVX512DQ-NEXT:    vmovdqa 48(%rsi), %xmm11
+; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
+; AVX512DQ-NEXT:    vpshufb %xmm7, %xmm12, %xmm7
+; AVX512DQ-NEXT:    vprold $16, %xmm11, %xmm11
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm11[2],xmm9[3,4],xmm11[5],xmm9[6,7]
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm7, %ymm9, %ymm7
 ; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm7[4,5,6,7]
-; AVX512DQ-NEXT:    vmovdqa (%rdx), %ymm7
-; AVX512DQ-NEXT:    vmovdqa 32(%rdx), %ymm8
-; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm10 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512DQ-NEXT:    vpshufb %ymm10, %ymm8, %ymm11
-; AVX512DQ-NEXT:    vpermd %ymm8, %ymm4, %ymm4
-; AVX512DQ-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm4
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm4, %zmm11, %zmm4
-; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm3 & mem)
-; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm3
-; AVX512DQ-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX512DQ-NEXT:    vmovdqa (%rsi), %ymm6
-; AVX512DQ-NEXT:    vpshufb %ymm9, %ymm6, %ymm6
-; AVX512DQ-NEXT:    vpor %ymm3, %ymm6, %ymm3
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm3
-; AVX512DQ-NEXT:    vprold $16, %xmm0, %xmm6
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm8 = xmm1[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm6 = xmm8[0,1],xmm6[2],xmm8[3,4],xmm6[5],xmm8[6,7]
+; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm7[4,5,6,7]
+; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm7 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512DQ-NEXT:    vpshufb %ymm7, %ymm5, %ymm9
+; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} ymm11 = [5,5,0,6,6,0,7,7]
+; AVX512DQ-NEXT:    vpermd %ymm5, %ymm11, %ymm5
+; AVX512DQ-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm5
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm5, %zmm9, %zmm5
+; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm4 & mem)
+; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm4
+; AVX512DQ-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512DQ-NEXT:    vmovdqa (%rsi), %ymm8
+; AVX512DQ-NEXT:    vpshufb %ymm10, %ymm8, %ymm8
+; AVX512DQ-NEXT:    vpor %ymm4, %ymm8, %ymm4
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm4
+; AVX512DQ-NEXT:    vprold $16, %xmm0, %xmm8
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm9 = xmm1[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
 ; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
 ; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
-; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[4,5,6,7]
-; AVX512DQ-NEXT:    vpshufb %ymm10, %ymm7, %ymm1
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm0
+; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
+; AVX512DQ-NEXT:    vpshufb %ymm7, %ymm3, %ymm1
 ; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [0,0,0,0,1,1,0,2]
-; AVX512DQ-NEXT:    vpermd %ymm7, %ymm2, %ymm2
+; AVX512DQ-NEXT:    vpermd %ymm3, %ymm2, %ymm2
 ; AVX512DQ-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
 ; AVX512DQ-NEXT:    vpandn %ymm2, %ymm3, %ymm2
 ; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm0 & zmm3)
 ; AVX512DQ-NEXT:    vmovdqa64 %zmm1, (%rcx)
-; AVX512DQ-NEXT:    vmovdqa64 %zmm4, 128(%rcx)
-; AVX512DQ-NEXT:    vmovdqa64 %zmm5, 64(%rcx)
+; AVX512DQ-NEXT:    vmovdqa64 %zmm5, 128(%rcx)
+; AVX512DQ-NEXT:    vmovdqa64 %zmm6, 64(%rcx)
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
@@ -1908,58 +1911,59 @@ define void @store_i16_stride3_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
 ; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4],xmm3[5],xmm5[6,7]
 ; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm3, %ymm3
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm4[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [5,5,0,6,6,0,7,7,0,8,8,0,9,9,0,10]
-; AVX512DQ-FCP-NEXT:    vpermd (%rdx), %zmm4, %zmm5
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm3))
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm8
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm8, %ymm8
-; AVX512DQ-FCP-NEXT:    vpor %ymm3, %ymm8, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rdi), %xmm8
-; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rsi), %xmm10
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm11, %xmm7
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm10, %xmm10
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm10[2],xmm8[3,4],xmm10[5],xmm8[6,7]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm7, %ymm8, %ymm7
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm3[0,1,2,3],zmm4[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm5
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [21,21,0,22,22,0,23,23,0,0,0,0,1,1,0,2]
+; AVX512DQ-FCP-NEXT:    vpermi2d (%rdx), %zmm5, %zmm6
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm4))
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm8 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm9
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT:    vpor %ymm4, %ymm9, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rdi), %xmm9
+; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rsi), %xmm11
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm12, %xmm7
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm11, %xmm11
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm11[2],xmm9[3,4],xmm11[5],xmm9[6,7]
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm7, %ymm9, %ymm7
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm7[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm7
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm8
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm8, %ymm11
-; AVX512DQ-FCP-NEXT:    vpermd %ymm8, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm11, %zmm4
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm3 & mem)
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm3
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm6
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm6, %ymm6
-; AVX512DQ-FCP-NEXT:    vpor %ymm3, %ymm6, %ymm3
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm3
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm0, %xmm6
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm1[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm6 = xmm8[0,1],xmm6[2],xmm8[3,4],xmm6[5],xmm8[6,7]
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm7[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm5, %ymm9
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm11 = [5,5,0,6,6,0,7,7]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm5, %ymm11, %ymm5
+; AVX512DQ-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm5
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm9, %zmm5
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm4 & mem)
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm4
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm8
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm8, %ymm8
+; AVX512DQ-FCP-NEXT:    vpor %ymm4, %ymm8, %ymm4
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm4
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm0, %xmm8
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm1[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
 ; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm7, %ymm1
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm3, %ymm1
 ; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [0,0,0,0,1,1,0,2]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm7, %ymm2, %ymm2
+; AVX512DQ-FCP-NEXT:    vpermd %ymm3, %ymm2, %ymm2
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
 ; AVX512DQ-FCP-NEXT:    vpandn %ymm2, %ymm3, %ymm2
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm0 & zmm3)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, (%rcx)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rcx)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rcx)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 128(%rcx)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rcx)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
 ;
@@ -3085,601 +3089,605 @@ define void @store_i16_stride3_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ;
 ; AVX512-LABEL: store_i16_stride3_vf64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmovdqa (%rdi), %ymm0
-; AVX512-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512-NEXT:    vmovdqa (%rsi), %ymm1
-; AVX512-NEXT:    vmovdqa {{.*#+}} ymm7 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
-; AVX512-NEXT:    vpshufb %ymm7, %ymm1, %ymm1
-; AVX512-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm3
-; AVX512-NEXT:    vmovdqa (%rsi), %xmm5
-; AVX512-NEXT:    vmovdqa64 16(%rsi), %xmm20
-; AVX512-NEXT:    vmovdqa64 32(%rsi), %xmm24
-; AVX512-NEXT:    vprold $16, %xmm5, %xmm8
-; AVX512-NEXT:    vmovdqa (%rdi), %xmm9
-; AVX512-NEXT:    vmovdqa64 16(%rdi), %xmm21
-; AVX512-NEXT:    vmovdqa 32(%rdi), %xmm4
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm10 = xmm9[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm8 = xmm10[0,1],xmm8[2],xmm10[3,4],xmm8[5],xmm10[6,7]
-; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
-; AVX512-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-NEXT:    vpshufb %xmm1, %xmm9, %xmm9
-; AVX512-NEXT:    vinserti128 $1, %xmm8, %ymm9, %ymm8
-; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm10 = zmm8[0,1,2,3],zmm3[4,5,6,7]
-; AVX512-NEXT:    vmovdqa (%rdx), %ymm3
-; AVX512-NEXT:    vmovdqa 32(%rdx), %ymm8
-; AVX512-NEXT:    vmovdqa 64(%rdx), %ymm14
-; AVX512-NEXT:    vmovdqa {{.*#+}} ymm9 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512-NEXT:    vpshufb %ymm9, %ymm3, %ymm11
-; AVX512-NEXT:    vpmovsxbd {{.*#+}} ymm16 = [0,0,0,0,1,1,0,2]
-; AVX512-NEXT:    vpermd %ymm3, %ymm16, %ymm3
-; AVX512-NEXT:    vmovdqa64 {{.*#+}} zmm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512-NEXT:    vpandn %ymm3, %ymm15, %ymm3
-; AVX512-NEXT:    vinserti64x4 $1, %ymm11, %zmm3, %zmm3
-; AVX512-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm10 & zmm15)
-; AVX512-NEXT:    vmovdqa 96(%rsi), %xmm10
-; AVX512-NEXT:    vprold $16, %xmm10, %xmm11
-; AVX512-NEXT:    vmovdqa 96(%rdi), %xmm12
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm13 = xmm12[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm11 = xmm13[0,1],xmm11[2],xmm13[3,4],xmm11[5],xmm13[6,7]
-; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3]
-; AVX512-NEXT:    vpshufb %xmm1, %xmm10, %xmm10
-; AVX512-NEXT:    vinserti128 $1, %xmm11, %ymm10, %ymm10
-; AVX512-NEXT:    vinserti64x4 $1, %ymm10, %zmm0, %zmm10
-; AVX512-NEXT:    vmovdqa 80(%rdi), %xmm12
-; AVX512-NEXT:    vmovdqa 80(%rsi), %xmm13
-; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512-NEXT:    vmovdqa {{.*#+}} xmm11 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
-; AVX512-NEXT:    vpshufb %xmm11, %xmm0, %xmm0
-; AVX512-NEXT:    vprold $16, %xmm13, %xmm13
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm12 = xmm12[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm13[2],xmm12[3,4],xmm13[5],xmm12[6,7]
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm12, %ymm0
-; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm10[4,5,6,7]
-; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm18 = [5,5,0,6,6,0,7,7,0,8,8,0,9,9,0,10]
-; AVX512-NEXT:    vpermd 64(%rdx), %zmm18, %zmm10
-; AVX512-NEXT:    vmovdqa64 {{.*#+}} zmm22 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
-; AVX512-NEXT:    vpternlogd {{.*#+}} zmm10 = zmm10 ^ (zmm22 & (zmm10 ^ zmm0))
-; AVX512-NEXT:    vmovdqa 96(%rdi), %ymm0
-; AVX512-NEXT:    vmovdqa %ymm6, %ymm2
-; AVX512-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512-NEXT:    vmovdqa 96(%rsi), %ymm5
-; AVX512-NEXT:    vmovdqa %ymm7, %ymm13
-; AVX512-NEXT:    vpshufb %ymm7, %ymm5, %ymm5
-; AVX512-NEXT:    vpor %ymm0, %ymm5, %ymm0
-; AVX512-NEXT:    vmovdqa 112(%rdi), %xmm5
-; AVX512-NEXT:    vmovdqa 112(%rsi), %xmm12
-; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm12[4],xmm5[4],xmm12[5],xmm5[5],xmm12[6],xmm5[6],xmm12[7],xmm5[7]
-; AVX512-NEXT:    vpshufb %xmm11, %xmm7, %xmm7
-; AVX512-NEXT:    vprold $16, %xmm12, %xmm12
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm12[2],xmm5[3,4],xmm12[5],xmm5[6,7]
-; AVX512-NEXT:    vinserti128 $1, %xmm7, %ymm5, %ymm5
-; AVX512-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm5
-; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm5[4,5,6,7]
-; AVX512-NEXT:    vmovdqa 96(%rdx), %ymm5
-; AVX512-NEXT:    vpermd %ymm5, %ymm18, %ymm7
-; AVX512-NEXT:    vpandnq %ymm7, %ymm22, %ymm7
-; AVX512-NEXT:    vpshufb %ymm9, %ymm5, %ymm5
-; AVX512-NEXT:    vinserti64x4 $1, %ymm7, %zmm5, %zmm17
-; AVX512-NEXT:    vmovdqa64 {{.*#+}} zmm19 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512-NEXT:    vpternlogq {{.*#+}} zmm17 = zmm17 | (zmm0 & zmm19)
-; AVX512-NEXT:    vmovdqa 64(%rdi), %ymm0
-; AVX512-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512-NEXT:    vmovdqa 64(%rsi), %ymm7
-; AVX512-NEXT:    vpshufb %ymm13, %ymm7, %ymm7
-; AVX512-NEXT:    vpor %ymm0, %ymm7, %ymm0
-; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-NEXT:    vmovdqa 64(%rsi), %xmm7
-; AVX512-NEXT:    vprold $16, %xmm7, %xmm12
-; AVX512-NEXT:    vmovdqa 64(%rdi), %xmm5
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm12[2],xmm6[3,4],xmm12[5],xmm6[6,7]
-; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
-; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
-; AVX512-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
-; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm5[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT:    vpshufb %ymm9, %ymm14, %ymm5
-; AVX512-NEXT:    vpermd %ymm14, %ymm16, %ymm6
-; AVX512-NEXT:    vpandn %ymm6, %ymm15, %ymm6
-; AVX512-NEXT:    vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm0 & zmm15)
-; AVX512-NEXT:    vmovdqa 32(%rdi), %ymm0
-; AVX512-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
-; AVX512-NEXT:    vmovdqa 32(%rsi), %ymm6
-; AVX512-NEXT:    vpshufb %ymm13, %ymm6, %ymm6
-; AVX512-NEXT:    vpor %ymm0, %ymm6, %ymm0
-; AVX512-NEXT:    vmovdqa 48(%rdi), %xmm6
-; AVX512-NEXT:    vmovdqa 48(%rsi), %xmm7
-; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
-; AVX512-NEXT:    vpshufb %xmm11, %xmm12, %xmm12
-; AVX512-NEXT:    vprold $16, %xmm7, %xmm7
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm7[2],xmm6[3,4],xmm7[5],xmm6[6,7]
+; AVX512-NEXT:    vmovdqa64 (%rdx), %zmm1
+; AVX512-NEXT:    vmovdqa64 64(%rdx), %zmm0
+; AVX512-NEXT:    vmovdqa (%rdi), %ymm2
+; AVX512-NEXT:    vmovdqa {{.*#+}} ymm5 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
+; AVX512-NEXT:    vmovdqa (%rsi), %ymm3
+; AVX512-NEXT:    vmovdqa {{.*#+}} ymm10 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
+; AVX512-NEXT:    vpshufb %ymm10, %ymm3, %ymm3
+; AVX512-NEXT:    vpor %ymm2, %ymm3, %ymm2
+; AVX512-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm4
+; AVX512-NEXT:    vmovdqa (%rsi), %xmm6
+; AVX512-NEXT:    vmovdqa64 16(%rsi), %xmm21
+; AVX512-NEXT:    vmovdqa64 32(%rsi), %xmm25
+; AVX512-NEXT:    vprold $16, %xmm6, %xmm8
+; AVX512-NEXT:    vmovdqa (%rdi), %xmm11
+; AVX512-NEXT:    vmovdqa64 16(%rdi), %xmm22
+; AVX512-NEXT:    vmovdqa 32(%rdi), %xmm7
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm12 = xmm11[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm8[2],xmm12[3,4],xmm8[5],xmm12[6,7]
+; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3]
+; AVX512-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
 ; AVX512-NEXT:    vinserti128 $1, %xmm12, %ymm6, %ymm6
+; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm12 = zmm6[0,1,2,3],zmm4[4,5,6,7]
+; AVX512-NEXT:    vmovdqa (%rdx), %ymm6
+; AVX512-NEXT:    vmovdqa 32(%rdx), %ymm4
+; AVX512-NEXT:    vmovdqa 64(%rdx), %ymm15
+; AVX512-NEXT:    vmovdqa {{.*#+}} ymm11 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512-NEXT:    vpshufb %ymm11, %ymm6, %ymm13
+; AVX512-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [0,0,0,0,1,1,0,2]
+; AVX512-NEXT:    vpermd %ymm6, %ymm17, %ymm6
+; AVX512-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
+; AVX512-NEXT:    vpandnq %ymm6, %ymm16, %ymm6
+; AVX512-NEXT:    vinserti64x4 $1, %ymm13, %zmm6, %zmm19
+; AVX512-NEXT:    vpternlogq {{.*#+}} zmm19 = zmm19 | (zmm12 & zmm16)
+; AVX512-NEXT:    vmovdqa 96(%rsi), %xmm12
+; AVX512-NEXT:    vprold $16, %xmm12, %xmm13
+; AVX512-NEXT:    vmovdqa 96(%rdi), %xmm14
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm8 = xmm14[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm13[2],xmm8[3,4],xmm13[5],xmm8[6,7]
+; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512-NEXT:    vpshufb %xmm3, %xmm12, %xmm12
+; AVX512-NEXT:    vinserti128 $1, %xmm8, %ymm12, %ymm8
+; AVX512-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm8
+; AVX512-NEXT:    vmovdqa 80(%rdi), %xmm13
+; AVX512-NEXT:    vmovdqa 80(%rsi), %xmm14
+; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512-NEXT:    vmovdqa {{.*#+}} xmm12 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
+; AVX512-NEXT:    vpshufb %xmm12, %xmm6, %xmm6
+; AVX512-NEXT:    vprold $16, %xmm14, %xmm14
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1],xmm14[2],xmm13[3,4],xmm14[5],xmm13[6,7]
+; AVX512-NEXT:    vinserti128 $1, %xmm6, %ymm13, %ymm6
+; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm8[4,5,6,7]
+; AVX512-NEXT:    vmovdqa 96(%rdx), %ymm8
+; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm20 = [5,5,0,6,6,0,7,7,0,16,16,0,17,17,0,18]
+; AVX512-NEXT:    vpermt2d %zmm8, %zmm20, %zmm0
+; AVX512-NEXT:    vmovdqa64 {{.*#+}} zmm23 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm23 & (zmm0 ^ zmm6))
+; AVX512-NEXT:    vmovdqa 96(%rdi), %ymm6
+; AVX512-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
+; AVX512-NEXT:    vmovdqa 96(%rsi), %ymm9
+; AVX512-NEXT:    vmovdqa %ymm10, %ymm14
+; AVX512-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
+; AVX512-NEXT:    vpor %ymm6, %ymm9, %ymm6
+; AVX512-NEXT:    vmovdqa 112(%rdi), %xmm9
+; AVX512-NEXT:    vmovdqa 112(%rsi), %xmm13
+; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm13[4],xmm9[4],xmm13[5],xmm9[5],xmm13[6],xmm9[6],xmm13[7],xmm9[7]
+; AVX512-NEXT:    vpshufb %xmm12, %xmm10, %xmm10
+; AVX512-NEXT:    vprold $16, %xmm13, %xmm13
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm13[2],xmm9[3,4],xmm13[5],xmm9[6,7]
+; AVX512-NEXT:    vinserti128 $1, %xmm10, %ymm9, %ymm9
+; AVX512-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm9[4,5,6,7]
+; AVX512-NEXT:    vpermd %ymm8, %ymm20, %ymm9
+; AVX512-NEXT:    vpandnq %ymm9, %ymm23, %ymm9
+; AVX512-NEXT:    vpshufb %ymm11, %ymm8, %ymm8
+; AVX512-NEXT:    vinserti64x4 $1, %ymm9, %zmm8, %zmm18
+; AVX512-NEXT:    vmovdqa64 {{.*#+}} zmm8 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512-NEXT:    vpternlogq {{.*#+}} zmm18 = zmm18 | (zmm6 & zmm8)
+; AVX512-NEXT:    vmovdqa 64(%rdi), %ymm6
+; AVX512-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
+; AVX512-NEXT:    vmovdqa 64(%rsi), %ymm9
+; AVX512-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
+; AVX512-NEXT:    vpor %ymm6, %ymm9, %ymm6
 ; AVX512-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm6
-; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm6[4,5,6,7]
-; AVX512-NEXT:    vpermd %ymm8, %ymm18, %ymm6
-; AVX512-NEXT:    vpandnq %ymm6, %ymm22, %ymm6
-; AVX512-NEXT:    vpshufb %ymm9, %ymm8, %ymm7
-; AVX512-NEXT:    vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm0 & zmm19)
-; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm2
-; AVX512-NEXT:    vprold $16, %xmm24, %xmm0
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm7 = xmm4[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm0 = xmm7[0,1],xmm0[2],xmm7[3,4],xmm0[5],xmm7[6,7]
-; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
-; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-NEXT:    vmovdqa64 %xmm20, %xmm4
-; AVX512-NEXT:    vmovdqa64 %xmm21, %xmm1
-; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
-; AVX512-NEXT:    vpshufb %xmm11, %xmm2, %xmm2
-; AVX512-NEXT:    vprold $16, %xmm20, %xmm4
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm21[1,1,2,2]
-; AVX512-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm4[2],xmm1[3,4],xmm4[5],xmm1[6,7]
-; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
-; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT:    vpermd (%rdx), %zmm18, %zmm1
-; AVX512-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm22 & (zmm1 ^ zmm0))
+; AVX512-NEXT:    vmovdqa 64(%rsi), %xmm9
+; AVX512-NEXT:    vprold $16, %xmm9, %xmm10
+; AVX512-NEXT:    vmovdqa 64(%rdi), %xmm13
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm2 = xmm13[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm10[2],xmm2[3,4],xmm10[5],xmm2[6,7]
+; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm13[0],xmm9[0],xmm13[1],xmm9[1],xmm13[2],xmm9[2],xmm13[3],xmm9[3]
+; AVX512-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
+; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm9, %ymm2
+; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm6[4,5,6,7]
+; AVX512-NEXT:    vpshufb %ymm11, %ymm15, %ymm6
+; AVX512-NEXT:    vpermd %ymm15, %ymm17, %ymm9
+; AVX512-NEXT:    vpandnq %ymm9, %ymm16, %ymm9
+; AVX512-NEXT:    vinserti64x4 $1, %ymm6, %zmm9, %zmm6
+; AVX512-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm2 & zmm16)
+; AVX512-NEXT:    vmovdqa 32(%rdi), %ymm2
+; AVX512-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
+; AVX512-NEXT:    vmovdqa 32(%rsi), %ymm9
+; AVX512-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
+; AVX512-NEXT:    vpor %ymm2, %ymm9, %ymm2
+; AVX512-NEXT:    vmovdqa 48(%rdi), %xmm9
+; AVX512-NEXT:    vmovdqa 48(%rsi), %xmm10
+; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512-NEXT:    vpshufb %xmm12, %xmm13, %xmm13
+; AVX512-NEXT:    vprold $16, %xmm10, %xmm10
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm10[2],xmm9[3,4],xmm10[5],xmm9[6,7]
+; AVX512-NEXT:    vinserti128 $1, %xmm13, %ymm9, %ymm9
+; AVX512-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm9[4,5,6,7]
+; AVX512-NEXT:    vpermd %ymm4, %ymm20, %ymm9
+; AVX512-NEXT:    vpandnq %ymm9, %ymm23, %ymm9
+; AVX512-NEXT:    vpshufb %ymm11, %ymm4, %ymm10
+; AVX512-NEXT:    vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; AVX512-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm9 | (zmm2 & zmm8)
+; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm5
+; AVX512-NEXT:    vprold $16, %xmm25, %xmm2
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm8 = xmm7[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm2 = xmm8[0,1],xmm2[2],xmm8[3,4],xmm2[5],xmm8[6,7]
+; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX512-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
+; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm5, %ymm2
+; AVX512-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
+; AVX512-NEXT:    vmovdqa64 %xmm21, %xmm7
+; AVX512-NEXT:    vmovdqa64 %xmm22, %xmm3
+; AVX512-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7]
+; AVX512-NEXT:    vpshufb %xmm12, %xmm5, %xmm5
+; AVX512-NEXT:    vprold $16, %xmm21, %xmm7
+; AVX512-NEXT:    vpshufd {{.*#+}} xmm3 = xmm22[1,1,2,2]
+; AVX512-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm7[2],xmm3[3,4],xmm7[5],xmm3[6,7]
+; AVX512-NEXT:    vinserti128 $1, %xmm5, %ymm3, %ymm3
+; AVX512-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[4,5,6,7]
+; AVX512-NEXT:    vpermt2d %zmm4, %zmm20, %zmm1
+; AVX512-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm23 & (zmm1 ^ zmm2))
 ; AVX512-NEXT:    vmovdqa64 %zmm1, 64(%rcx)
-; AVX512-NEXT:    vmovdqa64 %zmm6, 128(%rcx)
-; AVX512-NEXT:    vmovdqa64 %zmm5, 192(%rcx)
-; AVX512-NEXT:    vmovdqa64 %zmm17, 320(%rcx)
-; AVX512-NEXT:    vmovdqa64 %zmm10, 256(%rcx)
-; AVX512-NEXT:    vmovdqa64 %zmm3, (%rcx)
+; AVX512-NEXT:    vmovdqa64 %zmm9, 128(%rcx)
+; AVX512-NEXT:    vmovdqa64 %zmm6, 192(%rcx)
+; AVX512-NEXT:    vmovdqa64 %zmm18, 320(%rcx)
+; AVX512-NEXT:    vmovdqa64 %zmm0, 256(%rcx)
+; AVX512-NEXT:    vmovdqa64 %zmm19, (%rcx)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512-FCP-LABEL: store_i16_stride3_vf64:
 ; AVX512-FCP:       # %bb.0:
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm0
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm1
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
-; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm3
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm5
-; AVX512-FCP-NEXT:    vmovdqa64 16(%rsi), %xmm20
-; AVX512-FCP-NEXT:    vmovdqa64 32(%rsi), %xmm24
-; AVX512-FCP-NEXT:    vprold $16, %xmm5, %xmm8
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm9
-; AVX512-FCP-NEXT:    vmovdqa64 16(%rdi), %xmm21
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm4
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm10 = xmm9[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm10[0,1],xmm8[2],xmm10[3,4],xmm8[5],xmm10[6,7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm9, %xmm9
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm8, %ymm9, %ymm8
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm10 = zmm8[0,1,2,3],zmm3[4,5,6,7]
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm3
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm8
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %ymm14
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm3, %ymm11
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm16 = [0,0,0,0,1,1,0,2]
-; AVX512-FCP-NEXT:    vpermd %ymm3, %ymm16, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512-FCP-NEXT:    vpandn %ymm3, %ymm15, %ymm3
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm10 & zmm15)
-; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %xmm10
-; AVX512-FCP-NEXT:    vprold $16, %xmm10, %xmm11
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm12
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm12[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm11 = xmm13[0,1],xmm11[2],xmm13[3,4],xmm11[5],xmm13[6,7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm10
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm11, %ymm10, %ymm10
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm0, %zmm10
-; AVX512-FCP-NEXT:    vmovdqa 80(%rdi), %xmm12
-; AVX512-FCP-NEXT:    vmovdqa 80(%rsi), %xmm13
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
-; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm0
-; AVX512-FCP-NEXT:    vprold $16, %xmm13, %xmm13
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm12 = xmm12[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm13[2],xmm12[3,4],xmm13[5],xmm12[6,7]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm0, %ymm12, %ymm0
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm10[4,5,6,7]
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm18 = [5,5,0,6,6,0,7,7,0,8,8,0,9,9,0,10]
-; AVX512-FCP-NEXT:    vpermd 64(%rdx), %zmm18, %zmm10
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm22 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm10 = zmm10 ^ (zmm22 & (zmm10 ^ zmm0))
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %ymm0
-; AVX512-FCP-NEXT:    vmovdqa %ymm6, %ymm2
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %ymm5
-; AVX512-FCP-NEXT:    vmovdqa %ymm7, %ymm13
-; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm5, %ymm5
-; AVX512-FCP-NEXT:    vpor %ymm0, %ymm5, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 112(%rdi), %xmm5
-; AVX512-FCP-NEXT:    vmovdqa 112(%rsi), %xmm12
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm12[4],xmm5[4],xmm12[5],xmm5[5],xmm12[6],xmm5[6],xmm12[7],xmm5[7]
-; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm7, %xmm7
-; AVX512-FCP-NEXT:    vprold $16, %xmm12, %xmm12
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm12[2],xmm5[3,4],xmm12[5],xmm5[6,7]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm7, %ymm5, %ymm5
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm5
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm5[4,5,6,7]
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %ymm5
-; AVX512-FCP-NEXT:    vpermd %ymm5, %ymm18, %ymm7
-; AVX512-FCP-NEXT:    vpandnq %ymm7, %ymm22, %ymm7
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm5, %ymm5
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm5, %zmm17
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm19 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm17 = zmm17 | (zmm0 & zmm19)
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %ymm7
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm7, %ymm7
-; AVX512-FCP-NEXT:    vpor %ymm0, %ymm7, %ymm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %xmm7
-; AVX512-FCP-NEXT:    vprold $16, %xmm7, %xmm12
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm5
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm12[2],xmm6[3,4],xmm12[5],xmm6[6,7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm5[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm14, %ymm5
-; AVX512-FCP-NEXT:    vpermd %ymm14, %ymm16, %ymm6
-; AVX512-FCP-NEXT:    vpandn %ymm6, %ymm15, %ymm6
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm0 & zmm15)
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm6
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm6, %ymm6
-; AVX512-FCP-NEXT:    vpor %ymm0, %ymm6, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 48(%rdi), %xmm6
-; AVX512-FCP-NEXT:    vmovdqa 48(%rsi), %xmm7
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
-; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm12, %xmm12
-; AVX512-FCP-NEXT:    vprold $16, %xmm7, %xmm7
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm7[2],xmm6[3,4],xmm7[5],xmm6[6,7]
+; AVX512-FCP-NEXT:    vmovdqa64 (%rdx), %zmm1
+; AVX512-FCP-NEXT:    vmovdqa64 64(%rdx), %zmm0
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm2
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm5 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm3
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm3, %ymm3
+; AVX512-FCP-NEXT:    vpor %ymm2, %ymm3, %ymm2
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm4
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm6
+; AVX512-FCP-NEXT:    vmovdqa64 16(%rsi), %xmm21
+; AVX512-FCP-NEXT:    vmovdqa64 32(%rsi), %xmm25
+; AVX512-FCP-NEXT:    vprold $16, %xmm6, %xmm8
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm11
+; AVX512-FCP-NEXT:    vmovdqa64 16(%rdi), %xmm22
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm7
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm12 = xmm11[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm8[2],xmm12[3,4],xmm8[5],xmm12[6,7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
 ; AVX512-FCP-NEXT:    vinserti128 $1, %xmm12, %ymm6, %ymm6
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm12 = zmm6[0,1,2,3],zmm4[4,5,6,7]
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm6
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm4
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %ymm15
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm11 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm6, %ymm13
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [0,0,0,0,1,1,0,2]
+; AVX512-FCP-NEXT:    vpermd %ymm6, %ymm17, %ymm6
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
+; AVX512-FCP-NEXT:    vpandnq %ymm6, %ymm16, %ymm6
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm6, %zmm19
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm19 = zmm19 | (zmm12 & zmm16)
+; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %xmm12
+; AVX512-FCP-NEXT:    vprold $16, %xmm12, %xmm13
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm14
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm14[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm13[2],xmm8[3,4],xmm13[5],xmm8[6,7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm12, %xmm12
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm8, %ymm12, %ymm8
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm8
+; AVX512-FCP-NEXT:    vmovdqa 80(%rdi), %xmm13
+; AVX512-FCP-NEXT:    vmovdqa 80(%rsi), %xmm14
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm12 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
+; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm6, %xmm6
+; AVX512-FCP-NEXT:    vprold $16, %xmm14, %xmm14
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1],xmm14[2],xmm13[3,4],xmm14[5],xmm13[6,7]
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm13, %ymm6
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm8[4,5,6,7]
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %ymm8
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm20 = [5,5,0,6,6,0,7,7,0,16,16,0,17,17,0,18]
+; AVX512-FCP-NEXT:    vpermt2d %zmm8, %zmm20, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm23 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm23 & (zmm0 ^ zmm6))
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %ymm6
+; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
+; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %ymm9
+; AVX512-FCP-NEXT:    vmovdqa %ymm10, %ymm14
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
+; AVX512-FCP-NEXT:    vpor %ymm6, %ymm9, %ymm6
+; AVX512-FCP-NEXT:    vmovdqa 112(%rdi), %xmm9
+; AVX512-FCP-NEXT:    vmovdqa 112(%rsi), %xmm13
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm13[4],xmm9[4],xmm13[5],xmm9[5],xmm13[6],xmm9[6],xmm13[7],xmm9[7]
+; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm10, %xmm10
+; AVX512-FCP-NEXT:    vprold $16, %xmm13, %xmm13
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm13[2],xmm9[3,4],xmm13[5],xmm9[6,7]
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm10, %ymm9, %ymm9
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm9[4,5,6,7]
+; AVX512-FCP-NEXT:    vpermd %ymm8, %ymm20, %ymm9
+; AVX512-FCP-NEXT:    vpandnq %ymm9, %ymm23, %ymm9
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm8, %ymm8
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm8, %zmm18
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm8 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm18 = zmm18 | (zmm6 & zmm8)
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %ymm6
+; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
+; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %ymm9
+; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
+; AVX512-FCP-NEXT:    vpor %ymm6, %ymm9, %ymm6
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm6
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm6[4,5,6,7]
-; AVX512-FCP-NEXT:    vpermd %ymm8, %ymm18, %ymm6
-; AVX512-FCP-NEXT:    vpandnq %ymm6, %ymm22, %ymm6
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm8, %ymm7
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm0 & zmm19)
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm24, %xmm2
-; AVX512-FCP-NEXT:    vprold $16, %xmm24, %xmm0
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm7 = xmm4[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm0 = xmm7[0,1],xmm0[2],xmm7[3,4],xmm0[5],xmm7[6,7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm20, %xmm4
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm21, %xmm1
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
-; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm2
-; AVX512-FCP-NEXT:    vprold $16, %xmm20, %xmm4
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm1 = xmm21[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm4[2],xmm1[3,4],xmm4[5],xmm1[6,7]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-FCP-NEXT:    vpermd (%rdx), %zmm18, %zmm1
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm22 & (zmm1 ^ zmm0))
+; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %xmm9
+; AVX512-FCP-NEXT:    vprold $16, %xmm9, %xmm10
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm13
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm2 = xmm13[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm10[2],xmm2[3,4],xmm10[5],xmm2[6,7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm13[0],xmm9[0],xmm13[1],xmm9[1],xmm13[2],xmm9[2],xmm13[3],xmm9[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm9, %ymm2
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm6[4,5,6,7]
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm15, %ymm6
+; AVX512-FCP-NEXT:    vpermd %ymm15, %ymm17, %ymm9
+; AVX512-FCP-NEXT:    vpandnq %ymm9, %ymm16, %ymm9
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm9, %zmm6
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm2 & zmm16)
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm2
+; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm9
+; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
+; AVX512-FCP-NEXT:    vpor %ymm2, %ymm9, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 48(%rdi), %xmm9
+; AVX512-FCP-NEXT:    vmovdqa 48(%rsi), %xmm10
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm13, %xmm13
+; AVX512-FCP-NEXT:    vprold $16, %xmm10, %xmm10
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm10[2],xmm9[3,4],xmm10[5],xmm9[6,7]
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm13, %ymm9, %ymm9
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm9[4,5,6,7]
+; AVX512-FCP-NEXT:    vpermd %ymm4, %ymm20, %ymm9
+; AVX512-FCP-NEXT:    vpandnq %ymm9, %ymm23, %ymm9
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm4, %ymm10
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm9 | (zmm2 & zmm8)
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm25, %xmm5
+; AVX512-FCP-NEXT:    vprold $16, %xmm25, %xmm2
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm7[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm8[0,1],xmm2[2],xmm8[3,4],xmm2[5],xmm8[6,7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm5, %ymm2
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm21, %xmm7
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm22, %xmm3
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7]
+; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm5, %xmm5
+; AVX512-FCP-NEXT:    vprold $16, %xmm21, %xmm7
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm3 = xmm22[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm7[2],xmm3[3,4],xmm7[5],xmm3[6,7]
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm5, %ymm3, %ymm3
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2d %zmm4, %zmm20, %zmm1
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm23 & (zmm1 ^ zmm2))
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rcx)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 128(%rcx)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 192(%rcx)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, 320(%rcx)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 256(%rcx)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, (%rcx)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, 128(%rcx)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 192(%rcx)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, 320(%rcx)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 256(%rcx)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, (%rcx)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
 ; AVX512DQ-LABEL: store_i16_stride3_vf64:
 ; AVX512DQ:       # %bb.0:
-; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm0
-; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512DQ-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512DQ-NEXT:    vmovdqa (%rsi), %ymm1
-; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm7 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
-; AVX512DQ-NEXT:    vpshufb %ymm7, %ymm1, %ymm1
-; AVX512DQ-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm3
-; AVX512DQ-NEXT:    vmovdqa (%rsi), %xmm5
-; AVX512DQ-NEXT:    vmovdqa64 16(%rsi), %xmm20
-; AVX512DQ-NEXT:    vmovdqa64 32(%rsi), %xmm24
-; AVX512DQ-NEXT:    vprold $16, %xmm5, %xmm8
-; AVX512DQ-NEXT:    vmovdqa (%rdi), %xmm9
-; AVX512DQ-NEXT:    vmovdqa64 16(%rdi), %xmm21
-; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %xmm4
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm10 = xmm9[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm8 = xmm10[0,1],xmm8[2],xmm10[3,4],xmm8[5],xmm10[6,7]
-; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
-; AVX512DQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm9, %xmm9
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm8, %ymm9, %ymm8
-; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm10 = zmm8[0,1,2,3],zmm3[4,5,6,7]
-; AVX512DQ-NEXT:    vmovdqa (%rdx), %ymm3
-; AVX512DQ-NEXT:    vmovdqa 32(%rdx), %ymm8
-; AVX512DQ-NEXT:    vmovdqa 64(%rdx), %ymm14
-; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm9 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512DQ-NEXT:    vpshufb %ymm9, %ymm3, %ymm11
-; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} ymm16 = [0,0,0,0,1,1,0,2]
-; AVX512DQ-NEXT:    vpermd %ymm3, %ymm16, %ymm3
-; AVX512DQ-NEXT:    vmovdqa64 {{.*#+}} zmm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512DQ-NEXT:    vpandn %ymm3, %ymm15, %ymm3
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm11, %zmm3, %zmm3
-; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm10 & zmm15)
-; AVX512DQ-NEXT:    vmovdqa 96(%rsi), %xmm10
-; AVX512DQ-NEXT:    vprold $16, %xmm10, %xmm11
-; AVX512DQ-NEXT:    vmovdqa 96(%rdi), %xmm12
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm13 = xmm12[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm11 = xmm13[0,1],xmm11[2],xmm13[3,4],xmm11[5],xmm13[6,7]
-; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3]
-; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm10, %xmm10
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm11, %ymm10, %ymm10
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm10, %zmm0, %zmm10
-; AVX512DQ-NEXT:    vmovdqa 80(%rdi), %xmm12
-; AVX512DQ-NEXT:    vmovdqa 80(%rsi), %xmm13
-; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512DQ-NEXT:    vmovdqa {{.*#+}} xmm11 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
-; AVX512DQ-NEXT:    vpshufb %xmm11, %xmm0, %xmm0
-; AVX512DQ-NEXT:    vprold $16, %xmm13, %xmm13
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm12 = xmm12[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm13[2],xmm12[3,4],xmm13[5],xmm12[6,7]
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm12, %ymm0
-; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm10[4,5,6,7]
-; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm18 = [5,5,0,6,6,0,7,7,0,8,8,0,9,9,0,10]
-; AVX512DQ-NEXT:    vpermd 64(%rdx), %zmm18, %zmm10
-; AVX512DQ-NEXT:    vmovdqa64 {{.*#+}} zmm22 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
-; AVX512DQ-NEXT:    vpternlogd {{.*#+}} zmm10 = zmm10 ^ (zmm22 & (zmm10 ^ zmm0))
-; AVX512DQ-NEXT:    vmovdqa 96(%rdi), %ymm0
-; AVX512DQ-NEXT:    vmovdqa %ymm6, %ymm2
-; AVX512DQ-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512DQ-NEXT:    vmovdqa 96(%rsi), %ymm5
-; AVX512DQ-NEXT:    vmovdqa %ymm7, %ymm13
-; AVX512DQ-NEXT:    vpshufb %ymm7, %ymm5, %ymm5
-; AVX512DQ-NEXT:    vpor %ymm0, %ymm5, %ymm0
-; AVX512DQ-NEXT:    vmovdqa 112(%rdi), %xmm5
-; AVX512DQ-NEXT:    vmovdqa 112(%rsi), %xmm12
-; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm12[4],xmm5[4],xmm12[5],xmm5[5],xmm12[6],xmm5[6],xmm12[7],xmm5[7]
-; AVX512DQ-NEXT:    vpshufb %xmm11, %xmm7, %xmm7
-; AVX512DQ-NEXT:    vprold $16, %xmm12, %xmm12
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm12[2],xmm5[3,4],xmm12[5],xmm5[6,7]
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm7, %ymm5, %ymm5
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm5
-; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm5[4,5,6,7]
-; AVX512DQ-NEXT:    vmovdqa 96(%rdx), %ymm5
-; AVX512DQ-NEXT:    vpermd %ymm5, %ymm18, %ymm7
-; AVX512DQ-NEXT:    vpandnq %ymm7, %ymm22, %ymm7
-; AVX512DQ-NEXT:    vpshufb %ymm9, %ymm5, %ymm5
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm7, %zmm5, %zmm17
-; AVX512DQ-NEXT:    vmovdqa64 {{.*#+}} zmm19 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm17 = zmm17 | (zmm0 & zmm19)
-; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %ymm0
-; AVX512DQ-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512DQ-NEXT:    vmovdqa 64(%rsi), %ymm7
-; AVX512DQ-NEXT:    vpshufb %ymm13, %ymm7, %ymm7
-; AVX512DQ-NEXT:    vpor %ymm0, %ymm7, %ymm0
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512DQ-NEXT:    vmovdqa 64(%rsi), %xmm7
-; AVX512DQ-NEXT:    vprold $16, %xmm7, %xmm12
-; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %xmm5
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm12[2],xmm6[3,4],xmm12[5],xmm6[6,7]
-; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
-; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
-; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm5[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-NEXT:    vpshufb %ymm9, %ymm14, %ymm5
-; AVX512DQ-NEXT:    vpermd %ymm14, %ymm16, %ymm6
-; AVX512DQ-NEXT:    vpandn %ymm6, %ymm15, %ymm6
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm0 & zmm15)
-; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm0
-; AVX512DQ-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
-; AVX512DQ-NEXT:    vmovdqa 32(%rsi), %ymm6
-; AVX512DQ-NEXT:    vpshufb %ymm13, %ymm6, %ymm6
-; AVX512DQ-NEXT:    vpor %ymm0, %ymm6, %ymm0
-; AVX512DQ-NEXT:    vmovdqa 48(%rdi), %xmm6
-; AVX512DQ-NEXT:    vmovdqa 48(%rsi), %xmm7
-; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
-; AVX512DQ-NEXT:    vpshufb %xmm11, %xmm12, %xmm12
-; AVX512DQ-NEXT:    vprold $16, %xmm7, %xmm7
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm7[2],xmm6[3,4],xmm7[5],xmm6[6,7]
+; AVX512DQ-NEXT:    vmovdqa64 (%rdx), %zmm1
+; AVX512DQ-NEXT:    vmovdqa64 64(%rdx), %zmm0
+; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm2
+; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm5 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512DQ-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
+; AVX512DQ-NEXT:    vmovdqa (%rsi), %ymm3
+; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm10 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
+; AVX512DQ-NEXT:    vpshufb %ymm10, %ymm3, %ymm3
+; AVX512DQ-NEXT:    vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm4
+; AVX512DQ-NEXT:    vmovdqa (%rsi), %xmm6
+; AVX512DQ-NEXT:    vmovdqa64 16(%rsi), %xmm21
+; AVX512DQ-NEXT:    vmovdqa64 32(%rsi), %xmm25
+; AVX512DQ-NEXT:    vprold $16, %xmm6, %xmm8
+; AVX512DQ-NEXT:    vmovdqa (%rdi), %xmm11
+; AVX512DQ-NEXT:    vmovdqa64 16(%rdi), %xmm22
+; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %xmm7
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm12 = xmm11[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm8[2],xmm12[3,4],xmm8[5],xmm12[6,7]
+; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3]
+; AVX512DQ-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
 ; AVX512DQ-NEXT:    vinserti128 $1, %xmm12, %ymm6, %ymm6
+; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm12 = zmm6[0,1,2,3],zmm4[4,5,6,7]
+; AVX512DQ-NEXT:    vmovdqa (%rdx), %ymm6
+; AVX512DQ-NEXT:    vmovdqa 32(%rdx), %ymm4
+; AVX512DQ-NEXT:    vmovdqa 64(%rdx), %ymm15
+; AVX512DQ-NEXT:    vmovdqa {{.*#+}} ymm11 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512DQ-NEXT:    vpshufb %ymm11, %ymm6, %ymm13
+; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [0,0,0,0,1,1,0,2]
+; AVX512DQ-NEXT:    vpermd %ymm6, %ymm17, %ymm6
+; AVX512DQ-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
+; AVX512DQ-NEXT:    vpandnq %ymm6, %ymm16, %ymm6
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm13, %zmm6, %zmm19
+; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm19 = zmm19 | (zmm12 & zmm16)
+; AVX512DQ-NEXT:    vmovdqa 96(%rsi), %xmm12
+; AVX512DQ-NEXT:    vprold $16, %xmm12, %xmm13
+; AVX512DQ-NEXT:    vmovdqa 96(%rdi), %xmm14
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm8 = xmm14[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm13[2],xmm8[3,4],xmm13[5],xmm8[6,7]
+; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm12, %xmm12
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm8, %ymm12, %ymm8
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm8
+; AVX512DQ-NEXT:    vmovdqa 80(%rdi), %xmm13
+; AVX512DQ-NEXT:    vmovdqa 80(%rsi), %xmm14
+; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512DQ-NEXT:    vmovdqa {{.*#+}} xmm12 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
+; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm6, %xmm6
+; AVX512DQ-NEXT:    vprold $16, %xmm14, %xmm14
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1],xmm14[2],xmm13[3,4],xmm14[5],xmm13[6,7]
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm6, %ymm13, %ymm6
+; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm8[4,5,6,7]
+; AVX512DQ-NEXT:    vmovdqa 96(%rdx), %ymm8
+; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm20 = [5,5,0,6,6,0,7,7,0,16,16,0,17,17,0,18]
+; AVX512DQ-NEXT:    vpermt2d %zmm8, %zmm20, %zmm0
+; AVX512DQ-NEXT:    vmovdqa64 {{.*#+}} zmm23 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512DQ-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm23 & (zmm0 ^ zmm6))
+; AVX512DQ-NEXT:    vmovdqa 96(%rdi), %ymm6
+; AVX512DQ-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
+; AVX512DQ-NEXT:    vmovdqa 96(%rsi), %ymm9
+; AVX512DQ-NEXT:    vmovdqa %ymm10, %ymm14
+; AVX512DQ-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
+; AVX512DQ-NEXT:    vpor %ymm6, %ymm9, %ymm6
+; AVX512DQ-NEXT:    vmovdqa 112(%rdi), %xmm9
+; AVX512DQ-NEXT:    vmovdqa 112(%rsi), %xmm13
+; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm13[4],xmm9[4],xmm13[5],xmm9[5],xmm13[6],xmm9[6],xmm13[7],xmm9[7]
+; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm10, %xmm10
+; AVX512DQ-NEXT:    vprold $16, %xmm13, %xmm13
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm13[2],xmm9[3,4],xmm13[5],xmm9[6,7]
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm10, %ymm9, %ymm9
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm9[4,5,6,7]
+; AVX512DQ-NEXT:    vpermd %ymm8, %ymm20, %ymm9
+; AVX512DQ-NEXT:    vpandnq %ymm9, %ymm23, %ymm9
+; AVX512DQ-NEXT:    vpshufb %ymm11, %ymm8, %ymm8
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm9, %zmm8, %zmm18
+; AVX512DQ-NEXT:    vmovdqa64 {{.*#+}} zmm8 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm18 = zmm18 | (zmm6 & zmm8)
+; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %ymm6
+; AVX512DQ-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
+; AVX512DQ-NEXT:    vmovdqa 64(%rsi), %ymm9
+; AVX512DQ-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
+; AVX512DQ-NEXT:    vpor %ymm6, %ymm9, %ymm6
 ; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm6
-; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm6[4,5,6,7]
-; AVX512DQ-NEXT:    vpermd %ymm8, %ymm18, %ymm6
-; AVX512DQ-NEXT:    vpandnq %ymm6, %ymm22, %ymm6
-; AVX512DQ-NEXT:    vpshufb %ymm9, %ymm8, %ymm7
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm0 & zmm19)
-; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm2
-; AVX512DQ-NEXT:    vprold $16, %xmm24, %xmm0
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm7 = xmm4[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm0 = xmm7[0,1],xmm0[2],xmm7[3,4],xmm0[5],xmm7[6,7]
-; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
-; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512DQ-NEXT:    vmovdqa64 %xmm20, %xmm4
-; AVX512DQ-NEXT:    vmovdqa64 %xmm21, %xmm1
-; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
-; AVX512DQ-NEXT:    vpshufb %xmm11, %xmm2, %xmm2
-; AVX512DQ-NEXT:    vprold $16, %xmm20, %xmm4
-; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm1 = xmm21[1,1,2,2]
-; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm4[2],xmm1[3,4],xmm4[5],xmm1[6,7]
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
-; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-NEXT:    vpermd (%rdx), %zmm18, %zmm1
-; AVX512DQ-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm22 & (zmm1 ^ zmm0))
+; AVX512DQ-NEXT:    vmovdqa 64(%rsi), %xmm9
+; AVX512DQ-NEXT:    vprold $16, %xmm9, %xmm10
+; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %xmm13
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm2 = xmm13[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm10[2],xmm2[3,4],xmm10[5],xmm2[6,7]
+; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm13[0],xmm9[0],xmm13[1],xmm9[1],xmm13[2],xmm9[2],xmm13[3],xmm9[3]
+; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm9, %ymm2
+; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm6[4,5,6,7]
+; AVX512DQ-NEXT:    vpshufb %ymm11, %ymm15, %ymm6
+; AVX512DQ-NEXT:    vpermd %ymm15, %ymm17, %ymm9
+; AVX512DQ-NEXT:    vpandnq %ymm9, %ymm16, %ymm9
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm6, %zmm9, %zmm6
+; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm2 & zmm16)
+; AVX512DQ-NEXT:    vmovdqa 32(%rdi), %ymm2
+; AVX512DQ-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
+; AVX512DQ-NEXT:    vmovdqa 32(%rsi), %ymm9
+; AVX512DQ-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
+; AVX512DQ-NEXT:    vpor %ymm2, %ymm9, %ymm2
+; AVX512DQ-NEXT:    vmovdqa 48(%rdi), %xmm9
+; AVX512DQ-NEXT:    vmovdqa 48(%rsi), %xmm10
+; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm13, %xmm13
+; AVX512DQ-NEXT:    vprold $16, %xmm10, %xmm10
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm10[2],xmm9[3,4],xmm10[5],xmm9[6,7]
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm13, %ymm9, %ymm9
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm9[4,5,6,7]
+; AVX512DQ-NEXT:    vpermd %ymm4, %ymm20, %ymm9
+; AVX512DQ-NEXT:    vpandnq %ymm9, %ymm23, %ymm9
+; AVX512DQ-NEXT:    vpshufb %ymm11, %ymm4, %ymm10
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; AVX512DQ-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm9 | (zmm2 & zmm8)
+; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm5
+; AVX512DQ-NEXT:    vprold $16, %xmm25, %xmm2
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm8 = xmm7[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm2 = xmm8[0,1],xmm2[2],xmm8[3,4],xmm2[5],xmm8[6,7]
+; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm5, %ymm2
+; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
+; AVX512DQ-NEXT:    vmovdqa64 %xmm21, %xmm7
+; AVX512DQ-NEXT:    vmovdqa64 %xmm22, %xmm3
+; AVX512DQ-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7]
+; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm5, %xmm5
+; AVX512DQ-NEXT:    vprold $16, %xmm21, %xmm7
+; AVX512DQ-NEXT:    vpshufd {{.*#+}} xmm3 = xmm22[1,1,2,2]
+; AVX512DQ-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm7[2],xmm3[3,4],xmm7[5],xmm3[6,7]
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm5, %ymm3, %ymm3
+; AVX512DQ-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[4,5,6,7]
+; AVX512DQ-NEXT:    vpermt2d %zmm4, %zmm20, %zmm1
+; AVX512DQ-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm23 & (zmm1 ^ zmm2))
 ; AVX512DQ-NEXT:    vmovdqa64 %zmm1, 64(%rcx)
-; AVX512DQ-NEXT:    vmovdqa64 %zmm6, 128(%rcx)
-; AVX512DQ-NEXT:    vmovdqa64 %zmm5, 192(%rcx)
-; AVX512DQ-NEXT:    vmovdqa64 %zmm17, 320(%rcx)
-; AVX512DQ-NEXT:    vmovdqa64 %zmm10, 256(%rcx)
-; AVX512DQ-NEXT:    vmovdqa64 %zmm3, (%rcx)
+; AVX512DQ-NEXT:    vmovdqa64 %zmm9, 128(%rcx)
+; AVX512DQ-NEXT:    vmovdqa64 %zmm6, 192(%rcx)
+; AVX512DQ-NEXT:    vmovdqa64 %zmm18, 320(%rcx)
+; AVX512DQ-NEXT:    vmovdqa64 %zmm0, 256(%rcx)
+; AVX512DQ-NEXT:    vmovdqa64 %zmm19, (%rcx)
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
 ; AVX512DQ-FCP-LABEL: store_i16_stride3_vf64:
 ; AVX512DQ-FCP:       # %bb.0:
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa64 16(%rsi), %xmm20
-; AVX512DQ-FCP-NEXT:    vmovdqa64 32(%rsi), %xmm24
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm5, %xmm8
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm9
-; AVX512DQ-FCP-NEXT:    vmovdqa64 16(%rdi), %xmm21
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm4
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm10 = xmm9[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm10[0,1],xmm8[2],xmm10[3,4],xmm8[5],xmm10[6,7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm9, %xmm9
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm8, %ymm9, %ymm8
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm10 = zmm8[0,1,2,3],zmm3[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm8
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %ymm14
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm3, %ymm11
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm16 = [0,0,0,0,1,1,0,2]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm3, %ymm16, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpandn %ymm3, %ymm15, %ymm3
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm10 & zmm15)
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %xmm10
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm10, %xmm11
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm12
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm12[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm11 = xmm13[0,1],xmm11[2],xmm13[3,4],xmm11[5],xmm13[6,7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm10
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm11, %ymm10, %ymm10
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm0, %zmm10
-; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rdi), %xmm12
-; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rsi), %xmm13
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm13, %xmm13
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm12 = xmm12[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm13[2],xmm12[3,4],xmm13[5],xmm12[6,7]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm0, %ymm12, %ymm0
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm10[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm18 = [5,5,0,6,6,0,7,7,0,8,8,0,9,9,0,10]
-; AVX512DQ-FCP-NEXT:    vpermd 64(%rdx), %zmm18, %zmm10
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm22 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm10 = zmm10 ^ (zmm22 & (zmm10 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm6, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %ymm5
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm7, %ymm13
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm5, %ymm5
-; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm5, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rdi), %xmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rsi), %xmm12
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm12[4],xmm5[4],xmm12[5],xmm5[5],xmm12[6],xmm5[6],xmm12[7],xmm5[7]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm7, %xmm7
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm12, %xmm12
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm12[2],xmm5[3,4],xmm12[5],xmm5[6,7]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm7, %ymm5, %ymm5
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm5
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm5[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %ymm5
-; AVX512DQ-FCP-NEXT:    vpermd %ymm5, %ymm18, %ymm7
-; AVX512DQ-FCP-NEXT:    vpandnq %ymm7, %ymm22, %ymm7
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm5, %ymm5
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm5, %zmm17
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm19 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm17 = zmm17 | (zmm0 & zmm19)
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %ymm7
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm7, %ymm7
-; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm7, %ymm0
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %xmm7
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm7, %xmm12
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm5
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm12[2],xmm6[3,4],xmm12[5],xmm6[6,7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm5[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm14, %ymm5
-; AVX512DQ-FCP-NEXT:    vpermd %ymm14, %ymm16, %ymm6
-; AVX512DQ-FCP-NEXT:    vpandn %ymm6, %ymm15, %ymm6
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm0 & zmm15)
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm6
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm6, %ymm6
-; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm6, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rdi), %xmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rsi), %xmm7
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm12, %xmm12
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm7, %xmm7
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm7[2],xmm6[3,4],xmm7[5],xmm6[6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdx), %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdx), %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm5 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT:    vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm6
+; AVX512DQ-FCP-NEXT:    vmovdqa64 16(%rsi), %xmm21
+; AVX512DQ-FCP-NEXT:    vmovdqa64 32(%rsi), %xmm25
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm6, %xmm8
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm11
+; AVX512DQ-FCP-NEXT:    vmovdqa64 16(%rdi), %xmm22
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm7
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm12 = xmm11[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm8[2],xmm12[3,4],xmm8[5],xmm12[6,7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
 ; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm12, %ymm6, %ymm6
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm12 = zmm6[0,1,2,3],zmm4[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm6
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %ymm15
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm11 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm6, %ymm13
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [0,0,0,0,1,1,0,2]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm6, %ymm17, %ymm6
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
+; AVX512DQ-FCP-NEXT:    vpandnq %ymm6, %ymm16, %ymm6
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm6, %zmm19
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm19 = zmm19 | (zmm12 & zmm16)
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %xmm12
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm12, %xmm13
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm14
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm14[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm13[2],xmm8[3,4],xmm13[5],xmm8[6,7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm12, %xmm12
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm8, %ymm12, %ymm8
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm8
+; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rdi), %xmm13
+; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rsi), %xmm14
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm12 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm14, %xmm14
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1],xmm14[2],xmm13[3,4],xmm14[5],xmm13[6,7]
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm13, %ymm6
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm8[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %ymm8
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm20 = [5,5,0,6,6,0,7,7,0,16,16,0,17,17,0,18]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm8, %zmm20, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm23 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm23 & (zmm0 ^ zmm6))
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %ymm6
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %ymm9
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm10, %ymm14
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT:    vpor %ymm6, %ymm9, %ymm6
+; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rdi), %xmm9
+; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rsi), %xmm13
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm13[4],xmm9[4],xmm13[5],xmm9[5],xmm13[6],xmm9[6],xmm13[7],xmm9[7]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm10, %xmm10
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm13, %xmm13
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm13[2],xmm9[3,4],xmm13[5],xmm9[6,7]
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm10, %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm9[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm8, %ymm20, %ymm9
+; AVX512DQ-FCP-NEXT:    vpandnq %ymm9, %ymm23, %ymm9
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm8, %ymm8
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm8, %zmm18
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm8 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm18 = zmm18 | (zmm6 & zmm8)
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %ymm6
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %ymm9
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT:    vpor %ymm6, %ymm9, %ymm6
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm6
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm6[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm8, %ymm18, %ymm6
-; AVX512DQ-FCP-NEXT:    vpandnq %ymm6, %ymm22, %ymm6
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm8, %ymm7
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm0 & zmm19)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm24, %xmm2
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm24, %xmm0
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm7 = xmm4[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm0 = xmm7[0,1],xmm0[2],xmm7[3,4],xmm0[5],xmm7[6,7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm20, %xmm4
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm21, %xmm1
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm20, %xmm4
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm1 = xmm21[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm4[2],xmm1[3,4],xmm4[5],xmm1[6,7]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermd (%rdx), %zmm18, %zmm1
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm22 & (zmm1 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %xmm9
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm9, %xmm10
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm13
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm2 = xmm13[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm10[2],xmm2[3,4],xmm10[5],xmm2[6,7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm13[0],xmm9[0],xmm13[1],xmm9[1],xmm13[2],xmm9[2],xmm13[3],xmm9[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm9, %ymm2
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm6[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm15, %ymm6
+; AVX512DQ-FCP-NEXT:    vpermd %ymm15, %ymm17, %ymm9
+; AVX512DQ-FCP-NEXT:    vpandnq %ymm9, %ymm16, %ymm9
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm9, %zmm6
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm2 & zmm16)
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm9
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT:    vpor %ymm2, %ymm9, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rdi), %xmm9
+; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rsi), %xmm10
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm13, %xmm13
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm10, %xmm10
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm10[2],xmm9[3,4],xmm10[5],xmm9[6,7]
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm13, %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm9[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm4, %ymm20, %ymm9
+; AVX512DQ-FCP-NEXT:    vpandnq %ymm9, %ymm23, %ymm9
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm4, %ymm10
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm9 | (zmm2 & zmm8)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm25, %xmm5
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm25, %xmm2
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm7[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm8[0,1],xmm2[2],xmm8[3,4],xmm2[5],xmm8[6,7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm5, %ymm2
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm21, %xmm7
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm22, %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm5, %xmm5
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm21, %xmm7
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm3 = xmm22[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm7[2],xmm3[3,4],xmm7[5],xmm3[6,7]
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm5, %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm4, %zmm20, %zmm1
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm23 & (zmm1 ^ zmm2))
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rcx)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 128(%rcx)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 192(%rcx)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, 320(%rcx)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, 256(%rcx)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, (%rcx)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 128(%rcx)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 192(%rcx)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, 320(%rcx)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 256(%rcx)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, (%rcx)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-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 187a8102095ed..cec376e729205 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
@@ -1669,78 +1669,78 @@ define void @store_i16_stride4_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP:       # %bb.0:
 ; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm0
 ; AVX512-FCP-NEXT:    vmovdqa 16(%rsi), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm9
-; AVX512-FCP-NEXT:    vmovdqa 48(%rsi), %xmm5
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm10
+; AVX512-FCP-NEXT:    vmovdqa 48(%rsi), %xmm6
 ; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm2
 ; AVX512-FCP-NEXT:    vmovdqa 16(%rdi), %xmm3
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm10
-; AVX512-FCP-NEXT:    vmovdqa 48(%rdi), %xmm6
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm11
+; AVX512-FCP-NEXT:    vmovdqa 48(%rdi), %xmm7
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm7 = xmm4[0],zero,xmm4[1],zero
+; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm4, %ymm7, %ymm4
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm4, %ymm5, %ymm4
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
 ; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm3 = xmm1[0],zero,xmm1[1],zero
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
 ; AVX512-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm3, %ymm1
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm1, %zmm1
 ; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm3
-; AVX512-FCP-NEXT:    vmovdqa 16(%rcx), %xmm11
+; AVX512-FCP-NEXT:    vmovdqa 16(%rcx), %xmm5
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm12
-; AVX512-FCP-NEXT:    vmovdqa 48(%rcx), %xmm7
+; AVX512-FCP-NEXT:    vmovdqa 48(%rcx), %xmm8
 ; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm4
 ; AVX512-FCP-NEXT:    vmovdqa 16(%rdx), %xmm13
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm14
-; AVX512-FCP-NEXT:    vmovdqa 48(%rdx), %xmm8
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm11, %zmm11
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT:    vmovdqa 48(%rdx), %xmm9
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm5[4],xmm13[5],xmm5[5],xmm13[6],xmm5[6],xmm13[7],xmm5[7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,1,1,2,2,3,3,16,16,17,17,18,18,19,19]
+; AVX512-FCP-NEXT:    vpermt2d %zmm15, %zmm5, %zmm13
 ; AVX512-FCP-NEXT:    movw $-21846, %ax # imm = 0xAAAA
 ; AVX512-FCP-NEXT:    kmovw %eax, %k1
-; AVX512-FCP-NEXT:    vpermd %zmm11, %zmm13, %zmm1 {%k1}
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
-; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm15 = xmm11[0],zero,xmm11[1],zero
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm11, %ymm15, %ymm11
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm9, %ymm10, %ymm9
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm9, %zmm9
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm10, %zmm11, %zmm10
-; AVX512-FCP-NEXT:    vpermd %zmm10, %zmm13, %zmm9 {%k1}
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm13, %zmm1 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm15 = xmm13[0],zero,xmm13[1],zero
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm13, %ymm15, %ymm13
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
 ; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm11 = xmm10[0],zero,xmm10[1],zero
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
 ; AVX512-FCP-NEXT:    vinserti128 $1, %xmm10, %ymm11, %ymm10
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm5[0],zero,xmm5[1],zero
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm5, %zmm5
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm7, %zmm6
-; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm13, %zmm5 {%k1}
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm10, %zmm10
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512-FCP-NEXT:    vpermt2d %zmm11, %zmm5, %zmm12
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm12, %zmm10 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
+; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm11, %ymm12, %ymm11
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
 ; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm7 = xmm6[0],zero,xmm6[1],zero
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
 ; AVX512-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm7, %ymm6
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm6, %zmm6
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512-FCP-NEXT:    vpermt2d %zmm7, %zmm5, %zmm8
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm8, %zmm6 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm8 = xmm7[0],zero,xmm7[1],zero
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm7, %ymm8, %ymm7
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
 ; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
 ; AVX512-FCP-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm0
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[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:    vinserti32x4 $2, %xmm2, %zmm3, %zmm2
-; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm13, %zmm0 {%k1}
+; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm3
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm3, %zmm0 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%r8)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 192(%r8)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, 128(%r8)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 192(%r8)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 128(%r8)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 64(%r8)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
@@ -1829,78 +1829,78 @@ define void @store_i16_stride4_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP:       # %bb.0:
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa 16(%rsi), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm9
-; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rsi), %xmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm10
+; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rsi), %xmm6
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm2
 ; AVX512DQ-FCP-NEXT:    vmovdqa 16(%rdi), %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm10
-; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rdi), %xmm6
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm11
+; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rdi), %xmm7
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm7 = xmm4[0],zero,xmm4[1],zero
+; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm4, %ymm7, %ymm4
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm4, %ymm5, %ymm4
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
 ; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm3 = xmm1[0],zero,xmm1[1],zero
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
 ; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm3, %ymm1
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm1, %zmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 16(%rcx), %xmm11
+; AVX512DQ-FCP-NEXT:    vmovdqa 16(%rcx), %xmm5
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm12
-; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rcx), %xmm7
+; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rcx), %xmm8
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm4
 ; AVX512DQ-FCP-NEXT:    vmovdqa 16(%rdx), %xmm13
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm14
-; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rdx), %xmm8
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm11, %zmm11
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rdx), %xmm9
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm5[4],xmm13[5],xmm5[5],xmm13[6],xmm5[6],xmm13[7],xmm5[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,1,1,2,2,3,3,16,16,17,17,18,18,19,19]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm15, %zmm5, %zmm13
 ; AVX512DQ-FCP-NEXT:    movw $-21846, %ax # imm = 0xAAAA
 ; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT:    vpermd %zmm11, %zmm13, %zmm1 {%k1}
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
-; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm15 = xmm11[0],zero,xmm11[1],zero
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm11, %ymm15, %ymm11
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm9, %ymm10, %ymm9
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm9, %zmm9
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm10, %zmm11, %zmm10
-; AVX512DQ-FCP-NEXT:    vpermd %zmm10, %zmm13, %zmm9 {%k1}
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm13, %zmm1 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm15 = xmm13[0],zero,xmm13[1],zero
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm13, %ymm15, %ymm13
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
 ; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm11 = xmm10[0],zero,xmm10[1],zero
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
 ; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm10, %ymm11, %ymm10
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm5[0],zero,xmm5[1],zero
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm5, %zmm5
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm7, %zmm6
-; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm13, %zmm5 {%k1}
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm10, %zmm10
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm11, %zmm5, %zmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm12, %zmm10 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
+; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm11, %ymm12, %ymm11
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
 ; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm7 = xmm6[0],zero,xmm6[1],zero
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
 ; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm6, %ymm7, %ymm6
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm6, %zmm6
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm7, %zmm5, %zmm8
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm8, %zmm6 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm8 = xmm7[0],zero,xmm7[1],zero
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm7, %ymm8, %ymm7
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
 ; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
 ; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm0
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[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:    vinserti32x4 $2, %xmm2, %zmm3, %zmm2
-; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm13, %zmm0 {%k1}
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm3, %zmm0 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%r8)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 192(%r8)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 128(%r8)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 192(%r8)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, 128(%r8)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 64(%r8)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
@@ -3248,12 +3248,12 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm14
 ; AVX512-FCP-NEXT:    vmovdqa 48(%rdx), %xmm10
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm8[4],xmm5[4],xmm8[5],xmm5[5],xmm8[6],xmm5[6],xmm8[7],xmm5[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm5, %zmm8
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,1,1,2,2,3,3,16,16,17,17,18,18,19,19]
+; AVX512-FCP-NEXT:    vpermt2d %zmm15, %zmm5, %zmm8
 ; AVX512-FCP-NEXT:    movw $-21846, %ax # imm = 0xAAAA
 ; AVX512-FCP-NEXT:    kmovw %eax, %k1
-; AVX512-FCP-NEXT:    vpermd %zmm8, %zmm5, %zmm18 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm8, %zmm18 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %xmm8
 ; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm15
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm8[4],xmm15[5],xmm8[5],xmm15[6],xmm8[6],xmm15[7],xmm8[7]
@@ -3269,8 +3269,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %xmm15
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm5, %zmm8 {%k1}
+; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm5, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm8 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa 112(%rsi), %xmm0
 ; AVX512-FCP-NEXT:    vmovdqa 112(%rdi), %xmm1
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
@@ -3286,8 +3286,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vmovdqa 112(%rdx), %xmm1
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm5, %zmm15 {%k1}
+; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm15 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %xmm0
 ; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm1
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
@@ -3303,8 +3303,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %xmm1
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm5, %zmm16 {%k1}
+; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm16 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa 80(%rsi), %xmm0
 ; AVX512-FCP-NEXT:    vmovdqa 80(%rdi), %xmm1
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
@@ -3320,8 +3320,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vmovdqa 80(%rdx), %xmm1
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm5, %zmm17 {%k1}
+; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm17 {%k1}
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
 ; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
@@ -3333,8 +3333,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm2, %zmm1
-; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm5, %zmm0 {%k1}
+; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm5, %zmm2
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm2, %zmm0 {%k1}
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
 ; AVX512-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
@@ -3346,8 +3346,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm6, %zmm2
-; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm5, %zmm1 {%k1}
+; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm6
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm6, %zmm1 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa64 %xmm19, %xmm7
 ; AVX512-FCP-NEXT:    vmovdqa64 %xmm20, %xmm9
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
@@ -3361,8 +3361,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm6, %zmm2
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[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:    vinserti32x4 $2, %xmm6, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vpermd %zmm3, %zmm5, %zmm2 {%k1}
+; AVX512-FCP-NEXT:    vpermt2d %zmm6, %zmm5, %zmm3
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm3, %zmm2 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, (%r8)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 192(%r8)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 128(%r8)
@@ -3556,12 +3556,12 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm14
 ; AVX512DQ-FCP-NEXT:    vmovdqa 48(%rdx), %xmm10
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm8[4],xmm5[4],xmm8[5],xmm5[5],xmm8[6],xmm5[6],xmm8[7],xmm5[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm5, %zmm8
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,1,1,2,2,3,3,16,16,17,17,18,18,19,19]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm15, %zmm5, %zmm8
 ; AVX512DQ-FCP-NEXT:    movw $-21846, %ax # imm = 0xAAAA
 ; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT:    vpermd %zmm8, %zmm5, %zmm18 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm8, %zmm18 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %xmm8
 ; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm15
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm8[4],xmm15[5],xmm8[5],xmm15[6],xmm8[6],xmm15[7],xmm8[7]
@@ -3577,8 +3577,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %xmm15
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm5, %zmm8 {%k1}
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm0, %zmm8 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rsi), %xmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rdi), %xmm1
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
@@ -3594,8 +3594,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rdx), %xmm1
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm5, %zmm15 {%k1}
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm0, %zmm15 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %xmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm1
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
@@ -3611,8 +3611,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %xmm1
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm5, %zmm16 {%k1}
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm0, %zmm16 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rsi), %xmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rdi), %xmm1
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
@@ -3628,8 +3628,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rdx), %xmm1
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm5, %zmm17 {%k1}
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm0, %zmm17 {%k1}
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
 ; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
@@ -3641,8 +3641,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm2, %zmm1
-; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm5, %zmm0 {%k1}
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm5, %zmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm2, %zmm0 {%k1}
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
 ; AVX512DQ-FCP-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
@@ -3654,8 +3654,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm6, %zmm2
-; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm5, %zmm1 {%k1}
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm2, %zmm5, %zmm6
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm6, %zmm1 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm19, %xmm7
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm20, %xmm9
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
@@ -3669,8 +3669,8 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm6, %zmm2
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[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:    vinserti32x4 $2, %xmm6, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vpermd %zmm3, %zmm5, %zmm2 {%k1}
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm6, %zmm5, %zmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm3, %zmm2 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, (%r8)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 192(%r8)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 128(%r8)
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
index 1073a35353a63..244408c456dc4 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
@@ -231,12 +231,11 @@ define void @store_i16_stride6_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
 ; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm1 = [0,2,4,6,8,10,1,3,5,7,9,11,0,0,0,0]
-; AVX512BW-NEXT:    vpermw %ymm0, %ymm1, %ymm0
-; AVX512BW-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX512BW-NEXT:    vmovq %xmm1, 16(%rax)
-; AVX512BW-NEXT:    vmovdqa %xmm0, (%rax)
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,1,3,5,7,17,19,0,0,0,0]
+; AVX512BW-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512BW-NEXT:    vextracti128 $1, %ymm2, %xmm0
+; AVX512BW-NEXT:    vmovq %xmm0, 16(%rax)
+; AVX512BW-NEXT:    vmovdqa %xmm2, (%rax)
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
 ;
@@ -250,12 +249,11 @@ define void @store_i16_stride6_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
 ; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm1 = [0,2,4,6,8,10,1,3,5,7,9,11,0,0,0,0]
-; AVX512BW-FCP-NEXT:    vpermw %ymm0, %ymm1, %ymm0
-; AVX512BW-FCP-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX512BW-FCP-NEXT:    vmovq %xmm1, 16(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa %xmm0, (%rax)
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,1,3,5,7,17,19,0,0,0,0]
+; AVX512BW-FCP-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512BW-FCP-NEXT:    vextracti128 $1, %ymm2, %xmm0
+; AVX512BW-FCP-NEXT:    vmovq %xmm0, 16(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa %xmm2, (%rax)
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
 ;
@@ -269,12 +267,11 @@ define void @store_i16_stride6_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
 ; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm1 = [0,2,4,6,8,10,1,3,5,7,9,11,0,0,0,0]
-; AVX512DQ-BW-NEXT:    vpermw %ymm0, %ymm1, %ymm0
-; AVX512DQ-BW-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX512DQ-BW-NEXT:    vmovq %xmm1, 16(%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa %xmm0, (%rax)
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,1,3,5,7,17,19,0,0,0,0]
+; AVX512DQ-BW-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512DQ-BW-NEXT:    vextracti128 $1, %ymm2, %xmm0
+; AVX512DQ-BW-NEXT:    vmovq %xmm0, 16(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa %xmm2, (%rax)
 ; AVX512DQ-BW-NEXT:    vzeroupper
 ; AVX512DQ-BW-NEXT:    retq
 ;
@@ -288,12 +285,11 @@ define void @store_i16_stride6_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
 ; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm1 = [0,2,4,6,8,10,1,3,5,7,9,11,0,0,0,0]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm0, %ymm1, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovq %xmm1, 16(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm0, (%rax)
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,1,3,5,7,17,19,0,0,0,0]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vextracti128 $1, %ymm2, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vmovq %xmm0, 16(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm2, (%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
   %in.vec0 = load <2 x i16>, ptr %in.vecptr0, align 64
@@ -532,13 +528,13 @@ define void @store_i16_stride6_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm4 = xmm3[0],xmm2[0]
 ; AVX512-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm5
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,8,9,0,1,4,5,u,u,u,u,2,3,10,11,2,3,10,11,u,u,u,u,24,25,28,29,4,5,12,13]
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm7
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [4,6,1,3,4,6,1,3]
-; AVX512-FCP-NEXT:    # ymm8 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpermd %ymm5, %ymm8, %ymm5
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm5
-; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm7[0],ymm5[1],ymm7[2,3,4,5],ymm5[6],ymm7[7]
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [8,10,1,3,8,10,1,3]
+; AVX512-FCP-NEXT:    # ymm6 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm6
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,8,9,0,1,4,5,u,u,u,u,2,3,10,11,2,3,10,11,u,u,u,u,24,25,28,29,4,5,12,13]
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm6, %ymm6
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm5, %ymm5
+; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4,5],ymm6[6],ymm5[7]
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
 ; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [0,1,8,3,4,9,6,7]
 ; AVX512-FCP-NEXT:    vpermi2d %ymm2, %ymm5, %ymm3
@@ -601,13 +597,13 @@ define void @store_i16_stride6_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm4 = xmm3[0],xmm2[0]
 ; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm5
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,8,9,0,1,4,5,u,u,u,u,2,3,10,11,2,3,10,11,u,u,u,u,24,25,28,29,4,5,12,13]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm7
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [4,6,1,3,4,6,1,3]
-; AVX512DQ-FCP-NEXT:    # ymm8 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm5, %ymm8, %ymm5
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm5
-; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm7[0],ymm5[1],ymm7[2,3,4,5],ymm5[6],ymm7[7]
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [8,10,1,3,8,10,1,3]
+; AVX512DQ-FCP-NEXT:    # ymm6 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm6
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,8,9,0,1,4,5,u,u,u,u,2,3,10,11,2,3,10,11,u,u,u,u,24,25,28,29,4,5,12,13]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm6, %ymm6
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm5, %ymm5
+; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4,5],ymm6[6],ymm5[7]
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
 ; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [0,1,8,3,4,9,6,7]
 ; AVX512DQ-FCP-NEXT:    vpermi2d %ymm2, %ymm5, %ymm3
@@ -4138,210 +4134,222 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ;
 ; AVX512-FCP-LABEL: store_i16_stride6_vf32:
 ; AVX512-FCP:       # %bb.0:
-; AVX512-FCP-NEXT:    pushq %rax
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm2
+; AVX512-FCP-NEXT:    subq $40, %rsp
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm3
 ; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm0 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm2, %ymm1
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm3, %ymm1
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm4
-; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm4, %ymm3
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm3[0],ymm1[0],ymm3[1],ymm1[1],ymm3[2],ymm1[2],ymm3[3],ymm1[3],ymm3[8],ymm1[8],ymm3[9],ymm1[9],ymm3[10],ymm1[10],ymm3[11],ymm1[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm4[4],ymm2[4],ymm4[5],ymm2[5],ymm4[6],ymm2[6],ymm4[7],ymm2[7],ymm4[12],ymm2[12],ymm4[13],ymm2[13],ymm4[14],ymm2[14],ymm4[15],ymm2[15]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm4, %ymm23
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm24
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm6 = [2,1,2,3,11,11,11,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm6, %zmm1
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm11
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm13
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm13[4],ymm11[4],ymm13[5],ymm11[5],ymm13[6],ymm11[6],ymm13[7],ymm11[7],ymm13[12],ymm11[12],ymm13[13],ymm11[13],ymm13[14],ymm11[14],ymm13[15],ymm11[15]
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm10 = [5,6,5,6,5,6,7,7]
-; AVX512-FCP-NEXT:    vpermd %ymm3, %ymm10, %ymm3
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm5 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm4, %ymm2
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm23
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm4 = [2,1,2,3,11,11,11,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm4, %zmm2
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm6
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm7
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm5 = [5,6,5,6,5,6,7,7]
+; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm5, %ymm1
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm6, %ymm24
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm6 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm7, %ymm25
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm6[0],ymm3[0],ymm6[1],ymm3[1],ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[8],ymm3[8],ymm6[9],ymm3[9],ymm6[10],ymm3[10],ymm6[11],ymm3[11]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm3, %zmm1
 ; AVX512-FCP-NEXT:    movw $18724, %ax # imm = 0x4924
 ; AVX512-FCP-NEXT:    kmovw %eax, %k1
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm3, %zmm1 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm2
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm12 = [12,1,2,13,4,5,14,7]
-; AVX512-FCP-NEXT:    vmovdqa %ymm1, %ymm5
-; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm12, %ymm5
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [8,21,10,11,20,13,14,23]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
-; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm25
-; AVX512-FCP-NEXT:    vpermt2d %zmm3, %zmm17, %zmm1
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm5, %zmm16
-; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %ymm3
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm14 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
-; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm3, %ymm1
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm15 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512-FCP-NEXT:    # ymm15 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm3, %ymm2
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm26
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [2,2,0,3,10,0,10,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm18, %zmm2
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm3
-; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm3, %ymm1
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm2
-; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm2, %ymm0
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm27
-; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm0
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm1
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm6
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm6[4],ymm1[4],ymm6[5],ymm1[5],ymm6[6],ymm1[6],ymm6[7],ymm1[7],ymm6[12],ymm1[12],ymm6[13],ymm1[13],ymm6[14],ymm1[14],ymm6[15],ymm1[15]
-; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm10, %ymm2
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm10 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm10[0],ymm4[1],ymm10[1],ymm4[2],ymm10[2],ymm4[3],ymm10[3],ymm4[8],ymm10[8],ymm4[9],ymm10[9],ymm4[10],ymm10[10],ymm4[11],ymm10[11]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm2
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm2, %zmm0 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm10
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermi2d %ymm2, %ymm0, %ymm12
-; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm10, %ymm2
-; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm17, %zmm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm12, %zmm17
-; AVX512-FCP-NEXT:    vmovdqa (%r9), %ymm12
-; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm12, %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm12, %ymm2
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm18, %zmm2
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm1, %zmm2 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm9
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm6 = [12,1,2,13,4,5,14,7]
+; AVX512-FCP-NEXT:    vmovdqa %ymm2, %ymm3
+; AVX512-FCP-NEXT:    vpermt2d %ymm1, %ymm6, %ymm3
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm7 = [8,21,10,11,20,13,14,23]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
+; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm9, %ymm8
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm9, %ymm27
+; AVX512-FCP-NEXT:    vpermt2d %zmm8, %zmm7, %zmm2
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm14
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm15
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm6[0],ymm1[0],ymm6[1],ymm1[1],ymm6[2],ymm1[2],ymm6[3],ymm1[3],ymm6[8],ymm1[8],ymm6[9],ymm1[9],ymm6[10],ymm1[10],ymm6[11],ymm1[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm18
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm6
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm13[0],ymm11[0],ymm13[1],ymm11[1],ymm13[2],ymm11[2],ymm13[3],ymm11[3],ymm13[8],ymm11[8],ymm13[9],ymm11[9],ymm13[10],ymm11[10],ymm13[11],ymm11[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm20
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm13
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm13, %xmm4
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm0
-; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm0, %xmm5
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm22 = [0,0,2,1,8,9,8,9]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
-; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm22, %zmm11
+; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %ymm9
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm2 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
+; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm9, %ymm8
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm9, %ymm10
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm9, %ymm28
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm9 = [2,2,0,3,10,0,10,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm9, %zmm10
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm15
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm15, %ymm8
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm11
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm0
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[1],ymm8[1],ymm0[2],ymm8[2],ymm0[3],ymm8[3],ymm0[8],ymm8[8],ymm0[9],ymm8[9],ymm0[10],ymm8[10],ymm0[11],ymm8[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm8 = ymm11[4],ymm15[4],ymm11[5],ymm15[5],ymm11[6],ymm15[6],ymm11[7],ymm15[7],ymm11[12],ymm15[12],ymm11[13],ymm15[13],ymm11[14],ymm15[14],ymm11[15],ymm15[15]
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm11, %ymm21
+; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm4, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm8
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm14
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm14[4],ymm8[4],ymm14[5],ymm8[5],ymm14[6],ymm8[6],ymm14[7],ymm8[7],ymm14[12],ymm8[12],ymm14[13],ymm8[13],ymm14[14],ymm8[14],ymm14[15],ymm8[15]
+; AVX512-FCP-NEXT:    vpermd %ymm4, %ymm5, %ymm4
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm5 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm8, %ymm31
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm8[0],ymm5[0],ymm8[1],ymm5[1],ymm8[2],ymm5[2],ymm8[3],ymm5[3],ymm8[8],ymm5[8],ymm8[9],ymm5[9],ymm8[10],ymm5[10],ymm8[11],ymm5[11]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm4, %zmm0 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm5
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm4 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermi2d %ymm4, %ymm0, %ymm6
+; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm5, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm5, %ymm29
+; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm7, %zmm0
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm6, %zmm22
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %ymm1
+; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm1, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm1, %ymm11
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm30
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm9, %zmm11
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512-FCP-NEXT:    vpshufb %xmm0, %xmm4, %xmm1
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm5
+; AVX512-FCP-NEXT:    vpshufb %xmm0, %xmm5, %xmm2
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm1 = [0,0,2,1,8,9,8,9]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm5, %xmm16
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm4, %xmm17
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm1, %zmm2
 ; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [1,0,2,2,1,0,2,2]
 ; AVX512-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
-; AVX512-FCP-NEXT:    vpermd %ymm8, %ymm4, %ymm8
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
-; AVX512-FCP-NEXT:    vpbroadcastq %xmm1, %ymm1
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm8, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm13
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm12
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
+; AVX512-FCP-NEXT:    vpermd %ymm3, %ymm4, %ymm3
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512-FCP-NEXT:    vpbroadcastq %xmm5, %ymm5
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm3, %zmm3
 ; AVX512-FCP-NEXT:    movw $9362, %ax # imm = 0x2492
 ; AVX512-FCP-NEXT:    kmovw %eax, %k2
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm1, %zmm11 {%k2}
-; AVX512-FCP-NEXT:    vmovdqa %ymm11, %ymm6
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm8 = [16,9,10,17,12,13,18,15]
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm1
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm1[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2d %zmm9, %zmm8, %zmm11
-; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm9 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm21 = [0,1,8,3,4,9,6,7]
-; AVX512-FCP-NEXT:    vpermt2d %ymm9, %ymm21, %ymm6
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm6, %zmm19
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm3, %zmm2 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa %ymm2, %ymm3
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm20 = [16,9,10,17,12,13,18,15]
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm10
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm10[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2d %zmm6, %zmm20, %zmm2
+; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm19 = [0,1,8,3,4,9,6,7]
+; AVX512-FCP-NEXT:    vpermt2d %ymm6, %ymm19, %ymm3
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm18
 ; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm9
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm11
-; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm9, %xmm6
-; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
-; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm22, %zmm5
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
-; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm4, %ymm2
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm7
+; AVX512-FCP-NEXT:    vpshufb %xmm0, %xmm9, %xmm2
+; AVX512-FCP-NEXT:    vpshufb %xmm0, %xmm7, %xmm0
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm1, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm5
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
+; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm4, %ymm1
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm8[0],xmm4[0],xmm8[1],xmm4[1],xmm8[2],xmm4[2],xmm8[3],xmm4[3]
 ; AVX512-FCP-NEXT:    vpbroadcastq %xmm4, %ymm4
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm2, %zmm5 {%k2}
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm1, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k2}
 ; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm2
 ; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm4 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
-; AVX512-FCP-NEXT:    vpermi2d %ymm4, %ymm5, %ymm21
+; AVX512-FCP-NEXT:    vpermi2d %ymm4, %ymm0, %ymm19
 ; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm2[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2d %zmm4, %zmm8, %zmm5
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm4 = [0,0,0,0,8,8,0,9]
-; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm6
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm8 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm6, %xmm15
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm6[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm15, %zmm4, %zmm14
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm21, %zmm21
-; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm5
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm8
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm15 = xmm5[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm4, %zmm15
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm27, %ymm4
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm8 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm9 = [1,1,1,1,10,10,10,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm9, %zmm8
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
-; AVX512-FCP-NEXT:    vpermd %zmm18, %zmm4, %zmm11
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm8, %zmm11 {%k1}
+; AVX512-FCP-NEXT:    vpermt2d %zmm4, %zmm20, %zmm0
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [0,0,0,0,8,8,0,9]
+; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm8
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm4 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm8, %xmm1
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm8[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm20, %zmm6
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm19, %zmm19
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm0
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm0, %xmm1
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm0[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm20, %zmm4
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm21, %ymm1
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[2],ymm15[2],ymm1[3],ymm15[3],ymm1[8],ymm15[8],ymm1[9],ymm15[9],ymm1[10],ymm15[10],ymm1[11],ymm15[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm7[4],xmm9[4],xmm7[5],xmm9[5],xmm7[6],xmm9[6],xmm7[7],xmm9[7]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm21 = [1,1,1,1,10,10,10,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm21, %zmm7
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm31, %ymm1
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm14[0],ymm1[0],ymm14[1],ymm1[1],ymm14[2],ymm1[2],ymm14[3],ymm1[3],ymm14[8],ymm1[8],ymm14[9],ymm1[9],ymm14[10],ymm1[10],ymm14[11],ymm1[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm26 = [1,2,1,2,0,0,3,3,21,20,18,18,21,20,22,22]
+; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm26, %zmm5
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm7, %zmm5 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm18 = [0,9,2,3,8,5,6,11]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm11, %ymm22
-; AVX512-FCP-NEXT:    vpermt2d %ymm2, %ymm18, %ymm22
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [8,9,20,11,12,21,14,15]
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2d %zmm8, %zmm2, %zmm11
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm12 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm5, %xmm8
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm5 = [0,0,0,1,0,10,10,0]
-; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm5, %zmm8
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm23, %ymm10
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm1
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [0,9,2,3,8,5,6,11]
+; AVX512-FCP-NEXT:    vmovdqa %ymm5, %ymm7
+; AVX512-FCP-NEXT:    vpermt2d %ymm1, %ymm2, %ymm7
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm15 = [8,9,20,11,12,21,14,15]
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm29, %ymm1
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm15, %zmm5
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm7, %zmm5
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm30, %ymm1
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [0,0,0,1,0,10,10,0]
+; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm20, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm23, %ymm9
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm7[0],ymm9[0],ymm7[1],ymm9[1],ymm7[2],ymm9[2],ymm7[3],ymm9[3],ymm7[8],ymm9[8],ymm7[9],ymm9[9],ymm7[10],ymm9[10],ymm7[11],ymm9[11]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm16, %xmm9
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm17, %xmm14
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm9[4],xmm14[4],xmm9[5],xmm14[5],xmm9[6],xmm14[6],xmm9[7],xmm14[7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm21, %zmm9
 ; AVX512-FCP-NEXT:    vmovdqa64 %ymm24, %ymm7
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm10[0],ymm7[0],ymm10[1],ymm7[1],ymm10[2],ymm7[2],ymm10[3],ymm7[3],ymm10[8],ymm7[8],ymm10[9],ymm7[9],ymm10[10],ymm7[10],ymm10[11],ymm7[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm9, %zmm0
-; AVX512-FCP-NEXT:    vpermd %zmm20, %zmm4, %zmm4
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm4 {%k1}
-; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm0
-; AVX512-FCP-NEXT:    vpermi2d %ymm0, %ymm4, %ymm18
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm25, %ymm0
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm4
-; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm6, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm1
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm5, %zmm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm18, %zmm1
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm2 & (zmm0 ^ zmm1))
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm25, %ymm14
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm14[0],ymm7[0],ymm14[1],ymm7[1],ymm14[2],ymm7[2],ymm14[3],ymm7[3],ymm14[8],ymm7[8],ymm14[9],ymm7[9],ymm14[10],ymm7[10],ymm14[11],ymm7[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
+; AVX512-FCP-NEXT:    vpermt2d %zmm7, %zmm26, %zmm12
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm9, %zmm12 {%k1}
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm10, %xmm3
+; AVX512-FCP-NEXT:    vpermi2d %ymm3, %ymm12, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm27, %ymm3
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermt2d %zmm3, %zmm15, %zmm12
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm2, %zmm2
+; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm8, %xmm1
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm28, %ymm3
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm20, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm3 & (zmm1 ^ zmm2))
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 256(%rax)
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm22, %zmm0
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm8 = zmm8 ^ (zmm2 & (zmm8 ^ zmm0))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, 64(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 256(%rax)
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm3 & (zmm0 ^ zmm5))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 64(%rax)
 ; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm15 = zmm15 ^ (zmm0 & (zmm15 ^ zmm21))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, (%rax)
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm14 = zmm14 ^ (zmm0 & (zmm14 ^ zmm19))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, 192(%rax)
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm4 = zmm4 ^ (zmm0 & (zmm4 ^ zmm19))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (zmm0 & (zmm6 ^ zmm18))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 192(%rax)
 ; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm0 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm11 = zmm11 ^ (zmm0 & (zmm11 ^ zmm22))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, 128(%rax)
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm0 & (zmm1 ^ zmm17))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 128(%rax)
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm0 & (zmm1 ^ zmm16))
+; AVX512-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 64-byte Folded Reload
+; AVX512-FCP-NEXT:    # zmm1 = zmm1 ^ (zmm0 & (zmm1 ^ mem))
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 320(%rax)
-; AVX512-FCP-NEXT:    popq %rax
+; AVX512-FCP-NEXT:    addq $40, %rsp
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -4578,218 +4586,220 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ;
 ; AVX512DQ-FCP-LABEL: store_i16_stride6_vf32:
 ; AVX512DQ-FCP:       # %bb.0:
+; AVX512DQ-FCP-NEXT:    pushq %rax
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm1
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm2
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm31
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm25
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [1,1,1,1,10,10,10,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm18, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm3
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm26
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm27
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm6
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm20 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm20, %zmm19
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm29
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm2
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm3, %xmm21
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm23
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm2 = [1,1,1,1,10,10,10,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm4
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm4, %ymm28
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm30
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm5, %xmm22
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm3, %xmm20
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm17 = [1,2,1,2,0,0,3,3,21,20,18,18,21,20,22,22]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm17, %zmm4
 ; AVX512DQ-FCP-NEXT:    movw $18724, %ax # imm = 0x4924
 ; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm1, %zmm19 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm1, %zmm4 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm0, %xmm1
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm19, %ymm23
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm1, %ymm17, %ymm23
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm21 = [8,9,20,11,12,21,14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm9
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm4
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm4[0],ymm9[0],ymm4[1],ymm9[1],ymm4[2],ymm9[2],ymm4[3],ymm9[3],ymm4[8],ymm9[8],ymm4[9],ymm9[9],ymm4[10],ymm9[10],ymm4[11],ymm9[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm6, %xmm30
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm5, %xmm29
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm7
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm7[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm21, %zmm19
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %ymm8
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm6 = [0,0,0,1,0,10,10,0]
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm8[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm0 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm0, %xmm2, %xmm15
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm28
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm15
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm11
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm10
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm10[4],xmm11[4],xmm10[5],xmm11[5],xmm10[6],xmm11[6],xmm10[7],xmm11[7]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm12, %zmm18, %zmm13
-; AVX512DQ-FCP-NEXT:    vpermd %zmm5, %zmm20, %zmm20
-; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm13, %zmm20 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm12
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm12, %xmm3
-; AVX512DQ-FCP-NEXT:    vpermi2d %ymm3, %ymm20, %ymm17
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm24
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm7 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm4, %ymm1
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm0, %ymm7, %ymm1
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm19 = [8,9,20,11,12,21,14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm12
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm19, %zmm4
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm1, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %ymm15
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm15[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm4, %xmm25
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [0,0,0,1,0,10,10,0]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm18, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm14
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm13
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm4
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm13[4],xmm14[4],xmm13[5],xmm14[5],xmm13[6],xmm14[6],xmm13[7],xmm14[7]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm6
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm8
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm10
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm2
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm9, %zmm17, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm6, %zmm0 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm6
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm6, %xmm31
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm3, %ymm0, %ymm7
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm3
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm5 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm5, %zmm21, %zmm20
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %xmm5
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm0, %xmm5, %xmm13
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm5, %xmm24
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %ymm5
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm6, %zmm13
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm1, %ymm14
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm14[0],ymm0[0],ymm14[1],ymm0[1],ymm14[2],ymm0[2],ymm14[3],ymm0[3],ymm14[8],ymm0[8],ymm14[9],ymm0[9],ymm14[10],ymm0[10],ymm14[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm2[4],ymm1[5],ymm2[5],ymm1[6],ymm2[6],ymm1[7],ymm2[7],ymm1[12],ymm2[12],ymm1[13],ymm2[13],ymm1[14],ymm2[14],ymm1[15],ymm2[15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm2 = [2,1,2,3,11,11,11,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm21 = [5,6,5,6,5,6,7,7]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm9[4],ymm4[5],ymm9[5],ymm4[6],ymm9[6],ymm4[7],ymm9[7],ymm4[12],ymm9[12],ymm4[13],ymm9[13],ymm4[14],ymm9[14],ymm4[15],ymm9[15]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm1, %ymm21, %ymm1
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm9 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm9[0],ymm4[1],ymm9[1],ymm4[2],ymm9[2],ymm4[3],ymm9[3],ymm4[8],ymm9[8],ymm4[9],ymm9[9],ymm4[10],ymm9[10],ymm4[11],ymm9[11]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm1, %zmm4, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm4 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm0, %ymm9
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm6 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm6, %zmm19, %zmm0
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm7, %zmm17
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %xmm0
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm9
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm0, %xmm26
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %ymm6
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm18, %zmm9
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm5, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm4, %ymm11
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[1],ymm0[1],ymm11[2],ymm0[2],ymm11[3],ymm0[3],ymm11[8],ymm0[8],ymm11[9],ymm0[9],ymm11[10],ymm0[10],ymm11[11],ymm0[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm5 = [2,1,2,3,11,11,11,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm4 = [5,6,5,6,5,6,7,7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm11 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm11, %ymm4, %ymm11
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm11, %zmm1, %zmm0 {%k1}
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [12,1,2,13,4,5,14,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm0, %ymm11
 ; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm1, %ymm4, %ymm9
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm1, %ymm2, %ymm11
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm19 = [8,21,10,11,20,13,14,23]
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
 ; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm3, %ymm3
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm16
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm22 = [8,21,10,11,20,13,14,23]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm3, %zmm22, %zmm0
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm9, %zmm18
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm25, %ymm9
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm31, %ymm1
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm1, %ymm3
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm3, %zmm19, %zmm0
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm11, %zmm18
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm29, %ymm11
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm11, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm27, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm1, %ymm3
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[1],ymm0[1],ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[8],ymm0[8],ymm3[9],ymm0[9],ymm3[10],ymm0[10],ymm3[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm1[4],ymm9[4],ymm1[5],ymm9[5],ymm1[6],ymm9[6],ymm1[7],ymm9[7],ymm1[12],ymm9[12],ymm1[13],ymm9[13],ymm1[14],ymm9[14],ymm1[15],ymm9[15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm2 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm5, %ymm3
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512DQ-FCP-NEXT:    # ymm6 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm14
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm5 = [2,2,0,3,10,0,10,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm5, %zmm14
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm27, %ymm9
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm1[4],ymm9[4],ymm1[5],ymm9[5],ymm1[6],ymm9[6],ymm1[7],ymm9[7],ymm1[12],ymm9[12],ymm1[13],ymm9[13],ymm1[14],ymm9[14],ymm1[15],ymm9[15]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm3, %ymm21, %ymm3
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm9 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm1[4],ymm11[4],ymm1[5],ymm11[5],ymm1[6],ymm11[6],ymm1[7],ymm11[7],ymm1[12],ymm11[12],ymm1[13],ymm11[13],ymm1[14],ymm11[14],ymm1[15],ymm11[15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm3 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm6, %ymm5
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512DQ-FCP-NEXT:    # ymm7 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm6, %ymm11
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm27 = [2,2,0,3,10,0,10,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm27, %zmm11
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm28, %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm6
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm1[4],ymm6[4],ymm1[5],ymm6[5],ymm1[6],ymm6[6],ymm1[7],ymm6[7],ymm1[12],ymm6[12],ymm1[13],ymm6[13],ymm1[14],ymm6[14],ymm1[15],ymm6[15]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm5, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm5 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
 ; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm9[0],ymm1[1],ymm9[1],ymm1[2],ymm9[2],ymm1[3],ymm9[3],ymm1[8],ymm9[8],ymm1[9],ymm9[9],ymm1[10],ymm9[10],ymm1[11],ymm9[11]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm5[0],ymm1[1],ymm5[1],ymm1[2],ymm5[2],ymm1[3],ymm5[3],ymm1[8],ymm5[8],ymm1[9],ymm5[9],ymm1[10],ymm5[10],ymm1[11],ymm5[11]
 ; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm3, %zmm1, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm7[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm4
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm4, %zmm1, %zmm0 {%k1}
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm12[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm2
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm16, %ymm1
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm7, %ymm1
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm22, %zmm0
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm4, %zmm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm8, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm8, %ymm1
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm5, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm10, %xmm4
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm5 = [0,0,2,1,8,9,8,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm3
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [1,0,2,2,1,0,2,2]
-; AVX512DQ-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm29, %xmm7
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm30, %xmm8
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm6, %ymm4, %ymm6
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm7, %ymm7
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm12, %ymm1
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm19, %zmm0
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm15, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm15, %ymm12
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm27, %zmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm0, %xmm14, %xmm2
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm0, %xmm13, %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm4 = [0,0,2,1,8,9,8,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm4, %zmm2
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [1,0,2,2,1,0,2,2]
+; AVX512DQ-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm5, %ymm3, %ymm5
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm6, %ymm6
 ; AVX512DQ-FCP-NEXT:    movw $9362, %ax # imm = 0x2492
 ; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm7, %zmm6, %zmm3 {%k1}
-; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm12[0],zero,xmm12[1],zero,xmm12[2],zero,xmm12[3],zero
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm3, %ymm7
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm8 = [16,9,10,17,12,13,18,15]
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm12[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm9, %zmm8, %zmm3
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm9 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm6, %ymm9, %ymm7
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm7, %zmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm10, %xmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm7, %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm10[0],xmm7[1],xmm10[1],xmm7[2],xmm10[2],xmm7[3],xmm10[3]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm5, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm24, %xmm2
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm5
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm6, %zmm5, %zmm2 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm31, %xmm8
+; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm5 = xmm31[0],zero,xmm31[1],zero,xmm31[2],zero,xmm31[3],zero
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm6
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm7 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm8[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm8, %zmm7, %zmm2
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm8 = [0,1,8,3,4,9,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm5, %ymm8, %ymm6
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm6, %zmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm21, %xmm6
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm0, %xmm6, %xmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm23, %xmm10
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm0, %xmm10, %xmm0
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm6[0],xmm10[1],xmm6[1],xmm10[2],xmm6[2],xmm10[3],xmm6[3]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm4, %zmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm26, %xmm0
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm4
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,0,2,1,4,5,6,7]
 ; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm10 = [0,0,0,0,8,8,0,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm10, %zmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm5, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm11 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm5[0],xmm11[1],xmm5[1],xmm11[2],xmm5[2],xmm11[3],xmm5[3]
-; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm5, %ymm5
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm5, %zmm4, %zmm6 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm4 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512DQ-FCP-NEXT:    vpermi2d %ymm4, %ymm6, %ymm9
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm4, %zmm8, %zmm6
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm9, %zmm4
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm28, %xmm6
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm6, %xmm5
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm10, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (zmm5 & (zmm6 ^ zmm4))
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm10, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm22, %xmm13
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm20, %xmm14
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm4, %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm13 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm13[0],xmm4[0],xmm13[1],xmm4[1],xmm13[2],xmm4[2],xmm13[3],xmm4[3]
+; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm4, %ymm4
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm4, %zmm3, %zmm5 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm24, %xmm4
+; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm24[0],zero,xmm24[1],zero,xmm24[2],zero,xmm24[3],zero
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm3, %ymm5, %ymm8
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm3, %zmm7, %zmm5
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm8, %zmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm25, %xmm5
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm5, %xmm4
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm10, %zmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm9 = zmm9 ^ (zmm4 & (zmm9 ^ zmm17))
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, (%rax)
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (zmm5 & (zmm2 ^ zmm3))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm2 & (zmm1 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 128(%rax)
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm14 = zmm14 ^ (zmm2 & (zmm14 ^ zmm18))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, 320(%rax)
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm20, %zmm17, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm13 = zmm13 ^ (zmm1 & (zmm13 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 256(%rax)
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm23, %zmm0
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm15 = zmm15 ^ (zmm1 & (zmm15 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, 64(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 256(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm6 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # zmm6 = zmm6 ^ (zmm4 & (zmm6 ^ mem))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (zmm4 & (zmm5 ^ zmm3))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, (%rax)
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm4 & (zmm0 ^ zmm2))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 192(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm0 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm12 = zmm12 ^ (zmm0 & (zmm12 ^ zmm1))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, 128(%rax)
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm11 = zmm11 ^ (zmm0 & (zmm11 ^ zmm18))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, 320(%rax)
+; AVX512DQ-FCP-NEXT:    popq %rax
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
 ;
@@ -8737,415 +8747,422 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ;
 ; AVX512-FCP-LABEL: store_i16_stride6_vf64:
 ; AVX512-FCP:       # %bb.0:
-; AVX512-FCP-NEXT:    subq $1240, %rsp # imm = 0x4D8
-; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %ymm1
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %ymm2
-; AVX512-FCP-NEXT:    vmovdqa 64(%rcx), %ymm6
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %ymm5
-; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %xmm9
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %xmm10
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm7
-; AVX512-FCP-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 64(%rcx), %xmm13
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm8
-; AVX512-FCP-NEXT:    vmovdqa %xmm8, (%rsp) # 16-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm12
-; AVX512-FCP-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %xmm11
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm4
+; AVX512-FCP-NEXT:    subq $1320, %rsp # imm = 0x528
+; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %ymm3
+; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm0 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm3, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %ymm4
 ; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm3
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm4, %ymm2
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm22 = [2,1,2,3,11,11,11,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm22, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %ymm3
 ; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[8],ymm6[8],ymm5[9],ymm6[9],ymm5[10],ymm6[10],ymm5[11],ymm6[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm11[4],xmm13[4],xmm11[5],xmm13[5],xmm11[6],xmm13[6],xmm11[7],xmm13[7]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm13, %xmm18
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm3
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm7
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm7[0],ymm3[0],ymm7[1],ymm3[1],ymm7[2],ymm3[2],ymm7[3],ymm3[3],ymm7[8],ymm3[8],ymm7[9],ymm3[9],ymm7[10],ymm3[10],ymm7[11],ymm3[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm8 = xmm12[4],xmm0[4],xmm12[5],xmm0[5],xmm12[6],xmm0[6],xmm12[7],xmm0[7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm8, %zmm4
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %ymm12
-; AVX512-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm12, %ymm4
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %ymm0
-; AVX512-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm0, %ymm8
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm8[0],ymm4[0],ymm8[1],ymm4[1],ymm8[2],ymm4[2],ymm8[3],ymm4[3],ymm8[8],ymm4[8],ymm8[9],ymm4[9],ymm8[10],ymm4[10],ymm8[11],ymm4[11]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm27 = [2,1,2,3,11,11,11,11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm0[4],ymm12[4],ymm0[5],ymm12[5],ymm0[6],ymm12[6],ymm0[7],ymm12[7],ymm0[12],ymm12[12],ymm0[13],ymm12[13],ymm0[14],ymm12[14],ymm0[15],ymm12[15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm27, %zmm8
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm4 = [5,6,5,6,5,6,7,7]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm12 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; AVX512-FCP-NEXT:    vpermd %ymm12, %ymm4, %ymm12
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm23 = [5,6,5,6,5,6,7,7]
+; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm23, %ymm2
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
 ; AVX512-FCP-NEXT:    movw $18724, %ax # imm = 0x4924
 ; AVX512-FCP-NEXT:    kmovw %eax, %k1
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm8 {%k1}
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm25 = [12,1,2,13,4,5,14,7]
-; AVX512-FCP-NEXT:    vmovdqa %ymm8, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %ymm2
-; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2d %ymm1, %ymm25, %ymm0
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm28 = [8,21,10,11,20,13,14,23]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm12 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm2, %ymm1
-; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm28, %zmm8
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm29 = [2,2,0,3,10,0,10,11]
-; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %ymm1
-; AVX512-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm15 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
-; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm1, %ymm0
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm14 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512-FCP-NEXT:    # ymm14 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm29, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm21 = [12,1,2,13,4,5,14,7]
+; AVX512-FCP-NEXT:    vmovdqa %ymm1, %ymm3
+; AVX512-FCP-NEXT:    vpermt2d %ymm2, %ymm21, %ymm3
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm24 = [8,21,10,11,20,13,14,23]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm4, %ymm2
+; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm24, %zmm1
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm3, %zmm1
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %ymm2
 ; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm2, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %ymm8
-; AVX512-FCP-NEXT:    vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm8, %ymm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm8[4],ymm2[4],ymm8[5],ymm2[5],ymm8[6],ymm2[6],ymm8[7],ymm2[7],ymm8[12],ymm2[12],ymm8[13],ymm2[13],ymm8[14],ymm2[14],ymm8[15],ymm2[15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm27, %zmm0
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm5[4],ymm6[4],ymm5[5],ymm6[5],ymm5[6],ymm6[6],ymm5[7],ymm6[7],ymm5[12],ymm6[12],ymm5[13],ymm6[13],ymm5[14],ymm6[14],ymm5[15],ymm6[15]
-; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm4, %ymm1
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm6 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm5 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[8],ymm6[8],ymm5[9],ymm6[9],ymm5[10],ymm6[10],ymm5[11],ymm6[11]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm5, %zmm1
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa %ymm0, %ymm1
-; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %ymm2
-; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm5 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2d %ymm5, %ymm25, %ymm1
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm2, %ymm5
-; AVX512-FCP-NEXT:    vpermt2d %zmm5, %zmm28, %zmm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %ymm1
-; AVX512-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm1, %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm29, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm1
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm15 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512-FCP-NEXT:    # ymm15 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm2, %ymm2
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm25 = [2,2,0,3,10,0,10,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm25, %zmm2
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %ymm3
+; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm3, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm4, %ymm2
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm22, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa 64(%rcx), %ymm3
+; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm23, %ymm2
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa %ymm1, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm21, %ymm2
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm4, %ymm3
+; AVX512-FCP-NEXT:    vpermt2d %zmm3, %zmm24, %zmm1
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %xmm2
-; AVX512-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %ymm2
+; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm1
+; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm2, %ymm2
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm25, %zmm2
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %xmm3
+; AVX512-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm8 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm5
-; AVX512-FCP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm1
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm16 = [0,0,2,1,8,9,8,9]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm16, %zmm1
-; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} ymm30 = [1,0,2,2,1,0,2,2]
-; AVX512-FCP-NEXT:    # ymm30 = mem[0,1,2,3,0,1,2,3]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm30, %ymm0
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
-; AVX512-FCP-NEXT:    vpbroadcastq %xmm2, %ymm2
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm1
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm4, %xmm2
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm26 = [0,0,2,1,8,9,8,9]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm26, %zmm1
+; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} ymm27 = [1,0,2,2,1,0,2,2]
+; AVX512-FCP-NEXT:    # ymm27 = mem[0,1,2,3,0,1,2,3]
+; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %xmm3
+; AVX512-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, (%rsp) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm27, %ymm2
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpbroadcastq %xmm3, %ymm3
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
 ; AVX512-FCP-NEXT:    movw $9362, %ax # imm = 0x2492
 ; AVX512-FCP-NEXT:    kmovw %eax, %k2
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k2}
-; AVX512-FCP-NEXT:    vmovdqa %ymm1, %ymm0
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [16,9,10,17,12,13,18,15]
-; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %xmm5
-; AVX512-FCP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm17, %zmm1
-; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm2 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm26 = [0,1,8,3,4,9,6,7]
-; AVX512-FCP-NEXT:    vpermt2d %ymm2, %ymm26, %ymm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm31 = [0,0,0,0,8,8,0,9]
-; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm0
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm31, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %xmm2
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm5
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm1
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm5, %xmm23
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm24
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm16, %zmm1
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm18, %xmm2
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm11[0],xmm2[0],xmm11[1],xmm2[1],xmm11[2],xmm2[2],xmm11[3],xmm2[3]
-; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm30, %ymm0
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
-; AVX512-FCP-NEXT:    vpbroadcastq %xmm2, %ymm2
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k2}
-; AVX512-FCP-NEXT:    vmovdqa %ymm1, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %xmm5
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm17, %zmm1
-; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm2 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm5, %xmm19
-; AVX512-FCP-NEXT:    vpermt2d %ymm2, %ymm26, %ymm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm0
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm31, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa %ymm1, %ymm2
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm28 = [16,9,10,17,12,13,18,15]
+; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2d %zmm3, %zmm28, %zmm1
+; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm29 = [0,1,8,3,4,9,6,7]
+; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm29, %ymm2
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm2
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm2, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm5
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm5, %ymm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm5[4],ymm2[4],ymm5[5],ymm2[5],ymm5[6],ymm2[6],ymm5[7],ymm2[7],ymm5[12],ymm2[12],ymm5[13],ymm2[13],ymm5[14],ymm2[14],ymm5[15],ymm2[15]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm5, %ymm21
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm20
-; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm27, %zmm0
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm7[4],ymm3[4],ymm7[5],ymm3[5],ymm7[6],ymm3[6],ymm7[7],ymm3[7],ymm7[12],ymm3[12],ymm7[13],ymm3[13],ymm7[14],ymm3[14],ymm7[15],ymm3[15]
-; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm4, %ymm1
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm30 = [0,0,0,0,8,8,0,9]
+; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %xmm2
+; AVX512-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm5 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm1
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm30, %zmm2
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %xmm3
+; AVX512-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm1
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm4, %xmm2
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm26, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa 64(%rcx), %xmm3
+; AVX512-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm27, %ymm2
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpbroadcastq %xmm3, %ymm3
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa %ymm1, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2d %zmm3, %zmm28, %zmm1
+; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
+; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm29, %ymm2
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa %ymm0, %ymm1
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm3
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2d %ymm2, %ymm25, %ymm1
-; AVX512-FCP-NEXT:    vmovdqa %ymm12, %ymm5
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm3, %ymm2
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm18
-; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm28, %zmm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm12
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm12, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm11
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm11, %ymm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm11[4],ymm12[4],ymm11[5],ymm12[5],ymm11[6],ymm12[6],ymm11[7],ymm12[7],ymm11[12],ymm12[12],ymm11[13],ymm12[13],ymm11[14],ymm12[14],ymm11[15],ymm12[15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm27, %zmm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %ymm3
-; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm3, %ymm1
-; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm3, %ymm2
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm22
-; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm29, %zmm2
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %xmm2
+; AVX512-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm1
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm30, %zmm2
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15]
-; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm4, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm3
+; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm3, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm4, %ymm2
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm22, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm3
+; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm23, %ymm2
 ; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
 ; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa %ymm1, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm21, %ymm2
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm4, %ymm3
+; AVX512-FCP-NEXT:    vpermt2d %zmm3, %zmm24, %zmm1
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %ymm3
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm3, %ymm1
+; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm3, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm31
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm25, %zmm2
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm13
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm13, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm11
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm0
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm11[4],ymm13[4],ymm11[5],ymm13[5],ymm11[6],ymm13[6],ymm11[7],ymm13[7],ymm11[12],ymm13[12],ymm11[13],ymm13[13],ymm11[14],ymm13[14],ymm11[15],ymm13[15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm22, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm10
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm9
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[12],ymm10[12],ymm9[13],ymm10[13],ymm9[14],ymm10[14],ymm9[15],ymm10[15]
+; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm23, %ymm1
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm10[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512-FCP-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm10
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm25
-; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm10, %ymm1
-; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm28, %zmm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm25, %zmm25
-; AVX512-FCP-NEXT:    vmovdqa (%r9), %ymm7
-; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm7, %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm7, %ymm13
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm29, %zmm13
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm15
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm14
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm15, %xmm0
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm14, %xmm1
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm14
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm14[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm21
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm14, %ymm1
+; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm24, %zmm0
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm21, %zmm21
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %ymm12
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm12, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm12, %ymm1
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm25, %zmm1
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm2
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm3
+; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm0
+; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm1
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm16, %zmm4
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm30, %ymm5
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm6
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm3
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm6, %xmm9
-; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm8 = xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
-; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm16, %zmm0
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm9 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
-; AVX512-FCP-NEXT:    vpbroadcastq %xmm8, %ymm8
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm5, %zmm5
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm5, %zmm4 {%k2}
-; AVX512-FCP-NEXT:    vmovdqa %ymm4, %ymm5
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm8
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm8[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2d %zmm9, %zmm17, %zmm4
-; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm9 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero
-; AVX512-FCP-NEXT:    vpermt2d %ymm9, %ymm26, %ymm5
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm16
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-FCP-NEXT:    vmovdqa (%rsp), %xmm2 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512-FCP-NEXT:    vpermd %ymm4, %ymm30, %ymm4
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm9 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
-; AVX512-FCP-NEXT:    vpbroadcastq %xmm5, %ymm5
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm4, %zmm0 {%k2}
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm1
-; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm4 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
-; AVX512-FCP-NEXT:    vpermi2d %ymm4, %ymm0, %ymm26
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm1[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2d %zmm4, %zmm17, %zmm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm9
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm9, %xmm4
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm9[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm31, %zmm5
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm26, %zmm17
-; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm0
-; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm0, %xmm2
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm0[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm31, %zmm4
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm11[0],ymm12[0],ymm11[1],ymm12[1],ymm11[2],ymm12[2],ymm11[3],ymm12[3],ymm11[8],ymm12[8],ymm11[9],ymm12[9],ymm11[10],ymm12[10],ymm11[11],ymm12[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm12 = [1,1,1,1,10,10,10,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm12, %zmm3
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm26 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm3, %zmm2 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm3
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm27 = [0,9,2,3,8,5,6,11]
-; AVX512-FCP-NEXT:    vmovdqa %ymm2, %ymm1
-; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm27, %ymm1
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm30 = [8,9,20,11,12,21,14,15]
-; AVX512-FCP-NEXT:    vpermt2d %zmm3, %zmm30, %zmm2
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm3
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm6, %xmm31
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm28 = [0,0,0,1,0,10,10,0]
-; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm28, %zmm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm3, %xmm20
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm19
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm26, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm3
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm4
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm27, %ymm0
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm3, %xmm18
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm4, %xmm17
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-FCP-NEXT:    vpbroadcastq %xmm2, %ymm2
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa %ymm1, %ymm0
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm3
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm3[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2d %zmm2, %zmm28, %zmm1
+; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm2 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm3, %xmm16
+; AVX512-FCP-NEXT:    vpermt2d %ymm2, %ymm29, %ymm0
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm22
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm7
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm5
+; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm7, %xmm0
+; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm1
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm26, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm27, %ymm0
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
+; AVX512-FCP-NEXT:    vpbroadcastq %xmm6, %ymm6
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
+; AVX512-FCP-NEXT:    vpermi2d %ymm6, %ymm1, %ymm29
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm2[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2d %zmm6, %zmm28, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm15
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm0 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512-FCP-NEXT:    vpshufb %xmm0, %xmm15, %xmm6
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm15[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm30, %zmm8
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm29, %zmm23
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm1
+; AVX512-FCP-NEXT:    vpshufb %xmm0, %xmm1, %xmm0
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm1[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm30, %zmm6
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm11[0],ymm13[0],ymm11[1],ymm13[1],ymm11[2],ymm13[2],ymm11[3],ymm13[3],ymm11[8],ymm13[8],ymm11[9],ymm13[9],ymm11[10],ymm13[10],ymm11[11],ymm13[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm28 = [1,1,1,1,10,10,10,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm28, %zmm5
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[8],ymm10[8],ymm9[9],ymm10[9],ymm9[10],ymm10[10],ymm9[11],ymm10[11]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm24 = [1,2,1,2,0,0,3,3,21,20,18,18,21,20,22,22]
+; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm24, %zmm4
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm5, %zmm4 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm0
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [0,9,2,3,8,5,6,11]
+; AVX512-FCP-NEXT:    vmovdqa %ymm4, %ymm2
+; AVX512-FCP-NEXT:    vpermt2d %ymm0, %ymm3, %ymm2
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm14[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm25 = [8,9,20,11,12,21,14,15]
+; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm25, %zmm4
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm26
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm4 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm1, %xmm2
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm1 = [0,0,0,1,0,10,10,0]
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm2
 ; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    # ymm7 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512-FCP-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm12, %zmm0
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm7 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm0, %zmm7 {%k1}
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm5 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm28, %zmm5
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
+; AVX512-FCP-NEXT:    vmovdqa (%rsp), %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:    vpermt2d %zmm0, %zmm24, %zmm9
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm5, %zmm9 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm7, %ymm29
-; AVX512-FCP-NEXT:    vpermt2d %ymm0, %ymm27, %ymm29
+; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm0, %xmm0
+; AVX512-FCP-NEXT:    vmovdqa %ymm9, %ymm5
+; AVX512-FCP-NEXT:    vpermt2d %ymm0, %ymm3, %ymm5
 ; AVX512-FCP-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
 ; AVX512-FCP-NEXT:    # ymm0 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm30, %zmm7
+; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm25, %zmm9
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm5, %zmm27
+; AVX512-FCP-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm0 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm5, %xmm5
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm5
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
+; AVX512-FCP-NEXT:    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:    vpermt2q %zmm0, %zmm28, %zmm9
 ; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
 ; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
 ; AVX512-FCP-NEXT:    # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm23, %xmm10
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm24, %xmm6
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm10[4],xmm6[4],xmm10[5],xmm6[5],xmm10[6],xmm6[6],xmm10[7],xmm6[7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm12, %zmm10
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm0 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm10, %zmm0 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm19, %xmm6
-; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm6, %xmm10
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm19
-; AVX512-FCP-NEXT:    vpermt2d %ymm10, %ymm27, %ymm19
-; AVX512-FCP-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    # ymm10 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2d %zmm10, %zmm30, %zmm0
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm21, %ymm6
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm20, %ymm10
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm6[0],ymm10[0],ymm6[1],ymm10[1],ymm6[2],ymm10[2],ymm6[3],ymm10[3],ymm6[8],ymm10[8],ymm6[9],ymm10[9],ymm6[10],ymm10[10],ymm6[11],ymm10[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm14 = xmm14[4],xmm15[4],xmm14[5],xmm15[5],xmm14[6],xmm15[6],xmm14[7],xmm15[7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm12, %zmm14
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm10 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vmovdqa32 %zmm14, %zmm10 {%k1}
-; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm8, %xmm8
-; AVX512-FCP-NEXT:    vpermi2d %ymm8, %ymm10, %ymm27
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm18, %ymm6
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2d %zmm8, %zmm30, %zmm10
-; AVX512-FCP-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    # ymm8 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm31, %xmm14
-; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm11, %xmm11
-; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm28, %zmm11
-; AVX512-FCP-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    # ymm8 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm6, %xmm12
-; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm28, %zmm12
-; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm9, %xmm6
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm22, %ymm8
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm8[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm28, %zmm6
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm27, %zmm8
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (zmm9 & (zmm6 ^ zmm8))
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm10 = xmm10[4],mem[4],xmm10[5],mem[5],xmm10[6],mem[6],xmm10[7],mem[7]
+; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm24, %zmm10
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm9, %zmm10 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm0, %xmm0
+; AVX512-FCP-NEXT:    vmovdqa %ymm10, %ymm9
+; AVX512-FCP-NEXT:    vpermt2d %ymm0, %ymm3, %ymm9
+; AVX512-FCP-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm0 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm25, %zmm10
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512-FCP-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm10 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm0, %xmm0
+; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm1, %zmm0
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm10 = ymm10[0],mem[0],ymm10[1],mem[1],ymm10[2],mem[2],ymm10[3],mem[3],ymm10[8],mem[8],ymm10[9],mem[9],ymm10[10],mem[10],ymm10[11],mem[11]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm20, %xmm11
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm19, %xmm12
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm28, %zmm11
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm10 = ymm10[0],mem[0],ymm10[1],mem[1],ymm10[2],mem[2],ymm10[3],mem[3],ymm10[8],mem[8],ymm10[9],mem[9],ymm10[10],mem[10],ymm10[11],mem[11]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm18, %xmm12
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm17, %xmm13
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX512-FCP-NEXT:    vpermt2d %zmm10, %zmm24, %zmm12
+; AVX512-FCP-NEXT:    vmovdqa32 %zmm11, %zmm12 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm16, %xmm10
+; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm10, %xmm7
+; AVX512-FCP-NEXT:    vpermi2d %ymm7, %ymm12, %ymm3
+; AVX512-FCP-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm7 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermt2d %zmm7, %zmm25, %zmm12
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm3, %zmm3
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm15, %xmm4
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm31, %ymm7
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm1, %zmm4
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm4 = zmm4 ^ (zmm1 & (zmm4 ^ zmm3))
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 256(%rax)
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm19, %zmm0
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm12 = zmm12 ^ (zmm9 & (zmm12 ^ zmm0))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, 448(%rax)
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm29, %zmm0
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm11 = zmm11 ^ (zmm9 & (zmm11 ^ zmm0))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, 640(%rax)
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm0
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (zmm9 & (zmm3 ^ zmm0))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 256(%rax)
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm1 & (zmm0 ^ zmm9))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 448(%rax)
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (zmm1 & (zmm5 ^ zmm27))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 640(%rax)
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (zmm1 & (zmm2 ^ zmm26))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)
 ; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm4 = zmm4 ^ (zmm0 & (zmm4 ^ zmm17))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (zmm0 & (zmm5 ^ zmm16))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 192(%rax)
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (zmm0 & (zmm6 ^ zmm23))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, (%rax)
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm8 = zmm8 ^ (zmm0 & (zmm8 ^ zmm22))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, 192(%rax)
 ; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm13 = zmm13 ^ (zmm1 & (zmm13 ^ zmm25))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rax)
+; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (zmm1 & (zmm2 ^ zmm21))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 128(%rax)
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
 ; AVX512-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 64-byte Folded Reload
 ; AVX512-FCP-NEXT:    # zmm2 = zmm2 ^ (zmm1 & (zmm2 ^ mem))
@@ -9166,7 +9183,7 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
 ; AVX512-FCP-NEXT:    # zmm0 = zmm0 ^ (zmm1 & (zmm0 ^ mem))
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 704(%rax)
-; AVX512-FCP-NEXT:    addq $1240, %rsp # imm = 0x4D8
+; AVX512-FCP-NEXT:    addq $1320, %rsp # imm = 0x528
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -9661,448 +9678,451 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ;
 ; AVX512DQ-FCP-LABEL: store_i16_stride6_vf64:
 ; AVX512DQ-FCP:       # %bb.0:
-; AVX512DQ-FCP-NEXT:    subq $1176, %rsp # imm = 0x498
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm1
+; AVX512DQ-FCP-NEXT:    subq $1112, %rsp # imm = 0x458
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm1
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; 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:    vmovdqa (%rcx), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm4
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rcx), %xmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm2
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm22 = [1,1,1,1,10,10,10,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm22, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[8],ymm2[8],ymm0[9],ymm2[9],ymm0[10],ymm2[10],ymm0[11],ymm2[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm2
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %xmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm3
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %ymm7
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm7[0],ymm2[0],ymm7[1],ymm2[1],ymm7[2],ymm2[2],ymm7[3],ymm2[3],ymm7[8],ymm2[8],ymm7[9],ymm2[9],ymm7[10],ymm2[10],ymm7[11],ymm2[11]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm7, %ymm29
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm31
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %xmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %xmm7
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm7[4],xmm2[4],xmm7[5],xmm2[5],xmm7[6],xmm2[6],xmm7[7],xmm2[7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rcx), %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %ymm8
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm8[0],ymm1[0],ymm8[1],ymm1[1],ymm8[2],ymm1[2],ymm8[3],ymm1[3],ymm8[8],ymm1[8],ymm8[9],ymm1[9],ymm8[10],ymm1[10],ymm8[11],ymm1[11]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm8, %ymm30
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm3, %zmm20
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[8],ymm3[8],ymm1[9],ymm3[9],ymm1[10],ymm3[10],ymm1[11],ymm3[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm3, %zmm16
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm17 = [1,1,1,1,10,10,10,11]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[8],ymm3[8],ymm1[9],ymm3[9],ymm1[10],ymm3[10],ymm1[11],ymm3[11]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm4
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm17, %zmm3
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm19, %zmm21
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm18 = [1,2,1,2,0,0,3,3,21,20,18,18,21,20,22,22]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm18, %zmm2
 ; AVX512DQ-FCP-NEXT:    movw $18724, %ax # imm = 0x4924
 ; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm3, %zmm21 {%k1}
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm24 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm21, %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm1, %zmm2 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm8 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm0, %ymm24, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm18 = [8,9,20,11,12,21,14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm5 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm23 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm3
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm0, %ymm23, %ymm3
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm21 = [8,9,20,11,12,21,14,15]
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm18, %zmm21
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm22 = [0,0,0,1,0,10,10,0]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm21, %zmm2
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm14 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm22, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [0,0,0,1,0,10,10,0]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm20, %zmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %ymm9
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %ymm5
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm5[0],ymm9[0],ymm5[1],ymm9[1],ymm5[2],ymm9[2],ymm5[3],ymm9[3],ymm5[8],ymm9[8],ymm5[9],ymm9[9],ymm5[10],ymm9[10],ymm5[11],ymm9[11]
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm15
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm0, %xmm23
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm17, %zmm4
-; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm19, %zmm25
-; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm4, %zmm25 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm25, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %ymm9
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm9[0],ymm1[0],ymm9[1],ymm1[1],ymm9[2],ymm1[2],ymm9[3],ymm1[3],ymm9[8],ymm1[8],ymm9[9],ymm1[9],ymm9[10],ymm1[10],ymm9[11],ymm1[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm24
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm3, %xmm27
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm19
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm22, %zmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %ymm11
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %ymm1
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm11[0],ymm1[1],ymm11[1],ymm1[2],ymm11[2],ymm1[3],ymm11[3],ymm1[8],ymm11[8],ymm1[9],ymm11[9],ymm1[10],ymm11[10],ymm1[11],ymm11[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm25
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %xmm4
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm4, %xmm31
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm28
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm18, %zmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm2, %zmm3 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm3, %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r8), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm2
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm2, %ymm24, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm5, %xmm4
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm1, %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm30
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm2, %ymm23, %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r8), %ymm10
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm3, %zmm18, %zmm25
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %ymm11
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm4 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm22, %zmm0
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm2, %zmm21, %zmm3
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %ymm13
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm13[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm6, %xmm16
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm20, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %xmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %ymm13
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %ymm14
 ; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %ymm12
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[2],ymm13[2],ymm12[3],ymm13[3],ymm12[8],ymm13[8],ymm12[9],ymm13[9],ymm12[10],ymm13[10],ymm12[11],ymm13[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm17, %zmm6
-; AVX512DQ-FCP-NEXT:    vpermd %zmm20, %zmm19, %zmm28
-; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm6, %zmm28 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm28, %ymm20
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm12[0],ymm14[0],ymm12[1],ymm14[1],ymm12[2],ymm14[2],ymm12[3],ymm14[3],ymm12[8],ymm14[8],ymm12[9],ymm14[9],ymm12[10],ymm14[10],ymm12[11],ymm14[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm22, %zmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rcx), %xmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rcx), %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %ymm8
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm8[0],ymm5[0],ymm8[1],ymm5[1],ymm8[2],ymm5[2],ymm8[3],ymm5[3],ymm8[8],ymm5[8],ymm8[9],ymm5[9],ymm8[10],ymm5[10],ymm8[11],ymm5[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm18, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm1, %ymm2
 ; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r8), %xmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm6
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm6, %ymm24, %ymm20
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r8), %ymm6
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm7, %zmm18, %zmm28
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %xmm1
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm4, %xmm17
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm0, %ymm23, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r8), %ymm7
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm7[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm15, %zmm21, %zmm1
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %ymm6
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm16, %xmm2
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm16, %xmm26
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm20, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %ymm7
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm22, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, (%rsp) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm2
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm0
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm22, %zmm15
+; 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), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[8],ymm2[8],ymm0[9],ymm2[9],ymm0[10],ymm2[10],ymm0[11],ymm2[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm2
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm29
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm17, %zmm0
-; AVX512DQ-FCP-NEXT:    vpermd %zmm16, %zmm19, %zmm19
-; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm0, %zmm19 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT:    vpermi2d %ymm0, %ymm19, %ymm24
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm18, %zmm19
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm4, %zmm18, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa32 %zmm15, %zmm0 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, (%rsp) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm17, %xmm2
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm23
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm2
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm16
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm21, %zmm0
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm23, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %xmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm0, %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm22, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm26, %xmm1
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm0, %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %ymm2
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm2[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm18
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm20, %zmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm14 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm5, %ymm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm5[4],ymm9[4],ymm5[5],ymm9[5],ymm5[6],ymm9[6],ymm5[7],ymm9[7],ymm5[12],ymm9[12],ymm5[13],ymm9[13],ymm5[14],ymm9[14],ymm5[15],ymm9[15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm27 = [2,1,2,3,11,11,11,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm27, %zmm8
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm26 = [5,6,5,6,5,6,7,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm29, %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm31, %ymm1
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm26, %ymm0
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm5 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm5[0],ymm1[0],ymm5[1],ymm1[1],ymm5[2],ymm1[2],ymm5[3],ymm1[3],ymm5[8],ymm1[8],ymm5[9],ymm1[9],ymm5[10],ymm1[10],ymm5[11],ymm1[11]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm0, %zmm1, %zmm8 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm15 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm24, %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm9, %ymm1
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm9[4],ymm2[4],ymm9[5],ymm2[5],ymm9[6],ymm2[6],ymm9[7],ymm2[7],ymm9[12],ymm2[12],ymm9[13],ymm2[13],ymm9[14],ymm2[14],ymm9[15],ymm2[15]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm24 = [2,1,2,3,11,11,11,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm24, %zmm1
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm23 = [5,6,5,6,5,6,7,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm25, %ymm2
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm2[4],ymm11[4],ymm2[5],ymm11[5],ymm2[6],ymm11[6],ymm2[7],ymm11[7],ymm2[12],ymm11[12],ymm2[13],ymm11[13],ymm2[14],ymm11[14],ymm2[15],ymm11[15]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm23, %ymm0
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm9 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm11 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm11[0],ymm9[0],ymm11[1],ymm9[1],ymm11[2],ymm9[2],ymm11[3],ymm9[3],ymm11[8],ymm9[8],ymm11[9],ymm9[9],ymm11[10],ymm9[10],ymm11[11],ymm9[11]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm0, %zmm9, %zmm1 {%k1}
 ; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm22 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm1, %ymm22, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm9, %ymm22, %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm2 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm10, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm10, %ymm4
 ; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm10
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm29 = [8,21,10,11,20,13,14,23]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm29, %zmm8
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm5 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm11, %ymm0
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512DQ-FCP-NEXT:    # ymm8 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm11, %ymm1
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm31 = [2,2,0,3,10,0,10,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm31, %zmm1
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm12, %ymm2
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm12[4],ymm13[4],ymm12[5],ymm13[5],ymm12[6],ymm13[6],ymm12[7],ymm13[7],ymm12[12],ymm13[12],ymm12[13],ymm13[13],ymm12[14],ymm13[14],ymm12[15],ymm13[15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm27, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm4
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm2, %ymm26, %ymm2
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm2, %zmm3, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm0, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm3, %ymm22, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm6, %ymm3
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm3, %zmm29, %zmm0
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm25 = [8,21,10,11,20,13,14,23]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm4, %zmm25, %zmm1
+; 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 %ymm5, %ymm7, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm7, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm8, %ymm12
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm31, %zmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm8 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm23, %xmm4
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm4, %xmm0
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm15, %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm15[0],xmm4[0],xmm15[1],xmm4[1],xmm15[2],xmm4[2],xmm15[3],xmm4[3]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [0,0,2,1,8,9,8,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm18, %zmm2
-; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} ymm30 = [1,0,2,2,1,0,2,2]
-; AVX512DQ-FCP-NEXT:    # ymm30 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm30, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm13, %ymm0
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512DQ-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm13
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm26 = [2,2,0,3,10,0,10,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm26, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm14, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm12, %ymm1
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm12[4],ymm14[4],ymm12[5],ymm14[5],ymm12[6],ymm14[6],ymm12[7],ymm14[7],ymm12[12],ymm14[12],ymm12[13],ymm14[13],ymm12[14],ymm14[14],ymm12[15],ymm14[15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm24, %zmm0
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm8[4],ymm5[4],ymm8[5],ymm5[5],ymm8[6],ymm5[6],ymm8[7],ymm5[7],ymm8[12],ymm5[12],ymm8[13],ymm5[13],ymm8[14],ymm5[14],ymm8[15],ymm5[15]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm1, %ymm23, %ymm1
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm5 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm1, %zmm4, %zmm0 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm0, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm4 = ymm7[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm4, %ymm22, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm7, %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm10, %ymm12
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm3, %zmm25, %zmm0
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm21
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm6, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm6, %ymm0
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm26, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm19, %xmm5
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm27, %xmm4
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm4, %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm27 = [0,0,2,1,8,9,8,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm27, %zmm2
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [1,0,2,2,1,0,2,2]
+; AVX512DQ-FCP-NEXT:    # ymm7 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm31, %xmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm28, %xmm4
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm3, %ymm7, %ymm3
 ; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
 ; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
 ; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm4, %ymm4
 ; AVX512DQ-FCP-NEXT:    movw $9362, %ax # imm = 0x2492
 ; AVX512DQ-FCP-NEXT:    kmovw %eax, %k2
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm4, %zmm0, %zmm2 {%k2}
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm4 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm5
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [16,9,10,17,12,13,18,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm17, %zmm2
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm4, %ymm0, %ymm5
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm5, %zmm2
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm9 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm4, %zmm3, %zmm2 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm30, %xmm4
+; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm30[0],zero,xmm30[1],zero,xmm30[2],zero,xmm30[3],zero
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm6
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm31 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm4, %zmm31, %zmm2
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm28 = [0,1,8,3,4,9,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm3, %ymm28, %ymm6
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm6, %zmm20
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm3, %xmm4
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm17 = [0,0,0,0,8,8,0,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm17, %zmm3
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm5, %xmm2
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm5[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm16 = [0,0,0,0,8,8,0,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm16, %zmm4
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm6, %xmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm7, %xmm5
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm18, %zmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm2, %ymm30, %ymm2
-; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm6, %ymm6
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm6, %zmm2, %zmm5 {%k2}
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm5, %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm7[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm6, %zmm17, %zmm5
-; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm6, %ymm0, %ymm2
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm2, %zmm23
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm8, %xmm9
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm9[4],xmm4[4],xmm9[5],xmm4[5],xmm9[6],xmm4[6],xmm9[7],xmm4[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm27, %zmm9
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm5, %xmm2
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm16, %zmm5
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm9, %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm7, %ymm6
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm6[0],ymm2[0],ymm6[1],ymm2[1],ymm6[2],ymm2[2],ymm6[3],ymm2[3],ymm6[8],ymm2[8],ymm6[9],ymm2[9],ymm6[10],ymm2[10],ymm6[11],ymm2[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm7[4],ymm9[4],ymm7[5],ymm9[5],ymm7[6],ymm9[6],ymm7[7],ymm9[7],ymm7[12],ymm9[12],ymm7[13],ymm9[13],ymm7[14],ymm9[14],ymm7[15],ymm9[15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm27, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15]
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm7 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm9 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm2, %ymm26, %ymm2
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm2, %zmm7, %zmm6 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm6, %ymm7
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm2, %ymm22, %ymm7
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm10, %ymm13
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm9, %ymm2
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm2, %zmm29, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm11, %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm10, %ymm9
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm9[0],ymm2[0],ymm9[1],ymm2[1],ymm9[2],ymm2[2],ymm9[3],ymm2[3],ymm9[8],ymm2[8],ymm9[9],ymm2[9],ymm9[10],ymm2[10],ymm9[11],ymm2[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm10[4],ymm11[4],ymm10[5],ymm11[5],ymm10[6],ymm11[6],ymm10[7],ymm11[7],ymm10[12],ymm11[12],ymm10[13],ymm11[13],ymm10[14],ymm11[14],ymm10[15],ymm11[15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm27, %zmm2
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm7, %zmm27
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm14 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm6, %ymm7
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm10 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm11 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm4, %ymm7, %ymm4
+; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm10, %ymm10
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm10, %zmm4, %zmm9 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm9, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm5[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm10, %zmm31, %zmm9
+; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm10 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm10, %ymm28, %ymm4
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm4, %zmm30
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm5, %xmm4
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm17, %zmm14
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm8, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm5, %ymm9
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm9[0],ymm4[0],ymm9[1],ymm4[1],ymm9[2],ymm4[2],ymm9[3],ymm4[3],ymm9[8],ymm4[8],ymm9[9],ymm4[9],ymm9[10],ymm4[10],ymm9[11],ymm4[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm5[4],ymm8[4],ymm5[5],ymm8[5],ymm5[6],ymm8[6],ymm5[7],ymm8[7],ymm5[12],ymm8[12],ymm5[13],ymm8[13],ymm5[14],ymm8[14],ymm5[15],ymm8[15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm24, %zmm4
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm29, %ymm8
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm8[4],ymm5[4],ymm8[5],ymm5[5],ymm8[6],ymm5[6],ymm8[7],ymm5[7],ymm8[12],ymm5[12],ymm8[13],ymm5[13],ymm8[14],ymm5[14],ymm8[15],ymm5[15]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm9, %ymm23, %ymm9
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm10 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm11 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm9, %zmm10, %zmm4 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm4, %ymm9
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm16, %ymm5
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm10, %ymm22, %ymm9
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm12, %ymm8
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm5, %ymm10
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm10, %zmm25, %zmm4
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm12, %ymm10
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm5, %ymm11
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm11 = ymm5[4],ymm12[4],ymm5[5],ymm12[5],ymm5[6],ymm12[6],ymm5[7],ymm12[7],ymm5[12],ymm12[12],ymm5[13],ymm12[13],ymm5[14],ymm12[14],ymm5[15],ymm12[15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm24, %zmm10
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm9, %zmm24
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm18, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm15 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm4, %ymm11
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm13, %ymm6
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm4, %ymm9
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm26, %zmm9
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm11 = ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[12],ymm4[12],ymm5[13],ymm4[13],ymm5[14],ymm4[14],ymm5[15],ymm4[15]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm11, %ymm23, %ymm11
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm12 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm13 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm13[0],ymm12[0],ymm13[1],ymm12[1],ymm13[2],ymm12[2],ymm13[3],ymm12[3],ymm13[8],ymm12[8],ymm13[9],ymm12[9],ymm13[10],ymm12[10],ymm13[11],ymm12[11]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,2,2,2]
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm11, %zmm12, %zmm10 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm11 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm11, %ymm10, %ymm22
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm11
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm11, %zmm25, %zmm10
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm22, %zmm22
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm4, %ymm12
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm4, %ymm11
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm12, %zmm26, %zmm11
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm4, %xmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm13
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm27, %zmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm13, %ymm7, %ymm13
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm15 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm15[0],xmm5[1],xmm15[1],xmm5[2],xmm15[2],xmm5[3],xmm15[3]
+; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm5, %ymm5
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm5, %zmm13, %zmm12 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsp), %xmm4 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm4[2,1,3,3,4,5,6,7]
 ; AVX512DQ-FCP-NEXT:    vmovdqa %ymm12, %ymm15
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm6, %ymm6
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm31, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm7 = ymm10[4],ymm9[4],ymm10[5],ymm9[5],ymm10[6],ymm9[6],ymm10[7],ymm9[7],ymm10[12],ymm9[12],ymm10[13],ymm9[13],ymm10[14],ymm9[14],ymm10[15],ymm9[15]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm7, %ymm26, %ymm7
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm11 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} ymm12 = ymm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm10[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm12[0],ymm11[0],ymm12[1],ymm11[1],ymm12[2],ymm11[2],ymm12[3],ymm11[3],ymm12[8],ymm11[8],ymm12[9],ymm11[9],ymm12[10],ymm11[10],ymm12[11],ymm11[11]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm7, %zmm11, %zmm2 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermi2d %ymm7, %ymm2, %ymm22
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm9, %ymm7
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm7, %zmm29, %zmm2
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm7, %ymm11
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm7, %ymm7
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm31, %zmm7
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm9, %xmm11
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsp), %xmm10 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm10, %xmm12
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm12, %zmm18, %zmm11
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm12, %ymm30, %ymm12
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm13 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm14 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm13, %ymm13
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm13, %zmm31, %zmm12
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm10, %xmm14
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm9, %xmm8
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm14 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm14, %zmm18, %zmm8
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm13, %zmm12, %zmm11 {%k2}
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm12 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm9[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm11, %ymm14
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm13, %zmm17, %zmm11
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm13, %ymm30, %ymm13
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm15 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm10 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm15[0],xmm10[1],xmm15[1],xmm10[2],xmm15[2],xmm10[3],xmm15[3]
-; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm10, %ymm10
-; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm10, %zmm13, %zmm8 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermt2d %ymm12, %ymm0, %ymm14
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero
-; AVX512DQ-FCP-NEXT:    vpermi2d %ymm10, %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm9[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm10, %zmm17, %zmm8
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm22, %zmm2
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm14, %zmm10
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm13 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm9, %xmm11
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm9[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm16, %zmm12
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm9, %xmm8
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm9[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm16, %zmm11
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm11 = zmm11 ^ (zmm8 & (zmm11 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm13
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm4, %xmm1
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm13[4],xmm1[5],xmm13[5],xmm1[6],xmm13[6],xmm1[7],xmm13[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm4[0],xmm10[0],xmm4[1],xmm10[1],xmm4[2],xmm10[2],xmm4[3],xmm10[3]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm27, %zmm13
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm5, %ymm28, %ymm15
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm15, %zmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm1, %xmm12
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm12, %zmm17, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm4[0],xmm15[0],xmm4[1],xmm15[1],xmm4[2],xmm15[2],xmm4[3],xmm15[3]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm12, %ymm7, %ymm7
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm12 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpsrldq {{.*#+}} xmm15 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm15[0],xmm12[0],xmm15[1],xmm12[1],xmm15[2],xmm12[2],xmm15[3],xmm12[3]
+; AVX512DQ-FCP-NEXT:    vpbroadcastq %xmm12, %ymm12
+; AVX512DQ-FCP-NEXT:    vinserti32x8 $1, %ymm12, %zmm7, %zmm13 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpmovzxwd {{.*#+}} xmm7 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm7, %ymm13, %ymm28
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2d %zmm7, %zmm31, %zmm13
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm28, %zmm7
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm4[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm17, %zmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm4 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # zmm4 = zmm4 ^ (zmm2 & (zmm4 ^ mem))
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, (%rax)
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm12 = zmm12 ^ (zmm8 & (zmm12 ^ zmm10))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, 192(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm0 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (zmm0 & (zmm7 ^ zmm2))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 128(%rax)
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (zmm0 & (zmm6 ^ zmm27))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 320(%rax)
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm24, %zmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (zmm6 & (zmm7 ^ zmm2))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 256(%rax)
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm28, %zmm20, %zmm2
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (zmm6 & (zmm7 ^ zmm2))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 448(%rax)
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (zmm8 & (zmm5 ^ zmm23))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 384(%rax)
-; AVX512DQ-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm4 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    # zmm4 = zmm4 ^ (zmm8 & (zmm4 ^ mem))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 576(%rax)
-; AVX512DQ-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    # zmm3 = zmm3 ^ (zmm0 & (zmm3 ^ mem))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 512(%rax)
-; AVX512DQ-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    # zmm1 = zmm1 ^ (zmm0 & (zmm1 ^ mem))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 704(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm25, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm6 & (zmm1 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 640(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 256(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm4 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # zmm4 = zmm4 ^ (zmm2 & (zmm4 ^ mem))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 448(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm4 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # zmm4 = zmm4 ^ (zmm2 & (zmm4 ^ mem))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 640(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm4 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # zmm4 = zmm4 ^ (zmm2 & (zmm4 ^ mem))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 64(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm12 = zmm12 ^ (zmm2 & (zmm12 ^ zmm7))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, (%rax)
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm2 & (zmm1 ^ zmm5))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 192(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm11 = zmm11 ^ (zmm1 & (zmm11 ^ zmm22))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, 128(%rax)
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm9 = zmm9 ^ (zmm1 & (zmm9 ^ zmm24))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 320(%rax)
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm14 = zmm14 ^ (zmm2 & (zmm14 ^ zmm30))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, 384(%rax)
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (zmm2 & (zmm3 ^ zmm20))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 576(%rax)
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm1 & (zmm0 ^ zmm21))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 512(%rax)
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm21, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm6 & (zmm1 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rax)
-; AVX512DQ-FCP-NEXT:    addq $1176, %rsp # imm = 0x498
+; AVX512DQ-FCP-NEXT:    vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # zmm0 = zmm0 ^ (zmm1 & (zmm0 ^ mem))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 704(%rax)
+; AVX512DQ-FCP-NEXT:    addq $1112, %rsp # imm = 0x458
 ; 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 3f8f061f359f9..2aab84f0b9cc2 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
@@ -259,14 +259,14 @@ define void @store_i16_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512BW-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512BW-NEXT:    vinserti128 $1, (%r9), %ymm1, %ymm1
-; AVX512BW-NEXT:    vinserti128 $1, (%r8), %ymm0, %ymm0
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX512BW-NEXT:    vinserti128 $1, (%r10), %ymm2, %ymm1
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,16,18,8,10,24,1,3,17,19,9,11,25,0,0]
+; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512BW-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,1,3,5,7,17,19,21,0,0]
 ; AVX512BW-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
 ; AVX512BW-NEXT:    vextracti128 $1, %ymm2, %xmm0
 ; AVX512BW-NEXT:    vpextrd $2, %xmm0, 24(%rax)
@@ -280,14 +280,14 @@ define void @store_i16_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r9), %ymm1, %ymm1
-; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r8), %ymm0, %ymm0
-; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r10), %ymm2, %ymm1
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,16,18,8,10,24,1,3,17,19,9,11,25,0,0]
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,1,3,5,7,17,19,21,0,0]
 ; AVX512BW-FCP-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
 ; AVX512BW-FCP-NEXT:    vextracti128 $1, %ymm2, %xmm0
 ; AVX512BW-FCP-NEXT:    vpextrd $2, %xmm0, 24(%rax)
@@ -301,14 +301,14 @@ define void @store_i16_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r9), %ymm1, %ymm1
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r8), %ymm0, %ymm0
-; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r10), %ymm2, %ymm1
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,16,18,8,10,24,1,3,17,19,9,11,25,0,0]
+; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,1,3,5,7,17,19,21,0,0]
 ; AVX512DQ-BW-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
 ; AVX512DQ-BW-NEXT:    vextracti128 $1, %ymm2, %xmm0
 ; AVX512DQ-BW-NEXT:    vpextrd $2, %xmm0, 24(%rax)
@@ -322,14 +322,14 @@ define void @store_i16_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r9), %ymm1, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r8), %ymm0, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r10), %ymm2, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,16,18,8,10,24,1,3,17,19,9,11,25,0,0]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,1,3,5,7,17,19,21,0,0]
 ; AVX512DQ-BW-FCP-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
 ; AVX512DQ-BW-FCP-NEXT:    vextracti128 $1, %ymm2, %xmm0
 ; AVX512DQ-BW-FCP-NEXT:    vpextrd $2, %xmm0, 24(%rax)
@@ -680,13 +680,13 @@ define void @store_i16_stride7_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,ymm0[0,1,4,5,8,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[18,19,22,23,26,27],zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[0,1,8,9],zero,zero,zero,zero,ymm4[u,u,u,u,u,u,2,3],zero,zero,ymm4[18,19,26,27,u,u,u,u,u,u],zero,zero,zero,zero
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,ymm2[0,1,8,9,u,u,u,u,u,u],zero,zero,ymm2[26,27],zero,zero,zero,zero,ymm2[u,u,u,u,u,u,20,21,28,29]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm2 = mem & (ymm2 | ymm1)
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [5,7,1,3,7,0,0,0]
-; AVX512-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm1
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,3,0,1]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,ymm4[0,1,8,9,u,u,u,u,u,u],zero,zero,ymm4[26,27],zero,zero,zero,zero,ymm4[u,u,u,u,u,u,20,21,28,29]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm4 = mem & (ymm4 | ymm1)
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [9,11,1,3,11,0,0,0]
+; AVX512-FCP-NEXT:    vpermi2d %ymm3, %ymm2, %ymm1
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[0,1,4,5],zero,zero,zero,zero,zero,zero,ymm1[10,11,14,15,2,3,18,19],zero,zero,zero,zero,zero,zero,ymm1[u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm4, %zmm1
 ; AVX512-FCP-NEXT:    vporq %zmm0, %zmm1, %zmm0
 ; AVX512-FCP-NEXT:    vextracti32x4 $2, %zmm0, 32(%rax)
 ; AVX512-FCP-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
@@ -762,13 +762,13 @@ define void @store_i16_stride7_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,ymm0[0,1,4,5,8,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[18,19,22,23,26,27],zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[0,1,8,9],zero,zero,zero,zero,ymm4[u,u,u,u,u,u,2,3],zero,zero,ymm4[18,19,26,27,u,u,u,u,u,u],zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,ymm2[0,1,8,9,u,u,u,u,u,u],zero,zero,ymm2[26,27],zero,zero,zero,zero,ymm2[u,u,u,u,u,u,20,21,28,29]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm2 = mem & (ymm2 | ymm1)
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [5,7,1,3,7,0,0,0]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,3,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,ymm4[0,1,8,9,u,u,u,u,u,u],zero,zero,ymm4[26,27],zero,zero,zero,zero,ymm4[u,u,u,u,u,u,20,21,28,29]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm4 = mem & (ymm4 | ymm1)
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [9,11,1,3,11,0,0,0]
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm3, %ymm2, %ymm1
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[0,1,4,5],zero,zero,zero,zero,zero,zero,ymm1[10,11,14,15,2,3,18,19],zero,zero,zero,zero,zero,zero,ymm1[u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm4, %zmm1
 ; AVX512DQ-FCP-NEXT:    vporq %zmm0, %zmm1, %zmm0
 ; AVX512DQ-FCP-NEXT:    vextracti32x4 $2, %zmm0, 32(%rax)
 ; AVX512DQ-FCP-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
@@ -2963,122 +2963,122 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; 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:    vmovdqa (%r8), %ymm0
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %ymm1
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = 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:    vporq %ymm2, %ymm3, %ymm17
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm10
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm11
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = 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 = ymm2[u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[14,15,u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[16,17,u,u],zero,zero
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm1[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm1[u,u,u,u]
-; AVX512-FCP-NEXT:    vporq %ymm0, %ymm3, %ymm18
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,u,u,u,u,26,27,u,u,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm5[2,2,2,2,6,6,6,6]
-; 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:    vporq %ymm2, %ymm3, %ymm20
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm12
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm14
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm1[u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[14,15,u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[16,17,u,u],zero,zero
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = 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:    vporq %ymm2, %ymm3, %ymm21
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm15
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = 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 {{.*#+}} ymm9 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm9[0],ymm3[1],ymm9[2,3],ymm3[4],ymm9[5,6,7,8],ymm3[9],ymm9[10,11],ymm3[12],ymm9[13,14,15]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm4[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm9[2],ymm13[3,4],ymm9[5],ymm13[6,7,8,9],ymm9[10],ymm13[11,12],ymm9[13],ymm13[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 = [2,2,3,3,10,9,11,10]
-; 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:    vpermi2q %zmm3, %zmm13, %zmm9
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm13 = 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 {{.*#+}} ymm3 = ymm13[0,1],ymm3[2],ymm13[3,4],ymm3[5],ymm13[6,7,8,9],ymm3[10],ymm13[11,12],ymm3[13],ymm13[14,15]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = 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 {{.*#+}} ymm8 = ymm8[0,1],ymm13[2],ymm8[3,4],ymm13[5],ymm8[6,7,8,9],ymm13[10],ymm8[11,12],ymm13[13],ymm8[14,15]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm8, %zmm22
+; AVX512-FCP-NEXT:    vprold $16, %ymm1, %ymm3
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm0[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm8 = ymm8[0,1],ymm3[2],ymm8[3,4],ymm3[5],ymm8[6,7,8,9],ymm3[10],ymm8[11,12],ymm3[13],ymm8[14,15]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm13 = 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 {{.*#+}} ymm13 = 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 {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm16 = [2,2,3,3,10,9,11,10]
+; AVX512-FCP-NEXT:    vpermi2q %zmm8, %zmm13, %zmm16
+; AVX512-FCP-NEXT:    vprold $16, %xmm11, %xmm8
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm10[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 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm13 = xmm13[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [0,0,1,0,0,8,0,9]
+; AVX512-FCP-NEXT:    vpermi2q %zmm8, %zmm13, %zmm18
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm13 = 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 = [0,0,0,1,8,9,9,11]
-; 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:    vpmovsxbq {{.*#+}} zmm19 = [0,0,0,1,8,9,9,11]
+; AVX512-FCP-NEXT:    vpermi2q %zmm13, %zmm8, %zmm19
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = xmm12[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm14[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm13[0],xmm8[1],xmm13[2,3],xmm8[4],xmm13[5,6],xmm8[7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm13, %xmm13
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm13, %zmm8
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm15[4],xmm2[5],xmm15[5],xmm2[6],xmm15[6],xmm2[7],xmm15[7]
+; AVX512-FCP-NEXT:    vmovdqa (%rax), %ymm12
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
+; AVX512-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm3
+; AVX512-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm3
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm13 = ymm12[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm12[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm12[16,17,u,u]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm3, %zmm3
 ; 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:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
 ; AVX512-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
-; AVX512-FCP-NEXT:    vpermd %zmm8, %zmm6, %zmm6
+; AVX512-FCP-NEXT:    vpermd %zmm12, %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:    vpermd %ymm12, %ymm7, %ymm7
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm11 = xmm11[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,1,3]
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm10[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[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:    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:    movq {{[0-9]+}}(%rsp), %rcx
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm16, %zmm3, %zmm2
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} 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:    vpternlogd {{.*#+}} zmm0 = zmm9 ^ (mem & (zmm0 ^ zmm9))
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} 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:    vpternlogd {{.*#+}} zmm0 = zmm15 ^ (mem & (zmm0 ^ zmm15))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm11, %zmm1
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm20, %zmm10, %zmm10
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm1))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm21, %zmm2, %zmm1
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm1 & mem)
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm10))
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm22[2,2,2,3,6,6,6,7]
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm9 ^ (mem & (zmm1 ^ zmm9))
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm16))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm1))
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm8[0,0,1,1,4,4,5,5]
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm18 ^ (mem & (zmm1 ^ zmm18))
 ; AVX512-FCP-NEXT:    vpbroadcastd (%rax), %ymm2
-; AVX512-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm3
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm20))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
+; AVX512-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm8
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm2, %zmm2
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm19))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
 ; 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:    vpternlogq {{.*#+}} ymm0 = ymm7 ^ (mem & (ymm0 ^ ymm7))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm0 = ymm0 ^ (mem & (ymm0 ^ ymm4))
+; AVX512-FCP-NEXT:    vmovdqa %ymm0, 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:    vmovdqa64 %zmm3, 64(%rcx)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -3225,122 +3225,122 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; 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:    vmovdqa (%r8), %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = 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:    vporq %ymm2, %ymm3, %ymm17
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm10
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm11
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = 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 = ymm2[u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[14,15,u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[16,17,u,u],zero,zero
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm1[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm1[u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vporq %ymm0, %ymm3, %ymm18
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,u,u,u,u,26,27,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm5[2,2,2,2,6,6,6,6]
-; 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:    vporq %ymm2, %ymm3, %ymm20
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm14
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm1[u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[14,15,u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[16,17,u,u],zero,zero
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = 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:    vporq %ymm2, %ymm3, %ymm21
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm15
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = 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 {{.*#+}} ymm9 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm9[0],ymm3[1],ymm9[2,3],ymm3[4],ymm9[5,6,7,8],ymm3[9],ymm9[10,11],ymm3[12],ymm9[13,14,15]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm4[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm9[2],ymm13[3,4],ymm9[5],ymm13[6,7,8,9],ymm9[10],ymm13[11,12],ymm9[13],ymm13[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 = [2,2,3,3,10,9,11,10]
-; 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:    vpermi2q %zmm3, %zmm13, %zmm9
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm13 = 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 {{.*#+}} ymm3 = ymm13[0,1],ymm3[2],ymm13[3,4],ymm3[5],ymm13[6,7,8,9],ymm3[10],ymm13[11,12],ymm3[13],ymm13[14,15]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = 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 {{.*#+}} ymm8 = ymm8[0,1],ymm13[2],ymm8[3,4],ymm13[5],ymm8[6,7,8,9],ymm13[10],ymm8[11,12],ymm13[13],ymm8[14,15]
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm8, %zmm22
+; AVX512DQ-FCP-NEXT:    vprold $16, %ymm1, %ymm3
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm0[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm8 = ymm8[0,1],ymm3[2],ymm8[3,4],ymm3[5],ymm8[6,7,8,9],ymm3[10],ymm8[11,12],ymm3[13],ymm8[14,15]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm13 = 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 {{.*#+}} ymm13 = 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 {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm16 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm8, %zmm13, %zmm16
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm11, %xmm8
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm10[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 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm13 = xmm13[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [0,0,1,0,0,8,0,9]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm8, %zmm13, %zmm18
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm13 = 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 = [0,0,0,1,8,9,9,11]
-; 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:    vpmovsxbq {{.*#+}} zmm19 = [0,0,0,1,8,9,9,11]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm13, %zmm8, %zmm19
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = xmm12[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm14[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm13[0],xmm8[1],xmm13[2,3],xmm8[4],xmm13[5,6],xmm8[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm13, %xmm13
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm13, %zmm8
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm15[4],xmm2[5],xmm15[5],xmm2[6],xmm15[6],xmm2[7],xmm15[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %ymm12
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm3
+; AVX512DQ-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm13 = ymm12[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm12[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm12[16,17,u,u]
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm3, %zmm3
 ; 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:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
 ; AVX512DQ-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm8, %zmm6, %zmm6
+; AVX512DQ-FCP-NEXT:    vpermd %zmm12, %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:    vpermd %ymm12, %ymm7, %ymm7
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm11 = xmm11[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,1,3]
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm10[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[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:    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:    movq {{[0-9]+}}(%rsp), %rcx
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm16, %zmm3, %zmm2
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} 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:    vpternlogd {{.*#+}} zmm0 = zmm9 ^ (mem & (zmm0 ^ zmm9))
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} 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:    vpternlogd {{.*#+}} zmm0 = zmm15 ^ (mem & (zmm0 ^ zmm15))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm11, %zmm1
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm20, %zmm10, %zmm10
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm1))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm21, %zmm2, %zmm1
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm1 & mem)
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm10))
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm22[2,2,2,3,6,6,6,7]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm9 ^ (mem & (zmm1 ^ zmm9))
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm16))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm1))
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm8[0,0,1,1,4,4,5,5]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm18 ^ (mem & (zmm1 ^ zmm18))
 ; AVX512DQ-FCP-NEXT:    vpbroadcastd (%rax), %ymm2
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm3
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm20))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm8
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm2, %zmm2
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm19))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
 ; 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:    vpternlogq {{.*#+}} ymm0 = ymm7 ^ (mem & (ymm0 ^ ymm7))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm0 = ymm0 ^ (mem & (ymm0 ^ ymm4))
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm0, 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:    vmovdqa64 %zmm3, 64(%rcx)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
 ;
@@ -6202,286 +6202,289 @@ 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 $264, %rsp # imm = 0x108
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm10
+; AVX512-FCP-NEXT:    subq $328, %rsp # imm = 0x148
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm11
 ; 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, %ymm10, %ymm1
-; AVX512-FCP-NEXT:    vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm2
+; AVX512-FCP-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm4
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm2 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm4, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm4, %ymm18
+; AVX512-FCP-NEXT:    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:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpor %ymm1, %ymm3, %ymm1
-; AVX512-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm9
+; 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), %ymm10
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm13
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm1 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm9, %ymm3
-; AVX512-FCP-NEXT:    vmovdqu %ymm9, (%rsp) # 32-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm6
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm5 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm6, %ymm4
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm6, %ymm17
-; AVX512-FCP-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
-; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; 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, %ymm10, %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm10, (%rsp) # 32-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm12
+; 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, %ymm12, %ymm5
+; AVX512-FCP-NEXT:    vmovdqu %ymm12, {{[-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 (%r9), %ymm15
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm15, %ymm4
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm11
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} 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, %ymm4
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm6
 ; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm11, %ymm8
-; AVX512-FCP-NEXT:    vpor %ymm4, %ymm8, %ymm3
-; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm13, %ymm1
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm12
-; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm12, %ymm5
-; AVX512-FCP-NEXT:    vpor %ymm1, %ymm5, %ymm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm8
-; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm8, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm14
-; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm14, %ymm1
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm6, %ymm8
+; AVX512-FCP-NEXT:    vpor %ymm4, %ymm8, %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm9
+; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm9, %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), %ymm14
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm14, %ymm0
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm2
+; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm16
 ; 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 %ymm6, %ymm1, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm20
+; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %ymm0
+; AVX512-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm0, %ymm0
 ; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm2
 ; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm1
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm16
-; AVX512-FCP-NEXT:    vporq %ymm0, %ymm1, %ymm22
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm24
+; 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:    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:    vpshufb {{.*#+}} ymm2 = ymm13[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,10,0,11,0]
-; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm24
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm10[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm18[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    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:    vpmovsxbq {{.*#+}} zmm22 = [0,2,2,3,10,0,11,0]
+; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm22
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm11[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = 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 32(%rcx), %xmm8
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm5
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm8[0],xmm5[1],xmm8[1],xmm5[2],xmm8[2],xmm5[3],xmm8[3]
 ; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
 ; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
 ; AVX512-FCP-NEXT:    vmovdqa64 %xmm3, %xmm27
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm25 = [0,2,2,3,8,0,9,0]
-; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm25
-; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm9, %ymm0
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm1 = ymm17[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm19 = [0,2,2,3,8,0,9,0]
+; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm19
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm10, %ymm0
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm1 = ymm12[3,3,3,3,7,7,7,7]
 ; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm4 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm10
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm4
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm2
 ; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm21 = [2,1,3,3,8,8,9,9]
 ; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm0, %zmm21
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm12 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
 ; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm23 = [2,2,2,3,8,8,8,9]
 ; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm2
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm7
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm5 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm6
-; AVX512-FCP-NEXT:    vpermi2q %zmm6, %zmm0, %zmm23
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm10[4],xmm1[4],xmm10[5],xmm1[5],xmm10[6],xmm1[6],xmm10[7],xmm1[7]
-; AVX512-FCP-NEXT:    vprold $16, %xmm10, %xmm6
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm7
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm11, %xmm17
+; AVX512-FCP-NEXT:    vpermi2q %zmm7, %zmm12, %zmm23
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
+; AVX512-FCP-NEXT:    vprold $16, %xmm4, %xmm4
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm6[2],xmm1[3,4],xmm6[5],xmm1[6,7]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm0
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm4[2],xmm1[3,4],xmm4[5],xmm1[6,7]
+; 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, %xmm7, %xmm4
 ; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,8,9,0]
-; AVX512-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm26
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm2[4],xmm7[5],xmm2[5],xmm7[6],xmm2[6],xmm7[7],xmm2[7]
+; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm1, %zmm26
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm1
+; 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, %xmm0, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm1
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm29 = [0,1,1,3,8,8,9,9]
-; AVX512-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm29
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm2
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm28 = [0,1,1,3,8,8,9,9]
+; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm28
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm1
 ; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm0
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm2[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm3 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm2
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm3, %xmm20
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1],xmm3[2,3],xmm1[4],xmm3[5,6],xmm1[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm27, %xmm4
-; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3],xmm2[4],xmm3[5,6],xmm2[7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm27, %xmm7
+; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm3
 ; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm27 = [0,0,1,0,8,8,9,9]
-; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm3, %zmm27
-; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm6
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm9
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
-; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm28 = [0,0,0,1,8,9,9,11]
-; AVX512-FCP-NEXT:    vpermi2q %zmm7, %zmm3, %zmm28
+; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm27
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm12
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm12[0],xmm2[1],xmm12[1],xmm2[2],xmm12[2],xmm2[3],xmm12[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm17, %xmm7
+; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm3
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm29 = [0,0,0,1,8,9,9,11]
+; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm29
 ; AVX512-FCP-NEXT:    vprold $16, %ymm15, %ymm3
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm11[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm6[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 {{.*#+}} ymm4 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm15, %ymm5
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm7 = ymm11[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm31 = [2,2,3,3,10,9,11,10]
-; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm31
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm0, %xmm19
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm5
+; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm15, %ymm7
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm6[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3],ymm7[4,5],ymm6[6],ymm7[7,8,9,10],ymm6[11],ymm7[12,13],ymm6[14],ymm7[15]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm30 = [2,2,3,3,10,9,11,10]
+; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm6, %zmm30
+; 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, %xmm18
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm0
 ; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm3
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
-; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm18
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm1
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm17
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm0
-; AVX512-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa (%rax), %ymm10
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm1 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm10, %ymm2
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm30
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm12[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm6[2],ymm0[3,4],ymm6[5],ymm0[6,7,8,9],ymm6[10],ymm0[11,12],ymm6[13],ymm0[14,15]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm17
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm14[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm11 = ymm6[0,1],ymm0[2],ymm6[3,4],ymm0[5],ymm6[6,7,8,9],ymm0[10],ymm6[11,12],ymm0[13],ymm6[14,15]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm20, %ymm13
-; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm13, %ymm0
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm16[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm4[3],ymm0[4,5],ymm4[6],ymm0[7,8,9,10],ymm4[11],ymm0[12,13],ymm4[14],ymm0[15]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm9 = [0,1,0,3,10,10,11,11]
-; AVX512-FCP-NEXT:    vpermi2q %zmm0, %zmm22, %zmm9
-; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %ymm2
-; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm22 = [0,1,4,5,4,5,5,7,12,13,10,10,14,14,14,14]
-; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm22, %ymm4
-; AVX512-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm4
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm8[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm12[0],ymm4[1],ymm12[2,3],ymm4[4],ymm12[5,6,7,8],ymm4[9],ymm12[10,11],ymm4[12],ymm12[13,14,15]
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512-FCP-NEXT:    # ymm6 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm8, %ymm12
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm14[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm8[2],ymm12[3,4],ymm8[5],ymm12[6,7,8,9],ymm8[10],ymm12[11,12],ymm8[13],ymm12[14,15]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm20, %zmm12
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm16[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm8[2],ymm4[3,4],ymm8[5],ymm4[6,7,8,9],ymm8[10],ymm4[11,12],ymm8[13],ymm4[14,15]
-; AVX512-FCP-NEXT:    vprold $16, %ymm13, %ymm8
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm15 = ymm16[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm15 = ymm15[0,1],ymm8[2],ymm15[3,4],ymm8[5],ymm15[6,7,8,9],ymm8[10],ymm15[11,12],ymm8[13],ymm15[14,15]
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = xmm0[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm14 = xmm13[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm14[0],xmm8[1],xmm14[2,3],xmm8[4],xmm14[5,6],xmm8[7]
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm14 = [0,0,1,1,8,8,10,9]
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm14, %zmm8
-; AVX512-FCP-NEXT:    vprold $16, %xmm3, %xmm0
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm5[1,1,2,3]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1],xmm0[2],xmm13[3,4],xmm0[5],xmm13[6,7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm5 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm7, %xmm0
-; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
-; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm14, %zmm3
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm5 = [2,1,3,2,10,10,10,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm15
+; AVX512-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm1
+; AVX512-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa (%rax), %ymm15
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm15, %ymm11
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm1, %zmm31
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm2[4],xmm12[4],xmm2[5],xmm12[5],xmm2[6],xmm12[6],xmm2[7],xmm12[7]
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [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:    # ymm2 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm25
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm9[1,1,1,1,5,5,5,5]
+; 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:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm7, %ymm2
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm24[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm2[0,1,2],ymm4[3],ymm2[4,5],ymm4[6],ymm2[7,8,9,10],ymm4[11],ymm2[12,13],ymm4[14],ymm2[15]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm2 = [0,1,0,3,10,10,11,11]
+; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
+; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm9, %zmm2
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm14[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm9 = ymm16[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm9[0,1],ymm4[2],ymm9[3,4],ymm4[5],ymm9[6,7,8,9],ymm4[10],ymm9[11,12],ymm4[13],ymm9[14,15]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm20, %xmm9
+; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm8, %xmm9
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm8 = xmm5[4],xmm8[4],xmm5[5],xmm8[5],xmm5[6],xmm8[6],xmm5[7],xmm8[7]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0],xmm9[1],xmm5[2,3],xmm9[4],xmm5[5,6],xmm9[7]
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [0,0,1,1,8,8,10,9]
+; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm20, %zmm5
+; AVX512-FCP-NEXT:    vprold $16, %xmm3, %xmm8
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm0[1,1,2,3]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm11, %xmm11
+; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm8
+; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm20, %zmm8
+; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm3
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm20 = [0,1,4,5,4,5,5,7,4,5,2,2,6,6,6,6]
+; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm20, %ymm9
+; AVX512-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm9
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm3, %zmm3
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = 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 {{.*#+}} ymm12 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3],ymm9[4],ymm12[5,6,7,8],ymm9[9],ymm12[10,11],ymm9[12],ymm12[13,14,15]
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512-FCP-NEXT:    # ymm12 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm14, %ymm6
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm16[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm13[2],ymm6[3,4],ymm13[5],ymm6[6,7,8,9],ymm13[10],ymm6[11,12],ymm13[13],ymm6[14,15]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm16 = [2,2,2,3,8,10,10,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm16, %zmm6
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm7[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 {{.*#+}} ymm14 = ymm24[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm14[2],ymm9[3,4],ymm14[5],ymm9[6,7,8,9],ymm14[10],ymm9[11,12],ymm14[13],ymm9[14,15]
+; AVX512-FCP-NEXT:    vprold $16, %ymm7, %ymm14
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm10 = ymm24[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm10 = ymm10[0,1],ymm14[2],ymm10[3,4],ymm14[5],ymm10[6,7,8,9],ymm14[10],ymm10[11,12],ymm14[13],ymm10[14,15]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm14 = [2,1,3,2,10,10,10,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm14, %zmm10
 ; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm7, %ymm4
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm7, %ymm9
 ; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm13[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7,8,9],ymm6[10],ymm4[11,12],ymm6[13],ymm4[14,15]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm13[2,2,2,2,6,6,6,6]
+; 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:    vpunpcklwd {{.*#+}} ymm12 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm7 = ymm13[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm7[0,1],ymm6[2],ymm7[3,4],ymm6[5],ymm7[6,7,8,9],ymm6[10],ymm7[11,12],ymm6[13],ymm7[14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm6
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm12 = zmm24 ^ (zmm16 & (zmm12 ^ zmm24))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm12))
-; AVX512-FCP-NEXT:    vmovdqa64 (%rax), %zmm5
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,5,2,2,6,6,6,6,30,31,27,27,31,31,30,31]
-; AVX512-FCP-NEXT:    vpermi2d %zmm5, %zmm2, %zmm7
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm15))
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm26 ^ (zmm2 & (zmm8 ^ zmm26))
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm12[2],ymm7[3,4],ymm12[5],ymm7[6,7,8,9],ymm12[10],ymm7[11,12],ymm12[13],ymm7[14,15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm14, %zmm7
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [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 {{.*#+}} zmm6 = zmm22 ^ (zmm9 & (zmm6 ^ zmm22))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm6))
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [4,5,2,2,6,6,6,6,6,7,3,3,7,7,6,7]
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm6, %zmm0
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm10))
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm6 = [6,0,0,0,7,0,0,7]
+; AVX512-FCP-NEXT:    vpermd %ymm15, %ymm6, %ymm6
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [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 = zmm26 ^ (zmm10 & (zmm5 ^ zmm26))
 ; AVX512-FCP-NEXT:    vpbroadcastd 36(%rax), %ymm12
-; AVX512-FCP-NEXT:    vpbroadcastd 40(%rax), %ymm13
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm12, %zmm12
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm29))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm8))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm27 = zmm27 ^ (mem & (zmm27 ^ zmm3))
-; AVX512-FCP-NEXT:    vmovdqu (%rsp), %ymm4 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm4[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm24[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm8[0],ymm3[1],ymm8[2,3],ymm3[4],ymm8[5,6,7,8],ymm3[9],ymm8[10,11],ymm3[12],ymm8[13,14,15]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm19, %xmm8
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm13 = ymm18[0,0,1,3]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm14 = ymm17[2,2,2,3]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm15 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm24[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm15[0,1],ymm4[2],ymm15[3,4],ymm4[5],ymm15[6,7,8,9],ymm4[10],ymm15[11,12],ymm4[13],ymm15[14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm20, %zmm4
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm6))
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [6,0,0,0,7,0,0,7]
-; AVX512-FCP-NEXT:    vpermd %ymm10, %ymm3, %ymm3
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm5, %zmm5
-; AVX512-FCP-NEXT:    vpermd %zmm5, %zmm22, %zmm5
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm31))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ 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), %zmm13, %zmm6 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm4 ^ (zmm2 & (zmm6 ^ zmm4))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm21 = zmm25 ^ (zmm16 & (zmm21 ^ zmm25))
-; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = (zmm0 & mem) | zmm30
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm6))
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm14, %zmm2, %zmm2
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm4, %zmm4
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm2))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm9 & mem)
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm4))
-; AVX512-FCP-NEXT:    vpbroadcastd 32(%rax), %ymm2
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm23))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm21))
-; AVX512-FCP-NEXT:    vpbroadcastd (%rax), %ymm3
+; AVX512-FCP-NEXT:    vpbroadcastd 40(%rax), %ymm14
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm14, %zmm12, %zmm12
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm28))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm5))
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm18, %xmm5
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm14 = ymm17[0,0,1,3]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,1,1]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,1,3,2]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm27 = zmm27 ^ (mem & (zmm27 ^ zmm8))
+; AVX512-FCP-NEXT:    vmovdqu (%rsp), %ymm13 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = 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:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm15 = ymm17[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 %ymm25, %ymm15
+; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm13, %ymm15
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm17[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 %zmm8, %zmm16, %zmm13
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm7))
+; AVX512-FCP-NEXT:    vpermd (%rax), %zmm20, %zmm7
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm30))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm13))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm8 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm5 ^ (zmm10 & (zmm8 ^ zmm5))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm21 = zmm19 ^ (zmm9 & (zmm21 ^ zmm19))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm5 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = (zmm5 & mem) | zmm31
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm8))
+; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm8, %zmm1
+; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm8, %zmm4
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm1))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm2 & mem)
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm4))
+; AVX512-FCP-NEXT:    vpbroadcastd 32(%rax), %ymm1
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm6, %zmm1
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm23))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm21))
+; AVX512-FCP-NEXT:    vpbroadcastd (%rax), %ymm2
 ; AVX512-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm4
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm28))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm27))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm2
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm29))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm27))
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 128(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, (%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 320(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 128(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 320(%rax)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, 256(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 384(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 64(%rax)
-; AVX512-FCP-NEXT:    addq $264, %rsp # imm = 0x108
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 192(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 384(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rax)
+; AVX512-FCP-NEXT:    addq $328, %rsp # imm = 0x148
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -6826,286 +6829,289 @@ 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 $264, %rsp # imm = 0x108
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm10
+; AVX512DQ-FCP-NEXT:    subq $328, %rsp # imm = 0x148
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm11
 ; 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, %ymm10, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm4
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm2 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm4, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm4, %ymm18
+; AVX512DQ-FCP-NEXT:    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:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpor %ymm1, %ymm3, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm9
+; 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), %ymm10
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm13
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm1 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm9, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm9, (%rsp) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm6
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm5 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm6, %ymm4
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm6, %ymm17
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; 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, %ymm10, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm10, (%rsp) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm12
+; 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, %ymm12, %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm12, {{[-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 (%r9), %ymm15
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm15, %ymm4
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm11
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} 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, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm6
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm11, %ymm8
-; AVX512DQ-FCP-NEXT:    vpor %ymm4, %ymm8, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm13, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm12
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm12, %ymm5
-; AVX512DQ-FCP-NEXT:    vpor %ymm1, %ymm5, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm8
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm14
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm14, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm6, %ymm8
+; AVX512DQ-FCP-NEXT:    vpor %ymm4, %ymm8, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm9
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm9, %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), %ymm14
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm14, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm16
 ; 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 %ymm6, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm20
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm0, %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm2
 ; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm16
-; AVX512DQ-FCP-NEXT:    vporq %ymm0, %ymm1, %ymm22
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm24
+; 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:    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:    vpshufb {{.*#+}} ymm2 = ymm13[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,10,0,11,0]
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm24
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm10[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm18[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    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:    vpmovsxbq {{.*#+}} zmm22 = [0,2,2,3,10,0,11,0]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm22
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm11[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = 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 32(%rcx), %xmm8
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm5
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm8[0],xmm5[1],xmm8[1],xmm5[2],xmm8[2],xmm5[3],xmm8[3]
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
 ; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm3, %xmm27
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm25 = [0,2,2,3,8,0,9,0]
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm25
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm1 = ymm17[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm19 = [0,2,2,3,8,0,9,0]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm19
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm10, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm1 = ymm12[3,3,3,3,7,7,7,7]
 ; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm4 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm10
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm4
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm2
 ; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm21 = [2,1,3,3,8,8,9,9]
 ; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm0, %zmm21
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm12 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
 ; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm23 = [2,2,2,3,8,8,8,9]
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %xmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm7
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm5 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm6
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm6, %zmm0, %zmm23
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm10[4],xmm1[4],xmm10[5],xmm1[5],xmm10[6],xmm1[6],xmm10[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm10, %xmm6
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm7
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm11, %xmm17
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm7, %zmm12, %zmm23
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm4, %xmm4
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm6[2],xmm1[3,4],xmm6[5],xmm1[6,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm4[2],xmm1[3,4],xmm4[5],xmm1[6,7]
+; 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, %xmm7, %xmm4
 ; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,8,9,0]
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm26
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm2[4],xmm7[5],xmm2[5],xmm7[6],xmm2[6],xmm7[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm1, %zmm26
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm1
+; 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, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm1
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm29 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm29
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm2
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm28 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm28
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm0
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm2[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm3 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm3, %xmm20
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1],xmm3[2,3],xmm1[4],xmm3[5,6],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm27, %xmm4
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3],xmm2[4],xmm3[5,6],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm27, %xmm7
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm3
 ; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm27 = [0,0,1,0,8,8,9,9]
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm3, %zmm27
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm9
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm28 = [0,0,0,1,8,9,9,11]
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm7, %zmm3, %zmm28
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm27
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm12[0],xmm2[1],xmm12[1],xmm2[2],xmm12[2],xmm2[3],xmm12[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm17, %xmm7
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm29 = [0,0,0,1,8,9,9,11]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm29
 ; AVX512DQ-FCP-NEXT:    vprold $16, %ymm15, %ymm3
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm11[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm6[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 {{.*#+}} ymm4 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm15, %ymm5
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm7 = ymm11[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm31 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm31
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm0, %xmm19
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm5
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm15, %ymm7
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm6[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3],ymm7[4,5],ymm6[6],ymm7[7,8,9,10],ymm6[11],ymm7[12,13],ymm6[14],ymm7[15]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm30 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm6, %zmm30
+; 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, %xmm18
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm3
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm18
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm17
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm0
-; AVX512DQ-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %ymm10
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm1 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm10, %ymm2
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm30
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm12[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm6[2],ymm0[3,4],ymm6[5],ymm0[6,7,8,9],ymm6[10],ymm0[11,12],ymm6[13],ymm0[14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm17
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm14[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm11 = ymm6[0,1],ymm0[2],ymm6[3,4],ymm0[5],ymm6[6,7,8,9],ymm0[10],ymm6[11,12],ymm0[13],ymm6[14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm20, %ymm13
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm16[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm4[3],ymm0[4,5],ymm4[6],ymm0[7,8,9,10],ymm4[11],ymm0[12,13],ymm4[14],ymm0[15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm9 = [0,1,0,3,10,10,11,11]
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm0, %zmm22, %zmm9
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm22 = [0,1,4,5,4,5,5,7,12,13,10,10,14,14,14,14]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm2, %ymm22, %ymm4
-; AVX512DQ-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm8[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm12[0],ymm4[1],ymm12[2,3],ymm4[4],ymm12[5,6,7,8],ymm4[9],ymm12[10,11],ymm4[12],ymm12[13,14,15]
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512DQ-FCP-NEXT:    # ymm6 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm8, %ymm12
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm14[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm8[2],ymm12[3,4],ymm8[5],ymm12[6,7,8,9],ymm8[10],ymm12[11,12],ymm8[13],ymm12[14,15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm20, %zmm12
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm16[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm8[2],ymm4[3,4],ymm8[5],ymm4[6,7,8,9],ymm8[10],ymm4[11,12],ymm8[13],ymm4[14,15]
-; AVX512DQ-FCP-NEXT:    vprold $16, %ymm13, %ymm8
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm15 = ymm16[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm15 = ymm15[0,1],ymm8[2],ymm15[3,4],ymm8[5],ymm15[6,7,8,9],ymm8[10],ymm15[11,12],ymm8[13],ymm15[14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = xmm0[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm14 = xmm13[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm14[0],xmm8[1],xmm14[2,3],xmm8[4],xmm14[5,6],xmm8[7]
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm14 = [0,0,1,1,8,8,10,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm14, %zmm8
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm3, %xmm0
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm13 = xmm5[1,1,2,3]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1],xmm0[2],xmm13[3,4],xmm0[5],xmm13[6,7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm5 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm7, %xmm0
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm14, %zmm3
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm5 = [2,1,3,2,10,10,10,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm15
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm1
+; AVX512DQ-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %ymm15
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm15, %ymm11
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm1, %zmm31
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm2[4],xmm12[4],xmm2[5],xmm12[5],xmm2[6],xmm12[6],xmm2[7],xmm12[7]
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [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:    # ymm2 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm25
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm9[1,1,1,1,5,5,5,5]
+; 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:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm7, %ymm4
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm7, %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm24[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm2[0,1,2],ymm4[3],ymm2[4,5],ymm4[6],ymm2[7,8,9,10],ymm4[11],ymm2[12,13],ymm4[14],ymm2[15]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm2 = [0,1,0,3,10,10,11,11]
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm9, %zmm2
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm14[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm9 = ymm16[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm9[0,1],ymm4[2],ymm9[3,4],ymm4[5],ymm9[6,7,8,9],ymm4[10],ymm9[11,12],ymm4[13],ymm9[14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm20, %xmm9
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm8, %xmm9
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm8 = xmm5[4],xmm8[4],xmm5[5],xmm8[5],xmm5[6],xmm8[6],xmm5[7],xmm8[7]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0],xmm9[1],xmm5[2,3],xmm9[4],xmm5[5,6],xmm9[7]
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [0,0,1,1,8,8,10,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm20, %zmm5
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm3, %xmm8
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm0[1,1,2,3]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm11, %xmm11
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm8
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm20, %zmm8
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm3
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm20 = [0,1,4,5,4,5,5,7,4,5,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm20, %ymm9
+; AVX512DQ-FCP-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = 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 {{.*#+}} ymm12 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3],ymm9[4],ymm12[5,6,7,8],ymm9[9],ymm12[10,11],ymm9[12],ymm12[13,14,15]
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512DQ-FCP-NEXT:    # ymm12 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm14, %ymm6
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm16[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm13[2],ymm6[3,4],ymm13[5],ymm6[6,7,8,9],ymm13[10],ymm6[11,12],ymm13[13],ymm6[14,15]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm16 = [2,2,2,3,8,10,10,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm16, %zmm6
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm7[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 {{.*#+}} ymm14 = ymm24[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm14[2],ymm9[3,4],ymm14[5],ymm9[6,7,8,9],ymm14[10],ymm9[11,12],ymm14[13],ymm9[14,15]
+; AVX512DQ-FCP-NEXT:    vprold $16, %ymm7, %ymm14
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm10 = ymm24[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm10 = ymm10[0,1],ymm14[2],ymm10[3,4],ymm14[5],ymm10[6,7,8,9],ymm14[10],ymm10[11,12],ymm14[13],ymm10[14,15]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm14 = [2,1,3,2,10,10,10,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm14, %zmm10
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm7, %ymm9
 ; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm13[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7,8,9],ymm6[10],ymm4[11,12],ymm6[13],ymm4[14,15]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm13[2,2,2,2,6,6,6,6]
+; 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:    vpunpcklwd {{.*#+}} ymm12 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm7 = ymm13[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm7[0,1],ymm6[2],ymm7[3,4],ymm6[5],ymm7[6,7,8,9],ymm6[10],ymm7[11,12],ymm6[13],ymm7[14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm12 = zmm24 ^ (zmm16 & (zmm12 ^ zmm24))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm12))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rax), %zmm5
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,5,2,2,6,6,6,6,30,31,27,27,31,31,30,31]
-; AVX512DQ-FCP-NEXT:    vpermi2d %zmm5, %zmm2, %zmm7
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm15))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm26 ^ (zmm2 & (zmm8 ^ zmm26))
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm12[2],ymm7[3,4],ymm12[5],ymm7[6,7,8,9],ymm12[10],ymm7[11,12],ymm12[13],ymm7[14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm14, %zmm7
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [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 {{.*#+}} zmm6 = zmm22 ^ (zmm9 & (zmm6 ^ zmm22))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm6))
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [4,5,2,2,6,6,6,6,6,7,3,3,7,7,6,7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm6, %zmm0
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm10))
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm6 = [6,0,0,0,7,0,0,7]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm15, %ymm6, %ymm6
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [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 = zmm26 ^ (zmm10 & (zmm5 ^ zmm26))
 ; AVX512DQ-FCP-NEXT:    vpbroadcastd 36(%rax), %ymm12
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 40(%rax), %ymm13
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm12, %zmm12
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm29))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm8))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm27 = zmm27 ^ (mem & (zmm27 ^ zmm3))
-; AVX512DQ-FCP-NEXT:    vmovdqu (%rsp), %ymm4 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm4[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm24[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm8[0],ymm3[1],ymm8[2,3],ymm3[4],ymm8[5,6,7,8],ymm3[9],ymm8[10,11],ymm3[12],ymm8[13,14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm19, %xmm8
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm13 = ymm18[0,0,1,3]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm14 = ymm17[2,2,2,3]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm15 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm24[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm15[0,1],ymm4[2],ymm15[3,4],ymm4[5],ymm15[6,7,8,9],ymm4[10],ymm15[11,12],ymm4[13],ymm15[14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm20, %zmm4
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm6))
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [6,0,0,0,7,0,0,7]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm10, %ymm3, %ymm3
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm5, %zmm5
-; AVX512DQ-FCP-NEXT:    vpermd %zmm5, %zmm22, %zmm5
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm31))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ 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), %zmm13, %zmm6 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm4 ^ (zmm2 & (zmm6 ^ zmm4))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm21 = zmm25 ^ (zmm16 & (zmm21 ^ zmm25))
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = (zmm0 & mem) | zmm30
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm6))
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm14, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm4, %zmm4
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm2))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm9 & mem)
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm4))
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 32(%rax), %ymm2
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm23))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm21))
-; AVX512DQ-FCP-NEXT:    vpbroadcastd (%rax), %ymm3
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 40(%rax), %ymm14
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm14, %zmm12, %zmm12
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm28))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm5))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm18, %xmm5
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm14 = ymm17[0,0,1,3]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,1,1]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,1,3,2]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm27 = zmm27 ^ (mem & (zmm27 ^ zmm8))
+; AVX512DQ-FCP-NEXT:    vmovdqu (%rsp), %ymm13 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = 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:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm15 = ymm17[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 %ymm25, %ymm15
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm13, %ymm15
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm17[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 %zmm8, %zmm16, %zmm13
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm7))
+; AVX512DQ-FCP-NEXT:    vpermd (%rax), %zmm20, %zmm7
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm30))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm13))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm8 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm5 ^ (zmm10 & (zmm8 ^ zmm5))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm21 = zmm19 ^ (zmm9 & (zmm21 ^ zmm19))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm5 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = (zmm5 & mem) | zmm31
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm8))
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm8, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm8, %zmm4
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm1))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm2 & mem)
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm4))
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 32(%rax), %ymm1
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm6, %zmm1
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm23))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm21))
+; AVX512DQ-FCP-NEXT:    vpbroadcastd (%rax), %ymm2
 ; AVX512DQ-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm4
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm28))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm27))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm2
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm29))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm27))
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 128(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, (%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 320(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 128(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 320(%rax)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, 256(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 384(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 64(%rax)
-; AVX512DQ-FCP-NEXT:    addq $264, %rsp # imm = 0x108
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 192(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 384(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rax)
+; AVX512DQ-FCP-NEXT:    addq $328, %rsp # imm = 0x148
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
 ;
@@ -13236,70 +13242,70 @@ 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 $1496, %rsp # imm = 0x5D8
-; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %ymm3
+; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %ymm4
 ; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm8 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm3, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %ymm1
+; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm0
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %ymm5
 ; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm1, %ymm2
-; AVX512-FCP-NEXT:    vporq %ymm0, %ymm2, %ymm16
-; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %ymm4
+; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm5, %ymm1
+; AVX512-FCP-NEXT:    vporq %ymm0, %ymm1, %ymm16
+; 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, %ymm4, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %ymm5
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm2, %ymm0
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %ymm3
 ; 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, %ymm5, %ymm2
-; AVX512-FCP-NEXT:    vpor %ymm0, %ymm2, %ymm7
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm3, %ymm1
+; AVX512-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm7
 ; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %ymm2
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm26
-; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %ymm12
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm12, %ymm2
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm12, %ymm28
-; AVX512-FCP-NEXT:    vpor %ymm0, %ymm2, %ymm0
+; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %ymm1
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm1, %ymm0
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm23
+; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %ymm13
+; 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, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm13, %ymm25
+; AVX512-FCP-NEXT:    vpor %ymm0, %ymm1, %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 %ymm8, %ymm0, %ymm14
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm31
+; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm0, %ymm13
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm26
 ; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm0, %ymm15
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm21
-; AVX512-FCP-NEXT:    vpor %ymm14, %ymm15, %ymm12
-; AVX512-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm0, %ymm14
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm30
+; AVX512-FCP-NEXT:    vpor %ymm13, %ymm14, %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, %ymm29
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm13
+; 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, %ymm17
-; AVX512-FCP-NEXT:    vpor %ymm14, %ymm15, %ymm12
-; AVX512-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm0, %ymm14
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm27
+; AVX512-FCP-NEXT:    vpor %ymm13, %ymm14, %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 %ymm6, %ymm0, %ymm14
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm13
 ; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm0
 ; AVX512-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm0, %ymm15
-; AVX512-FCP-NEXT:    vpor %ymm14, %ymm15, %ymm12
-; AVX512-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm0, %ymm14
+; AVX512-FCP-NEXT:    vpor %ymm13, %ymm14, %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 %ymm8, %ymm0, %ymm14
+; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm0, %ymm13
 ; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm0
 ; AVX512-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm0, %ymm12
-; AVX512-FCP-NEXT:    vpor %ymm14, %ymm12, %ymm12
-; AVX512-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm0, %ymm14
+; AVX512-FCP-NEXT:    vpor %ymm13, %ymm14, %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, %ymm12
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm13
 ; 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, %ymm14
-; AVX512-FCP-NEXT:    vpor %ymm12, %ymm14, %ymm12
-; AVX512-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpor %ymm13, %ymm14, %ymm13
+; AVX512-FCP-NEXT:    vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm14
 ; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm14, %ymm8
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm15
@@ -13308,138 +13314,136 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm0
 ; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm8
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm2
-; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm2, %ymm9
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm1
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm1, %ymm9
 ; AVX512-FCP-NEXT:    vpor %ymm8, %ymm9, %ymm8
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm8
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm8, %ymm10
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm8, %ymm10
 ; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %ymm9
 ; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm9, %ymm11
-; AVX512-FCP-NEXT:    vporq %ymm11, %ymm10, %ymm22
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm11 = ymm12[0,1],ymm11[2],ymm12[3,4],ymm11[5],ymm12[6,7,8,9],ymm11[10],ymm12[11,12],ymm11[13],ymm12[14,15]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512-FCP-NEXT:    vporq %ymm11, %ymm10, %ymm17
+; 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 {{.*#+}} ymm12 = ymm5[0,1,1,3,4,5,5,7]
+; 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:    vpermq {{.*#+}} ymm10 = ymm10[2,1,3,2]
 ; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm13 = [18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
 ; AVX512-FCP-NEXT:    # ymm13 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm4, %ymm12
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm13, %ymm30
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm2, %ymm12
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm13, %ymm21
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm3[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, %ymm10, %zmm16, %zmm10
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm7, %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:    vpternlogq {{.*#+}} zmm12 = zmm10 ^ (mem & (zmm12 ^ zmm10))
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
 ; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %ymm7
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm7[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[16,17,u,u,u,u],zero,zero
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm11 = ymm11 | (ymm12 & ymm13)
-; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %ymm10
-; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm10, %ymm6
-; AVX512-FCP-NEXT:    vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm6 = ymm6 | (ymm11 & ymm16)
-; AVX512-FCP-NEXT:    vextracti64x4 $1, %zmm12, %ymm11
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm7[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[16,17,u,u,u,u],zero,zero
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm10 = ymm10 | (ymm12 & ymm18)
+; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %ymm11
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm11, %ymm6
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm11, %ymm16
+; AVX512-FCP-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; 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 {{.*#+}} ymm6 = ymm6 | (ymm10 & ymm11)
+; AVX512-FCP-NEXT:    vextracti64x4 $1, %zmm12, %ymm10
 ; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm12 = [0,4,0,0,0,5,0,0]
 ; AVX512-FCP-NEXT:    vpermd %ymm7, %ymm12, %ymm12
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm12 = ymm12 ^ (ymm16 & (ymm12 ^ ymm11))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm19
-; AVX512-FCP-NEXT:    vprold $16, %ymm10, %ymm11
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm11 = ymm11 ^ (ymm10 & (ymm11 ^ ymm12))
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm0, %zmm11
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm11[4,5,6,7]
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [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 ^ (ymm11 & (ymm12 ^ ymm10))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm20
+; AVX512-FCP-NEXT:    vprold $16, %ymm16, %ymm10
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [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 {{.*#+}} ymm10 = ymm10 ^ (ymm11 & (ymm10 ^ ymm12))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm0, %zmm10
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm10[4,5,6,7]
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm11 = [0,1,4,5,4,5,5,7]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm10 = [0,1,4,5,4,5,5,7]
 ; AVX512-FCP-NEXT:    vmovdqa 96(%rax), %ymm6
-; AVX512-FCP-NEXT:    vpermd %ymm6, %ymm11, %ymm11
-; AVX512-FCP-NEXT:    vpandn %ymm11, %ymm13, %ymm11
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm18
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm6, %ymm12
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm12, %zmm11
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm4, %ymm11
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm5[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm11 = ymm12[0,1,2],ymm11[3],ymm12[4,5],ymm11[6],ymm12[7,8,9,10],ymm11[11],ymm12[12,13],ymm11[14],ymm12[15]
+; AVX512-FCP-NEXT:    vpermd %ymm6, %ymm10, %ymm10
+; AVX512-FCP-NEXT:    vpandnq %ymm10, %ymm18, %ymm10
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm11 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm6, %ymm12
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm12, %zmm10
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm2, %ymm10
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm3[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm10 = ymm12[0,1,2],ymm10[3],ymm12[4,5],ymm10[6],ymm12[7,8,9,10],ymm10[11],ymm12[12,13],ymm10[14],ymm12[15]
 ; AVX512-FCP-NEXT:    vpmovsxdq {{.*#+}} ymm12 = [151522058,0,421010202,421010202]
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm4, %ymm4
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm12, %ymm23
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm5 = [0,2,2,3,10,9,11,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm5, %zmm4
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm2, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm12, %ymm19
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[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 {{.*#+}} zmm3 = [0,2,2,3,10,9,11,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm3, %zmm2
 ; AVX512-FCP-NEXT:    vpmovsxdq {{.*#+}} ymm12 = [218894094,0,488382238,488382238]
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm3, %ymm11
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm4, %ymm10
 ; AVX512-FCP-NEXT:    vmovdqa64 %ymm12, %ymm16
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm11 = ymm12[0],ymm11[1],ymm12[2,3],ymm11[4],ymm12[5,6,7,8],ymm11[9],ymm12[10,11],ymm11[12],ymm12[13,14,15]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm10 = ymm12[0],ymm10[1],ymm12[2,3],ymm10[4],ymm12[5,6,7,8],ymm10[9],ymm12[10,11],ymm10[12],ymm12[13,14,15]
 ; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
 ; AVX512-FCP-NEXT:    # ymm12 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm3, %ymm3
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4],ymm1[5],ymm3[6,7,8,9],ymm1[10],ymm3[11,12],ymm1[13],ymm3[14,15]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm20, %zmm1
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm4 ^ (zmm11 & (zmm1 ^ zmm4))
-; 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 %ymm7, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa64 64(%rax), %zmm1
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,5,0,0,0,6,0,0,30,0,0,0,31,0,0,31]
-; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm6, %zmm3
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm4, %ymm4
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; 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:    vpmovsxbq {{.*#+}} zmm24 = [2,2,2,3,8,10,10,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm24, %zmm4
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm31 = [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 {{.*#+}} zmm4 = zmm2 ^ (zmm31 & (zmm4 ^ zmm2))
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [5,0,0,0,6,0,0,6]
+; AVX512-FCP-NEXT:    vpermd %ymm7, %ymm2, %ymm2
+; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,5,0,0,0,6,0,0,6,0,0,0,7,0,0,7]
+; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm2, %zmm4
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512-FCP-NEXT:    vpbroadcastd 72(%rax), %ymm3
-; AVX512-FCP-NEXT:    vpandn %ymm3, %ymm6, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa 64(%rax), %ymm7
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm7, %ymm4
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm7, %ymm25
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm3
-; AVX512-FCP-NEXT:    vpandn %ymm3, %ymm6, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa (%rax), %ymm4
-; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm4, %ymm4
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm9, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm4, %ymm24
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm8[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm4 = [0,1,0,3,10,10,11,11]
-; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm22, %zmm4
+; AVX512-FCP-NEXT:    vpbroadcastd 72(%rax), %ymm4
+; AVX512-FCP-NEXT:    vpandn %ymm4, %ymm6, %ymm4
+; AVX512-FCP-NEXT:    vmovdqa 64(%rax), %ymm5
+; AVX512-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm5, %ymm5
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm4
+; AVX512-FCP-NEXT:    vpandn %ymm4, %ymm6, %ymm4
+; AVX512-FCP-NEXT:    vmovdqa (%rax), %ymm5
+; AVX512-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm5, %ymm5
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm11 = [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 %ymm11, %ymm9, %ymm4
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm8[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:    vpmovsxbq {{.*#+}} zmm5 = [0,1,0,3,10,10,11,11]
+; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm17, %zmm5
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm3
-; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm2[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5],ymm3[6],ymm4[7,8,9,10],ymm3[11],ymm4[12,13],ymm3[14],ymm4[15]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm23, %ymm7
-; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm4
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2,3],ymm4[4],ymm6[5,6,7,8],ymm4[9],ymm6[10,11],ymm4[12],ymm6[13,14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm5, %zmm4
+; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm0, %ymm4
+; AVX512-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8,9,10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm19, %ymm10
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm5
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2,3],ymm5[4],ymm6[5,6,7,8],ymm5[9],ymm6[10,11],ymm5[12],ymm6[13,14,15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm3, %zmm5
 ; AVX512-FCP-NEXT:    vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm16, %ymm2
-; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm14, %ymm3
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm16, %ymm7
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm14, %ymm3
 ; AVX512-FCP-NEXT:    vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm15[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6,7,8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14,15]
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm14, %ymm5
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm15[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6,7,8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14,15]
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm14, %ymm4
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm15[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7,8,9],ymm6[10],ymm5[11,12],ymm6[13],ymm5[14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm20, %zmm0
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm4 ^ (zmm11 & (zmm0 ^ zmm4))
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7,8,9],ymm6[10],ymm4[11,12],ymm6[13],ymm4[14,15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm24, %zmm0
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm5 ^ (zmm31 & (zmm0 ^ zmm5))
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm8[3,3,3,3,7,7,7,7]
@@ -13447,407 +13451,416 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512-FCP-NEXT:    vprold $16, %ymm9, %ymm4
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm8[1,2,2,3,5,6,6,7]
 ; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm11 = [2,1,3,2,10,10,10,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm11, %zmm4
-; AVX512-FCP-NEXT:    vmovdqa64 (%rax), %zmm16
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,21,0,0,0,22,0,0,14,0,0,0,15,0,0,15]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm14 = [2,1,3,2,10,10,10,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm14, %zmm4
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %ymm0
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm16, %zmm3
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (zmm19 & (zmm3 ^ zmm4))
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm29, %ymm10
-; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm10, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm17, %ymm19
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm17[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6,7,8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14,15]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm30, %ymm9
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm10, %ymm4
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm17[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm20, %zmm5
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm31, %ymm8
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm8, %ymm3
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm21[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm21[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm4[2],ymm6[3,4],ymm4[5],ymm6[6,7,8,9],ymm4[10],ymm6[11,12],ymm4[13],ymm6[14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm11, %zmm6
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm27 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm5 ^ (zmm27 & (zmm6 ^ zmm5))
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm12
-; AVX512-FCP-NEXT:    vprold $16, %ymm26, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm28, %ymm22
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm28[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm24, %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm12, %ymm5
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm30
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm7 = ymm28[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm23 = [2,2,3,3,10,9,11,10]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm23, %zmm5
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm25, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
-; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm18 & (zmm0 ^ zmm5))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm6))
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm2, %zmm0
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm20 & (zmm0 ^ zmm4))
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm8, %ymm0
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm1 = ymm21[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm22, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm2
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm27[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:    vmovdqa64 %ymm21, %ymm13
+; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm0, %ymm3
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm22, %ymm21
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm27[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm24, %zmm3
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm0, %ymm2
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm30[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:    vpunpcklwd {{.*#+}} ymm4 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm30[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm14, %zmm6
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm22 = [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 {{.*#+}} zmm6 = zmm3 ^ (zmm22 & (zmm6 ^ zmm3))
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm23, %ymm10
+; AVX512-FCP-NEXT:    vprold $16, %ymm23, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm25, %ymm20
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm25[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 %ymm11, %ymm10, %ymm3
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm23, %ymm19
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm25[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm1 = [2,2,3,3,10,9,11,10]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm1, %zmm3
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512-FCP-NEXT:    vpermd 64(%rax), %zmm1, %zmm1
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm18 & (zmm1 ^ zmm3))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm6))
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm1
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm0 = ymm30[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm9 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6,7,8],ymm1[9],ymm0[10,11],ymm1[12],ymm0[13,14,15]
 ; AVX512-FCP-NEXT:    vmovdqa 64(%rcx), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %xmm2
-; AVX512-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm14 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm0, %xmm3
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm2[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1],xmm5[2,3],xmm3[4],xmm5[5,6],xmm3[7]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512-FCP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm15 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm2, %xmm2
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm13 = [0,1,1,3,8,8,9,9]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm13, %zmm2
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %xmm3
+; AVX512-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm10 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm2
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm3[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm6[0],xmm2[1],xmm6[2,3],xmm2[4],xmm6[5,6],xmm2[7]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, (%rsp) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
+; 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, %xmm0, %xmm0
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm17 = [0,1,1,3,8,8,9,9]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm17, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %xmm6
 ; AVX512-FCP-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vprold $16, %xmm6, %xmm3
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm0[1,1,2,3]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4],xmm3[5],xmm5[6,7]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm5, %xmm5
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm28 = [0,0,1,1,8,8,10,9]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm28, %zmm5
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm2 ^ (zmm3 & (zmm5 ^ zmm2))
-; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %xmm2
+; AVX512-FCP-NEXT:    vprold $16, %xmm6, %xmm2
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm3 = xmm4[1,1,2,3]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2],xmm3[3,4],xmm2[5],xmm3[6,7]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [0,0,1,1,8,8,10,9]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm18, %zmm3
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm2 & (zmm3 ^ zmm0))
+; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %xmm0
 ; 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 {{.*#+}} xmm4 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm6
-; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm2
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,9,9,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm26, %zmm2
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm15 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm6
+; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm0, %xmm0
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm7 = [0,0,0,1,8,9,9,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm7, %zmm0
 ; AVX512-FCP-NEXT:    vpbroadcastd 64(%rax), %ymm6
 ; AVX512-FCP-NEXT:    vpbroadcastd 68(%rax), %ymm8
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm6, %zmm31
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm21 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm31 = zmm31 ^ (zmm21 & (zmm31 ^ 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 {{.*#+}} zmm31 = zmm31 ^ (zmm2 & (zmm31 ^ zmm5))
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm6, %zmm30
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm30 = zmm30 ^ (zmm12 & (zmm30 ^ zmm0))
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm0 & (zmm30 ^ zmm3))
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm8
 ; AVX512-FCP-NEXT:    vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm0, %xmm5
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm14, %xmm29
+; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm4, %xmm3
 ; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm8[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm6, %xmm6
-; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm13, %zmm6
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm13
-; AVX512-FCP-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vprold $16, %xmm13, %xmm5
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm0[1,1,2,3]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm5 = xmm8[0,1],xmm5[2],xmm8[3,4],xmm5[5],xmm8[6,7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm8, %xmm8
-; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm28, %zmm8
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm6 ^ (zmm3 & (zmm8 ^ zmm6))
-; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm3
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm5
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
-; AVX512-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm5
-; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm3
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm7, %xmm17
-; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm26, %zmm3
-; AVX512-FCP-NEXT:    vpbroadcastd (%rax), %ymm5
-; AVX512-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm6
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm5, %zmm26
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm26 = zmm26 ^ (zmm21 & (zmm26 ^ zmm3))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm26 = zmm26 ^ (zmm2 & (zmm26 ^ zmm8))
-; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm14[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm6[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm14, %ymm3
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm6[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT:    vmovdqa %ymm6, %ymm9
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm5[2],ymm3[3,4],ymm5[5],ymm3[6,7,8,9],ymm5[10],ymm3[11,12],ymm5[13],ymm3[14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm20, %zmm3
-; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm18[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm18[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4],ymm5[5],ymm6[6,7,8,9],ymm5[10],ymm6[11,12],ymm5[13],ymm6[14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm5
-; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %xmm11
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %xmm7
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm6[0],xmm3[1],xmm6[2,3],xmm3[4],xmm6[5,6],xmm3[7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm4[0],xmm8[1],xmm4[1],xmm8[2],xmm4[2],xmm8[3],xmm4[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm17, %zmm6
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm10
+; AVX512-FCP-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vprold $16, %xmm10, %xmm3
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm4[1,1,2,3]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm8[0,1],xmm3[2],xmm8[3,4],xmm3[5],xmm8[6,7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm4[0],xmm10[0],xmm4[1],xmm10[1],xmm4[2],xmm10[2],xmm4[3],xmm10[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm8, %xmm8
+; AVX512-FCP-NEXT:    vmovdqa %xmm1, %xmm10
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm18, %zmm8
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm6 ^ (zmm2 & (zmm8 ^ zmm6))
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm2
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm3
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm3
 ; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm2, %xmm2
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm6 = [0,2,2,3,8,9,9,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm6, %zmm1
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm10, %ymm2
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm19[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm10 = ymm8[0,1,2],ymm2[3],ymm8[4,5],ymm2[6],ymm8[7,8,9,10],ymm2[11],ymm8[12,13],ymm2[14],ymm8[15]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm27 & (zmm5 ^ zmm3))
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm15, %xmm17
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm7, %zmm2
+; AVX512-FCP-NEXT:    vpbroadcastd (%rax), %ymm3
+; AVX512-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm6
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm3, %zmm26
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm26 = zmm26 ^ (zmm12 & (zmm26 ^ zmm2))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm26 = zmm26 ^ (zmm0 & (zmm26 ^ zmm8))
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm15[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm11[2,2,2,2,6,6,6,6]
+; 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:    vpshufb %ymm13, %ymm15, %ymm2
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm11[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 %zmm0, %zmm24, %zmm2
 ; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512-FCP-NEXT:    vprold $16, %ymm4, %ymm2
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm19[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm4, %ymm24
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm19[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm8 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm23, %zmm8
-; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %xmm3
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm4
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm2
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [2,1,3,3,8,8,9,9]
-; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm20, %zmm10
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm21 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm1 ^ (zmm21 & (zmm10 ^ zmm1))
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm30, %ymm1
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm22[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = 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:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm0[0,1],ymm3[2],ymm0[3,4],ymm3[5],ymm0[6,7,8,9],ymm3[10],ymm0[11,12],ymm3[13],ymm0[14,15]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm4[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm7[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm6[0,1],ymm0[2],ymm6[3,4],ymm0[5],ymm6[6,7,8,9],ymm0[10],ymm6[11,12],ymm0[13],ymm6[14,15]
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
+; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %xmm13
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %xmm14
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm5, %xmm25
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,8,9,9,11]
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm1
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm24, %zmm1
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm12 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm21, %ymm3
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm3, %ymm3
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm27[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm6[0,1,2],ymm3[3],ymm6[4,5],ymm3[6],ymm6[7,8,9,10],ymm3[11],ymm6[12,13],ymm3[14],ymm6[15]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm2 ^ (zmm22 & (zmm0 ^ zmm2))
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512-FCP-NEXT:    vprold $16, %ymm5, %ymm2
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm23, %zmm16, %zmm1
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
-; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm16, %zmm22
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm22 = zmm22 ^ (mem & (zmm22 ^ zmm8))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm22 = zmm22 ^ (mem & (zmm22 ^ zmm5))
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm5
-; AVX512-FCP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm1, %xmm1
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm18[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5,6,7,8],ymm0[9],ymm5[10,11],ymm0[12],ymm5[13,14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm0
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm30 = [2,2,2,3,8,8,8,9]
-; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %xmm6
-; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %xmm5
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm23[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = 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, %ymm21
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm23[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 = [2,2,3,3,10,9,11,10]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm5, %zmm3
+; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %xmm5
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm9
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm2
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [2,1,3,3,8,8,9,9]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm18, %zmm6
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm1 ^ (zmm31 & (zmm6 ^ zmm1))
+; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} ymm16 = [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:    # ymm16 = mem[0,1,2,3,0,1,2,3]
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm16, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm19, %ymm2
+; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm20[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:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512-FCP-NEXT:    vpermd (%rax), %zmm1, %zmm20
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm20 = zmm20 ^ (mem & (zmm20 ^ zmm3))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm20 = zmm20 ^ (mem & (zmm20 ^ zmm0))
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm1
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm2
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm29
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm1, %xmm28
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm25, %xmm1
+; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm0, %xmm3
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = 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 %zmm3, %zmm24, %zmm0
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm15, %ymm1
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm12
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm11
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
+; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm2
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm10, %xmm19
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm18, %zmm3
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm24 = [2,2,2,3,8,8,8,9]
+; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %xmm7
+; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %xmm4
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3]
 ; AVX512-FCP-NEXT:    vmovdqa64 %xmm17, %xmm1
-; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm8, %xmm1
-; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm30, %zmm2
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm14, %ymm1
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm9 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm1 = ymm9[0,1,2],ymm1[3],ymm9[4,5],ymm1[6],ymm9[7,8,9,10],ymm1[11],ymm9[12,13],ymm1[14],ymm9[15]
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm15
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm9
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
-; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm14, %xmm14
-; AVX512-FCP-NEXT:    vpermt2q %zmm14, %zmm20, %zmm1
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [6,7,3,3,7,7,6,7]
-; AVX512-FCP-NEXT:    vpermd %ymm25, %ymm17, %ymm16
-; AVX512-FCP-NEXT:    vpbroadcastd 96(%rax), %ymm20
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm20, %zmm16, %zmm25
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm18 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm25 = zmm25 ^ (zmm18 & (zmm25 ^ zmm2))
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm27 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm25 = zmm25 ^ (zmm27 & (zmm25 ^ zmm10))
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512-FCP-NEXT:    vprold $16, %xmm3, %xmm3
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[1,1,2,3]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm0 ^ (zmm21 & (zmm1 ^ zmm0))
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = xmm2[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm21 = [0,0,2,1,8,8,9,11]
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm21, %zmm3
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm29, %xmm2
-; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm7 = xmm7[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0],xmm2[1],xmm7[2,3],xmm2[4],xmm7[5,6],xmm2[7]
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm28, %zmm7
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm3 ^ (zmm28 & (zmm7 ^ zmm3))
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Reload
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm16 = ymm16 ^ (ymm18 & (ymm16 ^ ymm0))
-; AVX512-FCP-NEXT:    vextracti64x4 $1, %zmm0, %ymm13
-; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm2
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm8[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm3 = [0,1,1,3,8,8,9,9]
-; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm3, %zmm6
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm24, %ymm2
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm19[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm8 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
-; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm3
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm11
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm3[0],xmm11[1],xmm3[1],xmm11[2],xmm3[2],xmm11[3],xmm3[3]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm2[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm30, %zmm8
-; AVX512-FCP-NEXT:    vpbroadcastd 100(%rax), %ymm5
-; AVX512-FCP-NEXT:    vpbroadcastd 104(%rax), %ymm19
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm5, %zmm5
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm24 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (zmm24 & (zmm5 ^ zmm6))
-; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufhw {{.*#+}} ymm6 = ymm4[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[3,3,3,3]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm15
+; AVX512-FCP-NEXT:    vpermt2q %zmm15, %zmm24, %zmm8
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm31 & (zmm3 ^ zmm0))
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm15 = [6,7,3,3,7,7,6,7]
+; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    vpbroadcastd 96(%rax), %ymm18
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm18, %zmm0, %zmm18
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm31 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm18 = zmm18 ^ (zmm31 & (zmm18 ^ zmm8))
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm25 = [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 {{.*#+}} zmm18 = zmm18 ^ (zmm25 & (zmm18 ^ zmm6))
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm9[4],xmm5[5],xmm9[5],xmm5[6],xmm9[6],xmm5[7],xmm9[7]
+; AVX512-FCP-NEXT:    vprold $16, %xmm5, %xmm0
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm9[1,1,2,3]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm0 = xmm5[0,1],xmm0[2],xmm5[3,4],xmm0[5],xmm5[6,7]
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm16, %ymm1
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm21, %ymm5
+; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm5, %ymm1
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm23[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), %xmm6
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm5
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm17, %xmm9
+; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm8, %xmm9
+; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm24, %zmm1
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm24 = [0,0,2,1,8,8,9,11]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm24, %zmm0
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm16 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm16, %xmm9
+; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm13, %xmm13
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm14 = xmm14[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm13 = xmm14[0],xmm13[1],xmm14[2,3],xmm13[4],xmm14[5,6],xmm13[7]
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm22 = [0,0,1,1,8,8,10,9]
+; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm22, %zmm13
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm13 = zmm0 ^ (zmm23 & (zmm13 ^ zmm0))
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
+; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm14 = ymm14 ^ (ymm31 & (ymm14 ^ ymm2))
+; AVX512-FCP-NEXT:    vextracti64x4 $1, %zmm2, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm19, %xmm9
+; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm0, %xmm0
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm17 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm17, %xmm4
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
+; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm21 = [0,1,1,3,8,8,9,9]
+; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm21, %zmm4
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm7[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm7[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 {{.*#+}} ymm7 = ymm7[2,2,2,2]
 ; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm10, %xmm14
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[0,0,1,1]
+; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm10, %xmm10
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm19 = ymm10[0,0,1,1]
 ; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm10, %xmm10
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm12, %xmm30
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
-; AVX512-FCP-NEXT:    vpermd %ymm23, %ymm17, %ymm17
-; AVX512-FCP-NEXT:    vpbroadcastd 32(%rax), %ymm19
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm17, %zmm20
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm20 = zmm20 ^ (zmm18 & (zmm20 ^ zmm8))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm20 = zmm20 ^ (zmm27 & (zmm20 ^ zmm1))
+; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm10, %xmm10
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm9, %xmm27
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm10[0,0,1,1]
+; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm10 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    vpbroadcastd 32(%rax), %ymm15
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm15, %zmm10, %zmm10
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm10 = zmm10 ^ (zmm31 & (zmm10 ^ 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 {{.*#+}} zmm31 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm15 = zmm15 ^ (zmm31 & (zmm15 ^ zmm4))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm10 ^ (zmm25 & (zmm10 ^ zmm3))
 ; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm1 & (zmm5 ^ zmm7))
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm15[4],xmm9[5],xmm15[5],xmm9[6],xmm15[6],xmm9[7],xmm15[7]
-; AVX512-FCP-NEXT:    vprold $16, %xmm9, %xmm8
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm15[1,1,2,3]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm15 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm7, %xmm7
-; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm21, %zmm8
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm29, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpshufb %xmm0, %xmm9, %xmm7
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm0[4],xmm9[4],xmm0[5],xmm9[5],xmm0[6],xmm9[6],xmm0[7],xmm9[7]
-; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm12 = xmm0[1,1,2,2]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm7 = xmm12[0],xmm7[1],xmm12[2,3],xmm7[4],xmm12[5,6],xmm7[7]
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm9[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm0 = [0,0,1,1,8,8,10,9]
-; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm0, %zmm7
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm8 ^ (zmm28 & (zmm7 ^ zmm8))
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm11[4],xmm3[4],xmm11[5],xmm3[5],xmm11[6],xmm3[6],xmm11[7],xmm3[7]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm30, %xmm0
-; AVX512-FCP-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm0 = [0,1,1,3,8,8,9,9]
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
-; AVX512-FCP-NEXT:    vpbroadcastd 36(%rax), %ymm3
-; AVX512-FCP-NEXT:    vpbroadcastd 40(%rax), %ymm8
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (zmm24 & (zmm3 ^ zmm2))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm1 & (zmm3 ^ zmm7))
-; 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 {{.*#+}} ymm13 = mem ^ (ymm0 & (ymm13 ^ mem))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm15 = zmm15 ^ (zmm1 & (zmm15 ^ zmm13))
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
+; AVX512-FCP-NEXT:    vprold $16, %xmm11, %xmm4
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm11 = xmm12[1,1,2,3]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm4 = xmm11[0,1],xmm4[2],xmm11[3,4],xmm4[5],xmm11[6,7]
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm25 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm25, %xmm11
+; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm3
+; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm24, %zmm4
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm16, %xmm3
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm28, %xmm11
+; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm11, %xmm3
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm29, %xmm12
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512-FCP-NEXT:    vpshufd {{.*#+}} xmm12 = xmm29[1,1,2,2]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm12[0],xmm3[1],xmm12[2,3],xmm3[4],xmm12[5,6],xmm3[7]
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm11[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm22, %zmm3
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm4 ^ (zmm23 & (zmm3 ^ zmm4))
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm27, %xmm5
+; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm17, %xmm5
+; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm8, %xmm5
+; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm21, %zmm5
+; AVX512-FCP-NEXT:    vpbroadcastd 36(%rax), %ymm4
+; AVX512-FCP-NEXT:    vpbroadcastd 40(%rax), %ymm6
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm4, %zmm4
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm4 = zmm4 ^ (zmm31 & (zmm4 ^ zmm5))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 ^ (zmm1 & (zmm4 ^ zmm3))
+; 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 {{.*#+}} ymm2 = mem ^ (ymm1 & (ymm2 ^ mem))
 ; 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 {{.*#+}} ymm6 = ymm6 ^ (ymm1 & (ymm6 ^ ymm13))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm0 = ymm0 ^ (ymm1 & (ymm0 ^ ymm2))
 ; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm4 = ymm4 ^ (ymm2 & (ymm4 ^ ymm16))
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm0[4,5,6,7]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm7 = ymm7 ^ (ymm2 & (ymm7 ^ ymm14))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm7[0,1,2,3],zmm0[4,5,6,7]
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
 ; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm0))
-; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm0 # 32-byte Folded Reload
 ; AVX512-FCP-NEXT:    vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm0 # 64-byte Folded Reload
 ; AVX512-FCP-NEXT:    # zmm0 = (zmm0 & zmm2) | mem
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4
-; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm2 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    # zmm2 = (zmm2 & zmm4) | mem
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpunpckhwd (%rsp), %xmm4, %xmm4 # 16-byte Folded Reload
-; AVX512-FCP-NEXT:    # xmm4 = xmm4[4],mem[4],xmm4[5],mem[5],xmm4[6],mem[6],xmm4[7],mem[7]
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3
+; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm2 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm2 # 64-byte Folded Reload
+; AVX512-FCP-NEXT:    # zmm2 = (zmm2 & zmm3) | mem
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm3 = 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), %xmm5 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm5 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
 ; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
 ; AVX512-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload
 ; AVX512-FCP-NEXT:    # xmm6 = xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload
-; AVX512-FCP-NEXT:    # xmm7 = xmm7[4],mem[4],xmm7[5],mem[5],xmm7[6],mem[6],xmm7[7],mem[7]
-; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm4, %xmm4
-; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm7, %xmm7
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm25, %xmm7
+; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm3
+; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm6, %xmm6
+; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm7[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), %ymm8 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm8 = mem[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm7 = ymm8[0,1],ymm7[2],ymm8[3,4],ymm7[5],ymm8[6,7,8,9],ymm7[10],ymm8[11,12],ymm7[13],ymm8[14,15]
 ; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT:    vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    # ymm9 = mem[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7,8,9],ymm8[10],ymm9[11,12],ymm8[13],ymm9[14,15]
-; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT:    vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    # ymm10 = mem[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT:    vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7,8,9],ymm10[10],ymm9[11,12],ymm10[13],ymm9[14,15]
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm10 = [0,1,4,5,4,5,5,7]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT:    vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    # ymm9 = mem[1,1,1,1,5,5,5,5]
+; 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:    vpmovsxbd {{.*#+}} ymm9 = [0,1,4,5,4,5,5,7]
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512-FCP-NEXT:    vpermd %ymm11, %ymm10, %ymm10
+; AVX512-FCP-NEXT:    vpermd %ymm11, %ymm9, %ymm9
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm11[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm11[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm11[16,17],zero,zero
 ; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT:    vpandn %ymm10, %ymm12, %ymm10
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm11, %zmm10
-; AVX512-FCP-NEXT:    vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    vpandn %ymm9, %ymm12, %ymm9
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm11, %zmm9
+; AVX512-FCP-NEXT:    vpshuflw $248, (%rsp), %xmm11 # 16-byte Folded Reload
 ; AVX512-FCP-NEXT:    # xmm11 = mem[0,2,3,3,4,5,6,7]
 ; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,1,3]
-; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,0,1,3]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,1,3]
+; AVX512-FCP-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,1,3,2]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
 ; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm12))
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm11 ^ (zmm28 & (zmm4 ^ zmm11))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm11 ^ (zmm23 & (zmm3 ^ zmm11))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 32-byte Folded Reload
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm6 ^ (zmm28 & (zmm7 ^ zmm6))
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm0 ^ (zmm6 & (zmm4 ^ zmm0))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm2 ^ (zmm6 & (zmm7 ^ zmm2))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm5 ^ (zmm23 & (zmm6 ^ 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 {{.*#+}} zmm3 = zmm0 ^ (zmm5 & (zmm3 ^ zmm0))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm2 ^ (zmm5 & (zmm6 ^ zmm2))
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm0
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm2, %zmm2
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm2, %zmm2
 ; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm0 ^ (mem & (zmm2 ^ zmm0))
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
 ; AVX512-FCP-NEXT:    vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
 ; AVX512-FCP-NEXT:    # zmm0 = zmm0 | (zmm1 & mem)
-; AVX512-FCP-NEXT:    vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm10 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    # zmm10 = zmm10 | (zmm1 & mem)
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm2))
+; AVX512-FCP-NEXT:    vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm9 # 64-byte Folded Reload
+; AVX512-FCP-NEXT:    # zmm9 = zmm9 | (zmm1 & mem)
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm2))
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 320(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 256(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, 192(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, 128(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, 320(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 256(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 192(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, 128(%rax)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, (%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, 448(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 704(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, 640(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, 448(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, 704(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, 640(%rax)
 ; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
 ; AVX512-FCP-NEXT:    vmovaps %zmm1, 576(%rax)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, 384(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 64(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 512(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 512(%rax)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 832(%rax)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 768(%rax)
 ; AVX512-FCP-NEXT:    addq $1496, %rsp # imm = 0x5D8
@@ -14562,70 +14575,70 @@ 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 $1496, %rsp # imm = 0x5D8
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %ymm4
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm8 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %ymm5
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm1, %ymm2
-; AVX512DQ-FCP-NEXT:    vporq %ymm0, %ymm2, %ymm16
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %ymm4
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT:    vporq %ymm0, %ymm1, %ymm16
+; 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, %ymm4, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %ymm5
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %ymm3
 ; 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, %ymm5, %ymm2
-; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm2, %ymm7
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm7
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm26
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r8), %ymm12
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm12, %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm12, %ymm28
-; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm23
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r8), %ymm13
+; 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, %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm13, %ymm25
+; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm1, %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 %ymm8, %ymm0, %ymm14
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm31
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm0, %ymm13
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm26
 ; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm0, %ymm15
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm21
-; AVX512DQ-FCP-NEXT:    vpor %ymm14, %ymm15, %ymm12
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm30
+; AVX512DQ-FCP-NEXT:    vpor %ymm13, %ymm14, %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, %ymm29
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm13
+; 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, %ymm17
-; AVX512DQ-FCP-NEXT:    vpor %ymm14, %ymm15, %ymm12
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm27
+; AVX512DQ-FCP-NEXT:    vpor %ymm13, %ymm14, %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 %ymm6, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm13
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm0, %ymm15
-; AVX512DQ-FCP-NEXT:    vpor %ymm14, %ymm15, %ymm12
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT:    vpor %ymm13, %ymm14, %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 %ymm8, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm0, %ymm13
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm0, %ymm12
-; AVX512DQ-FCP-NEXT:    vpor %ymm14, %ymm12, %ymm12
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT:    vpor %ymm13, %ymm14, %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, %ymm12
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm13
 ; 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, %ymm14
-; AVX512DQ-FCP-NEXT:    vpor %ymm12, %ymm14, %ymm12
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpor %ymm13, %ymm14, %ymm13
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm14
 ; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm14, %ymm8
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm15
@@ -14634,138 +14647,136 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm0
 ; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm8
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm2, %ymm9
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm1, %ymm9
 ; AVX512DQ-FCP-NEXT:    vpor %ymm8, %ymm9, %ymm8
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm8
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm8, %ymm10
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm8, %ymm10
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %ymm9
 ; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm9, %ymm11
-; AVX512DQ-FCP-NEXT:    vporq %ymm11, %ymm10, %ymm22
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm11 = ymm12[0,1],ymm11[2],ymm12[3,4],ymm11[5],ymm12[6,7,8,9],ymm11[10],ymm12[11,12],ymm11[13],ymm12[14,15]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512DQ-FCP-NEXT:    vporq %ymm11, %ymm10, %ymm17
+; 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 {{.*#+}} ymm12 = ymm5[0,1,1,3,4,5,5,7]
+; 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:    vpermq {{.*#+}} ymm10 = ymm10[2,1,3,2]
 ; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm13 = [18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
 ; AVX512DQ-FCP-NEXT:    # ymm13 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm4, %ymm12
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm13, %ymm30
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm2, %ymm12
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm13, %ymm21
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm13 = ymm3[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, %ymm10, %zmm16, %zmm10
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm7, %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:    vpternlogq {{.*#+}} zmm12 = zmm10 ^ (mem & (zmm12 ^ zmm10))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
 ; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r8), %ymm7
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm7[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[16,17,u,u,u,u],zero,zero
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm11 = ymm11 | (ymm12 & ymm13)
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %ymm10
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm10, %ymm6
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm6 = ymm6 | (ymm11 & ymm16)
-; AVX512DQ-FCP-NEXT:    vextracti64x4 $1, %zmm12, %ymm11
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm7[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[16,17,u,u,u,u],zero,zero
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm10 = ymm10 | (ymm12 & ymm18)
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %ymm11
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm11, %ymm6
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm11, %ymm16
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; 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 {{.*#+}} ymm6 = ymm6 | (ymm10 & ymm11)
+; AVX512DQ-FCP-NEXT:    vextracti64x4 $1, %zmm12, %ymm10
 ; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm12 = [0,4,0,0,0,5,0,0]
 ; AVX512DQ-FCP-NEXT:    vpermd %ymm7, %ymm12, %ymm12
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm12 = ymm12 ^ (ymm16 & (ymm12 ^ ymm11))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm19
-; AVX512DQ-FCP-NEXT:    vprold $16, %ymm10, %ymm11
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm11 = ymm11 ^ (ymm10 & (ymm11 ^ ymm12))
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm0, %zmm11
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm11[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [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 ^ (ymm11 & (ymm12 ^ ymm10))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm20
+; AVX512DQ-FCP-NEXT:    vprold $16, %ymm16, %ymm10
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [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 {{.*#+}} ymm10 = ymm10 ^ (ymm11 & (ymm10 ^ ymm12))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm0, %zmm10
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm10[4,5,6,7]
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm11 = [0,1,4,5,4,5,5,7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm10 = [0,1,4,5,4,5,5,7]
 ; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rax), %ymm6
-; AVX512DQ-FCP-NEXT:    vpermd %ymm6, %ymm11, %ymm11
-; AVX512DQ-FCP-NEXT:    vpandn %ymm11, %ymm13, %ymm11
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm18
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm6, %ymm12
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm12, %zmm11
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm4, %ymm11
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm5[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm11 = ymm12[0,1,2],ymm11[3],ymm12[4,5],ymm11[6],ymm12[7,8,9,10],ymm11[11],ymm12[12,13],ymm11[14],ymm12[15]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm6, %ymm10, %ymm10
+; AVX512DQ-FCP-NEXT:    vpandnq %ymm10, %ymm18, %ymm10
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm11 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm6, %ymm12
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm12, %zmm10
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm2, %ymm10
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm3[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm10 = ymm12[0,1,2],ymm10[3],ymm12[4,5],ymm10[6],ymm12[7,8,9,10],ymm10[11],ymm12[12,13],ymm10[14],ymm12[15]
 ; AVX512DQ-FCP-NEXT:    vpmovsxdq {{.*#+}} ymm12 = [151522058,0,421010202,421010202]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm12, %ymm23
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm5 = [0,2,2,3,10,9,11,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm5, %zmm4
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm2, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm12, %ymm19
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[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 {{.*#+}} zmm3 = [0,2,2,3,10,9,11,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm3, %zmm2
 ; AVX512DQ-FCP-NEXT:    vpmovsxdq {{.*#+}} ymm12 = [218894094,0,488382238,488382238]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm3, %ymm11
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm4, %ymm10
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm12, %ymm16
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm11 = ymm12[0],ymm11[1],ymm12[2,3],ymm11[4],ymm12[5,6,7,8],ymm11[9],ymm12[10,11],ymm11[12],ymm12[13,14,15]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm12 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm10 = ymm12[0],ymm10[1],ymm12[2,3],ymm10[4],ymm12[5,6,7,8],ymm10[9],ymm12[10,11],ymm10[12],ymm12[13,14,15]
 ; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
 ; AVX512DQ-FCP-NEXT:    # ymm12 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm3, %ymm3
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4],ymm1[5],ymm3[6,7,8,9],ymm1[10],ymm3[11,12],ymm1[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm20, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm4 ^ (zmm11 & (zmm1 ^ zmm4))
-; 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 %ymm7, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rax), %zmm1
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,5,0,0,0,6,0,0,30,0,0,0,31,0,0,31]
-; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm6, %zmm3
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; 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:    vpmovsxbq {{.*#+}} zmm24 = [2,2,2,3,8,10,10,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm24, %zmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm31 = [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 {{.*#+}} zmm4 = zmm2 ^ (zmm31 & (zmm4 ^ zmm2))
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [5,0,0,0,6,0,0,6]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm7, %ymm2, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,5,0,0,0,6,0,0,6,0,0,0,7,0,0,7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm2, %zmm4
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 72(%rax), %ymm3
-; AVX512DQ-FCP-NEXT:    vpandn %ymm3, %ymm6, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rax), %ymm7
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm7, %ymm4
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm7, %ymm25
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm3
-; AVX512DQ-FCP-NEXT:    vpandn %ymm3, %ymm6, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %ymm4
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm9, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm4, %ymm24
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm8[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm4 = [0,1,0,3,10,10,11,11]
-; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm22, %zmm4
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 72(%rax), %ymm4
+; AVX512DQ-FCP-NEXT:    vpandn %ymm4, %ymm6, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rax), %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm5, %ymm5
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 8(%rax), %ymm4
+; AVX512DQ-FCP-NEXT:    vpandn %ymm4, %ymm6, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm5, %ymm5
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm11 = [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 %ymm11, %ymm9, %ymm4
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm8[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:    vpmovsxbq {{.*#+}} zmm5 = [0,1,0,3,10,10,11,11]
+; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm17, %zmm5
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm2[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5],ymm3[6],ymm4[7,8,9,10],ymm3[11],ymm4[12,13],ymm3[14],ymm4[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm23, %ymm7
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm4
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2,3],ymm4[4],ymm6[5,6,7,8],ymm4[9],ymm6[10,11],ymm4[12],ymm6[13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm5, %zmm4
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm0, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8,9,10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm19, %ymm10
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm5
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2,3],ymm5[4],ymm6[5,6,7,8],ymm5[9],ymm6[10,11],ymm5[12],ymm6[13,14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm3, %zmm5
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm16, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm14, %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm16, %ymm7
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm14, %ymm3
 ; AVX512DQ-FCP-NEXT:    vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm15[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6,7,8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm14, %ymm5
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm15[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6,7,8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14,15]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm14, %ymm4
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm15[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7,8,9],ymm6[10],ymm5[11,12],ymm6[13],ymm5[14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm20, %zmm0
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm4 ^ (zmm11 & (zmm0 ^ zmm4))
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7,8,9],ymm6[10],ymm4[11,12],ymm6[13],ymm4[14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm24, %zmm0
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm5 ^ (zmm31 & (zmm0 ^ zmm5))
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm8[3,3,3,3,7,7,7,7]
@@ -14773,407 +14784,416 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ; AVX512DQ-FCP-NEXT:    vprold $16, %ymm9, %ymm4
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm8[1,2,2,3,5,6,6,7]
 ; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm11 = [2,1,3,2,10,10,10,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm11, %zmm4
-; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rax), %zmm16
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,21,0,0,0,22,0,0,14,0,0,0,15,0,0,15]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm14 = [2,1,3,2,10,10,10,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm14, %zmm4
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm16, %zmm3
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (zmm19 & (zmm3 ^ zmm4))
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm29, %ymm10
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm10, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm17, %ymm19
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm17[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6,7,8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm9
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm10, %ymm4
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm17[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm20, %zmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm31, %ymm8
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm8, %ymm3
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm21[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm21[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm4[2],ymm6[3,4],ymm4[5],ymm6[6,7,8,9],ymm4[10],ymm6[11,12],ymm4[13],ymm6[14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm11, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm27 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm5 ^ (zmm27 & (zmm6 ^ zmm5))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm12
-; AVX512DQ-FCP-NEXT:    vprold $16, %ymm26, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm28, %ymm22
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm28[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm24, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm12, %ymm5
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm30
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm7 = ymm28[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm23 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm23, %zmm5
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm25, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm18 & (zmm0 ^ zmm5))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm6))
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm2, %zmm0
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (zmm20 & (zmm0 ^ zmm4))
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm1 = ymm21[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm22, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm27[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:    vmovdqa64 %ymm21, %ymm13
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm0, %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm22, %ymm21
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm27[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm24, %zmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm0, %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm30[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:    vpunpcklwd {{.*#+}} ymm4 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm30[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm14, %zmm6
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm22 = [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 {{.*#+}} zmm6 = zmm3 ^ (zmm22 & (zmm6 ^ zmm3))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm23, %ymm10
+; AVX512DQ-FCP-NEXT:    vprold $16, %ymm23, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm25, %ymm20
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm25[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 %ymm11, %ymm10, %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm23, %ymm19
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm25[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm1 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm1, %zmm3
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512DQ-FCP-NEXT:    vpermd 64(%rax), %zmm1, %zmm1
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm1 = zmm1 ^ (zmm18 & (zmm1 ^ zmm3))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm6))
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm0 = ymm30[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm9 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6,7,8],ymm1[9],ymm0[10,11],ymm1[12],ymm0[13,14,15]
 ; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rcx), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %xmm2
-; AVX512DQ-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm14 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm0, %xmm3
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm2[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1],xmm5[2,3],xmm3[4],xmm5[5,6],xmm3[7]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm15 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm13 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm13, %zmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %xmm3
+; AVX512DQ-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm10 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm2
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm3[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm6[0],xmm2[1],xmm6[2,3],xmm2[4],xmm6[5,6],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm4, (%rsp) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
+; 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, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm17 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm17, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %xmm6
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm6, %xmm3
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm0[1,1,2,3]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4],xmm3[5],xmm5[6,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm5, %xmm5
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm28 = [0,0,1,1,8,8,10,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm28, %zmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm2 ^ (zmm3 & (zmm5 ^ zmm2))
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %xmm2
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm6, %xmm2
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm3 = xmm4[1,1,2,3]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2],xmm3[3,4],xmm2[5],xmm3[6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [0,0,1,1,8,8,10,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm18, %zmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm2 & (zmm3 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %xmm0
 ; 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 {{.*#+}} xmm4 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm6
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,9,9,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm26, %zmm2
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm15 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm6
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm7 = [0,0,0,1,8,9,9,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm7, %zmm0
 ; AVX512DQ-FCP-NEXT:    vpbroadcastd 64(%rax), %ymm6
 ; AVX512DQ-FCP-NEXT:    vpbroadcastd 68(%rax), %ymm8
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm6, %zmm31
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm21 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm31 = zmm31 ^ (zmm21 & (zmm31 ^ 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 {{.*#+}} zmm31 = zmm31 ^ (zmm2 & (zmm31 ^ zmm5))
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm6, %zmm30
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm30 = zmm30 ^ (zmm12 & (zmm30 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm0 & (zmm30 ^ zmm3))
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm8
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm0, %xmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm14, %xmm29
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm4, %xmm3
 ; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm6 = xmm8[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm6, %xmm6
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm13, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm13
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm13, %xmm5
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm0[1,1,2,3]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm5 = xmm8[0,1],xmm5[2],xmm8[3,4],xmm5[5],xmm8[6,7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm8, %xmm8
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm28, %zmm8
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm6 ^ (zmm3 & (zmm8 ^ zmm6))
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm5
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm5
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm7, %xmm17
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm26, %zmm3
-; AVX512DQ-FCP-NEXT:    vpbroadcastd (%rax), %ymm5
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm6
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm5, %zmm26
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm26 = zmm26 ^ (zmm21 & (zmm26 ^ zmm3))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm26 = zmm26 ^ (zmm2 & (zmm26 ^ zmm8))
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm14[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm6[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm14, %ymm3
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm6[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm6, %ymm9
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm5[2],ymm3[3,4],ymm5[5],ymm3[6,7,8,9],ymm5[10],ymm3[11,12],ymm5[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm20, %zmm3
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm18[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm18[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4],ymm5[5],ymm6[6,7,8,9],ymm5[10],ymm6[11,12],ymm5[13],ymm6[14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %xmm11
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %xmm7
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm6[0],xmm3[1],xmm6[2,3],xmm3[4],xmm6[5,6],xmm3[7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm4[0],xmm8[1],xmm4[1],xmm8[2],xmm4[2],xmm8[3],xmm4[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm17, %zmm6
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm10
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm10, %xmm3
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm8 = xmm4[1,1,2,3]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm8[0,1],xmm3[2],xmm8[3,4],xmm3[5],xmm8[6,7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm4[0],xmm10[0],xmm4[1],xmm10[1],xmm4[2],xmm10[2],xmm4[3],xmm10[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm8, %xmm8
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, %xmm10
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm18, %zmm8
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm6 ^ (zmm2 & (zmm8 ^ zmm6))
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm3
 ; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm6 = [0,2,2,3,8,9,9,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm6, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm10, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm8 = ymm19[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm10 = ymm8[0,1,2],ymm2[3],ymm8[4,5],ymm2[6],ymm8[7,8,9,10],ymm2[11],ymm8[12,13],ymm2[14],ymm8[15]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm27 & (zmm5 ^ zmm3))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm15, %xmm17
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm7, %zmm2
+; AVX512DQ-FCP-NEXT:    vpbroadcastd (%rax), %ymm3
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 4(%rax), %ymm6
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm3, %zmm26
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm26 = zmm26 ^ (zmm12 & (zmm26 ^ zmm2))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm26 = zmm26 ^ (zmm0 & (zmm26 ^ zmm8))
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm15[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm11[2,2,2,2,6,6,6,6]
+; 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:    vpshufb %ymm13, %ymm15, %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm11[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 %zmm0, %zmm24, %zmm2
 ; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vprold $16, %ymm4, %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm19[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm4, %ymm24
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm4 = ymm19[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm8 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm23, %zmm8
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm4
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm20 = [2,1,3,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm20, %zmm10
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm21 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm1 ^ (zmm21 & (zmm10 ^ zmm1))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm1
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm22[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = 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:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm0[0,1],ymm3[2],ymm0[3,4],ymm3[5],ymm0[6,7,8,9],ymm3[10],ymm0[11,12],ymm3[13],ymm0[14,15]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm4[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm7[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm6[0,1],ymm0[2],ymm6[3,4],ymm0[5],ymm6[6,7,8,9],ymm0[10],ymm6[11,12],ymm0[13],ymm6[14,15]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %xmm13
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %xmm14
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm5, %xmm25
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,8,9,9,11]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm1
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm24, %zmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm12 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm21, %ymm3
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm6 = ymm27[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm6 = ymm6[0,1,2],ymm3[3],ymm6[4,5],ymm3[6],ymm6[7,8,9,10],ymm3[11],ymm6[12,13],ymm3[14],ymm6[15]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm2 ^ (zmm22 & (zmm0 ^ zmm2))
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vprold $16, %ymm5, %ymm2
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm23, %zmm16, %zmm1
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm16, %zmm22
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm22 = zmm22 ^ (mem & (zmm22 ^ zmm8))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm22 = zmm22 ^ (mem & (zmm22 ^ zmm5))
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm18[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5,6,7,8],ymm0[9],ymm5[10,11],ymm0[12],ymm5[13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm0
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm30 = [2,2,2,3,8,8,8,9]
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %xmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r8), %xmm5
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm23[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = 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, %ymm21
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm23[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 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm5, %zmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %xmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm9
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm18 = [2,1,3,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm18, %zmm6
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm1 ^ (zmm31 & (zmm6 ^ zmm1))
+; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} ymm16 = [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:    # ymm16 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm16, %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm19, %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm20[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:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512DQ-FCP-NEXT:    vpermd (%rax), %zmm1, %zmm20
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm20 = zmm20 ^ (mem & (zmm20 ^ zmm3))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm20 = zmm20 ^ (mem & (zmm20 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm2
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm29
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm28
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm25, %xmm1
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm0, %xmm3
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm0 = 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 %zmm3, %zmm24, %zmm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm15, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm3 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm11
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm10, %xmm19
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm18, %zmm3
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm24 = [2,2,2,3,8,8,8,9]
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %xmm7
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r8), %xmm4
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3]
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm17, %xmm1
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm8, %xmm1
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm30, %zmm2
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm14, %ymm1
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm9 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm1 = ymm9[0,1,2],ymm1[3],ymm9[4,5],ymm1[6],ymm9[7,8,9,10],ymm1[11],ymm9[12,13],ymm1[14],ymm9[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm15
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm9
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm14, %xmm14
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm14, %zmm20, %zmm1
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm17 = [6,7,3,3,7,7,6,7]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm25, %ymm17, %ymm16
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 96(%rax), %ymm20
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm20, %zmm16, %zmm25
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm18 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm25 = zmm25 ^ (zmm18 & (zmm25 ^ zmm2))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm27 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm25 = zmm25 ^ (zmm27 & (zmm25 ^ zmm10))
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[1,1,2,3]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm0 ^ (zmm21 & (zmm1 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = xmm2[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm21 = [0,0,2,1,8,8,9,11]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm21, %zmm3
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm29, %xmm2
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm7 = xmm7[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0],xmm2[1],xmm7[2,3],xmm2[4],xmm7[5,6],xmm2[7]
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm28, %zmm7
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm3 ^ (zmm28 & (zmm7 ^ zmm3))
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm16 = ymm16 ^ (ymm18 & (ymm16 ^ ymm0))
-; AVX512DQ-FCP-NEXT:    vextracti64x4 $1, %zmm0, %ymm13
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm8[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm3 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm3, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm24, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm3 = ymm19[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm8 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm11
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm3[0],xmm11[1],xmm3[1],xmm11[2],xmm3[2],xmm11[3],xmm3[3]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm2[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm30, %zmm8
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 100(%rax), %ymm5
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 104(%rax), %ymm19
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm5, %zmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm24 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm5 = zmm5 ^ (zmm24 & (zmm5 ^ zmm6))
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufhw {{.*#+}} ymm6 = ymm4[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[3,3,3,3]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm15
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm15, %zmm24, %zmm8
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm31 & (zmm3 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm15 = [6,7,3,3,7,7,6,7]
+; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 96(%rax), %ymm18
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm18, %zmm0, %zmm18
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm31 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm18 = zmm18 ^ (zmm31 & (zmm18 ^ zmm8))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm25 = [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 {{.*#+}} zmm18 = zmm18 ^ (zmm25 & (zmm18 ^ zmm6))
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm9[4],xmm5[5],xmm9[5],xmm5[6],xmm9[6],xmm5[7],xmm9[7]
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm5, %xmm0
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm5 = xmm9[1,1,2,3]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm0 = xmm5[0,1],xmm0[2],xmm5[3,4],xmm0[5],xmm5[6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm16, %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm21, %ymm5
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} ymm5 = ymm23[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), %xmm6
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm5
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm17, %xmm9
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm8, %xmm9
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm24, %zmm1
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm24 = [0,0,2,1,8,8,9,11]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm24, %zmm0
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512DQ-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm16 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm16, %xmm9
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm13, %xmm13
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm14 = xmm14[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm13 = xmm14[0],xmm13[1],xmm14[2,3],xmm13[4],xmm14[5,6],xmm13[7]
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm22 = [0,0,1,1,8,8,10,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm22, %zmm13
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm13 = zmm0 ^ (zmm23 & (zmm13 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm14 = ymm14 ^ (ymm31 & (ymm14 ^ ymm2))
+; AVX512DQ-FCP-NEXT:    vextracti64x4 $1, %zmm2, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm19, %xmm9
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm17 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm17, %xmm4
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
+; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm21 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm21, %zmm4
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm7[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm7[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 {{.*#+}} ymm7 = ymm7[2,2,2,2]
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm10, %xmm14
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[0,0,1,1]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm10, %xmm10
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm19 = ymm10[0,0,1,1]
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm10, %xmm10
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm12, %xmm30
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm23, %ymm17, %ymm17
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 32(%rax), %ymm19
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm17, %zmm20
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm20 = zmm20 ^ (zmm18 & (zmm20 ^ zmm8))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm20 = zmm20 ^ (zmm27 & (zmm20 ^ zmm1))
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm10, %xmm10
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm9, %xmm27
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm10[0,0,1,1]
+; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm10 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 32(%rax), %ymm15
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm15, %zmm10, %zmm10
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm10 = zmm10 ^ (zmm31 & (zmm10 ^ 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 {{.*#+}} zmm31 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm15 = zmm15 ^ (zmm31 & (zmm15 ^ zmm4))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm10 ^ (zmm25 & (zmm10 ^ zmm3))
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm1 & (zmm5 ^ zmm7))
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm15[4],xmm9[5],xmm15[5],xmm9[6],xmm15[6],xmm9[7],xmm15[7]
-; AVX512DQ-FCP-NEXT:    vprold $16, %xmm9, %xmm8
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm9 = xmm15[1,1,2,3]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm15 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm7, %xmm7
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm21, %zmm8
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm29, %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm0, %xmm9, %xmm7
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm0[4],xmm9[4],xmm0[5],xmm9[5],xmm0[6],xmm9[6],xmm0[7],xmm9[7]
-; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm12 = xmm0[1,1,2,2]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm7 = xmm12[0],xmm7[1],xmm12[2,3],xmm7[4],xmm12[5,6],xmm7[7]
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm9[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm0 = [0,0,1,1,8,8,10,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm0, %zmm7
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm8 ^ (zmm28 & (zmm7 ^ zmm8))
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm11[4],xmm3[4],xmm11[5],xmm3[5],xmm11[6],xmm3[6],xmm11[7],xmm3[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm30, %xmm0
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm0 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 36(%rax), %ymm3
-; AVX512DQ-FCP-NEXT:    vpbroadcastd 40(%rax), %ymm8
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (zmm24 & (zmm3 ^ zmm2))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm1 & (zmm3 ^ zmm7))
-; 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 {{.*#+}} ymm13 = mem ^ (ymm0 & (ymm13 ^ mem))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm15 = zmm15 ^ (zmm1 & (zmm15 ^ zmm13))
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
+; AVX512DQ-FCP-NEXT:    vprold $16, %xmm11, %xmm4
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm11 = xmm12[1,1,2,3]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm4 = xmm11[0,1],xmm4[2],xmm11[3,4],xmm4[5],xmm11[6,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm25 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm25, %xmm11
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm24, %zmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm16, %xmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm28, %xmm11
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm11, %xmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm29, %xmm12
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512DQ-FCP-NEXT:    vpshufd {{.*#+}} xmm12 = xmm29[1,1,2,2]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} xmm3 = xmm12[0],xmm3[1],xmm12[2,3],xmm3[4],xmm12[5,6],xmm3[7]
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm11[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm22, %zmm3
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm4 ^ (zmm23 & (zmm3 ^ zmm4))
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm27, %xmm5
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm17, %xmm5
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm8, %xmm5
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm21, %zmm5
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 36(%rax), %ymm4
+; AVX512DQ-FCP-NEXT:    vpbroadcastd 40(%rax), %ymm6
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm4, %zmm4
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm4 = zmm4 ^ (zmm31 & (zmm4 ^ zmm5))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm4 ^ (zmm1 & (zmm4 ^ zmm3))
+; 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 {{.*#+}} ymm2 = mem ^ (ymm1 & (ymm2 ^ mem))
 ; 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 {{.*#+}} ymm6 = ymm6 ^ (ymm1 & (ymm6 ^ ymm13))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm0 = ymm0 ^ (ymm1 & (ymm0 ^ ymm2))
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm4 = ymm4 ^ (ymm2 & (ymm4 ^ ymm16))
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm0[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm7 = ymm7 ^ (ymm2 & (ymm7 ^ ymm14))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm7[0,1,2,3],zmm0[4,5,6,7]
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
 ; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm0 # 32-byte Folded Reload
 ; AVX512DQ-FCP-NEXT:    vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm0 # 64-byte Folded Reload
 ; AVX512DQ-FCP-NEXT:    # zmm0 = (zmm0 & zmm2) | mem
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm2 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm2 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    # zmm2 = (zmm2 & zmm4) | mem
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpckhwd (%rsp), %xmm4, %xmm4 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    # xmm4 = xmm4[4],mem[4],xmm4[5],mem[5],xmm4[6],mem[6],xmm4[7],mem[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm2 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm2 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # zmm2 = (zmm2 & zmm3) | mem
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # xmm3 = 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), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # xmm5 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload
 ; AVX512DQ-FCP-NEXT:    # xmm6 = xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    # xmm7 = xmm7[4],mem[4],xmm7[5],mem[5],xmm7[6],mem[6],xmm7[7],mem[7]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm4, %xmm4
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm25, %xmm7
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm7[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), %ymm8 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # ymm8 = mem[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm7 = ymm8[0,1],ymm7[2],ymm8[3,4],ymm7[5],ymm8[6,7,8,9],ymm7[10],ymm8[11,12],ymm7[13],ymm8[14,15]
 ; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT:    vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    # ymm9 = mem[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7,8,9],ymm8[10],ymm9[11,12],ymm8[13],ymm9[14,15]
-; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT:    vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    # ymm10 = mem[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT:    vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7,8,9],ymm10[10],ymm9[11,12],ymm10[13],ymm9[14,15]
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm10 = [0,1,4,5,4,5,5,7]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT:    vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # ymm9 = mem[1,1,1,1,5,5,5,5]
+; 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:    vpmovsxbd {{.*#+}} ymm9 = [0,1,4,5,4,5,5,7]
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vpermd %ymm11, %ymm10, %ymm10
+; AVX512DQ-FCP-NEXT:    vpermd %ymm11, %ymm9, %ymm9
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm11[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm11[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm11[16,17],zero,zero
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT:    vpandn %ymm10, %ymm12, %ymm10
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm11, %zmm10
-; AVX512DQ-FCP-NEXT:    vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    vpandn %ymm9, %ymm12, %ymm9
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm11, %zmm9
+; AVX512DQ-FCP-NEXT:    vpshuflw $248, (%rsp), %xmm11 # 16-byte Folded Reload
 ; AVX512DQ-FCP-NEXT:    # xmm11 = mem[0,2,3,3,4,5,6,7]
 ; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,1,3]
-; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,0,1,3]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,1,3]
+; AVX512DQ-FCP-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,1,3,2]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
 ; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm12))
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm11 ^ (zmm28 & (zmm4 ^ zmm11))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm11 ^ (zmm23 & (zmm3 ^ zmm11))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm6 ^ (zmm28 & (zmm7 ^ zmm6))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm0 ^ (zmm6 & (zmm4 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm7 = zmm2 ^ (zmm6 & (zmm7 ^ zmm2))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm5 ^ (zmm23 & (zmm6 ^ 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 {{.*#+}} zmm3 = zmm0 ^ (zmm5 & (zmm3 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm6 = zmm2 ^ (zmm5 & (zmm6 ^ zmm2))
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm2, %zmm2
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm2, %zmm2
 ; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm0 ^ (mem & (zmm2 ^ zmm0))
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
 ; AVX512DQ-FCP-NEXT:    vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
 ; AVX512DQ-FCP-NEXT:    # zmm0 = zmm0 | (zmm1 & mem)
-; AVX512DQ-FCP-NEXT:    vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm10 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    # zmm10 = zmm10 | (zmm1 & mem)
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm2))
+; AVX512DQ-FCP-NEXT:    vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm9 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # zmm9 = zmm9 | (zmm1 & mem)
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm2))
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, 320(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 256(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, 192(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, 128(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 320(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 256(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, 192(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, 128(%rax)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, (%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, 448(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 704(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, 640(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, 448(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, 704(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, 640(%rax)
 ; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
 ; AVX512DQ-FCP-NEXT:    vmovaps %zmm1, 576(%rax)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, 384(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 64(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 512(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 512(%rax)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 832(%rax)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 768(%rax)
 ; AVX512DQ-FCP-NEXT:    addq $1496, %rsp # imm = 0x5D8
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
index 4be6ccd2e3575..e03bd7d8e5378 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
@@ -231,18 +231,18 @@ define void @store_i16_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
 ; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512BW-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512BW-NEXT:    vmovdqa (%rcx), %xmm3
-; AVX512BW-NEXT:    vinserti128 $1, (%r11), %ymm3, %ymm3
-; AVX512BW-NEXT:    vinserti128 $1, (%r10), %ymm2, %ymm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]
-; AVX512BW-NEXT:    vinserti128 $1, (%r9), %ymm1, %ymm1
-; AVX512BW-NEXT:    vinserti128 $1, (%r8), %ymm0, %ymm0
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm1 = [0,2,16,18,8,10,24,26,1,3,17,19,9,11,25,27]
-; AVX512BW-NEXT:    vpermi2w %ymm2, %ymm0, %ymm1
-; AVX512BW-NEXT:    vmovdqa %ymm1, (%rax)
+; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512BW-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512BW-NEXT:    vmovdqa (%r11), %xmm3
+; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1]
+; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512BW-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512BW-NEXT:    vmovdqa %ymm2, (%rax)
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
 ;
@@ -252,18 +252,18 @@ define void @store_i16_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
 ; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %xmm3
-; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r11), %ymm3, %ymm3
-; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r10), %ymm2, %ymm2
-; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]
-; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r9), %ymm1, %ymm1
-; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r8), %ymm0, %ymm0
-; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm1 = [0,2,16,18,8,10,24,26,1,3,17,19,9,11,25,27]
-; AVX512BW-FCP-NEXT:    vpermi2w %ymm2, %ymm0, %ymm1
-; AVX512BW-FCP-NEXT:    vmovdqa %ymm1, (%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512BW-FCP-NEXT:    vmovdqa (%r11), %xmm3
+; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1]
+; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512BW-FCP-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512BW-FCP-NEXT:    vmovdqa %ymm2, (%rax)
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
 ;
@@ -273,18 +273,18 @@ define void @store_i16_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512DQ-BW-NEXT:    vmovdqa (%rcx), %xmm3
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r11), %ymm3, %ymm3
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r10), %ymm2, %ymm2
-; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r9), %ymm1, %ymm1
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r8), %ymm0, %ymm0
-; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm1 = [0,2,16,18,8,10,24,26,1,3,17,19,9,11,25,27]
-; AVX512DQ-BW-NEXT:    vpermi2w %ymm2, %ymm0, %ymm1
-; AVX512DQ-BW-NEXT:    vmovdqa %ymm1, (%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512DQ-BW-NEXT:    vmovdqa (%r11), %xmm3
+; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512DQ-BW-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1]
+; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512DQ-BW-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512DQ-BW-NEXT:    vmovdqa %ymm2, (%rax)
 ; AVX512DQ-BW-NEXT:    vzeroupper
 ; AVX512DQ-BW-NEXT:    retq
 ;
@@ -294,18 +294,18 @@ define void @store_i16_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %xmm3
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r11), %ymm3, %ymm3
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r10), %ymm2, %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r9), %ymm1, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r8), %ymm0, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm1 = [0,2,16,18,8,10,24,26,1,3,17,19,9,11,25,27]
-; AVX512DQ-BW-FCP-NEXT:    vpermi2w %ymm2, %ymm0, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm1, (%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %xmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r11), %xmm3
+; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1]
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm2, (%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
   %in.vec0 = load <2 x i16>, ptr %in.vecptr0, align 64
@@ -636,25 +636,24 @@ define void @store_i16_stride8_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [1,3,5,7,1,3,5,7]
-; AVX512-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm2, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm4 = [u,u,u,u,u,u,u,u,0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15]
-; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
-; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm2, %ymm2
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm5 = [0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
-; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [0,2,4,6,0,2,4,6]
-; AVX512-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm3, %ymm1
-; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm3, %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [1,3,9,11,1,3,9,11]
+; AVX512-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vmovdqa %ymm2, %ymm5
+; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm4, %ymm5
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15]
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm5
+; AVX512-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm4
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u]
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm4, %ymm4
+; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm5[2,3],ymm4[4,5],ymm5[6,7]
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm5 = [0,2,8,10,0,2,8,10]
+; AVX512-FCP-NEXT:    # ymm5 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm5, %ymm2
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
+; AVX512-FCP-NEXT:    vpermt2d %ymm1, %ymm5, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
+; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm0
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
@@ -724,25 +723,24 @@ define void @store_i16_stride8_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [1,3,5,7,1,3,5,7]
-; AVX512DQ-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm1, %ymm2, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm4 = [u,u,u,u,u,u,u,u,0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
-; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm5 = [0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [0,2,4,6,0,2,4,6]
-; AVX512DQ-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm1, %ymm3, %ymm1
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [1,3,9,11,1,3,9,11]
+; AVX512DQ-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm5
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm3, %ymm4, %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm5
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm5[2,3],ymm4[4,5],ymm5[6,7]
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm5 = [0,2,8,10,0,2,8,10]
+; AVX512DQ-FCP-NEXT:    # ymm5 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm3, %ymm5, %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm1, %ymm5, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
@@ -764,11 +762,11 @@ define void @store_i16_stride8_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512BW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [0,4,8,12,32,36,40,44,1,5,9,13,33,37,41,45,2,6,10,14,34,38,42,46,3,7,11,15,35,39,43,47]
-; AVX512BW-NEXT:    vpermi2w %zmm2, %zmm0, %zmm1
-; AVX512BW-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [0,4,8,12,32,36,40,44,1,5,9,13,33,37,41,45,2,6,10,14,34,38,42,46,3,7,11,15,35,39,43,47]
+; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
 ;
@@ -789,11 +787,11 @@ define void @store_i16_stride8_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [0,4,8,12,32,36,40,44,1,5,9,13,33,37,41,45,2,6,10,14,34,38,42,46,3,7,11,15,35,39,43,47]
-; AVX512BW-FCP-NEXT:    vpermi2w %zmm2, %zmm0, %zmm1
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [0,4,8,12,32,36,40,44,1,5,9,13,33,37,41,45,2,6,10,14,34,38,42,46,3,7,11,15,35,39,43,47]
+; AVX512BW-FCP-NEXT:    vpermi2w %zmm1, %zmm0, %zmm2
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
 ;
@@ -814,11 +812,11 @@ define void @store_i16_stride8_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [0,4,8,12,32,36,40,44,1,5,9,13,33,37,41,45,2,6,10,14,34,38,42,46,3,7,11,15,35,39,43,47]
-; AVX512DQ-BW-NEXT:    vpermi2w %zmm2, %zmm0, %zmm1
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [0,4,8,12,32,36,40,44,1,5,9,13,33,37,41,45,2,6,10,14,34,38,42,46,3,7,11,15,35,39,43,47]
+; AVX512DQ-BW-NEXT:    vpermi2w %zmm1, %zmm0, %zmm2
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512DQ-BW-NEXT:    vzeroupper
 ; AVX512DQ-BW-NEXT:    retq
 ;
@@ -839,11 +837,11 @@ define void @store_i16_stride8_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [0,4,8,12,32,36,40,44,1,5,9,13,33,37,41,45,2,6,10,14,34,38,42,46,3,7,11,15,35,39,43,47]
-; AVX512DQ-BW-FCP-NEXT:    vpermi2w %zmm2, %zmm0, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [0,4,8,12,32,36,40,44,1,5,9,13,33,37,41,45,2,6,10,14,34,38,42,46,3,7,11,15,35,39,43,47]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2w %zmm1, %zmm0, %zmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
   %in.vec0 = load <4 x i16>, ptr %in.vecptr0, align 64
@@ -4233,203 +4231,142 @@ define void @store_i16_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ;
 ; AVX512-FCP-LABEL: store_i16_stride8_vf32:
 ; AVX512-FCP:       # %bb.0:
-; AVX512-FCP-NEXT:    subq $520, %rsp # imm = 0x208
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512-FCP-NEXT:    vmovdqa (%r10), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa (%rax), %xmm2
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm25
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm1, %xmm26
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa (%rax), %xmm0
+; AVX512-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
 ; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm2
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm27
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm1, %xmm28
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm31
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm1, %xmm19
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm2
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm17
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm1, %xmm18
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm20
-; AVX512-FCP-NEXT:    vmovdqa 32(%r10), %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %ymm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm23
-; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %ymm3
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm4
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm24
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm5
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm6
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm21
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm7
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm8
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm22
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512-FCP-NEXT:    vmovdqa 32(%r10), %xmm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm16
-; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %xmm2
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm1
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm3
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm8[4],ymm7[4],ymm8[5],ymm7[5],ymm8[6],ymm7[6],ymm8[7],ymm7[7],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm5
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm26 = [0,0,0,0,0,0,1,1,2,2,2,2,0,0,3,3]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm26, %zmm25
+; AVX512-FCP-NEXT:    movw $-30584, %r11w # imm = 0x8888
+; AVX512-FCP-NEXT:    kmovw %r11d, %k1
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm26, %zmm25 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm7
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm8
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm9
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm10
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,1,1,1,1,0,0,2,0,3,0,3,0,0,0]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm6, %zmm29
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm24 = [0,0,1,1,1,1,0,0,2,2,3,3,3,3,0,0]
+; AVX512-FCP-NEXT:    movw $8738, %r11w # imm = 0x2222
+; AVX512-FCP-NEXT:    kmovw %r11d, %k2
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm24, %zmm29 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa 32(%r10), %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %ymm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
+; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %ymm12
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm13
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm13[0],ymm12[0],ymm13[1],ymm12[1],ymm13[2],ymm12[2],ymm13[3],ymm12[3],ymm13[8],ymm12[8],ymm13[9],ymm12[9],ymm13[10],ymm12[10],ymm13[11],ymm12[11]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [0,1,4,5,4,5,5,7,2,1,6,5,6,5,7,7]
+; AVX512-FCP-NEXT:    vpermd %zmm11, %zmm16, %zmm27
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm17 = [0,0,4,4,4,4,6,5,0,2,4,6,4,6,6,7]
+; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm17, %zmm27 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm14
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm15
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm1
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[2],ymm15[2],ymm1[3],ymm15[3],ymm1[8],ymm15[8],ymm1[9],ymm15[9],ymm1[10],ymm15[10],ymm1[11],ymm15[11]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm18 = [4,5,1,3,5,7,5,7,6,5,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm18, %zmm30
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm0
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [4,4,2,1,6,5,6,5,4,6,2,3,6,7,6,7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm0[0],ymm14[0],ymm0[1],ymm14[1],ymm0[2],ymm14[2],ymm0[3],ymm14[3],ymm0[8],ymm14[8],ymm0[9],ymm14[9],ymm0[10],ymm14[10],ymm0[11],ymm14[11]
+; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm19, %zmm30 {%k2}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
+; AVX512-FCP-NEXT:    vmovdqa 32(%r10), %xmm11
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm6 = ymm13[4],ymm12[4],ymm13[5],ymm12[5],ymm13[6],ymm12[6],ymm13[7],ymm12[7],ymm13[12],ymm12[12],ymm13[13],ymm12[13],ymm13[14],ymm12[14],ymm13[15],ymm12[15]
+; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %xmm3
+; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm16, %zmm28
+; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm6
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm17, %zmm28 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm2
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm14[4],ymm0[5],ymm14[5],ymm0[6],ymm14[6],ymm0[7],ymm14[7],ymm0[12],ymm14[12],ymm0[13],ymm14[13],ymm0[14],ymm14[14],ymm0[15],ymm14[15]
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm13
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm15[4],ymm1[5],ymm15[5],ymm1[6],ymm15[6],ymm1[7],ymm15[7],ymm1[12],ymm15[12],ymm1[13],ymm15[13],ymm1[14],ymm15[14],ymm1[15],ymm15[15]
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm15
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm18, %zmm31
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm12
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm19, %zmm31 {%k2}
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,0,0,1,0,2,2,2,2,0,0,3,3]
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm1, %zmm21
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm11[0],xmm3[1],xmm11[1],xmm3[2],xmm11[2],xmm3[3],xmm11[3]
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm26, %zmm21 {%k1}
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm12[0],xmm15[0],xmm12[1],xmm15[1],xmm12[2],xmm15[2],xmm12[3],xmm15[3]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,1,0,1,0,0,0,2,2,3,3,3,3,0,0]
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm1, %zmm20
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm0
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm24, %zmm20 {%k2}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm11[4],xmm3[5],xmm11[5],xmm3[6],xmm11[6],xmm3[7],xmm11[7]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm26, %zmm23
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm26, %zmm23 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm12[4],xmm15[4],xmm12[5],xmm15[5],xmm12[6],xmm15[6],xmm12[7],xmm15[7]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm24, %zmm22
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm24, %zmm22 {%k2}
 ; AVX512-FCP-NEXT:    vmovdqa (%r10), %ymm0
 ; AVX512-FCP-NEXT:    vmovdqa (%rax), %ymm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm13 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm11 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512-FCP-NEXT:    vmovdqa (%r9), %ymm0
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm8 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm0
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm0
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm25, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm26, %xmm1
-; 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, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm28, %xmm1
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm25
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm26
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm27
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm28
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm12, %zmm12
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm9, %zmm29
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm10, %zmm10, %zmm30
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm31, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm19, %xmm9
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3]
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm17, %xmm0
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm18, %xmm10
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3]
-; AVX512-FCP-NEXT:    movw $-30584, %ax # imm = 0x8888
-; AVX512-FCP-NEXT:    kmovw %eax, %k1
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,1,4,5,4,5,5,7,10,9,14,13,14,13,15,15]
-; AVX512-FCP-NEXT:    vpermd %zmm24, %zmm31, %zmm24
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,4,4,4,4,6,5,8,10,12,14,12,14,14,15]
-; AVX512-FCP-NEXT:    vpermd %zmm23, %zmm0, %zmm24 {%k1}
-; AVX512-FCP-NEXT:    vpermd %zmm25, %zmm31, %zmm23
-; AVX512-FCP-NEXT:    vpermd %zmm16, %zmm0, %zmm23 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm14 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm16
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm25
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm14
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm15, %zmm15
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm13, %zmm13
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm7, %zmm7
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm6, %zmm6
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm11, %zmm11
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm8, %zmm8
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm5, %zmm5
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm4, %zmm4
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm2, %zmm2
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm9, %zmm9
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm10, %zmm10, %zmm10
-; AVX512-FCP-NEXT:    vpermd %zmm7, %zmm31, %zmm7
-; AVX512-FCP-NEXT:    vpermd %zmm13, %zmm0, %zmm7 {%k1}
-; AVX512-FCP-NEXT:    vpermd %zmm8, %zmm31, %zmm8
-; AVX512-FCP-NEXT:    movw $8738, %ax # imm = 0x2222
-; AVX512-FCP-NEXT:    kmovw %eax, %k2
-; AVX512-FCP-NEXT:    vpermd %zmm11, %zmm0, %zmm8 {%k1}
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [4,5,1,3,5,7,5,7,14,13,11,11,15,15,15,15]
-; AVX512-FCP-NEXT:    vpermd %zmm22, %zmm0, %zmm11
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [4,4,2,1,6,5,6,5,12,14,10,11,14,15,14,15]
-; AVX512-FCP-NEXT:    vpermd %zmm21, %zmm13, %zmm11 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm27, %zmm0, %zmm21
-; AVX512-FCP-NEXT:    vpermd %zmm26, %zmm13, %zmm21 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm3, %zmm0, %zmm3
-; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm13, %zmm3 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpermd %zmm5, %zmm13, %zmm0 {%k2}
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,1,1,10,10,10,10,0,0,11,11]
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm5 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm5 {%k1} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,0,10,10,10,10,0,0,11,11]
-; AVX512-FCP-NEXT:    vpermd %zmm12, %zmm6, %zmm6
-; AVX512-FCP-NEXT:    vpermd %zmm28, %zmm4, %zmm6 {%k1}
-; AVX512-FCP-NEXT:    vpermd %zmm25, %zmm4, %zmm12
-; AVX512-FCP-NEXT:    vpermd %zmm16, %zmm4, %zmm12 {%k1}
-; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm4, %zmm1
-; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm4, %zmm1 {%k1}
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,1,1,1,1,0,0,10,0,11,0,11,0,0,0]
-; AVX512-FCP-NEXT:    vpermd %zmm20, %zmm2, %zmm2
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,1,1,1,1,0,0,10,10,11,11,11,11,0,0]
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm2 {%k2} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,1,0,1,0,0,0,10,10,11,11,11,11,0,0]
-; AVX512-FCP-NEXT:    vpermd %zmm30, %zmm13, %zmm13
-; AVX512-FCP-NEXT:    vpermd %zmm29, %zmm4, %zmm13 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm15, %zmm4, %zmm15
-; AVX512-FCP-NEXT:    vpermd %zmm14, %zmm4, %zmm15 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm10, %zmm4, %zmm10
-; AVX512-FCP-NEXT:    vpermd %zmm9, %zmm4, %zmm10 {%k2}
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %ymm3
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm6
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm6[0],ymm3[0],ymm6[1],ymm3[1],ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[8],ymm3[8],ymm6[9],ymm3[9],ymm6[10],ymm3[10],ymm6[11],ymm3[11]
+; AVX512-FCP-NEXT:    vpermd %zmm11, %zmm16, %zmm11
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm17, %zmm11 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm2
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm12
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm13 = ymm12[0],ymm2[0],ymm12[1],ymm2[1],ymm12[2],ymm2[2],ymm12[3],ymm2[3],ymm12[8],ymm2[8],ymm12[9],ymm2[9],ymm12[10],ymm2[10],ymm12[11],ymm2[11]
+; AVX512-FCP-NEXT:    vpermd %zmm13, %zmm18, %zmm13
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm15
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm14
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[8],ymm15[8],ymm14[9],ymm15[9],ymm14[10],ymm15[10],ymm14[11],ymm15[11]
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm19, %zmm13 {%k2}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm6[4],ymm3[4],ymm6[5],ymm3[5],ymm6[6],ymm3[6],ymm6[7],ymm3[7],ymm6[12],ymm3[12],ymm6[13],ymm3[13],ymm6[14],ymm3[14],ymm6[15],ymm3[15]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm16, %zmm1
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm17, %zmm1 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm14[4],ymm15[4],ymm14[5],ymm15[5],ymm14[6],ymm15[6],ymm14[7],ymm15[7],ymm14[12],ymm15[12],ymm14[13],ymm15[13],ymm14[14],ymm15[14],ymm14[15],ymm15[15]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm12[4],ymm2[4],ymm12[5],ymm2[5],ymm12[6],ymm2[6],ymm12[7],ymm2[7],ymm12[12],ymm2[12],ymm12[13],ymm2[13],ymm12[14],ymm2[14],ymm12[15],ymm2[15]
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm18, %zmm2
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm19, %zmm2 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm3 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm3 = xmm5[0],mem[0],xmm5[1],mem[1],xmm5[2],mem[2],xmm5[3],mem[3]
+; AVX512-FCP-NEXT:    vpermd %zmm3, %zmm26, %zmm3
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm26, %zmm3 {%k1}
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm24, %zmm4
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm24, %zmm4 {%k2}
 ; AVX512-FCP-NEXT:    movb $-86, %al
 ; AVX512-FCP-NEXT:    kmovw %eax, %k1
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm11 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm13 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm3 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm0 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm10 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm29 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm30 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm31 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm20 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm22 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm4 {%k1}
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 192(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 128(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, 320(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 256(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, 448(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, 384(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)
-; AVX512-FCP-NEXT:    addq $520, %rsp # imm = 0x208
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, 320(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, 256(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, 448(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, 384(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, 64(%rax)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -4575,203 +4512,142 @@ define void @store_i16_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ;
 ; AVX512DQ-FCP-LABEL: store_i16_stride8_vf32:
 ; AVX512DQ-FCP:       # %bb.0:
-; AVX512DQ-FCP-NEXT:    subq $520, %rsp # imm = 0x208
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r10), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm25
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm26
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm27
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm28
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm31
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm19
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm17
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm18
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm20
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r10), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %ymm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm23
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm4
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm24
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm5
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm6
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm21
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm7
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm8
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm22
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r10), %xmm0
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm16
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %xmm1
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm3
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm8[4],ymm7[4],ymm8[5],ymm7[5],ymm8[6],ymm7[6],ymm8[7],ymm7[7],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm5
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm26 = [0,0,0,0,0,0,1,1,2,2,2,2,0,0,3,3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm26, %zmm25
+; AVX512DQ-FCP-NEXT:    movw $-30584, %r11w # imm = 0x8888
+; AVX512DQ-FCP-NEXT:    kmovw %r11d, %k1
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm26, %zmm25 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm7
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm8
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm9
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm10
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,1,1,1,1,0,0,2,0,3,0,3,0,0,0]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm6, %zmm29
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm24 = [0,0,1,1,1,1,0,0,2,2,3,3,3,3,0,0]
+; AVX512DQ-FCP-NEXT:    movw $8738, %r11w # imm = 0x2222
+; AVX512DQ-FCP-NEXT:    kmovw %r11d, %k2
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm24, %zmm29 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r10), %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %ymm3
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %ymm12
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm13
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm13[0],ymm12[0],ymm13[1],ymm12[1],ymm13[2],ymm12[2],ymm13[3],ymm12[3],ymm13[8],ymm12[8],ymm13[9],ymm12[9],ymm13[10],ymm12[10],ymm13[11],ymm12[11]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [0,1,4,5,4,5,5,7,2,1,6,5,6,5,7,7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm11, %zmm16, %zmm27
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm17 = [0,0,4,4,4,4,6,5,0,2,4,6,4,6,6,7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm17, %zmm27 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm14
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm15
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm1
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[2],ymm15[2],ymm1[3],ymm15[3],ymm1[8],ymm15[8],ymm1[9],ymm15[9],ymm1[10],ymm15[10],ymm1[11],ymm15[11]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm18 = [4,5,1,3,5,7,5,7,6,5,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm18, %zmm30
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm0
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [4,4,2,1,6,5,6,5,4,6,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm0[0],ymm14[0],ymm0[1],ymm14[1],ymm0[2],ymm14[2],ymm0[3],ymm14[3],ymm0[8],ymm14[8],ymm0[9],ymm14[9],ymm0[10],ymm14[10],ymm0[11],ymm14[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm19, %zmm30 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r10), %xmm11
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm6 = ymm13[4],ymm12[4],ymm13[5],ymm12[5],ymm13[6],ymm12[6],ymm13[7],ymm12[7],ymm13[12],ymm12[12],ymm13[13],ymm12[13],ymm13[14],ymm12[14],ymm13[15],ymm12[15]
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %xmm3
+; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm16, %zmm28
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %xmm6
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm17, %zmm28 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm2
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm14[4],ymm0[5],ymm14[5],ymm0[6],ymm14[6],ymm0[7],ymm14[7],ymm0[12],ymm14[12],ymm0[13],ymm14[13],ymm0[14],ymm14[14],ymm0[15],ymm14[15]
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm13
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm15[4],ymm1[5],ymm15[5],ymm1[6],ymm15[6],ymm1[7],ymm15[7],ymm1[12],ymm15[12],ymm1[13],ymm15[13],ymm1[14],ymm15[14],ymm1[15],ymm15[15]
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm15
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm18, %zmm31
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm12
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm19, %zmm31 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,0,0,1,0,2,2,2,2,0,0,3,3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm1, %zmm21
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm11[0],xmm3[1],xmm11[1],xmm3[2],xmm11[2],xmm3[3],xmm11[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm26, %zmm21 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm12[0],xmm15[0],xmm12[1],xmm15[1],xmm12[2],xmm15[2],xmm12[3],xmm15[3]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,1,0,1,0,0,0,2,2,3,3,3,3,0,0]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm1, %zmm20
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm0
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm24, %zmm20 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm11[4],xmm3[5],xmm11[5],xmm3[6],xmm11[6],xmm3[7],xmm11[7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm26, %zmm23
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm26, %zmm23 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm12[4],xmm15[4],xmm12[5],xmm15[5],xmm12[6],xmm15[6],xmm12[7],xmm15[7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm24, %zmm22
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm24, %zmm22 {%k2}
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r10), %ymm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %ymm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm13 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm11 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm8 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm25, %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm26, %xmm1
-; 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, %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm28, %xmm1
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm25
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm26
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm27
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm28
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm12, %zmm12
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm9, %zmm29
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm10, %zmm10, %zmm30
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm31, %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm19, %xmm9
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm17, %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm18, %xmm10
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3]
-; AVX512DQ-FCP-NEXT:    movw $-30584, %ax # imm = 0x8888
-; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,1,4,5,4,5,5,7,10,9,14,13,14,13,15,15]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm24, %zmm31, %zmm24
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,4,4,4,4,6,5,8,10,12,14,12,14,14,15]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm23, %zmm0, %zmm24 {%k1}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm25, %zmm31, %zmm23
-; AVX512DQ-FCP-NEXT:    vpermd %zmm16, %zmm0, %zmm23 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm14 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm16
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm25
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm14
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm15, %zmm15
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm13, %zmm13
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm7, %zmm7
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm6, %zmm6
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm11, %zmm11
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm8, %zmm8
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm5, %zmm5
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm4, %zmm4
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm9, %zmm9
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm10, %zmm10, %zmm10
-; AVX512DQ-FCP-NEXT:    vpermd %zmm7, %zmm31, %zmm7
-; AVX512DQ-FCP-NEXT:    vpermd %zmm13, %zmm0, %zmm7 {%k1}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm8, %zmm31, %zmm8
-; AVX512DQ-FCP-NEXT:    movw $8738, %ax # imm = 0x2222
-; AVX512DQ-FCP-NEXT:    kmovw %eax, %k2
-; AVX512DQ-FCP-NEXT:    vpermd %zmm11, %zmm0, %zmm8 {%k1}
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [4,5,1,3,5,7,5,7,14,13,11,11,15,15,15,15]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm22, %zmm0, %zmm11
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [4,4,2,1,6,5,6,5,12,14,10,11,14,15,14,15]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm21, %zmm13, %zmm11 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm27, %zmm0, %zmm21
-; AVX512DQ-FCP-NEXT:    vpermd %zmm26, %zmm13, %zmm21 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm3, %zmm0, %zmm3
-; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm13, %zmm3 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpermd %zmm5, %zmm13, %zmm0 {%k2}
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,1,1,10,10,10,10,0,0,11,11]
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm5 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm5 {%k1} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,0,10,10,10,10,0,0,11,11]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm12, %zmm6, %zmm6
-; AVX512DQ-FCP-NEXT:    vpermd %zmm28, %zmm4, %zmm6 {%k1}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm25, %zmm4, %zmm12
-; AVX512DQ-FCP-NEXT:    vpermd %zmm16, %zmm4, %zmm12 {%k1}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm4, %zmm1
-; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm4, %zmm1 {%k1}
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,1,1,1,1,0,0,10,0,11,0,11,0,0,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm20, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,1,1,1,1,0,0,10,10,11,11,11,11,0,0]
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm2 {%k2} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,1,0,1,0,0,0,10,10,11,11,11,11,0,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm30, %zmm13, %zmm13
-; AVX512DQ-FCP-NEXT:    vpermd %zmm29, %zmm4, %zmm13 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm15, %zmm4, %zmm15
-; AVX512DQ-FCP-NEXT:    vpermd %zmm14, %zmm4, %zmm15 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm10, %zmm4, %zmm10
-; AVX512DQ-FCP-NEXT:    vpermd %zmm9, %zmm4, %zmm10 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm6
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm6[0],ymm3[0],ymm6[1],ymm3[1],ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[8],ymm3[8],ymm6[9],ymm3[9],ymm6[10],ymm3[10],ymm6[11],ymm3[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm11, %zmm16, %zmm11
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm17, %zmm11 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm12
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm13 = ymm12[0],ymm2[0],ymm12[1],ymm2[1],ymm12[2],ymm2[2],ymm12[3],ymm2[3],ymm12[8],ymm2[8],ymm12[9],ymm2[9],ymm12[10],ymm2[10],ymm12[11],ymm2[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm13, %zmm18, %zmm13
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm15
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm14
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[8],ymm15[8],ymm14[9],ymm15[9],ymm14[10],ymm15[10],ymm14[11],ymm15[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm19, %zmm13 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm6[4],ymm3[4],ymm6[5],ymm3[5],ymm6[6],ymm3[6],ymm6[7],ymm3[7],ymm6[12],ymm3[12],ymm6[13],ymm3[13],ymm6[14],ymm3[14],ymm6[15],ymm3[15]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm16, %zmm1
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm17, %zmm1 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm14[4],ymm15[4],ymm14[5],ymm15[5],ymm14[6],ymm15[6],ymm14[7],ymm15[7],ymm14[12],ymm15[12],ymm14[13],ymm15[13],ymm14[14],ymm15[14],ymm14[15],ymm15[15]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm12[4],ymm2[4],ymm12[5],ymm2[5],ymm12[6],ymm2[6],ymm12[7],ymm2[7],ymm12[12],ymm2[12],ymm12[13],ymm2[13],ymm12[14],ymm2[14],ymm12[15],ymm2[15]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm18, %zmm2
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm19, %zmm2 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm3 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # xmm3 = xmm5[0],mem[0],xmm5[1],mem[1],xmm5[2],mem[2],xmm5[3],mem[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm3, %zmm26, %zmm3
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm26, %zmm3 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm24, %zmm4
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm24, %zmm4 {%k2}
 ; AVX512DQ-FCP-NEXT:    movb $-86, %al
 ; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm11 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm13 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm3 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm10 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm29 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm30 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm31 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm20 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm22 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm4 {%k1}
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 192(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 128(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, 320(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 256(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, 448(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, 384(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)
-; AVX512DQ-FCP-NEXT:    addq $520, %rsp # imm = 0x208
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, 320(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, 256(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, 448(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, 384(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, 64(%rax)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
 ;
@@ -8175,377 +8051,274 @@ define void @store_i16_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ;
 ; AVX512-FCP-LABEL: store_i16_stride8_vf64:
 ; AVX512-FCP:       # %bb.0:
-; AVX512-FCP-NEXT:    subq $2312, %rsp # imm = 0x908
+; AVX512-FCP-NEXT:    subq $264, %rsp # imm = 0x108
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512-FCP-NEXT:    vmovdqa (%r10), %xmm12
-; AVX512-FCP-NEXT:    vmovdqa 64(%r10), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa (%rax), %xmm1
+; AVX512-FCP-NEXT:    vmovdqa (%r10), %xmm1
 ; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm12[4],xmm1[5],xmm12[5],xmm1[6],xmm12[6],xmm1[7],xmm12[7]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa (%rax), %xmm0
+; AVX512-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
 ; AVX512-FCP-NEXT:    vmovdqa (%r9), %xmm2
 ; AVX512-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa %xmm1, (%rsp) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,0,0,1,1,2,2,2,2,0,0,3,3]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm14, %zmm1
+; AVX512-FCP-NEXT:    movw $-30584, %r11w # imm = 0x8888
+; AVX512-FCP-NEXT:    kmovw %r11d, %k2
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm1 {%k2}
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%r10), %ymm1
-; AVX512-FCP-NEXT:    vmovdqa 96(%rax), %ymm2
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %ymm3
-; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %ymm4
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm5, %zmm5
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm1
+; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-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:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm2
+; AVX512-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm1
+; AVX512-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,1,1,1,1,0,0,2,0,3,0,3,0,0,0]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm2, %zmm17
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,1,1,1,1,0,0,2,2,3,3,3,3,0,0]
+; AVX512-FCP-NEXT:    movw $8738, %r11w # imm = 0x2222
+; AVX512-FCP-NEXT:    kmovw %r11d, %k1
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm7, %zmm17 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa 96(%r10), %ymm5
+; AVX512-FCP-NEXT:    vmovdqa 96(%rax), %ymm4
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11]
+; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %ymm13
+; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %ymm15
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm15[0],ymm13[0],ymm15[1],ymm13[1],ymm15[2],ymm13[2],ymm15[3],ymm13[3],ymm15[8],ymm13[8],ymm15[9],ymm13[9],ymm15[10],ymm13[10],ymm15[11],ymm13[11]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm23 = [0,1,4,5,4,5,5,7,2,1,6,5,6,5,7,7]
+; AVX512-FCP-NEXT:    vpermd %zmm11, %zmm23, %zmm0
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm22 = [0,0,4,4,4,4,6,5,0,2,4,6,4,6,6,7]
+; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm22, %zmm0 {%k2}
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %ymm3
+; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %ymm1
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm20 = [4,5,1,3,5,7,5,7,6,5,3,3,7,7,7,7]
+; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm20, %zmm10
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %ymm0
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm21 = [4,4,2,1,6,5,6,5,4,6,2,3,6,7,6,7]
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[8],ymm3[8],ymm0[9],ymm3[9],ymm0[10],ymm3[10],ymm0[11],ymm3[11]
+; AVX512-FCP-NEXT:    vpermd %zmm9, %zmm21, %zmm10 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm15[4],ymm13[4],ymm15[5],ymm13[5],ymm15[6],ymm13[6],ymm15[7],ymm13[7],ymm15[12],ymm13[12],ymm15[13],ymm13[13],ymm15[14],ymm13[14],ymm15[15],ymm13[15]
+; AVX512-FCP-NEXT:    vpermd %zmm5, %zmm23, %zmm5
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm22, %zmm5 {%k2}
 ; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %ymm5
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %ymm6
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm7, %zmm7
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %ymm7
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %ymm8
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm9, %zmm9
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm3[4],ymm0[5],ymm3[5],ymm0[6],ymm3[6],ymm0[7],ymm3[7],ymm0[12],ymm3[12],ymm0[13],ymm3[13],ymm0[14],ymm3[14],ymm0[15],ymm3[15]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm2[4],ymm1[5],ymm2[5],ymm1[6],ymm2[6],ymm1[7],ymm2[7],ymm1[12],ymm2[12],ymm1[13],ymm2[13],ymm1[14],ymm2[14],ymm1[15],ymm2[15]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm20, %zmm13
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm21, %zmm13 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa 64(%r10), %ymm0
+; AVX512-FCP-NEXT:    vmovdqa 64(%rax), %ymm1
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %ymm3
+; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %ymm4
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512-FCP-NEXT:    vpermd %zmm5, %zmm23, %zmm5
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm22, %zmm5 {%k2}
+; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, (%rsp) # 64-byte Spill
+; AVX512-FCP-NEXT:    vmovdqa 64(%rcx), %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %ymm5
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %ymm9
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm9[0],ymm5[0],ymm9[1],ymm5[1],ymm9[2],ymm5[2],ymm9[3],ymm5[3],ymm9[8],ymm5[8],ymm9[9],ymm5[9],ymm9[10],ymm5[10],ymm9[11],ymm5[11]
+; AVX512-FCP-NEXT:    vpermd %zmm8, %zmm20, %zmm16
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %ymm8
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm8[0],ymm2[0],ymm8[1],ymm2[1],ymm8[2],ymm2[2],ymm8[3],ymm2[3],ymm8[8],ymm2[8],ymm8[9],ymm2[9],ymm8[10],ymm2[10],ymm8[11],ymm2[11]
+; AVX512-FCP-NEXT:    vpermd %zmm12, %zmm21, %zmm16 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm8[4],ymm7[4],ymm8[5],ymm7[5],ymm8[6],ymm7[6],ymm8[7],ymm7[7],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%r10), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa 96(%rax), %xmm2
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm3, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %xmm3
-; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %xmm4
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm5, %zmm5, %zmm5
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %xmm5
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %xmm6
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %xmm7
-; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm8
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 64(%r10), %ymm2
-; AVX512-FCP-NEXT:    vmovdqa 64(%rax), %ymm3
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm25
-; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %ymm4
-; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %ymm5
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm24
-; AVX512-FCP-NEXT:    vmovdqa 64(%rcx), %ymm6
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %ymm7
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %ymm8
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[12],ymm4[12],ymm5[13],ymm4[13],ymm5[14],ymm4[14],ymm5[15],ymm4[15]
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %ymm5
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm5[0],ymm8[0],ymm5[1],ymm8[1],ymm5[2],ymm8[2],ymm5[3],ymm8[3],ymm5[8],ymm8[8],ymm5[9],ymm8[9],ymm5[10],ymm8[10],ymm5[11],ymm8[11]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm5[4],ymm8[4],ymm5[5],ymm8[5],ymm5[6],ymm8[6],ymm5[7],ymm8[7],ymm5[12],ymm8[12],ymm5[13],ymm8[13],ymm5[14],ymm8[14],ymm5[15],ymm8[15]
-; AVX512-FCP-NEXT:    vmovdqa 64(%rax), %xmm8
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm0[4],xmm8[5],xmm0[5],xmm8[6],xmm0[6],xmm8[7],xmm0[7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm7, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %xmm1
-; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %xmm7
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm30
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm4, %zmm18
-; AVX512-FCP-NEXT:    vmovdqa 64(%rcx), %xmm2
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %xmm4
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm5, %zmm31
-; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %xmm2
-; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm3
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm6, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm1
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm7, %zmm7, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm23, %zmm19
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm22, %zmm19 {%k2}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm8[4],ymm2[4],ymm8[5],ymm2[5],ymm8[6],ymm2[6],ymm8[7],ymm2[7],ymm8[12],ymm2[12],ymm8[13],ymm2[13],ymm8[14],ymm2[14],ymm8[15],ymm2[15]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm9[4],ymm5[4],ymm9[5],ymm5[5],ymm9[6],ymm5[6],ymm9[7],ymm5[7],ymm9[12],ymm5[12],ymm9[13],ymm5[13],ymm9[14],ymm5[14],ymm9[15],ymm5[15]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm20, %zmm18
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm21, %zmm18 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa 32(%r10), %ymm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %ymm2
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm17
+; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %ymm1
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
 ; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %ymm3
 ; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm4
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm5, %zmm15
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm5
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm6
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm7, %zmm29
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm7
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm8
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm9, %zmm28
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[12],ymm0[12],ymm2[13],ymm0[13],ymm2[14],ymm0[14],ymm2[15],ymm0[15]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm8[4],ymm7[4],ymm8[5],ymm7[5],ymm8[6],ymm7[6],ymm8[7],ymm7[7],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm9, %zmm13
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm7
+; AVX512-FCP-NEXT:    vpermd %zmm5, %zmm23, %zmm25
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm22, %zmm25 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %ymm2
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm5
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[2],ymm2[2],ymm5[3],ymm2[3],ymm5[8],ymm2[8],ymm5[9],ymm2[9],ymm5[10],ymm2[10],ymm5[11],ymm2[11]
+; AVX512-FCP-NEXT:    vpermd %zmm8, %zmm20, %zmm24
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm8
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm9
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
+; AVX512-FCP-NEXT:    vpermd %zmm12, %zmm21, %zmm24 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm23, %zmm27
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm22, %zmm27 {%k2}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[12],ymm8[12],ymm9[13],ymm8[13],ymm9[14],ymm8[14],ymm9[15],ymm8[15]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm5[4],ymm2[4],ymm5[5],ymm2[5],ymm5[6],ymm2[6],ymm5[7],ymm2[7],ymm5[12],ymm2[12],ymm5[13],ymm2[13],ymm5[14],ymm2[14],ymm5[15],ymm2[15]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm20, %zmm26
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm21, %zmm26 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa (%r10), %ymm0
+; AVX512-FCP-NEXT:    vmovdqa (%r9), %ymm1
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm2
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512-FCP-NEXT:    vpermd %zmm3, %zmm23, %zmm28
+; AVX512-FCP-NEXT:    vmovdqa (%rax), %ymm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm3[0],ymm0[0],ymm3[1],ymm0[1],ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[8],ymm0[8],ymm3[9],ymm0[9],ymm3[10],ymm0[10],ymm3[11],ymm0[11]
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm22, %zmm28 {%k2}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm0[4],ymm3[5],ymm0[5],ymm3[6],ymm0[6],ymm3[7],ymm0[7],ymm3[12],ymm0[12],ymm3[13],ymm0[13],ymm3[14],ymm0[14],ymm3[15],ymm0[15]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm2
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm23, %zmm23
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm1
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm22, %zmm23 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm0
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = 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:    vpermd %zmm3, %zmm20, %zmm22
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm21, %zmm22 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
+; AVX512-FCP-NEXT:    vmovdqa 96(%r9), %xmm1
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm20, %zmm20
+; AVX512-FCP-NEXT:    vmovdqa 96(%r8), %xmm0
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm21, %zmm20 {%k1}
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,0,2,2,2,2,0,0,3,3]
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm3, %zmm21
+; AVX512-FCP-NEXT:    vmovdqa 96(%r10), %xmm2
+; AVX512-FCP-NEXT:    vmovdqa 96(%rax), %xmm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm14, %zmm21 {%k2}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512-FCP-NEXT:    vmovdqa 64(%r9), %xmm1
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm29
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm14, %zmm29 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa 64(%r8), %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:    vpermd %zmm2, %zmm14, %zmm30
+; AVX512-FCP-NEXT:    vmovdqa 64(%r10), %xmm2
+; AVX512-FCP-NEXT:    vmovdqa 64(%rax), %xmm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm14, %zmm30 {%k2}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm31
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm14, %zmm31 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm0
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm2
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512-FCP-NEXT:    vpermd %zmm1, %zmm14, %zmm6
 ; AVX512-FCP-NEXT:    vmovdqa 32(%r10), %xmm3
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rax), %xmm4
 ; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpermd %zmm5, %zmm14, %zmm6 {%k2}
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm22
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm9
-; AVX512-FCP-NEXT:    vmovdqa 32(%r9), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %xmm2
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
 ; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm5, %zmm5, %zmm27
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm4, %zmm4, %zmm26
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm2
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm4
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm3, %zmm3, %zmm23
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm21
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm2
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vmovdqa (%r10), %ymm5
-; AVX512-FCP-NEXT:    vmovdqa (%rax), %ymm3
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm5[0],ymm3[1],ymm5[1],ymm3[2],ymm5[2],ymm3[3],ymm5[3],ymm3[8],ymm5[8],ymm3[9],ymm5[9],ymm3[10],ymm5[10],ymm3[11],ymm5[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm6
-; AVX512-FCP-NEXT:    vmovdqa (%r9), %ymm8
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm10
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm10[0],ymm8[0],ymm10[1],ymm8[1],ymm10[2],ymm8[2],ymm10[3],ymm8[3],ymm10[8],ymm8[8],ymm10[9],ymm8[9],ymm10[10],ymm8[10],ymm10[11],ymm8[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm4, %zmm19
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm11
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm14
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm14[0],ymm11[0],ymm14[1],ymm11[1],ymm14[2],ymm11[2],ymm14[3],ymm11[3],ymm14[8],ymm11[8],ymm14[9],ymm11[9],ymm14[10],ymm11[10],ymm14[11],ymm11[11]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm4, %zmm20
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm1
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm0
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = 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:    vinserti64x4 $1, %ymm4, %zmm4, %zmm16
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm3[4],ymm5[4],ymm3[5],ymm5[5],ymm3[6],ymm5[6],ymm3[7],ymm5[7],ymm3[12],ymm5[12],ymm3[13],ymm5[13],ymm3[14],ymm5[14],ymm3[15],ymm5[15]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm4
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm10[4],ymm8[4],ymm10[5],ymm8[5],ymm10[6],ymm8[6],ymm10[7],ymm8[7],ymm10[12],ymm8[12],ymm10[13],ymm8[13],ymm10[14],ymm8[14],ymm10[15],ymm8[15]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm2
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm14[4],ymm11[4],ymm14[5],ymm11[5],ymm14[6],ymm11[6],ymm14[7],ymm11[7],ymm14[12],ymm11[12],ymm14[13],ymm11[13],ymm14[14],ymm11[14],ymm14[15],ymm11[15]
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm12[0],xmm1[1],xmm12[1],xmm1[2],xmm12[2],xmm1[3],xmm12[3]
-; AVX512-FCP-NEXT:    vmovdqa (%rsp), %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:    vinserti64x4 $1, %ymm3, %zmm3, %zmm10
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm5
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm3
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm14
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm12
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm0
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm1
-; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    movw $-30584, %ax # imm = 0x8888
-; AVX512-FCP-NEXT:    kmovw %eax, %k2
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,1,4,5,4,5,5,7,10,9,14,13,14,13,15,15]
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,4,4,4,6,5,8,10,12,14,12,14,14,15]
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 {%k2} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, (%rsp) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 {%k2} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd %zmm24, %zmm0, %zmm24
-; AVX512-FCP-NEXT:    vpermd %zmm25, %zmm1, %zmm24 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm18, %zmm0, %zmm25
-; AVX512-FCP-NEXT:    vpermd %zmm30, %zmm1, %zmm25 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm15, %zmm0, %zmm8
-; AVX512-FCP-NEXT:    vpermd %zmm17, %zmm1, %zmm8 {%k2}
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpermd %zmm7, %zmm0, %zmm30
-; AVX512-FCP-NEXT:    vpermd %zmm13, %zmm1, %zmm30 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm19, %zmm0, %zmm7
-; AVX512-FCP-NEXT:    vpermd %zmm6, %zmm1, %zmm7 {%k2}
-; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm0, %zmm19
-; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm1, %zmm19 {%k2}
-; AVX512-FCP-NEXT:    movw $8738, %ax # imm = 0x2222
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm1
+; AVX512-FCP-NEXT:    vpermd %zmm3, %zmm14, %zmm1 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm2 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm3 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512-FCP-NEXT:    vpermd %zmm3, %zmm14, %zmm9
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm14, %zmm9 {%k2}
+; AVX512-FCP-NEXT:    vmovdqa 96(%rsi), %xmm2
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm3
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,1,0,1,0,0,0,2,2,3,3,3,3,0,0]
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm5, %zmm8
+; AVX512-FCP-NEXT:    vmovdqa 96(%rcx), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa 96(%rdx), %xmm5
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512-FCP-NEXT:    vpermd %zmm12, %zmm7, %zmm8 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512-FCP-NEXT:    vpermd %zmm2, %zmm7, %zmm2
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm7, %zmm2 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa 64(%rsi), %xmm3
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm4
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT:    vpermd %zmm5, %zmm7, %zmm5
+; AVX512-FCP-NEXT:    vmovdqa 64(%rcx), %xmm12
+; AVX512-FCP-NEXT:    vmovdqa 64(%rdx), %xmm14
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512-FCP-NEXT:    vpermd %zmm11, %zmm7, %zmm5 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512-FCP-NEXT:    vpermd %zmm3, %zmm7, %zmm3
+; AVX512-FCP-NEXT:    vpermd %zmm11, %zmm7, %zmm3 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm4
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm11
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm11[0],xmm4[0],xmm11[1],xmm4[1],xmm11[2],xmm4[2],xmm11[3],xmm4[3]
+; AVX512-FCP-NEXT:    vpermd %zmm12, %zmm7, %zmm12
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm14
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm0
+; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
+; AVX512-FCP-NEXT:    vpermd %zmm15, %zmm7, %zmm12 {%k1}
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm14[4],xmm0[5],xmm14[5],xmm0[6],xmm14[6],xmm0[7],xmm14[7]
+; AVX512-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm11[4],xmm4[4],xmm11[5],xmm4[5],xmm11[6],xmm4[6],xmm11[7],xmm4[7]
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm7, %zmm4
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm7, %zmm4 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX512-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
+; AVX512-FCP-NEXT:    # xmm11 = xmm11[0],mem[0],xmm11[1],mem[1],xmm11[2],mem[2],xmm11[3],mem[3]
+; AVX512-FCP-NEXT:    vpermd %zmm11, %zmm7, %zmm11
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm7, %zmm11 {%k1}
+; AVX512-FCP-NEXT:    movb $-86, %al
 ; AVX512-FCP-NEXT:    kmovw %eax, %k1
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [4,5,1,3,5,7,5,7,14,13,11,11,15,15,15,15]
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm6 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [4,4,2,1,6,5,6,5,12,14,10,11,14,15,14,15]
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm6 {%k1} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm7 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm7 {%k1} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm13 {%k1} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd %zmm31, %zmm0, %zmm15
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm15 {%k1} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd %zmm28, %zmm0, %zmm17
-; AVX512-FCP-NEXT:    vpermd %zmm29, %zmm1, %zmm17 {%k1}
-; AVX512-FCP-NEXT:    vpermd %zmm9, %zmm0, %zmm9
-; AVX512-FCP-NEXT:    vpermd %zmm22, %zmm1, %zmm9 {%k1}
-; AVX512-FCP-NEXT:    vpermd %zmm16, %zmm0, %zmm4
-; AVX512-FCP-NEXT:    vpermd %zmm20, %zmm1, %zmm4 {%k1}
-; AVX512-FCP-NEXT:    vpermd %zmm5, %zmm0, %zmm5
-; AVX512-FCP-NEXT:    vpermd %zmm10, %zmm1, %zmm5 {%k1}
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,10,10,10,10,0,0,11,11]
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 {%k2} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,1,0,10,10,10,10,0,0,11,11]
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm10 {%k2} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm16 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm16 {%k2} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm20 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm20 {%k2} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm22 # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm22 {%k2} # 64-byte Folded Reload
-; AVX512-FCP-NEXT:    vpermd %zmm26, %zmm0, %zmm26
-; AVX512-FCP-NEXT:    vpermd %zmm27, %zmm0, %zmm26 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm21, %zmm0, %zmm21
-; AVX512-FCP-NEXT:    vpermd %zmm23, %zmm0, %zmm21 {%k2}
-; AVX512-FCP-NEXT:    vpermd %zmm12, %zmm0, %zmm23
-; AVX512-FCP-NEXT:    vpermd %zmm14, %zmm0, %zmm23 {%k2}
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm11, %zmm11, %zmm0
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,1,1,1,1,0,0,10,0,11,0,11,0,0,0]
-; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm11, %zmm12
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,1,1,1,1,0,0,10,10,11,11,11,11,0,0]
-; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm11, %zmm12 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm17 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,1,0,1,0,0,0,10,10,11,11,11,11,0,0]
-; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm0
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm0 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm27
-; AVX512-FCP-NEXT:    vpermd %zmm27, %zmm11, %zmm27
-; AVX512-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm27 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm28
-; AVX512-FCP-NEXT:    vpermd %zmm28, %zmm11, %zmm28
-; AVX512-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm28 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm29
-; AVX512-FCP-NEXT:    vpermd %zmm29, %zmm11, %zmm29
-; AVX512-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm29 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm31
-; AVX512-FCP-NEXT:    vpermd %zmm31, %zmm11, %zmm31
-; AVX512-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm31 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm8
-; AVX512-FCP-NEXT:    vpermd %zmm8, %zmm11, %zmm8
-; AVX512-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm8 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm14
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm2, %zmm18
-; AVX512-FCP-NEXT:    vpermd %zmm18, %zmm11, %zmm18
-; AVX512-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm18 {%k1}
-; AVX512-FCP-NEXT:    movb $-86, %al
-; AVX512-FCP-NEXT:    kmovw %eax, %k1
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm7 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm0 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm27 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm15 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm28 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm29 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm17 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm9 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm31 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k1}
+; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm8 {%k1}
-; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm2 {%k1}
+; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm16 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm5 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm24 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm26 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm12 {%k1}
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm4 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm5 {%k1}
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm18 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm22 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm11 {%k1}
 ; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, (%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 192(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, 320(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, 256(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, 448(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, 384(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, 576(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, 512(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, 704(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 640(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, 832(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 768(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 960(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 896(%rax)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, 64(%rax)
-; AVX512-FCP-NEXT:    addq $2312, %rsp # imm = 0x908
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, (%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, 192(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, 128(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 320(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, 256(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, 448(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, 384(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 576(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 512(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, 704(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, 640(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 832(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, 768(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 960(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 896(%rax)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, 64(%rax)
+; AVX512-FCP-NEXT:    addq $264, %rsp # imm = 0x108
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -8825,377 +8598,274 @@ define void @store_i16_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
 ;
 ; AVX512DQ-FCP-LABEL: store_i16_stride8_vf64:
 ; AVX512DQ-FCP:       # %bb.0:
-; AVX512DQ-FCP-NEXT:    subq $2312, %rsp # imm = 0x908
+; AVX512DQ-FCP-NEXT:    subq $264, %rsp # imm = 0x108
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r10), %xmm12
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r10), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r10), %xmm1
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm12[4],xmm1[5],xmm12[5],xmm1[6],xmm12[6],xmm1[7],xmm12[7]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %xmm2
 ; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, (%rsp) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,0,0,1,1,2,2,2,2,0,0,3,3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm14, %zmm1
+; AVX512DQ-FCP-NEXT:    movw $-30584, %r11w # imm = 0x8888
+; AVX512DQ-FCP-NEXT:    kmovw %r11d, %k2
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm1 {%k2}
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r10), %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rax), %ymm2
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r8), %ymm4
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm5, %zmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-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:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,1,1,1,1,0,0,2,0,3,0,3,0,0,0]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm2, %zmm17
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,1,1,1,1,0,0,2,2,3,3,3,3,0,0]
+; AVX512DQ-FCP-NEXT:    movw $8738, %r11w # imm = 0x2222
+; AVX512DQ-FCP-NEXT:    kmovw %r11d, %k1
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm7, %zmm17 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r10), %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rax), %ymm4
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %ymm13
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r8), %ymm15
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm15[0],ymm13[0],ymm15[1],ymm13[1],ymm15[2],ymm13[2],ymm15[3],ymm13[3],ymm15[8],ymm13[8],ymm15[9],ymm13[9],ymm15[10],ymm13[10],ymm15[11],ymm13[11]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm23 = [0,1,4,5,4,5,5,7,2,1,6,5,6,5,7,7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm11, %zmm23, %zmm0
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm22 = [0,0,4,4,4,4,6,5,0,2,4,6,4,6,6,7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm22, %zmm0 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %ymm1
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm20 = [4,5,1,3,5,7,5,7,6,5,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm20, %zmm10
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %ymm0
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm21 = [4,4,2,1,6,5,6,5,4,6,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[8],ymm3[8],ymm0[9],ymm3[9],ymm0[10],ymm3[10],ymm0[11],ymm3[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm9, %zmm21, %zmm10 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm15[4],ymm13[4],ymm15[5],ymm13[5],ymm15[6],ymm13[6],ymm15[7],ymm13[7],ymm15[12],ymm13[12],ymm15[13],ymm13[13],ymm15[14],ymm13[14],ymm15[15],ymm13[15]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm5, %zmm23, %zmm5
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm22, %zmm5 {%k2}
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %ymm5
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %ymm6
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm7, %zmm7
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %ymm7
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %ymm8
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm9, %zmm9
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm3[4],ymm0[5],ymm3[5],ymm0[6],ymm3[6],ymm0[7],ymm3[7],ymm0[12],ymm3[12],ymm0[13],ymm3[13],ymm0[14],ymm3[14],ymm0[15],ymm3[15]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm2[4],ymm1[5],ymm2[5],ymm1[6],ymm2[6],ymm1[7],ymm2[7],ymm1[12],ymm2[12],ymm1[13],ymm2[13],ymm1[14],ymm2[14],ymm1[15],ymm2[15]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm20, %zmm13
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm21, %zmm13 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r10), %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rax), %ymm1
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r8), %ymm4
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm5, %zmm23, %zmm5
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm22, %zmm5 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, (%rsp) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rcx), %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %ymm9
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm9[0],ymm5[0],ymm9[1],ymm5[1],ymm9[2],ymm5[2],ymm9[3],ymm5[3],ymm9[8],ymm5[8],ymm9[9],ymm5[9],ymm9[10],ymm5[10],ymm9[11],ymm5[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm8, %zmm20, %zmm16
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %ymm8
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm8[0],ymm2[0],ymm8[1],ymm2[1],ymm8[2],ymm2[2],ymm8[3],ymm2[3],ymm8[8],ymm2[8],ymm8[9],ymm2[9],ymm8[10],ymm2[10],ymm8[11],ymm2[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm12, %zmm21, %zmm16 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm8[4],ymm7[4],ymm8[5],ymm7[5],ymm8[6],ymm7[6],ymm8[7],ymm7[7],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r10), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rax), %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm3, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %xmm3
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r8), %xmm4
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm5, %zmm5, %zmm5
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %xmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %xmm6
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %xmm7
-; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm8
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r10), %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rax), %ymm3
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm25
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %ymm4
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r8), %ymm5
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm24
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rcx), %ymm6
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %ymm7
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %ymm8
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[12],ymm4[12],ymm5[13],ymm4[13],ymm5[14],ymm4[14],ymm5[15],ymm4[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %ymm5
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm5[0],ymm8[0],ymm5[1],ymm8[1],ymm5[2],ymm8[2],ymm5[3],ymm8[3],ymm5[8],ymm8[8],ymm5[9],ymm8[9],ymm5[10],ymm8[10],ymm5[11],ymm8[11]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm5[4],ymm8[4],ymm5[5],ymm8[5],ymm5[6],ymm8[6],ymm5[7],ymm8[7],ymm5[12],ymm8[12],ymm5[13],ymm8[13],ymm5[14],ymm8[14],ymm5[15],ymm8[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rax), %xmm8
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm0[4],xmm8[5],xmm0[5],xmm8[6],xmm0[6],xmm8[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm7, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %xmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r8), %xmm7
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm30
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm4, %zmm18
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rcx), %xmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %xmm4
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm5, %zmm31
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %xmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm3
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm6, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm7, %zmm7, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm23, %zmm19
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm22, %zmm19 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm8[4],ymm2[4],ymm8[5],ymm2[5],ymm8[6],ymm2[6],ymm8[7],ymm2[7],ymm8[12],ymm2[12],ymm8[13],ymm2[13],ymm8[14],ymm2[14],ymm8[15],ymm2[15]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm9[4],ymm5[4],ymm9[5],ymm5[5],ymm9[6],ymm5[6],ymm9[7],ymm5[7],ymm9[12],ymm5[12],ymm9[13],ymm5[13],ymm9[14],ymm5[14],ymm9[15],ymm5[15]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm20, %zmm18
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm21, %zmm18 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r10), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %ymm2
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm17
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %ymm1
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %ymm3
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm4
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm5, %zmm15
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm5
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm6
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm7, %zmm29
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm7
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm8
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm9, %zmm28
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[12],ymm0[12],ymm2[13],ymm0[13],ymm2[14],ymm0[14],ymm2[15],ymm0[15]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm8[4],ymm7[4],ymm8[5],ymm7[5],ymm8[6],ymm7[6],ymm8[7],ymm7[7],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm9, %zmm13
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm7
+; AVX512DQ-FCP-NEXT:    vpermd %zmm5, %zmm23, %zmm25
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm22, %zmm25 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm5
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[2],ymm2[2],ymm5[3],ymm2[3],ymm5[8],ymm2[8],ymm5[9],ymm2[9],ymm5[10],ymm2[10],ymm5[11],ymm2[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm8, %zmm20, %zmm24
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm8
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm9
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm12, %zmm21, %zmm24 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm23, %zmm27
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm22, %zmm27 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[12],ymm8[12],ymm9[13],ymm8[13],ymm9[14],ymm8[14],ymm9[15],ymm8[15]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm5[4],ymm2[4],ymm5[5],ymm2[5],ymm5[6],ymm2[6],ymm5[7],ymm2[7],ymm5[12],ymm2[12],ymm5[13],ymm2[13],ymm5[14],ymm2[14],ymm5[15],ymm2[15]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm20, %zmm26
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm21, %zmm26 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r10), %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm2
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm3, %zmm23, %zmm28
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %ymm3
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm3[0],ymm0[0],ymm3[1],ymm0[1],ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[8],ymm0[8],ymm3[9],ymm0[9],ymm3[10],ymm0[10],ymm3[11],ymm0[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm22, %zmm28 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm0[4],ymm3[5],ymm0[5],ymm3[6],ymm0[6],ymm3[7],ymm0[7],ymm3[12],ymm0[12],ymm3[13],ymm0[13],ymm3[14],ymm0[14],ymm3[15],ymm0[15]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm2
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm23, %zmm23
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm1
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm22, %zmm23 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm0
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm3 = 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:    vpermd %zmm3, %zmm20, %zmm22
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm3
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm21, %zmm22 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r9), %xmm1
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm20, %zmm20
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r8), %xmm0
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm21, %zmm20 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,0,2,2,2,2,0,0,3,3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm3, %zmm21
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%r10), %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rax), %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm14, %zmm21 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r9), %xmm1
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm29
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm14, %zmm29 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r8), %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:    vpermd %zmm2, %zmm14, %zmm30
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%r10), %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rax), %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm14, %zmm30 {%k2}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm31
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm14, %zmm31 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %xmm0
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm2
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm1, %zmm14, %zmm6
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r10), %xmm3
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rax), %xmm4
 ; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm5, %zmm14, %zmm6 {%k2}
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm22
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm9
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r9), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm5, %zmm5, %zmm27
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm4, %zmm4, %zmm26
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm4
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm3, %zmm3, %zmm23
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm21
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm2
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r10), %ymm5
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rax), %ymm3
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm5[0],ymm3[1],ymm5[1],ymm3[2],ymm5[2],ymm3[3],ymm5[3],ymm3[8],ymm5[8],ymm3[9],ymm5[9],ymm3[10],ymm5[10],ymm3[11],ymm5[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm6
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r9), %ymm8
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm10
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm10[0],ymm8[0],ymm10[1],ymm8[1],ymm10[2],ymm8[2],ymm10[3],ymm8[3],ymm10[8],ymm8[8],ymm10[9],ymm8[9],ymm10[10],ymm8[10],ymm10[11],ymm8[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm4, %zmm19
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm11
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm14
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm14[0],ymm11[0],ymm14[1],ymm11[1],ymm14[2],ymm11[2],ymm14[3],ymm11[3],ymm14[8],ymm11[8],ymm14[9],ymm11[9],ymm14[10],ymm11[10],ymm14[11],ymm11[11]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm4, %zmm20
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm0
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} ymm4 = 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:    vinserti64x4 $1, %ymm4, %zmm4, %zmm16
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm3[4],ymm5[4],ymm3[5],ymm5[5],ymm3[6],ymm5[6],ymm3[7],ymm5[7],ymm3[12],ymm5[12],ymm3[13],ymm5[13],ymm3[14],ymm5[14],ymm3[15],ymm5[15]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm4
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm10[4],ymm8[4],ymm10[5],ymm8[5],ymm10[6],ymm8[6],ymm10[7],ymm8[7],ymm10[12],ymm8[12],ymm10[13],ymm8[13],ymm10[14],ymm8[14],ymm10[15],ymm8[15]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm3, %zmm2
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm14[4],ymm11[4],ymm14[5],ymm11[5],ymm14[6],ymm11[6],ymm14[7],ymm11[7],ymm14[12],ymm11[12],ymm14[13],ymm11[13],ymm14[14],ymm11[14],ymm14[15],ymm11[15]
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm12[0],xmm1[1],xmm12[1],xmm1[2],xmm12[2],xmm1[3],xmm12[3]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsp), %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:    vinserti64x4 $1, %ymm3, %zmm3, %zmm10
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm3
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm14
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm12
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm1
-; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    movw $-30584, %ax # imm = 0x8888
-; AVX512DQ-FCP-NEXT:    kmovw %eax, %k2
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,1,4,5,4,5,5,7,10,9,14,13,14,13,15,15]
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,4,4,4,6,5,8,10,12,14,12,14,14,15]
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 {%k2} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, (%rsp) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 {%k2} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd %zmm24, %zmm0, %zmm24
-; AVX512DQ-FCP-NEXT:    vpermd %zmm25, %zmm1, %zmm24 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm18, %zmm0, %zmm25
-; AVX512DQ-FCP-NEXT:    vpermd %zmm30, %zmm1, %zmm25 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm15, %zmm0, %zmm8
-; AVX512DQ-FCP-NEXT:    vpermd %zmm17, %zmm1, %zmm8 {%k2}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpermd %zmm7, %zmm0, %zmm30
-; AVX512DQ-FCP-NEXT:    vpermd %zmm13, %zmm1, %zmm30 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm19, %zmm0, %zmm7
-; AVX512DQ-FCP-NEXT:    vpermd %zmm6, %zmm1, %zmm7 {%k2}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm0, %zmm19
-; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm1, %zmm19 {%k2}
-; AVX512DQ-FCP-NEXT:    movw $8738, %ax # imm = 0x2222
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm1
+; AVX512DQ-FCP-NEXT:    vpermd %zmm3, %zmm14, %zmm1 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # xmm2 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # xmm3 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm3, %zmm14, %zmm9
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm14, %zmm9 {%k2}
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rsi), %xmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm3
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,1,0,1,0,0,0,2,2,3,3,3,3,0,0]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm5, %zmm8
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rcx), %xmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdx), %xmm5
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm12, %zmm7, %zmm8 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm2, %zmm7, %zmm2
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm7, %zmm2 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rsi), %xmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm4
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm5, %zmm7, %zmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rcx), %xmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdx), %xmm14
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm11, %zmm7, %zmm5 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm3, %zmm7, %zmm3
+; AVX512DQ-FCP-NEXT:    vpermd %zmm11, %zmm7, %zmm3 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm4
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm11
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm11[0],xmm4[0],xmm11[1],xmm4[1],xmm11[2],xmm4[2],xmm11[3],xmm4[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm12, %zmm7, %zmm12
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm14
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm0
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm15, %zmm7, %zmm12 {%k1}
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm14[4],xmm0[5],xmm14[5],xmm0[6],xmm14[6],xmm0[7],xmm14[7]
+; AVX512DQ-FCP-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm11[4],xmm4[4],xmm11[5],xmm4[5],xmm11[6],xmm4[6],xmm11[7],xmm4[7]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm7, %zmm4
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm7, %zmm4 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX512DQ-FCP-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT:    # xmm11 = xmm11[0],mem[0],xmm11[1],mem[1],xmm11[2],mem[2],xmm11[3],mem[3]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm11, %zmm7, %zmm11
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm7, %zmm11 {%k1}
+; AVX512DQ-FCP-NEXT:    movb $-86, %al
 ; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [4,5,1,3,5,7,5,7,14,13,11,11,15,15,15,15]
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm6 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [4,4,2,1,6,5,6,5,12,14,10,11,14,15,14,15]
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm6 {%k1} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm7 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm7 {%k1} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm13 {%k1} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd %zmm31, %zmm0, %zmm15
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm15 {%k1} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd %zmm28, %zmm0, %zmm17
-; AVX512DQ-FCP-NEXT:    vpermd %zmm29, %zmm1, %zmm17 {%k1}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm9, %zmm0, %zmm9
-; AVX512DQ-FCP-NEXT:    vpermd %zmm22, %zmm1, %zmm9 {%k1}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm16, %zmm0, %zmm4
-; AVX512DQ-FCP-NEXT:    vpermd %zmm20, %zmm1, %zmm4 {%k1}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm5, %zmm0, %zmm5
-; AVX512DQ-FCP-NEXT:    vpermd %zmm10, %zmm1, %zmm5 {%k1}
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,10,10,10,10,0,0,11,11]
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 {%k2} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,1,0,10,10,10,10,0,0,11,11]
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm10 {%k2} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm16 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm16 {%k2} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm20 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm20 {%k2} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm22 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm22 {%k2} # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT:    vpermd %zmm26, %zmm0, %zmm26
-; AVX512DQ-FCP-NEXT:    vpermd %zmm27, %zmm0, %zmm26 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm21, %zmm0, %zmm21
-; AVX512DQ-FCP-NEXT:    vpermd %zmm23, %zmm0, %zmm21 {%k2}
-; AVX512DQ-FCP-NEXT:    vpermd %zmm12, %zmm0, %zmm23
-; AVX512DQ-FCP-NEXT:    vpermd %zmm14, %zmm0, %zmm23 {%k2}
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm11, %zmm11, %zmm0
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,1,1,1,1,0,0,10,0,11,0,11,0,0,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm11, %zmm12
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,1,1,1,1,0,0,10,10,11,11,11,11,0,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm11, %zmm12 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm17 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,1,0,1,0,0,0,10,10,11,11,11,11,0,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm14, %zmm0
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512DQ-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm27
-; AVX512DQ-FCP-NEXT:    vpermd %zmm27, %zmm11, %zmm27
-; AVX512DQ-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm27 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm28
-; AVX512DQ-FCP-NEXT:    vpermd %zmm28, %zmm11, %zmm28
-; AVX512DQ-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm28 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm29
-; AVX512DQ-FCP-NEXT:    vpermd %zmm29, %zmm11, %zmm29
-; AVX512DQ-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm29 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm31
-; AVX512DQ-FCP-NEXT:    vpermd %zmm31, %zmm11, %zmm31
-; AVX512DQ-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm31 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm14
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm8
-; AVX512DQ-FCP-NEXT:    vpermd %zmm8, %zmm11, %zmm8
-; AVX512DQ-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm8 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm14
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm2, %zmm18
-; AVX512DQ-FCP-NEXT:    vpermd %zmm18, %zmm11, %zmm18
-; AVX512DQ-FCP-NEXT:    vpermd %zmm14, %zmm11, %zmm18 {%k1}
-; AVX512DQ-FCP-NEXT:    movb $-86, %al
-; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm7 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm27 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm15 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm28 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm29 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm17 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, %zmm9 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm31 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm8 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm2 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm16 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, %zmm5 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm24 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm26 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm12 {%k1}
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm4 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm5 {%k1}
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm18 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm22 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm11 {%k1}
 ; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, (%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 192(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, 320(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, 256(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 448(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, 384(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, 576(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, 512(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, 704(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 640(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, 832(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 768(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 960(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 896(%rax)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, 64(%rax)
-; AVX512DQ-FCP-NEXT:    addq $2312, %rsp # imm = 0x908
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, (%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, 192(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, 128(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 320(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, 256(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, 448(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, 384(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 576(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 512(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, 704(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, 640(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 832(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, 768(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 960(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, 896(%rax)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, 64(%rax)
+; AVX512DQ-FCP-NEXT:    addq $264, %rsp # imm = 0x108
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
 ;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i32-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i32-stride-8.ll
index 8d74c175f4bd2..fba3aa50af23d 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i32-stride-8.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i32-stride-8.ll
@@ -172,11 +172,11 @@ define void @store_i32_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512-NEXT:    vpermi2d %zmm2, %zmm0, %zmm1
-; AVX512-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
@@ -197,11 +197,11 @@ define void @store_i32_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512-FCP-NEXT:    vpermi2d %zmm2, %zmm0, %zmm1
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -222,11 +222,11 @@ define void @store_i32_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512DQ-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512DQ-NEXT:    vpermi2d %zmm2, %zmm0, %zmm1
-; AVX512DQ-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
@@ -247,11 +247,11 @@ define void @store_i32_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512DQ-FCP-NEXT:    vpermi2d %zmm2, %zmm0, %zmm1
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
 ;
@@ -272,11 +272,11 @@ define void @store_i32_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512BW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512BW-NEXT:    vpermi2d %zmm2, %zmm0, %zmm1
-; AVX512BW-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
 ;
@@ -297,11 +297,11 @@ define void @store_i32_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512BW-FCP-NEXT:    vpermi2d %zmm2, %zmm0, %zmm1
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
 ;
@@ -322,11 +322,11 @@ define void @store_i32_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512DQ-BW-NEXT:    vpermi2d %zmm2, %zmm0, %zmm1
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512DQ-BW-NEXT:    vzeroupper
 ; AVX512DQ-BW-NEXT:    retq
 ;
@@ -347,11 +347,11 @@ define void @store_i32_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm2, %zmm0, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
   %in.vec0 = load <2 x i32>, ptr %in.vecptr0, align 64
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
index 01a2f7f46f939..07be0ae8b41d5 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
@@ -1368,9 +1368,8 @@ define void @store_i8_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[2,6,u],zero,zero,ymm5[3,7,u],zero,zero,ymm5[8,12,u],zero,zero,ymm5[9,17,u],zero,zero,ymm5[22,18,u],zero,zero,ymm5[23,19,u],zero,zero,ymm5[24,28]
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
 ; AVX512-FCP-NEXT:    vporq %zmm3, %zmm4, %zmm3
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm4
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm5, %zmm4
+; 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 %zmm0, %zmm4, %zmm4
 ; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm3))
 ; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[1,3,2,3]
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[u],zero,zero,xmm2[5,13,u],zero,zero,xmm2[6,14,u],zero,zero,xmm2[7,15,u]
@@ -1444,9 +1443,8 @@ define void @store_i8_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[2,6,u],zero,zero,ymm5[3,7,u],zero,zero,ymm5[8,12,u],zero,zero,ymm5[9,17,u],zero,zero,ymm5[22,18,u],zero,zero,ymm5[23,19,u],zero,zero,ymm5[24,28]
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
 ; AVX512DQ-FCP-NEXT:    vporq %zmm3, %zmm4, %zmm3
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm4
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm5, %zmm4
+; 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 %zmm0, %zmm4, %zmm4
 ; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm3))
 ; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[1,3,2,3]
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[u],zero,zero,xmm2[5,13,u],zero,zero,xmm2[6,14,u],zero,zero,xmm2[7,15,u]
@@ -1519,19 +1517,18 @@ define void @store_i8_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm4 = zero,zero,zmm4[0,8],zero,zero,zero,zmm4[1,9],zero,zero,zero,zmm4[2,10],zero,zero,zero,zmm4[19,27],zero,zero,zero,zmm4[20,28],zero,zero,zero,zmm4[21,29],zero,zero,zero,zero,zero,zero,zmm4[35,39],zero,zero,zero,zmm4[40,44],zero,zero,zero,zmm4[41,45],zero,zero,zero,zmm4[50,54],zero,zero,zero,zmm4[51,55],zero,zero,zero,zmm4[56,60],zero,zero
 ; AVX512BW-FCP-NEXT:    vporq %zmm3, %zmm4, %zmm3
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm4
-; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
-; AVX512BW-FCP-NEXT:    vpermd %zmm4, %zmm5, %zmm4
+; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
+; AVX512BW-FCP-NEXT:    vpermd %zmm2, %zmm4, %zmm4
 ; AVX512BW-FCP-NEXT:    movabsq $595056260442243600, %rax # imm = 0x842108421084210
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k1
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm3 {%k1}
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[12],zero,zero,zero,zero,xmm2[13],zero,zero,zero,zero,xmm2[14],zero,zero,zero,zero,xmm2[15]
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,3,2,3]
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = zero,zero,zero,xmm1[5,13],zero,zero,zero,xmm1[6,14],zero,zero,zero,xmm1[7,15],zero
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,3,2,3]
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = zero,xmm0[5,13],zero,zero,zero,xmm0[6,14],zero,zero,zero,xmm0[7,15],zero,zero,zero
-; AVX512BW-FCP-NEXT:    vpternlogq {{.*#+}} xmm0 = xmm0 | xmm2 | xmm1
-; AVX512BW-FCP-NEXT:    vmovdqa %xmm0, 64(%r9)
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[12],zero,zero,zero,zero,xmm2[13],zero,zero,zero,zero,xmm2[14],zero,zero,zero,zero,xmm2[15]
+; AVX512BW-FCP-NEXT:    vpternlogq {{.*#+}} xmm2 = xmm2 | xmm0 | xmm1
+; AVX512BW-FCP-NEXT:    vmovdqa %xmm2, 64(%r9)
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, (%r9)
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
@@ -1594,19 +1591,18 @@ define void @store_i8_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm4 = zero,zero,zmm4[0,8],zero,zero,zero,zmm4[1,9],zero,zero,zero,zmm4[2,10],zero,zero,zero,zmm4[19,27],zero,zero,zero,zmm4[20,28],zero,zero,zero,zmm4[21,29],zero,zero,zero,zero,zero,zero,zmm4[35,39],zero,zero,zero,zmm4[40,44],zero,zero,zero,zmm4[41,45],zero,zero,zero,zmm4[50,54],zero,zero,zero,zmm4[51,55],zero,zero,zero,zmm4[56,60],zero,zero
 ; AVX512DQ-BW-FCP-NEXT:    vporq %zmm3, %zmm4, %zmm3
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm4
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %zmm4, %zmm5, %zmm4
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
+; AVX512DQ-BW-FCP-NEXT:    vpermd %zmm2, %zmm4, %zmm4
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $595056260442243600, %rax # imm = 0x842108421084210
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm3 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[12],zero,zero,zero,zero,xmm2[13],zero,zero,zero,zero,xmm2[14],zero,zero,zero,zero,xmm2[15]
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,3,2,3]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = zero,zero,zero,xmm1[5,13],zero,zero,zero,xmm1[6,14],zero,zero,zero,xmm1[7,15],zero
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,3,2,3]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = zero,xmm0[5,13],zero,zero,zero,xmm0[6,14],zero,zero,zero,xmm0[7,15],zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpternlogq {{.*#+}} xmm0 = xmm0 | xmm2 | xmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm0, 64(%r9)
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[12],zero,zero,zero,zero,xmm2[13],zero,zero,zero,zero,xmm2[14],zero,zero,zero,zero,xmm2[15]
+; AVX512DQ-BW-FCP-NEXT:    vpternlogq {{.*#+}} xmm2 = xmm2 | xmm0 | xmm1
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm2, 64(%r9)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, (%r9)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
@@ -2472,78 +2468,78 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ;
 ; AVX512-FCP-LABEL: store_i8_stride5_vf32:
 ; AVX512-FCP:       # %bb.0:
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm2
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm3
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm0
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm1
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm4
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm6
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9,u,11,u],zero,xmm6[10],zero,xmm6[12,u],zero
-; AVX512-FCP-NEXT:    vpor %xmm5, %xmm7, %xmm5
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,1,1]
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm7
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = zero,xmm7[6],zero,xmm7[8,u],zero,xmm7[7],zero,xmm7[9],zero,xmm7[11,u],zero,xmm7[10],zero,xmm7[12]
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm9
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm10 = xmm9[6],zero,xmm9[8],zero,xmm9[u,7],zero,xmm9[9],zero,xmm9[11],zero,xmm9[u,10],zero,xmm9[12],zero
-; AVX512-FCP-NEXT:    vpor %xmm8, %xmm10, %xmm8
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm3
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm4
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm1
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm2
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm0
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm5
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm5[8],zero,xmm5[u,7],zero,xmm5[9],zero,xmm5[u],zero,xmm5[u,10],zero,xmm5[12],zero,xmm5[u,11]
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm7
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = zero,xmm7[8,u],zero,xmm7[7],zero,xmm7[9,u,11,u],zero,xmm7[10],zero,xmm7[12,u],zero
+; AVX512-FCP-NEXT:    vpor %xmm6, %xmm8, %xmm6
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm8
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm9 = zero,xmm8[6],zero,xmm8[8,u],zero,xmm8[7],zero,xmm8[9],zero,xmm8[11,u],zero,xmm8[10],zero,xmm8[12]
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm10
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm11 = xmm10[6],zero,xmm10[8],zero,xmm10[u,7],zero,xmm10[9],zero,xmm10[11],zero,xmm10[u,10],zero,xmm10[12],zero
+; AVX512-FCP-NEXT:    vpor %xmm9, %xmm11, %xmm9
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm11 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm9 = ymm6 ^ (ymm11 & (ymm9 ^ ymm6))
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm6
+; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3],xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = xmm8[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
 ; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,1,1]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm8 = ymm5 ^ (ymm10 & (ymm8 ^ ymm5))
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm5
-; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3],xmm7[4],xmm9[4],xmm7[5],xmm9[5],xmm7[6],xmm9[6],xmm7[7],xmm9[7]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
-; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm4[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm4 = ymm7 ^ (ymm6 & (ymm4 ^ ymm7))
-; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm5 = zmm4[0,1,2,3],zmm5[4,5,6,7]
-; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm4 = mem[0,1,2,3,0,1,2,3]
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm7, %zmm7
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm5))
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm1[u,u,u],zero,ymm1[13,u,u,u],zero,ymm1[14,u,u,u],zero,ymm1[15,u,u,u],zero,ymm1[16,u,u,u],zero,ymm1[17,u,u,u],zero,ymm1[18,u,u]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm0[u,u,u,13],zero,ymm0[u,u,u,14],zero,ymm0[u,u,u,15],zero,ymm0[u,u,u,16],zero,ymm0[u,u,u,17],zero,ymm0[u,u,u,18],zero,ymm0[u,u]
+; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,1,1]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm5 = ymm8 ^ (ymm7 & (ymm5 ^ ymm8))
+; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
+; AVX512-FCP-NEXT:    vpermd %zmm0, %zmm6, %zmm6
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm5))
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm2[u,u,u],zero,ymm2[13,u,u,u],zero,ymm2[14,u,u,u],zero,ymm2[15,u,u,u],zero,ymm2[16,u,u,u],zero,ymm2[17,u,u,u],zero,ymm2[18,u,u]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,13],zero,ymm1[u,u,u,14],zero,ymm1[u,u,u,15],zero,ymm1[u,u,u,16],zero,ymm1[u,u,u,17],zero,ymm1[u,u,u,18],zero,ymm1[u,u]
 ; AVX512-FCP-NEXT:    vpor %ymm5, %ymm8, %ymm5
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm3[u],zero,ymm3[13,u,u,u],zero,ymm3[14,u,u,u],zero,ymm3[15,u,u,u],zero,ymm3[16,u,u,u],zero,ymm3[17,u,u,u],zero,ymm3[18,u,u,u],zero
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm2[u,13],zero,ymm2[u,u,u,14],zero,ymm2[u,u,u,15],zero,ymm2[u,u,u,16],zero,ymm2[u,u,u,17],zero,ymm2[u,u,u,18],zero,ymm2[u,u,u,19]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm4[u],zero,ymm4[13,u,u,u],zero,ymm4[14,u,u,u],zero,ymm4[15,u,u,u],zero,ymm4[16,u,u,u],zero,ymm4[17,u,u,u],zero,ymm4[18,u,u,u],zero
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm3[u,13],zero,ymm3[u,u,u,14],zero,ymm3[u,u,u,15],zero,ymm3[u,u,u,16],zero,ymm3[u,u,u,17],zero,ymm3[u,u,u,18],zero,ymm3[u,u,u,19]
 ; AVX512-FCP-NEXT:    vpor %ymm8, %ymm9, %ymm8
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm8 = ymm5 ^ (ymm10 & (ymm8 ^ ymm5))
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero,ymm2[25],zero,zero
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm3[21,u],zero,ymm3[20],zero,ymm3[22],zero,ymm3[24,u],zero,ymm3[23],zero,ymm3[25,u]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm8 = ymm5 ^ (ymm11 & (ymm8 ^ ymm5))
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero,ymm3[25],zero,zero
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21,u],zero,ymm4[20],zero,ymm4[22],zero,ymm4[24,u],zero,ymm4[23],zero,ymm4[25,u]
 ; AVX512-FCP-NEXT:    vpor %ymm5, %ymm9, %ymm5
 ; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[19],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
 ; AVX512-FCP-NEXT:    vpor %ymm9, %ymm10, %ymm9
 ; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
 ; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm9 = ymm9 ^ (mem & (ymm9 ^ ymm5))
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm5
 ; AVX512-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
 ; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm8 = [4,0,5,5,5,5,0,6]
-; AVX512-FCP-NEXT:    vpermd %ymm4, %ymm8, %ymm8
+; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm8, %ymm8
 ; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
 ; AVX512-FCP-NEXT:    vpandn %ymm8, %ymm9, %ymm8
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm4[12],zero,zero,zero,zero,ymm4[13],zero,zero,zero,zero,ymm4[14],zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,ymm4[18],zero
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm0[12],zero,zero,zero,zero,ymm0[13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,ymm0[15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,zero,ymm0[17],zero,zero,zero,zero,ymm0[18],zero
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm9, %zmm8
 ; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm8 | (zmm5 & mem)
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[26],zero,ymm2[28],zero,zero,ymm2[27],zero,ymm2[29],zero,ymm2[31],zero,zero,ymm2[30],zero
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u],zero,ymm3[26],zero,ymm3[28,u],zero,ymm3[u],zero,ymm3[29],zero,ymm3[31,u],zero,ymm3[30]
-; AVX512-FCP-NEXT:    vpor %ymm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,zero,ymm0[26],zero,ymm0[28],zero,ymm0[30],zero,zero,ymm0[29],zero,ymm0[31],zero,zero
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm1[27,u],zero,ymm1[26],zero,ymm1[28],zero,ymm1[30,u],zero,ymm1[29],zero,ymm1[31,u]
-; AVX512-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm0 = ymm2 ^ (ymm6 & (ymm0 ^ ymm2))
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [6,6,6,0,7,7,7,7]
-; AVX512-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm1 = ymm1 ^ (mem & (ymm1 ^ ymm0))
-; AVX512-FCP-NEXT:    vmovdqa %ymm1, 128(%r9)
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[26],zero,ymm3[28],zero,zero,ymm3[27],zero,ymm3[29],zero,ymm3[31],zero,zero,ymm3[30],zero
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u],zero,ymm4[26],zero,ymm4[28,u],zero,ymm4[u],zero,ymm4[29],zero,ymm4[31,u],zero,ymm4[30]
+; AVX512-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,ymm1[26],zero,ymm1[28],zero,ymm1[30],zero,zero,ymm1[29],zero,ymm1[31],zero,zero
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27,u],zero,ymm2[26],zero,ymm2[28],zero,ymm2[30,u],zero,ymm2[29],zero,ymm2[31,u]
+; AVX512-FCP-NEXT:    vpor %ymm1, %ymm2, %ymm1
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm1 = ymm3 ^ (ymm7 & (ymm1 ^ ymm3))
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [6,6,6,0,7,7,7,7]
+; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm2, %ymm0
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm0 = ymm0 ^ (mem & (ymm0 ^ ymm1))
+; AVX512-FCP-NEXT:    vmovdqa %ymm0, 128(%r9)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, 64(%r9)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, (%r9)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, (%r9)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -2627,78 +2623,78 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ;
 ; AVX512DQ-FCP-LABEL: store_i8_stride5_vf32:
 ; AVX512DQ-FCP:       # %bb.0:
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm4
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm6
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9,u,11,u],zero,xmm6[10],zero,xmm6[12,u],zero
-; AVX512DQ-FCP-NEXT:    vpor %xmm5, %xmm7, %xmm5
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,1,1]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm7
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = zero,xmm7[6],zero,xmm7[8,u],zero,xmm7[7],zero,xmm7[9],zero,xmm7[11,u],zero,xmm7[10],zero,xmm7[12]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm9
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm10 = xmm9[6],zero,xmm9[8],zero,xmm9[u,7],zero,xmm9[9],zero,xmm9[11],zero,xmm9[u,10],zero,xmm9[12],zero
-; AVX512DQ-FCP-NEXT:    vpor %xmm8, %xmm10, %xmm8
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm4
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm1
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm5
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm5[8],zero,xmm5[u,7],zero,xmm5[9],zero,xmm5[u],zero,xmm5[u,10],zero,xmm5[12],zero,xmm5[u,11]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm7
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = zero,xmm7[8,u],zero,xmm7[7],zero,xmm7[9,u,11,u],zero,xmm7[10],zero,xmm7[12,u],zero
+; AVX512DQ-FCP-NEXT:    vpor %xmm6, %xmm8, %xmm6
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm8
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm9 = zero,xmm8[6],zero,xmm8[8,u],zero,xmm8[7],zero,xmm8[9],zero,xmm8[11,u],zero,xmm8[10],zero,xmm8[12]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm10
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm11 = xmm10[6],zero,xmm10[8],zero,xmm10[u,7],zero,xmm10[9],zero,xmm10[11],zero,xmm10[u,10],zero,xmm10[12],zero
+; AVX512DQ-FCP-NEXT:    vpor %xmm9, %xmm11, %xmm9
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm11 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm9 = ymm6 ^ (ymm11 & (ymm9 ^ ymm6))
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm6
+; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3],xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = xmm8[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
 ; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,1,1]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm8 = ymm5 ^ (ymm10 & (ymm8 ^ ymm5))
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm5
-; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3],xmm7[4],xmm9[4],xmm7[5],xmm9[5],xmm7[6],xmm9[6],xmm7[7],xmm9[7]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
-; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm4[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm4 = ymm7 ^ (ymm6 & (ymm4 ^ ymm7))
-; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm5 = zmm4[0,1,2,3],zmm5[4,5,6,7]
-; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm4 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm7, %zmm7
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm5))
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm1[u,u,u],zero,ymm1[13,u,u,u],zero,ymm1[14,u,u,u],zero,ymm1[15,u,u,u],zero,ymm1[16,u,u,u],zero,ymm1[17,u,u,u],zero,ymm1[18,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm0[u,u,u,13],zero,ymm0[u,u,u,14],zero,ymm0[u,u,u,15],zero,ymm0[u,u,u,16],zero,ymm0[u,u,u,17],zero,ymm0[u,u,u,18],zero,ymm0[u,u]
+; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,1,1]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm5 = ymm8 ^ (ymm7 & (ymm5 ^ ymm8))
+; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm0, %zmm6, %zmm6
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm5))
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm2[u,u,u],zero,ymm2[13,u,u,u],zero,ymm2[14,u,u,u],zero,ymm2[15,u,u,u],zero,ymm2[16,u,u,u],zero,ymm2[17,u,u,u],zero,ymm2[18,u,u]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,13],zero,ymm1[u,u,u,14],zero,ymm1[u,u,u,15],zero,ymm1[u,u,u,16],zero,ymm1[u,u,u,17],zero,ymm1[u,u,u,18],zero,ymm1[u,u]
 ; AVX512DQ-FCP-NEXT:    vpor %ymm5, %ymm8, %ymm5
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm3[u],zero,ymm3[13,u,u,u],zero,ymm3[14,u,u,u],zero,ymm3[15,u,u,u],zero,ymm3[16,u,u,u],zero,ymm3[17,u,u,u],zero,ymm3[18,u,u,u],zero
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm2[u,13],zero,ymm2[u,u,u,14],zero,ymm2[u,u,u,15],zero,ymm2[u,u,u,16],zero,ymm2[u,u,u,17],zero,ymm2[u,u,u,18],zero,ymm2[u,u,u,19]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm4[u],zero,ymm4[13,u,u,u],zero,ymm4[14,u,u,u],zero,ymm4[15,u,u,u],zero,ymm4[16,u,u,u],zero,ymm4[17,u,u,u],zero,ymm4[18,u,u,u],zero
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm3[u,13],zero,ymm3[u,u,u,14],zero,ymm3[u,u,u,15],zero,ymm3[u,u,u,16],zero,ymm3[u,u,u,17],zero,ymm3[u,u,u,18],zero,ymm3[u,u,u,19]
 ; AVX512DQ-FCP-NEXT:    vpor %ymm8, %ymm9, %ymm8
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm8 = ymm5 ^ (ymm10 & (ymm8 ^ ymm5))
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero,ymm2[25],zero,zero
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm3[21,u],zero,ymm3[20],zero,ymm3[22],zero,ymm3[24,u],zero,ymm3[23],zero,ymm3[25,u]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm8 = ymm5 ^ (ymm11 & (ymm8 ^ ymm5))
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero,ymm3[25],zero,zero
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21,u],zero,ymm4[20],zero,ymm4[22],zero,ymm4[24,u],zero,ymm4[23],zero,ymm4[25,u]
 ; AVX512DQ-FCP-NEXT:    vpor %ymm5, %ymm9, %ymm5
 ; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[19],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
 ; AVX512DQ-FCP-NEXT:    vpor %ymm9, %ymm10, %ymm9
 ; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
 ; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm9 = ymm9 ^ (mem & (ymm9 ^ ymm5))
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm5
 ; AVX512DQ-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
 ; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm8 = [4,0,5,5,5,5,0,6]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm4, %ymm8, %ymm8
+; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm8, %ymm8
 ; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
 ; AVX512DQ-FCP-NEXT:    vpandn %ymm8, %ymm9, %ymm8
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm4[12],zero,zero,zero,zero,ymm4[13],zero,zero,zero,zero,ymm4[14],zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,ymm4[18],zero
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm0[12],zero,zero,zero,zero,ymm0[13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,ymm0[15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,zero,ymm0[17],zero,zero,zero,zero,ymm0[18],zero
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm9, %zmm8
 ; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm8 | (zmm5 & mem)
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[26],zero,ymm2[28],zero,zero,ymm2[27],zero,ymm2[29],zero,ymm2[31],zero,zero,ymm2[30],zero
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u],zero,ymm3[26],zero,ymm3[28,u],zero,ymm3[u],zero,ymm3[29],zero,ymm3[31,u],zero,ymm3[30]
-; AVX512DQ-FCP-NEXT:    vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,zero,ymm0[26],zero,ymm0[28],zero,ymm0[30],zero,zero,ymm0[29],zero,ymm0[31],zero,zero
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm1[27,u],zero,ymm1[26],zero,ymm1[28],zero,ymm1[30,u],zero,ymm1[29],zero,ymm1[31,u]
-; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm0 = ymm2 ^ (ymm6 & (ymm0 ^ ymm2))
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [6,6,6,0,7,7,7,7]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm1 = ymm1 ^ (mem & (ymm1 ^ ymm0))
-; AVX512DQ-FCP-NEXT:    vmovdqa %ymm1, 128(%r9)
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[26],zero,ymm3[28],zero,zero,ymm3[27],zero,ymm3[29],zero,ymm3[31],zero,zero,ymm3[30],zero
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u],zero,ymm4[26],zero,ymm4[28,u],zero,ymm4[u],zero,ymm4[29],zero,ymm4[31,u],zero,ymm4[30]
+; AVX512DQ-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,ymm1[26],zero,ymm1[28],zero,ymm1[30],zero,zero,ymm1[29],zero,ymm1[31],zero,zero
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27,u],zero,ymm2[26],zero,ymm2[28],zero,ymm2[30,u],zero,ymm2[29],zero,ymm2[31,u]
+; AVX512DQ-FCP-NEXT:    vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm1 = ymm3 ^ (ymm7 & (ymm1 ^ ymm3))
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [6,6,6,0,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm0 = ymm0 ^ (mem & (ymm0 ^ ymm1))
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm0, 128(%r9)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, 64(%r9)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, (%r9)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, (%r9)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
 ;
@@ -2795,82 +2791,82 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ;
 ; AVX512BW-FCP-LABEL: store_i8_stride5_vf32:
 ; AVX512BW-FCP:       # %bb.0:
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm1
-; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %ymm3
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %ymm0
-; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %ymm2
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm4
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
-; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %xmm6
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9,u,11,u],zero,xmm6[10],zero,xmm6[12,u],zero
-; AVX512BW-FCP-NEXT:    vpor %xmm5, %xmm7, %xmm5
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm7
-; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %xmm8
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm9 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm9 = xmm9[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm5, %zmm9, %zmm5
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm2
+; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %ymm4
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %ymm1
+; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %ymm3
+; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %ymm0
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm5
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm5[8],zero,xmm5[u,7],zero,xmm5[9],zero,xmm5[u],zero,xmm5[u,10],zero,xmm5[12],zero,xmm5[u,11]
+; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %xmm7
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = zero,xmm7[8,u],zero,xmm7[7],zero,xmm7[9,u,11,u],zero,xmm7[10],zero,xmm7[12,u],zero
+; AVX512BW-FCP-NEXT:    vpor %xmm6, %xmm8, %xmm6
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm8
+; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %xmm9
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm10 = xmm10[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm10, %zmm6
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm6 = zmm6[0,0,1,1,4,4,5,5]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm9 = zero,xmm9[6],zero,xmm9[8,u],zero,xmm9[7],zero,xmm9[9],zero,xmm9[11,u],zero,xmm9[10],zero,xmm9[12]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = xmm8[6],zero,xmm8[8],zero,xmm8[u,7],zero,xmm8[9],zero,xmm8[11],zero,xmm8[u,10],zero,xmm8[12],zero
+; AVX512BW-FCP-NEXT:    vpor %xmm9, %xmm8, %xmm8
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm5, %zmm5
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm5 = zmm5[0,0,1,1,4,4,5,5]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = zero,xmm8[6],zero,xmm8[8,u],zero,xmm8[7],zero,xmm8[9],zero,xmm8[11,u],zero,xmm8[10],zero,xmm8[12]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[6],zero,xmm7[8],zero,xmm7[u,7],zero,xmm7[9],zero,xmm7[11],zero,xmm7[u,10],zero,xmm7[12],zero
-; AVX512BW-FCP-NEXT:    vpor %xmm7, %xmm8, %xmm7
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm4[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm7, %zmm4, %zmm4
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm4 = zmm4[0,0,1,1,4,4,5,5]
 ; AVX512BW-FCP-NEXT:    movabsq $3570337559743967628, %rax # imm = 0x318C631818C6318C
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm5, %zmm4 {%k1}
-; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
-; AVX512BW-FCP-NEXT:    vpermd %zmm5, %zmm6, %zmm6
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm5 {%k1}
+; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
+; AVX512BW-FCP-NEXT:    vpermd %zmm0, %zmm6, %zmm6
 ; AVX512BW-FCP-NEXT:    movabsq $595056260442243600, %rax # imm = 0x842108421084210
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm4 {%k1}
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,ymm0[12,13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,ymm0[14,15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,ymm0[16,17],zero,zero,zero,zero,ymm0[18],zero,zero,zero
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm5 {%k1}
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,ymm1[12,13],zero,zero,zero,zero,ymm1[14],zero,zero,zero,ymm1[14,15],zero,zero,zero,zero,ymm1[16],zero,zero,zero,ymm1[16,17],zero,zero,zero,zero,ymm1[18],zero,zero,zero
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm3[13],zero,zero,zero,zero,ymm3[14],zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,ymm3[18],zero,zero
 ; AVX512BW-FCP-NEXT:    vpor %ymm6, %ymm7, %ymm6
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[19],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero
 ; AVX512BW-FCP-NEXT:    vpor %ymm7, %ymm8, %ymm7
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
 ; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm6, %zmm6
 ; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm7 = [3,3,3,0,4,4,4,4]
-; AVX512BW-FCP-NEXT:    vpermd %ymm1, %ymm7, %ymm7
+; AVX512BW-FCP-NEXT:    vpermd %ymm2, %ymm7, %ymm7
 ; AVX512BW-FCP-NEXT:    movl $138547332, %eax # imm = 0x8421084
 ; AVX512BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 {%k1} = ymm3[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero,ymm3[25],zero
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 {%k1} = ymm4[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero,ymm2[25],zero,zero
 ; AVX512BW-FCP-NEXT:    vpor %ymm8, %ymm9, %ymm8
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
 ; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm7, %zmm7
 ; AVX512BW-FCP-NEXT:    movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k1
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm7 {%k1}
-; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [3,3,3,3,0,4,4,4,12,14,13,13,13,13,12,14]
-; AVX512BW-FCP-NEXT:    vpermd %zmm5, %zmm6, %zmm6
+; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [3,3,3,3,0,4,4,4,4,6,5,5,5,5,4,6]
+; AVX512BW-FCP-NEXT:    vpermd %zmm0, %zmm6, %zmm6
 ; AVX512BW-FCP-NEXT:    movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k1
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm7 {%k1}
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm3[26],zero,ymm3[28],zero,zero,zero,zero,ymm3[29],zero,ymm3[31],zero,zero,ymm3[30]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,zero,zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[26],zero,ymm2[28],zero,zero,ymm2[27],zero,ymm2[29],zero,ymm2[31],zero,zero,ymm2[30],zero
+; AVX512BW-FCP-NEXT:    vpor %ymm4, %ymm2, %ymm2
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm3[27],zero,zero,ymm3[26],zero,ymm3[28],zero,ymm3[30],zero,zero,ymm3[29],zero,ymm3[31],zero
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,ymm1[26],zero,ymm1[28],zero,ymm1[30],zero,zero,ymm1[29],zero,ymm1[31],zero,zero
 ; AVX512BW-FCP-NEXT:    vpor %ymm3, %ymm1, %ymm1
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27],zero,zero,ymm2[26],zero,ymm2[28],zero,ymm2[30],zero,zero,ymm2[29],zero,ymm2[31],zero
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,zero,ymm0[26],zero,ymm0[28],zero,ymm0[30],zero,zero,ymm0[29],zero,ymm0[31],zero,zero
-; AVX512BW-FCP-NEXT:    vpor %ymm2, %ymm0, %ymm0
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
 ; AVX512BW-FCP-NEXT:    movl $415641996, %eax # imm = 0x18C6318C
 ; AVX512BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm1, %ymm0 {%k1}
-; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [6,6,6,6,7,7,7,7]
-; AVX512BW-FCP-NEXT:    vpermd %ymm5, %ymm1, %ymm1
+; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm2, %ymm1 {%k1}
+; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [6,6,6,6,7,7,7,7]
+; AVX512BW-FCP-NEXT:    vpermd %ymm0, %ymm2, %ymm0
 ; AVX512BW-FCP-NEXT:    movl $-2078209982, %eax # imm = 0x84210842
 ; AVX512BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm1, %ymm0 {%k1}
-; AVX512BW-FCP-NEXT:    vmovdqa %ymm0, 128(%r9)
+; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm0, %ymm1 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqa %ymm1, 128(%r9)
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, 64(%r9)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, (%r9)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, (%r9)
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
 ;
@@ -2967,82 +2963,82 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ;
 ; AVX512DQ-BW-FCP-LABEL: store_i8_stride5_vf32:
 ; AVX512DQ-BW-FCP:       # %bb.0:
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %ymm3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm4
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %xmm6
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9,u,11,u],zero,xmm6[10],zero,xmm6[12,u],zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %xmm5, %xmm7, %xmm5
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm7
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %xmm8
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm9 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm9 = xmm9[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm5, %zmm9, %zmm5
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %ymm3
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm5
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm5[8],zero,xmm5[u,7],zero,xmm5[9],zero,xmm5[u],zero,xmm5[u,10],zero,xmm5[12],zero,xmm5[u,11]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %xmm7
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = zero,xmm7[8,u],zero,xmm7[7],zero,xmm7[9,u,11,u],zero,xmm7[10],zero,xmm7[12,u],zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %xmm6, %xmm8, %xmm6
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm8
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %xmm9
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm10 = xmm10[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm10, %zmm6
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm6 = zmm6[0,0,1,1,4,4,5,5]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm9 = zero,xmm9[6],zero,xmm9[8,u],zero,xmm9[7],zero,xmm9[9],zero,xmm9[11,u],zero,xmm9[10],zero,xmm9[12]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = xmm8[6],zero,xmm8[8],zero,xmm8[u,7],zero,xmm8[9],zero,xmm8[11],zero,xmm8[u,10],zero,xmm8[12],zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %xmm9, %xmm8, %xmm8
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm8, %zmm5, %zmm5
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm5 = zmm5[0,0,1,1,4,4,5,5]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm8 = zero,xmm8[6],zero,xmm8[8,u],zero,xmm8[7],zero,xmm8[9],zero,xmm8[11,u],zero,xmm8[10],zero,xmm8[12]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[6],zero,xmm7[8],zero,xmm7[u,7],zero,xmm7[9],zero,xmm7[11],zero,xmm7[u,10],zero,xmm7[12],zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %xmm7, %xmm8, %xmm7
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm4[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm7, %zmm4, %zmm4
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm4 = zmm4[0,0,1,1,4,4,5,5]
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $3570337559743967628, %rax # imm = 0x318C631818C6318C
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm5, %zmm4 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %zmm5, %zmm6, %zmm6
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm5 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
+; AVX512DQ-BW-FCP-NEXT:    vpermd %zmm0, %zmm6, %zmm6
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $595056260442243600, %rax # imm = 0x842108421084210
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm4 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,ymm0[12,13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,ymm0[14,15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,ymm0[16,17],zero,zero,zero,zero,ymm0[18],zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm5 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,ymm1[12,13],zero,zero,zero,zero,ymm1[14],zero,zero,zero,ymm1[14,15],zero,zero,zero,zero,ymm1[16],zero,zero,zero,ymm1[16,17],zero,zero,zero,zero,ymm1[18],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm3[13],zero,zero,zero,zero,ymm3[14],zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,ymm3[18],zero,zero
 ; AVX512DQ-BW-FCP-NEXT:    vpor %ymm6, %ymm7, %ymm6
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[19],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero
 ; AVX512DQ-BW-FCP-NEXT:    vpor %ymm7, %ymm8, %ymm7
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
 ; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm6, %zmm6
 ; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm7 = [3,3,3,0,4,4,4,4]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm1, %ymm7, %ymm7
+; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm2, %ymm7, %ymm7
 ; AVX512DQ-BW-FCP-NEXT:    movl $138547332, %eax # imm = 0x8421084
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 {%k1} = ymm3[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero,ymm3[25],zero
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 {%k1} = ymm4[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero,ymm2[25],zero,zero
 ; AVX512DQ-BW-FCP-NEXT:    vpor %ymm8, %ymm9, %ymm8
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
 ; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm7, %zmm7
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm7 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [3,3,3,3,0,4,4,4,12,14,13,13,13,13,12,14]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %zmm5, %zmm6, %zmm6
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [3,3,3,3,0,4,4,4,4,6,5,5,5,5,4,6]
+; AVX512DQ-BW-FCP-NEXT:    vpermd %zmm0, %zmm6, %zmm6
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm7 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm3[26],zero,ymm3[28],zero,zero,zero,zero,ymm3[29],zero,ymm3[31],zero,zero,ymm3[30]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,zero,zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[26],zero,ymm2[28],zero,zero,ymm2[27],zero,ymm2[29],zero,ymm2[31],zero,zero,ymm2[30],zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm4, %ymm2, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm3[27],zero,zero,ymm3[26],zero,ymm3[28],zero,ymm3[30],zero,zero,ymm3[29],zero,ymm3[31],zero
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,ymm1[26],zero,ymm1[28],zero,ymm1[30],zero,zero,ymm1[29],zero,ymm1[31],zero,zero
 ; AVX512DQ-BW-FCP-NEXT:    vpor %ymm3, %ymm1, %ymm1
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27],zero,zero,ymm2[26],zero,ymm2[28],zero,ymm2[30],zero,zero,ymm2[29],zero,ymm2[31],zero
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,zero,ymm0[26],zero,ymm0[28],zero,ymm0[30],zero,zero,ymm0[29],zero,ymm0[31],zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm2, %ymm0, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
 ; AVX512DQ-BW-FCP-NEXT:    movl $415641996, %eax # imm = 0x18C6318C
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm1, %ymm0 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [6,6,6,6,7,7,7,7]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm5, %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm2, %ymm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [6,6,6,6,7,7,7,7]
+; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm0, %ymm2, %ymm0
 ; AVX512DQ-BW-FCP-NEXT:    movl $-2078209982, %eax # imm = 0x84210842
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm1, %ymm0 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm0, 128(%r9)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm0, %ymm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm1, 128(%r9)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, 64(%r9)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, (%r9)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, (%r9)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
   %in.vec0 = load <32 x i8>, ptr %in.vecptr0, align 64
@@ -4879,167 +4875,167 @@ 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), %ymm0
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm14 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm0, %ymm1
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm2
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm3
-; AVX512-FCP-NEXT:    vporq %ymm1, %ymm3, %ymm17
-; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm3
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
-; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm5
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
-; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm5, %xmm4
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm6, %xmm28
-; AVX512-FCP-NEXT:    vporq %xmm1, %xmm4, %xmm18
-; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm1
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm1, %ymm4
+; 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, %ymm0, %ymm2
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm16
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %ymm8
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm15 = [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 %ymm15, %ymm8, %ymm3
+; AVX512-FCP-NEXT:    vporq %ymm2, %ymm3, %ymm31
+; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm14
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm14[8],zero,xmm14[u,7],zero,xmm14[9],zero,xmm14[u],zero,xmm14[u,10],zero,xmm14[12],zero,xmm14[u,11]
+; AVX512-FCP-NEXT:    vmovdqa 32(%rsi), %xmm9
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm3 = zero,xmm9[8,u],zero,xmm9[7],zero,xmm9[9,u,11,u],zero,xmm9[10],zero,xmm9[12,u],zero
+; AVX512-FCP-NEXT:    vporq %xmm2, %xmm3, %xmm19
+; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %ymm10
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm12 = [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 %ymm12, %ymm10, %ymm2
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %ymm11
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm12 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm11, %ymm6
-; AVX512-FCP-NEXT:    vporq %ymm4, %ymm6, %ymm19
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm11, %ymm3
+; AVX512-FCP-NEXT:    vporq %ymm2, %ymm3, %ymm20
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rcx), %xmm6
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = zero,xmm6[6],zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9],zero,xmm6[11,u],zero,xmm6[10],zero,xmm6[12]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = zero,xmm6[6],zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9],zero,xmm6[11,u],zero,xmm6[10],zero,xmm6[12]
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rdx), %xmm7
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm13 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm8
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm13, %xmm29
-; AVX512-FCP-NEXT:    vporq %xmm4, %xmm8, %xmm20
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
-; AVX512-FCP-NEXT:    # ymm8 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm2, %ymm4
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm8, %ymm31
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm13 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
-; AVX512-FCP-NEXT:    # ymm13 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm2, %ymm2
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm21
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm7, %xmm3
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm1, %xmm28
+; AVX512-FCP-NEXT:    vporq %xmm2, %xmm3, %xmm21
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [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:    # ymm1 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm8, %ymm3
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm30
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [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:    # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm0, %ymm8
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm30
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
-; AVX512-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm22
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm8, %ymm4
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm22
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512-FCP-NEXT:    # ymm1 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm0, %ymm3
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm29
+; 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, %ymm11, %ymm8
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm0 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm0, %ymm0
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm23
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [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:    # ymm1 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm11, %ymm0
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm18
+; 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, %ymm10, %ymm5
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm5, %zmm24
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [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:    # ymm3 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm10, %ymm5
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm0 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
 ; AVX512-FCP-NEXT:    # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm1, %ymm15
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm15, %zmm24
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
-; AVX512-FCP-NEXT:    # ymm8 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm15 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512-FCP-NEXT:    # ymm15 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm11, %ymm11
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm11, %zmm26
-; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm1
-; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm1, %ymm10
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm11
-; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm11, %ymm12
-; AVX512-FCP-NEXT:    vporq %ymm10, %ymm12, %ymm23
-; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm1, %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm11, %ymm10
-; AVX512-FCP-NEXT:    vporq %ymm0, %ymm10, %ymm25
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm10
-; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm10, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm12
-; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm12, %ymm9
-; AVX512-FCP-NEXT:    vporq %ymm0, %ymm9, %ymm27
-; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm12, %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm10, %ymm2
-; AVX512-FCP-NEXT:    vporq %ymm0, %ymm2, %ymm16
-; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm9
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = xmm9[8],zero,xmm9[u,7],zero,xmm9[9],zero,xmm9[u],zero,xmm9[u,10],zero,xmm9[12],zero,xmm9[u,11]
-; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm14
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm28, %xmm2
-; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm14, %xmm2
-; AVX512-FCP-NEXT:    vporq %xmm0, %xmm2, %xmm28
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm10
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm10, %zmm25
+; AVX512-FCP-NEXT:    vmovdqa (%rcx), %ymm10
+; AVX512-FCP-NEXT:    vpshufb %ymm12, %ymm10, %ymm5
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm12
+; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm12, %ymm11
+; AVX512-FCP-NEXT:    vporq %ymm5, %ymm11, %ymm26
+; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm10, %ymm1
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm12, %ymm0
+; AVX512-FCP-NEXT:    vporq %ymm1, %ymm0, %ymm27
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %ymm11
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm16, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm0
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm5
+; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm5, %ymm1
+; AVX512-FCP-NEXT:    vporq %ymm0, %ymm1, %ymm16
+; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm5, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm11, %ymm1
+; AVX512-FCP-NEXT:    vporq %ymm0, %ymm1, %ymm17
+; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm4
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
+; AVX512-FCP-NEXT:    vmovdqa (%rsi), %xmm2
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = 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 %xmm0, %xmm1, %xmm15
 ; AVX512-FCP-NEXT:    vmovdqa (%rcx), %xmm13
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = zero,xmm13[6],zero,xmm13[8,u],zero,xmm13[7],zero,xmm13[9],zero,xmm13[11,u],zero,xmm13[10],zero,xmm13[12]
-; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm15
-; AVX512-FCP-NEXT:    vmovdqa64 %xmm29, %xmm2
-; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm15, %xmm2
-; AVX512-FCP-NEXT:    vporq %xmm0, %xmm2, %xmm29
-; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm11, %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm31, %ymm1
-; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm12, %ymm1
-; AVX512-FCP-NEXT:    vmovdqa64 %ymm30, %ymm2
-; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm10, %ymm2
-; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm4
-; AVX512-FCP-NEXT:    vpor %ymm1, %ymm2, %ymm1
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm8 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm2
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm10 = [1,1,2,2,2,2,2,2]
-; AVX512-FCP-NEXT:    vpermd %ymm4, %ymm10, %ymm10
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512-FCP-NEXT:    vpandn %ymm10, %ymm11, %ymm10
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm10, %zmm2
-; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm10
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm30 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512-FCP-NEXT:    vpermd %ymm10, %ymm12, %ymm31
-; AVX512-FCP-NEXT:    vpandnq %ymm31, %ymm30, %ymm31
-; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm10, %ymm8
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm31, %zmm8, %zmm8
-; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm14[0],xmm9[1],xmm14[1],xmm9[2],xmm14[2],xmm9[3],xmm14[3],xmm9[4],xmm14[4],xmm9[5],xmm14[5],xmm9[6],xmm14[6],xmm9[7],xmm14[7]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm14 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm9, %xmm9
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm28, %zmm9, %zmm9
-; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm13 = xmm13[0],xmm15[0],xmm13[1],xmm15[1],xmm13[2],xmm15[2],xmm13[3],xmm15[3],xmm13[4],xmm15[4],xmm13[5],xmm15[5],xmm13[6],xmm15[6],xmm13[7],xmm15[7]
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm15 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm13, %xmm13
-; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm29, %zmm13, %zmm13
-; AVX512-FCP-NEXT:    vmovdqa64 (%r8), %zmm28
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm28, %zmm10
+; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512-FCP-NEXT:    vmovdqa64 %xmm28, %xmm8
+; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
+; AVX512-FCP-NEXT:    vporq %xmm0, %xmm8, %xmm28
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm18, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm12, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm10, %ymm3
+; AVX512-FCP-NEXT:    vpor %ymm0, %ymm3, %ymm0
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm30, %ymm3
+; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm5, %ymm3
+; AVX512-FCP-NEXT:    vmovdqa64 %ymm29, %ymm5
+; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm11, %ymm5
+; AVX512-FCP-NEXT:    vmovdqa 32(%r8), %ymm8
+; AVX512-FCP-NEXT:    vpor %ymm3, %ymm5, %ymm3
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm8, %ymm5
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm11 = [1,1,2,2,2,2,2,2]
+; AVX512-FCP-NEXT:    vpermd %ymm8, %ymm11, %ymm11
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
+; AVX512-FCP-NEXT:    vpandn %ymm11, %ymm12, %ymm11
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm11, %zmm5
+; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm11
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [4,6,5,5,5,5,4,6,6,6,6,6,7,7,7,7]
+; AVX512-FCP-NEXT:    vpermd %ymm11, %ymm29, %ymm30
+; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm11, %ymm10
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512-FCP-NEXT:    vpandnq %ymm30, %ymm11, %ymm30
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm30, %zmm10, %zmm10
+; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm4 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
+; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm2, %zmm2
+; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3],xmm13[4],xmm1[4],xmm13[5],xmm1[5],xmm13[6],xmm1[6],xmm13[7],xmm1[7]
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} xmm13 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm1, %xmm1
+; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm28, %zmm1, %zmm1
 ; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
-; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm6, %xmm6
+; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm6, %xmm6
 ; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm7 = [2,2,3,3,8,8,9,9]
 ; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm7, %zmm0
-; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
-; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm3, %xmm3
-; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm7, %zmm1
-; AVX512-FCP-NEXT:    vporq %zmm21, %zmm22, %zmm3
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm3 = zmm3[2,2,3,3,6,6,7,7]
-; AVX512-FCP-NEXT:    vporq %zmm24, %zmm26, %zmm5
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm5 = zmm5[2,2,3,3,6,6,7,7]
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm6 & (zmm5 ^ zmm3))
-; AVX512-FCP-NEXT:    vpermt2d %zmm28, %zmm12, %zmm4
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm4 = zmm4 ^ (zmm30 & (zmm4 ^ zmm5))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 256(%r9)
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm18[0,0,1,1]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm20[0,0,1,1]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm4, %zmm4
-; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm3 ^ (zmm5 & (zmm4 ^ zmm3))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 | (zmm4 & zmm11)
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm25[2,2,3,3]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm23, %zmm3
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm16[2,2,3,3]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm27, %zmm4
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm3 ^ (zmm5 & (zmm4 ^ zmm3))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm0 ^ (zmm6 & (zmm1 ^ zmm0))
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm8 | (zmm4 & mem)
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm9[0,0,1,1,4,4,5,5]
-; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm3 = zmm13[0,0,1,1,4,4,5,5]
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm0))
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512-FCP-NEXT:    vpermd %zmm10, %zmm0, %zmm0
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm3))
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512-FCP-NEXT:    vpermd %zmm28, %zmm3, %zmm3
-; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm1))
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 128(%r9)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, 64(%r9)
+; AVX512-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm14[0],xmm9[0],xmm14[1],xmm9[1],xmm14[2],xmm9[2],xmm14[3],xmm9[3],xmm14[4],xmm9[4],xmm14[5],xmm9[5],xmm14[6],xmm9[6],xmm14[7],xmm9[7]
+; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm6, %xmm4
+; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm7, %zmm3
+; AVX512-FCP-NEXT:    vmovdqa64 (%r8), %zmm4
+; AVX512-FCP-NEXT:    vpermd %zmm8, %zmm29, %zmm6
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [6,6,6,0,7,7,7,7,0,16,16,16,16,0,17,17]
+; AVX512-FCP-NEXT:    vpermi2d %zmm8, %zmm4, %zmm7
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm19[0,0,1,1]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm31, %zmm8, %zmm8
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm21[0,0,1,1]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm20, %zmm9, %zmm9
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm13 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm8 ^ (zmm13 & (zmm9 ^ zmm8))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm9 & zmm12)
+; AVX512-FCP-NEXT:    vporq %zmm22, %zmm23, %zmm8
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT:    vporq %zmm24, %zmm25, %zmm9
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm9 = zmm9[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm8 ^ (zmm12 & (zmm9 ^ zmm8))
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (zmm11 & (zmm6 ^ zmm9))
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm27[2,2,3,3]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm26, %zmm8
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm17[2,2,3,3]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm16, %zmm9
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm8 ^ (zmm13 & (zmm9 ^ zmm8))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm12 & (zmm3 ^ zmm0))
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm10 | (zmm9 & mem)
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm2[0,0,1,1,4,4,5,5]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[0,0,1,1,4,4,5,5]
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
+; AVX512-FCP-NEXT:    vpermd %zmm4, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))
+; AVX512-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm3))
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 128(%r9)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 64(%r9)
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%r9)
-; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 192(%r9)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 256(%r9)
+; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 192(%r9)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
 ;
@@ -5222,167 +5218,167 @@ 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), %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm14 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm0, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm3
-; AVX512DQ-FCP-NEXT:    vporq %ymm1, %ymm3, %ymm17
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm3
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm5
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm5, %xmm4
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm6, %xmm28
-; AVX512DQ-FCP-NEXT:    vporq %xmm1, %xmm4, %xmm18
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm1, %ymm4
+; 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, %ymm0, %ymm2
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm16
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %ymm8
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm15 = [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 %ymm15, %ymm8, %ymm3
+; AVX512DQ-FCP-NEXT:    vporq %ymm2, %ymm3, %ymm31
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm14
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm14[8],zero,xmm14[u,7],zero,xmm14[9],zero,xmm14[u],zero,xmm14[u,10],zero,xmm14[12],zero,xmm14[u,11]
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rsi), %xmm9
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm3 = zero,xmm9[8,u],zero,xmm9[7],zero,xmm9[9,u,11,u],zero,xmm9[10],zero,xmm9[12,u],zero
+; AVX512DQ-FCP-NEXT:    vporq %xmm2, %xmm3, %xmm19
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %ymm10
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm12 = [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 %ymm12, %ymm10, %ymm2
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %ymm11
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm12 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm11, %ymm6
-; AVX512DQ-FCP-NEXT:    vporq %ymm4, %ymm6, %ymm19
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm11, %ymm3
+; AVX512DQ-FCP-NEXT:    vporq %ymm2, %ymm3, %ymm20
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rcx), %xmm6
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = zero,xmm6[6],zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9],zero,xmm6[11,u],zero,xmm6[10],zero,xmm6[12]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = zero,xmm6[6],zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9],zero,xmm6[11,u],zero,xmm6[10],zero,xmm6[12]
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdx), %xmm7
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm13 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm8
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm13, %xmm29
-; AVX512DQ-FCP-NEXT:    vporq %xmm4, %xmm8, %xmm20
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
-; AVX512DQ-FCP-NEXT:    # ymm8 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm2, %ymm4
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm8, %ymm31
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm13 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
-; AVX512DQ-FCP-NEXT:    # ymm13 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm21
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
-; AVX512DQ-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm0, %ymm8
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm30
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm7, %xmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm1, %xmm28
+; AVX512DQ-FCP-NEXT:    vporq %xmm2, %xmm3, %xmm21
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [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:    # ymm1 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm8, %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm30
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [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:    # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm22
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm8, %ymm4
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm22
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512DQ-FCP-NEXT:    # ymm1 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm0, %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm29
+; 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, %ymm11, %ymm8
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm0 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm23
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [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:    # ymm1 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm11, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm18
+; 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, %ymm10, %ymm5
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm5, %zmm24
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [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:    # ymm3 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm10, %ymm5
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm0 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
 ; AVX512DQ-FCP-NEXT:    # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm1, %ymm15
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm15, %zmm24
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
-; AVX512DQ-FCP-NEXT:    # ymm8 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm15 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512DQ-FCP-NEXT:    # ymm15 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm11, %ymm11
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm11, %zmm26
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm1
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm1, %ymm10
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm11
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm11, %ymm12
-; AVX512DQ-FCP-NEXT:    vporq %ymm10, %ymm12, %ymm23
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm11, %ymm10
-; AVX512DQ-FCP-NEXT:    vporq %ymm0, %ymm10, %ymm25
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm10
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm10, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm12
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm12, %ymm9
-; AVX512DQ-FCP-NEXT:    vporq %ymm0, %ymm9, %ymm27
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm12, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm10, %ymm2
-; AVX512DQ-FCP-NEXT:    vporq %ymm0, %ymm2, %ymm16
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm9
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = xmm9[8],zero,xmm9[u,7],zero,xmm9[9],zero,xmm9[u],zero,xmm9[u,10],zero,xmm9[12],zero,xmm9[u,11]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm14
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm28, %xmm2
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm14, %xmm2
-; AVX512DQ-FCP-NEXT:    vporq %xmm0, %xmm2, %xmm28
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm10
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm10, %zmm25
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %ymm10
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm12, %ymm10, %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm12
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm12, %ymm11
+; AVX512DQ-FCP-NEXT:    vporq %ymm5, %ymm11, %ymm26
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm10, %ymm1
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm12, %ymm0
+; AVX512DQ-FCP-NEXT:    vporq %ymm1, %ymm0, %ymm27
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %ymm11
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm16, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm5
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT:    vporq %ymm0, %ymm1, %ymm16
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm5, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm11, %ymm1
+; AVX512DQ-FCP-NEXT:    vporq %ymm0, %ymm1, %ymm17
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm4
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rsi), %xmm2
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = 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 %xmm0, %xmm1, %xmm15
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rcx), %xmm13
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = zero,xmm13[6],zero,xmm13[8,u],zero,xmm13[7],zero,xmm13[9],zero,xmm13[11,u],zero,xmm13[10],zero,xmm13[12]
-; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm15
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm29, %xmm2
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm15, %xmm2
-; AVX512DQ-FCP-NEXT:    vporq %xmm0, %xmm2, %xmm29
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm11, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm31, %ymm1
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm12, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm10, %ymm2
-; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm4
-; AVX512DQ-FCP-NEXT:    vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm8 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm2
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm10 = [1,1,2,2,2,2,2,2]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm4, %ymm10, %ymm10
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512DQ-FCP-NEXT:    vpandn %ymm10, %ymm11, %ymm10
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm10, %zmm2
-; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm10
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm30 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm10, %ymm12, %ymm31
-; AVX512DQ-FCP-NEXT:    vpandnq %ymm31, %ymm30, %ymm31
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm10, %ymm8
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm31, %zmm8, %zmm8
-; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm14[0],xmm9[1],xmm14[1],xmm9[2],xmm14[2],xmm9[3],xmm14[3],xmm9[4],xmm14[4],xmm9[5],xmm14[5],xmm9[6],xmm14[6],xmm9[7],xmm14[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm14 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm9, %xmm9
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm28, %zmm9, %zmm9
-; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm13 = xmm13[0],xmm15[0],xmm13[1],xmm15[1],xmm13[2],xmm15[2],xmm13[3],xmm15[3],xmm13[4],xmm15[4],xmm13[5],xmm15[5],xmm13[6],xmm15[6],xmm13[7],xmm15[7]
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm15 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm13, %xmm13
-; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm29, %zmm13, %zmm13
-; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r8), %zmm28
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm28, %zmm10
+; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm28, %xmm8
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
+; AVX512DQ-FCP-NEXT:    vporq %xmm0, %xmm8, %xmm28
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm18, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm12, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm10, %ymm3
+; AVX512DQ-FCP-NEXT:    vpor %ymm0, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm3
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm5, %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm29, %ymm5
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm11, %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqa 32(%r8), %ymm8
+; AVX512DQ-FCP-NEXT:    vpor %ymm3, %ymm5, %ymm3
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm10 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm8, %ymm5
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm11 = [1,1,2,2,2,2,2,2]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm8, %ymm11, %ymm11
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
+; AVX512DQ-FCP-NEXT:    vpandn %ymm11, %ymm12, %ymm11
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm11, %zmm5
+; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm11
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [4,6,5,5,5,5,4,6,6,6,6,6,7,7,7,7]
+; AVX512DQ-FCP-NEXT:    vpermd %ymm11, %ymm29, %ymm30
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm11, %ymm10
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512DQ-FCP-NEXT:    vpandnq %ymm30, %ymm11, %ymm30
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm30, %zmm10, %zmm10
+; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm4 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm2, %zmm2
+; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3],xmm13[4],xmm1[4],xmm13[5],xmm1[5],xmm13[6],xmm1[6],xmm13[7],xmm1[7]
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} xmm13 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm28, %zmm1, %zmm1
 ; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm6, %xmm6
 ; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} zmm7 = [2,2,3,3,8,8,9,9]
 ; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm7, %zmm0
-; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
-; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm7, %zmm1
-; AVX512DQ-FCP-NEXT:    vporq %zmm21, %zmm22, %zmm3
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm3 = zmm3[2,2,3,3,6,6,7,7]
-; AVX512DQ-FCP-NEXT:    vporq %zmm24, %zmm26, %zmm5
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm5 = zmm5[2,2,3,3,6,6,7,7]
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm6 & (zmm5 ^ zmm3))
-; AVX512DQ-FCP-NEXT:    vpermt2d %zmm28, %zmm12, %zmm4
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm4 = zmm4 ^ (zmm30 & (zmm4 ^ zmm5))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 256(%r9)
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm18[0,0,1,1]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm20[0,0,1,1]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm4, %zmm4
-; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm3 ^ (zmm5 & (zmm4 ^ zmm3))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm2 = zmm2 | (zmm4 & zmm11)
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm25[2,2,3,3]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm23, %zmm3
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm16[2,2,3,3]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm27, %zmm4
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm4 = zmm3 ^ (zmm5 & (zmm4 ^ zmm3))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm0 ^ (zmm6 & (zmm1 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm8 = zmm8 | (zmm4 & mem)
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm9[0,0,1,1,4,4,5,5]
-; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm3 = zmm13[0,0,1,1,4,4,5,5]
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm0))
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm10, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm3))
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512DQ-FCP-NEXT:    vpermd %zmm28, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm1))
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 128(%r9)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, 64(%r9)
+; AVX512DQ-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm14[0],xmm9[0],xmm14[1],xmm9[1],xmm14[2],xmm9[2],xmm14[3],xmm9[3],xmm14[4],xmm9[4],xmm14[5],xmm9[5],xmm14[6],xmm9[6],xmm14[7],xmm9[7]
+; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm6, %xmm4
+; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm7, %zmm3
+; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r8), %zmm4
+; AVX512DQ-FCP-NEXT:    vpermd %zmm8, %zmm29, %zmm6
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [6,6,6,0,7,7,7,7,0,16,16,16,16,0,17,17]
+; AVX512DQ-FCP-NEXT:    vpermi2d %zmm8, %zmm4, %zmm7
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm19[0,0,1,1]
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm31, %zmm8, %zmm8
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm21[0,0,1,1]
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm20, %zmm9, %zmm9
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm13 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm8 ^ (zmm13 & (zmm9 ^ zmm8))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm5 = zmm5 | (zmm9 & zmm12)
+; AVX512DQ-FCP-NEXT:    vporq %zmm22, %zmm23, %zmm8
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT:    vporq %zmm24, %zmm25, %zmm9
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm9 = zmm9[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm8 ^ (zmm12 & (zmm9 ^ zmm8))
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm6 = zmm6 ^ (zmm11 & (zmm6 ^ zmm9))
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm27[2,2,3,3]
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm26, %zmm8
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm17[2,2,3,3]
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm16, %zmm9
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm9 = zmm8 ^ (zmm13 & (zmm9 ^ zmm8))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm12 & (zmm3 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm10 = zmm10 | (zmm9 & mem)
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm2[0,0,1,1,4,4,5,5]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[0,0,1,1,4,4,5,5]
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
+; AVX512DQ-FCP-NEXT:    vpermd %zmm4, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))
+; AVX512DQ-FCP-NEXT:    vpternlogd {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm3))
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 128(%r9)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, 64(%r9)
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%r9)
-; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 192(%r9)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 256(%r9)
+; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 192(%r9)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
 ;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
index 852afee203ea1..290ad75a62749 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
@@ -2994,13 +2994,12 @@ define void @store_i8_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vmovdqa (%rcx), %ymm12
 ; AVX512BW-NEXT:    vmovdqa (%r8), %ymm7
 ; AVX512BW-NEXT:    vmovdqa (%r9), %ymm8
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[16],ymm10[16],ymm9[17],ymm10[17],ymm9[18],ymm10[18],ymm9[19],ymm10[19],ymm9[20],ymm10[20],ymm9[21],ymm10[21],ymm9[22],ymm10[22],ymm9[23],ymm10[23]
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm3 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[16],ymm10[16],ymm9[17],ymm10[17],ymm9[18],ymm10[18],ymm9[19],ymm10[19],ymm9[20],ymm10[20],ymm9[21],ymm10[21],ymm9[22],ymm10[22],ymm9[23],ymm10[23]
 ; AVX512BW-NEXT:    vmovdqa (%rsi), %xmm1
 ; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm2
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
-; AVX512BW-NEXT:    vpermw %zmm0, %zmm3, %zmm0
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm0 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512BW-NEXT:    vpermi2w %zmm3, %zmm4, %zmm0
 ; AVX512BW-NEXT:    vmovdqa (%rcx), %xmm3
 ; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm4
 ; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
@@ -3017,11 +3016,11 @@ define void @store_i8_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vmovdqa (%r9), %xmm5
 ; AVX512BW-NEXT:    vmovdqa (%r8), %xmm6
 ; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm14 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm13, %zmm14, %zmm13
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm14 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,26,25,24,27,26,25,24,27,26,25,24,27,28,28,28,28]
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,42,41,40,43,42,41,40,43,42,41,40,43,44,44,44,44]
+; AVX512BW-NEXT:    vpermi2w %zmm13, %zmm14, %zmm15
 ; AVX512BW-NEXT:    movl $1227133513, %ecx # imm = 0x49249249
 ; AVX512BW-NEXT:    kmovd %ecx, %k2
-; AVX512BW-NEXT:    vpermw %zmm13, %zmm14, %zmm0 {%k2}
+; AVX512BW-NEXT:    vmovdqu16 %zmm15, %zmm0 {%k2}
 ; AVX512BW-NEXT:    vpbroadcastq {{.*#+}} ymm13 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
 ; AVX512BW-NEXT:    vpshufb %ymm13, %ymm10, %ymm14
 ; AVX512BW-NEXT:    vpshufb %ymm13, %ymm9, %ymm13
@@ -3170,22 +3169,21 @@ define void @store_i8_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm7, %zmm8 {%k1}
 ; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[4],ymm6[4],ymm5[5],ymm6[5],ymm5[6],ymm6[6],ymm5[7],ymm6[7],ymm5[16],ymm6[16],ymm5[17],ymm6[17],ymm5[18],ymm6[18],ymm5[19],ymm6[19],ymm5[20],ymm6[20],ymm5[21],ymm6[21],ymm5[22],ymm6[22],ymm5[23],ymm6[23]
 ; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = 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:    vinserti64x4 $1, %ymm5, %zmm6, %zmm5
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm7 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
+; AVX512BW-FCP-NEXT:    vpermi2w %zmm5, %zmm6, %zmm7
 ; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[16],ymm3[16],ymm2[17],ymm3[17],ymm2[18],ymm3[18],ymm2[19],ymm3[19],ymm2[20],ymm3[20],ymm2[21],ymm3[21],ymm2[22],ymm3[22],ymm2[23],ymm3[23]
 ; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm12[8],xmm9[8],xmm12[9],xmm9[9],xmm12[10],xmm9[10],xmm12[11],xmm9[11],xmm12[12],xmm9[12],xmm12[13],xmm9[13],xmm12[14],xmm9[14],xmm12[15],xmm9[15]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
-; AVX512BW-FCP-NEXT:    vpermw %zmm2, %zmm3, %zmm2
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,25,24,27,26,25,24,27,26,25,24,27,26,29,28,31,30]
-; AVX512BW-FCP-NEXT:    vpermw %zmm5, %zmm3, %zmm2 {%k1}
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm5 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512BW-FCP-NEXT:    vpermi2w %zmm2, %zmm3, %zmm5
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm7, %zmm5 {%k1}
 ; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
 ; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm10[8],xmm15[9],xmm10[9],xmm15[10],xmm10[10],xmm15[11],xmm10[11],xmm15[12],xmm10[12],xmm15[13],xmm10[13],xmm15[14],xmm10[14],xmm15[15],xmm10[15]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,26,25,24,27,26,25,24,27,26,25,24,27,28,28,28,28]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,42,41,40,43,42,41,40,43,42,41,40,43,44,44,44,44]
+; AVX512BW-FCP-NEXT:    vpermi2w %zmm0, %zmm1, %zmm2
 ; AVX512BW-FCP-NEXT:    movl $1227133513, %ecx # imm = 0x49249249
 ; AVX512BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512BW-FCP-NEXT:    vpermw %zmm0, %zmm1, %zmm2 {%k1}
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm2, %zmm5 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rax)
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, (%rax)
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rax)
 ; AVX512BW-FCP-NEXT:    vzeroupper
@@ -3200,13 +3198,12 @@ define void @store_i8_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rcx), %ymm12
 ; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %ymm7
 ; AVX512DQ-BW-NEXT:    vmovdqa (%r9), %ymm8
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[16],ymm10[16],ymm9[17],ymm10[17],ymm9[18],ymm10[18],ymm9[19],ymm10[19],ymm9[20],ymm10[20],ymm9[21],ymm10[21],ymm9[22],ymm10[22],ymm9[23],ymm10[23]
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm3 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[16],ymm10[16],ymm9[17],ymm10[17],ymm9[18],ymm10[18],ymm9[19],ymm10[19],ymm9[20],ymm10[20],ymm9[21],ymm10[21],ymm9[22],ymm10[22],ymm9[23],ymm10[23]
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %xmm1
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm2
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
-; AVX512DQ-BW-NEXT:    vpermw %zmm0, %zmm3, %zmm0
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm0 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512DQ-BW-NEXT:    vpermi2w %zmm3, %zmm4, %zmm0
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rcx), %xmm3
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm4
 ; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
@@ -3223,11 +3220,11 @@ define void @store_i8_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vmovdqa (%r9), %xmm5
 ; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm6
 ; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm14 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm13, %zmm14, %zmm13
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm14 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,26,25,24,27,26,25,24,27,26,25,24,27,28,28,28,28]
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,42,41,40,43,42,41,40,43,42,41,40,43,44,44,44,44]
+; AVX512DQ-BW-NEXT:    vpermi2w %zmm13, %zmm14, %zmm15
 ; AVX512DQ-BW-NEXT:    movl $1227133513, %ecx # imm = 0x49249249
 ; AVX512DQ-BW-NEXT:    kmovd %ecx, %k2
-; AVX512DQ-BW-NEXT:    vpermw %zmm13, %zmm14, %zmm0 {%k2}
+; AVX512DQ-BW-NEXT:    vmovdqu16 %zmm15, %zmm0 {%k2}
 ; AVX512DQ-BW-NEXT:    vpbroadcastq {{.*#+}} ymm13 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
 ; AVX512DQ-BW-NEXT:    vpshufb %ymm13, %ymm10, %ymm14
 ; AVX512DQ-BW-NEXT:    vpshufb %ymm13, %ymm9, %ymm13
@@ -3376,22 +3373,21 @@ define void @store_i8_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm7, %zmm8 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[4],ymm6[4],ymm5[5],ymm6[5],ymm5[6],ymm6[6],ymm5[7],ymm6[7],ymm5[16],ymm6[16],ymm5[17],ymm6[17],ymm5[18],ymm6[18],ymm5[19],ymm6[19],ymm5[20],ymm6[20],ymm5[21],ymm6[21],ymm5[22],ymm6[22],ymm5[23],ymm6[23]
 ; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = 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:    vinserti64x4 $1, %ymm5, %zmm6, %zmm5
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm7 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2w %zmm5, %zmm6, %zmm7
 ; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[16],ymm3[16],ymm2[17],ymm3[17],ymm2[18],ymm3[18],ymm2[19],ymm3[19],ymm2[20],ymm3[20],ymm2[21],ymm3[21],ymm2[22],ymm3[22],ymm2[23],ymm3[23]
 ; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm12[8],xmm9[8],xmm12[9],xmm9[9],xmm12[10],xmm9[10],xmm12[11],xmm9[11],xmm12[12],xmm9[12],xmm12[13],xmm9[13],xmm12[14],xmm9[14],xmm12[15],xmm9[15]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm2, %zmm3, %zmm2
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,25,24,27,26,25,24,27,26,25,24,27,26,29,28,31,30]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm5, %zmm3, %zmm2 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm5 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2w %zmm2, %zmm3, %zmm5
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm7, %zmm5 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
 ; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm10[8],xmm15[9],xmm10[9],xmm15[10],xmm10[10],xmm15[11],xmm10[11],xmm15[12],xmm10[12],xmm15[13],xmm10[13],xmm15[14],xmm10[14],xmm15[15],xmm10[15]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,26,25,24,27,26,25,24,27,26,25,24,27,28,28,28,28]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,42,41,40,43,42,41,40,43,42,41,40,43,44,44,44,44]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2w %zmm0, %zmm1, %zmm2
 ; AVX512DQ-BW-FCP-NEXT:    movl $1227133513, %ecx # imm = 0x49249249
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm0, %zmm1, %zmm2 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm2, %zmm5 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, (%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
@@ -6420,22 +6416,21 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vmovdqa64 (%r9), %zmm12
 ; AVX512BW-NEXT:    vmovdqa64 (%rsi), %ymm16
 ; AVX512BW-NEXT:    vmovdqa64 (%rdi), %ymm17
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm17[0],ymm16[0],ymm17[1],ymm16[1],ymm17[2],ymm16[2],ymm17[3],ymm16[3],ymm17[4],ymm16[4],ymm17[5],ymm16[5],ymm17[6],ymm16[6],ymm17[7],ymm16[7],ymm17[16],ymm16[16],ymm17[17],ymm16[17],ymm17[18],ymm16[18],ymm17[19],ymm16[19],ymm17[20],ymm16[20],ymm17[21],ymm16[21],ymm17[22],ymm16[22],ymm17[23],ymm16[23]
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm2 = ymm17[0],ymm16[0],ymm17[1],ymm16[1],ymm17[2],ymm16[2],ymm17[3],ymm16[3],ymm17[4],ymm16[4],ymm17[5],ymm16[5],ymm17[6],ymm16[6],ymm17[7],ymm16[7],ymm17[16],ymm16[16],ymm17[17],ymm16[17],ymm17[18],ymm16[18],ymm17[19],ymm16[19],ymm17[20],ymm16[20],ymm17[21],ymm16[21],ymm17[22],ymm16[22],ymm17[23],ymm16[23]
 ; AVX512BW-NEXT:    vmovdqa (%rsi), %xmm1
 ; AVX512BW-NEXT:    vmovdqa 32(%rsi), %xmm9
 ; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm3
 ; AVX512BW-NEXT:    vmovdqa 32(%rdi), %xmm11
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm2 = 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:    vinserti64x4 $1, %ymm0, %zmm2, %zmm0
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm7 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
-; AVX512BW-NEXT:    vpermw %zmm0, %zmm7, %zmm0
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = 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:    vpmovsxbw {{.*#+}} zmm20 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512BW-NEXT:    vpermt2w %zmm2, %zmm20, %zmm0
 ; AVX512BW-NEXT:    vmovdqa (%rcx), %xmm2
 ; AVX512BW-NEXT:    vmovdqa 32(%rcx), %xmm8
 ; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm4
 ; AVX512BW-NEXT:    vmovdqa 32(%rdx), %xmm10
 ; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15]
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
-; AVX512BW-NEXT:    vpermw %ymm5, %ymm20, %ymm5
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm22 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
+; AVX512BW-NEXT:    vpermw %ymm5, %ymm22, %ymm5
 ; AVX512BW-NEXT:    vmovdqa64 (%rcx), %ymm18
 ; AVX512BW-NEXT:    vmovdqa64 (%rdx), %ymm19
 ; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm6 = ymm19[0],ymm18[0],ymm19[1],ymm18[1],ymm19[2],ymm18[2],ymm19[3],ymm18[3],ymm19[4],ymm18[4],ymm19[5],ymm18[5],ymm19[6],ymm18[6],ymm19[7],ymm18[7],ymm19[16],ymm18[16],ymm19[17],ymm18[17],ymm19[18],ymm18[18],ymm19[19],ymm18[19],ymm19[20],ymm18[20],ymm19[21],ymm18[21],ymm19[22],ymm18[22],ymm19[23],ymm18[23]
@@ -6448,79 +6443,78 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vmovdqa (%r8), %xmm5
 ; AVX512BW-NEXT:    vmovdqa 32(%r8), %xmm13
 ; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm23 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
-; AVX512BW-NEXT:    vpermw %ymm6, %ymm23, %ymm6
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm24 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
+; AVX512BW-NEXT:    vpermw %ymm6, %ymm24, %ymm6
 ; AVX512BW-NEXT:    vmovdqa64 (%r8), %ymm21
-; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} ymm24 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
-; AVX512BW-NEXT:    # ymm24 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT:    vpshufb %ymm24, %ymm21, %ymm15
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm15, %zmm6, %zmm6
+; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} ymm25 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
+; AVX512BW-NEXT:    # ymm25 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-NEXT:    vpshufb %ymm25, %ymm21, %ymm7
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm6, %zmm6
 ; AVX512BW-NEXT:    movl $1227133513, %r10d # imm = 0x49249249
 ; AVX512BW-NEXT:    kmovd %r10d, %k2
 ; AVX512BW-NEXT:    vmovdqu16 %zmm6, %zmm0 {%k2}
 ; AVX512BW-NEXT:    vmovdqa (%r9), %xmm6
 ; AVX512BW-NEXT:    vmovdqa 32(%r9), %xmm15
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm22 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-NEXT:    vpermw %ymm22, %ymm23, %ymm25
-; AVX512BW-NEXT:    vmovdqa64 (%r9), %ymm22
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-NEXT:    vpermw %ymm7, %ymm24, %ymm7
+; AVX512BW-NEXT:    vmovdqa64 (%r9), %ymm23
 ; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} ymm26 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
 ; AVX512BW-NEXT:    # ymm26 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT:    vpshufb %ymm26, %ymm22, %ymm27
+; AVX512BW-NEXT:    vpshufb %ymm26, %ymm23, %ymm27
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm27 = ymm27[2,2,2,3]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm27, %zmm25, %zmm25
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm27, %zmm7, %zmm7
 ; AVX512BW-NEXT:    movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082
 ; AVX512BW-NEXT:    kmovq %r10, %k3
-; AVX512BW-NEXT:    vmovdqu8 %zmm25, %zmm0 {%k3}
-; AVX512BW-NEXT:    vmovdqa64 32(%rsi), %ymm25
-; AVX512BW-NEXT:    vmovdqa64 32(%rdi), %ymm27
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm28 = ymm27[0],ymm25[0],ymm27[1],ymm25[1],ymm27[2],ymm25[2],ymm27[3],ymm25[3],ymm27[4],ymm25[4],ymm27[5],ymm25[5],ymm27[6],ymm25[6],ymm27[7],ymm25[7],ymm27[16],ymm25[16],ymm27[17],ymm25[17],ymm27[18],ymm25[18],ymm27[19],ymm25[19],ymm27[20],ymm25[20],ymm27[21],ymm25[21],ymm27[22],ymm25[22],ymm27[23],ymm25[23]
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm29 = xmm11[8],xmm9[8],xmm11[9],xmm9[9],xmm11[10],xmm9[10],xmm11[11],xmm9[11],xmm11[12],xmm9[12],xmm11[13],xmm9[13],xmm11[14],xmm9[14],xmm11[15],xmm9[15]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm28, %zmm29, %zmm28
-; AVX512BW-NEXT:    vpermw %zmm28, %zmm7, %zmm7
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm28 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
-; AVX512BW-NEXT:    vpermw %ymm28, %ymm20, %ymm20
-; AVX512BW-NEXT:    vmovdqa64 32(%rcx), %ymm28
-; AVX512BW-NEXT:    vmovdqa64 32(%rdx), %ymm29
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm30 = ymm29[0],ymm28[0],ymm29[1],ymm28[1],ymm29[2],ymm28[2],ymm29[3],ymm28[3],ymm29[4],ymm28[4],ymm29[5],ymm28[5],ymm29[6],ymm28[6],ymm29[7],ymm28[7],ymm29[16],ymm28[16],ymm29[17],ymm28[17],ymm29[18],ymm28[18],ymm29[19],ymm28[19],ymm29[20],ymm28[20],ymm29[21],ymm28[21],ymm29[22],ymm28[22],ymm29[23],ymm28[23]
-; AVX512BW-NEXT:    vprold $16, %ymm30, %ymm30
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm30 = ymm30[2,2,2,3]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm30, %zmm20, %zmm20
+; AVX512BW-NEXT:    vmovdqu8 %zmm7, %zmm0 {%k3}
+; AVX512BW-NEXT:    vmovdqa64 32(%rsi), %ymm27
+; AVX512BW-NEXT:    vmovdqa64 32(%rdi), %ymm28
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm29 = ymm28[0],ymm27[0],ymm28[1],ymm27[1],ymm28[2],ymm27[2],ymm28[3],ymm27[3],ymm28[4],ymm27[4],ymm28[5],ymm27[5],ymm28[6],ymm27[6],ymm28[7],ymm27[7],ymm28[16],ymm27[16],ymm28[17],ymm27[17],ymm28[18],ymm27[18],ymm28[19],ymm27[19],ymm28[20],ymm27[20],ymm28[21],ymm27[21],ymm28[22],ymm27[22],ymm28[23],ymm27[23]
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm11[8],xmm9[8],xmm11[9],xmm9[9],xmm11[10],xmm9[10],xmm11[11],xmm9[11],xmm11[12],xmm9[12],xmm11[13],xmm9[13],xmm11[14],xmm9[14],xmm11[15],xmm9[15]
+; AVX512BW-NEXT:    vpermt2w %zmm29, %zmm20, %zmm7
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
+; AVX512BW-NEXT:    vpermw %ymm20, %ymm22, %ymm20
+; AVX512BW-NEXT:    vmovdqa64 32(%rcx), %ymm29
+; AVX512BW-NEXT:    vmovdqa64 32(%rdx), %ymm30
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm22 = ymm30[0],ymm29[0],ymm30[1],ymm29[1],ymm30[2],ymm29[2],ymm30[3],ymm29[3],ymm30[4],ymm29[4],ymm30[5],ymm29[5],ymm30[6],ymm29[6],ymm30[7],ymm29[7],ymm30[16],ymm29[16],ymm30[17],ymm29[17],ymm30[18],ymm29[18],ymm30[19],ymm29[19],ymm30[20],ymm29[20],ymm30[21],ymm29[21],ymm30[22],ymm29[22],ymm30[23],ymm29[23]
+; AVX512BW-NEXT:    vprold $16, %ymm22, %ymm22
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm20, %zmm20
 ; AVX512BW-NEXT:    vmovdqu16 %zmm20, %zmm7 {%k1}
 ; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm13[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-NEXT:    vpermw %ymm20, %ymm23, %ymm20
-; AVX512BW-NEXT:    vmovdqa64 32(%r8), %ymm30
-; AVX512BW-NEXT:    vpshufb %ymm24, %ymm30, %ymm24
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm24, %zmm20, %zmm20
+; AVX512BW-NEXT:    vpermw %ymm20, %ymm24, %ymm20
+; AVX512BW-NEXT:    vmovdqa64 32(%r8), %ymm31
+; AVX512BW-NEXT:    vpshufb %ymm25, %ymm31, %ymm22
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm20, %zmm20
 ; AVX512BW-NEXT:    vmovdqu16 %zmm20, %zmm7 {%k2}
 ; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm15[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-NEXT:    vpermw %ymm20, %ymm23, %ymm20
+; AVX512BW-NEXT:    vpermw %ymm20, %ymm24, %ymm20
 ; AVX512BW-NEXT:    vmovdqa64 32(%r9), %ymm24
-; AVX512BW-NEXT:    vpshufb %ymm26, %ymm24, %ymm23
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm23 = ymm23[2,2,2,3]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm23, %zmm20, %zmm20
+; AVX512BW-NEXT:    vpshufb %ymm26, %ymm24, %ymm22
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm20, %zmm20
 ; AVX512BW-NEXT:    vmovdqu8 %zmm20, %zmm7 {%k3}
-; AVX512BW-NEXT:    vpbroadcastq {{.*#+}} ymm26 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512BW-NEXT:    vpshufb %ymm26, %ymm25, %ymm20
-; AVX512BW-NEXT:    vpshufb %ymm26, %ymm27, %ymm23
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm20 = ymm23[0],ymm20[0],ymm23[1],ymm20[1],ymm23[2],ymm20[2],ymm23[3],ymm20[3],ymm23[4],ymm20[4],ymm23[5],ymm20[5],ymm23[6],ymm20[6],ymm23[7],ymm20[7],ymm23[16],ymm20[16],ymm23[17],ymm20[17],ymm23[18],ymm20[18],ymm23[19],ymm20[19],ymm23[20],ymm20[20],ymm23[21],ymm20[21],ymm23[22],ymm20[22],ymm23[23],ymm20[23]
+; AVX512BW-NEXT:    vpbroadcastq {{.*#+}} ymm25 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512BW-NEXT:    vpshufb %ymm25, %ymm27, %ymm20
+; AVX512BW-NEXT:    vpshufb %ymm25, %ymm28, %ymm22
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm20 = ymm22[0],ymm20[0],ymm22[1],ymm20[1],ymm22[2],ymm20[2],ymm22[3],ymm20[3],ymm22[4],ymm20[4],ymm22[5],ymm20[5],ymm22[6],ymm20[6],ymm22[7],ymm20[7],ymm22[16],ymm20[16],ymm22[17],ymm20[17],ymm22[18],ymm20[18],ymm22[19],ymm20[19],ymm22[20],ymm20[20],ymm22[21],ymm20[21],ymm22[22],ymm20[22],ymm22[23],ymm20[23]
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} ymm23 = ymm27[8],ymm25[8],ymm27[9],ymm25[9],ymm27[10],ymm25[10],ymm27[11],ymm25[11],ymm27[12],ymm25[12],ymm27[13],ymm25[13],ymm27[14],ymm25[14],ymm27[15],ymm25[15],ymm27[24],ymm25[24],ymm27[25],ymm25[25],ymm27[26],ymm25[26],ymm27[27],ymm25[27],ymm27[28],ymm25[28],ymm27[29],ymm25[29],ymm27[30],ymm25[30],ymm27[31],ymm25[31]
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm25 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512BW-NEXT:    vpermw %ymm23, %ymm25, %ymm23
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm23, %zmm20, %zmm27
-; AVX512BW-NEXT:    vpbroadcastq {{.*#+}} ymm23 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
-; AVX512BW-NEXT:    vpshufb %ymm23, %ymm28, %ymm20
-; AVX512BW-NEXT:    vpshufb %ymm23, %ymm29, %ymm31
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm20 = ymm31[0],ymm20[0],ymm31[1],ymm20[1],ymm31[2],ymm20[2],ymm31[3],ymm20[3],ymm31[4],ymm20[4],ymm31[5],ymm20[5],ymm31[6],ymm20[6],ymm31[7],ymm20[7],ymm31[16],ymm20[16],ymm31[17],ymm20[17],ymm31[18],ymm20[18],ymm31[19],ymm20[19],ymm31[20],ymm20[20],ymm31[21],ymm20[21],ymm31[22],ymm20[22],ymm31[23],ymm20[23]
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} ymm22 = ymm28[8],ymm27[8],ymm28[9],ymm27[9],ymm28[10],ymm27[10],ymm28[11],ymm27[11],ymm28[12],ymm27[12],ymm28[13],ymm27[13],ymm28[14],ymm27[14],ymm28[15],ymm27[15],ymm28[24],ymm27[24],ymm28[25],ymm27[25],ymm28[26],ymm27[26],ymm28[27],ymm27[27],ymm28[28],ymm27[28],ymm28[29],ymm27[29],ymm28[30],ymm27[30],ymm28[31],ymm27[31]
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm26 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512BW-NEXT:    vpermw %ymm22, %ymm26, %ymm22
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm20, %zmm27
+; AVX512BW-NEXT:    vpbroadcastq {{.*#+}} ymm22 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
+; AVX512BW-NEXT:    vpshufb %ymm22, %ymm29, %ymm20
+; AVX512BW-NEXT:    vpshufb %ymm22, %ymm30, %ymm28
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm20 = ymm28[0],ymm20[0],ymm28[1],ymm20[1],ymm28[2],ymm20[2],ymm28[3],ymm20[3],ymm28[4],ymm20[4],ymm28[5],ymm20[5],ymm28[6],ymm20[6],ymm28[7],ymm20[7],ymm28[16],ymm20[16],ymm28[17],ymm20[17],ymm28[18],ymm20[18],ymm28[19],ymm20[19],ymm28[20],ymm20[20],ymm28[21],ymm20[21],ymm28[22],ymm20[22],ymm28[23],ymm20[23]
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} ymm28 = ymm29[8],ymm28[8],ymm29[9],ymm28[9],ymm29[10],ymm28[10],ymm29[11],ymm28[11],ymm29[12],ymm28[12],ymm29[13],ymm28[13],ymm29[14],ymm28[14],ymm29[15],ymm28[15],ymm29[24],ymm28[24],ymm29[25],ymm28[25],ymm29[26],ymm28[26],ymm29[27],ymm28[27],ymm29[28],ymm28[28],ymm29[29],ymm28[29],ymm29[30],ymm28[30],ymm29[31],ymm28[31]
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} ymm28 = ymm30[8],ymm29[8],ymm30[9],ymm29[9],ymm30[10],ymm29[10],ymm30[11],ymm29[11],ymm30[12],ymm29[12],ymm30[13],ymm29[13],ymm30[14],ymm29[14],ymm30[15],ymm29[15],ymm30[24],ymm29[24],ymm30[25],ymm29[25],ymm30[26],ymm29[26],ymm30[27],ymm29[27],ymm30[28],ymm29[28],ymm30[29],ymm29[29],ymm30[30],ymm29[30],ymm30[31],ymm29[31]
 ; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm29 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
 ; AVX512BW-NEXT:    vpermw %ymm28, %ymm29, %ymm28
 ; AVX512BW-NEXT:    vinserti64x4 $1, %ymm28, %zmm20, %zmm20
 ; AVX512BW-NEXT:    vmovdqu16 %zmm27, %zmm20 {%k1}
-; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm27 = zmm30[0,1,2,3],zmm14[4,5,6,7]
+; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm27 = zmm31[0,1,2,3],zmm14[4,5,6,7]
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm27 = zmm27[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm27 = zmm27[2,2,2,3,6,6,6,7]
 ; AVX512BW-NEXT:    movl $-1840700270, %ecx # imm = 0x92492492
@@ -6532,15 +6526,15 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    movabsq $-9076969306111049208, %rcx # imm = 0x8208208208208208
 ; AVX512BW-NEXT:    kmovq %rcx, %k3
 ; AVX512BW-NEXT:    vmovdqu8 %zmm24, %zmm20 {%k3}
-; AVX512BW-NEXT:    vpshufb %ymm26, %ymm16, %ymm24
-; AVX512BW-NEXT:    vpshufb %ymm26, %ymm17, %ymm26
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm24 = ymm26[0],ymm24[0],ymm26[1],ymm24[1],ymm26[2],ymm24[2],ymm26[3],ymm24[3],ymm26[4],ymm24[4],ymm26[5],ymm24[5],ymm26[6],ymm24[6],ymm26[7],ymm24[7],ymm26[16],ymm24[16],ymm26[17],ymm24[17],ymm26[18],ymm24[18],ymm26[19],ymm24[19],ymm26[20],ymm24[20],ymm26[21],ymm24[21],ymm26[22],ymm24[22],ymm26[23],ymm24[23]
+; AVX512BW-NEXT:    vpshufb %ymm25, %ymm16, %ymm24
+; AVX512BW-NEXT:    vpshufb %ymm25, %ymm17, %ymm25
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm24 = ymm25[0],ymm24[0],ymm25[1],ymm24[1],ymm25[2],ymm24[2],ymm25[3],ymm24[3],ymm25[4],ymm24[4],ymm25[5],ymm24[5],ymm25[6],ymm24[6],ymm25[7],ymm24[7],ymm25[16],ymm24[16],ymm25[17],ymm24[17],ymm25[18],ymm24[18],ymm25[19],ymm24[19],ymm25[20],ymm24[20],ymm25[21],ymm24[21],ymm25[22],ymm24[22],ymm25[23],ymm24[23]
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
 ; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} ymm16 = ymm17[8],ymm16[8],ymm17[9],ymm16[9],ymm17[10],ymm16[10],ymm17[11],ymm16[11],ymm17[12],ymm16[12],ymm17[13],ymm16[13],ymm17[14],ymm16[14],ymm17[15],ymm16[15],ymm17[24],ymm16[24],ymm17[25],ymm16[25],ymm17[26],ymm16[26],ymm17[27],ymm16[27],ymm17[28],ymm16[28],ymm17[29],ymm16[29],ymm17[30],ymm16[30],ymm17[31],ymm16[31]
-; AVX512BW-NEXT:    vpermw %ymm16, %ymm25, %ymm16
+; AVX512BW-NEXT:    vpermw %ymm16, %ymm26, %ymm16
 ; AVX512BW-NEXT:    vinserti64x4 $1, %ymm16, %zmm24, %zmm17
-; AVX512BW-NEXT:    vpshufb %ymm23, %ymm18, %ymm16
-; AVX512BW-NEXT:    vpshufb %ymm23, %ymm19, %ymm24
+; AVX512BW-NEXT:    vpshufb %ymm22, %ymm18, %ymm16
+; AVX512BW-NEXT:    vpshufb %ymm22, %ymm19, %ymm24
 ; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} ymm16 = ymm24[0],ymm16[0],ymm24[1],ymm16[1],ymm24[2],ymm16[2],ymm24[3],ymm16[3],ymm24[4],ymm16[4],ymm24[5],ymm16[5],ymm24[6],ymm16[6],ymm24[7],ymm16[7],ymm24[16],ymm16[16],ymm24[17],ymm16[17],ymm24[18],ymm16[18],ymm24[19],ymm16[19],ymm24[20],ymm16[20],ymm24[21],ymm16[21],ymm24[22],ymm16[22],ymm24[23],ymm16[23]
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm16 = ymm16[2,2,2,3]
 ; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} ymm18 = ymm19[8],ymm18[8],ymm19[9],ymm18[9],ymm19[10],ymm18[10],ymm19[11],ymm18[11],ymm19[12],ymm18[12],ymm19[13],ymm18[13],ymm19[14],ymm18[14],ymm19[15],ymm18[15],ymm19[24],ymm18[24],ymm19[25],ymm18[25],ymm19[26],ymm18[26],ymm19[27],ymm18[27],ymm19[28],ymm18[28],ymm19[29],ymm18[29],ymm19[30],ymm18[30],ymm19[31],ymm18[31]
@@ -6551,7 +6545,7 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm14 = zmm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm14 = zmm14[2,2,2,3,6,6,6,7]
 ; AVX512BW-NEXT:    vmovdqu16 %zmm14, %zmm16 {%k2}
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm12, %zmm12
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm23, %zmm12, %zmm12
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm12 = zmm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm12 = zmm12[2,2,2,3,6,6,6,7]
 ; AVX512BW-NEXT:    vmovdqu8 %zmm12, %zmm16 {%k3}
@@ -6564,8 +6558,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} ymm11 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
 ; AVX512BW-NEXT:    vpermw %ymm9, %ymm11, %ymm9
 ; AVX512BW-NEXT:    vinserti64x4 $1, %ymm14, %zmm9, %zmm9
-; AVX512BW-NEXT:    vpshufb %xmm23, %xmm8, %xmm14
-; AVX512BW-NEXT:    vpshufb %xmm23, %xmm10, %xmm17
+; AVX512BW-NEXT:    vpshufb %xmm22, %xmm8, %xmm14
+; AVX512BW-NEXT:    vpshufb %xmm22, %xmm10, %xmm17
 ; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm14 = xmm17[0],xmm14[0],xmm17[1],xmm14[1],xmm17[2],xmm14[2],xmm17[3],xmm14[3],xmm17[4],xmm14[4],xmm17[5],xmm14[5],xmm17[6],xmm14[6],xmm17[7],xmm14[7]
 ; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3],xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
 ; AVX512BW-NEXT:    vprold $16, %xmm8, %xmm8
@@ -6578,11 +6572,11 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vinserti32x4 $2, %xmm10, %zmm8, %zmm8
 ; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm10 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,22,21,16,23,22,21,16,23,22,21,16,23,17,17,17,17]
 ; AVX512BW-NEXT:    vpermw %zmm8, %zmm10, %zmm9 {%k1}
-; AVX512BW-NEXT:    vpshufd {{.*#+}} xmm8 = xmm15[2,1,2,3]
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm8[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm13 = xmm15[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-NEXT:    vinserti32x4 $2, %xmm8, %zmm13, %zmm8
-; AVX512BW-NEXT:    vpermw %zmm8, %zmm10, %zmm8
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm15[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-NEXT:    vpshufd {{.*#+}} xmm13 = xmm15[2,1,2,3]
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm13 = xmm13[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm14 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
+; AVX512BW-NEXT:    vpermt2w %zmm13, %zmm14, %zmm8
 ; AVX512BW-NEXT:    movabsq $585610922974906400, %rcx # imm = 0x820820820820820
 ; AVX512BW-NEXT:    kmovq %rcx, %k3
 ; AVX512BW-NEXT:    vmovdqu8 %zmm8, %zmm9 {%k3}
@@ -6593,8 +6587,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = 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:    vpermw %ymm1, %ymm11, %ymm1
 ; AVX512BW-NEXT:    vinserti64x4 $1, %ymm8, %zmm1, %zmm1
-; AVX512BW-NEXT:    vpshufb %xmm23, %xmm2, %xmm3
-; AVX512BW-NEXT:    vpshufb %xmm23, %xmm4, %xmm8
+; AVX512BW-NEXT:    vpshufb %xmm22, %xmm2, %xmm3
+; AVX512BW-NEXT:    vpshufb %xmm22, %xmm4, %xmm8
 ; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3],xmm8[4],xmm3[4],xmm8[5],xmm3[5],xmm8[6],xmm3[6],xmm8[7],xmm3[7]
 ; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
 ; AVX512BW-NEXT:    vprold $16, %xmm2, %xmm2
@@ -6606,11 +6600,10 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vpmovzxbw {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
 ; AVX512BW-NEXT:    vinserti32x4 $2, %xmm3, %zmm2, %zmm2
 ; AVX512BW-NEXT:    vpermw %zmm2, %zmm10, %zmm1 {%k1}
-; AVX512BW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm6[2,1,2,3]
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpermw %zmm2, %zmm10, %zmm2
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-NEXT:    vpshufd {{.*#+}} xmm3 = xmm6[2,1,2,3]
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-NEXT:    vpermt2w %zmm3, %zmm14, %zmm2
 ; AVX512BW-NEXT:    vmovdqu8 %zmm2, %zmm1 {%k3}
 ; AVX512BW-NEXT:    vmovdqa64 %zmm1, (%rax)
 ; AVX512BW-NEXT:    vmovdqa64 %zmm9, 192(%rax)
@@ -6623,8 +6616,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ;
 ; AVX512BW-FCP-LABEL: store_i8_stride6_vf64:
 ; AVX512BW-FCP:       # %bb.0:
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm8
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm10
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm11
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm13
 ; AVX512BW-FCP-NEXT:    vmovdqa 32(%rsi), %ymm3
 ; AVX512BW-FCP-NEXT:    vpbroadcastq {{.*#+}} ymm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
 ; AVX512BW-FCP-NEXT:    vpshufb %ymm5, %ymm3, %ymm0
@@ -6633,192 +6626,191 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
 ; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm1 = ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15],ymm4[24],ymm3[24],ymm4[25],ymm3[25],ymm4[26],ymm3[26],ymm4[27],ymm3[27],ymm4[28],ymm3[28],ymm4[29],ymm3[29],ymm4[30],ymm3[30],ymm4[31],ymm3[31]
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm9 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512BW-FCP-NEXT:    vpermw %ymm1, %ymm9, %ymm1
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm12 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512BW-FCP-NEXT:    vpermw %ymm1, %ymm12, %ymm1
 ; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rcx), %ymm6
-; AVX512BW-FCP-NEXT:    vpbroadcastq {{.*#+}} ymm17 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm17, %ymm6, %ymm0
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdx), %ymm7
-; AVX512BW-FCP-NEXT:    vpshufb %ymm17, %ymm7, %ymm2
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%rcx), %ymm8
+; AVX512BW-FCP-NEXT:    vpbroadcastq {{.*#+}} ymm6 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm8, %ymm0
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdx), %ymm10
+; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm10, %ymm2
 ; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[16],ymm0[16],ymm2[17],ymm0[17],ymm2[18],ymm0[18],ymm2[19],ymm0[19],ymm2[20],ymm0[20],ymm2[21],ymm0[21],ymm2[22],ymm0[22],ymm2[23],ymm0[23]
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm2 = ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15],ymm7[24],ymm6[24],ymm7[25],ymm6[25],ymm7[26],ymm6[26],ymm7[27],ymm6[27],ymm7[28],ymm6[28],ymm7[29],ymm6[29],ymm7[30],ymm6[30],ymm7[31],ymm6[31]
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm12 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
-; AVX512BW-FCP-NEXT:    vpermw %ymm2, %ymm12, %ymm2
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm2 = ymm10[8],ymm8[8],ymm10[9],ymm8[9],ymm10[10],ymm8[10],ymm10[11],ymm8[11],ymm10[12],ymm8[12],ymm10[13],ymm8[13],ymm10[14],ymm8[14],ymm10[15],ymm8[15],ymm10[24],ymm8[24],ymm10[25],ymm8[25],ymm10[26],ymm8[26],ymm10[27],ymm8[27],ymm10[28],ymm8[28],ymm10[29],ymm8[29],ymm10[30],ymm8[30],ymm10[31],ymm8[31]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm14 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
+; AVX512BW-FCP-NEXT:    vpermw %ymm2, %ymm14, %ymm2
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
 ; AVX512BW-FCP-NEXT:    movl $613566756, %eax # imm = 0x24924924
 ; AVX512BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm1, %zmm0 {%k1}
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%r8), %ymm2
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm8[4,5,6,7]
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm1, %zmm2 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqa64 32(%r8), %ymm17
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm17[0,1,2,3],zmm11[4,5,6,7]
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
 ; AVX512BW-FCP-NEXT:    movl $-1840700270, %eax # imm = 0x92492492
 ; AVX512BW-FCP-NEXT:    kmovd %eax, %k2
-; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm1, %zmm0 {%k2}
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%r9), %ymm1
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm11 = zmm1[0,1,2,3],zmm10[4,5,6,7]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm11 = zmm11[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm11 = zmm11[2,2,2,3,6,6,6,7]
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm1, %zmm2 {%k2}
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%r9), %ymm0
+; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm7 = zmm0[0,1,2,3],zmm13[4,5,6,7]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm7 = zmm7[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm7 = zmm7[2,2,2,3,6,6,6,7]
 ; AVX512BW-FCP-NEXT:    movabsq $-9076969306111049208, %rax # imm = 0x8208208208208208
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm11, %zmm0 {%k3}
-; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %ymm11
-; AVX512BW-FCP-NEXT:    vpshufb %ymm5, %ymm11, %ymm13
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm14
-; AVX512BW-FCP-NEXT:    vpshufb %ymm5, %ymm14, %ymm5
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm13[0],ymm5[1],ymm13[1],ymm5[2],ymm13[2],ymm5[3],ymm13[3],ymm5[4],ymm13[4],ymm5[5],ymm13[5],ymm5[6],ymm13[6],ymm5[7],ymm13[7],ymm5[16],ymm13[16],ymm5[17],ymm13[17],ymm5[18],ymm13[18],ymm5[19],ymm13[19],ymm5[20],ymm13[20],ymm5[21],ymm13[21],ymm5[22],ymm13[22],ymm5[23],ymm13[23]
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm7, %zmm2 {%k3}
+; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %ymm7
+; AVX512BW-FCP-NEXT:    vpshufb %ymm5, %ymm7, %ymm15
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm9
+; AVX512BW-FCP-NEXT:    vpshufb %ymm5, %ymm9, %ymm5
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm15[0],ymm5[1],ymm15[1],ymm5[2],ymm15[2],ymm5[3],ymm15[3],ymm5[4],ymm15[4],ymm5[5],ymm15[5],ymm5[6],ymm15[6],ymm5[7],ymm15[7],ymm5[16],ymm15[16],ymm5[17],ymm15[17],ymm5[18],ymm15[18],ymm5[19],ymm15[19],ymm5[20],ymm15[20],ymm5[21],ymm15[21],ymm5[22],ymm15[22],ymm5[23],ymm15[23]
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm13 = ymm14[8],ymm11[8],ymm14[9],ymm11[9],ymm14[10],ymm11[10],ymm14[11],ymm11[11],ymm14[12],ymm11[12],ymm14[13],ymm11[13],ymm14[14],ymm11[14],ymm14[15],ymm11[15],ymm14[24],ymm11[24],ymm14[25],ymm11[25],ymm14[26],ymm11[26],ymm14[27],ymm11[27],ymm14[28],ymm11[28],ymm14[29],ymm11[29],ymm14[30],ymm11[30],ymm14[31],ymm11[31]
-; AVX512BW-FCP-NEXT:    vpermw %ymm13, %ymm9, %ymm9
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm5, %zmm9
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rcx), %ymm19
-; AVX512BW-FCP-NEXT:    vpshufb %ymm17, %ymm19, %ymm5
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdx), %ymm20
-; AVX512BW-FCP-NEXT:    vpshufb %ymm17, %ymm20, %ymm13
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm13[0],ymm5[0],ymm13[1],ymm5[1],ymm13[2],ymm5[2],ymm13[3],ymm5[3],ymm13[4],ymm5[4],ymm13[5],ymm5[5],ymm13[6],ymm5[6],ymm13[7],ymm5[7],ymm13[16],ymm5[16],ymm13[17],ymm5[17],ymm13[18],ymm5[18],ymm13[19],ymm5[19],ymm13[20],ymm5[20],ymm13[21],ymm5[21],ymm13[22],ymm5[22],ymm13[23],ymm5[23]
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm15 = ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15],ymm9[24],ymm7[24],ymm9[25],ymm7[25],ymm9[26],ymm7[26],ymm9[27],ymm7[27],ymm9[28],ymm7[28],ymm9[29],ymm7[29],ymm9[30],ymm7[30],ymm9[31],ymm7[31]
+; AVX512BW-FCP-NEXT:    vpermw %ymm15, %ymm12, %ymm12
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm5, %zmm12
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%rcx), %ymm23
+; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm23, %ymm5
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdx), %ymm24
+; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm24, %ymm15
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm15[0],ymm5[0],ymm15[1],ymm5[1],ymm15[2],ymm5[2],ymm15[3],ymm5[3],ymm15[4],ymm5[4],ymm15[5],ymm5[5],ymm15[6],ymm5[6],ymm15[7],ymm5[7],ymm15[16],ymm5[16],ymm15[17],ymm5[17],ymm15[18],ymm5[18],ymm15[19],ymm5[19],ymm15[20],ymm5[20],ymm15[21],ymm5[21],ymm15[22],ymm5[22],ymm15[23],ymm5[23]
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm13 = ymm20[8],ymm19[8],ymm20[9],ymm19[9],ymm20[10],ymm19[10],ymm20[11],ymm19[11],ymm20[12],ymm19[12],ymm20[13],ymm19[13],ymm20[14],ymm19[14],ymm20[15],ymm19[15],ymm20[24],ymm19[24],ymm20[25],ymm19[25],ymm20[26],ymm19[26],ymm20[27],ymm19[27],ymm20[28],ymm19[28],ymm20[29],ymm19[29],ymm20[30],ymm19[30],ymm20[31],ymm19[31]
-; AVX512BW-FCP-NEXT:    vpermw %ymm13, %ymm12, %ymm12
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm5, %zmm5
-; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm9, %zmm5 {%k1}
-; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %ymm9
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm8, %zmm8
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm8 = zmm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7]
-; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm8, %zmm5 {%k2}
-; AVX512BW-FCP-NEXT:    vmovdqa (%r9), %ymm8
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm10, %zmm10
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm10 = zmm10[2,2,2,3,6,6,6,7]
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm10, %zmm5 {%k3}
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rcx), %xmm22
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rcx), %xmm13
-; AVX512BW-FCP-NEXT:    vpshufb %xmm17, %xmm13, %xmm10
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdx), %xmm23
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdx), %xmm15
-; AVX512BW-FCP-NEXT:    vpshufb %xmm17, %xmm15, %xmm12
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm12 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3],xmm15[4],xmm13[4],xmm15[5],xmm13[5],xmm15[6],xmm13[6],xmm15[7],xmm13[7]
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm24 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
-; AVX512BW-FCP-NEXT:    vpermw %ymm12, %ymm24, %ymm12
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm12, %zmm12
-; AVX512BW-FCP-NEXT:    vmovdqa64 32(%rsi), %xmm18
-; AVX512BW-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm25 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm25, %xmm18, %xmm10
-; AVX512BW-FCP-NEXT:    vmovdqa64 32(%rdi), %xmm21
-; AVX512BW-FCP-NEXT:    vpshufb %xmm25, %xmm21, %xmm16
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm16[8],xmm10[8],xmm16[9],xmm10[9],xmm16[10],xmm10[10],xmm16[11],xmm10[11],xmm16[12],xmm10[12],xmm16[13],xmm10[13],xmm16[14],xmm10[14],xmm16[15],xmm10[15]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm16 = xmm21[0],xmm18[0],xmm21[1],xmm18[1],xmm21[2],xmm18[2],xmm21[3],xmm18[3],xmm21[4],xmm18[4],xmm21[5],xmm18[5],xmm21[6],xmm18[6],xmm21[7],xmm18[7]
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm26 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
-; AVX512BW-FCP-NEXT:    vpermw %ymm16, %ymm26, %ymm16
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm16, %zmm10
-; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm12, %zmm10 {%k2}
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%r8), %xmm12
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm15 = ymm24[8],ymm23[8],ymm24[9],ymm23[9],ymm24[10],ymm23[10],ymm24[11],ymm23[11],ymm24[12],ymm23[12],ymm24[13],ymm23[13],ymm24[14],ymm23[14],ymm24[15],ymm23[15],ymm24[24],ymm23[24],ymm24[25],ymm23[25],ymm24[26],ymm23[26],ymm24[27],ymm23[27],ymm24[28],ymm23[28],ymm24[29],ymm23[29],ymm24[30],ymm23[30],ymm24[31],ymm23[31]
+; AVX512BW-FCP-NEXT:    vpermw %ymm15, %ymm14, %ymm14
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm14, %zmm5, %zmm5
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm12, %zmm5 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %ymm12
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm11, %zmm11
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm11 = zmm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm11 = zmm11[2,2,2,3,6,6,6,7]
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm11, %zmm5 {%k2}
+; AVX512BW-FCP-NEXT:    vmovdqa (%r9), %ymm11
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm13, %zmm13
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm13 = zmm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm13 = zmm13[2,2,2,3,6,6,6,7]
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm13, %zmm5 {%k3}
+; AVX512BW-FCP-NEXT:    vmovdqa64 32(%rcx), %xmm20
+; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm20, %xmm13
+; AVX512BW-FCP-NEXT:    vmovdqa64 32(%rdx), %xmm21
+; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm21, %xmm14
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm13 = 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:    vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1]
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm14 = xmm21[0],xmm20[0],xmm21[1],xmm20[1],xmm21[2],xmm20[2],xmm21[3],xmm20[3],xmm21[4],xmm20[4],xmm21[5],xmm20[5],xmm21[6],xmm20[6],xmm21[7],xmm20[7]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm0 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
+; AVX512BW-FCP-NEXT:    vpermw %ymm14, %ymm0, %ymm14
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm14, %zmm22
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%rsi), %xmm15
+; AVX512BW-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm19 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm19, %xmm15, %xmm30
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%r8), %xmm18
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%r8), %xmm13
 ; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} xmm27 = [8,9,0,0,0,5,6,7]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm27, %xmm12, %xmm16
-; AVX512BW-FCP-NEXT:    vpmovzxbw {{.*#+}} xmm28 = xmm12[0],zero,xmm12[1],zero,xmm12[2],zero,xmm12[3],zero,xmm12[4],zero,xmm12[5],zero,xmm12[6],zero,xmm12[7],zero
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm16, %zmm28, %zmm16
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm28 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,22,21,16,23,22,21,16,23,22,21,16,23,17,17,17,17]
-; AVX512BW-FCP-NEXT:    vpermw %zmm16, %zmm28, %zmm10 {%k1}
-; AVX512BW-FCP-NEXT:    vmovdqa64 32(%r9), %xmm16
+; AVX512BW-FCP-NEXT:    vpshufb %xmm27, %xmm13, %xmm14
+; AVX512BW-FCP-NEXT:    vpmovzxbw {{.*#+}} xmm25 = xmm13[0],zero,xmm13[1],zero,xmm13[2],zero,xmm13[3],zero,xmm13[4],zero,xmm13[5],zero,xmm13[6],zero,xmm13[7],zero
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm28 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
+; AVX512BW-FCP-NEXT:    vpermt2w %zmm14, %zmm28, %zmm25
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%r9), %xmm16
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%r9), %xmm14
 ; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm29 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm29, %xmm16, %xmm30
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm31 = xmm16[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm30, %zmm31, %zmm30
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rsi), %xmm31
-; AVX512BW-FCP-NEXT:    vpermw %zmm30, %zmm28, %zmm30
+; AVX512BW-FCP-NEXT:    vpshufb %xmm29, %xmm14, %xmm31
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm26 = xmm14[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-FCP-NEXT:    vpermt2w %zmm31, %zmm28, %zmm26
+; AVX512BW-FCP-NEXT:    vpshufb %xmm27, %xmm18, %xmm31
+; AVX512BW-FCP-NEXT:    vpmovzxbw {{.*#+}} xmm27 = xmm18[0],zero,xmm18[1],zero,xmm18[2],zero,xmm18[3],zero,xmm18[4],zero,xmm18[5],zero,xmm18[6],zero,xmm18[7],zero
+; AVX512BW-FCP-NEXT:    vpermt2w %zmm31, %zmm28, %zmm27
+; AVX512BW-FCP-NEXT:    vpshufb %xmm29, %xmm16, %xmm31
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm29 = xmm16[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-FCP-NEXT:    vpermt2w %zmm31, %zmm28, %zmm29
+; AVX512BW-FCP-NEXT:    vmovdqa64 32(%rdi), %xmm28
+; AVX512BW-FCP-NEXT:    vpshufb %xmm19, %xmm28, %xmm31
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm30 = xmm31[8],xmm30[8],xmm31[9],xmm30[9],xmm31[10],xmm30[10],xmm31[11],xmm30[11],xmm31[12],xmm30[12],xmm31[13],xmm30[13],xmm31[14],xmm30[14],xmm31[15],xmm30[15]
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%rcx), %xmm31
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm24 = ymm24[0],ymm23[0],ymm24[1],ymm23[1],ymm24[2],ymm23[2],ymm24[3],ymm23[3],ymm24[4],ymm23[4],ymm24[5],ymm23[5],ymm24[6],ymm23[6],ymm24[7],ymm23[7],ymm24[16],ymm23[16],ymm24[17],ymm23[17],ymm24[18],ymm23[18],ymm24[19],ymm23[19],ymm24[20],ymm23[20],ymm24[21],ymm23[21],ymm24[22],ymm23[22],ymm24[23],ymm23[23]
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm0
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm23 = xmm0[8],xmm31[8],xmm0[9],xmm31[9],xmm0[10],xmm31[10],xmm0[11],xmm31[11],xmm0[12],xmm31[12],xmm0[13],xmm31[13],xmm0[14],xmm31[14],xmm0[15],xmm31[15]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
+; AVX512BW-FCP-NEXT:    vpermt2w %zmm24, %zmm1, %zmm23
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm8 = ymm10[0],ymm8[0],ymm10[1],ymm8[1],ymm10[2],ymm8[2],ymm10[3],ymm8[3],ymm10[4],ymm8[4],ymm10[5],ymm8[5],ymm10[6],ymm8[6],ymm10[7],ymm8[7],ymm10[16],ymm8[16],ymm10[17],ymm8[17],ymm10[18],ymm8[18],ymm10[19],ymm8[19],ymm10[20],ymm8[20],ymm10[21],ymm8[21],ymm10[22],ymm8[22],ymm10[23],ymm8[23]
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm21[8],xmm20[8],xmm21[9],xmm20[9],xmm21[10],xmm20[10],xmm21[11],xmm20[11],xmm21[12],xmm20[12],xmm21[13],xmm20[13],xmm21[14],xmm20[14],xmm21[15],xmm20[15]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm20 = ymm30[0,0,0,1]
+; AVX512BW-FCP-NEXT:    vpermt2w %zmm8, %zmm1, %zmm10
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm28[0],xmm15[0],xmm28[1],xmm15[1],xmm28[2],xmm15[2],xmm28[3],xmm15[3],xmm28[4],xmm15[4],xmm28[5],xmm15[5],xmm28[6],xmm15[6],xmm28[7],xmm15[7]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm21 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
+; AVX512BW-FCP-NEXT:    vpermw %ymm1, %ymm21, %ymm1
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm20, %zmm1, %zmm8
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm22, %zmm8 {%k2}
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm25, %zmm8 {%k1}
 ; AVX512BW-FCP-NEXT:    movabsq $585610922974906400, %rax # imm = 0x820820820820820
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm30, %zmm10 {%k3}
-; AVX512BW-FCP-NEXT:    vpshufb %xmm17, %xmm22, %xmm30
-; AVX512BW-FCP-NEXT:    vpshufb %xmm17, %xmm23, %xmm17
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm17 = xmm17[0],xmm30[0],xmm17[1],xmm30[1],xmm17[2],xmm30[2],xmm17[3],xmm30[3],xmm17[4],xmm30[4],xmm17[5],xmm30[5],xmm17[6],xmm30[6],xmm17[7],xmm30[7]
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm30 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7]
-; AVX512BW-FCP-NEXT:    vpermw %ymm30, %ymm24, %ymm24
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %xmm30
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[0,0,0,1]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm24, %zmm24
-; AVX512BW-FCP-NEXT:    vpshufb %xmm25, %xmm31, %xmm17
-; AVX512BW-FCP-NEXT:    vpshufb %xmm25, %xmm30, %xmm25
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm17 = xmm25[8],xmm17[8],xmm25[9],xmm17[9],xmm25[10],xmm17[10],xmm25[11],xmm17[11],xmm25[12],xmm17[12],xmm25[13],xmm17[13],xmm25[14],xmm17[14],xmm25[15],xmm17[15]
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm25 = xmm30[0],xmm31[0],xmm30[1],xmm31[1],xmm30[2],xmm31[2],xmm30[3],xmm31[3],xmm30[4],xmm31[4],xmm30[5],xmm31[5],xmm30[6],xmm31[6],xmm30[7],xmm31[7]
-; AVX512BW-FCP-NEXT:    vpermw %ymm25, %ymm26, %ymm25
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[0,0,0,1]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm25, %zmm17
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%r8), %xmm25
-; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm24, %zmm17 {%k2}
-; AVX512BW-FCP-NEXT:    vpshufb %xmm27, %xmm25, %xmm24
-; AVX512BW-FCP-NEXT:    vpmovzxbw {{.*#+}} xmm26 = xmm25[0],zero,xmm25[1],zero,xmm25[2],zero,xmm25[3],zero,xmm25[4],zero,xmm25[5],zero,xmm25[6],zero,xmm25[7],zero
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm24, %zmm26, %zmm26
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%r9), %xmm24
-; AVX512BW-FCP-NEXT:    vpermw %zmm26, %zmm28, %zmm17 {%k1}
-; AVX512BW-FCP-NEXT:    vpshufb %xmm29, %xmm24, %xmm26
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm27 = xmm24[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm26, %zmm27, %zmm26
-; AVX512BW-FCP-NEXT:    vpermw %zmm26, %zmm28, %zmm26
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm26, %zmm17 {%k3}
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm19 = ymm20[0],ymm19[0],ymm20[1],ymm19[1],ymm20[2],ymm19[2],ymm20[3],ymm19[3],ymm20[4],ymm19[4],ymm20[5],ymm19[5],ymm20[6],ymm19[6],ymm20[7],ymm19[7],ymm20[16],ymm19[16],ymm20[17],ymm19[17],ymm20[18],ymm19[18],ymm20[19],ymm19[19],ymm20[20],ymm19[20],ymm20[21],ymm19[21],ymm20[22],ymm19[22],ymm20[23],ymm19[23]
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm20, %zmm19
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm11 = ymm14[0],ymm11[0],ymm14[1],ymm11[1],ymm14[2],ymm11[2],ymm14[3],ymm11[3],ymm14[4],ymm11[4],ymm14[5],ymm11[5],ymm14[6],ymm11[6],ymm14[7],ymm11[7],ymm14[16],ymm11[16],ymm14[17],ymm11[17],ymm14[18],ymm11[18],ymm14[19],ymm11[19],ymm14[20],ymm11[20],ymm14[21],ymm11[21],ymm14[22],ymm11[22],ymm14[23],ymm11[23]
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm14 = xmm30[8],xmm31[8],xmm30[9],xmm31[9],xmm30[10],xmm31[10],xmm30[11],xmm31[11],xmm30[12],xmm31[12],xmm30[13],xmm31[13],xmm30[14],xmm31[14],xmm30[15],xmm31[15]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm14, %zmm11
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm14 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
-; AVX512BW-FCP-NEXT:    vpermw %zmm11, %zmm14, %zmm11
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,25,24,27,26,25,24,27,26,25,24,27,26,29,28,31,30]
-; AVX512BW-FCP-NEXT:    vpermw %zmm19, %zmm20, %zmm11 {%k1}
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[16],ymm6[16],ymm7[17],ymm6[17],ymm7[18],ymm6[18],ymm7[19],ymm6[19],ymm7[20],ymm6[20],ymm7[21],ymm6[21],ymm7[22],ymm6[22],ymm7[23],ymm6[23]
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm15[8],xmm13[8],xmm15[9],xmm13[9],xmm15[10],xmm13[10],xmm15[11],xmm13[11],xmm15[12],xmm13[12],xmm15[13],xmm13[13],xmm15[14],xmm13[14],xmm15[15],xmm13[15]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[16],ymm3[16],ymm4[17],ymm3[17],ymm4[18],ymm3[18],ymm4[19],ymm3[19],ymm4[20],ymm3[20],ymm4[21],ymm3[21],ymm4[22],ymm3[22],ymm4[23],ymm3[23]
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm21[8],xmm18[8],xmm21[9],xmm18[9],xmm21[10],xmm18[10],xmm21[11],xmm18[11],xmm21[12],xmm18[12],xmm21[13],xmm18[13],xmm21[14],xmm18[14],xmm21[15],xmm18[15]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512BW-FCP-NEXT:    vpermw %zmm3, %zmm14, %zmm3
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm25[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FCP-NEXT:    vpermw %zmm6, %zmm20, %zmm3 {%k1}
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm6 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
-; AVX512BW-FCP-NEXT:    vpermw %ymm4, %ymm6, %ymm4
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm26, %zmm8 {%k3}
+; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm31, %xmm1
+; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm6
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3],xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm31[0],xmm0[1],xmm31[1],xmm0[2],xmm31[2],xmm0[3],xmm31[3],xmm0[4],xmm31[4],xmm0[5],xmm31[5],xmm0[6],xmm31[6],xmm0[7],xmm31[7]
+; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %xmm6
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm20 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
+; AVX512BW-FCP-NEXT:    vpermw %ymm0, %ymm20, %ymm0
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %xmm20
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512BW-FCP-NEXT:    vpshufb %xmm19, %xmm6, %xmm1
+; AVX512BW-FCP-NEXT:    vpshufb %xmm19, %xmm20, %xmm19
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm19[8],xmm1[8],xmm19[9],xmm1[9],xmm19[10],xmm1[10],xmm19[11],xmm1[11],xmm19[12],xmm1[12],xmm19[13],xmm1[13],xmm19[14],xmm1[14],xmm19[15],xmm1[15]
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm19 = xmm20[0],xmm6[0],xmm20[1],xmm6[1],xmm20[2],xmm6[2],xmm20[3],xmm6[3],xmm20[4],xmm6[4],xmm20[5],xmm6[5],xmm20[6],xmm6[6],xmm20[7],xmm6[7]
+; AVX512BW-FCP-NEXT:    vpermw %ymm19, %ymm21, %ymm19
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm19, %zmm1
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm0, %zmm1 {%k2}
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm27, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm29, %zmm1 {%k3}
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[16],ymm7[16],ymm9[17],ymm7[17],ymm9[18],ymm7[18],ymm9[19],ymm7[19],ymm9[20],ymm7[20],ymm9[21],ymm7[21],ymm9[22],ymm7[22],ymm9[23],ymm7[23]
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm20[8],xmm6[8],xmm20[9],xmm6[9],xmm20[10],xmm6[10],xmm20[11],xmm6[11],xmm20[12],xmm6[12],xmm20[13],xmm6[13],xmm20[14],xmm6[14],xmm20[15],xmm6[15]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm7 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512BW-FCP-NEXT:    vpermt2w %zmm0, %zmm7, %zmm6
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm23, %zmm6 {%k1}
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[16],ymm3[16],ymm4[17],ymm3[17],ymm4[18],ymm3[18],ymm4[19],ymm3[19],ymm4[20],ymm3[20],ymm4[21],ymm3[21],ymm4[22],ymm3[22],ymm4[23],ymm3[23]
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm28[8],xmm15[8],xmm28[9],xmm15[9],xmm28[10],xmm15[10],xmm28[11],xmm15[11],xmm28[12],xmm15[12],xmm28[13],xmm15[13],xmm28[14],xmm15[14],xmm28[15],xmm15[15]
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm18[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FCP-NEXT:    vpermt2w %zmm0, %zmm7, %zmm3
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm0 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
+; AVX512BW-FCP-NEXT:    vpermw %ymm4, %ymm0, %ymm4
 ; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
 ; AVX512BW-FCP-NEXT:    # ymm7 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm7, %ymm9, %ymm9
+; AVX512BW-FCP-NEXT:    vpshufb %ymm7, %ymm12, %ymm9
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
 ; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm4, %zmm4
 ; AVX512BW-FCP-NEXT:    movl $1227133513, %eax # imm = 0x49249249
-; AVX512BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm4, %zmm11 {%k1}
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm24[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FCP-NEXT:    vpermw %ymm4, %ymm6, %ymm4
+; AVX512BW-FCP-NEXT:    kmovd %eax, %k2
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm4, %zmm6 {%k2}
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm16[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FCP-NEXT:    vpermw %ymm4, %ymm0, %ymm4
 ; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm9 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
 ; AVX512BW-FCP-NEXT:    # ymm9 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm8, %ymm8
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm4, %zmm4
+; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm11, %ymm11
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm4, %zmm4
 ; AVX512BW-FCP-NEXT:    movabsq $2342443691899625602, %rax # imm = 0x2082082082082082
-; AVX512BW-FCP-NEXT:    kmovq %rax, %k2
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm11 {%k2}
-; AVX512BW-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm12[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FCP-NEXT:    vpermw %ymm4, %ymm6, %ymm4
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm2
-; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm2, %zmm3 {%k1}
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm16[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FCP-NEXT:    vpermw %ymm2, %ymm6, %ymm2
-; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm1, %ymm1
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k2}
+; AVX512BW-FCP-NEXT:    kmovq %rax, %k3
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm6 {%k3}
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm10, %zmm3 {%k1}
+; AVX512BW-FCP-NEXT:    vpshufb %ymm7, %ymm17, %ymm4
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm13[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FCP-NEXT:    vpermw %ymm7, %ymm0, %ymm7
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm7, %zmm4
+; AVX512BW-FCP-NEXT:    vmovdqu16 %zmm4, %zmm3 {%k2}
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm14[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FCP-NEXT:    vpermw %ymm4, %ymm0, %ymm0
+; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm4, %ymm4
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm3 {%k3}
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, 256(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, 64(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, 192(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, 192(%rax)
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, 128(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, 320(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, 320(%rax)
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
 ;
@@ -6829,22 +6821,21 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %zmm12
 ; AVX512DQ-BW-NEXT:    vmovdqa64 (%rsi), %ymm16
 ; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %ymm17
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm17[0],ymm16[0],ymm17[1],ymm16[1],ymm17[2],ymm16[2],ymm17[3],ymm16[3],ymm17[4],ymm16[4],ymm17[5],ymm16[5],ymm17[6],ymm16[6],ymm17[7],ymm16[7],ymm17[16],ymm16[16],ymm17[17],ymm16[17],ymm17[18],ymm16[18],ymm17[19],ymm16[19],ymm17[20],ymm16[20],ymm17[21],ymm16[21],ymm17[22],ymm16[22],ymm17[23],ymm16[23]
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm2 = ymm17[0],ymm16[0],ymm17[1],ymm16[1],ymm17[2],ymm16[2],ymm17[3],ymm16[3],ymm17[4],ymm16[4],ymm17[5],ymm16[5],ymm17[6],ymm16[6],ymm17[7],ymm16[7],ymm17[16],ymm16[16],ymm17[17],ymm16[17],ymm17[18],ymm16[18],ymm17[19],ymm16[19],ymm17[20],ymm16[20],ymm17[21],ymm16[21],ymm17[22],ymm16[22],ymm17[23],ymm16[23]
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %xmm1
 ; AVX512DQ-BW-NEXT:    vmovdqa 32(%rsi), %xmm9
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm3
 ; AVX512DQ-BW-NEXT:    vmovdqa 32(%rdi), %xmm11
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm2 = 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:    vinserti64x4 $1, %ymm0, %zmm2, %zmm0
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm7 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
-; AVX512DQ-BW-NEXT:    vpermw %zmm0, %zmm7, %zmm0
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = 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:    vpmovsxbw {{.*#+}} zmm20 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512DQ-BW-NEXT:    vpermt2w %zmm2, %zmm20, %zmm0
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rcx), %xmm2
 ; AVX512DQ-BW-NEXT:    vmovdqa 32(%rcx), %xmm8
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm4
 ; AVX512DQ-BW-NEXT:    vmovdqa 32(%rdx), %xmm10
 ; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15]
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
-; AVX512DQ-BW-NEXT:    vpermw %ymm5, %ymm20, %ymm5
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm22 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
+; AVX512DQ-BW-NEXT:    vpermw %ymm5, %ymm22, %ymm5
 ; AVX512DQ-BW-NEXT:    vmovdqa64 (%rcx), %ymm18
 ; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdx), %ymm19
 ; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm6 = ymm19[0],ymm18[0],ymm19[1],ymm18[1],ymm19[2],ymm18[2],ymm19[3],ymm18[3],ymm19[4],ymm18[4],ymm19[5],ymm18[5],ymm19[6],ymm18[6],ymm19[7],ymm18[7],ymm19[16],ymm18[16],ymm19[17],ymm18[17],ymm19[18],ymm18[18],ymm19[19],ymm18[19],ymm19[20],ymm18[20],ymm19[21],ymm18[21],ymm19[22],ymm18[22],ymm19[23],ymm18[23]
@@ -6857,79 +6848,78 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm5
 ; AVX512DQ-BW-NEXT:    vmovdqa 32(%r8), %xmm13
 ; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm23 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
-; AVX512DQ-BW-NEXT:    vpermw %ymm6, %ymm23, %ymm6
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm24 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
+; AVX512DQ-BW-NEXT:    vpermw %ymm6, %ymm24, %ymm6
 ; AVX512DQ-BW-NEXT:    vmovdqa64 (%r8), %ymm21
-; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} ymm24 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
-; AVX512DQ-BW-NEXT:    # ymm24 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm24, %ymm21, %ymm15
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm15, %zmm6, %zmm6
+; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} ymm25 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
+; AVX512DQ-BW-NEXT:    # ymm25 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm25, %ymm21, %ymm7
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm6, %zmm6
 ; AVX512DQ-BW-NEXT:    movl $1227133513, %r10d # imm = 0x49249249
 ; AVX512DQ-BW-NEXT:    kmovd %r10d, %k2
 ; AVX512DQ-BW-NEXT:    vmovdqu16 %zmm6, %zmm0 {%k2}
 ; AVX512DQ-BW-NEXT:    vmovdqa (%r9), %xmm6
 ; AVX512DQ-BW-NEXT:    vmovdqa 32(%r9), %xmm15
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm22 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-NEXT:    vpermw %ymm22, %ymm23, %ymm25
-; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %ymm22
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-NEXT:    vpermw %ymm7, %ymm24, %ymm7
+; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %ymm23
 ; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} ymm26 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
 ; AVX512DQ-BW-NEXT:    # ymm26 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm22, %ymm27
+; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm23, %ymm27
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm27 = ymm27[2,2,2,3]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm27, %zmm25, %zmm25
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm27, %zmm7, %zmm7
 ; AVX512DQ-BW-NEXT:    movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082
 ; AVX512DQ-BW-NEXT:    kmovq %r10, %k3
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm25, %zmm0 {%k3}
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rsi), %ymm25
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rdi), %ymm27
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm28 = ymm27[0],ymm25[0],ymm27[1],ymm25[1],ymm27[2],ymm25[2],ymm27[3],ymm25[3],ymm27[4],ymm25[4],ymm27[5],ymm25[5],ymm27[6],ymm25[6],ymm27[7],ymm25[7],ymm27[16],ymm25[16],ymm27[17],ymm25[17],ymm27[18],ymm25[18],ymm27[19],ymm25[19],ymm27[20],ymm25[20],ymm27[21],ymm25[21],ymm27[22],ymm25[22],ymm27[23],ymm25[23]
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm29 = xmm11[8],xmm9[8],xmm11[9],xmm9[9],xmm11[10],xmm9[10],xmm11[11],xmm9[11],xmm11[12],xmm9[12],xmm11[13],xmm9[13],xmm11[14],xmm9[14],xmm11[15],xmm9[15]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm28, %zmm29, %zmm28
-; AVX512DQ-BW-NEXT:    vpermw %zmm28, %zmm7, %zmm7
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm28 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
-; AVX512DQ-BW-NEXT:    vpermw %ymm28, %ymm20, %ymm20
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rcx), %ymm28
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rdx), %ymm29
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm30 = ymm29[0],ymm28[0],ymm29[1],ymm28[1],ymm29[2],ymm28[2],ymm29[3],ymm28[3],ymm29[4],ymm28[4],ymm29[5],ymm28[5],ymm29[6],ymm28[6],ymm29[7],ymm28[7],ymm29[16],ymm28[16],ymm29[17],ymm28[17],ymm29[18],ymm28[18],ymm29[19],ymm28[19],ymm29[20],ymm28[20],ymm29[21],ymm28[21],ymm29[22],ymm28[22],ymm29[23],ymm28[23]
-; AVX512DQ-BW-NEXT:    vprold $16, %ymm30, %ymm30
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm30 = ymm30[2,2,2,3]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm30, %zmm20, %zmm20
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm7, %zmm0 {%k3}
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rsi), %ymm27
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rdi), %ymm28
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm29 = ymm28[0],ymm27[0],ymm28[1],ymm27[1],ymm28[2],ymm27[2],ymm28[3],ymm27[3],ymm28[4],ymm27[4],ymm28[5],ymm27[5],ymm28[6],ymm27[6],ymm28[7],ymm27[7],ymm28[16],ymm27[16],ymm28[17],ymm27[17],ymm28[18],ymm27[18],ymm28[19],ymm27[19],ymm28[20],ymm27[20],ymm28[21],ymm27[21],ymm28[22],ymm27[22],ymm28[23],ymm27[23]
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm11[8],xmm9[8],xmm11[9],xmm9[9],xmm11[10],xmm9[10],xmm11[11],xmm9[11],xmm11[12],xmm9[12],xmm11[13],xmm9[13],xmm11[14],xmm9[14],xmm11[15],xmm9[15]
+; AVX512DQ-BW-NEXT:    vpermt2w %zmm29, %zmm20, %zmm7
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
+; AVX512DQ-BW-NEXT:    vpermw %ymm20, %ymm22, %ymm20
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rcx), %ymm29
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rdx), %ymm30
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm22 = ymm30[0],ymm29[0],ymm30[1],ymm29[1],ymm30[2],ymm29[2],ymm30[3],ymm29[3],ymm30[4],ymm29[4],ymm30[5],ymm29[5],ymm30[6],ymm29[6],ymm30[7],ymm29[7],ymm30[16],ymm29[16],ymm30[17],ymm29[17],ymm30[18],ymm29[18],ymm30[19],ymm29[19],ymm30[20],ymm29[20],ymm30[21],ymm29[21],ymm30[22],ymm29[22],ymm30[23],ymm29[23]
+; AVX512DQ-BW-NEXT:    vprold $16, %ymm22, %ymm22
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm20, %zmm20
 ; AVX512DQ-BW-NEXT:    vmovdqu16 %zmm20, %zmm7 {%k1}
 ; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm13[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-NEXT:    vpermw %ymm20, %ymm23, %ymm20
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%r8), %ymm30
-; AVX512DQ-BW-NEXT:    vpshufb %ymm24, %ymm30, %ymm24
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm24, %zmm20, %zmm20
+; AVX512DQ-BW-NEXT:    vpermw %ymm20, %ymm24, %ymm20
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%r8), %ymm31
+; AVX512DQ-BW-NEXT:    vpshufb %ymm25, %ymm31, %ymm22
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm20, %zmm20
 ; AVX512DQ-BW-NEXT:    vmovdqu16 %zmm20, %zmm7 {%k2}
 ; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm15[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-NEXT:    vpermw %ymm20, %ymm23, %ymm20
+; AVX512DQ-BW-NEXT:    vpermw %ymm20, %ymm24, %ymm20
 ; AVX512DQ-BW-NEXT:    vmovdqa64 32(%r9), %ymm24
-; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm24, %ymm23
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm23 = ymm23[2,2,2,3]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm23, %zmm20, %zmm20
+; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm24, %ymm22
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm20, %zmm20
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm20, %zmm7 {%k3}
-; AVX512DQ-BW-NEXT:    vpbroadcastq {{.*#+}} ymm26 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm25, %ymm20
-; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm27, %ymm23
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm20 = ymm23[0],ymm20[0],ymm23[1],ymm20[1],ymm23[2],ymm20[2],ymm23[3],ymm20[3],ymm23[4],ymm20[4],ymm23[5],ymm20[5],ymm23[6],ymm20[6],ymm23[7],ymm20[7],ymm23[16],ymm20[16],ymm23[17],ymm20[17],ymm23[18],ymm20[18],ymm23[19],ymm20[19],ymm23[20],ymm20[20],ymm23[21],ymm20[21],ymm23[22],ymm20[22],ymm23[23],ymm20[23]
+; AVX512DQ-BW-NEXT:    vpbroadcastq {{.*#+}} ymm25 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm25, %ymm27, %ymm20
+; AVX512DQ-BW-NEXT:    vpshufb %ymm25, %ymm28, %ymm22
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm20 = ymm22[0],ymm20[0],ymm22[1],ymm20[1],ymm22[2],ymm20[2],ymm22[3],ymm20[3],ymm22[4],ymm20[4],ymm22[5],ymm20[5],ymm22[6],ymm20[6],ymm22[7],ymm20[7],ymm22[16],ymm20[16],ymm22[17],ymm20[17],ymm22[18],ymm20[18],ymm22[19],ymm20[19],ymm22[20],ymm20[20],ymm22[21],ymm20[21],ymm22[22],ymm20[22],ymm22[23],ymm20[23]
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} ymm23 = ymm27[8],ymm25[8],ymm27[9],ymm25[9],ymm27[10],ymm25[10],ymm27[11],ymm25[11],ymm27[12],ymm25[12],ymm27[13],ymm25[13],ymm27[14],ymm25[14],ymm27[15],ymm25[15],ymm27[24],ymm25[24],ymm27[25],ymm25[25],ymm27[26],ymm25[26],ymm27[27],ymm25[27],ymm27[28],ymm25[28],ymm27[29],ymm25[29],ymm27[30],ymm25[30],ymm27[31],ymm25[31]
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm25 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512DQ-BW-NEXT:    vpermw %ymm23, %ymm25, %ymm23
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm23, %zmm20, %zmm27
-; AVX512DQ-BW-NEXT:    vpbroadcastq {{.*#+}} ymm23 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm23, %ymm28, %ymm20
-; AVX512DQ-BW-NEXT:    vpshufb %ymm23, %ymm29, %ymm31
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm20 = ymm31[0],ymm20[0],ymm31[1],ymm20[1],ymm31[2],ymm20[2],ymm31[3],ymm20[3],ymm31[4],ymm20[4],ymm31[5],ymm20[5],ymm31[6],ymm20[6],ymm31[7],ymm20[7],ymm31[16],ymm20[16],ymm31[17],ymm20[17],ymm31[18],ymm20[18],ymm31[19],ymm20[19],ymm31[20],ymm20[20],ymm31[21],ymm20[21],ymm31[22],ymm20[22],ymm31[23],ymm20[23]
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} ymm22 = ymm28[8],ymm27[8],ymm28[9],ymm27[9],ymm28[10],ymm27[10],ymm28[11],ymm27[11],ymm28[12],ymm27[12],ymm28[13],ymm27[13],ymm28[14],ymm27[14],ymm28[15],ymm27[15],ymm28[24],ymm27[24],ymm28[25],ymm27[25],ymm28[26],ymm27[26],ymm28[27],ymm27[27],ymm28[28],ymm27[28],ymm28[29],ymm27[29],ymm28[30],ymm27[30],ymm28[31],ymm27[31]
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm26 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512DQ-BW-NEXT:    vpermw %ymm22, %ymm26, %ymm22
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm20, %zmm27
+; AVX512DQ-BW-NEXT:    vpbroadcastq {{.*#+}} ymm22 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm22, %ymm29, %ymm20
+; AVX512DQ-BW-NEXT:    vpshufb %ymm22, %ymm30, %ymm28
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm20 = ymm28[0],ymm20[0],ymm28[1],ymm20[1],ymm28[2],ymm20[2],ymm28[3],ymm20[3],ymm28[4],ymm20[4],ymm28[5],ymm20[5],ymm28[6],ymm20[6],ymm28[7],ymm20[7],ymm28[16],ymm20[16],ymm28[17],ymm20[17],ymm28[18],ymm20[18],ymm28[19],ymm20[19],ymm28[20],ymm20[20],ymm28[21],ymm20[21],ymm28[22],ymm20[22],ymm28[23],ymm20[23]
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} ymm28 = ymm29[8],ymm28[8],ymm29[9],ymm28[9],ymm29[10],ymm28[10],ymm29[11],ymm28[11],ymm29[12],ymm28[12],ymm29[13],ymm28[13],ymm29[14],ymm28[14],ymm29[15],ymm28[15],ymm29[24],ymm28[24],ymm29[25],ymm28[25],ymm29[26],ymm28[26],ymm29[27],ymm28[27],ymm29[28],ymm28[28],ymm29[29],ymm28[29],ymm29[30],ymm28[30],ymm29[31],ymm28[31]
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} ymm28 = ymm30[8],ymm29[8],ymm30[9],ymm29[9],ymm30[10],ymm29[10],ymm30[11],ymm29[11],ymm30[12],ymm29[12],ymm30[13],ymm29[13],ymm30[14],ymm29[14],ymm30[15],ymm29[15],ymm30[24],ymm29[24],ymm30[25],ymm29[25],ymm30[26],ymm29[26],ymm30[27],ymm29[27],ymm30[28],ymm29[28],ymm30[29],ymm29[29],ymm30[30],ymm29[30],ymm30[31],ymm29[31]
 ; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm29 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
 ; AVX512DQ-BW-NEXT:    vpermw %ymm28, %ymm29, %ymm28
 ; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm28, %zmm20, %zmm20
 ; AVX512DQ-BW-NEXT:    vmovdqu16 %zmm27, %zmm20 {%k1}
-; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm27 = zmm30[0,1,2,3],zmm14[4,5,6,7]
+; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm27 = zmm31[0,1,2,3],zmm14[4,5,6,7]
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm27 = zmm27[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm27 = zmm27[2,2,2,3,6,6,6,7]
 ; AVX512DQ-BW-NEXT:    movl $-1840700270, %ecx # imm = 0x92492492
@@ -6941,15 +6931,15 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    movabsq $-9076969306111049208, %rcx # imm = 0x8208208208208208
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k3
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm24, %zmm20 {%k3}
-; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm16, %ymm24
-; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm17, %ymm26
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm24 = ymm26[0],ymm24[0],ymm26[1],ymm24[1],ymm26[2],ymm24[2],ymm26[3],ymm24[3],ymm26[4],ymm24[4],ymm26[5],ymm24[5],ymm26[6],ymm24[6],ymm26[7],ymm24[7],ymm26[16],ymm24[16],ymm26[17],ymm24[17],ymm26[18],ymm24[18],ymm26[19],ymm24[19],ymm26[20],ymm24[20],ymm26[21],ymm24[21],ymm26[22],ymm24[22],ymm26[23],ymm24[23]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm25, %ymm16, %ymm24
+; AVX512DQ-BW-NEXT:    vpshufb %ymm25, %ymm17, %ymm25
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm24 = ymm25[0],ymm24[0],ymm25[1],ymm24[1],ymm25[2],ymm24[2],ymm25[3],ymm24[3],ymm25[4],ymm24[4],ymm25[5],ymm24[5],ymm25[6],ymm24[6],ymm25[7],ymm24[7],ymm25[16],ymm24[16],ymm25[17],ymm24[17],ymm25[18],ymm24[18],ymm25[19],ymm24[19],ymm25[20],ymm24[20],ymm25[21],ymm24[21],ymm25[22],ymm24[22],ymm25[23],ymm24[23]
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
 ; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} ymm16 = ymm17[8],ymm16[8],ymm17[9],ymm16[9],ymm17[10],ymm16[10],ymm17[11],ymm16[11],ymm17[12],ymm16[12],ymm17[13],ymm16[13],ymm17[14],ymm16[14],ymm17[15],ymm16[15],ymm17[24],ymm16[24],ymm17[25],ymm16[25],ymm17[26],ymm16[26],ymm17[27],ymm16[27],ymm17[28],ymm16[28],ymm17[29],ymm16[29],ymm17[30],ymm16[30],ymm17[31],ymm16[31]
-; AVX512DQ-BW-NEXT:    vpermw %ymm16, %ymm25, %ymm16
+; AVX512DQ-BW-NEXT:    vpermw %ymm16, %ymm26, %ymm16
 ; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm16, %zmm24, %zmm17
-; AVX512DQ-BW-NEXT:    vpshufb %ymm23, %ymm18, %ymm16
-; AVX512DQ-BW-NEXT:    vpshufb %ymm23, %ymm19, %ymm24
+; AVX512DQ-BW-NEXT:    vpshufb %ymm22, %ymm18, %ymm16
+; AVX512DQ-BW-NEXT:    vpshufb %ymm22, %ymm19, %ymm24
 ; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} ymm16 = ymm24[0],ymm16[0],ymm24[1],ymm16[1],ymm24[2],ymm16[2],ymm24[3],ymm16[3],ymm24[4],ymm16[4],ymm24[5],ymm16[5],ymm24[6],ymm16[6],ymm24[7],ymm16[7],ymm24[16],ymm16[16],ymm24[17],ymm16[17],ymm24[18],ymm16[18],ymm24[19],ymm16[19],ymm24[20],ymm16[20],ymm24[21],ymm16[21],ymm24[22],ymm16[22],ymm24[23],ymm16[23]
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm16 = ymm16[2,2,2,3]
 ; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} ymm18 = ymm19[8],ymm18[8],ymm19[9],ymm18[9],ymm19[10],ymm18[10],ymm19[11],ymm18[11],ymm19[12],ymm18[12],ymm19[13],ymm18[13],ymm19[14],ymm18[14],ymm19[15],ymm18[15],ymm19[24],ymm18[24],ymm19[25],ymm18[25],ymm19[26],ymm18[26],ymm19[27],ymm18[27],ymm19[28],ymm18[28],ymm19[29],ymm18[29],ymm19[30],ymm18[30],ymm19[31],ymm18[31]
@@ -6960,7 +6950,7 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm14 = zmm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm14 = zmm14[2,2,2,3,6,6,6,7]
 ; AVX512DQ-BW-NEXT:    vmovdqu16 %zmm14, %zmm16 {%k2}
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm12, %zmm12
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm23, %zmm12, %zmm12
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm12 = zmm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm12 = zmm12[2,2,2,3,6,6,6,7]
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm12, %zmm16 {%k3}
@@ -6973,8 +6963,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} ymm11 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
 ; AVX512DQ-BW-NEXT:    vpermw %ymm9, %ymm11, %ymm9
 ; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm14, %zmm9, %zmm9
-; AVX512DQ-BW-NEXT:    vpshufb %xmm23, %xmm8, %xmm14
-; AVX512DQ-BW-NEXT:    vpshufb %xmm23, %xmm10, %xmm17
+; AVX512DQ-BW-NEXT:    vpshufb %xmm22, %xmm8, %xmm14
+; AVX512DQ-BW-NEXT:    vpshufb %xmm22, %xmm10, %xmm17
 ; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm14 = xmm17[0],xmm14[0],xmm17[1],xmm14[1],xmm17[2],xmm14[2],xmm17[3],xmm14[3],xmm17[4],xmm14[4],xmm17[5],xmm14[5],xmm17[6],xmm14[6],xmm17[7],xmm14[7]
 ; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3],xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
 ; AVX512DQ-BW-NEXT:    vprold $16, %xmm8, %xmm8
@@ -6987,11 +6977,11 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm10, %zmm8, %zmm8
 ; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm10 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,22,21,16,23,22,21,16,23,22,21,16,23,17,17,17,17]
 ; AVX512DQ-BW-NEXT:    vpermw %zmm8, %zmm10, %zmm9 {%k1}
-; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} xmm8 = xmm15[2,1,2,3]
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm8[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm13 = xmm15[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm8, %zmm13, %zmm8
-; AVX512DQ-BW-NEXT:    vpermw %zmm8, %zmm10, %zmm8
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm15[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} xmm13 = xmm15[2,1,2,3]
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm13 = xmm13[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm14 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
+; AVX512DQ-BW-NEXT:    vpermt2w %zmm13, %zmm14, %zmm8
 ; AVX512DQ-BW-NEXT:    movabsq $585610922974906400, %rcx # imm = 0x820820820820820
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k3
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm8, %zmm9 {%k3}
@@ -7002,8 +6992,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = 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:    vpermw %ymm1, %ymm11, %ymm1
 ; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm8, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT:    vpshufb %xmm23, %xmm2, %xmm3
-; AVX512DQ-BW-NEXT:    vpshufb %xmm23, %xmm4, %xmm8
+; AVX512DQ-BW-NEXT:    vpshufb %xmm22, %xmm2, %xmm3
+; AVX512DQ-BW-NEXT:    vpshufb %xmm22, %xmm4, %xmm8
 ; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3],xmm8[4],xmm3[4],xmm8[5],xmm3[5],xmm8[6],xmm3[6],xmm8[7],xmm3[7]
 ; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
 ; AVX512DQ-BW-NEXT:    vprold $16, %xmm2, %xmm2
@@ -7015,11 +7005,10 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vpmovzxbw {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
 ; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm3, %zmm2, %zmm2
 ; AVX512DQ-BW-NEXT:    vpermw %zmm2, %zmm10, %zmm1 {%k1}
-; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm6[2,1,2,3]
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm3, %zmm2
-; AVX512DQ-BW-NEXT:    vpermw %zmm2, %zmm10, %zmm2
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} xmm3 = xmm6[2,1,2,3]
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-NEXT:    vpermt2w %zmm3, %zmm14, %zmm2
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm2, %zmm1 {%k3}
 ; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, (%rax)
 ; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, 192(%rax)
@@ -7032,8 +7021,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ;
 ; AVX512DQ-BW-FCP-LABEL: store_i8_stride6_vf64:
 ; AVX512DQ-BW-FCP:       # %bb.0:
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm8
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm10
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm11
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm13
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rsi), %ymm3
 ; AVX512DQ-BW-FCP-NEXT:    vpbroadcastq {{.*#+}} ymm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm5, %ymm3, %ymm0
@@ -7042,192 +7031,191 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
 ; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm1 = ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15],ymm4[24],ymm3[24],ymm4[25],ymm3[25],ymm4[26],ymm3[26],ymm4[27],ymm3[27],ymm4[28],ymm3[28],ymm4[29],ymm3[29],ymm4[30],ymm3[30],ymm4[31],ymm3[31]
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm9 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm1, %ymm9, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm12 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm1, %ymm12, %ymm1
 ; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rcx), %ymm6
-; AVX512DQ-BW-FCP-NEXT:    vpbroadcastq {{.*#+}} ymm17 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm17, %ymm6, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdx), %ymm7
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm17, %ymm7, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rcx), %ymm8
+; AVX512DQ-BW-FCP-NEXT:    vpbroadcastq {{.*#+}} ymm6 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm8, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdx), %ymm10
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm10, %ymm2
 ; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[16],ymm0[16],ymm2[17],ymm0[17],ymm2[18],ymm0[18],ymm2[19],ymm0[19],ymm2[20],ymm0[20],ymm2[21],ymm0[21],ymm2[22],ymm0[22],ymm2[23],ymm0[23]
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm2 = ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15],ymm7[24],ymm6[24],ymm7[25],ymm6[25],ymm7[26],ymm6[26],ymm7[27],ymm6[27],ymm7[28],ymm6[28],ymm7[29],ymm6[29],ymm7[30],ymm6[30],ymm7[31],ymm6[31]
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm12 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm2, %ymm12, %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm2 = ymm10[8],ymm8[8],ymm10[9],ymm8[9],ymm10[10],ymm8[10],ymm10[11],ymm8[11],ymm10[12],ymm8[12],ymm10[13],ymm8[13],ymm10[14],ymm8[14],ymm10[15],ymm8[15],ymm10[24],ymm8[24],ymm10[25],ymm8[25],ymm10[26],ymm8[26],ymm10[27],ymm8[27],ymm10[28],ymm8[28],ymm10[29],ymm8[29],ymm10[30],ymm8[30],ymm10[31],ymm8[31]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm14 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm2, %ymm14, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
 ; AVX512DQ-BW-FCP-NEXT:    movl $613566756, %eax # imm = 0x24924924
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm1, %zmm0 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r8), %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm8[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm1, %zmm2 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%r8), %ymm17
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm17[0,1,2,3],zmm11[4,5,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    movl $-1840700270, %eax # imm = 0x92492492
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm1, %zmm0 {%k2}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r9), %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm11 = zmm1[0,1,2,3],zmm10[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm11 = zmm11[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm11 = zmm11[2,2,2,3,6,6,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm1, %zmm2 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r9), %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm7 = zmm0[0,1,2,3],zmm13[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm7 = zmm7[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm7 = zmm7[2,2,2,3,6,6,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $-9076969306111049208, %rax # imm = 0x8208208208208208
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm11, %zmm0 {%k3}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %ymm11
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm5, %ymm11, %ymm13
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm14
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm5, %ymm14, %ymm5
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm13[0],ymm5[1],ymm13[1],ymm5[2],ymm13[2],ymm5[3],ymm13[3],ymm5[4],ymm13[4],ymm5[5],ymm13[5],ymm5[6],ymm13[6],ymm5[7],ymm13[7],ymm5[16],ymm13[16],ymm5[17],ymm13[17],ymm5[18],ymm13[18],ymm5[19],ymm13[19],ymm5[20],ymm13[20],ymm5[21],ymm13[21],ymm5[22],ymm13[22],ymm5[23],ymm13[23]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm7, %zmm2 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %ymm7
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm5, %ymm7, %ymm15
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm9
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm5, %ymm9, %ymm5
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm15[0],ymm5[1],ymm15[1],ymm5[2],ymm15[2],ymm5[3],ymm15[3],ymm5[4],ymm15[4],ymm5[5],ymm15[5],ymm5[6],ymm15[6],ymm5[7],ymm15[7],ymm5[16],ymm15[16],ymm5[17],ymm15[17],ymm5[18],ymm15[18],ymm5[19],ymm15[19],ymm5[20],ymm15[20],ymm5[21],ymm15[21],ymm5[22],ymm15[22],ymm5[23],ymm15[23]
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm13 = ymm14[8],ymm11[8],ymm14[9],ymm11[9],ymm14[10],ymm11[10],ymm14[11],ymm11[11],ymm14[12],ymm11[12],ymm14[13],ymm11[13],ymm14[14],ymm11[14],ymm14[15],ymm11[15],ymm14[24],ymm11[24],ymm14[25],ymm11[25],ymm14[26],ymm11[26],ymm14[27],ymm11[27],ymm14[28],ymm11[28],ymm14[29],ymm11[29],ymm14[30],ymm11[30],ymm14[31],ymm11[31]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm13, %ymm9, %ymm9
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm5, %zmm9
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rcx), %ymm19
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm17, %ymm19, %ymm5
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdx), %ymm20
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm17, %ymm20, %ymm13
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm13[0],ymm5[0],ymm13[1],ymm5[1],ymm13[2],ymm5[2],ymm13[3],ymm5[3],ymm13[4],ymm5[4],ymm13[5],ymm5[5],ymm13[6],ymm5[6],ymm13[7],ymm5[7],ymm13[16],ymm5[16],ymm13[17],ymm5[17],ymm13[18],ymm5[18],ymm13[19],ymm5[19],ymm13[20],ymm5[20],ymm13[21],ymm5[21],ymm13[22],ymm5[22],ymm13[23],ymm5[23]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm15 = ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15],ymm9[24],ymm7[24],ymm9[25],ymm7[25],ymm9[26],ymm7[26],ymm9[27],ymm7[27],ymm9[28],ymm7[28],ymm9[29],ymm7[29],ymm9[30],ymm7[30],ymm9[31],ymm7[31]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm15, %ymm12, %ymm12
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm5, %zmm12
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rcx), %ymm23
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm23, %ymm5
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdx), %ymm24
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm24, %ymm15
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm5 = ymm15[0],ymm5[0],ymm15[1],ymm5[1],ymm15[2],ymm5[2],ymm15[3],ymm5[3],ymm15[4],ymm5[4],ymm15[5],ymm5[5],ymm15[6],ymm5[6],ymm15[7],ymm5[7],ymm15[16],ymm5[16],ymm15[17],ymm5[17],ymm15[18],ymm5[18],ymm15[19],ymm5[19],ymm15[20],ymm5[20],ymm15[21],ymm5[21],ymm15[22],ymm5[22],ymm15[23],ymm5[23]
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm13 = ymm20[8],ymm19[8],ymm20[9],ymm19[9],ymm20[10],ymm19[10],ymm20[11],ymm19[11],ymm20[12],ymm19[12],ymm20[13],ymm19[13],ymm20[14],ymm19[14],ymm20[15],ymm19[15],ymm20[24],ymm19[24],ymm20[25],ymm19[25],ymm20[26],ymm19[26],ymm20[27],ymm19[27],ymm20[28],ymm19[28],ymm20[29],ymm19[29],ymm20[30],ymm19[30],ymm20[31],ymm19[31]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm13, %ymm12, %ymm12
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm5, %zmm5
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm9, %zmm5 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %ymm9
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm8, %zmm8
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm8 = zmm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm8, %zmm5 {%k2}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r9), %ymm8
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm10, %zmm10
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm10 = zmm10[2,2,2,3,6,6,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm10, %zmm5 {%k3}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rcx), %xmm22
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rcx), %xmm13
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm17, %xmm13, %xmm10
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdx), %xmm23
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdx), %xmm15
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm17, %xmm15, %xmm12
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm12 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3],xmm15[4],xmm13[4],xmm15[5],xmm13[5],xmm15[6],xmm13[6],xmm15[7],xmm13[7]
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm24 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm12, %ymm24, %ymm12
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm12, %zmm12
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%rsi), %xmm18
-; AVX512DQ-BW-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm25 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm25, %xmm18, %xmm10
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%rdi), %xmm21
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm25, %xmm21, %xmm16
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm16[8],xmm10[8],xmm16[9],xmm10[9],xmm16[10],xmm10[10],xmm16[11],xmm10[11],xmm16[12],xmm10[12],xmm16[13],xmm10[13],xmm16[14],xmm10[14],xmm16[15],xmm10[15]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm16 = xmm21[0],xmm18[0],xmm21[1],xmm18[1],xmm21[2],xmm18[2],xmm21[3],xmm18[3],xmm21[4],xmm18[4],xmm21[5],xmm18[5],xmm21[6],xmm18[6],xmm21[7],xmm18[7]
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm26 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm16, %ymm26, %ymm16
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm16, %zmm10
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm12, %zmm10 {%k2}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r8), %xmm12
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} ymm15 = ymm24[8],ymm23[8],ymm24[9],ymm23[9],ymm24[10],ymm23[10],ymm24[11],ymm23[11],ymm24[12],ymm23[12],ymm24[13],ymm23[13],ymm24[14],ymm23[14],ymm24[15],ymm23[15],ymm24[24],ymm23[24],ymm24[25],ymm23[25],ymm24[26],ymm23[26],ymm24[27],ymm23[27],ymm24[28],ymm23[28],ymm24[29],ymm23[29],ymm24[30],ymm23[30],ymm24[31],ymm23[31]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm15, %ymm14, %ymm14
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm14, %zmm5, %zmm5
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm12, %zmm5 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %ymm12
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm11, %zmm11
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm11 = zmm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm11 = zmm11[2,2,2,3,6,6,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm11, %zmm5 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r9), %ymm11
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm13, %zmm13
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm13 = zmm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm13 = zmm13[2,2,2,3,6,6,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm13, %zmm5 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%rcx), %xmm20
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm20, %xmm13
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%rdx), %xmm21
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm21, %xmm14
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm13 = 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:    vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm14 = xmm21[0],xmm20[0],xmm21[1],xmm20[1],xmm21[2],xmm20[2],xmm21[3],xmm20[3],xmm21[4],xmm20[4],xmm21[5],xmm20[5],xmm21[6],xmm20[6],xmm21[7],xmm20[7]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm0 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm14, %ymm0, %ymm14
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm14, %zmm22
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rsi), %xmm15
+; AVX512DQ-BW-FCP-NEXT:    vpbroadcastq {{.*#+}} xmm19 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm19, %xmm15, %xmm30
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r8), %xmm18
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r8), %xmm13
 ; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} xmm27 = [8,9,0,0,0,5,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm27, %xmm12, %xmm16
-; AVX512DQ-BW-FCP-NEXT:    vpmovzxbw {{.*#+}} xmm28 = xmm12[0],zero,xmm12[1],zero,xmm12[2],zero,xmm12[3],zero,xmm12[4],zero,xmm12[5],zero,xmm12[6],zero,xmm12[7],zero
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm16, %zmm28, %zmm16
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm28 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,22,21,16,23,22,21,16,23,22,21,16,23,17,17,17,17]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm16, %zmm28, %zmm10 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%r9), %xmm16
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm27, %xmm13, %xmm14
+; AVX512DQ-BW-FCP-NEXT:    vpmovzxbw {{.*#+}} xmm25 = xmm13[0],zero,xmm13[1],zero,xmm13[2],zero,xmm13[3],zero,xmm13[4],zero,xmm13[5],zero,xmm13[6],zero,xmm13[7],zero
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm28 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
+; AVX512DQ-BW-FCP-NEXT:    vpermt2w %zmm14, %zmm28, %zmm25
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r9), %xmm16
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r9), %xmm14
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm29 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm29, %xmm16, %xmm30
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm31 = xmm16[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm30, %zmm31, %zmm30
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rsi), %xmm31
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm30, %zmm28, %zmm30
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm29, %xmm14, %xmm31
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm26 = xmm14[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-FCP-NEXT:    vpermt2w %zmm31, %zmm28, %zmm26
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm27, %xmm18, %xmm31
+; AVX512DQ-BW-FCP-NEXT:    vpmovzxbw {{.*#+}} xmm27 = xmm18[0],zero,xmm18[1],zero,xmm18[2],zero,xmm18[3],zero,xmm18[4],zero,xmm18[5],zero,xmm18[6],zero,xmm18[7],zero
+; AVX512DQ-BW-FCP-NEXT:    vpermt2w %zmm31, %zmm28, %zmm27
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm29, %xmm16, %xmm31
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm29 = xmm16[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-FCP-NEXT:    vpermt2w %zmm31, %zmm28, %zmm29
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%rdi), %xmm28
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm19, %xmm28, %xmm31
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm30 = xmm31[8],xmm30[8],xmm31[9],xmm30[9],xmm31[10],xmm30[10],xmm31[11],xmm30[11],xmm31[12],xmm30[12],xmm31[13],xmm30[13],xmm31[14],xmm30[14],xmm31[15],xmm30[15]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rcx), %xmm31
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm24 = ymm24[0],ymm23[0],ymm24[1],ymm23[1],ymm24[2],ymm23[2],ymm24[3],ymm23[3],ymm24[4],ymm23[4],ymm24[5],ymm23[5],ymm24[6],ymm23[6],ymm24[7],ymm23[7],ymm24[16],ymm23[16],ymm24[17],ymm23[17],ymm24[18],ymm23[18],ymm24[19],ymm23[19],ymm24[20],ymm23[20],ymm24[21],ymm23[21],ymm24[22],ymm23[22],ymm24[23],ymm23[23]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm23 = xmm0[8],xmm31[8],xmm0[9],xmm31[9],xmm0[10],xmm31[10],xmm0[11],xmm31[11],xmm0[12],xmm31[12],xmm0[13],xmm31[13],xmm0[14],xmm31[14],xmm0[15],xmm31[15]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
+; AVX512DQ-BW-FCP-NEXT:    vpermt2w %zmm24, %zmm1, %zmm23
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm8 = ymm10[0],ymm8[0],ymm10[1],ymm8[1],ymm10[2],ymm8[2],ymm10[3],ymm8[3],ymm10[4],ymm8[4],ymm10[5],ymm8[5],ymm10[6],ymm8[6],ymm10[7],ymm8[7],ymm10[16],ymm8[16],ymm10[17],ymm8[17],ymm10[18],ymm8[18],ymm10[19],ymm8[19],ymm10[20],ymm8[20],ymm10[21],ymm8[21],ymm10[22],ymm8[22],ymm10[23],ymm8[23]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm21[8],xmm20[8],xmm21[9],xmm20[9],xmm21[10],xmm20[10],xmm21[11],xmm20[11],xmm21[12],xmm20[12],xmm21[13],xmm20[13],xmm21[14],xmm20[14],xmm21[15],xmm20[15]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm20 = ymm30[0,0,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpermt2w %zmm8, %zmm1, %zmm10
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm28[0],xmm15[0],xmm28[1],xmm15[1],xmm28[2],xmm15[2],xmm28[3],xmm15[3],xmm28[4],xmm15[4],xmm28[5],xmm15[5],xmm28[6],xmm15[6],xmm28[7],xmm15[7]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm21 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm1, %ymm21, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm20, %zmm1, %zmm8
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm22, %zmm8 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm25, %zmm8 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $585610922974906400, %rax # imm = 0x820820820820820
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm30, %zmm10 {%k3}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm17, %xmm22, %xmm30
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm17, %xmm23, %xmm17
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm17 = xmm17[0],xmm30[0],xmm17[1],xmm30[1],xmm17[2],xmm30[2],xmm17[3],xmm30[3],xmm17[4],xmm30[4],xmm17[5],xmm30[5],xmm17[6],xmm30[6],xmm17[7],xmm30[7]
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm30 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm30, %ymm24, %ymm24
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %xmm30
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[0,0,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm24, %zmm24
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm25, %xmm31, %xmm17
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm25, %xmm30, %xmm25
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm17 = xmm25[8],xmm17[8],xmm25[9],xmm17[9],xmm25[10],xmm17[10],xmm25[11],xmm17[11],xmm25[12],xmm17[12],xmm25[13],xmm17[13],xmm25[14],xmm17[14],xmm25[15],xmm17[15]
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm25 = xmm30[0],xmm31[0],xmm30[1],xmm31[1],xmm30[2],xmm31[2],xmm30[3],xmm31[3],xmm30[4],xmm31[4],xmm30[5],xmm31[5],xmm30[6],xmm31[6],xmm30[7],xmm31[7]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm25, %ymm26, %ymm25
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[0,0,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm17, %zmm25, %zmm17
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r8), %xmm25
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm24, %zmm17 {%k2}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm27, %xmm25, %xmm24
-; AVX512DQ-BW-FCP-NEXT:    vpmovzxbw {{.*#+}} xmm26 = xmm25[0],zero,xmm25[1],zero,xmm25[2],zero,xmm25[3],zero,xmm25[4],zero,xmm25[5],zero,xmm25[6],zero,xmm25[7],zero
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm24, %zmm26, %zmm26
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r9), %xmm24
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm26, %zmm28, %zmm17 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm29, %xmm24, %xmm26
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm27 = xmm24[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm26, %zmm27, %zmm26
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm26, %zmm28, %zmm26
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm26, %zmm17 {%k3}
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm19 = ymm20[0],ymm19[0],ymm20[1],ymm19[1],ymm20[2],ymm19[2],ymm20[3],ymm19[3],ymm20[4],ymm19[4],ymm20[5],ymm19[5],ymm20[6],ymm19[6],ymm20[7],ymm19[7],ymm20[16],ymm19[16],ymm20[17],ymm19[17],ymm20[18],ymm19[18],ymm20[19],ymm19[19],ymm20[20],ymm19[20],ymm20[21],ymm19[21],ymm20[22],ymm19[22],ymm20[23],ymm19[23]
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm19, %zmm20, %zmm19
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm11 = ymm14[0],ymm11[0],ymm14[1],ymm11[1],ymm14[2],ymm11[2],ymm14[3],ymm11[3],ymm14[4],ymm11[4],ymm14[5],ymm11[5],ymm14[6],ymm11[6],ymm14[7],ymm11[7],ymm14[16],ymm11[16],ymm14[17],ymm11[17],ymm14[18],ymm11[18],ymm14[19],ymm11[19],ymm14[20],ymm11[20],ymm14[21],ymm11[21],ymm14[22],ymm11[22],ymm14[23],ymm11[23]
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm14 = xmm30[8],xmm31[8],xmm30[9],xmm31[9],xmm30[10],xmm31[10],xmm30[11],xmm31[11],xmm30[12],xmm31[12],xmm30[13],xmm31[13],xmm30[14],xmm31[14],xmm30[15],xmm31[15]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm14, %zmm11
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm14 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm11, %zmm14, %zmm11
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,25,24,27,26,25,24,27,26,25,24,27,26,29,28,31,30]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm19, %zmm20, %zmm11 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[16],ymm6[16],ymm7[17],ymm6[17],ymm7[18],ymm6[18],ymm7[19],ymm6[19],ymm7[20],ymm6[20],ymm7[21],ymm6[21],ymm7[22],ymm6[22],ymm7[23],ymm6[23]
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm15[8],xmm13[8],xmm15[9],xmm13[9],xmm15[10],xmm13[10],xmm15[11],xmm13[11],xmm15[12],xmm13[12],xmm15[13],xmm13[13],xmm15[14],xmm13[14],xmm15[15],xmm13[15]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[16],ymm3[16],ymm4[17],ymm3[17],ymm4[18],ymm3[18],ymm4[19],ymm3[19],ymm4[20],ymm3[20],ymm4[21],ymm3[21],ymm4[22],ymm3[22],ymm4[23],ymm3[23]
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm21[8],xmm18[8],xmm21[9],xmm18[9],xmm21[10],xmm18[10],xmm21[11],xmm18[11],xmm21[12],xmm18[12],xmm21[13],xmm18[13],xmm21[14],xmm18[14],xmm21[15],xmm18[15]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm3, %zmm14, %zmm3
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm25[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm6, %zmm20, %zmm3 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm6 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm4, %ymm6, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm26, %zmm8 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm31, %xmm1
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm6
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3],xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm31[0],xmm0[1],xmm31[1],xmm0[2],xmm31[2],xmm0[3],xmm31[3],xmm0[4],xmm31[4],xmm0[5],xmm31[5],xmm0[6],xmm31[6],xmm0[7],xmm31[7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %xmm6
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm20 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm0, %ymm20, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %xmm20
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm19, %xmm6, %xmm1
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm19, %xmm20, %xmm19
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm19[8],xmm1[8],xmm19[9],xmm1[9],xmm19[10],xmm1[10],xmm19[11],xmm1[11],xmm19[12],xmm1[12],xmm19[13],xmm1[13],xmm19[14],xmm1[14],xmm19[15],xmm1[15]
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm19 = xmm20[0],xmm6[0],xmm20[1],xmm6[1],xmm20[2],xmm6[2],xmm20[3],xmm6[3],xmm20[4],xmm6[4],xmm20[5],xmm6[5],xmm20[6],xmm6[6],xmm20[7],xmm6[7]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm19, %ymm21, %ymm19
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm19, %zmm1
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm0, %zmm1 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm27, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm29, %zmm1 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[16],ymm7[16],ymm9[17],ymm7[17],ymm9[18],ymm7[18],ymm9[19],ymm7[19],ymm9[20],ymm7[20],ymm9[21],ymm7[21],ymm9[22],ymm7[22],ymm9[23],ymm7[23]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm20[8],xmm6[8],xmm20[9],xmm6[9],xmm20[10],xmm6[10],xmm20[11],xmm6[11],xmm20[12],xmm6[12],xmm20[13],xmm6[13],xmm20[14],xmm6[14],xmm20[15],xmm6[15]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm7 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512DQ-BW-FCP-NEXT:    vpermt2w %zmm0, %zmm7, %zmm6
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm23, %zmm6 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[16],ymm3[16],ymm4[17],ymm3[17],ymm4[18],ymm3[18],ymm4[19],ymm3[19],ymm4[20],ymm3[20],ymm4[21],ymm3[21],ymm4[22],ymm3[22],ymm4[23],ymm3[23]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm28[8],xmm15[8],xmm28[9],xmm15[9],xmm28[10],xmm15[10],xmm28[11],xmm15[11],xmm28[12],xmm15[12],xmm28[13],xmm15[13],xmm28[14],xmm15[14],xmm28[15],xmm15[15]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm18[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-FCP-NEXT:    vpermt2w %zmm0, %zmm7, %zmm3
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm0 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm4, %ymm0, %ymm4
 ; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
 ; AVX512DQ-BW-FCP-NEXT:    # ymm7 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm7, %ymm9, %ymm9
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm7, %ymm12, %ymm9
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
 ; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm4, %zmm4
 ; AVX512DQ-BW-FCP-NEXT:    movl $1227133513, %eax # imm = 0x49249249
-; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm4, %zmm11 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm24[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm4, %ymm6, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm4, %zmm6 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm16[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm4, %ymm0, %ymm4
 ; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm9 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
 ; AVX512DQ-BW-FCP-NEXT:    # ymm9 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm8, %ymm8
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm8, %zmm4, %zmm4
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm11, %ymm11
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm4, %zmm4
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $2342443691899625602, %rax # imm = 0x2082082082082082
-; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm11 {%k2}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm12[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm4, %ymm6, %ymm4
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm2, %zmm3 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm16[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm2, %ymm6, %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm1, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k3
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm6 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm10, %zmm3 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm7, %ymm17, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm13[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm7, %ymm0, %ymm7
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm7, %zmm4
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu16 %zmm4, %zmm3 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm14[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm4, %ymm0, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm4, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm3 {%k3}
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, 256(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, 64(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, 192(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, (%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, 192(%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, 128(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, 320(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, 320(%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
   %in.vec0 = load <64 x i8>, ptr %in.vecptr0, align 64
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll
index 5b59762ae2175..d0ecb22853001 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll
@@ -1091,14 +1091,14 @@ define void @store_i8_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8],zero,zero,zero,zero,ymm0[1,5,9],zero,zero,zero,zero,ymm0[2,6,18],zero,zero,zero,zero,ymm0[23,27,19],zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[0,8],zero,zero,ymm4[u,u,u,1,9],zero,zero,ymm4[u,u,u,2,10,18,26,u,u,u],zero,zero,ymm4[19,27,u,u,u],zero,zero,ymm4[20,28]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,ymm2[0,8,u,u,u],zero,zero,ymm2[1,9,u,u,u],zero,zero,zero,zero,ymm2[u,u,u,19,27],zero,zero,ymm2[u,u,u,20,28],zero,zero
-; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm2 = mem & (ymm2 | ymm1)
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [1,3,5,7,1,3,5,7]
+; AVX512-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,3,0,1]
+; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,ymm4[0,8,u,u,u],zero,zero,ymm4[1,9,u,u,u],zero,zero,zero,zero,ymm4[u,u,u,19,27],zero,zero,ymm4[u,u,u,20,28],zero,zero
+; AVX512-FCP-NEXT:    vpternlogq {{.*#+}} ymm4 = mem & (ymm4 | ymm1)
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [1,3,9,11,1,3,9,11]
 ; AVX512-FCP-NEXT:    # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm1
+; AVX512-FCP-NEXT:    vpermi2d %ymm3, %ymm2, %ymm1
 ; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = zero,zero,zero,ymm1[1,5,9,13],zero,zero,zero,ymm1[2,6,10,14],zero,zero,zero,ymm1[19,23,27,31],zero,zero,zero,ymm1[u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm4, %zmm1
 ; AVX512-FCP-NEXT:    vporq %zmm0, %zmm1, %zmm0
 ; AVX512-FCP-NEXT:    vextracti32x4 $2, %zmm0, 32(%rax)
 ; AVX512-FCP-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
@@ -1173,14 +1173,14 @@ define void @store_i8_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8],zero,zero,zero,zero,ymm0[1,5,9],zero,zero,zero,zero,ymm0[2,6,18],zero,zero,zero,zero,ymm0[23,27,19],zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[0,8],zero,zero,ymm4[u,u,u,1,9],zero,zero,ymm4[u,u,u,2,10,18,26,u,u,u],zero,zero,ymm4[19,27,u,u,u],zero,zero,ymm4[20,28]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,ymm2[0,8,u,u,u],zero,zero,ymm2[1,9,u,u,u],zero,zero,zero,zero,ymm2[u,u,u,19,27],zero,zero,ymm2[u,u,u,20,28],zero,zero
-; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm2 = mem & (ymm2 | ymm1)
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [1,3,5,7,1,3,5,7]
+; AVX512DQ-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,3,0,1]
+; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,ymm4[0,8,u,u,u],zero,zero,ymm4[1,9,u,u,u],zero,zero,zero,zero,ymm4[u,u,u,19,27],zero,zero,ymm4[u,u,u,20,28],zero,zero
+; AVX512DQ-FCP-NEXT:    vpternlogq {{.*#+}} ymm4 = mem & (ymm4 | ymm1)
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [1,3,9,11,1,3,9,11]
 ; AVX512DQ-FCP-NEXT:    # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm3, %ymm2, %ymm1
 ; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = zero,zero,zero,ymm1[1,5,9,13],zero,zero,zero,ymm1[2,6,10,14],zero,zero,zero,ymm1[19,23,27,31],zero,zero,zero,ymm1[u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm4, %zmm1
 ; AVX512DQ-FCP-NEXT:    vporq %zmm0, %zmm1, %zmm0
 ; AVX512DQ-FCP-NEXT:    vextracti32x4 $2, %zmm0, 32(%rax)
 ; AVX512DQ-FCP-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
@@ -1242,28 +1242,28 @@ define void @store_i8_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
 ; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm5 = mem[0],zero
 ; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
 ; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]
-; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [1,3,5,7,1,3,5,7]
-; AVX512BW-FCP-NEXT:    # ymm1 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm1
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = zero,zero,zero,ymm1[1,5,9,13],zero,zero,zero,ymm1[2,6,10,14],zero,zero,zero,ymm1[19,23,27,31],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm5 = [1,5,3,0,3,1,5,0]
-; AVX512BW-FCP-NEXT:    vpermd %ymm0, %ymm5, %ymm5
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[0,4,8],zero,zero,zero,zero,ymm5[1,5,9],zero,zero,zero,zero,ymm5[2,6,18],zero,zero,zero,zero,ymm5[23,27,19],zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512BW-FCP-NEXT:    vpor %ymm1, %ymm5, %ymm1
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[0,8],zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,ymm4[2,10,18,26],zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,zero,ymm4[20,28]
-; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,ymm2[0,8],zero,zero,zero,zero,zero,ymm2[1,9],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[19,27],zero,zero,zero,zero,zero,ymm2[20,28],zero,zero
-; AVX512BW-FCP-NEXT:    vpor %ymm4, %ymm2, %ymm2
-; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [0,4,2,6,0,4,2,6]
-; AVX512BW-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpermd %ymm0, %ymm3, %ymm0
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[0,8],zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,ymm4[2,10,18,26],zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,zero,ymm4[20,28]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,3,0,1]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,ymm4[0,8],zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,zero,ymm4[20,28],zero,zero
+; AVX512BW-FCP-NEXT:    vpor %ymm1, %ymm4, %ymm1
+; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [0,4,2,6,0,4,2,6]
+; AVX512BW-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vpermd %ymm0, %ymm4, %ymm4
 ; AVX512BW-FCP-NEXT:    movl $236730480, %ecx # imm = 0xE1C3870
 ; AVX512BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 {%k1} = ymm0[u,u,u,u,0,4,8,u,u,u,u,1,5,9,u,u,u,u,18,22,26,u,u,u,u,19,23,27,u,u,u,u]
-; AVX512BW-FCP-NEXT:    vmovdqa %ymm2, (%rax)
-; AVX512BW-FCP-NEXT:    vextracti128 $1, %ymm1, %xmm0
-; AVX512BW-FCP-NEXT:    vmovq %xmm0, 48(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa %xmm1, 32(%rax)
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 {%k1} = ymm4[u,u,u,u,0,4,8,u,u,u,u,1,5,9,u,u,u,u,18,22,26,u,u,u,u,19,23,27,u,u,u,u]
+; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [1,3,9,11,1,3,9,11]
+; AVX512BW-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vpermi2d %ymm3, %ymm2, %ymm4
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,zero,ymm4[1,5,9,13],zero,zero,zero,ymm4[2,6,10,14],zero,zero,zero,ymm4[19,23,27,31],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [1,5,3,0,3,1,5,0]
+; AVX512BW-FCP-NEXT:    vpermd %ymm0, %ymm3, %ymm0
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8],zero,zero,zero,zero,ymm0[1,5,9],zero,zero,zero,zero,ymm0[2,6,18],zero,zero,zero,zero,ymm0[23,27,19],zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512BW-FCP-NEXT:    vpor %ymm2, %ymm0, %ymm0
+; AVX512BW-FCP-NEXT:    vextracti128 $1, %ymm0, %xmm2
+; AVX512BW-FCP-NEXT:    vmovq %xmm2, 48(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa %xmm0, 32(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa %ymm1, (%rax)
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
 ;
@@ -1320,28 +1320,28 @@ define void @store_i8_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
 ; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm5 = mem[0],zero
 ; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
 ; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]
-; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [1,3,5,7,1,3,5,7]
-; AVX512DQ-BW-FCP-NEXT:    # ymm1 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = zero,zero,zero,ymm1[1,5,9,13],zero,zero,zero,ymm1[2,6,10,14],zero,zero,zero,ymm1[19,23,27,31],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm5 = [1,5,3,0,3,1,5,0]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm0, %ymm5, %ymm5
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[0,4,8],zero,zero,zero,zero,ymm5[1,5,9],zero,zero,zero,zero,ymm5[2,6,18],zero,zero,zero,zero,ymm5[23,27,19],zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm1, %ymm5, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[0,8],zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,ymm4[2,10,18,26],zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,zero,ymm4[20,28]
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,ymm2[0,8],zero,zero,zero,zero,zero,ymm2[1,9],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[19,27],zero,zero,zero,zero,zero,ymm2[20,28],zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm4, %ymm2, %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [0,4,2,6,0,4,2,6]
-; AVX512DQ-BW-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm0, %ymm3, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[0,8],zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,ymm4[2,10,18,26],zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,zero,ymm4[20,28]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,3,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,ymm4[0,8],zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,zero,ymm4[20,28],zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm1, %ymm4, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [0,4,2,6,0,4,2,6]
+; AVX512DQ-BW-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm0, %ymm4, %ymm4
 ; AVX512DQ-BW-FCP-NEXT:    movl $236730480, %ecx # imm = 0xE1C3870
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 {%k1} = ymm0[u,u,u,u,0,4,8,u,u,u,u,1,5,9,u,u,u,u,18,22,26,u,u,u,u,19,23,27,u,u,u,u]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm2, (%rax)
-; AVX512DQ-BW-FCP-NEXT:    vextracti128 $1, %ymm1, %xmm0
-; AVX512DQ-BW-FCP-NEXT:    vmovq %xmm0, 48(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm1, 32(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 {%k1} = ymm4[u,u,u,u,0,4,8,u,u,u,u,1,5,9,u,u,u,u,18,22,26,u,u,u,u,19,23,27,u,u,u,u]
+; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [1,3,9,11,1,3,9,11]
+; AVX512DQ-BW-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2d %ymm3, %ymm2, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,zero,ymm4[1,5,9,13],zero,zero,zero,ymm4[2,6,10,14],zero,zero,zero,ymm4[19,23,27,31],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [1,5,3,0,3,1,5,0]
+; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm0, %ymm3, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8],zero,zero,zero,zero,ymm0[1,5,9],zero,zero,zero,zero,ymm0[2,6,18],zero,zero,zero,zero,ymm0[23,27,19],zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm2, %ymm0, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vextracti128 $1, %ymm0, %xmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovq %xmm2, 48(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm0, 32(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm1, (%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
   %in.vec0 = load <8 x i8>, ptr %in.vecptr0, align 64
@@ -2324,25 +2324,25 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW:       # %bb.0:
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm1
-; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm0
-; AVX512BW-NEXT:    vmovdqa (%r8), %xmm2
-; AVX512BW-NEXT:    vinserti128 $1, (%rcx), %ymm0, %ymm0
-; AVX512BW-NEXT:    vinserti128 $1, (%rsi), %ymm1, %ymm1
-; AVX512BW-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
-; AVX512BW-NEXT:    vinserti32x4 $2, (%r10), %zmm2, %zmm2
-; AVX512BW-NEXT:    vextracti128 $1, %ymm2, %xmm3
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm3[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
-; AVX512BW-NEXT:    vextracti64x4 $1, %zmm2, %ymm4
-; AVX512BW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm4[0,1,2,3,6,7,7,7]
+; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0
+; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512BW-NEXT:    vmovdqa (%r8), %xmm3
+; AVX512BW-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm2
+; AVX512BW-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm4
+; AVX512BW-NEXT:    vinserti128 $1, (%r9), %ymm3, %ymm0
+; AVX512BW-NEXT:    vinserti32x4 $2, (%r10), %zmm0, %zmm0
+; AVX512BW-NEXT:    vextracti128 $1, %ymm0, %xmm1
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm1[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
+; AVX512BW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512BW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm1[0,1,2,3,6,7,7,7]
 ; AVX512BW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,1,3,2]
 ; AVX512BW-NEXT:    movw $-32510, %cx # imm = 0x8102
 ; AVX512BW-NEXT:    kmovd %ecx, %k1
 ; AVX512BW-NEXT:    vmovdqu8 %xmm5, %xmm3 {%k1}
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm1[1,3,2,3]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm4[1,3,2,3]
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[6,14],zero,zero,zero,zero,zero,xmm5[7,15],zero,zero,zero,zero,zero
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm0[1,3,2,3]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm2[1,3,2,3]
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm6 = zero,zero,zero,zero,xmm6[6,14],zero,zero,zero,zero,zero,xmm6[7,15],zero,zero,zero
 ; AVX512BW-NEXT:    vpor %xmm5, %xmm6, %xmm5
 ; AVX512BW-NEXT:    movw $-7741, %cx # imm = 0xE1C3
@@ -2350,55 +2350,55 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vmovdqu8 %xmm3, %xmm5 {%k1}
 ; AVX512BW-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
 ; AVX512BW-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512BW-NEXT:    vpermw %ymm4, %ymm3, %ymm3
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm2[1,3,1,3]
+; AVX512BW-NEXT:    vpermw %ymm1, %ymm3, %ymm3
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm0[1,3,1,3]
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,1,9,u,u,u,u,u,2,10,u,u,u,u,u,19,27,u,u,u,u,u,20,28,u,u,u,u,u,21]
 ; AVX512BW-NEXT:    movl $67637280, %ecx # imm = 0x4081020
 ; AVX512BW-NEXT:    kmovd %ecx, %k1
 ; AVX512BW-NEXT:    vmovdqu8 %ymm3, %ymm6 {%k1}
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm3 = ymm0[1,3,3,1]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm3 = ymm2[1,3,3,1]
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm3 = zero,ymm3[1,9],zero,zero,zero,zero,zero,ymm3[2,10],zero,zero,zero,zero,zero,ymm3[3,19],zero,zero,zero,zero,zero,ymm3[28,20],zero,zero,zero,zero,zero,ymm3[29,21],zero
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm1[3,1,1,3]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm4[3,1,1,3]
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm7 = ymm7[1],zero,zero,zero,zero,zero,ymm7[10,2],zero,zero,zero,zero,zero,ymm7[11,3],zero,zero,zero,zero,zero,ymm7[20,28],zero,zero,zero,zero,zero,ymm7[21,29],zero,zero,zero
 ; AVX512BW-NEXT:    vpor %ymm3, %ymm7, %ymm3
 ; AVX512BW-NEXT:    movl $-2029118408, %ecx # imm = 0x870E1C38
 ; AVX512BW-NEXT:    kmovd %ecx, %k1
 ; AVX512BW-NEXT:    vmovdqu8 %ymm6, %ymm3 {%k1}
 ; AVX512BW-NEXT:    vinserti32x4 $2, %xmm5, %zmm3, %zmm3
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm6 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,50,50,50,52,50,50,50,52,51,51,51,51,50,50,50,52]
-; AVX512BW-NEXT:    vpermi2w %zmm2, %zmm4, %zmm6
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm2[4],zero,zero,zero,zero,zero,zero,ymm2[5],zero,zero,zero,zero,zero,zero,ymm2[6],zero,zero,zero,zero,zero,zero,zero,ymm2[23],zero,zero,zero,zero,zero,zero,ymm2[24],zero,zero
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,ymm2[5],zero,zero,zero,zero,zero,zero,ymm2[6],zero,zero,zero,zero,zero,zero,zero,ymm2[23],zero,zero,zero,zero,zero,zero,ymm2[24],zero,zero,zero,zero
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm2[2,3,0,1]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,ymm7[4],zero,zero,zero,zero,zero,zero,ymm7[5],zero,zero,zero,zero,zero,zero,ymm7[6],zero,zero,zero,zero,zero,ymm7[23],zero,zero,zero,zero,zero,zero,ymm7[24],zero,zero,zero
-; AVX512BW-NEXT:    vpor %ymm4, %ymm7, %ymm4
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,zero,zero,ymm7[5],zero,zero,zero,zero,zero,zero,ymm7[6],zero,zero,zero,zero,zero,ymm7[23],zero,zero,zero,zero,zero,zero,ymm7[24],zero,zero,zero,zero,zero
+; AVX512BW-NEXT:    vpor %ymm6, %ymm7, %ymm6
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,ymm4[5],zero,zero,zero,zero,zero,zero,ymm4[6],zero,zero,zero,zero,zero,zero,zero,ymm4[23],zero,zero,zero,zero,zero,zero,ymm4[24],zero,zero,zero,zero,zero,zero
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm8 = ymm4[2,3,0,1]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,ymm8[5],zero,zero,zero,zero,zero,zero,ymm8[6],zero,zero,zero,zero,zero,ymm8[23],zero,zero,zero,zero,zero,zero,ymm8[24],zero,zero,zero,zero,zero,zero,ymm8[25]
+; AVX512BW-NEXT:    vpor %ymm7, %ymm8, %ymm7
+; AVX512BW-NEXT:    movl $202911840, %ecx # imm = 0xC183060
+; AVX512BW-NEXT:    kmovd %ecx, %k1
+; AVX512BW-NEXT:    vmovdqu8 %ymm6, %ymm7 {%k1}
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,2,0,2]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[0,8],zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,ymm4[2,10],zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,zero,ymm4[20,28],zero,zero
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,0,2]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,0,8,u,u,u,u,u,1,9,u,u,u,u,u,18,26,u,u,u,u,u,19,27,u,u,u,u,u]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512BW-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
-; AVX512BW-NEXT:    kmovq %rcx, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm6, %zmm2 {%k1}
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,zero,ymm0[5],zero,zero,zero,zero,zero,zero,ymm0[6],zero,zero,zero,zero,zero,zero,zero,ymm0[23],zero,zero,zero,zero,zero,zero,ymm0[24],zero,zero,zero,zero
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,ymm2[0,8],zero,zero,zero,zero,zero,ymm2[1,9],zero,zero,zero,zero,zero,ymm2[18,26],zero,zero,zero,zero,zero,ymm2[19,27],zero,zero,zero,zero,zero,ymm2[20,28]
+; AVX512BW-NEXT:    vpor %ymm4, %ymm2, %ymm2
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm2, %zmm2
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm0[4],zero,zero,zero,zero,zero,zero,ymm0[5],zero,zero,zero,zero,zero,zero,ymm0[6],zero,zero,zero,zero,zero,zero,zero,ymm0[23],zero,zero,zero,zero,zero,zero,ymm0[24],zero,zero
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm0[2,3,0,1]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero,zero,zero,zero
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero,zero
 ; AVX512BW-NEXT:    vpor %ymm4, %ymm6, %ymm4
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm1[5],zero,zero,zero,zero,zero,zero,ymm1[6],zero,zero,zero,zero,zero,zero,zero,ymm1[23],zero,zero,zero,zero,zero,zero,ymm1[24],zero,zero,zero,zero,zero,zero
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm1[2,3,0,1]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm7[5],zero,zero,zero,zero,zero,zero,ymm7[6],zero,zero,zero,zero,zero,ymm7[23],zero,zero,zero,zero,zero,zero,ymm7[24],zero,zero,zero,zero,zero,zero,ymm7[25]
-; AVX512BW-NEXT:    vpor %ymm7, %ymm6, %ymm6
-; AVX512BW-NEXT:    movl $202911840, %ecx # imm = 0xC183060
-; AVX512BW-NEXT:    kmovd %ecx, %k1
-; AVX512BW-NEXT:    vmovdqu8 %ymm4, %ymm6 {%k1}
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,0,2]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[0,8],zero,zero,zero,zero,zero,ymm1[1,9],zero,zero,zero,zero,zero,ymm1[2,10],zero,zero,zero,zero,zero,ymm1[19,27],zero,zero,zero,zero,zero,ymm1[20,28],zero,zero
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,0,2]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm0 = zero,zero,ymm0[0,8],zero,zero,zero,zero,zero,ymm0[1,9],zero,zero,zero,zero,zero,ymm0[18,26],zero,zero,zero,zero,zero,ymm0[19,27],zero,zero,zero,zero,zero,ymm0[20,28]
-; AVX512BW-NEXT:    vpor %ymm1, %ymm0, %ymm0
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,0,8,u,u,u,u,u,1,9,u,u,u,u,u,18,26,u,u,u,u,u,19,27,u,u,u,u,u]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm4 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512BW-NEXT:    vpermw %zmm1, %zmm4, %zmm1
+; AVX512BW-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
+; AVX512BW-NEXT:    kmovq %rcx, %k1
+; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm0 {%k1}
 ; AVX512BW-NEXT:    movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
 ; AVX512BW-NEXT:    kmovq %rcx, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm2, %zmm0 {%k1}
+; AVX512BW-NEXT:    vmovdqu8 %zmm0, %zmm2 {%k1}
 ; AVX512BW-NEXT:    vmovdqa %xmm5, 96(%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512BW-NEXT:    vmovdqa %ymm3, 64(%rax)
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
@@ -2407,16 +2407,16 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP:       # %bb.0:
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm1
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm0
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm1
 ; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %xmm2
-; AVX512BW-FCP-NEXT:    vinserti128 $1, (%rcx), %ymm0, %ymm0
-; AVX512BW-FCP-NEXT:    vinserti128 $1, (%rsi), %ymm1, %ymm1
+; AVX512BW-FCP-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
+; AVX512BW-FCP-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
 ; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
 ; AVX512BW-FCP-NEXT:    vinserti32x4 $2, (%r10), %zmm2, %zmm2
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm1[1,3,2,3]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm0[1,3,2,3]
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm3 = zero,zero,xmm3[6,14],zero,zero,zero,zero,zero,xmm3[7,15],zero,zero,zero,zero,zero
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[1,3,2,3]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm1[1,3,2,3]
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = zero,zero,zero,zero,xmm4[6,14],zero,zero,zero,zero,zero,xmm4[7,15],zero,zero,zero
 ; AVX512BW-FCP-NEXT:    vpor %xmm3, %xmm4, %xmm3
 ; AVX512BW-FCP-NEXT:    vextracti128 $1, %ymm2, %xmm4
@@ -2436,38 +2436,38 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    movl $67637280, %ecx # imm = 0x4081020
 ; AVX512BW-FCP-NEXT:    kmovd %ecx, %k1
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm4, %ymm6 {%k1}
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[1,3,3,1]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm1[1,3,3,1]
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = zero,ymm4[1,9],zero,zero,zero,zero,zero,ymm4[2,10],zero,zero,zero,zero,zero,ymm4[3,19],zero,zero,zero,zero,zero,ymm4[28,20],zero,zero,zero,zero,zero,ymm4[29,21],zero
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm1[3,1,1,3]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm0[3,1,1,3]
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm7[1],zero,zero,zero,zero,zero,ymm7[10,2],zero,zero,zero,zero,zero,ymm7[11,3],zero,zero,zero,zero,zero,ymm7[20,28],zero,zero,zero,zero,zero,ymm7[21,29],zero,zero,zero
 ; AVX512BW-FCP-NEXT:    vpor %ymm4, %ymm7, %ymm4
 ; AVX512BW-FCP-NEXT:    movl $-2029118408, %ecx # imm = 0x870E1C38
 ; AVX512BW-FCP-NEXT:    kmovd %ecx, %k1
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm6, %ymm4 {%k1}
 ; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm3, %zmm4, %zmm4
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm6 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,50,50,50,52,50,50,50,52,51,51,51,51,50,50,50,52]
-; AVX512BW-FCP-NEXT:    vpermi2w %zmm2, %zmm5, %zmm6
-; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm5 = [1,5,2,6,1,5,2,6]
-; AVX512BW-FCP-NEXT:    # ymm5 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpermd %ymm2, %ymm5, %ymm7
+; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [1,5,2,6,1,5,2,6]
+; AVX512BW-FCP-NEXT:    # ymm6 = mem[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vpermd %ymm0, %ymm6, %ymm7
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,0,2]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm0
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[0,8],zero,zero,zero,zero,zero,zmm0[1,9],zero,zero,zero,zero,zero,zmm0[2,10],zero,zero,zero,zero,zero,zmm0[19,27],zero,zero,zero,zero,zero,zmm0[20,28],zero,zero,zero,zero,zero,zmm0[33,37],zero,zero,zero,zero,zero,zmm0[34,38],zero,zero,zero,zero,zero,zmm0[51,55],zero,zero,zero,zero,zero,zmm0[56,60],zero,zero,zero,zero,zero,zmm0[57]
+; AVX512BW-FCP-NEXT:    vpermd %ymm1, %ymm6, %ymm7
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,0,2]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm1, %zmm1
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zero,zero,zmm1[0,8],zero,zero,zero,zero,zero,zmm1[1,9],zero,zero,zero,zero,zero,zmm1[18,26],zero,zero,zero,zero,zero,zmm1[19,27],zero,zero,zero,zero,zero,zmm1[20,28],zero,zero,zero,zero,zero,zmm1[33,37],zero,zero,zero,zero,zero,zmm1[34,38],zero,zero,zero,zero,zero,zmm1[51,55],zero,zero,zero,zero,zero,zmm1[56,60],zero,zero,zero,zero
+; AVX512BW-FCP-NEXT:    vporq %zmm0, %zmm1, %zmm0
+; AVX512BW-FCP-NEXT:    vpermd %ymm2, %ymm6, %ymm1
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,0,2]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm2, %zmm2
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,0,8,u,u,u,u,u,1,9,u,u,u,u,u,18,26,u,u,u,u,u,19,27,u,u,u,u,u,32,36,u,u,u,u,u,33,37,u,u,u,u,u,34,38,u,u,u,u,u,51,55,u,u,u,u,u,56,60,u,u]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,0,8,u,u,u,u,u,1,9,u,u,u,u,u,18,26,u,u,u,u,u,19,27,u,u,u,u,u,32,36,u,u,u,u,u,33,37,u,u,u,u,u,34,38,u,u,u,u,u,51,55,u,u,u,u,u,56,60,u,u]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512BW-FCP-NEXT:    vpermw %zmm5, %zmm2, %zmm2
 ; AVX512BW-FCP-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
 ; AVX512BW-FCP-NEXT:    kmovq %rcx, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm2 {%k1}
-; AVX512BW-FCP-NEXT:    vpermd %ymm1, %ymm5, %ymm6
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,0,2]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm1, %zmm1
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[0,8],zero,zero,zero,zero,zero,zmm1[1,9],zero,zero,zero,zero,zero,zmm1[2,10],zero,zero,zero,zero,zero,zmm1[19,27],zero,zero,zero,zero,zero,zmm1[20,28],zero,zero,zero,zero,zero,zmm1[33,37],zero,zero,zero,zero,zero,zmm1[34,38],zero,zero,zero,zero,zero,zmm1[51,55],zero,zero,zero,zero,zero,zmm1[56,60],zero,zero,zero,zero,zero,zmm1[57]
-; AVX512BW-FCP-NEXT:    vpermd %ymm0, %ymm5, %ymm5
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,0,2]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm0
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[0,8],zero,zero,zero,zero,zero,zmm0[1,9],zero,zero,zero,zero,zero,zmm0[18,26],zero,zero,zero,zero,zero,zmm0[19,27],zero,zero,zero,zero,zero,zmm0[20,28],zero,zero,zero,zero,zero,zmm0[33,37],zero,zero,zero,zero,zero,zmm0[34,38],zero,zero,zero,zero,zero,zmm0[51,55],zero,zero,zero,zero,zero,zmm0[56,60],zero,zero,zero,zero
-; AVX512BW-FCP-NEXT:    vporq %zmm1, %zmm0, %zmm0
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm1 {%k1}
 ; AVX512BW-FCP-NEXT:    movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
 ; AVX512BW-FCP-NEXT:    kmovq %rcx, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm0 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm0 {%k1}
 ; AVX512BW-FCP-NEXT:    vmovdqa %xmm3, 96(%rax)
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
 ; AVX512BW-FCP-NEXT:    vmovdqa %ymm4, 64(%rax)
@@ -2478,25 +2478,25 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW:       # %bb.0:
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm1
-; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm0
-; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm2
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%rcx), %ymm0, %ymm0
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%rsi), %ymm1, %ymm1
-; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
-; AVX512DQ-BW-NEXT:    vinserti32x4 $2, (%r10), %zmm2, %zmm2
-; AVX512DQ-BW-NEXT:    vextracti128 $1, %ymm2, %xmm3
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm3[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
-; AVX512DQ-BW-NEXT:    vextracti64x4 $1, %zmm2, %ymm4
-; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm4[0,1,2,3,6,7,7,7]
+; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm0
+; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm1
+; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm3
+; AVX512DQ-BW-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm2
+; AVX512DQ-BW-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm4
+; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r9), %ymm3, %ymm0
+; AVX512DQ-BW-NEXT:    vinserti32x4 $2, (%r10), %zmm0, %zmm0
+; AVX512DQ-BW-NEXT:    vextracti128 $1, %ymm0, %xmm1
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm1[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
+; AVX512DQ-BW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm1[0,1,2,3,6,7,7,7]
 ; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,1,3,2]
 ; AVX512DQ-BW-NEXT:    movw $-32510, %cx # imm = 0x8102
 ; AVX512DQ-BW-NEXT:    kmovd %ecx, %k1
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %xmm5, %xmm3 {%k1}
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm1[1,3,2,3]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm4[1,3,2,3]
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[6,14],zero,zero,zero,zero,zero,xmm5[7,15],zero,zero,zero,zero,zero
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm0[1,3,2,3]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm2[1,3,2,3]
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm6 = zero,zero,zero,zero,xmm6[6,14],zero,zero,zero,zero,zero,xmm6[7,15],zero,zero,zero
 ; AVX512DQ-BW-NEXT:    vpor %xmm5, %xmm6, %xmm5
 ; AVX512DQ-BW-NEXT:    movw $-7741, %cx # imm = 0xE1C3
@@ -2504,55 +2504,55 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %xmm3, %xmm5 {%k1}
 ; AVX512DQ-BW-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
 ; AVX512DQ-BW-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vpermw %ymm4, %ymm3, %ymm3
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm2[1,3,1,3]
+; AVX512DQ-BW-NEXT:    vpermw %ymm1, %ymm3, %ymm3
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm0[1,3,1,3]
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,1,9,u,u,u,u,u,2,10,u,u,u,u,u,19,27,u,u,u,u,u,20,28,u,u,u,u,u,21]
 ; AVX512DQ-BW-NEXT:    movl $67637280, %ecx # imm = 0x4081020
 ; AVX512DQ-BW-NEXT:    kmovd %ecx, %k1
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm3, %ymm6 {%k1}
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm3 = ymm0[1,3,3,1]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm3 = ymm2[1,3,3,1]
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm3 = zero,ymm3[1,9],zero,zero,zero,zero,zero,ymm3[2,10],zero,zero,zero,zero,zero,ymm3[3,19],zero,zero,zero,zero,zero,ymm3[28,20],zero,zero,zero,zero,zero,ymm3[29,21],zero
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm1[3,1,1,3]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm4[3,1,1,3]
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm7 = ymm7[1],zero,zero,zero,zero,zero,ymm7[10,2],zero,zero,zero,zero,zero,ymm7[11,3],zero,zero,zero,zero,zero,ymm7[20,28],zero,zero,zero,zero,zero,ymm7[21,29],zero,zero,zero
 ; AVX512DQ-BW-NEXT:    vpor %ymm3, %ymm7, %ymm3
 ; AVX512DQ-BW-NEXT:    movl $-2029118408, %ecx # imm = 0x870E1C38
 ; AVX512DQ-BW-NEXT:    kmovd %ecx, %k1
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm6, %ymm3 {%k1}
 ; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm5, %zmm3, %zmm3
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm6 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,50,50,50,52,50,50,50,52,51,51,51,51,50,50,50,52]
-; AVX512DQ-BW-NEXT:    vpermi2w %zmm2, %zmm4, %zmm6
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm2[4],zero,zero,zero,zero,zero,zero,ymm2[5],zero,zero,zero,zero,zero,zero,ymm2[6],zero,zero,zero,zero,zero,zero,zero,ymm2[23],zero,zero,zero,zero,zero,zero,ymm2[24],zero,zero
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,ymm2[5],zero,zero,zero,zero,zero,zero,ymm2[6],zero,zero,zero,zero,zero,zero,zero,ymm2[23],zero,zero,zero,zero,zero,zero,ymm2[24],zero,zero,zero,zero
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm2[2,3,0,1]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,ymm7[4],zero,zero,zero,zero,zero,zero,ymm7[5],zero,zero,zero,zero,zero,zero,ymm7[6],zero,zero,zero,zero,zero,ymm7[23],zero,zero,zero,zero,zero,zero,ymm7[24],zero,zero,zero
-; AVX512DQ-BW-NEXT:    vpor %ymm4, %ymm7, %ymm4
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,zero,zero,ymm7[5],zero,zero,zero,zero,zero,zero,ymm7[6],zero,zero,zero,zero,zero,ymm7[23],zero,zero,zero,zero,zero,zero,ymm7[24],zero,zero,zero,zero,zero
+; AVX512DQ-BW-NEXT:    vpor %ymm6, %ymm7, %ymm6
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,ymm4[5],zero,zero,zero,zero,zero,zero,ymm4[6],zero,zero,zero,zero,zero,zero,zero,ymm4[23],zero,zero,zero,zero,zero,zero,ymm4[24],zero,zero,zero,zero,zero,zero
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm8 = ymm4[2,3,0,1]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,ymm8[5],zero,zero,zero,zero,zero,zero,ymm8[6],zero,zero,zero,zero,zero,ymm8[23],zero,zero,zero,zero,zero,zero,ymm8[24],zero,zero,zero,zero,zero,zero,ymm8[25]
+; AVX512DQ-BW-NEXT:    vpor %ymm7, %ymm8, %ymm7
+; AVX512DQ-BW-NEXT:    movl $202911840, %ecx # imm = 0xC183060
+; AVX512DQ-BW-NEXT:    kmovd %ecx, %k1
+; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm6, %ymm7 {%k1}
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,2,0,2]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[0,8],zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,ymm4[2,10],zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,zero,ymm4[20,28],zero,zero
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,0,2]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,0,8,u,u,u,u,u,1,9,u,u,u,u,u,18,26,u,u,u,u,u,19,27,u,u,u,u,u]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512DQ-BW-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
-; AVX512DQ-BW-NEXT:    kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm6, %zmm2 {%k1}
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,zero,ymm0[5],zero,zero,zero,zero,zero,zero,ymm0[6],zero,zero,zero,zero,zero,zero,zero,ymm0[23],zero,zero,zero,zero,zero,zero,ymm0[24],zero,zero,zero,zero
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,ymm2[0,8],zero,zero,zero,zero,zero,ymm2[1,9],zero,zero,zero,zero,zero,ymm2[18,26],zero,zero,zero,zero,zero,ymm2[19,27],zero,zero,zero,zero,zero,ymm2[20,28]
+; AVX512DQ-BW-NEXT:    vpor %ymm4, %ymm2, %ymm2
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm2, %zmm2
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm0[4],zero,zero,zero,zero,zero,zero,ymm0[5],zero,zero,zero,zero,zero,zero,ymm0[6],zero,zero,zero,zero,zero,zero,zero,ymm0[23],zero,zero,zero,zero,zero,zero,ymm0[24],zero,zero
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm0[2,3,0,1]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero,zero,zero,zero
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero,zero
 ; AVX512DQ-BW-NEXT:    vpor %ymm4, %ymm6, %ymm4
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm1[5],zero,zero,zero,zero,zero,zero,ymm1[6],zero,zero,zero,zero,zero,zero,zero,ymm1[23],zero,zero,zero,zero,zero,zero,ymm1[24],zero,zero,zero,zero,zero,zero
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm1[2,3,0,1]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm7[5],zero,zero,zero,zero,zero,zero,ymm7[6],zero,zero,zero,zero,zero,ymm7[23],zero,zero,zero,zero,zero,zero,ymm7[24],zero,zero,zero,zero,zero,zero,ymm7[25]
-; AVX512DQ-BW-NEXT:    vpor %ymm7, %ymm6, %ymm6
-; AVX512DQ-BW-NEXT:    movl $202911840, %ecx # imm = 0xC183060
-; AVX512DQ-BW-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm4, %ymm6 {%k1}
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,0,2]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[0,8],zero,zero,zero,zero,zero,ymm1[1,9],zero,zero,zero,zero,zero,ymm1[2,10],zero,zero,zero,zero,zero,ymm1[19,27],zero,zero,zero,zero,zero,ymm1[20,28],zero,zero
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,0,2]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm0 = zero,zero,ymm0[0,8],zero,zero,zero,zero,zero,ymm0[1,9],zero,zero,zero,zero,zero,ymm0[18,26],zero,zero,zero,zero,zero,ymm0[19,27],zero,zero,zero,zero,zero,ymm0[20,28]
-; AVX512DQ-BW-NEXT:    vpor %ymm1, %ymm0, %ymm0
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,0,8,u,u,u,u,u,1,9,u,u,u,u,u,18,26,u,u,u,u,u,19,27,u,u,u,u,u]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm4 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512DQ-BW-NEXT:    vpermw %zmm1, %zmm4, %zmm1
+; AVX512DQ-BW-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
+; AVX512DQ-BW-NEXT:    kmovq %rcx, %k1
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm0 {%k1}
 ; AVX512DQ-BW-NEXT:    movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm2, %zmm0 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm0, %zmm2 {%k1}
 ; AVX512DQ-BW-NEXT:    vmovdqa %xmm5, 96(%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, (%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, (%rax)
 ; AVX512DQ-BW-NEXT:    vmovdqa %ymm3, 64(%rax)
 ; AVX512DQ-BW-NEXT:    vzeroupper
 ; AVX512DQ-BW-NEXT:    retq
@@ -2561,16 +2561,16 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP:       # %bb.0:
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %xmm2
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%rcx), %ymm0, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%rsi), %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
 ; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
 ; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, (%r10), %zmm2, %zmm2
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm1[1,3,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm0[1,3,2,3]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm3 = zero,zero,xmm3[6,14],zero,zero,zero,zero,zero,xmm3[7,15],zero,zero,zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[1,3,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm1[1,3,2,3]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = zero,zero,zero,zero,xmm4[6,14],zero,zero,zero,zero,zero,xmm4[7,15],zero,zero,zero
 ; AVX512DQ-BW-FCP-NEXT:    vpor %xmm3, %xmm4, %xmm3
 ; AVX512DQ-BW-FCP-NEXT:    vextracti128 $1, %ymm2, %xmm4
@@ -2590,38 +2590,38 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    movl $67637280, %ecx # imm = 0x4081020
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm4, %ymm6 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[1,3,3,1]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm1[1,3,3,1]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = zero,ymm4[1,9],zero,zero,zero,zero,zero,ymm4[2,10],zero,zero,zero,zero,zero,ymm4[3,19],zero,zero,zero,zero,zero,ymm4[28,20],zero,zero,zero,zero,zero,ymm4[29,21],zero
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm1[3,1,1,3]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm0[3,1,1,3]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm7[1],zero,zero,zero,zero,zero,ymm7[10,2],zero,zero,zero,zero,zero,ymm7[11,3],zero,zero,zero,zero,zero,ymm7[20,28],zero,zero,zero,zero,zero,ymm7[21,29],zero,zero,zero
 ; AVX512DQ-BW-FCP-NEXT:    vpor %ymm4, %ymm7, %ymm4
 ; AVX512DQ-BW-FCP-NEXT:    movl $-2029118408, %ecx # imm = 0x870E1C38
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm6, %ymm4 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm3, %zmm4, %zmm4
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm6 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,50,50,50,52,50,50,50,52,51,51,51,51,50,50,50,52]
-; AVX512DQ-BW-FCP-NEXT:    vpermi2w %zmm2, %zmm5, %zmm6
-; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm5 = [1,5,2,6,1,5,2,6]
-; AVX512DQ-BW-FCP-NEXT:    # ymm5 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm2, %ymm5, %ymm7
+; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [1,5,2,6,1,5,2,6]
+; AVX512DQ-BW-FCP-NEXT:    # ymm6 = mem[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm0, %ymm6, %ymm7
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,0,2]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[0,8],zero,zero,zero,zero,zero,zmm0[1,9],zero,zero,zero,zero,zero,zmm0[2,10],zero,zero,zero,zero,zero,zmm0[19,27],zero,zero,zero,zero,zero,zmm0[20,28],zero,zero,zero,zero,zero,zmm0[33,37],zero,zero,zero,zero,zero,zmm0[34,38],zero,zero,zero,zero,zero,zmm0[51,55],zero,zero,zero,zero,zero,zmm0[56,60],zero,zero,zero,zero,zero,zmm0[57]
+; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm1, %ymm6, %ymm7
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,0,2]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm1, %zmm1
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zero,zero,zmm1[0,8],zero,zero,zero,zero,zero,zmm1[1,9],zero,zero,zero,zero,zero,zmm1[18,26],zero,zero,zero,zero,zero,zmm1[19,27],zero,zero,zero,zero,zero,zmm1[20,28],zero,zero,zero,zero,zero,zmm1[33,37],zero,zero,zero,zero,zero,zmm1[34,38],zero,zero,zero,zero,zero,zmm1[51,55],zero,zero,zero,zero,zero,zmm1[56,60],zero,zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vporq %zmm0, %zmm1, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm2, %ymm6, %ymm1
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,0,2]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm2, %zmm2
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,0,8,u,u,u,u,u,1,9,u,u,u,u,u,18,26,u,u,u,u,u,19,27,u,u,u,u,u,32,36,u,u,u,u,u,33,37,u,u,u,u,u,34,38,u,u,u,u,u,51,55,u,u,u,u,u,56,60,u,u]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,0,8,u,u,u,u,u,1,9,u,u,u,u,u,18,26,u,u,u,u,u,19,27,u,u,u,u,u,32,36,u,u,u,u,u,33,37,u,u,u,u,u,34,38,u,u,u,u,u,51,55,u,u,u,u,u,56,60,u,u]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm5, %zmm2, %zmm2
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm2 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm1, %ymm5, %ymm6
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,0,2]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm1, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[0,8],zero,zero,zero,zero,zero,zmm1[1,9],zero,zero,zero,zero,zero,zmm1[2,10],zero,zero,zero,zero,zero,zmm1[19,27],zero,zero,zero,zero,zero,zmm1[20,28],zero,zero,zero,zero,zero,zmm1[33,37],zero,zero,zero,zero,zero,zmm1[34,38],zero,zero,zero,zero,zero,zmm1[51,55],zero,zero,zero,zero,zero,zmm1[56,60],zero,zero,zero,zero,zero,zmm1[57]
-; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm0, %ymm5, %ymm5
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,0,2]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm0
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[0,8],zero,zero,zero,zero,zero,zmm0[1,9],zero,zero,zero,zero,zero,zmm0[18,26],zero,zero,zero,zero,zero,zmm0[19,27],zero,zero,zero,zero,zero,zmm0[20,28],zero,zero,zero,zero,zero,zmm0[33,37],zero,zero,zero,zero,zero,zmm0[34,38],zero,zero,zero,zero,zero,zmm0[51,55],zero,zero,zero,zero,zero,zmm0[56,60],zero,zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vporq %zmm1, %zmm0, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm1 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm0 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm0 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm3, 96(%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm4, 64(%rax)
@@ -4682,77 +4682,77 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW:       # %bb.0:
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512BW-NEXT:    vmovdqa (%rdi), %ymm4
-; AVX512BW-NEXT:    vmovdqa (%rsi), %ymm2
-; AVX512BW-NEXT:    vmovdqa (%rdx), %ymm1
-; AVX512BW-NEXT:    vmovdqa (%rcx), %ymm3
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm1[0,1,0,1,14],zero,ymm1[14,15,0,1,14,15],zero,ymm1[13,14,15,16,17,16],zero,ymm1[30,31,30,31,16,17],zero,ymm1[31,28,29,30,31]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm5 = zero,zero,zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero
-; AVX512BW-NEXT:    vpor %ymm0, %ymm5, %ymm0
-; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm8
+; AVX512BW-NEXT:    vmovdqa (%rdi), %ymm5
+; AVX512BW-NEXT:    vmovdqa (%rsi), %ymm3
+; AVX512BW-NEXT:    vmovdqa (%rdx), %ymm2
+; AVX512BW-NEXT:    vmovdqa (%rcx), %ymm4
+; AVX512BW-NEXT:    vmovdqa (%r10), %ymm0
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm2[0,1,0,1,14],zero,ymm2[14,15,0,1,14,15],zero,ymm2[13,14,15,16,17,16],zero,ymm2[30,31,30,31,16,17],zero,ymm2[31,28,29,30,31]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,ymm4[14],zero,zero,zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,zero
+; AVX512BW-NEXT:    vpor %ymm1, %ymm6, %ymm1
+; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm9
 ; AVX512BW-NEXT:    vmovdqa (%rsi), %xmm10
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm5, %zmm5
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,zero,zero,ymm2[18]
-; AVX512BW-NEXT:    vpor %ymm0, %ymm6, %ymm0
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm6, %zmm6
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm5[0,1,14],zero,ymm5[12,13,0,1,14,15],zero,ymm5[3,12,13,2,3,16],zero,ymm5[30,31,28,29,16,17],zero,ymm5[31,18,19,28,29,18],zero
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero,zero,ymm3[18]
+; AVX512BW-NEXT:    vpor %ymm1, %ymm7, %ymm1
 ; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm12
 ; AVX512BW-NEXT:    vmovdqa (%rcx), %xmm14
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm6, %zmm0
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm7, %zmm1
 ; AVX512BW-NEXT:    movabsq $435749858791416001, %rcx # imm = 0x60C1830183060C1
 ; AVX512BW-NEXT:    kmovq %rcx, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm5, %zmm0 {%k1}
-; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm7 = ymm5[13],zero,zero,zero,zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero
+; AVX512BW-NEXT:    vmovdqu8 %zmm6, %zmm1 {%k1}
 ; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm9 = zero,zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero,zero
-; AVX512BW-NEXT:    vpor %ymm7, %ymm9, %ymm7
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm8 = ymm6[13],zero,zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero
+; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm11 = zero,zero,zero,zero,zero,zero,ymm7[14],zero,zero,zero,zero,zero,zero,ymm7[15],zero,zero,zero,zero,zero,zero,ymm7[16],zero,zero,zero,zero,zero,zero,ymm7[17],zero,zero,zero,zero
+; AVX512BW-NEXT:    vpor %ymm8, %ymm11, %ymm8
 ; AVX512BW-NEXT:    vmovdqa (%r9), %xmm11
 ; AVX512BW-NEXT:    vmovdqa (%r8), %xmm13
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm9 = 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:    vpshufb {{.*#+}} xmm9 = xmm9[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm9, %zmm9
-; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm15 = 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:    vpshufb {{.*#+}} xmm15 = xmm15[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm8, %zmm15, %zmm8
 ; AVX512BW-NEXT:    vbroadcasti128 {{.*#+}} ymm15 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
 ; AVX512BW-NEXT:    # ymm15 = mem[0,1,0,1]
-; AVX512BW-NEXT:    vpermw %ymm7, %ymm15, %ymm15
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm16 = ymm7[12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512BW-NEXT:    vpermw %ymm0, %ymm15, %ymm15
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm16 = ymm0[12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
 ; AVX512BW-NEXT:    vinserti64x4 $1, %ymm16, %zmm15, %zmm15
 ; AVX512BW-NEXT:    movabsq $2323999253380730912, %rcx # imm = 0x2040810204081020
 ; AVX512BW-NEXT:    kmovq %rcx, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm15, %zmm9 {%k1}
+; AVX512BW-NEXT:    vmovdqu8 %zmm15, %zmm8 {%k1}
 ; AVX512BW-NEXT:    movabsq $4066998693416279096, %rcx # imm = 0x3870E1C3870E1C38
 ; AVX512BW-NEXT:    kmovq %rcx, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm9, %zmm0 {%k1}
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm9
-; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm9 = zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm9[18,19,20,21],zero,zmm9[19],zero,zmm9[25,26,27,22],zero,zmm9[20],zero,zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm9[55],zero,zero,zero,zero,zmm9[58],zero,zmm9[56],zero,zero,zero,zero,zmm9[59],zero
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm15
+; AVX512BW-NEXT:    vmovdqu8 %zmm8, %zmm1 {%k1}
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm8
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm8 = zmm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm8[18,19,20,21],zero,zmm8[19],zero,zmm8[25,26,27,22],zero,zmm8[20],zero,zmm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm8[55],zero,zero,zero,zero,zmm8[58],zero,zmm8[56],zero,zero,zero,zero,zmm8[59],zero
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm15
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm15 = zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm15[18],zero,zero,zero,zero,zmm15[21],zero,zmm15[19],zero,zero,zero,zero,zmm15[22],zero,zmm15[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm15[55],zero,zero,zero,zero,zmm15[58],zero,zmm15[56],zero,zero,zero,zero,zmm15[59],zero,zmm15[57]
-; AVX512BW-NEXT:    vporq %zmm9, %zmm15, %zmm9
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
-; AVX512BW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512BW-NEXT:    vporq %zmm8, %zmm15, %zmm8
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm5[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
 ; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm15 = ymm15[0,0,1,1,4,4,5,5]
 ; AVX512BW-NEXT:    movl $676341840, %ecx # imm = 0x28502850
 ; AVX512BW-NEXT:    kmovd %ecx, %k1
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm15 {%k1} = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm16 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm1[23],zero,ymm1[21,22,23,26],zero,ymm1[24],zero,ymm1[28,29,26,27]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm17 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm15 {%k1} = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm16 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm2[23],zero,ymm2[21,22,23,26],zero,ymm2[24],zero,ymm2[28,29,26,27]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm17 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[25],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero,zero,zero
 ; AVX512BW-NEXT:    vporq %ymm16, %ymm17, %ymm16
 ; AVX512BW-NEXT:    vinserti64x4 $1, %ymm16, %zmm15, %zmm15
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm15 = zmm15[2,3,2,3,6,7,6,7]
 ; AVX512BW-NEXT:    movabsq $-9005497107459067808, %rcx # imm = 0x83060C180C183060
 ; AVX512BW-NEXT:    kmovq %rcx, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm15, %zmm9 {%k2}
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512BW-NEXT:    vpermw %zmm7, %zmm15, %zmm15
-; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm16 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm6[18],zero,zmm6[20,21,20,21],zero,zmm6[19],zero,zmm6[19,20,21,22],zero,zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm6[55],zero,zmm6[55,56,57,58],zero,zmm6[56],zero,zmm6[62,63]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm17 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm5[20],zero,zmm5[18],zero,zero,zero,zero,zmm5[21],zero,zmm5[19],zero,zero,zero,zero,zmm5[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm5[57],zero,zmm5[55],zero,zero,zero,zero,zmm5[58],zero,zmm5[56],zero,zero
+; AVX512BW-NEXT:    vmovdqu8 %zmm15, %zmm8 {%k2}
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512BW-NEXT:    vpermw %zmm0, %zmm15, %zmm15
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm16 = zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm7[18],zero,zmm7[20,21,20,21],zero,zmm7[19],zero,zmm7[19,20,21,22],zero,zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm7[55],zero,zmm7[55,56,57,58],zero,zmm7[56],zero,zmm7[62,63]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm17 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm6[20],zero,zmm6[18],zero,zero,zero,zero,zmm6[21],zero,zmm6[19],zero,zero,zero,zero,zmm6[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm6[57],zero,zmm6[55],zero,zero,zero,zero,zmm6[58],zero,zmm6[56],zero,zero
 ; AVX512BW-NEXT:    vporq %zmm16, %zmm17, %zmm16
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
 ; AVX512BW-NEXT:    movabsq $1161999626690365456, %rcx # imm = 0x1020408102040810
@@ -4760,7 +4760,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vmovdqu8 %zmm15, %zmm16 {%k2}
 ; AVX512BW-NEXT:    movabsq $2033499346708139548, %rcx # imm = 0x1C3870E1C3870E1C
 ; AVX512BW-NEXT:    kmovq %rcx, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm16, %zmm9 {%k2}
+; AVX512BW-NEXT:    vmovdqu8 %zmm16, %zmm8 {%k2}
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm15 = xmm14[u,u,u],zero,xmm14[7],zero,xmm14[5,u,u,u],zero,xmm14[8],zero,xmm14[6,u,u]
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm16 = xmm12[u,u,u,7],zero,xmm12[5],zero,xmm12[u,u,u,8],zero,xmm12[6],zero,xmm12[u,u]
 ; AVX512BW-NEXT:    vporq %xmm15, %xmm16, %xmm15
@@ -4769,15 +4769,15 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vinserti32x4 $2, %xmm15, %zmm12, %zmm12
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm12 = zmm12[0,1,0,1,4,5,4,5]
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm14 = xmm10[u],zero,xmm10[7],zero,xmm10[5,u,u,u],zero,xmm10[8],zero,xmm10[6,u,u,u],zero
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm15 = xmm8[u,7],zero,xmm8[5],zero,xmm8[u,u,u,8],zero,xmm8[6],zero,xmm8[u,u,u,9]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm15 = xmm9[u,7],zero,xmm9[5],zero,xmm9[u,u,u,8],zero,xmm9[6],zero,xmm9[u,u,u,9]
 ; AVX512BW-NEXT:    vpor %xmm14, %xmm15, %xmm14
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3],xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm8 = xmm8[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512BW-NEXT:    vinserti32x4 $2, %xmm14, %zmm8, %zmm8
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[0,1,0,1,4,5,4,5]
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm9 = xmm9[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512BW-NEXT:    vinserti32x4 $2, %xmm14, %zmm9, %zmm9
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm9 = zmm9[0,1,0,1,4,5,4,5]
 ; AVX512BW-NEXT:    movabsq $871499720017774092, %rcx # imm = 0xC183060C183060C
 ; AVX512BW-NEXT:    kmovq %rcx, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm12, %zmm8 {%k2}
+; AVX512BW-NEXT:    vmovdqu8 %zmm12, %zmm9 {%k2}
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm10 = zero,xmm11[4,u,u,u],zero,xmm11[7],zero,xmm11[5,u,u,u],zero,xmm11[8],zero,xmm11[6]
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm12 = xmm13[4],zero,xmm13[u,u,u,7],zero,xmm13[5],zero,xmm13[u,u,u,8],zero,xmm13[6],zero
 ; AVX512BW-NEXT:    vpor %xmm10, %xmm12, %xmm10
@@ -4785,45 +4785,45 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm11 = xmm11[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
 ; AVX512BW-NEXT:    vinserti32x4 $2, %xmm10, %zmm11, %zmm10
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm10 = zmm10[0,1,0,1,4,5,4,5]
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm11 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512BW-NEXT:    vpermw %zmm7, %zmm11, %zmm11
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm11 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512BW-NEXT:    vpermw %zmm0, %zmm11, %zmm11
 ; AVX512BW-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
 ; AVX512BW-NEXT:    kmovq %rcx, %k2
 ; AVX512BW-NEXT:    vmovdqu8 %zmm11, %zmm10 {%k2}
 ; AVX512BW-NEXT:    movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
 ; AVX512BW-NEXT:    kmovq %rcx, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm10, %zmm8 {%k2}
-; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm4 = ymm4[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[2,2,3,3,6,6,7,7]
+; AVX512BW-NEXT:    vmovdqu8 %zmm10, %zmm9 {%k2}
+; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[2,2,3,3,6,6,7,7]
 ; AVX512BW-NEXT:    movl $338170920, %ecx # imm = 0x14281428
 ; AVX512BW-NEXT:    kmovd %ecx, %k2
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm4 {%k2} = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u,u,u]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm5 {%k2} = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u,u,u]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm3 = ymm5[2,3,2,3]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
+; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,3,3,4,6,7,7]
+; AVX512BW-NEXT:    vmovdqu8 %ymm2, %ymm4 {%k1}
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm2 = ymm4[2,3,2,3]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
-; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,3,3,4,6,7,7]
-; AVX512BW-NEXT:    vmovdqu8 %ymm1, %ymm3 {%k1}
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm1 = ymm3[2,3,2,3]
 ; AVX512BW-NEXT:    movl $101455920, %ecx # imm = 0x60C1830
 ; AVX512BW-NEXT:    kmovd %ecx, %k1
-; AVX512BW-NEXT:    vmovdqu8 %ymm2, %ymm1 {%k1}
-; AVX512BW-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
-; AVX512BW-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512BW-NEXT:    vpermw %ymm7, %ymm2, %ymm2
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
+; AVX512BW-NEXT:    vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512BW-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
+; AVX512BW-NEXT:    # ymm3 = mem[0,1,0,1]
+; AVX512BW-NEXT:    vpermw %ymm0, %ymm3, %ymm0
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[27],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29]
 ; AVX512BW-NEXT:    vpor %ymm3, %ymm4, %ymm3
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
 ; AVX512BW-NEXT:    movl $-2130574328, %ecx # imm = 0x81020408
 ; AVX512BW-NEXT:    kmovd %ecx, %k1
-; AVX512BW-NEXT:    vmovdqu8 %ymm2, %ymm3 {%k1}
+; AVX512BW-NEXT:    vmovdqu8 %ymm0, %ymm3 {%k1}
 ; AVX512BW-NEXT:    movl $-507279602, %ecx # imm = 0xE1C3870E
 ; AVX512BW-NEXT:    kmovd %ecx, %k1
-; AVX512BW-NEXT:    vmovdqu8 %ymm3, %ymm1 {%k1}
-; AVX512BW-NEXT:    vmovdqa %ymm1, 192(%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm8, (%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm9, 128(%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm0, 64(%rax)
+; AVX512BW-NEXT:    vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512BW-NEXT:    vmovdqa %ymm2, 192(%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm9, (%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm8, 128(%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm1, 64(%rax)
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
 ;
@@ -4831,73 +4831,73 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP:       # %bb.0:
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm3
-; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %ymm4
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %ymm1
-; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %ymm2
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm1[0,1,0,1,14],zero,ymm1[14,15,0,1,14,15],zero,ymm1[13,14,15,16,17,16],zero,ymm1[30,31,30,31,16,17],zero,ymm1[31,28,29,30,31]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,zero
-; AVX512BW-FCP-NEXT:    vpor %ymm0, %ymm5, %ymm0
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm4
+; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %ymm5
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %ymm2
+; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %ymm3
+; AVX512BW-FCP-NEXT:    vmovdqa (%r10), %ymm0
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm2[0,1,0,1,14],zero,ymm2[14,15,0,1,14,15],zero,ymm2[13,14,15,16,17,16],zero,ymm2[30,31,30,31,16,17],zero,ymm2[31,28,29,30,31]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero
+; AVX512BW-FCP-NEXT:    vpor %ymm1, %ymm6, %ymm1
 ; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm8
 ; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %xmm9
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm5, %zmm5
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm3[0,1,14],zero,ymm3[12,13,0,1,14,15],zero,ymm3[3,12,13,2,3,16],zero,ymm3[30,31,28,29,16,17],zero,ymm3[31,18,19,28,29,18],zero
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm4[14],zero,zero,zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,zero,zero,ymm4[18]
-; AVX512BW-FCP-NEXT:    vpor %ymm0, %ymm6, %ymm0
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm6, %zmm6
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero,zero,zero,zero,ymm5[18]
+; AVX512BW-FCP-NEXT:    vpor %ymm1, %ymm7, %ymm1
 ; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm12
 ; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %xmm14
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm6, %zmm0
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm7, %zmm1
 ; AVX512BW-FCP-NEXT:    movabsq $435749858791416001, %rcx # imm = 0x60C1830183060C1
 ; AVX512BW-FCP-NEXT:    kmovq %rcx, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm5, %zmm0 {%k1}
-; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm5[13],zero,zero,zero,zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm1 {%k1}
 ; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = zero,zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero,zero
-; AVX512BW-FCP-NEXT:    vpor %ymm7, %ymm10, %ymm7
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm6[13],zero,zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero
+; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = zero,zero,zero,zero,zero,zero,ymm7[14],zero,zero,zero,zero,zero,zero,ymm7[15],zero,zero,zero,zero,zero,zero,ymm7[16],zero,zero,zero,zero,zero,zero,ymm7[17],zero,zero,zero,zero
+; AVX512BW-FCP-NEXT:    vpor %ymm10, %ymm11, %ymm10
 ; AVX512BW-FCP-NEXT:    vmovdqa (%r9), %xmm11
 ; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %xmm13
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm10 = 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:    vpshufb {{.*#+}} xmm10 = xmm10[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm10, %zmm10
-; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm15 = 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:    vpshufb {{.*#+}} xmm15 = xmm15[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm15, %zmm10
 ; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm15 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
 ; AVX512BW-FCP-NEXT:    # ymm15 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpermw %ymm7, %ymm15, %ymm15
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm16 = ymm7[12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512BW-FCP-NEXT:    vpermw %ymm0, %ymm15, %ymm15
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm16 = ymm0[12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
 ; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm16, %zmm15, %zmm15
 ; AVX512BW-FCP-NEXT:    movabsq $2323999253380730912, %rcx # imm = 0x2040810204081020
 ; AVX512BW-FCP-NEXT:    kmovq %rcx, %k1
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm15, %zmm10 {%k1}
 ; AVX512BW-FCP-NEXT:    movabsq $4066998693416279096, %rcx # imm = 0x3870E1C3870E1C38
 ; AVX512BW-FCP-NEXT:    kmovq %rcx, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm10, %zmm0 {%k1}
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm3, %zmm10
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm10, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm10
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,zmm10[19],zero,zmm10[21,20,21,22],zero,zmm10[20],zero,zmm10[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57],zero,zmm10[55],zero,zmm10[53,54,55,58],zero,zmm10[56],zero,zmm10[60,61,58,59]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm15
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm5, %zmm15
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm15 = zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm15[21],zero,zmm15[19],zero,zero,zero,zero,zmm15[22],zero,zmm15[20],zero,zero,zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm15[57],zero,zmm15[55],zero,zero,zero,zero,zmm15[58],zero,zmm15[56],zero,zero,zero,zero
 ; AVX512BW-FCP-NEXT:    vporq %zmm10, %zmm15, %zmm10
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm15 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm1, %zmm10
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm2, %zmm10
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm10[18,19,20,21],zero,zmm10[19],zero,zmm10[25,26,27,22],zero,zmm10[20],zero,zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm10[55],zero,zero,zero,zero,zmm10[58],zero,zmm10[56],zero,zero,zero,zero,zmm10[59],zero
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm16
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm16
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm16 = zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[18],zero,zero,zero,zero,zmm16[21],zero,zmm16[19],zero,zero,zero,zero,zmm16[22],zero,zmm16[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[55],zero,zero,zero,zero,zmm16[58],zero,zmm16[56],zero,zero,zero,zero,zmm16[59],zero,zmm16[57]
 ; AVX512BW-FCP-NEXT:    vporq %zmm10, %zmm16, %zmm10
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
 ; AVX512BW-FCP-NEXT:    movabsq $-9005497107459067808, %rcx # imm = 0x83060C180C183060
 ; AVX512BW-FCP-NEXT:    kmovq %rcx, %k1
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm15, %zmm10 {%k1}
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512BW-FCP-NEXT:    vpermw %zmm7, %zmm15, %zmm15
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm16 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm6[18],zero,zmm6[20,21,20,21],zero,zmm6[19],zero,zmm6[19,20,21,22],zero,zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm6[55],zero,zmm6[55,56,57,58],zero,zmm6[56],zero,zmm6[62,63]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm17 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm5[20],zero,zmm5[18],zero,zero,zero,zero,zmm5[21],zero,zmm5[19],zero,zero,zero,zero,zmm5[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm5[57],zero,zmm5[55],zero,zero,zero,zero,zmm5[58],zero,zmm5[56],zero,zero
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512BW-FCP-NEXT:    vpermw %zmm0, %zmm15, %zmm15
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm16 = zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm7[18],zero,zmm7[20,21,20,21],zero,zmm7[19],zero,zmm7[19,20,21,22],zero,zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm7[55],zero,zmm7[55,56,57,58],zero,zmm7[56],zero,zmm7[62,63]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm17 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm6[20],zero,zmm6[18],zero,zero,zero,zero,zmm6[21],zero,zmm6[19],zero,zero,zero,zero,zmm6[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm6[57],zero,zmm6[55],zero,zero,zero,zero,zmm6[58],zero,zmm6[56],zero,zero
 ; AVX512BW-FCP-NEXT:    vporq %zmm16, %zmm17, %zmm16
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
 ; AVX512BW-FCP-NEXT:    movabsq $1161999626690365456, %rcx # imm = 0x1020408102040810
@@ -4930,42 +4930,42 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm11 = xmm11[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
 ; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm11, %zmm9
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm9 = zmm9[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm11 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512BW-FCP-NEXT:    vpermw %zmm7, %zmm11, %zmm11
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm11 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512BW-FCP-NEXT:    vpermw %zmm0, %zmm11, %zmm11
 ; AVX512BW-FCP-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
 ; AVX512BW-FCP-NEXT:    kmovq %rcx, %k1
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm11, %zmm9 {%k1}
 ; AVX512BW-FCP-NEXT:    movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
 ; AVX512BW-FCP-NEXT:    kmovq %rcx, %k1
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm9, %zmm8 {%k1}
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
-; AVX512BW-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
-; AVX512BW-FCP-NEXT:    vpor %ymm2, %ymm1, %ymm1
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm4[28],zero,ymm4[30,31,30,31],zero,ymm4[29],zero,ymm4[31,28,29]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero
+; AVX512BW-FCP-NEXT:    vpor %ymm4, %ymm5, %ymm4
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29],zero
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
+; AVX512BW-FCP-NEXT:    vpor %ymm3, %ymm2, %ymm2
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
 ; AVX512BW-FCP-NEXT:    movl $101455920, %ecx # imm = 0x60C1830
 ; AVX512BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm3, %ymm1 {%k1}
-; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
-; AVX512BW-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpermw %ymm7, %ymm2, %ymm2
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
+; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm4, %ymm2 {%k1}
+; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
+; AVX512BW-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vpermw %ymm0, %ymm3, %ymm0
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[27],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29]
 ; AVX512BW-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
 ; AVX512BW-FCP-NEXT:    movl $-2130574328, %ecx # imm = 0x81020408
 ; AVX512BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm2, %ymm3 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm0, %ymm3 {%k1}
 ; AVX512BW-FCP-NEXT:    movl $-507279602, %ecx # imm = 0xE1C3870E
 ; AVX512BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm3, %ymm1 {%k1}
-; AVX512BW-FCP-NEXT:    vmovdqa %ymm1, 192(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqa %ymm2, 192(%rax)
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, (%rax)
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, 128(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, 64(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rax)
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
 ;
@@ -4973,77 +4973,77 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW:       # %bb.0:
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %ymm4
-; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %ymm2
-; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %ymm1
-; AVX512DQ-BW-NEXT:    vmovdqa (%rcx), %ymm3
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm1[0,1,0,1,14],zero,ymm1[14,15,0,1,14,15],zero,ymm1[13,14,15,16,17,16],zero,ymm1[30,31,30,31,16,17],zero,ymm1[31,28,29,30,31]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm5 = zero,zero,zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero
-; AVX512DQ-BW-NEXT:    vpor %ymm0, %ymm5, %ymm0
-; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm8
+; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %ymm5
+; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %ymm3
+; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %ymm2
+; AVX512DQ-BW-NEXT:    vmovdqa (%rcx), %ymm4
+; AVX512DQ-BW-NEXT:    vmovdqa (%r10), %ymm0
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm2[0,1,0,1,14],zero,ymm2[14,15,0,1,14,15],zero,ymm2[13,14,15,16,17,16],zero,ymm2[30,31,30,31,16,17],zero,ymm2[31,28,29,30,31]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,ymm4[14],zero,zero,zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,zero
+; AVX512DQ-BW-NEXT:    vpor %ymm1, %ymm6, %ymm1
+; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm9
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %xmm10
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm5, %zmm5
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,zero,zero,ymm2[18]
-; AVX512DQ-BW-NEXT:    vpor %ymm0, %ymm6, %ymm0
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm6, %zmm6
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm5[0,1,14],zero,ymm5[12,13,0,1,14,15],zero,ymm5[3,12,13,2,3,16],zero,ymm5[30,31,28,29,16,17],zero,ymm5[31,18,19,28,29,18],zero
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero,zero,ymm3[18]
+; AVX512DQ-BW-NEXT:    vpor %ymm1, %ymm7, %ymm1
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm12
 ; AVX512DQ-BW-NEXT:    vmovdqa (%rcx), %xmm14
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm6, %zmm0
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm7, %zmm1
 ; AVX512DQ-BW-NEXT:    movabsq $435749858791416001, %rcx # imm = 0x60C1830183060C1
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm5, %zmm0 {%k1}
-; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm7 = ymm5[13],zero,zero,zero,zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm6, %zmm1 {%k1}
 ; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm9 = zero,zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero,zero
-; AVX512DQ-BW-NEXT:    vpor %ymm7, %ymm9, %ymm7
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm8 = ymm6[13],zero,zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero
+; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm11 = zero,zero,zero,zero,zero,zero,ymm7[14],zero,zero,zero,zero,zero,zero,ymm7[15],zero,zero,zero,zero,zero,zero,ymm7[16],zero,zero,zero,zero,zero,zero,ymm7[17],zero,zero,zero,zero
+; AVX512DQ-BW-NEXT:    vpor %ymm8, %ymm11, %ymm8
 ; AVX512DQ-BW-NEXT:    vmovdqa (%r9), %xmm11
 ; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm13
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm9 = 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:    vpshufb {{.*#+}} xmm9 = xmm9[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm9, %zmm9
-; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm15 = 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:    vpshufb {{.*#+}} xmm15 = xmm15[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm8, %zmm15, %zmm8
 ; AVX512DQ-BW-NEXT:    vbroadcasti128 {{.*#+}} ymm15 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
 ; AVX512DQ-BW-NEXT:    # ymm15 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vpermw %ymm7, %ymm15, %ymm15
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm16 = ymm7[12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512DQ-BW-NEXT:    vpermw %ymm0, %ymm15, %ymm15
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm16 = ymm0[12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
 ; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm16, %zmm15, %zmm15
 ; AVX512DQ-BW-NEXT:    movabsq $2323999253380730912, %rcx # imm = 0x2040810204081020
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm15, %zmm9 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm15, %zmm8 {%k1}
 ; AVX512DQ-BW-NEXT:    movabsq $4066998693416279096, %rcx # imm = 0x3870E1C3870E1C38
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm9, %zmm0 {%k1}
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm9
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm9 = zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm9[18,19,20,21],zero,zmm9[19],zero,zmm9[25,26,27,22],zero,zmm9[20],zero,zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm9[55],zero,zero,zero,zero,zmm9[58],zero,zmm9[56],zero,zero,zero,zero,zmm9[59],zero
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm15
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm8, %zmm1 {%k1}
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm8
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm8 = zmm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm8[18,19,20,21],zero,zmm8[19],zero,zmm8[25,26,27,22],zero,zmm8[20],zero,zmm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm8[55],zero,zero,zero,zero,zmm8[58],zero,zmm8[56],zero,zero,zero,zero,zmm8[59],zero
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm15
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm15 = zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm15[18],zero,zero,zero,zero,zmm15[21],zero,zmm15[19],zero,zero,zero,zero,zmm15[22],zero,zmm15[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm15[55],zero,zero,zero,zero,zmm15[58],zero,zmm15[56],zero,zero,zero,zero,zmm15[59],zero,zmm15[57]
-; AVX512DQ-BW-NEXT:    vporq %zmm9, %zmm15, %zmm9
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-BW-NEXT:    vporq %zmm8, %zmm15, %zmm8
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm5[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
 ; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm15 = ymm15[0,0,1,1,4,4,5,5]
 ; AVX512DQ-BW-NEXT:    movl $676341840, %ecx # imm = 0x28502850
 ; AVX512DQ-BW-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm15 {%k1} = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm16 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm1[23],zero,ymm1[21,22,23,26],zero,ymm1[24],zero,ymm1[28,29,26,27]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm17 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm15 {%k1} = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm16 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm2[23],zero,ymm2[21,22,23,26],zero,ymm2[24],zero,ymm2[28,29,26,27]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm17 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[25],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero,zero,zero
 ; AVX512DQ-BW-NEXT:    vporq %ymm16, %ymm17, %ymm16
 ; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm16, %zmm15, %zmm15
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm15 = zmm15[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-NEXT:    movabsq $-9005497107459067808, %rcx # imm = 0x83060C180C183060
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm15, %zmm9 {%k2}
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512DQ-BW-NEXT:    vpermw %zmm7, %zmm15, %zmm15
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm16 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm6[18],zero,zmm6[20,21,20,21],zero,zmm6[19],zero,zmm6[19,20,21,22],zero,zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm6[55],zero,zmm6[55,56,57,58],zero,zmm6[56],zero,zmm6[62,63]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm17 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm5[20],zero,zmm5[18],zero,zero,zero,zero,zmm5[21],zero,zmm5[19],zero,zero,zero,zero,zmm5[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm5[57],zero,zmm5[55],zero,zero,zero,zero,zmm5[58],zero,zmm5[56],zero,zero
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm15, %zmm8 {%k2}
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512DQ-BW-NEXT:    vpermw %zmm0, %zmm15, %zmm15
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm16 = zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm7[18],zero,zmm7[20,21,20,21],zero,zmm7[19],zero,zmm7[19,20,21,22],zero,zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm7[55],zero,zmm7[55,56,57,58],zero,zmm7[56],zero,zmm7[62,63]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm17 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm6[20],zero,zmm6[18],zero,zero,zero,zero,zmm6[21],zero,zmm6[19],zero,zero,zero,zero,zmm6[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm6[57],zero,zmm6[55],zero,zero,zero,zero,zmm6[58],zero,zmm6[56],zero,zero
 ; AVX512DQ-BW-NEXT:    vporq %zmm16, %zmm17, %zmm16
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-NEXT:    movabsq $1161999626690365456, %rcx # imm = 0x1020408102040810
@@ -5051,7 +5051,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm15, %zmm16 {%k2}
 ; AVX512DQ-BW-NEXT:    movabsq $2033499346708139548, %rcx # imm = 0x1C3870E1C3870E1C
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm16, %zmm9 {%k2}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm16, %zmm8 {%k2}
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm15 = xmm14[u,u,u],zero,xmm14[7],zero,xmm14[5,u,u,u],zero,xmm14[8],zero,xmm14[6,u,u]
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm16 = xmm12[u,u,u,7],zero,xmm12[5],zero,xmm12[u,u,u,8],zero,xmm12[6],zero,xmm12[u,u]
 ; AVX512DQ-BW-NEXT:    vporq %xmm15, %xmm16, %xmm15
@@ -5060,15 +5060,15 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm15, %zmm12, %zmm12
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm12 = zmm12[0,1,0,1,4,5,4,5]
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm14 = xmm10[u],zero,xmm10[7],zero,xmm10[5,u,u,u],zero,xmm10[8],zero,xmm10[6,u,u,u],zero
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm15 = xmm8[u,7],zero,xmm8[5],zero,xmm8[u,u,u,8],zero,xmm8[6],zero,xmm8[u,u,u,9]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm15 = xmm9[u,7],zero,xmm9[5],zero,xmm9[u,u,u,8],zero,xmm9[6],zero,xmm9[u,u,u,9]
 ; AVX512DQ-BW-NEXT:    vpor %xmm14, %xmm15, %xmm14
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3],xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm8 = xmm8[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm14, %zmm8, %zmm8
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm9 = xmm9[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm14, %zmm9, %zmm9
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm9 = zmm9[0,1,0,1,4,5,4,5]
 ; AVX512DQ-BW-NEXT:    movabsq $871499720017774092, %rcx # imm = 0xC183060C183060C
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm12, %zmm8 {%k2}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm12, %zmm9 {%k2}
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm10 = zero,xmm11[4,u,u,u],zero,xmm11[7],zero,xmm11[5,u,u,u],zero,xmm11[8],zero,xmm11[6]
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm12 = xmm13[4],zero,xmm13[u,u,u,7],zero,xmm13[5],zero,xmm13[u,u,u,8],zero,xmm13[6],zero
 ; AVX512DQ-BW-NEXT:    vpor %xmm10, %xmm12, %xmm10
@@ -5076,45 +5076,45 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm11 = xmm11[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
 ; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm10, %zmm11, %zmm10
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm10 = zmm10[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm11 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512DQ-BW-NEXT:    vpermw %zmm7, %zmm11, %zmm11
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm11 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512DQ-BW-NEXT:    vpermw %zmm0, %zmm11, %zmm11
 ; AVX512DQ-BW-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k2
 ; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm11, %zmm10 {%k2}
 ; AVX512DQ-BW-NEXT:    movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
 ; AVX512DQ-BW-NEXT:    kmovq %rcx, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm10, %zmm8 {%k2}
-; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm4 = ymm4[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[2,2,3,3,6,6,7,7]
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm10, %zmm9 {%k2}
+; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[2,2,3,3,6,6,7,7]
 ; AVX512DQ-BW-NEXT:    movl $338170920, %ecx # imm = 0x14281428
 ; AVX512DQ-BW-NEXT:    kmovd %ecx, %k2
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm4 {%k2} = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u,u,u]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm5 {%k2} = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u,u,u]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm3 = ymm5[2,3,2,3]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
+; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,3,3,4,6,7,7]
+; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm2, %ymm4 {%k1}
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm2 = ymm4[2,3,2,3]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
-; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,3,3,4,6,7,7]
-; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm1, %ymm3 {%k1}
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm1 = ymm3[2,3,2,3]
 ; AVX512DQ-BW-NEXT:    movl $101455920, %ecx # imm = 0x60C1830
 ; AVX512DQ-BW-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm2, %ymm1 {%k1}
-; AVX512DQ-BW-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
-; AVX512DQ-BW-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vpermw %ymm7, %ymm2, %ymm2
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
+; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512DQ-BW-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
+; AVX512DQ-BW-NEXT:    # ymm3 = mem[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vpermw %ymm0, %ymm3, %ymm0
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[27],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29]
 ; AVX512DQ-BW-NEXT:    vpor %ymm3, %ymm4, %ymm3
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
 ; AVX512DQ-BW-NEXT:    movl $-2130574328, %ecx # imm = 0x81020408
 ; AVX512DQ-BW-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm2, %ymm3 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm0, %ymm3 {%k1}
 ; AVX512DQ-BW-NEXT:    movl $-507279602, %ecx # imm = 0xE1C3870E
 ; AVX512DQ-BW-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm3, %ymm1 {%k1}
-; AVX512DQ-BW-NEXT:    vmovdqa %ymm1, 192(%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, (%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, 128(%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, 64(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqa %ymm2, 192(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, (%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, 128(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, 64(%rax)
 ; AVX512DQ-BW-NEXT:    vzeroupper
 ; AVX512DQ-BW-NEXT:    retq
 ;
@@ -5122,73 +5122,73 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP:       # %bb.0:
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %ymm4
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm1[0,1,0,1,14],zero,ymm1[14,15,0,1,14,15],zero,ymm1[13,14,15,16,17,16],zero,ymm1[30,31,30,31,16,17],zero,ymm1[31,28,29,30,31]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm0, %ymm5, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %ymm5
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %ymm3
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r10), %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm2[0,1,0,1,14],zero,ymm2[14,15,0,1,14,15],zero,ymm2[13,14,15,16,17,16],zero,ymm2[30,31,30,31,16,17],zero,ymm2[31,28,29,30,31]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm1, %ymm6, %ymm1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm8
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %xmm9
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm5, %zmm5
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm3[0,1,14],zero,ymm3[12,13,0,1,14,15],zero,ymm3[3,12,13,2,3,16],zero,ymm3[30,31,28,29,16,17],zero,ymm3[31,18,19,28,29,18],zero
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm4[14],zero,zero,zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,zero,zero,ymm4[18]
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm0, %ymm6, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm6, %zmm6
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero,zero,zero,zero,ymm5[18]
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm1, %ymm7, %ymm1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm12
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %xmm14
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm6, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm7, %zmm1
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $435749858791416001, %rcx # imm = 0x60C1830183060C1
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm5, %zmm0 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm5[13],zero,zero,zero,zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm1 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = zero,zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm7, %ymm10, %ymm7
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm6[13],zero,zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = zero,zero,zero,zero,zero,zero,ymm7[14],zero,zero,zero,zero,zero,zero,ymm7[15],zero,zero,zero,zero,zero,zero,ymm7[16],zero,zero,zero,zero,zero,zero,ymm7[17],zero,zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm10, %ymm11, %ymm10
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r9), %xmm11
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %xmm13
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm10 = 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:    vpshufb {{.*#+}} xmm10 = xmm10[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm10, %zmm10
-; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm15 = 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:    vpshufb {{.*#+}} xmm15 = xmm15[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm15, %zmm10
 ; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm15 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
 ; AVX512DQ-BW-FCP-NEXT:    # ymm15 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm7, %ymm15, %ymm15
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm16 = ymm7[12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm0, %ymm15, %ymm15
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm16 = ymm0[12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
 ; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm16, %zmm15, %zmm15
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $2323999253380730912, %rcx # imm = 0x2040810204081020
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm15, %zmm10 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $4066998693416279096, %rcx # imm = 0x3870E1C3870E1C38
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm10, %zmm0 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm3, %zmm10
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm10, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm10
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,zmm10[19],zero,zmm10[21,20,21,22],zero,zmm10[20],zero,zmm10[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57],zero,zmm10[55],zero,zmm10[53,54,55,58],zero,zmm10[56],zero,zmm10[60,61,58,59]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm15
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm5, %zmm15
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm15 = zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm15[21],zero,zmm15[19],zero,zero,zero,zero,zmm15[22],zero,zmm15[20],zero,zero,zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm15[57],zero,zmm15[55],zero,zero,zero,zero,zmm15[58],zero,zmm15[56],zero,zero,zero,zero
 ; AVX512DQ-BW-FCP-NEXT:    vporq %zmm10, %zmm15, %zmm10
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm15 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm1, %zmm10
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm2, %zmm10
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm10[18,19,20,21],zero,zmm10[19],zero,zmm10[25,26,27,22],zero,zmm10[20],zero,zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm10[55],zero,zero,zero,zero,zmm10[58],zero,zmm10[56],zero,zero,zero,zero,zmm10[59],zero
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm16
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm16
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm16 = zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[18],zero,zero,zero,zero,zmm16[21],zero,zmm16[19],zero,zero,zero,zero,zmm16[22],zero,zmm16[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[55],zero,zero,zero,zero,zmm16[58],zero,zmm16[56],zero,zero,zero,zero,zmm16[59],zero,zmm16[57]
 ; AVX512DQ-BW-FCP-NEXT:    vporq %zmm10, %zmm16, %zmm10
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $-9005497107459067808, %rcx # imm = 0x83060C180C183060
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm15, %zmm10 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm7, %zmm15, %zmm15
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm16 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm6[18],zero,zmm6[20,21,20,21],zero,zmm6[19],zero,zmm6[19,20,21,22],zero,zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm6[55],zero,zmm6[55,56,57,58],zero,zmm6[56],zero,zmm6[62,63]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm17 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm5[20],zero,zmm5[18],zero,zero,zero,zero,zmm5[21],zero,zmm5[19],zero,zero,zero,zero,zmm5[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm5[57],zero,zmm5[55],zero,zero,zero,zero,zmm5[58],zero,zmm5[56],zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm0, %zmm15, %zmm15
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm16 = zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm7[18],zero,zmm7[20,21,20,21],zero,zmm7[19],zero,zmm7[19,20,21,22],zero,zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm7[55],zero,zmm7[55,56,57,58],zero,zmm7[56],zero,zmm7[62,63]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm17 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm6[20],zero,zmm6[18],zero,zero,zero,zero,zmm6[21],zero,zmm6[19],zero,zero,zero,zero,zmm6[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm6[57],zero,zmm6[55],zero,zero,zero,zero,zmm6[58],zero,zmm6[56],zero,zero
 ; AVX512DQ-BW-FCP-NEXT:    vporq %zmm16, %zmm17, %zmm16
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $1161999626690365456, %rcx # imm = 0x1020408102040810
@@ -5221,42 +5221,42 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm11 = xmm11[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
 ; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm11, %zmm9
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm9 = zmm9[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm11 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm7, %zmm11, %zmm11
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm11 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm0, %zmm11, %zmm11
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm11, %zmm9 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm9, %zmm8 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm2, %ymm1, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm4[28],zero,ymm4[30,31,30,31],zero,ymm4[29],zero,ymm4[31,28,29]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm4, %ymm5, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29],zero
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm3, %ymm2, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
 ; AVX512DQ-BW-FCP-NEXT:    movl $101455920, %ecx # imm = 0x60C1830
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm3, %ymm1 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
-; AVX512DQ-BW-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm7, %ymm2, %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm4, %ymm2 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
+; AVX512DQ-BW-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm0, %ymm3, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[27],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29]
 ; AVX512DQ-BW-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
 ; AVX512DQ-BW-FCP-NEXT:    movl $-2130574328, %ecx # imm = 0x81020408
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm2, %ymm3 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm0, %ymm3 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    movl $-507279602, %ecx # imm = 0xE1C3870E
 ; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm3, %ymm1 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm1, 192(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm2, 192(%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, (%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, 128(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, 64(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
   %in.vec0 = load <32 x i8>, ptr %in.vecptr0, align 64
@@ -10052,1188 +10052,1186 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
 ; AVX512BW-LABEL: store_i8_stride7_vf64:
 ; AVX512BW:       # %bb.0:
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-NEXT:    vmovdqa (%rax), %ymm10
-; AVX512BW-NEXT:    vmovdqa 32(%rax), %ymm14
+; AVX512BW-NEXT:    vmovdqa (%rax), %ymm0
+; AVX512BW-NEXT:    vmovdqa64 32(%rax), %ymm23
 ; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm1 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX512BW-NEXT:    vpshufb %ymm1, %ymm10, %ymm0
-; AVX512BW-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
-; AVX512BW-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512BW-NEXT:    vpermw %ymm10, %ymm2, %ymm2
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm3
-; AVX512BW-NEXT:    vmovdqa64 (%r9), %ymm22
-; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm6 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
-; AVX512BW-NEXT:    vpshufb %ymm6, %ymm22, %ymm0
-; AVX512BW-NEXT:    vmovdqa64 (%r8), %ymm21
-; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm7 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512BW-NEXT:    vpshufb %ymm7, %ymm21, %ymm2
-; AVX512BW-NEXT:    vpor %ymm0, %ymm2, %ymm4
-; AVX512BW-NEXT:    vmovdqa (%r9), %xmm0
-; AVX512BW-NEXT:    vmovdqa (%r8), %xmm2
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm12
+; AVX512BW-NEXT:    vpshufb %ymm1, %ymm0, %ymm2
+; AVX512BW-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512BW-NEXT:    # ymm3 = mem[0,1,0,1]
+; AVX512BW-NEXT:    vpermw %ymm0, %ymm3, %ymm0
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
+; AVX512BW-NEXT:    vmovdqa64 (%r9), %ymm25
+; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm4 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
+; AVX512BW-NEXT:    vpshufb %ymm4, %ymm25, %ymm0
+; AVX512BW-NEXT:    vmovdqa64 (%r8), %ymm24
+; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
+; AVX512BW-NEXT:    vpshufb %ymm5, %ymm24, %ymm3
+; AVX512BW-NEXT:    vpor %ymm0, %ymm3, %ymm3
+; AVX512BW-NEXT:    vmovdqa64 (%r9), %xmm31
+; AVX512BW-NEXT:    vmovdqa (%r8), %xmm0
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm0[8],xmm31[8],xmm0[9],xmm31[9],xmm0[10],xmm31[10],xmm0[11],xmm31[11],xmm0[12],xmm31[12],xmm0[13],xmm31[13],xmm0[14],xmm31[14],xmm0[15],xmm31[15]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm3, %zmm6, %zmm14
 ; AVX512BW-NEXT:    movabsq $2323999253380730912, %r10 # imm = 0x2040810204081020
 ; AVX512BW-NEXT:    kmovq %r10, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm3, %zmm12 {%k1}
-; AVX512BW-NEXT:    vmovdqa64 (%rdi), %ymm19
-; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
-; AVX512BW-NEXT:    vpshufb %ymm3, %ymm19, %ymm4
-; AVX512BW-NEXT:    vmovdqa64 (%rsi), %ymm18
-; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm11 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512BW-NEXT:    vpshufb %ymm11, %ymm18, %ymm5
-; AVX512BW-NEXT:    vpor %ymm4, %ymm5, %ymm8
-; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm4
-; AVX512BW-NEXT:    vmovdqa (%rsi), %xmm5
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm9 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm9 = xmm9[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm8, %zmm9, %zmm13
-; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm15 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512BW-NEXT:    vpshufb %ymm15, %ymm8, %ymm16
-; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} ymm20 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512BW-NEXT:    vpshufb %ymm20, %ymm9, %ymm17
-; AVX512BW-NEXT:    vporq %ymm16, %ymm17, %ymm16
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm17 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm17 = xmm17[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[0,1,0,1]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm16, %zmm17, %zmm16
+; AVX512BW-NEXT:    vmovdqu8 %zmm2, %zmm14 {%k1}
+; AVX512BW-NEXT:    vmovdqa64 (%rdi), %ymm16
+; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
+; AVX512BW-NEXT:    vpshufb %ymm6, %ymm16, %ymm2
+; AVX512BW-NEXT:    vmovdqa (%rsi), %ymm13
+; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm9 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512BW-NEXT:    vpshufb %ymm9, %ymm13, %ymm3
+; AVX512BW-NEXT:    vpor %ymm2, %ymm3, %ymm7
+; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm2
+; AVX512BW-NEXT:    vmovdqa (%rsi), %xmm3
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm10 = xmm10[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm10, %zmm15
+; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} ymm17 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512BW-NEXT:    vpshufb %ymm17, %ymm10, %ymm18
+; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} ymm19 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512BW-NEXT:    vpshufb %ymm19, %ymm11, %ymm20
+; AVX512BW-NEXT:    vporq %ymm18, %ymm20, %ymm18
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm20 = xmm20[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm20 = ymm20[0,1,0,1]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm18, %zmm20, %zmm18
 ; AVX512BW-NEXT:    movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
 ; AVX512BW-NEXT:    kmovq %r10, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm16, %zmm13 {%k1}
+; AVX512BW-NEXT:    vmovdqu8 %zmm18, %zmm15 {%k1}
 ; AVX512BW-NEXT:    movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
 ; AVX512BW-NEXT:    kmovq %r10, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm12, %zmm13 {%k1}
-; AVX512BW-NEXT:    vmovdqa64 32(%rdx), %ymm16
-; AVX512BW-NEXT:    vpshufb %ymm15, %ymm16, %ymm12
-; AVX512BW-NEXT:    vmovdqa64 32(%rcx), %ymm17
-; AVX512BW-NEXT:    vpshufb %ymm20, %ymm17, %ymm15
-; AVX512BW-NEXT:    vpor %ymm12, %ymm15, %ymm12
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm15 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
-; AVX512BW-NEXT:    vpshufb %ymm15, %ymm16, %ymm15
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm20 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512BW-NEXT:    vpshufb %ymm20, %ymm17, %ymm20
-; AVX512BW-NEXT:    vporq %ymm15, %ymm20, %ymm15
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm15, %zmm12, %zmm12
-; AVX512BW-NEXT:    vmovdqa64 32(%rsi), %ymm25
-; AVX512BW-NEXT:    vmovdqa64 32(%rdi), %ymm24
-; AVX512BW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm24[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm15 = ymm15[0,0,1,1,4,4,5,5]
+; AVX512BW-NEXT:    vmovdqu8 %zmm14, %zmm15 {%k1}
+; AVX512BW-NEXT:    vmovdqa64 32(%rdx), %ymm21
+; AVX512BW-NEXT:    vpshufb %ymm17, %ymm21, %ymm14
+; AVX512BW-NEXT:    vmovdqa64 32(%rcx), %ymm22
+; AVX512BW-NEXT:    vpshufb %ymm19, %ymm22, %ymm17
+; AVX512BW-NEXT:    vporq %ymm14, %ymm17, %ymm14
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512BW-NEXT:    vpshufb %ymm7, %ymm21, %ymm17
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512BW-NEXT:    vpshufb %ymm7, %ymm22, %ymm18
+; AVX512BW-NEXT:    vporq %ymm17, %ymm18, %ymm17
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[2,3,2,3]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm17, %zmm14, %zmm14
+; AVX512BW-NEXT:    vmovdqa64 32(%rsi), %ymm30
+; AVX512BW-NEXT:    vmovdqa64 32(%rdi), %ymm29
+; AVX512BW-NEXT:    vpshuflw {{.*#+}} ymm17 = ymm29[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm17 = ymm17[0,0,1,1,4,4,5,5]
 ; AVX512BW-NEXT:    movl $676341840, %r10d # imm = 0x28502850
 ; AVX512BW-NEXT:    kmovd %r10d, %k1
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm15 {%k1} = ymm25[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512BW-NEXT:    vpshufb %ymm3, %ymm24, %ymm3
-; AVX512BW-NEXT:    vpshufb %ymm11, %ymm25, %ymm11
-; AVX512BW-NEXT:    vpor %ymm3, %ymm11, %ymm3
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm15, %zmm3, %zmm11
-; AVX512BW-NEXT:    movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
-; AVX512BW-NEXT:    kmovq %r10, %k3
-; AVX512BW-NEXT:    vmovdqu8 %zmm12, %zmm11 {%k3}
-; AVX512BW-NEXT:    vmovdqa64 32(%r9), %ymm23
-; AVX512BW-NEXT:    vpshufb %ymm6, %ymm23, %ymm3
-; AVX512BW-NEXT:    vmovdqa64 32(%r8), %ymm29
-; AVX512BW-NEXT:    vpshufb %ymm7, %ymm29, %ymm6
-; AVX512BW-NEXT:    vpor %ymm3, %ymm6, %ymm3
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm17 {%k1} = ymm30[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[2,3,2,3]
 ; AVX512BW-NEXT:    vpshufb %ymm6, %ymm29, %ymm6
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
-; AVX512BW-NEXT:    vpshufb %ymm7, %ymm23, %ymm7
-; AVX512BW-NEXT:    vpor %ymm6, %ymm7, %ymm6
-; AVX512BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm6 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512BW-NEXT:    vpermw %ymm14, %ymm6, %ymm6
-; AVX512BW-NEXT:    vpshufb %ymm1, %ymm14, %ymm1
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm6, %zmm1, %zmm1
-; AVX512BW-NEXT:    movabsq $145249953336295682, %r10 # imm = 0x204081020408102
+; AVX512BW-NEXT:    vpshufb %ymm9, %ymm30, %ymm9
+; AVX512BW-NEXT:    vpor %ymm6, %ymm9, %ymm6
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm17, %zmm6, %zmm9
+; AVX512BW-NEXT:    movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
 ; AVX512BW-NEXT:    kmovq %r10, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k2}
+; AVX512BW-NEXT:    vmovdqu8 %zmm14, %zmm9 {%k2}
+; AVX512BW-NEXT:    vmovdqa64 32(%r9), %ymm27
+; AVX512BW-NEXT:    vpshufb %ymm4, %ymm27, %ymm4
+; AVX512BW-NEXT:    vmovdqa64 32(%r8), %ymm28
+; AVX512BW-NEXT:    vpshufb %ymm5, %ymm28, %ymm5
+; AVX512BW-NEXT:    vpor %ymm4, %ymm5, %ymm4
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512BW-NEXT:    vpshufb %ymm5, %ymm28, %ymm5
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512BW-NEXT:    vpshufb %ymm6, %ymm27, %ymm6
+; AVX512BW-NEXT:    vpor %ymm5, %ymm6, %ymm5
+; AVX512BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm5 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512BW-NEXT:    vpermw %ymm23, %ymm5, %ymm5
+; AVX512BW-NEXT:    vpshufb %ymm1, %ymm23, %ymm1
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm5, %zmm1, %zmm1
+; AVX512BW-NEXT:    movabsq $145249953336295682, %r10 # imm = 0x204081020408102
+; AVX512BW-NEXT:    kmovq %r10, %k3
+; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm4 {%k3}
 ; AVX512BW-NEXT:    movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
-; AVX512BW-NEXT:    kmovq %r10, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm3, %zmm11 {%k2}
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm5[u],zero,xmm5[7],zero,xmm5[5,u,u,u],zero,xmm5[8],zero,xmm5[6,u,u,u],zero
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm4[u,7],zero,xmm4[5],zero,xmm4[u,u,u,8],zero,xmm4[6],zero,xmm4[u,u,u,9]
-; AVX512BW-NEXT:    vpor %xmm1, %xmm3, %xmm1
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512BW-NEXT:    vmovdqa {{.*#+}} xmm4 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512BW-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
-; AVX512BW-NEXT:    vinserti32x4 $2, %xmm1, %zmm3, %zmm1
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm9[u,u,u],zero,xmm9[7],zero,xmm9[5,u,u,u],zero,xmm9[8],zero,xmm9[6,u,u]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm8[u,u,u,7],zero,xmm8[5],zero,xmm8[u,u,u,8],zero,xmm8[6],zero,xmm8[u,u]
-; AVX512BW-NEXT:    vpor %xmm3, %xmm5, %xmm3
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
-; AVX512BW-NEXT:    vmovdqa {{.*#+}} xmm6 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512BW-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
-; AVX512BW-NEXT:    vinserti32x4 $2, %xmm3, %zmm5, %zmm3
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm15 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm3[0,1,0,1,4,5,4,5]
+; AVX512BW-NEXT:    kmovq %r10, %k3
+; AVX512BW-NEXT:    vmovdqu8 %zmm4, %zmm9 {%k3}
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm3[u],zero,xmm3[7],zero,xmm3[5,u,u,u],zero,xmm3[8],zero,xmm3[6,u,u,u],zero
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm4 = xmm2[u,7],zero,xmm2[5],zero,xmm2[u,u,u,8],zero,xmm2[6],zero,xmm2[u,u,u,9]
+; AVX512BW-NEXT:    vpor %xmm1, %xmm4, %xmm1
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512BW-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512BW-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
+; AVX512BW-NEXT:    vinserti32x4 $2, %xmm1, %zmm2, %zmm1
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} xmm17 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm2
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} xmm26 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512BW-NEXT:    vpshufb %xmm26, %xmm10, %xmm4
+; AVX512BW-NEXT:    vpor %xmm2, %xmm4, %xmm2
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3],xmm10[4],xmm11[4],xmm10[5],xmm11[5],xmm10[6],xmm11[6],xmm10[7],xmm11[7]
+; AVX512BW-NEXT:    vmovdqa {{.*#+}} xmm8 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512BW-NEXT:    vpshufb %xmm8, %xmm4, %xmm4
+; AVX512BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm4, %zmm2
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm14 = zmm1[0,1,0,1,4,5,4,5]
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm2[0,1,0,1,4,5,4,5]
 ; AVX512BW-NEXT:    movabsq $871499720017774092, %r10 # imm = 0xC183060C183060C
-; AVX512BW-NEXT:    kmovq %r10, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm15 {%k2}
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} xmm28 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512BW-NEXT:    vpshufb %xmm28, %xmm0, %xmm1
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} xmm30 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512BW-NEXT:    vpshufb %xmm30, %xmm2, %xmm3
-; AVX512BW-NEXT:    vpor %xmm1, %xmm3, %xmm1
-; 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:    vmovdqa {{.*#+}} xmm2 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512BW-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
+; AVX512BW-NEXT:    kmovq %r10, %k3
+; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm14 {%k3}
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm1 = zero,xmm31[4,u,u,u],zero,xmm31[7],zero,xmm31[5,u,u,u],zero,xmm31[8],zero,xmm31[6]
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} xmm18 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512BW-NEXT:    vpshufb %xmm18, %xmm0, %xmm2
+; AVX512BW-NEXT:    vpor %xmm1, %xmm2, %xmm1
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm31[0],xmm0[1],xmm31[1],xmm0[2],xmm31[2],xmm0[3],xmm31[3],xmm0[4],xmm31[4],xmm0[5],xmm31[5],xmm0[6],xmm31[6],xmm0[7],xmm31[7]
+; AVX512BW-NEXT:    vmovdqa {{.*#+}} xmm12 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512BW-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
 ; AVX512BW-NEXT:    vinserti32x4 $2, %xmm1, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512BW-NEXT:    vmovdqa64 (%rax), %zmm3
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm10, %zmm3, %zmm20
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512BW-NEXT:    vpermw %zmm20, %zmm1, %zmm1
+; AVX512BW-NEXT:    vmovdqa64 (%rax), %zmm5
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512BW-NEXT:    vpermw %zmm5, %zmm1, %zmm1
 ; AVX512BW-NEXT:    movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
-; AVX512BW-NEXT:    kmovq %rax, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm0 {%k2}
+; AVX512BW-NEXT:    kmovq %rax, %k3
+; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm0 {%k3}
 ; AVX512BW-NEXT:    movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
-; AVX512BW-NEXT:    kmovq %rax, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm0, %zmm15 {%k2}
-; AVX512BW-NEXT:    vmovdqa64 32(%rdi), %xmm31
-; AVX512BW-NEXT:    vmovdqa 32(%rsi), %xmm12
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm31[0],xmm12[0],xmm31[1],xmm12[1],xmm31[2],xmm12[2],xmm31[3],xmm12[3],xmm31[4],xmm12[4],xmm31[5],xmm12[5],xmm31[6],xmm12[6],xmm31[7],xmm12[7]
-; AVX512BW-NEXT:    vpshufb %xmm4, %xmm0, %xmm0
-; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm19[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512BW-NEXT:    kmovq %rax, %k3
+; AVX512BW-NEXT:    vmovdqu8 %zmm0, %zmm14 {%k3}
+; AVX512BW-NEXT:    vmovdqa 32(%rdi), %xmm6
+; AVX512BW-NEXT:    vmovdqa64 32(%rsi), %xmm31
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm6[0],xmm31[0],xmm6[1],xmm31[1],xmm6[2],xmm31[2],xmm6[3],xmm31[3],xmm6[4],xmm31[4],xmm6[5],xmm31[5],xmm6[6],xmm31[6],xmm6[7],xmm31[7]
+; AVX512BW-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
+; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm16[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
 ; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[2,2,3,3,6,6,7,7]
 ; AVX512BW-NEXT:    movl $338170920, %eax # imm = 0x14281428
-; AVX512BW-NEXT:    kmovd %eax, %k2
-; AVX512BW-NEXT:    vpbroadcastd {{.*#+}} ymm27 = [13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14]
-; AVX512BW-NEXT:    vpshufb %ymm27, %ymm18, %ymm1 {%k2}
-; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[2,3,2,3],zmm0[0,1,0,1]
-; AVX512BW-NEXT:    vmovdqa 32(%rcx), %xmm10
-; AVX512BW-NEXT:    vmovdqa 32(%rdx), %xmm7
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm7[0],xmm10[0],xmm7[1],xmm10[1],xmm7[2],xmm10[2],xmm7[3],xmm10[3],xmm7[4],xmm10[4],xmm7[5],xmm10[5],xmm7[6],xmm10[6],xmm7[7],xmm10[7]
-; AVX512BW-NEXT:    vpshufb %xmm6, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm4 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512BW-NEXT:    kmovd %eax, %k3
+; AVX512BW-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14]
+; AVX512BW-NEXT:    vpshufb %ymm7, %ymm13, %ymm1 {%k3}
+; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm1[2,3,2,3],zmm0[0,1,0,1]
+; AVX512BW-NEXT:    vmovdqa 32(%rcx), %xmm2
+; AVX512BW-NEXT:    vmovdqa 32(%rdx), %xmm1
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = 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]
+; AVX512BW-NEXT:    vpshufb %xmm8, %xmm0, %xmm0
+; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm4 = ymm10[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
 ; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,2,3,3,4,6,7,7]
-; AVX512BW-NEXT:    vpbroadcastq {{.*#+}} ymm26 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512BW-NEXT:    vpshufb %ymm26, %ymm9, %ymm5
-; AVX512BW-NEXT:    vmovdqu8 %ymm4, %ymm5 {%k1}
-; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm5[2,3,2,3],zmm1[0,1,0,1]
-; AVX512BW-NEXT:    vmovdqu8 %zmm0, %zmm4 {%k3}
-; AVX512BW-NEXT:    vmovdqa 32(%r9), %xmm6
-; AVX512BW-NEXT:    vmovdqa 32(%r8), %xmm5
-; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
-; AVX512BW-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm22[27],zero,zero,zero,zero,ymm22[30],zero,ymm22[28],zero,zero,zero,zero,ymm22[31],zero
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm2 = ymm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm21[27],zero,zero,zero,zero,ymm21[30],zero,ymm21[28],zero,zero,zero,zero,ymm21[31],zero,ymm21[29]
-; AVX512BW-NEXT:    vpor %ymm1, %ymm2, %ymm1
-; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm1[2,3,2,3],zmm0[0,1,0,1]
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
-; AVX512BW-NEXT:    vpermi2w %zmm3, %zmm14, %zmm1
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm0 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
-; AVX512BW-NEXT:    vpermi2w %zmm3, %zmm14, %zmm0
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm14 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
-; AVX512BW-NEXT:    vpermw %zmm3, %zmm14, %zmm3
+; AVX512BW-NEXT:    vpbroadcastq {{.*#+}} ymm8 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
+; AVX512BW-NEXT:    vpshufb %ymm8, %ymm11, %ymm20
+; AVX512BW-NEXT:    vmovdqu8 %ymm4, %ymm20 {%k1}
+; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm20[2,3,2,3],zmm0[0,1,0,1]
+; AVX512BW-NEXT:    vmovdqu8 %zmm3, %zmm0 {%k2}
+; AVX512BW-NEXT:    vmovdqa 32(%r9), %xmm4
+; AVX512BW-NEXT:    vmovdqa 32(%r8), %xmm3
+; AVX512BW-NEXT:    vpunpcklbw {{.*#+}} xmm20 = 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]
+; AVX512BW-NEXT:    vpshufb %xmm12, %xmm20, %xmm12
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm20 = ymm25[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm25[27],zero,zero,zero,zero,ymm25[30],zero,ymm25[28],zero,zero,zero,zero,ymm25[31],zero
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm19 = ymm24[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm24[27],zero,zero,zero,zero,ymm24[30],zero,ymm24[28],zero,zero,zero,zero,ymm24[31],zero,ymm24[29]
+; AVX512BW-NEXT:    vporq %ymm20, %ymm19, %ymm19
+; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm12 = zmm19[2,3,2,3],zmm12[0,1,0,1]
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm19 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,33,33,32,32,33,33,32,32,36,37,33,33,33,33,32,32]
+; AVX512BW-NEXT:    vpermi2w %zmm23, %zmm5, %zmm19
 ; AVX512BW-NEXT:    movabsq $580999813345182728, %rax # imm = 0x810204081020408
-; AVX512BW-NEXT:    kmovq %rax, %k3
-; AVX512BW-NEXT:    vmovdqu8 %zmm3, %zmm2 {%k3}
+; AVX512BW-NEXT:    kmovq %rax, %k2
+; AVX512BW-NEXT:    vmovdqu8 %zmm19, %zmm12 {%k2}
 ; AVX512BW-NEXT:    movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
-; AVX512BW-NEXT:    kmovq %rax, %k3
-; AVX512BW-NEXT:    vmovdqu8 %zmm2, %zmm4 {%k3}
-; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm2 = ymm24[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[2,2,3,3,6,6,7,7]
-; AVX512BW-NEXT:    vpshufb %ymm27, %ymm25, %ymm2 {%k2}
-; AVX512BW-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
-; AVX512BW-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512BW-NEXT:    vpshufb %ymm3, %ymm25, %ymm14
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} ymm25 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512BW-NEXT:    vpshufb %ymm25, %ymm24, %ymm24
-; AVX512BW-NEXT:    vporq %ymm14, %ymm24, %ymm14
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm2, %zmm14, %zmm2
-; AVX512BW-NEXT:    vpshufb %ymm26, %ymm17, %ymm14
-; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm24 = ymm16[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm24 = ymm24[0,2,3,3,4,6,7,7]
-; AVX512BW-NEXT:    vmovdqu8 %ymm24, %ymm14 {%k1}
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm16 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm16[24,25],zero,ymm16[23],zero,ymm16[21,22,23,26],zero,ymm16[24],zero,ymm16[28,29,26,27]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm17 = ymm17[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm17[25],zero,ymm17[23],zero,zero,zero,zero,ymm17[26],zero,ymm17[24],zero,zero,zero,zero
-; AVX512BW-NEXT:    vporq %ymm16, %ymm17, %ymm16
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm14, %zmm16, %zmm14
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm14 = zmm14[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT:    kmovq %rax, %k2
+; AVX512BW-NEXT:    vmovdqu8 %zmm12, %zmm0 {%k2}
+; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm12 = ymm29[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm12[2,2,3,3,6,6,7,7]
+; AVX512BW-NEXT:    vpshufb %ymm7, %ymm30, %ymm12 {%k3}
+; AVX512BW-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512BW-NEXT:    # ymm7 = mem[0,1,0,1]
+; AVX512BW-NEXT:    vpshufb %ymm7, %ymm30, %ymm19
+; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} ymm20 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512BW-NEXT:    vpshufb %ymm20, %ymm29, %ymm29
+; AVX512BW-NEXT:    vporq %ymm19, %ymm29, %ymm19
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm12, %zmm19, %zmm12
+; AVX512BW-NEXT:    vpshufb %ymm8, %ymm22, %ymm8
+; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm19 = ymm21[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm19 = ymm19[0,2,3,3,4,6,7,7]
+; AVX512BW-NEXT:    vmovdqu8 %ymm19, %ymm8 {%k1}
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm19 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm21[24,25],zero,ymm21[23],zero,ymm21[21,22,23,26],zero,ymm21[24],zero,ymm21[28,29,26,27]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm21 = ymm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm22[25],zero,ymm22[23],zero,zero,zero,zero,ymm22[26],zero,ymm22[24],zero,zero,zero,zero
+; AVX512BW-NEXT:    vporq %ymm19, %ymm21, %ymm19
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm8, %zmm19, %zmm8
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm21 = zmm12[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
 ; AVX512BW-NEXT:    movabsq $1742999440035548184, %rax # imm = 0x183060C183060C18
 ; AVX512BW-NEXT:    kmovq %rax, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm14, %zmm2 {%k2}
-; AVX512BW-NEXT:    vmovdqa64 (%r9), %zmm14
-; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm16 = zmm29[0,1,2,3],zmm14[4,5,6,7]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm16 = zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm16[23],zero,zmm16[23,24,25,26],zero,zmm16[24],zero,zmm16[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm16[59],zero,zero,zero,zero,zmm16[62],zero,zmm16[60],zero,zero,zero,zero,zmm16[63],zero
-; AVX512BW-NEXT:    vmovdqa64 (%r8), %zmm17
-; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm23 = zmm23[0,1,2,3],zmm17[4,5,6,7]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm23 = zmm23[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm23[25],zero,zmm23[23],zero,zero,zero,zero,zmm23[26],zero,zmm23[24],zero,zero,zmm23[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm23[59],zero,zero,zero,zero,zmm23[62],zero,zmm23[60],zero,zero,zero,zero,zmm23[63],zero,zmm23[61]
-; AVX512BW-NEXT:    vporq %zmm16, %zmm23, %zmm16
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT:    vmovdqu8 %zmm8, %zmm21 {%k2}
+; AVX512BW-NEXT:    vmovdqa64 (%r9), %zmm8
+; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm12 = zmm28[0,1,2,3],zmm8[4,5,6,7]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm12 = zmm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm12[23],zero,zmm12[23,24,25,26],zero,zmm12[24],zero,zmm12[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm12[59],zero,zero,zero,zero,zmm12[62],zero,zmm12[60],zero,zero,zero,zero,zmm12[63],zero
+; AVX512BW-NEXT:    vmovdqa64 (%r8), %zmm19
+; AVX512BW-NEXT:    vshufi64x2 {{.*#+}} zmm22 = zmm27[0,1,2,3],zmm19[4,5,6,7]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm22 = zmm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm22[25],zero,zmm22[23],zero,zero,zero,zero,zmm22[26],zero,zmm22[24],zero,zero,zmm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm22[59],zero,zero,zero,zero,zmm22[62],zero,zmm22[60],zero,zero,zero,zero,zmm22[63],zero,zmm22[61]
+; AVX512BW-NEXT:    vporq %zmm12, %zmm22, %zmm12
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm12 = zmm12[2,3,2,3,6,7,6,7]
 ; AVX512BW-NEXT:    movabsq $6971997760142192736, %rax # imm = 0x60C183060C183060
 ; AVX512BW-NEXT:    kmovq %rax, %k2
-; AVX512BW-NEXT:    vmovdqu8 %zmm16, %zmm2 {%k2}
+; AVX512BW-NEXT:    vmovdqu8 %zmm12, %zmm21 {%k2}
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm12 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
+; AVX512BW-NEXT:    vpermw %zmm23, %zmm12, %zmm12
 ; AVX512BW-NEXT:    movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
 ; AVX512BW-NEXT:    kmovq %rax, %k3
-; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k3}
-; AVX512BW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm19[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,0,1,1,4,4,5,5]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm1 {%k1} = ymm18[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512BW-NEXT:    vpshufb %ymm3, %ymm18, %ymm3
-; AVX512BW-NEXT:    vpshufb %ymm25, %ymm19, %ymm16
-; AVX512BW-NEXT:    vporq %ymm3, %ymm16, %ymm3
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm3, %zmm1, %zmm1
-; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm8[18],zero,zmm8[18,19,20,21],zero,zmm8[19],zero,zmm8[25,26,27,22],zero,zmm8[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm8[56,57],zero,zmm8[55],zero,zmm8[53,54,55,58],zero,zmm8[56],zero,zmm8[60,61,58,59]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm8 = zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm9[18],zero,zero,zero,zero,zmm9[21],zero,zmm9[19],zero,zero,zero,zero,zmm9[22],zero,zmm9[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm9[57],zero,zmm9[55],zero,zero,zero,zero,zmm9[58],zero,zmm9[56],zero,zero,zero,zero
-; AVX512BW-NEXT:    vporq %zmm3, %zmm8, %zmm3
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
-; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k2}
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm21, %zmm17, %zmm1
-; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[20],zero,zmm1[18],zero,zmm1[20,21,20,21],zero,zmm1[19],zero,zmm1[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[56,57,56,57],zero,zmm1[55],zero,zmm1[55,56,57,58],zero,zmm1[56],zero,zmm1[62,63]
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm14, %zmm8
+; AVX512BW-NEXT:    vmovdqu8 %zmm12, %zmm21 {%k3}
+; AVX512BW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm16[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512BW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm12[0,0,1,1,4,4,5,5]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm12 {%k1} = ymm13[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512BW-NEXT:    vpshufb %ymm7, %ymm13, %ymm7
+; AVX512BW-NEXT:    vpshufb %ymm20, %ymm16, %ymm13
+; AVX512BW-NEXT:    vpor %ymm7, %ymm13, %ymm7
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm12, %zmm7
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm10 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm10[18],zero,zmm10[18,19,20,21],zero,zmm10[19],zero,zmm10[25,26,27,22],zero,zmm10[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm10[56,57],zero,zmm10[55],zero,zmm10[53,54,55,58],zero,zmm10[56],zero,zmm10[60,61,58,59]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm11 = zmm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm11[18],zero,zero,zero,zero,zmm11[21],zero,zmm11[19],zero,zero,zero,zero,zmm11[22],zero,zmm11[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm11[57],zero,zmm11[55],zero,zero,zero,zero,zmm11[58],zero,zmm11[56],zero,zero,zero,zero
+; AVX512BW-NEXT:    vporq %zmm10, %zmm11, %zmm10
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT:    vmovdqu8 %zmm7, %zmm10 {%k2}
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm24, %zmm19, %zmm7
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm7 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm7[20],zero,zmm7[18],zero,zmm7[20,21,20,21],zero,zmm7[19],zero,zmm7[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm7[56,57,56,57],zero,zmm7[55],zero,zmm7[55,56,57,58],zero,zmm7[56],zero,zmm7[62,63]
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm25, %zmm8, %zmm8
 ; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm8 = zmm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm8[20],zero,zmm8[18],zero,zero,zero,zero,zmm8[21],zero,zmm8[19],zero,zero,zero,zero,zmm8[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm8[57],zero,zmm8[55],zero,zero,zero,zero,zmm8[58],zero,zmm8[56],zero,zero
-; AVX512BW-NEXT:    vporq %zmm1, %zmm8, %zmm1
-; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm8 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512BW-NEXT:    vpermw %zmm20, %zmm8, %zmm8
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT:    vporq %zmm7, %zmm8, %zmm7
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm8 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512BW-NEXT:    vpermw %zmm5, %zmm8, %zmm5
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
 ; AVX512BW-NEXT:    movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
 ; AVX512BW-NEXT:    kmovq %rax, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm8, %zmm1 {%k1}
+; AVX512BW-NEXT:    vmovdqu8 %zmm5, %zmm7 {%k1}
 ; AVX512BW-NEXT:    movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
 ; AVX512BW-NEXT:    kmovq %rax, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k1}
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm10[u,u,u],zero,xmm10[7],zero,xmm10[5,u,u,u],zero,xmm10[8],zero,xmm10[6,u,u]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm8 = xmm7[u,u,u,7],zero,xmm7[5],zero,xmm7[u,u,u,8],zero,xmm7[6],zero,xmm7[u,u]
-; AVX512BW-NEXT:    vpor %xmm1, %xmm8, %xmm1
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm10[8],xmm7[8],xmm10[9],xmm7[9],xmm10[10],xmm7[10],xmm10[11],xmm7[11],xmm10[12],xmm7[12],xmm10[13],xmm7[13],xmm10[14],xmm7[14],xmm10[15],xmm7[15]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-NEXT:    vinserti32x4 $2, %xmm7, %zmm1, %zmm1
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm7 = xmm12[u],zero,xmm12[7],zero,xmm12[5,u,u,u],zero,xmm12[8],zero,xmm12[6,u,u,u],zero
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm8 = xmm31[u,7],zero,xmm31[5],zero,xmm31[u,u,u,8],zero,xmm31[6],zero,xmm31[u,u,u,9]
-; AVX512BW-NEXT:    vpor %xmm7, %xmm8, %xmm7
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm8 = xmm12[8],xmm31[8],xmm12[9],xmm31[9],xmm12[10],xmm31[10],xmm12[11],xmm31[11],xmm12[12],xmm31[12],xmm12[13],xmm31[13],xmm12[14],xmm31[14],xmm12[15],xmm31[15]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm8 = xmm8[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-NEXT:    vinserti32x4 $2, %xmm8, %zmm7, %zmm7
+; AVX512BW-NEXT:    vmovdqu8 %zmm7, %zmm10 {%k1}
+; AVX512BW-NEXT:    vpshufb %xmm17, %xmm2, %xmm5
+; AVX512BW-NEXT:    vpshufb %xmm26, %xmm1, %xmm7
+; AVX512BW-NEXT:    vpor %xmm5, %xmm7, %xmm5
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512BW-NEXT:    vinserti32x4 $2, %xmm1, %zmm5, %zmm1
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm2 = xmm31[u],zero,xmm31[7],zero,xmm31[5,u,u,u],zero,xmm31[8],zero,xmm31[6,u,u,u],zero
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm6[u,7],zero,xmm6[5],zero,xmm6[u,u,u,8],zero,xmm6[6],zero,xmm6[u,u,u,9]
+; AVX512BW-NEXT:    vpor %xmm2, %xmm5, %xmm2
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm31[8],xmm6[8],xmm31[9],xmm6[9],xmm31[10],xmm6[10],xmm31[11],xmm6[11],xmm31[12],xmm6[12],xmm31[13],xmm6[13],xmm31[14],xmm6[14],xmm31[15],xmm6[15]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-NEXT:    vinserti32x4 $2, %xmm5, %zmm2, %zmm2
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512BW-NEXT:    vpermq {{.*#+}} zmm7 = zmm7[0,1,0,1,4,5,4,5]
-; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm7 {%k2}
-; AVX512BW-NEXT:    vpshufb %xmm28, %xmm6, %xmm1
-; AVX512BW-NEXT:    vpshufb %xmm30, %xmm5, %xmm8
-; AVX512BW-NEXT:    vpor %xmm1, %xmm8, %xmm1
-; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm6[8],xmm5[9],xmm6[9],xmm5[10],xmm6[10],xmm5[11],xmm6[11],xmm5[12],xmm6[12],xmm5[13],xmm6[13],xmm5[14],xmm6[14],xmm5[15],xmm6[15]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-NEXT:    vinserti32x4 $2, %xmm5, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
+; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k2}
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm1 = zero,xmm4[4,u,u,u],zero,xmm4[7],zero,xmm4[5,u,u,u],zero,xmm4[8],zero,xmm4[6]
+; AVX512BW-NEXT:    vpshufb %xmm18, %xmm3, %xmm5
+; AVX512BW-NEXT:    vpor %xmm1, %xmm5, %xmm1
+; AVX512BW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm3[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512BW-NEXT:    vinserti32x4 $2, %xmm3, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512BW-NEXT:    vpermw %zmm23, %zmm3, %zmm3
 ; AVX512BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
 ; AVX512BW-NEXT:    movabsq $290499906672591364, %rax # imm = 0x408102040810204
 ; AVX512BW-NEXT:    kmovq %rax, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512BW-NEXT:    vmovdqu8 %zmm3, %zmm1 {%k1}
 ; AVX512BW-NEXT:    movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
 ; AVX512BW-NEXT:    kmovq %rax, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm7 {%k1}
+; AVX512BW-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k1}
 ; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-NEXT:    vmovdqa64 %zmm11, 320(%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm7, 256(%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm3, 128(%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm2, 384(%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm4, 192(%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm15, (%rax)
-; AVX512BW-NEXT:    vmovdqa64 %zmm13, 64(%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm9, 320(%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm2, 256(%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm10, 128(%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm21, 384(%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm0, 192(%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm14, (%rax)
+; AVX512BW-NEXT:    vmovdqa64 %zmm15, 64(%rax)
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512BW-FCP-LABEL: store_i8_stride7_vf64:
 ; AVX512BW-FCP:       # %bb.0:
-; AVX512BW-FCP-NEXT:    subq $104, %rsp
+; AVX512BW-FCP-NEXT:    subq $72, %rsp
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-FCP-NEXT:    vmovdqa (%rax), %ymm2
-; AVX512BW-FCP-NEXT:    vmovdqu %ymm2, (%rsp) # 32-byte Spill
-; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} ymm20 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm20, %ymm2, %ymm0
-; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
-; AVX512BW-FCP-NEXT:    # ymm1 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpermw %ymm2, %ymm1, %ymm1
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm2
+; AVX512BW-FCP-NEXT:    vmovdqa (%rax), %ymm0
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} ymm19 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm19, %ymm0, %ymm1
+; AVX512BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512BW-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vpermw %ymm0, %ymm2, %ymm0
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512BW-FCP-NEXT:    vmovdqa (%r9), %ymm1
 ; AVX512BW-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm0 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm1, %ymm3
-; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %ymm4
-; AVX512BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm4, %ymm4
-; AVX512BW-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%r9), %xmm16
-; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %xmm15
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm15[8],xmm16[8],xmm15[9],xmm16[9],xmm15[10],xmm16[10],xmm15[11],xmm16[11],xmm15[12],xmm16[12],xmm15[13],xmm16[13],xmm15[14],xmm16[14],xmm15[15],xmm16[15]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm4[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm14 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm14, %ymm1, %ymm1
+; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %ymm2
+; AVX512BW-FCP-NEXT:    vmovdqu %ymm2, (%rsp) # 32-byte Spill
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm15 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm15, %ymm2, %ymm2
+; AVX512BW-FCP-NEXT:    vpor %ymm1, %ymm2, %ymm1
+; AVX512BW-FCP-NEXT:    vmovdqa (%r9), %xmm13
+; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %xmm12
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm12[8],xmm13[8],xmm12[9],xmm13[9],xmm12[10],xmm13[10],xmm12[11],xmm13[11],xmm12[12],xmm13[12],xmm12[13],xmm13[13],xmm12[14],xmm13[14],xmm12[15],xmm13[15]
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm16 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm16, %xmm2, %xmm2
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512BW-FCP-NEXT:    movabsq $2323999253380730912, %r10 # imm = 0x2040810204081020
 ; AVX512BW-FCP-NEXT:    kmovq %r10, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm3 {%k1}
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %ymm4
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %ymm2
+; AVX512BW-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm0 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm2, %ymm2
+; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %ymm4
 ; AVX512BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm2, %ymm4, %ymm4
-; AVX512BW-FCP-NEXT:    vmovdqa (%rcx), %ymm6
-; AVX512BW-FCP-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
-; AVX512BW-FCP-NEXT:    vpor %ymm4, %ymm6, %ymm4
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm3, %ymm4, %ymm4
+; AVX512BW-FCP-NEXT:    vpor %ymm2, %ymm4, %ymm2
 ; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdx), %xmm17
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rcx), %xmm19
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm19[8],xmm17[8],xmm19[9],xmm17[9],xmm19[10],xmm17[10],xmm19[11],xmm17[11],xmm19[12],xmm17[12],xmm19[13],xmm17[13],xmm19[14],xmm17[14],xmm19[15],xmm17[15]
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%rcx), %xmm18
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm18[8],xmm17[8],xmm18[9],xmm17[9],xmm18[10],xmm17[10],xmm18[11],xmm17[11],xmm18[12],xmm17[12],xmm18[13],xmm17[13],xmm18[14],xmm17[14],xmm18[15],xmm17[15]
 ; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm29 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm29, %xmm6, %xmm6
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm6, %zmm4
-; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm6
-; AVX512BW-FCP-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm6, %ymm6
-; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %ymm7
-; AVX512BW-FCP-NEXT:    vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} ymm23 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm23, %ymm7, %ymm21
-; AVX512BW-FCP-NEXT:    vporq %ymm6, %ymm21, %ymm6
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %xmm21
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rsi), %xmm22
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm24 = xmm22[8],xmm21[8],xmm22[9],xmm21[9],xmm22[10],xmm21[10],xmm22[11],xmm21[11],xmm22[12],xmm21[12],xmm22[13],xmm21[13],xmm22[14],xmm21[14],xmm22[15],xmm21[15]
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm24, %xmm24
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm24 = ymm24[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm24, %zmm14
+; AVX512BW-FCP-NEXT:    vpshufb %xmm29, %xmm4, %xmm4
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm2
+; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm4
+; AVX512BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm8 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512BW-FCP-NEXT:    vmovdqa (%rsi), %ymm5
+; AVX512BW-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} ymm22 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm22, %ymm5, %ymm20
+; AVX512BW-FCP-NEXT:    vporq %ymm4, %ymm20, %ymm4
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %xmm20
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%rsi), %xmm21
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm23 = xmm21[8],xmm20[8],xmm21[9],xmm20[9],xmm21[10],xmm20[10],xmm21[11],xmm20[11],xmm21[12],xmm20[12],xmm21[13],xmm20[13],xmm21[14],xmm20[14],xmm21[15],xmm20[15]
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm30 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm30, %xmm23, %xmm23
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm23 = ymm23[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm23, %zmm11
 ; AVX512BW-FCP-NEXT:    movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
 ; AVX512BW-FCP-NEXT:    kmovq %r10, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm14 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm11 {%k1}
 ; AVX512BW-FCP-NEXT:    movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
 ; AVX512BW-FCP-NEXT:    kmovq %r10, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm3, %zmm14 {%k1}
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdx), %ymm3
-; AVX512BW-FCP-NEXT:    vpshufb %ymm2, %ymm3, %ymm4
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rcx), %ymm2
-; AVX512BW-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm5
-; AVX512BW-FCP-NEXT:    vpor %ymm4, %ymm5, %ymm4
-; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm5, %ymm3, %ymm5
-; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm24
-; AVX512BW-FCP-NEXT:    vporq %ymm5, %ymm24, %ymm5
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdi), %ymm5
-; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm5, %ymm9
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm11 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdx), %ymm1
+; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm1, %ymm2
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%rcx), %ymm0
+; AVX512BW-FCP-NEXT:    vpshufb %ymm3, %ymm0, %ymm3
+; AVX512BW-FCP-NEXT:    vpor %ymm2, %ymm3, %ymm2
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm3, %ymm1, %ymm3
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm0, %ymm23
+; AVX512BW-FCP-NEXT:    vporq %ymm3, %ymm23, %ymm3
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdi), %ymm3
+; AVX512BW-FCP-NEXT:    vpshufb %ymm8, %ymm3, %ymm8
 ; AVX512BW-FCP-NEXT:    vmovdqa64 32(%rsi), %ymm25
-; AVX512BW-FCP-NEXT:    vpshufb %ymm23, %ymm25, %ymm23
-; AVX512BW-FCP-NEXT:    vporq %ymm9, %ymm23, %ymm9
-; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm23
-; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm25, %ymm24
-; AVX512BW-FCP-NEXT:    vporq %ymm23, %ymm24, %ymm23
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm23, %zmm9, %zmm9
+; AVX512BW-FCP-NEXT:    vpshufb %ymm22, %ymm25, %ymm22
+; AVX512BW-FCP-NEXT:    vporq %ymm8, %ymm22, %ymm8
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm3, %ymm22
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm25, %ymm23
+; AVX512BW-FCP-NEXT:    vporq %ymm22, %ymm23, %ymm22
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm22 = ymm22[2,3,2,3]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm22, %zmm8, %zmm8
 ; AVX512BW-FCP-NEXT:    movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
 ; AVX512BW-FCP-NEXT:    kmovq %r10, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm9 {%k1}
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%r9), %ymm4
-; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm8 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%r9), %ymm2
+; AVX512BW-FCP-NEXT:    vpshufb %ymm14, %ymm2, %ymm14
 ; AVX512BW-FCP-NEXT:    vmovdqa64 32(%r8), %ymm28
-; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm28, %ymm1
-; AVX512BW-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm28, %ymm1
-; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
-; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm4, %ymm23
-; AVX512BW-FCP-NEXT:    vporq %ymm1, %ymm23, %ymm1
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rax), %ymm6
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm23 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
-; AVX512BW-FCP-NEXT:    vpermw %ymm6, %ymm23, %ymm23
-; AVX512BW-FCP-NEXT:    vpshufb %ymm20, %ymm6, %ymm20
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm23, %zmm20, %zmm20
+; AVX512BW-FCP-NEXT:    vpshufb %ymm15, %ymm28, %ymm15
+; AVX512BW-FCP-NEXT:    vporq %ymm14, %ymm15, %ymm22
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm28, %ymm23
+; AVX512BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm2, %ymm24
+; AVX512BW-FCP-NEXT:    vporq %ymm23, %ymm24, %ymm23
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm23, %zmm22, %zmm23
+; AVX512BW-FCP-NEXT:    vmovdqa64 32(%rax), %ymm24
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm22 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
+; AVX512BW-FCP-NEXT:    vpermw %ymm24, %ymm22, %ymm22
+; AVX512BW-FCP-NEXT:    vpshufb %ymm19, %ymm24, %ymm19
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm22, %zmm19, %zmm19
 ; AVX512BW-FCP-NEXT:    movabsq $145249953336295682, %r10 # imm = 0x204081020408102
 ; AVX512BW-FCP-NEXT:    kmovq %r10, %k2
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm20, %zmm1 {%k2}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm19, %zmm23 {%k2}
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdx), %zmm22
 ; AVX512BW-FCP-NEXT:    movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
 ; AVX512BW-FCP-NEXT:    kmovq %r10, %k2
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm9 {%k2}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm23, %zmm8 {%k2}
 ; AVX512BW-FCP-NEXT:    vmovdqa64 (%rcx), %zmm23
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm23[4,5,6,7]
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm23[4,5,6,7]
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,zmm1[23],zero,zmm1[21,22,23,26],zero,zmm1[24],zero,zmm1[28,29,26,27,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero,zmm1[61],zero
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdx), %zmm24
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm24[4,5,6,7]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm2[25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zero,zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[62],zero,zmm2[60],zero,zero,zero,zero,zmm2[63],zero,zmm2[61],zero,zero
-; AVX512BW-FCP-NEXT:    vporq %zmm1, %zmm2, %zmm1
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm22[4,5,6,7]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm0[25],zero,zmm0[23],zero,zero,zero,zero,zmm0[26],zero,zmm0[24],zero,zero,zero,zero,zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm0[62],zero,zmm0[60],zero,zero,zero,zero,zmm0[63],zero,zmm0[61],zero,zero
+; AVX512BW-FCP-NEXT:    vporq %zmm1, %zmm0, %zmm0
 ; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm26
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm25[0,1,2,3],zmm26[4,5,6,7]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zero,zero,zmm2[27],zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,60,61,62],zero,zmm2[60],zero,zmm2[62,63,62,63],zero,zmm2[61],zero,zmm2[63,60,61]
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm25[0,1,2,3],zmm26[4,5,6,7]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,zmm1[23],zero,zero,zero,zero,zmm1[26],zero,zmm1[24],zero,zero,zero,zero,zmm1[27],zero,zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,60,61,62],zero,zmm1[60],zero,zmm1[62,63,62,63],zero,zmm1[61],zero,zmm1[63,60,61]
 ; AVX512BW-FCP-NEXT:    vmovdqa64 (%rsi), %zmm27
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm5[0,1,2,3],zmm27[4,5,6,7]
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm27[4,5,6,7]
 ; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[23],zero,zero,zero,zero,zmm3[26],zero,zmm3[24],zero,zero,zero,zero,zmm3[27],zero,zmm3[25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zmm3[62],zero,zmm3[60],zero,zero,zero,zero,zmm3[63],zero,zmm3[61],zero,zero,zero
-; AVX512BW-FCP-NEXT:    vporq %zmm2, %zmm3, %zmm2
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm20 = zmm2[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT:    vporq %zmm1, %zmm3, %zmm1
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm19 = zmm1[2,3,2,3,6,7,6,7]
 ; AVX512BW-FCP-NEXT:    movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
 ; AVX512BW-FCP-NEXT:    kmovq %r10, %k2
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm19 {%k2}
 ; AVX512BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm25
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm28[0,1,2,3],zmm25[4,5,6,7]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm1[23],zero,zmm1[23,24,25,26],zero,zmm1[24],zero,zmm1[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm1[59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm28[0,1,2,3],zmm25[4,5,6,7]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm0[23],zero,zmm0[23,24,25,26],zero,zmm0[24],zero,zmm0[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm0[59],zero,zero,zero,zero,zmm0[62],zero,zmm0[60],zero,zero,zero,zero,zmm0[63],zero
 ; AVX512BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm28
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm4[0,1,2,3],zmm28[4,5,6,7]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm2[59],zero,zero,zero,zero,zmm2[62],zero,zmm2[60],zero,zero,zero,zero,zmm2[63],zero,zmm2[61]
-; AVX512BW-FCP-NEXT:    vporq %zmm1, %zmm2, %zmm1
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm28[4,5,6,7]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm1[25],zero,zmm1[23],zero,zero,zero,zero,zmm1[26],zero,zmm1[24],zero,zero,zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm1[59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero,zmm1[61]
+; AVX512BW-FCP-NEXT:    vporq %zmm0, %zmm1, %zmm0
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
 ; AVX512BW-FCP-NEXT:    movabsq $6971997760142192736, %r10 # imm = 0x60C183060C183060
 ; AVX512BW-FCP-NEXT:    kmovq %r10, %k2
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm20 {%k2}
-; AVX512BW-FCP-NEXT:    vmovdqa64 (%rax), %zmm31
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
-; AVX512BW-FCP-NEXT:    vpermi2w %zmm31, %zmm6, %zmm1
-; AVX512BW-FCP-NEXT:    movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
-; AVX512BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm20 {%k3}
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdx), %xmm5
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm19 {%k2}
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm0 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
+; AVX512BW-FCP-NEXT:    vpermw %zmm24, %zmm0, %zmm0
+; AVX512BW-FCP-NEXT:    movabsq $-9150747060186627967, %r10 # imm = 0x8102040810204081
+; AVX512BW-FCP-NEXT:    kmovq %r10, %k3
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm19 {%k3}
+; AVX512BW-FCP-NEXT:    vmovdqa64 32(%rdx), %xmm31
 ; AVX512BW-FCP-NEXT:    vmovdqa 32(%rcx), %xmm4
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm13 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm4, %xmm1
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm12 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm5, %xmm2
-; AVX512BW-FCP-NEXT:    vpor %xmm1, %xmm2, %xmm1
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm29, %xmm2, %xmm2
-; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm1, %zmm3
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdi), %xmm2
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm4, %xmm0
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm9 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm31, %xmm1
+; AVX512BW-FCP-NEXT:    vpor %xmm0, %xmm1, %xmm0
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm31[8],xmm4[9],xmm31[9],xmm4[10],xmm31[10],xmm4[11],xmm31[11],xmm4[12],xmm31[12],xmm4[13],xmm31[13],xmm4[14],xmm31[14],xmm4[15],xmm31[15]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm29, %xmm1, %xmm1
+; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm0, %zmm2
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdi), %xmm3
 ; AVX512BW-FCP-NEXT:    vmovdqa 32(%rsi), %xmm1
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm29
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm8 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm30
-; AVX512BW-FCP-NEXT:    vporq %xmm29, %xmm30, %xmm29
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm30 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm30, %xmm30
-; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm30, %zmm29, %zmm29
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm3 = zmm3[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm29 = zmm29[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm3, %zmm29 {%k2}
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm29
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm3, %xmm14
+; AVX512BW-FCP-NEXT:    vporq %xmm29, %xmm14, %xmm14
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm29 = xmm1[8],xmm3[8],xmm1[9],xmm3[9],xmm1[10],xmm3[10],xmm1[11],xmm3[11],xmm1[12],xmm3[12],xmm1[13],xmm3[13],xmm1[14],xmm3[14],xmm1[15],xmm3[15]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm30, %xmm29, %xmm29
+; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm29, %zmm14, %zmm14
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm29 = zmm14[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm29 {%k2}
 ; AVX512BW-FCP-NEXT:    vmovdqa64 32(%r9), %xmm30
-; AVX512BW-FCP-NEXT:    vmovdqa 32(%r8), %xmm3
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm30, %xmm0
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm18
-; AVX512BW-FCP-NEXT:    vporq %xmm0, %xmm18, %xmm0
-; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm18 = xmm3[8],xmm30[8],xmm3[9],xmm30[9],xmm3[10],xmm30[10],xmm3[11],xmm30[11],xmm3[12],xmm30[12],xmm3[13],xmm30[13],xmm3[14],xmm30[14],xmm3[15],xmm30[15]
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm18 = xmm18[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm18, %zmm0, %zmm0
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm18 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
-; AVX512BW-FCP-NEXT:    vpermi2w %zmm31, %zmm6, %zmm18
+; AVX512BW-FCP-NEXT:    vmovdqa 32(%r8), %xmm2
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm14 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm14, %xmm30, %xmm0
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm5 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm15
+; AVX512BW-FCP-NEXT:    vpor %xmm0, %xmm15, %xmm0
+; AVX512BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm15 = xmm2[8],xmm30[8],xmm2[9],xmm30[9],xmm2[10],xmm30[10],xmm2[11],xmm30[11],xmm2[12],xmm30[12],xmm2[13],xmm30[13],xmm2[14],xmm30[14],xmm2[15],xmm30[15]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm16, %xmm15, %xmm15
+; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm0, %zmm0
 ; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT:    movabsq $290499906672591364, %rax # imm = 0x408102040810204
-; AVX512BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm18, %zmm0 {%k3}
-; AVX512BW-FCP-NEXT:    movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
-; AVX512BW-FCP-NEXT:    kmovq %rax, %k3
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512BW-FCP-NEXT:    vpermw %zmm24, %zmm15, %zmm15
+; AVX512BW-FCP-NEXT:    movabsq $290499906672591364, %rcx # imm = 0x408102040810204
+; AVX512BW-FCP-NEXT:    kmovq %rcx, %k3
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm15, %zmm0 {%k3}
+; AVX512BW-FCP-NEXT:    movabsq $-8714997200177740921, %rcx # imm = 0x870E1C3870E1C387
+; AVX512BW-FCP-NEXT:    kmovq %rcx, %k3
 ; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm29 {%k3}
-; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm19, %xmm0
-; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm17, %xmm6
-; AVX512BW-FCP-NEXT:    vpor %xmm0, %xmm6, %xmm0
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm17[0],xmm19[0],xmm17[1],xmm19[1],xmm17[2],xmm19[2],xmm17[3],xmm19[3],xmm17[4],xmm19[4],xmm17[5],xmm19[5],xmm17[6],xmm19[6],xmm17[7],xmm19[7]
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm12 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm6, %xmm6
-; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm6, %zmm0
-; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm22, %xmm6
-; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm21, %xmm8
-; AVX512BW-FCP-NEXT:    vpor %xmm6, %xmm8, %xmm6
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm21[0],xmm22[0],xmm21[1],xmm22[1],xmm21[2],xmm22[2],xmm21[3],xmm22[3],xmm21[4],xmm22[4],xmm21[5],xmm22[5],xmm21[6],xmm22[6],xmm21[7],xmm22[7]
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm8, %xmm8
-; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm8, %zmm6
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm8 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm6[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT:    movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
-; AVX512BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm8, %zmm0 {%k3}
-; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm16, %xmm6
-; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm15, %xmm7
-; AVX512BW-FCP-NEXT:    vpor %xmm6, %xmm7, %xmm6
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm15[0],xmm16[0],xmm15[1],xmm16[1],xmm15[2],xmm16[2],xmm15[3],xmm16[3],xmm15[4],xmm16[4],xmm15[5],xmm16[5],xmm15[6],xmm16[6],xmm15[7],xmm16[7]
-; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm7, %xmm7
+; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm18, %xmm0
+; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm17, %xmm9
+; AVX512BW-FCP-NEXT:    vpor %xmm0, %xmm9, %xmm0
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm9 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3],xmm17[4],xmm18[4],xmm17[5],xmm18[5],xmm17[6],xmm18[6],xmm17[7],xmm18[7]
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm9, %xmm9
+; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm9, %zmm0
+; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm21, %xmm7
+; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm20, %xmm6
+; AVX512BW-FCP-NEXT:    vpor %xmm7, %xmm6, %xmm6
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm20[0],xmm21[0],xmm20[1],xmm21[1],xmm20[2],xmm21[2],xmm20[3],xmm21[3],xmm20[4],xmm21[4],xmm20[5],xmm21[5],xmm20[6],xmm21[6],xmm20[7],xmm21[7]
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm9 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm7, %xmm7
 ; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm7, %zmm6
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm6 = zmm6[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, (%rsp), %zmm31, %zmm8 # 32-byte Folded Reload
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm7 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512BW-FCP-NEXT:    vpermw %zmm8, %zmm7, %zmm7
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm16 = zmm6[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT:    movabsq $871499720017774092, %rcx # imm = 0xC183060C183060C
+; AVX512BW-FCP-NEXT:    kmovq %rcx, %k3
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm16 {%k3}
+; AVX512BW-FCP-NEXT:    vpshufb %xmm14, %xmm13, %xmm0
+; AVX512BW-FCP-NEXT:    vpshufb %xmm5, %xmm12, %xmm5
+; AVX512BW-FCP-NEXT:    vpor %xmm0, %xmm5, %xmm0
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3],xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
+; AVX512BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
+; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm5, %zmm0
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT:    vmovdqa64 (%rax), %zmm7
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm5 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512BW-FCP-NEXT:    vpermw %zmm7, %zmm5, %zmm5
 ; AVX512BW-FCP-NEXT:    movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm7, %zmm6 {%k3}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm5, %zmm0 {%k3}
 ; AVX512BW-FCP-NEXT:    movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm0 {%k3}
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm1
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm16 {%k3}
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm0 = 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:    vpshufb %xmm9, %xmm0, %xmm0
+; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm12[28],zero,ymm12[30,31,30,31],zero,ymm12[29],zero,ymm12[31,28,29]
 ; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm13[28],zero,ymm13[30,31,30,31],zero,ymm13[29],zero,ymm13[31,28,29]
-; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm15[30],zero,ymm15[28],zero,zero,zero,zero,ymm15[31],zero,ymm15[29],zero,zero,zero
-; AVX512BW-FCP-NEXT:    vpor %ymm2, %ymm6, %ymm2
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm2[2,3,2,3],zmm1[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm2
-; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero,ymm11[29],zero
-; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29],zero,zero
-; AVX512BW-FCP-NEXT:    vpor %ymm4, %ymm5, %ymm4
-; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm4[2,3,2,3],zmm2[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k1}
-; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm30[0],xmm3[1],xmm30[1],xmm3[2],xmm30[2],xmm3[3],xmm30[3],xmm3[4],xmm30[4],xmm3[5],xmm30[5],xmm3[6],xmm30[6],xmm3[7],xmm30[7]
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm13[30],zero,ymm13[28],zero,zero,zero,zero,ymm13[31],zero,ymm13[29],zero,zero,zero
+; AVX512BW-FCP-NEXT:    vpor %ymm1, %ymm3, %ymm1
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[2,3,2,3],zmm0[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm31[0],xmm4[0],xmm31[1],xmm4[1],xmm31[2],xmm4[2],xmm31[3],xmm4[3],xmm31[4],xmm4[4],xmm31[5],xmm4[5],xmm31[6],xmm4[6],xmm31[7],xmm4[7]
 ; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm1
-; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
+; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm10[30],zero,ymm10[28],zero,zero,zero,zero,ymm10[31],zero,ymm10[29],zero
+; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero,ymm9[29],zero,zero
 ; AVX512BW-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
 ; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm3[2,3,2,3],zmm1[0,1,0,1]
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
-; AVX512BW-FCP-NEXT:    vpermw %zmm31, %zmm3, %zmm3
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm30[0],xmm2[1],xmm30[1],xmm2[2],xmm30[2],xmm2[3],xmm30[3],xmm2[4],xmm30[4],xmm2[5],xmm30[5],xmm2[6],xmm30[6],xmm2[7],xmm30[7]
+; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm0
+; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm4[27],zero,zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero
+; AVX512BW-FCP-NEXT:    vmovdqu (%rsp), %ymm5 # 32-byte Reload
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29]
+; AVX512BW-FCP-NEXT:    vpor %ymm2, %ymm3, %ymm2
+; AVX512BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm2[2,3,2,3],zmm0[0,1,0,1]
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,33,33,32,32,33,33,32,32,36,37,33,33,33,33,32,32]
+; AVX512BW-FCP-NEXT:    vpermi2w %zmm24, %zmm7, %zmm2
 ; AVX512BW-FCP-NEXT:    movabsq $580999813345182728, %rax # imm = 0x810204081020408
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm3, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm0 {%k1}
 ; AVX512BW-FCP-NEXT:    movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k1}
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm15, %zmm26, %zmm1
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[18,19,20,21],zero,zmm1[19],zero,zmm1[21,20,21,22],zero,zmm1[20],zero,zmm1[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm1[55],zero,zero,zero,zero,zmm1[58],zero,zmm1[56],zero,zero,zero,zero,zmm1[59],zero
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm27, %zmm3
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22],zero,zmm3[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero,zero,zero,zmm3[59],zero,zmm3[57]
-; AVX512BW-FCP-NEXT:    vporq %zmm1, %zmm3, %zmm1
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm24, %zmm3
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[18],zero,zmm3[18,19,20,21],zero,zmm3[19],zero,zmm3[25,26,27,22],zero,zmm3[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[56,57],zero,zmm3[55],zero,zmm3[53,54,55,58],zero,zmm3[56],zero,zmm3[60,61,58,59]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm23, %zmm4
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm4[18],zero,zero,zero,zero,zmm4[21],zero,zmm4[19],zero,zero,zero,zero,zmm4[22],zero,zmm4[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm4[57],zero,zmm4[55],zero,zero,zero,zero,zmm4[58],zero,zmm4[56],zero,zero,zero,zero
-; AVX512BW-FCP-NEXT:    vporq %zmm3, %zmm4, %zmm3
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k2}
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm28, %zmm1
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[20],zero,zmm1[18],zero,zmm1[20,21,20,21],zero,zmm1[19],zero,zmm1[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[56,57,56,57],zero,zmm1[55],zero,zmm1[55,56,57,58],zero,zmm1[56],zero,zmm1[62,63]
-; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm25, %zmm4
-; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm4[20],zero,zmm4[18],zero,zero,zero,zero,zmm4[21],zero,zmm4[19],zero,zero,zero,zero,zmm4[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm4[57],zero,zmm4[55],zero,zero,zero,zero,zmm4[58],zero,zmm4[56],zero,zero
-; AVX512BW-FCP-NEXT:    vporq %zmm1, %zmm4, %zmm1
-; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm4 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512BW-FCP-NEXT:    vpermw %zmm8, %zmm4, %zmm4
-; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm26, %zmm0
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[18,19,20,21],zero,zmm0[19],zero,zmm0[21,20,21,22],zero,zmm0[20],zero,zmm0[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm0[55],zero,zero,zero,zero,zmm0[58],zero,zmm0[56],zero,zero,zero,zero,zmm0[59],zero
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm27, %zmm2
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[21],zero,zmm2[19],zero,zero,zero,zero,zmm2[22],zero,zmm2[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[55],zero,zero,zero,zero,zmm2[58],zero,zmm2[56],zero,zero,zero,zero,zmm2[59],zero,zmm2[57]
+; AVX512BW-FCP-NEXT:    vporq %zmm0, %zmm2, %zmm0
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm22, %zmm2
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[18],zero,zmm2[18,19,20,21],zero,zmm2[19],zero,zmm2[25,26,27,22],zero,zmm2[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[56,57],zero,zmm2[55],zero,zmm2[53,54,55,58],zero,zmm2[56],zero,zmm2[60,61,58,59]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm23, %zmm3
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[18],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22],zero,zmm3[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm3[57],zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero,zero,zero
+; AVX512BW-FCP-NEXT:    vporq %zmm2, %zmm3, %zmm2
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm2 {%k2}
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm28, %zmm0
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[20],zero,zmm0[18],zero,zmm0[20,21,20,21],zero,zmm0[19],zero,zmm0[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[56,57,56,57],zero,zmm0[55],zero,zmm0[55,56,57,58],zero,zmm0[56],zero,zmm0[62,63]
+; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm25, %zmm3
+; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[20],zero,zmm3[18],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm3[57],zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero
+; AVX512BW-FCP-NEXT:    vporq %zmm0, %zmm3, %zmm0
+; AVX512BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512BW-FCP-NEXT:    vpermw %zmm7, %zmm3, %zmm3
+; AVX512BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
 ; AVX512BW-FCP-NEXT:    movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm3, %zmm0 {%k1}
 ; AVX512BW-FCP-NEXT:    movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
 ; AVX512BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k1}
+; AVX512BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm2 {%k1}
 ; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, 128(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, 320(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, 128(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, 320(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, 192(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, (%rax)
 ; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, 256(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, 64(%rax)
-; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, 384(%rax)
-; AVX512BW-FCP-NEXT:    addq $104, %rsp
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, 64(%rax)
+; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, 384(%rax)
+; AVX512BW-FCP-NEXT:    addq $72, %rsp
 ; AVX512BW-FCP-NEXT:    vzeroupper
 ; AVX512BW-FCP-NEXT:    retq
 ;
 ; AVX512DQ-BW-LABEL: store_i8_stride7_vf64:
 ; AVX512DQ-BW:       # %bb.0:
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-NEXT:    vmovdqa (%rax), %ymm10
-; AVX512DQ-BW-NEXT:    vmovdqa 32(%rax), %ymm14
+; AVX512DQ-BW-NEXT:    vmovdqa (%rax), %ymm0
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rax), %ymm23
 ; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm1 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm1, %ymm10, %ymm0
-; AVX512DQ-BW-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
-; AVX512DQ-BW-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vpermw %ymm10, %ymm2, %ymm2
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm3
-; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %ymm22
-; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm6 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm6, %ymm22, %ymm0
-; AVX512DQ-BW-NEXT:    vmovdqa64 (%r8), %ymm21
-; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm7 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm7, %ymm21, %ymm2
-; AVX512DQ-BW-NEXT:    vpor %ymm0, %ymm2, %ymm4
-; AVX512DQ-BW-NEXT:    vmovdqa (%r9), %xmm0
-; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm2
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm12
+; AVX512DQ-BW-NEXT:    vpshufb %ymm1, %ymm0, %ymm2
+; AVX512DQ-BW-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512DQ-BW-NEXT:    # ymm3 = mem[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vpermw %ymm0, %ymm3, %ymm0
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
+; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %ymm25
+; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm4 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm4, %ymm25, %ymm0
+; AVX512DQ-BW-NEXT:    vmovdqa64 (%r8), %ymm24
+; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm5, %ymm24, %ymm3
+; AVX512DQ-BW-NEXT:    vpor %ymm0, %ymm3, %ymm3
+; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %xmm31
+; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm0
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm0[8],xmm31[8],xmm0[9],xmm31[9],xmm0[10],xmm31[10],xmm0[11],xmm31[11],xmm0[12],xmm31[12],xmm0[13],xmm31[13],xmm0[14],xmm31[14],xmm0[15],xmm31[15]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm6 = xmm6[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm3, %zmm6, %zmm14
 ; AVX512DQ-BW-NEXT:    movabsq $2323999253380730912, %r10 # imm = 0x2040810204081020
 ; AVX512DQ-BW-NEXT:    kmovq %r10, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm3, %zmm12 {%k1}
-; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %ymm19
-; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm3, %ymm19, %ymm4
-; AVX512DQ-BW-NEXT:    vmovdqa64 (%rsi), %ymm18
-; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm11 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm11, %ymm18, %ymm5
-; AVX512DQ-BW-NEXT:    vpor %ymm4, %ymm5, %ymm8
-; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm4
-; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %xmm5
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm9 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm9 = xmm9[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm8, %zmm9, %zmm13
-; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm15 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm15, %ymm8, %ymm16
-; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} ymm20 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm20, %ymm9, %ymm17
-; AVX512DQ-BW-NEXT:    vporq %ymm16, %ymm17, %ymm16
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm17 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm17 = xmm17[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm16, %zmm17, %zmm16
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm2, %zmm14 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %ymm16
+; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm6, %ymm16, %ymm2
+; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %ymm13
+; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} ymm9 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm9, %ymm13, %ymm3
+; AVX512DQ-BW-NEXT:    vpor %ymm2, %ymm3, %ymm7
+; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm2
+; AVX512DQ-BW-NEXT:    vmovdqa (%rsi), %xmm3
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm10 = xmm10[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm10, %zmm15
+; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} ymm17 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm17, %ymm10, %ymm18
+; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} ymm19 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm19, %ymm11, %ymm20
+; AVX512DQ-BW-NEXT:    vporq %ymm18, %ymm20, %ymm18
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm20 = xmm20[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm20 = ymm20[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm18, %zmm20, %zmm18
 ; AVX512DQ-BW-NEXT:    movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
 ; AVX512DQ-BW-NEXT:    kmovq %r10, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm16, %zmm13 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm18, %zmm15 {%k1}
 ; AVX512DQ-BW-NEXT:    movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
 ; AVX512DQ-BW-NEXT:    kmovq %r10, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm12, %zmm13 {%k1}
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rdx), %ymm16
-; AVX512DQ-BW-NEXT:    vpshufb %ymm15, %ymm16, %ymm12
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rcx), %ymm17
-; AVX512DQ-BW-NEXT:    vpshufb %ymm20, %ymm17, %ymm15
-; AVX512DQ-BW-NEXT:    vpor %ymm12, %ymm15, %ymm12
-; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm15 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm15, %ymm16, %ymm15
-; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm20 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm20, %ymm17, %ymm20
-; AVX512DQ-BW-NEXT:    vporq %ymm15, %ymm20, %ymm15
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm15, %zmm12, %zmm12
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rsi), %ymm25
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rdi), %ymm24
-; AVX512DQ-BW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm24[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm15 = ymm15[0,0,1,1,4,4,5,5]
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm14, %zmm15 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rdx), %ymm21
+; AVX512DQ-BW-NEXT:    vpshufb %ymm17, %ymm21, %ymm14
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rcx), %ymm22
+; AVX512DQ-BW-NEXT:    vpshufb %ymm19, %ymm22, %ymm17
+; AVX512DQ-BW-NEXT:    vporq %ymm14, %ymm17, %ymm14
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm7, %ymm21, %ymm17
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm7, %ymm22, %ymm18
+; AVX512DQ-BW-NEXT:    vporq %ymm17, %ymm18, %ymm17
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[2,3,2,3]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm17, %zmm14, %zmm14
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rsi), %ymm30
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rdi), %ymm29
+; AVX512DQ-BW-NEXT:    vpshuflw {{.*#+}} ymm17 = ymm29[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm17 = ymm17[0,0,1,1,4,4,5,5]
 ; AVX512DQ-BW-NEXT:    movl $676341840, %r10d # imm = 0x28502850
 ; AVX512DQ-BW-NEXT:    kmovd %r10d, %k1
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm15 {%k1} = ymm25[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm3, %ymm24, %ymm3
-; AVX512DQ-BW-NEXT:    vpshufb %ymm11, %ymm25, %ymm11
-; AVX512DQ-BW-NEXT:    vpor %ymm3, %ymm11, %ymm3
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm15, %zmm3, %zmm11
-; AVX512DQ-BW-NEXT:    movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
-; AVX512DQ-BW-NEXT:    kmovq %r10, %k3
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm12, %zmm11 {%k3}
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%r9), %ymm23
-; AVX512DQ-BW-NEXT:    vpshufb %ymm6, %ymm23, %ymm3
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%r8), %ymm29
-; AVX512DQ-BW-NEXT:    vpshufb %ymm7, %ymm29, %ymm6
-; AVX512DQ-BW-NEXT:    vpor %ymm3, %ymm6, %ymm3
-; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm17 {%k1} = ymm30[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm17 = ymm17[2,3,2,3]
 ; AVX512DQ-BW-NEXT:    vpshufb %ymm6, %ymm29, %ymm6
-; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm7, %ymm23, %ymm7
-; AVX512DQ-BW-NEXT:    vpor %ymm6, %ymm7, %ymm6
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm6 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512DQ-BW-NEXT:    vpermw %ymm14, %ymm6, %ymm6
-; AVX512DQ-BW-NEXT:    vpshufb %ymm1, %ymm14, %ymm1
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm6, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT:    movabsq $145249953336295682, %r10 # imm = 0x204081020408102
+; AVX512DQ-BW-NEXT:    vpshufb %ymm9, %ymm30, %ymm9
+; AVX512DQ-BW-NEXT:    vpor %ymm6, %ymm9, %ymm6
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm17, %zmm6, %zmm9
+; AVX512DQ-BW-NEXT:    movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
 ; AVX512DQ-BW-NEXT:    kmovq %r10, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k2}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm14, %zmm9 {%k2}
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%r9), %ymm27
+; AVX512DQ-BW-NEXT:    vpshufb %ymm4, %ymm27, %ymm4
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%r8), %ymm28
+; AVX512DQ-BW-NEXT:    vpshufb %ymm5, %ymm28, %ymm5
+; AVX512DQ-BW-NEXT:    vpor %ymm4, %ymm5, %ymm4
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm5, %ymm28, %ymm5
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm6, %ymm27, %ymm6
+; AVX512DQ-BW-NEXT:    vpor %ymm5, %ymm6, %ymm5
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm5 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512DQ-BW-NEXT:    vpermw %ymm23, %ymm5, %ymm5
+; AVX512DQ-BW-NEXT:    vpshufb %ymm1, %ymm23, %ymm1
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm5, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT:    movabsq $145249953336295682, %r10 # imm = 0x204081020408102
+; AVX512DQ-BW-NEXT:    kmovq %r10, %k3
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm4 {%k3}
 ; AVX512DQ-BW-NEXT:    movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
-; AVX512DQ-BW-NEXT:    kmovq %r10, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm3, %zmm11 {%k2}
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm5[u],zero,xmm5[7],zero,xmm5[5,u,u,u],zero,xmm5[8],zero,xmm5[6,u,u,u],zero
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm4[u,7],zero,xmm4[5],zero,xmm4[u,u,u,8],zero,xmm4[6],zero,xmm4[u,u,u,9]
-; AVX512DQ-BW-NEXT:    vpor %xmm1, %xmm3, %xmm1
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} xmm4 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
-; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm1, %zmm3, %zmm1
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm9[u,u,u],zero,xmm9[7],zero,xmm9[5,u,u,u],zero,xmm9[8],zero,xmm9[6,u,u]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm8[u,u,u,7],zero,xmm8[5],zero,xmm8[u,u,u,8],zero,xmm8[6],zero,xmm8[u,u]
-; AVX512DQ-BW-NEXT:    vpor %xmm3, %xmm5, %xmm3
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
-; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} xmm6 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-BW-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
-; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm3, %zmm5, %zmm3
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm15 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm3[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-NEXT:    kmovq %r10, %k3
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm4, %zmm9 {%k3}
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm3[u],zero,xmm3[7],zero,xmm3[5,u,u,u],zero,xmm3[8],zero,xmm3[6,u,u,u],zero
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm4 = xmm2[u,7],zero,xmm2[5],zero,xmm2[u,u,u,8],zero,xmm2[6],zero,xmm2[u,u,u,9]
+; AVX512DQ-BW-NEXT:    vpor %xmm1, %xmm4, %xmm1
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
+; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm1, %zmm2, %zmm1
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} xmm17 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm2
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} xmm26 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-BW-NEXT:    vpshufb %xmm26, %xmm10, %xmm4
+; AVX512DQ-BW-NEXT:    vpor %xmm2, %xmm4, %xmm2
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3],xmm10[4],xmm11[4],xmm10[5],xmm11[5],xmm10[6],xmm11[6],xmm10[7],xmm11[7]
+; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} xmm8 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-BW-NEXT:    vpshufb %xmm8, %xmm4, %xmm4
+; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm4, %zmm2
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm14 = zmm1[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm2[0,1,0,1,4,5,4,5]
 ; AVX512DQ-BW-NEXT:    movabsq $871499720017774092, %r10 # imm = 0xC183060C183060C
-; AVX512DQ-BW-NEXT:    kmovq %r10, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm15 {%k2}
-; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} xmm28 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-BW-NEXT:    vpshufb %xmm28, %xmm0, %xmm1
-; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} xmm30 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm2, %xmm3
-; AVX512DQ-BW-NEXT:    vpor %xmm1, %xmm3, %xmm1
-; 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:    vmovdqa {{.*#+}} xmm2 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
+; AVX512DQ-BW-NEXT:    kmovq %r10, %k3
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm14 {%k3}
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm1 = zero,xmm31[4,u,u,u],zero,xmm31[7],zero,xmm31[5,u,u,u],zero,xmm31[8],zero,xmm31[6]
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} xmm18 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512DQ-BW-NEXT:    vpshufb %xmm18, %xmm0, %xmm2
+; AVX512DQ-BW-NEXT:    vpor %xmm1, %xmm2, %xmm1
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm31[0],xmm0[1],xmm31[1],xmm0[2],xmm31[2],xmm0[3],xmm31[3],xmm0[4],xmm31[4],xmm0[5],xmm31[5],xmm0[6],xmm31[6],xmm0[7],xmm31[7]
+; AVX512DQ-BW-NEXT:    vmovdqa {{.*#+}} xmm12 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
 ; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm1, %zmm0, %zmm0
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-NEXT:    vmovdqa64 (%rax), %zmm3
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm10, %zmm3, %zmm20
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512DQ-BW-NEXT:    vpermw %zmm20, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT:    vmovdqa64 (%rax), %zmm5
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512DQ-BW-NEXT:    vpermw %zmm5, %zmm1, %zmm1
 ; AVX512DQ-BW-NEXT:    movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
-; AVX512DQ-BW-NEXT:    kmovq %rax, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm0 {%k2}
+; AVX512DQ-BW-NEXT:    kmovq %rax, %k3
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm0 {%k3}
 ; AVX512DQ-BW-NEXT:    movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
-; AVX512DQ-BW-NEXT:    kmovq %rax, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm0, %zmm15 {%k2}
-; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rdi), %xmm31
-; AVX512DQ-BW-NEXT:    vmovdqa 32(%rsi), %xmm12
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm31[0],xmm12[0],xmm31[1],xmm12[1],xmm31[2],xmm12[2],xmm31[3],xmm12[3],xmm31[4],xmm12[4],xmm31[5],xmm12[5],xmm31[6],xmm12[6],xmm31[7],xmm12[7]
-; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm0, %xmm0
-; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm19[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512DQ-BW-NEXT:    kmovq %rax, %k3
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm0, %zmm14 {%k3}
+; AVX512DQ-BW-NEXT:    vmovdqa 32(%rdi), %xmm6
+; AVX512DQ-BW-NEXT:    vmovdqa64 32(%rsi), %xmm31
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm6[0],xmm31[0],xmm6[1],xmm31[1],xmm6[2],xmm31[2],xmm6[3],xmm31[3],xmm6[4],xmm31[4],xmm6[5],xmm31[5],xmm6[6],xmm31[6],xmm6[7],xmm31[7]
+; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
+; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm16[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
 ; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[2,2,3,3,6,6,7,7]
 ; AVX512DQ-BW-NEXT:    movl $338170920, %eax # imm = 0x14281428
-; AVX512DQ-BW-NEXT:    kmovd %eax, %k2
-; AVX512DQ-BW-NEXT:    vpbroadcastd {{.*#+}} ymm27 = [13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm27, %ymm18, %ymm1 {%k2}
-; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[2,3,2,3],zmm0[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vmovdqa 32(%rcx), %xmm10
-; AVX512DQ-BW-NEXT:    vmovdqa 32(%rdx), %xmm7
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm7[0],xmm10[0],xmm7[1],xmm10[1],xmm7[2],xmm10[2],xmm7[3],xmm10[3],xmm7[4],xmm10[4],xmm7[5],xmm10[5],xmm7[6],xmm10[6],xmm7[7],xmm10[7]
-; AVX512DQ-BW-NEXT:    vpshufb %xmm6, %xmm1, %xmm1
-; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm4 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-BW-NEXT:    kmovd %eax, %k3
+; AVX512DQ-BW-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm7, %ymm13, %ymm1 {%k3}
+; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm1[2,3,2,3],zmm0[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vmovdqa 32(%rcx), %xmm2
+; AVX512DQ-BW-NEXT:    vmovdqa 32(%rdx), %xmm1
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = 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-BW-NEXT:    vpshufb %xmm8, %xmm0, %xmm0
+; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm4 = ymm10[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
 ; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,2,3,3,4,6,7,7]
-; AVX512DQ-BW-NEXT:    vpbroadcastq {{.*#+}} ymm26 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm9, %ymm5
-; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm4, %ymm5 {%k1}
-; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm5[2,3,2,3],zmm1[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm0, %zmm4 {%k3}
-; AVX512DQ-BW-NEXT:    vmovdqa 32(%r9), %xmm6
-; AVX512DQ-BW-NEXT:    vmovdqa 32(%r8), %xmm5
-; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
-; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm22[27],zero,zero,zero,zero,ymm22[30],zero,ymm22[28],zero,zero,zero,zero,ymm22[31],zero
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm2 = ymm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm21[27],zero,zero,zero,zero,ymm21[30],zero,ymm21[28],zero,zero,zero,zero,ymm21[31],zero,ymm21[29]
-; AVX512DQ-BW-NEXT:    vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm1[2,3,2,3],zmm0[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
-; AVX512DQ-BW-NEXT:    vpermi2w %zmm3, %zmm14, %zmm1
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm0 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
-; AVX512DQ-BW-NEXT:    vpermi2w %zmm3, %zmm14, %zmm0
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm14 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
-; AVX512DQ-BW-NEXT:    vpermw %zmm3, %zmm14, %zmm3
+; AVX512DQ-BW-NEXT:    vpbroadcastq {{.*#+}} ymm8 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm8, %ymm11, %ymm20
+; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm4, %ymm20 {%k1}
+; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm20[2,3,2,3],zmm0[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm3, %zmm0 {%k2}
+; AVX512DQ-BW-NEXT:    vmovdqa 32(%r9), %xmm4
+; AVX512DQ-BW-NEXT:    vmovdqa 32(%r8), %xmm3
+; AVX512DQ-BW-NEXT:    vpunpcklbw {{.*#+}} xmm20 = 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-BW-NEXT:    vpshufb %xmm12, %xmm20, %xmm12
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm20 = ymm25[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm25[27],zero,zero,zero,zero,ymm25[30],zero,ymm25[28],zero,zero,zero,zero,ymm25[31],zero
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm19 = ymm24[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm24[27],zero,zero,zero,zero,ymm24[30],zero,ymm24[28],zero,zero,zero,zero,ymm24[31],zero,ymm24[29]
+; AVX512DQ-BW-NEXT:    vporq %ymm20, %ymm19, %ymm19
+; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm12 = zmm19[2,3,2,3],zmm12[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm19 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,33,33,32,32,33,33,32,32,36,37,33,33,33,33,32,32]
+; AVX512DQ-BW-NEXT:    vpermi2w %zmm23, %zmm5, %zmm19
 ; AVX512DQ-BW-NEXT:    movabsq $580999813345182728, %rax # imm = 0x810204081020408
-; AVX512DQ-BW-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm3, %zmm2 {%k3}
+; AVX512DQ-BW-NEXT:    kmovq %rax, %k2
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm19, %zmm12 {%k2}
 ; AVX512DQ-BW-NEXT:    movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
-; AVX512DQ-BW-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm2, %zmm4 {%k3}
-; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm2 = ymm24[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[2,2,3,3,6,6,7,7]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm27, %ymm25, %ymm2 {%k2}
-; AVX512DQ-BW-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
-; AVX512DQ-BW-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm3, %ymm25, %ymm14
-; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} ymm25 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm25, %ymm24, %ymm24
-; AVX512DQ-BW-NEXT:    vporq %ymm14, %ymm24, %ymm14
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm2, %zmm14, %zmm2
-; AVX512DQ-BW-NEXT:    vpshufb %ymm26, %ymm17, %ymm14
-; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm24 = ymm16[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm24 = ymm24[0,2,3,3,4,6,7,7]
-; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm24, %ymm14 {%k1}
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm16 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm16[24,25],zero,ymm16[23],zero,ymm16[21,22,23,26],zero,ymm16[24],zero,ymm16[28,29,26,27]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm17 = ymm17[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm17[25],zero,ymm17[23],zero,zero,zero,zero,ymm17[26],zero,ymm17[24],zero,zero,zero,zero
-; AVX512DQ-BW-NEXT:    vporq %ymm16, %ymm17, %ymm16
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm14, %zmm16, %zmm14
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm14 = zmm14[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT:    kmovq %rax, %k2
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm12, %zmm0 {%k2}
+; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm12 = ymm29[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm12[2,2,3,3,6,6,7,7]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm7, %ymm30, %ymm12 {%k3}
+; AVX512DQ-BW-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512DQ-BW-NEXT:    # ymm7 = mem[0,1,0,1]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm7, %ymm30, %ymm19
+; AVX512DQ-BW-NEXT:    vmovdqa64 {{.*#+}} ymm20 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm20, %ymm29, %ymm29
+; AVX512DQ-BW-NEXT:    vporq %ymm19, %ymm29, %ymm19
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm12, %zmm19, %zmm12
+; AVX512DQ-BW-NEXT:    vpshufb %ymm8, %ymm22, %ymm8
+; AVX512DQ-BW-NEXT:    vpshufhw {{.*#+}} ymm19 = ymm21[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm19 = ymm19[0,2,3,3,4,6,7,7]
+; AVX512DQ-BW-NEXT:    vmovdqu8 %ymm19, %ymm8 {%k1}
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm19 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm21[24,25],zero,ymm21[23],zero,ymm21[21,22,23,26],zero,ymm21[24],zero,ymm21[28,29,26,27]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm21 = ymm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm22[25],zero,ymm22[23],zero,zero,zero,zero,ymm22[26],zero,ymm22[24],zero,zero,zero,zero
+; AVX512DQ-BW-NEXT:    vporq %ymm19, %ymm21, %ymm19
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm8, %zmm19, %zmm8
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm21 = zmm12[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-NEXT:    movabsq $1742999440035548184, %rax # imm = 0x183060C183060C18
 ; AVX512DQ-BW-NEXT:    kmovq %rax, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm14, %zmm2 {%k2}
-; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %zmm14
-; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm16 = zmm29[0,1,2,3],zmm14[4,5,6,7]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm16 = zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm16[23],zero,zmm16[23,24,25,26],zero,zmm16[24],zero,zmm16[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm16[59],zero,zero,zero,zero,zmm16[62],zero,zmm16[60],zero,zero,zero,zero,zmm16[63],zero
-; AVX512DQ-BW-NEXT:    vmovdqa64 (%r8), %zmm17
-; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm23 = zmm23[0,1,2,3],zmm17[4,5,6,7]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm23 = zmm23[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm23[25],zero,zmm23[23],zero,zero,zero,zero,zmm23[26],zero,zmm23[24],zero,zero,zmm23[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm23[59],zero,zero,zero,zero,zmm23[62],zero,zmm23[60],zero,zero,zero,zero,zmm23[63],zero,zmm23[61]
-; AVX512DQ-BW-NEXT:    vporq %zmm16, %zmm23, %zmm16
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm8, %zmm21 {%k2}
+; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %zmm8
+; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm12 = zmm28[0,1,2,3],zmm8[4,5,6,7]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm12 = zmm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm12[23],zero,zmm12[23,24,25,26],zero,zmm12[24],zero,zmm12[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm12[59],zero,zero,zero,zero,zmm12[62],zero,zmm12[60],zero,zero,zero,zero,zmm12[63],zero
+; AVX512DQ-BW-NEXT:    vmovdqa64 (%r8), %zmm19
+; AVX512DQ-BW-NEXT:    vshufi64x2 {{.*#+}} zmm22 = zmm27[0,1,2,3],zmm19[4,5,6,7]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm22 = zmm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm22[25],zero,zmm22[23],zero,zero,zero,zero,zmm22[26],zero,zmm22[24],zero,zero,zmm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm22[59],zero,zero,zero,zero,zmm22[62],zero,zmm22[60],zero,zero,zero,zero,zmm22[63],zero,zmm22[61]
+; AVX512DQ-BW-NEXT:    vporq %zmm12, %zmm22, %zmm12
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm12 = zmm12[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-NEXT:    movabsq $6971997760142192736, %rax # imm = 0x60C183060C183060
 ; AVX512DQ-BW-NEXT:    kmovq %rax, %k2
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm16, %zmm2 {%k2}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm12, %zmm21 {%k2}
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm12 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
+; AVX512DQ-BW-NEXT:    vpermw %zmm23, %zmm12, %zmm12
 ; AVX512DQ-BW-NEXT:    movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
 ; AVX512DQ-BW-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k3}
-; AVX512DQ-BW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm19[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,0,1,1,4,4,5,5]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm1 {%k1} = ymm18[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512DQ-BW-NEXT:    vpshufb %ymm3, %ymm18, %ymm3
-; AVX512DQ-BW-NEXT:    vpshufb %ymm25, %ymm19, %ymm16
-; AVX512DQ-BW-NEXT:    vporq %ymm3, %ymm16, %ymm3
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm3, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm8[18],zero,zmm8[18,19,20,21],zero,zmm8[19],zero,zmm8[25,26,27,22],zero,zmm8[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm8[56,57],zero,zmm8[55],zero,zmm8[53,54,55,58],zero,zmm8[56],zero,zmm8[60,61,58,59]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm8 = zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm9[18],zero,zero,zero,zero,zmm9[21],zero,zmm9[19],zero,zero,zero,zero,zmm9[22],zero,zmm9[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm9[57],zero,zmm9[55],zero,zero,zero,zero,zmm9[58],zero,zmm9[56],zero,zero,zero,zero
-; AVX512DQ-BW-NEXT:    vporq %zmm3, %zmm8, %zmm3
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k2}
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm21, %zmm17, %zmm1
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[20],zero,zmm1[18],zero,zmm1[20,21,20,21],zero,zmm1[19],zero,zmm1[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[56,57,56,57],zero,zmm1[55],zero,zmm1[55,56,57,58],zero,zmm1[56],zero,zmm1[62,63]
-; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm22, %zmm14, %zmm8
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm12, %zmm21 {%k3}
+; AVX512DQ-BW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm16[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-BW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm12[0,0,1,1,4,4,5,5]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} ymm12 {%k1} = ymm13[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512DQ-BW-NEXT:    vpshufb %ymm7, %ymm13, %ymm7
+; AVX512DQ-BW-NEXT:    vpshufb %ymm20, %ymm16, %ymm13
+; AVX512DQ-BW-NEXT:    vpor %ymm7, %ymm13, %ymm7
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm12, %zmm7
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm10 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm10[18],zero,zmm10[18,19,20,21],zero,zmm10[19],zero,zmm10[25,26,27,22],zero,zmm10[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm10[56,57],zero,zmm10[55],zero,zmm10[53,54,55,58],zero,zmm10[56],zero,zmm10[60,61,58,59]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm11 = zmm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm11[18],zero,zero,zero,zero,zmm11[21],zero,zmm11[19],zero,zero,zero,zero,zmm11[22],zero,zmm11[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm11[57],zero,zmm11[55],zero,zero,zero,zero,zmm11[58],zero,zmm11[56],zero,zero,zero,zero
+; AVX512DQ-BW-NEXT:    vporq %zmm10, %zmm11, %zmm10
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm7, %zmm10 {%k2}
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm24, %zmm19, %zmm7
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm7 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm7[20],zero,zmm7[18],zero,zmm7[20,21,20,21],zero,zmm7[19],zero,zmm7[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm7[56,57,56,57],zero,zmm7[55],zero,zmm7[55,56,57,58],zero,zmm7[56],zero,zmm7[62,63]
+; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm25, %zmm8, %zmm8
 ; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} zmm8 = zmm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm8[20],zero,zmm8[18],zero,zero,zero,zero,zmm8[21],zero,zmm8[19],zero,zero,zero,zero,zmm8[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm8[57],zero,zmm8[55],zero,zero,zero,zero,zmm8[58],zero,zmm8[56],zero,zero
-; AVX512DQ-BW-NEXT:    vporq %zmm1, %zmm8, %zmm1
-; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm8 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512DQ-BW-NEXT:    vpermw %zmm20, %zmm8, %zmm8
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT:    vporq %zmm7, %zmm8, %zmm7
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm8 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512DQ-BW-NEXT:    vpermw %zmm5, %zmm8, %zmm5
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-NEXT:    movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
 ; AVX512DQ-BW-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm8, %zmm1 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm5, %zmm7 {%k1}
 ; AVX512DQ-BW-NEXT:    movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
 ; AVX512DQ-BW-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k1}
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm10[u,u,u],zero,xmm10[7],zero,xmm10[5,u,u,u],zero,xmm10[8],zero,xmm10[6,u,u]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm8 = xmm7[u,u,u,7],zero,xmm7[5],zero,xmm7[u,u,u,8],zero,xmm7[6],zero,xmm7[u,u]
-; AVX512DQ-BW-NEXT:    vpor %xmm1, %xmm8, %xmm1
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm10[8],xmm7[8],xmm10[9],xmm7[9],xmm10[10],xmm7[10],xmm10[11],xmm7[11],xmm10[12],xmm7[12],xmm10[13],xmm7[13],xmm10[14],xmm7[14],xmm10[15],xmm7[15]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm7 = xmm7[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm7, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm7 = xmm12[u],zero,xmm12[7],zero,xmm12[5,u,u,u],zero,xmm12[8],zero,xmm12[6,u,u,u],zero
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm8 = xmm31[u,7],zero,xmm31[5],zero,xmm31[u,u,u,8],zero,xmm31[6],zero,xmm31[u,u,u,9]
-; AVX512DQ-BW-NEXT:    vpor %xmm7, %xmm8, %xmm7
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm8 = xmm12[8],xmm31[8],xmm12[9],xmm31[9],xmm12[10],xmm31[10],xmm12[11],xmm31[11],xmm12[12],xmm31[12],xmm12[13],xmm31[13],xmm12[14],xmm31[14],xmm12[15],xmm31[15]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm8 = xmm8[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm8, %zmm7, %zmm7
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm7, %zmm10 {%k1}
+; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm2, %xmm5
+; AVX512DQ-BW-NEXT:    vpshufb %xmm26, %xmm1, %xmm7
+; AVX512DQ-BW-NEXT:    vpor %xmm5, %xmm7, %xmm5
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm1, %zmm5, %zmm1
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm2 = xmm31[u],zero,xmm31[7],zero,xmm31[5,u,u,u],zero,xmm31[8],zero,xmm31[6,u,u,u],zero
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm6[u,7],zero,xmm6[5],zero,xmm6[u,u,u,8],zero,xmm6[6],zero,xmm6[u,u,u,9]
+; AVX512DQ-BW-NEXT:    vpor %xmm2, %xmm5, %xmm2
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm31[8],xmm6[8],xmm31[9],xmm6[9],xmm31[10],xmm6[10],xmm31[11],xmm6[11],xmm31[12],xmm6[12],xmm31[13],xmm6[13],xmm31[14],xmm6[14],xmm31[15],xmm6[15]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm5, %zmm2, %zmm2
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm7 = zmm7[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm7 {%k2}
-; AVX512DQ-BW-NEXT:    vpshufb %xmm28, %xmm6, %xmm1
-; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm5, %xmm8
-; AVX512DQ-BW-NEXT:    vpor %xmm1, %xmm8, %xmm1
-; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm6[8],xmm5[9],xmm6[9],xmm5[10],xmm6[10],xmm5[11],xmm6[11],xmm5[12],xmm6[12],xmm5[13],xmm6[13],xmm5[14],xmm6[14],xmm5[15],xmm6[15]
-; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm5[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm5, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k2}
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm1 = zero,xmm4[4,u,u,u],zero,xmm4[7],zero,xmm4[5,u,u,u],zero,xmm4[8],zero,xmm4[6]
+; AVX512DQ-BW-NEXT:    vpshufb %xmm18, %xmm3, %xmm5
+; AVX512DQ-BW-NEXT:    vpor %xmm1, %xmm5, %xmm1
+; AVX512DQ-BW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
+; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm3[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm3, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512DQ-BW-NEXT:    vpermw %zmm23, %zmm3, %zmm3
 ; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
 ; AVX512DQ-BW-NEXT:    movabsq $290499906672591364, %rax # imm = 0x408102040810204
 ; AVX512DQ-BW-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm3, %zmm1 {%k1}
 ; AVX512DQ-BW-NEXT:    movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
 ; AVX512DQ-BW-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm7 {%k1}
+; AVX512DQ-BW-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k1}
 ; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, 320(%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, 256(%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, 128(%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, 384(%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, 192(%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, (%rax)
-; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, 64(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, 320(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, 256(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, 128(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, 384(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, 192(%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, (%rax)
+; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, 64(%rax)
 ; AVX512DQ-BW-NEXT:    vzeroupper
 ; AVX512DQ-BW-NEXT:    retq
 ;
 ; AVX512DQ-BW-FCP-LABEL: store_i8_stride7_vf64:
 ; AVX512DQ-BW-FCP:       # %bb.0:
-; AVX512DQ-BW-FCP-NEXT:    subq $104, %rsp
+; AVX512DQ-BW-FCP-NEXT:    subq $72, %rsp
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rax), %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm2, (%rsp) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} ymm20 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm20, %ymm2, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
-; AVX512DQ-BW-FCP-NEXT:    # ymm1 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm2, %ymm1, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rax), %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} ymm19 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm19, %ymm0, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512DQ-BW-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm0, %ymm2, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r9), %ymm1
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm0 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm1, %ymm3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %ymm4
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm4, %ymm4
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r9), %xmm16
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %xmm15
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm15[8],xmm16[8],xmm15[9],xmm16[9],xmm15[10],xmm16[10],xmm15[11],xmm16[11],xmm15[12],xmm16[12],xmm15[13],xmm16[13],xmm15[14],xmm16[14],xmm15[15],xmm16[15]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm4[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm14 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm14, %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm2, (%rsp) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm15 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm15, %ymm2, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r9), %xmm13
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %xmm12
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm12[8],xmm13[8],xmm12[9],xmm13[9],xmm12[10],xmm13[10],xmm12[11],xmm13[11],xmm12[12],xmm13[12],xmm12[13],xmm13[13],xmm12[14],xmm13[14],xmm12[15],xmm13[15]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm16 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm16, %xmm2, %xmm2
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $2323999253380730912, %r10 # imm = 0x2040810204081020
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm3 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm0 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm2, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %ymm4
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm2, %ymm4, %ymm4
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rcx), %ymm6
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm4, %ymm6, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm3, %ymm4, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm2, %ymm4, %ymm2
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdx), %xmm17
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rcx), %xmm19
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm19[8],xmm17[8],xmm19[9],xmm17[9],xmm19[10],xmm17[10],xmm19[11],xmm17[11],xmm19[12],xmm17[12],xmm19[13],xmm17[13],xmm19[14],xmm17[14],xmm19[15],xmm17[15]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rcx), %xmm18
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm18[8],xmm17[8],xmm18[9],xmm17[9],xmm18[10],xmm17[10],xmm18[11],xmm17[11],xmm18[12],xmm17[12],xmm18[13],xmm17[13],xmm18[14],xmm17[14],xmm18[15],xmm17[15]
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm29 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm29, %xmm6, %xmm6
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm6, %zmm4
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm6
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm9 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm6, %ymm6
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %ymm7
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} ymm23 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm23, %ymm7, %ymm21
-; AVX512DQ-BW-FCP-NEXT:    vporq %ymm6, %ymm21, %ymm6
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %xmm21
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rsi), %xmm22
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm24 = xmm22[8],xmm21[8],xmm22[9],xmm21[9],xmm22[10],xmm21[10],xmm22[11],xmm21[11],xmm22[12],xmm21[12],xmm22[13],xmm21[13],xmm22[14],xmm21[14],xmm22[15],xmm21[15]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm24, %xmm24
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm24 = ymm24[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm24, %zmm14
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm29, %xmm4, %xmm4
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} ymm8 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm8, %ymm4, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rsi), %ymm5
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} ymm22 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm22, %ymm5, %ymm20
+; AVX512DQ-BW-FCP-NEXT:    vporq %ymm4, %ymm20, %ymm4
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %xmm20
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rsi), %xmm21
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm23 = xmm21[8],xmm20[8],xmm21[9],xmm20[9],xmm21[10],xmm20[10],xmm21[11],xmm20[11],xmm21[12],xmm20[12],xmm21[13],xmm20[13],xmm21[14],xmm20[14],xmm21[15],xmm20[15]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} xmm30 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm30, %xmm23, %xmm23
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm23 = ymm23[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm23, %zmm11
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm14 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm11 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm3, %zmm14 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdx), %ymm3
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm2, %ymm3, %ymm4
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rcx), %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm5
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm4, %ymm5, %ymm4
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm5, %ymm3, %ymm5
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm24
-; AVX512DQ-BW-FCP-NEXT:    vporq %ymm5, %ymm24, %ymm5
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdi), %ymm5
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm5, %ymm9
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm11 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdx), %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm1, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rcx), %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm3, %ymm0, %ymm3
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm3, %ymm1, %ymm3
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm0, %ymm23
+; AVX512DQ-BW-FCP-NEXT:    vporq %ymm3, %ymm23, %ymm3
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdi), %ymm3
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm8, %ymm3, %ymm8
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%rsi), %ymm25
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm23, %ymm25, %ymm23
-; AVX512DQ-BW-FCP-NEXT:    vporq %ymm9, %ymm23, %ymm9
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm23
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm25, %ymm24
-; AVX512DQ-BW-FCP-NEXT:    vporq %ymm23, %ymm24, %ymm23
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm23, %zmm9, %zmm9
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm22, %ymm25, %ymm22
+; AVX512DQ-BW-FCP-NEXT:    vporq %ymm8, %ymm22, %ymm8
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm3, %ymm22
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm25, %ymm23
+; AVX512DQ-BW-FCP-NEXT:    vporq %ymm22, %ymm23, %ymm22
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm22 = ymm22[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm22, %zmm8, %zmm8
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm9 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r9), %ymm4
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm8 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r9), %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm14, %ymm2, %ymm14
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%r8), %ymm28
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm28, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm28, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm4, %ymm23
-; AVX512DQ-BW-FCP-NEXT:    vporq %ymm1, %ymm23, %ymm1
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rax), %ymm6
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm23 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm6, %ymm23, %ymm23
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm20, %ymm6, %ymm20
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm23, %zmm20, %zmm20
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm15, %ymm28, %ymm15
+; AVX512DQ-BW-FCP-NEXT:    vporq %ymm14, %ymm15, %ymm22
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm28, %ymm23
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm2, %ymm24
+; AVX512DQ-BW-FCP-NEXT:    vporq %ymm23, %ymm24, %ymm23
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm23, %zmm22, %zmm23
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%rax), %ymm24
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} ymm22 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %ymm24, %ymm22, %ymm22
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm19, %ymm24, %ymm19
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm22, %zmm19, %zmm19
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $145249953336295682, %r10 # imm = 0x204081020408102
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %r10, %k2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm20, %zmm1 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm19, %zmm23 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdx), %zmm22
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %r10, %k2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm9 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm23, %zmm8 {%k2}
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rcx), %zmm23
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm23[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm23[4,5,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,zmm1[23],zero,zmm1[21,22,23,26],zero,zmm1[24],zero,zmm1[28,29,26,27,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero,zmm1[61],zero
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdx), %zmm24
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm24[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm2[25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zero,zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[62],zero,zmm2[60],zero,zero,zero,zero,zmm2[63],zero,zmm2[61],zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vporq %zmm1, %zmm2, %zmm1
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm22[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm0[25],zero,zmm0[23],zero,zero,zero,zero,zmm0[26],zero,zmm0[24],zero,zero,zero,zero,zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm0[62],zero,zmm0[60],zero,zero,zero,zero,zmm0[63],zero,zmm0[61],zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vporq %zmm1, %zmm0, %zmm0
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm26
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm25[0,1,2,3],zmm26[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zero,zero,zmm2[27],zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,60,61,62],zero,zmm2[60],zero,zmm2[62,63,62,63],zero,zmm2[61],zero,zmm2[63,60,61]
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm25[0,1,2,3],zmm26[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,zmm1[23],zero,zero,zero,zero,zmm1[26],zero,zmm1[24],zero,zero,zero,zero,zmm1[27],zero,zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,60,61,62],zero,zmm1[60],zero,zmm1[62,63,62,63],zero,zmm1[61],zero,zmm1[63,60,61]
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rsi), %zmm27
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm5[0,1,2,3],zmm27[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm27[4,5,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[23],zero,zero,zero,zero,zmm3[26],zero,zmm3[24],zero,zero,zero,zero,zmm3[27],zero,zmm3[25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zmm3[62],zero,zmm3[60],zero,zero,zero,zero,zmm3[63],zero,zmm3[61],zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vporq %zmm2, %zmm3, %zmm2
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm20 = zmm2[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vporq %zmm1, %zmm3, %zmm1
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm19 = zmm1[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %r10, %k2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm19 {%k2}
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm25
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm28[0,1,2,3],zmm25[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm1[23],zero,zmm1[23,24,25,26],zero,zmm1[24],zero,zmm1[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm1[59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm28[0,1,2,3],zmm25[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm0[23],zero,zmm0[23,24,25,26],zero,zmm0[24],zero,zmm0[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm0[59],zero,zero,zero,zero,zmm0[62],zero,zmm0[60],zero,zero,zero,zero,zmm0[63],zero
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm28
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm4[0,1,2,3],zmm28[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm2[59],zero,zero,zero,zero,zmm2[62],zero,zmm2[60],zero,zero,zero,zero,zmm2[63],zero,zmm2[61]
-; AVX512DQ-BW-FCP-NEXT:    vporq %zmm1, %zmm2, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm28[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm1[25],zero,zmm1[23],zero,zero,zero,zero,zmm1[26],zero,zmm1[24],zero,zero,zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm1[59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero,zmm1[61]
+; AVX512DQ-BW-FCP-NEXT:    vporq %zmm0, %zmm1, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $6971997760142192736, %r10 # imm = 0x60C183060C183060
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %r10, %k2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm20 {%k2}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rax), %zmm31
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm1 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
-; AVX512DQ-BW-FCP-NEXT:    vpermi2w %zmm31, %zmm6, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
-; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm20 {%k3}
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdx), %xmm5
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm19 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm0 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm24, %zmm0, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    movabsq $-9150747060186627967, %r10 # imm = 0x8102040810204081
+; AVX512DQ-BW-FCP-NEXT:    kmovq %r10, %k3
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm19 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%rdx), %xmm31
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rcx), %xmm4
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm13 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm4, %xmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm12 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm5, %xmm2
-; AVX512DQ-BW-FCP-NEXT:    vpor %xmm1, %xmm2, %xmm1
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm29, %xmm2, %xmm2
-; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm1, %zmm3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdi), %xmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm4, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm9 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm31, %xmm1
+; AVX512DQ-BW-FCP-NEXT:    vpor %xmm0, %xmm1, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm31[8],xmm4[9],xmm31[9],xmm4[10],xmm31[10],xmm4[11],xmm31[11],xmm4[12],xmm31[12],xmm4[13],xmm31[13],xmm4[14],xmm31[14],xmm4[15],xmm31[15]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm29, %xmm1, %xmm1
+; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm1, %zmm0, %zmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdi), %xmm3
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rsi), %xmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm29
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm8 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm30
-; AVX512DQ-BW-FCP-NEXT:    vporq %xmm29, %xmm30, %xmm29
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm30 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm30, %xmm30
-; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm30, %zmm29, %zmm29
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm3 = zmm3[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm29 = zmm29[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm3, %zmm29 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm29
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm3, %xmm14
+; AVX512DQ-BW-FCP-NEXT:    vporq %xmm29, %xmm14, %xmm14
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm29 = xmm1[8],xmm3[8],xmm1[9],xmm3[9],xmm1[10],xmm3[10],xmm1[11],xmm3[11],xmm1[12],xmm3[12],xmm1[13],xmm3[13],xmm1[14],xmm3[14],xmm1[15],xmm3[15]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm30, %xmm29, %xmm29
+; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm29, %zmm14, %zmm14
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm29 = zmm14[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm29 {%k2}
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 32(%r9), %xmm30
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r8), %xmm3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm30, %xmm0
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm7 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm18
-; AVX512DQ-BW-FCP-NEXT:    vporq %xmm0, %xmm18, %xmm0
-; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm18 = xmm3[8],xmm30[8],xmm3[9],xmm30[9],xmm3[10],xmm30[10],xmm3[11],xmm30[11],xmm3[12],xmm30[12],xmm3[13],xmm30[13],xmm3[14],xmm30[14],xmm3[15],xmm30[15]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm18 = xmm18[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm18, %zmm0, %zmm0
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm18 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
-; AVX512DQ-BW-FCP-NEXT:    vpermi2w %zmm31, %zmm6, %zmm18
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%r8), %xmm2
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm14 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm14, %xmm30, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm5 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm15
+; AVX512DQ-BW-FCP-NEXT:    vpor %xmm0, %xmm15, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vpunpckhbw {{.*#+}} xmm15 = xmm2[8],xmm30[8],xmm2[9],xmm30[9],xmm2[10],xmm30[10],xmm2[11],xmm30[11],xmm2[12],xmm30[12],xmm2[13],xmm30[13],xmm2[14],xmm30[14],xmm2[15],xmm30[15]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm16, %xmm15, %xmm15
+; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm15, %zmm0, %zmm0
 ; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT:    movabsq $290499906672591364, %rax # imm = 0x408102040810204
-; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm18, %zmm0 {%k3}
-; AVX512DQ-BW-FCP-NEXT:    movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
-; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k3
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm15 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm24, %zmm15, %zmm15
+; AVX512DQ-BW-FCP-NEXT:    movabsq $290499906672591364, %rcx # imm = 0x408102040810204
+; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k3
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm15, %zmm0 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    movabsq $-8714997200177740921, %rcx # imm = 0x870E1C3870E1C387
+; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k3
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm29 {%k3}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm19, %xmm0
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm17, %xmm6
-; AVX512DQ-BW-FCP-NEXT:    vpor %xmm0, %xmm6, %xmm0
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm17[0],xmm19[0],xmm17[1],xmm19[1],xmm17[2],xmm19[2],xmm17[3],xmm19[3],xmm17[4],xmm19[4],xmm17[5],xmm19[5],xmm17[6],xmm19[6],xmm17[7],xmm19[7]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm12 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm6, %xmm6
-; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm6, %zmm0
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm22, %xmm6
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm21, %xmm8
-; AVX512DQ-BW-FCP-NEXT:    vpor %xmm6, %xmm8, %xmm6
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm21[0],xmm22[0],xmm21[1],xmm22[1],xmm21[2],xmm22[2],xmm21[3],xmm22[3],xmm21[4],xmm22[4],xmm21[5],xmm22[5],xmm21[6],xmm22[6],xmm21[7],xmm22[7]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm11 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm8, %xmm8
-; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm8, %zmm6
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm8 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm6[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT:    movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
-; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm8, %zmm0 {%k3}
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm16, %xmm6
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm15, %xmm7
-; AVX512DQ-BW-FCP-NEXT:    vpor %xmm6, %xmm7, %xmm6
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm15[0],xmm16[0],xmm15[1],xmm16[1],xmm15[2],xmm16[2],xmm15[3],xmm16[3],xmm15[4],xmm16[4],xmm15[5],xmm16[5],xmm15[6],xmm16[6],xmm15[7],xmm16[7]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm7, %xmm7
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm18, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm17, %xmm9
+; AVX512DQ-BW-FCP-NEXT:    vpor %xmm0, %xmm9, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm9 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3],xmm17[4],xmm18[4],xmm17[5],xmm18[5],xmm17[6],xmm18[6],xmm17[7],xmm18[7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm10 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm9, %xmm9
+; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm9, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm21, %xmm7
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm20, %xmm6
+; AVX512DQ-BW-FCP-NEXT:    vpor %xmm7, %xmm6, %xmm6
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm20[0],xmm21[0],xmm20[1],xmm21[1],xmm20[2],xmm21[2],xmm20[3],xmm21[3],xmm20[4],xmm21[4],xmm20[5],xmm21[5],xmm20[6],xmm21[6],xmm20[7],xmm21[7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm9 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm7, %xmm7
 ; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm7, %zmm6
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm6 = zmm6[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, (%rsp), %zmm31, %zmm8 # 32-byte Folded Reload
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm7 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm8, %zmm7, %zmm7
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm16 = zmm6[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT:    movabsq $871499720017774092, %rcx # imm = 0xC183060C183060C
+; AVX512DQ-BW-FCP-NEXT:    kmovq %rcx, %k3
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm16 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm14, %xmm13, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm5, %xmm12, %xmm5
+; AVX512DQ-BW-FCP-NEXT:    vpor %xmm0, %xmm5, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3],xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa {{.*#+}} xmm6 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
+; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm5, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rax), %zmm7
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm5 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm7, %zmm5, %zmm5
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm7, %zmm6 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm5, %zmm0 {%k3}
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm6, %zmm0 {%k3}
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm1
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm16 {%k3}
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm0 = 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:    vpshufb %xmm9, %xmm0, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm12[28],zero,ymm12[30,31,30,31],zero,ymm12[29],zero,ymm12[31,28,29]
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm13[28],zero,ymm13[30,31,30,31],zero,ymm13[29],zero,ymm13[31,28,29]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm6 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm15[30],zero,ymm15[28],zero,zero,zero,zero,ymm15[31],zero,ymm15[29],zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm2, %ymm6, %ymm2
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm2[2,3,2,3],zmm1[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm2
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero,ymm11[29],zero
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29],zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vpor %ymm4, %ymm5, %ymm4
-; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm4[2,3,2,3],zmm2[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm30[0],xmm3[1],xmm30[1],xmm3[2],xmm30[2],xmm3[3],xmm30[3],xmm3[4],xmm30[4],xmm3[5],xmm30[5],xmm3[6],xmm30[6],xmm3[7],xmm30[7]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm13[30],zero,ymm13[28],zero,zero,zero,zero,ymm13[31],zero,ymm13[29],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm1, %ymm3, %ymm1
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[2,3,2,3],zmm0[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm31[0],xmm4[0],xmm31[1],xmm4[1],xmm31[2],xmm4[2],xmm31[3],xmm4[3],xmm31[4],xmm4[4],xmm31[5],xmm4[5],xmm31[6],xmm4[6],xmm31[7],xmm4[7]
 ; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm10[30],zero,ymm10[28],zero,zero,zero,zero,ymm10[31],zero,ymm10[29],zero
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero,ymm9[29],zero,zero
 ; AVX512DQ-BW-FCP-NEXT:    vpor %ymm3, %ymm4, %ymm3
 ; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm3[2,3,2,3],zmm1[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm31, %zmm3, %zmm3
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm30[0],xmm2[1],xmm30[1],xmm2[2],xmm30[2],xmm2[3],xmm30[3],xmm2[4],xmm30[4],xmm2[5],xmm30[5],xmm2[6],xmm30[6],xmm2[7],xmm30[7]
+; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm0
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm4[27],zero,zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu (%rsp), %ymm5 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29]
+; AVX512DQ-BW-FCP-NEXT:    vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-BW-FCP-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm2[2,3,2,3],zmm0[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm2 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,33,33,32,32,33,33,32,32,36,37,33,33,33,33,32,32]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2w %zmm24, %zmm7, %zmm2
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $580999813345182728, %rax # imm = 0x810204081020408
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm3, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm2, %zmm0 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm2 {%k1}
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm15, %zmm26, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[18,19,20,21],zero,zmm1[19],zero,zmm1[21,20,21,22],zero,zmm1[20],zero,zmm1[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm1[55],zero,zero,zero,zero,zmm1[58],zero,zmm1[56],zero,zero,zero,zero,zmm1[59],zero
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm27, %zmm3
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22],zero,zmm3[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero,zero,zero,zmm3[59],zero,zmm3[57]
-; AVX512DQ-BW-FCP-NEXT:    vporq %zmm1, %zmm3, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm7, %zmm24, %zmm3
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[18],zero,zmm3[18,19,20,21],zero,zmm3[19],zero,zmm3[25,26,27,22],zero,zmm3[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[56,57],zero,zmm3[55],zero,zmm3[53,54,55,58],zero,zmm3[56],zero,zmm3[60,61,58,59]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm11, %zmm23, %zmm4
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm4[18],zero,zero,zero,zero,zmm4[21],zero,zmm4[19],zero,zero,zero,zero,zmm4[22],zero,zmm4[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm4[57],zero,zmm4[55],zero,zero,zero,zero,zmm4[58],zero,zmm4[56],zero,zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vporq %zmm3, %zmm4, %zmm3
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k2}
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm6, %zmm28, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[20],zero,zmm1[18],zero,zmm1[20,21,20,21],zero,zmm1[19],zero,zmm1[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[56,57,56,57],zero,zmm1[55],zero,zmm1[55,56,57,58],zero,zmm1[56],zero,zmm1[62,63]
-; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm25, %zmm4
-; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm4[20],zero,zmm4[18],zero,zero,zero,zero,zmm4[21],zero,zmm4[19],zero,zero,zero,zero,zmm4[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm4[57],zero,zmm4[55],zero,zero,zero,zero,zmm4[58],zero,zmm4[56],zero,zero
-; AVX512DQ-BW-FCP-NEXT:    vporq %zmm1, %zmm4, %zmm1
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm4 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm8, %zmm4, %zmm4
-; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm13, %zmm26, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[18,19,20,21],zero,zmm0[19],zero,zmm0[21,20,21,22],zero,zmm0[20],zero,zmm0[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm0[55],zero,zero,zero,zero,zmm0[58],zero,zmm0[56],zero,zero,zero,zero,zmm0[59],zero
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm12, %zmm27, %zmm2
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[21],zero,zmm2[19],zero,zero,zero,zero,zmm2[22],zero,zmm2[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[55],zero,zero,zero,zero,zmm2[58],zero,zmm2[56],zero,zero,zero,zero,zmm2[59],zero,zmm2[57]
+; AVX512DQ-BW-FCP-NEXT:    vporq %zmm0, %zmm2, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm22, %zmm2
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[18],zero,zmm2[18,19,20,21],zero,zmm2[19],zero,zmm2[25,26,27,22],zero,zmm2[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[56,57],zero,zmm2[55],zero,zmm2[53,54,55,58],zero,zmm2[56],zero,zmm2[60,61,58,59]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm10, %zmm23, %zmm3
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[18],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22],zero,zmm3[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm3[57],zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vporq %zmm2, %zmm3, %zmm2
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm2 {%k2}
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm5, %zmm28, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[20],zero,zmm0[18],zero,zmm0[20,21,20,21],zero,zmm0[19],zero,zmm0[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[56,57,56,57],zero,zmm0[55],zero,zmm0[55,56,57,58],zero,zmm0[56],zero,zmm0[62,63]
+; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm25, %zmm3
+; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[20],zero,zmm3[18],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm3[57],zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero
+; AVX512DQ-BW-FCP-NEXT:    vporq %zmm0, %zmm3, %zmm0
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbw {{.*#+}} zmm3 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512DQ-BW-FCP-NEXT:    vpermw %zmm7, %zmm3, %zmm3
+; AVX512DQ-BW-FCP-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm4, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm3, %zmm0 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
 ; AVX512DQ-BW-FCP-NEXT:    kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm1, %zmm3 {%k1}
+; AVX512DQ-BW-FCP-NEXT:    vmovdqu8 %zmm0, %zmm2 {%k1}
 ; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, 128(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, 320(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, 128(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, 320(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, 192(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, (%rax)
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, 256(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, 64(%rax)
-; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, 384(%rax)
-; AVX512DQ-BW-FCP-NEXT:    addq $104, %rsp
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, 64(%rax)
+; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, 384(%rax)
+; AVX512DQ-BW-FCP-NEXT:    addq $72, %rsp
 ; AVX512DQ-BW-FCP-NEXT:    vzeroupper
 ; AVX512DQ-BW-FCP-NEXT:    retq
   %in.vec0 = load <64 x i8>, ptr %in.vecptr0, align 64
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
index b0a91bfe690b7..2c1d219bec419 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
@@ -1033,25 +1033,24 @@ define void @store_i8_stride8_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
 ; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [1,3,5,7,1,3,5,7]
-; AVX512-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm2, %ymm3
-; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm4 = [u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15]
-; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
-; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm2, %ymm2
-; AVX512-FCP-NEXT:    vpmovsxdq {{.*#+}} ymm5 = [201851904,218694913,235537922,252380931]
-; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
-; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2],ymm3[3],ymm2[4],ymm3[5],ymm2[6],ymm3[7]
-; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [0,2,4,6,0,2,4,6]
-; AVX512-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm3, %ymm1
-; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm1, %ymm1
-; AVX512-FCP-NEXT:    vpermd %ymm0, %ymm3, %ymm0
-; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm0, %ymm0
-; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
-; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [1,3,9,11,1,3,9,11]
+; AVX512-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vmovdqa %ymm2, %ymm5
+; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm4, %ymm5
+; AVX512-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15]
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm5
+; AVX512-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm4
+; AVX512-FCP-NEXT:    vpmovsxdq {{.*#+}} ymm7 = [201851904,218694913,235537922,252380931]
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm4, %ymm4
+; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2],ymm5[3],ymm4[4],ymm5[5],ymm4[6],ymm5[7]
+; AVX512-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm5 = [0,2,8,10,0,2,8,10]
+; AVX512-FCP-NEXT:    # ymm5 = mem[0,1,0,1]
+; AVX512-FCP-NEXT:    vpermt2d %ymm3, %ymm5, %ymm2
+; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
+; AVX512-FCP-NEXT:    vpermt2d %ymm1, %ymm5, %ymm0
+; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
+; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7]
+; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm0
 ; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
 ; AVX512-FCP-NEXT:    vzeroupper
 ; AVX512-FCP-NEXT:    retq
@@ -1117,25 +1116,24 @@ define void @store_i8_stride8_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
 ; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
 ; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
 ; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [1,3,5,7,1,3,5,7]
-; AVX512DQ-FCP-NEXT:    # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm1, %ymm2, %ymm3
-; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm4 = [u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
-; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT:    vpmovsxdq {{.*#+}} ymm5 = [201851904,218694913,235537922,252380931]
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2],ymm3[3],ymm2[4],ymm3[5],ymm2[6],ymm3[7]
-; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [0,2,4,6,0,2,4,6]
-; AVX512DQ-FCP-NEXT:    # ymm3 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT:    vpermd %ymm1, %ymm3, %ymm1
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT:    vpermd %ymm0, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
-; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [1,3,9,11,1,3,9,11]
+; AVX512DQ-FCP-NEXT:    # ymm4 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm5
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm3, %ymm4, %ymm5
+; AVX512DQ-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm5
+; AVX512DQ-FCP-NEXT:    vpermi2d %ymm1, %ymm0, %ymm4
+; AVX512DQ-FCP-NEXT:    vpmovsxdq {{.*#+}} ymm7 = [201851904,218694913,235537922,252380931]
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2],ymm5[3],ymm4[4],ymm5[5],ymm4[6],ymm5[7]
+; AVX512DQ-FCP-NEXT:    vbroadcasti128 {{.*#+}} ymm5 = [0,2,8,10,0,2,8,10]
+; AVX512DQ-FCP-NEXT:    # ymm5 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm3, %ymm5, %ymm2
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
+; AVX512DQ-FCP-NEXT:    vpermt2d %ymm1, %ymm5, %ymm0
+; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7]
+; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm0
 ; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
 ; AVX512DQ-FCP-NEXT:    vzeroupper
 ; AVX512DQ-FCP-NEXT:    retq
diff --git a/llvm/test/CodeGen/X86/vector-reduce-ctpop.ll b/llvm/test/CodeGen/X86/vector-reduce-ctpop.ll
index 2511fbf99fc46..4d12edde75176 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-ctpop.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-ctpop.ll
@@ -1545,7 +1545,9 @@ define <8 x i32> @reduce_ctpop_v4i64_buildvector_v8i32(<4 x i64> %a0, <4 x i64>
 ; AVX512VL-NEXT:    vpmovqb %ymm1, %xmm1
 ; AVX512VL-NEXT:    vpsadbw %xmm1, %xmm9, %xmm1
 ; AVX512VL-NEXT:    vpmovqb %ymm2, %xmm2
+; AVX512VL-NEXT:    vpsadbw %xmm2, %xmm9, %xmm2
 ; AVX512VL-NEXT:    vpmovqb %ymm3, %xmm3
+; AVX512VL-NEXT:    vpsadbw %xmm3, %xmm9, %xmm3
 ; AVX512VL-NEXT:    vpmovqb %ymm4, %xmm4
 ; AVX512VL-NEXT:    vpmovqb %ymm5, %xmm5
 ; AVX512VL-NEXT:    vpmovqb %ymm6, %xmm6
@@ -1553,16 +1555,14 @@ define <8 x i32> @reduce_ctpop_v4i64_buildvector_v8i32(<4 x i64> %a0, <4 x i64>
 ; AVX512VL-NEXT:    vinserti128 $1, %xmm7, %ymm6, %ymm6
 ; AVX512VL-NEXT:    vpsadbw %ymm0, %ymm6, %ymm6
 ; AVX512VL-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm4
-; AVX512VL-NEXT:    vpsadbw %ymm0, %ymm4, %ymm4
-; AVX512VL-NEXT:    vpmovsxbd {{.*#+}} ymm5 = [0,0,0,0,0,4,8,12]
-; AVX512VL-NEXT:    vpermi2d %ymm6, %ymm4, %ymm5
-; AVX512VL-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1]
-; AVX512VL-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
-; AVX512VL-NEXT:    vpsadbw %ymm0, %ymm2, %ymm0
-; AVX512VL-NEXT:    vpbroadcastq {{.*#+}} xmm2 = [0,4,0,4]
-; AVX512VL-NEXT:    vpermd %ymm0, %ymm2, %ymm0
-; AVX512VL-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
-; AVX512VL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
+; AVX512VL-NEXT:    vpsadbw %ymm0, %ymm4, %ymm0
+; AVX512VL-NEXT:    vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,4,8,12]
+; AVX512VL-NEXT:    vpermi2d %ymm6, %ymm0, %ymm4
+; AVX512VL-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm8[0],xmm1[0],xmm8[1],xmm1[1]
+; AVX512VL-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [0,8,0,8]
+; AVX512VL-NEXT:    vpermi2d %ymm3, %ymm2, %ymm1
+; AVX512VL-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
+; AVX512VL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
 ; AVX512VL-NEXT:    retq
 ;
 ; AVX512VPOPCNT-LABEL: reduce_ctpop_v4i64_buildvector_v8i32:
@@ -1581,22 +1581,22 @@ define <8 x i32> @reduce_ctpop_v4i64_buildvector_v8i32(<4 x i64> %a0, <4 x i64>
 ; AVX512VPOPCNT-NEXT:    vpmovqb %ymm1, %xmm1
 ; AVX512VPOPCNT-NEXT:    vpsadbw %xmm1, %xmm8, %xmm1
 ; AVX512VPOPCNT-NEXT:    vpmovqb %ymm2, %xmm2
+; AVX512VPOPCNT-NEXT:    vpsadbw %xmm2, %xmm8, %xmm2
 ; AVX512VPOPCNT-NEXT:    vpmovqb %ymm3, %xmm3
+; AVX512VPOPCNT-NEXT:    vpsadbw %xmm3, %xmm8, %xmm3
 ; AVX512VPOPCNT-NEXT:    vpmovqb %ymm4, %xmm4
 ; AVX512VPOPCNT-NEXT:    vpmovqb %ymm5, %xmm5
 ; AVX512VPOPCNT-NEXT:    vpmovqb %ymm6, %xmm6
 ; AVX512VPOPCNT-NEXT:    vpmovqb %ymm7, %xmm7
+; AVX512VPOPCNT-NEXT:    vpbroadcastq {{.*#+}} xmm8 = [0,8,0,8]
+; AVX512VPOPCNT-NEXT:    vpermt2d %ymm3, %ymm8, %ymm2
 ; AVX512VPOPCNT-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; AVX512VPOPCNT-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [0,4,0,4]
-; AVX512VPOPCNT-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
-; AVX512VPOPCNT-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512VPOPCNT-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512VPOPCNT-NEXT:    vpermd %ymm2, %ymm1, %ymm1
-; AVX512VPOPCNT-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
+; AVX512VPOPCNT-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]
 ; AVX512VPOPCNT-NEXT:    vinserti128 $1, %xmm7, %ymm6, %ymm1
-; AVX512VPOPCNT-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512VPOPCNT-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm2
-; AVX512VPOPCNT-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
+; AVX512VPOPCNT-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512VPOPCNT-NEXT:    vpsadbw %ymm2, %ymm1, %ymm1
+; AVX512VPOPCNT-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm3
+; AVX512VPOPCNT-NEXT:    vpsadbw %ymm2, %ymm3, %ymm2
 ; AVX512VPOPCNT-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [0,0,0,0,0,4,8,12]
 ; AVX512VPOPCNT-NEXT:    vpermi2d %ymm1, %ymm2, %ymm3
 ; AVX512VPOPCNT-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]



More information about the llvm-commits mailing list