[llvm] [SelectionDAG] Deal with POISON for INSERT_VECTOR_ELT/INSERT_SUBVECTOR (part 1) (PR #143102)
Björn Pettersson via llvm-commits
llvm-commits at lists.llvm.org
Wed Jun 11 02:35:27 PDT 2025
https://github.com/bjope updated https://github.com/llvm/llvm-project/pull/143102
>From dc7004be2c61821b1b4f36f70ad8d87959adc4cb Mon Sep 17 00:00:00 2001
From: Bjorn Pettersson <bjorn.a.pettersson at ericsson.com>
Date: Sat, 31 May 2025 09:35:09 +0200
Subject: [PATCH 1/4] [SelectionDAG] Deal with POISON for
INSERT_VECTOR_ELT/INSERT_SUBVECTOR (part 1)
As reported in #141034 SelectionDAG::getNode had some unexpected
behaviors when trying to create vectors with UNDEF elements. Since
we treat both UNDEF and POISON as undefined (when using isUndef())
we can't just fold away INSERT_VECTOR_ELT/INSERT_SUBVECTOR based on
isUndef(), as that could make the resulting vector more poisonous.
Same kind of bug existed in DAGCombiner::visitINSERT_SUBVECTOR.
Here are some examples:
This fold was done even if vec[idx] was POISON:
INSERT_VECTOR_ELT vec, UNDEF, idx -> vec
This fold was done even if any of vec[idx..idx+size] was POISON:
INSERT_SUBVECTOR vec, UNDEF, idx -> vec
This fold was done even if the elements not extracted from vec could
be POISON:
sub = EXTRACT_SUBVECTOR vec, idx
INSERT_SUBVECTOR UNDEF, sub, idx -> vec
With this patch we avoid such folds unless we can prove that the
result isn't more poisonous when eliminating the insert.
This patch in itself result in some regressions. Goal is to try to
deal with those regressions in follow up commits.
Fixes #141034
---
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 64 +-
.../lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 65 +-
.../aarch64-neon-vector-insert-uaddlv.ll | 3 +-
.../CodeGen/AArch64/arm64-vector-insertion.ll | 2 +
.../AArch64/sve-fixed-length-fp-vselect.ll | 81 +-
.../AArch64/sve-fixed-length-int-vselect.ll | 108 +-
.../AArch64/sve-fixed-length-masked-gather.ll | 6 +-
...-streaming-mode-fixed-length-fp-vselect.ll | 21 +
...streaming-mode-fixed-length-int-vselect.ll | 28 +
.../AMDGPU/load-local-redundant-copies.ll | 45 +-
.../fixed-vectors-vfw-web-simplification.ll | 90 +-
.../fixed-vectors-vw-web-simplification.ll | 55 +-
.../CodeGen/RISCV/rvv/vector-deinterleave.ll | 36 +
llvm/test/CodeGen/Thumb2/active_lane_mask.ll | 3 +-
llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll | 36 +-
.../vector-interleaved-load-i16-stride-4.ll | 1290 +++++++++--------
.../vector-interleaved-store-i16-stride-5.ll | 4 +
.../vector-interleaved-store-i16-stride-7.ll | 32 +-
.../vector-interleaved-store-i8-stride-5.ll | 16 +-
.../vector-interleaved-store-i8-stride-7.ll | 380 ++---
20 files changed, 1437 insertions(+), 928 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index b65e8e06eae62..8da0bca158eda 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -22955,6 +22955,7 @@ SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
auto *IndexC = dyn_cast<ConstantSDNode>(EltNo);
// Insert into out-of-bounds element is undefined.
+ // Code below relies on that we handle this special case early.
if (IndexC && VT.isFixedLengthVector() &&
IndexC->getZExtValue() >= VT.getVectorNumElements())
return DAG.getUNDEF(VT);
@@ -22965,14 +22966,29 @@ SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
InVec == InVal.getOperand(0) && EltNo == InVal.getOperand(1))
return InVec;
- if (!IndexC) {
- // If this is variable insert to undef vector, it might be better to splat:
- // inselt undef, InVal, EltNo --> build_vector < InVal, InVal, ... >
- if (InVec.isUndef() && TLI.shouldSplatInsEltVarIndex(VT))
- return DAG.getSplat(VT, DL, InVal);
- return SDValue();
+ // If this is variable insert to undef vector, it might be better to splat:
+ // inselt undef, InVal, EltNo --> build_vector < InVal, InVal, ... >
+ if (!IndexC && InVec.isUndef() && TLI.shouldSplatInsEltVarIndex(VT))
+ return DAG.getSplat(VT, DL, InVal);
+
+ // Try to drop insert of UNDEF/POISON elements. This is also done in getNode,
+ // but we also do it as a DAG combine since for example simplifications into
+ // SPLAT_VECTOR/BUILD_VECTOR may turn poison elements into undef/zero etc, and
+ // then suddenly the InVec is guaranteed to not be poison.
+ if (InVal.isUndef()) {
+ if (IndexC && VT.isFixedLengthVector()) {
+ APInt EltMask = APInt::getOneBitSet(VT.getVectorNumElements(),
+ IndexC->getZExtValue());
+ if (DAG.isGuaranteedNotToBePoison(InVec, EltMask))
+ return InVec;
+ } else if (DAG.isGuaranteedNotToBePoison(InVec)) {
+ return InVec;
+ }
}
+ if (!IndexC)
+ return SDValue();
+
if (VT.isScalableVector())
return SDValue();
@@ -27405,18 +27421,40 @@ SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
SDValue N2 = N->getOperand(2);
uint64_t InsIdx = N->getConstantOperandVal(2);
- // If inserting an UNDEF, just return the original vector.
- if (N1.isUndef())
- return N0;
+ // If inserting an UNDEF, just return the original vector (unless it makes the
+ // result more poisonous).
+ if (N1.isUndef()){
+ if (VT.isFixedLengthVector()) {
+ unsigned SubVecNumElts = N1.getValueType().getVectorNumElements();
+ APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(),
+ InsIdx, InsIdx + SubVecNumElts);
+ if (DAG.isGuaranteedNotToBePoison(N0, EltMask))
+ return N0;
+ } else if (DAG.isGuaranteedNotToBePoison(N0))
+ return N0;
+ }
- // If this is an insert of an extracted vector into an undef vector, we can
- // just use the input to the extract if the types match, and can simplify
+ // If this is an insert of an extracted vector into an undef/poison vector, we
+ // can just use the input to the extract if the types match, and can simplify
// in some cases even if they don't.
if (N0.isUndef() && N1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
N1.getOperand(1) == N2) {
+ EVT N1VT = N1.getValueType();
EVT SrcVT = N1.getOperand(0).getValueType();
- if (SrcVT == VT)
- return N1.getOperand(0);
+ if (SrcVT == VT) {
+ // Need to ensure that result isn't more poisonous if skipping both the
+ // extract+insert.
+ if (N0.getOpcode() == ISD::POISON)
+ return N1.getOperand(0);
+ if (VT.isFixedLengthVector() && N1VT.isFixedLengthVector()) {
+ unsigned SubVecNumElts = N1VT.getVectorNumElements();
+ APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(),
+ InsIdx, InsIdx + SubVecNumElts);
+ if (DAG.isGuaranteedNotToBePoison(N1.getOperand(0), ~EltMask))
+ return N1.getOperand(0);
+ } else if (DAG.isGuaranteedNotToBePoison(N1.getOperand(0)))
+ return N1.getOperand(0);
+ }
// TODO: To remove the zero check, need to adjust the offset to
// a multiple of the new src type.
if (isNullConstant(N2)) {
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 4fc026ca562ba..4e2ad10ae5637 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -7911,23 +7911,42 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
// INSERT_VECTOR_ELT into out-of-bounds element is an UNDEF, except
// for scalable vectors where we will generate appropriate code to
// deal with out-of-bounds cases correctly.
- if (N3C && N1.getValueType().isFixedLengthVector() &&
- N3C->getZExtValue() >= N1.getValueType().getVectorNumElements())
+ if (N3C && VT.isFixedLengthVector() &&
+ N3C->getZExtValue() >= VT.getVectorNumElements())
return getUNDEF(VT);
// Undefined index can be assumed out-of-bounds, so that's UNDEF too.
if (N3.isUndef())
return getUNDEF(VT);
- // If the inserted element is an UNDEF, just use the input vector.
- if (N2.isUndef())
+ // If inserting poison, just use the input vector.
+ if (N2.getOpcode() == ISD::POISON)
return N1;
+ // Inserting undef into undef/poison is still undef.
+ if (N2.getOpcode() == ISD::UNDEF && N1.isUndef())
+ return getUNDEF(VT);
+
+ // If the inserted element is an UNDEF, just use the input vector.
+ // But not if skipping the insert could make the result more poisonous.
+ if (N2.isUndef()) {
+ if (N3C && VT.isFixedLengthVector()) {
+ APInt EltMask = APInt::getOneBitSet(VT.getVectorNumElements(),
+ N3C->getZExtValue());
+ if (isGuaranteedNotToBePoison(N1, EltMask))
+ return N1;
+ } else if (isGuaranteedNotToBePoison(N1))
+ return N1;
+ }
break;
}
case ISD::INSERT_SUBVECTOR: {
- // Inserting undef into undef is still undef.
- if (N1.isUndef() && N2.isUndef())
+ // If inserting poison, just use the input vector,
+ if (N2.getOpcode() == ISD::POISON)
+ return N1;
+
+ // Inserting undef into undef/poison is still undef.
+ if (N2.getOpcode() == ISD::UNDEF && N1.isUndef())
return getUNDEF(VT);
EVT N2VT = N2.getValueType();
@@ -7956,11 +7975,37 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
if (VT == N2VT)
return N2;
- // If this is an insert of an extracted vector into an undef vector, we
- // can just use the input to the extract.
+ // If this is an insert of an extracted vector into an undef/poison vector,
+ // we can just use the input to the extract. But not if skipping the
+ // extract+insert could make the result more poisonous.
if (N1.isUndef() && N2.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
- N2.getOperand(1) == N3 && N2.getOperand(0).getValueType() == VT)
- return N2.getOperand(0);
+ N2.getOperand(1) == N3 && N2.getOperand(0).getValueType() == VT) {
+ if (N1.getOpcode() == ISD::POISON)
+ return N2.getOperand(0);
+ if (VT.isFixedLengthVector() && N2VT.isFixedLengthVector()) {
+ unsigned LoBit = N3->getAsZExtVal();
+ unsigned HiBit = LoBit + N2VT.getVectorNumElements();
+ APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(),
+ LoBit, HiBit);
+ if (isGuaranteedNotToBePoison(N2.getOperand(0), ~EltMask))
+ return N2.getOperand(0);
+ } else if (isGuaranteedNotToBePoison(N2.getOperand(0)))
+ return N2.getOperand(0);
+ }
+
+ // If the inserted subvector is UNDEF, just use the input vector.
+ // But not if skipping the insert could make the result more poisonous.
+ if (N2.isUndef()) {
+ if (VT.isFixedLengthVector()) {
+ unsigned LoBit = N3->getAsZExtVal();
+ unsigned HiBit = LoBit + N2VT.getVectorNumElements();
+ APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(),
+ LoBit, HiBit);
+ if (isGuaranteedNotToBePoison(N1, EltMask))
+ return N1;
+ } else if (isGuaranteedNotToBePoison(N1))
+ return N1;
+ }
break;
}
case ISD::BITCAST:
diff --git a/llvm/test/CodeGen/AArch64/aarch64-neon-vector-insert-uaddlv.ll b/llvm/test/CodeGen/AArch64/aarch64-neon-vector-insert-uaddlv.ll
index 91eda8d552397..83e35599c4e9e 100644
--- a/llvm/test/CodeGen/AArch64/aarch64-neon-vector-insert-uaddlv.ll
+++ b/llvm/test/CodeGen/AArch64/aarch64-neon-vector-insert-uaddlv.ll
@@ -229,9 +229,10 @@ define void @insert_vec_v3i16_uaddlv_from_v8i16(ptr %0) {
; CHECK: ; %bb.0: ; %entry
; CHECK-NEXT: movi.2d v0, #0000000000000000
; CHECK-NEXT: movi.2d v1, #0000000000000000
-; CHECK-NEXT: add x8, x0, #8
; CHECK-NEXT: uaddlv.8h s0, v0
; CHECK-NEXT: mov.h v1[0], v0[0]
+; CHECK-NEXT: mov.h v1[3], w8
+; CHECK-NEXT: add x8, x0, #8
; CHECK-NEXT: ushll.4s v1, v1, #0
; CHECK-NEXT: ucvtf.4s v1, v1
; CHECK-NEXT: st1.s { v1 }[2], [x8]
diff --git a/llvm/test/CodeGen/AArch64/arm64-vector-insertion.ll b/llvm/test/CodeGen/AArch64/arm64-vector-insertion.ll
index ff28c7817d143..ee3b3d18df879 100644
--- a/llvm/test/CodeGen/AArch64/arm64-vector-insertion.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-vector-insertion.ll
@@ -84,6 +84,7 @@ define <16 x i8> @test_insert_v16i8_insert_2_undef_base_skip8(i32 %a0) {
; CHECK-NEXT: lsr w8, w0, #5
; CHECK-NEXT: dup.16b v0, w8
; CHECK-NEXT: mov.b v0[5], wzr
+; CHECK-NEXT: mov.b v0[8], w8
; CHECK-NEXT: mov.b v0[9], wzr
; CHECK-NEXT: ret
%a1 = lshr exact i32 %a0, 5
@@ -144,6 +145,7 @@ define <16 x i8> @test_insert_v16i8_insert_2_undef_base_different_values_skip8(i
; CHECK-NEXT: mov.b v0[2], w8
; CHECK-NEXT: mov.b v0[5], wzr
; CHECK-NEXT: mov.b v0[7], w8
+; CHECK-NEXT: mov.b v0[8], w8
; CHECK-NEXT: mov.b v0[9], wzr
; CHECK-NEXT: mov.b v0[12], w8
; CHECK-NEXT: mov.b v0[15], w8
diff --git a/llvm/test/CodeGen/AArch64/sve-fixed-length-fp-vselect.ll b/llvm/test/CodeGen/AArch64/sve-fixed-length-fp-vselect.ll
index 9efe0b33910c8..2905d707bdd09 100644
--- a/llvm/test/CodeGen/AArch64/sve-fixed-length-fp-vselect.ll
+++ b/llvm/test/CodeGen/AArch64/sve-fixed-length-fp-vselect.ll
@@ -37,6 +37,10 @@ define void @select_v16f16(ptr %a, ptr %b) vscale_range(2,0) #0 {
; CHECK-NEXT: ld1h { z0.h }, p0/z, [x0]
; CHECK-NEXT: ld1h { z1.h }, p0/z, [x1]
; CHECK-NEXT: fcmeq p1.h, p0/z, z0.h, z1.h
+; CHECK-NEXT: mov z2.h, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.h
+; CHECK-NEXT: and z2.h, z2.h, #0x1
+; CHECK-NEXT: cmpne p1.h, p1/z, z2.h, #0
; CHECK-NEXT: sel z0.h, p1, z0.h, z1.h
; CHECK-NEXT: st1h { z0.h }, p0, [x0]
; CHECK-NEXT: ret
@@ -59,8 +63,15 @@ define void @select_v32f16(ptr %a, ptr %b) #0 {
; VBITS_GE_256-NEXT: ld1h { z3.h }, p0/z, [x1]
; VBITS_GE_256-NEXT: fcmeq p1.h, p0/z, z0.h, z1.h
; VBITS_GE_256-NEXT: fcmeq p2.h, p0/z, z2.h, z3.h
-; VBITS_GE_256-NEXT: sel z0.h, p1, z0.h, z1.h
-; VBITS_GE_256-NEXT: sel z1.h, p2, z2.h, z3.h
+; VBITS_GE_256-NEXT: mov z4.h, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: ptrue p1.h
+; VBITS_GE_256-NEXT: mov z5.h, p2/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: and z4.h, z4.h, #0x1
+; VBITS_GE_256-NEXT: and z5.h, z5.h, #0x1
+; VBITS_GE_256-NEXT: cmpne p2.h, p1/z, z4.h, #0
+; VBITS_GE_256-NEXT: cmpne p1.h, p1/z, z5.h, #0
+; VBITS_GE_256-NEXT: sel z0.h, p2, z0.h, z1.h
+; VBITS_GE_256-NEXT: sel z1.h, p1, z2.h, z3.h
; VBITS_GE_256-NEXT: st1h { z0.h }, p0, [x0, x8, lsl #1]
; VBITS_GE_256-NEXT: st1h { z1.h }, p0, [x0]
; VBITS_GE_256-NEXT: ret
@@ -71,6 +82,10 @@ define void @select_v32f16(ptr %a, ptr %b) #0 {
; VBITS_GE_512-NEXT: ld1h { z0.h }, p0/z, [x0]
; VBITS_GE_512-NEXT: ld1h { z1.h }, p0/z, [x1]
; VBITS_GE_512-NEXT: fcmeq p1.h, p0/z, z0.h, z1.h
+; VBITS_GE_512-NEXT: mov z2.h, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_512-NEXT: ptrue p1.h
+; VBITS_GE_512-NEXT: and z2.h, z2.h, #0x1
+; VBITS_GE_512-NEXT: cmpne p1.h, p1/z, z2.h, #0
; VBITS_GE_512-NEXT: sel z0.h, p1, z0.h, z1.h
; VBITS_GE_512-NEXT: st1h { z0.h }, p0, [x0]
; VBITS_GE_512-NEXT: ret
@@ -89,6 +104,10 @@ define void @select_v64f16(ptr %a, ptr %b) vscale_range(8,0) #0 {
; CHECK-NEXT: ld1h { z0.h }, p0/z, [x0]
; CHECK-NEXT: ld1h { z1.h }, p0/z, [x1]
; CHECK-NEXT: fcmeq p1.h, p0/z, z0.h, z1.h
+; CHECK-NEXT: mov z2.h, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.h
+; CHECK-NEXT: and z2.h, z2.h, #0x1
+; CHECK-NEXT: cmpne p1.h, p1/z, z2.h, #0
; CHECK-NEXT: sel z0.h, p1, z0.h, z1.h
; CHECK-NEXT: st1h { z0.h }, p0, [x0]
; CHECK-NEXT: ret
@@ -107,6 +126,10 @@ define void @select_v128f16(ptr %a, ptr %b) vscale_range(16,0) #0 {
; CHECK-NEXT: ld1h { z0.h }, p0/z, [x0]
; CHECK-NEXT: ld1h { z1.h }, p0/z, [x1]
; CHECK-NEXT: fcmeq p1.h, p0/z, z0.h, z1.h
+; CHECK-NEXT: mov z2.h, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.h
+; CHECK-NEXT: and z2.h, z2.h, #0x1
+; CHECK-NEXT: cmpne p1.h, p1/z, z2.h, #0
; CHECK-NEXT: sel z0.h, p1, z0.h, z1.h
; CHECK-NEXT: st1h { z0.h }, p0, [x0]
; CHECK-NEXT: ret
@@ -150,6 +173,10 @@ define void @select_v8f32(ptr %a, ptr %b) vscale_range(2,0) #0 {
; CHECK-NEXT: ld1w { z0.s }, p0/z, [x0]
; CHECK-NEXT: ld1w { z1.s }, p0/z, [x1]
; CHECK-NEXT: fcmeq p1.s, p0/z, z0.s, z1.s
+; CHECK-NEXT: mov z2.s, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.s
+; CHECK-NEXT: and z2.s, z2.s, #0x1
+; CHECK-NEXT: cmpne p1.s, p1/z, z2.s, #0
; CHECK-NEXT: sel z0.s, p1, z0.s, z1.s
; CHECK-NEXT: st1w { z0.s }, p0, [x0]
; CHECK-NEXT: ret
@@ -172,8 +199,15 @@ define void @select_v16f32(ptr %a, ptr %b) #0 {
; VBITS_GE_256-NEXT: ld1w { z3.s }, p0/z, [x1]
; VBITS_GE_256-NEXT: fcmeq p1.s, p0/z, z0.s, z1.s
; VBITS_GE_256-NEXT: fcmeq p2.s, p0/z, z2.s, z3.s
-; VBITS_GE_256-NEXT: sel z0.s, p1, z0.s, z1.s
-; VBITS_GE_256-NEXT: sel z1.s, p2, z2.s, z3.s
+; VBITS_GE_256-NEXT: mov z4.s, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: ptrue p1.s
+; VBITS_GE_256-NEXT: mov z5.s, p2/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: and z4.s, z4.s, #0x1
+; VBITS_GE_256-NEXT: and z5.s, z5.s, #0x1
+; VBITS_GE_256-NEXT: cmpne p2.s, p1/z, z4.s, #0
+; VBITS_GE_256-NEXT: cmpne p1.s, p1/z, z5.s, #0
+; VBITS_GE_256-NEXT: sel z0.s, p2, z0.s, z1.s
+; VBITS_GE_256-NEXT: sel z1.s, p1, z2.s, z3.s
; VBITS_GE_256-NEXT: st1w { z0.s }, p0, [x0, x8, lsl #2]
; VBITS_GE_256-NEXT: st1w { z1.s }, p0, [x0]
; VBITS_GE_256-NEXT: ret
@@ -184,6 +218,10 @@ define void @select_v16f32(ptr %a, ptr %b) #0 {
; VBITS_GE_512-NEXT: ld1w { z0.s }, p0/z, [x0]
; VBITS_GE_512-NEXT: ld1w { z1.s }, p0/z, [x1]
; VBITS_GE_512-NEXT: fcmeq p1.s, p0/z, z0.s, z1.s
+; VBITS_GE_512-NEXT: mov z2.s, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_512-NEXT: ptrue p1.s
+; VBITS_GE_512-NEXT: and z2.s, z2.s, #0x1
+; VBITS_GE_512-NEXT: cmpne p1.s, p1/z, z2.s, #0
; VBITS_GE_512-NEXT: sel z0.s, p1, z0.s, z1.s
; VBITS_GE_512-NEXT: st1w { z0.s }, p0, [x0]
; VBITS_GE_512-NEXT: ret
@@ -202,6 +240,10 @@ define void @select_v32f32(ptr %a, ptr %b) vscale_range(8,0) #0 {
; CHECK-NEXT: ld1w { z0.s }, p0/z, [x0]
; CHECK-NEXT: ld1w { z1.s }, p0/z, [x1]
; CHECK-NEXT: fcmeq p1.s, p0/z, z0.s, z1.s
+; CHECK-NEXT: mov z2.s, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.s
+; CHECK-NEXT: and z2.s, z2.s, #0x1
+; CHECK-NEXT: cmpne p1.s, p1/z, z2.s, #0
; CHECK-NEXT: sel z0.s, p1, z0.s, z1.s
; CHECK-NEXT: st1w { z0.s }, p0, [x0]
; CHECK-NEXT: ret
@@ -220,6 +262,10 @@ define void @select_v64f32(ptr %a, ptr %b) vscale_range(16,0) #0 {
; CHECK-NEXT: ld1w { z0.s }, p0/z, [x0]
; CHECK-NEXT: ld1w { z1.s }, p0/z, [x1]
; CHECK-NEXT: fcmeq p1.s, p0/z, z0.s, z1.s
+; CHECK-NEXT: mov z2.s, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.s
+; CHECK-NEXT: and z2.s, z2.s, #0x1
+; CHECK-NEXT: cmpne p1.s, p1/z, z2.s, #0
; CHECK-NEXT: sel z0.s, p1, z0.s, z1.s
; CHECK-NEXT: st1w { z0.s }, p0, [x0]
; CHECK-NEXT: ret
@@ -264,6 +310,10 @@ define void @select_v4f64(ptr %a, ptr %b) vscale_range(2,0) #0 {
; CHECK-NEXT: ld1d { z0.d }, p0/z, [x0]
; CHECK-NEXT: ld1d { z1.d }, p0/z, [x1]
; CHECK-NEXT: fcmeq p1.d, p0/z, z0.d, z1.d
+; CHECK-NEXT: mov z2.d, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.d
+; CHECK-NEXT: and z2.d, z2.d, #0x1
+; CHECK-NEXT: cmpne p1.d, p1/z, z2.d, #0
; CHECK-NEXT: sel z0.d, p1, z0.d, z1.d
; CHECK-NEXT: st1d { z0.d }, p0, [x0]
; CHECK-NEXT: ret
@@ -286,8 +336,15 @@ define void @select_v8f64(ptr %a, ptr %b) #0 {
; VBITS_GE_256-NEXT: ld1d { z3.d }, p0/z, [x1]
; VBITS_GE_256-NEXT: fcmeq p1.d, p0/z, z0.d, z1.d
; VBITS_GE_256-NEXT: fcmeq p2.d, p0/z, z2.d, z3.d
-; VBITS_GE_256-NEXT: sel z0.d, p1, z0.d, z1.d
-; VBITS_GE_256-NEXT: sel z1.d, p2, z2.d, z3.d
+; VBITS_GE_256-NEXT: mov z4.d, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: ptrue p1.d
+; VBITS_GE_256-NEXT: mov z5.d, p2/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: and z4.d, z4.d, #0x1
+; VBITS_GE_256-NEXT: and z5.d, z5.d, #0x1
+; VBITS_GE_256-NEXT: cmpne p2.d, p1/z, z4.d, #0
+; VBITS_GE_256-NEXT: cmpne p1.d, p1/z, z5.d, #0
+; VBITS_GE_256-NEXT: sel z0.d, p2, z0.d, z1.d
+; VBITS_GE_256-NEXT: sel z1.d, p1, z2.d, z3.d
; VBITS_GE_256-NEXT: st1d { z0.d }, p0, [x0, x8, lsl #3]
; VBITS_GE_256-NEXT: st1d { z1.d }, p0, [x0]
; VBITS_GE_256-NEXT: ret
@@ -298,6 +355,10 @@ define void @select_v8f64(ptr %a, ptr %b) #0 {
; VBITS_GE_512-NEXT: ld1d { z0.d }, p0/z, [x0]
; VBITS_GE_512-NEXT: ld1d { z1.d }, p0/z, [x1]
; VBITS_GE_512-NEXT: fcmeq p1.d, p0/z, z0.d, z1.d
+; VBITS_GE_512-NEXT: mov z2.d, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_512-NEXT: ptrue p1.d
+; VBITS_GE_512-NEXT: and z2.d, z2.d, #0x1
+; VBITS_GE_512-NEXT: cmpne p1.d, p1/z, z2.d, #0
; VBITS_GE_512-NEXT: sel z0.d, p1, z0.d, z1.d
; VBITS_GE_512-NEXT: st1d { z0.d }, p0, [x0]
; VBITS_GE_512-NEXT: ret
@@ -316,6 +377,10 @@ define void @select_v16f64(ptr %a, ptr %b) vscale_range(8,0) #0 {
; CHECK-NEXT: ld1d { z0.d }, p0/z, [x0]
; CHECK-NEXT: ld1d { z1.d }, p0/z, [x1]
; CHECK-NEXT: fcmeq p1.d, p0/z, z0.d, z1.d
+; CHECK-NEXT: mov z2.d, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.d
+; CHECK-NEXT: and z2.d, z2.d, #0x1
+; CHECK-NEXT: cmpne p1.d, p1/z, z2.d, #0
; CHECK-NEXT: sel z0.d, p1, z0.d, z1.d
; CHECK-NEXT: st1d { z0.d }, p0, [x0]
; CHECK-NEXT: ret
@@ -334,6 +399,10 @@ define void @select_v32f64(ptr %a, ptr %b) vscale_range(16,0) #0 {
; CHECK-NEXT: ld1d { z0.d }, p0/z, [x0]
; CHECK-NEXT: ld1d { z1.d }, p0/z, [x1]
; CHECK-NEXT: fcmeq p1.d, p0/z, z0.d, z1.d
+; CHECK-NEXT: mov z2.d, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.d
+; CHECK-NEXT: and z2.d, z2.d, #0x1
+; CHECK-NEXT: cmpne p1.d, p1/z, z2.d, #0
; CHECK-NEXT: sel z0.d, p1, z0.d, z1.d
; CHECK-NEXT: st1d { z0.d }, p0, [x0]
; CHECK-NEXT: ret
diff --git a/llvm/test/CodeGen/AArch64/sve-fixed-length-int-vselect.ll b/llvm/test/CodeGen/AArch64/sve-fixed-length-int-vselect.ll
index 9cebbc4aab9b7..0e95da31c13cc 100644
--- a/llvm/test/CodeGen/AArch64/sve-fixed-length-int-vselect.ll
+++ b/llvm/test/CodeGen/AArch64/sve-fixed-length-int-vselect.ll
@@ -36,6 +36,10 @@ define void @select_v32i8(ptr %a, ptr %b) vscale_range(2,0) #0 {
; CHECK-NEXT: ld1b { z0.b }, p0/z, [x0]
; CHECK-NEXT: ld1b { z1.b }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.b, p0/z, z0.b, z1.b
+; CHECK-NEXT: mov z2.b, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.b
+; CHECK-NEXT: and z2.b, z2.b, #0x1
+; CHECK-NEXT: cmpne p1.b, p1/z, z2.b, #0
; CHECK-NEXT: sel z0.b, p1, z0.b, z1.b
; CHECK-NEXT: st1b { z0.b }, p0, [x0]
; CHECK-NEXT: ret
@@ -58,8 +62,15 @@ define void @select_v64i8(ptr %a, ptr %b) #0 {
; VBITS_GE_256-NEXT: ld1b { z3.b }, p0/z, [x1]
; VBITS_GE_256-NEXT: cmpeq p1.b, p0/z, z0.b, z1.b
; VBITS_GE_256-NEXT: cmpeq p2.b, p0/z, z2.b, z3.b
-; VBITS_GE_256-NEXT: sel z0.b, p1, z0.b, z1.b
-; VBITS_GE_256-NEXT: sel z1.b, p2, z2.b, z3.b
+; VBITS_GE_256-NEXT: mov z4.b, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: ptrue p1.b
+; VBITS_GE_256-NEXT: mov z5.b, p2/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: and z4.b, z4.b, #0x1
+; VBITS_GE_256-NEXT: and z5.b, z5.b, #0x1
+; VBITS_GE_256-NEXT: cmpne p2.b, p1/z, z4.b, #0
+; VBITS_GE_256-NEXT: cmpne p1.b, p1/z, z5.b, #0
+; VBITS_GE_256-NEXT: sel z0.b, p2, z0.b, z1.b
+; VBITS_GE_256-NEXT: sel z1.b, p1, z2.b, z3.b
; VBITS_GE_256-NEXT: st1b { z0.b }, p0, [x0, x8]
; VBITS_GE_256-NEXT: st1b { z1.b }, p0, [x0]
; VBITS_GE_256-NEXT: ret
@@ -70,6 +81,10 @@ define void @select_v64i8(ptr %a, ptr %b) #0 {
; VBITS_GE_512-NEXT: ld1b { z0.b }, p0/z, [x0]
; VBITS_GE_512-NEXT: ld1b { z1.b }, p0/z, [x1]
; VBITS_GE_512-NEXT: cmpeq p1.b, p0/z, z0.b, z1.b
+; VBITS_GE_512-NEXT: mov z2.b, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_512-NEXT: ptrue p1.b
+; VBITS_GE_512-NEXT: and z2.b, z2.b, #0x1
+; VBITS_GE_512-NEXT: cmpne p1.b, p1/z, z2.b, #0
; VBITS_GE_512-NEXT: sel z0.b, p1, z0.b, z1.b
; VBITS_GE_512-NEXT: st1b { z0.b }, p0, [x0]
; VBITS_GE_512-NEXT: ret
@@ -88,6 +103,10 @@ define void @select_v128i8(ptr %a, ptr %b) vscale_range(8,0) #0 {
; CHECK-NEXT: ld1b { z0.b }, p0/z, [x0]
; CHECK-NEXT: ld1b { z1.b }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.b, p0/z, z0.b, z1.b
+; CHECK-NEXT: mov z2.b, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.b
+; CHECK-NEXT: and z2.b, z2.b, #0x1
+; CHECK-NEXT: cmpne p1.b, p1/z, z2.b, #0
; CHECK-NEXT: sel z0.b, p1, z0.b, z1.b
; CHECK-NEXT: st1b { z0.b }, p0, [x0]
; CHECK-NEXT: ret
@@ -106,6 +125,10 @@ define void @select_v256i8(ptr %a, ptr %b) vscale_range(16,0) #0 {
; CHECK-NEXT: ld1b { z0.b }, p0/z, [x0]
; CHECK-NEXT: ld1b { z1.b }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.b, p0/z, z0.b, z1.b
+; CHECK-NEXT: mov z2.b, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.b
+; CHECK-NEXT: and z2.b, z2.b, #0x1
+; CHECK-NEXT: cmpne p1.b, p1/z, z2.b, #0
; CHECK-NEXT: sel z0.b, p1, z0.b, z1.b
; CHECK-NEXT: st1b { z0.b }, p0, [x0]
; CHECK-NEXT: ret
@@ -149,6 +172,10 @@ define void @select_v16i16(ptr %a, ptr %b) vscale_range(2,0) #0 {
; CHECK-NEXT: ld1h { z0.h }, p0/z, [x0]
; CHECK-NEXT: ld1h { z1.h }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.h, p0/z, z0.h, z1.h
+; CHECK-NEXT: mov z2.h, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.h
+; CHECK-NEXT: and z2.h, z2.h, #0x1
+; CHECK-NEXT: cmpne p1.h, p1/z, z2.h, #0
; CHECK-NEXT: sel z0.h, p1, z0.h, z1.h
; CHECK-NEXT: st1h { z0.h }, p0, [x0]
; CHECK-NEXT: ret
@@ -171,8 +198,15 @@ define void @select_v32i16(ptr %a, ptr %b) #0 {
; VBITS_GE_256-NEXT: ld1h { z3.h }, p0/z, [x1]
; VBITS_GE_256-NEXT: cmpeq p1.h, p0/z, z0.h, z1.h
; VBITS_GE_256-NEXT: cmpeq p2.h, p0/z, z2.h, z3.h
-; VBITS_GE_256-NEXT: sel z0.h, p1, z0.h, z1.h
-; VBITS_GE_256-NEXT: sel z1.h, p2, z2.h, z3.h
+; VBITS_GE_256-NEXT: mov z4.h, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: ptrue p1.h
+; VBITS_GE_256-NEXT: mov z5.h, p2/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: and z4.h, z4.h, #0x1
+; VBITS_GE_256-NEXT: and z5.h, z5.h, #0x1
+; VBITS_GE_256-NEXT: cmpne p2.h, p1/z, z4.h, #0
+; VBITS_GE_256-NEXT: cmpne p1.h, p1/z, z5.h, #0
+; VBITS_GE_256-NEXT: sel z0.h, p2, z0.h, z1.h
+; VBITS_GE_256-NEXT: sel z1.h, p1, z2.h, z3.h
; VBITS_GE_256-NEXT: st1h { z0.h }, p0, [x0, x8, lsl #1]
; VBITS_GE_256-NEXT: st1h { z1.h }, p0, [x0]
; VBITS_GE_256-NEXT: ret
@@ -183,6 +217,10 @@ define void @select_v32i16(ptr %a, ptr %b) #0 {
; VBITS_GE_512-NEXT: ld1h { z0.h }, p0/z, [x0]
; VBITS_GE_512-NEXT: ld1h { z1.h }, p0/z, [x1]
; VBITS_GE_512-NEXT: cmpeq p1.h, p0/z, z0.h, z1.h
+; VBITS_GE_512-NEXT: mov z2.h, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_512-NEXT: ptrue p1.h
+; VBITS_GE_512-NEXT: and z2.h, z2.h, #0x1
+; VBITS_GE_512-NEXT: cmpne p1.h, p1/z, z2.h, #0
; VBITS_GE_512-NEXT: sel z0.h, p1, z0.h, z1.h
; VBITS_GE_512-NEXT: st1h { z0.h }, p0, [x0]
; VBITS_GE_512-NEXT: ret
@@ -201,6 +239,10 @@ define void @select_v64i16(ptr %a, ptr %b) vscale_range(8,0) #0 {
; CHECK-NEXT: ld1h { z0.h }, p0/z, [x0]
; CHECK-NEXT: ld1h { z1.h }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.h, p0/z, z0.h, z1.h
+; CHECK-NEXT: mov z2.h, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.h
+; CHECK-NEXT: and z2.h, z2.h, #0x1
+; CHECK-NEXT: cmpne p1.h, p1/z, z2.h, #0
; CHECK-NEXT: sel z0.h, p1, z0.h, z1.h
; CHECK-NEXT: st1h { z0.h }, p0, [x0]
; CHECK-NEXT: ret
@@ -219,6 +261,10 @@ define void @select_v128i16(ptr %a, ptr %b) vscale_range(16,0) #0 {
; CHECK-NEXT: ld1h { z0.h }, p0/z, [x0]
; CHECK-NEXT: ld1h { z1.h }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.h, p0/z, z0.h, z1.h
+; CHECK-NEXT: mov z2.h, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.h
+; CHECK-NEXT: and z2.h, z2.h, #0x1
+; CHECK-NEXT: cmpne p1.h, p1/z, z2.h, #0
; CHECK-NEXT: sel z0.h, p1, z0.h, z1.h
; CHECK-NEXT: st1h { z0.h }, p0, [x0]
; CHECK-NEXT: ret
@@ -262,6 +308,10 @@ define void @select_v8i32(ptr %a, ptr %b) vscale_range(2,0) #0 {
; CHECK-NEXT: ld1w { z0.s }, p0/z, [x0]
; CHECK-NEXT: ld1w { z1.s }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.s, p0/z, z0.s, z1.s
+; CHECK-NEXT: mov z2.s, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.s
+; CHECK-NEXT: and z2.s, z2.s, #0x1
+; CHECK-NEXT: cmpne p1.s, p1/z, z2.s, #0
; CHECK-NEXT: sel z0.s, p1, z0.s, z1.s
; CHECK-NEXT: st1w { z0.s }, p0, [x0]
; CHECK-NEXT: ret
@@ -284,8 +334,15 @@ define void @select_v16i32(ptr %a, ptr %b) #0 {
; VBITS_GE_256-NEXT: ld1w { z3.s }, p0/z, [x1]
; VBITS_GE_256-NEXT: cmpeq p1.s, p0/z, z0.s, z1.s
; VBITS_GE_256-NEXT: cmpeq p2.s, p0/z, z2.s, z3.s
-; VBITS_GE_256-NEXT: sel z0.s, p1, z0.s, z1.s
-; VBITS_GE_256-NEXT: sel z1.s, p2, z2.s, z3.s
+; VBITS_GE_256-NEXT: mov z4.s, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: ptrue p1.s
+; VBITS_GE_256-NEXT: mov z5.s, p2/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: and z4.s, z4.s, #0x1
+; VBITS_GE_256-NEXT: and z5.s, z5.s, #0x1
+; VBITS_GE_256-NEXT: cmpne p2.s, p1/z, z4.s, #0
+; VBITS_GE_256-NEXT: cmpne p1.s, p1/z, z5.s, #0
+; VBITS_GE_256-NEXT: sel z0.s, p2, z0.s, z1.s
+; VBITS_GE_256-NEXT: sel z1.s, p1, z2.s, z3.s
; VBITS_GE_256-NEXT: st1w { z0.s }, p0, [x0, x8, lsl #2]
; VBITS_GE_256-NEXT: st1w { z1.s }, p0, [x0]
; VBITS_GE_256-NEXT: ret
@@ -296,6 +353,10 @@ define void @select_v16i32(ptr %a, ptr %b) #0 {
; VBITS_GE_512-NEXT: ld1w { z0.s }, p0/z, [x0]
; VBITS_GE_512-NEXT: ld1w { z1.s }, p0/z, [x1]
; VBITS_GE_512-NEXT: cmpeq p1.s, p0/z, z0.s, z1.s
+; VBITS_GE_512-NEXT: mov z2.s, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_512-NEXT: ptrue p1.s
+; VBITS_GE_512-NEXT: and z2.s, z2.s, #0x1
+; VBITS_GE_512-NEXT: cmpne p1.s, p1/z, z2.s, #0
; VBITS_GE_512-NEXT: sel z0.s, p1, z0.s, z1.s
; VBITS_GE_512-NEXT: st1w { z0.s }, p0, [x0]
; VBITS_GE_512-NEXT: ret
@@ -314,6 +375,10 @@ define void @select_v32i32(ptr %a, ptr %b) vscale_range(8,0) #0 {
; CHECK-NEXT: ld1w { z0.s }, p0/z, [x0]
; CHECK-NEXT: ld1w { z1.s }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.s, p0/z, z0.s, z1.s
+; CHECK-NEXT: mov z2.s, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.s
+; CHECK-NEXT: and z2.s, z2.s, #0x1
+; CHECK-NEXT: cmpne p1.s, p1/z, z2.s, #0
; CHECK-NEXT: sel z0.s, p1, z0.s, z1.s
; CHECK-NEXT: st1w { z0.s }, p0, [x0]
; CHECK-NEXT: ret
@@ -332,6 +397,10 @@ define void @select_v64i32(ptr %a, ptr %b) vscale_range(16,0) #0 {
; CHECK-NEXT: ld1w { z0.s }, p0/z, [x0]
; CHECK-NEXT: ld1w { z1.s }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.s, p0/z, z0.s, z1.s
+; CHECK-NEXT: mov z2.s, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.s
+; CHECK-NEXT: and z2.s, z2.s, #0x1
+; CHECK-NEXT: cmpne p1.s, p1/z, z2.s, #0
; CHECK-NEXT: sel z0.s, p1, z0.s, z1.s
; CHECK-NEXT: st1w { z0.s }, p0, [x0]
; CHECK-NEXT: ret
@@ -376,6 +445,10 @@ define void @select_v4i64(ptr %a, ptr %b) vscale_range(2,0) #0 {
; CHECK-NEXT: ld1d { z0.d }, p0/z, [x0]
; CHECK-NEXT: ld1d { z1.d }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.d, p0/z, z0.d, z1.d
+; CHECK-NEXT: mov z2.d, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.d
+; CHECK-NEXT: and z2.d, z2.d, #0x1
+; CHECK-NEXT: cmpne p1.d, p1/z, z2.d, #0
; CHECK-NEXT: sel z0.d, p1, z0.d, z1.d
; CHECK-NEXT: st1d { z0.d }, p0, [x0]
; CHECK-NEXT: ret
@@ -398,8 +471,15 @@ define void @select_v8i64(ptr %a, ptr %b) #0 {
; VBITS_GE_256-NEXT: ld1d { z3.d }, p0/z, [x1]
; VBITS_GE_256-NEXT: cmpeq p1.d, p0/z, z0.d, z1.d
; VBITS_GE_256-NEXT: cmpeq p2.d, p0/z, z2.d, z3.d
-; VBITS_GE_256-NEXT: sel z0.d, p1, z0.d, z1.d
-; VBITS_GE_256-NEXT: sel z1.d, p2, z2.d, z3.d
+; VBITS_GE_256-NEXT: mov z4.d, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: ptrue p1.d
+; VBITS_GE_256-NEXT: mov z5.d, p2/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_256-NEXT: and z4.d, z4.d, #0x1
+; VBITS_GE_256-NEXT: and z5.d, z5.d, #0x1
+; VBITS_GE_256-NEXT: cmpne p2.d, p1/z, z4.d, #0
+; VBITS_GE_256-NEXT: cmpne p1.d, p1/z, z5.d, #0
+; VBITS_GE_256-NEXT: sel z0.d, p2, z0.d, z1.d
+; VBITS_GE_256-NEXT: sel z1.d, p1, z2.d, z3.d
; VBITS_GE_256-NEXT: st1d { z0.d }, p0, [x0, x8, lsl #3]
; VBITS_GE_256-NEXT: st1d { z1.d }, p0, [x0]
; VBITS_GE_256-NEXT: ret
@@ -410,6 +490,10 @@ define void @select_v8i64(ptr %a, ptr %b) #0 {
; VBITS_GE_512-NEXT: ld1d { z0.d }, p0/z, [x0]
; VBITS_GE_512-NEXT: ld1d { z1.d }, p0/z, [x1]
; VBITS_GE_512-NEXT: cmpeq p1.d, p0/z, z0.d, z1.d
+; VBITS_GE_512-NEXT: mov z2.d, p1/z, #-1 // =0xffffffffffffffff
+; VBITS_GE_512-NEXT: ptrue p1.d
+; VBITS_GE_512-NEXT: and z2.d, z2.d, #0x1
+; VBITS_GE_512-NEXT: cmpne p1.d, p1/z, z2.d, #0
; VBITS_GE_512-NEXT: sel z0.d, p1, z0.d, z1.d
; VBITS_GE_512-NEXT: st1d { z0.d }, p0, [x0]
; VBITS_GE_512-NEXT: ret
@@ -428,6 +512,10 @@ define void @select_v16i64(ptr %a, ptr %b) vscale_range(8,0) #0 {
; CHECK-NEXT: ld1d { z0.d }, p0/z, [x0]
; CHECK-NEXT: ld1d { z1.d }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.d, p0/z, z0.d, z1.d
+; CHECK-NEXT: mov z2.d, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.d
+; CHECK-NEXT: and z2.d, z2.d, #0x1
+; CHECK-NEXT: cmpne p1.d, p1/z, z2.d, #0
; CHECK-NEXT: sel z0.d, p1, z0.d, z1.d
; CHECK-NEXT: st1d { z0.d }, p0, [x0]
; CHECK-NEXT: ret
@@ -446,6 +534,10 @@ define void @select_v32i64(ptr %a, ptr %b) vscale_range(16,0) #0 {
; CHECK-NEXT: ld1d { z0.d }, p0/z, [x0]
; CHECK-NEXT: ld1d { z1.d }, p0/z, [x1]
; CHECK-NEXT: cmpeq p1.d, p0/z, z0.d, z1.d
+; CHECK-NEXT: mov z2.d, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.d
+; CHECK-NEXT: and z2.d, z2.d, #0x1
+; CHECK-NEXT: cmpne p1.d, p1/z, z2.d, #0
; CHECK-NEXT: sel z0.d, p1, z0.d, z1.d
; CHECK-NEXT: st1d { z0.d }, p0, [x0]
; CHECK-NEXT: ret
diff --git a/llvm/test/CodeGen/AArch64/sve-fixed-length-masked-gather.ll b/llvm/test/CodeGen/AArch64/sve-fixed-length-masked-gather.ll
index 093e6cd9328c8..ebd32c73ec65b 100644
--- a/llvm/test/CodeGen/AArch64/sve-fixed-length-masked-gather.ll
+++ b/llvm/test/CodeGen/AArch64/sve-fixed-length-masked-gather.ll
@@ -1198,11 +1198,15 @@ define void @masked_gather_passthru(ptr %a, ptr %b, ptr %c) vscale_range(16,0) #
; CHECK-NEXT: ptrue p0.s, vl32
; CHECK-NEXT: ptrue p2.d, vl32
; CHECK-NEXT: ld1w { z0.s }, p0/z, [x0]
-; CHECK-NEXT: ld1w { z1.s }, p0/z, [x2]
; CHECK-NEXT: fcmeq p1.s, p0/z, z0.s, #0.0
; CHECK-NEXT: ld1d { z0.d }, p2/z, [x1]
; CHECK-NEXT: punpklo p2.h, p1.b
+; CHECK-NEXT: mov z1.s, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p1.s
; CHECK-NEXT: ld1w { z0.d }, p2/z, [z0.d]
+; CHECK-NEXT: and z1.s, z1.s, #0x1
+; CHECK-NEXT: cmpne p1.s, p1/z, z1.s, #0
+; CHECK-NEXT: ld1w { z1.s }, p0/z, [x2]
; CHECK-NEXT: uzp1 z0.s, z0.s, z0.s
; CHECK-NEXT: sel z0.s, p1, z0.s, z1.s
; CHECK-NEXT: st1w { z0.s }, p0, [x0]
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-vselect.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-vselect.ll
index ec0693a541e44..8b845dff64ffe 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-vselect.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-vselect.ll
@@ -199,6 +199,13 @@ define void @select_v16f16(ptr %a, ptr %b) {
; CHECK-NEXT: ldp q1, q2, [x0]
; CHECK-NEXT: fcmeq p1.h, p0/z, z1.h, z0.h
; CHECK-NEXT: fcmeq p0.h, p0/z, z2.h, z3.h
+; CHECK-NEXT: mov z4.h, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: mov z5.h, p0/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p0.h
+; CHECK-NEXT: and z4.h, z4.h, #0x1
+; CHECK-NEXT: and z5.h, z5.h, #0x1
+; CHECK-NEXT: cmpne p1.h, p0/z, z4.h, #0
+; CHECK-NEXT: cmpne p0.h, p0/z, z5.h, #0
; CHECK-NEXT: mov z0.h, p1/m, z1.h
; CHECK-NEXT: sel z1.h, p0, z2.h, z3.h
; CHECK-NEXT: stp q0, q1, [x0]
@@ -434,6 +441,13 @@ define void @select_v8f32(ptr %a, ptr %b) {
; CHECK-NEXT: ldp q1, q2, [x0]
; CHECK-NEXT: fcmeq p1.s, p0/z, z1.s, z0.s
; CHECK-NEXT: fcmeq p0.s, p0/z, z2.s, z3.s
+; CHECK-NEXT: mov z4.s, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: mov z5.s, p0/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p0.s
+; CHECK-NEXT: and z4.s, z4.s, #0x1
+; CHECK-NEXT: and z5.s, z5.s, #0x1
+; CHECK-NEXT: cmpne p1.s, p0/z, z4.s, #0
+; CHECK-NEXT: cmpne p0.s, p0/z, z5.s, #0
; CHECK-NEXT: mov z0.s, p1/m, z1.s
; CHECK-NEXT: sel z1.s, p0, z2.s, z3.s
; CHECK-NEXT: stp q0, q1, [x0]
@@ -558,6 +572,13 @@ define void @select_v4f64(ptr %a, ptr %b) {
; CHECK-NEXT: ldp q1, q2, [x0]
; CHECK-NEXT: fcmeq p1.d, p0/z, z1.d, z0.d
; CHECK-NEXT: fcmeq p0.d, p0/z, z2.d, z3.d
+; CHECK-NEXT: mov z4.d, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: mov z5.d, p0/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p0.d
+; CHECK-NEXT: and z4.d, z4.d, #0x1
+; CHECK-NEXT: and z5.d, z5.d, #0x1
+; CHECK-NEXT: cmpne p1.d, p0/z, z4.d, #0
+; CHECK-NEXT: cmpne p0.d, p0/z, z5.d, #0
; CHECK-NEXT: mov z0.d, p1/m, z1.d
; CHECK-NEXT: sel z1.d, p0, z2.d, z3.d
; CHECK-NEXT: stp q0, q1, [x0]
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-vselect.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-vselect.ll
index 39701131d7db6..12b7886d76c70 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-vselect.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-vselect.ll
@@ -293,6 +293,13 @@ define void @select_v32i8(ptr %a, ptr %b) {
; CHECK-NEXT: ldp q1, q2, [x0]
; CHECK-NEXT: cmpeq p1.b, p0/z, z1.b, z0.b
; CHECK-NEXT: cmpeq p0.b, p0/z, z2.b, z3.b
+; CHECK-NEXT: mov z4.b, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: mov z5.b, p0/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p0.b
+; CHECK-NEXT: and z4.b, z4.b, #0x1
+; CHECK-NEXT: and z5.b, z5.b, #0x1
+; CHECK-NEXT: cmpne p1.b, p0/z, z4.b, #0
+; CHECK-NEXT: cmpne p0.b, p0/z, z5.b, #0
; CHECK-NEXT: mov z0.b, p1/m, z1.b
; CHECK-NEXT: sel z1.b, p0, z2.b, z3.b
; CHECK-NEXT: stp q0, q1, [x0]
@@ -697,6 +704,13 @@ define void @select_v16i16(ptr %a, ptr %b) {
; CHECK-NEXT: ldp q1, q2, [x0]
; CHECK-NEXT: cmpeq p1.h, p0/z, z1.h, z0.h
; CHECK-NEXT: cmpeq p0.h, p0/z, z2.h, z3.h
+; CHECK-NEXT: mov z4.h, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: mov z5.h, p0/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p0.h
+; CHECK-NEXT: and z4.h, z4.h, #0x1
+; CHECK-NEXT: and z5.h, z5.h, #0x1
+; CHECK-NEXT: cmpne p1.h, p0/z, z4.h, #0
+; CHECK-NEXT: cmpne p0.h, p0/z, z5.h, #0
; CHECK-NEXT: mov z0.h, p1/m, z1.h
; CHECK-NEXT: sel z1.h, p0, z2.h, z3.h
; CHECK-NEXT: stp q0, q1, [x0]
@@ -911,6 +925,13 @@ define void @select_v8i32(ptr %a, ptr %b) {
; CHECK-NEXT: ldp q1, q2, [x0]
; CHECK-NEXT: cmpeq p1.s, p0/z, z1.s, z0.s
; CHECK-NEXT: cmpeq p0.s, p0/z, z2.s, z3.s
+; CHECK-NEXT: mov z4.s, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: mov z5.s, p0/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p0.s
+; CHECK-NEXT: and z4.s, z4.s, #0x1
+; CHECK-NEXT: and z5.s, z5.s, #0x1
+; CHECK-NEXT: cmpne p1.s, p0/z, z4.s, #0
+; CHECK-NEXT: cmpne p0.s, p0/z, z5.s, #0
; CHECK-NEXT: mov z0.s, p1/m, z1.s
; CHECK-NEXT: sel z1.s, p0, z2.s, z3.s
; CHECK-NEXT: stp q0, q1, [x0]
@@ -1044,6 +1065,13 @@ define void @select_v4i64(ptr %a, ptr %b) {
; CHECK-NEXT: ldp q1, q2, [x0]
; CHECK-NEXT: cmpeq p1.d, p0/z, z1.d, z0.d
; CHECK-NEXT: cmpeq p0.d, p0/z, z2.d, z3.d
+; CHECK-NEXT: mov z4.d, p1/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: mov z5.d, p0/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT: ptrue p0.d
+; CHECK-NEXT: and z4.d, z4.d, #0x1
+; CHECK-NEXT: and z5.d, z5.d, #0x1
+; CHECK-NEXT: cmpne p1.d, p0/z, z4.d, #0
+; CHECK-NEXT: cmpne p0.d, p0/z, z5.d, #0
; CHECK-NEXT: mov z0.d, p1/m, z1.d
; CHECK-NEXT: sel z1.d, p0, z2.d, z3.d
; CHECK-NEXT: stp q0, q1, [x0]
diff --git a/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll b/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll
index 8a3cc57e08579..2269c68b941e7 100644
--- a/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll
@@ -66,38 +66,39 @@ define amdgpu_vs void @test_3(i32 inreg %arg1, i32 inreg %arg2, ptr addrspace(8)
; CHECK-NEXT: s_mov_b32 s6, s4
; CHECK-NEXT: s_mov_b32 s5, s3
; CHECK-NEXT: s_mov_b32 s4, s2
-; CHECK-NEXT: v_add_i32_e32 v0, vcc, 12, v1
-; CHECK-NEXT: v_add_i32_e32 v3, vcc, 8, v1
-; CHECK-NEXT: v_add_i32_e32 v4, vcc, 4, v1
-; CHECK-NEXT: v_add_i32_e32 v7, vcc, 20, v1
-; CHECK-NEXT: v_add_i32_e32 v9, vcc, 16, v1
+; CHECK-NEXT: v_add_i32_e32 v0, vcc, 20, v1
+; CHECK-NEXT: v_add_i32_e32 v3, vcc, 16, v1
+; CHECK-NEXT: v_add_i32_e32 v4, vcc, 12, v1
+; CHECK-NEXT: v_add_i32_e32 v5, vcc, 8, v1
+; CHECK-NEXT: v_add_i32_e32 v9, vcc, 4, v1
; CHECK-NEXT: v_mov_b32_e32 v10, s0
-; CHECK-NEXT: v_add_i32_e32 v11, vcc, 12, v2
-; CHECK-NEXT: v_add_i32_e32 v12, vcc, 8, v2
+; CHECK-NEXT: v_add_i32_e32 v11, vcc, 20, v2
+; CHECK-NEXT: v_add_i32_e32 v12, vcc, 16, v2
; CHECK-NEXT: s_mov_b32 m0, -1
-; CHECK-NEXT: ds_read_b32 v6, v0
-; CHECK-NEXT: ds_read_b32 v5, v3
-; CHECK-NEXT: ds_read_b32 v4, v4
-; CHECK-NEXT: ds_read_b32 v8, v7
-; CHECK-NEXT: ds_read_b32 v7, v9
+; CHECK-NEXT: ds_read_b32 v8, v0
+; CHECK-NEXT: ds_read_b32 v7, v3
+; CHECK-NEXT: ds_read_b32 v6, v4
+; CHECK-NEXT: ds_read_b32 v5, v5
+; CHECK-NEXT: ds_read_b32 v4, v9
; CHECK-NEXT: ds_read_b32 v3, v1
-; CHECK-NEXT: v_add_i32_e32 v0, vcc, 4, v2
-; CHECK-NEXT: v_add_i32_e32 v1, vcc, 20, v2
-; CHECK-NEXT: v_add_i32_e32 v9, vcc, 16, v2
+; CHECK-NEXT: v_add_i32_e32 v1, vcc, 12, v2
+; CHECK-NEXT: v_add_i32_e32 v9, vcc, 8, v2
+; CHECK-NEXT: v_add_i32_e32 v13, vcc, 4, v2
; CHECK-NEXT: s_waitcnt lgkmcnt(0)
; CHECK-NEXT: tbuffer_store_format_xyzw v[3:6], v10, s[4:7], s1 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen offset:264 glc slc
; CHECK-NEXT: tbuffer_store_format_xy v[7:8], v10, s[4:7], s1 format:[BUF_DATA_FORMAT_INVALID,BUF_NUM_FORMAT_UINT] idxen offset:280 glc slc
+; CHECK-NEXT: ds_read_b32 v0, v12
; CHECK-NEXT: s_waitcnt expcnt(1)
-; CHECK-NEXT: ds_read_b32 v5, v11
-; CHECK-NEXT: ds_read_b32 v4, v12
-; CHECK-NEXT: ds_read_b32 v3, v0
-; CHECK-NEXT: ds_read_b32 v1, v1
-; CHECK-NEXT: ds_read_b32 v0, v9
+; CHECK-NEXT: ds_read_b32 v5, v1
+; CHECK-NEXT: ds_read_b32 v4, v9
+; CHECK-NEXT: ds_read_b32 v3, v13
; CHECK-NEXT: ds_read_b32 v2, v2
-; CHECK-NEXT: s_waitcnt lgkmcnt(1)
+; CHECK-NEXT: ds_read_b32 v1, v11
+; CHECK-NEXT: s_waitcnt lgkmcnt(5)
; CHECK-NEXT: exp mrt0 off, off, off, off
-; CHECK-NEXT: s_waitcnt lgkmcnt(0)
+; CHECK-NEXT: s_waitcnt lgkmcnt(1)
; CHECK-NEXT: tbuffer_store_format_xyzw v[2:5], v10, s[4:7], s1 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen offset:240 glc slc
+; CHECK-NEXT: s_waitcnt lgkmcnt(0)
; CHECK-NEXT: tbuffer_store_format_xy v[0:1], v10, s[4:7], s1 format:[BUF_DATA_FORMAT_INVALID,BUF_NUM_FORMAT_UINT] idxen offset:256 glc slc
; CHECK-NEXT: s_endpgm
%load1 = load <6 x float>, ptr addrspace(3) %arg5, align 4
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vfw-web-simplification.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vfw-web-simplification.ll
index aba9056c78cda..5aa3a246d7616 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vfw-web-simplification.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vfw-web-simplification.ll
@@ -1,6 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=1 | FileCheck %s --check-prefixes=NO_FOLDING
-; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=2 | FileCheck %s --check-prefixes=NO_FOLDING
+; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=1 | FileCheck %s --check-prefixes=NO_FOLDING,NO_FOLDING1
+; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=2 | FileCheck %s --check-prefixes=NO_FOLDING,NO_FOLDING2
; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=3 | FileCheck %s --check-prefixes=FOLDING,ZVFH
; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfhmin,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=3 | FileCheck %s --check-prefixes=FOLDING,ZVFHMIN
; Check that the default value enables the web folding and
@@ -8,20 +8,35 @@
; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=FOLDING
define void @vfwmul_v2f116_multiple_users(ptr %x, ptr %y, ptr %z, <2 x half> %a, <2 x half> %b, <2 x half> %b2) {
-; NO_FOLDING-LABEL: vfwmul_v2f116_multiple_users:
-; NO_FOLDING: # %bb.0:
-; NO_FOLDING-NEXT: vsetivli zero, 2, e16, mf4, ta, ma
-; NO_FOLDING-NEXT: vfwcvt.f.f.v v11, v8
-; NO_FOLDING-NEXT: vfwcvt.f.f.v v8, v9
-; NO_FOLDING-NEXT: vfwcvt.f.f.v v9, v10
-; NO_FOLDING-NEXT: vsetvli zero, zero, e32, mf2, ta, ma
-; NO_FOLDING-NEXT: vfmul.vv v10, v11, v8
-; NO_FOLDING-NEXT: vfadd.vv v11, v11, v9
-; NO_FOLDING-NEXT: vfsub.vv v8, v8, v9
-; NO_FOLDING-NEXT: vse32.v v10, (a0)
-; NO_FOLDING-NEXT: vse32.v v11, (a1)
-; NO_FOLDING-NEXT: vse32.v v8, (a2)
-; NO_FOLDING-NEXT: ret
+; NO_FOLDING1-LABEL: vfwmul_v2f116_multiple_users:
+; NO_FOLDING1: # %bb.0:
+; NO_FOLDING1-NEXT: vsetivli zero, 2, e16, mf4, ta, ma
+; NO_FOLDING1-NEXT: vfwcvt.f.f.v v11, v8
+; NO_FOLDING1-NEXT: vfwcvt.f.f.v v8, v9
+; NO_FOLDING1-NEXT: vfwcvt.f.f.v v9, v10
+; NO_FOLDING1-NEXT: vsetvli zero, zero, e32, mf2, ta, ma
+; NO_FOLDING1-NEXT: vfmul.vv v10, v11, v8
+; NO_FOLDING1-NEXT: vfadd.vv v11, v11, v9
+; NO_FOLDING1-NEXT: vfsub.vv v8, v8, v9
+; NO_FOLDING1-NEXT: vse32.v v10, (a0)
+; NO_FOLDING1-NEXT: vse32.v v11, (a1)
+; NO_FOLDING1-NEXT: vse32.v v8, (a2)
+; NO_FOLDING1-NEXT: ret
+;
+; NO_FOLDING2-LABEL: vfwmul_v2f116_multiple_users:
+; NO_FOLDING2: # %bb.0:
+; NO_FOLDING2-NEXT: vsetivli zero, 2, e16, mf4, ta, ma
+; NO_FOLDING2-NEXT: vfwcvt.f.f.v v11, v8
+; NO_FOLDING2-NEXT: vfwcvt.f.f.v v8, v9
+; NO_FOLDING2-NEXT: vsetvli zero, zero, e32, mf2, ta, ma
+; NO_FOLDING2-NEXT: vfmul.vv v9, v11, v8
+; NO_FOLDING2-NEXT: vsetvli zero, zero, e16, mf4, ta, ma
+; NO_FOLDING2-NEXT: vfwadd.wv v11, v11, v10
+; NO_FOLDING2-NEXT: vfwsub.wv v8, v8, v10
+; NO_FOLDING2-NEXT: vse32.v v9, (a0)
+; NO_FOLDING2-NEXT: vse32.v v11, (a1)
+; NO_FOLDING2-NEXT: vse32.v v8, (a2)
+; NO_FOLDING2-NEXT: ret
;
; ZVFH-LABEL: vfwmul_v2f116_multiple_users:
; ZVFH: # %bb.0:
@@ -61,20 +76,35 @@ define void @vfwmul_v2f116_multiple_users(ptr %x, ptr %y, ptr %z, <2 x half> %a,
}
define void @vfwmul_v2f32_multiple_users(ptr %x, ptr %y, ptr %z, <2 x float> %a, <2 x float> %b, <2 x float> %b2) {
-; NO_FOLDING-LABEL: vfwmul_v2f32_multiple_users:
-; NO_FOLDING: # %bb.0:
-; NO_FOLDING-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
-; NO_FOLDING-NEXT: vfwcvt.f.f.v v11, v8
-; NO_FOLDING-NEXT: vfwcvt.f.f.v v8, v9
-; NO_FOLDING-NEXT: vfwcvt.f.f.v v9, v10
-; NO_FOLDING-NEXT: vsetvli zero, zero, e64, m1, ta, ma
-; NO_FOLDING-NEXT: vfmul.vv v10, v11, v8
-; NO_FOLDING-NEXT: vfadd.vv v11, v11, v9
-; NO_FOLDING-NEXT: vfsub.vv v8, v8, v9
-; NO_FOLDING-NEXT: vse64.v v10, (a0)
-; NO_FOLDING-NEXT: vse64.v v11, (a1)
-; NO_FOLDING-NEXT: vse64.v v8, (a2)
-; NO_FOLDING-NEXT: ret
+; NO_FOLDING1-LABEL: vfwmul_v2f32_multiple_users:
+; NO_FOLDING1: # %bb.0:
+; NO_FOLDING1-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
+; NO_FOLDING1-NEXT: vfwcvt.f.f.v v11, v8
+; NO_FOLDING1-NEXT: vfwcvt.f.f.v v8, v9
+; NO_FOLDING1-NEXT: vfwcvt.f.f.v v9, v10
+; NO_FOLDING1-NEXT: vsetvli zero, zero, e64, m1, ta, ma
+; NO_FOLDING1-NEXT: vfmul.vv v10, v11, v8
+; NO_FOLDING1-NEXT: vfadd.vv v11, v11, v9
+; NO_FOLDING1-NEXT: vfsub.vv v8, v8, v9
+; NO_FOLDING1-NEXT: vse64.v v10, (a0)
+; NO_FOLDING1-NEXT: vse64.v v11, (a1)
+; NO_FOLDING1-NEXT: vse64.v v8, (a2)
+; NO_FOLDING1-NEXT: ret
+;
+; NO_FOLDING2-LABEL: vfwmul_v2f32_multiple_users:
+; NO_FOLDING2: # %bb.0:
+; NO_FOLDING2-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
+; NO_FOLDING2-NEXT: vfwcvt.f.f.v v11, v8
+; NO_FOLDING2-NEXT: vfwcvt.f.f.v v8, v9
+; NO_FOLDING2-NEXT: vsetvli zero, zero, e64, m1, ta, ma
+; NO_FOLDING2-NEXT: vfmul.vv v9, v11, v8
+; NO_FOLDING2-NEXT: vsetvli zero, zero, e32, mf2, ta, ma
+; NO_FOLDING2-NEXT: vfwadd.wv v11, v11, v10
+; NO_FOLDING2-NEXT: vfwsub.wv v8, v8, v10
+; NO_FOLDING2-NEXT: vse64.v v9, (a0)
+; NO_FOLDING2-NEXT: vse64.v v11, (a1)
+; NO_FOLDING2-NEXT: vse64.v v8, (a2)
+; NO_FOLDING2-NEXT: ret
;
; FOLDING-LABEL: vfwmul_v2f32_multiple_users:
; FOLDING: # %bb.0:
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vw-web-simplification.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vw-web-simplification.ll
index 227a428831b60..b093e9e35edad 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vw-web-simplification.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vw-web-simplification.ll
@@ -1,8 +1,8 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=riscv32 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=1 | FileCheck %s --check-prefixes=NO_FOLDING
-; RUN: llc -mtriple=riscv64 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=1 | FileCheck %s --check-prefixes=NO_FOLDING
-; RUN: llc -mtriple=riscv32 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=2 | FileCheck %s --check-prefixes=NO_FOLDING
-; RUN: llc -mtriple=riscv64 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=2 | FileCheck %s --check-prefixes=NO_FOLDING
+; RUN: llc -mtriple=riscv32 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=1 | FileCheck %s --check-prefixes=NO_FOLDING1
+; RUN: llc -mtriple=riscv64 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=1 | FileCheck %s --check-prefixes=NO_FOLDING1
+; RUN: llc -mtriple=riscv32 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=2 | FileCheck %s --check-prefixes=NO_FOLDING2
+; RUN: llc -mtriple=riscv64 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=2 | FileCheck %s --check-prefixes=NO_FOLDING2
; RUN: llc -mtriple=riscv32 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=3 | FileCheck %s --check-prefixes=FOLDING
; RUN: llc -mtriple=riscv64 -mattr=+v -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=3 | FileCheck %s --check-prefixes=FOLDING
; Check that the default value enables the web folding and
@@ -16,21 +16,38 @@
; We need the web size to be at least 3 for the folding to happen, because
; %c has 3 uses.
define <2 x i16> @vwmul_v2i16_multiple_users(ptr %x, ptr %y, ptr %z) {
-; NO_FOLDING-LABEL: vwmul_v2i16_multiple_users:
-; NO_FOLDING: # %bb.0:
-; NO_FOLDING-NEXT: vsetivli zero, 2, e16, mf4, ta, ma
-; NO_FOLDING-NEXT: vle8.v v8, (a0)
-; NO_FOLDING-NEXT: vle8.v v9, (a1)
-; NO_FOLDING-NEXT: vle8.v v10, (a2)
-; NO_FOLDING-NEXT: vsext.vf2 v11, v8
-; NO_FOLDING-NEXT: vsext.vf2 v8, v9
-; NO_FOLDING-NEXT: vsext.vf2 v9, v10
-; NO_FOLDING-NEXT: vmul.vv v8, v11, v8
-; NO_FOLDING-NEXT: vadd.vv v10, v11, v9
-; NO_FOLDING-NEXT: vsub.vv v9, v11, v9
-; NO_FOLDING-NEXT: vor.vv v8, v8, v10
-; NO_FOLDING-NEXT: vor.vv v8, v8, v9
-; NO_FOLDING-NEXT: ret
+; NO_FOLDING1-LABEL: vwmul_v2i16_multiple_users:
+; NO_FOLDING1: # %bb.0:
+; NO_FOLDING1-NEXT: vsetivli zero, 2, e16, mf4, ta, ma
+; NO_FOLDING1-NEXT: vle8.v v8, (a0)
+; NO_FOLDING1-NEXT: vle8.v v9, (a1)
+; NO_FOLDING1-NEXT: vle8.v v10, (a2)
+; NO_FOLDING1-NEXT: vsext.vf2 v11, v8
+; NO_FOLDING1-NEXT: vsext.vf2 v8, v9
+; NO_FOLDING1-NEXT: vsext.vf2 v9, v10
+; NO_FOLDING1-NEXT: vmul.vv v8, v11, v8
+; NO_FOLDING1-NEXT: vadd.vv v10, v11, v9
+; NO_FOLDING1-NEXT: vsub.vv v9, v11, v9
+; NO_FOLDING1-NEXT: vor.vv v8, v8, v10
+; NO_FOLDING1-NEXT: vor.vv v8, v8, v9
+; NO_FOLDING1-NEXT: ret
+;
+; NO_FOLDING2-LABEL: vwmul_v2i16_multiple_users:
+; NO_FOLDING2: # %bb.0:
+; NO_FOLDING2-NEXT: vsetivli zero, 2, e16, mf4, ta, ma
+; NO_FOLDING2-NEXT: vle8.v v8, (a0)
+; NO_FOLDING2-NEXT: vle8.v v9, (a1)
+; NO_FOLDING2-NEXT: vle8.v v10, (a2)
+; NO_FOLDING2-NEXT: vsext.vf2 v11, v8
+; NO_FOLDING2-NEXT: vsext.vf2 v8, v9
+; NO_FOLDING2-NEXT: vmul.vv v8, v11, v8
+; NO_FOLDING2-NEXT: vsetvli zero, zero, e8, mf8, ta, ma
+; NO_FOLDING2-NEXT: vwadd.wv v9, v11, v10
+; NO_FOLDING2-NEXT: vwsub.wv v11, v11, v10
+; NO_FOLDING2-NEXT: vsetvli zero, zero, e16, mf4, ta, ma
+; NO_FOLDING2-NEXT: vor.vv v8, v8, v9
+; NO_FOLDING2-NEXT: vor.vv v8, v8, v11
+; NO_FOLDING2-NEXT: ret
;
; FOLDING-LABEL: vwmul_v2i16_multiple_users:
; FOLDING: # %bb.0:
diff --git a/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave.ll b/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave.ll
index 6a08f5a28a295..c55bf684cc462 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave.ll
@@ -1936,6 +1936,8 @@ define {<vscale x 2 x half>, <vscale x 2 x half>, <vscale x 2 x half>} @vector_d
; CHECK-NEXT: srli a0, a0, 2
; CHECK-NEXT: vsetvli a1, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v10, v8, a0
+; CHECK-NEXT: add a1, a0, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v8, v10, a0
; CHECK-NEXT: addi a0, sp, 16
; CHECK-NEXT: vs2r.v v8, (a0)
@@ -2001,6 +2003,8 @@ define {<vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>} @ve
; CHECK-NEXT: srli a0, a0, 2
; CHECK-NEXT: vsetvli a1, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v10, v8, a0
+; CHECK-NEXT: add a1, a0, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v8, v10, a0
; CHECK-NEXT: addi a0, sp, 16
; CHECK-NEXT: vs2r.v v8, (a0)
@@ -2066,6 +2070,8 @@ define {<vscale x 1 x float>, <vscale x 1 x float>, <vscale x 1 x float>} @vecto
; CHECK-NEXT: srli a0, a0, 3
; CHECK-NEXT: vsetvli a1, zero, e32, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v10, v8, a0
+; CHECK-NEXT: add a1, a0, a0
+; CHECK-NEXT: vsetvli zero, a1, e32, m1, ta, ma
; CHECK-NEXT: vslideup.vx v8, v10, a0
; CHECK-NEXT: addi a0, sp, 16
; CHECK-NEXT: vs2r.v v8, (a0)
@@ -2391,8 +2397,12 @@ define {<vscale x 2 x half>, <vscale x 2 x half>, <vscale x 2 x half>, <vscale x
; CHECK-NEXT: srli a0, a0, 2
; CHECK-NEXT: vsetvli a1, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v11, v9, a0
+; CHECK-NEXT: add a1, a0, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v9, v11, a0
+; CHECK-NEXT: vsetvli a2, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v11, v8, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v8, v11, a0
; CHECK-NEXT: addi a0, sp, 16
; CHECK-NEXT: vs4r.v v8, (a0)
@@ -2478,8 +2488,12 @@ define {<vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vs
; CHECK-NEXT: srli a0, a0, 2
; CHECK-NEXT: vsetvli a1, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v11, v9, a0
+; CHECK-NEXT: add a1, a0, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v9, v11, a0
+; CHECK-NEXT: vsetvli a2, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v11, v8, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v8, v11, a0
; CHECK-NEXT: addi a0, sp, 16
; CHECK-NEXT: vs4r.v v8, (a0)
@@ -2565,8 +2579,12 @@ define {<vscale x 1 x float>, <vscale x 1 x float>, <vscale x 1 x float>, <vscal
; CHECK-NEXT: srli a0, a0, 3
; CHECK-NEXT: vsetvli a1, zero, e32, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v11, v9, a0
+; CHECK-NEXT: add a1, a0, a0
+; CHECK-NEXT: vsetvli zero, a1, e32, m1, ta, ma
; CHECK-NEXT: vslideup.vx v9, v11, a0
+; CHECK-NEXT: vsetvli a2, zero, e32, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v11, v8, a0
+; CHECK-NEXT: vsetvli zero, a1, e32, m1, ta, ma
; CHECK-NEXT: vslideup.vx v8, v11, a0
; CHECK-NEXT: addi a0, sp, 16
; CHECK-NEXT: vs4r.v v8, (a0)
@@ -3061,10 +3079,16 @@ define {<vscale x 2 x half>, <vscale x 2 x half>, <vscale x 2 x half>, <vscale x
; CHECK-NEXT: srli a0, a0, 2
; CHECK-NEXT: vsetvli a1, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v12, v9, a0
+; CHECK-NEXT: add a1, a0, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v9, v12, a0
+; CHECK-NEXT: vsetvli a2, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v12, v8, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v8, v12, a0
+; CHECK-NEXT: vsetvli a2, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v12, v10, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v10, v12, a0
; CHECK-NEXT: addi a0, sp, 16
; CHECK-NEXT: vs4r.v v8, (a0)
@@ -3156,10 +3180,16 @@ define {<vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vs
; CHECK-NEXT: srli a0, a0, 2
; CHECK-NEXT: vsetvli a1, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v12, v9, a0
+; CHECK-NEXT: add a1, a0, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v9, v12, a0
+; CHECK-NEXT: vsetvli a2, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v12, v8, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v8, v12, a0
+; CHECK-NEXT: vsetvli a2, zero, e16, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v12, v10, a0
+; CHECK-NEXT: vsetvli zero, a1, e16, m1, ta, ma
; CHECK-NEXT: vslideup.vx v10, v12, a0
; CHECK-NEXT: addi a0, sp, 16
; CHECK-NEXT: vs4r.v v8, (a0)
@@ -3251,10 +3281,16 @@ define {<vscale x 1 x float>, <vscale x 1 x float>, <vscale x 1 x float>, <vscal
; CHECK-NEXT: srli a0, a0, 3
; CHECK-NEXT: vsetvli a1, zero, e32, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v12, v9, a0
+; CHECK-NEXT: add a1, a0, a0
+; CHECK-NEXT: vsetvli zero, a1, e32, m1, ta, ma
; CHECK-NEXT: vslideup.vx v9, v12, a0
+; CHECK-NEXT: vsetvli a2, zero, e32, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v12, v8, a0
+; CHECK-NEXT: vsetvli zero, a1, e32, m1, ta, ma
; CHECK-NEXT: vslideup.vx v8, v12, a0
+; CHECK-NEXT: vsetvli a2, zero, e32, m1, ta, ma
; CHECK-NEXT: vslidedown.vx v12, v10, a0
+; CHECK-NEXT: vsetvli zero, a1, e32, m1, ta, ma
; CHECK-NEXT: vslideup.vx v10, v12, a0
; CHECK-NEXT: addi a0, sp, 16
; CHECK-NEXT: vs4r.v v8, (a0)
diff --git a/llvm/test/CodeGen/Thumb2/active_lane_mask.ll b/llvm/test/CodeGen/Thumb2/active_lane_mask.ll
index bcd92f81911b2..da13d07abf09b 100644
--- a/llvm/test/CodeGen/Thumb2/active_lane_mask.ll
+++ b/llvm/test/CodeGen/Thumb2/active_lane_mask.ll
@@ -306,9 +306,10 @@ define void @test_width2(ptr nocapture readnone %x, ptr nocapture %y, i8 zeroext
; CHECK-NEXT: ldrne.w r3, [r12]
; CHECK-NEXT: vmovne.32 q0[0], r3
; CHECK-NEXT: lsls r0, r0, #30
-; CHECK-NEXT: itt mi
+; CHECK-NEXT: ittt mi
; CHECK-NEXT: ldrmi.w r0, [r12, #4]
; CHECK-NEXT: vmovmi.32 q0[2], r0
+; CHECK-NEXT: vmovmi.32 q0[3], r0
; CHECK-NEXT: vmrs r3, p0
; CHECK-NEXT: and r0, r3, #1
; CHECK-NEXT: ubfx r3, r3, #8, #1
diff --git a/llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll b/llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll
index 67910e40e7705..53595bc5a5595 100644
--- a/llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll
@@ -110,9 +110,10 @@ define void @foo_sext_v2i64_v2i32(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: ldrne r3, [r2]
; CHECK-LE-NEXT: vmovne.32 q1[0], r3
; CHECK-LE-NEXT: lsls r1, r1, #30
-; CHECK-LE-NEXT: itt mi
+; CHECK-LE-NEXT: ittt mi
; CHECK-LE-NEXT: ldrmi r1, [r2, #4]
; CHECK-LE-NEXT: vmovmi.32 q1[2], r1
+; CHECK-LE-NEXT: vmovmi.32 q1[3], r0
; CHECK-LE-NEXT: vmov r2, s6
; CHECK-LE-NEXT: movs r1, #0
; CHECK-LE-NEXT: vmov r3, s0
@@ -219,9 +220,9 @@ define void @foo_sext_v2i64_v2i32_unaligned(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: sub sp, #4
; CHECK-LE-NEXT: ldrd r12, lr, [r1]
; CHECK-LE-NEXT: movs r1, #0
-; CHECK-LE-NEXT: @ implicit-def: $q0
+; CHECK-LE-NEXT: @ implicit-def: $q1
; CHECK-LE-NEXT: rsbs.w r3, r12, #0
-; CHECK-LE-NEXT: vmov q1[2], q1[0], r12, lr
+; CHECK-LE-NEXT: vmov q0[2], q0[0], r12, lr
; CHECK-LE-NEXT: sbcs.w r3, r1, r12, asr #31
; CHECK-LE-NEXT: csetm r3, lt
; CHECK-LE-NEXT: rsbs.w r4, lr, #0
@@ -232,23 +233,24 @@ define void @foo_sext_v2i64_v2i32_unaligned(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: lsls r3, r1, #31
; CHECK-LE-NEXT: itt ne
; CHECK-LE-NEXT: ldrne r3, [r2]
-; CHECK-LE-NEXT: vmovne.32 q0[0], r3
+; CHECK-LE-NEXT: vmovne.32 q1[0], r3
; CHECK-LE-NEXT: lsls r1, r1, #30
-; CHECK-LE-NEXT: itt mi
+; CHECK-LE-NEXT: ittt mi
; CHECK-LE-NEXT: ldrmi r1, [r2, #4]
-; CHECK-LE-NEXT: vmovmi.32 q0[2], r1
-; CHECK-LE-NEXT: vmov r2, s2
+; CHECK-LE-NEXT: vmovmi.32 q1[2], r1
+; CHECK-LE-NEXT: vmovmi.32 q1[3], r0
+; CHECK-LE-NEXT: vmov r2, s6
; CHECK-LE-NEXT: movs r1, #0
-; CHECK-LE-NEXT: vmov r3, s4
-; CHECK-LE-NEXT: vmov r4, s0
-; CHECK-LE-NEXT: vmov q0[2], q0[0], r4, r2
+; CHECK-LE-NEXT: vmov r3, s0
+; CHECK-LE-NEXT: vmov r4, s4
+; CHECK-LE-NEXT: vmov q1[2], q1[0], r4, r2
; CHECK-LE-NEXT: rsbs r5, r3, #0
; CHECK-LE-NEXT: asr.w r12, r2, #31
; CHECK-LE-NEXT: sbcs.w r2, r1, r3, asr #31
-; CHECK-LE-NEXT: vmov r3, s6
+; CHECK-LE-NEXT: vmov r3, s2
; CHECK-LE-NEXT: csetm r2, lt
; CHECK-LE-NEXT: asr.w lr, r4, #31
-; CHECK-LE-NEXT: vmov q0[3], q0[1], lr, r12
+; CHECK-LE-NEXT: vmov q1[3], q1[1], lr, r12
; CHECK-LE-NEXT: rsbs r5, r3, #0
; CHECK-LE-NEXT: sbcs.w r3, r1, r3, asr #31
; CHECK-LE-NEXT: bfi r1, r2, #0, #1
@@ -256,11 +258,11 @@ define void @foo_sext_v2i64_v2i32_unaligned(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: bfi r1, r2, #1, #1
; CHECK-LE-NEXT: lsls r2, r1, #31
; CHECK-LE-NEXT: itt ne
-; CHECK-LE-NEXT: vmovne r2, r3, d0
+; CHECK-LE-NEXT: vmovne r2, r3, d2
; CHECK-LE-NEXT: strdne r2, r3, [r0]
; CHECK-LE-NEXT: lsls r1, r1, #30
; CHECK-LE-NEXT: itt mi
-; CHECK-LE-NEXT: vmovmi r1, r2, d1
+; CHECK-LE-NEXT: vmovmi r1, r2, d3
; CHECK-LE-NEXT: strdmi r1, r2, [r0, #8]
; CHECK-LE-NEXT: add sp, #4
; CHECK-LE-NEXT: pop {r4, r5, r7, pc}
@@ -363,9 +365,10 @@ define void @foo_zext_v2i64_v2i32(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: ldrne r3, [r2]
; CHECK-LE-NEXT: vmovne.32 q0[0], r3
; CHECK-LE-NEXT: lsls r1, r1, #30
-; CHECK-LE-NEXT: itt mi
+; CHECK-LE-NEXT: ittt mi
; CHECK-LE-NEXT: ldrmi r1, [r2, #4]
; CHECK-LE-NEXT: vmovmi.32 q0[2], r1
+; CHECK-LE-NEXT: vmovmi.32 q0[3], r0
; CHECK-LE-NEXT: vmov r2, s4
; CHECK-LE-NEXT: movs r1, #0
; CHECK-LE-NEXT: vand q0, q0, q2
@@ -478,9 +481,10 @@ define void @foo_zext_v2i64_v2i32_unaligned(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: ldrne r3, [r2]
; CHECK-LE-NEXT: vmovne.32 q0[0], r3
; CHECK-LE-NEXT: lsls r1, r1, #30
-; CHECK-LE-NEXT: itt mi
+; CHECK-LE-NEXT: ittt mi
; CHECK-LE-NEXT: ldrmi r1, [r2, #4]
; CHECK-LE-NEXT: vmovmi.32 q0[2], r1
+; CHECK-LE-NEXT: vmovmi.32 q0[3], r0
; CHECK-LE-NEXT: vmov r2, s4
; CHECK-LE-NEXT: movs r1, #0
; CHECK-LE-NEXT: vand q0, q0, q2
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-4.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-4.ll
index ae4f85ce42a19..321aba7034498 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-4.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-4.ll
@@ -562,14 +562,18 @@ define void @load_i16_stride4_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
;
; AVX2-LABEL: load_i16_stride4_vf8:
; AVX2: # %bb.0:
-; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vmovdqa 32(%rdi), %ymm0
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2
; AVX2-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
+; AVX2-NEXT: vmovdqa (%rdi), %ymm2
+; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT: vpackusdw %xmm0, %xmm1, %xmm0
; AVX2-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-NEXT: vmovdqa 16(%rdi), %xmm2
; AVX2-NEXT: vmovdqa 32(%rdi), %xmm3
@@ -612,14 +616,18 @@ define void @load_i16_stride4_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
;
; AVX2-FP-LABEL: load_i16_stride4_vf8:
; AVX2-FP: # %bb.0:
-; AVX2-FP-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm0
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm2
; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
+; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpackusdw %xmm0, %xmm1, %xmm0
; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-FP-NEXT: vmovdqa 16(%rdi), %xmm2
; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm3
@@ -660,14 +668,18 @@ define void @load_i16_stride4_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
;
; AVX2-FCP-LABEL: load_i16_stride4_vf8:
; AVX2-FCP: # %bb.0:
-; AVX2-FCP-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm0
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm2
; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FCP-NEXT: vpackusdw %xmm0, %xmm1, %xmm0
; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm2
; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm3
@@ -1021,26 +1033,30 @@ define void @load_i16_stride4_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
;
; AVX2-LABEL: load_i16_stride4_vf16:
; AVX2: # %bb.0:
-; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
-; AVX2-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
-; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT: vmovdqa (%rdi), %ymm0
+; AVX2-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
+; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
+; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
+; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm5
+; AVX2-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
+; AVX2-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
+; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
+; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4],ymm2[5,6,7],ymm1[8],ymm2[9,10,11],ymm1[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm4
+; AVX2-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5,6,7],ymm0[8],ymm2[9,10,11],ymm0[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-NEXT: vmovdqa 16(%rdi), %xmm2
; AVX2-NEXT: vmovdqa 32(%rdi), %xmm3
@@ -1125,26 +1141,30 @@ define void @load_i16_stride4_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
;
; AVX2-FP-LABEL: load_i16_stride4_vf16:
; AVX2-FP: # %bb.0:
-; AVX2-FP-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
-; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm0
+; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX2-FP-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
+; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
+; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm5
+; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
+; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
+; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4],ymm2[5,6,7],ymm1[8],ymm2[9,10,11],ymm1[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm4
+; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5,6,7],ymm0[8],ymm2[9,10,11],ymm0[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-FP-NEXT: vmovdqa 16(%rdi), %xmm3
; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm4
@@ -1225,11 +1245,15 @@ define void @load_i16_stride4_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP: # %bb.0:
; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm1
; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm2
-; AVX2-FCP-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm4
-; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm0
+; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpxor %xmm4, %xmm4, %xmm4
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1,2,3],ymm3[4],ymm4[5,6,7],ymm3[8],ymm4[9,10,11],ymm3[12],ymm4[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm5
+; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm3, %xmm3
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1,2,3],ymm0[4],ymm4[5,6,7],ymm0[8],ymm4[9,10,11],ymm0[12],ymm4[13,14,15]
; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm4
; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm0, %xmm0
; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
@@ -2106,46 +2130,54 @@ define void @load_i16_stride4_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-LABEL: load_i16_stride4_vf32:
; AVX2: # %bb.0:
; AVX2-NEXT: subq $168, %rsp
-; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vmovdqa 128(%rdi), %ymm0
+; AVX2-NEXT: vmovdqa 160(%rdi), %ymm2
+; AVX2-NEXT: vmovdqa (%rdi), %ymm3
+; AVX2-NEXT: vmovdqa 32(%rdi), %ymm4
+; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm5, %xmm6
+; AVX2-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
+; AVX2-NEXT: vpackusdw %xmm5, %xmm5, %xmm5
+; AVX2-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm5
+; AVX2-NEXT: vpblendw {{.*#+}} ymm6 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm6, %xmm7
+; AVX2-NEXT: vpackusdw %xmm7, %xmm6, %xmm6
+; AVX2-NEXT: vpackusdw %xmm6, %xmm6, %xmm6
+; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
+; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm6
+; AVX2-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
+; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm1[1,2,3],ymm3[4],ymm1[5,6,7],ymm3[8],ymm1[9,10,11],ymm3[12],ymm1[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm6
+; AVX2-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
+; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
+; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
+; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
-; AVX2-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
-; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm5
+; AVX2-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
+; AVX2-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
+; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
+; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm4
+; AVX2-NEXT: vpackusdw %xmm4, %xmm2, %xmm2
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
+; AVX2-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-NEXT: vmovdqa %xmm1, (%rsp) # 16-byte Spill
@@ -2338,46 +2370,54 @@ define void @load_i16_stride4_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FP-LABEL: load_i16_stride4_vf32:
; AVX2-FP: # %bb.0:
; AVX2-FP-NEXT: subq $184, %rsp
-; AVX2-FP-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm0
+; AVX2-FP-NEXT: vmovdqa 160(%rdi), %ymm2
+; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm3
+; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm4
+; AVX2-FP-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm5, %xmm6
+; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
+; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm5, %xmm5
+; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm5
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm6, %xmm7
+; AVX2-FP-NEXT: vpackusdw %xmm7, %xmm6, %xmm6
+; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm6, %xmm6
+; AVX2-FP-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm6
+; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm1[1,2,3],ymm3[4],ymm1[5,6,7],ymm3[8],ymm1[9,10,11],ymm3[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm6
+; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
+; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
-; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm5
+; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
+; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
+; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm4
+; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm13
; AVX2-FP-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -2560,149 +2600,157 @@ define void @load_i16_stride4_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: subq $104, %rsp
; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %ymm3
; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm6
-; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm9
-; AVX2-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm5
+; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm0
+; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %ymm1
+; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm10
+; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm7
-; AVX2-FCP-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm4
+; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm4
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpxor %xmm8, %xmm8, %xmm8
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm8[1,2,3],ymm4[4],ymm8[5,6,7],ymm4[8],ymm8[9,10,11],ymm4[12],ymm8[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm4, %xmm6
+; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm8[1,2,3],ymm2[4],ymm8[5,6,7],ymm2[8],ymm8[9,10,11],ymm2[12],ymm8[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm6
+; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm2, %xmm2
; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm2, %xmm2
-; AVX2-FCP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,2,2,3,0,2,4,6]
-; AVX2-FCP-NEXT: vpermd %ymm7, %ymm2, %ymm5
+; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,2,2,3,0,2,4,6]
+; AVX2-FCP-NEXT: vpermd %ymm7, %ymm9, %ymm6
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,16,17,20,21,24,25,28,29]
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm8
-; AVX2-FCP-NEXT: vpermd %ymm9, %ymm2, %ymm10
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm9
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4,5],ymm8[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm8[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm8
-; AVX2-FCP-NEXT: vpackusdw %xmm8, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm8
-; AVX2-FCP-NEXT: vpackusdw %xmm8, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm11
+; AVX2-FCP-NEXT: vpermd %ymm10, %ymm9, %ymm10
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm12
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4,5],ymm11[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm8[1,2,3],ymm1[4],ymm8[5,6,7],ymm1[8],ymm8[9,10,11],ymm1[12],ymm8[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm8[1,2,3],ymm0[4],ymm8[5,6,7],ymm0[8],ymm8[9,10,11],ymm0[12],ymm8[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm2
+; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
; AVX2-FCP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vpermd %ymm6, %ymm2, %ymm9
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm1
-; AVX2-FCP-NEXT: vpermd %ymm3, %ymm2, %ymm6
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm8
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3,4,5],ymm1[6,7]
+; AVX2-FCP-NEXT: vpermd %ymm5, %ymm9, %ymm15
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm15, %ymm1
+; AVX2-FCP-NEXT: vpermd %ymm3, %ymm9, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm2
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm14
-; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm3
+; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm4
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,2,3,10,11,8,9,10,11,12,13,14,15]
-; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm3, %xmm0
-; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm14, %xmm8
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm8[0],xmm0[0],xmm8[1],xmm0[1]
+; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm4, %xmm0
+; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm14, %xmm2
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm13
-; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm2
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [2,3,10,11,8,9,10,11,8,9,10,11,12,13,14,15]
-; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm2, %xmm4
-; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm11
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm11[0],xmm4[0],xmm11[1],xmm4[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],xmm8[2,3]
+; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm5
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [2,3,10,11,8,9,10,11,8,9,10,11,12,13,14,15]
+; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm8
+; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm13, %xmm11
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm8[0,1],xmm2[2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm5
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm6, %ymm6
; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm10, %ymm10
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm10[0,1,2,3,4,5],ymm5[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm5[4,5,6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm10[0,1,2,3,4,5],ymm6[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm6[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %xmm11
-; AVX2-FCP-NEXT: vmovdqa 176(%rdi), %xmm15
-; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm15, %xmm4
-; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm10
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm10[0],xmm4[0],xmm10[1],xmm4[1]
+; AVX2-FCP-NEXT: vmovdqa 176(%rdi), %xmm2
+; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm10
+; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm12
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm12[0],xmm10[0],xmm12[1],xmm10[1]
; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %xmm10
; AVX2-FCP-NEXT: vmovdqa 144(%rdi), %xmm12
-; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm12, %xmm0
-; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm10, %xmm1
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm9, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm6, %ymm4
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
+; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm12, %xmm0
+; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm3
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm9, %ymm3
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [1,3,2,3,1,3,5,7]
-; AVX2-FCP-NEXT: vpermd %ymm7, %ymm6, %ymm4
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm7 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [1,3,2,3,1,3,5,7]
+; AVX2-FCP-NEXT: vpermd %ymm7, %ymm0, %ymm7
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,16,17,20,21,24,25,28,29]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm0
; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm1
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm3
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm1[6,7]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm14[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,1,2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm1[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm14[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[2,0,2,3,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm13[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm5[0,1],xmm3[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm3 # 32-byte Folded Reload
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm5 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm6
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm9
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0,1,2,3,4,5],ymm6[6,7]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm15[3,1,2,3]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm9[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm11[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm15[0],xmm0[1],xmm15[1]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm11[3,1,2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm9[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm12[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm12[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm11[2,0,2,3,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm10[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm8[0],xmm15[0],xmm8[1],xmm15[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm8[0,1],xmm4[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm4
; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm6
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3,4,5],ymm4[6,7]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm7
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,1,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm14[0,1,3,1,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm13[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm4[4,5,6,7]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm2
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm3
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm14[0,1,3,1,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm13[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm5[0,1],xmm1[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm3
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,1,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm9[0,1,3,1,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm11[0,1,3,1,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm12[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm11[3,1,2,3,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm10[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rsi)
-; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm3, (%rsi)
-; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rdx)
-; AVX2-FCP-NEXT: vmovups (%rsp), %ymm3 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm3, (%rdx)
-; AVX2-FCP-NEXT: vmovdqa %ymm0, 32(%rcx)
-; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm0, (%rcx)
-; AVX2-FCP-NEXT: vmovdqa %ymm2, 32(%r8)
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rsi)
+; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm2, (%rsi)
+; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rdx)
+; AVX2-FCP-NEXT: vmovups (%rsp), %ymm2 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm2, (%rdx)
+; AVX2-FCP-NEXT: vmovdqa %ymm4, 32(%rcx)
+; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm2, (%rcx)
+; AVX2-FCP-NEXT: vmovdqa %ymm0, 32(%r8)
; AVX2-FCP-NEXT: vmovdqa %ymm1, (%r8)
; AVX2-FCP-NEXT: addq $104, %rsp
; AVX2-FCP-NEXT: vzeroupper
@@ -4354,86 +4402,102 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-LABEL: load_i16_stride4_vf64:
; AVX2: # %bb.0:
; AVX2-NEXT: subq $696, %rsp # imm = 0x2B8
-; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
-; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
-; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
+; AVX2-NEXT: vmovdqa 128(%rdi), %ymm0
+; AVX2-NEXT: vmovdqa 160(%rdi), %ymm1
+; AVX2-NEXT: vmovdqa 384(%rdi), %ymm3
+; AVX2-NEXT: vmovdqa 416(%rdi), %ymm4
+; AVX2-NEXT: vmovdqa 256(%rdi), %ymm5
+; AVX2-NEXT: vmovdqa 288(%rdi), %ymm6
+; AVX2-NEXT: vmovdqa (%rdi), %ymm7
+; AVX2-NEXT: vmovdqa 32(%rdi), %ymm8
+; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT: vpblendw {{.*#+}} ymm9 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm9, %xmm10
+; AVX2-NEXT: vpackusdw %xmm10, %xmm9, %xmm9
+; AVX2-NEXT: vpackusdw %xmm9, %xmm9, %xmm9
+; AVX2-NEXT: vinserti128 $1, %xmm9, %ymm0, %ymm9
+; AVX2-NEXT: vpblendw {{.*#+}} ymm10 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm10, %xmm11
+; AVX2-NEXT: vpackusdw %xmm11, %xmm10, %xmm10
+; AVX2-NEXT: vpackusdw %xmm10, %xmm10, %xmm10
+; AVX2-NEXT: vinserti128 $1, %xmm10, %ymm0, %ymm10
+; AVX2-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm2[1,2,3],ymm8[4],ymm2[5,6,7],ymm8[8],ymm2[9,10,11],ymm8[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm8, %xmm10
+; AVX2-NEXT: vpackusdw %xmm10, %xmm8, %xmm8
+; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0],ymm2[1,2,3],ymm7[4],ymm2[5,6,7],ymm7[8],ymm2[9,10,11],ymm7[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm7, %xmm10
+; AVX2-NEXT: vpackusdw %xmm10, %xmm7, %xmm7
+; AVX2-NEXT: vpackusdw %xmm8, %xmm7, %xmm7
+; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm9[4,5,6,7]
+; AVX2-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpblendw {{.*#+}} ymm7 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm7, %xmm8
+; AVX2-NEXT: vpackusdw %xmm8, %xmm7, %xmm7
+; AVX2-NEXT: vpackusdw %xmm7, %xmm7, %xmm7
+; AVX2-NEXT: vinserti128 $1, %xmm7, %ymm0, %ymm7
+; AVX2-NEXT: vpblendw {{.*#+}} ymm8 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm8, %xmm9
+; AVX2-NEXT: vpackusdw %xmm9, %xmm8, %xmm8
+; AVX2-NEXT: vpackusdw %xmm8, %xmm8, %xmm8
+; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm0, %ymm8
+; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0],ymm2[1,2,3],ymm6[4],ymm2[5,6,7],ymm6[8],ymm2[9,10,11],ymm6[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm6, %xmm8
+; AVX2-NEXT: vpackusdw %xmm8, %xmm6, %xmm6
+; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7],ymm5[8],ymm2[9,10,11],ymm5[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm5, %xmm8
+; AVX2-NEXT: vpackusdw %xmm8, %xmm5, %xmm5
+; AVX2-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
+; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]
+; AVX2-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm5, %xmm6
+; AVX2-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
+; AVX2-NEXT: vpackusdw %xmm5, %xmm5, %xmm5
+; AVX2-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm5
+; AVX2-NEXT: vpblendw {{.*#+}} ymm6 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm6, %xmm7
+; AVX2-NEXT: vpackusdw %xmm7, %xmm6, %xmm6
+; AVX2-NEXT: vpackusdw %xmm6, %xmm6, %xmm6
+; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
+; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm2[1,2,3],ymm4[4],ymm2[5,6,7],ymm4[8],ymm2[9,10,11],ymm4[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm6
+; AVX2-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
+; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm2[1,2,3],ymm3[4],ymm2[5,6,7],ymm3[8],ymm2[9,10,11],ymm3[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm6
+; AVX2-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
-; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
+; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
-; AVX2-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
-; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm5
+; AVX2-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
+; AVX2-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
+; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
+; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4],ymm2[5,6,7],ymm1[8],ymm2[9,10,11],ymm1[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm4
+; AVX2-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5,6,7],ymm0[8],ymm2[9,10,11],ymm0[12],ymm2[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vmovdqa 256(%rdi), %xmm4
; AVX2-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -4858,86 +4922,102 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FP-LABEL: load_i16_stride4_vf64:
; AVX2-FP: # %bb.0:
; AVX2-FP-NEXT: subq $712, %rsp # imm = 0x2C8
-; AVX2-FP-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
-; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
-; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
+; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm0
+; AVX2-FP-NEXT: vmovdqa 160(%rdi), %ymm1
+; AVX2-FP-NEXT: vmovdqa 384(%rdi), %ymm3
+; AVX2-FP-NEXT: vmovdqa 416(%rdi), %ymm4
+; AVX2-FP-NEXT: vmovdqa 256(%rdi), %ymm5
+; AVX2-FP-NEXT: vmovdqa 288(%rdi), %ymm6
+; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm7
+; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm8
+; AVX2-FP-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm9 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm9, %xmm10
+; AVX2-FP-NEXT: vpackusdw %xmm10, %xmm9, %xmm9
+; AVX2-FP-NEXT: vpackusdw %xmm9, %xmm9, %xmm9
+; AVX2-FP-NEXT: vinserti128 $1, %xmm9, %ymm0, %ymm9
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm10 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm10, %xmm11
+; AVX2-FP-NEXT: vpackusdw %xmm11, %xmm10, %xmm10
+; AVX2-FP-NEXT: vpackusdw %xmm10, %xmm10, %xmm10
+; AVX2-FP-NEXT: vinserti128 $1, %xmm10, %ymm0, %ymm10
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm2[1,2,3],ymm8[4],ymm2[5,6,7],ymm8[8],ymm2[9,10,11],ymm8[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm8, %xmm10
+; AVX2-FP-NEXT: vpackusdw %xmm10, %xmm8, %xmm8
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0],ymm2[1,2,3],ymm7[4],ymm2[5,6,7],ymm7[8],ymm2[9,10,11],ymm7[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm7, %xmm10
+; AVX2-FP-NEXT: vpackusdw %xmm10, %xmm7, %xmm7
+; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm7, %xmm7
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm9[4,5,6,7]
+; AVX2-FP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm7 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm7, %xmm8
+; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm7, %xmm7
+; AVX2-FP-NEXT: vpackusdw %xmm7, %xmm7, %xmm7
+; AVX2-FP-NEXT: vinserti128 $1, %xmm7, %ymm0, %ymm7
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm8 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm8, %xmm9
+; AVX2-FP-NEXT: vpackusdw %xmm9, %xmm8, %xmm8
+; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm8, %xmm8
+; AVX2-FP-NEXT: vinserti128 $1, %xmm8, %ymm0, %ymm8
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0],ymm2[1,2,3],ymm6[4],ymm2[5,6,7],ymm6[8],ymm2[9,10,11],ymm6[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm6, %xmm8
+; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm6, %xmm6
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7],ymm5[8],ymm2[9,10,11],ymm5[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm5, %xmm8
+; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm5, %xmm5
+; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]
+; AVX2-FP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm5, %xmm6
+; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
+; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm5, %xmm5
+; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm5
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm6, %xmm7
+; AVX2-FP-NEXT: vpackusdw %xmm7, %xmm6, %xmm6
+; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm6, %xmm6
+; AVX2-FP-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm2[1,2,3],ymm4[4],ymm2[5,6,7],ymm4[8],ymm2[9,10,11],ymm4[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm6
+; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm2[1,2,3],ymm3[4],ymm2[5,6,7],ymm3[8],ymm2[9,10,11],ymm3[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm6
+; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
-; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
+; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
-; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
-; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm5
+; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
+; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
+; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4],ymm2[5,6,7],ymm1[8],ymm2[9,10,11],ymm1[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm4
+; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5,6,7],ymm0[8],ymm2[9,10,11],ymm0[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vmovdqa 256(%rdi), %xmm5
; AVX2-FP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -5333,246 +5413,262 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
;
; AVX2-FCP-LABEL: load_i16_stride4_vf64:
; AVX2-FCP: # %bb.0:
-; AVX2-FCP-NEXT: subq $680, %rsp # imm = 0x2A8
-; AVX2-FCP-NEXT: vmovdqa 320(%rdi), %ymm7
+; AVX2-FCP-NEXT: subq $664, %rsp # imm = 0x298
+; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %ymm0
+; AVX2-FCP-NEXT: vmovdqa 320(%rdi), %ymm9
+; AVX2-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 352(%rdi), %ymm10
+; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %ymm4
+; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %ymm5
+; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm8
+; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm7
; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 352(%rdi), %ymm6
-; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm5
-; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm4
-; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3,4,5,6,7]
; AVX2-FCP-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm2
-; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FCP-NEXT: vpackusdw %xmm0, %xmm2, %xmm2
-; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [0,2,2,3,0,2,4,6]
-; AVX2-FCP-NEXT: vpermd %ymm4, %ymm0, %ymm3
-; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,16,17,20,21,24,25,28,29]
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm3
-; AVX2-FCP-NEXT: vpermd %ymm5, %ymm0, %ymm5
-; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm5
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm3
-; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm5
-; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm3, %xmm3
-; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm3, %xmm3
-; AVX2-FCP-NEXT: vpermd %ymm6, %ymm0, %ymm2
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm5
-; AVX2-FCP-NEXT: vpermd %ymm7, %ymm0, %ymm6
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm7
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm5
-; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm3, %xmm3
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm5, %xmm7
-; AVX2-FCP-NEXT: vpackusdw %xmm7, %xmm5, %xmm5
-; AVX2-FCP-NEXT: vmovdqa 448(%rdi), %ymm7
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm1[1,2,3],ymm3[4],ymm1[5,6,7],ymm3[8],ymm1[9,10,11],ymm3[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm6
+; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm6
+; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm2, %xmm2
+; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm2, %xmm6
+; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,2,2,3,0,2,4,6]
+; AVX2-FCP-NEXT: vpermd %ymm7, %ymm2, %ymm7
; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm5, %xmm3
-; AVX2-FCP-NEXT: vmovdqa 480(%rdi), %ymm5
-; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermd %ymm5, %ymm0, %ymm5
-; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm5
-; AVX2-FCP-NEXT: vpermd %ymm7, %ymm0, %ymm14
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm14, %ymm7
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm5
-; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm3, %xmm3
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm5
-; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %ymm5
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,16,17,20,21,24,25,28,29]
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm7
+; AVX2-FCP-NEXT: vpermd %ymm8, %ymm2, %ymm8
+; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm8, %ymm8
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm1[1,2,3],ymm5[4],ymm1[5,6,7],ymm5[8],ymm1[9,10,11],ymm5[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm5, %xmm6
+; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm4, %xmm6
+; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
+; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
+; AVX2-FCP-NEXT: vpermd %ymm10, %ymm2, %ymm13
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm13, %ymm5
+; AVX2-FCP-NEXT: vpermd %ymm9, %ymm2, %ymm10
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm10, %ymm6
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
+; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %ymm6
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm4, %xmm5
+; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm5
+; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vmovdqa 448(%rdi), %ymm5
; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm3
-; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermd %ymm3, %ymm0, %ymm9
-; AVX2-FCP-NEXT: vpermd %ymm5, %ymm0, %ymm3
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm0
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm5
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vmovdqa 480(%rdi), %ymm4
+; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpermd %ymm4, %ymm2, %ymm4
+; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm4, %ymm4
+; AVX2-FCP-NEXT: vpermd %ymm5, %ymm2, %ymm12
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm12, %ymm5
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %xmm1
-; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 304(%rdi), %xmm12
+; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %ymm0
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm4
+; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm4
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm4
+; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
+; AVX2-FCP-NEXT: vpackusdw %xmm0, %xmm1, %xmm0
+; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %ymm1
+; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm4
+; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpermd %ymm4, %ymm2, %ymm6
+; AVX2-FCP-NEXT: vpermd %ymm1, %ymm2, %ymm4
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm4, %ymm2
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %xmm2
+; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 304(%rdi), %xmm0
+; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,2,3,10,11,8,9,10,11,12,13,14,15]
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm12, %xmm0
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm1
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm2, %xmm1
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %xmm10
-; AVX2-FCP-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %xmm0
+; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vmovdqa 272(%rdi), %xmm7
-; AVX2-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [2,3,10,11,8,9,10,11,8,9,10,11,12,13,14,15]
-; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm7
-; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm10, %xmm8
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm7[0,1],xmm1[2,3]
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm2
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm6, %ymm6
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm2
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [2,3,10,11,8,9,10,11,8,9,10,11,12,13,14,15]
+; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm7, %xmm8
+; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm14
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm14[0],xmm8[0],xmm14[1],xmm8[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm13, %ymm13
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm10, %ymm10
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm13[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm10[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm13
; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm1
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm6
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm2, %xmm7
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]
-; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm7
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm10
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm13, %xmm14
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm14 = xmm14[0],xmm10[0],xmm14[1],xmm10[1]
+; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm10
; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm15
-; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm15, %xmm11
-; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm13
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm6 = xmm11[0,1],xmm6[2,3]
+; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm15, %xmm0
+; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm10, %xmm9
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm14[2,3]
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm11
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm8, %ymm9
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm13
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm13[0,1,2,3,4,5],ymm11[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm11[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm8, %ymm14
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm14[0,1,2,3,4,5],ymm9[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %xmm8
; AVX2-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 176(%rdi), %xmm6
-; AVX2-FCP-NEXT: vmovdqa %xmm6, (%rsp) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm6
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm11
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm6 = xmm11[0],xmm6[0],xmm11[1],xmm6[1]
+; AVX2-FCP-NEXT: vmovdqa 176(%rdi), %xmm0
+; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm9
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]
; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %xmm8
; AVX2-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 144(%rdi), %xmm11
-; AVX2-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm11, %xmm11
-; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm8, %xmm13
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm6 = xmm11[0,1],xmm6[2,3]
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm9
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm3, %ymm3
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm9[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3],ymm3[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %xmm6
-; AVX2-FCP-NEXT: vmovdqa 432(%rdi), %xmm13
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm13, %xmm3
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm5
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm9 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
+; AVX2-FCP-NEXT: vmovdqa 144(%rdi), %xmm9
+; AVX2-FCP-NEXT: vmovdqa %xmm9, (%rsp) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm9, %xmm9
+; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm8, %xmm14
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm9 = xmm14[0],xmm9[0],xmm14[1],xmm9[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm9[0,1],xmm0[2,3]
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm6, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm4
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm9[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %xmm9
+; AVX2-FCP-NEXT: vmovdqa 432(%rdi), %xmm8
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm0
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm9, %xmm4
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1]
; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %xmm5
-; AVX2-FCP-NEXT: vmovdqa 400(%rdi), %xmm3
-; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm11
-; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm5, %xmm0
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm9[2,3]
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm9
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm14, %ymm11
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqa 400(%rdi), %xmm4
+; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm14
+; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm5, %xmm2
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm14[0],xmm2[1],xmm14[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
+; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm12, %ymm12
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm12[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [1,3,2,3,1,3,5,7]
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload
-; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm11 # 32-byte Folded Reload
-; AVX2-FCP-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm9
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm11
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm12
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm12[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,1,2,3]
-; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm13[3,1,2,3]
+; AVX2-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm15[3,1,2,3]
-; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[3,1,2,3]
-; AVX2-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm9[4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm6[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm12[0],xmm1[0],xmm12[1],xmm1[1]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm15[3,1,2,3]
+; AVX2-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm10[3,1,2,3]
+; AVX2-FCP-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm6[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm10[0,1],xmm1[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm2
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm12[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
+; AVX2-FCP-NEXT: # xmm2 = mem[3,1,2,3]
; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm14 = mem[3,1,2,3]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm14[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1]
-; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
-; AVX2-FCP-NEXT: # xmm11 = mem[3,1,2,3]
-; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
-; AVX2-FCP-NEXT: # xmm9 = mem[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm11[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm9[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm12 = xmm15[0],xmm12[0],xmm15[1],xmm12[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm12[0,1],xmm2[2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm14[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm7[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
+; AVX2-FCP-NEXT: # xmm10 = mem[3,1,2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm13 = xmm12[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm10[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm13 = xmm15[0],xmm13[0],xmm15[1],xmm13[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm13[0,1],xmm2[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm2
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm13[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm8[3,1,2,3]
; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm9[3,1,2,3]
; AVX2-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm3[3,1,2,3]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm5[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm13[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm12[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm6[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm9[0],xmm2[0],xmm9[1],xmm2[1]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm4[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm5[3,1,2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm13[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm9[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm4[0,1],xmm2[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 32-byte Folded Reload
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm0
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm7, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm8, %ymm0
+; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm1
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3,4,5],ymm0[6,7]
-; AVX2-FCP-NEXT: vpshufd $231, (%rsp), %xmm6 # 16-byte Folded Reload
+; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm6 = mem[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm5 = mem[3,1,2,3]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm6[0,1,2,0,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm5[0,1,2,0,4,5,6,7]
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
+; AVX2-FCP-NEXT: vpshufd $231, (%rsp), %xmm4 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm4 = mem[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm2 = mem[3,1,2,3]
@@ -5583,9 +5679,9 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm0
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm1
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FCP-NEXT: vpshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm1 = mem[0,1,3,1,4,5,6,7]
@@ -5600,21 +5696,21 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm15 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm0
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm1
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FCP-NEXT: vpshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm1 = mem[0,1,3,1,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm14[0,1,3,1,4,5,6,7]
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm11[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm9[0],xmm3[0],xmm9[1],xmm3[1]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm12[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm10[0],xmm3[0],xmm10[1],xmm3[1]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm3
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm8, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm3
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm6[0,1,3,1,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,3,1,4,5,6,7]
@@ -5625,9 +5721,9 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm2
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm3, %ymm3
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm3
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FCP-NEXT: vpshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm3 = mem[0,1,3,1,4,5,6,7]
@@ -5635,7 +5731,7 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: # xmm4 = mem[0,1,3,1,4,5,6,7]
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm13[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm12[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm9[3,1,2,3,4,5,6,7]
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
@@ -5667,7 +5763,7 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: vmovdqa %ymm1, 32(%r8)
; AVX2-FCP-NEXT: vmovdqa %ymm0, 64(%r8)
; AVX2-FCP-NEXT: vmovdqa %ymm15, (%r8)
-; AVX2-FCP-NEXT: addq $680, %rsp # imm = 0x2A8
+; AVX2-FCP-NEXT: addq $664, %rsp # imm = 0x298
; AVX2-FCP-NEXT: vzeroupper
; AVX2-FCP-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
index 418c987ab9a30..d282e051d35b3 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
@@ -444,6 +444,7 @@ define void @store_i16_stride5_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2,3],ymm5[4,5,6],ymm2[7],ymm5[8,9],ymm2[10,11],ymm5[12,13,14],ymm2[15]
; AVX512-NEXT: vpternlogq {{.*#+}} ymm2 = ymm3 ^ (mem & (ymm2 ^ ymm3))
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,1,3,4,5,6,7]
@@ -471,6 +472,7 @@ define void @store_i16_stride5_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2,3],ymm5[4,5,6],ymm2[7],ymm5[8,9],ymm2[10,11],ymm5[12,13,14],ymm2[15]
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm2 = ymm3 ^ (mem & (ymm2 ^ ymm3))
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,6,7,14,15,u,u,8,9,10,11,12,13,14,15]
; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
@@ -497,6 +499,7 @@ define void @store_i16_stride5_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2,3],ymm5[4,5,6],ymm2[7],ymm5[8,9],ymm2[10,11],ymm5[12,13,14],ymm2[15]
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm2 = ymm3 ^ (mem & (ymm2 ^ ymm3))
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,1,3,4,5,6,7]
@@ -524,6 +527,7 @@ define void @store_i16_stride5_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2,3],ymm5[4,5,6],ymm2[7],ymm5[8,9],ymm2[10,11],ymm5[12,13,14],ymm2[15]
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm2 = ymm3 ^ (mem & (ymm2 ^ ymm3))
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,6,7,14,15,u,u,8,9,10,11,12,13,14,15]
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
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 6f50d61f4d1f4..fc24976d651db 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
@@ -1390,18 +1390,19 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
; AVX512-NEXT: vpternlogd {{.*#+}} ymm8 = (mem & ~ymm8) | ymm7
; AVX512-NEXT: vpternlogq {{.*#+}} ymm8 = ymm2 ^ (mem & (ymm8 ^ ymm2))
-; AVX512-NEXT: vpsrlq $48, %xmm4, %xmm2
-; AVX512-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm3[1],xmm2[1]
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm8, %zmm2
+; AVX512-NEXT: vpsrlq $48, %xmm4, %xmm4
+; AVX512-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm2[3,4],xmm0[5,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm3[3,4],xmm0[5,6,7]
; AVX512-NEXT: vpsrld $16, %xmm6, %xmm1
; AVX512-NEXT: vpunpckhdq {{.*#+}} xmm1 = xmm5[2],xmm1[2],xmm5[3],xmm1[3]
-; AVX512-NEXT: vpbroadcastd 12(%r10), %xmm2
-; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3,4],xmm1[5,6],xmm2[7]
+; AVX512-NEXT: vpbroadcastd 12(%r10), %xmm3
+; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4],xmm1[5,6],xmm3[7]
; AVX512-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4],xmm1[5,6,7]
; AVX512-NEXT: vmovdqa %xmm0, 96(%rax)
-; AVX512-NEXT: vmovdqa %ymm8, 64(%rax)
+; AVX512-NEXT: vmovdqa %ymm2, 64(%rax)
; AVX512-NEXT: vmovdqa64 %zmm11, (%rax)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
@@ -1450,6 +1451,7 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
; AVX512-FCP-NEXT: vpternlogd {{.*#+}} ymm8 = (mem & ~ymm8) | ymm7
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm8 = ymm6 ^ (mem & (ymm8 ^ ymm6))
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm8, %zmm6
; AVX512-FCP-NEXT: vpsrlq $48, %xmm3, %xmm3
; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm3[1]
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
@@ -1461,8 +1463,8 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3,4],xmm1[5,6],xmm2[7]
; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4],xmm1[5,6,7]
; AVX512-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
-; AVX512-FCP-NEXT: vmovdqa %ymm8, 64(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm11, (%rax)
+; AVX512-FCP-NEXT: vmovdqa %ymm6, 64(%rax)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
@@ -1513,18 +1515,19 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
; AVX512DQ-NEXT: vpternlogd {{.*#+}} ymm8 = (mem & ~ymm8) | ymm7
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm8 = ymm2 ^ (mem & (ymm8 ^ ymm2))
-; AVX512DQ-NEXT: vpsrlq $48, %xmm4, %xmm2
-; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm3[1],xmm2[1]
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm8, %zmm2
+; AVX512DQ-NEXT: vpsrlq $48, %xmm4, %xmm4
+; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm2[3,4],xmm0[5,6,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm3[3,4],xmm0[5,6,7]
; AVX512DQ-NEXT: vpsrld $16, %xmm6, %xmm1
; AVX512DQ-NEXT: vpunpckhdq {{.*#+}} xmm1 = xmm5[2],xmm1[2],xmm5[3],xmm1[3]
-; AVX512DQ-NEXT: vpbroadcastd 12(%r10), %xmm2
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3,4],xmm1[5,6],xmm2[7]
+; AVX512DQ-NEXT: vpbroadcastd 12(%r10), %xmm3
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4],xmm1[5,6],xmm3[7]
; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4],xmm1[5,6,7]
; AVX512DQ-NEXT: vmovdqa %xmm0, 96(%rax)
-; AVX512DQ-NEXT: vmovdqa %ymm8, 64(%rax)
+; AVX512DQ-NEXT: vmovdqa %ymm2, 64(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm11, (%rax)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
@@ -1573,6 +1576,7 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
; AVX512DQ-FCP-NEXT: vpternlogd {{.*#+}} ymm8 = (mem & ~ymm8) | ymm7
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm8 = ymm6 ^ (mem & (ymm8 ^ ymm6))
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm8, %zmm6
; AVX512DQ-FCP-NEXT: vpsrlq $48, %xmm3, %xmm3
; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm3[1]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
@@ -1584,8 +1588,8 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3,4],xmm1[5,6],xmm2[7]
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4],xmm1[5,6,7]
; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm8, 64(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, (%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm6, 64(%rax)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
index 86efcf9c57616..7bbb082f07741 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
@@ -718,6 +718,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,0,1,1]
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX512-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm2))
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm2
; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -726,7 +727,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512-NEXT: vpbroadcastw %xmm1, %xmm1
; AVX512-NEXT: vpternlogq {{.*#+}} xmm1 = xmm1 ^ (mem & (xmm1 ^ xmm0))
; AVX512-NEXT: vmovq %xmm1, 32(%r9)
-; AVX512-NEXT: vmovdqa %ymm3, (%r9)
+; AVX512-NEXT: vmovdqa %ymm2, (%r9)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
@@ -748,6 +749,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,0,1,1]
; AVX512-FCP-NEXT: vpermd %ymm3, %ymm4, %ymm3
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm2))
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm2
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -756,7 +758,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512-FCP-NEXT: vpbroadcastw %xmm1, %xmm1
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} xmm1 = xmm1 ^ (mem & (xmm1 ^ xmm0))
; AVX512-FCP-NEXT: vmovq %xmm1, 32(%r9)
-; AVX512-FCP-NEXT: vmovdqa %ymm3, (%r9)
+; AVX512-FCP-NEXT: vmovdqa %ymm2, (%r9)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
@@ -778,6 +780,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,0,1,1]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm2))
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm2
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -786,7 +789,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-NEXT: vpbroadcastw %xmm1, %xmm1
; AVX512DQ-NEXT: vpternlogq {{.*#+}} xmm1 = xmm1 ^ (mem & (xmm1 ^ xmm0))
; AVX512DQ-NEXT: vmovq %xmm1, 32(%r9)
-; AVX512DQ-NEXT: vmovdqa %ymm3, (%r9)
+; AVX512DQ-NEXT: vmovdqa %ymm2, (%r9)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
@@ -808,6 +811,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,0,1,1]
; AVX512DQ-FCP-NEXT: vpermd %ymm3, %ymm4, %ymm3
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm2))
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm2
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -816,7 +820,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-FCP-NEXT: vpbroadcastw %xmm1, %xmm1
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} xmm1 = xmm1 ^ (mem & (xmm1 ^ xmm0))
; AVX512DQ-FCP-NEXT: vmovq %xmm1, 32(%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm3, (%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, (%r9)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
@@ -840,6 +844,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512BW-NEXT: movl $554189328, %ecx # imm = 0x21084210
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -873,6 +878,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512BW-FCP-NEXT: movl $554189328, %ecx # imm = 0x21084210
; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -906,6 +912,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-BW-NEXT: movl $554189328, %ecx # imm = 0x21084210
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -939,6 +946,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-BW-FCP-NEXT: movl $554189328, %ecx # imm = 0x21084210
; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
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 f4055a953badd..524161408bc0f 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
@@ -2088,6 +2088,7 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,1,9],zero,ymm6[u,u,u,u,2,10],zero,ymm6[u,u,u,u,19,27],zero,ymm6[u,u,u,u,20,28],zero,ymm6[u,u,u,u,21]
; AVX512-NEXT: vpternlogq {{.*#+}} ymm6 = ymm6 | (ymm7 & ~mem)
; AVX512-NEXT: vpternlogq {{.*#+}} ymm6 = ymm6 ^ (mem & (ymm6 ^ ymm5))
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm6, %zmm5
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,2,3]
; AVX512-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u],zero,zero,xmm3[6,14,u,u,u],zero,zero,xmm3[7,15,u,u,u]
; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,2,3]
@@ -2098,9 +2099,9 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm0[13,u,u,u,u],zero,zero,xmm0[14,u,u,u,u],zero,zero,xmm0[15]
; AVX512-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX512-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm3))
-; AVX512-NEXT: vmovdqa %ymm6, 64(%rax)
; AVX512-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512-NEXT: vmovdqa64 %zmm8, (%rax)
+; AVX512-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
@@ -2155,6 +2156,7 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,1,9],zero,ymm7[u,u,u,u,2,10],zero,ymm7[u,u,u,u,19,27],zero,ymm7[u,u,u,u,20,28],zero,ymm7[u,u,u,u,21]
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm7 = ymm7 | (ymm8 & ~mem)
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm7 = ymm7 ^ (mem & (ymm7 ^ ymm5))
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm7, %zmm5
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,2,3]
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u],zero,zero,xmm4[6,14,u,u,u],zero,zero,xmm4[7,15,u,u,u]
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,2,3]
@@ -2165,9 +2167,9 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm0[13,u,u,u,u],zero,zero,xmm0[14,u,u,u,u],zero,zero,xmm0[15]
; AVX512-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm3))
-; AVX512-FCP-NEXT: vmovdqa %ymm7, 64(%rax)
; AVX512-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm6, (%rax)
+; AVX512-FCP-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
@@ -2225,6 +2227,7 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,1,9],zero,ymm6[u,u,u,u,2,10],zero,ymm6[u,u,u,u,19,27],zero,ymm6[u,u,u,u,20,28],zero,ymm6[u,u,u,u,21]
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm6 = ymm6 | (ymm7 & ~mem)
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm6 = ymm6 ^ (mem & (ymm6 ^ ymm5))
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm6, %zmm5
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,2,3]
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u],zero,zero,xmm3[6,14,u,u,u],zero,zero,xmm3[7,15,u,u,u]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,2,3]
@@ -2235,9 +2238,9 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm0[13,u,u,u,u],zero,zero,xmm0[14,u,u,u,u],zero,zero,xmm0[15]
; AVX512DQ-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX512DQ-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm3))
-; AVX512DQ-NEXT: vmovdqa %ymm6, 64(%rax)
; AVX512DQ-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%rax)
+; AVX512DQ-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
@@ -2292,6 +2295,7 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,1,9],zero,ymm7[u,u,u,u,2,10],zero,ymm7[u,u,u,u,19,27],zero,ymm7[u,u,u,u,20,28],zero,ymm7[u,u,u,u,21]
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm7 = ymm7 | (ymm8 & ~mem)
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm7 = ymm7 ^ (mem & (ymm7 ^ ymm5))
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm7, %zmm5
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,2,3]
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u],zero,zero,xmm4[6,14,u,u,u],zero,zero,xmm4[7,15,u,u,u]
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,2,3]
@@ -2302,9 +2306,9 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm0[13,u,u,u,u],zero,zero,xmm0[14,u,u,u,u],zero,zero,xmm0[15]
; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm3))
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm7, 64(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, (%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
@@ -2312,80 +2316,81 @@ 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), %xmm3
-; AVX512BW-NEXT: vmovdqa (%r9), %xmm4
-; AVX512BW-NEXT: vmovdqa (%r10), %xmm2
-; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm0
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm5 = 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: 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: vpor %ymm5, %ymm6, %ymm5
-; AVX512BW-NEXT: vinserti128 $1, (%rsi), %ymm1, %ymm1
-; 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: vmovdqa (%rdi), %xmm3
+; AVX512BW-NEXT: vmovdqa (%rdx), %xmm4
+; AVX512BW-NEXT: vmovdqa (%r8), %xmm0
+; AVX512BW-NEXT: vmovdqa (%r9), %xmm2
+; AVX512BW-NEXT: vmovdqa (%r10), %xmm1
+; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512BW-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512BW-NEXT: vpermw %ymm1, %ymm5, %ymm5
+; AVX512BW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm6
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm6[1,3,1,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[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, %edx # imm = 0x4081020
+; AVX512BW-NEXT: kmovd %edx, %k1
+; AVX512BW-NEXT: vmovdqu8 %ymm5, %ymm7 {%k1}
+; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm4, %ymm4
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm4[1,3,3,1]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[3,19],zero,zero,zero,zero,zero,ymm5[28,20],zero,zero,zero,zero,zero,ymm5[29,21],zero
+; AVX512BW-NEXT: vinserti128 $1, (%rsi), %ymm3, %ymm3
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm3[3,1,1,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[1],zero,zero,zero,zero,zero,ymm8[10,2],zero,zero,zero,zero,zero,ymm8[11,3],zero,zero,zero,zero,zero,ymm8[20,28],zero,zero,zero,zero,zero,ymm8[21,29],zero,zero,zero
+; AVX512BW-NEXT: vpor %ymm5, %ymm8, %ymm5
+; AVX512BW-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu8 %ymm5, %ymm6 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,0,2]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[0,8],zero,zero,zero,zero,zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[19,27],zero,zero,zero,zero,zero,ymm5[20,28],zero,zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm0[0,2,0,2]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,ymm7[0,8],zero,zero,zero,zero,zero,ymm7[1,9],zero,zero,zero,zero,zero,ymm7[18,26],zero,zero,zero,zero,zero,ymm7[19,27],zero,zero,zero,zero,zero,ymm7[20,28]
-; AVX512BW-NEXT: vpor %ymm5, %ymm7, %ymm5
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5
-; AVX512BW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm6
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm6[2,3,0,1]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,ymm8[4],zero,zero,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
+; AVX512BW-NEXT: vmovdqu8 %ymm7, %ymm5 {%k1}
+; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm5
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,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
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm4[2,3,0,1]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,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
; AVX512BW-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm6[0,2,0,2]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[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, %ymm7, %zmm8, %zmm7
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,ymm3[5],zero,zero,zero,zero,zero,zero,ymm3[6],zero,zero,zero,zero,zero,zero,zero,ymm3[23],zero,zero,zero,zero,zero,zero,ymm3[24],zero,zero,zero,zero,zero,zero
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm3[2,3,0,1]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,zero,zero,ymm9[5],zero,zero,zero,zero,zero,zero,ymm9[6],zero,zero,zero,zero,zero,ymm9[23],zero,zero,zero,zero,zero,zero,ymm9[24],zero,zero,zero,zero,zero,zero,ymm9[25]
+; AVX512BW-NEXT: vpor %ymm9, %ymm8, %ymm8
+; AVX512BW-NEXT: movl $202911840, %ecx # imm = 0xC183060
+; AVX512BW-NEXT: kmovd %ecx, %k1
+; AVX512BW-NEXT: vmovdqu8 %ymm7, %ymm8 {%k1}
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm3[0,2,0,2]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[0,8],zero,zero,zero,zero,zero,ymm7[1,9],zero,zero,zero,zero,zero,ymm7[2,10],zero,zero,zero,zero,zero,ymm7[19,27],zero,zero,zero,zero,zero,ymm7[20,28],zero,zero
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm4[0,2,0,2]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,ymm9[0,8],zero,zero,zero,zero,zero,ymm9[1,9],zero,zero,zero,zero,zero,ymm9[18,26],zero,zero,zero,zero,zero,ymm9[19,27],zero,zero,zero,zero,zero,ymm9[20,28]
+; AVX512BW-NEXT: vpor %ymm7, %ymm9, %ymm7
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm6[2,3,0,1]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,ymm9[4],zero,zero,zero,zero,zero,zero,ymm9[5],zero,zero,zero,zero,zero,zero,ymm9[6],zero,zero,zero,zero,zero,ymm9[23],zero,zero,zero,zero,zero,zero,ymm9[24],zero,zero,zero
+; AVX512BW-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,0,2]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[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, %ymm8, %zmm6, %zmm6
; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm8 = [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 %zmm2, %zmm8, %zmm8
+; AVX512BW-NEXT: vpermw %zmm1, %zmm8, %zmm8
; AVX512BW-NEXT: movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
; AVX512BW-NEXT: kmovq %rcx, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm8, %zmm7 {%k1}
+; AVX512BW-NEXT: vmovdqu8 %zmm8, %zmm6 {%k1}
; AVX512BW-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
; AVX512BW-NEXT: kmovq %rcx, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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]
-; AVX512BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
-; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,6,7,7,7]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,1,3,2]
+; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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 {{.*#+}} xmm0 = xmm0[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,7,7,7]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,1,3,2]
; AVX512BW-NEXT: movw $-32510, %cx # imm = 0x8102
; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu8 %xmm4, %xmm3 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm1[1,3,2,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm4[6,14],zero,zero,zero,zero,zero,xmm4[7,15],zero,zero,zero,zero,zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm0[1,3,2,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,zero,zero,xmm7[6,14],zero,zero,zero,zero,zero,xmm7[7,15],zero,zero,zero
-; AVX512BW-NEXT: vpor %xmm4, %xmm7, %xmm4
+; AVX512BW-NEXT: vmovdqu8 %xmm1, %xmm0 {%k1}
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm3[1,3,2,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[6,14],zero,zero,zero,zero,zero,xmm1[7,15],zero,zero,zero,zero,zero
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm2 = ymm4[1,3,2,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,zero,zero,xmm2[6,14],zero,zero,zero,zero,zero,xmm2[7,15],zero,zero,zero
+; AVX512BW-NEXT: vpor %xmm1, %xmm2, %xmm1
; AVX512BW-NEXT: movw $-7741, %cx # imm = 0xE1C3
; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu8 %xmm3, %xmm4 {%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 %ymm2, %ymm3, %ymm2
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm6[1,3,1,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[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 %ymm2, %ymm3 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,3,1]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm0[1,9],zero,zero,zero,zero,zero,ymm0[2,10],zero,zero,zero,zero,zero,ymm0[3,19],zero,zero,zero,zero,zero,ymm0[28,20],zero,zero,zero,zero,zero,ymm0[29,21],zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,1,1,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[1],zero,zero,zero,zero,zero,ymm1[10,2],zero,zero,zero,zero,zero,ymm1[11,3],zero,zero,zero,zero,zero,ymm1[20,28],zero,zero,zero,zero,zero,ymm1[21,29],zero,zero,zero
-; AVX512BW-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512BW-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
-; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu8 %ymm3, %ymm0 {%k1}
-; AVX512BW-NEXT: vmovdqa %ymm0, 64(%rax)
-; AVX512BW-NEXT: vmovdqa %xmm4, 96(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512BW-NEXT: vmovdqu8 %xmm0, %xmm1 {%k1}
+; AVX512BW-NEXT: vmovdqa %xmm1, 96(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm7, (%rax)
+; AVX512BW-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
@@ -2398,45 +2403,46 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2
; AVX512BW-FCP-NEXT: vmovdqa (%r9), %xmm3
; AVX512BW-FCP-NEXT: vmovdqa (%r10), %xmm4
-; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [1,9,2,10,1,9,2,10]
+; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
; AVX512BW-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpermi2d %ymm3, %ymm2, %ymm5
+; AVX512BW-FCP-NEXT: vpermw %ymm4, %ymm5, %ymm5
; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm6
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm6[0,2,0,2]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm5 = zmm5[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: vpermq {{.*#+}} ymm7 = ymm6[1,3,1,3]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[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-FCP-NEXT: movl $67637280, %edx # imm = 0x4081020
+; AVX512BW-FCP-NEXT: kmovd %edx, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %ymm5, %ymm7 {%k1}
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm1[1,3,3,1]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[3,19],zero,zero,zero,zero,zero,ymm5[28,20],zero,zero,zero,zero,zero,ymm5[29,21],zero
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm0[3,1,1,3]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[1],zero,zero,zero,zero,zero,ymm8[10,2],zero,zero,zero,zero,zero,ymm8[11,3],zero,zero,zero,zero,zero,ymm8[20,28],zero,zero,zero,zero,zero,ymm8[21,29],zero,zero,zero
+; AVX512BW-FCP-NEXT: vpor %ymm5, %ymm8, %ymm5
+; AVX512BW-FCP-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
+; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %ymm7, %ymm5 {%k1}
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm5
+; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [1,9,2,10,1,9,2,10]
+; AVX512BW-FCP-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX512BW-FCP-NEXT: vpermi2d %ymm3, %ymm2, %ymm7
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,0,2]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm6 = zmm6[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 {{.*#+}} zmm7 = [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 %zmm4, %zmm7, %zmm7
-; AVX512BW-FCP-NEXT: movabsq $4647998506761461824, %rdx # imm = 0x4081020408102040
-; AVX512BW-FCP-NEXT: kmovq %rdx, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
-; AVX512BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
+; AVX512BW-FCP-NEXT: movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
+; AVX512BW-FCP-NEXT: kmovq %rcx, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm6 {%k1}
; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,0,4,0,0,1,4,5,1,5,0,0,1,5,2,6]
; AVX512BW-FCP-NEXT: vpermd %zmm0, %zmm7, %zmm8
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm8 = zmm8[0,8],zero,zero,zero,zero,zero,zmm8[1,9],zero,zero,zero,zero,zero,zmm8[2,10],zero,zero,zero,zero,zero,zmm8[19,27],zero,zero,zero,zero,zero,zmm8[20,28],zero,zero,zero,zero,zero,zmm8[33,37],zero,zero,zero,zero,zero,zmm8[34,38],zero,zero,zero,zero,zero,zmm8[51,55],zero,zero,zero,zero,zero,zmm8[56,60],zero,zero,zero,zero,zero,zmm8[57]
-; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512BW-FCP-NEXT: vpermd %zmm1, %zmm7, %zmm7
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm7 = zero,zero,zmm7[0,8],zero,zero,zero,zero,zero,zmm7[1,9],zero,zero,zero,zero,zero,zmm7[18,26],zero,zero,zero,zero,zero,zmm7[19,27],zero,zero,zero,zero,zero,zmm7[20,28],zero,zero,zero,zero,zero,zmm7[33,37],zero,zero,zero,zero,zero,zmm7[34,38],zero,zero,zero,zero,zero,zmm7[51,55],zero,zero,zero,zero,zero,zmm7[56,60],zero,zero,zero,zero
; AVX512BW-FCP-NEXT: vporq %zmm8, %zmm7, %zmm7
; AVX512BW-FCP-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
; AVX512BW-FCP-NEXT: kmovq %rcx, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm7 {%k1}
-; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
-; AVX512BW-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpermw %ymm4, %ymm5, %ymm5
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[1,3,1,3]
-; AVX512BW-FCP-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-FCP-NEXT: movl $67637280, %ecx # imm = 0x4081020
-; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %ymm5, %ymm6 {%k1}
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm1[1,3,3,1]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[3,19],zero,zero,zero,zero,zero,ymm5[28,20],zero,zero,zero,zero,zero,ymm5[29,21],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm0[3,1,1,3]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[1],zero,zero,zero,zero,zero,ymm8[10,2],zero,zero,zero,zero,zero,ymm8[11,3],zero,zero,zero,zero,zero,ymm8[20,28],zero,zero,zero,zero,zero,ymm8[21,29],zero,zero,zero
-; AVX512BW-FCP-NEXT: vpor %ymm5, %ymm8, %ymm5
-; AVX512BW-FCP-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
-; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %ymm6, %ymm5 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,2,3]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[6,14],zero,zero,zero,zero,zero,xmm0[7,15],zero,zero,zero,zero,zero
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,2,3]
@@ -2449,9 +2455,9 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: movw $-7741, %cx # imm = 0xE1C3
; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %xmm1, %xmm0 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512BW-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
+; AVX512BW-FCP-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512BW-FCP-NEXT: vzeroupper
; AVX512BW-FCP-NEXT: retq
;
@@ -2459,80 +2465,81 @@ 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), %xmm3
-; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm4
-; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm2
-; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm0
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm5 = 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: 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: vpor %ymm5, %ymm6, %ymm5
-; AVX512DQ-BW-NEXT: vinserti128 $1, (%rsi), %ymm1, %ymm1
-; 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: vmovdqa (%rdi), %xmm3
+; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm4
+; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm0
+; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm2
+; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm1
+; AVX512DQ-BW-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512DQ-BW-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpermw %ymm1, %ymm5, %ymm5
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm6
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm6[1,3,1,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[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, %edx # imm = 0x4081020
+; AVX512DQ-BW-NEXT: kmovd %edx, %k1
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm5, %ymm7 {%k1}
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm4, %ymm4
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm5 = ymm4[1,3,3,1]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[3,19],zero,zero,zero,zero,zero,ymm5[28,20],zero,zero,zero,zero,zero,ymm5[29,21],zero
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%rsi), %ymm3, %ymm3
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm3[3,1,1,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[1],zero,zero,zero,zero,zero,ymm8[10,2],zero,zero,zero,zero,zero,ymm8[11,3],zero,zero,zero,zero,zero,ymm8[20,28],zero,zero,zero,zero,zero,ymm8[21,29],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpor %ymm5, %ymm8, %ymm5
+; AVX512DQ-BW-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %ymm5, %ymm6 {%k1}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,0,2]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[0,8],zero,zero,zero,zero,zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[19,27],zero,zero,zero,zero,zero,ymm5[20,28],zero,zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm0[0,2,0,2]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,ymm7[0,8],zero,zero,zero,zero,zero,ymm7[1,9],zero,zero,zero,zero,zero,ymm7[18,26],zero,zero,zero,zero,zero,ymm7[19,27],zero,zero,zero,zero,zero,ymm7[20,28]
-; AVX512DQ-BW-NEXT: vpor %ymm5, %ymm7, %ymm5
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm6
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm6[2,3,0,1]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,ymm8[4],zero,zero,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
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm7, %ymm5 {%k1}
+; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm5
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,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
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm4[2,3,0,1]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,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
; AVX512DQ-BW-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm6[0,2,0,2]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[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, %ymm7, %zmm8, %zmm7
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,ymm3[5],zero,zero,zero,zero,zero,zero,ymm3[6],zero,zero,zero,zero,zero,zero,zero,ymm3[23],zero,zero,zero,zero,zero,zero,ymm3[24],zero,zero,zero,zero,zero,zero
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm3[2,3,0,1]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,zero,zero,ymm9[5],zero,zero,zero,zero,zero,zero,ymm9[6],zero,zero,zero,zero,zero,ymm9[23],zero,zero,zero,zero,zero,zero,ymm9[24],zero,zero,zero,zero,zero,zero,ymm9[25]
+; AVX512DQ-BW-NEXT: vpor %ymm9, %ymm8, %ymm8
+; AVX512DQ-BW-NEXT: movl $202911840, %ecx # imm = 0xC183060
+; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm7, %ymm8 {%k1}
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm3[0,2,0,2]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[0,8],zero,zero,zero,zero,zero,ymm7[1,9],zero,zero,zero,zero,zero,ymm7[2,10],zero,zero,zero,zero,zero,ymm7[19,27],zero,zero,zero,zero,zero,ymm7[20,28],zero,zero
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm4[0,2,0,2]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,ymm9[0,8],zero,zero,zero,zero,zero,ymm9[1,9],zero,zero,zero,zero,zero,ymm9[18,26],zero,zero,zero,zero,zero,ymm9[19,27],zero,zero,zero,zero,zero,ymm9[20,28]
+; AVX512DQ-BW-NEXT: vpor %ymm7, %ymm9, %ymm7
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm6[2,3,0,1]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,ymm9[4],zero,zero,zero,zero,zero,zero,ymm9[5],zero,zero,zero,zero,zero,zero,ymm9[6],zero,zero,zero,zero,zero,ymm9[23],zero,zero,zero,zero,zero,zero,ymm9[24],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,0,2]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[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, %ymm8, %zmm6, %zmm6
; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm8 = [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 %zmm2, %zmm8, %zmm8
+; AVX512DQ-BW-NEXT: vpermw %zmm1, %zmm8, %zmm8
; AVX512DQ-BW-NEXT: movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
; AVX512DQ-BW-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm8, %zmm7 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm8, %zmm6 {%k1}
; AVX512DQ-BW-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
; AVX512DQ-BW-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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]
-; 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: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,6,7,7,7]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,1,3,2]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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 {{.*#+}} xmm0 = xmm0[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,7,7,7]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,1,3,2]
; AVX512DQ-BW-NEXT: movw $-32510, %cx # imm = 0x8102
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %xmm4, %xmm3 {%k1}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm1[1,3,2,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm4[6,14],zero,zero,zero,zero,zero,xmm4[7,15],zero,zero,zero,zero,zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm0[1,3,2,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,zero,zero,xmm7[6,14],zero,zero,zero,zero,zero,xmm7[7,15],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpor %xmm4, %xmm7, %xmm4
+; AVX512DQ-BW-NEXT: vmovdqu8 %xmm1, %xmm0 {%k1}
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm3[1,3,2,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[6,14],zero,zero,zero,zero,zero,xmm1[7,15],zero,zero,zero,zero,zero
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm2 = ymm4[1,3,2,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,zero,zero,xmm2[6,14],zero,zero,zero,zero,zero,xmm2[7,15],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpor %xmm1, %xmm2, %xmm1
; AVX512DQ-BW-NEXT: movw $-7741, %cx # imm = 0xE1C3
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %xmm3, %xmm4 {%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 %ymm2, %ymm3, %ymm2
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm6[1,3,1,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[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 %ymm2, %ymm3 {%k1}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,3,1]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm0[1,9],zero,zero,zero,zero,zero,ymm0[2,10],zero,zero,zero,zero,zero,ymm0[3,19],zero,zero,zero,zero,zero,ymm0[28,20],zero,zero,zero,zero,zero,ymm0[29,21],zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,1,1,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[1],zero,zero,zero,zero,zero,ymm1[10,2],zero,zero,zero,zero,zero,ymm1[11,3],zero,zero,zero,zero,zero,ymm1[20,28],zero,zero,zero,zero,zero,ymm1[21,29],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-BW-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
-; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %ymm3, %ymm0 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa %ymm0, 64(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa %xmm4, 96(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512DQ-BW-NEXT: vmovdqu8 %xmm0, %xmm1 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqa %xmm1, 96(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, (%rax)
+; AVX512DQ-BW-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512DQ-BW-NEXT: vzeroupper
; AVX512DQ-BW-NEXT: retq
;
@@ -2545,45 +2552,46 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %xmm3
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r10), %xmm4
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [1,9,2,10,1,9,2,10]
+; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
; AVX512DQ-BW-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpermi2d %ymm3, %ymm2, %ymm5
+; AVX512DQ-BW-FCP-NEXT: vpermw %ymm4, %ymm5, %ymm5
; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm6
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm6[0,2,0,2]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm5 = zmm5[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: vpermq {{.*#+}} ymm7 = ymm6[1,3,1,3]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[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-FCP-NEXT: movl $67637280, %edx # imm = 0x4081020
+; AVX512DQ-BW-FCP-NEXT: kmovd %edx, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm5, %ymm7 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm1[1,3,3,1]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[3,19],zero,zero,zero,zero,zero,ymm5[28,20],zero,zero,zero,zero,zero,ymm5[29,21],zero
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm0[3,1,1,3]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[1],zero,zero,zero,zero,zero,ymm8[10,2],zero,zero,zero,zero,zero,ymm8[11,3],zero,zero,zero,zero,zero,ymm8[20,28],zero,zero,zero,zero,zero,ymm8[21,29],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm5, %ymm8, %ymm5
+; AVX512DQ-BW-FCP-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
+; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm7, %ymm5 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm5
+; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [1,9,2,10,1,9,2,10]
+; AVX512DQ-BW-FCP-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vpermi2d %ymm3, %ymm2, %ymm7
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,0,2]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm6 = zmm6[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 {{.*#+}} zmm7 = [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 %zmm4, %zmm7, %zmm7
-; AVX512DQ-BW-FCP-NEXT: movabsq $4647998506761461824, %rdx # imm = 0x4081020408102040
-; AVX512DQ-BW-FCP-NEXT: kmovq %rdx, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
+; AVX512DQ-BW-FCP-NEXT: movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
+; AVX512DQ-BW-FCP-NEXT: kmovq %rcx, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm6 {%k1}
; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,0,4,0,0,1,4,5,1,5,0,0,1,5,2,6]
; AVX512DQ-BW-FCP-NEXT: vpermd %zmm0, %zmm7, %zmm8
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm8 = zmm8[0,8],zero,zero,zero,zero,zero,zmm8[1,9],zero,zero,zero,zero,zero,zmm8[2,10],zero,zero,zero,zero,zero,zmm8[19,27],zero,zero,zero,zero,zero,zmm8[20,28],zero,zero,zero,zero,zero,zmm8[33,37],zero,zero,zero,zero,zero,zmm8[34,38],zero,zero,zero,zero,zero,zmm8[51,55],zero,zero,zero,zero,zero,zmm8[56,60],zero,zero,zero,zero,zero,zmm8[57]
-; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512DQ-BW-FCP-NEXT: vpermd %zmm1, %zmm7, %zmm7
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm7 = zero,zero,zmm7[0,8],zero,zero,zero,zero,zero,zmm7[1,9],zero,zero,zero,zero,zero,zmm7[18,26],zero,zero,zero,zero,zero,zmm7[19,27],zero,zero,zero,zero,zero,zmm7[20,28],zero,zero,zero,zero,zero,zmm7[33,37],zero,zero,zero,zero,zero,zmm7[34,38],zero,zero,zero,zero,zero,zmm7[51,55],zero,zero,zero,zero,zero,zmm7[56,60],zero,zero,zero,zero
; AVX512DQ-BW-FCP-NEXT: vporq %zmm8, %zmm7, %zmm7
; AVX512DQ-BW-FCP-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
; AVX512DQ-BW-FCP-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm7 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
-; AVX512DQ-BW-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpermw %ymm4, %ymm5, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[1,3,1,3]
-; AVX512DQ-BW-FCP-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-FCP-NEXT: movl $67637280, %ecx # imm = 0x4081020
-; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm5, %ymm6 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm1[1,3,3,1]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[3,19],zero,zero,zero,zero,zero,ymm5[28,20],zero,zero,zero,zero,zero,ymm5[29,21],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm0[3,1,1,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[1],zero,zero,zero,zero,zero,ymm8[10,2],zero,zero,zero,zero,zero,ymm8[11,3],zero,zero,zero,zero,zero,ymm8[20,28],zero,zero,zero,zero,zero,ymm8[21,29],zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm5, %ymm8, %ymm5
-; AVX512DQ-BW-FCP-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
-; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm6, %ymm5 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,2,3]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[6,14],zero,zero,zero,zero,zero,xmm0[7,15],zero,zero,zero,zero,zero
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,2,3]
@@ -2596,9 +2604,9 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: movw $-7741, %cx # imm = 0xE1C3
; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %xmm1, %xmm0 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm5, 64(%rax)
; AVX512DQ-BW-FCP-NEXT: vzeroupper
; AVX512DQ-BW-FCP-NEXT: retq
%in.vec0 = load <16 x i8>, ptr %in.vecptr0, align 64
>From 08b5c9471ca62408a893b9334600feb419723c14 Mon Sep 17 00:00:00 2001
From: Bjorn Pettersson <bjorn.a.pettersson at ericsson.com>
Date: Fri, 6 Jun 2025 12:17:09 +0200
Subject: [PATCH 2/4] clang-format fixes
---
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 10 +++++-----
llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 12 ++++++------
2 files changed, 11 insertions(+), 11 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 8da0bca158eda..d37163eb7cc1c 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -27423,11 +27423,11 @@ SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
// If inserting an UNDEF, just return the original vector (unless it makes the
// result more poisonous).
- if (N1.isUndef()){
+ if (N1.isUndef()) {
if (VT.isFixedLengthVector()) {
unsigned SubVecNumElts = N1.getValueType().getVectorNumElements();
- APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(),
- InsIdx, InsIdx + SubVecNumElts);
+ APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(), InsIdx,
+ InsIdx + SubVecNumElts);
if (DAG.isGuaranteedNotToBePoison(N0, EltMask))
return N0;
} else if (DAG.isGuaranteedNotToBePoison(N0))
@@ -27448,8 +27448,8 @@ SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
return N1.getOperand(0);
if (VT.isFixedLengthVector() && N1VT.isFixedLengthVector()) {
unsigned SubVecNumElts = N1VT.getVectorNumElements();
- APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(),
- InsIdx, InsIdx + SubVecNumElts);
+ APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(), InsIdx,
+ InsIdx + SubVecNumElts);
if (DAG.isGuaranteedNotToBePoison(N1.getOperand(0), ~EltMask))
return N1.getOperand(0);
} else if (DAG.isGuaranteedNotToBePoison(N1.getOperand(0)))
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 4e2ad10ae5637..88a0164769d4c 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -7931,8 +7931,8 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
// But not if skipping the insert could make the result more poisonous.
if (N2.isUndef()) {
if (N3C && VT.isFixedLengthVector()) {
- APInt EltMask = APInt::getOneBitSet(VT.getVectorNumElements(),
- N3C->getZExtValue());
+ APInt EltMask =
+ APInt::getOneBitSet(VT.getVectorNumElements(), N3C->getZExtValue());
if (isGuaranteedNotToBePoison(N1, EltMask))
return N1;
} else if (isGuaranteedNotToBePoison(N1))
@@ -7985,8 +7985,8 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
if (VT.isFixedLengthVector() && N2VT.isFixedLengthVector()) {
unsigned LoBit = N3->getAsZExtVal();
unsigned HiBit = LoBit + N2VT.getVectorNumElements();
- APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(),
- LoBit, HiBit);
+ APInt EltMask =
+ APInt::getBitsSet(VT.getVectorNumElements(), LoBit, HiBit);
if (isGuaranteedNotToBePoison(N2.getOperand(0), ~EltMask))
return N2.getOperand(0);
} else if (isGuaranteedNotToBePoison(N2.getOperand(0)))
@@ -7999,8 +7999,8 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
if (VT.isFixedLengthVector()) {
unsigned LoBit = N3->getAsZExtVal();
unsigned HiBit = LoBit + N2VT.getVectorNumElements();
- APInt EltMask = APInt::getBitsSet(VT.getVectorNumElements(),
- LoBit, HiBit);
+ APInt EltMask =
+ APInt::getBitsSet(VT.getVectorNumElements(), LoBit, HiBit);
if (isGuaranteedNotToBePoison(N1, EltMask))
return N1;
} else if (isGuaranteedNotToBePoison(N1))
>From 4ef20388672d3d046873954767850002e1d952f2 Mon Sep 17 00:00:00 2001
From: Bjorn Pettersson <bjorn.a.pettersson at ericsson.com>
Date: Sun, 8 Jun 2025 20:39:55 +0200
Subject: [PATCH 3/4] Fixup: Use getFreeze in DAGCombiner.
Also avoid some not needed isGuaranteedNotToBePoison calls in getNode.
---
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 7 +-
.../aarch64-neon-vector-insert-uaddlv.ll | 3 +-
.../CodeGen/AArch64/arm64-build-vector.ll | 4 +-
.../CodeGen/AArch64/arm64-vector-insertion.ll | 2 -
.../AArch64/concat-vector-add-combine.ll | 19 +-
llvm/test/CodeGen/Thumb2/active_lane_mask.ll | 3 +-
llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll | 36 +-
llvm/test/CodeGen/Thumb2/mve-vld3.ll | 435 +++------
.../any_extend_vector_inreg_of_broadcast.ll | 4 +-
.../X86/merge-consecutive-loads-128.ll | 78 +-
llvm/test/CodeGen/X86/mmx-build-vector.ll | 150 +--
llvm/test/CodeGen/X86/oddshuffles.ll | 196 ++--
.../vector-interleaved-load-i16-stride-5.ll | 915 +++++++++---------
.../vector-interleaved-load-i8-stride-3.ll | 212 ++--
.../vector-interleaved-store-i16-stride-5.ll | 4 -
.../vector-interleaved-store-i16-stride-7.ll | 81 +-
.../vector-interleaved-store-i64-stride-7.ll | 30 +-
.../vector-interleaved-store-i8-stride-3.ll | 210 ++--
.../vector-interleaved-store-i8-stride-4.ll | 62 +-
.../vector-interleaved-store-i8-stride-5.ll | 16 +-
.../vector-interleaved-store-i8-stride-7.ll | 243 +++--
.../CodeGen/X86/vector-shuffle-combining.ll | 89 +-
llvm/test/CodeGen/X86/vector-trunc.ll | 140 ++-
.../CodeGen/X86/x86-interleaved-access.ll | 446 ++++-----
.../CodeGen/X86/zero_extend_vector_inreg.ll | 21 +-
.../zero_extend_vector_inreg_of_broadcast.ll | 3 +-
26 files changed, 1708 insertions(+), 1701 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index d37163eb7cc1c..f7bd02c85b665 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -22981,9 +22981,8 @@ SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
IndexC->getZExtValue());
if (DAG.isGuaranteedNotToBePoison(InVec, EltMask))
return InVec;
- } else if (DAG.isGuaranteedNotToBePoison(InVec)) {
- return InVec;
}
+ return DAG.getFreeze(InVec);
}
if (!IndexC)
@@ -27430,8 +27429,8 @@ SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
InsIdx + SubVecNumElts);
if (DAG.isGuaranteedNotToBePoison(N0, EltMask))
return N0;
- } else if (DAG.isGuaranteedNotToBePoison(N0))
- return N0;
+ }
+ return DAG.getFreeze(N0);
}
// If this is an insert of an extracted vector into an undef/poison vector, we
diff --git a/llvm/test/CodeGen/AArch64/aarch64-neon-vector-insert-uaddlv.ll b/llvm/test/CodeGen/AArch64/aarch64-neon-vector-insert-uaddlv.ll
index 83e35599c4e9e..91eda8d552397 100644
--- a/llvm/test/CodeGen/AArch64/aarch64-neon-vector-insert-uaddlv.ll
+++ b/llvm/test/CodeGen/AArch64/aarch64-neon-vector-insert-uaddlv.ll
@@ -229,10 +229,9 @@ define void @insert_vec_v3i16_uaddlv_from_v8i16(ptr %0) {
; CHECK: ; %bb.0: ; %entry
; CHECK-NEXT: movi.2d v0, #0000000000000000
; CHECK-NEXT: movi.2d v1, #0000000000000000
+; CHECK-NEXT: add x8, x0, #8
; CHECK-NEXT: uaddlv.8h s0, v0
; CHECK-NEXT: mov.h v1[0], v0[0]
-; CHECK-NEXT: mov.h v1[3], w8
-; CHECK-NEXT: add x8, x0, #8
; CHECK-NEXT: ushll.4s v1, v1, #0
; CHECK-NEXT: ucvtf.4s v1, v1
; CHECK-NEXT: st1.s { v1 }[2], [x8]
diff --git a/llvm/test/CodeGen/AArch64/arm64-build-vector.ll b/llvm/test/CodeGen/AArch64/arm64-build-vector.ll
index 82802c79c7085..c6fff3e3d3181 100644
--- a/llvm/test/CodeGen/AArch64/arm64-build-vector.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-build-vector.ll
@@ -57,8 +57,8 @@ define void @widen_f16_build_vector(ptr %addr) {
; CHECK-LABEL: widen_f16_build_vector:
; CHECK: // %bb.0:
; CHECK-NEXT: mov w8, #13294 // =0x33ee
-; CHECK-NEXT: movk w8, #13294, lsl #16
-; CHECK-NEXT: str w8, [x0]
+; CHECK-NEXT: dup v0.4h, w8
+; CHECK-NEXT: str s0, [x0]
; CHECK-NEXT: ret
store <2 x half> <half 0xH33EE, half 0xH33EE>, ptr %addr, align 2
ret void
diff --git a/llvm/test/CodeGen/AArch64/arm64-vector-insertion.ll b/llvm/test/CodeGen/AArch64/arm64-vector-insertion.ll
index ee3b3d18df879..ff28c7817d143 100644
--- a/llvm/test/CodeGen/AArch64/arm64-vector-insertion.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-vector-insertion.ll
@@ -84,7 +84,6 @@ define <16 x i8> @test_insert_v16i8_insert_2_undef_base_skip8(i32 %a0) {
; CHECK-NEXT: lsr w8, w0, #5
; CHECK-NEXT: dup.16b v0, w8
; CHECK-NEXT: mov.b v0[5], wzr
-; CHECK-NEXT: mov.b v0[8], w8
; CHECK-NEXT: mov.b v0[9], wzr
; CHECK-NEXT: ret
%a1 = lshr exact i32 %a0, 5
@@ -145,7 +144,6 @@ define <16 x i8> @test_insert_v16i8_insert_2_undef_base_different_values_skip8(i
; CHECK-NEXT: mov.b v0[2], w8
; CHECK-NEXT: mov.b v0[5], wzr
; CHECK-NEXT: mov.b v0[7], w8
-; CHECK-NEXT: mov.b v0[8], w8
; CHECK-NEXT: mov.b v0[9], wzr
; CHECK-NEXT: mov.b v0[12], w8
; CHECK-NEXT: mov.b v0[15], w8
diff --git a/llvm/test/CodeGen/AArch64/concat-vector-add-combine.ll b/llvm/test/CodeGen/AArch64/concat-vector-add-combine.ll
index 34899cb47dba3..5abf3cce9ea39 100644
--- a/llvm/test/CodeGen/AArch64/concat-vector-add-combine.ll
+++ b/llvm/test/CodeGen/AArch64/concat-vector-add-combine.ll
@@ -93,17 +93,14 @@ define i32 @combine_add_8xi32(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, i3
define i32 @combine_undef_add_8xi32(i32 %a, i32 %b, i32 %c, i32 %d) local_unnamed_addr #0 {
; CHECK-LABEL: combine_undef_add_8xi32:
; CHECK: // %bb.0:
-; CHECK-NEXT: fmov s1, w0
-; CHECK-NEXT: movi v0.2d, #0000000000000000
-; CHECK-NEXT: mov v1.s[1], w1
-; CHECK-NEXT: uhadd v0.4h, v0.4h, v0.4h
-; CHECK-NEXT: mov v1.s[2], w2
-; CHECK-NEXT: mov v1.s[3], w3
-; CHECK-NEXT: xtn v2.4h, v1.4s
-; CHECK-NEXT: shrn v1.4h, v1.4s, #16
-; CHECK-NEXT: uhadd v1.4h, v2.4h, v1.4h
-; CHECK-NEXT: mov v1.d[1], v0.d[0]
-; CHECK-NEXT: uaddlv s0, v1.8h
+; CHECK-NEXT: fmov s0, w0
+; CHECK-NEXT: mov v0.s[1], w1
+; CHECK-NEXT: mov v0.s[2], w2
+; CHECK-NEXT: mov v0.s[3], w3
+; CHECK-NEXT: uzp2 v1.8h, v0.8h, v0.8h
+; CHECK-NEXT: uzp1 v0.8h, v0.8h, v0.8h
+; CHECK-NEXT: uhadd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT: uaddlv s0, v0.8h
; CHECK-NEXT: fmov w0, s0
; CHECK-NEXT: ret
%a1 = insertelement <8 x i32> poison, i32 %a, i32 0
diff --git a/llvm/test/CodeGen/Thumb2/active_lane_mask.ll b/llvm/test/CodeGen/Thumb2/active_lane_mask.ll
index da13d07abf09b..bcd92f81911b2 100644
--- a/llvm/test/CodeGen/Thumb2/active_lane_mask.ll
+++ b/llvm/test/CodeGen/Thumb2/active_lane_mask.ll
@@ -306,10 +306,9 @@ define void @test_width2(ptr nocapture readnone %x, ptr nocapture %y, i8 zeroext
; CHECK-NEXT: ldrne.w r3, [r12]
; CHECK-NEXT: vmovne.32 q0[0], r3
; CHECK-NEXT: lsls r0, r0, #30
-; CHECK-NEXT: ittt mi
+; CHECK-NEXT: itt mi
; CHECK-NEXT: ldrmi.w r0, [r12, #4]
; CHECK-NEXT: vmovmi.32 q0[2], r0
-; CHECK-NEXT: vmovmi.32 q0[3], r0
; CHECK-NEXT: vmrs r3, p0
; CHECK-NEXT: and r0, r3, #1
; CHECK-NEXT: ubfx r3, r3, #8, #1
diff --git a/llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll b/llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll
index 53595bc5a5595..67910e40e7705 100644
--- a/llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-masked-ldst.ll
@@ -110,10 +110,9 @@ define void @foo_sext_v2i64_v2i32(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: ldrne r3, [r2]
; CHECK-LE-NEXT: vmovne.32 q1[0], r3
; CHECK-LE-NEXT: lsls r1, r1, #30
-; CHECK-LE-NEXT: ittt mi
+; CHECK-LE-NEXT: itt mi
; CHECK-LE-NEXT: ldrmi r1, [r2, #4]
; CHECK-LE-NEXT: vmovmi.32 q1[2], r1
-; CHECK-LE-NEXT: vmovmi.32 q1[3], r0
; CHECK-LE-NEXT: vmov r2, s6
; CHECK-LE-NEXT: movs r1, #0
; CHECK-LE-NEXT: vmov r3, s0
@@ -220,9 +219,9 @@ define void @foo_sext_v2i64_v2i32_unaligned(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: sub sp, #4
; CHECK-LE-NEXT: ldrd r12, lr, [r1]
; CHECK-LE-NEXT: movs r1, #0
-; CHECK-LE-NEXT: @ implicit-def: $q1
+; CHECK-LE-NEXT: @ implicit-def: $q0
; CHECK-LE-NEXT: rsbs.w r3, r12, #0
-; CHECK-LE-NEXT: vmov q0[2], q0[0], r12, lr
+; CHECK-LE-NEXT: vmov q1[2], q1[0], r12, lr
; CHECK-LE-NEXT: sbcs.w r3, r1, r12, asr #31
; CHECK-LE-NEXT: csetm r3, lt
; CHECK-LE-NEXT: rsbs.w r4, lr, #0
@@ -233,24 +232,23 @@ define void @foo_sext_v2i64_v2i32_unaligned(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: lsls r3, r1, #31
; CHECK-LE-NEXT: itt ne
; CHECK-LE-NEXT: ldrne r3, [r2]
-; CHECK-LE-NEXT: vmovne.32 q1[0], r3
+; CHECK-LE-NEXT: vmovne.32 q0[0], r3
; CHECK-LE-NEXT: lsls r1, r1, #30
-; CHECK-LE-NEXT: ittt mi
+; CHECK-LE-NEXT: itt mi
; CHECK-LE-NEXT: ldrmi r1, [r2, #4]
-; CHECK-LE-NEXT: vmovmi.32 q1[2], r1
-; CHECK-LE-NEXT: vmovmi.32 q1[3], r0
-; CHECK-LE-NEXT: vmov r2, s6
+; CHECK-LE-NEXT: vmovmi.32 q0[2], r1
+; CHECK-LE-NEXT: vmov r2, s2
; CHECK-LE-NEXT: movs r1, #0
-; CHECK-LE-NEXT: vmov r3, s0
-; CHECK-LE-NEXT: vmov r4, s4
-; CHECK-LE-NEXT: vmov q1[2], q1[0], r4, r2
+; CHECK-LE-NEXT: vmov r3, s4
+; CHECK-LE-NEXT: vmov r4, s0
+; CHECK-LE-NEXT: vmov q0[2], q0[0], r4, r2
; CHECK-LE-NEXT: rsbs r5, r3, #0
; CHECK-LE-NEXT: asr.w r12, r2, #31
; CHECK-LE-NEXT: sbcs.w r2, r1, r3, asr #31
-; CHECK-LE-NEXT: vmov r3, s2
+; CHECK-LE-NEXT: vmov r3, s6
; CHECK-LE-NEXT: csetm r2, lt
; CHECK-LE-NEXT: asr.w lr, r4, #31
-; CHECK-LE-NEXT: vmov q1[3], q1[1], lr, r12
+; CHECK-LE-NEXT: vmov q0[3], q0[1], lr, r12
; CHECK-LE-NEXT: rsbs r5, r3, #0
; CHECK-LE-NEXT: sbcs.w r3, r1, r3, asr #31
; CHECK-LE-NEXT: bfi r1, r2, #0, #1
@@ -258,11 +256,11 @@ define void @foo_sext_v2i64_v2i32_unaligned(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: bfi r1, r2, #1, #1
; CHECK-LE-NEXT: lsls r2, r1, #31
; CHECK-LE-NEXT: itt ne
-; CHECK-LE-NEXT: vmovne r2, r3, d2
+; CHECK-LE-NEXT: vmovne r2, r3, d0
; CHECK-LE-NEXT: strdne r2, r3, [r0]
; CHECK-LE-NEXT: lsls r1, r1, #30
; CHECK-LE-NEXT: itt mi
-; CHECK-LE-NEXT: vmovmi r1, r2, d3
+; CHECK-LE-NEXT: vmovmi r1, r2, d1
; CHECK-LE-NEXT: strdmi r1, r2, [r0, #8]
; CHECK-LE-NEXT: add sp, #4
; CHECK-LE-NEXT: pop {r4, r5, r7, pc}
@@ -365,10 +363,9 @@ define void @foo_zext_v2i64_v2i32(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: ldrne r3, [r2]
; CHECK-LE-NEXT: vmovne.32 q0[0], r3
; CHECK-LE-NEXT: lsls r1, r1, #30
-; CHECK-LE-NEXT: ittt mi
+; CHECK-LE-NEXT: itt mi
; CHECK-LE-NEXT: ldrmi r1, [r2, #4]
; CHECK-LE-NEXT: vmovmi.32 q0[2], r1
-; CHECK-LE-NEXT: vmovmi.32 q0[3], r0
; CHECK-LE-NEXT: vmov r2, s4
; CHECK-LE-NEXT: movs r1, #0
; CHECK-LE-NEXT: vand q0, q0, q2
@@ -481,10 +478,9 @@ define void @foo_zext_v2i64_v2i32_unaligned(ptr %dest, ptr %mask, ptr %src) {
; CHECK-LE-NEXT: ldrne r3, [r2]
; CHECK-LE-NEXT: vmovne.32 q0[0], r3
; CHECK-LE-NEXT: lsls r1, r1, #30
-; CHECK-LE-NEXT: ittt mi
+; CHECK-LE-NEXT: itt mi
; CHECK-LE-NEXT: ldrmi r1, [r2, #4]
; CHECK-LE-NEXT: vmovmi.32 q0[2], r1
-; CHECK-LE-NEXT: vmovmi.32 q0[3], r0
; CHECK-LE-NEXT: vmov r2, s4
; CHECK-LE-NEXT: movs r1, #0
; CHECK-LE-NEXT: vand q0, q0, q2
diff --git a/llvm/test/CodeGen/Thumb2/mve-vld3.ll b/llvm/test/CodeGen/Thumb2/mve-vld3.ll
index b6c8056891f83..38e42c137e3a9 100644
--- a/llvm/test/CodeGen/Thumb2/mve-vld3.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-vld3.ll
@@ -56,7 +56,7 @@ define void @vld3_v4i32(ptr %src, ptr %dst) {
; CHECK-LV-NEXT: vstrw.32 q0, [r1]
; CHECK-LV-NEXT: vpop {d8, d9}
; CHECK-LV-NEXT: bx lr
-
+;
; CHECK-LIS-LABEL: vld3_v4i32:
; CHECK-LIS: @ %bb.0: @ %entry
; CHECK-LIS-NEXT: .vsave {d8, d9}
@@ -80,6 +80,7 @@ define void @vld3_v4i32(ptr %src, ptr %dst) {
; CHECK-LIS-NEXT: vstrw.32 q0, [r1]
; CHECK-LIS-NEXT: vpop {d8, d9}
; CHECK-LIS-NEXT: bx lr
+
entry:
%l1 = load <12 x i32>, ptr %src, align 4
%s1 = shufflevector <12 x i32> %l1, <12 x i32> undef, <4 x i32> <i32 0, i32 3, i32 6, i32 9>
@@ -132,7 +133,7 @@ define void @vld3_v8i32(ptr %src, ptr %dst) {
; CHECK-LV-NEXT: vstrw.32 q1, [r1]
; CHECK-LV-NEXT: vpop {d8, d9, d10, d11}
; CHECK-LV-NEXT: bx lr
-
+;
; CHECK-LIS-LABEL: vld3_v8i32:
; CHECK-LIS: @ %bb.0: @ %entry
; CHECK-LIS-NEXT: .vsave {d8, d9, d10, d11}
@@ -173,6 +174,7 @@ define void @vld3_v8i32(ptr %src, ptr %dst) {
; CHECK-LIS-NEXT: vstrw.32 q1, [r1]
; CHECK-LIS-NEXT: vpop {d8, d9, d10, d11}
; CHECK-LIS-NEXT: bx lr
+
entry:
%l1 = load <24 x i32>, ptr %src, align 4
%s1 = shufflevector <24 x i32> %l1, <24 x i32> undef, <8 x i32> <i32 0, i32 3, i32 6, i32 9, i32 12, i32 15, i32 18, i32 21>
@@ -259,7 +261,7 @@ define void @vld3_v16i32(ptr %src, ptr %dst) {
; CHECK-LV-NEXT: vstrw.32 q3, [r1, #32]
; CHECK-LV-NEXT: vpop {d8, d9, d10, d11, d12, d13, d14, d15}
; CHECK-LV-NEXT: bx lr
-
+;
; CHECK-LIS-LABEL: vld3_v16i32:
; CHECK-LIS: @ %bb.0: @ %entry
; CHECK-LIS-NEXT: .vsave {d8, d9, d10, d11, d12, d13, d14, d15}
@@ -334,6 +336,7 @@ define void @vld3_v16i32(ptr %src, ptr %dst) {
; CHECK-LIS-NEXT: vstrw.32 q3, [r1, #32]
; CHECK-LIS-NEXT: vpop {d8, d9, d10, d11, d12, d13, d14, d15}
; CHECK-LIS-NEXT: bx lr
+
entry:
%l1 = load <48 x i32>, ptr %src, align 4
%s1 = shufflevector <48 x i32> %l1, <48 x i32> undef, <16 x i32> <i32 0, i32 3, i32 6, i32 9, i32 12, i32 15, i32 18, i32 21, i32 24, i32 27, i32 30, i32 33, i32 36, i32 39, i32 42, i32 45>
@@ -660,8 +663,8 @@ define void @vld3_v2i8(ptr %src, ptr %dst) {
; CHECK: @ %bb.0: @ %entry
; CHECK-NEXT: .pad #8
; CHECK-NEXT: sub sp, #8
-; CHECK-NEXT: ldrd r0, r2, [r0]
-; CHECK-NEXT: strd r0, r2, [sp]
+; CHECK-NEXT: ldrd r2, r0, [r0]
+; CHECK-NEXT: strd r2, r0, [sp]
; CHECK-NEXT: mov r0, sp
; CHECK-NEXT: vldrb.u16 q0, [r0]
; CHECK-NEXT: vmov.u16 r0, q0[4]
@@ -961,7 +964,7 @@ define void @vld3_v2i64(ptr %src, ptr %dst) {
; CHECK-LV-NEXT: vmov q0[3], q0[1], r7, r2
; CHECK-LV-NEXT: vstrw.32 q0, [r1]
; CHECK-LV-NEXT: pop.w {r4, r5, r6, r7, r8, pc}
-
+;
; CHECK-LIS-LABEL: vld3_v2i64:
; CHECK-LIS: @ %bb.0: @ %entry
; CHECK-LIS-NEXT: .save {r4, r5, r6, r7, r8, lr}
@@ -991,6 +994,7 @@ define void @vld3_v2i64(ptr %src, ptr %dst) {
; CHECK-LIS-NEXT: vmov q0[3], q0[1], r7, r2
; CHECK-LIS-NEXT: vstrw.32 q0, [r1]
; CHECK-LIS-NEXT: pop.w {r4, r5, r6, r7, r8, pc}
+
entry:
%l1 = load <6 x i64>, ptr %src, align 4
%s1 = shufflevector <6 x i64> %l1, <6 x i64> undef, <2 x i32> <i32 0, i32 3>
@@ -1147,7 +1151,7 @@ define void @vld3_v2f32(ptr %src, ptr %dst) {
; CHECK-LV-NEXT: vadd.f32 q0, q0, q1
; CHECK-LV-NEXT: vstmia r1, {s0, s1}
; CHECK-LV-NEXT: bx lr
-
+;
; CHECK-LIS-LABEL: vld3_v2f32:
; CHECK-LIS: @ %bb.0: @ %entry
; CHECK-LIS-NEXT: vldrw.u32 q2, [r0]
@@ -1161,6 +1165,7 @@ define void @vld3_v2f32(ptr %src, ptr %dst) {
; CHECK-LIS-NEXT: vadd.f32 q0, q1, q0
; CHECK-LIS-NEXT: vstmia r1, {s0, s1}
; CHECK-LIS-NEXT: bx lr
+
entry:
%l1 = load <6 x float>, ptr %src, align 4
%s1 = shufflevector <6 x float> %l1, <6 x float> undef, <2 x i32> <i32 0, i32 3>
@@ -1173,53 +1178,30 @@ entry:
}
define void @vld3_v4f32(ptr %src, ptr %dst) {
-; CHECK-LV-LABEL: vld3_v4f32:
-; CHECK-LV: @ %bb.0: @ %entry
-; CHECK-LV-NEXT: .vsave {d8, d9}
-; CHECK-LV-NEXT: vpush {d8, d9}
-; CHECK-LV-NEXT: vldrw.u32 q0, [r0, #16]
-; CHECK-LV-NEXT: vldrw.u32 q1, [r0]
-; CHECK-LV-NEXT: vldrw.u32 q4, [r0, #32]
-; CHECK-LV-NEXT: vmov.f32 s10, s2
-; CHECK-LV-NEXT: vmov.f32 s13, s0
-; CHECK-LV-NEXT: vmov.f32 s14, s3
-; CHECK-LV-NEXT: vmov.f32 s8, s4
-; CHECK-LV-NEXT: vmov.f32 s9, s7
-; CHECK-LV-NEXT: vmov.f32 s12, s5
-; CHECK-LV-NEXT: vmov.f32 s15, s18
-; CHECK-LV-NEXT: vmov.f32 s11, s17
-; CHECK-LV-NEXT: vadd.f32 q2, q2, q3
-; CHECK-LV-NEXT: vmov.f32 s0, s6
-; CHECK-LV-NEXT: vmov.f32 s2, s16
-; CHECK-LV-NEXT: vmov.f32 s3, s19
-; CHECK-LV-NEXT: vadd.f32 q0, q2, q0
-; CHECK-LV-NEXT: vstrw.32 q0, [r1]
-; CHECK-LV-NEXT: vpop {d8, d9}
-; CHECK-LV-NEXT: bx lr
+; CHECK-LABEL: vld3_v4f32:
+; CHECK: @ %bb.0: @ %entry
+; CHECK-NEXT: .vsave {d8, d9}
+; CHECK-NEXT: vpush {d8, d9}
+; CHECK-NEXT: vldrw.u32 q0, [r0, #16]
+; CHECK-NEXT: vldrw.u32 q1, [r0]
+; CHECK-NEXT: vldrw.u32 q4, [r0, #32]
+; CHECK-NEXT: vmov.f32 s10, s2
+; CHECK-NEXT: vmov.f32 s13, s0
+; CHECK-NEXT: vmov.f32 s14, s3
+; CHECK-NEXT: vmov.f32 s8, s4
+; CHECK-NEXT: vmov.f32 s9, s7
+; CHECK-NEXT: vmov.f32 s12, s5
+; CHECK-NEXT: vmov.f32 s15, s18
+; CHECK-NEXT: vmov.f32 s11, s17
+; CHECK-NEXT: vadd.f32 q2, q2, q3
+; CHECK-NEXT: vmov.f32 s0, s6
+; CHECK-NEXT: vmov.f32 s2, s16
+; CHECK-NEXT: vmov.f32 s3, s19
+; CHECK-NEXT: vadd.f32 q0, q2, q0
+; CHECK-NEXT: vstrw.32 q0, [r1]
+; CHECK-NEXT: vpop {d8, d9}
+; CHECK-NEXT: bx lr
-; CHECK-LIS-LABEL: vld3_v4f32:
-; CHECK-LIS: @ %bb.0: @ %entry
-; CHECK-LIS-NEXT: .vsave {d8, d9}
-; CHECK-LIS-NEXT: vpush {d8, d9}
-; CHECK-LIS-NEXT: vldrw.u32 q0, [r0, #16]
-; CHECK-LIS-NEXT: vldrw.u32 q1, [r0]
-; CHECK-LIS-NEXT: vldrw.u32 q4, [r0, #32]
-; CHECK-LIS-NEXT: vmov.f32 s10, s2
-; CHECK-LIS-NEXT: vmov.f32 s13, s0
-; CHECK-LIS-NEXT: vmov.f32 s14, s3
-; CHECK-LIS-NEXT: vmov.f32 s8, s4
-; CHECK-LIS-NEXT: vmov.f32 s9, s7
-; CHECK-LIS-NEXT: vmov.f32 s12, s5
-; CHECK-LIS-NEXT: vmov.f32 s15, s18
-; CHECK-LIS-NEXT: vmov.f32 s11, s17
-; CHECK-LIS-NEXT: vadd.f32 q2, q2, q3
-; CHECK-LIS-NEXT: vmov.f32 s0, s6
-; CHECK-LIS-NEXT: vmov.f32 s2, s16
-; CHECK-LIS-NEXT: vmov.f32 s3, s19
-; CHECK-LIS-NEXT: vadd.f32 q0, q2, q0
-; CHECK-LIS-NEXT: vstrw.32 q0, [r1]
-; CHECK-LIS-NEXT: vpop {d8, d9}
-; CHECK-LIS-NEXT: bx lr
entry:
%l1 = load <12 x float>, ptr %src, align 4
%s1 = shufflevector <12 x float> %l1, <12 x float> undef, <4 x i32> <i32 0, i32 3, i32 6, i32 9>
@@ -1232,87 +1214,47 @@ entry:
}
define void @vld3_v8f32(ptr %src, ptr %dst) {
-; CHECK-LV-LABEL: vld3_v8f32:
-; CHECK-LV: @ %bb.0: @ %entry
-; CHECK-LV-NEXT: .vsave {d8, d9, d10, d11}
-; CHECK-LV-NEXT: vpush {d8, d9, d10, d11}
-; CHECK-LV-NEXT: vldrw.u32 q0, [r0, #64]
-; CHECK-LV-NEXT: vldrw.u32 q1, [r0, #48]
-; CHECK-LV-NEXT: vldrw.u32 q4, [r0, #80]
-; CHECK-LV-NEXT: vmov.f32 s10, s2
-; CHECK-LV-NEXT: vmov.f32 s13, s0
-; CHECK-LV-NEXT: vmov.f32 s14, s3
-; CHECK-LV-NEXT: vmov.f32 s8, s4
-; CHECK-LV-NEXT: vmov.f32 s9, s7
-; CHECK-LV-NEXT: vmov.f32 s12, s5
-; CHECK-LV-NEXT: vmov.f32 s15, s18
-; CHECK-LV-NEXT: vmov.f32 s11, s17
-; CHECK-LV-NEXT: vadd.f32 q2, q2, q3
-; CHECK-LV-NEXT: vmov.f32 s0, s6
-; CHECK-LV-NEXT: vmov.f32 s2, s16
-; CHECK-LV-NEXT: vldrw.u32 q1, [r0, #16]
-; CHECK-LV-NEXT: vmov.f32 s3, s19
-; CHECK-LV-NEXT: vldrw.u32 q3, [r0, #32]
-; CHECK-LV-NEXT: vadd.f32 q0, q2, q0
-; CHECK-LV-NEXT: vldrw.u32 q2, [r0]
-; CHECK-LV-NEXT: vmov.f32 s17, s4
-; CHECK-LV-NEXT: vstrw.32 q0, [r1, #16]
-; CHECK-LV-NEXT: vmov.f32 s18, s7
-; CHECK-LV-NEXT: vmov.f32 s22, s6
-; CHECK-LV-NEXT: vmov.f32 s16, s9
-; CHECK-LV-NEXT: vmov.f32 s19, s14
-; CHECK-LV-NEXT: vmov.f32 s20, s8
-; CHECK-LV-NEXT: vmov.f32 s21, s11
-; CHECK-LV-NEXT: vmov.f32 s23, s13
-; CHECK-LV-NEXT: vadd.f32 q4, q5, q4
-; CHECK-LV-NEXT: vmov.f32 s4, s10
-; CHECK-LV-NEXT: vmov.f32 s6, s12
-; CHECK-LV-NEXT: vmov.f32 s7, s15
-; CHECK-LV-NEXT: vadd.f32 q1, q4, q1
-; CHECK-LV-NEXT: vstrw.32 q1, [r1]
-; CHECK-LV-NEXT: vpop {d8, d9, d10, d11}
-; CHECK-LV-NEXT: bx lr
+; CHECK-LABEL: vld3_v8f32:
+; CHECK: @ %bb.0: @ %entry
+; CHECK-NEXT: .vsave {d8, d9, d10, d11}
+; CHECK-NEXT: vpush {d8, d9, d10, d11}
+; CHECK-NEXT: vldrw.u32 q0, [r0, #64]
+; CHECK-NEXT: vldrw.u32 q1, [r0, #48]
+; CHECK-NEXT: vldrw.u32 q4, [r0, #80]
+; CHECK-NEXT: vmov.f32 s10, s2
+; CHECK-NEXT: vmov.f32 s13, s0
+; CHECK-NEXT: vmov.f32 s14, s3
+; CHECK-NEXT: vmov.f32 s8, s4
+; CHECK-NEXT: vmov.f32 s9, s7
+; CHECK-NEXT: vmov.f32 s12, s5
+; CHECK-NEXT: vmov.f32 s15, s18
+; CHECK-NEXT: vmov.f32 s11, s17
+; CHECK-NEXT: vadd.f32 q2, q2, q3
+; CHECK-NEXT: vmov.f32 s0, s6
+; CHECK-NEXT: vmov.f32 s2, s16
+; CHECK-NEXT: vldrw.u32 q1, [r0, #16]
+; CHECK-NEXT: vmov.f32 s3, s19
+; CHECK-NEXT: vldrw.u32 q3, [r0, #32]
+; CHECK-NEXT: vadd.f32 q0, q2, q0
+; CHECK-NEXT: vldrw.u32 q2, [r0]
+; CHECK-NEXT: vmov.f32 s17, s4
+; CHECK-NEXT: vstrw.32 q0, [r1, #16]
+; CHECK-NEXT: vmov.f32 s18, s7
+; CHECK-NEXT: vmov.f32 s22, s6
+; CHECK-NEXT: vmov.f32 s16, s9
+; CHECK-NEXT: vmov.f32 s19, s14
+; CHECK-NEXT: vmov.f32 s20, s8
+; CHECK-NEXT: vmov.f32 s21, s11
+; CHECK-NEXT: vmov.f32 s23, s13
+; CHECK-NEXT: vadd.f32 q4, q5, q4
+; CHECK-NEXT: vmov.f32 s4, s10
+; CHECK-NEXT: vmov.f32 s6, s12
+; CHECK-NEXT: vmov.f32 s7, s15
+; CHECK-NEXT: vadd.f32 q1, q4, q1
+; CHECK-NEXT: vstrw.32 q1, [r1]
+; CHECK-NEXT: vpop {d8, d9, d10, d11}
+; CHECK-NEXT: bx lr
-; CHECK-LIS-LABEL: vld3_v8f32:
-; CHECK-LIS: @ %bb.0: @ %entry
-; CHECK-LIS-NEXT: .vsave {d8, d9, d10, d11}
-; CHECK-LIS-NEXT: vpush {d8, d9, d10, d11}
-; CHECK-LIS-NEXT: vldrw.u32 q0, [r0, #64]
-; CHECK-LIS-NEXT: vldrw.u32 q1, [r0, #48]
-; CHECK-LIS-NEXT: vldrw.u32 q4, [r0, #80]
-; CHECK-LIS-NEXT: vmov.f32 s10, s2
-; CHECK-LIS-NEXT: vmov.f32 s13, s0
-; CHECK-LIS-NEXT: vmov.f32 s14, s3
-; CHECK-LIS-NEXT: vmov.f32 s8, s4
-; CHECK-LIS-NEXT: vmov.f32 s9, s7
-; CHECK-LIS-NEXT: vmov.f32 s12, s5
-; CHECK-LIS-NEXT: vmov.f32 s15, s18
-; CHECK-LIS-NEXT: vmov.f32 s11, s17
-; CHECK-LIS-NEXT: vadd.f32 q2, q2, q3
-; CHECK-LIS-NEXT: vmov.f32 s0, s6
-; CHECK-LIS-NEXT: vmov.f32 s2, s16
-; CHECK-LIS-NEXT: vldrw.u32 q1, [r0, #16]
-; CHECK-LIS-NEXT: vmov.f32 s3, s19
-; CHECK-LIS-NEXT: vldrw.u32 q3, [r0, #32]
-; CHECK-LIS-NEXT: vadd.f32 q0, q2, q0
-; CHECK-LIS-NEXT: vldrw.u32 q2, [r0]
-; CHECK-LIS-NEXT: vmov.f32 s17, s4
-; CHECK-LIS-NEXT: vstrw.32 q0, [r1, #16]
-; CHECK-LIS-NEXT: vmov.f32 s18, s7
-; CHECK-LIS-NEXT: vmov.f32 s22, s6
-; CHECK-LIS-NEXT: vmov.f32 s16, s9
-; CHECK-LIS-NEXT: vmov.f32 s19, s14
-; CHECK-LIS-NEXT: vmov.f32 s20, s8
-; CHECK-LIS-NEXT: vmov.f32 s21, s11
-; CHECK-LIS-NEXT: vmov.f32 s23, s13
-; CHECK-LIS-NEXT: vadd.f32 q4, q5, q4
-; CHECK-LIS-NEXT: vmov.f32 s4, s10
-; CHECK-LIS-NEXT: vmov.f32 s6, s12
-; CHECK-LIS-NEXT: vmov.f32 s7, s15
-; CHECK-LIS-NEXT: vadd.f32 q1, q4, q1
-; CHECK-LIS-NEXT: vstrw.32 q1, [r1]
-; CHECK-LIS-NEXT: vpop {d8, d9, d10, d11}
-; CHECK-LIS-NEXT: bx lr
entry:
%l1 = load <24 x float>, ptr %src, align 4
%s1 = shufflevector <24 x float> %l1, <24 x float> undef, <8 x i32> <i32 0, i32 3, i32 6, i32 9, i32 12, i32 15, i32 18, i32 21>
@@ -1325,155 +1267,81 @@ entry:
}
define void @vld3_v16f32(ptr %src, ptr %dst) {
-; CHECK-LV-LABEL: vld3_v16f32:
-; CHECK-LV: @ %bb.0: @ %entry
-; CHECK-LV-NEXT: .vsave {d8, d9, d10, d11, d12, d13, d14, d15}
-; CHECK-LV-NEXT: vpush {d8, d9, d10, d11, d12, d13, d14, d15}
-; CHECK-LV-NEXT: vldrw.u32 q0, [r0, #64]
-; CHECK-LV-NEXT: vldrw.u32 q1, [r0, #48]
-; CHECK-LV-NEXT: vldrw.u32 q4, [r0, #80]
-; CHECK-LV-NEXT: vldrw.u32 q6, [r0, #176]
-; CHECK-LV-NEXT: vmov.f32 s10, s2
-; CHECK-LV-NEXT: vmov.f32 s13, s0
-; CHECK-LV-NEXT: vmov.f32 s14, s3
-; CHECK-LV-NEXT: vmov.f32 s8, s4
-; CHECK-LV-NEXT: vmov.f32 s9, s7
-; CHECK-LV-NEXT: vmov.f32 s12, s5
-; CHECK-LV-NEXT: vmov.f32 s15, s18
-; CHECK-LV-NEXT: vmov.f32 s11, s17
-; CHECK-LV-NEXT: vadd.f32 q2, q2, q3
-; CHECK-LV-NEXT: vmov.f32 s0, s6
-; CHECK-LV-NEXT: vmov.f32 s2, s16
-; CHECK-LV-NEXT: vldrw.u32 q1, [r0, #16]
-; CHECK-LV-NEXT: vmov.f32 s3, s19
-; CHECK-LV-NEXT: vldrw.u32 q3, [r0, #32]
-; CHECK-LV-NEXT: vadd.f32 q0, q2, q0
-; CHECK-LV-NEXT: vldrw.u32 q2, [r0]
-; CHECK-LV-NEXT: vmov.f32 s17, s4
-; CHECK-LV-NEXT: vmov.f32 s18, s7
-; CHECK-LV-NEXT: vmov.f32 s22, s6
-; CHECK-LV-NEXT: vmov.f32 s16, s9
-; CHECK-LV-NEXT: vmov.f32 s19, s14
-; CHECK-LV-NEXT: vmov.f32 s20, s8
-; CHECK-LV-NEXT: vmov.f32 s21, s11
-; CHECK-LV-NEXT: vmov.f32 s23, s13
-; CHECK-LV-NEXT: vmov.f32 s4, s10
-; CHECK-LV-NEXT: vldrw.u32 q2, [r0, #160]
-; CHECK-LV-NEXT: vmov.f32 s6, s12
-; CHECK-LV-NEXT: vadd.f32 q4, q5, q4
-; CHECK-LV-NEXT: vmov.f32 s7, s15
-; CHECK-LV-NEXT: vldrw.u32 q3, [r0, #144]
-; CHECK-LV-NEXT: vadd.f32 q1, q4, q1
-; CHECK-LV-NEXT: vmov.f32 s18, s10
-; CHECK-LV-NEXT: vmov.f32 s21, s8
-; CHECK-LV-NEXT: vmov.f32 s22, s11
-; CHECK-LV-NEXT: vmov.f32 s16, s12
-; CHECK-LV-NEXT: vmov.f32 s17, s15
-; CHECK-LV-NEXT: vmov.f32 s20, s13
-; CHECK-LV-NEXT: vmov.f32 s23, s26
-; CHECK-LV-NEXT: vmov.f32 s19, s25
-; CHECK-LV-NEXT: vadd.f32 q4, q4, q5
-; CHECK-LV-NEXT: vmov.f32 s8, s14
-; CHECK-LV-NEXT: vmov.f32 s10, s24
-; CHECK-LV-NEXT: vldrw.u32 q3, [r0, #112]
-; CHECK-LV-NEXT: vmov.f32 s11, s27
-; CHECK-LV-NEXT: vldrw.u32 q5, [r0, #128]
-; CHECK-LV-NEXT: vadd.f32 q2, q4, q2
-; CHECK-LV-NEXT: vldrw.u32 q4, [r0, #96]
-; CHECK-LV-NEXT: vmov.f32 s25, s12
-; CHECK-LV-NEXT: vstrw.32 q2, [r1, #48]
-; CHECK-LV-NEXT: vmov.f32 s26, s15
-; CHECK-LV-NEXT: vstrw.32 q0, [r1, #16]
-; CHECK-LV-NEXT: vmov.f32 s30, s14
-; CHECK-LV-NEXT: vstrw.32 q1, [r1]
-; CHECK-LV-NEXT: vmov.f32 s24, s17
-; CHECK-LV-NEXT: vmov.f32 s27, s22
-; CHECK-LV-NEXT: vmov.f32 s28, s16
-; CHECK-LV-NEXT: vmov.f32 s29, s19
-; CHECK-LV-NEXT: vmov.f32 s31, s21
-; CHECK-LV-NEXT: vadd.f32 q6, q7, q6
-; CHECK-LV-NEXT: vmov.f32 s12, s18
-; CHECK-LV-NEXT: vmov.f32 s14, s20
-; CHECK-LV-NEXT: vmov.f32 s15, s23
-; CHECK-LV-NEXT: vadd.f32 q3, q6, q3
-; CHECK-LV-NEXT: vstrw.32 q3, [r1, #32]
-; CHECK-LV-NEXT: vpop {d8, d9, d10, d11, d12, d13, d14, d15}
-; CHECK-LV-NEXT: bx lr
+; CHECK-LABEL: vld3_v16f32:
+; CHECK: @ %bb.0: @ %entry
+; CHECK-NEXT: .vsave {d8, d9, d10, d11, d12, d13, d14, d15}
+; CHECK-NEXT: vpush {d8, d9, d10, d11, d12, d13, d14, d15}
+; CHECK-NEXT: vldrw.u32 q0, [r0, #64]
+; CHECK-NEXT: vldrw.u32 q1, [r0, #48]
+; CHECK-NEXT: vldrw.u32 q4, [r0, #80]
+; CHECK-NEXT: vldrw.u32 q6, [r0, #176]
+; CHECK-NEXT: vmov.f32 s10, s2
+; CHECK-NEXT: vmov.f32 s13, s0
+; CHECK-NEXT: vmov.f32 s14, s3
+; CHECK-NEXT: vmov.f32 s8, s4
+; CHECK-NEXT: vmov.f32 s9, s7
+; CHECK-NEXT: vmov.f32 s12, s5
+; CHECK-NEXT: vmov.f32 s15, s18
+; CHECK-NEXT: vmov.f32 s11, s17
+; CHECK-NEXT: vadd.f32 q2, q2, q3
+; CHECK-NEXT: vmov.f32 s0, s6
+; CHECK-NEXT: vmov.f32 s2, s16
+; CHECK-NEXT: vldrw.u32 q1, [r0, #16]
+; CHECK-NEXT: vmov.f32 s3, s19
+; CHECK-NEXT: vldrw.u32 q3, [r0, #32]
+; CHECK-NEXT: vadd.f32 q0, q2, q0
+; CHECK-NEXT: vldrw.u32 q2, [r0]
+; CHECK-NEXT: vmov.f32 s17, s4
+; CHECK-NEXT: vmov.f32 s18, s7
+; CHECK-NEXT: vmov.f32 s22, s6
+; CHECK-NEXT: vmov.f32 s16, s9
+; CHECK-NEXT: vmov.f32 s19, s14
+; CHECK-NEXT: vmov.f32 s20, s8
+; CHECK-NEXT: vmov.f32 s21, s11
+; CHECK-NEXT: vmov.f32 s23, s13
+; CHECK-NEXT: vmov.f32 s4, s10
+; CHECK-NEXT: vldrw.u32 q2, [r0, #160]
+; CHECK-NEXT: vmov.f32 s6, s12
+; CHECK-NEXT: vadd.f32 q4, q5, q4
+; CHECK-NEXT: vmov.f32 s7, s15
+; CHECK-NEXT: vldrw.u32 q3, [r0, #144]
+; CHECK-NEXT: vadd.f32 q1, q4, q1
+; CHECK-NEXT: vmov.f32 s18, s10
+; CHECK-NEXT: vmov.f32 s21, s8
+; CHECK-NEXT: vmov.f32 s22, s11
+; CHECK-NEXT: vmov.f32 s16, s12
+; CHECK-NEXT: vmov.f32 s17, s15
+; CHECK-NEXT: vmov.f32 s20, s13
+; CHECK-NEXT: vmov.f32 s23, s26
+; CHECK-NEXT: vmov.f32 s19, s25
+; CHECK-NEXT: vadd.f32 q4, q4, q5
+; CHECK-NEXT: vmov.f32 s8, s14
+; CHECK-NEXT: vmov.f32 s10, s24
+; CHECK-NEXT: vldrw.u32 q3, [r0, #112]
+; CHECK-NEXT: vmov.f32 s11, s27
+; CHECK-NEXT: vldrw.u32 q5, [r0, #128]
+; CHECK-NEXT: vadd.f32 q2, q4, q2
+; CHECK-NEXT: vldrw.u32 q4, [r0, #96]
+; CHECK-NEXT: vmov.f32 s25, s12
+; CHECK-NEXT: vstrw.32 q2, [r1, #48]
+; CHECK-NEXT: vmov.f32 s26, s15
+; CHECK-NEXT: vstrw.32 q0, [r1, #16]
+; CHECK-NEXT: vmov.f32 s30, s14
+; CHECK-NEXT: vstrw.32 q1, [r1]
+; CHECK-NEXT: vmov.f32 s24, s17
+; CHECK-NEXT: vmov.f32 s27, s22
+; CHECK-NEXT: vmov.f32 s28, s16
+; CHECK-NEXT: vmov.f32 s29, s19
+; CHECK-NEXT: vmov.f32 s31, s21
+; CHECK-NEXT: vadd.f32 q6, q7, q6
+; CHECK-NEXT: vmov.f32 s12, s18
+; CHECK-NEXT: vmov.f32 s14, s20
+; CHECK-NEXT: vmov.f32 s15, s23
+; CHECK-NEXT: vadd.f32 q3, q6, q3
+; CHECK-NEXT: vstrw.32 q3, [r1, #32]
+; CHECK-NEXT: vpop {d8, d9, d10, d11, d12, d13, d14, d15}
+; CHECK-NEXT: bx lr
-; CHECK-LIS-LABEL: vld3_v16f32:
-; CHECK-LIS: @ %bb.0: @ %entry
-; CHECK-LIS-NEXT: .vsave {d8, d9, d10, d11, d12, d13, d14, d15}
-; CHECK-LIS-NEXT: vpush {d8, d9, d10, d11, d12, d13, d14, d15}
-; CHECK-LIS-NEXT: vldrw.u32 q0, [r0, #64]
-; CHECK-LIS-NEXT: vldrw.u32 q1, [r0, #48]
-; CHECK-LIS-NEXT: vldrw.u32 q4, [r0, #80]
-; CHECK-LIS-NEXT: vldrw.u32 q6, [r0, #176]
-; CHECK-LIS-NEXT: vmov.f32 s10, s2
-; CHECK-LIS-NEXT: vmov.f32 s13, s0
-; CHECK-LIS-NEXT: vmov.f32 s14, s3
-; CHECK-LIS-NEXT: vmov.f32 s8, s4
-; CHECK-LIS-NEXT: vmov.f32 s9, s7
-; CHECK-LIS-NEXT: vmov.f32 s12, s5
-; CHECK-LIS-NEXT: vmov.f32 s15, s18
-; CHECK-LIS-NEXT: vmov.f32 s11, s17
-; CHECK-LIS-NEXT: vadd.f32 q2, q2, q3
-; CHECK-LIS-NEXT: vmov.f32 s0, s6
-; CHECK-LIS-NEXT: vmov.f32 s2, s16
-; CHECK-LIS-NEXT: vldrw.u32 q1, [r0, #16]
-; CHECK-LIS-NEXT: vmov.f32 s3, s19
-; CHECK-LIS-NEXT: vldrw.u32 q3, [r0, #32]
-; CHECK-LIS-NEXT: vadd.f32 q0, q2, q0
-; CHECK-LIS-NEXT: vldrw.u32 q2, [r0]
-; CHECK-LIS-NEXT: vmov.f32 s17, s4
-; CHECK-LIS-NEXT: vmov.f32 s18, s7
-; CHECK-LIS-NEXT: vmov.f32 s22, s6
-; CHECK-LIS-NEXT: vmov.f32 s16, s9
-; CHECK-LIS-NEXT: vmov.f32 s19, s14
-; CHECK-LIS-NEXT: vmov.f32 s20, s8
-; CHECK-LIS-NEXT: vmov.f32 s21, s11
-; CHECK-LIS-NEXT: vmov.f32 s23, s13
-; CHECK-LIS-NEXT: vmov.f32 s4, s10
-; CHECK-LIS-NEXT: vldrw.u32 q2, [r0, #160]
-; CHECK-LIS-NEXT: vmov.f32 s6, s12
-; CHECK-LIS-NEXT: vadd.f32 q4, q5, q4
-; CHECK-LIS-NEXT: vmov.f32 s7, s15
-; CHECK-LIS-NEXT: vldrw.u32 q3, [r0, #144]
-; CHECK-LIS-NEXT: vadd.f32 q1, q4, q1
-; CHECK-LIS-NEXT: vmov.f32 s18, s10
-; CHECK-LIS-NEXT: vmov.f32 s21, s8
-; CHECK-LIS-NEXT: vmov.f32 s22, s11
-; CHECK-LIS-NEXT: vmov.f32 s16, s12
-; CHECK-LIS-NEXT: vmov.f32 s17, s15
-; CHECK-LIS-NEXT: vmov.f32 s20, s13
-; CHECK-LIS-NEXT: vmov.f32 s23, s26
-; CHECK-LIS-NEXT: vmov.f32 s19, s25
-; CHECK-LIS-NEXT: vadd.f32 q4, q4, q5
-; CHECK-LIS-NEXT: vmov.f32 s8, s14
-; CHECK-LIS-NEXT: vmov.f32 s10, s24
-; CHECK-LIS-NEXT: vldrw.u32 q3, [r0, #112]
-; CHECK-LIS-NEXT: vmov.f32 s11, s27
-; CHECK-LIS-NEXT: vldrw.u32 q5, [r0, #128]
-; CHECK-LIS-NEXT: vadd.f32 q2, q4, q2
-; CHECK-LIS-NEXT: vldrw.u32 q4, [r0, #96]
-; CHECK-LIS-NEXT: vmov.f32 s25, s12
-; CHECK-LIS-NEXT: vstrw.32 q2, [r1, #48]
-; CHECK-LIS-NEXT: vmov.f32 s26, s15
-; CHECK-LIS-NEXT: vstrw.32 q0, [r1, #16]
-; CHECK-LIS-NEXT: vmov.f32 s30, s14
-; CHECK-LIS-NEXT: vstrw.32 q1, [r1]
-; CHECK-LIS-NEXT: vmov.f32 s24, s17
-; CHECK-LIS-NEXT: vmov.f32 s27, s22
-; CHECK-LIS-NEXT: vmov.f32 s28, s16
-; CHECK-LIS-NEXT: vmov.f32 s29, s19
-; CHECK-LIS-NEXT: vmov.f32 s31, s21
-; CHECK-LIS-NEXT: vadd.f32 q6, q7, q6
-; CHECK-LIS-NEXT: vmov.f32 s12, s18
-; CHECK-LIS-NEXT: vmov.f32 s14, s20
-; CHECK-LIS-NEXT: vmov.f32 s15, s23
-; CHECK-LIS-NEXT: vadd.f32 q3, q6, q3
-; CHECK-LIS-NEXT: vstrw.32 q3, [r1, #32]
-; CHECK-LIS-NEXT: vpop {d8, d9, d10, d11, d12, d13, d14, d15}
-; CHECK-LIS-NEXT: bx lr
entry:
%l1 = load <48 x float>, ptr %src, align 4
%s1 = shufflevector <48 x float> %l1, <48 x float> undef, <16 x i32> <i32 0, i32 3, i32 6, i32 9, i32 12, i32 15, i32 18, i32 21, i32 24, i32 27, i32 30, i32 33, i32 36, i32 39, i32 42, i32 45>
@@ -1692,7 +1560,7 @@ define void @vld3_v16f16(ptr %src, ptr %dst) {
; CHECK-LV-NEXT: vstrw.32 q0, [r1]
; CHECK-LV-NEXT: vpop {d8, d9}
; CHECK-LV-NEXT: bx lr
-
+;
; CHECK-LIS-LABEL: vld3_v16f16:
; CHECK-LIS: @ %bb.0: @ %entry
; CHECK-LIS-NEXT: .vsave {d8, d9}
@@ -1773,6 +1641,7 @@ define void @vld3_v16f16(ptr %src, ptr %dst) {
; CHECK-LIS-NEXT: vstrw.32 q0, [r1]
; CHECK-LIS-NEXT: vpop {d8, d9}
; CHECK-LIS-NEXT: bx lr
+
entry:
%l1 = load <48 x half>, ptr %src, align 4
%s1 = shufflevector <48 x half> %l1, <48 x half> undef, <16 x i32> <i32 0, i32 3, i32 6, i32 9, i32 12, i32 15, i32 18, i32 21, i32 24, i32 27, i32 30, i32 33, i32 36, i32 39, i32 42, i32 45>
@@ -1841,4 +1710,4 @@ entry:
%a = fadd <4 x double> %a1, %s3
store <4 x double> %a, ptr %dst
ret void
-}
\ No newline at end of file
+}
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 7d2915ddc75b1..5f2c0108a2286 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
@@ -4451,8 +4451,8 @@ define void @vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3(ptr %i
; AVX-NEXT: vmovdqa 16(%rdi), %xmm1
; AVX-NEXT: vmovdqa 48(%rdi), %xmm2
; AVX-NEXT: vpaddb 16(%rsi), %xmm1, %xmm1
-; AVX-NEXT: vpaddb (%rsi), %xmm0, %xmm0
; AVX-NEXT: vpaddb 48(%rsi), %xmm2, %xmm2
+; AVX-NEXT: vpaddb (%rsi), %xmm0, %xmm0
; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm0[0,1],xmm2[2,3,4,5,6,7]
; AVX-NEXT: vpaddb (%rdx), %xmm2, %xmm2
; AVX-NEXT: vpaddb 48(%rdx), %xmm1, %xmm1
@@ -4462,6 +4462,7 @@ define void @vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3(ptr %i
; AVX-NEXT: vmovdqa %xmm3, 32(%rcx)
; AVX-NEXT: vmovdqa %xmm1, 48(%rcx)
; AVX-NEXT: vmovdqa %xmm2, (%rcx)
+; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3:
@@ -4716,6 +4717,7 @@ define void @vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3(ptr %i
; AVX-NEXT: vmovdqa %xmm3, 32(%rcx)
; AVX-NEXT: vmovdqa %xmm1, 48(%rcx)
; AVX-NEXT: vmovdqa %xmm2, (%rcx)
+; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3:
diff --git a/llvm/test/CodeGen/X86/merge-consecutive-loads-128.ll b/llvm/test/CodeGen/X86/merge-consecutive-loads-128.ll
index 595f8491b405c..1df4e9f47f21b 100644
--- a/llvm/test/CodeGen/X86/merge-consecutive-loads-128.ll
+++ b/llvm/test/CodeGen/X86/merge-consecutive-loads-128.ll
@@ -262,37 +262,54 @@ define <4 x float> @merge_4f32_f32_45zz(ptr %ptr) nounwind uwtable noinline ssp
define <4 x float> @merge_4f32_f32_012u(ptr %ptr) nounwind uwtable noinline ssp {
; SSE2-LABEL: merge_4f32_f32_012u:
; SSE2: # %bb.0:
+; SSE2-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; SSE2-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
-; SSE2-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE2-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
+; SSE2-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; SSE2-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]
; SSE2-NEXT: retq
;
; SSE41-LABEL: merge_4f32_f32_012u:
; SSE41: # %bb.0:
-; SSE41-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
-; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
+; SSE41-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE41-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; SSE41-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],xmm2[0],xmm0[3]
+; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1,2,0]
; SSE41-NEXT: retq
;
; AVX-LABEL: merge_4f32_f32_012u:
; AVX: # %bb.0:
-; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
+; AVX-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; AVX-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; AVX-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],xmm2[0],xmm0[3]
+; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2,0]
; AVX-NEXT: retq
;
; X86-SSE1-LABEL: merge_4f32_f32_012u:
; X86-SSE1: # %bb.0:
; X86-SSE1-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-SSE1-NEXT: xorps %xmm0, %xmm0
-; X86-SSE1-NEXT: movlps {{.*#+}} xmm0 = mem[0,1],xmm0[2,3]
+; X86-SSE1-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; X86-SSE1-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; X86-SSE1-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; X86-SSE1-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; X86-SSE1-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
+; X86-SSE1-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; X86-SSE1-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]
; X86-SSE1-NEXT: retl
;
; X86-SSE41-LABEL: merge_4f32_f32_012u:
; X86-SSE41: # %bb.0:
; X86-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-SSE41-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
-; X86-SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
+; X86-SSE41-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-SSE41-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X86-SSE41-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; X86-SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+; X86-SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],xmm2[0],xmm0[3]
+; X86-SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1,2,0]
; X86-SSE41-NEXT: retl
%ptr1 = getelementptr inbounds float, ptr %ptr, i64 1
%ptr2 = getelementptr inbounds float, ptr %ptr, i64 2
@@ -309,37 +326,54 @@ define <4 x float> @merge_4f32_f32_012u(ptr %ptr) nounwind uwtable noinline ssp
define <4 x float> @merge_4f32_f32_019u(ptr %ptr) nounwind uwtable noinline ssp {
; SSE2-LABEL: merge_4f32_f32_019u:
; SSE2: # %bb.0:
+; SSE2-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; SSE2-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
-; SSE2-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE2-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
+; SSE2-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; SSE2-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]
; SSE2-NEXT: retq
;
; SSE41-LABEL: merge_4f32_f32_019u:
; SSE41: # %bb.0:
-; SSE41-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
-; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
+; SSE41-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE41-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; SSE41-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],xmm2[0],xmm0[3]
+; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1,2,0]
; SSE41-NEXT: retq
;
; AVX-LABEL: merge_4f32_f32_019u:
; AVX: # %bb.0:
-; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
+; AVX-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; AVX-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; AVX-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],xmm2[0],xmm0[3]
+; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2,0]
; AVX-NEXT: retq
;
; X86-SSE1-LABEL: merge_4f32_f32_019u:
; X86-SSE1: # %bb.0:
; X86-SSE1-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-SSE1-NEXT: xorps %xmm0, %xmm0
-; X86-SSE1-NEXT: movlps {{.*#+}} xmm0 = mem[0,1],xmm0[2,3]
+; X86-SSE1-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; X86-SSE1-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; X86-SSE1-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; X86-SSE1-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; X86-SSE1-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
+; X86-SSE1-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; X86-SSE1-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]
; X86-SSE1-NEXT: retl
;
; X86-SSE41-LABEL: merge_4f32_f32_019u:
; X86-SSE41: # %bb.0:
; X86-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-SSE41-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
-; X86-SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
+; X86-SSE41-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-SSE41-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X86-SSE41-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; X86-SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+; X86-SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],xmm2[0],xmm0[3]
+; X86-SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1,2,0]
; X86-SSE41-NEXT: retl
%ptr1 = getelementptr inbounds float, ptr %ptr, i64 1
%ptr2 = getelementptr inbounds float, ptr %ptr, i64 9
diff --git a/llvm/test/CodeGen/X86/mmx-build-vector.ll b/llvm/test/CodeGen/X86/mmx-build-vector.ll
index d8a010bacc683..1d2b730d54a38 100644
--- a/llvm/test/CodeGen/X86/mmx-build-vector.ll
+++ b/llvm/test/CodeGen/X86/mmx-build-vector.ll
@@ -290,15 +290,21 @@ define void @build_v4i16_0zuz(ptr%p0, i16 %a0, i16 %a1, i16 %a2, i16 %a3) nounwi
define void @build_v4i16_012u(ptr%p0, i16 %a0, i16 %a1, i16 %a2, i16 %a3) nounwind {
; X86-LABEL: build_v4i16_012u:
; X86: # %bb.0:
+; X86-NEXT: pushl %esi
+; X86-NEXT: movzwl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: movzwl {{[0-9]+}}(%esp), %edx
+; X86-NEXT: movzwl {{[0-9]+}}(%esp), %esi
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: movd {{[0-9]+}}(%esp), %mm0
-; X86-NEXT: punpcklwd %mm0, %mm0 # mm0 = mm0[0,0,1,1]
-; X86-NEXT: movd {{[0-9]+}}(%esp), %mm1
-; X86-NEXT: movd {{[0-9]+}}(%esp), %mm2
-; X86-NEXT: punpcklwd %mm1, %mm2 # mm2 = mm2[0],mm1[0],mm2[1],mm1[1]
-; X86-NEXT: punpckldq %mm0, %mm2 # mm2 = mm2[0],mm0[0]
+; X86-NEXT: movd %eax, %mm0
+; X86-NEXT: movd %esi, %mm1
+; X86-NEXT: punpcklwd %mm0, %mm1 # mm1 = mm1[0],mm0[0],mm1[1],mm0[1]
+; X86-NEXT: movd %edx, %mm0
+; X86-NEXT: movd %ecx, %mm2
+; X86-NEXT: punpcklwd %mm0, %mm2 # mm2 = mm2[0],mm0[0],mm2[1],mm0[1]
+; X86-NEXT: punpckldq %mm1, %mm2 # mm2 = mm2[0],mm1[0]
; X86-NEXT: paddd %mm2, %mm2
; X86-NEXT: movq %mm2, (%eax)
+; X86-NEXT: popl %esi
; X86-NEXT: retl
;
; X64-LABEL: build_v4i16_012u:
@@ -475,45 +481,61 @@ define void @build_v8i8_0u2345z7(ptr%p0, i8 %a0, i8 %a1, i8 %a2, i8 %a3, i8 %a4,
ret void
}
+
+; Recursion depth limit in isGuaranteedNotToBeUndefOrPoison prevents llc from
+; detecting that we insert an "undef" element in a position that already is
+; undef. OTOH, opt would optimize away that insertelement operation from the
+; IR, so maybe that isn't a problem in reality.
define void @build_v8i8_0123zzzu(ptr%p0, i8 %a0, i8 %a1, i8 %a2, i8 %a3, i8 %a4, i8 %a5, i8 %a6, i8 %a7) nounwind {
-; X86-LABEL: build_v8i8_0123zzzu:
-; X86: # %bb.0:
-; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: movd {{[0-9]+}}(%esp), %mm0
-; X86-NEXT: movd {{[0-9]+}}(%esp), %mm1
-; X86-NEXT: punpcklbw %mm0, %mm1 # mm1 = mm1[0],mm0[0],mm1[1],mm0[1],mm1[2],mm0[2],mm1[3],mm0[3]
-; X86-NEXT: movd {{[0-9]+}}(%esp), %mm0
-; X86-NEXT: movd {{[0-9]+}}(%esp), %mm2
-; X86-NEXT: punpcklbw %mm0, %mm2 # mm2 = mm2[0],mm0[0],mm2[1],mm0[1],mm2[2],mm0[2],mm2[3],mm0[3]
-; X86-NEXT: punpcklwd %mm1, %mm2 # mm2 = mm2[0],mm1[0],mm2[1],mm1[1]
-; X86-NEXT: pxor %mm0, %mm0
-; X86-NEXT: pxor %mm1, %mm1
-; X86-NEXT: punpcklbw %mm1, %mm1 # mm1 = mm1[0,0,1,1,2,2,3,3]
-; X86-NEXT: punpcklbw %mm0, %mm0 # mm0 = mm0[0,0,1,1,2,2,3,3]
-; X86-NEXT: punpcklwd %mm1, %mm0 # mm0 = mm0[0],mm1[0],mm0[1],mm1[1]
-; X86-NEXT: punpckldq %mm0, %mm2 # mm2 = mm2[0],mm0[0]
-; X86-NEXT: paddd %mm2, %mm2
-; X86-NEXT: movq %mm2, (%eax)
-; X86-NEXT: retl
+; X86-MMX-LABEL: build_v8i8_0123zzzu:
+; X86-MMX: # %bb.0:
+; X86-MMX-NEXT: pushl %ebp
+; X86-MMX-NEXT: movl %esp, %ebp
+; X86-MMX-NEXT: pushl %esi
+; X86-MMX-NEXT: andl $-8, %esp
+; X86-MMX-NEXT: subl $16, %esp
+; X86-MMX-NEXT: movl 8(%ebp), %eax
+; X86-MMX-NEXT: movzbl 20(%ebp), %edx
+; X86-MMX-NEXT: movzbl 24(%ebp), %ecx
+; X86-MMX-NEXT: shll $8, %ecx
+; X86-MMX-NEXT: orl %edx, %ecx
+; X86-MMX-NEXT: shll $16, %ecx
+; X86-MMX-NEXT: movzbl 12(%ebp), %edx
+; X86-MMX-NEXT: movzbl 16(%ebp), %esi
+; X86-MMX-NEXT: shll $8, %esi
+; X86-MMX-NEXT: orl %edx, %esi
+; X86-MMX-NEXT: movzwl %si, %edx
+; X86-MMX-NEXT: orl %ecx, %edx
+; X86-MMX-NEXT: movzbl %al, %ecx
+; X86-MMX-NEXT: shll $24, %ecx
+; X86-MMX-NEXT: movl %ecx, {{[0-9]+}}(%esp)
+; X86-MMX-NEXT: movl %edx, (%esp)
+; X86-MMX-NEXT: movq (%esp), %mm0
+; X86-MMX-NEXT: paddd %mm0, %mm0
+; X86-MMX-NEXT: movq %mm0, (%eax)
+; X86-MMX-NEXT: leal -4(%ebp), %esp
+; X86-MMX-NEXT: popl %esi
+; X86-MMX-NEXT: popl %ebp
+; X86-MMX-NEXT: retl
;
-; X64-LABEL: build_v8i8_0123zzzu:
-; X64: # %bb.0:
-; X64-NEXT: movd %r8d, %mm0
-; X64-NEXT: movd %ecx, %mm1
-; X64-NEXT: punpcklbw %mm0, %mm1 # mm1 = mm1[0],mm0[0],mm1[1],mm0[1],mm1[2],mm0[2],mm1[3],mm0[3]
-; X64-NEXT: movd %edx, %mm0
-; X64-NEXT: movd %esi, %mm2
-; X64-NEXT: punpcklbw %mm0, %mm2 # mm2 = mm2[0],mm0[0],mm2[1],mm0[1],mm2[2],mm0[2],mm2[3],mm0[3]
-; X64-NEXT: punpcklwd %mm1, %mm2 # mm2 = mm2[0],mm1[0],mm2[1],mm1[1]
-; X64-NEXT: pxor %mm0, %mm0
-; X64-NEXT: pxor %mm1, %mm1
-; X64-NEXT: punpcklbw %mm1, %mm1 # mm1 = mm1[0,0,1,1,2,2,3,3]
-; X64-NEXT: punpcklbw %mm0, %mm0 # mm0 = mm0[0,0,1,1,2,2,3,3]
-; X64-NEXT: punpcklwd %mm1, %mm0 # mm0 = mm0[0],mm1[0],mm0[1],mm1[1]
-; X64-NEXT: punpckldq %mm0, %mm2 # mm2 = mm2[0],mm0[0]
-; X64-NEXT: paddd %mm2, %mm2
-; X64-NEXT: movq %mm2, (%rdi)
-; X64-NEXT: retq
+; X86-SSE-LABEL: build_v8i8_0123zzzu:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
+; X86-SSE-NEXT: movzbl {{[0-9]+}}(%esp), %edx
+; X86-SSE-NEXT: shll $8, %edx
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
+; X86-SSE-NEXT: shll $16, %ecx
+; X86-SSE-NEXT: orl %edx, %ecx
+; X86-SSE-NEXT: movzbl {{[0-9]+}}(%esp), %edx
+; X86-SSE-NEXT: shll $24, %edx
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: movd %edx, %xmm0
+; X86-SSE-NEXT: movdq2q %xmm0, %mm0
+; X86-SSE-NEXT: paddd %mm0, %mm0
+; X86-SSE-NEXT: movq %mm0, (%eax)
+; X86-SSE-NEXT: retl
%1 = insertelement <8 x i8> undef, i8 %a0, i32 0
%2 = insertelement <8 x i8> %1, i8 %a1, i32 1
%3 = insertelement <8 x i8> %2, i8 %a2, i32 2
@@ -558,22 +580,34 @@ define void @build_v8i8_0uuuuzzz(ptr%p0, i8 %a0, i8 %a1, i8 %a2, i8 %a3, i8 %a4,
}
define void @build_v8i8_0zzzzzzu(ptr%p0, i8 %a0, i8 %a1, i8 %a2, i8 %a3, i8 %a4, i8 %a5, i8 %a6, i8 %a7) nounwind {
-; X86-LABEL: build_v8i8_0zzzzzzu:
-; X86: # %bb.0:
-; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: movd %eax, %mm0
-; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: paddd %mm0, %mm0
-; X86-NEXT: movq %mm0, (%eax)
-; X86-NEXT: retl
+; X86-MMX-LABEL: build_v8i8_0zzzzzzu:
+; X86-MMX: # %bb.0:
+; X86-MMX-NEXT: pushl %ebp
+; X86-MMX-NEXT: movl %esp, %ebp
+; X86-MMX-NEXT: andl $-8, %esp
+; X86-MMX-NEXT: subl $8, %esp
+; X86-MMX-NEXT: movl 8(%ebp), %eax
+; X86-MMX-NEXT: movzbl 12(%ebp), %ecx
+; X86-MMX-NEXT: movl %ecx, (%esp)
+; X86-MMX-NEXT: movzbl %al, %ecx
+; X86-MMX-NEXT: shll $24, %ecx
+; X86-MMX-NEXT: movl %ecx, {{[0-9]+}}(%esp)
+; X86-MMX-NEXT: movq (%esp), %mm0
+; X86-MMX-NEXT: paddd %mm0, %mm0
+; X86-MMX-NEXT: movq %mm0, (%eax)
+; X86-MMX-NEXT: movl %ebp, %esp
+; X86-MMX-NEXT: popl %ebp
+; X86-MMX-NEXT: retl
;
-; X64-LABEL: build_v8i8_0zzzzzzu:
-; X64: # %bb.0:
-; X64-NEXT: movzbl %sil, %eax
-; X64-NEXT: movd %eax, %mm0
-; X64-NEXT: paddd %mm0, %mm0
-; X64-NEXT: movq %mm0, (%rdi)
-; X64-NEXT: retq
+; X86-SSE-LABEL: build_v8i8_0zzzzzzu:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
+; X86-SSE-NEXT: movd %ecx, %xmm0
+; X86-SSE-NEXT: movdq2q %xmm0, %mm0
+; X86-SSE-NEXT: paddd %mm0, %mm0
+; X86-SSE-NEXT: movq %mm0, (%eax)
+; X86-SSE-NEXT: retl
%1 = insertelement <8 x i8> undef, i8 %a0, i32 0
%2 = insertelement <8 x i8> %1, i8 0, i32 1
%3 = insertelement <8 x i8> %2, i8 0, i32 2
diff --git a/llvm/test/CodeGen/X86/oddshuffles.ll b/llvm/test/CodeGen/X86/oddshuffles.ll
index 6b9a86343ea10..1428fa0924a83 100644
--- a/llvm/test/CodeGen/X86/oddshuffles.ll
+++ b/llvm/test/CodeGen/X86/oddshuffles.ll
@@ -1920,35 +1920,36 @@ define void @splat3_128(<16 x i8> %a0, <16 x i8> %a1, ptr%a2) {
;
; AVX1-LABEL: splat3_128:
; AVX1: # %bb.0:
-; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm1[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm0[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm1[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm0[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm6 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm7 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm0[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm1[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm0[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm1[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm6 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm7 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm1 = xmm7[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm7[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm1 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
; AVX1-NEXT: vmovdqa {{.*#+}} xmm6 = [0,11,6,1,12,7,2,13,8,3,14,9,4,15,10,5]
-; AVX1-NEXT: vpshufb %xmm6, %xmm3, %xmm3
-; AVX1-NEXT: vpshufb %xmm6, %xmm0, %xmm0
; AVX1-NEXT: vpshufb %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vpshufb %xmm6, %xmm1, %xmm1
; AVX1-NEXT: vpshufb %xmm6, %xmm2, %xmm2
+; AVX1-NEXT: vpshufb %xmm6, %xmm0, %xmm0
+; AVX1-NEXT: vpshufb %xmm6, %xmm1, %xmm1
+; AVX1-NEXT: vpshufb %xmm6, %xmm3, %xmm3
; AVX1-NEXT: vpshufb %xmm6, %xmm4, %xmm4
; AVX1-NEXT: vmovdqa %xmm4, 80(%rdi)
-; AVX1-NEXT: vmovdqa %xmm2, 64(%rdi)
+; AVX1-NEXT: vmovdqa %xmm3, 64(%rdi)
; AVX1-NEXT: vmovdqa %xmm1, 48(%rdi)
-; AVX1-NEXT: vmovdqa %xmm5, 32(%rdi)
-; AVX1-NEXT: vmovdqa %xmm3, 16(%rdi)
-; AVX1-NEXT: vmovdqa %xmm0, (%rdi)
+; AVX1-NEXT: vmovdqa %xmm0, 32(%rdi)
+; AVX1-NEXT: vmovdqa %xmm2, 16(%rdi)
+; AVX1-NEXT: vmovdqa %xmm5, (%rdi)
+; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-LABEL: splat3_128:
@@ -1988,20 +1989,21 @@ define void @splat3_128(<16 x i8> %a0, <16 x i8> %a1, ptr%a2) {
; XOP-NEXT: vpalignr {{.*#+}} xmm3 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
; XOP-NEXT: vpalignr {{.*#+}} xmm2 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
; XOP-NEXT: vpalignr {{.*#+}} xmm5 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; XOP-NEXT: vpalignr {{.*#+}} xmm4 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; XOP-NEXT: vmovdqa {{.*#+}} xmm8 = [5,16,11,6,17,12,7,18,13,8,19,14,9,20,15,10]
-; XOP-NEXT: vpperm %xmm8, %xmm4, %xmm2, %xmm2
-; XOP-NEXT: vpperm %xmm8, %xmm0, %xmm7, %xmm0
-; XOP-NEXT: vpperm %xmm8, %xmm7, %xmm4, %xmm4
-; XOP-NEXT: vpperm %xmm8, %xmm1, %xmm6, %xmm1
-; XOP-NEXT: vpperm %xmm8, %xmm5, %xmm3, %xmm3
-; XOP-NEXT: vpperm %xmm8, %xmm6, %xmm5, %xmm5
-; XOP-NEXT: vmovdqa %xmm5, 80(%rdi)
+; XOP-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; XOP-NEXT: vmovdqa {{.*#+}} xmm4 = [5,16,11,6,17,12,7,18,13,8,19,14,9,20,15,10]
+; XOP-NEXT: vpperm %xmm4, %xmm2, %xmm7, %xmm8
+; XOP-NEXT: vpperm %xmm4, %xmm0, %xmm2, %xmm2
+; XOP-NEXT: vpperm %xmm4, %xmm7, %xmm0, %xmm0
+; XOP-NEXT: vpperm %xmm4, %xmm1, %xmm6, %xmm1
+; XOP-NEXT: vpperm %xmm4, %xmm5, %xmm3, %xmm3
+; XOP-NEXT: vpperm %xmm4, %xmm6, %xmm5, %xmm4
+; XOP-NEXT: vmovdqa %xmm4, 80(%rdi)
; XOP-NEXT: vmovdqa %xmm3, 64(%rdi)
; XOP-NEXT: vmovdqa %xmm1, 48(%rdi)
-; XOP-NEXT: vmovdqa %xmm4, 32(%rdi)
+; XOP-NEXT: vmovdqa %xmm0, 32(%rdi)
; XOP-NEXT: vmovdqa %xmm2, 16(%rdi)
-; XOP-NEXT: vmovdqa %xmm0, (%rdi)
+; XOP-NEXT: vmovdqa %xmm8, (%rdi)
+; XOP-NEXT: vzeroupper
; XOP-NEXT: retq
%1 = shufflevector <16 x i8> %a0, <16 x i8> %a1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
%2 = shufflevector <16 x i8> %a0, <16 x i8> %a1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
@@ -2092,36 +2094,36 @@ define void @splat3_256(<32 x i8> %a0, ptr%a1) {
;
; AVX1-LABEL: splat3_256:
; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm1[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm0[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm1[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm0[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm6 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm7 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm1 = xmm0[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
+; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm2[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm0[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm2[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm6 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm7 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm1 = xmm7[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm1 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm7[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
; AVX1-NEXT: vmovdqa {{.*#+}} xmm6 = [0,11,6,1,12,7,2,13,8,3,14,9,4,15,10,5]
-; AVX1-NEXT: vpshufb %xmm6, %xmm3, %xmm3
-; AVX1-NEXT: vpshufb %xmm6, %xmm0, %xmm0
; AVX1-NEXT: vpshufb %xmm6, %xmm5, %xmm5
; AVX1-NEXT: vpshufb %xmm6, %xmm1, %xmm1
+; AVX1-NEXT: vpshufb %xmm6, %xmm0, %xmm0
; AVX1-NEXT: vpshufb %xmm6, %xmm2, %xmm2
+; AVX1-NEXT: vpshufb %xmm6, %xmm3, %xmm3
; AVX1-NEXT: vpshufb %xmm6, %xmm4, %xmm4
; AVX1-NEXT: vmovdqa %xmm4, 80(%rdi)
-; AVX1-NEXT: vmovdqa %xmm2, 64(%rdi)
-; AVX1-NEXT: vmovdqa %xmm1, 48(%rdi)
-; AVX1-NEXT: vmovdqa %xmm5, 32(%rdi)
-; AVX1-NEXT: vmovdqa %xmm3, 16(%rdi)
-; AVX1-NEXT: vmovdqa %xmm0, (%rdi)
+; AVX1-NEXT: vmovdqa %xmm3, 64(%rdi)
+; AVX1-NEXT: vmovdqa %xmm2, 48(%rdi)
+; AVX1-NEXT: vmovdqa %xmm0, 32(%rdi)
+; AVX1-NEXT: vmovdqa %xmm1, 16(%rdi)
+; AVX1-NEXT: vmovdqa %xmm5, (%rdi)
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
@@ -2161,20 +2163,20 @@ define void @splat3_256(<32 x i8> %a0, ptr%a1) {
; XOP-NEXT: vpalignr {{.*#+}} xmm3 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
; XOP-NEXT: vpalignr {{.*#+}} xmm1 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
; XOP-NEXT: vpalignr {{.*#+}} xmm5 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; XOP-NEXT: vpalignr {{.*#+}} xmm4 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; XOP-NEXT: vmovdqa {{.*#+}} xmm8 = [5,16,11,6,17,12,7,18,13,8,19,14,9,20,15,10]
-; XOP-NEXT: vpperm %xmm8, %xmm4, %xmm1, %xmm1
-; XOP-NEXT: vpperm %xmm8, %xmm0, %xmm7, %xmm0
-; XOP-NEXT: vpperm %xmm8, %xmm7, %xmm4, %xmm4
-; XOP-NEXT: vpperm %xmm8, %xmm2, %xmm6, %xmm2
-; XOP-NEXT: vpperm %xmm8, %xmm5, %xmm3, %xmm3
-; XOP-NEXT: vpperm %xmm8, %xmm6, %xmm5, %xmm5
-; XOP-NEXT: vmovdqa %xmm5, 80(%rdi)
+; XOP-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; XOP-NEXT: vmovdqa {{.*#+}} xmm4 = [5,16,11,6,17,12,7,18,13,8,19,14,9,20,15,10]
+; XOP-NEXT: vpperm %xmm4, %xmm1, %xmm7, %xmm8
+; XOP-NEXT: vpperm %xmm4, %xmm0, %xmm1, %xmm1
+; XOP-NEXT: vpperm %xmm4, %xmm7, %xmm0, %xmm0
+; XOP-NEXT: vpperm %xmm4, %xmm2, %xmm6, %xmm2
+; XOP-NEXT: vpperm %xmm4, %xmm5, %xmm3, %xmm3
+; XOP-NEXT: vpperm %xmm4, %xmm6, %xmm5, %xmm4
+; XOP-NEXT: vmovdqa %xmm4, 80(%rdi)
; XOP-NEXT: vmovdqa %xmm3, 64(%rdi)
; XOP-NEXT: vmovdqa %xmm2, 48(%rdi)
-; XOP-NEXT: vmovdqa %xmm4, 32(%rdi)
+; XOP-NEXT: vmovdqa %xmm0, 32(%rdi)
; XOP-NEXT: vmovdqa %xmm1, 16(%rdi)
-; XOP-NEXT: vmovdqa %xmm0, (%rdi)
+; XOP-NEXT: vmovdqa %xmm8, (%rdi)
; XOP-NEXT: vzeroupper
; XOP-NEXT: retq
%1 = shufflevector <32 x i8> %a0, <32 x i8> undef, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
@@ -2398,11 +2400,12 @@ define void @D107009(ptr %input, ptr %output) {
;
; AVX1-LABEL: D107009:
; AVX1: # %bb.0:
-; AVX1-NEXT: vmovups 128(%rdi), %ymm0
-; AVX1-NEXT: vmovups 224(%rdi), %ymm1
-; AVX1-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],mem[0],ymm1[2],mem[2]
-; AVX1-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5]
-; AVX1-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,0],ymm0[4,5],ymm1[6,4]
+; AVX1-NEXT: vmovups 224(%rdi), %ymm0
+; AVX1-NEXT: vmovups 128(%rdi), %ymm1
+; AVX1-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[4],mem[4],ymm1[5],mem[5]
+; AVX1-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],mem[0],ymm0[2],mem[2]
+; AVX1-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4]
+; AVX1-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX1-NEXT: vmovaps 112(%rdi), %xmm1
; AVX1-NEXT: vunpcklpd {{.*#+}} xmm1 = xmm1[0],mem[0]
; AVX1-NEXT: vmovaps 16(%rdi), %xmm2
@@ -2412,22 +2415,22 @@ define void @D107009(ptr %input, ptr %output) {
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
; AVX1-NEXT: vpsrld $16, %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,1,1,3]
-; AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm1[2,1,3,3]
-; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX1-NEXT: vshufps {{.*#+}} ymm3 = ymm0[3,3,3,3,7,7,7,7]
-; AVX1-NEXT: vshufpd {{.*#+}} ymm4 = ymm0[0,0,3,2]
-; AVX1-NEXT: vmovshdup {{.*#+}} ymm5 = ymm0[1,1,3,3,5,5,7,7]
+; AVX1-NEXT: vshufps {{.*#+}} ymm2 = ymm0[3,3,3,3,7,7,7,7]
+; AVX1-NEXT: vshufpd {{.*#+}} ymm3 = ymm0[0,0,3,2]
+; AVX1-NEXT: vmovshdup {{.*#+}} ymm4 = ymm0[1,1,3,3,5,5,7,7]
+; AVX1-NEXT: vpshufd {{.*#+}} xmm5 = xmm1[0,1,1,3]
+; AVX1-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[2,1,3,3]
+; AVX1-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[3,3,3,3]
; AVX1-NEXT: vpshufd {{.*#+}} xmm7 = xmm1[1,1,1,1]
; AVX1-NEXT: vmovdqa %xmm1, 16(%rsi)
; AVX1-NEXT: vmovdqa %xmm7, 48(%rsi)
; AVX1-NEXT: vmovdqa %xmm6, 112(%rsi)
+; AVX1-NEXT: vmovups %ymm5, 64(%rsi)
; AVX1-NEXT: vmovups %ymm0, 128(%rsi)
-; AVX1-NEXT: vmovups %ymm5, 160(%rsi)
-; AVX1-NEXT: vmovupd %ymm4, 192(%rsi)
-; AVX1-NEXT: vmovupd %ymm3, 224(%rsi)
-; AVX1-NEXT: vmovups %ymm2, 64(%rsi)
+; AVX1-NEXT: vmovups %ymm4, 160(%rsi)
+; AVX1-NEXT: vmovupd %ymm3, 192(%rsi)
+; AVX1-NEXT: vmovupd %ymm2, 224(%rsi)
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
@@ -2466,11 +2469,12 @@ define void @D107009(ptr %input, ptr %output) {
;
; XOP-LABEL: D107009:
; XOP: # %bb.0:
-; XOP-NEXT: vmovups 128(%rdi), %ymm0
-; XOP-NEXT: vmovups 224(%rdi), %ymm1
-; XOP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],mem[0],ymm1[2],mem[2]
-; XOP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5]
-; XOP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,0],ymm0[4,5],ymm1[6,4]
+; XOP-NEXT: vmovups 224(%rdi), %ymm0
+; XOP-NEXT: vmovups 128(%rdi), %ymm1
+; XOP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[4],mem[4],ymm1[5],mem[5]
+; XOP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],mem[0],ymm0[2],mem[2]
+; XOP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4]
+; XOP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; XOP-NEXT: vmovaps 112(%rdi), %xmm1
; XOP-NEXT: vunpcklpd {{.*#+}} xmm1 = xmm1[0],mem[0]
; XOP-NEXT: vmovaps 16(%rdi), %xmm2
@@ -2480,22 +2484,22 @@ define void @D107009(ptr %input, ptr %output) {
; XOP-NEXT: vextractf128 $1, %ymm0, %xmm0
; XOP-NEXT: vpsrld $16, %xmm0, %xmm0
; XOP-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; XOP-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,1,1,3]
-; XOP-NEXT: vpshufd {{.*#+}} xmm3 = xmm1[2,1,3,3]
-; XOP-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; XOP-NEXT: vshufps {{.*#+}} ymm3 = ymm0[3,3,3,3,7,7,7,7]
-; XOP-NEXT: vshufpd {{.*#+}} ymm4 = ymm0[0,0,3,2]
-; XOP-NEXT: vmovshdup {{.*#+}} ymm5 = ymm0[1,1,3,3,5,5,7,7]
+; XOP-NEXT: vshufps {{.*#+}} ymm2 = ymm0[3,3,3,3,7,7,7,7]
+; XOP-NEXT: vshufpd {{.*#+}} ymm3 = ymm0[0,0,3,2]
+; XOP-NEXT: vmovshdup {{.*#+}} ymm4 = ymm0[1,1,3,3,5,5,7,7]
+; XOP-NEXT: vpshufd {{.*#+}} xmm5 = xmm1[0,1,1,3]
+; XOP-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[2,1,3,3]
+; XOP-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm5
; XOP-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[3,3,3,3]
; XOP-NEXT: vpshufd {{.*#+}} xmm7 = xmm1[1,1,1,1]
; XOP-NEXT: vmovdqa %xmm1, 16(%rsi)
; XOP-NEXT: vmovdqa %xmm7, 48(%rsi)
; XOP-NEXT: vmovdqa %xmm6, 112(%rsi)
+; XOP-NEXT: vmovups %ymm5, 64(%rsi)
; XOP-NEXT: vmovups %ymm0, 128(%rsi)
-; XOP-NEXT: vmovups %ymm5, 160(%rsi)
-; XOP-NEXT: vmovupd %ymm4, 192(%rsi)
-; XOP-NEXT: vmovupd %ymm3, 224(%rsi)
-; XOP-NEXT: vmovups %ymm2, 64(%rsi)
+; XOP-NEXT: vmovups %ymm4, 160(%rsi)
+; XOP-NEXT: vmovupd %ymm3, 192(%rsi)
+; XOP-NEXT: vmovupd %ymm2, 224(%rsi)
; XOP-NEXT: vzeroupper
; XOP-NEXT: retq
%i = load <64 x i32>, ptr %input, align 16
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-5.ll
index 9b19ec15c6f55..0797a83ddaa96 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-5.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-5.ll
@@ -1528,26 +1528,26 @@ define void @load_i16_stride5_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vmovdqa 128(%rdi), %xmm7
; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm7[0,1],xmm8[2,3],xmm7[4,5,6,7]
; AVX-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,u,u,u,u,u,u,u,u,2,3,12,13,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm3[0,1,2,3,4],xmm4[5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm3[0,1,2,3,4],xmm4[5,6,7]
; AVX-NEXT: vmovdqa (%rdi), %xmm3
; AVX-NEXT: vmovdqa 16(%rdi), %xmm4
; AVX-NEXT: vmovdqa 32(%rdi), %xmm5
; AVX-NEXT: vmovdqa 48(%rdi), %xmm6
-; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm6[0,1,0,3]
-; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm9[0,1,2,3],xmm5[4],xmm9[5,6,7]
+; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm6[0,1,0,3]
+; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm10[0,1,2,3],xmm5[4],xmm10[5,6,7]
; AVX-NEXT: vpshufd {{.*#+}} xmm11 = xmm4[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm11 = xmm11[2,1,2,3,4,5,6,7]
; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm3[0,2,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,3,2,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1]
-; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm11[0,1,2,3],xmm9[4,5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm11[0,1,2,3],xmm10[4,5,6,7]
; AVX-NEXT: vmovaps {{.*#+}} ymm11 = [65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,65535,65535,65535]
-; AVX-NEXT: vandps %ymm11, %ymm9, %ymm12
-; AVX-NEXT: vmovaps 64(%rdi), %xmm9
-; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm9[0,1,0,1]
+; AVX-NEXT: vandps %ymm11, %ymm10, %ymm12
+; AVX-NEXT: vmovaps 64(%rdi), %xmm10
+; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm10[0,1,0,1]
; AVX-NEXT: vandnps %ymm13, %ymm11, %ymm13
; AVX-NEXT: vorps %ymm13, %ymm12, %ymm12
-; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm12, %ymm10
+; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm12, %ymm9
; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,1,3,0,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,5,5,5,5]
@@ -1557,7 +1557,7 @@ define void @load_i16_stride5_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vpblendw {{.*#+}} xmm13 = xmm7[0,1,2,3],xmm8[4,5],xmm7[6,7]
; AVX-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u,u,u,u,u,u,u,u,u,4,5,14,15,8,9]
; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm12[0,1,2,3,4],xmm13[5,6,7]
-; AVX-NEXT: vpsllq $48, %xmm9, %xmm13
+; AVX-NEXT: vpsllq $48, %xmm10, %xmm13
; AVX-NEXT: vandnps %ymm13, %ymm11, %ymm13
; AVX-NEXT: vpsrlq $48, %xmm4, %xmm14
; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm3[0,3,2,3]
@@ -1583,7 +1583,7 @@ define void @load_i16_stride5_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vpblendw {{.*#+}} xmm14 = xmm5[0,1],xmm6[2,3],xmm5[4,5,6,7]
; AVX-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,u,u,u,u,2,3,12,13,6,7,u,u,u,u]
; AVX-NEXT: vpblendw {{.*#+}} xmm13 = xmm13[0,1,2],xmm14[3,4,5],xmm13[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm14 = xmm9[0,1,2,0]
+; AVX-NEXT: vpshufd {{.*#+}} xmm14 = xmm10[0,1,2,0]
; AVX-NEXT: vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,5,6,5]
; AVX-NEXT: vpblendw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,5],xmm14[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm13, %ymm12
@@ -1600,7 +1600,7 @@ define void @load_i16_stride5_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[2,2,2,2,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,7,4,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm14 = xmm14[0,1,2],xmm15[3,4,5],xmm14[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm9[0,1,0,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm10[0,1,0,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,5,5,6]
; AVX-NEXT: vpblendw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,5],xmm15[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm14, %ymm13
@@ -1621,11 +1621,11 @@ define void @load_i16_stride5_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0,1,2,3],xmm3[4,5],xmm4[6,7]
; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[8,9,2,3,12,13,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3,4,5],xmm2[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm9[0,1,1,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm10[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5],xmm2[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX-NEXT: vmovaps %ymm10, (%rsi)
+; AVX-NEXT: vmovaps %ymm9, (%rsi)
; AVX-NEXT: vmovaps %ymm11, (%rdx)
; AVX-NEXT: vmovaps %ymm12, (%rcx)
; AVX-NEXT: vmovaps %ymm13, (%r8)
@@ -2891,61 +2891,63 @@ define void @load_i16_stride5_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-LABEL: load_i16_stride5_vf32:
; AVX: # %bb.0:
; AVX-NEXT: subq $424, %rsp # imm = 0x1A8
-; AVX-NEXT: vmovdqa 144(%rdi), %xmm9
-; AVX-NEXT: vmovdqa 128(%rdi), %xmm7
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm7[0,1],xmm9[2,3],xmm7[4,5,6,7]
-; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 144(%rdi), %xmm7
+; AVX-NEXT: vmovdqa 128(%rdi), %xmm15
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm15[0,1],xmm7[2,3],xmm15[4,5,6,7]
+; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [6,7,2,3,4,5,6,7,6,7,2,3,12,13,6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovdqa 96(%rdi), %xmm11
-; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm11[0,1,1,3]
+; AVX-NEXT: vmovdqa 96(%rdi), %xmm2
+; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
-; AVX-NEXT: vmovdqa 112(%rdi), %xmm10
-; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm10[1]
-; AVX-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa 80(%rdi), %xmm3
+; AVX-NEXT: vmovdqa 112(%rdi), %xmm3
; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,2,2,3]
+; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm3[1]
+; AVX-NEXT: vmovdqa 80(%rdi), %xmm9
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm9[0,2,2,3]
+; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,3,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3,4],xmm0[5,6,7]
-; AVX-NEXT: vmovdqa (%rdi), %xmm5
-; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa 16(%rdi), %xmm12
+; AVX-NEXT: vmovdqa (%rdi), %xmm4
+; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 16(%rdi), %xmm10
; AVX-NEXT: vmovdqa 32(%rdi), %xmm3
; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa 48(%rdi), %xmm15
-; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm15[0,1,0,3]
+; AVX-NEXT: vmovdqa 48(%rdi), %xmm0
+; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm3[4],xmm0[5,6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm12[3,1,2,3]
-; AVX-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm10[3,1,2,3]
+; AVX-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[2,1,2,3,4,5,6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm5[0,2,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,2,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,3,2,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3],xmm0[4,5,6,7]
-; AVX-NEXT: vmovaps {{.*#+}} ymm6 = [65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,65535,65535,65535]
-; AVX-NEXT: vandps %ymm6, %ymm3, %ymm3
-; AVX-NEXT: vmovaps 64(%rdi), %xmm5
-; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm5[0,1,0,1]
-; AVX-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vandnps %ymm4, %ymm6, %ymm4
+; AVX-NEXT: vmovaps {{.*#+}} ymm5 = [65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,65535,65535,65535]
+; AVX-NEXT: vandps %ymm5, %ymm3, %ymm3
+; AVX-NEXT: vmovaps 64(%rdi), %xmm14
+; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm14[0,1,0,1]
+; AVX-NEXT: vandnps %ymm4, %ymm5, %ymm4
; AVX-NEXT: vorps %ymm4, %ymm3, %ymm3
; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa 304(%rdi), %xmm2
-; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa 288(%rdi), %xmm13
-; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm13[0,1],xmm2[2,3],xmm13[4,5,6,7]
-; AVX-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 304(%rdi), %xmm12
+; AVX-NEXT: vmovdqa 288(%rdi), %xmm0
+; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm0[0,1],xmm12[2,3],xmm0[4,5,6,7]
+; AVX-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufb %xmm1, %xmm2, %xmm1
; AVX-NEXT: vmovdqa 256(%rdi), %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
-; AVX-NEXT: vmovdqa 272(%rdi), %xmm0
-; AVX-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
+; AVX-NEXT: vmovdqa 272(%rdi), %xmm8
+; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm8[1]
+; AVX-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vmovdqa 240(%rdi), %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,2,2,3]
@@ -2956,257 +2958,252 @@ define void @load_i16_stride5_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,1,2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa 160(%rdi), %xmm0
-; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[0,2,2,3]
+; AVX-NEXT: vmovdqa 160(%rdi), %xmm11
+; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm11[0,2,2,3]
+; AVX-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,3,2,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; AVX-NEXT: vmovdqa 208(%rdi), %xmm0
-; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
-; AVX-NEXT: vmovdqa 192(%rdi), %xmm14
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm14[4],xmm0[5,6,7]
-; AVX-NEXT: vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 208(%rdi), %xmm13
+; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm13[0,1,0,3]
+; AVX-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 192(%rdi), %xmm2
+; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4],xmm0[5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
-; AVX-NEXT: vandps %ymm6, %ymm0, %ymm0
+; AVX-NEXT: vandps %ymm5, %ymm0, %ymm0
; AVX-NEXT: vmovaps 224(%rdi), %xmm1
-; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm1[0,1,0,1]
-; AVX-NEXT: vandnps %ymm8, %ymm6, %ymm8
-; AVX-NEXT: vorps %ymm0, %ymm8, %ymm0
+; AVX-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
+; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm1[0,1,0,1]
+; AVX-NEXT: vandnps %ymm6, %ymm5, %ymm6
+; AVX-NEXT: vorps %ymm6, %ymm0, %ymm0
; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa %xmm11, %xmm6
-; AVX-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm10[0,1],xmm11[2,3],xmm10[4,5,6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0,1],xmm4[2,3],xmm2[4,5,6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,0,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm11[0,3,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm9[0,3,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[1,2,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm7[0,1,2,3],xmm9[4,5],xmm7[6,7]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,2,3,4,5,6,7,8,9,4,5,14,15,8,9]
-; AVX-NEXT: vpshufb %xmm8, %xmm3, %xmm3
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm15[0,1,2,3],xmm7[4,5],xmm15[6,7]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,2,3,4,5,6,7,8,9,4,5,14,15,8,9]
+; AVX-NEXT: vpshufb %xmm6, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm3[5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm4[0,3,2,3]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm15[0,3,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[1,2,2,3,4,5,6,7]
-; AVX-NEXT: vpsrlq $48, %xmm12, %xmm9
+; AVX-NEXT: vpsrlq $48, %xmm10, %xmm9
; AVX-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1]
-; AVX-NEXT: vmovdqa %xmm15, %xmm12
-; AVX-NEXT: vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm2[0,1],xmm15[2,3],xmm2[4,5],xmm15[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm10[0,1],xmm7[2,3],xmm10[4,5],xmm7[6,7]
; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,0,1,0,1,0,1,10,11,4,5,14,15,6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm9, %xmm9
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2],xmm9[3,4,5,6,7]
-; AVX-NEXT: vmovaps {{.*#+}} ymm10 = [65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,65535,65535,65535]
-; AVX-NEXT: vandps %ymm3, %ymm10, %ymm3
-; AVX-NEXT: vpsllq $48, %xmm5, %xmm9
-; AVX-NEXT: vandnps %ymm9, %ymm10, %ymm9
+; AVX-NEXT: vandps %ymm5, %ymm3, %ymm3
+; AVX-NEXT: vpsllq $48, %xmm14, %xmm9
+; AVX-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vandnps %ymm9, %ymm5, %ymm9
; AVX-NEXT: vorps %ymm3, %ymm9, %ymm3
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm13[0,1,2,3],xmm5[4,5],xmm13[6,7]
-; AVX-NEXT: vpshufb %xmm8, %xmm0, %xmm0
-; AVX-NEXT: vmovdqa (%rsp), %xmm15 # 16-byte Reload
-; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = xmm15[0,1],mem[2,3],xmm15[4,5,6,7]
+; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT: # xmm0 = mem[0,1,2,3],xmm12[4,5],mem[6,7]
+; AVX-NEXT: vpshufb %xmm6, %xmm0, %xmm0
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm8[0,1],xmm12[2,3],xmm8[4,5,6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,1,3,0,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm8 = xmm10[0,3,2,3]
-; AVX-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[1,2,2,3,4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm8[0,1],xmm3[2,3,4,5,6,7]
+; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
+; AVX-NEXT: # xmm6 = mem[0,3,2,3]
+; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[1,2,2,3,4,5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm6[0,1],xmm3[2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1,2,3,4],xmm0[5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm14[0,1],xmm13[2,3],xmm14[4,5],xmm13[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm8[0,1],xmm13[2,3],xmm8[4,5],xmm13[6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm1
-; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = mem[0,3,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm11[0,3,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[1,2,2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX-NEXT: vpsrlq $48, %xmm8, %xmm8
-; AVX-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX-NEXT: vpsrlq $48, %xmm6, %xmm6
+; AVX-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3,4,5,6,7]
-; AVX-NEXT: vmovaps {{.*#+}} ymm8 = [65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,65535,65535,65535]
-; AVX-NEXT: vandps %ymm1, %ymm8, %ymm1
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX-NEXT: vpsllq $48, %xmm14, %xmm3
-; AVX-NEXT: vandnps %ymm3, %ymm8, %ymm3
+; AVX-NEXT: vandps %ymm5, %ymm1, %ymm1
+; AVX-NEXT: vmovdqa (%rsp), %xmm13 # 16-byte Reload
+; AVX-NEXT: vpsllq $48, %xmm13, %xmm3
+; AVX-NEXT: vandnps %ymm3, %ymm5, %ymm3
; AVX-NEXT: vorps %ymm3, %ymm1, %ymm1
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm0 # 16-byte Folded Reload
-; AVX-NEXT: # xmm0 = mem[0,1,2,3],xmm6[4,5],mem[6,7]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [u,u,u,u,8,9,2,3,12,13,12,13,12,13,12,13]
-; AVX-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0,1,2,3],xmm4[4,5],xmm2[6,7]
+; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,8,9,2,3,12,13,12,13,12,13,12,13]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm11[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[2,1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3,4,5,6,7]
-; AVX-NEXT: vpblendw $243, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = mem[0,1],xmm7[2,3],mem[4,5,6,7]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,6,7,4,5,6,7,8,9,6,7,0,1,10,11]
-; AVX-NEXT: vpshufb %xmm7, %xmm3, %xmm3
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm4[2,3],xmm5[4,5,6,7]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,6,7,4,5,6,7,8,9,6,7,0,1,10,11]
+; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm3[5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm2[0,1],xmm12[2,3],xmm2[4,5,6,7]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm8 = [2,3,2,3,2,3,2,3,12,13,6,7,12,13,14,15]
-; AVX-NEXT: vpshufb %xmm8, %xmm3, %xmm3
-; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm4[0,1,1,3]
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm10[0,1],xmm7[2,3],xmm10[4,5,6,7]
+; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[2,3,2,3,2,3,2,3,12,13,6,7,12,13,14,15]
+; AVX-NEXT: vmovdqa %xmm15, %xmm2
+; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm15[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,7,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX-NEXT: vpunpckhdq {{.*#+}} xmm9 = xmm9[2],xmm11[2],xmm9[3],xmm11[3]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX-NEXT: vpunpckhdq {{.*#+}} xmm9 = xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm9[0,1,2],xmm3[3,4,5],xmm9[6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm12[0,1,2,0]
+; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm14[0,1,2,0]
; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,6,5]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5],xmm9[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX-NEXT: vmovdqa %xmm5, %xmm9
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm5[0,1],xmm6[2,3],xmm5[4,5,6,7]
-; AVX-NEXT: vpshufb %xmm7, %xmm0, %xmm4
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX-NEXT: vmovdqa %xmm15, %xmm0
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm15[0,1,2,3],xmm5[4,5],xmm15[6,7]
-; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm1
-; AVX-NEXT: vmovdqa %xmm10, %xmm2
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm10[3,1,2,3]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm6[0,1],xmm15[2,3],xmm6[4,5,6,7]
+; AVX-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm7[0,1,2,3],xmm12[4,5],xmm7[6,7]
+; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm3[u,u,u,u,8,9,2,3,12,13,12,13,12,13,12,13]
+; AVX-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = mem[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[2,1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3,4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm1[0,1,2,3,4],xmm4[5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm15[0,1],xmm13[2,3],xmm15[4,5,6,7]
-; AVX-NEXT: vpshufb %xmm8, %xmm1, %xmm1
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm4[0,1,1,3]
-; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4],xmm0[5,6,7]
+; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm1 # 16-byte Folded Reload
+; AVX-NEXT: # xmm1 = xmm8[0,1],mem[2,3],xmm8[4,5,6,7]
+; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[2,3,2,3,2,3,2,3,12,13,6,7,12,13,14,15]
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX-NEXT: vpunpckhdq {{.*#+}} xmm3 = xmm3[2],xmm8[2],xmm3[3],xmm8[3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm8[0,1,1,3]
+; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX-NEXT: vpunpckhdq {{.*#+}} xmm3 = xmm3[2],xmm14[2],xmm3[3],xmm14[3]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3,4,5],xmm3[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm14[0,1,2,0]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm13[0,1,2,0]
; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,6,5]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5],xmm3[6,7]
-; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm1, %ymm1
-; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm0[2,3],xmm5[4,5],xmm0[6,7]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [u,u,0,1,10,11,4,5,14,15,14,15,14,15,14,15]
-; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm0
-; AVX-NEXT: vpsrlq $48, %xmm2, %xmm3
+; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm12[0,1],xmm13[2,3],xmm12[4,5],xmm13[6,7]
+; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,0,1,10,11,4,5,14,15,14,15,14,15,14,15]
+; AVX-NEXT: vpsrlq $48, %xmm11, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0],xmm0[1,2,3,4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm9[0,1,2,3],xmm6[4,5],xmm9[6,7]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,4,5,6,7,8,9,8,9,2,3,12,13]
-; AVX-NEXT: vpshufb %xmm5, %xmm3, %xmm3
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1,2,3],xmm4[4,5],xmm5[6,7]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,6,7,8,9,8,9,2,3,12,13]
+; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm3[5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm8[0,1],xmm4[2,3],xmm8[4,5,6,7]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm6 = [6,7,0,1,10,11,10,11,8,9,10,11,12,13,14,15]
-; AVX-NEXT: vpshufb %xmm6, %xmm3, %xmm3
-; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm15[0,1,2,3],xmm13[4,5],xmm15[6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm10[0,1],xmm2[2,3],xmm10[4,5,6,7]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm11 = [6,7,0,1,10,11,10,11,8,9,10,11,12,13,14,15]
+; AVX-NEXT: vpshufb %xmm11, %xmm3, %xmm3
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm9 # 16-byte Folded Reload
+; AVX-NEXT: # xmm9 = xmm2[0,1,2,3],mem[4,5],xmm2[6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[2,2,2,2,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,7,4,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2],xmm9[3,4,5],xmm3[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm14[0,1,0,3]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm4[0,1,0,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,5,6]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5],xmm9[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm6[0,1,2,3],xmm15[4,5],xmm6[6,7]
+; AVX-NEXT: vmovdqa %xmm6, %xmm15
+; AVX-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm8[0,1,2,3],xmm9[4,5],xmm8[6,7]
-; AVX-NEXT: vpshufb %xmm5, %xmm0, %xmm0
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm4[2,3],xmm5[4,5],xmm4[6,7]
-; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm1
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
-; AVX-NEXT: vpsrlq $48, %xmm2, %xmm3
-; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1,2,3,4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm1[0,1,2,3,4],xmm0[5,6,7]
-; AVX-NEXT: vmovdqa %xmm11, %xmm0
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm9[0,1],xmm7[2,3],xmm9[4,5],xmm7[6,7]
+; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm3[u,u,0,1,10,11,4,5,14,15,14,15,14,15,14,15]
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm11[0,1],xmm10[2,3],xmm11[4,5,6,7]
-; AVX-NEXT: vpshufb %xmm6, %xmm1, %xmm1
+; AVX-NEXT: vpsrlq $48, %xmm10, %xmm3
+; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1,2,3,4,5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4],xmm0[5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm14[0,1],xmm8[2,3],xmm14[4,5,6,7]
+; AVX-NEXT: vmovdqa %xmm8, %xmm14
+; AVX-NEXT: vpshufb %xmm11, %xmm1, %xmm1
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm6[0,1,2,3],xmm11[4,5],xmm6[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm7[0,1,2,3],xmm11[4,5],xmm7[6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[2,2,2,2,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,4,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm3[3,4,5],xmm1[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm12[0,1,0,3]
+; AVX-NEXT: vmovdqa (%rsp), %xmm8 # 16-byte Reload
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm8[0,1,0,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,5,6]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5],xmm3[6,7]
-; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm1, %ymm7
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm8[3,1,2,3]
+; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm5[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,2,1,4,5,6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm9[0,2,2,3]
+; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = mem[0,2,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,1,0,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm5[0,3,2,3]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2],xmm4[3],xmm3[4,5,6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[2,3,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm12[0,3,2,3]
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2],xmm13[3],xmm3[4,5,6,7]
+; AVX-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
+; AVX-NEXT: # xmm5 = mem[2,3,2,3]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3],xmm1[4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm0[0,1,2,3],xmm10[4,5],xmm0[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm6[1,1,1,1]
-; AVX-NEXT: vpshufd {{.*#+}} xmm8 = xmm11[0,2,2,3]
-; AVX-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,1,0,3,4,5,6,7]
-; AVX-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm8[0],xmm5[0],xmm8[1],xmm5[1]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm8 = [8,9,2,3,12,13,12,13,8,9,12,13,12,13,14,15]
-; AVX-NEXT: vpshufb %xmm8, %xmm3, %xmm3
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = xmm3[0,1,2,3],mem[4,5],xmm3[6,7]
+; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[1,1,1,1]
+; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
+; AVX-NEXT: # xmm6 = mem[0,2,2,3]
+; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,1,0,3,4,5,6,7]
+; AVX-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,2,3,12,13,12,13,8,9,12,13,12,13,14,15]
+; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2],xmm5[3,4,5],xmm3[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm12[0,1,1,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,4,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5],xmm5[6,7]
-; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm3, %ymm1
-; AVX-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = mem[3,1,2,3]
-; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,1,2,1,4,5,6,7]
-; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = mem[0,2,2,3]
+; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm3, %ymm6
+; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm15[3,1,2,3]
+; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,2,1,4,5,6,7]
+; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = mem[0,2,2,3]
+; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,1,0,3,4,5,6,7]
+; AVX-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm9[0,3,2,3]
+; AVX-NEXT: vpblendw $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = xmm3[0,1,2],mem[3],xmm3[4,5,6,7]
+; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm10[2,3,2,3]
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1,2,3,4,5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3],xmm1[4,5,6,7]
+; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = mem[0,1,2,3],xmm14[4,5],mem[6,7]
+; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm3
+; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm7[1,1,1,1]
+; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm11[0,2,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,0,3,4,5,6,7]
-; AVX-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
-; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = mem[0,3,2,3]
-; AVX-NEXT: vpblendw $8, (%rsp), %xmm5, %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = xmm5[0,1,2],mem[3],xmm5[4,5,6,7]
-; AVX-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
-; AVX-NEXT: # xmm9 = mem[2,3,2,3]
-; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm9[0],xmm5[1,2,3,4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1,2,3],xmm3[4,5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = xmm0[0,1,2,3],mem[4,5],xmm0[6,7]
-; AVX-NEXT: vpshufb %xmm8, %xmm5, %xmm5
-; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm15[1,1,1,1]
-; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm13[0,2,2,3]
-; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,1,0,3,4,5,6,7]
-; AVX-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1]
-; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm5[0,1,2],xmm2[3,4,5],xmm5[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm14[0,1,1,3]
-; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,5,4,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5],xmm4[6,7]
-; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX-NEXT: vmovaps %ymm3, 32(%rsi)
-; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX-NEXT: vmovaps %ymm3, (%rsi)
-; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX-NEXT: vmovaps %ymm3, 32(%rdx)
-; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX-NEXT: vmovaps %ymm3, (%rdx)
-; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX-NEXT: vmovaps %ymm0, 32(%rcx)
-; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX-NEXT: vmovaps %ymm0, (%rcx)
-; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]
+; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3,4,5],xmm3[6,7]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm8[0,1,1,3]
+; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5],xmm3[6,7]
+; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
+; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX-NEXT: vmovaps %ymm2, 32(%rsi)
+; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX-NEXT: vmovaps %ymm2, (%rsi)
+; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX-NEXT: vmovaps %ymm2, 32(%rdx)
+; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX-NEXT: vmovaps %ymm2, (%rdx)
+; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX-NEXT: vmovaps %ymm2, 32(%rcx)
+; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX-NEXT: vmovaps %ymm2, (%rcx)
; AVX-NEXT: vmovaps %ymm0, 32(%r8)
-; AVX-NEXT: vmovaps %ymm7, (%r8)
-; AVX-NEXT: vmovaps %ymm2, 32(%r9)
-; AVX-NEXT: vmovaps %ymm1, (%r9)
+; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX-NEXT: vmovaps %ymm0, (%r8)
+; AVX-NEXT: vmovaps %ymm1, 32(%r9)
+; AVX-NEXT: vmovaps %ymm6, (%r9)
; AVX-NEXT: addq $424, %rsp # imm = 0x1A8
; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
@@ -5862,14 +5859,13 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: subq $1032, %rsp # imm = 0x408
; AVX-NEXT: vmovdqa 304(%rdi), %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa 288(%rdi), %xmm1
-; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
+; AVX-NEXT: vmovdqa 288(%rdi), %xmm11
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm11[0,1],xmm0[2,3],xmm11[4,5,6,7]
; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [6,7,2,3,4,5,6,7,6,7,2,3,12,13,6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovdqa 256(%rdi), %xmm2
-; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,1,1,3]
+; AVX-NEXT: vmovdqa 256(%rdi), %xmm13
+; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm13[0,1,1,3]
+; AVX-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
; AVX-NEXT: vmovdqa 272(%rdi), %xmm15
; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm15[1]
@@ -5883,16 +5879,15 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vmovdqa 208(%rdi), %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
-; AVX-NEXT: vmovdqa 192(%rdi), %xmm3
-; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm3[4],xmm0[5,6,7]
-; AVX-NEXT: vmovdqa 176(%rdi), %xmm3
-; AVX-NEXT: vmovdqa %xmm3, (%rsp) # 16-byte Spill
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[3,1,2,3]
+; AVX-NEXT: vmovdqa 192(%rdi), %xmm8
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm8[4],xmm0[5,6,7]
+; AVX-NEXT: vmovdqa %xmm8, (%rsp) # 16-byte Spill
+; AVX-NEXT: vmovdqa 176(%rdi), %xmm7
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm7[3,1,2,3]
+; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[2,1,2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa 160(%rdi), %xmm4
-; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,2,2,3]
+; AVX-NEXT: vmovdqa 160(%rdi), %xmm12
+; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm12[0,2,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,3,2,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3],xmm0[4,5,6,7]
@@ -5909,9 +5904,9 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
-; AVX-NEXT: vmovdqa 592(%rdi), %xmm12
-; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm12[1]
-; AVX-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 592(%rdi), %xmm0
+; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; AVX-NEXT: vmovdqa 560(%rdi), %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,2,2,3]
@@ -5928,43 +5923,43 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[2,1,2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa 480(%rdi), %xmm9
-; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm9[0,2,2,3]
-; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 480(%rdi), %xmm10
+; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm10[0,2,2,3]
+; AVX-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,3,2,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
; AVX-NEXT: vmovdqa 528(%rdi), %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[0,1,0,3]
-; AVX-NEXT: vmovdqa 512(%rdi), %xmm13
-; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1,2,3],xmm13[4],xmm4[5,6,7]
-; AVX-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 512(%rdi), %xmm0
+; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1,2,3],xmm0[4],xmm4[5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT: vandps %ymm5, %ymm3, %ymm3
-; AVX-NEXT: vmovaps 544(%rdi), %xmm11
-; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm11[0,1,0,1]
-; AVX-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovaps 544(%rdi), %xmm0
+; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm0[0,1,0,1]
; AVX-NEXT: vandnps %ymm4, %ymm5, %ymm4
; AVX-NEXT: vorps %ymm4, %ymm3, %ymm3
; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa 96(%rdi), %xmm10
-; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm10[0,1,1,3]
-; AVX-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
-; AVX-NEXT: vmovdqa 112(%rdi), %xmm0
-; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
-; AVX-NEXT: vmovdqa 80(%rdi), %xmm0
+; AVX-NEXT: vmovdqa 96(%rdi), %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,2,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[0,1,1,3]
+; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
+; AVX-NEXT: vmovdqa 112(%rdi), %xmm6
+; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm6[1]
+; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 80(%rdi), %xmm9
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm9[0,2,2,3]
+; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,3,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa 144(%rdi), %xmm7
-; AVX-NEXT: vmovdqa 128(%rdi), %xmm6
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm6[0,1],xmm7[2,3],xmm6[4,5,6,7]
-; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 144(%rdi), %xmm3
+; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 128(%rdi), %xmm0
+; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm0[0,1],xmm3[2,3],xmm0[4,5,6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3,4],xmm3[5,6,7]
; AVX-NEXT: vmovdqa 16(%rdi), %xmm0
@@ -5991,11 +5986,11 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vorps %ymm4, %ymm3, %ymm3
; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa 464(%rdi), %xmm8
+; AVX-NEXT: vmovdqa 464(%rdi), %xmm2
+; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vmovdqa 448(%rdi), %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm0[0,1],xmm8[2,3],xmm0[4,5,6,7]
-; AVX-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm0[0,1],xmm2[2,3],xmm0[4,5,6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm2, %xmm1
; AVX-NEXT: vmovdqa 416(%rdi), %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -6034,102 +6029,104 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vorps %ymm3, %ymm2, %ymm2
; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm1 # 16-byte Folded Reload
-; AVX-NEXT: # xmm1 = xmm15[0,1],mem[2,3],xmm15[4,5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm15[0,1],xmm13[2,3],xmm15[4,5,6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,0,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
-; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
-; AVX-NEXT: # xmm2 = mem[0,3,2,3]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm13[0,3,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[1,2,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = mem[0,1,2,3],xmm0[4,5],mem[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm11[0,1,2,3],xmm4[4,5],xmm11[6,7]
; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,4,5,14,15,8,9]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm1[0,1,2,3,4],xmm3[5,6,7]
-; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
-; AVX-NEXT: # xmm1 = mem[0,3,2,3]
+; AVX-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm12[0,3,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,2,2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa (%rsp), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpsrlq $48, %xmm0, %xmm15
+; AVX-NEXT: vpsrlq $48, %xmm7, %xmm15
; AVX-NEXT: vpunpckldq {{.*#+}} xmm15 = xmm1[0],xmm15[0],xmm1[1],xmm15[1]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpblendw $204, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3],xmm0[4,5],mem[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm8[0,1],xmm14[2,3],xmm8[4,5],xmm14[6,7]
; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,0,1,0,1,0,1,10,11,4,5,14,15,6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm15[0,1,2],xmm0[3,4,5,6,7]
; AVX-NEXT: vandps %ymm5, %ymm0, %ymm0
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX-NEXT: vpsllq $48, %xmm4, %xmm15
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; AVX-NEXT: vpsllq $48, %xmm8, %xmm15
; AVX-NEXT: vandnps %ymm15, %ymm5, %ymm15
; AVX-NEXT: vorps %ymm0, %ymm15, %ymm0
; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm0 # 16-byte Folded Reload
-; AVX-NEXT: # xmm0 = xmm12[0,1],mem[2,3],xmm12[4,5,6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3],xmm0[4,5,6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,0,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; AVX-NEXT: # xmm3 = mem[0,3,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[1,2,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = mem[0,1,2,3],xmm14[4,5],mem[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = mem[0,1,2,3],xmm3[4,5],mem[6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm3[5,6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm9[0,3,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm10[0,3,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[1,2,2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX-NEXT: vpsrlq $48, %xmm4, %xmm15
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX-NEXT: vpsrlq $48, %xmm7, %xmm15
; AVX-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],xmm15[0],xmm3[1],xmm15[1]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm15 = xmm13[0,1],xmm9[2,3],xmm13[4,5],xmm9[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX-NEXT: vpblendw $204, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm15 # 16-byte Folded Reload
+; AVX-NEXT: # xmm15 = xmm7[0,1],mem[2,3],xmm7[4,5],mem[6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm15, %xmm15
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2],xmm15[3,4,5,6,7]
; AVX-NEXT: vandps %ymm5, %ymm3, %ymm3
-; AVX-NEXT: vpsllq $48, %xmm11, %xmm15
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX-NEXT: vpsllq $48, %xmm10, %xmm15
; AVX-NEXT: vandnps %ymm15, %ymm5, %ymm15
; AVX-NEXT: vorps %ymm3, %ymm15, %ymm3
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm13[0,1],xmm10[2,3],xmm13[4,5,6,7]
+; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT: # xmm0 = xmm6[0,1],mem[2,3],xmm6[4,5,6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,0,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm12[0,3,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm9[0,3,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[1,2,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3,4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm6[0,1,2,3],xmm7[4,5],xmm6[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = xmm3[0,1,2,3],mem[4,5],xmm3[6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm3[5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm11[0,3,2,3]
+; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = mem[0,3,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[1,2,2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX-NEXT: vpsrlq $48, %xmm10, %xmm15
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX-NEXT: vpsrlq $48, %xmm6, %xmm15
; AVX-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],xmm15[0],xmm3[1],xmm15[1]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm15 = xmm4[0,1],xmm6[2,3],xmm4[4,5],xmm6[6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm15 = xmm9[0,1],xmm6[2,3],xmm9[4,5],xmm6[6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm15, %xmm15
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2],xmm15[3,4,5,6,7]
; AVX-NEXT: vandps %ymm5, %ymm3, %ymm3
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX-NEXT: vpsllq $48, %xmm7, %xmm15
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
+; AVX-NEXT: vpsllq $48, %xmm15, %xmm15
; AVX-NEXT: vandnps %ymm15, %ymm5, %ymm15
; AVX-NEXT: vorps %ymm3, %ymm15, %ymm3
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0
; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm0 # 16-byte Folded Reload
-; AVX-NEXT: # xmm0 = mem[0,1,2,3],xmm8[4,5],mem[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT: # xmm0 = xmm0[0,1,2,3],mem[4,5],xmm0[6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX-NEXT: vpblendw $243, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm2 # 16-byte Folded Reload
-; AVX-NEXT: # xmm2 = mem[0,1],xmm8[2,3],mem[4,5,6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
+; AVX-NEXT: # xmm2 = xmm2[0,1],mem[2,3],xmm2[4,5,6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,0,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5]
; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
@@ -6160,64 +6157,60 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: # xmm1 = mem[0,1,2,3],xmm0[4,5],mem[6,7]
; AVX-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,u,u,8,9,2,3,12,13,12,13,12,13,12,13]
; AVX-NEXT: vpshufb %xmm0, %xmm1, %xmm1
-; AVX-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
-; AVX-NEXT: # xmm2 = mem[3,1,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm13[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[2,1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
-; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = xmm2[0,1],mem[2,3],xmm2[4,5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0,1],xmm11[2,3],xmm4[4,5,6,7]
; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,6,7,4,5,6,7,8,9,6,7,0,1,10,11]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm1[0,1,2,3,4],xmm3[5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT: vpblendw $243, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = mem[0,1],xmm1[2,3],mem[4,5,6,7]
+; AVX-NEXT: vpblendw $243, (%rsp), %xmm14, %xmm5 # 16-byte Folded Reload
+; AVX-NEXT: # xmm5 = mem[0,1],xmm14[2,3],mem[4,5,6,7]
; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [2,3,2,3,2,3,2,3,12,13,6,7,12,13,14,15]
; AVX-NEXT: vpshufb %xmm1, %xmm5, %xmm5
-; AVX-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
-; AVX-NEXT: # xmm15 = mem[0,1,1,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm12[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,7,6,7]
-; AVX-NEXT: vpunpckhdq (%rsp), %xmm15, %xmm15 # 16-byte Folded Reload
-; AVX-NEXT: # xmm15 = xmm15[2],mem[2],xmm15[3],mem[3]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX-NEXT: vpunpckhdq {{.*#+}} xmm15 = xmm15[2],xmm4[2],xmm15[3],xmm4[3]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm15[0,1,2],xmm5[3,4,5],xmm15[6,7]
-; AVX-NEXT: vpshufd $36, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
-; AVX-NEXT: # xmm15 = mem[0,1,2,0]
+; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm8[0,1,2,0]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,5,6,5]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5],xmm15[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3
; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = mem[0,1,2,3],xmm3[4,5],mem[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm13[0,1,2,3],xmm11[4,5],xmm13[6,7]
; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm3
; AVX-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; AVX-NEXT: # xmm5 = mem[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[2,1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2,3,4,5,6,7]
-; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = xmm14[0,1],mem[2,3],xmm14[4,5,6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
+; AVX-NEXT: vpblendw $243, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm5 # 16-byte Folded Reload
+; AVX-NEXT: # xmm5 = mem[0,1],xmm12[2,3],mem[4,5,6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm5[5,6,7]
-; AVX-NEXT: vpblendw $243, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = mem[0,1],xmm9[2,3],mem[4,5,6,7]
+; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm5 # 16-byte Folded Reload
+; AVX-NEXT: # xmm5 = xmm7[0,1],mem[2,3],xmm7[4,5,6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm5, %xmm5
-; AVX-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
-; AVX-NEXT: # xmm15 = mem[0,1,1,3]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm8[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,7,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX-NEXT: vpunpckhdq {{.*#+}} xmm15 = xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX-NEXT: vpunpckhdq {{.*#+}} xmm15 = xmm15[2],xmm7[2],xmm15[3],xmm7[3]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm15[0,1,2],xmm5[3,4,5],xmm15[6,7]
-; AVX-NEXT: vpshufd $36, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
-; AVX-NEXT: # xmm15 = mem[0,1,2,0]
+; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm10[0,1,2,0]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,5,6,5]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5],xmm15[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3
; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = xmm13[0,1,2,3],mem[4,5],xmm13[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm10[0,1,2,3],xmm14[4,5],xmm10[6,7]
; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm3
-; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm12[3,1,2,3]
+; AVX-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
+; AVX-NEXT: # xmm5 = mem[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[2,1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2,3,4,5,6,7]
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
@@ -6225,41 +6218,45 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: # xmm5 = mem[0,1],xmm5[2,3],mem[4,5,6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm5[5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm4[0,1],xmm6[2,3],xmm4[4,5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm9[0,1],xmm6[2,3],xmm9[4,5,6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm5, %xmm5
-; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm11[0,1,1,3]
+; AVX-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX-NEXT: # xmm15 = mem[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,7,6,7]
-; AVX-NEXT: vpunpckhdq {{.*#+}} xmm15 = xmm15[2],xmm10[2],xmm15[3],xmm10[3]
+; AVX-NEXT: vpunpckhdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
+; AVX-NEXT: # xmm15 = xmm15[2],mem[2],xmm15[3],mem[3]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm15[0,1,2],xmm5[3,4,5],xmm15[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm7[0,1,2,0]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm9[0,1,2,0]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,5,6,5]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5],xmm15[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3
; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm12[0,1],xmm13[2,3],xmm12[4,5,6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX-NEXT: vpblendw $243, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = mem[0,1],xmm3[2,3],mem[4,5,6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm2
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm7[0,1,2,3],xmm8[4,5],xmm7[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = xmm3[0,1,2,3],mem[4,5],xmm3[6,7]
; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm0
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm8[3,1,2,3]
+; AVX-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = mem[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[2,1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm2[5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm10[0,1],xmm11[2,3],xmm10[4,5,6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
+; AVX-NEXT: # xmm2 = xmm2[0,1],mem[2,3],xmm2[4,5,6,7]
; AVX-NEXT: vpshufb %xmm1, %xmm2, %xmm1
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm6[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]
+; AVX-NEXT: vpunpckhdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
+; AVX-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3,4,5],xmm2[6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm9[0,1,2,0]
+; AVX-NEXT: vpshufd $36, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
+; AVX-NEXT: # xmm2 = mem[0,1,2,0]
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,6,5]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5],xmm2[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
@@ -6267,7 +6264,8 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX-NEXT: vpblendw $204, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload
; AVX-NEXT: # xmm1 = xmm0[0,1],mem[2,3],xmm0[4,5],mem[6,7]
-; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,0,1,10,11,4,5,14,15,14,15,14,15,14,15]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,0,1,10,11,4,5,14,15,14,15,14,15,14,15]
+; AVX-NEXT: vpshufb %xmm0, %xmm1, %xmm1
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT: vpsrlq $48, %xmm2, %xmm2
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1,2,3,4,5,6,7]
@@ -6277,13 +6275,13 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,8,9,2,3,12,13]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm1[0,1,2,3,4],xmm3[5,6,7]
-; AVX-NEXT: vmovdqa (%rsp), %xmm1 # 16-byte Reload
-; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = xmm1[0,1],mem[2,3],xmm1[4,5,6,7]
-; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[6,7,0,1,10,11,10,11,8,9,10,11,12,13,14,15]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm15 # 16-byte Folded Reload
-; AVX-NEXT: # xmm15 = xmm0[0,1,2,3],mem[4,5],xmm0[6,7]
+; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm5 # 16-byte Folded Reload
+; AVX-NEXT: # xmm5 = xmm4[0,1],mem[2,3],xmm4[4,5,6,7]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [6,7,0,1,10,11,10,11,8,9,10,11,12,13,14,15]
+; AVX-NEXT: vpshufb %xmm1, %xmm5, %xmm5
+; AVX-NEXT: vmovdqa (%rsp), %xmm4 # 16-byte Reload
+; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm15 # 16-byte Folded Reload
+; AVX-NEXT: # xmm15 = xmm4[0,1,2,3],mem[4,5],xmm4[6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[2,2,2,2,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,7,4,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm15[3,4,5],xmm5[6,7]
@@ -6293,47 +6291,46 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5],xmm15[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3
; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpblendw $204, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = xmm0[0,1],mem[2,3],xmm0[4,5],mem[6,7]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [u,u,0,1,10,11,4,5,14,15,14,15,14,15,14,15]
-; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm3
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpsrlq $48, %xmm0, %xmm5
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm11[0,1],xmm13[2,3],xmm11[4,5],xmm13[6,7]
+; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX-NEXT: vpsrlq $48, %xmm4, %xmm5
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1,2,3,4,5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = xmm0[0,1,2,3],mem[4,5],xmm0[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm13[0,1,2,3],xmm12[4,5],xmm13[6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm5[5,6,7]
-; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = xmm14[0,1],mem[2,3],xmm14[4,5,6,7]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm0 = [6,7,0,1,10,11,10,11,8,9,10,11,12,13,14,15]
-; AVX-NEXT: vpshufb %xmm0, %xmm5, %xmm5
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm15 # 16-byte Folded Reload
-; AVX-NEXT: # xmm15 = mem[0,1,2,3],xmm14[4,5],mem[6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm7[0,1],xmm8[2,3],xmm7[4,5,6,7]
+; AVX-NEXT: vpshufb %xmm1, %xmm5, %xmm5
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm15 = xmm7[0,1,2,3],xmm8[4,5],xmm7[6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[2,2,2,2,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,7,4,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm15[3,4,5],xmm5[6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm14[0,1,0,3]
+; AVX-NEXT: vpshufd $196, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX-NEXT: # xmm15 = mem[0,1,0,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,5,5,6]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5],xmm15[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3
; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vpblendw $51, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = mem[0,1],xmm7[2,3],mem[4,5],xmm7[6,7]
-; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm3
-; AVX-NEXT: vpsrlq $48, %xmm8, %xmm5
+; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm14[0,1],xmm10[2,3],xmm14[4,5],xmm10[6,7]
+; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX-NEXT: vpsrlq $48, %xmm4, %xmm5
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1,2,3,4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm12[0,1,2,3],xmm13[4,5],xmm12[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm4[0,1,2,3],xmm14[4,5],xmm4[6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm5[5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm4[0,1],xmm6[2,3],xmm4[4,5,6,7]
-; AVX-NEXT: vpshufb %xmm0, %xmm5, %xmm5
-; AVX-NEXT: vmovdqa %xmm0, %xmm4
-; AVX-NEXT: vpblendw {{.*#+}} xmm15 = xmm10[0,1,2,3],xmm11[4,5],xmm10[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX-NEXT: vpblendw $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm5 # 16-byte Folded Reload
+; AVX-NEXT: # xmm5 = xmm10[0,1],mem[2,3],xmm10[4,5,6,7]
+; AVX-NEXT: vpshufb %xmm1, %xmm5, %xmm5
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm15 # 16-byte Folded Reload
+; AVX-NEXT: # xmm15 = mem[0,1,2,3],xmm11[4,5],mem[6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[2,2,2,2,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,7,4,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm15[3,4,5],xmm5[6,7]
@@ -6342,30 +6339,29 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5],xmm15[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3
; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm3 # 16-byte Folded Reload
-; AVX-NEXT: # xmm3 = xmm13[0,1,2,3],mem[4,5],xmm13[6,7]
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = xmm9[0,1,2,3],mem[4,5],xmm9[6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm2
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm10[0,1],xmm11[2,3],xmm10[4,5],xmm11[6,7]
-; AVX-NEXT: vpshufb %xmm1, %xmm3, %xmm0
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX-NEXT: vpsrlq $48, %xmm7, %xmm3
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX-NEXT: vpblendw $204, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX-NEXT: # xmm3 = xmm3[0,1],mem[2,3],xmm3[4,5],mem[6,7]
+; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm0
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX-NEXT: vpsrlq $48, %xmm3, %xmm3
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0],xmm0[1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm2[5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm6[0,1],xmm3[2,3],xmm6[4,5,6,7]
-; AVX-NEXT: vpshufb %xmm4, %xmm2, %xmm1
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0,1,2,3],xmm8[4,5],xmm4[6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm6[2,3],xmm3[4,5,6,7]
+; AVX-NEXT: vpshufb %xmm1, %xmm2, %xmm1
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX-NEXT: vpblendw $207, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm2 # 16-byte Folded Reload
+; AVX-NEXT: # xmm2 = mem[0,1,2,3],xmm6[4,5],mem[6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[2,2,2,2,4,5,6,7]
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,4,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm2[3,4,5],xmm1[6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm9[0,1,0,3]
+; AVX-NEXT: vpshufd $196, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
+; AVX-NEXT: # xmm2 = mem[0,1,0,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,5,6]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5],xmm2[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
@@ -6385,10 +6381,10 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: # xmm2 = mem[2,3,2,3]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
-; AVX-NEXT: vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT: # xmm1 = xmm1[0,1,2,3],mem[4,5],xmm1[6,7]
-; AVX-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
+; AVX-NEXT: vpshufd $85, (%rsp), %xmm2 # 16-byte Folded Reload
; AVX-NEXT: # xmm2 = mem[1,1,1,1]
; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; AVX-NEXT: # xmm5 = mem[0,2,2,3]
@@ -6402,11 +6398,9 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,4,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5],xmm5[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm1
-; AVX-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
-; AVX-NEXT: # xmm0 = mem[3,1,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm13[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,2,1,4,5,6,7]
-; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = mem[0,2,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm12[0,2,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,0,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]
; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
@@ -6420,63 +6414,64 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
; AVX-NEXT: # xmm5 = xmm5[0,1,2,3],mem[4,5],xmm5[6,7]
-; AVX-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
-; AVX-NEXT: # xmm15 = mem[1,1,1,1]
-; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
-; AVX-NEXT: # xmm12 = mem[0,2,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm7[1,1,1,1]
+; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm8[0,2,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,1,0,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm12 = xmm12[0],xmm15[0],xmm12[1],xmm15[1]
; AVX-NEXT: vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm12[3,4,5],xmm5[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm14[0,1,1,3]
+; AVX-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
+; AVX-NEXT: # xmm12 = mem[0,1,1,3]
; AVX-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,5,4,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5],xmm12[6,7]
; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm5, %ymm0
-; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm13[3,1,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,2,1,4,5,6,7]
-; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
-; AVX-NEXT: # xmm12 = mem[0,2,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm14[0,2,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,1,0,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm12[0],xmm5[0],xmm12[1],xmm5[1]
-; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm10[0,3,2,3]
-; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm12[0,1,2],xmm11[3],xmm12[4,5,6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm7[2,3,2,3]
+; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
+; AVX-NEXT: # xmm12 = mem[0,3,2,3]
+; AVX-NEXT: vpblendw $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload
+; AVX-NEXT: # xmm12 = xmm12[0,1,2],mem[3],xmm12[4,5,6,7]
+; AVX-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX-NEXT: # xmm15 = mem[2,3,2,3]
; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm15[0],xmm12[1,2,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm12[0,1,2,3],xmm5[4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm6[0,1,2,3],xmm3[4,5],xmm6[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm4[1,1,1,1]
-; AVX-NEXT: vpshufd {{.*#+}} xmm14 = xmm8[0,2,2,3]
-; AVX-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,1,0,3,4,5,6,7]
-; AVX-NEXT: vpunpckldq {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1]
+; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm12 # 16-byte Folded Reload
+; AVX-NEXT: # xmm12 = xmm10[0,1,2,3],mem[4,5],xmm10[6,7]
+; AVX-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX-NEXT: # xmm15 = mem[1,1,1,1]
+; AVX-NEXT: vpshufd {{.*#+}} xmm13 = xmm11[0,2,2,3]
+; AVX-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,1,0,3,4,5,6,7]
+; AVX-NEXT: vpunpckldq {{.*#+}} xmm13 = xmm13[0],xmm15[0],xmm13[1],xmm15[1]
; AVX-NEXT: vpshufb %xmm2, %xmm12, %xmm12
-; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm12[0,1,2],xmm14[3,4,5],xmm12[6,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm14 = xmm9[0,1,1,3]
-; AVX-NEXT: vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,5,4,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,5],xmm14[6,7]
-; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm12, %ymm12
-; AVX-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
-; AVX-NEXT: # xmm5 = mem[3,1,2,3]
+; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm12[0,1,2],xmm13[3,4,5],xmm12[6,7]
+; AVX-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
+; AVX-NEXT: # xmm10 = mem[0,1,1,3]
+; AVX-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,5,4,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm12[0,1,2,3,4,5],xmm10[6,7]
+; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm10, %ymm10
+; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm9[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,2,1,4,5,6,7]
-; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload
-; AVX-NEXT: # xmm13 = mem[0,2,2,3]
-; AVX-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,1,0,3,4,5,6,7]
-; AVX-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm13[0],xmm5[0],xmm13[1],xmm5[1]
-; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload
-; AVX-NEXT: # xmm13 = mem[0,3,2,3]
-; AVX-NEXT: vpblendw $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm11 # 16-byte Folded Reload
-; AVX-NEXT: # xmm11 = xmm13[0,1,2],mem[3],xmm13[4,5,6,7]
-; AVX-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
-; AVX-NEXT: # xmm10 = mem[2,3,2,3]
-; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm10[0],xmm11[1,2,3,4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm10[0,1,2,3],xmm5[4,5,6,7]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
+; AVX-NEXT: # xmm12 = mem[0,2,2,3]
+; AVX-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,1,0,3,4,5,6,7]
+; AVX-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm12[0],xmm5[0],xmm12[1],xmm5[1]
+; AVX-NEXT: vpshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
+; AVX-NEXT: # xmm12 = mem[0,3,2,3]
+; AVX-NEXT: vpblendw $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload
+; AVX-NEXT: # xmm12 = xmm12[0,1,2],mem[3],xmm12[4,5,6,7]
+; AVX-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
+; AVX-NEXT: # xmm11 = mem[2,3,2,3]
+; AVX-NEXT: vpblendw {{.*#+}} xmm11 = xmm11[0],xmm12[1,2,3,4,5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm11[0,1,2,3],xmm5[4,5,6,7]
; AVX-NEXT: vpblendw $48, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm8 # 16-byte Folded Reload
; AVX-NEXT: # xmm8 = xmm3[0,1,2,3],mem[4,5],xmm3[6,7]
; AVX-NEXT: vpshufb %xmm2, %xmm8, %xmm2
; AVX-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
; AVX-NEXT: # xmm4 = mem[1,1,1,1]
-; AVX-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
-; AVX-NEXT: # xmm7 = mem[0,2,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[0,2,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,1,0,3,4,5,6,7]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm7[0],xmm4[0],xmm7[1],xmm4[1]
; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1,2],xmm4[3,4,5],xmm2[6,7]
@@ -6518,7 +6513,7 @@ define void @load_i16_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX-NEXT: vmovaps %ymm3, 32(%r8)
; AVX-NEXT: vmovaps %ymm2, 64(%r9)
-; AVX-NEXT: vmovaps %ymm12, (%r9)
+; AVX-NEXT: vmovaps %ymm10, (%r9)
; AVX-NEXT: vmovaps %ymm0, 96(%r9)
; AVX-NEXT: vmovaps %ymm1, 32(%r9)
; AVX-NEXT: addq $1032, %rsp # imm = 0x408
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-3.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-3.ll
index d1d7cb0a34332..17484ddad70a5 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-3.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-3.ll
@@ -2110,116 +2110,116 @@ define void @load_i8_stride3_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
;
; AVX-LABEL: load_i8_stride3_vf64:
; AVX: # %bb.0:
-; AVX-NEXT: vmovdqa (%rdi), %xmm6
-; AVX-NEXT: vmovdqa 16(%rdi), %xmm0
-; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa 32(%rdi), %xmm4
-; AVX-NEXT: vmovdqa 48(%rdi), %xmm7
-; AVX-NEXT: vmovdqa 64(%rdi), %xmm2
-; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa 96(%rdi), %xmm11
-; AVX-NEXT: vmovdqa 112(%rdi), %xmm3
-; AVX-NEXT: vmovdqa 144(%rdi), %xmm10
+; AVX-NEXT: vmovdqa (%rdi), %xmm11
+; AVX-NEXT: vmovdqa 32(%rdi), %xmm2
+; AVX-NEXT: vmovdqa 48(%rdi), %xmm14
+; AVX-NEXT: vmovdqa 64(%rdi), %xmm10
+; AVX-NEXT: vmovdqa 96(%rdi), %xmm13
+; AVX-NEXT: vmovdqa 128(%rdi), %xmm4
+; AVX-NEXT: vmovdqa 144(%rdi), %xmm7
; AVX-NEXT: vmovdqa 160(%rdi), %xmm1
; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa {{.*#+}} xmm12 = [128,128,128,128,128,0,3,6,9,12,15,2,5,8,11,14]
-; AVX-NEXT: vpshufb %xmm12, %xmm6, %xmm5
-; AVX-NEXT: vpshufb %xmm12, %xmm10, %xmm8
-; AVX-NEXT: vpshufb %xmm12, %xmm11, %xmm9
-; AVX-NEXT: vpshufb %xmm12, %xmm7, %xmm13
-; AVX-NEXT: vmovdqa {{.*#+}} xmm14 = [1,4,7,10,13,128,128,128,128,128,128,u,u,u,u,u]
-; AVX-NEXT: vpshufb %xmm14, %xmm6, %xmm6
-; AVX-NEXT: vmovdqa {{.*#+}} xmm15 = [128,128,128,128,128,0,3,6,9,12,15,u,u,u,u,u]
-; AVX-NEXT: vpshufb %xmm15, %xmm0, %xmm12
-; AVX-NEXT: vpor %xmm6, %xmm12, %xmm6
-; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufb %xmm14, %xmm10, %xmm10
-; AVX-NEXT: vpshufb %xmm15, %xmm1, %xmm12
-; AVX-NEXT: vpor %xmm10, %xmm12, %xmm6
-; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufb %xmm14, %xmm11, %xmm11
-; AVX-NEXT: vpshufb %xmm15, %xmm3, %xmm12
-; AVX-NEXT: vpor %xmm11, %xmm12, %xmm6
-; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufb %xmm14, %xmm7, %xmm7
-; AVX-NEXT: vpshufb %xmm15, %xmm2, %xmm12
-; AVX-NEXT: vpor %xmm7, %xmm12, %xmm6
-; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufb %xmm14, %xmm0, %xmm12
-; AVX-NEXT: vpshufb %xmm15, %xmm4, %xmm6
-; AVX-NEXT: vpor %xmm6, %xmm12, %xmm6
-; AVX-NEXT: vmovdqa 176(%rdi), %xmm12
-; AVX-NEXT: vpshufb %xmm14, %xmm1, %xmm0
-; AVX-NEXT: vpshufb %xmm15, %xmm12, %xmm10
-; AVX-NEXT: vpor %xmm0, %xmm10, %xmm1
-; AVX-NEXT: vmovdqa 128(%rdi), %xmm10
-; AVX-NEXT: vpshufb %xmm14, %xmm3, %xmm0
-; AVX-NEXT: vpshufb %xmm15, %xmm10, %xmm11
-; AVX-NEXT: vpor %xmm0, %xmm11, %xmm0
-; AVX-NEXT: vpshufb %xmm14, %xmm2, %xmm11
-; AVX-NEXT: vmovdqa 80(%rdi), %xmm14
-; AVX-NEXT: vpshufb %xmm15, %xmm14, %xmm15
-; AVX-NEXT: vpor %xmm11, %xmm15, %xmm11
-; AVX-NEXT: vmovdqa {{.*#+}} xmm15 = [1,4,7,10,13,128,128,128,128,128,128,128,128,128,128,128]
-; AVX-NEXT: vpshufb %xmm15, %xmm14, %xmm7
-; AVX-NEXT: vpor %xmm7, %xmm13, %xmm7
-; AVX-NEXT: vpalignr {{.*#+}} xmm13 = xmm13[11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7,8,9,10]
-; AVX-NEXT: vpshufb %xmm15, %xmm10, %xmm11
-; AVX-NEXT: vpor %xmm11, %xmm9, %xmm11
-; AVX-NEXT: vpalignr {{.*#+}} xmm9 = xmm9[11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7,8,9,10]
-; AVX-NEXT: vpshufb %xmm15, %xmm12, %xmm0
-; AVX-NEXT: vpor %xmm0, %xmm8, %xmm2
-; AVX-NEXT: vpalignr {{.*#+}} xmm8 = xmm8[11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7,8,9,10]
-; AVX-NEXT: vpshufb %xmm15, %xmm4, %xmm1
+; AVX-NEXT: vmovdqa {{.*#+}} xmm3 = [1,4,7,10,13,128,128,128,128,128,128,128,128,128,128,128]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm5 = [128,128,128,128,128,0,3,6,9,12,15,2,5,8,11,14]
+; AVX-NEXT: vpshufb %xmm5, %xmm7, %xmm8
+; AVX-NEXT: vpshufb %xmm5, %xmm14, %xmm9
+; AVX-NEXT: vpshufb %xmm3, %xmm2, %xmm6
+; AVX-NEXT: vpshufb %xmm5, %xmm11, %xmm12
+; AVX-NEXT: vpor %xmm6, %xmm12, %xmm0
+; AVX-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX-NEXT: vpshufb %xmm3, %xmm4, %xmm12
+; AVX-NEXT: vpshufb %xmm5, %xmm13, %xmm5
+; AVX-NEXT: vpor %xmm5, %xmm12, %xmm6
+; AVX-NEXT: vmovdqa {{.*#+}} xmm0 = [1,4,7,10,13,128,128,128,128,128,128,u,u,u,u,u]
+; AVX-NEXT: vpshufb %xmm0, %xmm7, %xmm7
+; AVX-NEXT: vmovdqa {{.*#+}} xmm3 = [128,128,128,128,128,0,3,6,9,12,15,u,u,u,u,u]
+; AVX-NEXT: vpshufb %xmm3, %xmm1, %xmm12
+; AVX-NEXT: vpor %xmm7, %xmm12, %xmm5
+; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpshufb %xmm0, %xmm14, %xmm7
+; AVX-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpshufb %xmm3, %xmm10, %xmm12
+; AVX-NEXT: vpor %xmm7, %xmm12, %xmm5
+; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 16(%rdi), %xmm7
+; AVX-NEXT: vpshufb %xmm0, %xmm11, %xmm11
+; AVX-NEXT: vpshufb %xmm3, %xmm7, %xmm14
+; AVX-NEXT: vpor %xmm11, %xmm14, %xmm5
+; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpshufb %xmm0, %xmm13, %xmm13
+; AVX-NEXT: vmovdqa 112(%rdi), %xmm11
+; AVX-NEXT: vpshufb %xmm3, %xmm11, %xmm15
+; AVX-NEXT: vpor %xmm13, %xmm15, %xmm5
+; AVX-NEXT: vmovdqa 176(%rdi), %xmm13
+; AVX-NEXT: vpshufb %xmm0, %xmm1, %xmm15
+; AVX-NEXT: vpshufb %xmm3, %xmm13, %xmm12
+; AVX-NEXT: vpor %xmm15, %xmm12, %xmm12
+; AVX-NEXT: vmovdqa 80(%rdi), %xmm1
+; AVX-NEXT: vpshufb %xmm0, %xmm10, %xmm15
+; AVX-NEXT: vpshufb %xmm3, %xmm1, %xmm14
+; AVX-NEXT: vpor %xmm15, %xmm14, %xmm14
+; AVX-NEXT: vpshufb %xmm0, %xmm7, %xmm15
+; AVX-NEXT: vpshufb %xmm3, %xmm2, %xmm10
+; AVX-NEXT: vpor %xmm15, %xmm10, %xmm15
+; AVX-NEXT: vpshufb %xmm0, %xmm11, %xmm0
+; AVX-NEXT: vpshufb %xmm3, %xmm4, %xmm3
+; AVX-NEXT: vpor %xmm0, %xmm3, %xmm0
+; AVX-NEXT: vmovdqa {{.*#+}} xmm10 = [1,4,7,10,13,128,128,128,128,128,128,128,128,128,128,128]
+; AVX-NEXT: vpshufb %xmm10, %xmm1, %xmm3
+; AVX-NEXT: vpor %xmm3, %xmm9, %xmm3
+; AVX-NEXT: vpalignr {{.*#+}} xmm9 = xmm9[11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7,8,9,10]
+; AVX-NEXT: vpshufb %xmm10, %xmm13, %xmm10
+; AVX-NEXT: vpor %xmm10, %xmm8, %xmm10
+; AVX-NEXT: vpalignr {{.*#+}} xmm8 = xmm8[11,12,13,14,15],xmm12[0,1,2,3,4,5,6,7,8,9,10]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,10,128,128,128,128,128]
+; AVX-NEXT: vpshufb %xmm12, %xmm5, %xmm5
+; AVX-NEXT: vmovdqa {{.*#+}} xmm14 = [128,128,128,128,128,128,128,128,128,128,128,2,5,8,11,14]
+; AVX-NEXT: vpshufb %xmm14, %xmm4, %xmm4
+; AVX-NEXT: vpor %xmm4, %xmm5, %xmm4
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX-NEXT: vpshufb %xmm12, %xmm5, %xmm5
+; AVX-NEXT: vpshufb %xmm14, %xmm2, %xmm2
+; AVX-NEXT: vpor %xmm2, %xmm5, %xmm2
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX-NEXT: vpshufb %xmm12, %xmm5, %xmm5
+; AVX-NEXT: vpshufb %xmm14, %xmm1, %xmm1
; AVX-NEXT: vpor %xmm1, %xmm5, %xmm1
-; AVX-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[11,12,13,14,15],xmm6[0,1,2,3,4,5,6,7,8,9,10]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,2,3,4,5,6,7,8,9,10,128,128,128,128,128]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm6, %xmm0, %xmm15
-; AVX-NEXT: vmovdqa {{.*#+}} xmm0 = [128,128,128,128,128,128,128,128,128,128,128,2,5,8,11,14]
-; AVX-NEXT: vpshufb %xmm0, %xmm14, %xmm14
-; AVX-NEXT: vpor %xmm14, %xmm15, %xmm14
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm6, %xmm15, %xmm15
-; AVX-NEXT: vpshufb %xmm0, %xmm10, %xmm10
-; AVX-NEXT: vpor %xmm10, %xmm15, %xmm10
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm6, %xmm15, %xmm15
-; AVX-NEXT: vpshufb %xmm0, %xmm12, %xmm12
-; AVX-NEXT: vpor %xmm12, %xmm15, %xmm12
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm6, %xmm15, %xmm6
-; AVX-NEXT: vpshufb %xmm0, %xmm4, %xmm0
-; AVX-NEXT: vpor %xmm0, %xmm6, %xmm0
-; AVX-NEXT: vmovdqa {{.*#+}} xmm4 = [5,6,7,8,9,10,128,128,128,128,128,0,1,2,3,4]
-; AVX-NEXT: vpshufb %xmm4, %xmm7, %xmm6
-; AVX-NEXT: vmovdqa {{.*#+}} xmm7 = [128,128,128,128,128,128,2,5,8,11,14,128,128,128,128,128]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm7, %xmm15, %xmm15
-; AVX-NEXT: vpor %xmm6, %xmm15, %xmm6
-; AVX-NEXT: vpshufb %xmm4, %xmm11, %xmm11
-; AVX-NEXT: vpshufb %xmm7, %xmm3, %xmm15
-; AVX-NEXT: vpor %xmm15, %xmm11, %xmm11
-; AVX-NEXT: vpshufb %xmm4, %xmm2, %xmm2
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm7, %xmm3, %xmm3
-; AVX-NEXT: vpor %xmm3, %xmm2, %xmm2
-; AVX-NEXT: vpshufb %xmm4, %xmm1, %xmm1
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm7, %xmm3, %xmm3
-; AVX-NEXT: vpor %xmm3, %xmm1, %xmm1
-; AVX-NEXT: vmovdqa %xmm1, (%rsi)
-; AVX-NEXT: vmovdqa %xmm2, 48(%rsi)
-; AVX-NEXT: vmovdqa %xmm11, 32(%rsi)
-; AVX-NEXT: vmovdqa %xmm6, 16(%rsi)
-; AVX-NEXT: vmovdqa %xmm0, (%rdx)
-; AVX-NEXT: vmovdqa %xmm12, 48(%rdx)
-; AVX-NEXT: vmovdqa %xmm10, 32(%rdx)
-; AVX-NEXT: vmovdqa %xmm14, 16(%rdx)
-; AVX-NEXT: vmovdqa %xmm5, (%rcx)
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX-NEXT: vpshufb %xmm12, %xmm5, %xmm5
+; AVX-NEXT: vpshufb %xmm14, %xmm13, %xmm12
+; AVX-NEXT: vpor %xmm5, %xmm12, %xmm5
+; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm6[11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7,8,9,10]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm12 = [5,6,7,8,9,10,128,128,128,128,128,0,1,2,3,4]
+; AVX-NEXT: vpshufb %xmm12, %xmm6, %xmm6
+; AVX-NEXT: vmovdqa {{.*#+}} xmm13 = [128,128,128,128,128,128,2,5,8,11,14,128,128,128,128,128]
+; AVX-NEXT: vpshufb %xmm13, %xmm11, %xmm11
+; AVX-NEXT: vpor %xmm6, %xmm11, %xmm6
+; AVX-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX-NEXT: vpalignr {{.*#+}} xmm11 = xmm14[11,12,13,14,15],xmm15[0,1,2,3,4,5,6,7,8,9,10]
+; AVX-NEXT: vpshufb %xmm12, %xmm14, %xmm14
+; AVX-NEXT: vpshufb %xmm13, %xmm7, %xmm7
+; AVX-NEXT: vpor %xmm7, %xmm14, %xmm7
+; AVX-NEXT: vpshufb %xmm12, %xmm3, %xmm3
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX-NEXT: vpshufb %xmm13, %xmm14, %xmm14
+; AVX-NEXT: vpor %xmm3, %xmm14, %xmm3
+; AVX-NEXT: vpshufb %xmm12, %xmm10, %xmm10
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
+; AVX-NEXT: vpshufb %xmm13, %xmm12, %xmm12
+; AVX-NEXT: vpor %xmm12, %xmm10, %xmm10
+; AVX-NEXT: vmovdqa %xmm10, 48(%rsi)
+; AVX-NEXT: vmovdqa %xmm3, 16(%rsi)
+; AVX-NEXT: vmovdqa %xmm7, (%rsi)
+; AVX-NEXT: vmovdqa %xmm6, 32(%rsi)
+; AVX-NEXT: vmovdqa %xmm5, 48(%rdx)
+; AVX-NEXT: vmovdqa %xmm1, 16(%rdx)
+; AVX-NEXT: vmovdqa %xmm2, (%rdx)
+; AVX-NEXT: vmovdqa %xmm4, 32(%rdx)
; AVX-NEXT: vmovdqa %xmm8, 48(%rcx)
-; AVX-NEXT: vmovdqa %xmm9, 32(%rcx)
-; AVX-NEXT: vmovdqa %xmm13, 16(%rcx)
+; AVX-NEXT: vmovdqa %xmm9, 16(%rcx)
+; AVX-NEXT: vmovdqa %xmm11, (%rcx)
+; AVX-NEXT: vmovdqa %xmm0, 32(%rcx)
+; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: load_i8_stride3_vf64:
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
index d282e051d35b3..418c987ab9a30 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
@@ -444,7 +444,6 @@ define void @store_i16_stride5_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2,3],ymm5[4,5,6],ymm2[7],ymm5[8,9],ymm2[10,11],ymm5[12,13,14],ymm2[15]
; AVX512-NEXT: vpternlogq {{.*#+}} ymm2 = ymm3 ^ (mem & (ymm2 ^ ymm3))
-; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,1,3,4,5,6,7]
@@ -472,7 +471,6 @@ define void @store_i16_stride5_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2,3],ymm5[4,5,6],ymm2[7],ymm5[8,9],ymm2[10,11],ymm5[12,13,14],ymm2[15]
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm2 = ymm3 ^ (mem & (ymm2 ^ ymm3))
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,6,7,14,15,u,u,8,9,10,11,12,13,14,15]
; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
@@ -499,7 +497,6 @@ define void @store_i16_stride5_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2,3],ymm5[4,5,6],ymm2[7],ymm5[8,9],ymm2[10,11],ymm5[12,13,14],ymm2[15]
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm2 = ymm3 ^ (mem & (ymm2 ^ ymm3))
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,1,3,4,5,6,7]
@@ -527,7 +524,6 @@ define void @store_i16_stride5_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2,3],ymm5[4,5,6],ymm2[7],ymm5[8,9],ymm2[10,11],ymm5[12,13,14],ymm2[15]
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm2 = ymm3 ^ (mem & (ymm2 ^ ymm3))
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,6,7,14,15,u,u,8,9,10,11,12,13,14,15]
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
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 fc24976d651db..a6ac90ab56ea3 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
@@ -439,50 +439,51 @@ define void @store_i16_stride7_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX: # %bb.0:
; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
-; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
+; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero
-; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
-; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm1[0,1,1,3]
-; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm2[3,1,2,1]
+; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm3[0],xmm0[0]
+; AVX-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero
+; AVX-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero
+; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
+; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[0,1,1,3]
+; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,7,7]
+; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[3,1,2,1]
; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[2,0,2,3,4,5,6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0,1,2,3,4],xmm3[5,6],xmm4[7]
-; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[3,1,2,3]
+; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0,1,2,3,4],xmm2[5,6],xmm4[7]
+; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm1[3,1,2,3]
; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm4[0,1,2,0,4,5,6,7]
; AVX-NEXT: vmovddup {{.*#+}} xmm6 = mem[0,0]
; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm6[1,1,1,1]
; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],xmm7[0]
-; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm5[2,3,4],xmm3[5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3,4],xmm2[5,6,7]
; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,3,1,3,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1,2],xmm6[3],xmm4[4,5,6,7]
-; AVX-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0],xmm4[1,2,3,4,5,6,7]
-; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm1[0,1,8,9,u,u,u,u,u,u,u,u,2,3,2,3]
-; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
-; AVX-NEXT: vpshuflw {{.*#+}} xmm7 = xmm2[0,1,0,2,4,5,6,7]
+; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm0[0,1,8,9,u,u,u,u,u,u,u,u,2,3,2,3]
+; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,2,2,3]
+; AVX-NEXT: vpshuflw {{.*#+}} xmm7 = xmm3[0,1,0,2,4,5,6,7]
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm7[2,3],xmm5[4,5,6,7]
; AVX-NEXT: vpxor %xmm7, %xmm7, %xmm7
-; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm0[0],xmm7[1,2,3],xmm0[4],xmm7[5,6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm1[0],xmm7[1,2,3],xmm1[4],xmm7[5,6,7]
; AVX-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX-NEXT: vpackusdw %xmm8, %xmm7, %xmm7
; AVX-NEXT: vpackusdw %xmm7, %xmm7, %xmm7
; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3],xmm7[4,5,6],xmm5[7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4,5,6,7]
-; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[10,11,2,3,6,7,u,u,u,u,u,u,4,5,12,13]
-; AVX-NEXT: vmovsldup {{.*#+}} xmm0 = xmm0[0,0,2,2]
-; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm6[5],xmm0[6,7]
-; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3,4,5],xmm1[6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm3[1],xmm0[2],xmm3[3],xmm0[4,5,6,7]
+; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[10,11,2,3,6,7,u,u,u,u,u,u,4,5,12,13]
+; AVX-NEXT: vmovsldup {{.*#+}} xmm1 = xmm1[0,0,2,2]
+; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
+; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm6[5],xmm1[6,7]
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3,4,5],xmm0[6,7]
; AVX-NEXT: vmovdqa %xmm0, 16(%rax)
; AVX-NEXT: vmovdqa %xmm5, (%rax)
; AVX-NEXT: vmovq %xmm4, 48(%rax)
-; AVX-NEXT: vmovdqa %xmm3, 32(%rax)
+; AVX-NEXT: vmovdqa %xmm2, 32(%rax)
+; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: store_i16_stride7_vf4:
@@ -1390,19 +1391,18 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
; AVX512-NEXT: vpternlogd {{.*#+}} ymm8 = (mem & ~ymm8) | ymm7
; AVX512-NEXT: vpternlogq {{.*#+}} ymm8 = ymm2 ^ (mem & (ymm8 ^ ymm2))
-; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm8, %zmm2
-; AVX512-NEXT: vpsrlq $48, %xmm4, %xmm4
-; AVX512-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm4[1]
+; AVX512-NEXT: vpsrlq $48, %xmm4, %xmm2
+; AVX512-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm3[3,4],xmm0[5,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm2[3,4],xmm0[5,6,7]
; AVX512-NEXT: vpsrld $16, %xmm6, %xmm1
; AVX512-NEXT: vpunpckhdq {{.*#+}} xmm1 = xmm5[2],xmm1[2],xmm5[3],xmm1[3]
-; AVX512-NEXT: vpbroadcastd 12(%r10), %xmm3
-; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4],xmm1[5,6],xmm3[7]
+; AVX512-NEXT: vpbroadcastd 12(%r10), %xmm2
+; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3,4],xmm1[5,6],xmm2[7]
; AVX512-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4],xmm1[5,6,7]
; AVX512-NEXT: vmovdqa %xmm0, 96(%rax)
-; AVX512-NEXT: vmovdqa %ymm2, 64(%rax)
+; AVX512-NEXT: vmovdqa %ymm8, 64(%rax)
; AVX512-NEXT: vmovdqa64 %zmm11, (%rax)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
@@ -1451,7 +1451,6 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
; AVX512-FCP-NEXT: vpternlogd {{.*#+}} ymm8 = (mem & ~ymm8) | ymm7
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm8 = ymm6 ^ (mem & (ymm8 ^ ymm6))
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm8, %zmm6
; AVX512-FCP-NEXT: vpsrlq $48, %xmm3, %xmm3
; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm3[1]
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
@@ -1464,7 +1463,7 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4],xmm1[5,6,7]
; AVX512-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm11, (%rax)
-; AVX512-FCP-NEXT: vmovdqa %ymm6, 64(%rax)
+; AVX512-FCP-NEXT: vmovdqa %ymm8, 64(%rax)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
@@ -1515,19 +1514,18 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
; AVX512DQ-NEXT: vpternlogd {{.*#+}} ymm8 = (mem & ~ymm8) | ymm7
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm8 = ymm2 ^ (mem & (ymm8 ^ ymm2))
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm8, %zmm2
-; AVX512DQ-NEXT: vpsrlq $48, %xmm4, %xmm4
-; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm4[1]
+; AVX512DQ-NEXT: vpsrlq $48, %xmm4, %xmm2
+; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm3[3,4],xmm0[5,6,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm2[3,4],xmm0[5,6,7]
; AVX512DQ-NEXT: vpsrld $16, %xmm6, %xmm1
; AVX512DQ-NEXT: vpunpckhdq {{.*#+}} xmm1 = xmm5[2],xmm1[2],xmm5[3],xmm1[3]
-; AVX512DQ-NEXT: vpbroadcastd 12(%r10), %xmm3
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4],xmm1[5,6],xmm3[7]
+; AVX512DQ-NEXT: vpbroadcastd 12(%r10), %xmm2
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3,4],xmm1[5,6],xmm2[7]
; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4],xmm1[5,6,7]
; AVX512DQ-NEXT: vmovdqa %xmm0, 96(%rax)
-; AVX512DQ-NEXT: vmovdqa %ymm2, 64(%rax)
+; AVX512DQ-NEXT: vmovdqa %ymm8, 64(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm11, (%rax)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
@@ -1576,7 +1574,6 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
; AVX512DQ-FCP-NEXT: vpternlogd {{.*#+}} ymm8 = (mem & ~ymm8) | ymm7
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm8 = ymm6 ^ (mem & (ymm8 ^ ymm6))
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm8, %zmm6
; AVX512DQ-FCP-NEXT: vpsrlq $48, %xmm3, %xmm3
; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm3[1]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
@@ -1589,7 +1586,7 @@ define void @store_i16_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4],xmm1[5,6,7]
; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm6, 64(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm8, 64(%rax)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i64-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i64-stride-7.ll
index d1fd4a360036b..14cdfb38212c0 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i64-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i64-stride-7.ll
@@ -51,22 +51,22 @@ define void @store_i64_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX-NEXT: vmovaps (%rdi), %xmm0
; AVX-NEXT: vmovaps (%rsi), %xmm1
; AVX-NEXT: vmovaps (%rdx), %xmm2
-; AVX-NEXT: vmovaps (%r9), %xmm3
-; AVX-NEXT: vmovaps (%r10), %xmm4
+; AVX-NEXT: vmovaps (%r8), %xmm3
+; AVX-NEXT: vmovaps (%r9), %xmm4
+; AVX-NEXT: vmovaps (%r10), %xmm5
+; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm4, %ymm6
+; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm3, %ymm7
+; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[2],ymm6[3]
; AVX-NEXT: vinsertf128 $1, (%rcx), %ymm1, %ymm1
-; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm5
-; AVX-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm5[0],ymm1[0],ymm5[2],ymm1[2]
-; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0
-; AVX-NEXT: vmovaps (%r8), %xmm6
-; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm6, %ymm7
-; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[2],ymm0[3]
-; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm2, %ymm2
-; AVX-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
-; AVX-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm4[1]
-; AVX-NEXT: vmovaps %xmm2, 96(%rax)
-; AVX-NEXT: vmovaps %ymm1, 64(%rax)
-; AVX-NEXT: vmovapd %ymm0, 32(%rax)
-; AVX-NEXT: vmovaps %ymm5, (%rax)
+; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm3
+; AVX-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]
+; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
+; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm5[1]
+; AVX-NEXT: vmovaps %xmm1, 96(%rax)
+; AVX-NEXT: vmovaps %ymm0, (%rax)
+; AVX-NEXT: vmovaps %ymm3, 64(%rax)
+; AVX-NEXT: vmovapd %ymm6, 32(%rax)
; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-3.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-3.ll
index 674bad2c7aa87..3cc019e920a76 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-3.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-3.ll
@@ -1040,39 +1040,40 @@ define void @store_i8_stride3_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX: # %bb.0:
; AVX-NEXT: vmovdqa (%rdi), %xmm0
; AVX-NEXT: vmovdqa 16(%rdi), %xmm1
-; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
+; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
; AVX-NEXT: vmovdqa (%rsi), %xmm2
; AVX-NEXT: vmovdqa 16(%rsi), %xmm3
+; AVX-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
; AVX-NEXT: vpalignr {{.*#+}} xmm4 = xmm3[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
-; AVX-NEXT: vpalignr {{.*#+}} xmm5 = xmm2[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
-; AVX-NEXT: vmovdqa (%rdx), %xmm6
-; AVX-NEXT: vmovdqa 16(%rdx), %xmm7
-; AVX-NEXT: vpalignr {{.*#+}} xmm8 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm9 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
+; AVX-NEXT: vmovdqa (%rdx), %xmm5
+; AVX-NEXT: vmovdqa 16(%rdx), %xmm6
+; AVX-NEXT: vpalignr {{.*#+}} xmm7 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm8 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm5 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm4 = xmm7[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm3 = xmm9[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm2 = xmm8[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm2 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm4 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm5 = xmm8[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm3 = xmm7[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4]
; AVX-NEXT: vmovdqa {{.*#+}} xmm6 = [0,11,6,1,12,7,2,13,8,3,14,9,4,15,10,5]
; AVX-NEXT: vpshufb %xmm6, %xmm0, %xmm0
-; AVX-NEXT: vpshufb %xmm6, %xmm2, %xmm2
; AVX-NEXT: vpshufb %xmm6, %xmm5, %xmm5
+; AVX-NEXT: vpshufb %xmm6, %xmm2, %xmm2
; AVX-NEXT: vpshufb %xmm6, %xmm3, %xmm3
; AVX-NEXT: vpshufb %xmm6, %xmm4, %xmm4
; AVX-NEXT: vpshufb %xmm6, %xmm1, %xmm1
; AVX-NEXT: vmovdqa %xmm1, 64(%rcx)
; AVX-NEXT: vmovdqa %xmm4, 80(%rcx)
; AVX-NEXT: vmovdqa %xmm3, 48(%rcx)
-; AVX-NEXT: vmovdqa %xmm5, 32(%rcx)
-; AVX-NEXT: vmovdqa %xmm2, (%rcx)
+; AVX-NEXT: vmovdqa %xmm2, 32(%rcx)
+; AVX-NEXT: vmovdqa %xmm5, (%rcx)
; AVX-NEXT: vmovdqa %xmm0, 16(%rcx)
+; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: store_i8_stride3_vf32:
@@ -1637,116 +1638,113 @@ define void @store_i8_stride3_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX-NEXT: vmovdqa 32(%rdi), %xmm6
; AVX-NEXT: vmovdqa 48(%rdi), %xmm3
; AVX-NEXT: vmovdqa {{.*#+}} xmm7 = [11,12,13,14,15,0,1,2,3,4,5,128,128,128,128,128]
-; AVX-NEXT: vpshufb %xmm7, %xmm8, %xmm0
-; AVX-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT: vmovdqa (%rdx), %xmm5
-; AVX-NEXT: vmovdqa 16(%rdx), %xmm1
-; AVX-NEXT: vpshufb %xmm7, %xmm9, %xmm0
+; AVX-NEXT: vpshufb %xmm7, %xmm8, %xmm2
+; AVX-NEXT: vmovdqa (%rdx), %xmm1
+; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vmovdqa 32(%rdx), %xmm0
+; AVX-NEXT: vpslldq {{.*#+}} xmm4 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm1[0,1,2,3,4]
+; AVX-NEXT: vpor %xmm4, %xmm2, %xmm1
+; AVX-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
+; AVX-NEXT: vpshufb %xmm7, %xmm9, %xmm1
+; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpshufb %xmm7, %xmm6, %xmm5
+; AVX-NEXT: vpslldq {{.*#+}} xmm10 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4]
+; AVX-NEXT: vmovdqa %xmm0, %xmm1
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpshufb %xmm7, %xmm6, %xmm4
+; AVX-NEXT: vpor %xmm5, %xmm10, %xmm5
; AVX-NEXT: vpshufb %xmm7, %xmm3, %xmm7
; AVX-NEXT: vmovdqa {{.*#+}} xmm10 = [u,u,u,u,u,128,128,128,128,128,128,6,7,8,9,10]
; AVX-NEXT: vpshufb %xmm10, %xmm3, %xmm3
-; AVX-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,u,u,5,6,7,8,9,10,128,128,128,128,128]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm12 = [u,u,u,u,u,5,6,7,8,9,10,128,128,128,128,128]
; AVX-NEXT: vmovdqa 16(%rsi), %xmm13
-; AVX-NEXT: vmovdqa 32(%rsi), %xmm14
-; AVX-NEXT: vmovdqa 48(%rsi), %xmm15
-; AVX-NEXT: vpshufb %xmm11, %xmm15, %xmm12
-; AVX-NEXT: vpor %xmm3, %xmm12, %xmm0
+; AVX-NEXT: vmovdqa 32(%rsi), %xmm11
+; AVX-NEXT: vmovdqa 48(%rsi), %xmm14
+; AVX-NEXT: vpshufb %xmm12, %xmm14, %xmm15
+; AVX-NEXT: vpor %xmm3, %xmm15, %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufb %xmm10, %xmm6, %xmm6
-; AVX-NEXT: vpshufb %xmm11, %xmm14, %xmm12
-; AVX-NEXT: vpor %xmm6, %xmm12, %xmm0
+; AVX-NEXT: vpshufb %xmm12, %xmm11, %xmm15
+; AVX-NEXT: vpor %xmm6, %xmm15, %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT: vpshufb %xmm10, %xmm9, %xmm9
-; AVX-NEXT: vpshufb %xmm11, %xmm13, %xmm12
-; AVX-NEXT: vpor %xmm9, %xmm12, %xmm0
+; AVX-NEXT: vpshufb %xmm12, %xmm13, %xmm15
+; AVX-NEXT: vpor %xmm9, %xmm15, %xmm0
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa (%rsi), %xmm2
+; AVX-NEXT: vmovdqa (%rsi), %xmm15
; AVX-NEXT: vpshufb %xmm10, %xmm8, %xmm8
-; AVX-NEXT: vpshufb %xmm11, %xmm2, %xmm10
-; AVX-NEXT: vpor %xmm8, %xmm10, %xmm0
-; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vmovdqa %xmm5, %xmm0
-; AVX-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm5[8],xmm2[8],xmm5[9],xmm2[9],xmm5[10],xmm2[10],xmm5[11],xmm2[11],xmm5[12],xmm2[12],xmm5[13],xmm2[13],xmm5[14],xmm2[14],xmm5[15],xmm2[15]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,u,u,4,6,8,10,12,14,7,9,11,13,15]
-; AVX-NEXT: vpshufb %xmm11, %xmm8, %xmm8
-; AVX-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm1[8],xmm13[8],xmm1[9],xmm13[9],xmm1[10],xmm13[10],xmm1[11],xmm13[11],xmm1[12],xmm13[12],xmm1[13],xmm13[13],xmm1[14],xmm13[14],xmm1[15],xmm13[15]
-; AVX-NEXT: vpshufb %xmm11, %xmm10, %xmm6
-; AVX-NEXT: vmovdqa 32(%rdx), %xmm12
-; AVX-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm12[8],xmm14[8],xmm12[9],xmm14[9],xmm12[10],xmm14[10],xmm12[11],xmm14[11],xmm12[12],xmm14[12],xmm12[13],xmm14[13],xmm12[14],xmm14[14],xmm12[15],xmm14[15]
-; AVX-NEXT: vpshufb %xmm11, %xmm10, %xmm5
+; AVX-NEXT: vpshufb %xmm12, %xmm15, %xmm10
+; AVX-NEXT: vpor %xmm8, %xmm10, %xmm9
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm2[8],xmm15[8],xmm2[9],xmm15[9],xmm2[10],xmm15[10],xmm2[11],xmm15[11],xmm2[12],xmm15[12],xmm2[13],xmm15[13],xmm2[14],xmm15[14],xmm2[15],xmm15[15]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,u,u,u,4,6,8,10,12,14,7,9,11,13,15]
+; AVX-NEXT: vpshufb %xmm0, %xmm10, %xmm8
+; AVX-NEXT: vmovdqa 16(%rdx), %xmm12
+; AVX-NEXT: vpunpckhbw {{.*#+}} xmm10 = 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]
+; AVX-NEXT: vpshufb %xmm0, %xmm10, %xmm4
+; AVX-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm1[8],xmm11[8],xmm1[9],xmm11[9],xmm1[10],xmm11[10],xmm1[11],xmm11[11],xmm1[12],xmm11[12],xmm1[13],xmm11[13],xmm1[14],xmm11[14],xmm1[15],xmm11[15]
+; AVX-NEXT: vpshufb %xmm0, %xmm10, %xmm3
; AVX-NEXT: vmovdqa 48(%rdx), %xmm10
-; AVX-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm10[8],xmm15[8],xmm10[9],xmm15[9],xmm10[10],xmm15[10],xmm10[11],xmm15[11],xmm10[12],xmm15[12],xmm10[13],xmm15[13],xmm10[14],xmm15[14],xmm10[15],xmm15[15]
-; AVX-NEXT: vpshufb %xmm11, %xmm9, %xmm9
-; AVX-NEXT: vpslldq {{.*#+}} xmm11 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm10[0,1,2,3,4]
-; AVX-NEXT: vpor %xmm7, %xmm11, %xmm11
-; AVX-NEXT: vpalignr {{.*#+}} xmm3 = xmm11[5,6,7,8,9,10,11,12,13,14,15],xmm15[0,1,2,3,4]
-; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpslldq {{.*#+}} xmm11 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm12[0,1,2,3,4]
-; AVX-NEXT: vpor %xmm4, %xmm11, %xmm11
-; AVX-NEXT: vpalignr {{.*#+}} xmm3 = xmm11[5,6,7,8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4]
-; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpslldq {{.*#+}} xmm14 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm1[0,1,2,3,4]
-; AVX-NEXT: vmovdqa %xmm1, %xmm11
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT: vpor %xmm1, %xmm14, %xmm14
-; AVX-NEXT: vpalignr {{.*#+}} xmm3 = xmm14[5,6,7,8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4]
-; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpslldq {{.*#+}} xmm14 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4]
-; AVX-NEXT: vmovdqa %xmm0, %xmm3
-; AVX-NEXT: vmovdqa (%rsp), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpor %xmm0, %xmm14, %xmm14
-; AVX-NEXT: vpalignr {{.*#+}} xmm14 = xmm14[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm2 = xmm9[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
-; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpalignr {{.*#+}} xmm13 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm15 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
-; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm8[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
-; AVX-NEXT: vmovdqa {{.*#+}} xmm8 = [5,128,11,6,128,12,7,128,13,8,128,14,9,128,15,10]
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm8, %xmm0, %xmm6
-; AVX-NEXT: vmovdqa {{.*#+}} xmm9 = [128,5,128,128,6,128,128,7,128,128,8,128,128,9,128,128]
-; AVX-NEXT: vpshufb %xmm9, %xmm3, %xmm2
-; AVX-NEXT: vpor %xmm2, %xmm6, %xmm2
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm8, %xmm0, %xmm6
-; AVX-NEXT: vpshufb %xmm9, %xmm11, %xmm4
-; AVX-NEXT: vpor %xmm4, %xmm6, %xmm4
-; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm8, %xmm0, %xmm6
-; AVX-NEXT: vpshufb %xmm9, %xmm12, %xmm12
-; AVX-NEXT: vpor %xmm6, %xmm12, %xmm6
+; AVX-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm10[8],xmm14[8],xmm10[9],xmm14[9],xmm10[10],xmm14[10],xmm10[11],xmm14[11],xmm10[12],xmm14[12],xmm10[13],xmm14[13],xmm10[14],xmm14[14],xmm10[15],xmm14[15]
+; AVX-NEXT: vpshufb %xmm0, %xmm6, %xmm1
+; AVX-NEXT: vpslldq {{.*#+}} xmm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm10[0,1,2,3,4]
+; AVX-NEXT: vpor %xmm6, %xmm7, %xmm6
+; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4]
+; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpslldq {{.*#+}} xmm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm12[0,1,2,3,4]
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm8, %xmm0, %xmm8
-; AVX-NEXT: vpshufb %xmm9, %xmm10, %xmm9
-; AVX-NEXT: vpor %xmm9, %xmm8, %xmm8
-; AVX-NEXT: vmovdqa {{.*#+}} xmm9 = [0,11,6,1,12,7,2,13,8,3,14,9,4,15,10,5]
-; AVX-NEXT: vpshufb %xmm9, %xmm14, %xmm0
-; AVX-NEXT: vpshufb %xmm9, %xmm1, %xmm5
+; AVX-NEXT: vpor %xmm6, %xmm0, %xmm6
+; AVX-NEXT: vpalignr {{.*#+}} xmm14 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
+; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT: vpalignr {{.*#+}} xmm13 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm7 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
+; AVX-NEXT: vmovdqu (%rsp), %ymm3 # 32-byte Reload
+; AVX-NEXT: vpalignr {{.*#+}} xmm15 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm15[0,1,2,3,4]
+; AVX-NEXT: vpalignr {{.*#+}} xmm3 = xmm8[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
+; AVX-NEXT: vmovdqa {{.*#+}} xmm5 = [5,128,11,6,128,12,7,128,13,8,128,14,9,128,15,10]
+; AVX-NEXT: vpshufb %xmm5, %xmm9, %xmm6
+; AVX-NEXT: vmovdqa {{.*#+}} xmm8 = [128,5,128,128,6,128,128,7,128,128,8,128,128,9,128,128]
+; AVX-NEXT: vpshufb %xmm8, %xmm2, %xmm11
+; AVX-NEXT: vpor %xmm6, %xmm11, %xmm6
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT: vpshufb %xmm5, %xmm1, %xmm9
+; AVX-NEXT: vpshufb %xmm8, %xmm12, %xmm11
+; AVX-NEXT: vpor %xmm11, %xmm9, %xmm9
+; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT: vpshufb %xmm5, %xmm1, %xmm11
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm9, %xmm1, %xmm10
-; AVX-NEXT: vpshufb %xmm9, %xmm15, %xmm3
+; AVX-NEXT: vpshufb %xmm8, %xmm1, %xmm12
+; AVX-NEXT: vpor %xmm12, %xmm11, %xmm11
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm9, %xmm1, %xmm11
-; AVX-NEXT: vpshufb %xmm9, %xmm13, %xmm1
+; AVX-NEXT: vpshufb %xmm5, %xmm1, %xmm5
+; AVX-NEXT: vpshufb %xmm8, %xmm10, %xmm8
+; AVX-NEXT: vpor %xmm5, %xmm8, %xmm5
+; AVX-NEXT: vmovdqa {{.*#+}} xmm8 = [0,11,6,1,12,7,2,13,8,3,14,9,4,15,10,5]
+; AVX-NEXT: vpshufb %xmm8, %xmm15, %xmm4
+; AVX-NEXT: vpshufb %xmm8, %xmm3, %xmm3
+; AVX-NEXT: vpshufb %xmm8, %xmm14, %xmm10
+; AVX-NEXT: vpshufb %xmm8, %xmm0, %xmm2
+; AVX-NEXT: vpshufb %xmm8, %xmm13, %xmm0
+; AVX-NEXT: vpshufb %xmm8, %xmm7, %xmm1
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm9, %xmm7, %xmm12
+; AVX-NEXT: vpshufb %xmm8, %xmm7, %xmm12
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX-NEXT: vpshufb %xmm9, %xmm7, %xmm7
-; AVX-NEXT: vmovdqa %xmm3, 80(%rcx)
-; AVX-NEXT: vmovdqa %xmm4, 64(%rcx)
-; AVX-NEXT: vmovdqa %xmm0, (%rcx)
-; AVX-NEXT: vmovdqa %xmm2, 16(%rcx)
-; AVX-NEXT: vmovdqa %xmm5, 32(%rcx)
+; AVX-NEXT: vpshufb %xmm8, %xmm7, %xmm7
+; AVX-NEXT: vmovdqa %xmm2, 80(%rcx)
+; AVX-NEXT: vmovdqa %xmm9, 64(%rcx)
+; AVX-NEXT: vmovdqa %xmm6, 16(%rcx)
; AVX-NEXT: vmovdqa %xmm10, 48(%rcx)
; AVX-NEXT: vmovdqa %xmm7, 176(%rcx)
-; AVX-NEXT: vmovdqa %xmm8, 160(%rcx)
-; AVX-NEXT: vmovdqa %xmm11, 96(%rcx)
-; AVX-NEXT: vmovdqa %xmm6, 112(%rcx)
-; AVX-NEXT: vmovdqa %xmm1, 128(%rcx)
+; AVX-NEXT: vmovdqa %xmm5, 160(%rcx)
+; AVX-NEXT: vmovdqa %xmm11, 112(%rcx)
; AVX-NEXT: vmovdqa %xmm12, 144(%rcx)
+; AVX-NEXT: vmovdqa %xmm4, (%rcx)
+; AVX-NEXT: vmovdqa %xmm3, 32(%rcx)
+; AVX-NEXT: vmovdqa %xmm0, 96(%rcx)
+; AVX-NEXT: vmovdqa %xmm1, 128(%rcx)
; AVX-NEXT: addq $40, %rsp
+; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: store_i8_stride3_vf64:
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-4.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-4.ll
index 21b98dbb3843e..7c0a1588db8b8 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-4.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-4.ll
@@ -1294,30 +1294,30 @@ define void @store_i8_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX-NEXT: vmovdqa (%rsi), %xmm0
; AVX-NEXT: vmovdqa 16(%rsi), %xmm1
; AVX-NEXT: vmovdqa 32(%rsi), %xmm2
-; AVX-NEXT: vmovdqa 48(%rsi), %xmm4
-; AVX-NEXT: vmovdqa (%rdi), %xmm6
-; AVX-NEXT: vmovdqa 16(%rdi), %xmm8
+; AVX-NEXT: vmovdqa 48(%rsi), %xmm3
+; AVX-NEXT: vmovdqa (%rdi), %xmm5
+; AVX-NEXT: vmovdqa 16(%rdi), %xmm7
; AVX-NEXT: vmovdqa 32(%rdi), %xmm9
; AVX-NEXT: vmovdqa 48(%rdi), %xmm10
-; AVX-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3],xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
-; AVX-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3],xmm8[4],xmm1[4],xmm8[5],xmm1[5],xmm8[6],xmm1[6],xmm8[7],xmm1[7]
-; AVX-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm9[0],xmm2[0],xmm9[1],xmm2[1],xmm9[2],xmm2[2],xmm9[3],xmm2[3],xmm9[4],xmm2[4],xmm9[5],xmm2[5],xmm9[6],xmm2[6],xmm9[7],xmm2[7]
-; AVX-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm10[0],xmm4[0],xmm10[1],xmm4[1],xmm10[2],xmm4[2],xmm10[3],xmm4[3],xmm10[4],xmm4[4],xmm10[5],xmm4[5],xmm10[6],xmm4[6],xmm10[7],xmm4[7]
-; AVX-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm0[8],xmm6[9],xmm0[9],xmm6[10],xmm0[10],xmm6[11],xmm0[11],xmm6[12],xmm0[12],xmm6[13],xmm0[13],xmm6[14],xmm0[14],xmm6[15],xmm0[15]
+; AVX-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3],xmm5[4],xmm0[4],xmm5[5],xmm0[5],xmm5[6],xmm0[6],xmm5[7],xmm0[7]
+; AVX-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3],xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
+; AVX-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm9[0],xmm2[0],xmm9[1],xmm2[1],xmm9[2],xmm2[2],xmm9[3],xmm2[3],xmm9[4],xmm2[4],xmm9[5],xmm2[5],xmm9[6],xmm2[6],xmm9[7],xmm2[7]
+; AVX-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3],xmm10[4],xmm3[4],xmm10[5],xmm3[5],xmm10[6],xmm3[6],xmm10[7],xmm3[7]
+; AVX-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm8[8],xmm1[8],xmm8[9],xmm1[9],xmm8[10],xmm1[10],xmm8[11],xmm1[11],xmm8[12],xmm1[12],xmm8[13],xmm1[13],xmm8[14],xmm1[14],xmm8[15],xmm1[15]
+; AVX-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm7[8],xmm1[8],xmm7[9],xmm1[9],xmm7[10],xmm1[10],xmm7[11],xmm1[11],xmm7[12],xmm1[12],xmm7[13],xmm1[13],xmm7[14],xmm1[14],xmm7[15],xmm1[15]
; AVX-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm9[8],xmm2[8],xmm9[9],xmm2[9],xmm9[10],xmm2[10],xmm9[11],xmm2[11],xmm9[12],xmm2[12],xmm9[13],xmm2[13],xmm9[14],xmm2[14],xmm9[15],xmm2[15]
-; AVX-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm10[8],xmm4[8],xmm10[9],xmm4[9],xmm10[10],xmm4[10],xmm10[11],xmm4[11],xmm10[12],xmm4[12],xmm10[13],xmm4[13],xmm10[14],xmm4[14],xmm10[15],xmm4[15]
-; AVX-NEXT: vmovdqa (%rcx), %xmm6
+; AVX-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm10[8],xmm3[8],xmm10[9],xmm3[9],xmm10[10],xmm3[10],xmm10[11],xmm3[11],xmm10[12],xmm3[12],xmm10[13],xmm3[13],xmm10[14],xmm3[14],xmm10[15],xmm3[15]
+; AVX-NEXT: vmovdqa (%rcx), %xmm5
; AVX-NEXT: vmovdqa 16(%rcx), %xmm9
; AVX-NEXT: vmovdqa 32(%rcx), %xmm10
; AVX-NEXT: vmovdqa 48(%rcx), %xmm12
-; AVX-NEXT: vmovdqa (%rdx), %xmm8
+; AVX-NEXT: vmovdqa (%rdx), %xmm7
; AVX-NEXT: vmovdqa 16(%rdx), %xmm13
; AVX-NEXT: vmovdqa 32(%rdx), %xmm14
-; AVX-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15]
-; AVX-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm13[0],xmm9[0],xmm13[1],xmm9[1],xmm13[2],xmm9[2],xmm13[3],xmm9[3],xmm13[4],xmm9[4],xmm13[5],xmm9[5],xmm13[6],xmm9[6],xmm13[7],xmm9[7]
+; AVX-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
+; AVX-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15]
+; AVX-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm13[0],xmm9[0],xmm13[1],xmm9[1],xmm13[2],xmm9[2],xmm13[3],xmm9[3],xmm13[4],xmm9[4],xmm13[5],xmm9[5],xmm13[6],xmm9[6],xmm13[7],xmm9[7]
; AVX-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm13[8],xmm9[8],xmm13[9],xmm9[9],xmm13[10],xmm9[10],xmm13[11],xmm9[11],xmm13[12],xmm9[12],xmm13[13],xmm9[13],xmm13[14],xmm9[14],xmm13[15],xmm9[15]
; AVX-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3],xmm14[4],xmm10[4],xmm14[5],xmm10[5],xmm14[6],xmm10[6],xmm14[7],xmm10[7]
; AVX-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm14[8],xmm10[8],xmm14[9],xmm10[9],xmm14[10],xmm10[10],xmm14[11],xmm10[11],xmm14[12],xmm10[12],xmm14[13],xmm10[13],xmm14[14],xmm10[14],xmm14[15],xmm10[15]
@@ -1327,38 +1327,38 @@ define void @store_i8_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm0[0],xmm11[1],xmm0[1],xmm11[2],xmm0[2],xmm11[3],xmm0[3]
; AVX-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm11[4],xmm0[4],xmm11[5],xmm0[5],xmm11[6],xmm0[6],xmm11[7],xmm0[7]
-; AVX-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3]
-; AVX-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm7[4],xmm9[4],xmm7[5],xmm9[5],xmm7[6],xmm9[6],xmm7[7],xmm9[7]
-; AVX-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
-; AVX-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
-; AVX-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm15[0],xmm3[1],xmm15[1],xmm3[2],xmm15[2],xmm3[3],xmm15[3]
-; AVX-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm15[4],xmm3[5],xmm15[5],xmm3[6],xmm15[6],xmm3[7],xmm15[7]
-; AVX-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3]
-; AVX-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm12[4],xmm4[5],xmm12[5],xmm4[6],xmm12[6],xmm4[7],xmm12[7]
+; AVX-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
+; AVX-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
+; AVX-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm15[0],xmm4[1],xmm15[1],xmm4[2],xmm15[2],xmm4[3],xmm15[3]
+; AVX-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm15[4],xmm4[5],xmm15[5],xmm4[6],xmm15[6],xmm4[7],xmm15[7]
+; AVX-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3]
+; AVX-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm12[4],xmm3[5],xmm12[5],xmm3[6],xmm12[6],xmm3[7],xmm12[7]
; AVX-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm2[0],xmm14[0],xmm2[1],xmm14[1],xmm2[2],xmm14[2],xmm2[3],xmm14[3]
; AVX-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm14[4],xmm2[5],xmm14[5],xmm2[6],xmm14[6],xmm2[7],xmm14[7]
; AVX-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm1[0],xmm13[0],xmm1[1],xmm13[1],xmm1[2],xmm13[2],xmm1[3],xmm13[3]
; AVX-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm13[4],xmm1[5],xmm13[5],xmm1[6],xmm13[6],xmm1[7],xmm13[7]
; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
-; AVX-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
-; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm6, %ymm3
-; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm13, %ymm6
+; AVX-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3]
+; AVX-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm0[4],xmm7[4],xmm0[5],xmm7[5],xmm0[6],xmm7[6],xmm0[7],xmm7[7]
+; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm6, %ymm4
+; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm13, %ymm6
; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm9, %ymm5
; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm14, %ymm1
-; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm11, %ymm7
+; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm11, %ymm7
; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm12, %ymm2
; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm0
-; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm15, %ymm4
+; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm15, %ymm3
; AVX-NEXT: vmovaps %ymm5, 64(%r8)
; AVX-NEXT: vmovaps %ymm1, 96(%r8)
; AVX-NEXT: vmovaps %ymm7, 128(%r8)
; AVX-NEXT: vmovaps %ymm0, 192(%r8)
; AVX-NEXT: vmovaps %ymm2, 160(%r8)
-; AVX-NEXT: vmovaps %ymm4, 224(%r8)
-; AVX-NEXT: vmovaps %ymm3, (%r8)
+; AVX-NEXT: vmovaps %ymm3, 224(%r8)
; AVX-NEXT: vmovaps %ymm6, 32(%r8)
+; AVX-NEXT: vmovaps %ymm4, (%r8)
; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
index 7bbb082f07741..86efcf9c57616 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
@@ -718,7 +718,6 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,0,1,1]
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX512-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm2))
-; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm2
; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -727,7 +726,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512-NEXT: vpbroadcastw %xmm1, %xmm1
; AVX512-NEXT: vpternlogq {{.*#+}} xmm1 = xmm1 ^ (mem & (xmm1 ^ xmm0))
; AVX512-NEXT: vmovq %xmm1, 32(%r9)
-; AVX512-NEXT: vmovdqa %ymm2, (%r9)
+; AVX512-NEXT: vmovdqa %ymm3, (%r9)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
@@ -749,7 +748,6 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,0,1,1]
; AVX512-FCP-NEXT: vpermd %ymm3, %ymm4, %ymm3
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm2))
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm2
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -758,7 +756,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512-FCP-NEXT: vpbroadcastw %xmm1, %xmm1
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} xmm1 = xmm1 ^ (mem & (xmm1 ^ xmm0))
; AVX512-FCP-NEXT: vmovq %xmm1, 32(%r9)
-; AVX512-FCP-NEXT: vmovdqa %ymm2, (%r9)
+; AVX512-FCP-NEXT: vmovdqa %ymm3, (%r9)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
@@ -780,7 +778,6 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,0,1,1]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm2))
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm2
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -789,7 +786,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-NEXT: vpbroadcastw %xmm1, %xmm1
; AVX512DQ-NEXT: vpternlogq {{.*#+}} xmm1 = xmm1 ^ (mem & (xmm1 ^ xmm0))
; AVX512DQ-NEXT: vmovq %xmm1, 32(%r9)
-; AVX512DQ-NEXT: vmovdqa %ymm2, (%r9)
+; AVX512DQ-NEXT: vmovdqa %ymm3, (%r9)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
@@ -811,7 +808,6 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,0,1,1]
; AVX512DQ-FCP-NEXT: vpermd %ymm3, %ymm4, %ymm3
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm2))
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm2
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -820,7 +816,7 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-FCP-NEXT: vpbroadcastw %xmm1, %xmm1
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} xmm1 = xmm1 ^ (mem & (xmm1 ^ xmm0))
; AVX512DQ-FCP-NEXT: vmovq %xmm1, 32(%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, (%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm3, (%r9)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
@@ -844,7 +840,6 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512BW-NEXT: movl $554189328, %ecx # imm = 0x21084210
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -878,7 +873,6 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512BW-FCP-NEXT: movl $554189328, %ecx # imm = 0x21084210
; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -912,7 +906,6 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-BW-NEXT: movl $554189328, %ecx # imm = 0x21084210
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT: vpor %xmm0, %xmm1, %xmm0
@@ -946,7 +939,6 @@ define void @store_i8_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX512DQ-BW-FCP-NEXT: movl $554189328, %ecx # imm = 0x21084210
; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm2
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,14,u],zero,zero,xmm1[7,15,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
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 524161408bc0f..266e1aa65ea98 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
@@ -852,52 +852,53 @@ define void @store_i8_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecp
; AVX: # %bb.0:
; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
-; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
+; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero
-; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
+; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm3[0],xmm0[0]
; AVX-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero
-; AVX-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[u,u,u],zero,zero,xmm2[5,13,u,u,u],zero,zero,xmm2[6,14,u,u]
-; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm1[u,u,u,5,13],zero,zero,xmm1[u,u,u,6,14],zero,zero,xmm1[u,u]
+; AVX-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero
+; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
+; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; AVX-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
+; AVX-NEXT: vpshufb {{.*#+}} xmm4 = xmm3[u,u,u],zero,zero,xmm3[5,13,u,u,u],zero,zero,xmm3[6,14,u,u]
+; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm0[u,u,u,5,13],zero,zero,xmm0[u,u,u,6,14],zero,zero,xmm0[u,u]
; AVX-NEXT: vpor %xmm4, %xmm5, %xmm4
-; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm0[4,12],zero,xmm0[u,u,u,u,5,13],zero,xmm0[u,u,u,u,6,14]
-; AVX-NEXT: vpshufb {{.*#+}} xmm6 = zero,zero,xmm3[4,u,u,u,u],zero,zero,xmm3[5,u,u,u,u],zero,zero
+; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm1[4,12],zero,xmm1[u,u,u,u,5,13],zero,xmm1[u,u,u,u,6,14]
+; AVX-NEXT: vpshufb {{.*#+}} xmm6 = zero,zero,xmm2[4,u,u,u,u],zero,zero,xmm2[5,u,u,u,u],zero,zero
; AVX-NEXT: vpor %xmm6, %xmm5, %xmm5
; AVX-NEXT: vmovdqa {{.*#+}} xmm6 = [255,255,255,0,0,0,0,255,255,255,0,0,0,0,255,255]
; AVX-NEXT: vpblendvb %xmm6, %xmm5, %xmm4, %xmm4
-; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm2[u],zero,zero,xmm2[7,15,u,u,u,u,u,u,u,u,u,u,u]
-; AVX-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[u,7,15],zero,zero,xmm1[u,u,u,u,u,u,u,u,u,u,u]
+; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm3[u],zero,zero,xmm3[7,15,u,u,u,u,u,u,u,u,u,u,u]
+; AVX-NEXT: vpshufb {{.*#+}} xmm6 = xmm0[u,7,15],zero,zero,xmm0[u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT: vpor %xmm5, %xmm6, %xmm5
-; AVX-NEXT: vpshufb {{.*#+}} xmm6 = zero,xmm0[u,u,u,u,7,15],zero,xmm0[u,u,u,u,u,u,u,u]
-; AVX-NEXT: vpshufb {{.*#+}} xmm7 = xmm3[6,u,u,u,u],zero,zero,xmm3[7,u,u,u,u,u,u,u,u]
+; AVX-NEXT: vpshufb {{.*#+}} xmm6 = zero,xmm1[u,u,u,u,7,15],zero,xmm1[u,u,u,u,u,u,u,u]
+; AVX-NEXT: vpshufb {{.*#+}} xmm7 = xmm2[6,u,u,u,u],zero,zero,xmm2[7,u,u,u,u,u,u,u,u]
; AVX-NEXT: vpor %xmm6, %xmm7, %xmm6
; AVX-NEXT: vmovq {{.*#+}} xmm7 = [0,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT: vpblendvb %xmm7, %xmm5, %xmm6, %xmm5
-; AVX-NEXT: vpshufb {{.*#+}} xmm6 = zero,zero,xmm2[0,8,u,u,u],zero,zero,xmm2[1,9,u,u,u],zero,zero
-; AVX-NEXT: vpshufb {{.*#+}} xmm7 = xmm1[0,8],zero,zero,xmm1[u,u,u,1,9],zero,zero,xmm1[u,u,u,2,10]
+; AVX-NEXT: vpshufb {{.*#+}} xmm6 = zero,zero,xmm3[0,8,u,u,u],zero,zero,xmm3[1,9,u,u,u],zero,zero
+; AVX-NEXT: vpshufb {{.*#+}} xmm7 = xmm0[0,8],zero,zero,xmm0[u,u,u,1,9],zero,zero,xmm0[u,u,u,2,10]
; AVX-NEXT: vpor %xmm6, %xmm7, %xmm6
-; AVX-NEXT: vpshufb {{.*#+}} xmm7 = xmm0[u,u,u,u,0,8],zero,xmm0[u,u,u,u,1,9],zero,xmm0[u,u]
-; AVX-NEXT: vpshufb {{.*#+}} xmm8 = xmm3[u,u,u,u],zero,zero,xmm3[0,u,u,u,u],zero,zero,xmm3[1,u,u]
+; AVX-NEXT: vpshufb {{.*#+}} xmm7 = xmm1[u,u,u,u,0,8],zero,xmm1[u,u,u,u,1,9],zero,xmm1[u,u]
+; AVX-NEXT: vpshufb {{.*#+}} xmm8 = xmm2[u,u,u,u],zero,zero,xmm2[0,u,u,u,u],zero,zero,xmm2[1,u,u]
; AVX-NEXT: vpor %xmm7, %xmm8, %xmm7
; AVX-NEXT: vmovdqa {{.*#+}} xmm8 = [255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255]
; AVX-NEXT: vpblendvb %xmm8, %xmm6, %xmm7, %xmm6
-; AVX-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[u,u,u,3,11],zero,zero,xmm1[u,u,u,4,12],zero,zero
-; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[2,10,u,u,u],zero,zero,xmm2[3,11,u,u,u],zero,zero,xmm2[4,12]
-; AVX-NEXT: vpor %xmm1, %xmm2, %xmm1
-; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,2,10],zero,xmm0[u,u,u,u,3,11],zero,xmm0[u,u,u,u]
-; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm3[u,u],zero,zero,xmm3[2,u,u,u,u],zero,zero,xmm3[3,u,u,u,u]
-; AVX-NEXT: vpor %xmm2, %xmm0, %xmm0
+; AVX-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,u,3,11],zero,zero,xmm0[u,u,u,4,12],zero,zero
+; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[2,10,u,u,u],zero,zero,xmm3[3,11,u,u,u],zero,zero,xmm3[4,12]
+; AVX-NEXT: vpor %xmm0, %xmm3, %xmm0
+; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,2,10],zero,xmm1[u,u,u,u,3,11],zero,xmm1[u,u,u,u]
+; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u],zero,zero,xmm2[2,u,u,u,u],zero,zero,xmm2[3,u,u,u,u]
+; AVX-NEXT: vpor %xmm2, %xmm1, %xmm1
; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255]
-; AVX-NEXT: vpblendvb %xmm2, %xmm1, %xmm0, %xmm0
+; AVX-NEXT: vpblendvb %xmm2, %xmm0, %xmm1, %xmm0
; AVX-NEXT: vmovdqa %xmm0, 16(%rax)
; AVX-NEXT: vmovdqa %xmm6, (%rax)
; AVX-NEXT: vmovq %xmm5, 48(%rax)
; AVX-NEXT: vmovdqa %xmm4, 32(%rax)
+; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: store_i8_stride7_vf8:
@@ -2088,7 +2089,6 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,1,9],zero,ymm6[u,u,u,u,2,10],zero,ymm6[u,u,u,u,19,27],zero,ymm6[u,u,u,u,20,28],zero,ymm6[u,u,u,u,21]
; AVX512-NEXT: vpternlogq {{.*#+}} ymm6 = ymm6 | (ymm7 & ~mem)
; AVX512-NEXT: vpternlogq {{.*#+}} ymm6 = ymm6 ^ (mem & (ymm6 ^ ymm5))
-; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm6, %zmm5
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,2,3]
; AVX512-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u],zero,zero,xmm3[6,14,u,u,u],zero,zero,xmm3[7,15,u,u,u]
; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,2,3]
@@ -2101,7 +2101,7 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm3))
; AVX512-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512-NEXT: vmovdqa64 %zmm8, (%rax)
-; AVX512-NEXT: vmovdqa %ymm5, 64(%rax)
+; AVX512-NEXT: vmovdqa %ymm6, 64(%rax)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
@@ -2156,7 +2156,6 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,1,9],zero,ymm7[u,u,u,u,2,10],zero,ymm7[u,u,u,u,19,27],zero,ymm7[u,u,u,u,20,28],zero,ymm7[u,u,u,u,21]
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm7 = ymm7 | (ymm8 & ~mem)
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm7 = ymm7 ^ (mem & (ymm7 ^ ymm5))
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm7, %zmm5
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,2,3]
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u],zero,zero,xmm4[6,14,u,u,u],zero,zero,xmm4[7,15,u,u,u]
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,2,3]
@@ -2169,7 +2168,7 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm3))
; AVX512-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm6, (%rax)
-; AVX512-FCP-NEXT: vmovdqa %ymm5, 64(%rax)
+; AVX512-FCP-NEXT: vmovdqa %ymm7, 64(%rax)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
@@ -2227,7 +2226,6 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,1,9],zero,ymm6[u,u,u,u,2,10],zero,ymm6[u,u,u,u,19,27],zero,ymm6[u,u,u,u,20,28],zero,ymm6[u,u,u,u,21]
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm6 = ymm6 | (ymm7 & ~mem)
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm6 = ymm6 ^ (mem & (ymm6 ^ ymm5))
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm6, %zmm5
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,2,3]
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u],zero,zero,xmm3[6,14,u,u,u],zero,zero,xmm3[7,15,u,u,u]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,2,3]
@@ -2240,7 +2238,7 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm3))
; AVX512DQ-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%rax)
-; AVX512DQ-NEXT: vmovdqa %ymm5, 64(%rax)
+; AVX512DQ-NEXT: vmovdqa %ymm6, 64(%rax)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
@@ -2295,7 +2293,6 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,1,9],zero,ymm7[u,u,u,u,2,10],zero,ymm7[u,u,u,u,19,27],zero,ymm7[u,u,u,u,20,28],zero,ymm7[u,u,u,u,21]
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm7 = ymm7 | (ymm8 & ~mem)
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm7 = ymm7 ^ (mem & (ymm7 ^ ymm5))
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm7, %zmm5
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,2,3]
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u],zero,zero,xmm4[6,14,u,u,u],zero,zero,xmm4[7,15,u,u,u]
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,2,3]
@@ -2308,7 +2305,7 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm3))
; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, 96(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm5, 64(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm7, 64(%rax)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
@@ -2316,15 +2313,15 @@ 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), %xmm3
+; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0
; AVX512BW-NEXT: vmovdqa (%rdx), %xmm4
-; AVX512BW-NEXT: vmovdqa (%r8), %xmm0
-; AVX512BW-NEXT: vmovdqa (%r9), %xmm2
-; AVX512BW-NEXT: vmovdqa (%r10), %xmm1
+; AVX512BW-NEXT: vmovdqa (%r8), %xmm1
+; AVX512BW-NEXT: vmovdqa (%r9), %xmm3
+; AVX512BW-NEXT: vmovdqa (%r10), %xmm2
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
; AVX512BW-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512BW-NEXT: vpermw %ymm1, %ymm5, %ymm5
-; AVX512BW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm6
+; AVX512BW-NEXT: vpermw %ymm2, %ymm5, %ymm5
+; AVX512BW-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm6
; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm6[1,3,1,3]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[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, %edx # imm = 0x4081020
@@ -2333,64 +2330,63 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm4, %ymm4
; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm4[1,3,3,1]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[3,19],zero,zero,zero,zero,zero,ymm5[28,20],zero,zero,zero,zero,zero,ymm5[29,21],zero
-; AVX512BW-NEXT: vinserti128 $1, (%rsi), %ymm3, %ymm3
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm3[3,1,1,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[1],zero,zero,zero,zero,zero,ymm8[10,2],zero,zero,zero,zero,zero,ymm8[11,3],zero,zero,zero,zero,zero,ymm8[20,28],zero,zero,zero,zero,zero,ymm8[21,29],zero,zero,zero
-; AVX512BW-NEXT: vpor %ymm5, %ymm8, %ymm5
+; AVX512BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm8
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm8[3,1,1,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[1],zero,zero,zero,zero,zero,ymm0[10,2],zero,zero,zero,zero,zero,ymm0[11,3],zero,zero,zero,zero,zero,ymm0[20,28],zero,zero,zero,zero,zero,ymm0[21,29],zero,zero,zero
+; AVX512BW-NEXT: vpor %ymm5, %ymm0, %ymm0
; AVX512BW-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu8 %ymm7, %ymm5 {%k1}
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm5
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,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
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm4[2,3,0,1]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,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
-; AVX512BW-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,ymm3[5],zero,zero,zero,zero,zero,zero,ymm3[6],zero,zero,zero,zero,zero,zero,zero,ymm3[23],zero,zero,zero,zero,zero,zero,ymm3[24],zero,zero,zero,zero,zero,zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm3[2,3,0,1]
+; AVX512BW-NEXT: vmovdqu8 %ymm7, %ymm0 {%k1}
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm5 = zero,zero,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
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm4[2,3,0,1]
+; 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 %ymm5, %ymm7, %ymm5
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,zero,ymm8[5],zero,zero,zero,zero,zero,zero,ymm8[6],zero,zero,zero,zero,zero,zero,zero,ymm8[23],zero,zero,zero,zero,zero,zero,ymm8[24],zero,zero,zero,zero,zero,zero
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm8[2,3,0,1]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,zero,zero,ymm9[5],zero,zero,zero,zero,zero,zero,ymm9[6],zero,zero,zero,zero,zero,ymm9[23],zero,zero,zero,zero,zero,zero,ymm9[24],zero,zero,zero,zero,zero,zero,ymm9[25]
-; AVX512BW-NEXT: vpor %ymm9, %ymm8, %ymm8
+; AVX512BW-NEXT: vpor %ymm7, %ymm9, %ymm7
; AVX512BW-NEXT: movl $202911840, %ecx # imm = 0xC183060
; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu8 %ymm7, %ymm8 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm3[0,2,0,2]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[0,8],zero,zero,zero,zero,zero,ymm7[1,9],zero,zero,zero,zero,zero,ymm7[2,10],zero,zero,zero,zero,zero,ymm7[19,27],zero,zero,zero,zero,zero,ymm7[20,28],zero,zero
+; AVX512BW-NEXT: vmovdqu8 %ymm5, %ymm7 {%k1}
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm8[0,2,0,2]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[0,8],zero,zero,zero,zero,zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[19,27],zero,zero,zero,zero,zero,ymm5[20,28],zero,zero
; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm4[0,2,0,2]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,ymm9[0,8],zero,zero,zero,zero,zero,ymm9[1,9],zero,zero,zero,zero,zero,ymm9[18,26],zero,zero,zero,zero,zero,ymm9[19,27],zero,zero,zero,zero,zero,ymm9[20,28]
-; AVX512BW-NEXT: vpor %ymm7, %ymm9, %ymm7
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero
+; AVX512BW-NEXT: vpor %ymm5, %ymm9, %ymm5
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm7, %zmm5, %zmm5
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero
; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm6[2,3,0,1]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,ymm9[4],zero,zero,zero,zero,zero,zero,ymm9[5],zero,zero,zero,zero,zero,zero,ymm9[6],zero,zero,zero,zero,zero,ymm9[23],zero,zero,zero,zero,zero,zero,ymm9[24],zero,zero,zero
-; AVX512BW-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512BW-NEXT: vpor %ymm7, %ymm9, %ymm7
; AVX512BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,0,2]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[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, %ymm8, %zmm6, %zmm6
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm8 = [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, %zmm8, %zmm8
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm7 = [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 %zmm2, %zmm7, %zmm7
; AVX512BW-NEXT: movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
; AVX512BW-NEXT: kmovq %rcx, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm8, %zmm6 {%k1}
+; AVX512BW-NEXT: vmovdqu8 %zmm7, %zmm6 {%k1}
; AVX512BW-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
; AVX512BW-NEXT: kmovq %rcx, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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 {{.*#+}} xmm0 = xmm0[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
-; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,7,7,7]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,1,3,2]
+; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm5 {%k1}
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,7,7,7]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,1,3,2]
; AVX512BW-NEXT: movw $-32510, %cx # imm = 0x8102
; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu8 %xmm1, %xmm0 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm3[1,3,2,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[6,14],zero,zero,zero,zero,zero,xmm1[7,15],zero,zero,zero,zero,zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm2 = ymm4[1,3,2,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,zero,zero,xmm2[6,14],zero,zero,zero,zero,zero,xmm2[7,15],zero,zero,zero
-; AVX512BW-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512BW-NEXT: vmovdqu8 %xmm2, %xmm1 {%k1}
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm2 = ymm8[1,3,2,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[6,14],zero,zero,zero,zero,zero,xmm2[7,15],zero,zero,zero,zero,zero
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm4[1,3,2,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,zero,zero,xmm3[6,14],zero,zero,zero,zero,zero,xmm3[7,15],zero,zero,zero
+; AVX512BW-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512BW-NEXT: movw $-7741, %cx # imm = 0xE1C3
; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu8 %xmm0, %xmm1 {%k1}
-; AVX512BW-NEXT: vmovdqa %xmm1, 96(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm7, (%rax)
-; AVX512BW-NEXT: vmovdqa %ymm5, 64(%rax)
+; AVX512BW-NEXT: vmovdqu8 %xmm1, %xmm2 {%k1}
+; AVX512BW-NEXT: vmovdqa %xmm2, 96(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512BW-NEXT: vmovdqa %ymm0, 64(%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
@@ -2422,7 +2418,6 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %ymm7, %ymm5 {%k1}
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm5
; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [1,9,2,10,1,9,2,10]
; AVX512BW-FCP-NEXT: # ymm7 = mem[0,1,0,1]
; AVX512BW-FCP-NEXT: vpermi2d %ymm3, %ymm2, %ymm7
@@ -2465,15 +2460,15 @@ 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), %xmm3
+; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0
; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm4
-; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm0
-; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm2
-; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm1
+; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm1
+; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm3
+; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm2
; AVX512DQ-BW-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
; AVX512DQ-BW-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpermw %ymm1, %ymm5, %ymm5
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm6
+; AVX512DQ-BW-NEXT: vpermw %ymm2, %ymm5, %ymm5
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm6
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm6[1,3,1,3]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[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, %edx # imm = 0x4081020
@@ -2482,64 +2477,63 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm4, %ymm4
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm5 = ymm4[1,3,3,1]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[3,19],zero,zero,zero,zero,zero,ymm5[28,20],zero,zero,zero,zero,zero,ymm5[29,21],zero
-; AVX512DQ-BW-NEXT: vinserti128 $1, (%rsi), %ymm3, %ymm3
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm3[3,1,1,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[1],zero,zero,zero,zero,zero,ymm8[10,2],zero,zero,zero,zero,zero,ymm8[11,3],zero,zero,zero,zero,zero,ymm8[20,28],zero,zero,zero,zero,zero,ymm8[21,29],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpor %ymm5, %ymm8, %ymm5
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm8
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm0 = ymm8[3,1,1,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[1],zero,zero,zero,zero,zero,ymm0[10,2],zero,zero,zero,zero,zero,ymm0[11,3],zero,zero,zero,zero,zero,ymm0[20,28],zero,zero,zero,zero,zero,ymm0[21,29],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpor %ymm5, %ymm0, %ymm0
; AVX512DQ-BW-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %ymm7, %ymm5 {%k1}
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm5
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,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
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm4[2,3,0,1]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,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
-; AVX512DQ-BW-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,ymm3[5],zero,zero,zero,zero,zero,zero,ymm3[6],zero,zero,zero,zero,zero,zero,zero,ymm3[23],zero,zero,zero,zero,zero,zero,ymm3[24],zero,zero,zero,zero,zero,zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm3[2,3,0,1]
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm7, %ymm0 {%k1}
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm5 = zero,zero,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
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm4[2,3,0,1]
+; 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 %ymm5, %ymm7, %ymm5
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,zero,ymm8[5],zero,zero,zero,zero,zero,zero,ymm8[6],zero,zero,zero,zero,zero,zero,zero,ymm8[23],zero,zero,zero,zero,zero,zero,ymm8[24],zero,zero,zero,zero,zero,zero
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm8[2,3,0,1]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,zero,zero,ymm9[5],zero,zero,zero,zero,zero,zero,ymm9[6],zero,zero,zero,zero,zero,ymm9[23],zero,zero,zero,zero,zero,zero,ymm9[24],zero,zero,zero,zero,zero,zero,ymm9[25]
-; AVX512DQ-BW-NEXT: vpor %ymm9, %ymm8, %ymm8
+; AVX512DQ-BW-NEXT: vpor %ymm7, %ymm9, %ymm7
; AVX512DQ-BW-NEXT: movl $202911840, %ecx # imm = 0xC183060
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %ymm7, %ymm8 {%k1}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm3[0,2,0,2]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[0,8],zero,zero,zero,zero,zero,ymm7[1,9],zero,zero,zero,zero,zero,ymm7[2,10],zero,zero,zero,zero,zero,ymm7[19,27],zero,zero,zero,zero,zero,ymm7[20,28],zero,zero
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm5, %ymm7 {%k1}
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm5 = ymm8[0,2,0,2]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[0,8],zero,zero,zero,zero,zero,ymm5[1,9],zero,zero,zero,zero,zero,ymm5[2,10],zero,zero,zero,zero,zero,ymm5[19,27],zero,zero,zero,zero,zero,ymm5[20,28],zero,zero
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm4[0,2,0,2]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,ymm9[0,8],zero,zero,zero,zero,zero,ymm9[1,9],zero,zero,zero,zero,zero,ymm9[18,26],zero,zero,zero,zero,zero,ymm9[19,27],zero,zero,zero,zero,zero,ymm9[20,28]
-; AVX512DQ-BW-NEXT: vpor %ymm7, %ymm9, %ymm7
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero
+; AVX512DQ-BW-NEXT: vpor %ymm5, %ymm9, %ymm5
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm7, %zmm5, %zmm5
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[4],zero,zero,zero,zero,zero,zero,ymm6[5],zero,zero,zero,zero,zero,zero,ymm6[6],zero,zero,zero,zero,zero,zero,zero,ymm6[23],zero,zero,zero,zero,zero,zero,ymm6[24],zero,zero
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm6[2,3,0,1]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 = zero,ymm9[4],zero,zero,zero,zero,zero,zero,ymm9[5],zero,zero,zero,zero,zero,zero,ymm9[6],zero,zero,zero,zero,zero,ymm9[23],zero,zero,zero,zero,zero,zero,ymm9[24],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512DQ-BW-NEXT: vpor %ymm7, %ymm9, %ymm7
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,0,2]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[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, %ymm8, %zmm6, %zmm6
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm8 = [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, %zmm8, %zmm8
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm7 = [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 %zmm2, %zmm7, %zmm7
; AVX512DQ-BW-NEXT: movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
; AVX512DQ-BW-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm8, %zmm6 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm7, %zmm6 {%k1}
; AVX512DQ-BW-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
; AVX512DQ-BW-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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 {{.*#+}} xmm0 = xmm0[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
-; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,7,7,7]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,1,3,2]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm5 {%k1}
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[10,u,u,u,u,u,13,12,u,u,u,u,u,15,14,u]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,7,7,7]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,1,3,2]
; AVX512DQ-BW-NEXT: movw $-32510, %cx # imm = 0x8102
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %xmm1, %xmm0 {%k1}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm3[1,3,2,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[6,14],zero,zero,zero,zero,zero,xmm1[7,15],zero,zero,zero,zero,zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm2 = ymm4[1,3,2,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,zero,zero,xmm2[6,14],zero,zero,zero,zero,zero,xmm2[7,15],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512DQ-BW-NEXT: vmovdqu8 %xmm2, %xmm1 {%k1}
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm2 = ymm8[1,3,2,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[6,14],zero,zero,zero,zero,zero,xmm2[7,15],zero,zero,zero,zero,zero
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm4[1,3,2,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,zero,zero,xmm3[6,14],zero,zero,zero,zero,zero,xmm3[7,15],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512DQ-BW-NEXT: movw $-7741, %cx # imm = 0xE1C3
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %xmm0, %xmm1 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa %xmm1, 96(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, (%rax)
-; AVX512DQ-BW-NEXT: vmovdqa %ymm5, 64(%rax)
+; AVX512DQ-BW-NEXT: vmovdqu8 %xmm1, %xmm2 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqa %xmm2, 96(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512DQ-BW-NEXT: vmovdqa %ymm0, 64(%rax)
; AVX512DQ-BW-NEXT: vzeroupper
; AVX512DQ-BW-NEXT: retq
;
@@ -2571,7 +2565,6 @@ define void @store_i8_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: movl $-2029118408, %ecx # imm = 0x870E1C38
; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm7, %ymm5 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm5
; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [1,9,2,10,1,9,2,10]
; AVX512DQ-BW-FCP-NEXT: # ymm7 = mem[0,1,0,1]
; AVX512DQ-BW-FCP-NEXT: vpermi2d %ymm3, %ymm2, %ymm7
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining.ll
index efa6c16fbf4eb..c16985e081334 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining.ll
@@ -3062,6 +3062,7 @@ define <8 x i16> @shuffle_scalar_to_vector_extract(ptr %p0, ptr %p1, ptr %p2) {
; SSE2-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: shuffle_scalar_to_vector_extract:
@@ -3079,13 +3080,16 @@ define <8 x i16> @shuffle_scalar_to_vector_extract(ptr %p0, ptr %p1, ptr %p2) {
; SSSE3-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; SSSE3-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
; SSSE3-NEXT: retq
;
; SSE41-LABEL: shuffle_scalar_to_vector_extract:
; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbw (%rdi), %xmm0
+; SSE41-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
+; SSE41-NEXT: pmovsxbw %xmm0, %xmm0
; SSE41-NEXT: pextrw $4, %xmm0, %eax
; SSE41-NEXT: pextrw $7, %xmm0, %ecx
+; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: pxor %xmm0, %xmm0
; SSE41-NEXT: pinsrw $1, %eax, %xmm0
; SSE41-NEXT: movl $65531, %eax # imm = 0xFFFB
@@ -3095,22 +3099,25 @@ define <8 x i16> @shuffle_scalar_to_vector_extract(ptr %p0, ptr %p1, ptr %p2) {
; SSE41-NEXT: pinsrw $5, %eax, %xmm0
; SSE41-NEXT: movsbl (%rdx), %eax
; SSE41-NEXT: pinsrw $6, %eax, %xmm0
+; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3],xmm0[4,5,6],xmm1[7]
; SSE41-NEXT: retq
;
; AVX-LABEL: shuffle_scalar_to_vector_extract:
; AVX: # %bb.0:
-; AVX-NEXT: vpmovsxbw (%rdi), %xmm0
+; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
+; AVX-NEXT: vpmovsxbw %xmm0, %xmm0
; AVX-NEXT: vpextrw $4, %xmm0, %eax
; AVX-NEXT: vpextrw $7, %xmm0, %ecx
; AVX-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0
+; AVX-NEXT: vpinsrw $1, %eax, %xmm0, %xmm1
; AVX-NEXT: movl $65531, %eax # imm = 0xFFFB
-; AVX-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
-; AVX-NEXT: vpinsrw $4, %ecx, %xmm0, %xmm0
+; AVX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1
+; AVX-NEXT: vpinsrw $4, %ecx, %xmm1, %xmm1
; AVX-NEXT: movsbl (%rsi), %eax
-; AVX-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0
+; AVX-NEXT: vpinsrw $5, %eax, %xmm1, %xmm1
; AVX-NEXT: movsbl (%rdx), %eax
-; AVX-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
+; AVX-NEXT: vpinsrw $6, %eax, %xmm1, %xmm1
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2],xmm0[3],xmm1[4,5,6],xmm0[7]
; AVX-NEXT: retq
%tmp = load <8 x i8>, ptr %p0, align 1
%tmp1 = sext <8 x i8> %tmp to <8 x i16>
@@ -3572,45 +3579,53 @@ define void @SpinningCube() {
; SSE2-LABEL: SpinningCube:
; SSE2: # %bb.0: # %entry
; SSE2-NEXT: movl $1065353216, (%rax) # imm = 0x3F800000
-; SSE2-NEXT: movaps {{.*#+}} xmm0 = [u,u,u,1.0E+0]
-; SSE2-NEXT: movss {{.*#+}} xmm1 = [NaN,0.0E+0,0.0E+0,0.0E+0]
-; SSE2-NEXT: movapd {{.*#+}} xmm2 = [u,u,-2.0E+0,u]
-; SSE2-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]
-; SSE2-NEXT: xorps %xmm3, %xmm3
-; SSE2-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]
-; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,3]
-; SSE2-NEXT: addps %xmm3, %xmm1
-; SSE2-NEXT: movaps %xmm1, (%rax)
-; SSE2-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0,0,0]
-; SSE2-NEXT: mulps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT: addps %xmm0, %xmm1
-; SSE2-NEXT: movaps %xmm1, (%rax)
+; SSE2-NEXT: xorps %xmm0, %xmm0
+; SSE2-NEXT: movss {{.*#+}} xmm1 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
+; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[0,1]
+; SSE2-NEXT: xorps %xmm2, %xmm2
+; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,0]
+; SSE2-NEXT: movss {{.*#+}} xmm3 = [NaN,0.0E+0,0.0E+0,0.0E+0]
+; SSE2-NEXT: movapd {{.*#+}} xmm4 = [u,u,-2.0E+0,u]
+; SSE2-NEXT: movsd {{.*#+}} xmm4 = xmm3[0],xmm4[1]
+; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,0]
+; SSE2-NEXT: movq {{.*#+}} xmm3 = xmm3[0],zero
+; SSE2-NEXT: shufps {{.*#+}} xmm3 = xmm3[2,0],xmm1[2,0]
+; SSE2-NEXT: addps %xmm0, %xmm3
+; SSE2-NEXT: movaps %xmm3, (%rax)
+; SSE2-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,0,0,0]
+; SSE2-NEXT: mulps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT: addps %xmm2, %xmm0
+; SSE2-NEXT: movaps %xmm0, (%rax)
; SSE2-NEXT: retq
;
; SSSE3-LABEL: SpinningCube:
; SSSE3: # %bb.0: # %entry
; SSSE3-NEXT: movl $1065353216, (%rax) # imm = 0x3F800000
-; SSSE3-NEXT: movaps {{.*#+}} xmm0 = [u,u,u,1.0E+0]
-; SSSE3-NEXT: movss {{.*#+}} xmm1 = [NaN,0.0E+0,0.0E+0,0.0E+0]
-; SSSE3-NEXT: movapd {{.*#+}} xmm2 = [u,u,-2.0E+0,u]
-; SSSE3-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]
-; SSSE3-NEXT: xorps %xmm3, %xmm3
-; SSSE3-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]
-; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,3]
-; SSSE3-NEXT: addps %xmm3, %xmm1
-; SSSE3-NEXT: movaps %xmm1, (%rax)
-; SSSE3-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0,0,2]
-; SSSE3-NEXT: mulps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSSE3-NEXT: addps %xmm0, %xmm1
-; SSSE3-NEXT: movaps %xmm1, (%rax)
+; SSSE3-NEXT: xorps %xmm0, %xmm0
+; SSSE3-NEXT: movss {{.*#+}} xmm1 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
+; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[0,1]
+; SSSE3-NEXT: xorps %xmm2, %xmm2
+; SSSE3-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,0]
+; SSSE3-NEXT: movss {{.*#+}} xmm3 = [NaN,0.0E+0,0.0E+0,0.0E+0]
+; SSSE3-NEXT: movapd {{.*#+}} xmm4 = [u,u,-2.0E+0,u]
+; SSSE3-NEXT: movsd {{.*#+}} xmm4 = xmm3[0],xmm4[1]
+; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,0]
+; SSSE3-NEXT: movq {{.*#+}} xmm3 = xmm3[0],zero
+; SSSE3-NEXT: shufps {{.*#+}} xmm3 = xmm3[2,0],xmm1[2,0]
+; SSSE3-NEXT: addps %xmm0, %xmm3
+; SSSE3-NEXT: movaps %xmm3, (%rax)
+; SSSE3-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,0,0,2]
+; SSSE3-NEXT: mulps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT: addps %xmm2, %xmm0
+; SSSE3-NEXT: movaps %xmm0, (%rax)
; SSSE3-NEXT: retq
;
; SSE41-LABEL: SpinningCube:
; SSE41: # %bb.0: # %entry
; SSE41-NEXT: movl $1065353216, (%rax) # imm = 0x3F800000
-; SSE41-NEXT: movaps {{.*#+}} xmm0 = [u,u,u,1.0E+0]
+; SSE41-NEXT: insertps {{.*#+}} xmm0 = zero,zero,zero,mem[0]
; SSE41-NEXT: movaps {{.*#+}} xmm1 = [0.0E+0,0.0E+0,-2.0E+0,u]
; SSE41-NEXT: movss {{.*#+}} xmm2 = [NaN,0.0E+0,0.0E+0,0.0E+0]
; SSE41-NEXT: movaps %xmm1, %xmm3
@@ -3629,7 +3644,7 @@ define void @SpinningCube() {
; AVX-LABEL: SpinningCube:
; AVX: # %bb.0: # %entry
; AVX-NEXT: movl $1065353216, (%rax) # imm = 0x3F800000
-; AVX-NEXT: vbroadcastss {{.*#+}} xmm0 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
+; AVX-NEXT: vmovaps {{.*#+}} xmm0 = [0.0E+0,0.0E+0,0.0E+0,1.0E+0]
; AVX-NEXT: vmovaps {{.*#+}} xmm1 = [0.0E+0,0.0E+0,-2.0E+0,u]
; AVX-NEXT: vmovss {{.*#+}} xmm2 = [NaN,0.0E+0,0.0E+0,0.0E+0]
; AVX-NEXT: vinsertps {{.*#+}} xmm3 = xmm1[0,1,2],xmm2[0]
diff --git a/llvm/test/CodeGen/X86/vector-trunc.ll b/llvm/test/CodeGen/X86/vector-trunc.ll
index 46f770a349d96..27f4a3ecb206f 100644
--- a/llvm/test/CodeGen/X86/vector-trunc.ll
+++ b/llvm/test/CodeGen/X86/vector-trunc.ll
@@ -2018,7 +2018,8 @@ define i16 @PR66194(i8 %q) {
; AVX2-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0
; AVX2-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
; AVX2-NEXT: vpinsrw $7, %ecx, %xmm0, %xmm0
-; AVX2-NEXT: vpaddw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; AVX2-NEXT: vpbroadcastw {{.*#+}} xmm1 = [1,1,1,1,1,1,1,1]
+; AVX2-NEXT: vpaddw %xmm0, %xmm1, %xmm0
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
@@ -2027,33 +2028,120 @@ define i16 @PR66194(i8 %q) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
-; AVX512-LABEL: PR66194:
-; AVX512: # %bb.0: # %entry
-; AVX512-NEXT: xorl %eax, %eax
-; AVX512-NEXT: xorl %ecx, %ecx
-; AVX512-NEXT: testb %dil, %dil
-; AVX512-NEXT: setne %al
-; AVX512-NEXT: sete %cl
-; AVX512-NEXT: vmovd %eax, %xmm0
-; AVX512-NEXT: vpinsrw $1, %ecx, %xmm0, %xmm0
-; AVX512-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
-; AVX512-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0
-; AVX512-NEXT: vpinsrw $4, %ecx, %xmm0, %xmm0
-; AVX512-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0
-; AVX512-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
-; AVX512-NEXT: vpinsrw $7, %ecx, %xmm0, %xmm0
-; AVX512-NEXT: vpaddw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX512-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX512-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX512-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX512-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX512-NEXT: vmovd %xmm0, %eax
-; AVX512-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX512-NEXT: retq
+; AVX512F-LABEL: PR66194:
+; AVX512F: # %bb.0: # %entry
+; AVX512F-NEXT: xorl %eax, %eax
+; AVX512F-NEXT: xorl %ecx, %ecx
+; AVX512F-NEXT: testb %dil, %dil
+; AVX512F-NEXT: setne %al
+; AVX512F-NEXT: sete %cl
+; AVX512F-NEXT: vmovd %eax, %xmm0
+; AVX512F-NEXT: vpinsrw $1, %ecx, %xmm0, %xmm0
+; AVX512F-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
+; AVX512F-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0
+; AVX512F-NEXT: vpinsrw $4, %ecx, %xmm0, %xmm0
+; AVX512F-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0
+; AVX512F-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
+; AVX512F-NEXT: vpinsrw $7, %ecx, %xmm0, %xmm0
+; AVX512F-NEXT: vpbroadcastw {{.*#+}} xmm1 = [1,1,1,1,1,1,1,1]
+; AVX512F-NEXT: vpaddw %xmm0, %xmm1, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512F-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512F-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512F-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vmovd %xmm0, %eax
+; AVX512F-NEXT: # kill: def $ax killed $ax killed $eax
+; AVX512F-NEXT: vzeroupper
+; AVX512F-NEXT: retq
+;
+; AVX512VL-LABEL: PR66194:
+; AVX512VL: # %bb.0: # %entry
+; AVX512VL-NEXT: xorl %eax, %eax
+; AVX512VL-NEXT: xorl %ecx, %ecx
+; AVX512VL-NEXT: testb %dil, %dil
+; AVX512VL-NEXT: setne %al
+; AVX512VL-NEXT: sete %cl
+; AVX512VL-NEXT: vmovd %eax, %xmm0
+; AVX512VL-NEXT: vpinsrw $1, %ecx, %xmm0, %xmm0
+; AVX512VL-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
+; AVX512VL-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0
+; AVX512VL-NEXT: vpinsrw $4, %ecx, %xmm0, %xmm0
+; AVX512VL-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0
+; AVX512VL-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
+; AVX512VL-NEXT: vpinsrw $7, %ecx, %xmm0, %xmm0
+; AVX512VL-NEXT: vpbroadcastd {{.*#+}} xmm1 = [1,1,1,1,1,1,1,1]
+; AVX512VL-NEXT: vpaddw %xmm0, %xmm1, %xmm0
+; AVX512VL-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512VL-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512VL-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512VL-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512VL-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512VL-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512VL-NEXT: vmovd %xmm0, %eax
+; AVX512VL-NEXT: # kill: def $ax killed $ax killed $eax
+; AVX512VL-NEXT: vzeroupper
+; AVX512VL-NEXT: retq
+;
+; AVX512BW-LABEL: PR66194:
+; AVX512BW: # %bb.0: # %entry
+; AVX512BW-NEXT: xorl %eax, %eax
+; AVX512BW-NEXT: xorl %ecx, %ecx
+; AVX512BW-NEXT: testb %dil, %dil
+; AVX512BW-NEXT: setne %al
+; AVX512BW-NEXT: sete %cl
+; AVX512BW-NEXT: vmovd %eax, %xmm0
+; AVX512BW-NEXT: vpinsrw $1, %ecx, %xmm0, %xmm0
+; AVX512BW-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
+; AVX512BW-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0
+; AVX512BW-NEXT: vpinsrw $4, %ecx, %xmm0, %xmm0
+; AVX512BW-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0
+; AVX512BW-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
+; AVX512BW-NEXT: vpinsrw $7, %ecx, %xmm0, %xmm0
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [1,1,1,1,1,1,1,1]
+; AVX512BW-NEXT: vpaddw %xmm0, %xmm1, %xmm0
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512BW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512BW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512BW-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512BW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512BW-NEXT: vmovd %xmm0, %eax
+; AVX512BW-NEXT: # kill: def $ax killed $ax killed $eax
+; AVX512BW-NEXT: vzeroupper
+; AVX512BW-NEXT: retq
+;
+; AVX512BWVL-LABEL: PR66194:
+; AVX512BWVL: # %bb.0: # %entry
+; AVX512BWVL-NEXT: xorl %eax, %eax
+; AVX512BWVL-NEXT: xorl %ecx, %ecx
+; AVX512BWVL-NEXT: testb %dil, %dil
+; AVX512BWVL-NEXT: setne %al
+; AVX512BWVL-NEXT: sete %cl
+; AVX512BWVL-NEXT: vmovd %eax, %xmm0
+; AVX512BWVL-NEXT: vpinsrw $1, %ecx, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vpinsrw $4, %ecx, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vpinsrw $7, %ecx, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vpbroadcastw {{.*#+}} xmm1 = [1,1,1,1,1,1,1,1]
+; AVX512BWVL-NEXT: vpaddw %xmm0, %xmm1, %xmm0
+; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512BWVL-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512BWVL-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512BWVL-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; AVX512BWVL-NEXT: vmovd %xmm0, %eax
+; AVX512BWVL-NEXT: # kill: def $ax killed $ax killed $eax
+; AVX512BWVL-NEXT: vzeroupper
+; AVX512BWVL-NEXT: retq
entry:
%cmp12.i.13 = icmp ne i8 %q, 0
%cond.i15.13 = zext i1 %cmp12.i.13 to i16
diff --git a/llvm/test/CodeGen/X86/x86-interleaved-access.ll b/llvm/test/CodeGen/X86/x86-interleaved-access.ll
index 7cddebdca5cca..88a47fa33e5b6 100644
--- a/llvm/test/CodeGen/X86/x86-interleaved-access.ll
+++ b/llvm/test/CodeGen/X86/x86-interleaved-access.ll
@@ -982,38 +982,38 @@ ret void
define void @interleaved_store_vf32_i8_stride3(<32 x i8> %a, <32 x i8> %b, <32 x i8> %c, ptr %p) nounwind {
; AVX1-LABEL: interleaved_store_vf32_i8_stride3:
; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
-; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm0[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
+; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15,0,1,2,3,4,5]
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm4[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm6 = xmm1[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm7 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm8
-; AVX1-NEXT: vpalignr {{.*#+}} xmm9 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm8[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm9[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm1[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm1[11,12,13,14,15,0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm6
+; AVX1-NEXT: vpalignr {{.*#+}} xmm7 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm8 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm8[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
; AVX1-NEXT: vpalignr {{.*#+}} xmm1 = xmm7[5,6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm3[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[5,6,7,8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4]
; AVX1-NEXT: vmovdqa {{.*#+}} xmm6 = [0,11,6,1,12,7,2,13,8,3,14,9,4,15,10,5]
-; AVX1-NEXT: vpshufb %xmm6, %xmm0, %xmm0
-; AVX1-NEXT: vpshufb %xmm6, %xmm1, %xmm1
+; AVX1-NEXT: vpshufb %xmm6, %xmm5, %xmm5
+; AVX1-NEXT: vpshufb %xmm6, %xmm3, %xmm3
; AVX1-NEXT: vpshufb %xmm6, %xmm2, %xmm2
+; AVX1-NEXT: vpshufb %xmm6, %xmm1, %xmm1
+; AVX1-NEXT: vpshufb %xmm6, %xmm0, %xmm0
; AVX1-NEXT: vpshufb %xmm6, %xmm4, %xmm4
-; AVX1-NEXT: vpshufb %xmm6, %xmm3, %xmm3
-; AVX1-NEXT: vpshufb %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vmovdqu %xmm5, 80(%rdi)
-; AVX1-NEXT: vmovdqu %xmm3, 64(%rdi)
-; AVX1-NEXT: vmovdqu %xmm4, 48(%rdi)
+; AVX1-NEXT: vmovdqu %xmm4, 80(%rdi)
+; AVX1-NEXT: vmovdqu %xmm0, 64(%rdi)
+; AVX1-NEXT: vmovdqu %xmm1, 48(%rdi)
; AVX1-NEXT: vmovdqu %xmm2, 32(%rdi)
-; AVX1-NEXT: vmovdqu %xmm0, 16(%rdi)
-; AVX1-NEXT: vmovdqu %xmm1, (%rdi)
+; AVX1-NEXT: vmovdqu %xmm3, 16(%rdi)
+; AVX1-NEXT: vmovdqu %xmm5, (%rdi)
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
@@ -1073,109 +1073,103 @@ ret void
define void @interleaved_store_vf64_i8_stride3(<64 x i8> %a, <64 x i8> %b, <64 x i8> %c, ptr %p) nounwind {
; AVX1-LABEL: interleaved_store_vf64_i8_stride3:
; AVX1: # %bb.0:
-; AVX1-NEXT: subq $24, %rsp
-; AVX1-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX1-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX1-NEXT: vmovdqa %ymm2, %ymm4
-; AVX1-NEXT: vmovdqa %ymm0, %ymm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm9
+; AVX1-NEXT: vmovdqa %ymm5, %ymm10
+; AVX1-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm7 = [11,12,13,14,15,0,1,2,3,4,5,128,128,128,128,128]
+; AVX1-NEXT: vpshufb %xmm7, %xmm0, %xmm6
+; AVX1-NEXT: vpslldq {{.*#+}} xmm8 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm4[0,1,2,3,4]
+; AVX1-NEXT: vpor %xmm6, %xmm8, %xmm5
+; AVX1-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm11
+; AVX1-NEXT: vpshufb %xmm7, %xmm11, %xmm8
+; AVX1-NEXT: vpshufb %xmm7, %xmm1, %xmm9
+; AVX1-NEXT: vpslldq {{.*#+}} xmm10 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm10[0,1,2,3,4]
+; AVX1-NEXT: vpor %xmm10, %xmm9, %xmm9
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm12
+; AVX1-NEXT: vpshufb %xmm7, %xmm12, %xmm10
; AVX1-NEXT: vmovdqa {{.*#+}} xmm13 = [u,u,u,u,u,128,128,128,128,128,128,6,7,8,9,10]
-; AVX1-NEXT: vpshufb %xmm13, %xmm12, %xmm6
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm14 = [u,u,u,u,u,5,6,7,8,9,10,128,128,128,128,128]
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm11
-; AVX1-NEXT: vpshufb %xmm14, %xmm11, %xmm7
-; AVX1-NEXT: vpor %xmm6, %xmm7, %xmm0
-; AVX1-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vpshufb %xmm13, %xmm1, %xmm7
-; AVX1-NEXT: vpshufb %xmm14, %xmm3, %xmm8
-; AVX1-NEXT: vpor %xmm7, %xmm8, %xmm0
-; AVX1-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vpshufb %xmm13, %xmm9, %xmm8
-; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm15
-; AVX1-NEXT: vpshufb %xmm14, %xmm15, %xmm10
-; AVX1-NEXT: vpor %xmm8, %xmm10, %xmm0
-; AVX1-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm0 = [11,12,13,14,15,0,1,2,3,4,5,128,128,128,128,128]
-; AVX1-NEXT: vpshufb %xmm0, %xmm9, %xmm10
-; AVX1-NEXT: vpshufb %xmm0, %xmm1, %xmm5
-; AVX1-NEXT: vpshufb %xmm0, %xmm12, %xmm12
-; AVX1-NEXT: vpshufb %xmm0, %xmm2, %xmm0
-; AVX1-NEXT: vpshufb %xmm13, %xmm2, %xmm1
-; AVX1-NEXT: vpshufb %xmm14, %xmm4, %xmm2
-; AVX1-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX1-NEXT: vpshufb %xmm13, %xmm12, %xmm7
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm12 = [u,u,u,u,u,5,6,7,8,9,10,128,128,128,128,128]
+; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm14
+; AVX1-NEXT: vpshufb %xmm12, %xmm14, %xmm15
+; AVX1-NEXT: vpor %xmm7, %xmm15, %xmm5
+; AVX1-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT: vpshufb %xmm13, %xmm1, %xmm1
+; AVX1-NEXT: vpshufb %xmm12, %xmm3, %xmm15
+; AVX1-NEXT: vpor %xmm1, %xmm15, %xmm1
; AVX1-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15]
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [u,u,u,u,u,4,6,8,10,12,14,7,9,11,13,15]
-; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX1-NEXT: vpshufb %xmm13, %xmm11, %xmm11
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm15
+; AVX1-NEXT: vpshufb %xmm12, %xmm15, %xmm7
+; AVX1-NEXT: vpor %xmm7, %xmm11, %xmm1
; AVX1-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vextractf128 $1, %ymm6, %xmm9
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm9[8],xmm15[8],xmm9[9],xmm15[9],xmm9[10],xmm15[10],xmm9[11],xmm15[11],xmm9[12],xmm15[12],xmm9[13],xmm15[13],xmm9[14],xmm15[14],xmm9[15],xmm15[15]
-; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm8
-; AVX1-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm3[8],xmm14[9],xmm3[9],xmm14[10],xmm3[10],xmm14[11],xmm3[11],xmm14[12],xmm3[12],xmm14[13],xmm3[13],xmm14[14],xmm3[14],xmm14[15],xmm3[15]
-; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm7
-; AVX1-NEXT: vextractf128 $1, %ymm14, %xmm1
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm1[8],xmm11[8],xmm1[9],xmm11[9],xmm1[10],xmm11[10],xmm1[11],xmm11[11],xmm1[12],xmm11[12],xmm1[13],xmm11[13],xmm1[14],xmm11[14],xmm1[15],xmm11[15]
-; AVX1-NEXT: vpshufb %xmm2, %xmm13, %xmm2
-; AVX1-NEXT: vpslldq {{.*#+}} xmm13 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm1[0,1,2,3,4]
-; AVX1-NEXT: vpor %xmm13, %xmm12, %xmm13
-; AVX1-NEXT: vpalignr {{.*#+}} xmm13 = xmm13[5,6,7,8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4]
-; AVX1-NEXT: vpslldq {{.*#+}} xmm11 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm14[0,1,2,3,4]
-; AVX1-NEXT: vpor %xmm5, %xmm11, %xmm11
-; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm11[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
-; AVX1-NEXT: vpslldq {{.*#+}} xmm11 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm9[0,1,2,3,4]
-; AVX1-NEXT: vpor %xmm11, %xmm10, %xmm11
-; AVX1-NEXT: vpalignr {{.*#+}} xmm11 = xmm11[5,6,7,8,9,10,11,12,13,14,15],xmm15[0,1,2,3,4]
-; AVX1-NEXT: vpslldq {{.*#+}} xmm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm6[0,1,2,3,4]
-; AVX1-NEXT: vpor %xmm0, %xmm15, %xmm15
-; AVX1-NEXT: vpalignr {{.*#+}} xmm4 = xmm15[5,6,7,8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm12[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm7[5,6,7,8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4]
-; AVX1-NEXT: vpalignr {{.*#+}} xmm15 = xmm8[5,6,7,8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4]
-; AVX1-NEXT: vpalignr $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX1-NEXT: # xmm0 = mem[5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4]
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm7 = [5,128,11,6,128,12,7,128,13,8,128,14,9,128,15,10]
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm7, %xmm8, %xmm8
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm10 = [128,5,128,128,6,128,128,7,128,128,8,128,128,9,128,128]
-; AVX1-NEXT: vpshufb %xmm10, %xmm6, %xmm12
-; AVX1-NEXT: vpor %xmm12, %xmm8, %xmm8
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm7, %xmm6, %xmm12
-; AVX1-NEXT: vpshufb %xmm10, %xmm9, %xmm9
-; AVX1-NEXT: vpor %xmm9, %xmm12, %xmm9
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm7, %xmm6, %xmm12
-; AVX1-NEXT: vpshufb %xmm10, %xmm14, %xmm14
-; AVX1-NEXT: vpor %xmm14, %xmm12, %xmm12
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm7, %xmm6, %xmm7
-; AVX1-NEXT: vpshufb %xmm10, %xmm1, %xmm1
-; AVX1-NEXT: vpor %xmm1, %xmm7, %xmm1
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm7 = [0,11,6,1,12,7,2,13,8,3,14,9,4,15,10,5]
-; AVX1-NEXT: vpshufb %xmm7, %xmm4, %xmm4
-; AVX1-NEXT: vpshufb %xmm7, %xmm0, %xmm0
-; AVX1-NEXT: vpshufb %xmm7, %xmm11, %xmm10
-; AVX1-NEXT: vpshufb %xmm7, %xmm15, %xmm6
-; AVX1-NEXT: vpshufb %xmm7, %xmm3, %xmm3
-; AVX1-NEXT: vpshufb %xmm7, %xmm5, %xmm5
-; AVX1-NEXT: vpshufb %xmm7, %xmm13, %xmm11
-; AVX1-NEXT: vpshufb %xmm7, %xmm2, %xmm2
-; AVX1-NEXT: vmovdqu %xmm6, 80(%rdi)
-; AVX1-NEXT: vmovdqu %xmm9, 64(%rdi)
-; AVX1-NEXT: vmovdqu %xmm8, 16(%rdi)
-; AVX1-NEXT: vmovdqu %xmm4, (%rdi)
-; AVX1-NEXT: vmovdqu %xmm10, 48(%rdi)
-; AVX1-NEXT: vmovdqu %xmm0, 32(%rdi)
-; AVX1-NEXT: vmovdqu %xmm2, 176(%rdi)
-; AVX1-NEXT: vmovdqu %xmm1, 160(%rdi)
-; AVX1-NEXT: vmovdqu %xmm12, 112(%rdi)
+; AVX1-NEXT: vpshufb %xmm13, %xmm0, %xmm0
+; AVX1-NEXT: vpshufb %xmm12, %xmm2, %xmm7
+; AVX1-NEXT: vpor %xmm0, %xmm7, %xmm7
+; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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]
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm1 = [u,u,u,u,u,4,6,8,10,12,14,7,9,11,13,15]
+; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm6
+; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm13
+; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm13[8],xmm15[8],xmm13[9],xmm15[9],xmm13[10],xmm15[10],xmm13[11],xmm15[11],xmm13[12],xmm15[12],xmm13[13],xmm15[13],xmm13[14],xmm15[14],xmm13[15],xmm15[15]
+; AVX1-NEXT: vpshufb %xmm1, %xmm12, %xmm5
+; AVX1-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
+; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm11[8],xmm3[8],xmm11[9],xmm3[9],xmm11[10],xmm3[10],xmm11[11],xmm3[11],xmm11[12],xmm3[12],xmm11[13],xmm3[13],xmm11[14],xmm3[14],xmm11[15],xmm3[15]
+; AVX1-NEXT: vpshufb %xmm1, %xmm12, %xmm0
+; AVX1-NEXT: vextractf128 $1, %ymm11, %xmm12
+; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm12[8],xmm14[8],xmm12[9],xmm14[9],xmm12[10],xmm14[10],xmm12[11],xmm14[11],xmm12[12],xmm14[12],xmm12[13],xmm14[13],xmm12[14],xmm14[14],xmm12[15],xmm14[15]
+; AVX1-NEXT: vpshufb %xmm1, %xmm11, %xmm11
+; AVX1-NEXT: vpslldq {{.*#+}} xmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm12[0,1,2,3,4]
+; AVX1-NEXT: vpor %xmm1, %xmm10, %xmm1
+; AVX1-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[5,6,7,8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4]
+; AVX1-NEXT: vpslldq {{.*#+}} xmm14 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm13[0,1,2,3,4]
+; AVX1-NEXT: vpor %xmm14, %xmm8, %xmm14
+; AVX1-NEXT: vpalignr {{.*#+}} xmm14 = xmm14[5,6,7,8,9,10,11,12,13,14,15],xmm15[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm10 = xmm11[5,6,7,8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm3 = xmm9[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[5,6,7,8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[5,6,7,8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4]
+; AVX1-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX1-NEXT: vpalignr {{.*#+}} xmm2 = xmm8[5,6,7,8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4]
+; AVX1-NEXT: vpalignr {{.*#+}} xmm6 = xmm6[5,6,7,8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4]
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm8 = [5,128,11,6,128,12,7,128,13,8,128,14,9,128,15,10]
+; AVX1-NEXT: vpshufb %xmm8, %xmm7, %xmm7
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm9 = [128,5,128,128,6,128,128,7,128,128,8,128,128,9,128,128]
+; AVX1-NEXT: vpshufb %xmm9, %xmm4, %xmm4
+; AVX1-NEXT: vpor %xmm4, %xmm7, %xmm4
+; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX1-NEXT: vpshufb %xmm8, %xmm7, %xmm7
+; AVX1-NEXT: vpshufb %xmm9, %xmm13, %xmm11
+; AVX1-NEXT: vpor %xmm7, %xmm11, %xmm7
+; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX1-NEXT: vpshufb %xmm8, %xmm11, %xmm11
+; AVX1-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX1-NEXT: vpshufb %xmm9, %xmm13, %xmm13
+; AVX1-NEXT: vpor %xmm13, %xmm11, %xmm11
+; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX1-NEXT: vpshufb %xmm8, %xmm13, %xmm8
+; AVX1-NEXT: vpshufb %xmm9, %xmm12, %xmm9
+; AVX1-NEXT: vpor %xmm9, %xmm8, %xmm8
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm9 = [0,11,6,1,12,7,2,13,8,3,14,9,4,15,10,5]
+; AVX1-NEXT: vpshufb %xmm9, %xmm2, %xmm2
+; AVX1-NEXT: vpshufb %xmm9, %xmm6, %xmm6
+; AVX1-NEXT: vpshufb %xmm9, %xmm14, %xmm12
+; AVX1-NEXT: vpshufb %xmm9, %xmm5, %xmm5
+; AVX1-NEXT: vpshufb %xmm9, %xmm3, %xmm3
+; AVX1-NEXT: vpshufb %xmm9, %xmm0, %xmm0
+; AVX1-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX1-NEXT: vpshufb %xmm9, %xmm10, %xmm9
+; AVX1-NEXT: vmovdqu %xmm5, 80(%rdi)
+; AVX1-NEXT: vmovdqu %xmm7, 64(%rdi)
+; AVX1-NEXT: vmovdqu %xmm4, 16(%rdi)
+; AVX1-NEXT: vmovdqu %xmm12, 48(%rdi)
+; AVX1-NEXT: vmovdqu %xmm9, 176(%rdi)
+; AVX1-NEXT: vmovdqu %xmm8, 160(%rdi)
+; AVX1-NEXT: vmovdqu %xmm11, 112(%rdi)
+; AVX1-NEXT: vmovdqu %xmm1, 144(%rdi)
+; AVX1-NEXT: vmovdqu %xmm2, (%rdi)
+; AVX1-NEXT: vmovdqu %xmm6, 32(%rdi)
; AVX1-NEXT: vmovdqu %xmm3, 96(%rdi)
-; AVX1-NEXT: vmovdqu %xmm11, 144(%rdi)
-; AVX1-NEXT: vmovdqu %xmm5, 128(%rdi)
-; AVX1-NEXT: addq $24, %rsp
+; AVX1-NEXT: vmovdqu %xmm0, 128(%rdi)
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
@@ -1274,116 +1268,122 @@ ret void
define <64 x i8> @interleaved_load_vf64_i8_stride3(ptr %ptr){
; AVX1-LABEL: interleaved_load_vf64_i8_stride3:
; AVX1: # %bb.0:
-; AVX1-NEXT: vmovdqu (%rdi), %xmm11
-; AVX1-NEXT: vmovdqu 16(%rdi), %xmm1
-; AVX1-NEXT: vmovdqu 48(%rdi), %xmm13
-; AVX1-NEXT: vmovups 64(%rdi), %xmm0
-; AVX1-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vmovdqu 80(%rdi), %xmm4
-; AVX1-NEXT: vmovdqu 96(%rdi), %xmm5
+; AVX1-NEXT: subq $40, %rsp
+; AVX1-NEXT: .cfi_def_cfa_offset 48
+; AVX1-NEXT: vmovdqu (%rdi), %xmm9
+; AVX1-NEXT: vmovdqu 32(%rdi), %xmm7
+; AVX1-NEXT: vmovdqu 48(%rdi), %xmm11
+; AVX1-NEXT: vmovdqu 96(%rdi), %xmm6
; AVX1-NEXT: vmovdqu 112(%rdi), %xmm2
-; AVX1-NEXT: vmovdqu 144(%rdi), %xmm10
-; AVX1-NEXT: vmovdqu 160(%rdi), %xmm3
-; AVX1-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm9 = [128,128,128,128,128,0,3,6,9,12,15,2,5,8,11,14]
-; AVX1-NEXT: vpshufb %xmm9, %xmm5, %xmm6
-; AVX1-NEXT: vpshufb %xmm9, %xmm10, %xmm7
-; AVX1-NEXT: vpshufb %xmm9, %xmm11, %xmm8
-; AVX1-NEXT: vpshufb %xmm9, %xmm13, %xmm9
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm14 = [1,4,7,10,13,128,128,128,128,128,128,u,u,u,u,u]
-; AVX1-NEXT: vpshufb %xmm14, %xmm5, %xmm5
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm15 = [128,128,128,128,128,0,3,6,9,12,15,u,u,u,u,u]
-; AVX1-NEXT: vpshufb %xmm15, %xmm2, %xmm12
-; AVX1-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vpor %xmm5, %xmm12, %xmm0
+; AVX1-NEXT: vmovdqu 128(%rdi), %xmm0
; AVX1-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vpshufb %xmm14, %xmm10, %xmm10
-; AVX1-NEXT: vpshufb %xmm15, %xmm3, %xmm12
-; AVX1-NEXT: vpor %xmm10, %xmm12, %xmm0
-; AVX1-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vpshufb %xmm14, %xmm11, %xmm11
-; AVX1-NEXT: vmovdqa %xmm1, %xmm0
+; AVX1-NEXT: vmovdqu 144(%rdi), %xmm12
+; AVX1-NEXT: vmovdqu 160(%rdi), %xmm1
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm8 = [1,4,7,10,13,128,128,128,128,128,128,128,128,128,128,128]
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm13 = [128,128,128,128,128,0,3,6,9,12,15,2,5,8,11,14]
+; AVX1-NEXT: vpshufb %xmm8, %xmm7, %xmm4
+; AVX1-NEXT: vpshufb %xmm13, %xmm9, %xmm5
+; AVX1-NEXT: vpor %xmm4, %xmm5, %xmm3
+; AVX1-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT: vpshufb %xmm8, %xmm0, %xmm4
+; AVX1-NEXT: vpshufb %xmm13, %xmm6, %xmm8
+; AVX1-NEXT: vpor %xmm4, %xmm8, %xmm0
+; AVX1-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm15 = [1,4,7,10,13,128,128,128,128,128,128,u,u,u,u,u]
+; AVX1-NEXT: vpshufb %xmm15, %xmm12, %xmm8
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm0 = [128,128,128,128,128,0,3,6,9,12,15,u,u,u,u,u]
+; AVX1-NEXT: vmovdqa %xmm1, %xmm5
+; AVX1-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT: vpshufb %xmm0, %xmm1, %xmm14
+; AVX1-NEXT: vpor %xmm8, %xmm14, %xmm1
; AVX1-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vpshufb %xmm15, %xmm1, %xmm12
-; AVX1-NEXT: vpor %xmm11, %xmm12, %xmm1
+; AVX1-NEXT: vpshufb %xmm15, %xmm6, %xmm6
+; AVX1-NEXT: vpshufb %xmm0, %xmm2, %xmm8
+; AVX1-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT: vpor %xmm6, %xmm8, %xmm1
; AVX1-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT: vpshufb %xmm14, %xmm13, %xmm11
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm15, %xmm1, %xmm13
-; AVX1-NEXT: vpor %xmm11, %xmm13, %xmm11
-; AVX1-NEXT: vpshufb %xmm14, %xmm1, %xmm13
-; AVX1-NEXT: vpshufb %xmm15, %xmm4, %xmm5
-; AVX1-NEXT: vpor %xmm5, %xmm13, %xmm5
-; AVX1-NEXT: vmovdqu 32(%rdi), %xmm1
-; AVX1-NEXT: vpshufb %xmm14, %xmm0, %xmm13
-; AVX1-NEXT: vpshufb %xmm15, %xmm1, %xmm10
-; AVX1-NEXT: vpor %xmm13, %xmm10, %xmm10
-; AVX1-NEXT: vmovdqu 176(%rdi), %xmm13
-; AVX1-NEXT: vpshufb %xmm14, %xmm3, %xmm0
-; AVX1-NEXT: vpshufb %xmm15, %xmm13, %xmm12
-; AVX1-NEXT: vpor %xmm0, %xmm12, %xmm3
-; AVX1-NEXT: vpshufb %xmm14, %xmm2, %xmm12
-; AVX1-NEXT: vmovdqu 128(%rdi), %xmm14
-; AVX1-NEXT: vpshufb %xmm15, %xmm14, %xmm15
-; AVX1-NEXT: vpor %xmm12, %xmm15, %xmm15
+; AVX1-NEXT: vmovdqu 64(%rdi), %xmm6
+; AVX1-NEXT: vpshufb %xmm13, %xmm12, %xmm12
+; AVX1-NEXT: vpshufb %xmm13, %xmm11, %xmm13
+; AVX1-NEXT: vpshufb %xmm15, %xmm11, %xmm11
+; AVX1-NEXT: vpshufb %xmm0, %xmm6, %xmm14
+; AVX1-NEXT: vpor %xmm11, %xmm14, %xmm14
+; AVX1-NEXT: vpshufb %xmm15, %xmm9, %xmm11
+; AVX1-NEXT: vmovdqu 16(%rdi), %xmm9
+; AVX1-NEXT: vpshufb %xmm0, %xmm9, %xmm8
+; AVX1-NEXT: vpor %xmm11, %xmm8, %xmm8
+; AVX1-NEXT: vmovdqu 80(%rdi), %xmm11
+; AVX1-NEXT: vpshufb %xmm15, %xmm6, %xmm1
+; AVX1-NEXT: vpshufb %xmm0, %xmm11, %xmm4
+; AVX1-NEXT: vpor %xmm1, %xmm4, %xmm3
+; AVX1-NEXT: vmovdqu 176(%rdi), %xmm4
+; AVX1-NEXT: vpshufb %xmm15, %xmm5, %xmm1
+; AVX1-NEXT: vpshufb %xmm0, %xmm4, %xmm5
+; AVX1-NEXT: vpor %xmm1, %xmm5, %xmm1
+; AVX1-NEXT: vpshufb %xmm15, %xmm9, %xmm5
+; AVX1-NEXT: vpshufb %xmm0, %xmm7, %xmm10
+; AVX1-NEXT: vpor %xmm5, %xmm10, %xmm5
+; AVX1-NEXT: vpshufb %xmm15, %xmm2, %xmm10
+; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX1-NEXT: vpshufb %xmm0, %xmm2, %xmm0
+; AVX1-NEXT: vpor %xmm0, %xmm10, %xmm15
; AVX1-NEXT: vmovdqa {{.*#+}} xmm0 = [1,4,7,10,13,128,128,128,128,128,128,128,128,128,128,128]
-; AVX1-NEXT: vpshufb %xmm0, %xmm14, %xmm12
-; AVX1-NEXT: vpor %xmm6, %xmm12, %xmm12
-; AVX1-NEXT: vpalignr {{.*#+}} xmm6 = xmm6[11,12,13,14,15],xmm15[0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vpshufb %xmm0, %xmm13, %xmm15
-; AVX1-NEXT: vpor %xmm7, %xmm15, %xmm15
-; AVX1-NEXT: vpalignr {{.*#+}} xmm7 = xmm7[11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX1-NEXT: vpor %xmm2, %xmm8, %xmm2
-; AVX1-NEXT: vpalignr {{.*#+}} xmm8 = xmm8[11,12,13,14,15],xmm10[0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vpshufb %xmm0, %xmm4, %xmm0
-; AVX1-NEXT: vpor %xmm0, %xmm9, %xmm0
-; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm9[11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7,8,9,10]
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,6,7,8,9,10,128,128,128,128,128]
-; AVX1-NEXT: vpshufb %xmm9, %xmm11, %xmm10
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm11 = [128,128,128,128,128,128,128,128,128,128,128,2,5,8,11,14]
-; AVX1-NEXT: vpshufb %xmm11, %xmm4, %xmm4
-; AVX1-NEXT: vpor %xmm4, %xmm10, %xmm4
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm9, %xmm3, %xmm10
-; AVX1-NEXT: vpshufb %xmm11, %xmm1, %xmm1
-; AVX1-NEXT: vpor %xmm1, %xmm10, %xmm1
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm9, %xmm3, %xmm10
-; AVX1-NEXT: vpshufb %xmm11, %xmm13, %xmm13
-; AVX1-NEXT: vpor %xmm13, %xmm10, %xmm10
+; AVX1-NEXT: vpshufb %xmm0, %xmm4, %xmm10
+; AVX1-NEXT: vpor %xmm10, %xmm12, %xmm10
+; AVX1-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT: vpalignr {{.*#+}} xmm12 = xmm12[11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vpshufb %xmm0, %xmm11, %xmm1
+; AVX1-NEXT: vpor %xmm1, %xmm13, %xmm1
+; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm13[11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm10 = [0,1,2,3,4,5,6,7,8,9,10,128,128,128,128,128]
+; AVX1-NEXT: vpshufb %xmm10, %xmm8, %xmm8
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm13 = [128,128,128,128,128,128,128,128,128,128,128,2,5,8,11,14]
+; AVX1-NEXT: vpshufb %xmm13, %xmm7, %xmm7
+; AVX1-NEXT: vpor %xmm7, %xmm8, %xmm7
+; AVX1-NEXT: vpshufb %xmm10, %xmm14, %xmm8
+; AVX1-NEXT: vpshufb %xmm13, %xmm11, %xmm11
+; AVX1-NEXT: vpor %xmm11, %xmm8, %xmm8
; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm9, %xmm3, %xmm9
-; AVX1-NEXT: vpshufb %xmm11, %xmm14, %xmm11
-; AVX1-NEXT: vpor %xmm11, %xmm9, %xmm9
-; AVX1-NEXT: vmovdqa {{.*#+}} xmm11 = [5,6,7,8,9,10,128,128,128,128,128,0,1,2,3,4]
-; AVX1-NEXT: vpshufb %xmm11, %xmm0, %xmm0
+; AVX1-NEXT: vpshufb %xmm10, %xmm3, %xmm11
+; AVX1-NEXT: vpshufb %xmm13, %xmm2, %xmm3
+; AVX1-NEXT: vpor %xmm3, %xmm11, %xmm3
+; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX1-NEXT: vpshufb %xmm10, %xmm2, %xmm10
+; AVX1-NEXT: vpshufb %xmm13, %xmm4, %xmm4
+; AVX1-NEXT: vpor %xmm4, %xmm10, %xmm4
+; AVX1-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
+; AVX1-NEXT: vpalignr {{.*#+}} xmm5 = xmm11[11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vmovdqa {{.*#+}} xmm10 = [5,6,7,8,9,10,128,128,128,128,128,0,1,2,3,4]
+; AVX1-NEXT: vpshufb %xmm10, %xmm11, %xmm11
; AVX1-NEXT: vmovdqa {{.*#+}} xmm13 = [128,128,128,128,128,128,2,5,8,11,14,128,128,128,128,128]
+; AVX1-NEXT: vpshufb %xmm13, %xmm9, %xmm9
+; AVX1-NEXT: vpor %xmm9, %xmm11, %xmm9
+; AVX1-NEXT: vpaddb %xmm7, %xmm9, %xmm7
+; AVX1-NEXT: vpaddb %xmm7, %xmm5, %xmm5
+; AVX1-NEXT: vpshufb %xmm10, %xmm1, %xmm1
+; AVX1-NEXT: vpshufb %xmm13, %xmm6, %xmm6
+; AVX1-NEXT: vpor %xmm6, %xmm1, %xmm1
+; AVX1-NEXT: vpaddb %xmm1, %xmm8, %xmm1
+; AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm1
+; AVX1-NEXT: vmovdqu (%rsp), %ymm2 # 32-byte Reload
+; AVX1-NEXT: vpalignr {{.*#+}} xmm0 = xmm2[11,12,13,14,15],xmm15[0,1,2,3,4,5,6,7,8,9,10]
+; AVX1-NEXT: vpshufb %xmm10, %xmm2, %xmm2
+; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX1-NEXT: vpshufb %xmm13, %xmm6, %xmm6
+; AVX1-NEXT: vpor %xmm6, %xmm2, %xmm2
+; AVX1-NEXT: vpaddb %xmm3, %xmm2, %xmm2
+; AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm2
+; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT: vpshufb %xmm10, %xmm0, %xmm0
; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-NEXT: vpshufb %xmm13, %xmm3, %xmm3
; AVX1-NEXT: vpor %xmm3, %xmm0, %xmm0
; AVX1-NEXT: vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT: vpaddb %xmm0, %xmm5, %xmm0
-; AVX1-NEXT: vpshufb %xmm11, %xmm2, %xmm2
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm13, %xmm3, %xmm3
-; AVX1-NEXT: vpor %xmm3, %xmm2, %xmm2
-; AVX1-NEXT: vpaddb %xmm1, %xmm2, %xmm1
-; AVX1-NEXT: vpaddb %xmm1, %xmm8, %xmm1
-; AVX1-NEXT: vpshufb %xmm11, %xmm15, %xmm2
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm13, %xmm3, %xmm3
-; AVX1-NEXT: vpor %xmm3, %xmm2, %xmm2
-; AVX1-NEXT: vpaddb %xmm2, %xmm10, %xmm2
-; AVX1-NEXT: vpaddb %xmm2, %xmm7, %xmm2
-; AVX1-NEXT: vpshufb %xmm11, %xmm12, %xmm3
-; AVX1-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX1-NEXT: vpshufb %xmm13, %xmm4, %xmm4
-; AVX1-NEXT: vpor %xmm4, %xmm3, %xmm3
-; AVX1-NEXT: vpaddb %xmm3, %xmm9, %xmm3
-; AVX1-NEXT: vpaddb %xmm3, %xmm6, %xmm3
-; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm1
+; AVX1-NEXT: vpaddb %xmm0, %xmm12, %xmm3
+; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm5, %ymm0
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT: addq $40, %rsp
+; AVX1-NEXT: .cfi_def_cfa_offset 8
; AVX1-NEXT: retq
;
; AVX2-LABEL: interleaved_load_vf64_i8_stride3:
diff --git a/llvm/test/CodeGen/X86/zero_extend_vector_inreg.ll b/llvm/test/CodeGen/X86/zero_extend_vector_inreg.ll
index ddd7f10168936..88b7ae0bf552a 100644
--- a/llvm/test/CodeGen/X86/zero_extend_vector_inreg.ll
+++ b/llvm/test/CodeGen/X86/zero_extend_vector_inreg.ll
@@ -7058,21 +7058,22 @@ define void @vec512_v16i32_to_v8i64_factor2(ptr %in.vec.base.ptr, ptr %in.vec.bi
; AVX: # %bb.0:
; AVX-NEXT: vmovdqa (%rdi), %xmm0
; AVX-NEXT: vmovdqa 16(%rdi), %xmm1
-; AVX-NEXT: vpaddb 16(%rsi), %xmm1, %xmm1
; AVX-NEXT: vpaddb (%rsi), %xmm0, %xmm0
-; AVX-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
+; AVX-NEXT: vpaddb 16(%rsi), %xmm1, %xmm1
+; AVX-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
; AVX-NEXT: vpxor %xmm3, %xmm3, %xmm3
-; AVX-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3]
-; AVX-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm1[0],zero,xmm1[1],zero
; AVX-NEXT: vpunpckhdq {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]
-; AVX-NEXT: vpaddb 48(%rdx), %xmm1, %xmm1
-; AVX-NEXT: vpaddb 32(%rdx), %xmm4, %xmm3
+; AVX-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm0[0],zero,xmm0[1],zero
+; AVX-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX-NEXT: vpaddb 16(%rdx), %xmm0, %xmm0
-; AVX-NEXT: vpaddb (%rdx), %xmm2, %xmm2
-; AVX-NEXT: vmovdqa %xmm2, (%rcx)
-; AVX-NEXT: vmovdqa %xmm0, 16(%rcx)
-; AVX-NEXT: vmovdqa %xmm3, 32(%rcx)
+; AVX-NEXT: vpaddb (%rdx), %xmm4, %xmm3
+; AVX-NEXT: vpaddb 48(%rdx), %xmm1, %xmm1
+; AVX-NEXT: vpaddb 32(%rdx), %xmm2, %xmm2
+; AVX-NEXT: vmovdqa %xmm2, 32(%rcx)
; AVX-NEXT: vmovdqa %xmm1, 48(%rcx)
+; AVX-NEXT: vmovdqa %xmm3, (%rcx)
+; AVX-NEXT: vmovdqa %xmm0, 16(%rcx)
+; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: vec512_v16i32_to_v8i64_factor2:
diff --git a/llvm/test/CodeGen/X86/zero_extend_vector_inreg_of_broadcast.ll b/llvm/test/CodeGen/X86/zero_extend_vector_inreg_of_broadcast.ll
index 7ad9fb0c27170..f03db5fa2ceef 100644
--- a/llvm/test/CodeGen/X86/zero_extend_vector_inreg_of_broadcast.ll
+++ b/llvm/test/CodeGen/X86/zero_extend_vector_inreg_of_broadcast.ll
@@ -3283,9 +3283,10 @@ define void @vec384_i8_widen_to_i128_factor16_broadcast_to_v3i128_factor3(ptr %i
; AVX-NEXT: vpaddb (%rdx), %xmm1, %xmm1
; AVX-NEXT: vpaddb 32(%rdx), %xmm0, %xmm2
; AVX-NEXT: vpaddb 16(%rdx), %xmm0, %xmm0
-; AVX-NEXT: vmovdqa %xmm1, (%rcx)
; AVX-NEXT: vmovdqa %xmm0, 16(%rcx)
; AVX-NEXT: vmovdqa %xmm2, 32(%rcx)
+; AVX-NEXT: vmovdqa %xmm1, (%rcx)
+; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: vec384_i8_widen_to_i128_factor16_broadcast_to_v3i128_factor3:
>From bf22ac60f72b2bbf843846c98a830403d54e2268 Mon Sep 17 00:00:00 2001
From: Bjorn Pettersson <bjorn.a.pettersson at ericsson.com>
Date: Wed, 11 Jun 2025 09:53:00 +0200
Subject: [PATCH 4/4] Fixup: Update test checks after rebase
---
.../vector-interleaved-load-i16-stride-4.ll | 1290 ++++++++---------
1 file changed, 597 insertions(+), 693 deletions(-)
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-4.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-4.ll
index 321aba7034498..ae4f85ce42a19 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-4.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-4.ll
@@ -562,18 +562,14 @@ define void @load_i16_stride4_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
;
; AVX2-LABEL: load_i16_stride4_vf8:
; AVX2: # %bb.0:
-; AVX2-NEXT: vmovdqa 32(%rdi), %ymm0
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vmovdqa (%rdi), %ymm2
-; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
+; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-NEXT: vpackusdw %xmm0, %xmm1, %xmm0
+; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-NEXT: vmovdqa 16(%rdi), %xmm2
; AVX2-NEXT: vmovdqa 32(%rdi), %xmm3
@@ -616,18 +612,14 @@ define void @load_i16_stride4_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
;
; AVX2-FP-LABEL: load_i16_stride4_vf8:
; AVX2-FP: # %bb.0:
-; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm0
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm2
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpackusdw %xmm0, %xmm1, %xmm0
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-FP-NEXT: vmovdqa 16(%rdi), %xmm2
; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm3
@@ -668,18 +660,14 @@ define void @load_i16_stride4_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
;
; AVX2-FCP-LABEL: load_i16_stride4_vf8:
; AVX2-FCP: # %bb.0:
-; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm0
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm2
-; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm2
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpackusdw %xmm0, %xmm1, %xmm0
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm2
+; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm2
; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm3
@@ -1033,30 +1021,26 @@ define void @load_i16_stride4_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
;
; AVX2-LABEL: load_i16_stride4_vf16:
; AVX2: # %bb.0:
-; AVX2-NEXT: vmovdqa (%rdi), %ymm0
-; AVX2-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
-; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
-; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
-; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm5
-; AVX2-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
-; AVX2-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
-; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
-; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4],ymm2[5,6,7],ymm1[8],ymm2[9,10,11],ymm1[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm4
-; AVX2-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5,6,7],ymm0[8],ymm2[9,10,11],ymm0[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
+; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
+; AVX2-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
+; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-NEXT: vmovdqa 16(%rdi), %xmm2
; AVX2-NEXT: vmovdqa 32(%rdi), %xmm3
@@ -1141,30 +1125,26 @@ define void @load_i16_stride4_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
;
; AVX2-FP-LABEL: load_i16_stride4_vf16:
; AVX2-FP: # %bb.0:
-; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm0
-; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX2-FP-NEXT: vpxor %xmm2, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
-; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
-; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm5
-; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
-; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
-; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4],ymm2[5,6,7],ymm1[8],ymm2[9,10,11],ymm1[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm4
-; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5,6,7],ymm0[8],ymm2[9,10,11],ymm0[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
+; AVX2-FP-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
+; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-FP-NEXT: vmovdqa 16(%rdi), %xmm3
; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm4
@@ -1245,15 +1225,11 @@ define void @load_i16_stride4_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP: # %bb.0:
; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm1
; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm2
-; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm0
-; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpxor %xmm4, %xmm4, %xmm4
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1,2,3],ymm3[4],ymm4[5,6,7],ymm3[8],ymm4[9,10,11],ymm3[12],ymm4[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm5
-; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm3, %xmm3
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1,2,3],ymm0[4],ymm4[5,6,7],ymm0[8],ymm4[9,10,11],ymm0[12],ymm4[13,14,15]
+; AVX2-FCP-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm4
+; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm4
; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm0, %xmm0
; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
@@ -2130,54 +2106,46 @@ define void @load_i16_stride4_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-LABEL: load_i16_stride4_vf32:
; AVX2: # %bb.0:
; AVX2-NEXT: subq $168, %rsp
-; AVX2-NEXT: vmovdqa 128(%rdi), %ymm0
-; AVX2-NEXT: vmovdqa 160(%rdi), %ymm2
-; AVX2-NEXT: vmovdqa (%rdi), %ymm3
-; AVX2-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm5, %xmm6
-; AVX2-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vpackusdw %xmm5, %xmm5, %xmm5
-; AVX2-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm5
-; AVX2-NEXT: vpblendw {{.*#+}} ymm6 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm6, %xmm7
-; AVX2-NEXT: vpackusdw %xmm7, %xmm6, %xmm6
-; AVX2-NEXT: vpackusdw %xmm6, %xmm6, %xmm6
-; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
-; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm6
-; AVX2-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
-; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm1[1,2,3],ymm3[4],ymm1[5,6,7],ymm3[8],ymm1[9,10,11],ymm3[12],ymm1[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm6
-; AVX2-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
-; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
-; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm5
-; AVX2-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
-; AVX2-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
-; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
-; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm4
-; AVX2-NEXT: vpackusdw %xmm4, %xmm2, %xmm2
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
+; AVX2-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
+; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vmovdqa (%rdi), %xmm1
; AVX2-NEXT: vmovdqa %xmm1, (%rsp) # 16-byte Spill
@@ -2370,54 +2338,46 @@ define void @load_i16_stride4_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FP-LABEL: load_i16_stride4_vf32:
; AVX2-FP: # %bb.0:
; AVX2-FP-NEXT: subq $184, %rsp
-; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm0
-; AVX2-FP-NEXT: vmovdqa 160(%rdi), %ymm2
-; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm3
-; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX2-FP-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm5, %xmm6
-; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
-; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm5, %xmm5
-; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm5
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm6, %xmm7
-; AVX2-FP-NEXT: vpackusdw %xmm7, %xmm6, %xmm6
-; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm6, %xmm6
-; AVX2-FP-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm6
-; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm1[1,2,3],ymm3[4],ymm1[5,6,7],ymm3[8],ymm1[9,10,11],ymm3[12],ymm1[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm6
-; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
-; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
-; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FP-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm5
-; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
-; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
-; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm4
-; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
+; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm13
; AVX2-FP-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -2600,157 +2560,149 @@ define void @load_i16_stride4_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: subq $104, %rsp
; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %ymm3
; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm5
-; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm0
-; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %ymm1
-; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm10
-; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm6
+; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm9
+; AVX2-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm7
-; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm2
-; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpxor %xmm8, %xmm8, %xmm8
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm8[1,2,3],ymm4[4],ymm8[5,6,7],ymm4[8],ymm8[9,10,11],ymm4[12],ymm8[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm4, %xmm6
-; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm8[1,2,3],ymm2[4],ymm8[5,6,7],ymm2[8],ymm8[9,10,11],ymm2[12],ymm8[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm6
-; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm2, %xmm2
-; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm2, %xmm2
-; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,2,2,3,0,2,4,6]
-; AVX2-FCP-NEXT: vpermd %ymm7, %ymm9, %ymm6
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,16,17,20,21,24,25,28,29]
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm11
-; AVX2-FCP-NEXT: vpermd %ymm10, %ymm9, %ymm10
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm12
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4,5],ymm11[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm8[1,2,3],ymm1[4],ymm8[5,6,7],ymm1[8],ymm8[9,10,11],ymm1[12],ymm8[13,14,15]
+; AVX2-FCP-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm8[1,2,3],ymm0[4],ymm8[5,6,7],ymm0[8],ymm8[9,10,11],ymm0[12],ymm8[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm2
-; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm4
+; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm2, %xmm2
+; AVX2-FCP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,2,2,3,0,2,4,6]
+; AVX2-FCP-NEXT: vpermd %ymm7, %ymm2, %ymm5
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,16,17,20,21,24,25,28,29]
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm8
+; AVX2-FCP-NEXT: vpermd %ymm9, %ymm2, %ymm10
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm9
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4,5],ymm8[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm8[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm8
+; AVX2-FCP-NEXT: vpackusdw %xmm8, %xmm1, %xmm1
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm8
+; AVX2-FCP-NEXT: vpackusdw %xmm8, %xmm0, %xmm0
; AVX2-FCP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vpermd %ymm5, %ymm9, %ymm15
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm15, %ymm1
-; AVX2-FCP-NEXT: vpermd %ymm3, %ymm9, %ymm9
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm2
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
+; AVX2-FCP-NEXT: vpermd %ymm6, %ymm2, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm1
+; AVX2-FCP-NEXT: vpermd %ymm3, %ymm2, %ymm6
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm8
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm14
-; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm4
+; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm3
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,2,3,10,11,8,9,10,11,12,13,14,15]
-; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm4, %xmm0
-; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm14, %xmm2
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
+; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm3, %xmm0
+; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm14, %xmm8
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm8[0],xmm0[0],xmm8[1],xmm0[1]
; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm13
-; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm5
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [2,3,10,11,8,9,10,11,8,9,10,11,12,13,14,15]
-; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm8
-; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm13, %xmm11
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm8[0,1],xmm2[2,3]
+; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm2
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [2,3,10,11,8,9,10,11,8,9,10,11,12,13,14,15]
+; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm2, %xmm4
+; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm11
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm11[0],xmm4[0],xmm11[1],xmm4[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],xmm8[2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm6, %ymm6
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm5
; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm10, %ymm10
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm10[0,1,2,3,4,5],ymm6[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm6[4,5,6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm10[0,1,2,3,4,5],ymm5[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %xmm11
-; AVX2-FCP-NEXT: vmovdqa 176(%rdi), %xmm2
-; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm10
-; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm12
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm12[0],xmm10[0],xmm12[1],xmm10[1]
+; AVX2-FCP-NEXT: vmovdqa 176(%rdi), %xmm15
+; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm15, %xmm4
+; AVX2-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm10
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm10[0],xmm4[0],xmm10[1],xmm4[1]
; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %xmm10
; AVX2-FCP-NEXT: vmovdqa 144(%rdi), %xmm12
-; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm12, %xmm0
-; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm3
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm9, %ymm3
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
+; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm12, %xmm0
+; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm9, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm6, %ymm4
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [1,3,2,3,1,3,5,7]
-; AVX2-FCP-NEXT: vpermd %ymm7, %ymm0, %ymm7
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [1,3,2,3,1,3,5,7]
+; AVX2-FCP-NEXT: vpermd %ymm7, %ymm6, %ymm4
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm7 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,16,17,20,21,24,25,28,29]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm0
; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm3
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm1[6,7]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[3,1,2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm1[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm14[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[3,1,2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm14[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[2,0,2,3,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm13[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm5[0,1],xmm3[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm3 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm5 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm6
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm9
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0,1,2,3,4,5],ymm6[6,7]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,1,2,3]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm11[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm9[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm12[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm15[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[3,1,2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm9[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm11[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm15[0],xmm0[1],xmm15[1]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm11[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm12[2,0,2,3,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm10[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm8[0,1],xmm4[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm8[0],xmm15[0],xmm8[1],xmm15[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm4
; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm6
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm7
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3,4,5],ymm4[6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,1,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm14[0,1,3,1,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm13[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm5[0,1],xmm1[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm3
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,1,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm14[0,1,3,1,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm13[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm4[4,5,6,7]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm3
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm9[0,1,3,1,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm11[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm11[0,1,3,1,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm12[3,1,2,3,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm10[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
-; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rsi)
-; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm2, (%rsi)
-; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rdx)
-; AVX2-FCP-NEXT: vmovups (%rsp), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm2, (%rdx)
-; AVX2-FCP-NEXT: vmovdqa %ymm4, 32(%rcx)
-; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm2, (%rcx)
-; AVX2-FCP-NEXT: vmovdqa %ymm0, 32(%r8)
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rsi)
+; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm3, (%rsi)
+; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rdx)
+; AVX2-FCP-NEXT: vmovups (%rsp), %ymm3 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm3, (%rdx)
+; AVX2-FCP-NEXT: vmovdqa %ymm0, 32(%rcx)
+; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FCP-NEXT: vmovaps %ymm0, (%rcx)
+; AVX2-FCP-NEXT: vmovdqa %ymm2, 32(%r8)
; AVX2-FCP-NEXT: vmovdqa %ymm1, (%r8)
; AVX2-FCP-NEXT: addq $104, %rsp
; AVX2-FCP-NEXT: vzeroupper
@@ -4402,102 +4354,86 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-LABEL: load_i16_stride4_vf64:
; AVX2: # %bb.0:
; AVX2-NEXT: subq $696, %rsp # imm = 0x2B8
-; AVX2-NEXT: vmovdqa 128(%rdi), %ymm0
-; AVX2-NEXT: vmovdqa 160(%rdi), %ymm1
-; AVX2-NEXT: vmovdqa 384(%rdi), %ymm3
-; AVX2-NEXT: vmovdqa 416(%rdi), %ymm4
-; AVX2-NEXT: vmovdqa 256(%rdi), %ymm5
-; AVX2-NEXT: vmovdqa 288(%rdi), %ymm6
-; AVX2-NEXT: vmovdqa (%rdi), %ymm7
-; AVX2-NEXT: vmovdqa 32(%rdi), %ymm8
-; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
-; AVX2-NEXT: vpblendw {{.*#+}} ymm9 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm9, %xmm10
-; AVX2-NEXT: vpackusdw %xmm10, %xmm9, %xmm9
-; AVX2-NEXT: vpackusdw %xmm9, %xmm9, %xmm9
-; AVX2-NEXT: vinserti128 $1, %xmm9, %ymm0, %ymm9
-; AVX2-NEXT: vpblendw {{.*#+}} ymm10 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm10, %xmm11
-; AVX2-NEXT: vpackusdw %xmm11, %xmm10, %xmm10
-; AVX2-NEXT: vpackusdw %xmm10, %xmm10, %xmm10
-; AVX2-NEXT: vinserti128 $1, %xmm10, %ymm0, %ymm10
-; AVX2-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm2[1,2,3],ymm8[4],ymm2[5,6,7],ymm8[8],ymm2[9,10,11],ymm8[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm8, %xmm10
-; AVX2-NEXT: vpackusdw %xmm10, %xmm8, %xmm8
-; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0],ymm2[1,2,3],ymm7[4],ymm2[5,6,7],ymm7[8],ymm2[9,10,11],ymm7[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm7, %xmm10
-; AVX2-NEXT: vpackusdw %xmm10, %xmm7, %xmm7
-; AVX2-NEXT: vpackusdw %xmm8, %xmm7, %xmm7
-; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm9[4,5,6,7]
-; AVX2-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpblendw {{.*#+}} ymm7 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm7, %xmm8
-; AVX2-NEXT: vpackusdw %xmm8, %xmm7, %xmm7
-; AVX2-NEXT: vpackusdw %xmm7, %xmm7, %xmm7
-; AVX2-NEXT: vinserti128 $1, %xmm7, %ymm0, %ymm7
-; AVX2-NEXT: vpblendw {{.*#+}} ymm8 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm8, %xmm9
-; AVX2-NEXT: vpackusdw %xmm9, %xmm8, %xmm8
-; AVX2-NEXT: vpackusdw %xmm8, %xmm8, %xmm8
-; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm0, %ymm8
-; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0],ymm2[1,2,3],ymm6[4],ymm2[5,6,7],ymm6[8],ymm2[9,10,11],ymm6[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm6, %xmm8
-; AVX2-NEXT: vpackusdw %xmm8, %xmm6, %xmm6
-; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7],ymm5[8],ymm2[9,10,11],ymm5[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm5, %xmm8
-; AVX2-NEXT: vpackusdw %xmm8, %xmm5, %xmm5
-; AVX2-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]
-; AVX2-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm5, %xmm6
-; AVX2-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vpackusdw %xmm5, %xmm5, %xmm5
-; AVX2-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm5
-; AVX2-NEXT: vpblendw {{.*#+}} ymm6 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm6, %xmm7
-; AVX2-NEXT: vpackusdw %xmm7, %xmm6, %xmm6
-; AVX2-NEXT: vpackusdw %xmm6, %xmm6, %xmm6
-; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
-; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm2[1,2,3],ymm4[4],ymm2[5,6,7],ymm4[8],ymm2[9,10,11],ymm4[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm6
-; AVX2-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
-; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm2[1,2,3],ymm3[4],ymm2[5,6,7],ymm3[8],ymm2[9,10,11],ymm3[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm6
-; AVX2-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
+; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
-; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
+; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
+; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm4, %xmm5
-; AVX2-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
-; AVX2-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
-; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
-; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
-; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4],ymm2[5,6,7],ymm1[8],ymm2[9,10,11],ymm1[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm4
-; AVX2-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5,6,7],ymm0[8],ymm2[9,10,11],ymm0[12],ymm2[13,14,15]
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2
-; AVX2-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
+; AVX2-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
+; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
+; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm3
+; AVX2-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
+; AVX2-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
+; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vmovdqa 256(%rdi), %xmm4
; AVX2-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -4922,102 +4858,86 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FP-LABEL: load_i16_stride4_vf64:
; AVX2-FP: # %bb.0:
; AVX2-FP-NEXT: subq $712, %rsp # imm = 0x2C8
-; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm0
-; AVX2-FP-NEXT: vmovdqa 160(%rdi), %ymm1
-; AVX2-FP-NEXT: vmovdqa 384(%rdi), %ymm3
-; AVX2-FP-NEXT: vmovdqa 416(%rdi), %ymm4
-; AVX2-FP-NEXT: vmovdqa 256(%rdi), %ymm5
-; AVX2-FP-NEXT: vmovdqa 288(%rdi), %ymm6
-; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm7
-; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm8
-; AVX2-FP-NEXT: vpxor %xmm2, %xmm2, %xmm2
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm9 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm9, %xmm10
-; AVX2-FP-NEXT: vpackusdw %xmm10, %xmm9, %xmm9
-; AVX2-FP-NEXT: vpackusdw %xmm9, %xmm9, %xmm9
-; AVX2-FP-NEXT: vinserti128 $1, %xmm9, %ymm0, %ymm9
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm10 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm10, %xmm11
-; AVX2-FP-NEXT: vpackusdw %xmm11, %xmm10, %xmm10
-; AVX2-FP-NEXT: vpackusdw %xmm10, %xmm10, %xmm10
-; AVX2-FP-NEXT: vinserti128 $1, %xmm10, %ymm0, %ymm10
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm2[1,2,3],ymm8[4],ymm2[5,6,7],ymm8[8],ymm2[9,10,11],ymm8[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm8, %xmm10
-; AVX2-FP-NEXT: vpackusdw %xmm10, %xmm8, %xmm8
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0],ymm2[1,2,3],ymm7[4],ymm2[5,6,7],ymm7[8],ymm2[9,10,11],ymm7[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm7, %xmm10
-; AVX2-FP-NEXT: vpackusdw %xmm10, %xmm7, %xmm7
-; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm7, %xmm7
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm9[4,5,6,7]
-; AVX2-FP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm7 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm7, %xmm8
-; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm7, %xmm7
-; AVX2-FP-NEXT: vpackusdw %xmm7, %xmm7, %xmm7
-; AVX2-FP-NEXT: vinserti128 $1, %xmm7, %ymm0, %ymm7
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm8 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm8, %xmm9
-; AVX2-FP-NEXT: vpackusdw %xmm9, %xmm8, %xmm8
-; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm8, %xmm8
-; AVX2-FP-NEXT: vinserti128 $1, %xmm8, %ymm0, %ymm8
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0],ymm2[1,2,3],ymm6[4],ymm2[5,6,7],ymm6[8],ymm2[9,10,11],ymm6[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm6, %xmm8
-; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm6, %xmm6
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7],ymm5[8],ymm2[9,10,11],ymm5[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm5, %xmm8
-; AVX2-FP-NEXT: vpackusdw %xmm8, %xmm5, %xmm5
-; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]
-; AVX2-FP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm5, %xmm6
-; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
-; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm5, %xmm5
-; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm5
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm6, %xmm7
-; AVX2-FP-NEXT: vpackusdw %xmm7, %xmm6, %xmm6
-; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm6, %xmm6
-; AVX2-FP-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm2[1,2,3],ymm4[4],ymm2[5,6,7],ymm4[8],ymm2[9,10,11],ymm4[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm6
-; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm2[1,2,3],ymm3[4],ymm2[5,6,7],ymm3[8],ymm2[9,10,11],ymm3[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm6
-; AVX2-FP-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
+; AVX2-FP-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
+; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
+; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
-; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
+; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
; AVX2-FP-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm3, %xmm3
; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm3, %xmm3
; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = mem[0],ymm2[1,2,3],mem[4],ymm2[5,6,7],mem[8],ymm2[9,10,11],mem[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm4, %xmm5
-; AVX2-FP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
-; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm4, %xmm4
-; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4],ymm2[5,6,7],ymm1[8],ymm2[9,10,11],ymm1[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm4
-; AVX2-FP-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5,6,7],ymm0[8],ymm2[9,10,11],ymm0[12],ymm2[13,14,15]
-; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm2
-; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
-; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm1, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm2, %xmm2
+; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
+; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1,2,3],mem[4],ymm0[5,6,7],mem[8],ymm0[9,10,11],mem[12],ymm0[13,14,15]
+; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm3
+; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
+; AVX2-FP-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
+; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vmovdqa 256(%rdi), %xmm5
; AVX2-FP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -5413,262 +5333,246 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
;
; AVX2-FCP-LABEL: load_i16_stride4_vf64:
; AVX2-FCP: # %bb.0:
-; AVX2-FCP-NEXT: subq $664, %rsp # imm = 0x298
-; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %ymm0
-; AVX2-FCP-NEXT: vmovdqa 320(%rdi), %ymm9
-; AVX2-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 352(%rdi), %ymm10
-; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %ymm4
-; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %ymm5
-; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm8
-; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm7
-; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm2
-; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm1[1,2,3],ymm3[4],ymm1[5,6,7],ymm3[8],ymm1[9,10,11],ymm3[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm6
-; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm3, %xmm3
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm1[1,2,3],ymm2[4],ymm1[5,6,7],ymm2[8],ymm1[9,10,11],ymm2[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm6
-; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm2, %xmm2
-; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm2, %xmm6
-; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,2,2,3,0,2,4,6]
-; AVX2-FCP-NEXT: vpermd %ymm7, %ymm2, %ymm7
+; AVX2-FCP-NEXT: subq $680, %rsp # imm = 0x2A8
+; AVX2-FCP-NEXT: vmovdqa 320(%rdi), %ymm7
; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,16,17,20,21,24,25,28,29]
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm7
-; AVX2-FCP-NEXT: vpermd %ymm8, %ymm2, %ymm8
-; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm8, %ymm8
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqa 352(%rdi), %ymm6
; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm1[1,2,3],ymm5[4],ymm1[5,6,7],ymm5[8],ymm1[9,10,11],ymm5[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm5, %xmm6
-; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm5, %xmm5
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm4, %xmm6
-; AVX2-FCP-NEXT: vpackusdw %xmm6, %xmm4, %xmm4
-; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
-; AVX2-FCP-NEXT: vpermd %ymm10, %ymm2, %ymm13
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm13, %ymm5
-; AVX2-FCP-NEXT: vpermd %ymm9, %ymm2, %ymm10
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm10, %ymm6
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
-; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %ymm6
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm4, %xmm5
-; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm4, %xmm4
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm5
-; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vmovdqa 448(%rdi), %ymm5
+; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm5
; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vmovdqa 480(%rdi), %ymm4
-; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermd %ymm4, %ymm2, %ymm4
+; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm4
; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm4, %ymm4
-; AVX2-FCP-NEXT: vpermd %ymm5, %ymm2, %ymm12
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm12, %ymm5
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %ymm0
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm4
-; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm4
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0],ymm1[1,2,3],ymm4[4],ymm1[5,6,7],ymm4[8],ymm1[9,10,11],ymm4[12],ymm1[13,14,15]
-; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm4
-; AVX2-FCP-NEXT: vpackusdw %xmm4, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vpackusdw %xmm0, %xmm1, %xmm0
-; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %ymm1
-; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm4
-; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermd %ymm4, %ymm2, %ymm6
-; AVX2-FCP-NEXT: vpermd %ymm1, %ymm2, %ymm4
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm4, %ymm2
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
+; AVX2-FCP-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm2
+; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FCP-NEXT: vpackusdw %xmm0, %xmm2, %xmm2
+; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [0,2,2,3,0,2,4,6]
+; AVX2-FCP-NEXT: vpermd %ymm4, %ymm0, %ymm3
+; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,16,17,20,21,24,25,28,29]
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm3
+; AVX2-FCP-NEXT: vpermd %ymm5, %ymm0, %ymm5
+; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm5
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm2, %xmm3
+; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm5
+; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm3, %xmm3
+; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm3, %xmm3
+; AVX2-FCP-NEXT: vpermd %ymm6, %ymm0, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm5
+; AVX2-FCP-NEXT: vpermd %ymm7, %ymm0, %ymm6
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm7
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm5
+; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm3, %xmm3
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm5, %xmm7
+; AVX2-FCP-NEXT: vpackusdw %xmm7, %xmm5, %xmm5
+; AVX2-FCP-NEXT: vmovdqa 448(%rdi), %ymm7
+; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm5, %xmm3
+; AVX2-FCP-NEXT: vmovdqa 480(%rdi), %ymm5
+; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpermd %ymm5, %ymm0, %ymm5
+; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm5
+; AVX2-FCP-NEXT: vpermd %ymm7, %ymm0, %ymm14
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm14, %ymm7
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm3, %xmm5
+; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm3, %xmm3
+; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = mem[0],ymm1[1,2,3],mem[4],ymm1[5,6,7],mem[8],ymm1[9,10,11],mem[12],ymm1[13,14,15]
+; AVX2-FCP-NEXT: vextracti128 $1, %ymm1, %xmm5
+; AVX2-FCP-NEXT: vpackusdw %xmm5, %xmm1, %xmm1
+; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm1, %xmm1
+; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %ymm5
+; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm3
+; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpermd %ymm3, %ymm0, %ymm9
+; AVX2-FCP-NEXT: vpermd %ymm5, %ymm0, %ymm3
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm0
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm5
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %xmm2
-; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 304(%rdi), %xmm0
-; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %xmm1
+; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 304(%rdi), %xmm12
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,2,3,10,11,8,9,10,11,12,13,14,15]
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm2, %xmm1
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm12, %xmm0
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm1
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %xmm0
-; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %xmm10
+; AVX2-FCP-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vmovdqa 272(%rdi), %xmm7
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [2,3,10,11,8,9,10,11,8,9,10,11,12,13,14,15]
-; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm7, %xmm8
-; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm14
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm14[0],xmm8[0],xmm14[1],xmm8[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm13, %ymm13
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm10, %ymm10
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm13[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm10[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm13
+; AVX2-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [2,3,10,11,8,9,10,11,8,9,10,11,12,13,14,15]
+; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm7
+; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm10, %xmm8
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm7[0,1],xmm1[2,3]
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm6, %ymm6
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm2
; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm1
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm10
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm13, %xmm14
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm14 = xmm14[0],xmm10[0],xmm14[1],xmm10[1]
-; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm10
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm6
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm2, %xmm7
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]
+; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm7
; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm15
-; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm15, %xmm0
-; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm10, %xmm9
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm14[2,3]
+; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm15, %xmm11
+; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm13
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm6 = xmm11[0,1],xmm6[2,3]
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm8, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm11
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm8, %ymm14
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm14[0,1,2,3,4,5],ymm9[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm13
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm13[0,1,2,3,4,5],ymm11[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm11[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %xmm8
; AVX2-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 176(%rdi), %xmm0
-; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm9
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]
+; AVX2-FCP-NEXT: vmovdqa 176(%rdi), %xmm6
+; AVX2-FCP-NEXT: vmovdqa %xmm6, (%rsp) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm6
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm11
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm6 = xmm11[0],xmm6[0],xmm11[1],xmm6[1]
; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %xmm8
; AVX2-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 144(%rdi), %xmm9
-; AVX2-FCP-NEXT: vmovdqa %xmm9, (%rsp) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm9, %xmm9
-; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm8, %xmm14
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm9 = xmm14[0],xmm9[0],xmm14[1],xmm9[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm9[0,1],xmm0[2,3]
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm6, %ymm9
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm4
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm9[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
-; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %xmm9
-; AVX2-FCP-NEXT: vmovdqa 432(%rdi), %xmm8
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm0
-; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm9, %xmm4
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1]
+; AVX2-FCP-NEXT: vmovdqa 144(%rdi), %xmm11
+; AVX2-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm11, %xmm11
+; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm8, %xmm13
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm6 = xmm11[0,1],xmm6[2,3]
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm3, %ymm3
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm9[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3],ymm3[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %xmm6
+; AVX2-FCP-NEXT: vmovdqa 432(%rdi), %xmm13
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm13, %xmm3
+; AVX2-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm5
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm9 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %xmm5
-; AVX2-FCP-NEXT: vmovdqa 400(%rdi), %xmm4
-; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm14
-; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm5, %xmm2
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm14[0],xmm2[1],xmm14[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm2
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm12, %ymm12
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm12[0,1,2,3,4,5],ymm2[6,7]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-FCP-NEXT: vmovdqa 400(%rdi), %xmm3
+; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm11
+; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm5, %xmm0
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm9[2,3]
+; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm14, %ymm11
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [1,3,2,3,1,3,5,7]
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
-; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
-; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm2
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm12
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm12[0,1,2,3,4,5],ymm2[6,7]
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm11 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm11
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]
; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm13[3,1,2,3]
-; AVX2-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,1,2,3]
+; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm6[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm12[0],xmm1[0],xmm12[1],xmm1[1]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm15[3,1,2,3]
-; AVX2-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm10[3,1,2,3]
-; AVX2-FCP-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm6[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm10[0,1],xmm1[2,3]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm15[3,1,2,3]
+; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[3,1,2,3]
+; AVX2-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm9[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
-; AVX2-FCP-NEXT: # xmm2 = mem[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm12[3,1,2,3]
; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm14 = mem[3,1,2,3]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm14[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm7[3,1,2,3]
-; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
-; AVX2-FCP-NEXT: # xmm10 = mem[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm13 = xmm12[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm10[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm13 = xmm15[0],xmm13[0],xmm15[1],xmm13[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm13[0,1],xmm2[2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm14[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1]
+; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
+; AVX2-FCP-NEXT: # xmm11 = mem[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
+; AVX2-FCP-NEXT: # xmm9 = mem[3,1,2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm11[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm9[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm12 = xmm15[0],xmm12[0],xmm15[1],xmm12[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm12[0,1],xmm2[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm8[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm13[3,1,2,3]
; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm9[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[3,1,2,3]
; AVX2-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm6[0,1,2,0,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm9[0],xmm2[0],xmm9[1],xmm2[1]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm4[3,1,2,3]
-; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm5[3,1,2,3]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm13[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm9[2,0,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
-; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm4[0,1],xmm2[2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,1,2,0,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm3[3,1,2,3]
+; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm5[3,1,2,3]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm13[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm12[2,0,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
+; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload
; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 32-byte Folded Reload
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm8, %ymm0
-; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm0
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm7, %ymm1
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3,4,5],ymm0[6,7]
-; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
+; AVX2-FCP-NEXT: vpshufd $231, (%rsp), %xmm6 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm6 = mem[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm5 = mem[3,1,2,3]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm6[0,1,2,0,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm5[0,1,2,0,4,5,6,7]
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; AVX2-FCP-NEXT: vpshufd $231, (%rsp), %xmm4 # 16-byte Folded Reload
+; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm4 = mem[3,1,2,3]
; AVX2-FCP-NEXT: vpshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm2 = mem[3,1,2,3]
@@ -5679,9 +5583,9 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm0
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FCP-NEXT: vpshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm1 = mem[0,1,3,1,4,5,6,7]
@@ -5696,21 +5600,21 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm15 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm0
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FCP-NEXT: vpshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm1 = mem[0,1,3,1,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm14[0,1,3,1,4,5,6,7]
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm12[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm10[0],xmm3[0],xmm10[1],xmm3[1]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm11[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm9[0],xmm3[0],xmm9[1],xmm3[1]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm8, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm3
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm3
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm6[0,1,3,1,4,5,6,7]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,3,1,4,5,6,7]
@@ -5721,9 +5625,9 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm2
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm3
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm3, %ymm3
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FCP-NEXT: vpshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm3 = mem[0,1,3,1,4,5,6,7]
@@ -5731,7 +5635,7 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: # xmm4 = mem[0,1,3,1,4,5,6,7]
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm13[3,1,2,3,4,5,6,7]
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm9[3,1,2,3,4,5,6,7]
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm12[3,1,2,3,4,5,6,7]
; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]
; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
@@ -5763,7 +5667,7 @@ define void @load_i16_stride4_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX2-FCP-NEXT: vmovdqa %ymm1, 32(%r8)
; AVX2-FCP-NEXT: vmovdqa %ymm0, 64(%r8)
; AVX2-FCP-NEXT: vmovdqa %ymm15, (%r8)
-; AVX2-FCP-NEXT: addq $664, %rsp # imm = 0x298
+; AVX2-FCP-NEXT: addq $680, %rsp # imm = 0x2A8
; AVX2-FCP-NEXT: vzeroupper
; AVX2-FCP-NEXT: retq
;
More information about the llvm-commits
mailing list