[x86] lower shuffle of extracts to AVX2 vperm instructions

[spatel]

Summary

I was trying to prevent shuffle regressions while matching more horizontal ops and ended up here:

shuf (extract X, 0), (extract X, 4), Mask --> extract (shuf X, undef, Mask'), 0

I think the affected tests were added for:
https://bugs.llvm.org/show_bug.cgi?id=34380

This patch won't change the examples in the bug report itself, but we might be able to extend this to catch more types.

[RKSimon]

LGTM with a few minors

lib/Target/X86/X86ISelLowering.cpp

11683 ↗

(On Diff #181907)

Really these should be safe now, but it might be better to use:

SDValue ExtIdx0 = N0.getOperand(1);
const APInt &ExtIndex0 = cast<ConstantSDNode>(ExtIdx0)->getAPIntValue();

I keep wondering whether we should have made getConstantOperandVal return APInt&

11698 ↗

(On Diff #181907)

NewMask.append(NumElts, -1);

test/CodeGen/X86/avx512-shuffles/partial_permute.ll

1052 ↗

(On Diff #181907)

This looks like a missed opportunity to merge a avx512 mask select across an extract_subvector(vec,0) ? There a few more below as well, please can you raise a bug about them.

vmovdqa {{.*#+}} ymm2 = <5,3,2,5,u,u,u,u>
vptestnmd %xmm1, %xmm1, %k1
vpermd %ymm0, %ymm2, %ymm0 {%k1} {z}

[spatel]

test/CodeGen/X86/avx512-shuffles/partial_permute.ll

1052 ↗

(On Diff #181907)

Filed here:
https://bugs.llvm.org/show_bug.cgi?id=40332

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

static SDValue lowerVectorShuffleAsTruncBroadcast(const SDLoc &DL, MVT VT,
   if (const int OffsetIdx = BroadcastIdx % Scale)
     Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
                          DAG.getConstant(OffsetIdx * EltSize, DL, MVT::i8));
 
   return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
                      DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
 }
 
+/// Test whether this can be lowered with a single SHUFPS instruction.
+///
+/// This is used to disable more specialized lowerings when the shufps lowering
+/// will happen to be efficient.
+static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
+  // This routine only handles 128-bit shufps.
+  assert(Mask.size() == 4 && "Unsupported mask size!");
+  assert(Mask[0] >= -1 && Mask[0] < 8 && "Out of bound mask element!");
+  assert(Mask[1] >= -1 && Mask[1] < 8 && "Out of bound mask element!");
+  assert(Mask[2] >= -1 && Mask[2] < 8 && "Out of bound mask element!");
+  assert(Mask[3] >= -1 && Mask[3] < 8 && "Out of bound mask element!");
+
+  // To lower with a single SHUFPS we need to have the low half and high half
+  // each requiring a single input.
+  if (Mask[0] >= 0 && Mask[1] >= 0 && (Mask[0] < 4) != (Mask[1] < 4))
+    return false;
+  if (Mask[2] >= 0 && Mask[3] >= 0 && (Mask[2] < 4) != (Mask[3] < 4))
+    return false;
+
+  return true;
+}
+
+/// If we are extracting two 128-bit halves of a vector and shuffling the
+/// result, match that to a 256-bit AVX2 vperm* instruction to avoid a
+/// multi-shuffle lowering.
+static SDValue lowerShuffleOfExtractsAsVperm(const SDLoc &DL, SDValue N0,
+                                             SDValue N1, ArrayRef<int> Mask,
+                                             SelectionDAG &DAG) {
+  EVT VT = N0.getValueType();
+  assert((VT.is128BitVector() &&
+          (VT.getScalarSizeInBits() == 32 || VT.getScalarSizeInBits() == 64)) &&
+         "VPERM* family of shuffles requires 32-bit or 64-bit elements");
+
+  // Check that both sources are extracts of the same source vector.
+  if (!N0.hasOneUse() || !N1.hasOneUse() ||
+      N0.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
+      N1.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
+      N0.getOperand(0) != N1.getOperand(0))
+    return SDValue();
+
+  SDValue WideVec = N0.getOperand(0);
+  EVT WideVT = WideVec.getValueType();
+  if (!WideVT.is256BitVector() || !isa<ConstantSDNode>(N0.getOperand(1)) ||
+      !isa<ConstantSDNode>(N1.getOperand(1)))
+    return SDValue();
+
+  // Match extracts of each half of the wide source vector. Commute the shuffle
+  // if the extract of the low half is N1.
+  unsigned NumElts = VT.getVectorNumElements();
+  SmallVector<int, 4> NewMask(Mask.begin(), Mask.end());
+  APInt ExtIndex0 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
+  APInt ExtIndex1 = cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
+  if (ExtIndex1 == 0 && ExtIndex0 == NumElts) {
+    std::swap(ExtIndex0, ExtIndex1);
+    ShuffleVectorSDNode::commuteMask(NewMask);
+  }
+  if (ExtIndex0 != 0 || ExtIndex1 != NumElts)
+    return SDValue();
+
+  // Final bailout: if the mask is simple, we are better off using an extract
+  // and a simple narrow shuffle.
+  if (NumElts == 4 && isSingleSHUFPSMask(NewMask))
+    return SDValue();
+
+  // Extend the shuffle mask with undef elements.
+  NewMask.append(NumElts, -1);
+
+  // shuf (extract X, 0), (extract X, 4), M --> extract (shuf X, undef, M'), 0
+  SDValue Shuf = DAG.getVectorShuffle(WideVT, DL, WideVec, DAG.getUNDEF(WideVT),
+                                      NewMask);
+  // This is free: ymm -> xmm.
+  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Shuf,
+                     DAG.getIntPtrConstant(0, DL));
+}
+
 /// Try to lower broadcast of a single element.
 ///
 /// For convenience, this code also bundles all of the subtarget feature set
 /// filtering. While a little annoying to re-dispatch on type here, there isn't
 /// a convenient way to factor it out.
 static SDValue lowerVectorShuffleAsBroadcast(const SDLoc &DL, MVT VT,
                                              SDValue V1, SDValue V2,
                                              ArrayRef<int> Mask,
return DAG.getNode(
         Mask[1] == SM_SentinelUndef ? DAG.getUNDEF(MVT::v2f64) : V1,
         DAG.getConstant(SHUFPDMask, DL, MVT::i8));
   }
   assert(Mask[0] >= 0 && "No undef lanes in multi-input v2 shuffles!");
   assert(Mask[1] >= 0 && "No undef lanes in multi-input v2 shuffles!");
   assert(Mask[0] < 2 && "We sort V1 to be the first input.");
   assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
 
+  if (Subtarget.hasAVX2())
+    if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG))
+      return Extract;
+
   // When loading a scalar and then shuffling it into a vector we can often do
   // the insertion cheaply.
   if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
           DL, MVT::v2f64, V1, V2, Mask, Zeroable, Subtarget, DAG))
     return Insertion;
   // Try inverting the insertion since for v2 masks it is easy to do and we
   // can't reliably sort the mask one way or the other.
   int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
return DAG.getBitcast(
         DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
                     getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
   }
   assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
   assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
   assert(Mask[0] < 2 && "We sort V1 to be the first input.");
   assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
 
+  if (Subtarget.hasAVX2())
+    if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG))
+      return Extract;
+
   // Try to use shift instructions.
   if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask,
                                                 Zeroable, Subtarget, DAG))
     return Shift;
 
   // When loading a scalar and then shuffling it into a vector we can often do
   // the insertion cheaply.
   if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
static SDValue lowerV2I64VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask,
   // However, all the alternatives are still more cycles and newer chips don't
   // have this problem. It would be really nice if x86 had better shuffles here.
   V1 = DAG.getBitcast(MVT::v2f64, V1);
   V2 = DAG.getBitcast(MVT::v2f64, V2);
   return DAG.getBitcast(MVT::v2i64,
                         DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
 }
 
-/// Test whether this can be lowered with a single SHUFPS instruction.
-///
-/// This is used to disable more specialized lowerings when the shufps lowering
-/// will happen to be efficient.
-static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
-  // This routine only handles 128-bit shufps.
-  assert(Mask.size() == 4 && "Unsupported mask size!");
-  assert(Mask[0] >= -1 && Mask[0] < 8 && "Out of bound mask element!");
-  assert(Mask[1] >= -1 && Mask[1] < 8 && "Out of bound mask element!");
-  assert(Mask[2] >= -1 && Mask[2] < 8 && "Out of bound mask element!");
-  assert(Mask[3] >= -1 && Mask[3] < 8 && "Out of bound mask element!");
-
-  // To lower with a single SHUFPS we need to have the low half and high half
-  // each requiring a single input.
-  if (Mask[0] >= 0 && Mask[1] >= 0 && (Mask[0] < 4) != (Mask[1] < 4))
-    return false;
-  if (Mask[2] >= 0 && Mask[3] >= 0 && (Mask[2] < 4) != (Mask[3] < 4))
-    return false;
-
-  return true;
-}
-
 /// Lower a vector shuffle using the SHUFPS instruction.
 ///
 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
 /// It makes no assumptions about whether this is the *best* lowering, it simply
 /// uses it.
 static SDValue lowerVectorShuffleWithSHUFPS(const SDLoc &DL, MVT VT,
                                             ArrayRef<int> Mask, SDValue V1,
                                             SDValue V2, SelectionDAG &DAG) {
if (NumV2Elements == 0) {
     }
 
     // Otherwise, use a straight shuffle of a single input vector. We pass the
     // input vector to both operands to simulate this with a SHUFPS.
     return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
                        getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
   }
 
+  if (Subtarget.hasAVX2())
+    if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG))
+      return Extract;
+
   // There are special ways we can lower some single-element blends. However, we
   // have custom ways we can lower more complex single-element blends below that
   // we defer to if both this and BLENDPS fail to match, so restrict this to
   // when the V2 input is targeting element 0 of the mask -- that is the fast
   // case here.
   if (NumV2Elements == 1 && Mask[0] >= 4)
     if (SDValue V = lowerVectorShuffleAsElementInsertion(
             DL, MVT::v4f32, V1, V2, Mask, Zeroable, Subtarget, DAG))
if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
       Mask = UnpackLoMask;
     else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
       Mask = UnpackHiMask;
 
     return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
                        getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
   }
 
+  if (Subtarget.hasAVX2())
+    if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG))
+      return Extract;
+
   // Try to use shift instructions.
   if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask,
                                                 Zeroable, Subtarget, DAG))
     return Shift;
 
   // There are special ways we can lower some single-element blends.
   if (NumV2Elements == 1)
     if (SDValue V = lowerVectorShuffleAsElementInsertion(

llvm/trunk/test/CodeGen/X86/avx512-shuffles/partial_permute.ll

; CHECK-NEXT:    retq
   %cmp = icmp eq <8 x i16> %mask, zeroinitializer
   %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> zeroinitializer
   ret <8 x i16> %res
 }
 
 define <4 x i32> @test_8xi32_to_4xi32_perm_mask0(<8 x i32> %vec) {
 ; CHECK-LABEL: test_8xi32_to_4xi32_perm_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,3,2]
-; CHECK-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm1 = <4,0,3,2,u,u,u,u>
+; CHECK-NEXT:    vpermps %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %res = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 4, i32 0, i32 3, i32 2>
   ret <4 x i32> %res
 }
 define <4 x i32> @test_masked_8xi32_to_4xi32_perm_mask0(<8 x i32> %vec, <4 x i32> %vec2, <4 x i32> %mask) {
 ; CHECK-LABEL: test_masked_8xi32_to_4xi32_perm_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm3
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,3,2]
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0],xmm0[1,2,3]
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm3 = <4,0,3,2,u,u,u,u>
+; CHECK-NEXT:    vpermd %ymm0, %ymm3, %ymm0
 ; CHECK-NEXT:    vptestnmd %xmm2, %xmm2, %k1
 ; CHECK-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 4, i32 0, i32 3, i32 2>
   %cmp = icmp eq <4 x i32> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %vec2
   ret <4 x i32> %res
 }
 
 define <4 x i32> @test_masked_z_8xi32_to_4xi32_perm_mask0(<8 x i32> %vec, <4 x i32> %mask) {
 ; CHECK-LABEL: test_masked_z_8xi32_to_4xi32_perm_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,3,2]
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0],xmm0[1,2,3]
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm2 = <4,0,3,2,u,u,u,u>
+; CHECK-NEXT:    vpermd %ymm0, %ymm2, %ymm0
 ; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
 ; CHECK-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 4, i32 0, i32 3, i32 2>
   %cmp = icmp eq <4 x i32> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
   ret <4 x i32> %res
 }
 define <4 x i32> @test_masked_8xi32_to_4xi32_perm_mask1(<8 x i32> %vec, <4 x i32> %vec2, <4 x i32> %mask) {
 ; CHECK-LABEL: test_masked_8xi32_to_4xi32_perm_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm3
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,0,2,3]
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm3[2],xmm0[3]
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm3 = <3,0,7,3,u,u,u,u>
+; CHECK-NEXT:    vpermd %ymm0, %ymm3, %ymm0
 ; CHECK-NEXT:    vptestnmd %xmm2, %xmm2, %k1
 ; CHECK-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 3, i32 0, i32 7, i32 3>
   %cmp = icmp eq <4 x i32> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %vec2
   ret <4 x i32> %res
 }
 
 define <4 x i32> @test_masked_z_8xi32_to_4xi32_perm_mask1(<8 x i32> %vec, <4 x i32> %mask) {
 ; CHECK-LABEL: test_masked_z_8xi32_to_4xi32_perm_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,0,2,3]
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm2[2],xmm0[3]
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm2 = <3,0,7,3,u,u,u,u>
+; CHECK-NEXT:    vpermd %ymm0, %ymm2, %ymm0
 ; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
 ; CHECK-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 3, i32 0, i32 7, i32 3>
   %cmp = icmp eq <4 x i32> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
   ret <4 x i32> %res
; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 6, i32 7, i32 2, i32 3>
   %cmp = icmp eq <4 x i32> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
   ret <4 x i32> %res
 }
 define <4 x i32> @test_8xi32_to_4xi32_perm_mask3(<8 x i32> %vec) {
 ; CHECK-LABEL: test_8xi32_to_4xi32_perm_mask3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; CHECK-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,3,2,1]
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm1 = <5,3,2,5,u,u,u,u>
+; CHECK-NEXT:    vpermps %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %res = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 5, i32 3, i32 2, i32 5>
   ret <4 x i32> %res
 }
 define <4 x i32> @test_masked_8xi32_to_4xi32_perm_mask3(<8 x i32> %vec, <4 x i32> %vec2, <4 x i32> %mask) {
 ; CHECK-LABEL: test_masked_8xi32_to_4xi32_perm_mask3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm3
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm3 = <5,3,2,5,u,u,u,u>
+; CHECK-NEXT:    vpermd %ymm0, %ymm3, %ymm0
 ; CHECK-NEXT:    vptestnmd %xmm2, %xmm2, %k1
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm1 {%k1} = xmm0[1,3,2,1]
-; CHECK-NEXT:    vmovdqa %xmm1, %xmm0
+; CHECK-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 5, i32 3, i32 2, i32 5>
   %cmp = icmp eq <4 x i32> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %vec2
   ret <4 x i32> %res
 }
 
 define <4 x i32> @test_masked_z_8xi32_to_4xi32_perm_mask3(<8 x i32> %vec, <4 x i32> %mask) {
 ; CHECK-LABEL: test_masked_z_8xi32_to_4xi32_perm_mask3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm2 = <5,3,2,5,u,u,u,u>
+; CHECK-NEXT:    vpermd %ymm0, %ymm2, %ymm0
 ; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm0 {%k1} {z} = xmm0[1,3,2,1]
+; CHECK-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 5, i32 3, i32 2, i32 5>
   %cmp = icmp eq <4 x i32> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
   ret <4 x i32> %res
 }
 define <4 x i32> @test_8xi32_to_4xi32_perm_mem_mask0(<8 x i32>* %vp) {
; CHECK-NEXT:    retq
   %cmp = icmp eq <4 x i32> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
   ret <4 x i32> %res
 }
 
 define <2 x i64> @test_4xi64_to_2xi64_perm_mask0(<4 x i64> %vec) {
 ; CHECK-LABEL: test_4xi64_to_2xi64_perm_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; CHECK-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; CHECK-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,0,2,3]
+; CHECK-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %res = shufflevector <4 x i64> %vec, <4 x i64> undef, <2 x i32> <i32 2, i32 0>
   ret <2 x i64> %res
 }
 define <2 x i64> @test_masked_4xi64_to_2xi64_perm_mask0(<4 x i64> %vec, <2 x i64> %vec2, <2 x i64> %mask) {
 ; CHECK-LABEL: test_masked_4xi64_to_2xi64_perm_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm3
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,0,2,3]
 ; CHECK-NEXT:    vptestnmq %xmm2, %xmm2, %k1
-; CHECK-NEXT:    vpunpcklqdq {{.*#+}} xmm1 {%k1} = xmm3[0],xmm0[0]
-; CHECK-NEXT:    vmovdqa %xmm1, %xmm0
+; CHECK-NEXT:    vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <2 x i32> <i32 2, i32 0>
   %cmp = icmp eq <2 x i64> %mask, zeroinitializer
   %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> %vec2
   ret <2 x i64> %res
 }
 
 define <2 x i64> @test_masked_z_4xi64_to_2xi64_perm_mask0(<4 x i64> %vec, <2 x i64> %mask) {
 ; CHECK-LABEL: test_masked_z_4xi64_to_2xi64_perm_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm2
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,0,2,3]
 ; CHECK-NEXT:    vptestnmq %xmm1, %xmm1, %k1
-; CHECK-NEXT:    vpunpcklqdq {{.*#+}} xmm0 {%k1} {z} = xmm2[0],xmm0[0]
+; CHECK-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <2 x i32> <i32 2, i32 0>
   %cmp = icmp eq <2 x i64> %mask, zeroinitializer
   %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> zeroinitializer
   ret <2 x i64> %res
 }
 define <2 x i64> @test_masked_4xi64_to_2xi64_perm_mask1(<4 x i64> %vec, <2 x i64> %vec2, <2 x i64> %mask) {
 ; CHECK-LABEL: test_masked_4xi64_to_2xi64_perm_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm3
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
 ; CHECK-NEXT:    vptestnmq %xmm2, %xmm2, %k1
 ; CHECK-NEXT:    vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <2 x i32> <i32 2, i32 1>
   %cmp = icmp eq <2 x i64> %mask, zeroinitializer
   %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> %vec2
   ret <2 x i64> %res
 }
 
 define <2 x i64> @test_masked_z_4xi64_to_2xi64_perm_mask1(<4 x i64> %vec, <2 x i64> %mask) {
 ; CHECK-LABEL: test_masked_z_4xi64_to_2xi64_perm_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
 ; CHECK-NEXT:    vptestnmq %xmm1, %xmm1, %k1
 ; CHECK-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <2 x i32> <i32 2, i32 1>
   %cmp = icmp eq <2 x i64> %mask, zeroinitializer
   %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> zeroinitializer
   ret <2 x i64> %res
; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x float> %vec, <8 x float> undef, <4 x i32> <i32 0, i32 3, i32 4, i32 5>
   %cmp = fcmp oeq <4 x float> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x float> %shuf, <4 x float> zeroinitializer
   ret <4 x float> %res
 }
 define <4 x float> @test_masked_8xfloat_to_4xfloat_perm_mask1(<8 x float> %vec, <4 x float> %vec2, <4 x float> %mask) {
 ; CHECK-LABEL: test_masked_8xfloat_to_4xfloat_perm_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm3
-; CHECK-NEXT:    vshufps {{.*#+}} xmm3 = xmm3[1,0],xmm0[0,0]
-; CHECK-NEXT:    vxorps %xmm4, %xmm4, %xmm4
-; CHECK-NEXT:    vcmpeqps %xmm4, %xmm2, %k1
-; CHECK-NEXT:    vshufps {{.*#+}} xmm1 {%k1} = xmm0[1,3],xmm3[0,2]
-; CHECK-NEXT:    vmovaps %xmm1, %xmm0
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm3 = <1,3,5,0,u,u,u,u>
+; CHECK-NEXT:    vpermps %ymm0, %ymm3, %ymm0
+; CHECK-NEXT:    vxorps %xmm3, %xmm3, %xmm3
+; CHECK-NEXT:    vcmpeqps %xmm3, %xmm2, %k1
+; CHECK-NEXT:    vblendmps %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x float> %vec, <8 x float> undef, <4 x i32> <i32 1, i32 3, i32 5, i32 0>
   %cmp = fcmp oeq <4 x float> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x float> %shuf, <4 x float> %vec2
   ret <4 x float> %res
 }
 
 define <4 x float> @test_masked_z_8xfloat_to_4xfloat_perm_mask1(<8 x float> %vec, <4 x float> %mask) {
 ; CHECK-LABEL: test_masked_z_8xfloat_to_4xfloat_perm_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[1,0],xmm0[0,0]
-; CHECK-NEXT:    vxorps %xmm3, %xmm3, %xmm3
-; CHECK-NEXT:    vcmpeqps %xmm3, %xmm1, %k1
-; CHECK-NEXT:    vshufps {{.*#+}} xmm0 {%k1} {z} = xmm0[1,3],xmm2[0,2]
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm2 = <1,3,5,0,u,u,u,u>
+; CHECK-NEXT:    vpermps %ymm0, %ymm2, %ymm0
+; CHECK-NEXT:    vxorps %xmm2, %xmm2, %xmm2
+; CHECK-NEXT:    vcmpeqps %xmm2, %xmm1, %k1
+; CHECK-NEXT:    vmovaps %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x float> %vec, <8 x float> undef, <4 x i32> <i32 1, i32 3, i32 5, i32 0>
   %cmp = fcmp oeq <4 x float> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x float> %shuf, <4 x float> zeroinitializer
   ret <4 x float> %res
 }
 define <4 x float> @test_masked_8xfloat_to_4xfloat_perm_mask2(<8 x float> %vec, <4 x float> %vec2, <4 x float> %mask) {
 ; CHECK-LABEL: test_masked_8xfloat_to_4xfloat_perm_mask2:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm3
-; CHECK-NEXT:    vshufps {{.*#+}} xmm3 = xmm3[3,0],xmm0[0,0]
-; CHECK-NEXT:    vxorps %xmm4, %xmm4, %xmm4
-; CHECK-NEXT:    vcmpeqps %xmm4, %xmm2, %k1
-; CHECK-NEXT:    vshufps {{.*#+}} xmm1 {%k1} = xmm0[3,2],xmm3[0,2]
-; CHECK-NEXT:    vmovaps %xmm1, %xmm0
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm3 = <3,2,7,0,u,u,u,u>
+; CHECK-NEXT:    vpermps %ymm0, %ymm3, %ymm0
+; CHECK-NEXT:    vxorps %xmm3, %xmm3, %xmm3
+; CHECK-NEXT:    vcmpeqps %xmm3, %xmm2, %k1
+; CHECK-NEXT:    vblendmps %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x float> %vec, <8 x float> undef, <4 x i32> <i32 3, i32 2, i32 7, i32 0>
   %cmp = fcmp oeq <4 x float> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x float> %shuf, <4 x float> %vec2
   ret <4 x float> %res
 }
 
 define <4 x float> @test_masked_z_8xfloat_to_4xfloat_perm_mask2(<8 x float> %vec, <4 x float> %mask) {
 ; CHECK-LABEL: test_masked_z_8xfloat_to_4xfloat_perm_mask2:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[3,0],xmm0[0,0]
-; CHECK-NEXT:    vxorps %xmm3, %xmm3, %xmm3
-; CHECK-NEXT:    vcmpeqps %xmm3, %xmm1, %k1
-; CHECK-NEXT:    vshufps {{.*#+}} xmm0 {%k1} {z} = xmm0[3,2],xmm2[0,2]
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm2 = <3,2,7,0,u,u,u,u>
+; CHECK-NEXT:    vpermps %ymm0, %ymm2, %ymm0
+; CHECK-NEXT:    vxorps %xmm2, %xmm2, %xmm2
+; CHECK-NEXT:    vcmpeqps %xmm2, %xmm1, %k1
+; CHECK-NEXT:    vmovaps %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x float> %vec, <8 x float> undef, <4 x i32> <i32 3, i32 2, i32 7, i32 0>
   %cmp = fcmp oeq <4 x float> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x float> %shuf, <4 x float> zeroinitializer
   ret <4 x float> %res
 }
 define <4 x float> @test_8xfloat_to_4xfloat_perm_mask3(<8 x float> %vec) {
 ; CHECK-LABEL: test_8xfloat_to_4xfloat_perm_mask3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; CHECK-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,1,2]
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm1 = <3,3,5,2,u,u,u,u>
+; CHECK-NEXT:    vpermps %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %res = shufflevector <8 x float> %vec, <8 x float> undef, <4 x i32> <i32 3, i32 3, i32 5, i32 2>
   ret <4 x float> %res
 }
 define <4 x float> @test_masked_8xfloat_to_4xfloat_perm_mask3(<8 x float> %vec, <4 x float> %vec2, <4 x float> %mask) {
 ; CHECK-LABEL: test_masked_8xfloat_to_4xfloat_perm_mask3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm3
-; CHECK-NEXT:    vblendps {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm3 = <3,3,5,2,u,u,u,u>
+; CHECK-NEXT:    vpermps %ymm0, %ymm3, %ymm0
 ; CHECK-NEXT:    vxorps %xmm3, %xmm3, %xmm3
 ; CHECK-NEXT:    vcmpeqps %xmm3, %xmm2, %k1
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm1 {%k1} = xmm0[3,3,1,2]
-; CHECK-NEXT:    vmovaps %xmm1, %xmm0
+; CHECK-NEXT:    vblendmps %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x float> %vec, <8 x float> undef, <4 x i32> <i32 3, i32 3, i32 5, i32 2>
   %cmp = fcmp oeq <4 x float> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x float> %shuf, <4 x float> %vec2
   ret <4 x float> %res
 }
 
 define <4 x float> @test_masked_z_8xfloat_to_4xfloat_perm_mask3(<8 x float> %vec, <4 x float> %mask) {
 ; CHECK-LABEL: test_masked_z_8xfloat_to_4xfloat_perm_mask3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm2 = <3,3,5,2,u,u,u,u>
+; CHECK-NEXT:    vpermps %ymm0, %ymm2, %ymm0
 ; CHECK-NEXT:    vxorps %xmm2, %xmm2, %xmm2
 ; CHECK-NEXT:    vcmpeqps %xmm2, %xmm1, %k1
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 {%k1} {z} = xmm0[3,3,1,2]
+; CHECK-NEXT:    vmovaps %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <8 x float> %vec, <8 x float> undef, <4 x i32> <i32 3, i32 3, i32 5, i32 2>
   %cmp = fcmp oeq <4 x float> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x float> %shuf, <4 x float> zeroinitializer
   ret <4 x float> %res
 }
 define <4 x float> @test_8xfloat_to_4xfloat_perm_mem_mask0(<8 x float>* %vp) {
; CHECK-NEXT:    retq
   %cmp = fcmp oeq <4 x float> %mask, zeroinitializer
   %res = select <4 x i1> %cmp, <4 x float> %shuf, <4 x float> zeroinitializer
   ret <4 x float> %res
 }
 
 define <2 x double> @test_4xdouble_to_2xdouble_perm_mask0(<4 x double> %vec) {
 ; CHECK-LABEL: test_4xdouble_to_2xdouble_perm_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; CHECK-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; CHECK-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,0,2,3]
+; CHECK-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %res = shufflevector <4 x double> %vec, <4 x double> undef, <2 x i32> <i32 2, i32 0>
   ret <2 x double> %res
 }
 define <2 x double> @test_masked_4xdouble_to_2xdouble_perm_mask0(<4 x double> %vec, <2 x double> %vec2, <2 x double> %mask) {
 ; CHECK-LABEL: test_masked_4xdouble_to_2xdouble_perm_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm3
-; CHECK-NEXT:    vxorpd %xmm4, %xmm4, %xmm4
-; CHECK-NEXT:    vcmpeqpd %xmm4, %xmm2, %k1
-; CHECK-NEXT:    vunpcklpd {{.*#+}} xmm1 {%k1} = xmm3[0],xmm0[0]
-; CHECK-NEXT:    vmovapd %xmm1, %xmm0
+; CHECK-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,0,2,3]
+; CHECK-NEXT:    vxorpd %xmm3, %xmm3, %xmm3
+; CHECK-NEXT:    vcmpeqpd %xmm3, %xmm2, %k1
+; CHECK-NEXT:    vblendmpd %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <4 x double> %vec, <4 x double> undef, <2 x i32> <i32 2, i32 0>
   %cmp = fcmp oeq <2 x double> %mask, zeroinitializer
   %res = select <2 x i1> %cmp, <2 x double> %shuf, <2 x double> %vec2
   ret <2 x double> %res
 }
 
 define <2 x double> @test_masked_z_4xdouble_to_2xdouble_perm_mask0(<4 x double> %vec, <2 x double> %mask) {
 ; CHECK-LABEL: test_masked_z_4xdouble_to_2xdouble_perm_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vxorpd %xmm3, %xmm3, %xmm3
-; CHECK-NEXT:    vcmpeqpd %xmm3, %xmm1, %k1
-; CHECK-NEXT:    vunpcklpd {{.*#+}} xmm0 {%k1} {z} = xmm2[0],xmm0[0]
+; CHECK-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,0,2,3]
+; CHECK-NEXT:    vxorpd %xmm2, %xmm2, %xmm2
+; CHECK-NEXT:    vcmpeqpd %xmm2, %xmm1, %k1
+; CHECK-NEXT:    vmovapd %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <4 x double> %vec, <4 x double> undef, <2 x i32> <i32 2, i32 0>
   %cmp = fcmp oeq <2 x double> %mask, zeroinitializer
   %res = select <2 x i1> %cmp, <2 x double> %shuf, <2 x double> zeroinitializer
   ret <2 x double> %res
 }
 define <2 x double> @test_masked_4xdouble_to_2xdouble_perm_mask1(<4 x double> %vec, <2 x double> %vec2, <2 x double> %mask) {
 ; CHECK-LABEL: test_masked_4xdouble_to_2xdouble_perm_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm3
-; CHECK-NEXT:    vxorpd %xmm4, %xmm4, %xmm4
-; CHECK-NEXT:    vcmpeqpd %xmm4, %xmm2, %k1
-; CHECK-NEXT:    vunpckhpd {{.*#+}} xmm1 {%k1} = xmm0[1],xmm3[1]
-; CHECK-NEXT:    vmovapd %xmm1, %xmm0
+; CHECK-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[1,3,2,3]
+; CHECK-NEXT:    vxorpd %xmm3, %xmm3, %xmm3
+; CHECK-NEXT:    vcmpeqpd %xmm3, %xmm2, %k1
+; CHECK-NEXT:    vblendmpd %xmm0, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <4 x double> %vec, <4 x double> undef, <2 x i32> <i32 1, i32 3>
   %cmp = fcmp oeq <2 x double> %mask, zeroinitializer
   %res = select <2 x i1> %cmp, <2 x double> %shuf, <2 x double> %vec2
   ret <2 x double> %res
 }
 
 define <2 x double> @test_masked_z_4xdouble_to_2xdouble_perm_mask1(<4 x double> %vec, <2 x double> %mask) {
 ; CHECK-LABEL: test_masked_z_4xdouble_to_2xdouble_perm_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vxorpd %xmm3, %xmm3, %xmm3
-; CHECK-NEXT:    vcmpeqpd %xmm3, %xmm1, %k1
-; CHECK-NEXT:    vunpckhpd {{.*#+}} xmm0 {%k1} {z} = xmm0[1],xmm2[1]
+; CHECK-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[1,3,2,3]
+; CHECK-NEXT:    vxorpd %xmm2, %xmm2, %xmm2
+; CHECK-NEXT:    vcmpeqpd %xmm2, %xmm1, %k1
+; CHECK-NEXT:    vmovapd %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %shuf = shufflevector <4 x double> %vec, <4 x double> undef, <2 x i32> <i32 1, i32 3>
   %cmp = fcmp oeq <2 x double> %mask, zeroinitializer
   %res = select <2 x i1> %cmp, <2 x double> %shuf, <2 x double> zeroinitializer
   ret <2 x double> %res
 }
 define <2 x double> @test_4xdouble_to_2xdouble_perm_mem_mask0(<4 x double>* %vp) {