[DAGCombiner] eliminate shuffle of insert element

[spatel]

Summary

I noticed this pattern in D38316 / D38388. We failed to combine a shuffle that is either repeating a scalar insertion at the same position in a vector or translated to a different element index.

Like the earlier patch, this could be an instcombine too, but since we opted to make this a DAG transform earlier, I've made this one a DAG patch too.

I don't think we need any legality checking because the new insert is identical to the existing insert except that it may have a different constant insertion operand.

The constant insertion test in test/CodeGen/X86/vector-shuffle-combining.ll was the motivation for D38756.

[spatel]

Ping.

[efriedma]

I don't think we need any legality checking because the new insert is identical to the existing insert except that it may have a different constant insertion operand.

We should probably document this rule somewhere... I mean, it makes sense, but it's not obvious given that we allow the legality to vary based on whether the index is constant, and some targets have operations like movss which could be modeled as an insert_vector_elt.

[spatel]

Patch updated:
Added a code comment to explain why we shouldn't need a legality check or hook to perform this fold. Not sure if that's sufficient. Let me know if there's another place or way to document the logic. Thanks!

[efriedma]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

15737 ↗

(On Diff #124484)

Do we need to handle the case where Mask[i] < 0?

15778 ↗

(On Diff #124484)

I guess there's sort of a family of related transforms here... does it make sense to optimize other cases where we know the value? A SCALAR_TO_VECTOR? A constant? A BUILD_VECTOR? Another shuffle we could examine recursively? (Not that you need to implement any of that right now, but it would make sense to note the direction you expect for future generalizations.)

[spatel]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

15737 ↗

(On Diff #124484)

I was going to postpone that to a follow-up because I know we can get into some gray areas with undefs in shuffle masks. Ie, is it better to preserve the undef-ness of some vector lanes or eliminate an op but then have all lanes defined?

Ok, if I add a TODO in this patch?

15778 ↗

(On Diff #124484)

Good point - I haven't seen the other patterns in real code, but it's worth noting here.

[spatel]

Patch updated:

1. Added TODO comment to check for undef shuffle mask elements.
2. Added TODO comment to extend the search for a scalar element.

[spatel]

Ping.

[spatel]

Patch updated:
Offline, @RKSimon noted that Mask could point to CommutedMask after it has gone out-of-scope, so I moved that out of the 'if' clause.

[efriedma]

LGTM

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

static SDValue combineShuffleOfSplat(ArrayRef<int> UserMask,
   for (int Idx : UserMask)
     NewMask.push_back(Idx == -1 ? -1 : SplatMask[Idx]);
 
   return DAG.getVectorShuffle(Splat->getValueType(0), SDLoc(Splat),
                               Splat->getOperand(0), Splat->getOperand(1),
                               NewMask);
 }
 
+/// If the shuffle mask is taking exactly one element from the first vector
+/// operand and passing through all other elements from the second vector
+/// operand, return the index of the mask element that is choosing an element
+/// from the first operand. Otherwise, return -1.
+static int getShuffleMaskIndexOfOneElementFromOp0IntoOp1(ArrayRef<int> Mask) {
+  int MaskSize = Mask.size();
+  int EltFromOp0 = -1;
+  // TODO: This does not match if there are undef elements in the shuffle mask.
+  // Should we ignore undefs in the shuffle mask instead? The trade-off is
+  // removing an instruction (a shuffle), but losing the knowledge that some
+  // vector lanes are not needed.
+  for (int i = 0; i != MaskSize; ++i) {
+    if (Mask[i] >= 0 && Mask[i] < MaskSize) {
+      // We're looking for a shuffle of exactly one element from operand 0.
+      if (EltFromOp0 != -1)
+        return -1;
+      EltFromOp0 = i;
+    } else if (Mask[i] != i + MaskSize) {
+      // Nothing from operand 1 can change lanes.
+      return -1;
+    }
+  }
+  return EltFromOp0;
+}
+
+/// If a shuffle inserts exactly one element from a source vector operand into
+/// another vector operand and we can access the specified element as a scalar,
+/// then we can eliminate the shuffle.
+static SDValue replaceShuffleOfInsert(ShuffleVectorSDNode *Shuf,
+                                      SelectionDAG &DAG) {
+  // First, check if we are taking one element of a vector and shuffling that
+  // element into another vector.
+  ArrayRef<int> Mask = Shuf->getMask();
+  SmallVector<int, 16> CommutedMask(Mask.begin(), Mask.end());
+  SDValue Op0 = Shuf->getOperand(0);
+  SDValue Op1 = Shuf->getOperand(1);
+  int ShufOp0Index = getShuffleMaskIndexOfOneElementFromOp0IntoOp1(Mask);
+  if (ShufOp0Index == -1) {
+    // Commute mask and check again.
+    ShuffleVectorSDNode::commuteMask(CommutedMask);
+    ShufOp0Index = getShuffleMaskIndexOfOneElementFromOp0IntoOp1(CommutedMask);
+    if (ShufOp0Index == -1)
+      return SDValue();
+    // Commute operands to match the commuted shuffle mask.
+    std::swap(Op0, Op1);
+    Mask = CommutedMask;
+  }
+
+  // The shuffle inserts exactly one element from operand 0 into operand 1.
+  // Now see if we can access that element as a scalar via a real insert element
+  // instruction.
+  // TODO: We can try harder to locate the element as a scalar. Examples: it
+  // could be an operand of SCALAR_TO_VECTOR, BUILD_VECTOR, or a constant.
+  assert(Mask[ShufOp0Index] >= 0 && Mask[ShufOp0Index] < (int)Mask.size() &&
+         "Shuffle mask value must be from operand 0");
+  if (Op0.getOpcode() != ISD::INSERT_VECTOR_ELT)
+    return SDValue();
+
+  auto *InsIndexC = dyn_cast<ConstantSDNode>(Op0.getOperand(2));
+  if (!InsIndexC || InsIndexC->getSExtValue() != Mask[ShufOp0Index])
+    return SDValue();
+
+  // There's an existing insertelement with constant insertion index, so we
+  // don't need to check the legality/profitability of a replacement operation
+  // that differs at most in the constant value. The target should be able to
+  // lower any of those in a similar way. If not, legalization will expand this
+  // to a scalar-to-vector plus shuffle.
+  //
+  // Note that the shuffle may move the scalar from the position that the insert
+  // element used. Therefore, our new insert element occurs at the shuffle's
+  // mask index value, not the insert's index value.
+  // shuffle (insertelt v1, x, C), v2, mask --> insertelt v2, x, C'
+  SDValue NewInsIndex = DAG.getConstant(ShufOp0Index, SDLoc(Shuf),
+                                        Op0.getOperand(2).getValueType());
+  return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(Shuf), Op0.getValueType(),
+                     Op1, Op0.getOperand(1), NewInsIndex);
+}
+
 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
   EVT VT = N->getValueType(0);
   unsigned NumElts = VT.getVectorNumElements();
 
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
 
   assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
if (N1.isUndef()) {
     if (Changed)
       return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, NewMask);
   }
 
   // Simplify shuffle mask if a referenced element is UNDEF.
   if (SDValue V = simplifyShuffleMask(SVN, N0, N1, DAG))
     return V;
 
+  if (SDValue InsElt = replaceShuffleOfInsert(SVN, DAG))
+    return InsElt;
+
   // A shuffle of a single vector that is a splat can always be folded.
   if (auto *N0Shuf = dyn_cast<ShuffleVectorSDNode>(N0))
     if (N1->isUndef() && N0Shuf->isSplat())
       return combineShuffleOfSplat(SVN->getMask(), N0Shuf, DAG);
 
   // If it is a splat, check if the argument vector is another splat or a
   // build_vector.
   if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {

llvm/trunk/test/CodeGen/X86/shuffle-of-insert.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=sse2   | FileCheck %s --check-prefix=ALL --check-prefix=SSEANY --check-prefix=SSE2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=sse4.1 | FileCheck %s --check-prefix=ALL --check-prefix=SSEANY --check-prefix=SSE4
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx    | FileCheck %s --check-prefix=ALL --check-prefix=AVX
 
 define <4 x i32> @ins_elt_0(i32 %x, <4 x i32> %v1, <4 x i32> %v2) {
 ; SSE2-LABEL: ins_elt_0:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movd %edi, %xmm0
 ; SSE2-NEXT:    movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
 ; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: ins_elt_0:
 ; SSE4:       # %bb.0:
-; SSE4-NEXT:    pinsrd $0, %edi, %xmm0
-; SSE4-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
+; SSE4-NEXT:    pinsrd $0, %edi, %xmm1
+; SSE4-NEXT:    movdqa %xmm1, %xmm0
 ; SSE4-NEXT:    retq
 ;
 ; AVX-LABEL: ins_elt_0:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrd $0, %edi, %xmm0, %xmm0
-; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
+; AVX-NEXT:    vpinsrd $0, %edi, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %ins = insertelement <4 x i32> %v1, i32 %x, i32 0
   %shuf = shufflevector <4 x i32> %ins, <4 x i32> %v2, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
   ret <4 x i32> %shuf
 }
 
 define <4 x i32> @ins_elt_1(i32 %x, <4 x i32> %v1, <4 x i32> %v2) {
 ; SSE2-LABEL: ins_elt_1:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movd %edi, %xmm2
-; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm0[0,0]
-; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,2],xmm1[0,0]
-; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm1[2,3]
-; SSE2-NEXT:    movaps %xmm2, %xmm0
+; SSE2-NEXT:    movd %edi, %xmm0
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0]
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: ins_elt_1:
 ; SSE4:       # %bb.0:
-; SSE4-NEXT:    pinsrd $1, %edi, %xmm0
-; SSE4-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
+; SSE4-NEXT:    pinsrd $1, %edi, %xmm1
+; SSE4-NEXT:    movdqa %xmm1, %xmm0
 ; SSE4-NEXT:    retq
 ;
 ; AVX-LABEL: ins_elt_1:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0
-; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
+; AVX-NEXT:    vpinsrd $1, %edi, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %ins = insertelement <4 x i32> %v1, i32 %x, i32 1
   %shuf = shufflevector <4 x i32> %ins, <4 x i32> %v2, <4 x i32> <i32 4, i32 1, i32 6, i32 7>
   ret <4 x i32> %shuf
 }
 
 ; Verify that the transform still works when the insert element is the 2nd operand to the shuffle.
 
 define <4 x i32> @ins_elt_2_commute(i32 %x, <4 x i32> %v1, <4 x i32> %v2) {
 ; SSE2-LABEL: ins_elt_2_commute:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movd %edi, %xmm2
-; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm0[3,0]
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[0,2]
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[3,0]
+; SSE2-NEXT:    movd %edi, %xmm0
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[3,0]
 ; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]
 ; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: ins_elt_2_commute:
 ; SSE4:       # %bb.0:
-; SSE4-NEXT:    pinsrd $2, %edi, %xmm0
-; SSE4-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5],xmm1[6,7]
+; SSE4-NEXT:    pinsrd $2, %edi, %xmm1
+; SSE4-NEXT:    movdqa %xmm1, %xmm0
 ; SSE4-NEXT:    retq
 ;
 ; AVX-LABEL: ins_elt_2_commute:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrd $2, %edi, %xmm0, %xmm0
-; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5],xmm1[6,7]
+; AVX-NEXT:    vpinsrd $2, %edi, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %ins = insertelement <4 x i32> %v1, i32 %x, i32 2
   %shuf = shufflevector <4 x i32> %v2, <4 x i32> %ins, <4 x i32> <i32 0, i32 1, i32 6, i32 3>
   ret <4 x i32> %shuf
 }
 
 define <4 x i32> @ins_elt_3_commute(i32 %x, <4 x i32> %v1, <4 x i32> %v2) {
 ; SSE2-LABEL: ins_elt_3_commute:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movd %edi, %xmm2
-; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm0[2,0]
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,0]
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,0],xmm1[2,0]
+; SSE2-NEXT:    movd %edi, %xmm0
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[2,0]
 ; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0]
 ; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: ins_elt_3_commute:
 ; SSE4:       # %bb.0:
-; SSE4-NEXT:    pinsrd $3, %edi, %xmm0
-; SSE4-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
+; SSE4-NEXT:    pinsrd $3, %edi, %xmm1
+; SSE4-NEXT:    movdqa %xmm1, %xmm0
 ; SSE4-NEXT:    retq
 ;
 ; AVX-LABEL: ins_elt_3_commute:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrd $3, %edi, %xmm0, %xmm0
-; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
+; AVX-NEXT:    vpinsrd $3, %edi, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %ins = insertelement <4 x i32> %v1, i32 %x, i32 3
   %shuf = shufflevector <4 x i32> %v2, <4 x i32> %ins, <4 x i32> <i32 0, i32 1, i32 2, i32 7>
   ret <4 x i32> %shuf
 }
 
 ; In the next 4 tests, the shuffle moves the inserted scalar to a different position in the output vector.
 
 define <4 x i32> @ins_elt_0_to_2(i32 %x, <4 x i32> %v1, <4 x i32> %v2) {
 ; SSE2-LABEL: ins_elt_0_to_2:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movd %edi, %xmm0
 ; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[3,0]
 ; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]
 ; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: ins_elt_0_to_2:
 ; SSE4:       # %bb.0:
-; SSE4-NEXT:    pinsrd $0, %edi, %xmm0
-; SSE4-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
-; SSE4-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5],xmm1[6,7]
+; SSE4-NEXT:    pinsrd $2, %edi, %xmm1
+; SSE4-NEXT:    movdqa %xmm1, %xmm0
 ; SSE4-NEXT:    retq
 ;
 ; AVX-LABEL: ins_elt_0_to_2:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrd $0, %edi, %xmm0, %xmm0
-; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
-; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5],xmm1[6,7]
+; AVX-NEXT:    vpinsrd $2, %edi, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %ins = insertelement <4 x i32> %v1, i32 %x, i32 0
   %shuf = shufflevector <4 x i32> %ins, <4 x i32> %v2, <4 x i32> <i32 4, i32 5, i32 0, i32 7>
   ret <4 x i32> %shuf
 }
 
 define <4 x i32> @ins_elt_1_to_0(i32 %x, <4 x i32> %v1, <4 x i32> %v2) {
 ; SSE2-LABEL: ins_elt_1_to_0:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movd %edi, %xmm0
 ; SSE2-NEXT:    movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
 ; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: ins_elt_1_to_0:
 ; SSE4:       # %bb.0:
-; SSE4-NEXT:    pinsrd $1, %edi, %xmm0
-; SSE4-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; SSE4-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
+; SSE4-NEXT:    pinsrd $0, %edi, %xmm1
+; SSE4-NEXT:    movdqa %xmm1, %xmm0
 ; SSE4-NEXT:    retq
 ;
 ; AVX-LABEL: ins_elt_1_to_0:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0
-; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
+; AVX-NEXT:    vpinsrd $0, %edi, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %ins = insertelement <4 x i32> %v1, i32 %x, i32 1
   %shuf = shufflevector <4 x i32> %ins, <4 x i32> %v2, <4 x i32> <i32 1, i32 5, i32 6, i32 7>
   ret <4 x i32> %shuf
 }
 
 define <4 x i32> @ins_elt_2_to_3(i32 %x, <4 x i32> %v1, <4 x i32> %v2) {
 ; SSE2-LABEL: ins_elt_2_to_3:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movd %edi, %xmm2
-; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm0[3,0]
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[0,2]
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,0]
+; SSE2-NEXT:    movd %edi, %xmm0
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[2,0]
 ; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0]
 ; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: ins_elt_2_to_3:
 ; SSE4:       # %bb.0:
-; SSE4-NEXT:    pinsrd $2, %edi, %xmm0
-; SSE4-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,2,2]
-; SSE4-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
+; SSE4-NEXT:    pinsrd $3, %edi, %xmm1
+; SSE4-NEXT:    movdqa %xmm1, %xmm0
 ; SSE4-NEXT:    retq
 ;
 ; AVX-LABEL: ins_elt_2_to_3:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrd $2, %edi, %xmm0, %xmm0
-; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,2]
-; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
+; AVX-NEXT:    vpinsrd $3, %edi, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %ins = insertelement <4 x i32> %v1, i32 %x, i32 2
   %shuf = shufflevector <4 x i32> %v2, <4 x i32> %ins, <4 x i32> <i32 0, i32 1, i32 2, i32 6>
   ret <4 x i32> %shuf
 }
 
 define <4 x i32> @ins_elt_3_to_1(i32 %x, <4 x i32> %v1, <4 x i32> %v2) {
 ; SSE2-LABEL: ins_elt_3_to_1:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movd %edi, %xmm2
-; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm0[2,0]
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,0]
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,0],xmm1[0,0]
+; SSE2-NEXT:    movd %edi, %xmm0
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0]
 ; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: ins_elt_3_to_1:
 ; SSE4:       # %bb.0:
-; SSE4-NEXT:    pinsrd $3, %edi, %xmm0
-; SSE4-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
-; SSE4-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
+; SSE4-NEXT:    pinsrd $1, %edi, %xmm1
+; SSE4-NEXT:    movdqa %xmm1, %xmm0
 ; SSE4-NEXT:    retq
 ;
 ; AVX-LABEL: ins_elt_3_to_1:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrd $3, %edi, %xmm0, %xmm0
-; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
-; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
+; AVX-NEXT:    vpinsrd $1, %edi, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %ins = insertelement <4 x i32> %v1, i32 %x, i32 3
   %shuf = shufflevector <4 x i32> %v2, <4 x i32> %ins, <4 x i32> <i32 0, i32 7, i32 2, i32 3>
   ret <4 x i32> %shuf
 }
 

llvm/trunk/test/CodeGen/X86/sse41.ll

; X64-NEXT:    retq
   ret <4 x float> %result
 }
 
 ; TODO: Like on pinsrd_from_shufflevector_i32, remove this mov instr
 define <4 x i32> @insertps_from_load_ins_elt_undef_i32(<4 x i32> %a, i32* %b) {
 ; X32-LABEL: insertps_from_load_ins_elt_undef_i32:
 ; X32:       ## %bb.0:
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; X32-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]
-; X32-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5],xmm0[6,7]
+; X32-NEXT:    pinsrd $2, (%eax), %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: insertps_from_load_ins_elt_undef_i32:
 ; X64:       ## %bb.0:
-; X64-NEXT:    movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; X64-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]
-; X64-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5],xmm0[6,7]
+; X64-NEXT:    pinsrd $2, (%rdi), %xmm0
 ; X64-NEXT:    retq
   %1 = load i32, i32* %b, align 4
   %2 = insertelement <4 x i32> undef, i32 %1, i32 0
   %result = shufflevector <4 x i32> %a, <4 x i32> %2, <4 x i32> <i32 0, i32 1, i32 4, i32 3>
   ret <4 x i32> %result
 }
 
 ;;;;;; Shuffles optimizable with a single insertps or blend instruction

llvm/trunk/test/CodeGen/X86/vector-shuffle-128-v2.ll

 ; SSSE3-LABEL: insert_reg_lo_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movq %rdi, %xmm1
 ; SSSE3-NEXT:    movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: insert_reg_lo_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movq %rdi, %xmm1
-; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
+; SSE41-NEXT:    pinsrq $0, %rdi, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: insert_reg_lo_v2i64:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovq %rdi, %xmm1
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: insert_reg_lo_v2i64:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovq %rdi, %xmm1
-; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
-; AVX2-NEXT:    retq
-;
-; AVX512VL-LABEL: insert_reg_lo_v2i64:
-; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    vmovq %rdi, %xmm1
-; AVX512VL-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
-; AVX512VL-NEXT:    retq
+; AVX-LABEL: insert_reg_lo_v2i64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpinsrq $0, %rdi, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %v = insertelement <2 x i64> undef, i64 %a, i32 0
   %shuffle = shufflevector <2 x i64> %v, <2 x i64> %b, <2 x i32> <i32 0, i32 3>
   ret <2 x i64> %shuffle
 }
 
 define <2 x i64> @insert_mem_lo_v2i64(i64* %ptr, <2 x i64> %b) {
 ; SSE2-LABEL: insert_mem_lo_v2i64:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movlpd {{.*#+}} xmm0 = mem[0],xmm0[1]
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: insert_mem_lo_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movlpd {{.*#+}} xmm0 = mem[0],xmm0[1]
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: insert_mem_lo_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movlpd {{.*#+}} xmm0 = mem[0],xmm0[1]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: insert_mem_lo_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
-; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
+; SSE41-NEXT:    pinsrq $0, (%rdi), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: insert_mem_lo_v2i64:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: insert_mem_lo_v2i64:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
-; AVX2-NEXT:    retq
-;
-; AVX512VL-LABEL: insert_mem_lo_v2i64:
-; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX512VL-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
-; AVX512VL-NEXT:    retq
+; AVX-LABEL: insert_mem_lo_v2i64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpinsrq $0, (%rdi), %xmm0, %xmm0
+; AVX-NEXT:    retq
   %a = load i64, i64* %ptr
   %v = insertelement <2 x i64> undef, i64 %a, i32 0
   %shuffle = shufflevector <2 x i64> %v, <2 x i64> %b, <2 x i32> <i32 0, i32 3>
   ret <2 x i64> %shuffle
 }
 
 define <2 x i64> @insert_reg_hi_v2i64(i64 %a, <2 x i64> %b) {
-; SSE-LABEL: insert_reg_hi_v2i64:
-; SSE:       # %bb.0:
-; SSE-NEXT:    movq %rdi, %xmm1
-; SSE-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE-NEXT:    retq
+; SSE2-LABEL: insert_reg_hi_v2i64:
+; SSE2:       # %bb.0:
+; SSE2-NEXT:    movq %rdi, %xmm1
+; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT:    retq
+;
+; SSE3-LABEL: insert_reg_hi_v2i64:
+; SSE3:       # %bb.0:
+; SSE3-NEXT:    movq %rdi, %xmm1
+; SSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE3-NEXT:    retq
+;
+; SSSE3-LABEL: insert_reg_hi_v2i64:
+; SSSE3:       # %bb.0:
+; SSSE3-NEXT:    movq %rdi, %xmm1
+; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSSE3-NEXT:    retq
+;
+; SSE41-LABEL: insert_reg_hi_v2i64:
+; SSE41:       # %bb.0:
+; SSE41-NEXT:    pinsrq $1, %rdi, %xmm0
+; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: insert_reg_hi_v2i64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vmovq %rdi, %xmm1
-; AVX-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX-NEXT:    vpinsrq $1, %rdi, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %v = insertelement <2 x i64> undef, i64 %a, i32 0
   %shuffle = shufflevector <2 x i64> %v, <2 x i64> %b, <2 x i32> <i32 2, i32 0>
   ret <2 x i64> %shuffle
 }
 
 define <2 x i64> @insert_mem_hi_v2i64(i64* %ptr, <2 x i64> %b) {
-; SSE-LABEL: insert_mem_hi_v2i64:
-; SSE:       # %bb.0:
-; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
-; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE-NEXT:    retq
+; SSE2-LABEL: insert_mem_hi_v2i64:
+; SSE2:       # %bb.0:
+; SSE2-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
+; SSE2-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT:    retq
+;
+; SSE3-LABEL: insert_mem_hi_v2i64:
+; SSE3:       # %bb.0:
+; SSE3-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
+; SSE3-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE3-NEXT:    retq
+;
+; SSSE3-LABEL: insert_mem_hi_v2i64:
+; SSSE3:       # %bb.0:
+; SSSE3-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
+; SSSE3-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSSE3-NEXT:    retq
+;
+; SSE41-LABEL: insert_mem_hi_v2i64:
+; SSE41:       # %bb.0:
+; SSE41-NEXT:    pinsrq $1, (%rdi), %xmm0
+; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: insert_mem_hi_v2i64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX-NEXT:    vpinsrq $1, (%rdi), %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %a = load i64, i64* %ptr
   %v = insertelement <2 x i64> undef, i64 %a, i32 0
   %shuffle = shufflevector <2 x i64> %v, <2 x i64> %b, <2 x i32> <i32 2, i32 0>
   ret <2 x i64> %shuffle
 }
 
 define <2 x double> @insert_reg_lo_v2f64(double %a, <2 x double> %b) {

llvm/trunk/test/CodeGen/X86/vector-shuffle-combining.ll

 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movd %edi, %xmm1
 ; SSSE3-NEXT:    movaps {{.*#+}} xmm0 = <u,4,5,30>
 ; SSSE3-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: combine_constant_insertion_v4i32:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movd %edi, %xmm0
-; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1],mem[2,3,4,5,6,7]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = <u,4,5,30>
+; SSE41-NEXT:    pinsrd $0, %edi, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: combine_constant_insertion_v4i32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovd %edi, %xmm0
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],mem[2,3,4,5,6,7]
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: combine_constant_insertion_v4i32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovd %edi, %xmm0
-; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],mem[1,2,3]
-; AVX2-NEXT:    retq
+; AVX-LABEL: combine_constant_insertion_v4i32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vmovdqa {{.*#+}} xmm0 = <u,4,5,30>
+; AVX-NEXT:    vpinsrd $0, %edi, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %a0 = insertelement <4 x i32> undef, i32 %f, i32 0
   %ret = shufflevector <4 x i32> %a0, <4 x i32> <i32 undef, i32 4, i32 5, i32 30>, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
   ret <4 x i32> %ret
 }
 
 define <4 x float> @PR22377(<4 x float> %a, <4 x float> %b) {
 ; SSE-LABEL: PR22377:
 ; SSE:       # %bb.0: # %entry