[X86] combineX86ShufflesRecursively(): call SimplifyMultipleUseDemandedVectorElts() on after finishing recursing

[lebedev.ri]

Summary

This was suggested in https://reviews.llvm.org/D108382#inline-1039018,
maybe hopefully this will help avoid regressions in that patch.

[RKSimon]

llvm/lib/Target/X86/X86ISelLowering.cpp

37912

Op might be a different width to the Root - see the "Widen any subvector shuffle inputs we've collected." code below.

[lebedev.ri]

Move the code after widening.
TBN, i don't believe this helps D108382.

[lebedev.ri]

Rebased ontop of D109074, NFC.

[lebedev.ri]

llvm/lib/Target/X86/X86ISelLowering.cpp

37912

I keep hitting the same pitfail.

[lebedev.ri]

Rebased, NFC.

[lebedev.ri]

ping

[RKSimon]

llvm/lib/Target/X86/X86ISelLowering.cpp

37912

We still need to do this before the widenSubVector() code - otherwise we'll never be able to simplify any input that doesn't match RootSizeInBits, which are likely to be the most interesting cases imo.

37964

This seems to be really bulky for what its actually doing. I don't think we need to create this shuffle mask for instance, we should be able to create a demanded elts mask directly and then trunc/scale it for the input's size.

I keep meaning to create a scaleDemandedMask() common helper method as we have several places that would use it (e.g. SelectionDAG.computeKnownBits bitcast handling and other parts of value tracking).

[lebedev.ri]

llvm/lib/Target/X86/X86ISelLowering.cpp

37964

That is what what i initially came up with, and it's *much* uglier than this code :)
I can do that again, but i'm not sure that will be be better.

[lebedev.ri]

Introduce ScaleDemandedEltsMask() and use it.

llvm/lib/Target/X86/X86ISelLowering.cpp

37912

I agree, but is this a correctness concern for this patch?

37964

Ok, how about this?

[RKSimon]

llvm/lib/Target/X86/X86ISelLowering.cpp

37599

We should assert that (NumElts % NumSrcElts) == 0 || (NumSrcElts % NumElts) == 0 - or return true/false on success/failure.

37912

what correctness?

[lebedev.ri]

llvm/lib/Target/X86/X86ISelLowering.cpp

37599

Oops, i meant to do that, but forgot to in the end.

37912

I mean, if we don't do this in this patch, will that lead to miscompiles, or simply to missed optimizations?

[lebedev.ri]

Add forgotten assert.

[RKSimon]

llvm/lib/Target/X86/X86ISelLowering.cpp

37595

This can probably move to the APIntOps helpers

37973

To move this before widening, we should just need to truncate OpDemandedElts based on its size vs RootSizeInBits - we should assert that no lost elts were demanded. Then we can scale it.

[lebedev.ri]

Try to move before widenSubVector() - now without miscompiles?

llvm/lib/Target/X86/X86ISelLowering.cpp

37595

Let's do that afterwards?

37973

Ok, i admit i've tried to avoid doing that because i don't quite understand all of the logic here.
Does this look right? It avoids the miscompiles that were visible in some previous attempt at least.

[lebedev.ri]

Did i get it right this time? :)

[RKSimon]

llvm/test/CodeGen/X86/insertelement-ones.ll

381

Any luck on improving this?

[lebedev.ri]

llvm/test/CodeGen/X86/insertelement-ones.ll

381

This one is obscure.
I believe the problem is X86ISelLowering.cpp's matchBinaryShuffle()'s ISD::OR lowering.

We have:

mask:  0 1 2 3 4 5 6 7 8 9 10 11 12 13 30 -2

matchBinaryShuffle()
EltSizeInBits: 8
V1:
t4: v16i8,ch = CopyFromReg t0, Register:v16i8 %1
  t3: v16i8 = Register %1
V2:
t74: v16i8 = X86ISD::VSHLDQ t51, TargetConstant:i8<14>
  t51: v16i8 = bitcast t50
    t50: v4i32 = scalar_to_vector Constant:i32<255>
      t49: i32 = Constant<255>
  t73: i8 = TargetConstant<14>

We can't say anything about t4, but i think it's obvious that t74 is actually
an all-zeros except the 14'th element, which is all-ones.
So we of course can lower that as an or blend, and we do not care what t4 is.
But the code fails to do that.

I think we'd basically have to do computeKnownBits() for each element of V1/V2 separately.

Should i keep looking?

[lebedev.ri]

llvm/test/CodeGen/X86/insertelement-ones.ll

381

Ok, got it: D109726

[lebedev.ri]

Rebased ontop main+D109726.
The noted regression is gone (but many more took it's place.)

[lebedev.ri]

Rebased, NFC.

[lebedev.ri]

llvm/test/CodeGen/X86/insertelement-ones.ll

311

Here we have:

Optimized legalized selection DAG: %bb.0 'insert_v16i8_x123456789ABCDEx:'
SelectionDAG has 20 nodes:
  t0: ch = EntryToken
          t2: v16i8,ch = CopyFromReg t0, Register:v16i8 %0
        t19: v16i8 = and t2, t36
        t20: v16i8 = X86ISD::ANDNP t36, t27
      t21: v16i8 = or t19, t20
      t33: v16i8 = X86ISD::VSHLDQ t27, TargetConstant:i8<15>
    t45: v16i8 = or t21, t33
  t12: ch,glue = CopyToReg t0, Register:v16i8 $xmm0, t45
    t26: v4i32 = scalar_to_vector Constant:i32<255>
  t27: v16i8 = bitcast t26
    t38: i64 = X86ISD::Wrapper TargetConstantPool:i64<<16 x i8> <i8 0, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>> 0
  t36: v16i8,ch = load<(load (s128) from constant-pool)> t0, t38, undef:i64
  t13: ch = X86ISD::RET_FLAG t12, TargetConstant:i32<0>, Register:v16i8 $xmm0, t12:1

... so matchBinaryShuffle() again fails to omit the masking,
even though it's obviously redundant here for the reasons seen in D109726.
I would suspect that is because around scalar_to_vector we operate on i32 elt type,
so we don't have all-ones elements until after bitcast.
Without changing computeKnownBits to operate on a specified element width,
i'm not sure it can help us further, and that does not sound like the right fix.

[RKSimon]

llvm/test/CodeGen/X86/insertelement-ones.ll

315

We're going to have to improve INSERT_VECTOR_ELT handling of 0/-1 elements - just AND/OR if we don't have a legal PINSRB instruction (pre-SSE41).

llvm/test/CodeGen/X86/oddshuffles.ll

2268

Looks like we're missing a fold to share scalar_to_vector(x) and scalar_to_vector(trunc(x)) (maybe worth supporting scalar_to_vector(ext(x)) as well)?

[RKSimon]

llvm/test/CodeGen/X86/insertelement-ones.ll

315

It looks like we might be able to do this more easily by extending lowerShuffleAsBitMask to handle the allones elements case as well as the zero elements case.

[lebedev.ri]

llvm/test/CodeGen/X86/insertelement-ones.ll

315

Note that X86TargetLowering::LowerINSERT_VECTOR_ELT isn't even called for this test,
since we expand, not legalize, in this case.
Marking it as legalize causes crashes "don't know how to legalize",
i guess it doesn't retry to legalize via the generic expansion.

[lebedev.ri]

Rebased ontop of D109989 - llvm/test/CodeGen/X86/insertelement-ones.ll is all good now.

[lebedev.ri]

llvm/test/CodeGen/X86/insertelement-ones.ll

315

Done, D109989.

[lebedev.ri]

llvm/test/CodeGen/X86/oddshuffles.ll

2268

This stuff is broken.
We've in AVX1-more, and only have broadcast-from-mem,
yet we've successfully obscured the load via the truncation.

I suppose we could look past ext/trunc of scalar_to_vector operand,
and change it to bitcast/ext of scalar_to_vector itself,
let me see.

Optimized legalized selection DAG: %bb.0 'splat_v3i32:'
SelectionDAG has 28 nodes:
  t0: ch = EntryToken
    t2: i64,ch = CopyFromReg t0, Register:i64 %0
  t27: i64,ch = load<(load (s64) from %ir.ptr, align 1)> t0, t2, undef:i64
      t24: v8i32 = BUILD_VECTOR Constant:i32<0>, undef:i32, Constant:i32<0>, Constant:i32<0>, Constant:i32<0>, Constant:i32<0>, Constant:i32<0>, Constant:i32<0>
          t30: v2i64 = scalar_to_vector t27
        t107: v4i64 = insert_subvector undef:v4i64, t30, Constant:i64<0>
      t108: v8i32 = bitcast t107
    t101: v8i32 = X86ISD::BLENDI t24, t108, TargetConstant:i8<2>
  t19: ch,glue = CopyToReg t0, Register:v8i32 $ymm0, t101
      t25: v8i32 = BUILD_VECTOR Constant:i32<0>, Constant:i32<0>, undef:i32, Constant:i32<0>, Constant:i32<0>, Constant:i32<0>, Constant:i32<0>, Constant:i32<0>
          t91: i32 = truncate t27
        t92: v4i32 = X86ISD::VBROADCAST t91
      t94: v8i32 = insert_subvector undef:v8i32, t92, Constant:i64<0>
    t97: v8i32 = X86ISD::BLENDI t25, t94, TargetConstant:i8<4>
  t21: ch,glue = CopyToReg t19, Register:v8i32 $ymm1, t97, t19:1
  t22: ch = X86ISD::RET_FLAG t21, TargetConstant:i32<0>, Register:v8i32 $ymm0, Register:v8i32 $ymm1, t21:1

[lebedev.ri]

llvm/test/CodeGen/X86/oddshuffles.ll

2268

While something like this should work:

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 5a49f33e46fe..4d7c2c2a8651 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -21824,6 +21824,13 @@ SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
     }
   }
 
+  // Fold SCALAR_TO_VECTOR(TRUNCATE(V)) to SCALAR_TO_VECTOR(V),
+  // by making trucation of the operand implicit.
+  if (InVal.getOpcode() == ISD::TRUNCATE && VT.isFixedLengthVector() &&
+      Level < AfterLegalizeDAG)
+    return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), VT,
+                       InVal->getOperand(0));
+
   return SDValue();
 }
 
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 09ba7af6e38a..695cc8303cc1 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -14076,10 +14076,12 @@ static SDValue lowerShuffleAsBroadcast(const SDLoc &DL, MVT VT, SDValue V1,
 
   // If this is a scalar, do the broadcast on this type and bitcast.
   if (!V.getValueType().isVector()) {
-    assert(V.getScalarValueSizeInBits() == NumEltBits &&
-           "Unexpected scalar size");
-    MVT BroadcastVT = MVT::getVectorVT(V.getSimpleValueType(),
-                                       VT.getVectorNumElements());
+    if(V.getValueType().isInteger() &&
+       V.getScalarValueSizeInBits() > NumEltBits)
+      V = DAG.getNode(ISD::TRUNCATE, DL, VT.getScalarType(), V);
+    assert(V.getScalarValueSizeInBits() == NumEltBits && "Unexpected scalar size");
+    MVT BroadcastVT =
+        MVT::getVectorVT(V.getSimpleValueType(), VT.getVectorNumElements());
     return DAG.getBitcast(VT, DAG.getNode(Opcode, DL, BroadcastVT, V));
   }

it doesn't catch anything with the cut-off,
and without it. it exposes numerous places that don't expect this truncation to be implicit.

[lebedev.ri]

llvm/test/CodeGen/X86/oddshuffles.ll

2268

I've looked again, and i'm not sure i have enough motivation to tackle all the fallout from the
scalar_to_vector(trunc(x)) --> scalar_to_vector(x) fold, unless i misunderstood the suggestion.

[lebedev.ri]

Rebased, NFC.

[RKSimon]

LGTM

llvm/lib/Target/X86/X86ISelLowering.cpp

37907

Lo + Hi

llvm/test/CodeGen/X86/oddshuffles.ll

2268

That's OK - I'll take a look when I get the chance.

[lebedev.ri]

In D109065#3008188, @RKSimon wrote:

LGTM

:/
Thank you for the review!

llvm/test/CodeGen/X86/oddshuffles.ll

2268

Sorry :/

llvm/lib/Target/X86/X86ISelLowering.cpp

namespace llvm {
   namespace X86 {
     enum {
       MaxShuffleCombineDepth = 8
     };
   }
 } // namespace llvm
 
 /// Fully generic combining of x86 shuffle instructions.
 ///
 /// This should be the last combine run over the x86 shuffle instructions. Once
 /// they have been fully optimized, this will recursively consider all chains
 /// of single-use shuffle instructions, build a generic model of the cumulative
 /// shuffle operation, and check for simpler instructions which implement this
 /// operation. We use this primarily for two purposes:
 ///
 /// 1) Collapse generic shuffles to specialized single instructions when
 ///    equivalent. In most cases, this is just an encoding size win, but
 ///    sometimes we will collapse multiple generic shuffles into a single
 ///    special-purpose shuffle.
 /// 2) Look for sequences of shuffle instructions with 3 or more total
 ///    instructions, and replace them with the slightly more expensive SSSE3
static SDValue combineX86ShufflesRecursively(
   }
 
   // Canonicalize the combined shuffle mask chain with horizontal ops.
   // NOTE: This will update the Ops and Mask.
   if (SDValue HOp = canonicalizeShuffleMaskWithHorizOp(
           Ops, Mask, RootSizeInBits, SDLoc(Root), DAG, Subtarget))
     return DAG.getBitcast(Root.getValueType(), HOp);
 
+  // Try to refine our inputs given our knowledge of target shuffle mask.
+  for (auto I : enumerate(Ops)) {
+    int OpIdx = I.index();
+    SDValue &Op = I.value();
+
+    // What range of shuffle mask element values results in picking from Op?
+    int lo = OpIdx * Mask.size();
+    int hi = lo + Mask.size();
+
+    // Which elements of Op do we demand, given the mask's granularity?
+    APInt OpDemandedElts(Mask.size(), 0);
+    for (int MaskElt : Mask) {
+      if (isInRange(MaskElt, lo, hi)) { // Picks from Op?
+        int OpEltIdx = MaskElt - lo;
+        OpDemandedElts.setBit(OpEltIdx);
+      }
+    }
+
+    // Is the shuffle result smaller than the root?
+    if (Op.getValueSizeInBits() < RootSizeInBits) {
+      // We padded the mask with undefs. But we now need to undo that.
+      unsigned NumExpectedVectorElts = Mask.size();
+      unsigned EltSizeInBits = RootSizeInBits / NumExpectedVectorElts;
+      unsigned NumOpVectorElts = Op.getValueSizeInBits() / EltSizeInBits;
+      assert(!OpDemandedElts.extractBits(
+                 NumExpectedVectorElts - NumOpVectorElts, NumOpVectorElts) &&
+             "Demanding the virtual undef widening padding?");
+      OpDemandedElts = OpDemandedElts.trunc(NumOpVectorElts); // NUW
+    }
+
+    // The Op itself may be of different VT, so we need to scale the mask.
+    unsigned NumOpElts = Op.getValueType().getVectorNumElements();
+    APInt OpScaledDemandedElts = APIntOps::ScaleBitMask(OpDemandedElts, NumOpElts);
+
+    // Can this operand be simplified any further, given it's demanded elements?
+    if (SDValue NewOp =
+            DAG.getTargetLoweringInfo().SimplifyMultipleUseDemandedVectorElts(
+                Op, OpScaledDemandedElts, DAG))
+      Op = NewOp;
+  }
+  // FIXME: should we rerun resolveTargetShuffleInputsAndMask() now?
+
   // Widen any subvector shuffle inputs we've collected.
   if (any_of(Ops, [RootSizeInBits](SDValue Op) {
         return Op.getValueSizeInBits() < RootSizeInBits;
       })) {
     for (SDValue &Op : Ops)
       if (Op.getValueSizeInBits() < RootSizeInBits)
         Op = widenSubVector(Op, false, Subtarget, DAG, SDLoc(Op),
                             RootSizeInBits);
     // Reresolve - we might have repeated subvector sources.
     resolveTargetShuffleInputsAndMask(Ops, Mask);
   }
 
   // We can only combine unary and binary shuffle mask cases.
   if (Ops.size() <= 2) {
     // Minor canonicalization of the accumulated shuffle mask to make it easier
     // to match below. All this does is detect masks with sequential pairs of
     // elements, and shrink them to the half-width mask. It does this in a loop
     // so it will reduce the size of the mask to the minimal width mask which
     // performs an equivalent shuffle.
     while (Mask.size() > 1) {
       SmallVector<int, 64> WidenedMask;
       if (!canWidenShuffleElements(Mask, WidenedMask))
         break;
       Mask = std::move(WidenedMask);
     }
 
     // Canonicalization of binary shuffle masks to improve pattern matching by
     // commuting the inputs.
     if (Ops.size() == 2 && canonicalizeShuffleMaskWithCommute(Mask)) {
       ShuffleVectorSDNode::commuteMask(Mask);
       std::swap(Ops[0], Ops[1]);
     }
 
     // Finally, try to combine into a single shuffle instruction.
     return combineX86ShuffleChain(Ops, Root, Mask, Depth, HasVariableMask,
                                   AllowVariableCrossLaneMask,
                                   AllowVariablePerLaneMask, DAG, Subtarget);
   }
 
   // If that failed and any input is extracted then try to combine as a

llvm/test/CodeGen/X86/insertelement-ones.ll

 ; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
 ; AVX512-NEXT:    retq
   %1 = insertelement <16 x i16> %a, i16 -1, i32 0
   %2 = insertelement <16 x i16> %1, i16 -1, i32 6
   %3 = insertelement <16 x i16> %2, i16 -1, i32 15
   ret <16 x i16> %3
 }
 
 define <16 x i8> @insert_v16i8_x123456789ABCDEx(<16 x i8> %a) {
 ; SSE2-LABEL: insert_v16i8_x123456789ABCDEx:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
+; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    movl $255, %eax
 ; SSE2-NEXT:    movd %eax, %xmm2
 ; SSE2-NEXT:    pandn %xmm2, %xmm1
 ; SSE2-NEXT:    pslldq {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm2[0]
 ; SSE2-NEXT:    por %xmm1, %xmm2
 ; SSE2-NEXT:    por %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: insert_v16i8_x123456789ABCDEx:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
+; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    movl $255, %eax
 ; SSE3-NEXT:    movd %eax, %xmm2
 ; SSE3-NEXT:    pandn %xmm2, %xmm1
 ; SSE3-NEXT:    pslldq {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm2[0]
 ; SSE3-NEXT:    por %xmm1, %xmm2
 ; SSE3-NEXT:    por %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
; AVX-NEXT:    retq
   %2 = insertelement <16 x i8> %1, i8 -1, i32 15
   ret <16 x i8> %2
 }
 
 define <32 x i8> @insert_v32i8_x123456789ABCDEzGHIJKLMNOPQRSTxx(<32 x i8> %a) {
 ; SSE2-LABEL: insert_v32i8_x123456789ABCDEzGHIJKLMNOPQRSTxx:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
+; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    movl $255, %eax
 ; SSE2-NEXT:    movd %eax, %xmm3
 ; SSE2-NEXT:    pandn %xmm3, %xmm2
 ; SSE2-NEXT:    movdqa %xmm3, %xmm4
 ; SSE2-NEXT:    pslldq {{.*#+}} xmm4 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm4[0]
 ; SSE2-NEXT:    por %xmm4, %xmm2
 ; SSE2-NEXT:    por %xmm2, %xmm0
 ; SSE2-NEXT:    pslldq {{.*#+}} xmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1]
 ; SSE2-NEXT:    por %xmm4, %xmm3
 ; SSE2-NEXT:    por %xmm3, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: insert_v32i8_x123456789ABCDEzGHIJKLMNOPQRSTxx:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm2 = [0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
+; SSE3-NEXT:    pand %xmm2, %xmm0
 ; SSE3-NEXT:    movl $255, %eax
 ; SSE3-NEXT:    movd %eax, %xmm3
 ; SSE3-NEXT:    pandn %xmm3, %xmm2
 ; SSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSE3-NEXT:    pslldq {{.*#+}} xmm4 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm4[0]
 ; SSE3-NEXT:    por %xmm4, %xmm2
 ; SSE3-NEXT:    por %xmm2, %xmm0
 ; SSE3-NEXT:    pslldq {{.*#+}} xmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1]

llvm/test/CodeGen/X86/oddshuffles.ll

 ; SSE42-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
 ; SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,0,1]
 ; SSE42-NEXT:    pxor %xmm1, %xmm1
 ; SSE42-NEXT:    xorps %xmm3, %xmm3
 ; SSE42-NEXT:    retq
 ;
 ; AVX1-LABEL: splat_v3i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX1-NEXT:    vpinsrd $2, 8(%rdi), %xmm0, %xmm1
-; AVX1-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3,4,5,6,7]
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]
-; AVX1-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4,5,6,7]
+; AVX1-NEXT:    movq (%rdi), %rax
+; AVX1-NEXT:    vmovq %rax, %xmm0
+; AVX1-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4,5,6,7]
+; AVX1-NEXT:    vmovd %eax, %xmm2
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,0,0,0]
+; AVX1-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6,7]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-SLOW-LABEL: splat_v3i32:
 ; AVX2-SLOW:       # %bb.0:
 ; AVX2-SLOW-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
 ; AVX2-SLOW-NEXT:    vxorps %xmm2, %xmm2, %xmm2
 ; AVX2-SLOW-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0],ymm1[1],ymm2[2,3,4,5,6,7]
 ; AVX2-SLOW-NEXT:    vbroadcastss %xmm1, %xmm1
 ; AVX2-SLOW-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4,5,6,7]
 ; AVX2-SLOW-NEXT:    retq
 ;
 ; AVX2-FAST-LABEL: splat_v3i32:
 ; AVX2-FAST:       # %bb.0:
 ; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
 ; AVX2-FAST-NEXT:    vpxor %xmm0, %xmm0, %xmm0
 ; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7]
 ; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,zero,zero,ymm1[0,1,2,3],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX2-FAST-NEXT:    retq
 ;
 ; XOP-LABEL: splat_v3i32:
 ; XOP:       # %bb.0:
-; XOP-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; XOP-NEXT:    vpinsrd $2, 8(%rdi), %xmm0, %xmm1
-; XOP-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; XOP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3,4,5,6,7]
-; XOP-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]
-; XOP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4,5,6,7]
+; XOP-NEXT:    movq (%rdi), %rax
+; XOP-NEXT:    vmovq %rax, %xmm0
+; XOP-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; XOP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4,5,6,7]
+; XOP-NEXT:    vmovd %eax, %xmm2
+; XOP-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,0,0,0]
+; XOP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6,7]
 ; XOP-NEXT:    retq
   %1 = load <3 x i32>, <3 x i32>* %ptr, align 1
   %2 = shufflevector <3 x i32> %1, <3 x i32> undef, <16 x i32> <i32 0, i32 1, 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>
   %3 = shufflevector <16 x i32> <i32 0, i32 undef, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 undef, i32 0, i32 0, i32 0, i32 0, i32 0>, <16 x i32> %2, <16 x i32> <i32 0, i32 17, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 16, i32 11, i32 12, i32 13, i32 14, i32 15>
   ret <16 x i32 > %3
 }
 
 define <2 x double> @wrongorder(<4 x double> %A, <8 x double>* %P) #0 {

llvm/test/CodeGen/X86/vselect.ll

 define <2 x i32> @simplify_select(i32 %x, <2 x i1> %z) {
 ; SSE2-LABEL: simplify_select:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SSE2-NEXT:    pslld $31, %xmm0
 ; SSE2-NEXT:    psrad $31, %xmm0
 ; SSE2-NEXT:    movd %edi, %xmm1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,0,1,1]
-; SSE2-NEXT:    por %xmm1, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm3
+; SSE2-NEXT:    por %xmm1, %xmm3
 ; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm2[1,3]
 ; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0],xmm2[1,1]
-; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm3
 ; SSE2-NEXT:    pandn %xmm1, %xmm0
-; SSE2-NEXT:    por %xmm2, %xmm0
+; SSE2-NEXT:    por %xmm3, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: simplify_select:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SSE41-NEXT:    pslld $31, %xmm0
 ; SSE41-NEXT:    movd %edi, %xmm1
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,0,1,1]