static SDValue lowerVectorShuffleAsInsertPS(const SDLoc &DL, SDValue V1,
SDValue V2, ArrayRef<int> Mask,
+ const SmallBitVector &Zeroable,
SelectionDAG &DAG) {
assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
- SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
// Attempt to match the insertps pattern.
unsigned InsertPSMask;
/// domain crossing penalties, as these are sufficient to implement all v4f32
/// shuffles.
static SDValue lowerV4F32VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask,
+ const SmallBitVector &Zeroable,
SDValue V1, SDValue V2,
const X86Subtarget &Subtarget,
SelectionDAG &DAG) {
return Blend;
// Use INSERTPS if we can complete the shuffle efficiently.
- if (SDValue V = lowerVectorShuffleAsInsertPS(DL, V1, V2, Mask, DAG))
+ if (SDValue V =
+ lowerVectorShuffleAsInsertPS(DL, V1, V2, Mask, Zeroable, DAG))
return V;
if (!isSingleSHUFPSMask(Mask))
case MVT::v4i32:
return lowerV4I32VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
case MVT::v4f32:
- return lowerV4F32VectorShuffle(DL, Mask, V1, V2, Subtarget, DAG);
+ return lowerV4F32VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
case MVT::v8i16:
return lowerV8I16VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
case MVT::v16i8: