// 2. [ssse3] 1 x pshufb
// 3. [ssse3] 2 x pshufb + 1 x por
// 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
-static
-SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
- SelectionDAG &DAG,
- const X86TargetLowering &TLI) {
+SDValue
+X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
+ SelectionDAG &DAG) const {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
SDValue V1 = SVOp->getOperand(0);
SDValue V2 = SVOp->getOperand(1);
DebugLoc dl = SVOp->getDebugLoc();
// quads, disable the next transformation since it does not help SSSE3.
bool V1Used = InputQuads[0] || InputQuads[1];
bool V2Used = InputQuads[2] || InputQuads[3];
- if (TLI.getSubtarget()->hasSSSE3()) {
+ if (Subtarget->hasSSSE3()) {
if (InputQuads.count() == 2 && V1Used && V2Used) {
BestLoQuad = InputQuads.find_first();
BestHiQuad = InputQuads.find_next(BestLoQuad);
NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
+ if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE)
+ NewV = LowerVECTOR_SHUFFLE(NewV, DAG);
NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
// Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
// If we have SSSE3, and all words of the result are from 1 input vector,
// case 2 is generated, otherwise case 3 is generated. If no SSSE3
// is present, fall back to case 4.
- if (TLI.getSubtarget()->hasSSSE3()) {
+ if (Subtarget->hasSSSE3()) {
SmallVector<SDValue,16> pshufbMask;
// If we have elements from both input vectors, set the high bit of the
// Handle v8i16 specifically since SSE can do byte extraction and insertion.
if (VT == MVT::v8i16) {
- SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
+ SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
if (NewOp.getNode())
return NewOp;
}