diff options
Diffstat (limited to 'lib')
| -rw-r--r-- | lib/Target/X86/X86ISelLowering.cpp | 20 |
1 files changed, 17 insertions, 3 deletions
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp index 55de3d3..dbc0e01 100644 --- a/lib/Target/X86/X86ISelLowering.cpp +++ b/lib/Target/X86/X86ISelLowering.cpp @@ -6662,9 +6662,10 @@ X86TargetLowering::LowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const { return SDValue(); } + LLVMContext *Context = DAG.getContext(); unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift; - EVT NeVT = EVT::getIntegerVT(*DAG.getContext(), NBits); - EVT NVT = EVT::getVectorVT(*DAG.getContext(), NeVT, NumElems >> Shift); + EVT NeVT = EVT::getIntegerVT(*Context, NBits); + EVT NVT = EVT::getVectorVT(*Context, NeVT, NumElems >> Shift); if (!isTypeLegal(NVT)) return SDValue(); @@ -6683,8 +6684,21 @@ X86TargetLowering::LowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const { // If it's foldable, i.e. normal load with single use, we will let code // selection to fold it. Otherwise, we will short the conversion sequence. if (CIdx && CIdx->getZExtValue() == 0 && - (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse())) + (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse())) { + if (V.getValueSizeInBits() > V1.getValueSizeInBits()) { + // The "ext_vec_elt" node is wider than the result node. + // In this case we should extract subvector from V. + // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast (extract_subvector x)). + unsigned Ratio = V.getValueSizeInBits() / V1.getValueSizeInBits(); + EVT FullVT = V.getValueType(); + EVT SubVecVT = EVT::getVectorVT(*Context, + FullVT.getVectorElementType(), + FullVT.getVectorNumElements()/Ratio); + V = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, V, + DAG.getIntPtrConstant(0)); + } V1 = DAG.getNode(ISD::BITCAST, DL, V1.getValueType(), V); + } } return DAG.getNode(ISD::BITCAST, DL, VT, |