X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FSelectionDAG%2FLegalizeVectorOps.cpp;h=eddf666c9c3099e4306dbb9768a310feebad9b18;hb=15c5be1ee58a67965bee79832441f1136a7698dc;hp=54a904143c86a339c999bebfc1f95d7e1df9eafa;hpb=75af7b361a780501bf8f31ec274e0ccd028582c6;p=oota-llvm.git diff --git a/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp b/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp index 54a904143c8..eddf666c9c3 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp @@ -59,6 +59,12 @@ class VectorLegalizer { /// \brief Implements unrolling a VSETCC. SDValue UnrollVSETCC(SDValue Op); + /// \brief Implement expand-based legalization of vector operations. + /// + /// This is just a high-level routine to dispatch to specific code paths for + /// operations to legalize them. + SDValue Expand(SDValue Op); + /// \brief Implements expansion for FNEG; falls back to UnrollVectorOp if /// FSUB isn't legal. /// @@ -69,6 +75,26 @@ class VectorLegalizer { /// \brief Implement expansion for SIGN_EXTEND_INREG using SRL and SRA. SDValue ExpandSEXTINREG(SDValue Op); + /// \brief Implement expansion for ANY_EXTEND_VECTOR_INREG. + /// + /// Shuffles the low lanes of the operand into place and bitcasts to the proper + /// type. The contents of the bits in the extended part of each element are + /// undef. + SDValue ExpandANY_EXTEND_VECTOR_INREG(SDValue Op); + + /// \brief Implement expansion for SIGN_EXTEND_VECTOR_INREG. + /// + /// Shuffles the low lanes of the operand into place, bitcasts to the proper + /// type, then shifts left and arithmetic shifts right to introduce a sign + /// extension. + SDValue ExpandSIGN_EXTEND_VECTOR_INREG(SDValue Op); + + /// \brief Implement expansion for ZERO_EXTEND_VECTOR_INREG. + /// + /// Shuffles the low lanes of the operand into place and blends zeros into + /// the remaining lanes, finally bitcasting to the proper type. + SDValue ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op); + /// \brief Expand bswap of vectors into a shuffle if legal. SDValue ExpandBSWAP(SDValue Op); @@ -84,18 +110,18 @@ class VectorLegalizer { /// /// This is essentially just bitcasting the operands to a different type and /// bitcasting the result back to the original type. - SDValue PromoteVectorOp(SDValue Op); + SDValue Promote(SDValue Op); /// \brief Implements [SU]INT_TO_FP vector promotion. /// /// This is a [zs]ext of the input operand to the next size up. - SDValue PromoteVectorOpINT_TO_FP(SDValue Op); + SDValue PromoteINT_TO_FP(SDValue Op); /// \brief Implements FP_TO_[SU]INT vector promotion of the result type. /// /// It is promoted to the next size up integer type. The result is then /// truncated back to the original type. - SDValue PromoteVectorOpFP_TO_INT(SDValue Op, bool isSigned); + SDValue PromoteFP_TO_INT(SDValue Op, bool isSigned); public: /// \brief Begin legalizer the vector operations in the DAG. @@ -133,7 +159,7 @@ bool VectorLegalizer::Run() { DAG.AssignTopologicalOrder(); for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(), E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I) - LegalizeOp(SDValue(I, 0)); + LegalizeOp(SDValue(&*I, 0)); // Finally, it's possible the root changed. Get the new root. SDValue OldRoot = DAG.getRoot(); @@ -165,39 +191,61 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) { // Legalize the operands SmallVector Ops; - for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) - Ops.push_back(LegalizeOp(Node->getOperand(i))); + for (const SDValue &Op : Node->op_values()) + Ops.push_back(LegalizeOp(Op)); SDValue Result = SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops), 0); + bool HasVectorValue = false; if (Op.getOpcode() == ISD::LOAD) { LoadSDNode *LD = cast(Op.getNode()); ISD::LoadExtType ExtType = LD->getExtensionType(); - if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD) { - if (TLI.isLoadExtLegal(LD->getExtensionType(), LD->getMemoryVT())) + if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD) + switch (TLI.getLoadExtAction(LD->getExtensionType(), LD->getValueType(0), + LD->getMemoryVT())) { + default: llvm_unreachable("This action is not supported yet!"); + case TargetLowering::Legal: return TranslateLegalizeResults(Op, Result); - Changed = true; - return LegalizeOp(ExpandLoad(Op)); - } + case TargetLowering::Custom: + if (SDValue Lowered = TLI.LowerOperation(Result, DAG)) { + if (Lowered == Result) + return TranslateLegalizeResults(Op, Lowered); + Changed = true; + if (Lowered->getNumValues() != Op->getNumValues()) { + // This expanded to something other than the load. Assume the + // lowering code took care of any chain values, and just handle the + // returned value. + assert(Result.getValue(1).use_empty() && + "There are still live users of the old chain!"); + return LegalizeOp(Lowered); + } + return TranslateLegalizeResults(Op, Lowered); + } + case TargetLowering::Expand: + Changed = true; + return LegalizeOp(ExpandLoad(Op)); + } } else if (Op.getOpcode() == ISD::STORE) { StoreSDNode *ST = cast(Op.getNode()); EVT StVT = ST->getMemoryVT(); MVT ValVT = ST->getValue().getSimpleValueType(); if (StVT.isVector() && ST->isTruncatingStore()) - switch (TLI.getTruncStoreAction(ValVT, StVT.getSimpleVT())) { + switch (TLI.getTruncStoreAction(ValVT, StVT)) { default: llvm_unreachable("This action is not supported yet!"); case TargetLowering::Legal: return TranslateLegalizeResults(Op, Result); - case TargetLowering::Custom: - Changed = true; - return TranslateLegalizeResults(Op, TLI.LowerOperation(Result, DAG)); + case TargetLowering::Custom: { + SDValue Lowered = TLI.LowerOperation(Result, DAG); + Changed = Lowered != Result; + return TranslateLegalizeResults(Op, Lowered); + } case TargetLowering::Expand: Changed = true; return LegalizeOp(ExpandStore(Op)); } - } + } else if (Op.getOpcode() == ISD::MSCATTER || Op.getOpcode() == ISD::MSTORE) + HasVectorValue = true; - bool HasVectorValue = false; for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end(); J != E; ++J) @@ -216,6 +264,8 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) { case ISD::UDIV: case ISD::SREM: case ISD::UREM: + case ISD::SDIVREM: + case ISD::UDIVREM: case ISD::FADD: case ISD::FSUB: case ISD::FMUL: @@ -247,6 +297,10 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) { case ISD::FP_TO_UINT: case ISD::FNEG: case ISD::FABS: + case ISD::FMINNUM: + case ISD::FMAXNUM: + case ISD::FMINNAN: + case ISD::FMAXNAN: case ISD::FCOPYSIGN: case ISD::FSQRT: case ISD::FSIN: @@ -268,6 +322,13 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) { case ISD::FP_EXTEND: case ISD::FMA: case ISD::SIGN_EXTEND_INREG: + case ISD::ANY_EXTEND_VECTOR_INREG: + case ISD::SIGN_EXTEND_VECTOR_INREG: + case ISD::ZERO_EXTEND_VECTOR_INREG: + case ISD::SMIN: + case ISD::SMAX: + case ISD::UMIN: + case ISD::UMAX: QueryType = Node->getValueType(0); break; case ISD::FP_ROUND_INREG: @@ -277,31 +338,22 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) { case ISD::UINT_TO_FP: QueryType = Node->getOperand(0).getValueType(); break; + case ISD::MSCATTER: + QueryType = cast(Node)->getValue().getValueType(); + break; + case ISD::MSTORE: + QueryType = cast(Node)->getValue().getValueType(); + break; } switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) { + default: llvm_unreachable("This action is not supported yet!"); case TargetLowering::Promote: - switch (Op.getOpcode()) { - default: - // "Promote" the operation by bitcasting - Result = PromoteVectorOp(Op); - Changed = true; - break; - case ISD::SINT_TO_FP: - case ISD::UINT_TO_FP: - // "Promote" the operation by extending the operand. - Result = PromoteVectorOpINT_TO_FP(Op); - Changed = true; - break; - case ISD::FP_TO_UINT: - case ISD::FP_TO_SINT: - // Promote the operation by extending the operand. - Result = PromoteVectorOpFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT); - Changed = true; - break; - } + Result = Promote(Op); + Changed = true; + break; + case TargetLowering::Legal: break; - case TargetLowering::Legal: break; case TargetLowering::Custom: { SDValue Tmp1 = TLI.LowerOperation(Op, DAG); if (Tmp1.getNode()) { @@ -311,23 +363,7 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) { // FALL THROUGH } case TargetLowering::Expand: - if (Node->getOpcode() == ISD::SIGN_EXTEND_INREG) - Result = ExpandSEXTINREG(Op); - else if (Node->getOpcode() == ISD::BSWAP) - Result = ExpandBSWAP(Op); - else if (Node->getOpcode() == ISD::VSELECT) - Result = ExpandVSELECT(Op); - else if (Node->getOpcode() == ISD::SELECT) - Result = ExpandSELECT(Op); - else if (Node->getOpcode() == ISD::UINT_TO_FP) - Result = ExpandUINT_TO_FLOAT(Op); - else if (Node->getOpcode() == ISD::FNEG) - Result = ExpandFNEG(Op); - else if (Node->getOpcode() == ISD::SETCC) - Result = UnrollVSETCC(Op); - else - Result = DAG.UnrollVectorOp(Op.getNode()); - break; + Result = Expand(Op); } // Make sure that the generated code is itself legal. @@ -342,10 +378,25 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) { return Result; } -SDValue VectorLegalizer::PromoteVectorOp(SDValue Op) { - // Vector "promotion" is basically just bitcasting and doing the operation - // in a different type. For example, x86 promotes ISD::AND on v2i32 to - // v1i64. +SDValue VectorLegalizer::Promote(SDValue Op) { + // For a few operations there is a specific concept for promotion based on + // the operand's type. + switch (Op.getOpcode()) { + case ISD::SINT_TO_FP: + case ISD::UINT_TO_FP: + // "Promote" the operation by extending the operand. + return PromoteINT_TO_FP(Op); + case ISD::FP_TO_UINT: + case ISD::FP_TO_SINT: + // Promote the operation by extending the operand. + return PromoteFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT); + } + + // There are currently two cases of vector promotion: + // 1) Bitcasting a vector of integers to a different type to a vector of the + // same overall length. For example, x86 promotes ISD::AND v2i32 to v1i64. + // 2) Extending a vector of floats to a vector of the same number of larger + // floats. For example, AArch64 promotes ISD::FADD on v4f16 to v4f32. MVT VT = Op.getSimpleValueType(); assert(Op.getNode()->getNumValues() == 1 && "Can't promote a vector with multiple results!"); @@ -355,17 +406,28 @@ SDValue VectorLegalizer::PromoteVectorOp(SDValue Op) { for (unsigned j = 0; j != Op.getNumOperands(); ++j) { if (Op.getOperand(j).getValueType().isVector()) - Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j)); + if (Op.getOperand(j) + .getValueType() + .getVectorElementType() + .isFloatingPoint() && + NVT.isVector() && NVT.getVectorElementType().isFloatingPoint()) + Operands[j] = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Op.getOperand(j)); + else + Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j)); else Operands[j] = Op.getOperand(j); } - - Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands); - - return DAG.getNode(ISD::BITCAST, dl, VT, Op); + + Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands, Op.getNode()->getFlags()); + if ((VT.isFloatingPoint() && NVT.isFloatingPoint()) || + (VT.isVector() && VT.getVectorElementType().isFloatingPoint() && + NVT.isVector() && NVT.getVectorElementType().isFloatingPoint())) + return DAG.getNode(ISD::FP_ROUND, dl, VT, Op, DAG.getIntPtrConstant(0, dl)); + else + return DAG.getNode(ISD::BITCAST, dl, VT, Op); } -SDValue VectorLegalizer::PromoteVectorOpINT_TO_FP(SDValue Op) { +SDValue VectorLegalizer::PromoteINT_TO_FP(SDValue Op) { // INT_TO_FP operations may require the input operand be promoted even // when the type is otherwise legal. EVT VT = Op.getOperand(0).getValueType(); @@ -401,7 +463,7 @@ SDValue VectorLegalizer::PromoteVectorOpINT_TO_FP(SDValue Op) { // elements and then truncate the result. This is different from the default // PromoteVector which uses bitcast to promote thus assumning that the // promoted vector type has the same overall size. -SDValue VectorLegalizer::PromoteVectorOpFP_TO_INT(SDValue Op, bool isSigned) { +SDValue VectorLegalizer::PromoteFP_TO_INT(SDValue Op, bool isSigned) { assert(Op.getNode()->getNumValues() == 1 && "Can't promote a vector with multiple results!"); EVT VT = Op.getValueType(); @@ -448,7 +510,7 @@ SDValue VectorLegalizer::ExpandLoad(SDValue Op) { // Instead, we load all significant words, mask bits off, and concatenate // them to form each element. Finally, they are extended to destination // scalar type to build the destination vector. - EVT WideVT = TLI.getPointerTy(); + EVT WideVT = TLI.getPointerTy(DAG.getDataLayout()); assert(WideVT.isRound() && "Could not handle the sophisticated case when the widest integer is" @@ -469,8 +531,9 @@ SDValue VectorLegalizer::ExpandLoad(SDValue Op) { ScalarLoad = DAG.getLoad(WideVT, dl, Chain, BasePTR, LD->getPointerInfo().getWithOffset(Offset), LD->isVolatile(), LD->isNonTemporal(), - LD->isInvariant(), LD->getAlignment(), - LD->getTBAAInfo()); + LD->isInvariant(), + MinAlign(LD->getAlignment(), Offset), + LD->getAAInfo()); } else { EVT LoadVT = WideVT; while (RemainingBytes < LoadBytes) { @@ -480,14 +543,16 @@ SDValue VectorLegalizer::ExpandLoad(SDValue Op) { ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR, LD->getPointerInfo().getWithOffset(Offset), LoadVT, LD->isVolatile(), - LD->isNonTemporal(), LD->getAlignment(), - LD->getTBAAInfo()); + LD->isNonTemporal(), LD->isInvariant(), + MinAlign(LD->getAlignment(), Offset), + LD->getAAInfo()); } RemainingBytes -= LoadBytes; Offset += LoadBytes; BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR, - DAG.getConstant(LoadBytes, BasePTR.getValueType())); + DAG.getConstant(LoadBytes, dl, + BasePTR.getValueType())); LoadVals.push_back(ScalarLoad.getValue(0)); LoadChains.push_back(ScalarLoad.getValue(1)); @@ -495,7 +560,7 @@ SDValue VectorLegalizer::ExpandLoad(SDValue Op) { // Extract bits, pack and extend/trunc them into destination type. unsigned SrcEltBits = SrcEltVT.getSizeInBits(); - SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, WideVT); + SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, dl, WideVT); unsigned BitOffset = 0; unsigned WideIdx = 0; @@ -505,7 +570,8 @@ SDValue VectorLegalizer::ExpandLoad(SDValue Op) { SDValue Lo, Hi, ShAmt; if (BitOffset < WideBits) { - ShAmt = DAG.getConstant(BitOffset, TLI.getShiftAmountTy(WideVT)); + ShAmt = DAG.getConstant( + BitOffset, dl, TLI.getShiftAmountTy(WideVT, DAG.getDataLayout())); Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt); Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask); } @@ -513,10 +579,11 @@ SDValue VectorLegalizer::ExpandLoad(SDValue Op) { BitOffset += SrcEltBits; if (BitOffset >= WideBits) { WideIdx++; - Offset -= WideBits; - if (Offset > 0) { - ShAmt = DAG.getConstant(SrcEltBits - Offset, - TLI.getShiftAmountTy(WideVT)); + BitOffset -= WideBits; + if (BitOffset > 0) { + ShAmt = DAG.getConstant( + SrcEltBits - BitOffset, dl, + TLI.getShiftAmountTy(WideVT, DAG.getDataLayout())); Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt); Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask); } @@ -534,8 +601,9 @@ SDValue VectorLegalizer::ExpandLoad(SDValue Op) { Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT); break; case ISD::SEXTLOAD: - ShAmt = DAG.getConstant(WideBits - SrcEltBits, - TLI.getShiftAmountTy(WideVT)); + ShAmt = + DAG.getConstant(WideBits - SrcEltBits, dl, + TLI.getShiftAmountTy(WideVT, DAG.getDataLayout())); Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt); Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt); Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT); @@ -551,11 +619,11 @@ SDValue VectorLegalizer::ExpandLoad(SDValue Op) { Op.getNode()->getValueType(0).getScalarType(), Chain, BasePTR, LD->getPointerInfo().getWithOffset(Idx * Stride), SrcVT.getScalarType(), - LD->isVolatile(), LD->isNonTemporal(), - LD->getAlignment(), LD->getTBAAInfo()); + LD->isVolatile(), LD->isNonTemporal(), LD->isInvariant(), + MinAlign(LD->getAlignment(), Idx * Stride), LD->getAAInfo()); BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR, - DAG.getConstant(Stride, BasePTR.getValueType())); + DAG.getConstant(Stride, dl, BasePTR.getValueType())); Vals.push_back(ScalarLoad.getValue(0)); LoadChains.push_back(ScalarLoad.getValue(1)); @@ -583,7 +651,7 @@ SDValue VectorLegalizer::ExpandStore(SDValue Op) { unsigned Alignment = ST->getAlignment(); bool isVolatile = ST->isVolatile(); bool isNonTemporal = ST->isNonTemporal(); - const MDNode *TBAAInfo = ST->getTBAAInfo(); + AAMDNodes AAInfo = ST->getAAInfo(); unsigned NumElem = StVT.getVectorNumElements(); // The type of the data we want to save @@ -605,16 +673,18 @@ SDValue VectorLegalizer::ExpandStore(SDValue Op) { // and save them into memory individually. SmallVector Stores; for (unsigned Idx = 0; Idx < NumElem; Idx++) { - SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, - RegSclVT, Value, DAG.getConstant(Idx, TLI.getVectorIdxTy())); + SDValue Ex = DAG.getNode( + ISD::EXTRACT_VECTOR_ELT, dl, RegSclVT, Value, + DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))); // This scalar TruncStore may be illegal, but we legalize it later. SDValue Store = DAG.getTruncStore(Chain, dl, Ex, BasePTR, ST->getPointerInfo().getWithOffset(Idx*Stride), MemSclVT, - isVolatile, isNonTemporal, Alignment, TBAAInfo); + isVolatile, isNonTemporal, MinAlign(Alignment, Idx*Stride), + AAInfo); BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR, - DAG.getConstant(Stride, BasePTR.getValueType())); + DAG.getConstant(Stride, dl, BasePTR.getValueType())); Stores.push_back(Store); } @@ -623,6 +693,33 @@ SDValue VectorLegalizer::ExpandStore(SDValue Op) { return TF; } +SDValue VectorLegalizer::Expand(SDValue Op) { + switch (Op->getOpcode()) { + case ISD::SIGN_EXTEND_INREG: + return ExpandSEXTINREG(Op); + case ISD::ANY_EXTEND_VECTOR_INREG: + return ExpandANY_EXTEND_VECTOR_INREG(Op); + case ISD::SIGN_EXTEND_VECTOR_INREG: + return ExpandSIGN_EXTEND_VECTOR_INREG(Op); + case ISD::ZERO_EXTEND_VECTOR_INREG: + return ExpandZERO_EXTEND_VECTOR_INREG(Op); + case ISD::BSWAP: + return ExpandBSWAP(Op); + case ISD::VSELECT: + return ExpandVSELECT(Op); + case ISD::SELECT: + return ExpandSELECT(Op); + case ISD::UINT_TO_FP: + return ExpandUINT_TO_FLOAT(Op); + case ISD::FNEG: + return ExpandFNEG(Op); + case ISD::SETCC: + return UnrollVSETCC(Op); + default: + return DAG.UnrollVectorOp(Op.getNode()); + } +} + SDValue VectorLegalizer::ExpandSELECT(SDValue Op) { // Lower a select instruction where the condition is a scalar and the // operands are vectors. Lower this select to VSELECT and implement it @@ -657,8 +754,9 @@ SDValue VectorLegalizer::ExpandSELECT(SDValue Op) { EVT BitTy = MaskTy.getScalarType(); Mask = DAG.getSelect(DL, BitTy, Mask, - DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), BitTy), - DAG.getConstant(0, BitTy)); + DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), DL, + BitTy), + DAG.getConstant(0, DL, BitTy)); // Broadcast the mask so that the entire vector is all-one or all zero. SmallVector Ops(NumElem, Mask); @@ -671,7 +769,7 @@ SDValue VectorLegalizer::ExpandSELECT(SDValue Op) { Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2); SDValue AllOnes = DAG.getConstant( - APInt::getAllOnesValue(BitTy.getSizeInBits()), MaskTy); + APInt::getAllOnesValue(BitTy.getSizeInBits()), DL, MaskTy); SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes); Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask); @@ -693,13 +791,92 @@ SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) { unsigned BW = VT.getScalarType().getSizeInBits(); unsigned OrigBW = OrigTy.getScalarType().getSizeInBits(); - SDValue ShiftSz = DAG.getConstant(BW - OrigBW, VT); + SDValue ShiftSz = DAG.getConstant(BW - OrigBW, DL, VT); Op = Op.getOperand(0); Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz); return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz); } +// Generically expand a vector anyext in register to a shuffle of the relevant +// lanes into the appropriate locations, with other lanes left undef. +SDValue VectorLegalizer::ExpandANY_EXTEND_VECTOR_INREG(SDValue Op) { + SDLoc DL(Op); + EVT VT = Op.getValueType(); + int NumElements = VT.getVectorNumElements(); + SDValue Src = Op.getOperand(0); + EVT SrcVT = Src.getValueType(); + int NumSrcElements = SrcVT.getVectorNumElements(); + + // Build a base mask of undef shuffles. + SmallVector ShuffleMask; + ShuffleMask.resize(NumSrcElements, -1); + + // Place the extended lanes into the correct locations. + int ExtLaneScale = NumSrcElements / NumElements; + int EndianOffset = DAG.getDataLayout().isBigEndian() ? ExtLaneScale - 1 : 0; + for (int i = 0; i < NumElements; ++i) + ShuffleMask[i * ExtLaneScale + EndianOffset] = i; + + return DAG.getNode( + ISD::BITCAST, DL, VT, + DAG.getVectorShuffle(SrcVT, DL, Src, DAG.getUNDEF(SrcVT), ShuffleMask)); +} + +SDValue VectorLegalizer::ExpandSIGN_EXTEND_VECTOR_INREG(SDValue Op) { + SDLoc DL(Op); + EVT VT = Op.getValueType(); + SDValue Src = Op.getOperand(0); + EVT SrcVT = Src.getValueType(); + + // First build an any-extend node which can be legalized above when we + // recurse through it. + Op = DAG.getAnyExtendVectorInReg(Src, DL, VT); + + // Now we need sign extend. Do this by shifting the elements. Even if these + // aren't legal operations, they have a better chance of being legalized + // without full scalarization than the sign extension does. + unsigned EltWidth = VT.getVectorElementType().getSizeInBits(); + unsigned SrcEltWidth = SrcVT.getVectorElementType().getSizeInBits(); + SDValue ShiftAmount = DAG.getConstant(EltWidth - SrcEltWidth, DL, VT); + return DAG.getNode(ISD::SRA, DL, VT, + DAG.getNode(ISD::SHL, DL, VT, Op, ShiftAmount), + ShiftAmount); +} + +// Generically expand a vector zext in register to a shuffle of the relevant +// lanes into the appropriate locations, a blend of zero into the high bits, +// and a bitcast to the wider element type. +SDValue VectorLegalizer::ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op) { + SDLoc DL(Op); + EVT VT = Op.getValueType(); + int NumElements = VT.getVectorNumElements(); + SDValue Src = Op.getOperand(0); + EVT SrcVT = Src.getValueType(); + int NumSrcElements = SrcVT.getVectorNumElements(); + + // Build up a zero vector to blend into this one. + EVT SrcScalarVT = SrcVT.getScalarType(); + SDValue ScalarZero = DAG.getTargetConstant(0, DL, SrcScalarVT); + SmallVector BuildVectorOperands(NumSrcElements, ScalarZero); + SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, DL, SrcVT, BuildVectorOperands); + + // Shuffle the incoming lanes into the correct position, and pull all other + // lanes from the zero vector. + SmallVector ShuffleMask; + ShuffleMask.reserve(NumSrcElements); + for (int i = 0; i < NumSrcElements; ++i) + ShuffleMask.push_back(i); + + int ExtLaneScale = NumSrcElements / NumElements; + int EndianOffset = DAG.getDataLayout().isBigEndian() ? ExtLaneScale - 1 : 0; + for (int i = 0; i < NumElements; ++i) + ShuffleMask[i * ExtLaneScale + EndianOffset] = NumSrcElements + i; + + return DAG.getNode(ISD::BITCAST, DL, VT, + DAG.getVectorShuffle(SrcVT, DL, Zero, Src, ShuffleMask)); +} + SDValue VectorLegalizer::ExpandBSWAP(SDValue Op) { EVT VT = Op.getValueType(); @@ -743,9 +920,9 @@ SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) { // FIXME: Sign extend 1 to all ones if thats legal on the target. if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand || TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand || - TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand || - TLI.getBooleanContents(true) != - TargetLowering::ZeroOrNegativeOneBooleanContent) + TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand || + TLI.getBooleanContents(Op1.getValueType()) != + TargetLowering::ZeroOrNegativeOneBooleanContent) return DAG.UnrollVectorOp(Op.getNode()); // If the mask and the type are different sizes, unroll the vector op. This @@ -761,7 +938,7 @@ SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) { Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2); SDValue AllOnes = DAG.getConstant( - APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), VT); + APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), DL, VT); SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes); Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask); @@ -784,16 +961,16 @@ SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) { "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide"); unsigned BW = SVT.getSizeInBits(); - SDValue HalfWord = DAG.getConstant(BW/2, VT); + SDValue HalfWord = DAG.getConstant(BW/2, DL, VT); // Constants to clear the upper part of the word. // Notice that we can also use SHL+SHR, but using a constant is slightly // faster on x86. uint64_t HWMask = (SVT.getSizeInBits()==64)?0x00000000FFFFFFFF:0x0000FFFF; - SDValue HalfWordMask = DAG.getConstant(HWMask, VT); + SDValue HalfWordMask = DAG.getConstant(HWMask, DL, VT); // Two to the power of half-word-size. - SDValue TWOHW = DAG.getConstantFP((1<<(BW/2)), Op.getValueType()); + SDValue TWOHW = DAG.getConstantFP(1 << (BW/2), DL, Op.getValueType()); // Clear upper part of LO, lower HI SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord); @@ -801,6 +978,7 @@ SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) { // Convert hi and lo to floats // Convert the hi part back to the upper values + // TODO: Can any fast-math-flags be set on these nodes? SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI); fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW); SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO); @@ -812,8 +990,10 @@ SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) { SDValue VectorLegalizer::ExpandFNEG(SDValue Op) { if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) { - SDValue Zero = DAG.getConstantFP(-0.0, Op.getValueType()); - return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), + SDLoc DL(Op); + SDValue Zero = DAG.getConstantFP(-0.0, DL, Op.getValueType()); + // TODO: If FNEG had fast-math-flags, they'd get propagated to this FSUB. + return DAG.getNode(ISD::FSUB, DL, Op.getValueType(), Zero, Op.getOperand(0)); } return DAG.UnrollVectorOp(Op.getNode()); @@ -828,17 +1008,20 @@ SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) { SDLoc dl(Op); SmallVector Ops(NumElems); for (unsigned i = 0; i < NumElems; ++i) { - SDValue LHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS, - DAG.getConstant(i, TLI.getVectorIdxTy())); - SDValue RHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS, - DAG.getConstant(i, TLI.getVectorIdxTy())); + SDValue LHSElem = DAG.getNode( + ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS, + DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))); + SDValue RHSElem = DAG.getNode( + ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS, + DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))); Ops[i] = DAG.getNode(ISD::SETCC, dl, - TLI.getSetCCResultType(*DAG.getContext(), TmpEltVT), + TLI.getSetCCResultType(DAG.getDataLayout(), + *DAG.getContext(), TmpEltVT), LHSElem, RHSElem, CC); Ops[i] = DAG.getSelect(dl, EltVT, Ops[i], DAG.getConstant(APInt::getAllOnesValue - (EltVT.getSizeInBits()), EltVT), - DAG.getConstant(0, EltVT)); + (EltVT.getSizeInBits()), dl, EltVT), + DAG.getConstant(0, dl, EltVT)); } return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops); }