diff options
Diffstat (limited to 'contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp')
-rw-r--r-- | contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp | 638 |
1 files changed, 371 insertions, 267 deletions
diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp index 458e865..e657445 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp @@ -70,22 +70,29 @@ LimitFPPrecision("limit-float-precision", cl::location(LimitFloatPrecision), cl::init(0)); +static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL, + const SDValue *Parts, unsigned NumParts, + EVT PartVT, EVT ValueVT); + /// getCopyFromParts - Create a value that contains the specified legal parts /// combined into the value they represent. If the parts combine to a type /// larger then ValueVT then AssertOp can be used to specify whether the extra /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT /// (ISD::AssertSext). -static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, +static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc DL, const SDValue *Parts, unsigned NumParts, EVT PartVT, EVT ValueVT, ISD::NodeType AssertOp = ISD::DELETED_NODE) { + if (ValueVT.isVector()) + return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT); + assert(NumParts > 0 && "No parts to assemble!"); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); SDValue Val = Parts[0]; if (NumParts > 1) { // Assemble the value from multiple parts. - if (!ValueVT.isVector() && ValueVT.isInteger()) { + if (ValueVT.isInteger()) { unsigned PartBits = PartVT.getSizeInBits(); unsigned ValueBits = ValueVT.getSizeInBits(); @@ -100,25 +107,25 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2); if (RoundParts > 2) { - Lo = getCopyFromParts(DAG, dl, Parts, RoundParts / 2, + Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2, PartVT, HalfVT); - Hi = getCopyFromParts(DAG, dl, Parts + RoundParts / 2, + Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2, RoundParts / 2, PartVT, HalfVT); } else { - Lo = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[0]); - Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[1]); + Lo = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[0]); + Hi = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[1]); } if (TLI.isBigEndian()) std::swap(Lo, Hi); - Val = DAG.getNode(ISD::BUILD_PAIR, dl, RoundVT, Lo, Hi); + Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi); if (RoundParts < NumParts) { // Assemble the trailing non-power-of-2 part. unsigned OddParts = NumParts - RoundParts; EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits); - Hi = getCopyFromParts(DAG, dl, + Hi = getCopyFromParts(DAG, DL, Parts + RoundParts, OddParts, PartVT, OddVT); // Combine the round and odd parts. @@ -126,68 +133,29 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, if (TLI.isBigEndian()) std::swap(Lo, Hi); EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); - Hi = DAG.getNode(ISD::ANY_EXTEND, dl, TotalVT, Hi); - Hi = DAG.getNode(ISD::SHL, dl, TotalVT, Hi, + Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi); + Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi, DAG.getConstant(Lo.getValueType().getSizeInBits(), TLI.getPointerTy())); - Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, TotalVT, Lo); - Val = DAG.getNode(ISD::OR, dl, TotalVT, Lo, Hi); - } - } else if (ValueVT.isVector()) { - // Handle a multi-element vector. - EVT IntermediateVT, RegisterVT; - unsigned NumIntermediates; - unsigned NumRegs = - TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT, - NumIntermediates, RegisterVT); - assert(NumRegs == NumParts - && "Part count doesn't match vector breakdown!"); - NumParts = NumRegs; // Silence a compiler warning. - assert(RegisterVT == PartVT - && "Part type doesn't match vector breakdown!"); - assert(RegisterVT == Parts[0].getValueType() && - "Part type doesn't match part!"); - - // Assemble the parts into intermediate operands. - SmallVector<SDValue, 8> Ops(NumIntermediates); - if (NumIntermediates == NumParts) { - // If the register was not expanded, truncate or copy the value, - // as appropriate. - for (unsigned i = 0; i != NumParts; ++i) - Ops[i] = getCopyFromParts(DAG, dl, &Parts[i], 1, - PartVT, IntermediateVT); - } else if (NumParts > 0) { - // If the intermediate type was expanded, build the intermediate - // operands from the parts. - assert(NumParts % NumIntermediates == 0 && - "Must expand into a divisible number of parts!"); - unsigned Factor = NumParts / NumIntermediates; - for (unsigned i = 0; i != NumIntermediates; ++i) - Ops[i] = getCopyFromParts(DAG, dl, &Parts[i * Factor], Factor, - PartVT, IntermediateVT); + Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo); + Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi); } - - // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the - // intermediate operands. - Val = DAG.getNode(IntermediateVT.isVector() ? - ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, dl, - ValueVT, &Ops[0], NumIntermediates); } else if (PartVT.isFloatingPoint()) { // FP split into multiple FP parts (for ppcf128) assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) && "Unexpected split"); SDValue Lo, Hi; - Lo = DAG.getNode(ISD::BIT_CONVERT, dl, EVT(MVT::f64), Parts[0]); - Hi = DAG.getNode(ISD::BIT_CONVERT, dl, EVT(MVT::f64), Parts[1]); + Lo = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[0]); + Hi = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[1]); if (TLI.isBigEndian()) std::swap(Lo, Hi); - Val = DAG.getNode(ISD::BUILD_PAIR, dl, ValueVT, Lo, Hi); + Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi); } else { // FP split into integer parts (soft fp) assert(ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && "Unexpected split"); EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits()); - Val = getCopyFromParts(DAG, dl, Parts, NumParts, PartVT, IntVT); + Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT); } } @@ -197,219 +165,315 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, if (PartVT == ValueVT) return Val; - if (PartVT.isVector()) { - assert(ValueVT.isVector() && "Unknown vector conversion!"); - return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val); - } - - if (ValueVT.isVector()) { - assert(ValueVT.getVectorElementType() == PartVT && - ValueVT.getVectorNumElements() == 1 && - "Only trivial scalar-to-vector conversions should get here!"); - return DAG.getNode(ISD::BUILD_VECTOR, dl, ValueVT, Val); - } - - if (PartVT.isInteger() && - ValueVT.isInteger()) { + if (PartVT.isInteger() && ValueVT.isInteger()) { if (ValueVT.bitsLT(PartVT)) { // For a truncate, see if we have any information to // indicate whether the truncated bits will always be // zero or sign-extension. if (AssertOp != ISD::DELETED_NODE) - Val = DAG.getNode(AssertOp, dl, PartVT, Val, + Val = DAG.getNode(AssertOp, DL, PartVT, Val, DAG.getValueType(ValueVT)); - return DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val); - } else { - return DAG.getNode(ISD::ANY_EXTEND, dl, ValueVT, Val); + return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); } + return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val); } if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { - if (ValueVT.bitsLT(Val.getValueType())) { - // FP_ROUND's are always exact here. - return DAG.getNode(ISD::FP_ROUND, dl, ValueVT, Val, + // FP_ROUND's are always exact here. + if (ValueVT.bitsLT(Val.getValueType())) + return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val, DAG.getIntPtrConstant(1)); - } - return DAG.getNode(ISD::FP_EXTEND, dl, ValueVT, Val); + return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val); } if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) - return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val); + return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val); llvm_unreachable("Unknown mismatch!"); return SDValue(); } +/// getCopyFromParts - Create a value that contains the specified legal parts +/// combined into the value they represent. If the parts combine to a type +/// larger then ValueVT then AssertOp can be used to specify whether the extra +/// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT +/// (ISD::AssertSext). +static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL, + const SDValue *Parts, unsigned NumParts, + EVT PartVT, EVT ValueVT) { + assert(ValueVT.isVector() && "Not a vector value"); + assert(NumParts > 0 && "No parts to assemble!"); + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); + SDValue Val = Parts[0]; + + // Handle a multi-element vector. + if (NumParts > 1) { + EVT IntermediateVT, RegisterVT; + unsigned NumIntermediates; + unsigned NumRegs = + TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT, + NumIntermediates, RegisterVT); + assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!"); + NumParts = NumRegs; // Silence a compiler warning. + assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!"); + assert(RegisterVT == Parts[0].getValueType() && + "Part type doesn't match part!"); + + // Assemble the parts into intermediate operands. + SmallVector<SDValue, 8> Ops(NumIntermediates); + if (NumIntermediates == NumParts) { + // If the register was not expanded, truncate or copy the value, + // as appropriate. + for (unsigned i = 0; i != NumParts; ++i) + Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1, + PartVT, IntermediateVT); + } else if (NumParts > 0) { + // If the intermediate type was expanded, build the intermediate + // operands from the parts. + assert(NumParts % NumIntermediates == 0 && + "Must expand into a divisible number of parts!"); + unsigned Factor = NumParts / NumIntermediates; + for (unsigned i = 0; i != NumIntermediates; ++i) + Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor, + PartVT, IntermediateVT); + } + + // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the + // intermediate operands. + Val = DAG.getNode(IntermediateVT.isVector() ? + ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL, + ValueVT, &Ops[0], NumIntermediates); + } + + // There is now one part, held in Val. Correct it to match ValueVT. + PartVT = Val.getValueType(); + + if (PartVT == ValueVT) + return Val; + + if (PartVT.isVector()) { + // If the element type of the source/dest vectors are the same, but the + // parts vector has more elements than the value vector, then we have a + // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the + // elements we want. + if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) { + assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() && + "Cannot narrow, it would be a lossy transformation"); + return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val, + DAG.getIntPtrConstant(0)); + } + + // Vector/Vector bitcast. + return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val); + } + + assert(ValueVT.getVectorElementType() == PartVT && + ValueVT.getVectorNumElements() == 1 && + "Only trivial scalar-to-vector conversions should get here!"); + return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val); +} + + + + +static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl, + SDValue Val, SDValue *Parts, unsigned NumParts, + EVT PartVT); + /// getCopyToParts - Create a series of nodes that contain the specified value /// split into legal parts. If the parts contain more bits than Val, then, for /// integers, ExtendKind can be used to specify how to generate the extra bits. -static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, +static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL, SDValue Val, SDValue *Parts, unsigned NumParts, EVT PartVT, ISD::NodeType ExtendKind = ISD::ANY_EXTEND) { - const TargetLowering &TLI = DAG.getTargetLoweringInfo(); - EVT PtrVT = TLI.getPointerTy(); EVT ValueVT = Val.getValueType(); + + // Handle the vector case separately. + if (ValueVT.isVector()) + return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT); + + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); unsigned PartBits = PartVT.getSizeInBits(); unsigned OrigNumParts = NumParts; assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!"); - if (!NumParts) + if (NumParts == 0) return; - if (!ValueVT.isVector()) { - if (PartVT == ValueVT) { - assert(NumParts == 1 && "No-op copy with multiple parts!"); - Parts[0] = Val; - return; - } - - if (NumParts * PartBits > ValueVT.getSizeInBits()) { - // If the parts cover more bits than the value has, promote the value. - if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { - assert(NumParts == 1 && "Do not know what to promote to!"); - Val = DAG.getNode(ISD::FP_EXTEND, dl, PartVT, Val); - } else if (PartVT.isInteger() && ValueVT.isInteger()) { - ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); - Val = DAG.getNode(ExtendKind, dl, ValueVT, Val); - } else { - llvm_unreachable("Unknown mismatch!"); - } - } else if (PartBits == ValueVT.getSizeInBits()) { - // Different types of the same size. - assert(NumParts == 1 && PartVT != ValueVT); - Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val); - } else if (NumParts * PartBits < ValueVT.getSizeInBits()) { - // If the parts cover less bits than value has, truncate the value. - if (PartVT.isInteger() && ValueVT.isInteger()) { - ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); - Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val); - } else { - llvm_unreachable("Unknown mismatch!"); - } - } - - // The value may have changed - recompute ValueVT. - ValueVT = Val.getValueType(); - assert(NumParts * PartBits == ValueVT.getSizeInBits() && - "Failed to tile the value with PartVT!"); - - if (NumParts == 1) { - assert(PartVT == ValueVT && "Type conversion failed!"); - Parts[0] = Val; - return; - } + assert(!ValueVT.isVector() && "Vector case handled elsewhere"); + if (PartVT == ValueVT) { + assert(NumParts == 1 && "No-op copy with multiple parts!"); + Parts[0] = Val; + return; + } - // Expand the value into multiple parts. - if (NumParts & (NumParts - 1)) { - // The number of parts is not a power of 2. Split off and copy the tail. + if (NumParts * PartBits > ValueVT.getSizeInBits()) { + // If the parts cover more bits than the value has, promote the value. + if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { + assert(NumParts == 1 && "Do not know what to promote to!"); + Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val); + } else { assert(PartVT.isInteger() && ValueVT.isInteger() && - "Do not know what to expand to!"); - unsigned RoundParts = 1 << Log2_32(NumParts); - unsigned RoundBits = RoundParts * PartBits; - unsigned OddParts = NumParts - RoundParts; - SDValue OddVal = DAG.getNode(ISD::SRL, dl, ValueVT, Val, - DAG.getConstant(RoundBits, - TLI.getPointerTy())); - getCopyToParts(DAG, dl, OddVal, Parts + RoundParts, - OddParts, PartVT); - - if (TLI.isBigEndian()) - // The odd parts were reversed by getCopyToParts - unreverse them. - std::reverse(Parts + RoundParts, Parts + NumParts); - - NumParts = RoundParts; + "Unknown mismatch!"); ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); - Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val); + Val = DAG.getNode(ExtendKind, DL, ValueVT, Val); } + } else if (PartBits == ValueVT.getSizeInBits()) { + // Different types of the same size. + assert(NumParts == 1 && PartVT != ValueVT); + Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val); + } else if (NumParts * PartBits < ValueVT.getSizeInBits()) { + // If the parts cover less bits than value has, truncate the value. + assert(PartVT.isInteger() && ValueVT.isInteger() && + "Unknown mismatch!"); + ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); + Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); + } + + // The value may have changed - recompute ValueVT. + ValueVT = Val.getValueType(); + assert(NumParts * PartBits == ValueVT.getSizeInBits() && + "Failed to tile the value with PartVT!"); - // The number of parts is a power of 2. Repeatedly bisect the value using - // EXTRACT_ELEMENT. - Parts[0] = DAG.getNode(ISD::BIT_CONVERT, dl, - EVT::getIntegerVT(*DAG.getContext(), - ValueVT.getSizeInBits()), - Val); - - for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) { - for (unsigned i = 0; i < NumParts; i += StepSize) { - unsigned ThisBits = StepSize * PartBits / 2; - EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits); - SDValue &Part0 = Parts[i]; - SDValue &Part1 = Parts[i+StepSize/2]; - - Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, - ThisVT, Part0, - DAG.getConstant(1, PtrVT)); - Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, - ThisVT, Part0, - DAG.getConstant(0, PtrVT)); - - if (ThisBits == PartBits && ThisVT != PartVT) { - Part0 = DAG.getNode(ISD::BIT_CONVERT, dl, - PartVT, Part0); - Part1 = DAG.getNode(ISD::BIT_CONVERT, dl, - PartVT, Part1); - } + if (NumParts == 1) { + assert(PartVT == ValueVT && "Type conversion failed!"); + Parts[0] = Val; + return; + } + + // Expand the value into multiple parts. + if (NumParts & (NumParts - 1)) { + // The number of parts is not a power of 2. Split off and copy the tail. + assert(PartVT.isInteger() && ValueVT.isInteger() && + "Do not know what to expand to!"); + unsigned RoundParts = 1 << Log2_32(NumParts); + unsigned RoundBits = RoundParts * PartBits; + unsigned OddParts = NumParts - RoundParts; + SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val, + DAG.getIntPtrConstant(RoundBits)); + getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT); + + if (TLI.isBigEndian()) + // The odd parts were reversed by getCopyToParts - unreverse them. + std::reverse(Parts + RoundParts, Parts + NumParts); + + NumParts = RoundParts; + ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); + Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); + } + + // The number of parts is a power of 2. Repeatedly bisect the value using + // EXTRACT_ELEMENT. + Parts[0] = DAG.getNode(ISD::BIT_CONVERT, DL, + EVT::getIntegerVT(*DAG.getContext(), + ValueVT.getSizeInBits()), + Val); + + for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) { + for (unsigned i = 0; i < NumParts; i += StepSize) { + unsigned ThisBits = StepSize * PartBits / 2; + EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits); + SDValue &Part0 = Parts[i]; + SDValue &Part1 = Parts[i+StepSize/2]; + + Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, + ThisVT, Part0, DAG.getIntPtrConstant(1)); + Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, + ThisVT, Part0, DAG.getIntPtrConstant(0)); + + if (ThisBits == PartBits && ThisVT != PartVT) { + Part0 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part0); + Part1 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part1); } } + } - if (TLI.isBigEndian()) - std::reverse(Parts, Parts + OrigNumParts); + if (TLI.isBigEndian()) + std::reverse(Parts, Parts + OrigNumParts); +} - return; - } - // Vector ValueVT. +/// getCopyToPartsVector - Create a series of nodes that contain the specified +/// value split into legal parts. +static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL, + SDValue Val, SDValue *Parts, unsigned NumParts, + EVT PartVT) { + EVT ValueVT = Val.getValueType(); + assert(ValueVT.isVector() && "Not a vector"); + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); + if (NumParts == 1) { - if (PartVT != ValueVT) { - if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) { - Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val); - } else { - assert(ValueVT.getVectorElementType() == PartVT && - ValueVT.getVectorNumElements() == 1 && - "Only trivial vector-to-scalar conversions should get here!"); - Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, - PartVT, Val, - DAG.getConstant(0, PtrVT)); - } - } + if (PartVT == ValueVT) { + // Nothing to do. + } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) { + // Bitconvert vector->vector case. + Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val); + } else if (PartVT.isVector() && + PartVT.getVectorElementType() == ValueVT.getVectorElementType()&& + PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) { + EVT ElementVT = PartVT.getVectorElementType(); + // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in + // undef elements. + SmallVector<SDValue, 16> Ops; + for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i) + Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, + ElementVT, Val, DAG.getIntPtrConstant(i))); + + for (unsigned i = ValueVT.getVectorNumElements(), + e = PartVT.getVectorNumElements(); i != e; ++i) + Ops.push_back(DAG.getUNDEF(ElementVT)); + + Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size()); + // FIXME: Use CONCAT for 2x -> 4x. + + //SDValue UndefElts = DAG.getUNDEF(VectorTy); + //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts); + } else { + // Vector -> scalar conversion. + assert(ValueVT.getVectorElementType() == PartVT && + ValueVT.getVectorNumElements() == 1 && + "Only trivial vector-to-scalar conversions should get here!"); + Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, + PartVT, Val, DAG.getIntPtrConstant(0)); + } + Parts[0] = Val; return; } - + // Handle a multi-element vector. EVT IntermediateVT, RegisterVT; unsigned NumIntermediates; unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, - IntermediateVT, NumIntermediates, RegisterVT); + IntermediateVT, + NumIntermediates, RegisterVT); unsigned NumElements = ValueVT.getVectorNumElements(); - + assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!"); NumParts = NumRegs; // Silence a compiler warning. assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!"); - + // Split the vector into intermediate operands. SmallVector<SDValue, 8> Ops(NumIntermediates); for (unsigned i = 0; i != NumIntermediates; ++i) { if (IntermediateVT.isVector()) - Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, + Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, IntermediateVT, Val, - DAG.getConstant(i * (NumElements / NumIntermediates), - PtrVT)); + DAG.getIntPtrConstant(i * (NumElements / NumIntermediates))); else - Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, - IntermediateVT, Val, - DAG.getConstant(i, PtrVT)); + Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, + IntermediateVT, Val, DAG.getIntPtrConstant(i)); } - + // Split the intermediate operands into legal parts. if (NumParts == NumIntermediates) { // If the register was not expanded, promote or copy the value, // as appropriate. for (unsigned i = 0; i != NumParts; ++i) - getCopyToParts(DAG, dl, Ops[i], &Parts[i], 1, PartVT); + getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT); } else if (NumParts > 0) { // If the intermediate type was expanded, split each the value into // legal parts. @@ -417,10 +481,13 @@ static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, "Must expand into a divisible number of parts!"); unsigned Factor = NumParts / NumIntermediates; for (unsigned i = 0; i != NumIntermediates; ++i) - getCopyToParts(DAG, dl, Ops[i], &Parts[i*Factor], Factor, PartVT); + getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT); } } + + + namespace { /// RegsForValue - This struct represents the registers (physical or virtual) /// that a particular set of values is assigned, and the type information @@ -460,11 +527,6 @@ namespace { EVT regvt, EVT valuevt) : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {} - RegsForValue(const SmallVector<unsigned, 4> ®s, - const SmallVector<EVT, 4> ®vts, - const SmallVector<EVT, 4> &valuevts) - : ValueVTs(valuevts), RegVTs(regvts), Regs(regs) {} - RegsForValue(LLVMContext &Context, const TargetLowering &tli, unsigned Reg, const Type *Ty) { ComputeValueVTs(tli, Ty, ValueVTs); @@ -530,6 +592,10 @@ SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo, DebugLoc dl, SDValue &Chain, SDValue *Flag) const { + // A Value with type {} or [0 x %t] needs no registers. + if (ValueVTs.empty()) + return SDValue(); + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); // Assemble the legal parts into the final values. @@ -623,8 +689,7 @@ void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT); EVT RegisterVT = RegVTs[Value]; - getCopyToParts(DAG, dl, - Val.getValue(Val.getResNo() + Value), + getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value), &Parts[Part], NumParts, RegisterVT); Part += NumParts; } @@ -701,6 +766,7 @@ void SelectionDAGBuilder::clear() { UnusedArgNodeMap.clear(); PendingLoads.clear(); PendingExports.clear(); + DanglingDebugInfoMap.clear(); CurDebugLoc = DebugLoc(); HasTailCall = false; } @@ -805,6 +871,33 @@ void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) { } } +// resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V, +// generate the debug data structures now that we've seen its definition. +void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V, + SDValue Val) { + DanglingDebugInfo &DDI = DanglingDebugInfoMap[V]; + if (DDI.getDI()) { + const DbgValueInst *DI = DDI.getDI(); + DebugLoc dl = DDI.getdl(); + unsigned DbgSDNodeOrder = DDI.getSDNodeOrder(); + MDNode *Variable = DI->getVariable(); + uint64_t Offset = DI->getOffset(); + SDDbgValue *SDV; + if (Val.getNode()) { + if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) { + SDV = DAG.getDbgValue(Variable, Val.getNode(), + Val.getResNo(), Offset, dl, DbgSDNodeOrder); + DAG.AddDbgValue(SDV, Val.getNode(), false); + } + } else { + SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()), + Offset, dl, SDNodeOrder); + DAG.AddDbgValue(SDV, 0, false); + } + DanglingDebugInfoMap[V] = DanglingDebugInfo(); + } +} + // getValue - Return an SDValue for the given Value. SDValue SelectionDAGBuilder::getValue(const Value *V) { // If we already have an SDValue for this value, use it. It's important @@ -826,6 +919,7 @@ SDValue SelectionDAGBuilder::getValue(const Value *V) { // Otherwise create a new SDValue and remember it. SDValue Val = getValueImpl(V); NodeMap[V] = Val; + resolveDanglingDebugInfo(V, Val); return Val; } @@ -839,10 +933,11 @@ SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) { // Otherwise create a new SDValue and remember it. SDValue Val = getValueImpl(V); NodeMap[V] = Val; + resolveDanglingDebugInfo(V, Val); return Val; } -/// getValueImpl - Helper function for getValue and getMaterializedValue. +/// getValueImpl - Helper function for getValue and getNonRegisterValue. /// Create an SDValue for the given value. SDValue SelectionDAGBuilder::getValueImpl(const Value *V) { if (const Constant *C = dyn_cast<Constant>(V)) { @@ -986,10 +1081,10 @@ void SelectionDAGBuilder::visitRet(const ReturnInst &I) { unsigned NumValues = ValueVTs.size(); SmallVector<SDValue, 4> Chains(NumValues); - EVT PtrVT = PtrValueVTs[0]; for (unsigned i = 0; i != NumValues; ++i) { - SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, RetPtr, - DAG.getConstant(Offsets[i], PtrVT)); + SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), + RetPtr.getValueType(), RetPtr, + DAG.getIntPtrConstant(Offsets[i])); Chains[i] = DAG.getStore(Chain, getCurDebugLoc(), SDValue(RetOp.getNode(), RetOp.getResNo() + i), @@ -2709,11 +2804,6 @@ void SelectionDAGBuilder::visitGetElementPtr(const User &I) { } Ty = StTy->getElementType(Field); - } else if (const UnionType *UnTy = dyn_cast<UnionType>(Ty)) { - unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); - - // Offset canonically 0 for unions, but type changes - Ty = UnTy->getElementType(Field); } else { Ty = cast<SequentialType>(Ty)->getElementType(); @@ -2818,7 +2908,7 @@ void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) { // Inform the Frame Information that we have just allocated a variable-sized // object. - FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(); + FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1); } void SelectionDAGBuilder::visitLoad(const LoadInst &I) { @@ -3824,11 +3914,11 @@ static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS, /// argument, create the corresponding DBG_VALUE machine instruction for it now. /// At the end of instruction selection, they will be inserted to the entry BB. bool -SelectionDAGBuilder::EmitFuncArgumentDbgValue(const DbgValueInst &DI, - const Value *V, MDNode *Variable, - uint64_t Offset, +SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable, + int64_t Offset, const SDValue &N) { - if (!isa<Argument>(V)) + const Argument *Arg = dyn_cast<Argument>(V); + if (!Arg) return false; MachineFunction &MF = DAG.getMachineFunction(); @@ -3842,7 +3932,15 @@ SelectionDAGBuilder::EmitFuncArgumentDbgValue(const DbgValueInst &DI, return false; unsigned Reg = 0; - if (N.getOpcode() == ISD::CopyFromReg) { + if (Arg->hasByValAttr()) { + // Byval arguments' frame index is recorded during argument lowering. + // Use this info directly. + const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo(); + Reg = TRI->getFrameRegister(MF); + Offset = FuncInfo.getByValArgumentFrameIndex(Arg); + } + + if (N.getNode() && N.getOpcode() == ISD::CopyFromReg) { Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg(); if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) { MachineRegisterInfo &RegInfo = MF.getRegInfo(); @@ -3966,42 +4064,40 @@ SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { } case Intrinsic::dbg_declare: { const DbgDeclareInst &DI = cast<DbgDeclareInst>(I); - if (!DIVariable(DI.getVariable()).Verify()) - return 0; - MDNode *Variable = DI.getVariable(); - // Parameters are handled specially. - bool isParameter = - DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable; const Value *Address = DI.getAddress(); - if (!Address) + if (!Address || !DIVariable(DI.getVariable()).Verify()) return 0; - if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) - Address = BCI->getOperand(0); - const AllocaInst *AI = dyn_cast<AllocaInst>(Address); - if (AI) { - // Don't handle byval arguments or VLAs, for example. - // Non-byval arguments are handled here (they refer to the stack temporary - // alloca at this point). - DenseMap<const AllocaInst*, int>::iterator SI = - FuncInfo.StaticAllocaMap.find(AI); - if (SI == FuncInfo.StaticAllocaMap.end()) - return 0; // VLAs. - int FI = SI->second; - - MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); - if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo()) - MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc()); - } // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder // but do not always have a corresponding SDNode built. The SDNodeOrder // absolute, but not relative, values are different depending on whether // debug info exists. ++SDNodeOrder; + + // Check if address has undef value. + if (isa<UndefValue>(Address) || + (Address->use_empty() && !isa<Argument>(Address))) { + SDDbgValue*SDV = + DAG.getDbgValue(Variable, UndefValue::get(Address->getType()), + 0, dl, SDNodeOrder); + DAG.AddDbgValue(SDV, 0, false); + return 0; + } + SDValue &N = NodeMap[Address]; + if (!N.getNode() && isa<Argument>(Address)) + // Check unused arguments map. + N = UnusedArgNodeMap[Address]; SDDbgValue *SDV; if (N.getNode()) { + // Parameters are handled specially. + bool isParameter = + DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable; + if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) + Address = BCI->getOperand(0); + const AllocaInst *AI = dyn_cast<AllocaInst>(Address); + if (isParameter && !AI) { FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode()); if (FINode) @@ -4020,10 +4116,14 @@ SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { return 0; DAG.AddDbgValue(SDV, N.getNode(), isParameter); } else { - // This isn't useful, but it shows what we're missing. - SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()), - 0, dl, SDNodeOrder); - DAG.AddDbgValue(SDV, 0, isParameter); + // If Address is an arugment then try to emits its dbg value using + // virtual register info from the FuncInfo.ValueMap. Otherwise add undef + // to help track missing debug info. + if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) { + SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()), + 0, dl, SDNodeOrder); + DAG.AddDbgValue(SDV, 0, false); + } } return 0; } @@ -4048,31 +4148,24 @@ SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder); DAG.AddDbgValue(SDV, 0, false); } else { - bool createUndef = false; - // FIXME : Why not use getValue() directly ? + // Do not use getValue() in here; we don't want to generate code at + // this point if it hasn't been done yet. SDValue N = NodeMap[V]; if (!N.getNode() && isa<Argument>(V)) // Check unused arguments map. N = UnusedArgNodeMap[V]; if (N.getNode()) { - if (!EmitFuncArgumentDbgValue(DI, V, Variable, Offset, N)) { + if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) { SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(), Offset, dl, SDNodeOrder); DAG.AddDbgValue(SDV, N.getNode(), false); } - } else if (isa<PHINode>(V) && !V->use_empty()) { - SDValue N = getValue(V); - if (N.getNode()) { - if (!EmitFuncArgumentDbgValue(DI, V, Variable, Offset, N)) { - SDV = DAG.getDbgValue(Variable, N.getNode(), - N.getResNo(), Offset, dl, SDNodeOrder); - DAG.AddDbgValue(SDV, N.getNode(), false); - } - } else - createUndef = true; - } else - createUndef = true; - if (createUndef) { + } else if (isa<PHINode>(V) && !V->use_empty() ) { + // Do not call getValue(V) yet, as we don't want to generate code. + // Remember it for later. + DanglingDebugInfo DDI(&DI, dl, SDNodeOrder); + DanglingDebugInfoMap[V] = DDI; + } else { // We may expand this to cover more cases. One case where we have no // data available is an unreferenced parameter; we need this fallback. SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()), @@ -4572,6 +4665,11 @@ void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee, !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI)) isTailCall = false; + // If there's a possibility that fast-isel has already selected some amount + // of the current basic block, don't emit a tail call. + if (isTailCall && EnableFastISel) + isTailCall = false; + std::pair<SDValue,SDValue> Result = TLI.LowerCallTo(getRoot(), RetTy, CS.paramHasAttr(0, Attribute::SExt), @@ -6054,6 +6152,12 @@ void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) { i += NumParts; } + // Note down frame index for byval arguments. + if (I->hasByValAttr() && !ArgValues.empty()) + if (FrameIndexSDNode *FI = + dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode())) + FuncInfo->setByValArgumentFrameIndex(I, FI->getIndex()); + if (!I->use_empty()) { SDValue Res; if (!ArgValues.empty()) |