X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FVMCore%2FConstantFold.cpp;h=b743287adf3941fd3a31f3fee8e9063d99672d01;hb=6bbab86af959a8819ddadfe8909bcfa5aa53ce9f;hp=194a6d4d8c06c8113f92d26fedd7a2fc0f608cd6;hpb=9ee0d8a917a90380afc65c5b8fcea9d9975e6749;p=oota-llvm.git diff --git a/lib/VMCore/ConstantFold.cpp b/lib/VMCore/ConstantFold.cpp index 194a6d4d8c0..b743287adf3 100644 --- a/lib/VMCore/ConstantFold.cpp +++ b/lib/VMCore/ConstantFold.cpp @@ -24,6 +24,7 @@ #include "llvm/Function.h" #include "llvm/GlobalAlias.h" #include "llvm/GlobalVariable.h" +#include "llvm/Operator.h" #include "llvm/ADT/SmallVector.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/ErrorHandling.h" @@ -37,31 +38,33 @@ using namespace llvm; // ConstantFold*Instruction Implementations //===----------------------------------------------------------------------===// -/// BitCastConstantVector - Convert the specified ConstantVector node to the +/// BitCastConstantVector - Convert the specified vector Constant node to the /// specified vector type. At this point, we know that the elements of the /// input vector constant are all simple integer or FP values. -static Constant *BitCastConstantVector(ConstantVector *CV, - const VectorType *DstTy) { +static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { + + if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); + if (CV->isNullValue()) return Constant::getNullValue(DstTy); + // If this cast changes element count then we can't handle it here: // doing so requires endianness information. This should be handled by // Analysis/ConstantFolding.cpp unsigned NumElts = DstTy->getNumElements(); - if (NumElts != CV->getNumOperands()) + if (NumElts != CV->getType()->getVectorNumElements()) return 0; + + Type *DstEltTy = DstTy->getElementType(); // Check to verify that all elements of the input are simple. + SmallVector Result; for (unsigned i = 0; i != NumElts; ++i) { - if (!isa(CV->getOperand(i)) && - !isa(CV->getOperand(i))) - return 0; + Constant *C = CV->getAggregateElement(i); + if (C == 0) return 0; + C = ConstantExpr::getBitCast(C, DstEltTy); + if (isa(C)) return 0; + Result.push_back(C); } - // Bitcast each element now. - std::vector Result; - const Type *DstEltTy = DstTy->getElementType(); - for (unsigned i = 0; i != NumElts; ++i) - Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i), - DstEltTy)); return ConstantVector::get(Result); } @@ -73,15 +76,15 @@ static unsigned foldConstantCastPair( unsigned opc, ///< opcode of the second cast constant expression ConstantExpr *Op, ///< the first cast constant expression - const Type *DstTy ///< desintation type of the first cast + Type *DstTy ///< desintation type of the first cast ) { assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); assert(CastInst::isCast(opc) && "Invalid cast opcode"); // The the types and opcodes for the two Cast constant expressions - const Type *SrcTy = Op->getOperand(0)->getType(); - const Type *MidTy = Op->getType(); + Type *SrcTy = Op->getOperand(0)->getType(); + Type *MidTy = Op->getType(); Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); Instruction::CastOps secondOp = Instruction::CastOps(opc); @@ -90,27 +93,28 @@ foldConstantCastPair( Type::getInt64Ty(DstTy->getContext())); } -static Constant *FoldBitCast(Constant *V, const Type *DestTy) { - const Type *SrcTy = V->getType(); +static Constant *FoldBitCast(Constant *V, Type *DestTy) { + Type *SrcTy = V->getType(); if (SrcTy == DestTy) return V; // no-op cast // Check to see if we are casting a pointer to an aggregate to a pointer to // the first element. If so, return the appropriate GEP instruction. - if (const PointerType *PTy = dyn_cast(V->getType())) - if (const PointerType *DPTy = dyn_cast(DestTy)) - if (PTy->getAddressSpace() == DPTy->getAddressSpace()) { + if (PointerType *PTy = dyn_cast(V->getType())) + if (PointerType *DPTy = dyn_cast(DestTy)) + if (PTy->getAddressSpace() == DPTy->getAddressSpace() + && DPTy->getElementType()->isSized()) { SmallVector IdxList; Value *Zero = Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); IdxList.push_back(Zero); - const Type *ElTy = PTy->getElementType(); + Type *ElTy = PTy->getElementType(); while (ElTy != DPTy->getElementType()) { - if (const StructType *STy = dyn_cast(ElTy)) { + if (StructType *STy = dyn_cast(ElTy)) { if (STy->getNumElements() == 0) break; ElTy = STy->getElementType(0); IdxList.push_back(Zero); - } else if (const SequentialType *STy = + } else if (SequentialType *STy = dyn_cast(ElTy)) { if (ElTy->isPointerTy()) break; // Can't index into pointers! ElTy = STy->getElementType(); @@ -122,14 +126,13 @@ static Constant *FoldBitCast(Constant *V, const Type *DestTy) { if (ElTy == DPTy->getElementType()) // This GEP is inbounds because all indices are zero. - return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0], - IdxList.size()); + return ConstantExpr::getInBoundsGetElementPtr(V, IdxList); } // Handle casts from one vector constant to another. We know that the src // and dest type have the same size (otherwise its an illegal cast). - if (const VectorType *DestPTy = dyn_cast(DestTy)) { - if (const VectorType *SrcTy = dyn_cast(V->getType())) { + if (VectorType *DestPTy = dyn_cast(DestTy)) { + if (VectorType *SrcTy = dyn_cast(V->getType())) { assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && "Not cast between same sized vectors!"); SrcTy = NULL; @@ -137,15 +140,15 @@ static Constant *FoldBitCast(Constant *V, const Type *DestTy) { if (isa(V)) return Constant::getNullValue(DestTy); - if (ConstantVector *CV = dyn_cast(V)) - return BitCastConstantVector(CV, DestPTy); + // Handle ConstantVector and ConstantAggregateVector. + return BitCastConstantVector(V, DestPTy); } // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts // This allows for other simplifications (although some of them // can only be handled by Analysis/ConstantFolding.cpp). if (isa(V) || isa(V)) - return ConstantExpr::getBitCast(ConstantVector::get(&V, 1), DestPTy); + return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); } // Finally, implement bitcast folding now. The code below doesn't handle @@ -202,7 +205,7 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, APInt V = CI->getValue(); if (ByteStart) V = V.lshr(ByteStart*8); - V.trunc(ByteSize*8); + V = V.trunc(ByteSize*8); return ConstantInt::get(CI->getContext(), V); } @@ -327,15 +330,15 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, /// return null if no factoring was possible, to avoid endlessly /// bouncing an unfoldable expression back into the top-level folder. /// -static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy, +static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, bool Folded) { - if (const ArrayType *ATy = dyn_cast(Ty)) { + if (ArrayType *ATy = dyn_cast(Ty)) { Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); return ConstantExpr::getNUWMul(E, N); } - if (const StructType *STy = dyn_cast(Ty)) + if (StructType *STy = dyn_cast(Ty)) if (!STy->isPacked()) { unsigned NumElems = STy->getNumElements(); // An empty struct has size zero. @@ -357,25 +360,9 @@ static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy, } } - if (const UnionType *UTy = dyn_cast(Ty)) { - unsigned NumElems = UTy->getNumElements(); - // Check for a union with all members having the same size. - Constant *MemberSize = - getFoldedSizeOf(UTy->getElementType(0), DestTy, true); - bool AllSame = true; - for (unsigned i = 1; i != NumElems; ++i) - if (MemberSize != - getFoldedSizeOf(UTy->getElementType(i), DestTy, true)) { - AllSame = false; - break; - } - if (AllSame) - return MemberSize; - } - // Pointer size doesn't depend on the pointee type, so canonicalize them // to an arbitrary pointee. - if (const PointerType *PTy = dyn_cast(Ty)) + if (PointerType *PTy = dyn_cast(Ty)) if (!PTy->getElementType()->isIntegerTy(1)) return getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), @@ -400,11 +387,11 @@ static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy, /// return null if no factoring was possible, to avoid endlessly /// bouncing an unfoldable expression back into the top-level folder. /// -static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy, +static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, bool Folded) { // The alignment of an array is equal to the alignment of the // array element. Note that this is not always true for vectors. - if (const ArrayType *ATy = dyn_cast(Ty)) { + if (ArrayType *ATy = dyn_cast(Ty)) { Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, DestTy, @@ -413,7 +400,7 @@ static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy, return C; } - if (const StructType *STy = dyn_cast(Ty)) { + if (StructType *STy = dyn_cast(Ty)) { // Packed structs always have an alignment of 1. if (STy->isPacked()) return ConstantInt::get(DestTy, 1); @@ -438,27 +425,9 @@ static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy, return MemberAlign; } - if (const UnionType *UTy = dyn_cast(Ty)) { - // Union alignment is the maximum alignment of any member. - // Without target data, we can't compare much, but we can check to see - // if all the members have the same alignment. - unsigned NumElems = UTy->getNumElements(); - // Check for a union with all members having the same alignment. - Constant *MemberAlign = - getFoldedAlignOf(UTy->getElementType(0), DestTy, true); - bool AllSame = true; - for (unsigned i = 1; i != NumElems; ++i) - if (MemberAlign != getFoldedAlignOf(UTy->getElementType(i), DestTy, true)) { - AllSame = false; - break; - } - if (AllSame) - return MemberAlign; - } - // Pointer alignment doesn't depend on the pointee type, so canonicalize them // to an arbitrary pointee. - if (const PointerType *PTy = dyn_cast(Ty)) + if (PointerType *PTy = dyn_cast(Ty)) if (!PTy->getElementType()->isIntegerTy(1)) return getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), @@ -484,10 +453,10 @@ static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy, /// return null if no factoring was possible, to avoid endlessly /// bouncing an unfoldable expression back into the top-level folder. /// -static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo, - const Type *DestTy, +static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, + Type *DestTy, bool Folded) { - if (const ArrayType *ATy = dyn_cast(Ty)) { + if (ArrayType *ATy = dyn_cast(Ty)) { Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, DestTy, false), FieldNo, DestTy); @@ -495,7 +464,7 @@ static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo, return ConstantExpr::getNUWMul(E, N); } - if (const StructType *STy = dyn_cast(Ty)) + if (StructType *STy = dyn_cast(Ty)) if (!STy->isPacked()) { unsigned NumElems = STy->getNumElements(); // An empty struct has no members. @@ -535,7 +504,7 @@ static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo, } Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, - const Type *DestTy) { + Type *DestTy) { if (isa(V)) { // zext(undef) = 0, because the top bits will be zero. // sext(undef) = 0, because the top bits will all be the same. @@ -545,10 +514,14 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, return Constant::getNullValue(DestTy); return UndefValue::get(DestTy); } + // No compile-time operations on this type yet. if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty()) return 0; + if (V->isNullValue() && !DestTy->isX86_MMXTy()) + return Constant::getNullValue(DestTy); + // If the cast operand is a constant expression, there's a few things we can // do to try to simplify it. if (ConstantExpr *CE = dyn_cast(V)) { @@ -574,18 +547,17 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, // If the cast operand is a constant vector, perform the cast by // operating on each element. In the cast of bitcasts, the element // count may be mismatched; don't attempt to handle that here. - if (ConstantVector *CV = dyn_cast(V)) - if (DestTy->isVectorTy() && - cast(DestTy)->getNumElements() == - CV->getType()->getNumElements()) { - std::vector res; - const VectorType *DestVecTy = cast(DestTy); - const Type *DstEltTy = DestVecTy->getElementType(); - for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) - res.push_back(ConstantExpr::getCast(opc, - CV->getOperand(i), DstEltTy)); - return ConstantVector::get(DestVecTy, res); - } + if ((isa(V) || isa(V)) && + DestTy->isVectorTy() && + DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { + SmallVector res; + VectorType *DestVecTy = cast(DestTy); + Type *DstEltTy = DestVecTy->getElementType(); + for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) + res.push_back(ConstantExpr::getCast(opc, + V->getAggregateElement(i), DstEltTy)); + return ConstantVector::get(res); + } // We actually have to do a cast now. Perform the cast according to the // opcode specified. @@ -597,7 +569,8 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, if (ConstantFP *FPC = dyn_cast(V)) { bool ignored; APFloat Val = FPC->getValueAPF(); - Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle : + Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf : + DestTy->isFloatTy() ? APFloat::IEEEsingle : DestTy->isDoubleTy() ? APFloat::IEEEdouble : DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : DestTy->isFP128Ty() ? APFloat::IEEEquad : @@ -615,7 +588,7 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, uint32_t DestBitWidth = cast(DestTy)->getBitWidth(); (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, APFloat::rmTowardZero, &ignored); - APInt Val(DestBitWidth, 2, x); + APInt Val(DestBitWidth, x); return ConstantInt::get(FPC->getContext(), Val); } return 0; // Can't fold. @@ -633,7 +606,7 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, if (ConstantExpr *CE = dyn_cast(V)) if (CE->getOpcode() == Instruction::GetElementPtr && CE->getOperand(0)->isNullValue()) { - const Type *Ty = + Type *Ty = cast(CE->getOperand(0)->getType())->getElementType(); if (CE->getNumOperands() == 2) { // Handle a sizeof-like expression. @@ -648,7 +621,7 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, } else if (CE->getNumOperands() == 3 && CE->getOperand(1)->isNullValue()) { // Handle an alignof-like expression. - if (const StructType *STy = dyn_cast(Ty)) + if (StructType *STy = dyn_cast(Ty)) if (!STy->isPacked()) { ConstantInt *CI = cast(CE->getOperand(2)); if (CI->isOne() && @@ -658,7 +631,7 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, } } // Handle an offsetof-like expression. - if (Ty->isStructTy() || Ty->isArrayTy() || Ty->isVectorTy()){ + if (Ty->isStructTy() || Ty->isArrayTy()) { if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), DestTy, false)) return C; @@ -671,9 +644,7 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, case Instruction::SIToFP: if (ConstantInt *CI = dyn_cast(V)) { APInt api = CI->getValue(); - const uint64_t zero[] = {0, 0}; - APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(), - 2, zero)); + APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true); (void)apf.convertFromAPInt(api, opc==Instruction::SIToFP, APFloat::rmNearestTiesToEven); @@ -683,25 +654,22 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, case Instruction::ZExt: if (ConstantInt *CI = dyn_cast(V)) { uint32_t BitWidth = cast(DestTy)->getBitWidth(); - APInt Result(CI->getValue()); - Result.zext(BitWidth); - return ConstantInt::get(V->getContext(), Result); + return ConstantInt::get(V->getContext(), + CI->getValue().zext(BitWidth)); } return 0; case Instruction::SExt: if (ConstantInt *CI = dyn_cast(V)) { uint32_t BitWidth = cast(DestTy)->getBitWidth(); - APInt Result(CI->getValue()); - Result.sext(BitWidth); - return ConstantInt::get(V->getContext(), Result); + return ConstantInt::get(V->getContext(), + CI->getValue().sext(BitWidth)); } return 0; case Instruction::Trunc: { uint32_t DestBitWidth = cast(DestTy)->getBitWidth(); if (ConstantInt *CI = dyn_cast(V)) { - APInt Result(CI->getValue()); - Result.trunc(DestBitWidth); - return ConstantInt::get(V->getContext(), Result); + return ConstantInt::get(V->getContext(), + CI->getValue().trunc(DestBitWidth)); } // The input must be a constantexpr. See if we can simplify this based on @@ -721,31 +689,65 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, Constant *V1, Constant *V2) { - if (ConstantInt *CB = dyn_cast(Cond)) - return CB->getZExtValue() ? V1 : V2; + // Check for i1 and vector true/false conditions. + if (Cond->isNullValue()) return V2; + if (Cond->isAllOnesValue()) return V1; + + // If the condition is a vector constant, fold the result elementwise. + if (ConstantVector *CondV = dyn_cast(Cond)) { + SmallVector Result; + for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ + ConstantInt *Cond = dyn_cast(CondV->getOperand(i)); + if (Cond == 0) break; + + Constant *Res = (Cond->getZExtValue() ? V2 : V1)->getAggregateElement(i); + if (Res == 0) break; + Result.push_back(Res); + } + + // If we were able to build the vector, return it. + if (Result.size() == V1->getType()->getVectorNumElements()) + return ConstantVector::get(Result); + } + if (isa(Cond)) { + if (isa(V1)) return V1; + return V2; + } if (isa(V1)) return V2; if (isa(V2)) return V1; - if (isa(Cond)) return V1; if (V1 == V2) return V1; + + if (ConstantExpr *TrueVal = dyn_cast(V1)) { + if (TrueVal->getOpcode() == Instruction::Select) + if (TrueVal->getOperand(0) == Cond) + return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); + } + if (ConstantExpr *FalseVal = dyn_cast(V2)) { + if (FalseVal->getOpcode() == Instruction::Select) + if (FalseVal->getOperand(0) == Cond) + return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); + } + return 0; } Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, Constant *Idx) { if (isa(Val)) // ee(undef, x) -> undef - return UndefValue::get(cast(Val->getType())->getElementType()); + return UndefValue::get(Val->getType()->getVectorElementType()); if (Val->isNullValue()) // ee(zero, x) -> zero - return Constant::getNullValue( - cast(Val->getType())->getElementType()); - - if (ConstantVector *CVal = dyn_cast(Val)) { - if (ConstantInt *CIdx = dyn_cast(Idx)) { - return CVal->getOperand(CIdx->getZExtValue()); - } else if (isa(Idx)) { - // ee({w,x,y,z}, undef) -> w (an arbitrary value). - return CVal->getOperand(0); - } + return Constant::getNullValue(Val->getType()->getVectorElementType()); + // ee({w,x,y,z}, undef) -> undef + if (isa(Idx)) + return UndefValue::get(Val->getType()->getVectorElementType()); + + if (ConstantInt *CIdx = dyn_cast(Idx)) { + uint64_t Index = CIdx->getZExtValue(); + // ee({w,x,y,z}, wrong_value) -> undef + if (Index >= Val->getType()->getVectorNumElements()) + return UndefValue::get(Val->getType()->getVectorElementType()); + return Val->getAggregateElement(Index); } return 0; } @@ -755,232 +757,104 @@ Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, Constant *Idx) { ConstantInt *CIdx = dyn_cast(Idx); if (!CIdx) return 0; - APInt idxVal = CIdx->getValue(); - if (isa(Val)) { - // Insertion of scalar constant into vector undef - // Optimize away insertion of undef - if (isa(Elt)) - return Val; - // Otherwise break the aggregate undef into multiple undefs and do - // the insertion - unsigned numOps = - cast(Val->getType())->getNumElements(); - std::vector Ops; - Ops.reserve(numOps); - for (unsigned i = 0; i < numOps; ++i) { - Constant *Op = - (idxVal == i) ? Elt : UndefValue::get(Elt->getType()); - Ops.push_back(Op); - } - return ConstantVector::get(Ops); - } - if (isa(Val)) { - // Insertion of scalar constant into vector aggregate zero - // Optimize away insertion of zero - if (Elt->isNullValue()) - return Val; - // Otherwise break the aggregate zero into multiple zeros and do - // the insertion - unsigned numOps = - cast(Val->getType())->getNumElements(); - std::vector Ops; - Ops.reserve(numOps); - for (unsigned i = 0; i < numOps; ++i) { - Constant *Op = - (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType()); - Ops.push_back(Op); - } - return ConstantVector::get(Ops); - } - if (ConstantVector *CVal = dyn_cast(Val)) { - // Insertion of scalar constant into vector constant - std::vector Ops; - Ops.reserve(CVal->getNumOperands()); - for (unsigned i = 0; i < CVal->getNumOperands(); ++i) { - Constant *Op = - (idxVal == i) ? Elt : cast(CVal->getOperand(i)); - Ops.push_back(Op); + const APInt &IdxVal = CIdx->getValue(); + + SmallVector Result; + for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){ + if (i == IdxVal) { + Result.push_back(Elt); + continue; } - return ConstantVector::get(Ops); + + if (Constant *C = Val->getAggregateElement(i)) + Result.push_back(C); + else + return 0; } - - return 0; -} - -/// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef -/// return the specified element value. Otherwise return null. -static Constant *GetVectorElement(Constant *C, unsigned EltNo) { - if (ConstantVector *CV = dyn_cast(C)) - return CV->getOperand(EltNo); - - const Type *EltTy = cast(C->getType())->getElementType(); - if (isa(C)) - return Constant::getNullValue(EltTy); - if (isa(C)) - return UndefValue::get(EltTy); - return 0; + + return ConstantVector::get(Result); } Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, Constant *V2, Constant *Mask) { + unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); + Type *EltTy = V1->getType()->getVectorElementType(); + // Undefined shuffle mask -> undefined value. - if (isa(Mask)) return UndefValue::get(V1->getType()); + if (isa(Mask)) + return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); - unsigned MaskNumElts = cast(Mask->getType())->getNumElements(); - unsigned SrcNumElts = cast(V1->getType())->getNumElements(); - const Type *EltTy = cast(V1->getType())->getElementType(); + // Don't break the bitcode reader hack. + if (isa(Mask)) return 0; + + unsigned SrcNumElts = V1->getType()->getVectorNumElements(); // Loop over the shuffle mask, evaluating each element. SmallVector Result; for (unsigned i = 0; i != MaskNumElts; ++i) { - Constant *InElt = GetVectorElement(Mask, i); - if (InElt == 0) return 0; - - if (isa(InElt)) - InElt = UndefValue::get(EltTy); - else if (ConstantInt *CI = dyn_cast(InElt)) { - unsigned Elt = CI->getZExtValue(); - if (Elt >= SrcNumElts*2) - InElt = UndefValue::get(EltTy); - else if (Elt >= SrcNumElts) - InElt = GetVectorElement(V2, Elt - SrcNumElts); - else - InElt = GetVectorElement(V1, Elt); - if (InElt == 0) return 0; - } else { - // Unknown value. - return 0; + int Elt = ShuffleVectorInst::getMaskValue(Mask, i); + if (Elt == -1) { + Result.push_back(UndefValue::get(EltTy)); + continue; } + Constant *InElt; + if (unsigned(Elt) >= SrcNumElts*2) + InElt = UndefValue::get(EltTy); + else if (unsigned(Elt) >= SrcNumElts) + InElt = V2->getAggregateElement(Elt - SrcNumElts); + else + InElt = V1->getAggregateElement(Elt); + if (InElt == 0) return 0; Result.push_back(InElt); } - return ConstantVector::get(&Result[0], Result.size()); + return ConstantVector::get(Result); } Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, - const unsigned *Idxs, - unsigned NumIdx) { + ArrayRef Idxs) { // Base case: no indices, so return the entire value. - if (NumIdx == 0) + if (Idxs.empty()) return Agg; - if (isa(Agg)) // ev(undef, x) -> undef - return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(), - Idxs, - Idxs + NumIdx)); - - if (isa(Agg)) // ev(0, x) -> 0 - return - Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), - Idxs, - Idxs + NumIdx)); - - // Otherwise recurse. - if (ConstantStruct *CS = dyn_cast(Agg)) - return ConstantFoldExtractValueInstruction(CS->getOperand(*Idxs), - Idxs+1, NumIdx-1); - - if (ConstantArray *CA = dyn_cast(Agg)) - return ConstantFoldExtractValueInstruction(CA->getOperand(*Idxs), - Idxs+1, NumIdx-1); - ConstantVector *CV = cast(Agg); - return ConstantFoldExtractValueInstruction(CV->getOperand(*Idxs), - Idxs+1, NumIdx-1); + if (Constant *C = Agg->getAggregateElement(Idxs[0])) + return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); + + return 0; } Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, Constant *Val, - const unsigned *Idxs, - unsigned NumIdx) { + ArrayRef Idxs) { // Base case: no indices, so replace the entire value. - if (NumIdx == 0) + if (Idxs.empty()) return Val; - if (isa(Agg)) { - // Insertion of constant into aggregate undef - // Optimize away insertion of undef. - if (isa(Val)) - return Agg; - - // Otherwise break the aggregate undef into multiple undefs and do - // the insertion. - const CompositeType *AggTy = cast(Agg->getType()); - unsigned numOps; - if (const ArrayType *AR = dyn_cast(AggTy)) - numOps = AR->getNumElements(); - else if (AggTy->isUnionTy()) - numOps = 1; - else - numOps = cast(AggTy)->getNumElements(); - - std::vector Ops(numOps); - for (unsigned i = 0; i < numOps; ++i) { - const Type *MemberTy = AggTy->getTypeAtIndex(i); - Constant *Op = - (*Idxs == i) ? - ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy), - Val, Idxs+1, NumIdx-1) : - UndefValue::get(MemberTy); - Ops[i] = Op; - } - - if (const StructType* ST = dyn_cast(AggTy)) - return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked()); - if (const UnionType* UT = dyn_cast(AggTy)) { - assert(Ops.size() == 1 && "Union can only contain a single value!"); - return ConstantUnion::get(UT, Ops[0]); - } - return ConstantArray::get(cast(AggTy), Ops); - } + unsigned NumElts; + if (StructType *ST = dyn_cast(Agg->getType())) + NumElts = ST->getNumElements(); + else if (ArrayType *AT = dyn_cast(Agg->getType())) + NumElts = AT->getNumElements(); + else + NumElts = AT->getVectorNumElements(); - if (isa(Agg)) { - // Insertion of constant into aggregate zero - // Optimize away insertion of zero. - if (Val->isNullValue()) - return Agg; - - // Otherwise break the aggregate zero into multiple zeros and do - // the insertion. - const CompositeType *AggTy = cast(Agg->getType()); - unsigned numOps; - if (const ArrayType *AR = dyn_cast(AggTy)) - numOps = AR->getNumElements(); - else - numOps = cast(AggTy)->getNumElements(); + SmallVector Result; + for (unsigned i = 0; i != NumElts; ++i) { + Constant *C = Agg->getAggregateElement(i); + if (C == 0) return 0; - std::vector Ops(numOps); - for (unsigned i = 0; i < numOps; ++i) { - const Type *MemberTy = AggTy->getTypeAtIndex(i); - Constant *Op = - (*Idxs == i) ? - ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy), - Val, Idxs+1, NumIdx-1) : - Constant::getNullValue(MemberTy); - Ops[i] = Op; - } + if (Idxs[0] == i) + C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); - if (const StructType *ST = dyn_cast(AggTy)) - return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked()); - return ConstantArray::get(cast(AggTy), Ops); + Result.push_back(C); } - if (isa(Agg) || isa(Agg)) { - // Insertion of constant into aggregate constant. - std::vector Ops(Agg->getNumOperands()); - for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { - Constant *Op = cast(Agg->getOperand(i)); - if (*Idxs == i) - Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs+1, NumIdx-1); - Ops[i] = Op; - } - - if (const StructType* ST = dyn_cast(Agg->getType())) - return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked()); - return ConstantArray::get(cast(Agg->getType()), Ops); - } - - return 0; + if (StructType *ST = dyn_cast(Agg->getType())) + return ConstantStruct::get(ST, Result); + if (ArrayType *AT = dyn_cast(Agg->getType())) + return ConstantArray::get(AT, Result); + return ConstantVector::get(Result); } @@ -1002,33 +876,53 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, case Instruction::Add: case Instruction::Sub: return UndefValue::get(C1->getType()); - case Instruction::Mul: case Instruction::And: + if (isa(C1) && isa(C2)) // undef & undef -> undef + return C1; + return Constant::getNullValue(C1->getType()); // undef & X -> 0 + case Instruction::Mul: { + ConstantInt *CI; + // X * undef -> undef if X is odd or undef + if (((CI = dyn_cast(C1)) && CI->getValue()[0]) || + ((CI = dyn_cast(C2)) && CI->getValue()[0]) || + (isa(C1) && isa(C2))) + return UndefValue::get(C1->getType()); + + // X * undef -> 0 otherwise return Constant::getNullValue(C1->getType()); + } case Instruction::UDiv: case Instruction::SDiv: + // undef / 1 -> undef + if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv) + if (ConstantInt *CI2 = dyn_cast(C2)) + if (CI2->isOne()) + return C1; + // FALL THROUGH case Instruction::URem: case Instruction::SRem: if (!isa(C2)) // undef / X -> 0 return Constant::getNullValue(C1->getType()); return C2; // X / undef -> undef case Instruction::Or: // X | undef -> -1 - if (const VectorType *PTy = dyn_cast(C1->getType())) - return Constant::getAllOnesValue(PTy); - return Constant::getAllOnesValue(C1->getType()); + if (isa(C1) && isa(C2)) // undef | undef -> undef + return C1; + return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 case Instruction::LShr: if (isa(C2) && isa(C1)) return C1; // undef lshr undef -> undef return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 // undef lshr X -> 0 case Instruction::AShr: - if (!isa(C2)) - return C1; // undef ashr X --> undef + if (!isa(C2)) // undef ashr X --> all ones + return Constant::getAllOnesValue(C1->getType()); else if (isa(C1)) return C1; // undef ashr undef -> undef else return C1; // X ashr undef --> X case Instruction::Shl: + if (isa(C2) && isa(C1)) + return C1; // undef shl undef -> undef // undef << X -> 0 or X << undef -> 0 return Constant::getNullValue(C1->getType()); } @@ -1138,7 +1032,6 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, // At this point we know neither constant is an UndefValue. if (ConstantInt *CI1 = dyn_cast(C1)) { if (ConstantInt *CI2 = dyn_cast(C2)) { - using namespace APIntOps; const APInt &C1V = CI1->getValue(); const APInt &C2V = CI2->getValue(); switch (Opcode) { @@ -1234,146 +1127,19 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, return ConstantFP::get(C1->getContext(), C3V); } } - } else if (const VectorType *VTy = dyn_cast(C1->getType())) { - ConstantVector *CP1 = dyn_cast(C1); - ConstantVector *CP2 = dyn_cast(C2); - if ((CP1 != NULL || isa(C1)) && - (CP2 != NULL || isa(C2))) { - std::vector Res; - const Type* EltTy = VTy->getElementType(); - Constant *C1 = 0; - Constant *C2 = 0; - switch (Opcode) { - default: - break; - case Instruction::Add: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getAdd(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::FAdd: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getFAdd(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::Sub: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getSub(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::FSub: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getFSub(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::Mul: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getMul(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::FMul: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getFMul(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::UDiv: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getUDiv(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::SDiv: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getSDiv(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::FDiv: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getFDiv(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::URem: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getURem(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::SRem: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getSRem(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::FRem: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getFRem(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::And: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getAnd(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::Or: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getOr(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::Xor: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getXor(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::LShr: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getLShr(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::AShr: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getAShr(C1, C2)); - } - return ConstantVector::get(Res); - case Instruction::Shl: - for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { - C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); - C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); - Res.push_back(ConstantExpr::getShl(C1, C2)); - } - return ConstantVector::get(Res); - } + } else if (VectorType *VTy = dyn_cast(C1->getType())) { + // Perform elementwise folding. + SmallVector Result; + for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { + Constant *LHS = C1->getAggregateElement(i); + Constant *RHS = C2->getAggregateElement(i); + if (LHS == 0 || RHS == 0) break; + + Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); } + + if (Result.size() == VTy->getNumElements()) + return ConstantVector::get(Result); } if (ConstantExpr *CE1 = dyn_cast(C1)) { @@ -1383,8 +1149,7 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, // Given ((a + b) + c), if (b + c) folds to something interesting, return // (a + (b + c)). - if (Instruction::isAssociative(Opcode, C1->getType()) && - CE1->getOpcode() == Opcode) { + if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); if (!isa(T) || cast(T)->getOpcode() != Opcode) return ConstantExpr::get(Opcode, CE1->getOperand(0), T); @@ -1431,16 +1196,16 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, /// isZeroSizedType - This type is zero sized if its an array or structure of /// zero sized types. The only leaf zero sized type is an empty structure. -static bool isMaybeZeroSizedType(const Type *Ty) { - if (Ty->isOpaqueTy()) return true; // Can't say. - if (const StructType *STy = dyn_cast(Ty)) { +static bool isMaybeZeroSizedType(Type *Ty) { + if (StructType *STy = dyn_cast(Ty)) { + if (STy->isOpaque()) return true; // Can't say. // If all of elements have zero size, this does too. for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; return true; - } else if (const ArrayType *ATy = dyn_cast(Ty)) { + } else if (ArrayType *ATy = dyn_cast(Ty)) { return isMaybeZeroSizedType(ATy->getElementType()); } return false; @@ -1453,7 +1218,7 @@ static bool isMaybeZeroSizedType(const Type *Ty) { /// first is less than the second, return -1, if the second is less than the /// first, return 1. If the constants are not integral, return -2. /// -static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) { +static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { if (C1 == C2) return 0; // Ok, we found a different index. If they are not ConstantInt, we can't do @@ -1725,7 +1490,7 @@ static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, // with a single zero index, it must be nonzero. assert(CE1->getNumOperands() == 2 && !CE1->getOperand(1)->isNullValue() && - "Suprising getelementptr!"); + "Surprising getelementptr!"); return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; } else { // If they are different globals, we don't know what the value is, @@ -1802,8 +1567,8 @@ static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, Constant *C1, Constant *C2) { - const Type *ResultTy; - if (const VectorType *VT = dyn_cast(C1->getType())) + Type *ResultTy; + if (VectorType *VT = dyn_cast(C1->getType())) ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), VT->getNumElements()); else @@ -1817,8 +1582,17 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, return Constant::getAllOnesValue(ResultTy); // Handle some degenerate cases first - if (isa(C1) || isa(C2)) - return UndefValue::get(ResultTy); + if (isa(C1) || isa(C2)) { + // For EQ and NE, we can always pick a value for the undef to make the + // predicate pass or fail, so we can return undef. + // Also, if both operands are undef, we can return undef. + if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) || + (isa(C1) && isa(C2))) + return UndefValue::get(ResultTy); + // Otherwise, pick the same value as the non-undef operand, and fold + // it to true or false. + return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); + } // No compile-time operations on this type yet. if (C1->getType()->isPPC_FP128Ty()) @@ -1864,7 +1638,7 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, APInt V1 = cast(C1)->getValue(); APInt V2 = cast(C2)->getValue(); switch (pred) { - default: llvm_unreachable("Invalid ICmp Predicate"); return 0; + default: llvm_unreachable("Invalid ICmp Predicate"); case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); @@ -1881,7 +1655,7 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, APFloat C2V = cast(C2)->getValueAPF(); APFloat::cmpResult R = C1V.compare(C2V); switch (pred) { - default: llvm_unreachable("Invalid FCmp Predicate"); return 0; + default: llvm_unreachable("Invalid FCmp Predicate"); case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); case FCmpInst::FCMP_UNO: @@ -1920,20 +1694,20 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, R==APFloat::cmpEqual); } } else if (C1->getType()->isVectorTy()) { - SmallVector C1Elts, C2Elts; - C1->getVectorElements(C1Elts); - C2->getVectorElements(C2Elts); - if (C1Elts.empty() || C2Elts.empty()) - return 0; - // If we can constant fold the comparison of each element, constant fold // the whole vector comparison. SmallVector ResElts; - for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) { - // Compare the elements, producing an i1 result or constant expr. - ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); + // Compare the elements, producing an i1 result or constant expr. + for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ + Constant *C1E = C1->getAggregateElement(i); + Constant *C2E = C2->getAggregateElement(i); + if (C1E == 0 || C2E == 0) break; + + ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); } - return ConstantVector::get(&ResElts[0], ResElts.size()); + + if (ResElts.size() == C1->getType()->getVectorNumElements()) + return ConstantVector::get(ResElts); } if (C1->getType()->isFloatingPointTy()) { @@ -1981,7 +1755,7 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) Result = 1; break; - case ICmpInst::ICMP_NE: // We know that C1 != C2 + case FCmpInst::FCMP_ONE: // We know that C1 != C2 // We can only partially decide this relation. if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) Result = 0; @@ -2070,7 +1844,7 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, if (ConstantExpr *CE2 = dyn_cast(C2)) { Constant *CE2Op0 = CE2->getOperand(0); if (CE2->getOpcode() == Instruction::BitCast && - CE2->getType()->isVectorTy()==CE2Op0->getType()->isVectorTy()) { + CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); return ConstantExpr::getICmp(pred, Inverse, CE2Op0); } @@ -2078,8 +1852,8 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, // If the left hand side is an extension, try eliminating it. if (ConstantExpr *CE1 = dyn_cast(C1)) { - if (CE1->getOpcode() == Instruction::SExt || - CE1->getOpcode() == Instruction::ZExt) { + if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || + (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ Constant *CE1Op0 = CE1->getOperand(0); Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); if (CE1Inverse == CE1Op0) { @@ -2097,27 +1871,8 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, // If C2 is a constant expr and C1 isn't, flip them around and fold the // other way if possible. // Also, if C1 is null and C2 isn't, flip them around. - switch (pred) { - case ICmpInst::ICMP_EQ: - case ICmpInst::ICMP_NE: - // No change of predicate required. - return ConstantExpr::getICmp(pred, C2, C1); - - case ICmpInst::ICMP_ULT: - case ICmpInst::ICMP_SLT: - case ICmpInst::ICMP_UGT: - case ICmpInst::ICMP_SGT: - case ICmpInst::ICMP_ULE: - case ICmpInst::ICMP_SLE: - case ICmpInst::ICMP_UGE: - case ICmpInst::ICMP_SGE: - // Change the predicate as necessary to swap the operands. - pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); - return ConstantExpr::getICmp(pred, C2, C1); - - default: // These predicates cannot be flopped around. - break; - } + pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); + return ConstantExpr::getICmp(pred, C2, C1); } } return 0; @@ -2125,56 +1880,53 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, /// isInBoundsIndices - Test whether the given sequence of *normalized* indices /// is "inbounds". -static bool isInBoundsIndices(Constant *const *Idxs, size_t NumIdx) { +template +static bool isInBoundsIndices(ArrayRef Idxs) { // No indices means nothing that could be out of bounds. - if (NumIdx == 0) return true; + if (Idxs.empty()) return true; // If the first index is zero, it's in bounds. - if (Idxs[0]->isNullValue()) return true; + if (cast(Idxs[0])->isNullValue()) return true; // If the first index is one and all the rest are zero, it's in bounds, // by the one-past-the-end rule. if (!cast(Idxs[0])->isOne()) return false; - for (unsigned i = 1, e = NumIdx; i != e; ++i) - if (!Idxs[i]->isNullValue()) + for (unsigned i = 1, e = Idxs.size(); i != e; ++i) + if (!cast(Idxs[i])->isNullValue()) return false; return true; } -Constant *llvm::ConstantFoldGetElementPtr(Constant *C, - bool inBounds, - Constant* const *Idxs, - unsigned NumIdx) { - if (NumIdx == 0 || - (NumIdx == 1 && Idxs[0]->isNullValue())) +template +static Constant *ConstantFoldGetElementPtrImpl(Constant *C, + bool inBounds, + ArrayRef Idxs) { + if (Idxs.empty()) return C; + Constant *Idx0 = cast(Idxs[0]); + if ((Idxs.size() == 1 && Idx0->isNullValue())) return C; if (isa(C)) { - const PointerType *Ptr = cast(C->getType()); - const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, - (Value **)Idxs, - (Value **)Idxs+NumIdx); + PointerType *Ptr = cast(C->getType()); + Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); assert(Ty != 0 && "Invalid indices for GEP!"); return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); } - Constant *Idx0 = Idxs[0]; if (C->isNullValue()) { bool isNull = true; - for (unsigned i = 0, e = NumIdx; i != e; ++i) - if (!Idxs[i]->isNullValue()) { + for (unsigned i = 0, e = Idxs.size(); i != e; ++i) + if (!cast(Idxs[i])->isNullValue()) { isNull = false; break; } if (isNull) { - const PointerType *Ptr = cast(C->getType()); - const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, - (Value**)Idxs, - (Value**)Idxs+NumIdx); + PointerType *Ptr = cast(C->getType()); + Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); assert(Ty != 0 && "Invalid indices for GEP!"); - return ConstantPointerNull::get( - PointerType::get(Ty,Ptr->getAddressSpace())); + return ConstantPointerNull::get(PointerType::get(Ty, + Ptr->getAddressSpace())); } } @@ -2184,14 +1936,14 @@ Constant *llvm::ConstantFoldGetElementPtr(Constant *C, // getelementptr instructions into a single instruction. // if (CE->getOpcode() == Instruction::GetElementPtr) { - const Type *LastTy = 0; + Type *LastTy = 0; for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); I != E; ++I) LastTy = *I; - if ((LastTy && LastTy->isArrayTy()) || Idx0->isNullValue()) { + if ((LastTy && isa(LastTy)) || Idx0->isNullValue()) { SmallVector NewIndices; - NewIndices.reserve(NumIdx + CE->getNumOperands()); + NewIndices.reserve(Idxs.size() + CE->getNumOperands()); for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) NewIndices.push_back(CE->getOperand(i)); @@ -2200,9 +1952,9 @@ Constant *llvm::ConstantFoldGetElementPtr(Constant *C, Constant *Combined = CE->getOperand(CE->getNumOperands()-1); // Otherwise it must be an array. if (!Idx0->isNullValue()) { - const Type *IdxTy = Combined->getType(); + Type *IdxTy = Combined->getType(); if (IdxTy != Idx0->getType()) { - const Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); + Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); Combined = ConstantExpr::get(Instruction::Add, C1, C2); @@ -2213,34 +1965,29 @@ Constant *llvm::ConstantFoldGetElementPtr(Constant *C, } NewIndices.push_back(Combined); - NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx); - return (inBounds && cast(CE)->isInBounds()) ? - ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0), - &NewIndices[0], - NewIndices.size()) : - ConstantExpr::getGetElementPtr(CE->getOperand(0), - &NewIndices[0], - NewIndices.size()); + NewIndices.append(Idxs.begin() + 1, Idxs.end()); + return + ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices, + inBounds && + cast(CE)->isInBounds()); } } // Implement folding of: - // int* getelementptr ([2 x int]* bitcast ([3 x int]* %X to [2 x int]*), - // long 0, long 0) - // To: int* getelementptr ([3 x int]* %X, long 0, long 0) + // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), + // i64 0, i64 0) + // To: i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) // - if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) { - if (const PointerType *SPT = + if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { + if (PointerType *SPT = dyn_cast(CE->getOperand(0)->getType())) - if (const ArrayType *SAT = dyn_cast(SPT->getElementType())) - if (const ArrayType *CAT = + if (ArrayType *SAT = dyn_cast(SPT->getElementType())) + if (ArrayType *CAT = dyn_cast(cast(C->getType())->getElementType())) if (CAT->getElementType() == SAT->getElementType()) - return inBounds ? - ConstantExpr::getInBoundsGetElementPtr( - (Constant*)CE->getOperand(0), Idxs, NumIdx) : - ConstantExpr::getGetElementPtr( - (Constant*)CE->getOperand(0), Idxs, NumIdx); + return + ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0), + Idxs, inBounds); } } @@ -2249,24 +1996,24 @@ Constant *llvm::ConstantFoldGetElementPtr(Constant *C, // out into preceding dimensions. bool Unknown = false; SmallVector NewIdxs; - const Type *Ty = C->getType(); - const Type *Prev = 0; - for (unsigned i = 0; i != NumIdx; + Type *Ty = C->getType(); + Type *Prev = 0; + for (unsigned i = 0, e = Idxs.size(); i != e; Prev = Ty, Ty = cast(Ty)->getTypeAtIndex(Idxs[i]), ++i) { if (ConstantInt *CI = dyn_cast(Idxs[i])) { - if (const ArrayType *ATy = dyn_cast(Ty)) + if (ArrayType *ATy = dyn_cast(Ty)) if (ATy->getNumElements() <= INT64_MAX && ATy->getNumElements() != 0 && CI->getSExtValue() >= (int64_t)ATy->getNumElements()) { if (isa(Prev)) { // It's out of range, but we can factor it into the prior // dimension. - NewIdxs.resize(NumIdx); + NewIdxs.resize(Idxs.size()); ConstantInt *Factor = ConstantInt::get(CI->getType(), ATy->getNumElements()); NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); - Constant *PrevIdx = Idxs[i-1]; + Constant *PrevIdx = cast(Idxs[i-1]); Constant *Div = ConstantExpr::getSDiv(CI, Factor); // Before adding, extend both operands to i64 to avoid @@ -2293,19 +2040,28 @@ Constant *llvm::ConstantFoldGetElementPtr(Constant *C, // If we did any factoring, start over with the adjusted indices. if (!NewIdxs.empty()) { - for (unsigned i = 0; i != NumIdx; ++i) - if (!NewIdxs[i]) NewIdxs[i] = Idxs[i]; - return inBounds ? - ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(), - NewIdxs.size()) : - ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size()); + for (unsigned i = 0, e = Idxs.size(); i != e; ++i) + if (!NewIdxs[i]) NewIdxs[i] = cast(Idxs[i]); + return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds); } // If all indices are known integers and normalized, we can do a simple // check for the "inbounds" property. if (!Unknown && !inBounds && - isa(C) && isInBoundsIndices(Idxs, NumIdx)) - return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx); + isa(C) && isInBoundsIndices(Idxs)) + return ConstantExpr::getInBoundsGetElementPtr(C, Idxs); return 0; } + +Constant *llvm::ConstantFoldGetElementPtr(Constant *C, + bool inBounds, + ArrayRef Idxs) { + return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); +} + +Constant *llvm::ConstantFoldGetElementPtr(Constant *C, + bool inBounds, + ArrayRef Idxs) { + return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); +}