X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FVMCore%2FConstantFold.cpp;h=5138031da9186494fef609420be860e0d75eef1d;hb=efe65369a74871c3140a540a6c95ce5d1f080954;hp=4dc1340f908cc0ffc9c40b23471e81e4543fa666;hpb=a0ef5ed742a81b134ac4438c3a4adc0c9a151b64;p=oota-llvm.git diff --git a/lib/VMCore/ConstantFold.cpp b/lib/VMCore/ConstantFold.cpp index 4dc1340f908..5138031da91 100644 --- a/lib/VMCore/ConstantFold.cpp +++ b/lib/VMCore/ConstantFold.cpp @@ -2,8 +2,8 @@ // // The LLVM Compiler Infrastructure // -// This file was developed by the LLVM research group and is distributed under -// the University of Illinois Open Source License. See LICENSE.TXT for details. +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // @@ -36,82 +36,31 @@ using namespace llvm; // ConstantFold*Instruction Implementations //===----------------------------------------------------------------------===// -/// CastConstantVector - Convert the specified ConstantVector node to the +/// BitCastConstantVector - Convert the specified ConstantVector 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 *CastConstantVector(ConstantVector *CV, - const VectorType *DstTy) { - unsigned SrcNumElts = CV->getType()->getNumElements(); - unsigned DstNumElts = DstTy->getNumElements(); - const Type *SrcEltTy = CV->getType()->getElementType(); - const Type *DstEltTy = DstTy->getElementType(); +static Constant *BitCastConstantVector(ConstantVector *CV, + const VectorType *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()) + return 0; - // If both vectors have the same number of elements (thus, the elements - // are the same size), perform the conversion now. - if (SrcNumElts == DstNumElts) { - std::vector Result; - - // If the src and dest elements are both integers, or both floats, we can - // just BitCast each element because the elements are the same size. - if ((SrcEltTy->isInteger() && DstEltTy->isInteger()) || - (SrcEltTy->isFloatingPoint() && DstEltTy->isFloatingPoint())) { - for (unsigned i = 0; i != SrcNumElts; ++i) - Result.push_back( - ConstantExpr::getBitCast(CV->getOperand(i), DstEltTy)); - return ConstantVector::get(Result); - } - - // If this is an int-to-fp cast .. - if (SrcEltTy->isInteger()) { - // Ensure that it is int-to-fp cast - assert(DstEltTy->isFloatingPoint()); - if (DstEltTy->getTypeID() == Type::DoubleTyID) { - for (unsigned i = 0; i != SrcNumElts; ++i) { - ConstantInt *CI = cast(CV->getOperand(i)); - double V = CI->getValue().bitsToDouble(); - Result.push_back(ConstantFP::get(Type::DoubleTy, APFloat(V))); - } - return ConstantVector::get(Result); - } - assert(DstEltTy == Type::FloatTy && "Unknown fp type!"); - for (unsigned i = 0; i != SrcNumElts; ++i) { - ConstantInt *CI = cast(CV->getOperand(i)); - float V = CI->getValue().bitsToFloat(); - Result.push_back(ConstantFP::get(Type::FloatTy, APFloat(V))); - } - return ConstantVector::get(Result); - } - - // Otherwise, this is an fp-to-int cast. - assert(SrcEltTy->isFloatingPoint() && DstEltTy->isInteger()); - - if (SrcEltTy->getTypeID() == Type::DoubleTyID) { - for (unsigned i = 0; i != SrcNumElts; ++i) { - uint64_t V = - DoubleToBits(cast(CV->getOperand(i))-> - getValueAPF().convertToDouble()); - Constant *C = ConstantInt::get(Type::Int64Ty, V); - Result.push_back(ConstantExpr::getBitCast(C, DstEltTy )); - } - return ConstantVector::get(Result); - } - - assert(SrcEltTy->getTypeID() == Type::FloatTyID); - for (unsigned i = 0; i != SrcNumElts; ++i) { - uint32_t V = FloatToBits(cast(CV->getOperand(i))-> - getValueAPF().convertToFloat()); - Constant *C = ConstantInt::get(Type::Int32Ty, V); - Result.push_back(ConstantExpr::getBitCast(C, DstEltTy)); - } - return ConstantVector::get(Result); + // Check to verify that all elements of the input are simple. + for (unsigned i = 0; i != NumElts; ++i) { + if (!isa(CV->getOperand(i)) && + !isa(CV->getOperand(i))) + return 0; } - - // Otherwise, this is a cast that changes element count and size. Handle - // casts which shrink the elements here. - - // FIXME: We need to know endianness to do this! - - return 0; + + // 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); } /// This function determines which opcode to use to fold two constant cast @@ -139,17 +88,102 @@ foldConstantCastPair( Type::Int64Ty); } -Constant *llvm::ConstantFoldCastInstruction(unsigned opc, const Constant *V, - const Type *DestTy) { +static Constant *FoldBitCast(Constant *V, const Type *DestTy) { const 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()) { + SmallVector IdxList; + IdxList.push_back(Constant::getNullValue(Type::Int32Ty)); + const Type *ElTy = PTy->getElementType(); + while (ElTy != DPTy->getElementType()) { + if (const StructType *STy = dyn_cast(ElTy)) { + if (STy->getNumElements() == 0) break; + ElTy = STy->getElementType(0); + IdxList.push_back(Constant::getNullValue(Type::Int32Ty)); + } else if (const SequentialType *STy = + dyn_cast(ElTy)) { + if (isa(ElTy)) break; // Can't index into pointers! + ElTy = STy->getElementType(); + IdxList.push_back(IdxList[0]); + } else { + break; + } + } + + if (ElTy == DPTy->getElementType()) + return ConstantExpr::getGetElementPtr(V, &IdxList[0], IdxList.size()); + } + + // 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())) { + assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && + "Not cast between same sized vectors!"); + // First, check for null. Undef is already handled. + if (isa(V)) + return Constant::getNullValue(DestTy); + + if (ConstantVector *CV = dyn_cast(V)) + return BitCastConstantVector(CV, DestPTy); + } + } + + // Finally, implement bitcast folding now. The code below doesn't handle + // bitcast right. + if (isa(V)) // ptr->ptr cast. + return ConstantPointerNull::get(cast(DestTy)); + + // Handle integral constant input. + if (const ConstantInt *CI = dyn_cast(V)) { + if (DestTy->isInteger()) + // Integral -> Integral. This is a no-op because the bit widths must + // be the same. Consequently, we just fold to V. + return V; + + if (DestTy->isFloatingPoint()) { + assert((DestTy == Type::DoubleTy || DestTy == Type::FloatTy) && + "Unknown FP type!"); + return ConstantFP::get(APFloat(CI->getValue())); + } + // Otherwise, can't fold this (vector?) + return 0; + } + + // Handle ConstantFP input. + if (const ConstantFP *FP = dyn_cast(V)) { + // FP -> Integral. + if (DestTy == Type::Int32Ty) { + return ConstantInt::get(FP->getValueAPF().convertToAPInt()); + } else { + assert(DestTy == Type::Int64Ty && "only support f32/f64 for now!"); + return ConstantInt::get(FP->getValueAPF().convertToAPInt()); + } + } + return 0; +} + +Constant *llvm::ConstantFoldCastInstruction(unsigned opc, const Constant *V, + const 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. - if (opc == Instruction::ZExt || opc == Instruction::SExt) + // [us]itofp(undef) = 0, because the result value is bounded. + if (opc == Instruction::ZExt || opc == Instruction::SExt || + opc == Instruction::UIToFP || opc == Instruction::SIToFP) return Constant::getNullValue(DestTy); return UndefValue::get(DestTy); } + // No compile-time operations on this type yet. + if (V->getType() == Type::PPC_FP128Ty || DestTy == Type::PPC_FP128Ty) + return 0; // If the cast operand is a constant expression, there's a few things we can // do to try to simplify it. @@ -179,32 +213,36 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, const Constant *V, case Instruction::FPTrunc: case Instruction::FPExt: if (const ConstantFP *FPC = dyn_cast(V)) { - APFloat Val = FPC->getValueAPF(); - Val.convert(DestTy==Type::FloatTy ? APFloat::IEEEsingle : - APFloat::IEEEdouble, + APFloat Val = FPC->getValueAPF(); + Val.convert(DestTy == Type::FloatTy ? APFloat::IEEEsingle : + DestTy == Type::DoubleTy ? APFloat::IEEEdouble : + DestTy == Type::X86_FP80Ty ? APFloat::x87DoubleExtended : + DestTy == Type::FP128Ty ? APFloat::IEEEquad : + APFloat::Bogus, APFloat::rmNearestTiesToEven); - return ConstantFP::get(DestTy, Val); + return ConstantFP::get(Val); } return 0; // Can't fold. case Instruction::FPToUI: - if (const ConstantFP *FPC = dyn_cast(V)) { - APFloat V = FPC->getValueAPF(); - bool isDouble = &V.getSemantics()==&APFloat::IEEEdouble; - uint32_t DestBitWidth = cast(DestTy)->getBitWidth(); - APInt Val(APIntOps::RoundDoubleToAPInt(isDouble ? V.convertToDouble() : - (double)V.convertToFloat(), DestBitWidth)); - return ConstantInt::get(Val); - } - return 0; // Can't fold. case Instruction::FPToSI: if (const ConstantFP *FPC = dyn_cast(V)) { - APFloat V = FPC->getValueAPF(); - bool isDouble = &V.getSemantics()==&APFloat::IEEEdouble; + const APFloat &V = FPC->getValueAPF(); + uint64_t x[2]; uint32_t DestBitWidth = cast(DestTy)->getBitWidth(); - APInt Val(APIntOps::RoundDoubleToAPInt(isDouble ? V.convertToDouble() : - (double)V.convertToFloat(), DestBitWidth)); + (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, + APFloat::rmTowardZero); + APInt Val(DestBitWidth, 2, x); return ConstantInt::get(Val); } + if (const ConstantVector *CV = dyn_cast(V)) { + 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(ConstantFoldCastInstruction(opc, V->getOperand(i), + DstEltTy)); + return ConstantVector::get(DestVecTy, res); + } return 0; // Can't fold. case Instruction::IntToPtr: //always treated as unsigned if (V->isNullValue()) // Is it an integral null value? @@ -215,21 +253,25 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, const Constant *V, return ConstantInt::get(DestTy, 0); return 0; // Other pointer types cannot be casted case Instruction::UIToFP: - if (const ConstantInt *CI = dyn_cast(V)) { - if (DestTy==Type::FloatTy) - return ConstantFP::get(DestTy, - APFloat((float)CI->getValue().roundToDouble())); - else - return ConstantFP::get(DestTy, APFloat(CI->getValue().roundToDouble())); - } - return 0; case Instruction::SIToFP: if (const ConstantInt *CI = dyn_cast(V)) { - double d = CI->getValue().signedRoundToDouble(); - if (DestTy==Type::FloatTy) - return ConstantFP::get(DestTy, APFloat((float)d)); - else - return ConstantFP::get(DestTy, APFloat(d)); + APInt api = CI->getValue(); + const uint64_t zero[] = {0, 0}; + APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(), + 2, zero)); + (void)apf.convertFromAPInt(api, + opc==Instruction::SIToFP, + APFloat::rmNearestTiesToEven); + return ConstantFP::get(apf); + } + if (const ConstantVector *CV = dyn_cast(V)) { + 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(ConstantFoldCastInstruction(opc, V->getOperand(i), + DstEltTy)); + return ConstantVector::get(DestVecTy, res); } return 0; case Instruction::ZExt: @@ -257,105 +299,7 @@ Constant *llvm::ConstantFoldCastInstruction(unsigned opc, const Constant *V, } return 0; case Instruction::BitCast: - if (SrcTy == DestTy) - return (Constant*)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)) { - SmallVector IdxList; - IdxList.push_back(Constant::getNullValue(Type::Int32Ty)); - const Type *ElTy = PTy->getElementType(); - while (ElTy != DPTy->getElementType()) { - if (const StructType *STy = dyn_cast(ElTy)) { - if (STy->getNumElements() == 0) break; - ElTy = STy->getElementType(0); - IdxList.push_back(Constant::getNullValue(Type::Int32Ty)); - } else if (const SequentialType *STy = - dyn_cast(ElTy)) { - if (isa(ElTy)) break; // Can't index into pointers! - ElTy = STy->getElementType(); - IdxList.push_back(IdxList[0]); - } else { - break; - } - } - - if (ElTy == DPTy->getElementType()) - return ConstantExpr::getGetElementPtr( - const_cast(V), &IdxList[0], IdxList.size()); - } - - // 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())) { - assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && - "Not cast between same sized vectors!"); - // First, check for null and undef - if (isa(V)) - return Constant::getNullValue(DestTy); - if (isa(V)) - return UndefValue::get(DestTy); - - if (const ConstantVector *CV = dyn_cast(V)) { - // This is a cast from a ConstantVector of one type to a - // ConstantVector of another type. Check to see if all elements of - // the input are simple. - bool AllSimpleConstants = true; - for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { - if (!isa(CV->getOperand(i)) && - !isa(CV->getOperand(i))) { - AllSimpleConstants = false; - break; - } - } - - // If all of the elements are simple constants, we can fold this. - if (AllSimpleConstants) - return CastConstantVector(const_cast(CV), DestPTy); - } - } - } - - // Finally, implement bitcast folding now. The code below doesn't handle - // bitcast right. - if (isa(V)) // ptr->ptr cast. - return ConstantPointerNull::get(cast(DestTy)); - - // Handle integral constant input. - if (const ConstantInt *CI = dyn_cast(V)) { - if (DestTy->isInteger()) - // Integral -> Integral. This is a no-op because the bit widths must - // be the same. Consequently, we just fold to V. - return const_cast(V); - - if (DestTy->isFloatingPoint()) { - if (DestTy == Type::FloatTy) - return ConstantFP::get(DestTy, APFloat(CI->getValue().bitsToFloat())); - assert(DestTy == Type::DoubleTy && "Unknown FP type!"); - return ConstantFP::get(DestTy, APFloat(CI->getValue().bitsToDouble())); - } - // Otherwise, can't fold this (vector?) - return 0; - } - - // Handle ConstantFP input. - if (const ConstantFP *FP = dyn_cast(V)) { - // FP -> Integral. - if (DestTy == Type::Int32Ty) { - APInt Val(32, 0); - return ConstantInt::get(Val.floatToBits(FP-> - getValueAPF().convertToFloat())); - } else { - assert(DestTy == Type::Int64Ty && "only support f32/f64 for now!"); - APInt Val(64, 0); - return ConstantInt::get(Val.doubleToBits(FP-> - getValueAPF().convertToDouble())); - } - } - return 0; + return FoldBitCast(const_cast(V), DestTy); default: assert(!"Invalid CE CastInst opcode"); break; @@ -388,10 +332,10 @@ Constant *llvm::ConstantFoldExtractElementInstruction(const Constant *Val, if (const ConstantVector *CVal = dyn_cast(Val)) { if (const ConstantInt *CIdx = dyn_cast(Idx)) { - return const_cast(CVal->getOperand(CIdx->getZExtValue())); + return CVal->getOperand(CIdx->getZExtValue()); } else if (isa(Idx)) { // ee({w,x,y,z}, undef) -> w (an arbitrary value). - return const_cast(CVal->getOperand(0)); + return CVal->getOperand(0); } } return 0; @@ -453,35 +397,93 @@ Constant *llvm::ConstantFoldInsertElementInstruction(const Constant *Val, return 0; } +/// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef +/// return the specified element value. Otherwise return null. +static Constant *GetVectorElement(const Constant *C, unsigned EltNo) { + if (const 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; +} + Constant *llvm::ConstantFoldShuffleVectorInstruction(const Constant *V1, const Constant *V2, const Constant *Mask) { - // TODO: - return 0; + // Undefined shuffle mask -> undefined value. + if (isa(Mask)) return UndefValue::get(V1->getType()); + + unsigned NumElts = cast(V1->getType())->getNumElements(); + const Type *EltTy = cast(V1->getType())->getElementType(); + + // Loop over the shuffle mask, evaluating each element. + SmallVector Result; + for (unsigned i = 0; i != NumElts; ++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 >= NumElts*2) + InElt = UndefValue::get(EltTy); + else if (Elt >= NumElts) + InElt = GetVectorElement(V2, Elt-NumElts); + else + InElt = GetVectorElement(V1, Elt); + if (InElt == 0) return 0; + } else { + // Unknown value. + return 0; + } + Result.push_back(InElt); + } + + return ConstantVector::get(&Result[0], Result.size()); } /// EvalVectorOp - Given two vector constants and a function pointer, apply the /// function pointer to each element pair, producing a new ConstantVector -/// constant. +/// constant. Either or both of V1 and V2 may be NULL, meaning a +/// ConstantAggregateZero operand. static Constant *EvalVectorOp(const ConstantVector *V1, const ConstantVector *V2, + const VectorType *VTy, Constant *(*FP)(Constant*, Constant*)) { std::vector Res; - for (unsigned i = 0, e = V1->getNumOperands(); i != e; ++i) - Res.push_back(FP(const_cast(V1->getOperand(i)), - const_cast(V2->getOperand(i)))); + const Type *EltTy = VTy->getElementType(); + for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { + const Constant *C1 = V1 ? V1->getOperand(i) : Constant::getNullValue(EltTy); + const Constant *C2 = V2 ? V2->getOperand(i) : Constant::getNullValue(EltTy); + Res.push_back(FP(const_cast(C1), + const_cast(C2))); + } return ConstantVector::get(Res); } Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, const Constant *C1, const Constant *C2) { + // No compile-time operations on this type yet. + if (C1->getType() == Type::PPC_FP128Ty) + return 0; + // Handle UndefValue up front if (isa(C1) || isa(C2)) { switch (Opcode) { + case Instruction::Xor: + if (isa(C1) && isa(C2)) + // Handle undef ^ undef -> 0 special case. This is a common + // idiom (misuse). + return Constant::getNullValue(C1->getType()); + // Fallthrough case Instruction::Add: case Instruction::Sub: - case Instruction::Xor: return UndefValue::get(C1->getType()); case Instruction::Mul: case Instruction::And: @@ -517,115 +519,92 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, } } - if (const ConstantExpr *CE1 = dyn_cast(C1)) { - if (isa(C2)) { - // There are many possible foldings we could do here. We should probably - // at least fold add of a pointer with an integer into the appropriate - // getelementptr. This will improve alias analysis a bit. - } else { - // Just implement a couple of simple identities. - switch (Opcode) { - case Instruction::Add: - if (C2->isNullValue()) return const_cast(C1); // X + 0 == X - break; - case Instruction::Sub: - if (C2->isNullValue()) return const_cast(C1); // X - 0 == X - break; - case Instruction::Mul: - if (C2->isNullValue()) return const_cast(C2); // X * 0 == 0 - if (const ConstantInt *CI = dyn_cast(C2)) - if (CI->equalsInt(1)) - return const_cast(C1); // X * 1 == X - break; - case Instruction::UDiv: - case Instruction::SDiv: - if (const ConstantInt *CI = dyn_cast(C2)) - if (CI->equalsInt(1)) - return const_cast(C1); // X / 1 == X - break; - case Instruction::URem: - case Instruction::SRem: - if (const ConstantInt *CI = dyn_cast(C2)) - if (CI->equalsInt(1)) - return Constant::getNullValue(CI->getType()); // X % 1 == 0 - break; - case Instruction::And: - if (const ConstantInt *CI = dyn_cast(C2)) { - if (CI->isZero()) return const_cast(C2); // X & 0 == 0 - if (CI->isAllOnesValue()) - return const_cast(C1); // X & -1 == X - - // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) - if (CE1->getOpcode() == Instruction::ZExt) { - APInt PossiblySetBits - = cast(CE1->getOperand(0)->getType())->getMask(); - PossiblySetBits.zext(C1->getType()->getPrimitiveSizeInBits()); - if ((PossiblySetBits & CI->getValue()) == PossiblySetBits) - return const_cast(C1); - } - } - if (CE1->isCast() && isa(CE1->getOperand(0))) { - GlobalValue *CPR = cast(CE1->getOperand(0)); - - // Functions are at least 4-byte aligned. If and'ing the address of a - // function with a constant < 4, fold it to zero. - if (const ConstantInt *CI = dyn_cast(C2)) - if (CI->getValue().ult(APInt(CI->getType()->getBitWidth(),4)) && - isa(CPR)) - return Constant::getNullValue(CI->getType()); - } - break; - case Instruction::Or: - if (C2->isNullValue()) return const_cast(C1); // X | 0 == X - if (const ConstantInt *CI = dyn_cast(C2)) - if (CI->isAllOnesValue()) - return const_cast(C2); // X | -1 == -1 - break; - case Instruction::Xor: - if (C2->isNullValue()) return const_cast(C1); // X ^ 0 == X - break; - case Instruction::AShr: - // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 - if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. - return ConstantExpr::getLShr(const_cast(C1), - const_cast(C2)); - break; - } - } - } else if (isa(C2)) { - // If C2 is a constant expr and C1 isn't, flop them around and fold the - // other way if possible. + // Handle simplifications of the RHS when a constant int. + if (const ConstantInt *CI2 = dyn_cast(C2)) { switch (Opcode) { case Instruction::Add: + if (CI2->equalsInt(0)) return const_cast(C1); // X + 0 == X + break; + case Instruction::Sub: + if (CI2->equalsInt(0)) return const_cast(C1); // X - 0 == X + break; case Instruction::Mul: + if (CI2->equalsInt(0)) return const_cast(C2); // X * 0 == 0 + if (CI2->equalsInt(1)) + return const_cast(C1); // X * 1 == X + break; + case Instruction::UDiv: + case Instruction::SDiv: + if (CI2->equalsInt(1)) + return const_cast(C1); // X / 1 == X + break; + case Instruction::URem: + case Instruction::SRem: + if (CI2->equalsInt(1)) + return Constant::getNullValue(CI2->getType()); // X % 1 == 0 + break; case Instruction::And: + if (CI2->isZero()) return const_cast(C2); // X & 0 == 0 + if (CI2->isAllOnesValue()) + return const_cast(C1); // X & -1 == X + + if (const ConstantExpr *CE1 = dyn_cast(C1)) { + // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) + if (CE1->getOpcode() == Instruction::ZExt) { + unsigned DstWidth = CI2->getType()->getBitWidth(); + unsigned SrcWidth = + CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); + APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); + if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) + return const_cast(C1); + } + + // If and'ing the address of a global with a constant, fold it. + if (CE1->getOpcode() == Instruction::PtrToInt && + isa(CE1->getOperand(0))) { + GlobalValue *GV = cast(CE1->getOperand(0)); + + // Functions are at least 4-byte aligned. + unsigned GVAlign = GV->getAlignment(); + if (isa(GV)) + GVAlign = std::max(GVAlign, 4U); + + if (GVAlign > 1) { + unsigned DstWidth = CI2->getType()->getBitWidth(); + unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); + APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); + + // If checking bits we know are clear, return zero. + if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) + return Constant::getNullValue(CI2->getType()); + } + } + } + break; case Instruction::Or: + if (CI2->equalsInt(0)) return const_cast(C1); // X | 0 == X + if (CI2->isAllOnesValue()) + return const_cast(C2); // X | -1 == -1 + break; case Instruction::Xor: - // No change of opcode required. - return ConstantFoldBinaryInstruction(Opcode, C2, C1); - - case Instruction::Shl: - case Instruction::LShr: + if (CI2->equalsInt(0)) return const_cast(C1); // X ^ 0 == X + break; case Instruction::AShr: - case Instruction::Sub: - case Instruction::SDiv: - case Instruction::UDiv: - case Instruction::FDiv: - case Instruction::URem: - case Instruction::SRem: - case Instruction::FRem: - default: // These instructions cannot be flopped around. - return 0; + // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 + if (const ConstantExpr *CE1 = dyn_cast(C1)) + if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. + return ConstantExpr::getLShr(const_cast(C1), + const_cast(C2)); + break; } } - - // At this point we know neither constant is an UndefValue nor a ConstantExpr - // so look at directly computing the value. + + // At this point we know neither constant is an UndefValue. if (const ConstantInt *CI1 = dyn_cast(C1)) { if (const ConstantInt *CI2 = dyn_cast(C2)) { using namespace APIntOps; - APInt C1V = CI1->getValue(); - APInt C2V = CI2->getValue(); + const APInt &C1V = CI1->getValue(); + const APInt &C2V = CI2->getValue(); switch (Opcode) { default: break; @@ -661,27 +640,27 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, return ConstantInt::get(C1V | C2V); case Instruction::Xor: return ConstantInt::get(C1V ^ C2V); - case Instruction::Shl: - if (uint32_t shiftAmt = C2V.getZExtValue()) - if (shiftAmt < C1V.getBitWidth()) - return ConstantInt::get(C1V.shl(shiftAmt)); - else - return UndefValue::get(C1->getType()); // too big shift is undef - return const_cast(CI1); // Zero shift is identity - case Instruction::LShr: - if (uint32_t shiftAmt = C2V.getZExtValue()) - if (shiftAmt < C1V.getBitWidth()) - return ConstantInt::get(C1V.lshr(shiftAmt)); - else - return UndefValue::get(C1->getType()); // too big shift is undef - return const_cast(CI1); // Zero shift is identity - case Instruction::AShr: - if (uint32_t shiftAmt = C2V.getZExtValue()) - if (shiftAmt < C1V.getBitWidth()) - return ConstantInt::get(C1V.ashr(shiftAmt)); - else - return UndefValue::get(C1->getType()); // too big shift is undef - return const_cast(CI1); // Zero shift is identity + case Instruction::Shl: { + uint32_t shiftAmt = C2V.getZExtValue(); + if (shiftAmt < C1V.getBitWidth()) + return ConstantInt::get(C1V.shl(shiftAmt)); + else + return UndefValue::get(C1->getType()); // too big shift is undef + } + case Instruction::LShr: { + uint32_t shiftAmt = C2V.getZExtValue(); + if (shiftAmt < C1V.getBitWidth()) + return ConstantInt::get(C1V.lshr(shiftAmt)); + else + return UndefValue::get(C1->getType()); // too big shift is undef + } + case Instruction::AShr: { + uint32_t shiftAmt = C2V.getZExtValue(); + if (shiftAmt < C1V.getBitWidth()) + return ConstantInt::get(C1V.ashr(shiftAmt)); + else + return UndefValue::get(C1->getType()); // too big shift is undef + } } } } else if (const ConstantFP *CFP1 = dyn_cast(C1)) { @@ -689,83 +668,104 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, APFloat C1V = CFP1->getValueAPF(); APFloat C2V = CFP2->getValueAPF(); APFloat C3V = C1V; // copy for modification - bool isDouble = CFP1->getType()==Type::DoubleTy; switch (Opcode) { default: break; case Instruction::Add: (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(CFP1->getType(), C3V); + return ConstantFP::get(C3V); case Instruction::Sub: (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(CFP1->getType(), C3V); + return ConstantFP::get(C3V); case Instruction::Mul: (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(CFP1->getType(), C3V); + return ConstantFP::get(C3V); case Instruction::FDiv: - // FIXME better to look at the return code - if (C2V.isZero()) - if (C1V.isZero()) - // IEEE 754, Section 7.1, #4 - return ConstantFP::get(CFP1->getType(), isDouble ? - APFloat(std::numeric_limits::quiet_NaN()) : - APFloat(std::numeric_limits::quiet_NaN())); - else if (C2V.isNegZero() || C1V.isNegative()) - // IEEE 754, Section 7.2, negative infinity case - return ConstantFP::get(CFP1->getType(), isDouble ? - APFloat(-std::numeric_limits::infinity()) : - APFloat(-std::numeric_limits::infinity())); - else - // IEEE 754, Section 7.2, positive infinity case - return ConstantFP::get(CFP1->getType(), isDouble ? - APFloat(std::numeric_limits::infinity()) : - APFloat(std::numeric_limits::infinity())); (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(CFP1->getType(), C3V); + return ConstantFP::get(C3V); case Instruction::FRem: - if (C2V.isZero()) + if (C2V.isZero()) { // IEEE 754, Section 7.1, #5 - return ConstantFP::get(CFP1->getType(), isDouble ? - APFloat(std::numeric_limits::quiet_NaN()) : - APFloat(std::numeric_limits::quiet_NaN())); + if (CFP1->getType() == Type::DoubleTy) + return ConstantFP::get(APFloat(std::numeric_limits:: + quiet_NaN())); + if (CFP1->getType() == Type::FloatTy) + return ConstantFP::get(APFloat(std::numeric_limits:: + quiet_NaN())); + break; + } (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(CFP1->getType(), C3V); + return ConstantFP::get(C3V); } } - } else if (const ConstantVector *CP1 = dyn_cast(C1)) { - if (const ConstantVector *CP2 = dyn_cast(C2)) { + } else if (const VectorType *VTy = dyn_cast(C1->getType())) { + const ConstantVector *CP1 = dyn_cast(C1); + const ConstantVector *CP2 = dyn_cast(C2); + if ((CP1 != NULL || isa(C1)) && + (CP2 != NULL || isa(C2))) { switch (Opcode) { - default: - break; - case Instruction::Add: - return EvalVectorOp(CP1, CP2, ConstantExpr::getAdd); - case Instruction::Sub: - return EvalVectorOp(CP1, CP2, ConstantExpr::getSub); - case Instruction::Mul: - return EvalVectorOp(CP1, CP2, ConstantExpr::getMul); - case Instruction::UDiv: - return EvalVectorOp(CP1, CP2, ConstantExpr::getUDiv); - case Instruction::SDiv: - return EvalVectorOp(CP1, CP2, ConstantExpr::getSDiv); - case Instruction::FDiv: - return EvalVectorOp(CP1, CP2, ConstantExpr::getFDiv); - case Instruction::URem: - return EvalVectorOp(CP1, CP2, ConstantExpr::getURem); - case Instruction::SRem: - return EvalVectorOp(CP1, CP2, ConstantExpr::getSRem); - case Instruction::FRem: - return EvalVectorOp(CP1, CP2, ConstantExpr::getFRem); - case Instruction::And: - return EvalVectorOp(CP1, CP2, ConstantExpr::getAnd); - case Instruction::Or: - return EvalVectorOp(CP1, CP2, ConstantExpr::getOr); - case Instruction::Xor: - return EvalVectorOp(CP1, CP2, ConstantExpr::getXor); + default: + break; + case Instruction::Add: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getAdd); + case Instruction::Sub: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSub); + case Instruction::Mul: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getMul); + case Instruction::UDiv: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getUDiv); + case Instruction::SDiv: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSDiv); + case Instruction::FDiv: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFDiv); + case Instruction::URem: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getURem); + case Instruction::SRem: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSRem); + case Instruction::FRem: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFRem); + case Instruction::And: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getAnd); + case Instruction::Or: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getOr); + case Instruction::Xor: + return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getXor); } } } - // We don't know how to fold this + if (isa(C1)) { + // There are many possible foldings we could do here. We should probably + // at least fold add of a pointer with an integer into the appropriate + // getelementptr. This will improve alias analysis a bit. + } else if (isa(C2)) { + // If C2 is a constant expr and C1 isn't, flop them around and fold the + // other way if possible. + switch (Opcode) { + case Instruction::Add: + case Instruction::Mul: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + // No change of opcode required. + return ConstantFoldBinaryInstruction(Opcode, C2, C1); + + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + case Instruction::Sub: + case Instruction::SDiv: + case Instruction::UDiv: + case Instruction::FDiv: + case Instruction::URem: + case Instruction::SRem: + case Instruction::FRem: + default: // These instructions cannot be flopped around. + break; + } + } + + // We don't know how to fold this. return 0; } @@ -840,6 +840,11 @@ static FCmpInst::Predicate evaluateFCmpRelation(const Constant *V1, const Constant *V2) { assert(V1->getType() == V2->getType() && "Cannot compare values of different types!"); + + // No compile-time operations on this type yet. + if (V1->getType() == Type::PPC_FP128Ty) + return FCmpInst::BAD_FCMP_PREDICATE; + // Handle degenerate case quickly if (V1 == V2) return FCmpInst::FCMP_OEQ; @@ -978,20 +983,19 @@ static ICmpInst::Predicate evaluateICmpRelation(const Constant *V1, case Instruction::UIToFP: case Instruction::SIToFP: - case Instruction::IntToPtr: case Instruction::BitCast: case Instruction::ZExt: case Instruction::SExt: - case Instruction::PtrToInt: // If the cast is not actually changing bits, and the second operand is a // null pointer, do the comparison with the pre-casted value. if (V2->isNullValue() && (isa(CE1->getType()) || CE1->getType()->isInteger())) { - bool sgnd = CE1->getOpcode() == Instruction::ZExt ? false : - (CE1->getOpcode() == Instruction::SExt ? true : - (CE1->getOpcode() == Instruction::PtrToInt ? false : isSigned)); - return evaluateICmpRelation( - CE1Op0, Constant::getNullValue(CE1Op0->getType()), sgnd); + bool sgnd = isSigned; + if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; + if (CE1->getOpcode() == Instruction::SExt) isSigned = true; + return evaluateICmpRelation(CE1Op0, + Constant::getNullValue(CE1Op0->getType()), + sgnd); } // If the dest type is a pointer type, and the RHS is a constantexpr cast @@ -1002,11 +1006,11 @@ static ICmpInst::Predicate evaluateICmpRelation(const Constant *V1, if (CE2->isCast() && isa(CE1->getType()) && CE1->getOperand(0)->getType() == CE2->getOperand(0)->getType() && CE1->getOperand(0)->getType()->isInteger()) { - bool sgnd = CE1->getOpcode() == Instruction::ZExt ? false : - (CE1->getOpcode() == Instruction::SExt ? true : - (CE1->getOpcode() == Instruction::PtrToInt ? false : isSigned)); + bool sgnd = isSigned; + if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; + if (CE1->getOpcode() == Instruction::SExt) isSigned = true; return evaluateICmpRelation(CE1->getOperand(0), CE2->getOperand(0), - sgnd); + sgnd); } break; @@ -1095,18 +1099,20 @@ static ICmpInst::Predicate evaluateICmpRelation(const Constant *V1, // Ok, we ran out of things they have in common. If any leftovers // are non-zero then we have a difference, otherwise we are equal. for (; i < CE1->getNumOperands(); ++i) - if (!CE1->getOperand(i)->isNullValue()) + if (!CE1->getOperand(i)->isNullValue()) { if (isa(CE1->getOperand(i))) return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; else return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. + } for (; i < CE2->getNumOperands(); ++i) - if (!CE2->getOperand(i)->isNullValue()) + if (!CE2->getOperand(i)->isNullValue()) { if (isa(CE2->getOperand(i))) return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; else return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. + } return ICmpInst::ICMP_EQ; } } @@ -1127,22 +1133,30 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, if (isa(C1) || isa(C2)) return UndefValue::get(Type::Int1Ty); + // No compile-time operations on this type yet. + if (C1->getType() == Type::PPC_FP128Ty) + return 0; + // icmp eq/ne(null,GV) -> false/true if (C1->isNullValue()) { if (const GlobalValue *GV = dyn_cast(C2)) - if (!GV->hasExternalWeakLinkage()) // External weak GV can be null + // Don't try to evaluate aliases. External weak GV can be null. + if (!isa(GV) && !GV->hasExternalWeakLinkage()) { if (pred == ICmpInst::ICMP_EQ) return ConstantInt::getFalse(); else if (pred == ICmpInst::ICMP_NE) return ConstantInt::getTrue(); + } // icmp eq/ne(GV,null) -> false/true } else if (C2->isNullValue()) { if (const GlobalValue *GV = dyn_cast(C1)) - if (!GV->hasExternalWeakLinkage()) // External weak GV can be null + // Don't try to evaluate aliases. External weak GV can be null. + if (!isa(GV) && !GV->hasExternalWeakLinkage()) { if (pred == ICmpInst::ICMP_EQ) return ConstantInt::getFalse(); else if (pred == ICmpInst::ICMP_NE) return ConstantInt::getTrue(); + } } if (isa(C1) && isa(C2)) { @@ -1208,9 +1222,9 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, if (const ConstantVector *CP2 = dyn_cast(C2)) { if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) { for (unsigned i = 0, e = CP1->getNumOperands(); i != e; ++i) { - Constant *C= ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, - const_cast(CP1->getOperand(i)), - const_cast(CP2->getOperand(i))); + Constant *C = ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, + CP1->getOperand(i), + CP2->getOperand(i)); if (ConstantInt *CB = dyn_cast(C)) return CB; } @@ -1219,8 +1233,8 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, } else if (pred == ICmpInst::ICMP_EQ) { for (unsigned i = 0, e = CP1->getNumOperands(); i != e; ++i) { Constant *C = ConstantExpr::getICmp(ICmpInst::ICMP_EQ, - const_cast(CP1->getOperand(i)), - const_cast(CP2->getOperand(i))); + CP1->getOperand(i), + CP2->getOperand(i)); if (ConstantInt *CB = dyn_cast(C)) return CB; } @@ -1391,12 +1405,13 @@ Constant *llvm::ConstantFoldGetElementPtr(const Constant *C, return const_cast(C); if (isa(C)) { - const Type *Ty = GetElementPtrInst::getIndexedType(C->getType(), + const PointerType *Ptr = cast(C->getType()); + const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, (Value **)Idxs, (Value **)Idxs+NumIdx, true); assert(Ty != 0 && "Invalid indices for GEP!"); - return UndefValue::get(PointerType::get(Ty)); + return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); } Constant *Idx0 = Idxs[0]; @@ -1408,12 +1423,14 @@ Constant *llvm::ConstantFoldGetElementPtr(const Constant *C, break; } if (isNull) { - const Type *Ty = GetElementPtrInst::getIndexedType(C->getType(), + const PointerType *Ptr = cast(C->getType()); + const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, (Value**)Idxs, (Value**)Idxs+NumIdx, true); assert(Ty != 0 && "Invalid indices for GEP!"); - return ConstantPointerNull::get(PointerType::get(Ty)); + return + ConstantPointerNull::get(PointerType::get(Ty,Ptr->getAddressSpace())); } }