X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FVMCore%2FConstantFold.cpp;h=c22c3e975ba1002de3f66d608f3eb8d1bd98bf03;hb=4f8eea82d8967cffa85b9df6c9255717b059009e;hp=da9b6aa9784731844450de21ddda4c44f5245fd0;hpb=d7f2a6cb3fbc012763adb42fd967f6fefbb22a37;p=oota-llvm.git diff --git a/lib/VMCore/ConstantFold.cpp b/lib/VMCore/ConstantFold.cpp index da9b6aa9784..c22c3e975ba 100644 --- a/lib/VMCore/ConstantFold.cpp +++ b/lib/VMCore/ConstantFold.cpp @@ -12,9 +12,8 @@ // ConstantExpr::get* methods to automatically fold constants when possible. // // The current constant folding implementation is implemented in two pieces: the -// template-based folder for simple primitive constants like ConstantInt, and -// the special case hackery that we use to symbolically evaluate expressions -// that use ConstantExprs. +// pieces that don't need TargetData, and the pieces that do. This is to avoid +// a dependence in VMCore on Target. // //===----------------------------------------------------------------------===// @@ -24,6 +23,7 @@ #include "llvm/DerivedTypes.h" #include "llvm/Function.h" #include "llvm/GlobalAlias.h" +#include "llvm/GlobalVariable.h" #include "llvm/LLVMContext.h" #include "llvm/ADT/SmallVector.h" #include "llvm/Support/Compiler.h" @@ -49,7 +49,7 @@ static Constant *BitCastConstantVector(LLVMContext &Context, ConstantVector *CV, unsigned NumElts = DstTy->getNumElements(); if (NumElts != CV->getNumOperands()) return 0; - + // Check to verify that all elements of the input are simple. for (unsigned i = 0; i != NumElts; ++i) { if (!isa(CV->getOperand(i)) && @@ -73,13 +73,13 @@ static Constant *BitCastConstantVector(LLVMContext &Context, ConstantVector *CV, static unsigned foldConstantCastPair( unsigned opc, ///< opcode of the second cast constant expression - const ConstantExpr*Op, ///< the first cast constant expression + ConstantExpr *Op, ///< the first cast constant expression const 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(); @@ -88,7 +88,7 @@ foldConstantCastPair( // Let CastInst::isEliminableCastPair do the heavy lifting. return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, - Type::Int64Ty); + Type::getInt64Ty(DstTy->getContext())); } static Constant *FoldBitCast(LLVMContext &Context, @@ -96,35 +96,37 @@ static Constant *FoldBitCast(LLVMContext &Context, 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)); + Value *Zero = Constant::getNullValue(Type::getInt32Ty(Context)); + IdxList.push_back(Zero); 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)); + IdxList.push_back(Zero); } 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]); + IdxList.push_back(Zero); } else { break; } } - + if (ElTy == DPTy->getElementType()) - return ConstantExpr::getGetElementPtr(V, &IdxList[0], - IdxList.size()); + // This GEP is inbounds because all indices are zero. + return ConstantExpr::getInBoundsGetElementPtr(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)) { @@ -135,7 +137,7 @@ static Constant *FoldBitCast(LLVMContext &Context, // First, check for null. Undef is already handled. if (isa(V)) return Constant::getNullValue(DestTy); - + if (ConstantVector *CV = dyn_cast(V)) return BitCastConstantVector(Context, CV, DestPTy); } @@ -147,14 +149,14 @@ static Constant *FoldBitCast(LLVMContext &Context, return ConstantExpr::getBitCast( ConstantVector::get(&V, 1), 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 (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. @@ -162,14 +164,14 @@ static Constant *FoldBitCast(LLVMContext &Context, if (DestTy->isFloatingPoint()) return ConstantFP::get(Context, APFloat(CI->getValue(), - DestTy != Type::PPC_FP128Ty)); + DestTy != Type::getPPC_FP128Ty(Context))); // Otherwise, can't fold this (vector?) return 0; } // Handle ConstantFP input. - if (const ConstantFP *FP = dyn_cast(V)) + if (ConstantFP *FP = dyn_cast(V)) // FP -> Integral. return ConstantInt::get(Context, FP->getValueAPF().bitcastToAPInt()); @@ -177,8 +179,263 @@ static Constant *FoldBitCast(LLVMContext &Context, } +/// ExtractConstantBytes - V is an integer constant which only has a subset of +/// its bytes used. The bytes used are indicated by ByteStart (which is the +/// first byte used, counting from the least significant byte) and ByteSize, +/// which is the number of bytes used. +/// +/// This function analyzes the specified constant to see if the specified byte +/// range can be returned as a simplified constant. If so, the constant is +/// returned, otherwise null is returned. +/// +static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, + unsigned ByteSize) { + assert(isa(C->getType()) && + (cast(C->getType())->getBitWidth() & 7) == 0 && + "Non-byte sized integer input"); + unsigned CSize = cast(C->getType())->getBitWidth()/8; + assert(ByteSize && "Must be accessing some piece"); + assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); + assert(ByteSize != CSize && "Should not extract everything"); + + // Constant Integers are simple. + if (ConstantInt *CI = dyn_cast(C)) { + APInt V = CI->getValue(); + if (ByteStart) + V = V.lshr(ByteStart*8); + V.trunc(ByteSize*8); + return ConstantInt::get(CI->getContext(), V); + } + + // In the input is a constant expr, we might be able to recursively simplify. + // If not, we definitely can't do anything. + ConstantExpr *CE = dyn_cast(C); + if (CE == 0) return 0; + + switch (CE->getOpcode()) { + default: return 0; + case Instruction::Or: { + Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); + if (RHS == 0) + return 0; + + // X | -1 -> -1. + if (ConstantInt *RHSC = dyn_cast(RHS)) + if (RHSC->isAllOnesValue()) + return RHSC; + + Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); + if (LHS == 0) + return 0; + return ConstantExpr::getOr(LHS, RHS); + } + case Instruction::And: { + Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); + if (RHS == 0) + return 0; + + // X & 0 -> 0. + if (RHS->isNullValue()) + return RHS; + + Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); + if (LHS == 0) + return 0; + return ConstantExpr::getAnd(LHS, RHS); + } + case Instruction::LShr: { + ConstantInt *Amt = dyn_cast(CE->getOperand(1)); + if (Amt == 0) + return 0; + unsigned ShAmt = Amt->getZExtValue(); + // Cannot analyze non-byte shifts. + if ((ShAmt & 7) != 0) + return 0; + ShAmt >>= 3; + + // If the extract is known to be all zeros, return zero. + if (ByteStart >= CSize-ShAmt) + return Constant::getNullValue(IntegerType::get(CE->getContext(), + ByteSize*8)); + // If the extract is known to be fully in the input, extract it. + if (ByteStart+ByteSize+ShAmt <= CSize) + return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); + + // TODO: Handle the 'partially zero' case. + return 0; + } + + case Instruction::Shl: { + ConstantInt *Amt = dyn_cast(CE->getOperand(1)); + if (Amt == 0) + return 0; + unsigned ShAmt = Amt->getZExtValue(); + // Cannot analyze non-byte shifts. + if ((ShAmt & 7) != 0) + return 0; + ShAmt >>= 3; + + // If the extract is known to be all zeros, return zero. + if (ByteStart+ByteSize <= ShAmt) + return Constant::getNullValue(IntegerType::get(CE->getContext(), + ByteSize*8)); + // If the extract is known to be fully in the input, extract it. + if (ByteStart >= ShAmt) + return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); + + // TODO: Handle the 'partially zero' case. + return 0; + } + + case Instruction::ZExt: { + unsigned SrcBitSize = + cast(CE->getOperand(0)->getType())->getBitWidth(); + + // If extracting something that is completely zero, return 0. + if (ByteStart*8 >= SrcBitSize) + return Constant::getNullValue(IntegerType::get(CE->getContext(), + ByteSize*8)); + + // If exactly extracting the input, return it. + if (ByteStart == 0 && ByteSize*8 == SrcBitSize) + return CE->getOperand(0); + + // If extracting something completely in the input, if if the input is a + // multiple of 8 bits, recurse. + if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) + return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); + + // Otherwise, if extracting a subset of the input, which is not multiple of + // 8 bits, do a shift and trunc to get the bits. + if ((ByteStart+ByteSize)*8 < SrcBitSize) { + assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); + Constant *Res = CE->getOperand(0); + if (ByteStart) + Res = ConstantExpr::getLShr(Res, + ConstantInt::get(Res->getType(), ByteStart*8)); + return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), + ByteSize*8)); + } + + // TODO: Handle the 'partially zero' case. + return 0; + } + } +} + +/// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof +/// on Ty, with any known factors factored out. If Folded is false, +/// 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, + bool Folded) { + if (const 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 VectorType *VTy = dyn_cast(Ty)) { + Constant *N = ConstantInt::get(DestTy, VTy->getNumElements()); + Constant *E = getFoldedSizeOf(VTy->getElementType(), DestTy, true); + return ConstantExpr::getNUWMul(E, N); + } + if (const StructType *STy = dyn_cast(Ty)) + if (!STy->isPacked()) { + unsigned NumElems = STy->getNumElements(); + // An empty struct has size zero. + if (NumElems == 0) + return ConstantExpr::getNullValue(DestTy); + // Check for a struct with all members having the same type. + const Type *MemberTy = STy->getElementType(0); + bool AllSame = true; + for (unsigned i = 1; i != NumElems; ++i) + if (MemberTy != STy->getElementType(i)) { + AllSame = false; + break; + } + if (AllSame) { + Constant *N = ConstantInt::get(DestTy, NumElems); + Constant *E = getFoldedSizeOf(MemberTy, DestTy, true); + return ConstantExpr::getNUWMul(E, N); + } + } + + // If there's no interesting folding happening, bail so that we don't create + // a constant that looks like it needs folding but really doesn't. + if (!Folded) + return 0; + + // Base case: Get a regular sizeof expression. + Constant *C = ConstantExpr::getSizeOf(Ty); + C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, + DestTy, false), + C, DestTy); + return C; +} + +/// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof +/// on Ty and FieldNo, with any known factors factored out. If Folded is false, +/// 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, + bool Folded) { + if (const ArrayType *ATy = dyn_cast(Ty)) { + Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, + DestTy, false), + FieldNo, DestTy); + Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); + return ConstantExpr::getNUWMul(E, N); + } + if (const VectorType *VTy = dyn_cast(Ty)) { + Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, + DestTy, false), + FieldNo, DestTy); + Constant *E = getFoldedSizeOf(VTy->getElementType(), DestTy, true); + return ConstantExpr::getNUWMul(E, N); + } + if (const StructType *STy = dyn_cast(Ty)) + if (!STy->isPacked()) { + unsigned NumElems = STy->getNumElements(); + // An empty struct has no members. + if (NumElems == 0) + return 0; + // Check for a struct with all members having the same type. + const Type *MemberTy = STy->getElementType(0); + bool AllSame = true; + for (unsigned i = 1; i != NumElems; ++i) + if (MemberTy != STy->getElementType(i)) { + AllSame = false; + break; + } + if (AllSame) { + Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, + false, + DestTy, + false), + FieldNo, DestTy); + Constant *E = getFoldedSizeOf(MemberTy, DestTy, true); + return ConstantExpr::getNUWMul(E, N); + } + } + + // If there's no interesting folding happening, bail so that we don't create + // a constant that looks like it needs folding but really doesn't. + if (!Folded) + return 0; + + // Base case: Get a regular offsetof expression. + Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); + C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, + DestTy, false), + C, DestTy); + return C; +} + Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, - unsigned opc, const Constant *V, + unsigned opc, Constant *V, const Type *DestTy) { if (isa(V)) { // zext(undef) = 0, because the top bits will be zero. @@ -190,12 +447,12 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, return UndefValue::get(DestTy); } // No compile-time operations on this type yet. - if (V->getType() == Type::PPC_FP128Ty || DestTy == Type::PPC_FP128Ty) + if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty()) return 0; // If the cast operand is a constant expression, there's a few things we can // do to try to simplify it. - if (const ConstantExpr *CE = dyn_cast(V)) { + if (ConstantExpr *CE = dyn_cast(V)) { if (CE->isCast()) { // Try hard to fold cast of cast because they are often eliminable. if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) @@ -218,7 +475,7 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, // 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 (const ConstantVector *CV = dyn_cast(V)) + if (ConstantVector *CV = dyn_cast(V)) if (isa(DestTy) && cast(DestTy)->getNumElements() == CV->getType()->getNumElements()) { @@ -234,15 +491,17 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, // We actually have to do a cast now. Perform the cast according to the // opcode specified. switch (opc) { + default: + llvm_unreachable("Failed to cast constant expression"); case Instruction::FPTrunc: case Instruction::FPExt: - if (const ConstantFP *FPC = dyn_cast(V)) { + if (ConstantFP *FPC = dyn_cast(V)) { bool ignored; 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 : + Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle : + DestTy->isDoubleTy() ? APFloat::IEEEdouble : + DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : + DestTy->isFP128Ty() ? APFloat::IEEEquad : APFloat::Bogus, APFloat::rmNearestTiesToEven, &ignored); return ConstantFP::get(Context, Val); @@ -250,7 +509,7 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, return 0; // Can't fold. case Instruction::FPToUI: case Instruction::FPToSI: - if (const ConstantFP *FPC = dyn_cast(V)) { + if (ConstantFP *FPC = dyn_cast(V)) { const APFloat &V = FPC->getValueAPF(); bool ignored; uint64_t x[2]; @@ -266,12 +525,67 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, return ConstantPointerNull::get(cast(DestTy)); return 0; // Other pointer types cannot be casted case Instruction::PtrToInt: // always treated as unsigned - if (V->isNullValue()) // is it a null pointer value? + // Is it a null pointer value? + if (V->isNullValue()) return ConstantInt::get(DestTy, 0); - return 0; // Other pointer types cannot be casted + // If this is a sizeof-like expression, pull out multiplications by + // known factors to expose them to subsequent folding. If it's an + // alignof-like expression, factor out known factors. + if (ConstantExpr *CE = dyn_cast(V)) + if (CE->getOpcode() == Instruction::GetElementPtr && + CE->getOperand(0)->isNullValue()) { + const Type *Ty = + cast(CE->getOperand(0)->getType())->getElementType(); + if (CE->getNumOperands() == 2) { + // Handle a sizeof-like expression. + Constant *Idx = CE->getOperand(1); + bool isOne = isa(Idx) && cast(Idx)->isOne(); + if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { + Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, + DestTy, false), + Idx, DestTy); + return ConstantExpr::getMul(C, Idx); + } + } else if (CE->getNumOperands() == 3 && + CE->getOperand(1)->isNullValue()) { + // Handle an alignof-like expression. + if (const StructType *STy = dyn_cast(Ty)) + if (!STy->isPacked()) { + ConstantInt *CI = cast(CE->getOperand(2)); + if (CI->isOne() && + STy->getNumElements() == 2 && + STy->getElementType(0)->isInteger(1)) { + // 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(STy->getElementType(1))) { + Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); + C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, + DestTy, + false), + C, DestTy); + return C; + } + // Packed structs always have an alignment of 1. + if (const StructType *InnerSTy = + dyn_cast(STy->getElementType(1))) + if (InnerSTy->isPacked()) + return ConstantInt::get(DestTy, 1); + } + } + // Handle an offsetof-like expression. + if (isa(Ty) || isa(Ty) || isa(Ty)){ + if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), + DestTy, false)) + return C; + } + } + } + // Other pointer types cannot be casted + return 0; case Instruction::UIToFP: case Instruction::SIToFP: - if (const ConstantInt *CI = dyn_cast(V)) { + if (ConstantInt *CI = dyn_cast(V)) { APInt api = CI->getValue(); const uint64_t zero[] = {0, 0}; APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(), @@ -283,7 +597,7 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, } return 0; case Instruction::ZExt: - if (const ConstantInt *CI = dyn_cast(V)) { + if (ConstantInt *CI = dyn_cast(V)) { uint32_t BitWidth = cast(DestTy)->getBitWidth(); APInt Result(CI->getValue()); Result.zext(BitWidth); @@ -291,57 +605,60 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, } return 0; case Instruction::SExt: - if (const ConstantInt *CI = dyn_cast(V)) { + if (ConstantInt *CI = dyn_cast(V)) { uint32_t BitWidth = cast(DestTy)->getBitWidth(); APInt Result(CI->getValue()); Result.sext(BitWidth); return ConstantInt::get(Context, Result); } return 0; - case Instruction::Trunc: - if (const ConstantInt *CI = dyn_cast(V)) { - uint32_t BitWidth = cast(DestTy)->getBitWidth(); + case Instruction::Trunc: { + uint32_t DestBitWidth = cast(DestTy)->getBitWidth(); + if (ConstantInt *CI = dyn_cast(V)) { APInt Result(CI->getValue()); - Result.trunc(BitWidth); + Result.trunc(DestBitWidth); return ConstantInt::get(Context, Result); } + + // The input must be a constantexpr. See if we can simplify this based on + // the bytes we are demanding. Only do this if the source and dest are an + // even multiple of a byte. + if ((DestBitWidth & 7) == 0 && + (cast(V->getType())->getBitWidth() & 7) == 0) + if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) + return Res; + return 0; + } case Instruction::BitCast: - return FoldBitCast(Context, const_cast(V), DestTy); - default: - assert(!"Invalid CE CastInst opcode"); - break; + return FoldBitCast(Context, V, DestTy); } - - llvm_unreachable("Failed to cast constant expression"); - return 0; } Constant *llvm::ConstantFoldSelectInstruction(LLVMContext&, - const Constant *Cond, - const Constant *V1, - const Constant *V2) { - if (const ConstantInt *CB = dyn_cast(Cond)) - return const_cast(CB->getZExtValue() ? V1 : V2); - - if (isa(V1)) return const_cast(V2); - if (isa(V2)) return const_cast(V1); - if (isa(Cond)) return const_cast(V1); - if (V1 == V2) return const_cast(V1); + Constant *Cond, + Constant *V1, Constant *V2) { + if (ConstantInt *CB = dyn_cast(Cond)) + return CB->getZExtValue() ? V1 : V2; + + if (isa(V1)) return V2; + if (isa(V2)) return V1; + if (isa(Cond)) return V1; + if (V1 == V2) return V1; return 0; } Constant *llvm::ConstantFoldExtractElementInstruction(LLVMContext &Context, - const Constant *Val, - const Constant *Idx) { + Constant *Val, + Constant *Idx) { if (isa(Val)) // ee(undef, x) -> undef return UndefValue::get(cast(Val->getType())->getElementType()); if (Val->isNullValue()) // ee(zero, x) -> zero return Constant::getNullValue( cast(Val->getType())->getElementType()); - - if (const ConstantVector *CVal = dyn_cast(Val)) { - if (const ConstantInt *CIdx = dyn_cast(Idx)) { + + 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). @@ -352,17 +669,17 @@ Constant *llvm::ConstantFoldExtractElementInstruction(LLVMContext &Context, } Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, - const Constant *Val, - const Constant *Elt, - const Constant *Idx) { - const ConstantInt *CIdx = dyn_cast(Idx); + Constant *Val, + Constant *Elt, + 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 const_cast(Val); + return Val; // Otherwise break the aggregate undef into multiple undefs and do // the insertion unsigned numOps = @@ -370,9 +687,9 @@ Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, std::vector Ops; Ops.reserve(numOps); for (unsigned i = 0; i < numOps; ++i) { - const Constant *Op = + Constant *Op = (idxVal == i) ? Elt : UndefValue::get(Elt->getType()); - Ops.push_back(const_cast(Op)); + Ops.push_back(Op); } return ConstantVector::get(Ops); } @@ -380,7 +697,7 @@ Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, // Insertion of scalar constant into vector aggregate zero // Optimize away insertion of zero if (Elt->isNullValue()) - return const_cast(Val); + return Val; // Otherwise break the aggregate zero into multiple zeros and do // the insertion unsigned numOps = @@ -388,20 +705,20 @@ Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, std::vector Ops; Ops.reserve(numOps); for (unsigned i = 0; i < numOps; ++i) { - const Constant *Op = + Constant *Op = (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType()); - Ops.push_back(const_cast(Op)); + Ops.push_back(Op); } return ConstantVector::get(Ops); } - if (const ConstantVector *CVal = dyn_cast(Val)) { + 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) { - const Constant *Op = + Constant *Op = (idxVal == i) ? Elt : cast(CVal->getOperand(i)); - Ops.push_back(const_cast(Op)); + Ops.push_back(Op); } return ConstantVector::get(Ops); } @@ -411,11 +728,11 @@ Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef /// return the specified element value. Otherwise return null. -static Constant *GetVectorElement(LLVMContext &Context, const Constant *C, +static Constant *GetVectorElement(LLVMContext &Context, Constant *C, unsigned EltNo) { - if (const ConstantVector *CV = dyn_cast(C)) + if (ConstantVector *CV = dyn_cast(C)) return CV->getOperand(EltNo); - + const Type *EltTy = cast(C->getType())->getElementType(); if (isa(C)) return Constant::getNullValue(EltTy); @@ -425,9 +742,9 @@ static Constant *GetVectorElement(LLVMContext &Context, const Constant *C, } Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, - const Constant *V1, - const Constant *V2, - const Constant *Mask) { + Constant *V1, + Constant *V2, + Constant *Mask) { // Undefined shuffle mask -> undefined value. if (isa(Mask)) return UndefValue::get(V1->getType()); @@ -463,12 +780,12 @@ Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, } Constant *llvm::ConstantFoldExtractValueInstruction(LLVMContext &Context, - const Constant *Agg, + Constant *Agg, const unsigned *Idxs, unsigned NumIdx) { // Base case: no indices, so return the entire value. if (NumIdx == 0) - return const_cast(Agg); + return Agg; if (isa(Agg)) // ev(undef, x) -> undef return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(), @@ -482,93 +799,104 @@ Constant *llvm::ConstantFoldExtractValueInstruction(LLVMContext &Context, Idxs + NumIdx)); // Otherwise recurse. - return ConstantFoldExtractValueInstruction(Context, Agg->getOperand(*Idxs), + if (ConstantStruct *CS = dyn_cast(Agg)) + return ConstantFoldExtractValueInstruction(Context, CS->getOperand(*Idxs), + Idxs+1, NumIdx-1); + + if (ConstantArray *CA = dyn_cast(Agg)) + return ConstantFoldExtractValueInstruction(Context, CA->getOperand(*Idxs), + Idxs+1, NumIdx-1); + ConstantVector *CV = cast(Agg); + return ConstantFoldExtractValueInstruction(Context, CV->getOperand(*Idxs), Idxs+1, NumIdx-1); } Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, - const Constant *Agg, - const Constant *Val, + Constant *Agg, + Constant *Val, const unsigned *Idxs, unsigned NumIdx) { // Base case: no indices, so replace the entire value. if (NumIdx == 0) - return const_cast(Val); + return Val; if (isa(Agg)) { // Insertion of constant into aggregate undef - // Optimize away insertion of undef + // Optimize away insertion of undef. if (isa(Val)) - return const_cast(Agg); + return Agg; + // Otherwise break the aggregate undef into multiple undefs and do - // the insertion + // 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(); + std::vector Ops(numOps); for (unsigned i = 0; i < numOps; ++i) { const Type *MemberTy = AggTy->getTypeAtIndex(i); - const Constant *Op = + Constant *Op = (*Idxs == i) ? ConstantFoldInsertValueInstruction(Context, UndefValue::get(MemberTy), Val, Idxs+1, NumIdx-1) : UndefValue::get(MemberTy); - Ops[i] = const_cast(Op); + Ops[i] = Op; } - if (isa(AggTy)) - return ConstantStruct::get(Context, Ops); - else - return ConstantArray::get(cast(AggTy), Ops); + + if (const StructType* ST = dyn_cast(AggTy)) + return ConstantStruct::get(Context, Ops, ST->isPacked()); + return ConstantArray::get(cast(AggTy), Ops); } + if (isa(Agg)) { // Insertion of constant into aggregate zero - // Optimize away insertion of zero + // Optimize away insertion of zero. if (Val->isNullValue()) - return const_cast(Agg); + return Agg; + // Otherwise break the aggregate zero into multiple zeros and do - // the insertion + // 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(); + std::vector Ops(numOps); for (unsigned i = 0; i < numOps; ++i) { const Type *MemberTy = AggTy->getTypeAtIndex(i); - const Constant *Op = + Constant *Op = (*Idxs == i) ? ConstantFoldInsertValueInstruction(Context, Constant::getNullValue(MemberTy), Val, Idxs+1, NumIdx-1) : Constant::getNullValue(MemberTy); - Ops[i] = const_cast(Op); + Ops[i] = Op; } - if (isa(AggTy)) - return ConstantStruct::get(Context, Ops); - else - return ConstantArray::get(cast(AggTy), Ops); + + if (const StructType* ST = dyn_cast(AggTy)) + return ConstantStruct::get(Context, Ops, ST->isPacked()); + return ConstantArray::get(cast(AggTy), Ops); } + if (isa(Agg) || isa(Agg)) { - // Insertion of constant into aggregate constant + // Insertion of constant into aggregate constant. std::vector Ops(Agg->getNumOperands()); for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { - const Constant *Op = - (*Idxs == i) ? - ConstantFoldInsertValueInstruction(Context, Agg->getOperand(i), - Val, Idxs+1, NumIdx-1) : - Agg->getOperand(i); - Ops[i] = const_cast(Op); + Constant *Op = cast(Agg->getOperand(i)); + if (*Idxs == i) + Op = ConstantFoldInsertValueInstruction(Context, Op, + Val, Idxs+1, NumIdx-1); + Ops[i] = Op; } - Constant *C; - if (isa(Agg->getType())) - C = ConstantStruct::get(Context, Ops); - else - C = ConstantArray::get(cast(Agg->getType()), Ops); - return C; + + if (const StructType* ST = dyn_cast(Agg->getType())) + return ConstantStruct::get(Context, Ops, ST->isPacked()); + return ConstantArray::get(cast(Agg->getType()), Ops); } return 0; @@ -577,13 +905,12 @@ Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, unsigned Opcode, - const Constant *C1, - const Constant *C2) { + Constant *C1, Constant *C2) { // No compile-time operations on this type yet. - if (C1->getType() == Type::PPC_FP128Ty) + if (C1->getType()->isPPC_FP128Ty()) return 0; - // Handle UndefValue up front + // Handle UndefValue up front. if (isa(C1) || isa(C2)) { switch (Opcode) { case Instruction::Xor: @@ -604,23 +931,23 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, case Instruction::SRem: if (!isa(C2)) // undef / X -> 0 return Constant::getNullValue(C1->getType()); - return const_cast(C2); // X / undef -> undef + 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()); case Instruction::LShr: if (isa(C2) && isa(C1)) - return const_cast(C1); // undef lshr undef -> undef + 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 const_cast(C1); // undef ashr X --> undef + return C1; // undef ashr X --> undef else if (isa(C1)) - return const_cast(C1); // undef ashr undef -> undef + return C1; // undef ashr undef -> undef else - return const_cast(C1); // X ashr undef --> X + return C1; // X ashr undef --> X case Instruction::Shl: // undef << X -> 0 or X << undef -> 0 return Constant::getNullValue(C1->getType()); @@ -628,23 +955,23 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, } // Handle simplifications when the RHS is a constant int. - if (const ConstantInt *CI2 = dyn_cast(C2)) { + if (ConstantInt *CI2 = dyn_cast(C2)) { switch (Opcode) { case Instruction::Add: - if (CI2->equalsInt(0)) return const_cast(C1); // X + 0 == X + if (CI2->equalsInt(0)) return C1; // X + 0 == X break; case Instruction::Sub: - if (CI2->equalsInt(0)) return const_cast(C1); // X - 0 == X + if (CI2->equalsInt(0)) return C1; // X - 0 == X break; case Instruction::Mul: - if (CI2->equalsInt(0)) return const_cast(C2); // X * 0 == 0 + if (CI2->equalsInt(0)) return C2; // X * 0 == 0 if (CI2->equalsInt(1)) - return const_cast(C1); // X * 1 == X + return C1; // X * 1 == X break; case Instruction::UDiv: case Instruction::SDiv: if (CI2->equalsInt(1)) - return const_cast(C1); // X / 1 == X + return C1; // X / 1 == X if (CI2->equalsInt(0)) return UndefValue::get(CI2->getType()); // X / 0 == undef break; @@ -656,11 +983,11 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, return UndefValue::get(CI2->getType()); // X % 0 == undef break; case Instruction::And: - if (CI2->isZero()) return const_cast(C2); // X & 0 == 0 + if (CI2->isZero()) return C2; // X & 0 == 0 if (CI2->isAllOnesValue()) - return const_cast(C1); // X & -1 == X - - if (const ConstantExpr *CE1 = dyn_cast(C1)) { + return C1; // X & -1 == X + + if (ConstantExpr *CE1 = dyn_cast(C1)) { // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) if (CE1->getOpcode() == Instruction::ZExt) { unsigned DstWidth = CI2->getType()->getBitWidth(); @@ -668,19 +995,19 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) - return const_cast(C1); + return 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)); @@ -694,26 +1021,39 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, } break; case Instruction::Or: - if (CI2->equalsInt(0)) return const_cast(C1); // X | 0 == X + if (CI2->equalsInt(0)) return C1; // X | 0 == X if (CI2->isAllOnesValue()) - return const_cast(C2); // X | -1 == -1 + return C2; // X | -1 == -1 break; case Instruction::Xor: - if (CI2->equalsInt(0)) return const_cast(C1); // X ^ 0 == X + if (CI2->equalsInt(0)) return C1; // X ^ 0 == X + + if (ConstantExpr *CE1 = dyn_cast(C1)) { + switch (CE1->getOpcode()) { + default: break; + case Instruction::ICmp: + case Instruction::FCmp: + // cmp pred ^ true -> cmp !pred + assert(CI2->equalsInt(1)); + CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); + pred = CmpInst::getInversePredicate(pred); + return ConstantExpr::getCompare(pred, CE1->getOperand(0), + CE1->getOperand(1)); + } + } break; case Instruction::AShr: // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 - if (const ConstantExpr *CE1 = dyn_cast(C1)) + if (ConstantExpr *CE1 = dyn_cast(C1)) if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. - return ConstantExpr::getLShr(const_cast(C1), - const_cast(C2)); + return ConstantExpr::getLShr(C1, C2); break; } } - + // At this point we know neither constant is an UndefValue. - if (const ConstantInt *CI1 = dyn_cast(C1)) { - if (const ConstantInt *CI2 = dyn_cast(C2)) { + 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(); @@ -780,13 +1120,13 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, case Instruction::LShr: case Instruction::AShr: case Instruction::Shl: - if (CI1->equalsInt(0)) return const_cast(C1); + if (CI1->equalsInt(0)) return C1; break; default: break; } - } else if (const ConstantFP *CFP1 = dyn_cast(C1)) { - if (const ConstantFP *CFP2 = dyn_cast(C2)) { + } else if (ConstantFP *CFP1 = dyn_cast(C1)) { + if (ConstantFP *CFP2 = dyn_cast(C2)) { APFloat C1V = CFP1->getValueAPF(); APFloat C2V = CFP2->getValueAPF(); APFloat C3V = C1V; // copy for modification @@ -811,14 +1151,14 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, } } } else if (const VectorType *VTy = dyn_cast(C1->getType())) { - const ConstantVector *CP1 = dyn_cast(C1); - const ConstantVector *CP2 = dyn_cast(C2); + 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(); - const Constant *C1 = 0; - const Constant *C2 = 0; + Constant *C1 = 0; + Constant *C2 = 0; switch (Opcode) { default: break; @@ -826,154 +1166,145 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + 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(const_cast(C1), - const_cast(C2))); + Res.push_back(ConstantExpr::getShl(C1, C2)); } return ConstantVector::get(Res); } } } - if (isa(C1)) { + if (ConstantExpr *CE1 = dyn_cast(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. + + // Given ((a + b) + c), if (b + c) folds to something interesting, return + // (a + (b + c)). + if (Instruction::isAssociative(Opcode, C1->getType()) && + 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); + } } else if (isa(C2)) { // If C2 is a constant expr and C1 isn't, flop them around and fold the // other way if possible. @@ -987,7 +1318,7 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, case Instruction::Xor: // No change of opcode required. return ConstantFoldBinaryInstruction(Context, Opcode, C2, C1); - + case Instruction::Shl: case Instruction::LShr: case Instruction::AShr: @@ -1003,7 +1334,36 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, break; } } - + + // i1 can be simplified in many cases. + if (C1->getType()->isInteger(1)) { + switch (Opcode) { + case Instruction::Add: + case Instruction::Sub: + return ConstantExpr::getXor(C1, C2); + case Instruction::Mul: + return ConstantExpr::getAnd(C1, C2); + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + // We can assume that C2 == 0. If it were one the result would be + // undefined because the shift value is as large as the bitwidth. + return C1; + case Instruction::SDiv: + case Instruction::UDiv: + // We can assume that C2 == 1. If it were zero the result would be + // undefined through division by zero. + return C1; + case Instruction::URem: + case Instruction::SRem: + // We can assume that C2 == 1. If it were zero the result would be + // undefined through division by zero. + return ConstantInt::getFalse(Context); + default: + break; + } + } + // We don't know how to fold this. return 0; } @@ -1043,11 +1403,11 @@ static int IdxCompare(LLVMContext &Context, Constant *C1, Constant *C2, // Ok, we have two differing integer indices. Sign extend them to be the same // type. Long is always big enough, so we use it. - if (C1->getType() != Type::Int64Ty) - C1 = ConstantExpr::getSExt(C1, Type::Int64Ty); + if (!C1->getType()->isInteger(64)) + C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(Context)); - if (C2->getType() != Type::Int64Ty) - C2 = ConstantExpr::getSExt(C2, Type::Int64Ty); + if (!C2->getType()->isInteger(64)) + C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(Context)); if (C1 == C2) return 0; // They are equal @@ -1077,13 +1437,12 @@ static int IdxCompare(LLVMContext &Context, Constant *C1, Constant *C2, /// operand is always the most "complex" of the two. We consider ConstantFP /// to be the simplest, and ConstantExprs to be the most complex. static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, - const Constant *V1, - const Constant *V2) { + Constant *V1, 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) + if (V1->getType()->isPPC_FP128Ty()) return FCmpInst::BAD_FCMP_PREDICATE; // Handle degenerate case quickly @@ -1093,25 +1452,23 @@ static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, if (!isa(V2)) { // We distilled thisUse the standard constant folder for a few cases ConstantInt *R = 0; - Constant *C1 = const_cast(V1); - Constant *C2 = const_cast(V2); R = dyn_cast( - ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, C1, C2)); + ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); if (R && !R->isZero()) return FCmpInst::FCMP_OEQ; R = dyn_cast( - ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, C1, C2)); + ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); if (R && !R->isZero()) return FCmpInst::FCMP_OLT; R = dyn_cast( - ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, C1, C2)); + ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); if (R && !R->isZero()) return FCmpInst::FCMP_OGT; // Nothing more we can do return FCmpInst::BAD_FCMP_PREDICATE; } - + // If the first operand is simple and second is ConstantExpr, swap operands. FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(Context, V2, V1); if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) @@ -1119,7 +1476,7 @@ static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, } else { // Ok, the LHS is known to be a constantexpr. The RHS can be any of a // constantexpr or a simple constant. - const ConstantExpr *CE1 = cast(V1); + ConstantExpr *CE1 = cast(V1); switch (CE1->getOpcode()) { case Instruction::FPTrunc: case Instruction::FPExt: @@ -1149,8 +1506,8 @@ static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, /// GlobalValues, followed by ConstantExpr's (the most complex). /// static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, - const Constant *V1, - const Constant *V2, + Constant *V1, + Constant *V2, bool isSigned) { assert(V1->getType() == V2->getType() && "Cannot compare different types of values!"); @@ -1161,25 +1518,23 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, // We distilled this down to a simple case, use the standard constant // folder. ConstantInt *R = 0; - Constant *C1 = const_cast(V1); - Constant *C2 = const_cast(V2); ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; - R = dyn_cast(ConstantExpr::getICmp(pred, C1, C2)); + R = dyn_cast(ConstantExpr::getICmp(pred, V1, V2)); if (R && !R->isZero()) return pred; pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; - R = dyn_cast(ConstantExpr::getICmp(pred, C1, C2)); + R = dyn_cast(ConstantExpr::getICmp(pred, V1, V2)); if (R && !R->isZero()) return pred; pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; - R = dyn_cast(ConstantExpr::getICmp(pred, C1, C2)); + R = dyn_cast(ConstantExpr::getICmp(pred, V1, V2)); if (R && !R->isZero()) return pred; - + // If we couldn't figure it out, bail. return ICmpInst::BAD_ICMP_PREDICATE; } - + // If the first operand is simple, swap operands. ICmpInst::Predicate SwappedRelation = evaluateICmpRelation(Context, V2, V1, isSigned); @@ -1212,8 +1567,8 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, } else { // Ok, the LHS is known to be a constantexpr. The RHS can be any of a // constantexpr, a CPR, or a simple constant. - const ConstantExpr *CE1 = cast(V1); - const Constant *CE1Op0 = CE1->getOperand(0); + ConstantExpr *CE1 = cast(V1); + Constant *CE1Op0 = CE1->getOperand(0); switch (CE1->getOpcode()) { case Instruction::Trunc: @@ -1232,28 +1587,12 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, // null pointer, do the comparison with the pre-casted value. if (V2->isNullValue() && (isa(CE1->getType()) || CE1->getType()->isInteger())) { - bool sgnd = isSigned; if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; if (CE1->getOpcode() == Instruction::SExt) isSigned = true; return evaluateICmpRelation(Context, CE1Op0, Constant::getNullValue(CE1Op0->getType()), - sgnd); + isSigned); } - - // If the dest type is a pointer type, and the RHS is a constantexpr cast - // from the same type as the src of the LHS, evaluate the inputs. This is - // important for things like "icmp eq (cast 4 to int*), (cast 5 to int*)", - // which happens a lot in compilers with tagged integers. - if (const ConstantExpr *CE2 = dyn_cast(V2)) - if (CE2->isCast() && isa(CE1->getType()) && - CE1->getOperand(0)->getType() == CE2->getOperand(0)->getType() && - CE1->getOperand(0)->getType()->isInteger()) { - bool sgnd = isSigned; - if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; - if (CE1->getOpcode() == Instruction::SExt) isSigned = true; - return evaluateICmpRelation(Context, CE1->getOperand(0), - CE2->getOperand(0), sgnd); - } break; case Instruction::GetElementPtr: @@ -1270,7 +1609,7 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, else // If its not weak linkage, the GVal must have a non-zero address // so the result is greater-than - return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; + return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; } else if (isa(CE1Op0)) { // If we are indexing from a null pointer, check to see if we have any // non-zero indices. @@ -1309,8 +1648,8 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, } } } else { - const ConstantExpr *CE2 = cast(V2); - const Constant *CE2Op0 = CE2->getOperand(0); + ConstantExpr *CE2 = cast(V2); + Constant *CE2Op0 = CE2->getOperand(0); // There are MANY other foldings that we could perform here. They will // probably be added on demand, as they seem needed. @@ -1327,6 +1666,14 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, // ordering of the resultant pointers. unsigned i = 1; + // The logic below assumes that the result of the comparison + // can be determined by finding the first index that differs. + // This doesn't work if there is over-indexing in any + // subsequent indices, so check for that case first. + if (!CE1->isGEPWithNoNotionalOverIndexing() || + !CE2->isGEPWithNoNotionalOverIndexing()) + return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. + // Compare all of the operands the GEP's have in common. gep_type_iterator GTI = gep_type_begin(CE1); for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); @@ -1369,13 +1716,12 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, unsigned short pred, - const Constant *C1, - const Constant *C2) { + Constant *C1, Constant *C2) { const Type *ResultTy; if (const VectorType *VT = dyn_cast(C1->getType())) - ResultTy = VectorType::get(Type::Int1Ty, VT->getNumElements()); + ResultTy = VectorType::get(Type::getInt1Ty(Context), VT->getNumElements()); else - ResultTy = Type::Int1Ty; + ResultTy = Type::getInt1Ty(Context); // Fold FCMP_FALSE/FCMP_TRUE unconditionally. if (pred == FCmpInst::FCMP_FALSE) @@ -1389,7 +1735,7 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, return UndefValue::get(ResultTy); // No compile-time operations on this type yet. - if (C1->getType() == Type::PPC_FP128Ty) + if (C1->getType()->isPPC_FP128Ty()) return 0; // icmp eq/ne(null,GV) -> false/true @@ -1414,31 +1760,45 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, } } + // If the comparison is a comparison between two i1's, simplify it. + if (C1->getType()->isInteger(1)) { + switch(pred) { + case ICmpInst::ICMP_EQ: + if (isa(C2)) + return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); + return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); + case ICmpInst::ICMP_NE: + return ConstantExpr::getXor(C1, C2); + default: + break; + } + } + if (isa(C1) && isa(C2)) { APInt V1 = cast(C1)->getValue(); APInt V2 = cast(C2)->getValue(); switch (pred) { default: llvm_unreachable("Invalid ICmp Predicate"); return 0; case ICmpInst::ICMP_EQ: - return ConstantInt::get(Type::Int1Ty, V1 == V2); + return ConstantInt::get(Type::getInt1Ty(Context), V1 == V2); case ICmpInst::ICMP_NE: - return ConstantInt::get(Type::Int1Ty, V1 != V2); + return ConstantInt::get(Type::getInt1Ty(Context), V1 != V2); case ICmpInst::ICMP_SLT: - return ConstantInt::get(Type::Int1Ty, V1.slt(V2)); + return ConstantInt::get(Type::getInt1Ty(Context), V1.slt(V2)); case ICmpInst::ICMP_SGT: - return ConstantInt::get(Type::Int1Ty, V1.sgt(V2)); + return ConstantInt::get(Type::getInt1Ty(Context), V1.sgt(V2)); case ICmpInst::ICMP_SLE: - return ConstantInt::get(Type::Int1Ty, V1.sle(V2)); + return ConstantInt::get(Type::getInt1Ty(Context), V1.sle(V2)); case ICmpInst::ICMP_SGE: - return ConstantInt::get(Type::Int1Ty, V1.sge(V2)); + return ConstantInt::get(Type::getInt1Ty(Context), V1.sge(V2)); case ICmpInst::ICMP_ULT: - return ConstantInt::get(Type::Int1Ty, V1.ult(V2)); + return ConstantInt::get(Type::getInt1Ty(Context), V1.ult(V2)); case ICmpInst::ICMP_UGT: - return ConstantInt::get(Type::Int1Ty, V1.ugt(V2)); + return ConstantInt::get(Type::getInt1Ty(Context), V1.ugt(V2)); case ICmpInst::ICMP_ULE: - return ConstantInt::get(Type::Int1Ty, V1.ule(V2)); + return ConstantInt::get(Type::getInt1Ty(Context), V1.ule(V2)); case ICmpInst::ICMP_UGE: - return ConstantInt::get(Type::Int1Ty, V1.uge(V2)); + return ConstantInt::get(Type::getInt1Ty(Context), V1.uge(V2)); } } else if (isa(C1) && isa(C2)) { APFloat C1V = cast(C1)->getValueAPF(); @@ -1449,52 +1809,53 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, case FCmpInst::FCMP_FALSE: return ConstantInt::getFalse(Context); case FCmpInst::FCMP_TRUE: return ConstantInt::getTrue(Context); case FCmpInst::FCMP_UNO: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered); + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered); case FCmpInst::FCMP_ORD: - return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpUnordered); + return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpUnordered); case FCmpInst::FCMP_UEQ: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered || + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || R==APFloat::cmpEqual); case FCmpInst::FCMP_OEQ: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpEqual); + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpEqual); case FCmpInst::FCMP_UNE: - return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpEqual); + return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpEqual); case FCmpInst::FCMP_ONE: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan || + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan || R==APFloat::cmpGreaterThan); case FCmpInst::FCMP_ULT: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered || + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || R==APFloat::cmpLessThan); case FCmpInst::FCMP_OLT: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan); + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan); case FCmpInst::FCMP_UGT: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered || + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || R==APFloat::cmpGreaterThan); case FCmpInst::FCMP_OGT: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan); + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpGreaterThan); case FCmpInst::FCMP_ULE: - return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpGreaterThan); + return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpGreaterThan); case FCmpInst::FCMP_OLE: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan || + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan || R==APFloat::cmpEqual); case FCmpInst::FCMP_UGE: - return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpLessThan); + return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpLessThan); case FCmpInst::FCMP_OGE: - return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan || + return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpGreaterThan || R==APFloat::cmpEqual); } } else if (isa(C1->getType())) { SmallVector C1Elts, C2Elts; C1->getVectorElements(Context, C1Elts); C2->getVectorElements(Context, 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])); + ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); } return ConstantVector::get(&ResElts[0], ResElts.size()); } @@ -1552,10 +1913,10 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, Result = 1; break; } - + // If we evaluated the result, return it now. if (Result != -1) - return ConstantInt::get(Type::Int1Ty, Result); + return ConstantInt::get(Type::getInt1Ty(Context), Result); } else { // Evaluate the relation between the two constants, per the predicate. @@ -1567,81 +1928,102 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, case ICmpInst::ICMP_EQ: // We know the constants are equal! // If we know the constants are equal, we can decide the result of this // computation precisely. - Result = (pred == ICmpInst::ICMP_EQ || - pred == ICmpInst::ICMP_ULE || - pred == ICmpInst::ICMP_SLE || - pred == ICmpInst::ICMP_UGE || - pred == ICmpInst::ICMP_SGE); + Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); break; case ICmpInst::ICMP_ULT: - // If we know that C1 < C2, we can decide the result of this computation - // precisely. - Result = (pred == ICmpInst::ICMP_ULT || - pred == ICmpInst::ICMP_NE || - pred == ICmpInst::ICMP_ULE); + switch (pred) { + case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: + Result = 1; break; + case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: + Result = 0; break; + } break; case ICmpInst::ICMP_SLT: - // If we know that C1 < C2, we can decide the result of this computation - // precisely. - Result = (pred == ICmpInst::ICMP_SLT || - pred == ICmpInst::ICMP_NE || - pred == ICmpInst::ICMP_SLE); + switch (pred) { + case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: + Result = 1; break; + case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: + Result = 0; break; + } break; case ICmpInst::ICMP_UGT: - // If we know that C1 > C2, we can decide the result of this computation - // precisely. - Result = (pred == ICmpInst::ICMP_UGT || - pred == ICmpInst::ICMP_NE || - pred == ICmpInst::ICMP_UGE); + switch (pred) { + case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: + Result = 1; break; + case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: + Result = 0; break; + } break; case ICmpInst::ICMP_SGT: - // If we know that C1 > C2, we can decide the result of this computation - // precisely. - Result = (pred == ICmpInst::ICMP_SGT || - pred == ICmpInst::ICMP_NE || - pred == ICmpInst::ICMP_SGE); + switch (pred) { + case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: + Result = 1; break; + case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: + Result = 0; break; + } break; case ICmpInst::ICMP_ULE: - // If we know that C1 <= C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_UGT) Result = 0; - if (pred == ICmpInst::ICMP_ULT) Result = 1; + if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; break; case ICmpInst::ICMP_SLE: - // If we know that C1 <= C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_SGT) Result = 0; - if (pred == ICmpInst::ICMP_SLT) Result = 1; + if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; break; - case ICmpInst::ICMP_UGE: - // If we know that C1 >= C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_ULT) Result = 0; - if (pred == ICmpInst::ICMP_UGT) Result = 1; + if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; break; case ICmpInst::ICMP_SGE: - // If we know that C1 >= C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_SLT) Result = 0; - if (pred == ICmpInst::ICMP_SGT) Result = 1; + if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; break; - case ICmpInst::ICMP_NE: - // If we know that C1 != C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_EQ) Result = 0; if (pred == ICmpInst::ICMP_NE) Result = 1; break; } - + // If we evaluated the result, return it now. if (Result != -1) - return ConstantInt::get(Type::Int1Ty, Result); - - if (!isa(C1) && isa(C2)) { + return ConstantInt::get(Type::getInt1Ty(Context), Result); + + // If the right hand side is a bitcast, try using its inverse to simplify + // it by moving it to the left hand side. + if (ConstantExpr *CE2 = dyn_cast(C2)) { + if (CE2->getOpcode() == Instruction::BitCast) { + Constant *CE2Op0 = CE2->getOperand(0); + Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); + return ConstantExpr::getICmp(pred, Inverse, CE2Op0); + } + } + + // 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) { + Constant *CE1Op0 = CE1->getOperand(0); + Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); + if (CE1Inverse == CE1Op0) { + // Check whether we can safely truncate the right hand side. + Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); + if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) { + return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); + } + } + } + } + + if ((!isa(C1) && isa(C2)) || + (C1->isNullValue() && !C2->isNullValue())) { // 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 ConstantFoldCompareInstruction(Context, pred, C2, C1); + return ConstantExpr::getICmp(pred, C2, C1); case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_SLT: @@ -1653,7 +2035,7 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, case ICmpInst::ICMP_SGE: // Change the predicate as necessary to swap the operands. pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); - return ConstantFoldCompareInstruction(Context, pred, C2, C1); + return ConstantExpr::getICmp(pred, C2, C1); default: // These predicates cannot be flopped around. break; @@ -1661,15 +2043,35 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, } } return 0; - } +} + +/// isInBoundsIndices - Test whether the given sequence of *normalized* indices +/// is "inbounds". +static bool isInBoundsIndices(Constant *const *Idxs, size_t NumIdx) { + // No indices means nothing that could be out of bounds. + if (NumIdx == 0) return true; + + // If the first index is zero, it's in bounds. + if (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()) + return false; + return true; +} Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, - const Constant *C, + Constant *C, + bool inBounds, Constant* const *Idxs, unsigned NumIdx) { if (NumIdx == 0 || (NumIdx == 1 && Idxs[0]->isNullValue())) - return const_cast(C); + return C; if (isa(C)) { const PointerType *Ptr = cast(C->getType()); @@ -1699,7 +2101,7 @@ Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, } } - if (ConstantExpr *CE = dyn_cast(const_cast(C))) { + if (ConstantExpr *CE = dyn_cast(C)) { // Combine Indices - If the source pointer to this getelementptr instruction // is a getelementptr instruction, combine the indices of the two // getelementptr instructions into a single instruction. @@ -1724,9 +2126,9 @@ Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, const Type *IdxTy = Combined->getType(); if (IdxTy != Idx0->getType()) { Constant *C1 = - ConstantExpr::getSExtOrBitCast(Idx0, Type::Int64Ty); + ConstantExpr::getSExtOrBitCast(Idx0, Type::getInt64Ty(Context)); Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, - Type::Int64Ty); + Type::getInt64Ty(Context)); Combined = ConstantExpr::get(Instruction::Add, C1, C2); } else { Combined = @@ -1736,14 +2138,18 @@ Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, NewIndices.push_back(Combined); NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx); - return ConstantExpr::getGetElementPtr(CE->getOperand(0), - &NewIndices[0], - NewIndices.size()); + return (inBounds && cast(CE)->isInBounds()) ? + ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0), + &NewIndices[0], + NewIndices.size()) : + ConstantExpr::getGetElementPtr(CE->getOperand(0), + &NewIndices[0], + NewIndices.size()); } } // Implement folding of: - // int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*), + // 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) // @@ -1754,31 +2160,76 @@ Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, if (const ArrayType *CAT = dyn_cast(cast(C->getType())->getElementType())) if (CAT->getElementType() == SAT->getElementType()) - return ConstantExpr::getGetElementPtr( + return inBounds ? + ConstantExpr::getInBoundsGetElementPtr( + (Constant*)CE->getOperand(0), Idxs, NumIdx) : + ConstantExpr::getGetElementPtr( (Constant*)CE->getOperand(0), Idxs, NumIdx); } - - // Fold: getelementptr (i8* inttoptr (i64 1 to i8*), i32 -1) - // Into: inttoptr (i64 0 to i8*) - // This happens with pointers to member functions in C++. - if (CE->getOpcode() == Instruction::IntToPtr && NumIdx == 1 && - isa(CE->getOperand(0)) && isa(Idxs[0]) && - cast(CE->getType())->getElementType() == Type::Int8Ty) { - Constant *Base = CE->getOperand(0); - Constant *Offset = Idxs[0]; - - // Convert the smaller integer to the larger type. - if (Offset->getType()->getPrimitiveSizeInBits() < - Base->getType()->getPrimitiveSizeInBits()) - Offset = ConstantExpr::getSExt(Offset, Base->getType()); - else if (Base->getType()->getPrimitiveSizeInBits() < - Offset->getType()->getPrimitiveSizeInBits()) - Base = ConstantExpr::getZExt(Base, Offset->getType()); - - Base = ConstantExpr::getAdd(Base, Offset); - return ConstantExpr::getIntToPtr(Base, CE->getType()); + } + + // Check to see if any array indices are not within the corresponding + // notional array bounds. If so, try to determine if they can be factored + // out into preceding dimensions. + bool Unknown = false; + SmallVector NewIdxs; + const Type *Ty = C->getType(); + const Type *Prev = 0; + for (unsigned i = 0; i != NumIdx; + Prev = Ty, Ty = cast(Ty)->getTypeAtIndex(Idxs[i]), ++i) { + if (ConstantInt *CI = dyn_cast(Idxs[i])) { + if (const 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); + ConstantInt *Factor = ConstantInt::get(CI->getType(), + ATy->getNumElements()); + NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); + + Constant *PrevIdx = Idxs[i-1]; + Constant *Div = ConstantExpr::getSDiv(CI, Factor); + + // Before adding, extend both operands to i64 to avoid + // overflow trouble. + if (!PrevIdx->getType()->isInteger(64)) + PrevIdx = ConstantExpr::getSExt(PrevIdx, + Type::getInt64Ty(Context)); + if (!Div->getType()->isInteger(64)) + Div = ConstantExpr::getSExt(Div, + Type::getInt64Ty(Context)); + + NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); + } else { + // It's out of range, but the prior dimension is a struct + // so we can't do anything about it. + Unknown = true; + } + } + } else { + // We don't know if it's in range or not. + Unknown = true; } } + + // 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()); + } + + // 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); + return 0; } -