X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FVMCore%2FConstantFold.cpp;h=47c693b0306fa7c2b7560425643dc495d64cac65;hb=a9203109f4ac95aa7e9624f2838e3d89623ec902;hp=2c0a67f1d0435da860085a245830c643adc49e12;hpb=59c4eba4169e40b36719c521c25d29115ff6b206;p=oota-llvm.git diff --git a/lib/VMCore/ConstantFold.cpp b/lib/VMCore/ConstantFold.cpp index 2c0a67f1d04..47c693b0306 100644 --- a/lib/VMCore/ConstantFold.cpp +++ b/lib/VMCore/ConstantFold.cpp @@ -24,7 +24,7 @@ #include "llvm/Function.h" #include "llvm/GlobalAlias.h" #include "llvm/GlobalVariable.h" -#include "llvm/LLVMContext.h" +#include "llvm/Operator.h" #include "llvm/ADT/SmallVector.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/ErrorHandling.h" @@ -41,8 +41,12 @@ using namespace llvm; /// 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 *BitCastConstantVector(LLVMContext &Context, ConstantVector *CV, - const VectorType *DstTy) { +static Constant *BitCastConstantVector(ConstantVector *CV, + VectorType *DstTy) { + + if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); + if (CV->isNullValue()) return Constant::getNullValue(DstTy); + // If this cast changes element count then we can't handle it here: // doing so requires endianness information. This should be handled by // Analysis/ConstantFolding.cpp @@ -59,7 +63,7 @@ static Constant *BitCastConstantVector(LLVMContext &Context, ConstantVector *CV, // Bitcast each element now. std::vector Result; - const Type *DstEltTy = DstTy->getElementType(); + Type *DstEltTy = DstTy->getElementType(); for (unsigned i = 0; i != NumElts; ++i) Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i), DstEltTy)); @@ -74,15 +78,15 @@ static unsigned foldConstantCastPair( unsigned opc, ///< opcode of the second cast constant expression ConstantExpr *Op, ///< the first cast constant expression - const Type *DstTy ///< desintation type of the first cast + Type *DstTy ///< desintation type of the first cast ) { assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); assert(CastInst::isCast(opc) && "Invalid cast opcode"); // The the types and opcodes for the two Cast constant expressions - const Type *SrcTy = Op->getOperand(0)->getType(); - const Type *MidTy = Op->getType(); + Type *SrcTy = Op->getOperand(0)->getType(); + Type *MidTy = Op->getType(); Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); Instruction::CastOps secondOp = Instruction::CastOps(opc); @@ -91,29 +95,29 @@ foldConstantCastPair( Type::getInt64Ty(DstTy->getContext())); } -static Constant *FoldBitCast(LLVMContext &Context, - Constant *V, const Type *DestTy) { - const Type *SrcTy = V->getType(); +static Constant *FoldBitCast(Constant *V, Type *DestTy) { + Type *SrcTy = V->getType(); if (SrcTy == DestTy) return V; // no-op cast // Check to see if we are casting a pointer to an aggregate to a pointer to // the first element. If so, return the appropriate GEP instruction. - if (const PointerType *PTy = dyn_cast(V->getType())) - if (const PointerType *DPTy = dyn_cast(DestTy)) + if (PointerType *PTy = dyn_cast(V->getType())) + if (PointerType *DPTy = dyn_cast(DestTy)) if (PTy->getAddressSpace() == DPTy->getAddressSpace()) { SmallVector IdxList; - Value *Zero = Constant::getNullValue(Type::getInt32Ty(Context)); + Value *Zero = + Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); IdxList.push_back(Zero); - const Type *ElTy = PTy->getElementType(); + Type *ElTy = PTy->getElementType(); while (ElTy != DPTy->getElementType()) { - if (const StructType *STy = dyn_cast(ElTy)) { + if (StructType *STy = dyn_cast(ElTy)) { if (STy->getNumElements() == 0) break; ElTy = STy->getElementType(0); IdxList.push_back(Zero); - } else if (const SequentialType *STy = + } else if (SequentialType *STy = dyn_cast(ElTy)) { - if (isa(ElTy)) break; // Can't index into pointers! + if (ElTy->isPointerTy()) break; // Can't index into pointers! ElTy = STy->getElementType(); IdxList.push_back(Zero); } else { @@ -123,14 +127,13 @@ static Constant *FoldBitCast(LLVMContext &Context, if (ElTy == DPTy->getElementType()) // This GEP is inbounds because all indices are zero. - return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0], - IdxList.size()); + return ConstantExpr::getInBoundsGetElementPtr(V, IdxList); } // Handle casts from one vector constant to another. We know that the src // and dest type have the same size (otherwise its an illegal cast). - if (const VectorType *DestPTy = dyn_cast(DestTy)) { - if (const VectorType *SrcTy = dyn_cast(V->getType())) { + if (VectorType *DestPTy = dyn_cast(DestTy)) { + if (VectorType *SrcTy = dyn_cast(V->getType())) { assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && "Not cast between same sized vectors!"); SrcTy = NULL; @@ -139,15 +142,14 @@ static Constant *FoldBitCast(LLVMContext &Context, return Constant::getNullValue(DestTy); if (ConstantVector *CV = dyn_cast(V)) - return BitCastConstantVector(Context, CV, DestPTy); + return BitCastConstantVector(CV, DestPTy); } // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts // This allows for other simplifications (although some of them // can only be handled by Analysis/ConstantFolding.cpp). if (isa(V) || isa(V)) - return ConstantExpr::getBitCast( - ConstantVector::get(&V, 1), DestPTy); + return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); } // Finally, implement bitcast folding now. The code below doesn't handle @@ -157,23 +159,24 @@ static Constant *FoldBitCast(LLVMContext &Context, // Handle integral constant input. if (ConstantInt *CI = dyn_cast(V)) { - if (DestTy->isInteger()) + if (DestTy->isIntegerTy()) // 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()) - return ConstantFP::get(Context, APFloat(CI->getValue(), - DestTy != Type::getPPC_FP128Ty(Context))); + if (DestTy->isFloatingPointTy()) + return ConstantFP::get(DestTy->getContext(), + APFloat(CI->getValue(), + !DestTy->isPPC_FP128Ty())); // Otherwise, can't fold this (vector?) return 0; } - // Handle ConstantFP input. + // Handle ConstantFP input: FP -> Integral. if (ConstantFP *FP = dyn_cast(V)) - // FP -> Integral. - return ConstantInt::get(Context, FP->getValueAPF().bitcastToAPInt()); + return ConstantInt::get(FP->getContext(), + FP->getValueAPF().bitcastToAPInt()); return 0; } @@ -190,7 +193,7 @@ static Constant *FoldBitCast(LLVMContext &Context, /// static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, unsigned ByteSize) { - assert(isa(C->getType()) && + assert(C->getType()->isIntegerTy() && (cast(C->getType())->getBitWidth() & 7) == 0 && "Non-byte sized integer input"); unsigned CSize = cast(C->getType())->getBitWidth()/8; @@ -203,7 +206,7 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, APInt V = CI->getValue(); if (ByteStart) V = V.lshr(ByteStart*8); - V.trunc(ByteSize*8); + V = V.trunc(ByteSize*8); return ConstantInt::get(CI->getContext(), V); } @@ -215,7 +218,7 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, switch (CE->getOpcode()) { default: return 0; case Instruction::Or: { - Constant *RHS = ExtractConstantBytes(C->getOperand(1), ByteStart, ByteSize); + Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); if (RHS == 0) return 0; @@ -224,13 +227,13 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, if (RHSC->isAllOnesValue()) return RHSC; - Constant *LHS = ExtractConstantBytes(C->getOperand(0), ByteStart, ByteSize); + Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); if (LHS == 0) return 0; return ConstantExpr::getOr(LHS, RHS); } case Instruction::And: { - Constant *RHS = ExtractConstantBytes(C->getOperand(1), ByteStart, ByteSize); + Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); if (RHS == 0) return 0; @@ -238,7 +241,7 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, if (RHS->isNullValue()) return RHS; - Constant *LHS = ExtractConstantBytes(C->getOperand(0), ByteStart, ByteSize); + Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); if (LHS == 0) return 0; return ConstantExpr::getAnd(LHS, RHS); @@ -259,7 +262,7 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, ByteSize*8)); // If the extract is known to be fully in the input, extract it. if (ByteStart+ByteSize+ShAmt <= CSize) - return ExtractConstantBytes(C->getOperand(0), ByteStart+ShAmt, ByteSize); + return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); // TODO: Handle the 'partially zero' case. return 0; @@ -281,7 +284,7 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, ByteSize*8)); // If the extract is known to be fully in the input, extract it. if (ByteStart >= ShAmt) - return ExtractConstantBytes(C->getOperand(0), ByteStart-ShAmt, ByteSize); + return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); // TODO: Handle the 'partially zero' case. return 0; @@ -289,7 +292,7 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, case Instruction::ZExt: { unsigned SrcBitSize = - cast(C->getOperand(0)->getType())->getBitWidth(); + cast(CE->getOperand(0)->getType())->getBitWidth(); // If extracting something that is completely zero, return 0. if (ByteStart*8 >= SrcBitSize) @@ -298,18 +301,18 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, // If exactly extracting the input, return it. if (ByteStart == 0 && ByteSize*8 == SrcBitSize) - return C->getOperand(0); + 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(C->getOperand(0), ByteStart, ByteSize); + 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 = C->getOperand(0); + Constant *Res = CE->getOperand(0); if (ByteStart) Res = ConstantExpr::getLShr(Res, ConstantInt::get(Res->getType(), ByteStart*8)); @@ -323,10 +326,186 @@ static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, } } +/// 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(Type *Ty, Type *DestTy, + bool Folded) { + if (ArrayType *ATy = dyn_cast(Ty)) { + Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); + Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); + return ConstantExpr::getNUWMul(E, N); + } + + if (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 size. + Constant *MemberSize = + getFoldedSizeOf(STy->getElementType(0), DestTy, true); + bool AllSame = true; + for (unsigned i = 1; i != NumElems; ++i) + if (MemberSize != + getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { + AllSame = false; + break; + } + if (AllSame) { + Constant *N = ConstantInt::get(DestTy, NumElems); + return ConstantExpr::getNUWMul(MemberSize, N); + } + } + + // Pointer size doesn't depend on the pointee type, so canonicalize them + // to an arbitrary pointee. + if (PointerType *PTy = dyn_cast(Ty)) + if (!PTy->getElementType()->isIntegerTy(1)) + return + getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), + PTy->getAddressSpace()), + DestTy, true); + + // 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; +} + +/// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof +/// 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 *getFoldedAlignOf(Type *Ty, Type *DestTy, + bool Folded) { + // The alignment of an array is equal to the alignment of the + // array element. Note that this is not always true for vectors. + if (ArrayType *ATy = dyn_cast(Ty)) { + Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); + C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, + DestTy, + false), + C, DestTy); + return C; + } + + if (StructType *STy = dyn_cast(Ty)) { + // Packed structs always have an alignment of 1. + if (STy->isPacked()) + return ConstantInt::get(DestTy, 1); + + // Otherwise, struct alignment is the maximum alignment of any member. + // Without target data, we can't compare much, but we can check to see + // if all the members have the same alignment. + unsigned NumElems = STy->getNumElements(); + // An empty struct has minimal alignment. + if (NumElems == 0) + return ConstantInt::get(DestTy, 1); + // Check for a struct with all members having the same alignment. + Constant *MemberAlign = + getFoldedAlignOf(STy->getElementType(0), DestTy, true); + bool AllSame = true; + for (unsigned i = 1; i != NumElems; ++i) + if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { + AllSame = false; + break; + } + if (AllSame) + return MemberAlign; + } + + // Pointer alignment doesn't depend on the pointee type, so canonicalize them + // to an arbitrary pointee. + if (PointerType *PTy = dyn_cast(Ty)) + if (!PTy->getElementType()->isIntegerTy(1)) + return + getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), + 1), + PTy->getAddressSpace()), + DestTy, true); + + // 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 alignof expression. + Constant *C = ConstantExpr::getAlignOf(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(Type *Ty, Constant *FieldNo, + Type *DestTy, + bool Folded) { + if (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 (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 size. + Constant *MemberSize = + getFoldedSizeOf(STy->getElementType(0), DestTy, true); + bool AllSame = true; + for (unsigned i = 1; i != NumElems; ++i) + if (MemberSize != + getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { + AllSame = false; + break; + } + if (AllSame) { + Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, + false, + DestTy, + false), + FieldNo, DestTy); + return ConstantExpr::getNUWMul(MemberSize, 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, Constant *V, - const Type *DestTy) { +Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, + 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. @@ -336,10 +515,14 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, return Constant::getNullValue(DestTy); return UndefValue::get(DestTy); } + // No compile-time operations on this type yet. if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty()) return 0; + if (V->isNullValue() && !DestTy->isX86_MMXTy()) + return Constant::getNullValue(DestTy); + // If the cast operand is a constant expression, there's a few things we can // do to try to simplify it. if (ConstantExpr *CE = dyn_cast(V)) { @@ -366,16 +549,16 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, // operating on each element. In the cast of bitcasts, the element // count may be mismatched; don't attempt to handle that here. if (ConstantVector *CV = dyn_cast(V)) - if (isa(DestTy) && + if (DestTy->isVectorTy() && cast(DestTy)->getNumElements() == CV->getType()->getNumElements()) { std::vector res; - const VectorType *DestVecTy = cast(DestTy); - const Type *DstEltTy = DestVecTy->getElementType(); + VectorType *DestVecTy = cast(DestTy); + Type *DstEltTy = DestVecTy->getElementType(); for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) res.push_back(ConstantExpr::getCast(opc, CV->getOperand(i), DstEltTy)); - return ConstantVector::get(DestVecTy, res); + return ConstantVector::get(res); } // We actually have to do a cast now. Perform the cast according to the @@ -394,7 +577,7 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, DestTy->isFP128Ty() ? APFloat::IEEEquad : APFloat::Bogus, APFloat::rmNearestTiesToEven, &ignored); - return ConstantFP::get(Context, Val); + return ConstantFP::get(V->getContext(), Val); } return 0; // Can't fold. case Instruction::FPToUI: @@ -406,8 +589,8 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, uint32_t DestBitWidth = cast(DestTy)->getBitWidth(); (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, APFloat::rmTowardZero, &ignored); - APInt Val(DestBitWidth, 2, x); - return ConstantInt::get(Context, Val); + APInt Val(DestBitWidth, x); + return ConstantInt::get(FPC->getContext(), Val); } return 0; // Can't fold. case Instruction::IntToPtr: //always treated as unsigned @@ -415,44 +598,79 @@ 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()) { + 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 (StructType *STy = dyn_cast(Ty)) + if (!STy->isPacked()) { + ConstantInt *CI = cast(CE->getOperand(2)); + if (CI->isOne() && + STy->getNumElements() == 2 && + STy->getElementType(0)->isIntegerTy(1)) { + return getFoldedAlignOf(STy->getElementType(1), DestTy, false); + } + } + // Handle an offsetof-like expression. + if (Ty->isStructTy() || Ty->isArrayTy()) { + 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 (ConstantInt *CI = dyn_cast(V)) { APInt api = CI->getValue(); - const uint64_t zero[] = {0, 0}; - APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(), - 2, zero)); + APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true); (void)apf.convertFromAPInt(api, opc==Instruction::SIToFP, APFloat::rmNearestTiesToEven); - return ConstantFP::get(Context, apf); + return ConstantFP::get(V->getContext(), apf); } return 0; case Instruction::ZExt: if (ConstantInt *CI = dyn_cast(V)) { uint32_t BitWidth = cast(DestTy)->getBitWidth(); - APInt Result(CI->getValue()); - Result.zext(BitWidth); - return ConstantInt::get(Context, Result); + return ConstantInt::get(V->getContext(), + CI->getValue().zext(BitWidth)); } return 0; case Instruction::SExt: if (ConstantInt *CI = dyn_cast(V)) { uint32_t BitWidth = cast(DestTy)->getBitWidth(); - APInt Result(CI->getValue()); - Result.sext(BitWidth); - return ConstantInt::get(Context, Result); + return ConstantInt::get(V->getContext(), + CI->getValue().sext(BitWidth)); } return 0; case Instruction::Trunc: { uint32_t DestBitWidth = cast(DestTy)->getBitWidth(); if (ConstantInt *CI = dyn_cast(V)) { - APInt Result(CI->getValue()); - Result.trunc(DestBitWidth); - return ConstantInt::get(Context, Result); + return ConstantInt::get(V->getContext(), + CI->getValue().trunc(DestBitWidth)); } // The input must be a constantexpr. See if we can simplify this based on @@ -466,25 +684,74 @@ Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, return 0; } case Instruction::BitCast: - return FoldBitCast(Context, V, DestTy); + return FoldBitCast(V, DestTy); } } -Constant *llvm::ConstantFoldSelectInstruction(LLVMContext&, - Constant *Cond, +Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, Constant *V1, Constant *V2) { if (ConstantInt *CB = dyn_cast(Cond)) return CB->getZExtValue() ? V1 : V2; + // Check for zero aggregate and ConstantVector of zeros + if (Cond->isNullValue()) return V2; + + if (ConstantVector* CondV = dyn_cast(Cond)) { + + if (CondV->isAllOnesValue()) return V1; + + VectorType *VTy = cast(V1->getType()); + ConstantVector *CP1 = dyn_cast(V1); + ConstantVector *CP2 = dyn_cast(V2); + + if ((CP1 || isa(V1)) && + (CP2 || isa(V2))) { + + // Find the element type of the returned vector + Type *EltTy = VTy->getElementType(); + unsigned NumElem = VTy->getNumElements(); + std::vector Res(NumElem); + + bool Valid = true; + for (unsigned i = 0; i < NumElem; ++i) { + ConstantInt* c = dyn_cast(CondV->getOperand(i)); + if (!c) { + Valid = false; + break; + } + Constant *C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); + Constant *C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); + Res[i] = c->getZExtValue() ? C1 : C2; + } + // If we were able to build the vector, return it + if (Valid) return ConstantVector::get(Res); + } + } + + + if (isa(Cond)) { + if (isa(V1)) return V1; + return V2; + } if (isa(V1)) return V2; if (isa(V2)) return V1; - if (isa(Cond)) return V1; if (V1 == V2) return V1; + + if (ConstantExpr *TrueVal = dyn_cast(V1)) { + if (TrueVal->getOpcode() == Instruction::Select) + if (TrueVal->getOperand(0) == Cond) + return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); + } + if (ConstantExpr *FalseVal = dyn_cast(V2)) { + if (FalseVal->getOpcode() == Instruction::Select) + if (FalseVal->getOperand(0) == Cond) + return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); + } + return 0; } -Constant *llvm::ConstantFoldExtractElementInstruction(LLVMContext &Context, - Constant *Val, +Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, Constant *Idx) { if (isa(Val)) // ee(undef, x) -> undef return UndefValue::get(cast(Val->getType())->getElementType()); @@ -503,8 +770,7 @@ Constant *llvm::ConstantFoldExtractElementInstruction(LLVMContext &Context, return 0; } -Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, - Constant *Val, +Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, Constant *Elt, Constant *Idx) { ConstantInt *CIdx = dyn_cast(Idx); @@ -563,12 +829,11 @@ 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, Constant *C, - unsigned EltNo) { +static Constant *GetVectorElement(Constant *C, unsigned EltNo) { if (ConstantVector *CV = dyn_cast(C)) return CV->getOperand(EltNo); - const Type *EltTy = cast(C->getType())->getElementType(); + Type *EltTy = cast(C->getType())->getElementType(); if (isa(C)) return Constant::getNullValue(EltTy); if (isa(C)) @@ -576,8 +841,7 @@ static Constant *GetVectorElement(LLVMContext &Context, Constant *C, return 0; } -Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, - Constant *V1, +Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, Constant *V2, Constant *Mask) { // Undefined shuffle mask -> undefined value. @@ -585,12 +849,12 @@ Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, unsigned MaskNumElts = cast(Mask->getType())->getNumElements(); unsigned SrcNumElts = cast(V1->getType())->getNumElements(); - const Type *EltTy = cast(V1->getType())->getElementType(); + Type *EltTy = cast(V1->getType())->getElementType(); // Loop over the shuffle mask, evaluating each element. SmallVector Result; for (unsigned i = 0; i != MaskNumElts; ++i) { - Constant *InElt = GetVectorElement(Context, Mask, i); + Constant *InElt = GetVectorElement(Mask, i); if (InElt == 0) return 0; if (isa(InElt)) @@ -600,9 +864,9 @@ Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, if (Elt >= SrcNumElts*2) InElt = UndefValue::get(EltTy); else if (Elt >= SrcNumElts) - InElt = GetVectorElement(Context, V2, Elt - SrcNumElts); + InElt = GetVectorElement(V2, Elt - SrcNumElts); else - InElt = GetVectorElement(Context, V1, Elt); + InElt = GetVectorElement(V1, Elt); if (InElt == 0) return 0; } else { // Unknown value. @@ -611,40 +875,42 @@ Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, Result.push_back(InElt); } - return ConstantVector::get(&Result[0], Result.size()); + return ConstantVector::get(Result); } -Constant *llvm::ConstantFoldExtractValueInstruction(LLVMContext &Context, - Constant *Agg, - const unsigned *Idxs, - unsigned NumIdx) { +Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, + ArrayRef Idxs) { // Base case: no indices, so return the entire value. - if (NumIdx == 0) + if (Idxs.empty()) return Agg; if (isa(Agg)) // ev(undef, x) -> undef return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(), - Idxs, - Idxs + NumIdx)); + Idxs)); if (isa(Agg)) // ev(0, x) -> 0 return Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), - Idxs, - Idxs + NumIdx)); + Idxs)); // Otherwise recurse. - return ConstantFoldExtractValueInstruction(Context, Agg->getOperand(*Idxs), - Idxs+1, NumIdx-1); + if (ConstantStruct *CS = dyn_cast(Agg)) + return ConstantFoldExtractValueInstruction(CS->getOperand(Idxs[0]), + Idxs.slice(1)); + + if (ConstantArray *CA = dyn_cast(Agg)) + return ConstantFoldExtractValueInstruction(CA->getOperand(Idxs[0]), + Idxs.slice(1)); + ConstantVector *CV = cast(Agg); + return ConstantFoldExtractValueInstruction(CV->getOperand(Idxs[0]), + Idxs.slice(1)); } -Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, - Constant *Agg, +Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, Constant *Val, - const unsigned *Idxs, - unsigned NumIdx) { + ArrayRef Idxs) { // Base case: no indices, so replace the entire value. - if (NumIdx == 0) + if (Idxs.empty()) return Val; if (isa(Agg)) { @@ -655,26 +921,26 @@ Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, // Otherwise break the aggregate undef into multiple undefs and do // the insertion. - const CompositeType *AggTy = cast(Agg->getType()); + CompositeType *AggTy = cast(Agg->getType()); unsigned numOps; - if (const ArrayType *AR = dyn_cast(AggTy)) + if (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); + Type *MemberTy = AggTy->getTypeAtIndex(i); Constant *Op = - (*Idxs == i) ? - ConstantFoldInsertValueInstruction(Context, UndefValue::get(MemberTy), - Val, Idxs+1, NumIdx-1) : + (Idxs[0] == i) ? + ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy), + Val, Idxs.slice(1)) : UndefValue::get(MemberTy); Ops[i] = Op; } - if (const StructType* ST = dyn_cast(AggTy)) - return ConstantStruct::get(Context, Ops, ST->isPacked()); + if (StructType* ST = dyn_cast(AggTy)) + return ConstantStruct::get(ST, Ops); return ConstantArray::get(cast(AggTy), Ops); } @@ -686,27 +952,26 @@ Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, // Otherwise break the aggregate zero into multiple zeros and do // the insertion. - const CompositeType *AggTy = cast(Agg->getType()); + CompositeType *AggTy = cast(Agg->getType()); unsigned numOps; - if (const ArrayType *AR = dyn_cast(AggTy)) + if (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); + Type *MemberTy = AggTy->getTypeAtIndex(i); Constant *Op = - (*Idxs == i) ? - ConstantFoldInsertValueInstruction(Context, - Constant::getNullValue(MemberTy), - Val, Idxs+1, NumIdx-1) : + (Idxs[0] == i) ? + ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy), + Val, Idxs.slice(1)) : Constant::getNullValue(MemberTy); Ops[i] = Op; } - if (const StructType* ST = dyn_cast(AggTy)) - return ConstantStruct::get(Context, Ops, ST->isPacked()); + if (StructType *ST = dyn_cast(AggTy)) + return ConstantStruct::get(ST, Ops); return ConstantArray::get(cast(AggTy), Ops); } @@ -714,16 +979,14 @@ Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, // Insertion of constant into aggregate constant. std::vector Ops(Agg->getNumOperands()); for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { - Constant *Op = - (*Idxs == i) ? - ConstantFoldInsertValueInstruction(Context, Agg->getOperand(i), - Val, Idxs+1, NumIdx-1) : - Agg->getOperand(i); + Constant *Op = cast(Agg->getOperand(i)); + if (Idxs[0] == i) + Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs.slice(1)); Ops[i] = Op; } - if (const StructType* ST = dyn_cast(Agg->getType())) - return ConstantStruct::get(Context, Ops, ST->isPacked()); + if (StructType* ST = dyn_cast(Agg->getType())) + return ConstantStruct::get(ST, Ops); return ConstantArray::get(cast(Agg->getType()), Ops); } @@ -731,8 +994,7 @@ Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, } -Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, - unsigned Opcode, +Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1, Constant *C2) { // No compile-time operations on this type yet. if (C1->getType()->isPPC_FP128Ty()) @@ -750,33 +1012,53 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, case Instruction::Add: case Instruction::Sub: return UndefValue::get(C1->getType()); - case Instruction::Mul: case Instruction::And: + if (isa(C1) && isa(C2)) // undef & undef -> undef + return C1; + return Constant::getNullValue(C1->getType()); // undef & X -> 0 + case Instruction::Mul: { + ConstantInt *CI; + // X * undef -> undef if X is odd or undef + if (((CI = dyn_cast(C1)) && CI->getValue()[0]) || + ((CI = dyn_cast(C2)) && CI->getValue()[0]) || + (isa(C1) && isa(C2))) + return UndefValue::get(C1->getType()); + + // X * undef -> 0 otherwise return Constant::getNullValue(C1->getType()); + } case Instruction::UDiv: case Instruction::SDiv: + // undef / 1 -> undef + if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv) + if (ConstantInt *CI2 = dyn_cast(C2)) + if (CI2->isOne()) + return C1; + // FALL THROUGH case Instruction::URem: case Instruction::SRem: if (!isa(C2)) // undef / X -> 0 return Constant::getNullValue(C1->getType()); return C2; // X / undef -> undef case Instruction::Or: // X | undef -> -1 - if (const VectorType *PTy = dyn_cast(C1->getType())) - return Constant::getAllOnesValue(PTy); - return Constant::getAllOnesValue(C1->getType()); + if (isa(C1) && isa(C2)) // undef | undef -> undef + return C1; + return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 case Instruction::LShr: if (isa(C2) && isa(C1)) return C1; // undef lshr undef -> undef return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 // undef lshr X -> 0 case Instruction::AShr: - if (!isa(C2)) - return C1; // undef ashr X --> undef + if (!isa(C2)) // undef ashr X --> all ones + return Constant::getAllOnesValue(C1->getType()); else if (isa(C1)) return C1; // undef ashr undef -> undef else return C1; // X ashr undef --> X case Instruction::Shl: + if (isa(C2) && isa(C1)) + return C1; // undef shl undef -> undef // undef << X -> 0 or X << undef -> 0 return Constant::getNullValue(C1->getType()); } @@ -877,6 +1159,10 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, return ConstantExpr::getLShr(C1, C2); break; } + } else if (isa(C1)) { + // If C1 is a ConstantInt and C2 is not, swap the operands. + if (Instruction::isCommutative(Opcode)) + return ConstantExpr::get(Opcode, C2, C1); } // At this point we know neither constant is an UndefValue. @@ -889,51 +1175,51 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, default: break; case Instruction::Add: - return ConstantInt::get(Context, C1V + C2V); + return ConstantInt::get(CI1->getContext(), C1V + C2V); case Instruction::Sub: - return ConstantInt::get(Context, C1V - C2V); + return ConstantInt::get(CI1->getContext(), C1V - C2V); case Instruction::Mul: - return ConstantInt::get(Context, C1V * C2V); + return ConstantInt::get(CI1->getContext(), C1V * C2V); case Instruction::UDiv: assert(!CI2->isNullValue() && "Div by zero handled above"); - return ConstantInt::get(Context, C1V.udiv(C2V)); + return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); case Instruction::SDiv: assert(!CI2->isNullValue() && "Div by zero handled above"); if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef - return ConstantInt::get(Context, C1V.sdiv(C2V)); + return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); case Instruction::URem: assert(!CI2->isNullValue() && "Div by zero handled above"); - return ConstantInt::get(Context, C1V.urem(C2V)); + return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); case Instruction::SRem: assert(!CI2->isNullValue() && "Div by zero handled above"); if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef - return ConstantInt::get(Context, C1V.srem(C2V)); + return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); case Instruction::And: - return ConstantInt::get(Context, C1V & C2V); + return ConstantInt::get(CI1->getContext(), C1V & C2V); case Instruction::Or: - return ConstantInt::get(Context, C1V | C2V); + return ConstantInt::get(CI1->getContext(), C1V | C2V); case Instruction::Xor: - return ConstantInt::get(Context, C1V ^ C2V); + return ConstantInt::get(CI1->getContext(), C1V ^ C2V); case Instruction::Shl: { uint32_t shiftAmt = C2V.getZExtValue(); if (shiftAmt < C1V.getBitWidth()) - return ConstantInt::get(Context, C1V.shl(shiftAmt)); + return ConstantInt::get(CI1->getContext(), 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(Context, C1V.lshr(shiftAmt)); + return ConstantInt::get(CI1->getContext(), 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(Context, C1V.ashr(shiftAmt)); + return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); else return UndefValue::get(C1->getType()); // too big shift is undef } @@ -963,28 +1249,28 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, break; case Instruction::FAdd: (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(Context, C3V); + return ConstantFP::get(C1->getContext(), C3V); case Instruction::FSub: (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(Context, C3V); + return ConstantFP::get(C1->getContext(), C3V); case Instruction::FMul: (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(Context, C3V); + return ConstantFP::get(C1->getContext(), C3V); case Instruction::FDiv: (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(Context, C3V); + return ConstantFP::get(C1->getContext(), C3V); case Instruction::FRem: (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); - return ConstantFP::get(Context, C3V); + return ConstantFP::get(C1->getContext(), C3V); } } - } else if (const VectorType *VTy = dyn_cast(C1->getType())) { + } else if (VectorType *VTy = dyn_cast(C1->getType())) { ConstantVector *CP1 = dyn_cast(C1); ConstantVector *CP2 = dyn_cast(C2); if ((CP1 != NULL || isa(C1)) && (CP2 != NULL || isa(C2))) { std::vector Res; - const Type* EltTy = VTy->getElementType(); + Type* EltTy = VTy->getElementType(); Constant *C1 = 0; Constant *C2 = 0; switch (Opcode) { @@ -1120,42 +1406,27 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, } } - 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) && 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. - switch (Opcode) { - case Instruction::Add: - case Instruction::FAdd: - case Instruction::Mul: - case Instruction::FMul: - case Instruction::And: - case Instruction::Or: - case Instruction::Xor: - // No change of opcode required. - return ConstantFoldBinaryInstruction(Context, Opcode, C2, C1); - - case Instruction::Shl: - case Instruction::LShr: - case Instruction::AShr: - case Instruction::Sub: - case Instruction::FSub: - 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; - } + if (Instruction::isCommutative(Opcode)) + return ConstantFoldBinaryInstruction(Opcode, C2, C1); } // i1 can be simplified in many cases. - if (C1->getType() == Type::getInt1Ty(Context)) { + if (C1->getType()->isIntegerTy(1)) { switch (Opcode) { case Instruction::Add: case Instruction::Sub: @@ -1177,7 +1448,7 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, 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); + return ConstantInt::getFalse(C1->getContext()); default: break; } @@ -1189,16 +1460,16 @@ Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, /// isZeroSizedType - This type is zero sized if its an array or structure of /// zero sized types. The only leaf zero sized type is an empty structure. -static bool isMaybeZeroSizedType(const Type *Ty) { - if (isa(Ty)) return true; // Can't say. - if (const StructType *STy = dyn_cast(Ty)) { +static bool isMaybeZeroSizedType(Type *Ty) { + if (StructType *STy = dyn_cast(Ty)) { + if (STy->isOpaque()) return true; // Can't say. // If all of elements have zero size, this does too. for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; return true; - } else if (const ArrayType *ATy = dyn_cast(Ty)) { + } else if (ArrayType *ATy = dyn_cast(Ty)) { return isMaybeZeroSizedType(ATy->getElementType()); } return false; @@ -1211,8 +1482,7 @@ static bool isMaybeZeroSizedType(const Type *Ty) { /// first is less than the second, return -1, if the second is less than the /// first, return 1. If the constants are not integral, return -2. /// -static int IdxCompare(LLVMContext &Context, Constant *C1, Constant *C2, - const Type *ElTy) { +static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { if (C1 == C2) return 0; // Ok, we found a different index. If they are not ConstantInt, we can't do @@ -1222,11 +1492,11 @@ 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::getInt64Ty(Context)) - C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(Context)); + if (!C1->getType()->isIntegerTy(64)) + C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); - if (C2->getType() != Type::getInt64Ty(Context)) - C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(Context)); + if (!C2->getType()->isIntegerTy(64)) + C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); if (C1 == C2) return 0; // They are equal @@ -1255,8 +1525,7 @@ static int IdxCompare(LLVMContext &Context, Constant *C1, Constant *C2, /// To simplify this code we canonicalize the relation so that the first /// 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, - Constant *V1, Constant *V2) { +static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { assert(V1->getType() == V2->getType() && "Cannot compare values of different types!"); @@ -1289,7 +1558,7 @@ static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, } // If the first operand is simple and second is ConstantExpr, swap operands. - FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(Context, V2, V1); + FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) return FCmpInst::getSwappedPredicate(SwappedRelation); } else { @@ -1324,16 +1593,16 @@ static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, /// constants (like ConstantInt) to be the simplest, followed by /// GlobalValues, followed by ConstantExpr's (the most complex). /// -static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, - Constant *V1, - Constant *V2, +static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, bool isSigned) { assert(V1->getType() == V2->getType() && "Cannot compare different types of values!"); if (V1 == V2) return ICmpInst::ICMP_EQ; - if (!isa(V1) && !isa(V1)) { - if (!isa(V2) && !isa(V2)) { + if (!isa(V1) && !isa(V1) && + !isa(V1)) { + if (!isa(V2) && !isa(V2) && + !isa(V2)) { // We distilled this down to a simple case, use the standard constant // folder. ConstantInt *R = 0; @@ -1356,36 +1625,63 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, // If the first operand is simple, swap operands. ICmpInst::Predicate SwappedRelation = - evaluateICmpRelation(Context, V2, V1, isSigned); + evaluateICmpRelation(V2, V1, isSigned); if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) return ICmpInst::getSwappedPredicate(SwappedRelation); - } else if (const GlobalValue *CPR1 = dyn_cast(V1)) { + } else if (const GlobalValue *GV = dyn_cast(V1)) { if (isa(V2)) { // Swap as necessary. ICmpInst::Predicate SwappedRelation = - evaluateICmpRelation(Context, V2, V1, isSigned); + evaluateICmpRelation(V2, V1, isSigned); if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) return ICmpInst::getSwappedPredicate(SwappedRelation); - else - return ICmpInst::BAD_ICMP_PREDICATE; + return ICmpInst::BAD_ICMP_PREDICATE; } - // Now we know that the RHS is a GlobalValue or simple constant, - // which (since the types must match) means that it's a ConstantPointerNull. - if (const GlobalValue *CPR2 = dyn_cast(V2)) { + // Now we know that the RHS is a GlobalValue, BlockAddress or simple + // constant (which, since the types must match, means that it's a + // ConstantPointerNull). + if (const GlobalValue *GV2 = dyn_cast(V2)) { // Don't try to decide equality of aliases. - if (!isa(CPR1) && !isa(CPR2)) - if (!CPR1->hasExternalWeakLinkage() || !CPR2->hasExternalWeakLinkage()) + if (!isa(GV) && !isa(GV2)) + if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage()) return ICmpInst::ICMP_NE; + } else if (isa(V2)) { + return ICmpInst::ICMP_NE; // Globals never equal labels. } else { assert(isa(V2) && "Canonicalization guarantee!"); - // GlobalVals can never be null. Don't try to evaluate aliases. - if (!CPR1->hasExternalWeakLinkage() && !isa(CPR1)) + // GlobalVals can never be null unless they have external weak linkage. + // We don't try to evaluate aliases here. + if (!GV->hasExternalWeakLinkage() && !isa(GV)) + return ICmpInst::ICMP_NE; + } + } else if (const BlockAddress *BA = dyn_cast(V1)) { + if (isa(V2)) { // Swap as necessary. + ICmpInst::Predicate SwappedRelation = + evaluateICmpRelation(V2, V1, isSigned); + if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) + return ICmpInst::getSwappedPredicate(SwappedRelation); + return ICmpInst::BAD_ICMP_PREDICATE; + } + + // Now we know that the RHS is a GlobalValue, BlockAddress or simple + // constant (which, since the types must match, means that it is a + // ConstantPointerNull). + if (const BlockAddress *BA2 = dyn_cast(V2)) { + // Block address in another function can't equal this one, but block + // addresses in the current function might be the same if blocks are + // empty. + if (BA2->getFunction() != BA->getFunction()) return ICmpInst::ICMP_NE; + } else { + // Block addresses aren't null, don't equal the address of globals. + assert((isa(V2) || isa(V2)) && + "Canonicalization guarantee!"); + return ICmpInst::ICMP_NE; } } else { // Ok, the LHS is known to be a constantexpr. The RHS can be any of a - // constantexpr, a CPR, or a simple constant. + // constantexpr, a global, block address, or a simple constant. ConstantExpr *CE1 = cast(V1); Constant *CE1Op0 = CE1->getOperand(0); @@ -1405,10 +1701,10 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, // 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())) { + (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; if (CE1->getOpcode() == Instruction::SExt) isSigned = true; - return evaluateICmpRelation(Context, CE1Op0, + return evaluateICmpRelation(CE1Op0, Constant::getNullValue(CE1Op0->getType()), isSigned); } @@ -1440,9 +1736,9 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, return ICmpInst::ICMP_EQ; } // Otherwise, we can't really say if the first operand is null or not. - } else if (const GlobalValue *CPR2 = dyn_cast(V2)) { + } else if (const GlobalValue *GV2 = dyn_cast(V2)) { if (isa(CE1Op0)) { - if (CPR2->hasExternalWeakLinkage()) + if (GV2->hasExternalWeakLinkage()) // Weak linkage GVals could be zero or not. We're comparing it to // a null pointer, so its less-or-equal return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; @@ -1450,15 +1746,15 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, // If its not weak linkage, the GVal must have a non-zero address // so the result is less-than return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; - } else if (const GlobalValue *CPR1 = dyn_cast(CE1Op0)) { - if (CPR1 == CPR2) { + } else if (const GlobalValue *GV = dyn_cast(CE1Op0)) { + if (GV == GV2) { // If this is a getelementptr of the same global, then it must be // different. Because the types must match, the getelementptr could // only have at most one index, and because we fold getelementptr's // with a single zero index, it must be nonzero. assert(CE1->getNumOperands() == 2 && !CE1->getOperand(1)->isNullValue() && - "Suprising getelementptr!"); + "Surprising getelementptr!"); return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; } else { // If they are different globals, we don't know what the value is, @@ -1497,7 +1793,7 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, gep_type_iterator GTI = gep_type_begin(CE1); for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); ++i, ++GTI) - switch (IdxCompare(Context, CE1->getOperand(i), + switch (IdxCompare(CE1->getOperand(i), CE2->getOperand(i), GTI.getIndexedType())) { case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; @@ -1533,14 +1829,14 @@ static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, return ICmpInst::BAD_ICMP_PREDICATE; } -Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, - unsigned short pred, +Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, Constant *C1, Constant *C2) { - const Type *ResultTy; - if (const VectorType *VT = dyn_cast(C1->getType())) - ResultTy = VectorType::get(Type::getInt1Ty(Context), VT->getNumElements()); + Type *ResultTy; + if (VectorType *VT = dyn_cast(C1->getType())) + ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), + VT->getNumElements()); else - ResultTy = Type::getInt1Ty(Context); + ResultTy = Type::getInt1Ty(C1->getContext()); // Fold FCMP_FALSE/FCMP_TRUE unconditionally. if (pred == FCmpInst::FCMP_FALSE) @@ -1550,8 +1846,17 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, return Constant::getAllOnesValue(ResultTy); // Handle some degenerate cases first - if (isa(C1) || isa(C2)) - return UndefValue::get(ResultTy); + if (isa(C1) || isa(C2)) { + // For EQ and NE, we can always pick a value for the undef to make the + // predicate pass or fail, so we can return undef. + // Also, if both operands are undef, we can return undef. + if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) || + (isa(C1) && isa(C2))) + return UndefValue::get(ResultTy); + // Otherwise, pick the same value as the non-undef operand, and fold + // it to true or false. + return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); + } // No compile-time operations on this type yet. if (C1->getType()->isPPC_FP128Ty()) @@ -1563,9 +1868,9 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, // 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(Context); + return ConstantInt::getFalse(C1->getContext()); else if (pred == ICmpInst::ICMP_NE) - return ConstantInt::getTrue(Context); + return ConstantInt::getTrue(C1->getContext()); } // icmp eq/ne(GV,null) -> false/true } else if (C2->isNullValue()) { @@ -1573,14 +1878,14 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, // 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(Context); + return ConstantInt::getFalse(C1->getContext()); else if (pred == ICmpInst::ICMP_NE) - return ConstantInt::getTrue(Context); + return ConstantInt::getTrue(C1->getContext()); } } // If the comparison is a comparison between two i1's, simplify it. - if (C1->getType() == Type::getInt1Ty(Context)) { + if (C1->getType()->isIntegerTy(1)) { switch(pred) { case ICmpInst::ICMP_EQ: if (isa(C2)) @@ -1598,26 +1903,16 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, APInt V2 = cast(C2)->getValue(); switch (pred) { default: llvm_unreachable("Invalid ICmp Predicate"); return 0; - case ICmpInst::ICMP_EQ: - return ConstantInt::get(Type::getInt1Ty(Context), V1 == V2); - case ICmpInst::ICMP_NE: - return ConstantInt::get(Type::getInt1Ty(Context), V1 != V2); - case ICmpInst::ICMP_SLT: - return ConstantInt::get(Type::getInt1Ty(Context), V1.slt(V2)); - case ICmpInst::ICMP_SGT: - return ConstantInt::get(Type::getInt1Ty(Context), V1.sgt(V2)); - case ICmpInst::ICMP_SLE: - return ConstantInt::get(Type::getInt1Ty(Context), V1.sle(V2)); - case ICmpInst::ICMP_SGE: - return ConstantInt::get(Type::getInt1Ty(Context), V1.sge(V2)); - case ICmpInst::ICMP_ULT: - return ConstantInt::get(Type::getInt1Ty(Context), V1.ult(V2)); - case ICmpInst::ICMP_UGT: - return ConstantInt::get(Type::getInt1Ty(Context), V1.ugt(V2)); - case ICmpInst::ICMP_ULE: - return ConstantInt::get(Type::getInt1Ty(Context), V1.ule(V2)); - case ICmpInst::ICMP_UGE: - return ConstantInt::get(Type::getInt1Ty(Context), V1.uge(V2)); + case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); + case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); + case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); + case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); + case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); + case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); + case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); + case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); + case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); + case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); } } else if (isa(C1) && isa(C2)) { APFloat C1V = cast(C1)->getValueAPF(); @@ -1625,62 +1920,63 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, APFloat::cmpResult R = C1V.compare(C2V); switch (pred) { default: llvm_unreachable("Invalid FCmp Predicate"); return 0; - case FCmpInst::FCMP_FALSE: return ConstantInt::getFalse(Context); - case FCmpInst::FCMP_TRUE: return ConstantInt::getTrue(Context); + case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); + case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); case FCmpInst::FCMP_UNO: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered); + return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); case FCmpInst::FCMP_ORD: - return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpUnordered); + return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); case FCmpInst::FCMP_UEQ: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || - R==APFloat::cmpEqual); + return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || + R==APFloat::cmpEqual); case FCmpInst::FCMP_OEQ: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpEqual); + return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); case FCmpInst::FCMP_UNE: - return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpEqual); + return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); case FCmpInst::FCMP_ONE: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan || - R==APFloat::cmpGreaterThan); + return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || + R==APFloat::cmpGreaterThan); case FCmpInst::FCMP_ULT: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || - R==APFloat::cmpLessThan); + return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || + R==APFloat::cmpLessThan); case FCmpInst::FCMP_OLT: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan); + return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); case FCmpInst::FCMP_UGT: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || - R==APFloat::cmpGreaterThan); + return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || + R==APFloat::cmpGreaterThan); case FCmpInst::FCMP_OGT: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpGreaterThan); + return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); case FCmpInst::FCMP_ULE: - return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpGreaterThan); + return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); case FCmpInst::FCMP_OLE: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan || - R==APFloat::cmpEqual); + return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || + R==APFloat::cmpEqual); case FCmpInst::FCMP_UGE: - return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpLessThan); + return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); case FCmpInst::FCMP_OGE: - return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpGreaterThan || - R==APFloat::cmpEqual); + return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || + R==APFloat::cmpEqual); } - } else if (isa(C1->getType())) { + } else if (C1->getType()->isVectorTy()) { SmallVector C1Elts, C2Elts; - C1->getVectorElements(Context, C1Elts); - C2->getVectorElements(Context, C2Elts); + C1->getVectorElements(C1Elts); + C2->getVectorElements(C2Elts); + if (C1Elts.empty() || C2Elts.empty()) + return 0; // If we can constant fold the comparison of each element, constant fold // the whole vector comparison. SmallVector ResElts; - for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) { - // Compare the elements, producing an i1 result or constant expr. - ResElts.push_back( - ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); - } - return ConstantVector::get(&ResElts[0], ResElts.size()); + // Compare the elements, producing an i1 result or constant expr. + for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) + ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); + + return ConstantVector::get(ResElts); } - if (C1->getType()->isFloatingPoint()) { + if (C1->getType()->isFloatingPointTy()) { int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. - switch (evaluateFCmpRelation(Context, C1, C2)) { + switch (evaluateFCmpRelation(C1, C2)) { default: llvm_unreachable("Unknown relation!"); case FCmpInst::FCMP_UNO: case FCmpInst::FCMP_ORD: @@ -1723,7 +2019,7 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) Result = 1; break; - case ICmpInst::ICMP_NE: // We know that C1 != C2 + case FCmpInst::FCMP_ONE: // We know that C1 != C2 // We can only partially decide this relation. if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) Result = 0; @@ -1734,12 +2030,12 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, // If we evaluated the result, return it now. if (Result != -1) - return ConstantInt::get(Type::getInt1Ty(Context), Result); + return ConstantInt::get(ResultTy, Result); } else { // Evaluate the relation between the two constants, per the predicate. int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. - switch (evaluateICmpRelation(Context, C1, C2, CmpInst::isSigned(pred))) { + switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { default: llvm_unreachable("Unknown relational!"); case ICmpInst::BAD_ICMP_PREDICATE: break; // Couldn't determine anything about these constants. @@ -1804,13 +2100,15 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, // If we evaluated the result, return it now. if (Result != -1) - return ConstantInt::get(Type::getInt1Ty(Context), Result); + return ConstantInt::get(ResultTy, Result); // If the right hand side is a bitcast, try using its inverse to simplify - // it by moving it to the left hand side. + // it by moving it to the left hand side. We can't do this if it would turn + // a vector compare into a scalar compare or visa versa. if (ConstantExpr *CE2 = dyn_cast(C2)) { - if (CE2->getOpcode() == Instruction::BitCast) { - Constant *CE2Op0 = CE2->getOperand(0); + Constant *CE2Op0 = CE2->getOperand(0); + if (CE2->getOpcode() == Instruction::BitCast && + CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); return ConstantExpr::getICmp(pred, Inverse, CE2Op0); } @@ -1818,8 +2116,8 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, // If the left hand side is an extension, try eliminating it. if (ConstantExpr *CE1 = dyn_cast(C1)) { - if (CE1->getOpcode() == Instruction::SExt || - CE1->getOpcode() == Instruction::ZExt) { + if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || + (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ Constant *CE1Op0 = CE1->getOperand(0); Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); if (CE1Inverse == CE1Op0) { @@ -1832,30 +2130,13 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, } } - if (!isa(C1) && isa(C2)) { + 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. - switch (pred) { - case ICmpInst::ICMP_EQ: - case ICmpInst::ICMP_NE: - // No change of predicate required. - return ConstantFoldCompareInstruction(Context, pred, C2, C1); - - case ICmpInst::ICMP_ULT: - case ICmpInst::ICMP_SLT: - case ICmpInst::ICMP_UGT: - case ICmpInst::ICMP_SGT: - case ICmpInst::ICMP_ULE: - case ICmpInst::ICMP_SLE: - case ICmpInst::ICMP_UGE: - case ICmpInst::ICMP_SGE: - // Change the predicate as necessary to swap the operands. - pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); - return ConstantFoldCompareInstruction(Context, pred, C2, C1); - - default: // These predicates cannot be flopped around. - break; - } + // Also, if C1 is null and C2 isn't, flip them around. + pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); + return ConstantExpr::getICmp(pred, C2, C1); } } return 0; @@ -1863,57 +2144,53 @@ Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, /// isInBoundsIndices - Test whether the given sequence of *normalized* indices /// is "inbounds". -static bool isInBoundsIndices(Constant *const *Idxs, size_t NumIdx) { +template +static bool isInBoundsIndices(ArrayRef Idxs) { // No indices means nothing that could be out of bounds. - if (NumIdx == 0) return true; + if (Idxs.empty()) return true; // If the first index is zero, it's in bounds. - if (Idxs[0]->isNullValue()) return true; + if (cast(Idxs[0])->isNullValue()) return true; // If the first index is one and all the rest are zero, it's in bounds, // by the one-past-the-end rule. if (!cast(Idxs[0])->isOne()) return false; - for (unsigned i = 1, e = NumIdx; i != e; ++i) - if (!Idxs[i]->isNullValue()) + for (unsigned i = 1, e = Idxs.size(); i != e; ++i) + if (!cast(Idxs[i])->isNullValue()) return false; return true; } -Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, - Constant *C, - bool inBounds, - Constant* const *Idxs, - unsigned NumIdx) { - if (NumIdx == 0 || - (NumIdx == 1 && Idxs[0]->isNullValue())) +template +static Constant *ConstantFoldGetElementPtrImpl(Constant *C, + bool inBounds, + ArrayRef Idxs) { + if (Idxs.empty()) return C; + Constant *Idx0 = cast(Idxs[0]); + if ((Idxs.size() == 1 && Idx0->isNullValue())) return C; if (isa(C)) { - const PointerType *Ptr = cast(C->getType()); - const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, - (Value **)Idxs, - (Value **)Idxs+NumIdx); + PointerType *Ptr = cast(C->getType()); + Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); assert(Ty != 0 && "Invalid indices for GEP!"); return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); } - Constant *Idx0 = Idxs[0]; if (C->isNullValue()) { bool isNull = true; - for (unsigned i = 0, e = NumIdx; i != e; ++i) - if (!Idxs[i]->isNullValue()) { + for (unsigned i = 0, e = Idxs.size(); i != e; ++i) + if (!cast(Idxs[i])->isNullValue()) { isNull = false; break; } if (isNull) { - const PointerType *Ptr = cast(C->getType()); - const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, - (Value**)Idxs, - (Value**)Idxs+NumIdx); + PointerType *Ptr = cast(C->getType()); + Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); assert(Ty != 0 && "Invalid indices for GEP!"); - return ConstantPointerNull::get( - PointerType::get(Ty,Ptr->getAddressSpace())); + return ConstantPointerNull::get(PointerType::get(Ty, + Ptr->getAddressSpace())); } } @@ -1923,14 +2200,14 @@ Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, // getelementptr instructions into a single instruction. // if (CE->getOpcode() == Instruction::GetElementPtr) { - const Type *LastTy = 0; + Type *LastTy = 0; for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); I != E; ++I) LastTy = *I; - if ((LastTy && isa(LastTy)) || Idx0->isNullValue()) { + if ((LastTy && LastTy->isArrayTy()) || Idx0->isNullValue()) { SmallVector NewIndices; - NewIndices.reserve(NumIdx + CE->getNumOperands()); + NewIndices.reserve(Idxs.size() + CE->getNumOperands()); for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) NewIndices.push_back(CE->getOperand(i)); @@ -1939,12 +2216,11 @@ Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, Constant *Combined = CE->getOperand(CE->getNumOperands()-1); // Otherwise it must be an array. if (!Idx0->isNullValue()) { - const Type *IdxTy = Combined->getType(); + Type *IdxTy = Combined->getType(); if (IdxTy != Idx0->getType()) { - Constant *C1 = - ConstantExpr::getSExtOrBitCast(Idx0, Type::getInt64Ty(Context)); - Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, - Type::getInt64Ty(Context)); + Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); + Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); + Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); Combined = ConstantExpr::get(Instruction::Add, C1, C2); } else { Combined = @@ -1953,56 +2229,29 @@ Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, } NewIndices.push_back(Combined); - NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx); - return (inBounds && cast(CE)->isInBounds()) ? - ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0), - &NewIndices[0], - NewIndices.size()) : - ConstantExpr::getGetElementPtr(CE->getOperand(0), - &NewIndices[0], - NewIndices.size()); + NewIndices.append(Idxs.begin() + 1, Idxs.end()); + return + ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices, + inBounds && + cast(CE)->isInBounds()); } } // Implement folding of: - // int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*), - // long 0, long 0) - // To: int* getelementptr ([3 x int]* %X, long 0, long 0) + // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), + // i64 0, i64 0) + // To: i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) // - if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) { - if (const PointerType *SPT = + if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { + if (PointerType *SPT = dyn_cast(CE->getOperand(0)->getType())) - if (const ArrayType *SAT = dyn_cast(SPT->getElementType())) - if (const ArrayType *CAT = + if (ArrayType *SAT = dyn_cast(SPT->getElementType())) + if (ArrayType *CAT = dyn_cast(cast(C->getType())->getElementType())) if (CAT->getElementType() == SAT->getElementType()) - return inBounds ? - ConstantExpr::getInBoundsGetElementPtr( - (Constant*)CE->getOperand(0), Idxs, NumIdx) : - ConstantExpr::getGetElementPtr( - (Constant*)CE->getOperand(0), Idxs, NumIdx); - } - - // 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::getInt8Ty(Context)) { - 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()); + return + ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0), + Idxs, inBounds); } } @@ -2011,34 +2260,34 @@ Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, // out into preceding dimensions. bool Unknown = false; SmallVector NewIdxs; - const Type *Ty = C->getType(); - const Type *Prev = 0; - for (unsigned i = 0; i != NumIdx; + Type *Ty = C->getType(); + Type *Prev = 0; + for (unsigned i = 0, e = Idxs.size(); i != e; Prev = Ty, Ty = cast(Ty)->getTypeAtIndex(Idxs[i]), ++i) { if (ConstantInt *CI = dyn_cast(Idxs[i])) { - if (const ArrayType *ATy = dyn_cast(Ty)) + if (ArrayType *ATy = dyn_cast(Ty)) if (ATy->getNumElements() <= INT64_MAX && ATy->getNumElements() != 0 && CI->getSExtValue() >= (int64_t)ATy->getNumElements()) { if (isa(Prev)) { // It's out of range, but we can factor it into the prior // dimension. - NewIdxs.resize(NumIdx); + NewIdxs.resize(Idxs.size()); ConstantInt *Factor = ConstantInt::get(CI->getType(), ATy->getNumElements()); NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); - Constant *PrevIdx = Idxs[i-1]; + Constant *PrevIdx = cast(Idxs[i-1]); Constant *Div = ConstantExpr::getSDiv(CI, Factor); // Before adding, extend both operands to i64 to avoid // overflow trouble. - if (PrevIdx->getType() != Type::getInt64Ty(Context)) + if (!PrevIdx->getType()->isIntegerTy(64)) PrevIdx = ConstantExpr::getSExt(PrevIdx, - Type::getInt64Ty(Context)); - if (Div->getType() != Type::getInt64Ty(Context)) + Type::getInt64Ty(Div->getContext())); + if (!Div->getType()->isIntegerTy(64)) Div = ConstantExpr::getSExt(Div, - Type::getInt64Ty(Context)); + Type::getInt64Ty(Div->getContext())); NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); } else { @@ -2055,19 +2304,28 @@ Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, // If we did any factoring, start over with the adjusted indices. if (!NewIdxs.empty()) { - for (unsigned i = 0; i != NumIdx; ++i) - if (!NewIdxs[i]) NewIdxs[i] = Idxs[i]; - return inBounds ? - ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(), - NewIdxs.size()) : - ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size()); + for (unsigned i = 0, e = Idxs.size(); i != e; ++i) + if (!NewIdxs[i]) NewIdxs[i] = cast(Idxs[i]); + return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds); } // If all indices are known integers and normalized, we can do a simple // check for the "inbounds" property. if (!Unknown && !inBounds && - isa(C) && isInBoundsIndices(Idxs, NumIdx)) - return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx); + isa(C) && isInBoundsIndices(Idxs)) + return ConstantExpr::getInBoundsGetElementPtr(C, Idxs); return 0; } + +Constant *llvm::ConstantFoldGetElementPtr(Constant *C, + bool inBounds, + ArrayRef Idxs) { + return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); +} + +Constant *llvm::ConstantFoldGetElementPtr(Constant *C, + bool inBounds, + ArrayRef Idxs) { + return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); +}