X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FConstantFolding.cpp;h=08a6065b31ac7a357f53253062532dd2884c6a89;hb=491a13691d3b30b8288dfc6e01ad6a58f69a4ce6;hp=109eaad4584e4537b1737e5bf89bc33bd8cefbec;hpb=f19f9347b85279100f7d549154e37b6d726c3c94;p=oota-llvm.git diff --git a/lib/Analysis/ConstantFolding.cpp b/lib/Analysis/ConstantFolding.cpp index 109eaad4584..08a6065b31a 100644 --- a/lib/Analysis/ConstantFolding.cpp +++ b/lib/Analysis/ConstantFolding.cpp @@ -1,4 +1,4 @@ -//===-- ConstantFolding.cpp - Analyze constant folding possibilities ------===// +//===-- ConstantFolding.cpp - Fold instructions into constants ------------===// // // The LLVM Compiler Infrastructure // @@ -7,8 +7,12 @@ // //===----------------------------------------------------------------------===// // -// This family of functions determines the possibility of performing constant -// folding. +// This file defines routines for folding instructions into constants. +// +// Also, to supplement the basic VMCore ConstantExpr simplifications, +// this file defines some additional folding routines that can make use of +// TargetData information. These functions cannot go in VMCore due to library +// dependency issues. // //===----------------------------------------------------------------------===// @@ -19,13 +23,15 @@ #include "llvm/GlobalVariable.h" #include "llvm/Instructions.h" #include "llvm/Intrinsics.h" -#include "llvm/LLVMContext.h" +#include "llvm/Operator.h" +#include "llvm/Analysis/ValueTracking.h" +#include "llvm/Target/TargetData.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/StringMap.h" -#include "llvm/Target/TargetData.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/MathExtras.h" +#include "llvm/Support/FEnv.h" #include #include using namespace llvm; @@ -34,6 +40,138 @@ using namespace llvm; // Constant Folding internal helper functions //===----------------------------------------------------------------------===// +/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with +/// TargetData. This always returns a non-null constant, but it may be a +/// ConstantExpr if unfoldable. +static Constant *FoldBitCast(Constant *C, const Type *DestTy, + const TargetData &TD) { + + // This only handles casts to vectors currently. + const VectorType *DestVTy = dyn_cast(DestTy); + if (DestVTy == 0) + return ConstantExpr::getBitCast(C, DestTy); + + // If this is a scalar -> vector cast, convert the input into a <1 x scalar> + // vector so the code below can handle it uniformly. + if (isa(C) || isa(C)) { + Constant *Ops = C; // don't take the address of C! + return FoldBitCast(ConstantVector::get(Ops), DestTy, TD); + } + + // If this is a bitcast from constant vector -> vector, fold it. + ConstantVector *CV = dyn_cast(C); + if (CV == 0) + return ConstantExpr::getBitCast(C, DestTy); + + // If the element types match, VMCore can fold it. + unsigned NumDstElt = DestVTy->getNumElements(); + unsigned NumSrcElt = CV->getNumOperands(); + if (NumDstElt == NumSrcElt) + return ConstantExpr::getBitCast(C, DestTy); + + const Type *SrcEltTy = CV->getType()->getElementType(); + const Type *DstEltTy = DestVTy->getElementType(); + + // Otherwise, we're changing the number of elements in a vector, which + // requires endianness information to do the right thing. For example, + // bitcast (<2 x i64> to <4 x i32>) + // folds to (little endian): + // <4 x i32> + // and to (big endian): + // <4 x i32> + + // First thing is first. We only want to think about integer here, so if + // we have something in FP form, recast it as integer. + if (DstEltTy->isFloatingPointTy()) { + // Fold to an vector of integers with same size as our FP type. + unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); + const Type *DestIVTy = + VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt); + // Recursively handle this integer conversion, if possible. + C = FoldBitCast(C, DestIVTy, TD); + if (!C) return ConstantExpr::getBitCast(C, DestTy); + + // Finally, VMCore can handle this now that #elts line up. + return ConstantExpr::getBitCast(C, DestTy); + } + + // Okay, we know the destination is integer, if the input is FP, convert + // it to integer first. + if (SrcEltTy->isFloatingPointTy()) { + unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); + const Type *SrcIVTy = + VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt); + // Ask VMCore to do the conversion now that #elts line up. + C = ConstantExpr::getBitCast(C, SrcIVTy); + CV = dyn_cast(C); + if (!CV) // If VMCore wasn't able to fold it, bail out. + return C; + } + + // Now we know that the input and output vectors are both integer vectors + // of the same size, and that their #elements is not the same. Do the + // conversion here, which depends on whether the input or output has + // more elements. + bool isLittleEndian = TD.isLittleEndian(); + + SmallVector Result; + if (NumDstElt < NumSrcElt) { + // Handle: bitcast (<4 x i32> to <2 x i64>) + Constant *Zero = Constant::getNullValue(DstEltTy); + unsigned Ratio = NumSrcElt/NumDstElt; + unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); + unsigned SrcElt = 0; + for (unsigned i = 0; i != NumDstElt; ++i) { + // Build each element of the result. + Constant *Elt = Zero; + unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); + for (unsigned j = 0; j != Ratio; ++j) { + Constant *Src = dyn_cast(CV->getOperand(SrcElt++)); + if (!Src) // Reject constantexpr elements. + return ConstantExpr::getBitCast(C, DestTy); + + // Zero extend the element to the right size. + Src = ConstantExpr::getZExt(Src, Elt->getType()); + + // Shift it to the right place, depending on endianness. + Src = ConstantExpr::getShl(Src, + ConstantInt::get(Src->getType(), ShiftAmt)); + ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; + + // Mix it in. + Elt = ConstantExpr::getOr(Elt, Src); + } + Result.push_back(Elt); + } + } else { + // Handle: bitcast (<2 x i64> to <4 x i32>) + unsigned Ratio = NumDstElt/NumSrcElt; + unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits(); + + // Loop over each source value, expanding into multiple results. + for (unsigned i = 0; i != NumSrcElt; ++i) { + Constant *Src = dyn_cast(CV->getOperand(i)); + if (!Src) // Reject constantexpr elements. + return ConstantExpr::getBitCast(C, DestTy); + + unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); + for (unsigned j = 0; j != Ratio; ++j) { + // Shift the piece of the value into the right place, depending on + // endianness. + Constant *Elt = ConstantExpr::getLShr(Src, + ConstantInt::get(Src->getType(), ShiftAmt)); + ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; + + // Truncate and remember this piece. + Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); + } + } + } + + return ConstantVector::get(Result); +} + + /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset /// from a global, return the global and the constant. Because of /// constantexprs, this function is recursive. @@ -72,7 +210,7 @@ static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, i != e; ++i, ++GTI) { ConstantInt *CI = dyn_cast(*i); if (!CI) return false; // Index isn't a simple constant? - if (CI->getZExtValue() == 0) continue; // Not adding anything. + if (CI->isZero()) continue; // Not adding anything. if (const StructType *ST = dyn_cast(*GTI)) { // N = N + Offset @@ -88,14 +226,289 @@ static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, return false; } +/// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the +/// constant being copied out of. ByteOffset is an offset into C. CurPtr is the +/// pointer to copy results into and BytesLeft is the number of bytes left in +/// the CurPtr buffer. TD is the target data. +static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, + unsigned char *CurPtr, unsigned BytesLeft, + const TargetData &TD) { + assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) && + "Out of range access"); + + // If this element is zero or undefined, we can just return since *CurPtr is + // zero initialized. + if (isa(C) || isa(C)) + return true; + + if (ConstantInt *CI = dyn_cast(C)) { + if (CI->getBitWidth() > 64 || + (CI->getBitWidth() & 7) != 0) + return false; + + uint64_t Val = CI->getZExtValue(); + unsigned IntBytes = unsigned(CI->getBitWidth()/8); + + for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { + CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8)); + ++ByteOffset; + } + return true; + } + + if (ConstantFP *CFP = dyn_cast(C)) { + if (CFP->getType()->isDoubleTy()) { + C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD); + return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); + } + if (CFP->getType()->isFloatTy()){ + C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD); + return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); + } + return false; + } + + if (ConstantStruct *CS = dyn_cast(C)) { + const StructLayout *SL = TD.getStructLayout(CS->getType()); + unsigned Index = SL->getElementContainingOffset(ByteOffset); + uint64_t CurEltOffset = SL->getElementOffset(Index); + ByteOffset -= CurEltOffset; + + while (1) { + // If the element access is to the element itself and not to tail padding, + // read the bytes from the element. + uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType()); + + if (ByteOffset < EltSize && + !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr, + BytesLeft, TD)) + return false; + + ++Index; + + // Check to see if we read from the last struct element, if so we're done. + if (Index == CS->getType()->getNumElements()) + return true; + + // If we read all of the bytes we needed from this element we're done. + uint64_t NextEltOffset = SL->getElementOffset(Index); + + if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset) + return true; + + // Move to the next element of the struct. + CurPtr += NextEltOffset-CurEltOffset-ByteOffset; + BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset; + ByteOffset = 0; + CurEltOffset = NextEltOffset; + } + // not reached. + } + + if (ConstantArray *CA = dyn_cast(C)) { + uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType()); + uint64_t Index = ByteOffset / EltSize; + uint64_t Offset = ByteOffset - Index * EltSize; + for (; Index != CA->getType()->getNumElements(); ++Index) { + if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr, + BytesLeft, TD)) + return false; + if (EltSize >= BytesLeft) + return true; + + Offset = 0; + BytesLeft -= EltSize; + CurPtr += EltSize; + } + return true; + } + + if (ConstantVector *CV = dyn_cast(C)) { + uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType()); + uint64_t Index = ByteOffset / EltSize; + uint64_t Offset = ByteOffset - Index * EltSize; + for (; Index != CV->getType()->getNumElements(); ++Index) { + if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr, + BytesLeft, TD)) + return false; + if (EltSize >= BytesLeft) + return true; + + Offset = 0; + BytesLeft -= EltSize; + CurPtr += EltSize; + } + return true; + } + + if (ConstantExpr *CE = dyn_cast(C)) { + if (CE->getOpcode() == Instruction::IntToPtr && + CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext())) + return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr, + BytesLeft, TD); + } + + // Otherwise, unknown initializer type. + return false; +} + +static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, + const TargetData &TD) { + const Type *LoadTy = cast(C->getType())->getElementType(); + const IntegerType *IntType = dyn_cast(LoadTy); + + // If this isn't an integer load we can't fold it directly. + if (!IntType) { + // If this is a float/double load, we can try folding it as an int32/64 load + // and then bitcast the result. This can be useful for union cases. Note + // that address spaces don't matter here since we're not going to result in + // an actual new load. + const Type *MapTy; + if (LoadTy->isFloatTy()) + MapTy = Type::getInt32PtrTy(C->getContext()); + else if (LoadTy->isDoubleTy()) + MapTy = Type::getInt64PtrTy(C->getContext()); + else if (LoadTy->isVectorTy()) { + MapTy = IntegerType::get(C->getContext(), + TD.getTypeAllocSizeInBits(LoadTy)); + MapTy = PointerType::getUnqual(MapTy); + } else + return 0; + + C = FoldBitCast(C, MapTy, TD); + if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD)) + return FoldBitCast(Res, LoadTy, TD); + return 0; + } + + unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; + if (BytesLoaded > 32 || BytesLoaded == 0) return 0; + + GlobalValue *GVal; + int64_t Offset; + if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD)) + return 0; + + GlobalVariable *GV = dyn_cast(GVal); + if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || + !GV->getInitializer()->getType()->isSized()) + return 0; + + // If we're loading off the beginning of the global, some bytes may be valid, + // but we don't try to handle this. + if (Offset < 0) return 0; + + // If we're not accessing anything in this constant, the result is undefined. + if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType())) + return UndefValue::get(IntType); + + unsigned char RawBytes[32] = {0}; + if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes, + BytesLoaded, TD)) + return 0; + + APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]); + for (unsigned i = 1; i != BytesLoaded; ++i) { + ResultVal <<= 8; + ResultVal |= RawBytes[BytesLoaded-1-i]; + } + + return ConstantInt::get(IntType->getContext(), ResultVal); +} + +/// ConstantFoldLoadFromConstPtr - Return the value that a load from C would +/// produce if it is constant and determinable. If this is not determinable, +/// return null. +Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, + const TargetData *TD) { + // First, try the easy cases: + if (GlobalVariable *GV = dyn_cast(C)) + if (GV->isConstant() && GV->hasDefinitiveInitializer()) + return GV->getInitializer(); + + // If the loaded value isn't a constant expr, we can't handle it. + ConstantExpr *CE = dyn_cast(C); + if (!CE) return 0; + + if (CE->getOpcode() == Instruction::GetElementPtr) { + if (GlobalVariable *GV = dyn_cast(CE->getOperand(0))) + if (GV->isConstant() && GV->hasDefinitiveInitializer()) + if (Constant *V = + ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) + return V; + } + + // Instead of loading constant c string, use corresponding integer value + // directly if string length is small enough. + std::string Str; + if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) { + unsigned StrLen = Str.length(); + const Type *Ty = cast(CE->getType())->getElementType(); + unsigned NumBits = Ty->getPrimitiveSizeInBits(); + // Replace load with immediate integer if the result is an integer or fp + // value. + if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 && + (isa(Ty) || Ty->isFloatingPointTy())) { + APInt StrVal(NumBits, 0); + APInt SingleChar(NumBits, 0); + if (TD->isLittleEndian()) { + for (signed i = StrLen-1; i >= 0; i--) { + SingleChar = (uint64_t) Str[i] & UCHAR_MAX; + StrVal = (StrVal << 8) | SingleChar; + } + } else { + for (unsigned i = 0; i < StrLen; i++) { + SingleChar = (uint64_t) Str[i] & UCHAR_MAX; + StrVal = (StrVal << 8) | SingleChar; + } + // Append NULL at the end. + SingleChar = 0; + StrVal = (StrVal << 8) | SingleChar; + } + + Constant *Res = ConstantInt::get(CE->getContext(), StrVal); + if (Ty->isFloatingPointTy()) + Res = ConstantExpr::getBitCast(Res, Ty); + return Res; + } + } + + // If this load comes from anywhere in a constant global, and if the global + // is all undef or zero, we know what it loads. + if (GlobalVariable *GV = + dyn_cast(GetUnderlyingObject(CE, TD))) { + if (GV->isConstant() && GV->hasDefinitiveInitializer()) { + const Type *ResTy = cast(C->getType())->getElementType(); + if (GV->getInitializer()->isNullValue()) + return Constant::getNullValue(ResTy); + if (isa(GV->getInitializer())) + return UndefValue::get(ResTy); + } + } + + // Try hard to fold loads from bitcasted strange and non-type-safe things. We + // currently don't do any of this for big endian systems. It can be + // generalized in the future if someone is interested. + if (TD && TD->isLittleEndian()) + return FoldReinterpretLoadFromConstPtr(CE, *TD); + return 0; +} + +static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){ + if (LI->isVolatile()) return 0; + + if (Constant *C = dyn_cast(LI->getOperand(0))) + return ConstantFoldLoadFromConstPtr(C, TD); + + return 0; +} /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression. /// Attempt to symbolically evaluate the result of a binary operator merging /// these together. If target data info is available, it is provided as TD, /// otherwise TD is null. static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, - Constant *Op1, const TargetData *TD, - LLVMContext &Context){ + Constant *Op1, const TargetData *TD){ // SROA // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. @@ -120,46 +533,115 @@ static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, return 0; } +/// CastGEPIndices - If array indices are not pointer-sized integers, +/// explicitly cast them so that they aren't implicitly casted by the +/// getelementptr. +static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps, + const Type *ResultTy, + const TargetData *TD) { + if (!TD) return 0; + const Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext()); + + bool Any = false; + SmallVector NewIdxs; + for (unsigned i = 1; i != NumOps; ++i) { + if ((i == 1 || + !isa(GetElementPtrInst::getIndexedType(Ops[0]->getType(), + reinterpret_cast(Ops+1), + i-1))) && + Ops[i]->getType() != IntPtrTy) { + Any = true; + NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i], + true, + IntPtrTy, + true), + Ops[i], IntPtrTy)); + } else + NewIdxs.push_back(Ops[i]); + } + if (!Any) return 0; + + Constant *C = + ConstantExpr::getGetElementPtr(Ops[0], &NewIdxs[0], NewIdxs.size()); + if (ConstantExpr *CE = dyn_cast(C)) + if (Constant *Folded = ConstantFoldConstantExpression(CE, TD)) + C = Folded; + return C; +} + /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP /// constant expression, do so. -static Constant *SymbolicallyEvaluateGEP(Constant* const* Ops, unsigned NumOps, +static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, const Type *ResultTy, - LLVMContext &Context, const TargetData *TD) { Constant *Ptr = Ops[0]; if (!TD || !cast(Ptr->getType())->getElementType()->isSized()) return 0; - - unsigned BitWidth = TD->getTypeSizeInBits(TD->getIntPtrType(Context)); - APInt BasePtr(BitWidth, 0); - bool BaseIsInt = true; - if (!Ptr->isNullValue()) { - // If this is a inttoptr from a constant int, we can fold this as the base, - // otherwise we can't. - if (ConstantExpr *CE = dyn_cast(Ptr)) - if (CE->getOpcode() == Instruction::IntToPtr) - if (ConstantInt *Base = dyn_cast(CE->getOperand(0))) { - BasePtr = Base->getValue(); - BasePtr.zextOrTrunc(BitWidth); - } - - if (BasePtr == 0) - BaseIsInt = false; - } + + const Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext()); // If this is a constant expr gep that is effectively computing an // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12' for (unsigned i = 1; i != NumOps; ++i) - if (!isa(Ops[i])) + if (!isa(Ops[i])) { + + // If this is "gep i8* Ptr, (sub 0, V)", fold this as: + // "inttoptr (sub (ptrtoint Ptr), V)" + if (NumOps == 2 && + cast(ResultTy)->getElementType()->isIntegerTy(8)) { + ConstantExpr *CE = dyn_cast(Ops[1]); + assert((CE == 0 || CE->getType() == IntPtrTy) && + "CastGEPIndices didn't canonicalize index types!"); + if (CE && CE->getOpcode() == Instruction::Sub && + CE->getOperand(0)->isNullValue()) { + Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType()); + Res = ConstantExpr::getSub(Res, CE->getOperand(1)); + Res = ConstantExpr::getIntToPtr(Res, ResultTy); + if (ConstantExpr *ResCE = dyn_cast(Res)) + Res = ConstantFoldConstantExpression(ResCE, TD); + return Res; + } + } return 0; + } + unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy); APInt Offset = APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(), (Value**)Ops+1, NumOps-1)); + Ptr = cast(Ptr->stripPointerCasts()); + + // If this is a GEP of a GEP, fold it all into a single GEP. + while (GEPOperator *GEP = dyn_cast(Ptr)) { + SmallVector NestedOps(GEP->op_begin()+1, GEP->op_end()); + + // Do not try the incorporate the sub-GEP if some index is not a number. + bool AllConstantInt = true; + for (unsigned i = 0, e = NestedOps.size(); i != e; ++i) + if (!isa(NestedOps[i])) { + AllConstantInt = false; + break; + } + if (!AllConstantInt) + break; + + Ptr = cast(GEP->getOperand(0)); + Offset += APInt(BitWidth, + TD->getIndexedOffset(Ptr->getType(), + (Value**)NestedOps.data(), + NestedOps.size())); + Ptr = cast(Ptr->stripPointerCasts()); + } + // If the base value for this address is a literal integer value, fold the // getelementptr to the resulting integer value casted to the pointer type. - if (BaseIsInt) { - Constant *C = ConstantInt::get(Context, Offset+BasePtr); + APInt BasePtr(BitWidth, 0); + if (ConstantExpr *CE = dyn_cast(Ptr)) + if (CE->getOpcode() == Instruction::IntToPtr) + if (ConstantInt *Base = dyn_cast(CE->getOperand(0))) + BasePtr = Base->getValue().zextOrTrunc(BitWidth); + if (Ptr->isNullValue() || BasePtr != 0) { + Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr); return ConstantExpr::getIntToPtr(C, ResultTy); } @@ -171,16 +653,31 @@ static Constant *SymbolicallyEvaluateGEP(Constant* const* Ops, unsigned NumOps, SmallVector NewIdxs; do { if (const SequentialType *ATy = dyn_cast(Ty)) { - // The only pointer indexing we'll do is on the first index of the GEP. - if (isa(ATy) && !NewIdxs.empty()) - break; + if (ATy->isPointerTy()) { + // The only pointer indexing we'll do is on the first index of the GEP. + if (!NewIdxs.empty()) + break; + + // Only handle pointers to sized types, not pointers to functions. + if (!ATy->getElementType()->isSized()) + return 0; + } + // Determine which element of the array the offset points into. APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType())); + const IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext()); if (ElemSize == 0) - return 0; - APInt NewIdx = Offset.udiv(ElemSize); - Offset -= NewIdx * ElemSize; - NewIdxs.push_back(ConstantInt::get(TD->getIntPtrType(Context), NewIdx)); + // The element size is 0. This may be [0 x Ty]*, so just use a zero + // index for this level and proceed to the next level to see if it can + // accommodate the offset. + NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0)); + else { + // The element size is non-zero divide the offset by the element + // size (rounding down), to compute the index at this level. + APInt NewIdx = Offset.udiv(ElemSize); + Offset -= NewIdx * ElemSize; + NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx)); + } Ty = ATy->getElementType(); } else if (const StructType *STy = dyn_cast(Ty)) { // Determine which field of the struct the offset points into. The @@ -188,7 +685,8 @@ static Constant *SymbolicallyEvaluateGEP(Constant* const* Ops, unsigned NumOps, // know the offset is within the struct at this point. const StructLayout &SL = *TD->getStructLayout(STy); unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); - NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Context), ElIdx)); + NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), + ElIdx)); Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx)); Ty = STy->getTypeAtIndex(ElIdx); } else { @@ -203,12 +701,8 @@ static Constant *SymbolicallyEvaluateGEP(Constant* const* Ops, unsigned NumOps, if (Offset != 0) return 0; - // If the base is the start of a GlobalVariable and all the array indices - // remain in their static bounds, the GEP is inbounds. We can check that - // all indices are in bounds by just checking the first index only - // because we've just normalized all the indices. - Constant *C = isa(Ptr) && NewIdxs[0]->isNullValue() ? - ConstantExpr::getInBoundsGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size()) : + // Create a GEP. + Constant *C = ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size()); assert(cast(C->getType())->getElementType() == Ty && "Computed GetElementPtr has unexpected type!"); @@ -216,154 +710,45 @@ static Constant *SymbolicallyEvaluateGEP(Constant* const* Ops, unsigned NumOps, // If we ended up indexing a member with a type that doesn't match // the type of what the original indices indexed, add a cast. if (Ty != cast(ResultTy)->getElementType()) - C = ConstantExpr::getBitCast(C, ResultTy); + C = FoldBitCast(C, ResultTy, *TD); return C; } -/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with -/// targetdata. Return 0 if unfoldable. -static Constant *FoldBitCast(Constant *C, const Type *DestTy, - const TargetData &TD, LLVMContext &Context) { - // If this is a bitcast from constant vector -> vector, fold it. - if (ConstantVector *CV = dyn_cast(C)) { - if (const VectorType *DestVTy = dyn_cast(DestTy)) { - // If the element types match, VMCore can fold it. - unsigned NumDstElt = DestVTy->getNumElements(); - unsigned NumSrcElt = CV->getNumOperands(); - if (NumDstElt == NumSrcElt) - return 0; - - const Type *SrcEltTy = CV->getType()->getElementType(); - const Type *DstEltTy = DestVTy->getElementType(); - - // Otherwise, we're changing the number of elements in a vector, which - // requires endianness information to do the right thing. For example, - // bitcast (<2 x i64> to <4 x i32>) - // folds to (little endian): - // <4 x i32> - // and to (big endian): - // <4 x i32> - - // First thing is first. We only want to think about integer here, so if - // we have something in FP form, recast it as integer. - if (DstEltTy->isFloatingPoint()) { - // Fold to an vector of integers with same size as our FP type. - unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); - const Type *DestIVTy = VectorType::get( - IntegerType::get(Context, FPWidth), NumDstElt); - // Recursively handle this integer conversion, if possible. - C = FoldBitCast(C, DestIVTy, TD, Context); - if (!C) return 0; - - // Finally, VMCore can handle this now that #elts line up. - return ConstantExpr::getBitCast(C, DestTy); - } - - // Okay, we know the destination is integer, if the input is FP, convert - // it to integer first. - if (SrcEltTy->isFloatingPoint()) { - unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); - const Type *SrcIVTy = VectorType::get( - IntegerType::get(Context, FPWidth), NumSrcElt); - // Ask VMCore to do the conversion now that #elts line up. - C = ConstantExpr::getBitCast(C, SrcIVTy); - CV = dyn_cast(C); - if (!CV) return 0; // If VMCore wasn't able to fold it, bail out. - } - - // Now we know that the input and output vectors are both integer vectors - // of the same size, and that their #elements is not the same. Do the - // conversion here, which depends on whether the input or output has - // more elements. - bool isLittleEndian = TD.isLittleEndian(); - - SmallVector Result; - if (NumDstElt < NumSrcElt) { - // Handle: bitcast (<4 x i32> to <2 x i64>) - Constant *Zero = Constant::getNullValue(DstEltTy); - unsigned Ratio = NumSrcElt/NumDstElt; - unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); - unsigned SrcElt = 0; - for (unsigned i = 0; i != NumDstElt; ++i) { - // Build each element of the result. - Constant *Elt = Zero; - unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); - for (unsigned j = 0; j != Ratio; ++j) { - Constant *Src = dyn_cast(CV->getOperand(SrcElt++)); - if (!Src) return 0; // Reject constantexpr elements. - - // Zero extend the element to the right size. - Src = ConstantExpr::getZExt(Src, Elt->getType()); - - // Shift it to the right place, depending on endianness. - Src = ConstantExpr::getShl(Src, - ConstantInt::get(Src->getType(), ShiftAmt)); - ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; - - // Mix it in. - Elt = ConstantExpr::getOr(Elt, Src); - } - Result.push_back(Elt); - } - } else { - // Handle: bitcast (<2 x i64> to <4 x i32>) - unsigned Ratio = NumDstElt/NumSrcElt; - unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits(); - - // Loop over each source value, expanding into multiple results. - for (unsigned i = 0; i != NumSrcElt; ++i) { - Constant *Src = dyn_cast(CV->getOperand(i)); - if (!Src) return 0; // Reject constantexpr elements. - - unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); - for (unsigned j = 0; j != Ratio; ++j) { - // Shift the piece of the value into the right place, depending on - // endianness. - Constant *Elt = ConstantExpr::getLShr(Src, - ConstantInt::get(Src->getType(), ShiftAmt)); - ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; - - // Truncate and remember this piece. - Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); - } - } - } - - return ConstantVector::get(Result.data(), Result.size()); - } - } - - return 0; -} //===----------------------------------------------------------------------===// // Constant Folding public APIs //===----------------------------------------------------------------------===// - -/// ConstantFoldInstruction - Attempt to constant fold the specified -/// instruction. If successful, the constant result is returned, if not, null -/// is returned. Note that this function can only fail when attempting to fold -/// instructions like loads and stores, which have no constant expression form. -/// -Constant *llvm::ConstantFoldInstruction(Instruction *I, LLVMContext &Context, - const TargetData *TD) { +/// ConstantFoldInstruction - Try to constant fold the specified instruction. +/// If successful, the constant result is returned, if not, null is returned. +/// Note that this fails if not all of the operands are constant. Otherwise, +/// this function can only fail when attempting to fold instructions like loads +/// and stores, which have no constant expression form. +Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { + // Handle PHI nodes quickly here... if (PHINode *PN = dyn_cast(I)) { - if (PN->getNumIncomingValues() == 0) - return UndefValue::get(PN->getType()); + Constant *CommonValue = 0; - Constant *Result = dyn_cast(PN->getIncomingValue(0)); - if (Result == 0) return 0; - - // Handle PHI nodes specially here... - for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) - if (PN->getIncomingValue(i) != Result && PN->getIncomingValue(i) != PN) - return 0; // Not all the same incoming constants... + for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { + Value *Incoming = PN->getIncomingValue(i); + // If the incoming value is undef then skip it. Note that while we could + // skip the value if it is equal to the phi node itself we choose not to + // because that would break the rule that constant folding only applies if + // all operands are constants. + if (isa(Incoming)) + continue; + // If the incoming value is not a constant, or is a different constant to + // the one we saw previously, then give up. + Constant *C = dyn_cast(Incoming); + if (!C || (CommonValue && C != CommonValue)) + return 0; + CommonValue = C; + } - // If we reach here, all incoming values are the same constant. - return Result; + // If we reach here, all incoming values are the same constant or undef. + return CommonValue ? CommonValue : UndefValue::get(PN->getType()); } // Scan the operand list, checking to see if they are all constants, if so, @@ -376,31 +761,47 @@ Constant *llvm::ConstantFoldInstruction(Instruction *I, LLVMContext &Context, return 0; // All operands not constant! if (const CmpInst *CI = dyn_cast(I)) - return ConstantFoldCompareInstOperands(CI->getPredicate(), - Ops.data(), Ops.size(), - Context, TD); - else - return ConstantFoldInstOperands(I->getOpcode(), I->getType(), - Ops.data(), Ops.size(), Context, TD); + return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], + TD); + + if (const LoadInst *LI = dyn_cast(I)) + return ConstantFoldLoadInst(LI, TD); + + if (InsertValueInst *IVI = dyn_cast(I)) + return ConstantExpr::getInsertValue( + cast(IVI->getAggregateOperand()), + cast(IVI->getInsertedValueOperand()), + IVI->idx_begin(), IVI->getNumIndices()); + + if (ExtractValueInst *EVI = dyn_cast(I)) + return ConstantExpr::getExtractValue( + cast(EVI->getAggregateOperand()), + EVI->idx_begin(), EVI->getNumIndices()); + + return ConstantFoldInstOperands(I->getOpcode(), I->getType(), + Ops.data(), Ops.size(), TD); } /// ConstantFoldConstantExpression - Attempt to fold the constant expression /// using the specified TargetData. If successful, the constant result is /// result is returned, if not, null is returned. -Constant *llvm::ConstantFoldConstantExpression(ConstantExpr *CE, - LLVMContext &Context, +Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, const TargetData *TD) { SmallVector Ops; - for (User::op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; ++i) - Ops.push_back(cast(*i)); + for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); + i != e; ++i) { + Constant *NewC = cast(*i); + // Recursively fold the ConstantExpr's operands. + if (ConstantExpr *NewCE = dyn_cast(NewC)) + NewC = ConstantFoldConstantExpression(NewCE, TD); + Ops.push_back(NewC); + } if (CE->isCompare()) - return ConstantFoldCompareInstOperands(CE->getPredicate(), - Ops.data(), Ops.size(), - Context, TD); - else - return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), - Ops.data(), Ops.size(), Context, TD); + return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], + TD); + return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), + Ops.data(), Ops.size(), TD); } /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the @@ -409,15 +810,17 @@ Constant *llvm::ConstantFoldConstantExpression(ConstantExpr *CE, /// attempting to fold instructions like loads and stores, which have no /// constant expression form. /// +/// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc +/// information, due to only being passed an opcode and operands. Constant +/// folding using this function strips this information. +/// Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, Constant* const* Ops, unsigned NumOps, - LLVMContext &Context, const TargetData *TD) { // Handle easy binops first. if (Instruction::isBinaryOp(Opcode)) { if (isa(Ops[0]) || isa(Ops[1])) - if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD, - Context)) + if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD)) return C; return ConstantExpr::get(Opcode, Ops[0], Ops[1]); @@ -425,14 +828,13 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, switch (Opcode) { default: return 0; + case Instruction::ICmp: + case Instruction::FCmp: assert(0 && "Invalid for compares"); case Instruction::Call: - if (Function *F = dyn_cast(Ops[0])) + if (Function *F = dyn_cast(Ops[NumOps - 1])) if (canConstantFoldCallTo(F)) - return ConstantFoldCall(F, Ops+1, NumOps-1); + return ConstantFoldCall(F, Ops, NumOps - 1); return 0; - case Instruction::ICmp: - case Instruction::FCmp: - llvm_unreachable("This function is invalid for compares: no predicate specified"); case Instruction::PtrToInt: // If the input is a inttoptr, eliminate the pair. This requires knowing // the width of a pointer, so it can't be done in ConstantExpr::getCast. @@ -442,7 +844,7 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, unsigned InWidth = Input->getType()->getScalarSizeInBits(); if (TD->getPointerSizeInBits() < InWidth) { Constant *Mask = - ConstantInt::get(Context, APInt::getLowBitsSet(InWidth, + ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, TD->getPointerSizeInBits())); Input = ConstantExpr::getAnd(Input, Mask); } @@ -455,47 +857,12 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if // the int size is >= the ptr size. This requires knowing the width of a // pointer, so it can't be done in ConstantExpr::getCast. - if (ConstantExpr *CE = dyn_cast(Ops[0])) { + if (ConstantExpr *CE = dyn_cast(Ops[0])) if (TD && - TD->getPointerSizeInBits() <= - CE->getType()->getScalarSizeInBits()) { - if (CE->getOpcode() == Instruction::PtrToInt) { - Constant *Input = CE->getOperand(0); - Constant *C = FoldBitCast(Input, DestTy, *TD, Context); - return C ? C : ConstantExpr::getBitCast(Input, DestTy); - } - // If there's a constant offset added to the integer value before - // it is casted back to a pointer, see if the expression can be - // converted into a GEP. - if (CE->getOpcode() == Instruction::Add) - if (ConstantInt *L = dyn_cast(CE->getOperand(0))) - if (ConstantExpr *R = dyn_cast(CE->getOperand(1))) - if (R->getOpcode() == Instruction::PtrToInt) - if (GlobalVariable *GV = - dyn_cast(R->getOperand(0))) { - const PointerType *GVTy = cast(GV->getType()); - if (const ArrayType *AT = - dyn_cast(GVTy->getElementType())) { - const Type *ElTy = AT->getElementType(); - uint64_t AllocSize = TD->getTypeAllocSize(ElTy); - APInt PSA(L->getValue().getBitWidth(), AllocSize); - if (ElTy == cast(DestTy)->getElementType() && - L->getValue().urem(PSA) == 0) { - APInt ElemIdx = L->getValue().udiv(PSA); - if (ElemIdx.ult(APInt(ElemIdx.getBitWidth(), - AT->getNumElements()))) { - Constant *Index[] = { - Constant::getNullValue(CE->getType()), - ConstantInt::get(Context, ElemIdx) - }; - return - ConstantExpr::getGetElementPtr(GV, &Index[0], 2); - } - } - } - } - } - } + TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() && + CE->getOpcode() == Instruction::PtrToInt) + return FoldBitCast(CE->getOperand(0), DestTy, *TD); + return ConstantExpr::getCast(Opcode, Ops[0], DestTy); case Instruction::Trunc: case Instruction::ZExt: @@ -509,8 +876,7 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, return ConstantExpr::getCast(Opcode, Ops[0], DestTy); case Instruction::BitCast: if (TD) - if (Constant *C = FoldBitCast(Ops[0], DestTy, *TD, Context)) - return C; + return FoldBitCast(Ops[0], DestTy, *TD); return ConstantExpr::getBitCast(Ops[0], DestTy); case Instruction::Select: return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]); @@ -521,7 +887,9 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, case Instruction::ShuffleVector: return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]); case Instruction::GetElementPtr: - if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, Context, TD)) + if (Constant *C = CastGEPIndices(Ops, NumOps, DestTy, TD)) + return C; + if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD)) return C; return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1); @@ -533,9 +901,7 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, /// returns a constant expression of the specified operands. /// Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, - Constant*const * Ops, - unsigned NumOps, - LLVMContext &Context, + Constant *Ops0, Constant *Ops1, const TargetData *TD) { // fold: icmp (inttoptr x), null -> icmp x, 0 // fold: icmp (ptrtoint x), 0 -> icmp x, null @@ -544,17 +910,16 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, // // ConstantExpr::getCompare cannot do this, because it doesn't have TD // around to know if bit truncation is happening. - if (ConstantExpr *CE0 = dyn_cast(Ops[0])) { - if (TD && Ops[1]->isNullValue()) { - const Type *IntPtrTy = TD->getIntPtrType(Context); + if (ConstantExpr *CE0 = dyn_cast(Ops0)) { + if (TD && Ops1->isNullValue()) { + const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); if (CE0->getOpcode() == Instruction::IntToPtr) { // Convert the integer value to the right size to ensure we get the // proper extension or truncation. Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0), IntPtrTy, false); - Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) }; - return ConstantFoldCompareInstOperands(Predicate, NewOps, 2, - Context, TD); + Constant *Null = Constant::getNullValue(C->getType()); + return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); } // Only do this transformation if the int is intptrty in size, otherwise @@ -562,16 +927,14 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, if (CE0->getOpcode() == Instruction::PtrToInt && CE0->getType() == IntPtrTy) { Constant *C = CE0->getOperand(0); - Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) }; - // FIXME! - return ConstantFoldCompareInstOperands(Predicate, NewOps, 2, - Context, TD); + Constant *Null = Constant::getNullValue(C->getType()); + return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); } } - if (ConstantExpr *CE1 = dyn_cast(Ops[1])) { + if (ConstantExpr *CE1 = dyn_cast(Ops1)) { if (TD && CE0->getOpcode() == CE1->getOpcode()) { - const Type *IntPtrTy = TD->getIntPtrType(Context); + const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); if (CE0->getOpcode() == Instruction::IntToPtr) { // Convert the integer value to the right size to ensure we get the @@ -580,26 +943,35 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, IntPtrTy, false); Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), IntPtrTy, false); - Constant *NewOps[] = { C0, C1 }; - return ConstantFoldCompareInstOperands(Predicate, NewOps, 2, - Context, TD); + return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD); } // Only do this transformation if the int is intptrty in size, otherwise // there is a truncation or extension that we aren't modeling. if ((CE0->getOpcode() == Instruction::PtrToInt && CE0->getType() == IntPtrTy && - CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) { - Constant *NewOps[] = { - CE0->getOperand(0), CE1->getOperand(0) - }; - return ConstantFoldCompareInstOperands(Predicate, NewOps, 2, - Context, TD); - } + CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) + return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), + CE1->getOperand(0), TD); } } + + // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0) + // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0) + if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && + CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { + Constant *LHS = + ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD); + Constant *RHS = + ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD); + unsigned OpC = + Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; + Constant *Ops[] = { LHS, RHS }; + return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, 2, TD); + } } - return ConstantExpr::getCompare(Predicate, Ops[0], Ops[1]); + + return ConstantExpr::getCompare(Predicate, Ops0, Ops1); } @@ -607,8 +979,7 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, /// getelementptr constantexpr, return the constant value being addressed by the /// constant expression, or null if something is funny and we can't decide. Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, - ConstantExpr *CE, - LLVMContext &Context) { + ConstantExpr *CE) { if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType())) return 0; // Do not allow stepping over the value! @@ -642,15 +1013,15 @@ Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, C = UndefValue::get(ATy->getElementType()); else return 0; - } else if (const VectorType *PTy = dyn_cast(*I)) { - if (CI->getZExtValue() >= PTy->getNumElements()) + } else if (const VectorType *VTy = dyn_cast(*I)) { + if (CI->getZExtValue() >= VTy->getNumElements()) return 0; if (ConstantVector *CP = dyn_cast(C)) C = CP->getOperand(CI->getZExtValue()); else if (isa(C)) - C = Constant::getNullValue(PTy->getElementType()); + C = Constant::getNullValue(VTy->getElementType()); else if (isa(C)) - C = UndefValue::get(PTy->getElementType()); + C = UndefValue::get(VTy->getElementType()); else return 0; } else { @@ -678,8 +1049,26 @@ llvm::canConstantFoldCallTo(const Function *F) { case Intrinsic::ctpop: case Intrinsic::ctlz: case Intrinsic::cttz: + case Intrinsic::sadd_with_overflow: + case Intrinsic::uadd_with_overflow: + case Intrinsic::ssub_with_overflow: + case Intrinsic::usub_with_overflow: + case Intrinsic::smul_with_overflow: + case Intrinsic::umul_with_overflow: + case Intrinsic::convert_from_fp16: + case Intrinsic::convert_to_fp16: + case Intrinsic::x86_sse_cvtss2si: + case Intrinsic::x86_sse_cvtss2si64: + case Intrinsic::x86_sse_cvttss2si: + case Intrinsic::x86_sse_cvttss2si64: + case Intrinsic::x86_sse2_cvtsd2si: + case Intrinsic::x86_sse2_cvtsd2si64: + case Intrinsic::x86_sse2_cvttsd2si: + case Intrinsic::x86_sse2_cvttsd2si64: return true; - default: break; + default: + return false; + case 0: break; } if (!F->hasName()) return false; @@ -696,7 +1085,7 @@ llvm::canConstantFoldCallTo(const Function *F) { case 'c': return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh"; case 'e': - return Name == "exp"; + return Name == "exp" || Name == "exp2"; case 'f': return Name == "fabs" || Name == "fmod" || Name == "floor"; case 'l': @@ -712,168 +1101,312 @@ llvm::canConstantFoldCallTo(const Function *F) { } static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, - const Type *Ty, LLVMContext &Context) { - errno = 0; + const Type *Ty) { + sys::llvm_fenv_clearexcept(); V = NativeFP(V); - if (errno != 0) { - errno = 0; + if (sys::llvm_fenv_testexcept()) { + sys::llvm_fenv_clearexcept(); return 0; } - if (Ty == Type::getFloatTy(Context)) - return ConstantFP::get(Context, APFloat((float)V)); - if (Ty == Type::getDoubleTy(Context)) - return ConstantFP::get(Context, APFloat(V)); + if (Ty->isFloatTy()) + return ConstantFP::get(Ty->getContext(), APFloat((float)V)); + if (Ty->isDoubleTy()) + return ConstantFP::get(Ty->getContext(), APFloat(V)); llvm_unreachable("Can only constant fold float/double"); return 0; // dummy return to suppress warning } static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), - double V, double W, - const Type *Ty, - LLVMContext &Context) { - errno = 0; + double V, double W, const Type *Ty) { + sys::llvm_fenv_clearexcept(); V = NativeFP(V, W); - if (errno != 0) { - errno = 0; + if (sys::llvm_fenv_testexcept()) { + sys::llvm_fenv_clearexcept(); return 0; } - if (Ty == Type::getFloatTy(Context)) - return ConstantFP::get(Context, APFloat((float)V)); - if (Ty == Type::getDoubleTy(Context)) - return ConstantFP::get(Context, APFloat(V)); + if (Ty->isFloatTy()) + return ConstantFP::get(Ty->getContext(), APFloat((float)V)); + if (Ty->isDoubleTy()) + return ConstantFP::get(Ty->getContext(), APFloat(V)); llvm_unreachable("Can only constant fold float/double"); return 0; // dummy return to suppress warning } +/// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer +/// conversion of a constant floating point. If roundTowardZero is false, the +/// default IEEE rounding is used (toward nearest, ties to even). This matches +/// the behavior of the non-truncating SSE instructions in the default rounding +/// mode. The desired integer type Ty is used to select how many bits are +/// available for the result. Returns null if the conversion cannot be +/// performed, otherwise returns the Constant value resulting from the +/// conversion. +static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero, + const Type *Ty) { + assert(Op && "Called with NULL operand"); + APFloat Val(Op->getValueAPF()); + + // All of these conversion intrinsics form an integer of at most 64bits. + unsigned ResultWidth = cast(Ty)->getBitWidth(); + assert(ResultWidth <= 64 && + "Can only constant fold conversions to 64 and 32 bit ints"); + + uint64_t UIntVal; + bool isExact = false; + APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero + : APFloat::rmNearestTiesToEven; + APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth, + /*isSigned=*/true, mode, + &isExact); + if (status != APFloat::opOK && status != APFloat::opInexact) + return 0; + return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true); +} + /// ConstantFoldCall - Attempt to constant fold a call to the specified function /// with the specified arguments, returning null if unsuccessful. - Constant * llvm::ConstantFoldCall(Function *F, - Constant* const* Operands, unsigned NumOperands) { + Constant *const *Operands, unsigned NumOperands) { if (!F->hasName()) return 0; - LLVMContext &Context = F->getContext(); StringRef Name = F->getName(); - + const Type *Ty = F->getReturnType(); if (NumOperands == 1) { if (ConstantFP *Op = dyn_cast(Operands[0])) { - if (Ty!=Type::getFloatTy(F->getContext()) && - Ty!=Type::getDoubleTy(Context)) + if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) { + APFloat Val(Op->getValueAPF()); + + bool lost = false; + Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost); + + return ConstantInt::get(F->getContext(), Val.bitcastToAPInt()); + } + + if (!Ty->isFloatTy() && !Ty->isDoubleTy()) + return 0; + + /// We only fold functions with finite arguments. Folding NaN and inf is + /// likely to be aborted with an exception anyway, and some host libms + /// have known errors raising exceptions. + if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity()) return 0; + /// Currently APFloat versions of these functions do not exist, so we use /// the host native double versions. Float versions are not called /// directly but for all these it is true (float)(f((double)arg)) == /// f(arg). Long double not supported yet. - double V = Ty==Type::getFloatTy(F->getContext()) ? - (double)Op->getValueAPF().convertToFloat(): + double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() : Op->getValueAPF().convertToDouble(); switch (Name[0]) { case 'a': if (Name == "acos") - return ConstantFoldFP(acos, V, Ty, Context); + return ConstantFoldFP(acos, V, Ty); else if (Name == "asin") - return ConstantFoldFP(asin, V, Ty, Context); + return ConstantFoldFP(asin, V, Ty); else if (Name == "atan") - return ConstantFoldFP(atan, V, Ty, Context); + return ConstantFoldFP(atan, V, Ty); break; case 'c': if (Name == "ceil") - return ConstantFoldFP(ceil, V, Ty, Context); + return ConstantFoldFP(ceil, V, Ty); else if (Name == "cos") - return ConstantFoldFP(cos, V, Ty, Context); + return ConstantFoldFP(cos, V, Ty); else if (Name == "cosh") - return ConstantFoldFP(cosh, V, Ty, Context); + return ConstantFoldFP(cosh, V, Ty); else if (Name == "cosf") - return ConstantFoldFP(cos, V, Ty, Context); + return ConstantFoldFP(cos, V, Ty); break; case 'e': if (Name == "exp") - return ConstantFoldFP(exp, V, Ty, Context); + return ConstantFoldFP(exp, V, Ty); + + if (Name == "exp2") { + // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a + // C99 library. + return ConstantFoldBinaryFP(pow, 2.0, V, Ty); + } break; case 'f': if (Name == "fabs") - return ConstantFoldFP(fabs, V, Ty, Context); + return ConstantFoldFP(fabs, V, Ty); else if (Name == "floor") - return ConstantFoldFP(floor, V, Ty, Context); + return ConstantFoldFP(floor, V, Ty); break; case 'l': if (Name == "log" && V > 0) - return ConstantFoldFP(log, V, Ty, Context); + return ConstantFoldFP(log, V, Ty); else if (Name == "log10" && V > 0) - return ConstantFoldFP(log10, V, Ty, Context); - else if (Name == "llvm.sqrt.f32" || - Name == "llvm.sqrt.f64") { + return ConstantFoldFP(log10, V, Ty); + else if (F->getIntrinsicID() == Intrinsic::sqrt && + (Ty->isFloatTy() || Ty->isDoubleTy())) { if (V >= -0.0) - return ConstantFoldFP(sqrt, V, Ty, Context); + return ConstantFoldFP(sqrt, V, Ty); else // Undefined return Constant::getNullValue(Ty); } break; case 's': if (Name == "sin") - return ConstantFoldFP(sin, V, Ty, Context); + return ConstantFoldFP(sin, V, Ty); else if (Name == "sinh") - return ConstantFoldFP(sinh, V, Ty, Context); + return ConstantFoldFP(sinh, V, Ty); else if (Name == "sqrt" && V >= 0) - return ConstantFoldFP(sqrt, V, Ty, Context); + return ConstantFoldFP(sqrt, V, Ty); else if (Name == "sqrtf" && V >= 0) - return ConstantFoldFP(sqrt, V, Ty, Context); + return ConstantFoldFP(sqrt, V, Ty); else if (Name == "sinf") - return ConstantFoldFP(sin, V, Ty, Context); + return ConstantFoldFP(sin, V, Ty); break; case 't': if (Name == "tan") - return ConstantFoldFP(tan, V, Ty, Context); + return ConstantFoldFP(tan, V, Ty); else if (Name == "tanh") - return ConstantFoldFP(tanh, V, Ty, Context); + return ConstantFoldFP(tanh, V, Ty); break; default: break; } - } else if (ConstantInt *Op = dyn_cast(Operands[0])) { - if (Name.startswith("llvm.bswap")) - return ConstantInt::get(Context, Op->getValue().byteSwap()); - else if (Name.startswith("llvm.ctpop")) + return 0; + } + + if (ConstantInt *Op = dyn_cast(Operands[0])) { + switch (F->getIntrinsicID()) { + case Intrinsic::bswap: + return ConstantInt::get(F->getContext(), Op->getValue().byteSwap()); + case Intrinsic::ctpop: return ConstantInt::get(Ty, Op->getValue().countPopulation()); - else if (Name.startswith("llvm.cttz")) + case Intrinsic::cttz: return ConstantInt::get(Ty, Op->getValue().countTrailingZeros()); - else if (Name.startswith("llvm.ctlz")) + case Intrinsic::ctlz: return ConstantInt::get(Ty, Op->getValue().countLeadingZeros()); + case Intrinsic::convert_from_fp16: { + APFloat Val(Op->getValue()); + + bool lost = false; + APFloat::opStatus status = + Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost); + + // Conversion is always precise. + (void)status; + assert(status == APFloat::opOK && !lost && + "Precision lost during fp16 constfolding"); + + return ConstantFP::get(F->getContext(), Val); + } + default: + return 0; + } + } + + if (ConstantVector *Op = dyn_cast(Operands[0])) { + switch (F->getIntrinsicID()) { + default: break; + case Intrinsic::x86_sse_cvtss2si: + case Intrinsic::x86_sse_cvtss2si64: + case Intrinsic::x86_sse2_cvtsd2si: + case Intrinsic::x86_sse2_cvtsd2si64: + if (ConstantFP *FPOp = dyn_cast(Op->getOperand(0))) + return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty); + case Intrinsic::x86_sse_cvttss2si: + case Intrinsic::x86_sse_cvttss2si64: + case Intrinsic::x86_sse2_cvttsd2si: + case Intrinsic::x86_sse2_cvttsd2si64: + if (ConstantFP *FPOp = dyn_cast(Op->getOperand(0))) + return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty); + } + } + + if (isa(Operands[0])) { + if (F->getIntrinsicID() == Intrinsic::bswap) + return Operands[0]; + return 0; } - } else if (NumOperands == 2) { + + return 0; + } + + if (NumOperands == 2) { if (ConstantFP *Op1 = dyn_cast(Operands[0])) { - if (Ty!=Type::getFloatTy(F->getContext()) && - Ty!=Type::getDoubleTy(Context)) + if (!Ty->isFloatTy() && !Ty->isDoubleTy()) return 0; - double Op1V = Ty==Type::getFloatTy(F->getContext()) ? - (double)Op1->getValueAPF().convertToFloat(): + double Op1V = Ty->isFloatTy() ? + (double)Op1->getValueAPF().convertToFloat() : Op1->getValueAPF().convertToDouble(); if (ConstantFP *Op2 = dyn_cast(Operands[1])) { - double Op2V = Ty==Type::getFloatTy(F->getContext()) ? + if (Op2->getType() != Op1->getType()) + return 0; + + double Op2V = Ty->isFloatTy() ? (double)Op2->getValueAPF().convertToFloat(): Op2->getValueAPF().convertToDouble(); - if (Name == "pow") { - return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty, Context); - } else if (Name == "fmod") { - return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty, Context); - } else if (Name == "atan2") { - return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty, Context); - } + if (Name == "pow") + return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); + if (Name == "fmod") + return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); + if (Name == "atan2") + return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); } else if (ConstantInt *Op2C = dyn_cast(Operands[1])) { - if (Name == "llvm.powi.f32") { - return ConstantFP::get(Context, APFloat((float)std::pow((float)Op1V, - (int)Op2C->getZExtValue()))); - } else if (Name == "llvm.powi.f64") { - return ConstantFP::get(Context, APFloat((double)std::pow((double)Op1V, + if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy()) + return ConstantFP::get(F->getContext(), + APFloat((float)std::pow((float)Op1V, (int)Op2C->getZExtValue()))); + if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy()) + return ConstantFP::get(F->getContext(), + APFloat((double)std::pow((double)Op1V, + (int)Op2C->getZExtValue()))); + } + return 0; + } + + + if (ConstantInt *Op1 = dyn_cast(Operands[0])) { + if (ConstantInt *Op2 = dyn_cast(Operands[1])) { + switch (F->getIntrinsicID()) { + default: break; + case Intrinsic::sadd_with_overflow: + case Intrinsic::uadd_with_overflow: + case Intrinsic::ssub_with_overflow: + case Intrinsic::usub_with_overflow: + case Intrinsic::smul_with_overflow: + case Intrinsic::umul_with_overflow: { + APInt Res; + bool Overflow; + switch (F->getIntrinsicID()) { + default: assert(0 && "Invalid case"); + case Intrinsic::sadd_with_overflow: + Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow); + break; + case Intrinsic::uadd_with_overflow: + Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow); + break; + case Intrinsic::ssub_with_overflow: + Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow); + break; + case Intrinsic::usub_with_overflow: + Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow); + break; + case Intrinsic::smul_with_overflow: + Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow); + break; + case Intrinsic::umul_with_overflow: + Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow); + break; + } + Constant *Ops[] = { + ConstantInt::get(F->getContext(), Res), + ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow) + }; + return ConstantStruct::get(F->getContext(), Ops, 2, false); + } } } + + return 0; } + return 0; } return 0; } -