X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FConstantFolding.cpp;h=bc0dffc473626481a1adc664b2be86285ce62ced;hb=08368387a450dc2b5681000e2728ec702a8f1197;hp=b7bf044a3687f7519323360eda4d1d1b137ff1b3;hpb=426c2bf5cdd2173e4a33aea8cb92cf684a724f4b;p=oota-llvm.git diff --git a/lib/Analysis/ConstantFolding.cpp b/lib/Analysis/ConstantFolding.cpp index b7bf044a368..bc0dffc4736 100644 --- a/lib/Analysis/ConstantFolding.cpp +++ b/lib/Analysis/ConstantFolding.cpp @@ -9,30 +9,31 @@ // // This file defines routines for folding instructions into constants. // -// Also, to supplement the basic VMCore ConstantExpr simplifications, +// Also, to supplement the basic IR ConstantExpr simplifications, // this file defines some additional folding routines that can make use of -// DataLayout information. These functions cannot go in VMCore due to library +// DataLayout information. These functions cannot go in IR due to library // dependency issues. // //===----------------------------------------------------------------------===// #include "llvm/Analysis/ConstantFolding.h" -#include "llvm/Constants.h" -#include "llvm/DerivedTypes.h" -#include "llvm/Function.h" -#include "llvm/GlobalVariable.h" -#include "llvm/Instructions.h" -#include "llvm/Intrinsics.h" -#include "llvm/Operator.h" -#include "llvm/Analysis/ValueTracking.h" -#include "llvm/DataLayout.h" -#include "llvm/Target/TargetLibraryInfo.h" +#include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/StringMap.h" +#include "llvm/Analysis/ValueTracking.h" +#include "llvm/IR/Constants.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/DerivedTypes.h" +#include "llvm/IR/Function.h" +#include "llvm/IR/GlobalVariable.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/Intrinsics.h" +#include "llvm/IR/Operator.h" #include "llvm/Support/ErrorHandling.h" +#include "llvm/Support/FEnv.h" #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/MathExtras.h" -#include "llvm/Support/FEnv.h" +#include "llvm/Target/TargetLibraryInfo.h" #include #include using namespace llvm; @@ -41,7 +42,7 @@ using namespace llvm; // Constant Folding internal helper functions //===----------------------------------------------------------------------===// -/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with +/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with /// DataLayout. This always returns a non-null constant, but it may be a /// ConstantExpr if unfoldable. static Constant *FoldBitCast(Constant *C, Type *DestTy, @@ -54,25 +55,27 @@ static Constant *FoldBitCast(Constant *C, Type *DestTy, // Handle a vector->integer cast. if (IntegerType *IT = dyn_cast(DestTy)) { - ConstantDataVector *CDV = dyn_cast(C); - if (CDV == 0) + VectorType *VTy = dyn_cast(C->getType()); + if (VTy == 0) return ConstantExpr::getBitCast(C, DestTy); - unsigned NumSrcElts = CDV->getType()->getNumElements(); - - Type *SrcEltTy = CDV->getType()->getElementType(); - + unsigned NumSrcElts = VTy->getNumElements(); + Type *SrcEltTy = VTy->getElementType(); + // If the vector is a vector of floating point, convert it to vector of int // to simplify things. if (SrcEltTy->isFloatingPointTy()) { unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); Type *SrcIVTy = VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts); - // Ask VMCore to do the conversion now that #elts line up. + // Ask IR to do the conversion now that #elts line up. C = ConstantExpr::getBitCast(C, SrcIVTy); - CDV = cast(C); } - + + ConstantDataVector *CDV = dyn_cast(C); + if (CDV == 0) + return ConstantExpr::getBitCast(C, DestTy); + // Now that we know that the input value is a vector of integers, just shift // and insert them into our result. unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy); @@ -84,43 +87,43 @@ static Constant *FoldBitCast(Constant *C, Type *DestTy, else Result |= CDV->getElementAsInteger(i); } - + return ConstantInt::get(IT, Result); } - + // The code below only handles casts to vectors currently. 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. if (!isa(C) && !isa(C)) return ConstantExpr::getBitCast(C, DestTy); - - // If the element types match, VMCore can fold it. + + // If the element types match, IR can fold it. unsigned NumDstElt = DestVTy->getNumElements(); unsigned NumSrcElt = C->getType()->getVectorNumElements(); if (NumDstElt == NumSrcElt) return ConstantExpr::getBitCast(C, DestTy); - + Type *SrcEltTy = C->getType()->getVectorElementType(); Type *DstEltTy = DestVTy->getElementType(); - - // Otherwise, we're changing the number of elements in a vector, which + + // 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()) { @@ -130,31 +133,31 @@ static Constant *FoldBitCast(Constant *C, Type *DestTy, VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt); // Recursively handle this integer conversion, if possible. C = FoldBitCast(C, DestIVTy, TD); - - // Finally, VMCore can handle this now that #elts line up. + + // Finally, IR 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(); Type *SrcIVTy = VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt); - // Ask VMCore to do the conversion now that #elts line up. + // Ask IR to do the conversion now that #elts line up. C = ConstantExpr::getBitCast(C, SrcIVTy); - // If VMCore wasn't able to fold it, bail out. + // If IR wasn't able to fold it, bail out. if (!isa(C) && // FIXME: Remove ConstantVector. !isa(C)) 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>) @@ -170,15 +173,15 @@ static Constant *FoldBitCast(Constant *C, Type *DestTy, Constant *Src =dyn_cast(C->getAggregateElement(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, + Src = ConstantExpr::getShl(Src, ConstantInt::get(Src->getType(), ShiftAmt)); ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; - + // Mix it in. Elt = ConstantExpr::getOr(Elt, Src); } @@ -186,30 +189,30 @@ static Constant *FoldBitCast(Constant *C, Type *DestTy, } return ConstantVector::get(Result); } - + // 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(C->getAggregateElement(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, + 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); } @@ -218,53 +221,32 @@ static Constant *FoldBitCast(Constant *C, Type *DestTy, /// from a global, return the global and the constant. Because of /// constantexprs, this function is recursive. static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, - int64_t &Offset, const DataLayout &TD) { + APInt &Offset, const DataLayout &TD) { // Trivial case, constant is the global. if ((GV = dyn_cast(C))) { - Offset = 0; + Offset.clearAllBits(); return true; } - + // Otherwise, if this isn't a constant expr, bail out. ConstantExpr *CE = dyn_cast(C); if (!CE) return false; - + // Look through ptr->int and ptr->ptr casts. if (CE->getOpcode() == Instruction::PtrToInt || CE->getOpcode() == Instruction::BitCast) return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD); - - // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) - if (CE->getOpcode() == Instruction::GetElementPtr) { - // Cannot compute this if the element type of the pointer is missing size - // info. - if (!cast(CE->getOperand(0)->getType()) - ->getElementType()->isSized()) - return false; - + + // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) + if (GEPOperator *GEP = dyn_cast(CE)) { // If the base isn't a global+constant, we aren't either. if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD)) return false; - + // Otherwise, add any offset that our operands provide. - gep_type_iterator GTI = gep_type_begin(CE); - for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end(); - i != e; ++i, ++GTI) { - ConstantInt *CI = dyn_cast(*i); - if (!CI) return false; // Index isn't a simple constant? - if (CI->isZero()) continue; // Not adding anything. - - if (StructType *ST = dyn_cast(*GTI)) { - // N = N + Offset - Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue()); - } else { - SequentialType *SQT = cast(*GTI); - Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue(); - } - } - return true; + return GEP->accumulateConstantOffset(TD, Offset); } - + return false; } @@ -277,27 +259,30 @@ static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, const DataLayout &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)); + int n = ByteOffset; + if (!TD.isLittleEndian()) + n = IntBytes - n - 1; + CurPtr[i] = (unsigned char)(Val >> (n * 8)); ++ByteOffset; } return true; } - + if (ConstantFP *CFP = dyn_cast(C)) { if (CFP->getType()->isDoubleTy()) { C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD); @@ -307,15 +292,19 @@ static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD); return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); } + if (CFP->getType()->isHalfTy()){ + C = FoldBitCast(C, Type::getInt16Ty(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. @@ -325,9 +314,9 @@ static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, !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; @@ -375,11 +364,11 @@ static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, } 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, + CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext())) + return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr, BytesLeft, TD); } @@ -391,7 +380,7 @@ static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, const DataLayout &TD) { Type *LoadTy = cast(C->getType())->getElementType(); 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 @@ -399,7 +388,9 @@ static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, // that address spaces don't matter here since we're not going to result in // an actual new load. Type *MapTy; - if (LoadTy->isFloatTy()) + if (LoadTy->isHalfTy()) + MapTy = Type::getInt16PtrTy(C->getContext()); + else if (LoadTy->isFloatTy()) MapTy = Type::getInt32PtrTy(C->getContext()); else if (LoadTy->isDoubleTy()) MapTy = Type::getInt64PtrTy(C->getContext()); @@ -415,15 +406,15 @@ static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, 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; + APInt Offset(TD.getPointerSizeInBits(), 0); if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD)) return 0; - + GlobalVariable *GV = dyn_cast(GVal); if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || !GV->getInitializer()->getType()->isSized()) @@ -431,21 +422,31 @@ static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, // 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 (Offset.isNegative()) return 0; + // If we're not accessing anything in this constant, the result is undefined. - if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType())) + if (Offset.getZExtValue() >= + TD.getTypeAllocSize(GV->getInitializer()->getType())) return UndefValue::get(IntType); - + unsigned char RawBytes[32] = {0}; - if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes, + if (!ReadDataFromGlobal(GV->getInitializer(), Offset.getZExtValue(), 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]; + APInt ResultVal = APInt(IntType->getBitWidth(), 0); + if (TD.isLittleEndian()) { + ResultVal = RawBytes[BytesLoaded - 1]; + for (unsigned i = 1; i != BytesLoaded; ++i) { + ResultVal <<= 8; + ResultVal |= RawBytes[BytesLoaded-1-i]; + } + } else { + ResultVal = RawBytes[0]; + for (unsigned i = 1; i != BytesLoaded; ++i) { + ResultVal <<= 8; + ResultVal |= RawBytes[i]; + } } return ConstantInt::get(IntType->getContext(), ResultVal); @@ -464,15 +465,15 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, // 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 = + 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. StringRef Str; @@ -500,14 +501,14 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, 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 = @@ -520,18 +521,16 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, 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()) + + // Try hard to fold loads from bitcasted strange and non-type-safe things. + if (TD) return FoldReinterpretLoadFromConstPtr(CE, *TD); return 0; } static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){ if (LI->isVolatile()) return 0; - + if (Constant *C = dyn_cast(LI->getOperand(0))) return ConstantFoldLoadFromConstPtr(C, TD); @@ -540,31 +539,58 @@ static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){ /// 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. +/// these together. If target data info is available, it is provided as DL, +/// otherwise DL is null. static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, - Constant *Op1, const DataLayout *TD){ + Constant *Op1, const DataLayout *DL){ // SROA - + // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute // bits. - - + + + if (Opc == Instruction::And && DL) { + unsigned BitWidth = DL->getTypeSizeInBits(Op0->getType()->getScalarType()); + APInt KnownZero0(BitWidth, 0), KnownOne0(BitWidth, 0); + APInt KnownZero1(BitWidth, 0), KnownOne1(BitWidth, 0); + ComputeMaskedBits(Op0, KnownZero0, KnownOne0, DL); + ComputeMaskedBits(Op1, KnownZero1, KnownOne1, DL); + if ((KnownOne1 | KnownZero0).isAllOnesValue()) { + // All the bits of Op0 that the 'and' could be masking are already zero. + return Op0; + } + if ((KnownOne0 | KnownZero1).isAllOnesValue()) { + // All the bits of Op1 that the 'and' could be masking are already zero. + return Op1; + } + + APInt KnownZero = KnownZero0 | KnownZero1; + APInt KnownOne = KnownOne0 & KnownOne1; + if ((KnownZero | KnownOne).isAllOnesValue()) { + return ConstantInt::get(Op0->getType(), KnownOne); + } + } + // If the constant expr is something like &A[123] - &A[4].f, fold this into a // constant. This happens frequently when iterating over a global array. - if (Opc == Instruction::Sub && TD) { + if (Opc == Instruction::Sub && DL) { GlobalValue *GV1, *GV2; - int64_t Offs1, Offs2; - - if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD)) - if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) && + unsigned PtrSize = DL->getPointerSizeInBits(); + unsigned OpSize = DL->getTypeSizeInBits(Op0->getType()); + APInt Offs1(PtrSize, 0), Offs2(PtrSize, 0); + + if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *DL)) + if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *DL) && GV1 == GV2) { // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. - return ConstantInt::get(Op0->getType(), Offs1-Offs2); + // PtrToInt may change the bitwidth so we have convert to the right size + // first. + return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) - + Offs2.zextOrTrunc(OpSize)); } } - + return 0; } @@ -628,14 +654,14 @@ static Constant *SymbolicallyEvaluateGEP(ArrayRef Ops, if (!TD || !cast(Ptr->getType())->getElementType()->isSized() || !Ptr->getType()->isPointerTy()) return 0; - + 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, e = Ops.size(); i != e; ++i) if (!isa(Ops[i])) { - + // If this is "gep i8* Ptr, (sub 0, V)", fold this as: // "inttoptr (sub (ptrtoint Ptr), V)" if (Ops.size() == 2 && @@ -709,12 +735,12 @@ static Constant *SymbolicallyEvaluateGEP(ArrayRef Ops, // 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())); IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext()); @@ -837,7 +863,7 @@ Constant *llvm::ConstantFoldInstruction(Instruction *I, if (const CmpInst *CI = dyn_cast(I)) return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], TD, TLI); - + if (const LoadInst *LI = dyn_cast(I)) return ConstantFoldLoadInst(LI, TD); @@ -855,19 +881,20 @@ Constant *llvm::ConstantFoldInstruction(Instruction *I, return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI); } -/// ConstantFoldConstantExpression - Attempt to fold the constant expression -/// using the specified DataLayout. If successful, the constant result is -/// result is returned, if not, null is returned. -Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, - const DataLayout *TD, - const TargetLibraryInfo *TLI) { - SmallVector Ops; - for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); - i != e; ++i) { +static Constant * +ConstantFoldConstantExpressionImpl(const ConstantExpr *CE, const DataLayout *TD, + const TargetLibraryInfo *TLI, + SmallPtrSet &FoldedOps) { + SmallVector Ops; + 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, TLI); + // Recursively fold the ConstantExpr's operands. If we have already folded + // a ConstantExpr, we don't have to process it again. + if (ConstantExpr *NewCE = dyn_cast(NewC)) { + if (FoldedOps.insert(NewCE)) + NewC = ConstantFoldConstantExpressionImpl(NewCE, TD, TLI, FoldedOps); + } Ops.push_back(NewC); } @@ -877,6 +904,16 @@ Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI); } +/// ConstantFoldConstantExpression - Attempt to fold the constant expression +/// using the specified DataLayout. If successful, the constant result is +/// result is returned, if not, null is returned. +Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, + const DataLayout *TD, + const TargetLibraryInfo *TLI) { + SmallPtrSet FoldedOps; + return ConstantFoldConstantExpressionImpl(CE, TD, TLI, FoldedOps); +} + /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the /// specified opcode and operands. If successful, the constant result is /// returned, if not, null is returned. Note that this function can fail when @@ -887,19 +924,19 @@ Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, /// information, due to only being passed an opcode and operands. Constant /// folding using this function strips this information. /// -Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy, +Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy, ArrayRef Ops, const DataLayout *TD, - const TargetLibraryInfo *TLI) { + const TargetLibraryInfo *TLI) { // 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)) return C; - + return ConstantExpr::get(Opcode, Ops[0], Ops[1]); } - + switch (Opcode) { default: return 0; case Instruction::ICmp: @@ -917,7 +954,7 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy, Constant *Input = CE->getOperand(0); unsigned InWidth = Input->getType()->getScalarSizeInBits(); if (TD->getPointerSizeInBits() < InWidth) { - Constant *Mask = + Constant *Mask = ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, TD->getPointerSizeInBits())); Input = ConstantExpr::getAnd(Input, Mask); @@ -965,7 +1002,7 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy, return C; if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI)) return C; - + return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1)); } } @@ -975,7 +1012,7 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy, /// returns a constant expression of the specified operands. /// Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, - Constant *Ops0, Constant *Ops1, + Constant *Ops0, Constant *Ops1, const DataLayout *TD, const TargetLibraryInfo *TLI) { // fold: icmp (inttoptr x), null -> icmp x, 0 @@ -996,17 +1033,17 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, Constant *Null = Constant::getNullValue(C->getType()); return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI); } - + // 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 && + if (CE0->getOpcode() == Instruction::PtrToInt && CE0->getType() == IntPtrTy) { Constant *C = CE0->getOperand(0); Constant *Null = Constant::getNullValue(C->getType()); return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI); } } - + if (ConstantExpr *CE1 = dyn_cast(Ops1)) { if (TD && CE0->getOpcode() == CE1->getOpcode()) { Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); @@ -1030,24 +1067,24 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, CE1->getOperand(0), TD, TLI); } } - + // 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 = + Constant *LHS = ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1, TD, TLI); - Constant *RHS = + Constant *RHS = ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1, TD, TLI); - unsigned OpC = + unsigned OpC = Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; Constant *Ops[] = { LHS, RHS }; return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI); } } - + return ConstantExpr::getCompare(Predicate, Ops0, Ops1); } @@ -1055,7 +1092,7 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a /// 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, +Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, ConstantExpr *CE) { if (!CE->getOperand(1)->isNullValue()) return 0; // Do not allow stepping over the value! @@ -1094,6 +1131,13 @@ Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C, bool llvm::canConstantFoldCallTo(const Function *F) { switch (F->getIntrinsicID()) { + case Intrinsic::fabs: + case Intrinsic::log: + case Intrinsic::log2: + case Intrinsic::log10: + case Intrinsic::exp: + case Intrinsic::exp2: + case Intrinsic::floor: case Intrinsic::sqrt: case Intrinsic::pow: case Intrinsic::powi: @@ -1125,15 +1169,14 @@ llvm::canConstantFoldCallTo(const Function *F) { if (!F->hasName()) return false; StringRef Name = F->getName(); - + // In these cases, the check of the length is required. We don't want to // return true for a name like "cos\0blah" which strcmp would return equal to // "cos", but has length 8. switch (Name[0]) { default: return false; case 'a': - return Name == "acos" || Name == "asin" || - Name == "atan" || Name == "atan2"; + return Name == "acos" || Name == "asin" || Name == "atan" || Name =="atan2"; case 'c': return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh"; case 'e': @@ -1152,7 +1195,7 @@ llvm::canConstantFoldCallTo(const Function *F) { } } -static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, +static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, Type *Ty) { sys::llvm_fenv_clearexcept(); V = NativeFP(V); @@ -1160,12 +1203,18 @@ static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, sys::llvm_fenv_clearexcept(); return 0; } - + + if (Ty->isHalfTy()) { + APFloat APF(V); + bool unused; + APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused); + return ConstantFP::get(Ty->getContext(), APF); + } 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"); + llvm_unreachable("Can only constant fold half/float/double"); } static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), @@ -1176,12 +1225,18 @@ static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), sys::llvm_fenv_clearexcept(); return 0; } - + + if (Ty->isHalfTy()) { + APFloat APF(V); + bool unused; + APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused); + return ConstantFP::get(Ty->getContext(), APF); + } 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"); + llvm_unreachable("Can only constant fold half/float/double"); } /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer @@ -1233,7 +1288,7 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, if (!TLI) return 0; - if (!Ty->isFloatTy() && !Ty->isDoubleTy()) + if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy()) return 0; /// We only fold functions with finite arguments. Folding NaN and inf is @@ -1246,8 +1301,46 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, /// 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->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() : - Op->getValueAPF().convertToDouble(); + double V; + if (Ty->isFloatTy()) + V = Op->getValueAPF().convertToFloat(); + else if (Ty->isDoubleTy()) + V = Op->getValueAPF().convertToDouble(); + else { + bool unused; + APFloat APF = Op->getValueAPF(); + APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused); + V = APF.convertToDouble(); + } + + switch (F->getIntrinsicID()) { + default: break; + case Intrinsic::fabs: + return ConstantFoldFP(fabs, V, Ty); +#if HAVE_LOG2 + case Intrinsic::log2: + return ConstantFoldFP(log2, V, Ty); +#endif +#if HAVE_LOG + case Intrinsic::log: + return ConstantFoldFP(log, V, Ty); +#endif +#if HAVE_LOG10 + case Intrinsic::log10: + return ConstantFoldFP(log10, V, Ty); +#endif +#if HAVE_EXP + case Intrinsic::exp: + return ConstantFoldFP(exp, V, Ty); +#endif +#if HAVE_EXP2 + case Intrinsic::exp2: + return ConstantFoldFP(exp2, V, Ty); +#endif + case Intrinsic::floor: + return ConstantFoldFP(floor, V, Ty); + } + switch (Name[0]) { case 'a': if (Name == "acos" && TLI->has(LibFunc::acos)) @@ -1270,7 +1363,7 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, case 'e': if (Name == "exp" && TLI->has(LibFunc::exp)) return ConstantFoldFP(exp, V, Ty); - + if (Name == "exp2" && TLI->has(LibFunc::exp2)) { // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a // C99 library. @@ -1289,7 +1382,7 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10)) return ConstantFoldFP(log10, V, Ty); else if (F->getIntrinsicID() == Intrinsic::sqrt && - (Ty->isFloatTy() || Ty->isDoubleTy())) { + (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) { if (V >= -0.0) return ConstantFoldFP(sqrt, V, Ty); else // Undefined @@ -1327,7 +1420,7 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, case Intrinsic::ctpop: return ConstantInt::get(Ty, Op->getValue().countPopulation()); case Intrinsic::convert_from_fp16: { - APFloat Val(Op->getValue()); + APFloat Val(APFloat::IEEEhalf, Op->getValue()); bool lost = false; APFloat::opStatus status = @@ -1346,7 +1439,7 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, } // Support ConstantVector in case we have an Undef in the top. - if (isa(Operands[0]) || + if (isa(Operands[0]) || isa(Operands[0])) { Constant *Op = cast(Operands[0]); switch (F->getIntrinsicID()) { @@ -1365,11 +1458,11 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, case Intrinsic::x86_sse2_cvttsd2si64: if (ConstantFP *FPOp = dyn_cast_or_null(Op->getAggregateElement(0U))) - return ConstantFoldConvertToInt(FPOp->getValueAPF(), + return ConstantFoldConvertToInt(FPOp->getValueAPF(), /*roundTowardZero=*/true, Ty); } } - + if (isa(Operands[0])) { if (F->getIntrinsicID() == Intrinsic::bswap) return Operands[0]; @@ -1381,18 +1474,35 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, if (Operands.size() == 2) { if (ConstantFP *Op1 = dyn_cast(Operands[0])) { - if (!Ty->isFloatTy() && !Ty->isDoubleTy()) + if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy()) return 0; - double Op1V = Ty->isFloatTy() ? - (double)Op1->getValueAPF().convertToFloat() : - Op1->getValueAPF().convertToDouble(); + double Op1V; + if (Ty->isFloatTy()) + Op1V = Op1->getValueAPF().convertToFloat(); + else if (Ty->isDoubleTy()) + Op1V = Op1->getValueAPF().convertToDouble(); + else { + bool unused; + APFloat APF = Op1->getValueAPF(); + APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused); + Op1V = APF.convertToDouble(); + } + if (ConstantFP *Op2 = dyn_cast(Operands[1])) { if (Op2->getType() != Op1->getType()) return 0; - double Op2V = Ty->isFloatTy() ? - (double)Op2->getValueAPF().convertToFloat(): - Op2->getValueAPF().convertToDouble(); + double Op2V; + if (Ty->isFloatTy()) + Op2V = Op2->getValueAPF().convertToFloat(); + else if (Ty->isDoubleTy()) + Op2V = Op2->getValueAPF().convertToDouble(); + else { + bool unused; + APFloat APF = Op2->getValueAPF(); + APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused); + Op2V = APF.convertToDouble(); + } if (F->getIntrinsicID() == Intrinsic::pow) { return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); @@ -1406,6 +1516,10 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, if (Name == "atan2" && TLI->has(LibFunc::atan2)) return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); } else if (ConstantInt *Op2C = dyn_cast(Operands[1])) { + if (F->getIntrinsicID() == Intrinsic::powi && Ty->isHalfTy()) + return ConstantFP::get(F->getContext(), + APFloat((float)std::pow((float)Op1V, + (int)Op2C->getZExtValue()))); if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy()) return ConstantFP::get(F->getContext(), APFloat((float)std::pow((float)Op1V, @@ -1417,7 +1531,7 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, } return 0; } - + if (ConstantInt *Op1 = dyn_cast(Operands[0])) { if (ConstantInt *Op2 = dyn_cast(Operands[1])) { switch (F->getIntrinsicID()) { @@ -1458,16 +1572,16 @@ llvm::ConstantFoldCall(Function *F, ArrayRef Operands, return ConstantStruct::get(cast(F->getReturnType()), Ops); } case Intrinsic::cttz: - // FIXME: This should check for Op2 == 1, and become unreachable if - // Op1 == 0. + if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef. + return UndefValue::get(Ty); return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros()); case Intrinsic::ctlz: - // FIXME: This should check for Op2 == 1, and become unreachable if - // Op1 == 0. + if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef. + return UndefValue::get(Ty); return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros()); } } - + return 0; } return 0;