X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FConstantFolding.cpp;h=146897ad675b057bb27b124765d302db6cb382e6;hb=847a9c6d778b3209683a92fcb37708b2e8b08f3f;hp=b5fafd685cd7b7f6aaa63263c291ba081df1af20;hpb=83471853b1d5a9efe6b9e7565c457fc901f84926;p=oota-llvm.git diff --git a/lib/Analysis/ConstantFolding.cpp b/lib/Analysis/ConstantFolding.cpp index b5fafd685cd..146897ad675 100644 --- a/lib/Analysis/ConstantFolding.cpp +++ b/lib/Analysis/ConstantFolding.cpp @@ -11,7 +11,7 @@ // // 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 +// DataLayout information. These functions cannot go in VMCore due to library // dependency issues. // //===----------------------------------------------------------------------===// @@ -25,7 +25,8 @@ #include "llvm/Intrinsics.h" #include "llvm/Operator.h" #include "llvm/Analysis/ValueTracking.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" +#include "llvm/Target/TargetLibraryInfo.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/StringMap.h" #include "llvm/Support/ErrorHandling.h" @@ -41,13 +42,54 @@ using namespace llvm; //===----------------------------------------------------------------------===// /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with -/// TargetData. This always returns a non-null constant, but it may be a +/// DataLayout. 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) { +static Constant *FoldBitCast(Constant *C, Type *DestTy, + const DataLayout &TD) { + // Catch the obvious splat cases. + if (C->isNullValue() && !DestTy->isX86_MMXTy()) + return Constant::getNullValue(DestTy); + if (C->isAllOnesValue() && !DestTy->isX86_MMXTy()) + return Constant::getAllOnesValue(DestTy); + + // Handle a vector->integer cast. + if (IntegerType *IT = dyn_cast(DestTy)) { + ConstantDataVector *CDV = dyn_cast(C); + if (CDV == 0) + return ConstantExpr::getBitCast(C, DestTy); + + unsigned NumSrcElts = CDV->getType()->getNumElements(); + + Type *SrcEltTy = CDV->getType()->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. + C = ConstantExpr::getBitCast(C, SrcIVTy); + CDV = cast(C); + } + + // 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); + APInt Result(IT->getBitWidth(), 0); + for (unsigned i = 0; i != NumSrcElts; ++i) { + Result <<= BitShift; + if (TD.isLittleEndian()) + Result |= CDV->getElementAsInteger(NumSrcElts-i-1); + else + Result |= CDV->getElementAsInteger(i); + } + + return ConstantInt::get(IT, Result); + } - // This only handles casts to vectors currently. - const VectorType *DestVTy = dyn_cast(DestTy); + // The code below only handles casts to vectors currently. + VectorType *DestVTy = dyn_cast(DestTy); if (DestVTy == 0) return ConstantExpr::getBitCast(C, DestTy); @@ -59,18 +101,17 @@ static Constant *FoldBitCast(Constant *C, const Type *DestTy, } // If this is a bitcast from constant vector -> vector, fold it. - ConstantVector *CV = dyn_cast(C); - if (CV == 0) + if (!isa(C) && !isa(C)) return ConstantExpr::getBitCast(C, DestTy); // If the element types match, VMCore can fold it. unsigned NumDstElt = DestVTy->getNumElements(); - unsigned NumSrcElt = CV->getNumOperands(); + unsigned NumSrcElt = C->getType()->getVectorNumElements(); if (NumDstElt == NumSrcElt) return ConstantExpr::getBitCast(C, DestTy); - const Type *SrcEltTy = CV->getType()->getElementType(); - const Type *DstEltTy = DestVTy->getElementType(); + Type *SrcEltTy = C->getType()->getVectorElementType(); + 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, @@ -85,11 +126,10 @@ static Constant *FoldBitCast(Constant *C, const Type *DestTy, if (DstEltTy->isFloatingPointTy()) { // Fold to an vector of integers with same size as our FP type. unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); - const Type *DestIVTy = + 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); @@ -99,12 +139,13 @@ static Constant *FoldBitCast(Constant *C, const Type *DestTy, // it to integer first. if (SrcEltTy->isFloatingPointTy()) { unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); - const Type *SrcIVTy = + 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. + // If VMCore wasn't able to fold it, bail out. + if (!isa(C) && // FIXME: Remove ConstantVector. + !isa(C)) return C; } @@ -126,7 +167,7 @@ static Constant *FoldBitCast(Constant *C, const Type *DestTy, Constant *Elt = Zero; unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); for (unsigned j = 0; j != Ratio; ++j) { - Constant *Src = dyn_cast(CV->getOperand(SrcElt++)); + Constant *Src =dyn_cast(C->getAggregateElement(SrcElt++)); if (!Src) // Reject constantexpr elements. return ConstantExpr::getBitCast(C, DestTy); @@ -143,28 +184,29 @@ static Constant *FoldBitCast(Constant *C, const Type *DestTy, } Result.push_back(Elt); } - } else { - // Handle: bitcast (<2 x i64> to <4 x i32>) - unsigned Ratio = NumDstElt/NumSrcElt; - unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits(); + 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); - // 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; - 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)); - } + // Truncate and remember this piece. + Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); } } @@ -176,7 +218,7 @@ static Constant *FoldBitCast(Constant *C, const 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 TargetData &TD) { + int64_t &Offset, const DataLayout &TD) { // Trivial case, constant is the global. if ((GV = dyn_cast(C))) { Offset = 0; @@ -212,11 +254,11 @@ static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, if (!CI) return false; // Index isn't a simple constant? if (CI->isZero()) continue; // Not adding anything. - if (const StructType *ST = dyn_cast(*GTI)) { + if (StructType *ST = dyn_cast(*GTI)) { // N = N + Offset Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue()); } else { - const SequentialType *SQT = cast(*GTI); + SequentialType *SQT = cast(*GTI); Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue(); } } @@ -232,7 +274,7 @@ static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, /// the CurPtr buffer. TD is the target data. static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr, unsigned BytesLeft, - const TargetData &TD) { + const DataLayout &TD) { assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) && "Out of range access"); @@ -267,7 +309,7 @@ static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, } return false; } - + if (ConstantStruct *CS = dyn_cast(C)) { const StructLayout *SL = TD.getStructLayout(CS->getType()); unsigned Index = SL->getElementContainingOffset(ByteOffset); @@ -305,47 +347,40 @@ static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, // not reached. } - if (ConstantArray *CA = dyn_cast(C)) { - uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType()); + if (isa(C) || isa(C) || + isa(C)) { + Type *EltTy = cast(C->getType())->getElementType(); + uint64_t EltSize = TD.getTypeAllocSize(EltTy); uint64_t Index = ByteOffset / EltSize; uint64_t Offset = ByteOffset - Index * EltSize; - for (; Index != CA->getType()->getNumElements(); ++Index) { - if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr, + uint64_t NumElts; + if (ArrayType *AT = dyn_cast(C->getType())) + NumElts = AT->getNumElements(); + else + NumElts = cast(C->getType())->getNumElements(); + + for (; Index != NumElts; ++Index) { + if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr, BytesLeft, TD)) return false; - if (EltSize >= BytesLeft) + + uint64_t BytesWritten = EltSize - Offset; + assert(BytesWritten <= EltSize && "Not indexing into this element?"); + if (BytesWritten >= BytesLeft) return true; - + Offset = 0; - BytesLeft -= EltSize; - CurPtr += EltSize; + BytesLeft -= BytesWritten; + CurPtr += BytesWritten; } 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); + return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr, + BytesLeft, TD); } // Otherwise, unknown initializer type. @@ -353,9 +388,9 @@ static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, } static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, - const TargetData &TD) { - const Type *LoadTy = cast(C->getType())->getElementType(); - const IntegerType *IntType = dyn_cast(LoadTy); + 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) { @@ -363,7 +398,7 @@ static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, // 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; + Type *MapTy; if (LoadTy->isFloatTy()) MapTy = Type::getInt32PtrTy(C->getContext()); else if (LoadTy->isDoubleTy()) @@ -420,7 +455,7 @@ static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, /// produce if it is constant and determinable. If this is not determinable, /// return null. Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, - const TargetData *TD) { + const DataLayout *TD) { // First, try the easy cases: if (GlobalVariable *GV = dyn_cast(C)) if (GV->isConstant() && GV->hasDefinitiveInitializer()) @@ -440,10 +475,10 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, // 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(); + StringRef Str; + if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) { + unsigned StrLen = Str.size(); + Type *Ty = cast(CE->getType())->getElementType(); unsigned NumBits = Ty->getPrimitiveSizeInBits(); // Replace load with immediate integer if the result is an integer or fp // value. @@ -478,7 +513,7 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, if (GlobalVariable *GV = dyn_cast(GetUnderlyingObject(CE, TD))) { if (GV->isConstant() && GV->hasDefinitiveInitializer()) { - const Type *ResTy = cast(C->getType())->getElementType(); + Type *ResTy = cast(C->getType())->getElementType(); if (GV->getInitializer()->isNullValue()) return Constant::getNullValue(ResTy); if (isa(GV->getInitializer())) @@ -494,7 +529,7 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, return 0; } -static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){ +static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){ if (LI->isVolatile()) return 0; if (Constant *C = dyn_cast(LI->getOperand(0))) @@ -508,7 +543,7 @@ static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){ /// 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){ + Constant *Op1, const DataLayout *TD){ // SROA // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. @@ -536,19 +571,18 @@ static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, /// 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) { +static Constant *CastGEPIndices(ArrayRef Ops, + Type *ResultTy, const DataLayout *TD, + const TargetLibraryInfo *TLI) { if (!TD) return 0; - const Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext()); + Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext()); bool Any = false; SmallVector NewIdxs; - for (unsigned i = 1; i != NumOps; ++i) { + for (unsigned i = 1, e = Ops.size(); i != e; ++i) { if ((i == 1 || !isa(GetElementPtrInst::getIndexedType(Ops[0]->getType(), - reinterpret_cast(Ops+1), - i-1))) && + Ops.slice(1, i-1)))) && Ops[i]->getType() != IntPtrTy) { Any = true; NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i], @@ -562,32 +596,49 @@ static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps, if (!Any) return 0; Constant *C = - ConstantExpr::getGetElementPtr(Ops[0], &NewIdxs[0], NewIdxs.size()); + ConstantExpr::getGetElementPtr(Ops[0], NewIdxs); if (ConstantExpr *CE = dyn_cast(C)) - if (Constant *Folded = ConstantFoldConstantExpression(CE, TD)) + if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI)) C = Folded; return C; } +/// Strip the pointer casts, but preserve the address space information. +static Constant* StripPtrCastKeepAS(Constant* Ptr) { + assert(Ptr->getType()->isPointerTy() && "Not a pointer type"); + PointerType *OldPtrTy = cast(Ptr->getType()); + Ptr = cast(Ptr->stripPointerCasts()); + PointerType *NewPtrTy = cast(Ptr->getType()); + + // Preserve the address space number of the pointer. + if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) { + NewPtrTy = NewPtrTy->getElementType()->getPointerTo( + OldPtrTy->getAddressSpace()); + Ptr = ConstantExpr::getBitCast(Ptr, NewPtrTy); + } + return Ptr; +} + /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP /// constant expression, do so. -static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, - const Type *ResultTy, - const TargetData *TD) { +static Constant *SymbolicallyEvaluateGEP(ArrayRef Ops, + Type *ResultTy, const DataLayout *TD, + const TargetLibraryInfo *TLI) { Constant *Ptr = Ops[0]; - if (!TD || !cast(Ptr->getType())->getElementType()->isSized()) + if (!TD || !cast(Ptr->getType())->getElementType()->isSized() || + !Ptr->getType()->isPointerTy()) return 0; - const Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext()); + 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) + 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 (NumOps == 2 && + if (Ops.size() == 2 && cast(ResultTy)->getElementType()->isIntegerTy(8)) { ConstantExpr *CE = dyn_cast(Ops[1]); assert((CE == 0 || CE->getType() == IntPtrTy) && @@ -598,18 +649,20 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, Res = ConstantExpr::getSub(Res, CE->getOperand(1)); Res = ConstantExpr::getIntToPtr(Res, ResultTy); if (ConstantExpr *ResCE = dyn_cast(Res)) - Res = ConstantFoldConstantExpression(ResCE, TD); + Res = ConstantFoldConstantExpression(ResCE, TD, TLI); 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()); + APInt Offset = + APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(), + makeArrayRef((Value *const*) + Ops.data() + 1, + Ops.size() - 1))); + Ptr = StripPtrCastKeepAS(Ptr); // If this is a GEP of a GEP, fold it all into a single GEP. while (GEPOperator *GEP = dyn_cast(Ptr)) { @@ -627,10 +680,8 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, Ptr = cast(GEP->getOperand(0)); Offset += APInt(BitWidth, - TD->getIndexedOffset(Ptr->getType(), - (Value**)NestedOps.data(), - NestedOps.size())); - Ptr = cast(Ptr->stripPointerCasts()); + TD->getIndexedOffset(Ptr->getType(), NestedOps)); + Ptr = StripPtrCastKeepAS(Ptr); } // If the base value for this address is a literal integer value, fold the @@ -649,10 +700,11 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, // we eliminate over-indexing of the notional static type array bounds. // This makes it easy to determine if the getelementptr is "inbounds". // Also, this helps GlobalOpt do SROA on GlobalVariables. - const Type *Ty = Ptr->getType(); + Type *Ty = Ptr->getType(); + assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type"); SmallVector NewIdxs; do { - if (const SequentialType *ATy = dyn_cast(Ty)) { + if (SequentialType *ATy = dyn_cast(Ty)) { if (ATy->isPointerTy()) { // The only pointer indexing we'll do is on the first index of the GEP. if (!NewIdxs.empty()) @@ -665,7 +717,7 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, // Determine which element of the array the offset points into. APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType())); - const IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext()); + IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext()); if (ElemSize == 0) // 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 @@ -679,11 +731,18 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, 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 - // getZExtValue is at least as safe as the StructLayout API because we - // know the offset is within the struct at this point. + } else if (StructType *STy = dyn_cast(Ty)) { + // If we end up with an offset that isn't valid for this struct type, we + // can't re-form this GEP in a regular form, so bail out. The pointer + // operand likely went through casts that are necessary to make the GEP + // sensible. const StructLayout &SL = *TD->getStructLayout(STy); + if (Offset.uge(SL.getSizeInBytes())) + break; + + // Determine which field of the struct the offset points into. The + // getZExtValue is fine as we've already ensured that the offset is + // within the range representable by the StructLayout API. unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); @@ -703,7 +762,7 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, // Create a GEP. Constant *C = - ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size()); + ConstantExpr::getGetElementPtr(Ptr, NewIdxs); assert(cast(C->getType())->getElementType() == Ty && "Computed GetElementPtr has unexpected type!"); @@ -726,7 +785,9 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, /// 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) { +Constant *llvm::ConstantFoldInstruction(Instruction *I, + const DataLayout *TD, + const TargetLibraryInfo *TLI) { // Handle PHI nodes quickly here... if (PHINode *PN = dyn_cast(I)) { Constant *CommonValue = 0; @@ -739,14 +800,21 @@ Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { // 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. + // If the incoming value is not a constant, then give up. Constant *C = dyn_cast(Incoming); - if (!C || (CommonValue && C != CommonValue)) + if (!C) + return 0; + // Fold the PHI's operands. + if (ConstantExpr *NewC = dyn_cast(C)) + C = ConstantFoldConstantExpression(NewC, TD, TLI); + // If the incoming value is a different constant to + // the one we saw previously, then give up. + if (CommonValue && C != CommonValue) return 0; CommonValue = C; } + // If we reach here, all incoming values are the same constant or undef. return CommonValue ? CommonValue : UndefValue::get(PN->getType()); } @@ -754,15 +822,21 @@ Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { // Scan the operand list, checking to see if they are all constants, if so, // hand off to ConstantFoldInstOperands. SmallVector Ops; - for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) - if (Constant *Op = dyn_cast(*i)) - Ops.push_back(Op); - else + for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) { + Constant *Op = dyn_cast(*i); + if (!Op) return 0; // All operands not constant! + // Fold the Instruction's operands. + if (ConstantExpr *NewCE = dyn_cast(Op)) + Op = ConstantFoldConstantExpression(NewCE, TD, TLI); + + Ops.push_back(Op); + } + if (const CmpInst *CI = dyn_cast(I)) return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], - TD); + TD, TLI); if (const LoadInst *LI = dyn_cast(I)) return ConstantFoldLoadInst(LI, TD); @@ -771,37 +845,36 @@ Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { return ConstantExpr::getInsertValue( cast(IVI->getAggregateOperand()), cast(IVI->getInsertedValueOperand()), - IVI->idx_begin(), IVI->getNumIndices()); + IVI->getIndices()); if (ExtractValueInst *EVI = dyn_cast(I)) return ConstantExpr::getExtractValue( cast(EVI->getAggregateOperand()), - EVI->idx_begin(), EVI->getNumIndices()); + EVI->getIndices()); - return ConstantFoldInstOperands(I->getOpcode(), I->getType(), - Ops.data(), Ops.size(), TD); + return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI); } /// ConstantFoldConstantExpression - Attempt to fold the constant expression -/// using the specified TargetData. If successful, the constant result is +/// 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 TargetData *TD) { + const DataLayout *TD, + const TargetLibraryInfo *TLI) { 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); + NewC = ConstantFoldConstantExpression(NewCE, TD, TLI); Ops.push_back(NewC); } if (CE->isCompare()) return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], - TD); - return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), - Ops.data(), Ops.size(), TD); + TD, TLI); + return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI); } /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the @@ -814,9 +887,10 @@ 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, const Type *DestTy, - Constant* const* Ops, unsigned NumOps, - const TargetData *TD) { +Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy, + ArrayRef Ops, + const DataLayout *TD, + const TargetLibraryInfo *TLI) { // Handle easy binops first. if (Instruction::isBinaryOp(Opcode)) { if (isa(Ops[0]) || isa(Ops[1])) @@ -829,11 +903,11 @@ 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::FCmp: llvm_unreachable("Invalid for compares"); case Instruction::Call: - if (Function *F = dyn_cast(Ops[NumOps - 1])) + if (Function *F = dyn_cast(Ops.back())) if (canConstantFoldCallTo(F)) - return ConstantFoldCall(F, Ops, NumOps - 1); + return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI); return 0; case Instruction::PtrToInt: // If the input is a inttoptr, eliminate the pair. This requires knowing @@ -842,10 +916,11 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, if (TD && CE->getOpcode() == Instruction::IntToPtr) { Constant *Input = CE->getOperand(0); unsigned InWidth = Input->getType()->getScalarSizeInBits(); - if (TD->getPointerSizeInBits() < InWidth) { + unsigned AS = cast(CE->getType())->getAddressSpace(); + if (TD->getPointerSizeInBits(AS) < InWidth) { Constant *Mask = ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, - TD->getPointerSizeInBits())); + TD->getPointerSizeInBits(AS))); Input = ConstantExpr::getAnd(Input, Mask); } // Do a zext or trunc to get to the dest size. @@ -858,9 +933,10 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, // 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 (TD && - TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() && - CE->getOpcode() == Instruction::PtrToInt) + if (TD && CE->getOpcode() == Instruction::PtrToInt && + TD->getPointerSizeInBits( + cast(CE->getOperand(0)->getType())->getAddressSpace()) + <= CE->getType()->getScalarSizeInBits()) return FoldBitCast(CE->getOperand(0), DestTy, *TD); return ConstantExpr::getCast(Opcode, Ops[0], DestTy); @@ -887,12 +963,12 @@ 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 = CastGEPIndices(Ops, NumOps, DestTy, TD)) + if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI)) return C; - if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD)) + if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI)) return C; - return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1); + return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1)); } } @@ -902,7 +978,8 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, /// Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, Constant *Ops0, Constant *Ops1, - const TargetData *TD) { + const DataLayout *TD, + const TargetLibraryInfo *TLI) { // fold: icmp (inttoptr x), null -> icmp x, 0 // fold: icmp (ptrtoint x), 0 -> icmp x, null // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y @@ -912,14 +989,14 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, // around to know if bit truncation is happening. if (ConstantExpr *CE0 = dyn_cast(Ops0)) { if (TD && Ops1->isNullValue()) { - const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); + 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 *Null = Constant::getNullValue(C->getType()); - return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); + return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI); } // Only do this transformation if the int is intptrty in size, otherwise @@ -928,13 +1005,13 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, CE0->getType() == IntPtrTy) { Constant *C = CE0->getOperand(0); Constant *Null = Constant::getNullValue(C->getType()); - return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); + return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI); } } if (ConstantExpr *CE1 = dyn_cast(Ops1)) { if (TD && CE0->getOpcode() == CE1->getOpcode()) { - const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); + Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); if (CE0->getOpcode() == Instruction::IntToPtr) { // Convert the integer value to the right size to ensure we get the @@ -943,7 +1020,7 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, IntPtrTy, false); Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), IntPtrTy, false); - return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD); + return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI); } // Only do this transformation if the int is intptrty in size, otherwise @@ -952,7 +1029,7 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, CE0->getType() == IntPtrTy && CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), - CE1->getOperand(0), TD); + CE1->getOperand(0), TD, TLI); } } @@ -961,13 +1038,15 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { Constant *LHS = - ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD); + ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1, + TD, TLI); Constant *RHS = - ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD); + ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1, + TD, TLI); unsigned OpC = Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; Constant *Ops[] = { LHS, RHS }; - return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, 2, TD); + return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI); } } @@ -980,56 +1059,30 @@ Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, /// constant expression, or null if something is funny and we can't decide. Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, ConstantExpr *CE) { - if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType())) + if (!CE->getOperand(1)->isNullValue()) return 0; // Do not allow stepping over the value! - + // Loop over all of the operands, tracking down which value we are - // addressing... - gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); - for (++I; I != E; ++I) - if (const StructType *STy = dyn_cast(*I)) { - ConstantInt *CU = cast(I.getOperand()); - assert(CU->getZExtValue() < STy->getNumElements() && - "Struct index out of range!"); - unsigned El = (unsigned)CU->getZExtValue(); - if (ConstantStruct *CS = dyn_cast(C)) { - C = CS->getOperand(El); - } else if (isa(C)) { - C = Constant::getNullValue(STy->getElementType(El)); - } else if (isa(C)) { - C = UndefValue::get(STy->getElementType(El)); - } else { - return 0; - } - } else if (ConstantInt *CI = dyn_cast(I.getOperand())) { - if (const ArrayType *ATy = dyn_cast(*I)) { - if (CI->getZExtValue() >= ATy->getNumElements()) - return 0; - if (ConstantArray *CA = dyn_cast(C)) - C = CA->getOperand(CI->getZExtValue()); - else if (isa(C)) - C = Constant::getNullValue(ATy->getElementType()); - else if (isa(C)) - C = UndefValue::get(ATy->getElementType()); - else - return 0; - } 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(VTy->getElementType()); - else if (isa(C)) - C = UndefValue::get(VTy->getElementType()); - else - return 0; - } else { - return 0; - } - } else { - return 0; - } + // addressing. + for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) { + C = C->getAggregateElement(CE->getOperand(i)); + if (C == 0) return 0; + } + return C; +} + +/// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr +/// indices (with an *implied* zero pointer index that is not in the list), +/// return the constant value being addressed by a virtual load, or null if +/// something is funny and we can't decide. +Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C, + ArrayRef Indices) { + // Loop over all of the operands, tracking down which value we are + // addressing. + for (unsigned i = 0, e = Indices.size(); i != e; ++i) { + C = C->getAggregateElement(Indices[i]); + if (C == 0) return 0; + } return C; } @@ -1044,6 +1097,7 @@ bool llvm::canConstantFoldCallTo(const Function *F) { switch (F->getIntrinsicID()) { case Intrinsic::sqrt: + case Intrinsic::pow: case Intrinsic::powi: case Intrinsic::bswap: case Intrinsic::ctpop: @@ -1101,7 +1155,7 @@ llvm::canConstantFoldCallTo(const Function *F) { } static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, - const Type *Ty) { + Type *Ty) { sys::llvm_fenv_clearexcept(); V = NativeFP(V); if (sys::llvm_fenv_testexcept()) { @@ -1114,11 +1168,10 @@ static Constant *ConstantFoldFP(double (*NativeFP)(double), double 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) { + double V, double W, Type *Ty) { sys::llvm_fenv_clearexcept(); V = NativeFP(V, W); if (sys::llvm_fenv_testexcept()) { @@ -1131,7 +1184,6 @@ static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), 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 @@ -1142,11 +1194,8 @@ static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), /// 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()); - +static Constant *ConstantFoldConvertToInt(const APFloat &Val, + bool roundTowardZero, Type *Ty) { // All of these conversion intrinsics form an integer of at most 64bits. unsigned ResultWidth = cast(Ty)->getBitWidth(); assert(ResultWidth <= 64 && @@ -1167,13 +1216,13 @@ static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero, /// 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) { +llvm::ConstantFoldCall(Function *F, ArrayRef Operands, + const TargetLibraryInfo *TLI) { if (!F->hasName()) return 0; StringRef Name = F->getName(); - const Type *Ty = F->getReturnType(); - if (NumOperands == 1) { + Type *Ty = F->getReturnType(); + if (Operands.size() == 1) { if (ConstantFP *Op = dyn_cast(Operands[0])) { if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) { APFloat Val(Op->getValueAPF()); @@ -1183,6 +1232,8 @@ llvm::ConstantFoldCall(Function *F, return ConstantInt::get(F->getContext(), Val.bitcastToAPInt()); } + if (!TLI) + return 0; if (!Ty->isFloatTy() && !Ty->isDoubleTy()) return 0; @@ -1201,43 +1252,43 @@ llvm::ConstantFoldCall(Function *F, Op->getValueAPF().convertToDouble(); switch (Name[0]) { case 'a': - if (Name == "acos") + if (Name == "acos" && TLI->has(LibFunc::acos)) return ConstantFoldFP(acos, V, Ty); - else if (Name == "asin") + else if (Name == "asin" && TLI->has(LibFunc::asin)) return ConstantFoldFP(asin, V, Ty); - else if (Name == "atan") + else if (Name == "atan" && TLI->has(LibFunc::atan)) return ConstantFoldFP(atan, V, Ty); break; case 'c': - if (Name == "ceil") + if (Name == "ceil" && TLI->has(LibFunc::ceil)) return ConstantFoldFP(ceil, V, Ty); - else if (Name == "cos") + else if (Name == "cos" && TLI->has(LibFunc::cos)) return ConstantFoldFP(cos, V, Ty); - else if (Name == "cosh") + else if (Name == "cosh" && TLI->has(LibFunc::cosh)) return ConstantFoldFP(cosh, V, Ty); - else if (Name == "cosf") + else if (Name == "cosf" && TLI->has(LibFunc::cosf)) return ConstantFoldFP(cos, V, Ty); break; case 'e': - if (Name == "exp") + if (Name == "exp" && TLI->has(LibFunc::exp)) return ConstantFoldFP(exp, V, Ty); - if (Name == "exp2") { + 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. return ConstantFoldBinaryFP(pow, 2.0, V, Ty); } break; case 'f': - if (Name == "fabs") + if (Name == "fabs" && TLI->has(LibFunc::fabs)) return ConstantFoldFP(fabs, V, Ty); - else if (Name == "floor") + else if (Name == "floor" && TLI->has(LibFunc::floor)) return ConstantFoldFP(floor, V, Ty); break; case 'l': - if (Name == "log" && V > 0) + if (Name == "log" && V > 0 && TLI->has(LibFunc::log)) return ConstantFoldFP(log, V, Ty); - else if (Name == "log10" && V > 0) + 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())) { @@ -1248,21 +1299,21 @@ llvm::ConstantFoldCall(Function *F, } break; case 's': - if (Name == "sin") + if (Name == "sin" && TLI->has(LibFunc::sin)) return ConstantFoldFP(sin, V, Ty); - else if (Name == "sinh") + else if (Name == "sinh" && TLI->has(LibFunc::sinh)) return ConstantFoldFP(sinh, V, Ty); - else if (Name == "sqrt" && V >= 0) + else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt)) return ConstantFoldFP(sqrt, V, Ty); - else if (Name == "sqrtf" && V >= 0) + else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf)) return ConstantFoldFP(sqrt, V, Ty); - else if (Name == "sinf") + else if (Name == "sinf" && TLI->has(LibFunc::sinf)) return ConstantFoldFP(sin, V, Ty); break; case 't': - if (Name == "tan") + if (Name == "tan" && TLI->has(LibFunc::tan)) return ConstantFoldFP(tan, V, Ty); - else if (Name == "tanh") + else if (Name == "tanh" && TLI->has(LibFunc::tanh)) return ConstantFoldFP(tanh, V, Ty); break; default: @@ -1277,10 +1328,6 @@ llvm::ConstantFoldCall(Function *F, return ConstantInt::get(F->getContext(), Op->getValue().byteSwap()); case Intrinsic::ctpop: return ConstantInt::get(Ty, Op->getValue().countPopulation()); - case Intrinsic::cttz: - return ConstantInt::get(Ty, Op->getValue().countTrailingZeros()); - case Intrinsic::ctlz: - return ConstantInt::get(Ty, Op->getValue().countLeadingZeros()); case Intrinsic::convert_from_fp16: { APFloat Val(Op->getValue()); @@ -1300,24 +1347,31 @@ llvm::ConstantFoldCall(Function *F, } } - if (ConstantVector *Op = dyn_cast(Operands[0])) { + // Support ConstantVector in case we have an Undef in the top. + if (isa(Operands[0]) || + isa(Operands[0])) { + Constant *Op = 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); + if (ConstantFP *FPOp = + dyn_cast_or_null(Op->getAggregateElement(0U))) + return ConstantFoldConvertToInt(FPOp->getValueAPF(), + /*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 (ConstantFP *FPOp = + dyn_cast_or_null(Op->getAggregateElement(0U))) + return ConstantFoldConvertToInt(FPOp->getValueAPF(), + /*roundTowardZero=*/true, Ty); } } - + if (isa(Operands[0])) { if (F->getIntrinsicID() == Intrinsic::bswap) return Operands[0]; @@ -1327,7 +1381,7 @@ llvm::ConstantFoldCall(Function *F, return 0; } - if (NumOperands == 2) { + if (Operands.size() == 2) { if (ConstantFP *Op1 = dyn_cast(Operands[0])) { if (!Ty->isFloatTy() && !Ty->isDoubleTy()) return 0; @@ -1337,16 +1391,21 @@ llvm::ConstantFoldCall(Function *F, 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(); - if (Name == "pow") + if (F->getIntrinsicID() == Intrinsic::pow) { return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); - if (Name == "fmod") + } + if (!TLI) + return 0; + if (Name == "pow" && TLI->has(LibFunc::pow)) + return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); + if (Name == "fmod" && TLI->has(LibFunc::fmod)) return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); - if (Name == "atan2") + 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->isFloatTy()) @@ -1361,7 +1420,6 @@ llvm::ConstantFoldCall(Function *F, return 0; } - if (ConstantInt *Op1 = dyn_cast(Operands[0])) { if (ConstantInt *Op2 = dyn_cast(Operands[1])) { switch (F->getIntrinsicID()) { @@ -1375,7 +1433,7 @@ llvm::ConstantFoldCall(Function *F, APInt Res; bool Overflow; switch (F->getIntrinsicID()) { - default: assert(0 && "Invalid case"); + default: llvm_unreachable("Invalid case"); case Intrinsic::sadd_with_overflow: Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow); break; @@ -1401,6 +1459,14 @@ llvm::ConstantFoldCall(Function *F, }; return ConstantStruct::get(cast(F->getReturnType()), Ops); } + case Intrinsic::cttz: + // FIXME: This should check for Op2 == 1, and become unreachable if + // Op1 == 0. + return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros()); + case Intrinsic::ctlz: + // FIXME: This should check for Op2 == 1, and become unreachable if + // Op1 == 0. + return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros()); } }