X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FConstantFolding.cpp;h=68a55b6b94d50b80b682522c29f8ec6c9b3b5ce2;hb=48f17ba2a611d197082d4de730b646a4ecf68df4;hp=1d23fe0b6c911e0498d139560e99fc4d4467b8d2;hpb=0e488b3d1c9293bbf2d64d0ecc4d6339f9100351;p=oota-llvm.git diff --git a/lib/Analysis/ConstantFolding.cpp b/lib/Analysis/ConstantFolding.cpp index 1d23fe0b6c9..68a55b6b94d 100644 --- a/lib/Analysis/ConstantFolding.cpp +++ b/lib/Analysis/ConstantFolding.cpp @@ -30,6 +30,7 @@ #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; @@ -208,7 +209,7 @@ 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 @@ -401,7 +402,7 @@ static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]); for (unsigned i = 1; i != BytesLoaded; ++i) { ResultVal <<= 8; - ResultVal |= APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1-i]); + ResultVal |= RawBytes[BytesLoaded-1-i]; } return ConstantInt::get(IntType->getContext(), ResultVal); @@ -436,8 +437,10 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, unsigned StrLen = Str.length(); const Type *Ty = cast(CE->getType())->getElementType(); unsigned NumBits = Ty->getPrimitiveSizeInBits(); - // Replace LI with immediate integer store. - if ((NumBits >> 3) == StrLen + 1) { + // 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()) { @@ -454,7 +457,11 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, SingleChar = 0; StrVal = (StrVal << 8) | SingleChar; } - return ConstantInt::get(CE->getContext(), StrVal); + + Constant *Res = ConstantInt::get(CE->getContext(), StrVal); + if (Ty->isFloatingPointTy()) + Res = ConstantExpr::getBitCast(Res, Ty); + return Res; } } @@ -531,7 +538,7 @@ static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps, for (unsigned i = 1; i != NumOps; ++i) { if ((i == 1 || !isa(GetElementPtrInst::getIndexedType(Ops[0]->getType(), - reinterpret_cast(Ops+1), + reinterpret_cast(Ops+1), i-1))) && Ops[i]->getType() != IntPtrTy) { Any = true; @@ -564,21 +571,6 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, unsigned BitWidth = TD->getTypeSizeInBits(TD->getIntPtrType(Ptr->getContext())); - 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; - } // 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' @@ -589,9 +581,38 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, 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) { + 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); } @@ -600,7 +621,6 @@ 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. - Ptr = cast(Ptr->stripPointerCasts()); const Type *Ty = Ptr->getType(); SmallVector NewIdxs; do { @@ -617,12 +637,19 @@ 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()); if (ElemSize == 0) - return 0; - APInt NewIdx = Offset.udiv(ElemSize); - Offset -= NewIdx * ElemSize; - NewIdxs.push_back(ConstantInt::get(TD->getIntPtrType(Ty->getContext()), - 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 @@ -666,27 +693,34 @@ static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, // 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. -/// +/// 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 *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... + Constant *CommonValue = 0; + + 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, @@ -704,7 +738,18 @@ Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *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); } @@ -715,7 +760,8 @@ Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, const TargetData *TD) { SmallVector Ops; - for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; ++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)) @@ -757,9 +803,9 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, 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::PtrToInt: // If the input is a inttoptr, eliminate the pair. This requires knowing @@ -783,44 +829,12 @@ 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) - return FoldBitCast(CE->getOperand(0), DestTy, *TD); - - // 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(1))) - if (ConstantExpr *R = dyn_cast(CE->getOperand(0))) - 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(ElTy->getContext(), 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: @@ -1011,6 +1025,9 @@ llvm::canConstantFoldCallTo(const Function *F) { case Intrinsic::usub_with_overflow: case Intrinsic::sadd_with_overflow: case Intrinsic::ssub_with_overflow: + case Intrinsic::smul_with_overflow: + case Intrinsic::convert_from_fp16: + case Intrinsic::convert_to_fp16: return true; default: return false; @@ -1048,10 +1065,10 @@ llvm::canConstantFoldCallTo(const Function *F) { static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, const Type *Ty) { - errno = 0; + sys::llvm_fenv_clearexcept(); V = NativeFP(V); - if (errno != 0) { - errno = 0; + if (sys::llvm_fenv_testexcept()) { + sys::llvm_fenv_clearexcept(); return 0; } @@ -1065,10 +1082,10 @@ static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), double V, double W, const Type *Ty) { - errno = 0; + sys::llvm_fenv_clearexcept(); V = NativeFP(V, W); - if (errno != 0) { - errno = 0; + if (sys::llvm_fenv_testexcept()) { + sys::llvm_fenv_clearexcept(); return 0; } @@ -1091,8 +1108,24 @@ llvm::ConstantFoldCall(Function *F, const Type *Ty = F->getReturnType(); if (NumOperands == 1) { if (ConstantFP *Op = dyn_cast(Operands[0])) { + if (Name == "llvm.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)) == @@ -1175,6 +1208,20 @@ llvm::ConstantFoldCall(Function *F, return ConstantInt::get(Ty, Op->getValue().countTrailingZeros()); else if (Name.startswith("llvm.ctlz")) return ConstantInt::get(Ty, Op->getValue().countLeadingZeros()); + else if (Name == "llvm.convert.from.fp16") { + APFloat Val(Op->getValue()); + + bool lost = false; + APFloat::opStatus status = + Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost); + + // Conversion is always precise. + status = status; + assert(status == APFloat::opOK && !lost && + "Precision lost during fp16 constfolding"); + + return ConstantFP::get(F->getContext(), Val); + } return 0; } @@ -1226,42 +1273,37 @@ llvm::ConstantFoldCall(Function *F, if (ConstantInt *Op2 = dyn_cast(Operands[1])) { switch (F->getIntrinsicID()) { default: break; - case Intrinsic::uadd_with_overflow: { - Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result. + 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: { + 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; + } Constant *Ops[] = { - Res, ConstantExpr::getICmp(CmpInst::ICMP_ULT, Res, Op1) // overflow. + ConstantInt::get(F->getContext(), Res), + ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow) }; return ConstantStruct::get(F->getContext(), Ops, 2, false); } - case Intrinsic::usub_with_overflow: { - Constant *Res = ConstantExpr::getSub(Op1, Op2); // result. - Constant *Ops[] = { - Res, ConstantExpr::getICmp(CmpInst::ICMP_UGT, Res, Op1) // overflow. - }; - return ConstantStruct::get(F->getContext(), Ops, 2, false); - } - case Intrinsic::sadd_with_overflow: { - Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result. - Constant *Overflow = ConstantExpr::getSelect( - ConstantExpr::getICmp(CmpInst::ICMP_SGT, - ConstantInt::get(Op1->getType(), 0), Op1), - ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op2), - ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op2)); // overflow. - - Constant *Ops[] = { Res, Overflow }; - return ConstantStruct::get(F->getContext(), Ops, 2, false); - } - case Intrinsic::ssub_with_overflow: { - Constant *Res = ConstantExpr::getSub(Op1, Op2); // result. - Constant *Overflow = ConstantExpr::getSelect( - ConstantExpr::getICmp(CmpInst::ICMP_SGT, - ConstantInt::get(Op2->getType(), 0), Op2), - ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op1), - ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op1)); // overflow. - - Constant *Ops[] = { Res, Overflow }; - return ConstantStruct::get(F->getContext(), Ops, 2, false); - } } }