//===- ConstantFold.cpp - LLVM constant folder ----------------------------===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // // This file implements folding of constants for LLVM. This implements the // (internal) ConstantFold.h interface, which is used by the // ConstantExpr::get* methods to automatically fold constants when possible. // // The current constant folding implementation is implemented in two pieces: the // template-based folder for simple primitive constants like ConstantInt, and // the special case hackery that we use to symbolically evaluate expressions // that use ConstantExprs. // //===----------------------------------------------------------------------===// #include "ConstantFold.h" #include "llvm/Constants.h" #include "llvm/Instructions.h" #include "llvm/DerivedTypes.h" #include "llvm/Function.h" #include "llvm/GlobalAlias.h" #include "llvm/LLVMContext.h" #include "llvm/ADT/SmallVector.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/ManagedStatic.h" #include "llvm/Support/MathExtras.h" #include using namespace llvm; //===----------------------------------------------------------------------===// // ConstantFold*Instruction Implementations //===----------------------------------------------------------------------===// /// BitCastConstantVector - Convert the specified ConstantVector node to the /// specified vector type. At this point, we know that the elements of the /// input vector constant are all simple integer or FP values. static Constant *BitCastConstantVector(LLVMContext &Context, ConstantVector *CV, const VectorType *DstTy) { // If this cast changes element count then we can't handle it here: // doing so requires endianness information. This should be handled by // Analysis/ConstantFolding.cpp unsigned NumElts = DstTy->getNumElements(); if (NumElts != CV->getNumOperands()) return 0; // Check to verify that all elements of the input are simple. for (unsigned i = 0; i != NumElts; ++i) { if (!isa(CV->getOperand(i)) && !isa(CV->getOperand(i))) return 0; } // Bitcast each element now. std::vector Result; const Type *DstEltTy = DstTy->getElementType(); for (unsigned i = 0; i != NumElts; ++i) Result.push_back(Context.getConstantExprBitCast(CV->getOperand(i), DstEltTy)); return Context.getConstantVector(Result); } /// This function determines which opcode to use to fold two constant cast /// expressions together. It uses CastInst::isEliminableCastPair to determine /// the opcode. Consequently its just a wrapper around that function. /// @brief Determine if it is valid to fold a cast of a cast static unsigned foldConstantCastPair( unsigned opc, ///< opcode of the second cast constant expression const ConstantExpr*Op, ///< the first cast constant expression const Type *DstTy ///< desintation type of the first cast ) { assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); assert(CastInst::isCast(opc) && "Invalid cast opcode"); // The the types and opcodes for the two Cast constant expressions const Type *SrcTy = Op->getOperand(0)->getType(); const Type *MidTy = Op->getType(); Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); Instruction::CastOps secondOp = Instruction::CastOps(opc); // Let CastInst::isEliminableCastPair do the heavy lifting. return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, Type::Int64Ty); } static Constant *FoldBitCast(LLVMContext &Context, Constant *V, const Type *DestTy) { const Type *SrcTy = V->getType(); if (SrcTy == DestTy) return V; // no-op cast // Check to see if we are casting a pointer to an aggregate to a pointer to // the first element. If so, return the appropriate GEP instruction. if (const PointerType *PTy = dyn_cast(V->getType())) if (const PointerType *DPTy = dyn_cast(DestTy)) if (PTy->getAddressSpace() == DPTy->getAddressSpace()) { SmallVector IdxList; IdxList.push_back(Context.getNullValue(Type::Int32Ty)); const Type *ElTy = PTy->getElementType(); while (ElTy != DPTy->getElementType()) { if (const StructType *STy = dyn_cast(ElTy)) { if (STy->getNumElements() == 0) break; ElTy = STy->getElementType(0); IdxList.push_back(Context.getNullValue(Type::Int32Ty)); } else if (const SequentialType *STy = dyn_cast(ElTy)) { if (isa(ElTy)) break; // Can't index into pointers! ElTy = STy->getElementType(); IdxList.push_back(IdxList[0]); } else { break; } } if (ElTy == DPTy->getElementType()) return Context.getConstantExprGetElementPtr(V, &IdxList[0], IdxList.size()); } // Handle casts from one vector constant to another. We know that the src // and dest type have the same size (otherwise its an illegal cast). if (const VectorType *DestPTy = dyn_cast(DestTy)) { if (const VectorType *SrcTy = dyn_cast(V->getType())) { assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && "Not cast between same sized vectors!"); SrcTy = NULL; // First, check for null. Undef is already handled. if (isa(V)) return Context.getNullValue(DestTy); if (ConstantVector *CV = dyn_cast(V)) return BitCastConstantVector(Context, CV, DestPTy); } // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts // This allows for other simplifications (although some of them // can only be handled by Analysis/ConstantFolding.cpp). if (isa(V) || isa(V)) return Context.getConstantExprBitCast( Context.getConstantVector(&V, 1), DestPTy); } // Finally, implement bitcast folding now. The code below doesn't handle // bitcast right. if (isa(V)) // ptr->ptr cast. return Context.getConstantPointerNull(cast(DestTy)); // Handle integral constant input. if (const ConstantInt *CI = dyn_cast(V)) { if (DestTy->isInteger()) // Integral -> Integral. This is a no-op because the bit widths must // be the same. Consequently, we just fold to V. return V; if (DestTy->isFloatingPoint()) return Context.getConstantFP(APFloat(CI->getValue(), DestTy != Type::PPC_FP128Ty)); // Otherwise, can't fold this (vector?) return 0; } // Handle ConstantFP input. if (const ConstantFP *FP = dyn_cast(V)) // FP -> Integral. return Context.getConstantInt(FP->getValueAPF().bitcastToAPInt()); return 0; } Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, unsigned opc, const Constant *V, const Type *DestTy) { if (isa(V)) { // zext(undef) = 0, because the top bits will be zero. // sext(undef) = 0, because the top bits will all be the same. // [us]itofp(undef) = 0, because the result value is bounded. if (opc == Instruction::ZExt || opc == Instruction::SExt || opc == Instruction::UIToFP || opc == Instruction::SIToFP) return Context.getNullValue(DestTy); return Context.getUndef(DestTy); } // No compile-time operations on this type yet. if (V->getType() == Type::PPC_FP128Ty || DestTy == Type::PPC_FP128Ty) return 0; // If the cast operand is a constant expression, there's a few things we can // do to try to simplify it. if (const ConstantExpr *CE = dyn_cast(V)) { if (CE->isCast()) { // Try hard to fold cast of cast because they are often eliminable. if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) return Context.getConstantExprCast(newOpc, CE->getOperand(0), DestTy); } else if (CE->getOpcode() == Instruction::GetElementPtr) { // If all of the indexes in the GEP are null values, there is no pointer // adjustment going on. We might as well cast the source pointer. bool isAllNull = true; for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) if (!CE->getOperand(i)->isNullValue()) { isAllNull = false; break; } if (isAllNull) // This is casting one pointer type to another, always BitCast return Context.getConstantExprPointerCast(CE->getOperand(0), DestTy); } } // If the cast operand is a constant vector, perform the cast by // operating on each element. In the cast of bitcasts, the element // count may be mismatched; don't attempt to handle that here. if (const ConstantVector *CV = dyn_cast(V)) if (isa(DestTy) && cast(DestTy)->getNumElements() == CV->getType()->getNumElements()) { std::vector res; const VectorType *DestVecTy = cast(DestTy); const Type *DstEltTy = DestVecTy->getElementType(); for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) res.push_back(Context.getConstantExprCast(opc, CV->getOperand(i), DstEltTy)); return Context.getConstantVector(DestVecTy, res); } // We actually have to do a cast now. Perform the cast according to the // opcode specified. switch (opc) { case Instruction::FPTrunc: case Instruction::FPExt: if (const ConstantFP *FPC = dyn_cast(V)) { bool ignored; APFloat Val = FPC->getValueAPF(); Val.convert(DestTy == Type::FloatTy ? APFloat::IEEEsingle : DestTy == Type::DoubleTy ? APFloat::IEEEdouble : DestTy == Type::X86_FP80Ty ? APFloat::x87DoubleExtended : DestTy == Type::FP128Ty ? APFloat::IEEEquad : APFloat::Bogus, APFloat::rmNearestTiesToEven, &ignored); return Context.getConstantFP(Val); } return 0; // Can't fold. case Instruction::FPToUI: case Instruction::FPToSI: if (const ConstantFP *FPC = dyn_cast(V)) { const APFloat &V = FPC->getValueAPF(); bool ignored; uint64_t x[2]; uint32_t DestBitWidth = cast(DestTy)->getBitWidth(); (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, APFloat::rmTowardZero, &ignored); APInt Val(DestBitWidth, 2, x); return Context.getConstantInt(Val); } return 0; // Can't fold. case Instruction::IntToPtr: //always treated as unsigned if (V->isNullValue()) // Is it an integral null value? return Context.getConstantPointerNull(cast(DestTy)); return 0; // Other pointer types cannot be casted case Instruction::PtrToInt: // always treated as unsigned if (V->isNullValue()) // is it a null pointer value? return Context.getConstantInt(DestTy, 0); return 0; // Other pointer types cannot be casted case Instruction::UIToFP: case Instruction::SIToFP: if (const ConstantInt *CI = dyn_cast(V)) { APInt api = CI->getValue(); const uint64_t zero[] = {0, 0}; APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(), 2, zero)); (void)apf.convertFromAPInt(api, opc==Instruction::SIToFP, APFloat::rmNearestTiesToEven); return Context.getConstantFP(apf); } return 0; case Instruction::ZExt: if (const ConstantInt *CI = dyn_cast(V)) { uint32_t BitWidth = cast(DestTy)->getBitWidth(); APInt Result(CI->getValue()); Result.zext(BitWidth); return Context.getConstantInt(Result); } return 0; case Instruction::SExt: if (const ConstantInt *CI = dyn_cast(V)) { uint32_t BitWidth = cast(DestTy)->getBitWidth(); APInt Result(CI->getValue()); Result.sext(BitWidth); return Context.getConstantInt(Result); } return 0; case Instruction::Trunc: if (const ConstantInt *CI = dyn_cast(V)) { uint32_t BitWidth = cast(DestTy)->getBitWidth(); APInt Result(CI->getValue()); Result.trunc(BitWidth); return Context.getConstantInt(Result); } return 0; case Instruction::BitCast: return FoldBitCast(Context, const_cast(V), DestTy); default: assert(!"Invalid CE CastInst opcode"); break; } LLVM_UNREACHABLE("Failed to cast constant expression"); return 0; } Constant *llvm::ConstantFoldSelectInstruction(LLVMContext&, const Constant *Cond, const Constant *V1, const Constant *V2) { if (const ConstantInt *CB = dyn_cast(Cond)) return const_cast(CB->getZExtValue() ? V1 : V2); if (isa(V1)) return const_cast(V2); if (isa(V2)) return const_cast(V1); if (isa(Cond)) return const_cast(V1); if (V1 == V2) return const_cast(V1); return 0; } Constant *llvm::ConstantFoldExtractElementInstruction(LLVMContext &Context, const Constant *Val, const Constant *Idx) { if (isa(Val)) // ee(undef, x) -> undef return Context.getUndef(cast(Val->getType())->getElementType()); if (Val->isNullValue()) // ee(zero, x) -> zero return Context.getNullValue( cast(Val->getType())->getElementType()); if (const ConstantVector *CVal = dyn_cast(Val)) { if (const ConstantInt *CIdx = dyn_cast(Idx)) { return CVal->getOperand(CIdx->getZExtValue()); } else if (isa(Idx)) { // ee({w,x,y,z}, undef) -> w (an arbitrary value). return CVal->getOperand(0); } } return 0; } Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, const Constant *Val, const Constant *Elt, const Constant *Idx) { const ConstantInt *CIdx = dyn_cast(Idx); if (!CIdx) return 0; APInt idxVal = CIdx->getValue(); if (isa(Val)) { // Insertion of scalar constant into vector undef // Optimize away insertion of undef if (isa(Elt)) return const_cast(Val); // Otherwise break the aggregate undef into multiple undefs and do // the insertion unsigned numOps = cast(Val->getType())->getNumElements(); std::vector Ops; Ops.reserve(numOps); for (unsigned i = 0; i < numOps; ++i) { const Constant *Op = (idxVal == i) ? Elt : Context.getUndef(Elt->getType()); Ops.push_back(const_cast(Op)); } return Context.getConstantVector(Ops); } if (isa(Val)) { // Insertion of scalar constant into vector aggregate zero // Optimize away insertion of zero if (Elt->isNullValue()) return const_cast(Val); // Otherwise break the aggregate zero into multiple zeros and do // the insertion unsigned numOps = cast(Val->getType())->getNumElements(); std::vector Ops; Ops.reserve(numOps); for (unsigned i = 0; i < numOps; ++i) { const Constant *Op = (idxVal == i) ? Elt : Context.getNullValue(Elt->getType()); Ops.push_back(const_cast(Op)); } return Context.getConstantVector(Ops); } if (const ConstantVector *CVal = dyn_cast(Val)) { // Insertion of scalar constant into vector constant std::vector Ops; Ops.reserve(CVal->getNumOperands()); for (unsigned i = 0; i < CVal->getNumOperands(); ++i) { const Constant *Op = (idxVal == i) ? Elt : cast(CVal->getOperand(i)); Ops.push_back(const_cast(Op)); } return Context.getConstantVector(Ops); } return 0; } /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef /// return the specified element value. Otherwise return null. static Constant *GetVectorElement(LLVMContext &Context, const Constant *C, unsigned EltNo) { if (const ConstantVector *CV = dyn_cast(C)) return CV->getOperand(EltNo); const Type *EltTy = cast(C->getType())->getElementType(); if (isa(C)) return Context.getNullValue(EltTy); if (isa(C)) return Context.getUndef(EltTy); return 0; } Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, const Constant *V1, const Constant *V2, const Constant *Mask) { // Undefined shuffle mask -> undefined value. if (isa(Mask)) return Context.getUndef(V1->getType()); unsigned MaskNumElts = cast(Mask->getType())->getNumElements(); unsigned SrcNumElts = cast(V1->getType())->getNumElements(); const Type *EltTy = cast(V1->getType())->getElementType(); // Loop over the shuffle mask, evaluating each element. SmallVector Result; for (unsigned i = 0; i != MaskNumElts; ++i) { Constant *InElt = GetVectorElement(Context, Mask, i); if (InElt == 0) return 0; if (isa(InElt)) InElt = Context.getUndef(EltTy); else if (ConstantInt *CI = dyn_cast(InElt)) { unsigned Elt = CI->getZExtValue(); if (Elt >= SrcNumElts*2) InElt = Context.getUndef(EltTy); else if (Elt >= SrcNumElts) InElt = GetVectorElement(Context, V2, Elt - SrcNumElts); else InElt = GetVectorElement(Context, V1, Elt); if (InElt == 0) return 0; } else { // Unknown value. return 0; } Result.push_back(InElt); } return Context.getConstantVector(&Result[0], Result.size()); } Constant *llvm::ConstantFoldExtractValueInstruction(LLVMContext &Context, const Constant *Agg, const unsigned *Idxs, unsigned NumIdx) { // Base case: no indices, so return the entire value. if (NumIdx == 0) return const_cast(Agg); if (isa(Agg)) // ev(undef, x) -> undef return Context.getUndef(ExtractValueInst::getIndexedType(Agg->getType(), Idxs, Idxs + NumIdx)); if (isa(Agg)) // ev(0, x) -> 0 return Context.getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), Idxs, Idxs + NumIdx)); // Otherwise recurse. return ConstantFoldExtractValueInstruction(Context, Agg->getOperand(*Idxs), Idxs+1, NumIdx-1); } Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, const Constant *Agg, const Constant *Val, const unsigned *Idxs, unsigned NumIdx) { // Base case: no indices, so replace the entire value. if (NumIdx == 0) return const_cast(Val); if (isa(Agg)) { // Insertion of constant into aggregate undef // Optimize away insertion of undef if (isa(Val)) return const_cast(Agg); // Otherwise break the aggregate undef into multiple undefs and do // the insertion const CompositeType *AggTy = cast(Agg->getType()); unsigned numOps; if (const ArrayType *AR = dyn_cast(AggTy)) numOps = AR->getNumElements(); else numOps = cast(AggTy)->getNumElements(); std::vector Ops(numOps); for (unsigned i = 0; i < numOps; ++i) { const Type *MemberTy = AggTy->getTypeAtIndex(i); const Constant *Op = (*Idxs == i) ? ConstantFoldInsertValueInstruction(Context, Context.getUndef(MemberTy), Val, Idxs+1, NumIdx-1) : Context.getUndef(MemberTy); Ops[i] = const_cast(Op); } if (isa(AggTy)) return Context.getConstantStruct(Ops); else return Context.getConstantArray(cast(AggTy), Ops); } if (isa(Agg)) { // Insertion of constant into aggregate zero // Optimize away insertion of zero if (Val->isNullValue()) return const_cast(Agg); // Otherwise break the aggregate zero into multiple zeros and do // the insertion const CompositeType *AggTy = cast(Agg->getType()); unsigned numOps; if (const ArrayType *AR = dyn_cast(AggTy)) numOps = AR->getNumElements(); else numOps = cast(AggTy)->getNumElements(); std::vector Ops(numOps); for (unsigned i = 0; i < numOps; ++i) { const Type *MemberTy = AggTy->getTypeAtIndex(i); const Constant *Op = (*Idxs == i) ? ConstantFoldInsertValueInstruction(Context, Context.getNullValue(MemberTy), Val, Idxs+1, NumIdx-1) : Context.getNullValue(MemberTy); Ops[i] = const_cast(Op); } if (isa(AggTy)) return Context.getConstantStruct(Ops); else return Context.getConstantArray(cast(AggTy), Ops); } if (isa(Agg) || isa(Agg)) { // Insertion of constant into aggregate constant std::vector Ops(Agg->getNumOperands()); for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { const Constant *Op = (*Idxs == i) ? ConstantFoldInsertValueInstruction(Context, Agg->getOperand(i), Val, Idxs+1, NumIdx-1) : Agg->getOperand(i); Ops[i] = const_cast(Op); } Constant *C; if (isa(Agg->getType())) C = Context.getConstantStruct(Ops); else C = Context.getConstantArray(cast(Agg->getType()), Ops); return C; } return 0; } Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, unsigned Opcode, const Constant *C1, const Constant *C2) { // No compile-time operations on this type yet. if (C1->getType() == Type::PPC_FP128Ty) return 0; // Handle UndefValue up front if (isa(C1) || isa(C2)) { switch (Opcode) { case Instruction::Xor: if (isa(C1) && isa(C2)) // Handle undef ^ undef -> 0 special case. This is a common // idiom (misuse). return Context.getNullValue(C1->getType()); // Fallthrough case Instruction::Add: case Instruction::Sub: return Context.getUndef(C1->getType()); case Instruction::Mul: case Instruction::And: return Context.getNullValue(C1->getType()); case Instruction::UDiv: case Instruction::SDiv: case Instruction::URem: case Instruction::SRem: if (!isa(C2)) // undef / X -> 0 return Context.getNullValue(C1->getType()); return const_cast(C2); // X / undef -> undef case Instruction::Or: // X | undef -> -1 if (const VectorType *PTy = dyn_cast(C1->getType())) return Context.getAllOnesValue(PTy); return Context.getAllOnesValue(C1->getType()); case Instruction::LShr: if (isa(C2) && isa(C1)) return const_cast(C1); // undef lshr undef -> undef return Context.getNullValue(C1->getType()); // X lshr undef -> 0 // undef lshr X -> 0 case Instruction::AShr: if (!isa(C2)) return const_cast(C1); // undef ashr X --> undef else if (isa(C1)) return const_cast(C1); // undef ashr undef -> undef else return const_cast(C1); // X ashr undef --> X case Instruction::Shl: // undef << X -> 0 or X << undef -> 0 return Context.getNullValue(C1->getType()); } } // Handle simplifications when the RHS is a constant int. if (const ConstantInt *CI2 = dyn_cast(C2)) { switch (Opcode) { case Instruction::Add: if (CI2->equalsInt(0)) return const_cast(C1); // X + 0 == X break; case Instruction::Sub: if (CI2->equalsInt(0)) return const_cast(C1); // X - 0 == X break; case Instruction::Mul: if (CI2->equalsInt(0)) return const_cast(C2); // X * 0 == 0 if (CI2->equalsInt(1)) return const_cast(C1); // X * 1 == X break; case Instruction::UDiv: case Instruction::SDiv: if (CI2->equalsInt(1)) return const_cast(C1); // X / 1 == X if (CI2->equalsInt(0)) return Context.getUndef(CI2->getType()); // X / 0 == undef break; case Instruction::URem: case Instruction::SRem: if (CI2->equalsInt(1)) return Context.getNullValue(CI2->getType()); // X % 1 == 0 if (CI2->equalsInt(0)) return Context.getUndef(CI2->getType()); // X % 0 == undef break; case Instruction::And: if (CI2->isZero()) return const_cast(C2); // X & 0 == 0 if (CI2->isAllOnesValue()) return const_cast(C1); // X & -1 == X if (const ConstantExpr *CE1 = dyn_cast(C1)) { // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) if (CE1->getOpcode() == Instruction::ZExt) { unsigned DstWidth = CI2->getType()->getBitWidth(); unsigned SrcWidth = CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) return const_cast(C1); } // If and'ing the address of a global with a constant, fold it. if (CE1->getOpcode() == Instruction::PtrToInt && isa(CE1->getOperand(0))) { GlobalValue *GV = cast(CE1->getOperand(0)); // Functions are at least 4-byte aligned. unsigned GVAlign = GV->getAlignment(); if (isa(GV)) GVAlign = std::max(GVAlign, 4U); if (GVAlign > 1) { unsigned DstWidth = CI2->getType()->getBitWidth(); unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); // If checking bits we know are clear, return zero. if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) return Context.getNullValue(CI2->getType()); } } } break; case Instruction::Or: if (CI2->equalsInt(0)) return const_cast(C1); // X | 0 == X if (CI2->isAllOnesValue()) return const_cast(C2); // X | -1 == -1 break; case Instruction::Xor: if (CI2->equalsInt(0)) return const_cast(C1); // X ^ 0 == X break; case Instruction::AShr: // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 if (const ConstantExpr *CE1 = dyn_cast(C1)) if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. return Context.getConstantExprLShr(const_cast(C1), const_cast(C2)); break; } } // At this point we know neither constant is an UndefValue. if (const ConstantInt *CI1 = dyn_cast(C1)) { if (const ConstantInt *CI2 = dyn_cast(C2)) { using namespace APIntOps; const APInt &C1V = CI1->getValue(); const APInt &C2V = CI2->getValue(); switch (Opcode) { default: break; case Instruction::Add: return Context.getConstantInt(C1V + C2V); case Instruction::Sub: return Context.getConstantInt(C1V - C2V); case Instruction::Mul: return Context.getConstantInt(C1V * C2V); case Instruction::UDiv: assert(!CI2->isNullValue() && "Div by zero handled above"); return Context.getConstantInt(C1V.udiv(C2V)); case Instruction::SDiv: assert(!CI2->isNullValue() && "Div by zero handled above"); if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) return Context.getUndef(CI1->getType()); // MIN_INT / -1 -> undef return Context.getConstantInt(C1V.sdiv(C2V)); case Instruction::URem: assert(!CI2->isNullValue() && "Div by zero handled above"); return Context.getConstantInt(C1V.urem(C2V)); case Instruction::SRem: assert(!CI2->isNullValue() && "Div by zero handled above"); if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) return Context.getUndef(CI1->getType()); // MIN_INT % -1 -> undef return Context.getConstantInt(C1V.srem(C2V)); case Instruction::And: return Context.getConstantInt(C1V & C2V); case Instruction::Or: return Context.getConstantInt(C1V | C2V); case Instruction::Xor: return Context.getConstantInt(C1V ^ C2V); case Instruction::Shl: { uint32_t shiftAmt = C2V.getZExtValue(); if (shiftAmt < C1V.getBitWidth()) return Context.getConstantInt(C1V.shl(shiftAmt)); else return Context.getUndef(C1->getType()); // too big shift is undef } case Instruction::LShr: { uint32_t shiftAmt = C2V.getZExtValue(); if (shiftAmt < C1V.getBitWidth()) return Context.getConstantInt(C1V.lshr(shiftAmt)); else return Context.getUndef(C1->getType()); // too big shift is undef } case Instruction::AShr: { uint32_t shiftAmt = C2V.getZExtValue(); if (shiftAmt < C1V.getBitWidth()) return Context.getConstantInt(C1V.ashr(shiftAmt)); else return Context.getUndef(C1->getType()); // too big shift is undef } } } switch (Opcode) { case Instruction::SDiv: case Instruction::UDiv: case Instruction::URem: case Instruction::SRem: case Instruction::LShr: case Instruction::AShr: case Instruction::Shl: if (CI1->equalsInt(0)) return const_cast(C1); break; default: break; } } else if (const ConstantFP *CFP1 = dyn_cast(C1)) { if (const ConstantFP *CFP2 = dyn_cast(C2)) { APFloat C1V = CFP1->getValueAPF(); APFloat C2V = CFP2->getValueAPF(); APFloat C3V = C1V; // copy for modification switch (Opcode) { default: break; case Instruction::FAdd: (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); return Context.getConstantFP(C3V); case Instruction::FSub: (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); return Context.getConstantFP(C3V); case Instruction::FMul: (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); return Context.getConstantFP(C3V); case Instruction::FDiv: (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); return Context.getConstantFP(C3V); case Instruction::FRem: (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); return Context.getConstantFP(C3V); } } } else if (const VectorType *VTy = dyn_cast(C1->getType())) { const ConstantVector *CP1 = dyn_cast(C1); const ConstantVector *CP2 = dyn_cast(C2); if ((CP1 != NULL || isa(C1)) && (CP2 != NULL || isa(C2))) { std::vector Res; const Type* EltTy = VTy->getElementType(); const Constant *C1 = 0; const Constant *C2 = 0; switch (Opcode) { default: break; case Instruction::Add: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprAdd(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::FAdd: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprFAdd(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::Sub: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprSub(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::FSub: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprFSub(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::Mul: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprMul(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::FMul: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprFMul(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::UDiv: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprUDiv(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::SDiv: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprSDiv(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::FDiv: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprFDiv(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::URem: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprURem(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::SRem: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprSRem(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::FRem: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprFRem(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::And: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprAnd(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::Or: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprOr(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::Xor: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprXor(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::LShr: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprLShr(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::AShr: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprAShr(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); case Instruction::Shl: for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); Res.push_back(Context.getConstantExprShl(const_cast(C1), const_cast(C2))); } return Context.getConstantVector(Res); } } } if (isa(C1)) { // There are many possible foldings we could do here. We should probably // at least fold add of a pointer with an integer into the appropriate // getelementptr. This will improve alias analysis a bit. } else if (isa(C2)) { // If C2 is a constant expr and C1 isn't, flop them around and fold the // other way if possible. switch (Opcode) { case Instruction::Add: case Instruction::FAdd: case Instruction::Mul: case Instruction::FMul: case Instruction::And: case Instruction::Or: case Instruction::Xor: // No change of opcode required. return ConstantFoldBinaryInstruction(Context, Opcode, C2, C1); case Instruction::Shl: case Instruction::LShr: case Instruction::AShr: case Instruction::Sub: case Instruction::FSub: case Instruction::SDiv: case Instruction::UDiv: case Instruction::FDiv: case Instruction::URem: case Instruction::SRem: case Instruction::FRem: default: // These instructions cannot be flopped around. break; } } // We don't know how to fold this. return 0; } /// isZeroSizedType - This type is zero sized if its an array or structure of /// zero sized types. The only leaf zero sized type is an empty structure. static bool isMaybeZeroSizedType(const Type *Ty) { if (isa(Ty)) return true; // Can't say. if (const StructType *STy = dyn_cast(Ty)) { // If all of elements have zero size, this does too. for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; return true; } else if (const ArrayType *ATy = dyn_cast(Ty)) { return isMaybeZeroSizedType(ATy->getElementType()); } return false; } /// IdxCompare - Compare the two constants as though they were getelementptr /// indices. This allows coersion of the types to be the same thing. /// /// If the two constants are the "same" (after coersion), return 0. If the /// first is less than the second, return -1, if the second is less than the /// first, return 1. If the constants are not integral, return -2. /// static int IdxCompare(LLVMContext &Context, Constant *C1, Constant *C2, const Type *ElTy) { if (C1 == C2) return 0; // Ok, we found a different index. If they are not ConstantInt, we can't do // anything with them. if (!isa(C1) || !isa(C2)) return -2; // don't know! // Ok, we have two differing integer indices. Sign extend them to be the same // type. Long is always big enough, so we use it. if (C1->getType() != Type::Int64Ty) C1 = Context.getConstantExprSExt(C1, Type::Int64Ty); if (C2->getType() != Type::Int64Ty) C2 = Context.getConstantExprSExt(C2, Type::Int64Ty); if (C1 == C2) return 0; // They are equal // If the type being indexed over is really just a zero sized type, there is // no pointer difference being made here. if (isMaybeZeroSizedType(ElTy)) return -2; // dunno. // If they are really different, now that they are the same type, then we // found a difference! if (cast(C1)->getSExtValue() < cast(C2)->getSExtValue()) return -1; else return 1; } /// evaluateFCmpRelation - This function determines if there is anything we can /// decide about the two constants provided. This doesn't need to handle simple /// things like ConstantFP comparisons, but should instead handle ConstantExprs. /// If we can determine that the two constants have a particular relation to /// each other, we should return the corresponding FCmpInst predicate, /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in /// ConstantFoldCompareInstruction. /// /// To simplify this code we canonicalize the relation so that the first /// operand is always the most "complex" of the two. We consider ConstantFP /// to be the simplest, and ConstantExprs to be the most complex. static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, const Constant *V1, const Constant *V2) { assert(V1->getType() == V2->getType() && "Cannot compare values of different types!"); // No compile-time operations on this type yet. if (V1->getType() == Type::PPC_FP128Ty) return FCmpInst::BAD_FCMP_PREDICATE; // Handle degenerate case quickly if (V1 == V2) return FCmpInst::FCMP_OEQ; if (!isa(V1)) { if (!isa(V2)) { // We distilled thisUse the standard constant folder for a few cases ConstantInt *R = 0; Constant *C1 = const_cast(V1); Constant *C2 = const_cast(V2); R = dyn_cast( Context.getConstantExprFCmp(FCmpInst::FCMP_OEQ, C1, C2)); if (R && !R->isZero()) return FCmpInst::FCMP_OEQ; R = dyn_cast( Context.getConstantExprFCmp(FCmpInst::FCMP_OLT, C1, C2)); if (R && !R->isZero()) return FCmpInst::FCMP_OLT; R = dyn_cast( Context.getConstantExprFCmp(FCmpInst::FCMP_OGT, C1, C2)); if (R && !R->isZero()) return FCmpInst::FCMP_OGT; // Nothing more we can do return FCmpInst::BAD_FCMP_PREDICATE; } // If the first operand is simple and second is ConstantExpr, swap operands. FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(Context, V2, V1); if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) return FCmpInst::getSwappedPredicate(SwappedRelation); } else { // Ok, the LHS is known to be a constantexpr. The RHS can be any of a // constantexpr or a simple constant. const ConstantExpr *CE1 = cast(V1); switch (CE1->getOpcode()) { case Instruction::FPTrunc: case Instruction::FPExt: case Instruction::UIToFP: case Instruction::SIToFP: // We might be able to do something with these but we don't right now. break; default: break; } } // There are MANY other foldings that we could perform here. They will // probably be added on demand, as they seem needed. return FCmpInst::BAD_FCMP_PREDICATE; } /// evaluateICmpRelation - This function determines if there is anything we can /// decide about the two constants provided. This doesn't need to handle simple /// things like integer comparisons, but should instead handle ConstantExprs /// and GlobalValues. If we can determine that the two constants have a /// particular relation to each other, we should return the corresponding ICmp /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. /// /// To simplify this code we canonicalize the relation so that the first /// operand is always the most "complex" of the two. We consider simple /// constants (like ConstantInt) to be the simplest, followed by /// GlobalValues, followed by ConstantExpr's (the most complex). /// static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, const Constant *V1, const Constant *V2, bool isSigned) { assert(V1->getType() == V2->getType() && "Cannot compare different types of values!"); if (V1 == V2) return ICmpInst::ICMP_EQ; if (!isa(V1) && !isa(V1)) { if (!isa(V2) && !isa(V2)) { // We distilled this down to a simple case, use the standard constant // folder. ConstantInt *R = 0; Constant *C1 = const_cast(V1); Constant *C2 = const_cast(V2); ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; R = dyn_cast(Context.getConstantExprICmp(pred, C1, C2)); if (R && !R->isZero()) return pred; pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; R = dyn_cast(Context.getConstantExprICmp(pred, C1, C2)); if (R && !R->isZero()) return pred; pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; R = dyn_cast(Context.getConstantExprICmp(pred, C1, C2)); if (R && !R->isZero()) return pred; // If we couldn't figure it out, bail. return ICmpInst::BAD_ICMP_PREDICATE; } // If the first operand is simple, swap operands. ICmpInst::Predicate SwappedRelation = evaluateICmpRelation(Context, V2, V1, isSigned); if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) return ICmpInst::getSwappedPredicate(SwappedRelation); } else if (const GlobalValue *CPR1 = dyn_cast(V1)) { if (isa(V2)) { // Swap as necessary. ICmpInst::Predicate SwappedRelation = evaluateICmpRelation(Context, V2, V1, isSigned); if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) return ICmpInst::getSwappedPredicate(SwappedRelation); else return ICmpInst::BAD_ICMP_PREDICATE; } // Now we know that the RHS is a GlobalValue or simple constant, // which (since the types must match) means that it's a ConstantPointerNull. if (const GlobalValue *CPR2 = dyn_cast(V2)) { // Don't try to decide equality of aliases. if (!isa(CPR1) && !isa(CPR2)) if (!CPR1->hasExternalWeakLinkage() || !CPR2->hasExternalWeakLinkage()) return ICmpInst::ICMP_NE; } else { assert(isa(V2) && "Canonicalization guarantee!"); // GlobalVals can never be null. Don't try to evaluate aliases. if (!CPR1->hasExternalWeakLinkage() && !isa(CPR1)) return ICmpInst::ICMP_NE; } } else { // Ok, the LHS is known to be a constantexpr. The RHS can be any of a // constantexpr, a CPR, or a simple constant. const ConstantExpr *CE1 = cast(V1); const Constant *CE1Op0 = CE1->getOperand(0); switch (CE1->getOpcode()) { case Instruction::Trunc: case Instruction::FPTrunc: case Instruction::FPExt: case Instruction::FPToUI: case Instruction::FPToSI: break; // We can't evaluate floating point casts or truncations. case Instruction::UIToFP: case Instruction::SIToFP: case Instruction::BitCast: case Instruction::ZExt: case Instruction::SExt: // If the cast is not actually changing bits, and the second operand is a // null pointer, do the comparison with the pre-casted value. if (V2->isNullValue() && (isa(CE1->getType()) || CE1->getType()->isInteger())) { bool sgnd = isSigned; if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; if (CE1->getOpcode() == Instruction::SExt) isSigned = true; return evaluateICmpRelation(Context, CE1Op0, Context.getNullValue(CE1Op0->getType()), sgnd); } // If the dest type is a pointer type, and the RHS is a constantexpr cast // from the same type as the src of the LHS, evaluate the inputs. This is // important for things like "icmp eq (cast 4 to int*), (cast 5 to int*)", // which happens a lot in compilers with tagged integers. if (const ConstantExpr *CE2 = dyn_cast(V2)) if (CE2->isCast() && isa(CE1->getType()) && CE1->getOperand(0)->getType() == CE2->getOperand(0)->getType() && CE1->getOperand(0)->getType()->isInteger()) { bool sgnd = isSigned; if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; if (CE1->getOpcode() == Instruction::SExt) isSigned = true; return evaluateICmpRelation(Context, CE1->getOperand(0), CE2->getOperand(0), sgnd); } break; case Instruction::GetElementPtr: // Ok, since this is a getelementptr, we know that the constant has a // pointer type. Check the various cases. if (isa(V2)) { // If we are comparing a GEP to a null pointer, check to see if the base // of the GEP equals the null pointer. if (const GlobalValue *GV = dyn_cast(CE1Op0)) { if (GV->hasExternalWeakLinkage()) // Weak linkage GVals could be zero or not. We're comparing that // to null pointer so its greater-or-equal return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; else // If its not weak linkage, the GVal must have a non-zero address // so the result is greater-than return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; } else if (isa(CE1Op0)) { // If we are indexing from a null pointer, check to see if we have any // non-zero indices. for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) if (!CE1->getOperand(i)->isNullValue()) // Offsetting from null, must not be equal. return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; // Only zero indexes from null, must still be zero. return ICmpInst::ICMP_EQ; } // Otherwise, we can't really say if the first operand is null or not. } else if (const GlobalValue *CPR2 = dyn_cast(V2)) { if (isa(CE1Op0)) { if (CPR2->hasExternalWeakLinkage()) // Weak linkage GVals could be zero or not. We're comparing it to // a null pointer, so its less-or-equal return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; else // If its not weak linkage, the GVal must have a non-zero address // so the result is less-than return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; } else if (const GlobalValue *CPR1 = dyn_cast(CE1Op0)) { if (CPR1 == CPR2) { // If this is a getelementptr of the same global, then it must be // different. Because the types must match, the getelementptr could // only have at most one index, and because we fold getelementptr's // with a single zero index, it must be nonzero. assert(CE1->getNumOperands() == 2 && !CE1->getOperand(1)->isNullValue() && "Suprising getelementptr!"); return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; } else { // If they are different globals, we don't know what the value is, // but they can't be equal. return ICmpInst::ICMP_NE; } } } else { const ConstantExpr *CE2 = cast(V2); const Constant *CE2Op0 = CE2->getOperand(0); // There are MANY other foldings that we could perform here. They will // probably be added on demand, as they seem needed. switch (CE2->getOpcode()) { default: break; case Instruction::GetElementPtr: // By far the most common case to handle is when the base pointers are // obviously to the same or different globals. if (isa(CE1Op0) && isa(CE2Op0)) { if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal return ICmpInst::ICMP_NE; // Ok, we know that both getelementptr instructions are based on the // same global. From this, we can precisely determine the relative // ordering of the resultant pointers. unsigned i = 1; // Compare all of the operands the GEP's have in common. gep_type_iterator GTI = gep_type_begin(CE1); for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); ++i, ++GTI) switch (IdxCompare(Context, CE1->getOperand(i), CE2->getOperand(i), GTI.getIndexedType())) { case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; case -2: return ICmpInst::BAD_ICMP_PREDICATE; } // Ok, we ran out of things they have in common. If any leftovers // are non-zero then we have a difference, otherwise we are equal. for (; i < CE1->getNumOperands(); ++i) if (!CE1->getOperand(i)->isNullValue()) { if (isa(CE1->getOperand(i))) return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; else return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. } for (; i < CE2->getNumOperands(); ++i) if (!CE2->getOperand(i)->isNullValue()) { if (isa(CE2->getOperand(i))) return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; else return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. } return ICmpInst::ICMP_EQ; } } } default: break; } } return ICmpInst::BAD_ICMP_PREDICATE; } Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, unsigned short pred, const Constant *C1, const Constant *C2) { const Type *ResultTy; if (const VectorType *VT = dyn_cast(C1->getType())) ResultTy = Context.getVectorType(Type::Int1Ty, VT->getNumElements()); else ResultTy = Type::Int1Ty; // Fold FCMP_FALSE/FCMP_TRUE unconditionally. if (pred == FCmpInst::FCMP_FALSE) return Context.getNullValue(ResultTy); if (pred == FCmpInst::FCMP_TRUE) return Context.getAllOnesValue(ResultTy); // Handle some degenerate cases first if (isa(C1) || isa(C2)) return Context.getUndef(ResultTy); // No compile-time operations on this type yet. if (C1->getType() == Type::PPC_FP128Ty) return 0; // icmp eq/ne(null,GV) -> false/true if (C1->isNullValue()) { if (const GlobalValue *GV = dyn_cast(C2)) // Don't try to evaluate aliases. External weak GV can be null. if (!isa(GV) && !GV->hasExternalWeakLinkage()) { if (pred == ICmpInst::ICMP_EQ) return Context.getConstantIntFalse(); else if (pred == ICmpInst::ICMP_NE) return Context.getConstantIntTrue(); } // icmp eq/ne(GV,null) -> false/true } else if (C2->isNullValue()) { if (const GlobalValue *GV = dyn_cast(C1)) // Don't try to evaluate aliases. External weak GV can be null. if (!isa(GV) && !GV->hasExternalWeakLinkage()) { if (pred == ICmpInst::ICMP_EQ) return Context.getConstantIntFalse(); else if (pred == ICmpInst::ICMP_NE) return Context.getConstantIntTrue(); } } if (isa(C1) && isa(C2)) { APInt V1 = cast(C1)->getValue(); APInt V2 = cast(C2)->getValue(); switch (pred) { default: LLVM_UNREACHABLE("Invalid ICmp Predicate"); return 0; case ICmpInst::ICMP_EQ: return Context.getConstantInt(Type::Int1Ty, V1 == V2); case ICmpInst::ICMP_NE: return Context.getConstantInt(Type::Int1Ty, V1 != V2); case ICmpInst::ICMP_SLT: return Context.getConstantInt(Type::Int1Ty, V1.slt(V2)); case ICmpInst::ICMP_SGT: return Context.getConstantInt(Type::Int1Ty, V1.sgt(V2)); case ICmpInst::ICMP_SLE: return Context.getConstantInt(Type::Int1Ty, V1.sle(V2)); case ICmpInst::ICMP_SGE: return Context.getConstantInt(Type::Int1Ty, V1.sge(V2)); case ICmpInst::ICMP_ULT: return Context.getConstantInt(Type::Int1Ty, V1.ult(V2)); case ICmpInst::ICMP_UGT: return Context.getConstantInt(Type::Int1Ty, V1.ugt(V2)); case ICmpInst::ICMP_ULE: return Context.getConstantInt(Type::Int1Ty, V1.ule(V2)); case ICmpInst::ICMP_UGE: return Context.getConstantInt(Type::Int1Ty, V1.uge(V2)); } } else if (isa(C1) && isa(C2)) { APFloat C1V = cast(C1)->getValueAPF(); APFloat C2V = cast(C2)->getValueAPF(); APFloat::cmpResult R = C1V.compare(C2V); switch (pred) { default: LLVM_UNREACHABLE("Invalid FCmp Predicate"); return 0; case FCmpInst::FCMP_FALSE: return Context.getConstantIntFalse(); case FCmpInst::FCMP_TRUE: return Context.getConstantIntTrue(); case FCmpInst::FCMP_UNO: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpUnordered); case FCmpInst::FCMP_ORD: return Context.getConstantInt(Type::Int1Ty, R!=APFloat::cmpUnordered); case FCmpInst::FCMP_UEQ: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpUnordered || R==APFloat::cmpEqual); case FCmpInst::FCMP_OEQ: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpEqual); case FCmpInst::FCMP_UNE: return Context.getConstantInt(Type::Int1Ty, R!=APFloat::cmpEqual); case FCmpInst::FCMP_ONE: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpLessThan || R==APFloat::cmpGreaterThan); case FCmpInst::FCMP_ULT: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpUnordered || R==APFloat::cmpLessThan); case FCmpInst::FCMP_OLT: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpLessThan); case FCmpInst::FCMP_UGT: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpUnordered || R==APFloat::cmpGreaterThan); case FCmpInst::FCMP_OGT: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpGreaterThan); case FCmpInst::FCMP_ULE: return Context.getConstantInt(Type::Int1Ty, R!=APFloat::cmpGreaterThan); case FCmpInst::FCMP_OLE: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpLessThan || R==APFloat::cmpEqual); case FCmpInst::FCMP_UGE: return Context.getConstantInt(Type::Int1Ty, R!=APFloat::cmpLessThan); case FCmpInst::FCMP_OGE: return Context.getConstantInt(Type::Int1Ty, R==APFloat::cmpGreaterThan || R==APFloat::cmpEqual); } } else if (isa(C1->getType())) { SmallVector C1Elts, C2Elts; C1->getVectorElements(Context, C1Elts); C2->getVectorElements(Context, C2Elts); // If we can constant fold the comparison of each element, constant fold // the whole vector comparison. SmallVector ResElts; for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) { // Compare the elements, producing an i1 result or constant expr. ResElts.push_back( Context.getConstantExprCompare(pred, C1Elts[i], C2Elts[i])); } return Context.getConstantVector(&ResElts[0], ResElts.size()); } if (C1->getType()->isFloatingPoint()) { int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. switch (evaluateFCmpRelation(Context, C1, C2)) { default: LLVM_UNREACHABLE("Unknown relation!"); case FCmpInst::FCMP_UNO: case FCmpInst::FCMP_ORD: case FCmpInst::FCMP_UEQ: case FCmpInst::FCMP_UNE: case FCmpInst::FCMP_ULT: case FCmpInst::FCMP_UGT: case FCmpInst::FCMP_ULE: case FCmpInst::FCMP_UGE: case FCmpInst::FCMP_TRUE: case FCmpInst::FCMP_FALSE: case FCmpInst::BAD_FCMP_PREDICATE: break; // Couldn't determine anything about these constants. case FCmpInst::FCMP_OEQ: // We know that C1 == C2 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); break; case FCmpInst::FCMP_OLT: // We know that C1 < C2 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); break; case FCmpInst::FCMP_OGT: // We know that C1 > C2 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); break; case FCmpInst::FCMP_OLE: // We know that C1 <= C2 // We can only partially decide this relation. if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) Result = 0; else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) Result = 1; break; case FCmpInst::FCMP_OGE: // We known that C1 >= C2 // We can only partially decide this relation. if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) Result = 0; else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) Result = 1; break; case ICmpInst::ICMP_NE: // We know that C1 != C2 // We can only partially decide this relation. if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) Result = 0; else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) Result = 1; break; } // If we evaluated the result, return it now. if (Result != -1) return Context.getConstantInt(Type::Int1Ty, Result); } else { // Evaluate the relation between the two constants, per the predicate. int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. switch (evaluateICmpRelation(Context, C1, C2, CmpInst::isSigned(pred))) { default: LLVM_UNREACHABLE("Unknown relational!"); case ICmpInst::BAD_ICMP_PREDICATE: break; // Couldn't determine anything about these constants. case ICmpInst::ICMP_EQ: // We know the constants are equal! // If we know the constants are equal, we can decide the result of this // computation precisely. Result = (pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_ULE || pred == ICmpInst::ICMP_SLE || pred == ICmpInst::ICMP_UGE || pred == ICmpInst::ICMP_SGE); break; case ICmpInst::ICMP_ULT: // If we know that C1 < C2, we can decide the result of this computation // precisely. Result = (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_NE || pred == ICmpInst::ICMP_ULE); break; case ICmpInst::ICMP_SLT: // If we know that C1 < C2, we can decide the result of this computation // precisely. Result = (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_NE || pred == ICmpInst::ICMP_SLE); break; case ICmpInst::ICMP_UGT: // If we know that C1 > C2, we can decide the result of this computation // precisely. Result = (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_NE || pred == ICmpInst::ICMP_UGE); break; case ICmpInst::ICMP_SGT: // If we know that C1 > C2, we can decide the result of this computation // precisely. Result = (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_NE || pred == ICmpInst::ICMP_SGE); break; case ICmpInst::ICMP_ULE: // If we know that C1 <= C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_UGT) Result = 0; if (pred == ICmpInst::ICMP_ULT) Result = 1; break; case ICmpInst::ICMP_SLE: // If we know that C1 <= C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_SGT) Result = 0; if (pred == ICmpInst::ICMP_SLT) Result = 1; break; case ICmpInst::ICMP_UGE: // If we know that C1 >= C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_ULT) Result = 0; if (pred == ICmpInst::ICMP_UGT) Result = 1; break; case ICmpInst::ICMP_SGE: // If we know that C1 >= C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_SLT) Result = 0; if (pred == ICmpInst::ICMP_SGT) Result = 1; break; case ICmpInst::ICMP_NE: // If we know that C1 != C2, we can only partially decide this relation. if (pred == ICmpInst::ICMP_EQ) Result = 0; if (pred == ICmpInst::ICMP_NE) Result = 1; break; } // If we evaluated the result, return it now. if (Result != -1) return Context.getConstantInt(Type::Int1Ty, Result); if (!isa(C1) && isa(C2)) { // If C2 is a constant expr and C1 isn't, flip them around and fold the // other way if possible. switch (pred) { case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_NE: // No change of predicate required. return ConstantFoldCompareInstruction(Context, pred, C2, C1); case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_ULE: case ICmpInst::ICMP_SLE: case ICmpInst::ICMP_UGE: case ICmpInst::ICMP_SGE: // Change the predicate as necessary to swap the operands. pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); return ConstantFoldCompareInstruction(Context, pred, C2, C1); default: // These predicates cannot be flopped around. break; } } } return 0; } Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, const Constant *C, Constant* const *Idxs, unsigned NumIdx) { if (NumIdx == 0 || (NumIdx == 1 && Idxs[0]->isNullValue())) return const_cast(C); if (isa(C)) { const PointerType *Ptr = cast(C->getType()); const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, (Value **)Idxs, (Value **)Idxs+NumIdx); assert(Ty != 0 && "Invalid indices for GEP!"); return Context.getUndef(Context.getPointerType(Ty, Ptr->getAddressSpace())); } Constant *Idx0 = Idxs[0]; if (C->isNullValue()) { bool isNull = true; for (unsigned i = 0, e = NumIdx; i != e; ++i) if (!Idxs[i]->isNullValue()) { isNull = false; break; } if (isNull) { const PointerType *Ptr = cast(C->getType()); const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, (Value**)Idxs, (Value**)Idxs+NumIdx); assert(Ty != 0 && "Invalid indices for GEP!"); return Context.getConstantPointerNull( Context.getPointerType(Ty,Ptr->getAddressSpace())); } } if (ConstantExpr *CE = dyn_cast(const_cast(C))) { // Combine Indices - If the source pointer to this getelementptr instruction // is a getelementptr instruction, combine the indices of the two // getelementptr instructions into a single instruction. // if (CE->getOpcode() == Instruction::GetElementPtr) { const Type *LastTy = 0; for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); I != E; ++I) LastTy = *I; if ((LastTy && isa(LastTy)) || Idx0->isNullValue()) { SmallVector NewIndices; NewIndices.reserve(NumIdx + CE->getNumOperands()); for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) NewIndices.push_back(CE->getOperand(i)); // Add the last index of the source with the first index of the new GEP. // Make sure to handle the case when they are actually different types. Constant *Combined = CE->getOperand(CE->getNumOperands()-1); // Otherwise it must be an array. if (!Idx0->isNullValue()) { const Type *IdxTy = Combined->getType(); if (IdxTy != Idx0->getType()) { Constant *C1 = Context.getConstantExprSExtOrBitCast(Idx0, Type::Int64Ty); Constant *C2 = Context.getConstantExprSExtOrBitCast(Combined, Type::Int64Ty); Combined = Context.getConstantExpr(Instruction::Add, C1, C2); } else { Combined = Context.getConstantExpr(Instruction::Add, Idx0, Combined); } } NewIndices.push_back(Combined); NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx); return Context.getConstantExprGetElementPtr(CE->getOperand(0), &NewIndices[0], NewIndices.size()); } } // Implement folding of: // int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*), // long 0, long 0) // To: int* getelementptr ([3 x int]* %X, long 0, long 0) // if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) { if (const PointerType *SPT = dyn_cast(CE->getOperand(0)->getType())) if (const ArrayType *SAT = dyn_cast(SPT->getElementType())) if (const ArrayType *CAT = dyn_cast(cast(C->getType())->getElementType())) if (CAT->getElementType() == SAT->getElementType()) return Context.getConstantExprGetElementPtr( (Constant*)CE->getOperand(0), Idxs, NumIdx); } // Fold: getelementptr (i8* inttoptr (i64 1 to i8*), i32 -1) // Into: inttoptr (i64 0 to i8*) // This happens with pointers to member functions in C++. if (CE->getOpcode() == Instruction::IntToPtr && NumIdx == 1 && isa(CE->getOperand(0)) && isa(Idxs[0]) && cast(CE->getType())->getElementType() == Type::Int8Ty) { Constant *Base = CE->getOperand(0); Constant *Offset = Idxs[0]; // Convert the smaller integer to the larger type. if (Offset->getType()->getPrimitiveSizeInBits() < Base->getType()->getPrimitiveSizeInBits()) Offset = Context.getConstantExprSExt(Offset, Base->getType()); else if (Base->getType()->getPrimitiveSizeInBits() < Offset->getType()->getPrimitiveSizeInBits()) Base = Context.getConstantExprZExt(Base, Offset->getType()); Base = Context.getConstantExprAdd(Base, Offset); return Context.getConstantExprIntToPtr(Base, CE->getType()); } } return 0; }