X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTransforms%2FInstCombine%2FInstCombineCasts.cpp;h=b41cdc65202f6dc6edbc00765a58b0c7dea9785b;hp=6c42c0c66003112a9b6c3ca68755d57fcd09c02a;hb=fe58be3733715dd30807e9aaa31500be456ecf53;hpb=be6ceb6ebcfb48e09ffab297c1537c87383f0ce9 diff --git a/lib/Transforms/InstCombine/InstCombineCasts.cpp b/lib/Transforms/InstCombine/InstCombineCasts.cpp index 6c42c0c6600..b41cdc65202 100644 --- a/lib/Transforms/InstCombine/InstCombineCasts.cpp +++ b/lib/Transforms/InstCombine/InstCombineCasts.cpp @@ -13,11 +13,14 @@ #include "InstCombine.h" #include "llvm/Analysis/ConstantFolding.h" -#include "llvm/Target/TargetData.h" -#include "llvm/Support/PatternMatch.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/PatternMatch.h" +#include "llvm/Target/TargetLibraryInfo.h" using namespace llvm; using namespace PatternMatch; +#define DEBUG_TYPE "instcombine" + /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear /// expression. If so, decompose it, returning some value X, such that Val is /// X*Scale+Offset. @@ -29,11 +32,11 @@ static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale, Scale = 0; return ConstantInt::get(Val->getType(), 0); } - + if (BinaryOperator *I = dyn_cast(Val)) { // Cannot look past anything that might overflow. OverflowingBinaryOperator *OBI = dyn_cast(Val); - if (OBI && !OBI->hasNoUnsignedWrap()) { + if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) { Scale = 1; Offset = 0; return Val; @@ -46,19 +49,19 @@ static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale, Offset = 0; return I->getOperand(0); } - + if (I->getOpcode() == Instruction::Mul) { // This value is scaled by 'RHS'. Scale = RHS->getZExtValue(); Offset = 0; return I->getOperand(0); } - + if (I->getOpcode() == Instruction::Add) { - // We have X+C. Check to see if we really have (X*C2)+C1, + // We have X+C. Check to see if we really have (X*C2)+C1, // where C1 is divisible by C2. unsigned SubScale; - Value *SubVal = + Value *SubVal = DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset); Offset += RHS->getZExtValue(); Scale = SubScale; @@ -77,31 +80,37 @@ static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale, /// try to eliminate the cast by moving the type information into the alloc. Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI) { - // This requires TargetData to get the alloca alignment and size information. - if (!TD) return 0; + // This requires DataLayout to get the alloca alignment and size information. + if (!DL) return nullptr; PointerType *PTy = cast(CI.getType()); - + BuilderTy AllocaBuilder(*Builder); AllocaBuilder.SetInsertPoint(AI.getParent(), &AI); // Get the type really allocated and the type casted to. Type *AllocElTy = AI.getAllocatedType(); Type *CastElTy = PTy->getElementType(); - if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0; + if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr; - unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy); - unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy); - if (CastElTyAlign < AllocElTyAlign) return 0; + unsigned AllocElTyAlign = DL->getABITypeAlignment(AllocElTy); + unsigned CastElTyAlign = DL->getABITypeAlignment(CastElTy); + if (CastElTyAlign < AllocElTyAlign) return nullptr; // If the allocation has multiple uses, only promote it if we are strictly // increasing the alignment of the resultant allocation. If we keep it the // same, we open the door to infinite loops of various kinds. - if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0; + if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr; + + uint64_t AllocElTySize = DL->getTypeAllocSize(AllocElTy); + uint64_t CastElTySize = DL->getTypeAllocSize(CastElTy); + if (CastElTySize == 0 || AllocElTySize == 0) return nullptr; - uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy); - uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy); - if (CastElTySize == 0 || AllocElTySize == 0) return 0; + // If the allocation has multiple uses, only promote it if we're not + // shrinking the amount of memory being allocated. + uint64_t AllocElTyStoreSize = DL->getTypeStoreSize(AllocElTy); + uint64_t CastElTyStoreSize = DL->getTypeStoreSize(CastElTy); + if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr; // See if we can satisfy the modulus by pulling a scale out of the array // size argument. @@ -109,32 +118,33 @@ Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, uint64_t ArrayOffset; Value *NumElements = // See if the array size is a decomposable linear expr. DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); - + // If we can now satisfy the modulus, by using a non-1 scale, we really can // do the xform. if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || - (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0; + (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return nullptr; unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; - Value *Amt = 0; + Value *Amt = nullptr; if (Scale == 1) { Amt = NumElements; } else { Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale); // Insert before the alloca, not before the cast. - Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp"); + Amt = AllocaBuilder.CreateMul(Amt, NumElements); } - + if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { Value *Off = ConstantInt::get(AI.getArraySize()->getType(), Offset, true); - Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp"); + Amt = AllocaBuilder.CreateAdd(Amt, Off); } - + AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); New->setAlignment(AI.getAlignment()); New->takeName(&AI); - + New->setUsedWithInAlloca(AI.isUsedWithInAlloca()); + // If the allocation has multiple real uses, insert a cast and change all // things that used it to use the new cast. This will also hack on CI, but it // will die soon. @@ -147,24 +157,22 @@ Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, return ReplaceInstUsesWith(CI, New); } - - -/// EvaluateInDifferentType - Given an expression that +/// EvaluateInDifferentType - Given an expression that /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually /// insert the code to evaluate the expression. -Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, +Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned) { if (Constant *C = dyn_cast(V)) { C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); - // If we got a constantexpr back, try to simplify it with TD info. + // If we got a constantexpr back, try to simplify it with DL info. if (ConstantExpr *CE = dyn_cast(C)) - C = ConstantFoldConstantExpression(CE, TD); + C = ConstantFoldConstantExpression(CE, DL, TLI); return C; } // Otherwise, it must be an instruction. Instruction *I = cast(V); - Instruction *Res = 0; + Instruction *Res = nullptr; unsigned Opc = I->getOpcode(); switch (Opc) { case Instruction::Add: @@ -182,7 +190,7 @@ Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); break; - } + } case Instruction::Trunc: case Instruction::ZExt: case Instruction::SExt: @@ -191,7 +199,7 @@ Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, // new. if (I->getOperand(0)->getType() == Ty) return I->getOperand(0); - + // Otherwise, must be the same type of cast, so just reinsert a new one. // This also handles the case of zext(trunc(x)) -> zext(x). Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, @@ -213,12 +221,11 @@ Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, Res = NPN; break; } - default: + default: // TODO: Can handle more cases here. llvm_unreachable("Unreachable!"); - break; } - + Res->takeName(I); return InsertNewInstWith(Res, *I); } @@ -226,12 +233,12 @@ Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, /// This function is a wrapper around CastInst::isEliminableCastPair. It /// simply extracts arguments and returns what that function returns. -static Instruction::CastOps +static Instruction::CastOps isEliminableCastPair( const CastInst *CI, ///< The first cast instruction unsigned opcode, ///< The opcode of the second cast instruction Type *DstTy, ///< The target type for the second cast instruction - TargetData *TD ///< The target data for pointer size + const DataLayout *DL ///< The target data for pointer size ) { Type *SrcTy = CI->getOperand(0)->getType(); // A from above @@ -240,19 +247,22 @@ isEliminableCastPair( // Get the opcodes of the two Cast instructions Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); Instruction::CastOps secondOp = Instruction::CastOps(opcode); - + Type *SrcIntPtrTy = DL && SrcTy->isPtrOrPtrVectorTy() ? + DL->getIntPtrType(SrcTy) : nullptr; + Type *MidIntPtrTy = DL && MidTy->isPtrOrPtrVectorTy() ? + DL->getIntPtrType(MidTy) : nullptr; + Type *DstIntPtrTy = DL && DstTy->isPtrOrPtrVectorTy() ? + DL->getIntPtrType(DstTy) : nullptr; unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, - DstTy, - TD ? TD->getIntPtrType(CI->getContext()) : 0); - + DstTy, SrcIntPtrTy, MidIntPtrTy, + DstIntPtrTy); + // We don't want to form an inttoptr or ptrtoint that converts to an integer // type that differs from the pointer size. - if ((Res == Instruction::IntToPtr && - (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) || - (Res == Instruction::PtrToInt && - (!TD || DstTy != TD->getIntPtrType(CI->getContext())))) + if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) || + (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy)) Res = 0; - + return Instruction::CastOps(Res); } @@ -264,18 +274,18 @@ bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, Type *Ty) { // Noop casts and casts of constants should be eliminated trivially. if (V->getType() == Ty || isa(V)) return false; - + // If this is another cast that can be eliminated, we prefer to have it // eliminated. if (const CastInst *CI = dyn_cast(V)) - if (isEliminableCastPair(CI, opc, Ty, TD)) + if (isEliminableCastPair(CI, opc, Ty, DL)) return false; - + // If this is a vector sext from a compare, then we don't want to break the // idiom where each element of the extended vector is either zero or all ones. if (opc == Instruction::SExt && isa(V) && Ty->isVectorTy()) return false; - + return true; } @@ -287,8 +297,8 @@ Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { // Many cases of "cast of a cast" are eliminable. If it's eliminable we just // eliminate it now. if (CastInst *CSrc = dyn_cast(Src)) { // A->B->C cast - if (Instruction::CastOps opc = - isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { + if (Instruction::CastOps opc = + isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) { // The first cast (CSrc) is eliminable so we need to fix up or replace // the second cast (CI). CSrc will then have a good chance of being dead. return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); @@ -310,8 +320,8 @@ Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { if (Instruction *NV = FoldOpIntoPhi(CI)) return NV; } - - return 0; + + return nullptr; } /// CanEvaluateTruncated - Return true if we can evaluate the specified @@ -325,19 +335,20 @@ Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { /// /// This function works on both vectors and scalars. /// -static bool CanEvaluateTruncated(Value *V, Type *Ty) { +static bool CanEvaluateTruncated(Value *V, Type *Ty, InstCombiner &IC, + Instruction *CxtI) { // We can always evaluate constants in another type. if (isa(V)) return true; - + Instruction *I = dyn_cast(V); if (!I) return false; - + Type *OrigTy = V->getType(); - + // If this is an extension from the dest type, we can eliminate it, even if it // has multiple uses. - if ((isa(I) || isa(I)) && + if ((isa(I) || isa(I)) && I->getOperand(0)->getType() == Ty) return true; @@ -354,8 +365,8 @@ static bool CanEvaluateTruncated(Value *V, Type *Ty) { case Instruction::Or: case Instruction::Xor: // These operators can all arbitrarily be extended or truncated. - return CanEvaluateTruncated(I->getOperand(0), Ty) && - CanEvaluateTruncated(I->getOperand(1), Ty); + return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) && + CanEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI); case Instruction::UDiv: case Instruction::URem: { @@ -364,10 +375,10 @@ static bool CanEvaluateTruncated(Value *V, Type *Ty) { uint32_t BitWidth = Ty->getScalarSizeInBits(); if (BitWidth < OrigBitWidth) { APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); - if (MaskedValueIsZero(I->getOperand(0), Mask) && - MaskedValueIsZero(I->getOperand(1), Mask)) { - return CanEvaluateTruncated(I->getOperand(0), Ty) && - CanEvaluateTruncated(I->getOperand(1), Ty); + if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) && + IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) { + return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) && + CanEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI); } } break; @@ -378,7 +389,7 @@ static bool CanEvaluateTruncated(Value *V, Type *Ty) { if (ConstantInt *CI = dyn_cast(I->getOperand(1))) { uint32_t BitWidth = Ty->getScalarSizeInBits(); if (CI->getLimitedValue(BitWidth) < BitWidth) - return CanEvaluateTruncated(I->getOperand(0), Ty); + return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI); } break; case Instruction::LShr: @@ -388,10 +399,10 @@ static bool CanEvaluateTruncated(Value *V, Type *Ty) { if (ConstantInt *CI = dyn_cast(I->getOperand(1))) { uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); uint32_t BitWidth = Ty->getScalarSizeInBits(); - if (MaskedValueIsZero(I->getOperand(0), - APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && + if (IC.MaskedValueIsZero(I->getOperand(0), + APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth), 0, CxtI) && CI->getLimitedValue(BitWidth) < BitWidth) { - return CanEvaluateTruncated(I->getOperand(0), Ty); + return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI); } } break; @@ -405,8 +416,8 @@ static bool CanEvaluateTruncated(Value *V, Type *Ty) { return true; case Instruction::Select: { SelectInst *SI = cast(I); - return CanEvaluateTruncated(SI->getTrueValue(), Ty) && - CanEvaluateTruncated(SI->getFalseValue(), Ty); + return CanEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) && + CanEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI); } case Instruction::PHI: { // We can change a phi if we can change all operands. Note that we never @@ -414,7 +425,7 @@ static bool CanEvaluateTruncated(Value *V, Type *Ty) { // instructions with a single use. PHINode *PN = cast(I); for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) - if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) + if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty, IC, CxtI)) return false; return true; } @@ -422,29 +433,29 @@ static bool CanEvaluateTruncated(Value *V, Type *Ty) { // TODO: Can handle more cases here. break; } - + return false; } Instruction *InstCombiner::visitTrunc(TruncInst &CI) { if (Instruction *Result = commonCastTransforms(CI)) return Result; - - // See if we can simplify any instructions used by the input whose sole + + // See if we can simplify any instructions used by the input whose sole // purpose is to compute bits we don't care about. if (SimplifyDemandedInstructionBits(CI)) return &CI; - + Value *Src = CI.getOperand(0); Type *DestTy = CI.getType(), *SrcTy = Src->getType(); - + // Attempt to truncate the entire input expression tree to the destination // type. Only do this if the dest type is a simple type, don't convert the // expression tree to something weird like i93 unless the source is also // strange. if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && - CanEvaluateTruncated(Src, DestTy)) { - + CanEvaluateTruncated(Src, DestTy, *this, &CI)) { + // If this cast is a truncate, evaluting in a different type always // eliminates the cast, so it is always a win. DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" @@ -457,13 +468,13 @@ Instruction *InstCombiner::visitTrunc(TruncInst &CI) { // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. if (DestTy->getScalarSizeInBits() == 1) { Constant *One = ConstantInt::get(Src->getType(), 1); - Src = Builder->CreateAnd(Src, One, "tmp"); + Src = Builder->CreateAnd(Src, One); Value *Zero = Constant::getNullValue(Src->getType()); return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); } - + // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion. - Value *A = 0; ConstantInt *Cst = 0; + Value *A = nullptr; ConstantInt *Cst = nullptr; if (Src->hasOneUse() && match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) { // We have three types to worry about here, the type of A, the source of @@ -471,7 +482,7 @@ Instruction *InstCombiner::visitTrunc(TruncInst &CI) { // ASize < MidSize and MidSize > ResultSize, but don't know the relation // between ASize and ResultSize. unsigned ASize = A->getType()->getPrimitiveSizeInBits(); - + // If the shift amount is larger than the size of A, then the result is // known to be zero because all the input bits got shifted out. if (Cst->getZExtValue() >= ASize) @@ -484,7 +495,7 @@ Instruction *InstCombiner::visitTrunc(TruncInst &CI) { Shift->takeName(Src); return CastInst::CreateIntegerCast(Shift, CI.getType(), false); } - + // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest // type isn't non-native. if (Src->hasOneUse() && isa(Src->getType()) && @@ -495,7 +506,7 @@ Instruction *InstCombiner::visitTrunc(TruncInst &CI) { ConstantExpr::getTrunc(Cst, CI.getType())); } - return 0; + return nullptr; } /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations @@ -507,7 +518,7 @@ Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, // cast to integer to avoid the comparison. if (ConstantInt *Op1C = dyn_cast(ICI->getOperand(1))) { const APInt &Op1CV = Op1C->getValue(); - + // zext (x x>>u31 true if signbit set. // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || @@ -519,7 +530,7 @@ Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, In->getType()->getScalarSizeInBits()-1); In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); if (In->getType() != CI.getType()) - In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp"); + In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/); if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { Constant *One = ConstantInt::get(In->getType(), 1); @@ -528,9 +539,7 @@ Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, return ReplaceInstUsesWith(CI, In); } - - - + // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. // zext (X == 1) to i32 --> X iff X has only the low bit set. @@ -539,15 +548,14 @@ Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. - if ((Op1CV == 0 || Op1CV.isPowerOf2()) && + if ((Op1CV == 0 || Op1CV.isPowerOf2()) && // This only works for EQ and NE ICI->isEquality()) { // If Op1C some other power of two, convert: uint32_t BitWidth = Op1C->getType()->getBitWidth(); APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); - APInt TypeMask(APInt::getAllOnesValue(BitWidth)); - ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne); - + computeKnownBits(ICI->getOperand(0), KnownZero, KnownOne, 0, &CI); + APInt KnownZeroMask(~KnownZero); if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? if (!DoXform) return ICI; @@ -561,7 +569,7 @@ Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, Res = ConstantExpr::getZExt(Res, CI.getType()); return ReplaceInstUsesWith(CI, Res); } - + uint32_t ShiftAmt = KnownZeroMask.logBase2(); Value *In = ICI->getOperand(0); if (ShiftAmt) { @@ -570,12 +578,12 @@ Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), In->getName()+".lobit"); } - + if ((Op1CV != 0) == isNE) { // Toggle the low bit. Constant *One = ConstantInt::get(In->getType(), 1); - In = Builder->CreateXor(In, One, "tmp"); + In = Builder->CreateXor(In, One); } - + if (CI.getType() == In->getType()) return ReplaceInstUsesWith(CI, In); return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); @@ -594,9 +602,8 @@ Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); - APInt TypeMask(APInt::getAllOnesValue(BitWidth)); - ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS); - ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS); + computeKnownBits(LHS, KnownZeroLHS, KnownOneLHS, 0, &CI); + computeKnownBits(RHS, KnownZeroRHS, KnownOneRHS, 0, &CI); if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { APInt KnownBits = KnownZeroLHS | KnownOneLHS; @@ -624,7 +631,7 @@ Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, } } - return 0; + return nullptr; } /// CanEvaluateZExtd - Determine if the specified value can be computed in the @@ -645,25 +652,24 @@ Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, /// clear the top bits anyway, doing this has no extra cost. /// /// This function works on both vectors and scalars. -static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { +static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear, + InstCombiner &IC, Instruction *CxtI) { BitsToClear = 0; if (isa(V)) return true; - + Instruction *I = dyn_cast(V); if (!I) return false; - + // If the input is a truncate from the destination type, we can trivially - // eliminate it, even if it has multiple uses. - // FIXME: This is currently disabled until codegen can handle this without - // pessimizing code, PR5997. - if (0 && isa(I) && I->getOperand(0)->getType() == Ty) + // eliminate it. + if (isa(I) && I->getOperand(0)->getType() == Ty) return true; - + // We can't extend or shrink something that has multiple uses: doing so would // require duplicating the instruction in general, which isn't profitable. if (!I->hasOneUse()) return false; - + unsigned Opc = I->getOpcode(), Tmp; switch (Opc) { case Instruction::ZExt: // zext(zext(x)) -> zext(x). @@ -676,14 +682,13 @@ static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { case Instruction::Add: case Instruction::Sub: case Instruction::Mul: - case Instruction::Shl: - if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || - !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) + if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) || + !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI)) return false; // These can all be promoted if neither operand has 'bits to clear'. if (BitsToClear == 0 && Tmp == 0) return true; - + // If the operation is an AND/OR/XOR and the bits to clear are zero in the // other side, BitsToClear is ok. if (Tmp == 0 && @@ -692,19 +697,31 @@ static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { // We use MaskedValueIsZero here for generality, but the case we care // about the most is constant RHS. unsigned VSize = V->getType()->getScalarSizeInBits(); - if (MaskedValueIsZero(I->getOperand(1), - APInt::getHighBitsSet(VSize, BitsToClear))) + if (IC.MaskedValueIsZero(I->getOperand(1), + APInt::getHighBitsSet(VSize, BitsToClear), + 0, CxtI)) return true; } - + // Otherwise, we don't know how to analyze this BitsToClear case yet. return false; - + + case Instruction::Shl: + // We can promote shl(x, cst) if we can promote x. Since shl overwrites the + // upper bits we can reduce BitsToClear by the shift amount. + if (ConstantInt *Amt = dyn_cast(I->getOperand(1))) { + if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI)) + return false; + uint64_t ShiftAmt = Amt->getZExtValue(); + BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0; + return true; + } + return false; case Instruction::LShr: // We can promote lshr(x, cst) if we can promote x. This requires the // ultimate 'and' to clear out the high zero bits we're clearing out though. if (ConstantInt *Amt = dyn_cast(I->getOperand(1))) { - if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) + if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI)) return false; BitsToClear += Amt->getZExtValue(); if (BitsToClear > V->getType()->getScalarSizeInBits()) @@ -714,23 +731,23 @@ static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { // Cannot promote variable LSHR. return false; case Instruction::Select: - if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || - !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || + if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) || + !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) || // TODO: If important, we could handle the case when the BitsToClear are // known zero in the disagreeing side. Tmp != BitsToClear) return false; return true; - + case Instruction::PHI: { // We can change a phi if we can change all operands. Note that we never // get into trouble with cyclic PHIs here because we only consider // instructions with a single use. PHINode *PN = cast(I); - if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) + if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI)) return false; for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) - if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || + if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) || // TODO: If important, we could handle the case when the BitsToClear // are known zero in the disagreeing input. Tmp != BitsToClear) @@ -744,48 +761,50 @@ static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { } Instruction *InstCombiner::visitZExt(ZExtInst &CI) { - // If this zero extend is only used by a truncate, let the truncate by + // If this zero extend is only used by a truncate, let the truncate be // eliminated before we try to optimize this zext. - if (CI.hasOneUse() && isa(CI.use_back())) - return 0; - + if (CI.hasOneUse() && isa(CI.user_back())) + return nullptr; + // If one of the common conversion will work, do it. if (Instruction *Result = commonCastTransforms(CI)) return Result; - // See if we can simplify any instructions used by the input whose sole + // See if we can simplify any instructions used by the input whose sole // purpose is to compute bits we don't care about. if (SimplifyDemandedInstructionBits(CI)) return &CI; - + Value *Src = CI.getOperand(0); Type *SrcTy = Src->getType(), *DestTy = CI.getType(); - + // Attempt to extend the entire input expression tree to the destination // type. Only do this if the dest type is a simple type, don't convert the // expression tree to something weird like i93 unless the source is also // strange. unsigned BitsToClear; if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && - CanEvaluateZExtd(Src, DestTy, BitsToClear)) { + CanEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) { assert(BitsToClear < SrcTy->getScalarSizeInBits() && "Unreasonable BitsToClear"); - + // Okay, we can transform this! Insert the new expression now. DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" " to avoid zero extend: " << CI); Value *Res = EvaluateInDifferentType(Src, DestTy, false); assert(Res->getType() == DestTy); - + uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; uint32_t DestBitSize = DestTy->getScalarSizeInBits(); - + // If the high bits are already filled with zeros, just replace this // cast with the result. - if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, - DestBitSize-SrcBitsKept))) + if (MaskedValueIsZero(Res, + APInt::getHighBitsSet(DestBitSize, + DestBitSize-SrcBitsKept), + 0, &CI)) return ReplaceInstUsesWith(CI, Res); - + // We need to emit an AND to clear the high bits. Constant *C = ConstantInt::get(Res->getType(), APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); @@ -797,7 +816,7 @@ Instruction *InstCombiner::visitZExt(ZExtInst &CI) { // 'and' which will be much cheaper than the pair of casts. if (TruncInst *CSrc = dyn_cast(Src)) { // A->B->C cast // TODO: Subsume this into EvaluateInDifferentType. - + // Get the sizes of the types involved. We know that the intermediate type // will be smaller than A or C, but don't know the relation between A and C. Value *A = CSrc->getOperand(0); @@ -814,16 +833,16 @@ Instruction *InstCombiner::visitZExt(ZExtInst &CI) { Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); return new ZExtInst(And, CI.getType()); } - + if (SrcSize == DstSize) { APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), AndValue)); } if (SrcSize > DstSize) { - Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp"); + Value *Trunc = Builder->CreateTrunc(A, CI.getType()); APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); - return BinaryOperator::CreateAnd(Trunc, + return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(Trunc->getType(), AndValue)); } @@ -847,42 +866,32 @@ Instruction *InstCombiner::visitZExt(ZExtInst &CI) { } } - // zext(trunc(t) & C) -> (t & zext(C)). - if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) - if (ConstantInt *C = dyn_cast(SrcI->getOperand(1))) - if (TruncInst *TI = dyn_cast(SrcI->getOperand(0))) { - Value *TI0 = TI->getOperand(0); - if (TI0->getType() == CI.getType()) - return - BinaryOperator::CreateAnd(TI0, - ConstantExpr::getZExt(C, CI.getType())); - } - - // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). - if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) - if (ConstantInt *C = dyn_cast(SrcI->getOperand(1))) - if (BinaryOperator *And = dyn_cast(SrcI->getOperand(0))) - if (And->getOpcode() == Instruction::And && And->hasOneUse() && - And->getOperand(1) == C) - if (TruncInst *TI = dyn_cast(And->getOperand(0))) { - Value *TI0 = TI->getOperand(0); - if (TI0->getType() == CI.getType()) { - Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); - Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp"); - return BinaryOperator::CreateXor(NewAnd, ZC); - } - } + // zext(trunc(X) & C) -> (X & zext(C)). + Constant *C; + Value *X; + if (SrcI && + match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) && + X->getType() == CI.getType()) + return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType())); + + // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)). + Value *And; + if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) && + match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) && + X->getType() == CI.getType()) { + Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); + return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC); + } // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 - Value *X; - if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) && - match(SrcI, m_Not(m_Value(X))) && - (!X->hasOneUse() || !isa(X))) { + if (SrcI && SrcI->hasOneUse() && + SrcI->getType()->getScalarType()->isIntegerTy(1) && + match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa(X))) { Value *New = Builder->CreateZExt(X, CI.getType()); return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); } - - return 0; + + return nullptr; } /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations @@ -891,23 +900,29 @@ Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) { Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1); ICmpInst::Predicate Pred = ICI->getPredicate(); - if (ConstantInt *Op1C = dyn_cast(Op1)) { + // Don't bother if Op1 isn't of vector or integer type. + if (!Op1->getType()->isIntOrIntVectorTy()) + return nullptr; + + if (Constant *Op1C = dyn_cast(Op1)) { // (x ashr x, 31 -> all ones if negative // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive - if ((Pred == ICmpInst::ICMP_SLT && Op1C->isZero()) || + if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) || (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) { Value *Sh = ConstantInt::get(Op0->getType(), Op0->getType()->getScalarSizeInBits()-1); Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit"); if (In->getType() != CI.getType()) - In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp"); + In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/); if (Pred == ICmpInst::ICMP_SGT) In = Builder->CreateNot(In, In->getName()+".not"); return ReplaceInstUsesWith(CI, In); } + } + if (ConstantInt *Op1C = dyn_cast(Op1)) { // If we know that only one bit of the LHS of the icmp can be set and we // have an equality comparison with zero or a power of 2, we can transform // the icmp and sext into bitwise/integer operations. @@ -915,8 +930,7 @@ Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) { ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){ unsigned BitWidth = Op1C->getType()->getBitWidth(); APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); - APInt TypeMask(APInt::getAllOnesValue(BitWidth)); - ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne); + computeKnownBits(Op0, KnownZero, KnownOne, 0, &CI); APInt KnownZeroMask(~KnownZero); if (KnownZeroMask.isPowerOf2()) { @@ -965,20 +979,7 @@ Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) { } } - // vector (x ashr x, 31 -> all ones if signed. - if (VectorType *VTy = dyn_cast(CI.getType())) { - if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_Zero()) && - Op0->getType() == CI.getType()) { - Type *EltTy = VTy->getElementType(); - - // splat the shift constant to a constant vector. - Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1); - Value *In = Builder->CreateAShr(Op0, VSh, Op0->getName()+".lobit"); - return ReplaceInstUsesWith(CI, In); - } - } - - return 0; + return nullptr; } /// CanEvaluateSExtd - Return true if we can take the specified value @@ -995,17 +996,14 @@ static bool CanEvaluateSExtd(Value *V, Type *Ty) { // If this is a constant, it can be trivially promoted. if (isa(V)) return true; - + Instruction *I = dyn_cast(V); if (!I) return false; - - // If this is a truncate from the dest type, we can trivially eliminate it, - // even if it has multiple uses. - // FIXME: This is currently disabled until codegen can handle this without - // pessimizing code, PR5997. - if (0 && isa(I) && I->getOperand(0)->getType() == Ty) + + // If this is a truncate from the dest type, we can trivially eliminate it. + if (isa(I) && I->getOperand(0)->getType() == Ty) return true; - + // We can't extend or shrink something that has multiple uses: doing so would // require duplicating the instruction in general, which isn't profitable. if (!I->hasOneUse()) return false; @@ -1024,14 +1022,14 @@ static bool CanEvaluateSExtd(Value *V, Type *Ty) { // These operators can all arbitrarily be extended if their inputs can. return CanEvaluateSExtd(I->getOperand(0), Ty) && CanEvaluateSExtd(I->getOperand(1), Ty); - + //case Instruction::Shl: TODO //case Instruction::LShr: TODO - + case Instruction::Select: return CanEvaluateSExtd(I->getOperand(1), Ty) && CanEvaluateSExtd(I->getOperand(2), Ty); - + case Instruction::PHI: { // We can change a phi if we can change all operands. Note that we never // get into trouble with cyclic PHIs here because we only consider @@ -1045,24 +1043,24 @@ static bool CanEvaluateSExtd(Value *V, Type *Ty) { // TODO: Can handle more cases here. break; } - + return false; } Instruction *InstCombiner::visitSExt(SExtInst &CI) { - // If this sign extend is only used by a truncate, let the truncate by - // eliminated before we try to optimize this zext. - if (CI.hasOneUse() && isa(CI.use_back())) - return 0; - + // If this sign extend is only used by a truncate, let the truncate be + // eliminated before we try to optimize this sext. + if (CI.hasOneUse() && isa(CI.user_back())) + return nullptr; + if (Instruction *I = commonCastTransforms(CI)) return I; - - // See if we can simplify any instructions used by the input whose sole + + // See if we can simplify any instructions used by the input whose sole // purpose is to compute bits we don't care about. if (SimplifyDemandedInstructionBits(CI)) return &CI; - + Value *Src = CI.getOperand(0); Type *SrcTy = Src->getType(), *DestTy = CI.getType(); @@ -1083,9 +1081,9 @@ Instruction *InstCombiner::visitSExt(SExtInst &CI) { // If the high bits are already filled with sign bit, just replace this // cast with the result. - if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) + if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize) return ReplaceInstUsesWith(CI, Res); - + // We need to emit a shl + ashr to do the sign extend. Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), @@ -1098,7 +1096,7 @@ Instruction *InstCombiner::visitSExt(SExtInst &CI) { if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); uint32_t DestBitSize = DestTy->getScalarSizeInBits(); - + // We need to emit a shl + ashr to do the sign extend. Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); @@ -1121,9 +1119,9 @@ Instruction *InstCombiner::visitSExt(SExtInst &CI) { // into: // %a = shl i32 %i, 30 // %d = ashr i32 %a, 30 - Value *A = 0; + Value *A = nullptr; // TODO: Eventually this could be subsumed by EvaluateInDifferentType. - ConstantInt *BA = 0, *CA = 0; + ConstantInt *BA = nullptr, *CA = nullptr; if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), m_ConstantInt(CA))) && BA == CA && A->getType() == CI.getType()) { @@ -1134,8 +1132,8 @@ Instruction *InstCombiner::visitSExt(SExtInst &CI) { A = Builder->CreateShl(A, ShAmtV, CI.getName()); return BinaryOperator::CreateAShr(A, ShAmtV); } - - return 0; + + return nullptr; } @@ -1147,7 +1145,7 @@ static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); if (!losesInfo) return ConstantFP::get(CFP->getContext(), F); - return 0; + return nullptr; } /// LookThroughFPExtensions - If this is an fp extension instruction, look @@ -1156,13 +1154,16 @@ static Value *LookThroughFPExtensions(Value *V) { if (Instruction *I = dyn_cast(V)) if (I->getOpcode() == Instruction::FPExt) return LookThroughFPExtensions(I->getOperand(0)); - + // If this value is a constant, return the constant in the smallest FP type // that can accurately represent it. This allows us to turn // (float)((double)X+2.0) into x+2.0f. if (ConstantFP *CFP = dyn_cast(V)) { if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) return V; // No constant folding of this. + // See if the value can be truncated to half and then reextended. + if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf)) + return V; // See if the value can be truncated to float and then reextended. if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) return V; @@ -1172,79 +1173,155 @@ static Value *LookThroughFPExtensions(Value *V) { return V; // Don't try to shrink to various long double types. } - + return V; } Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { if (Instruction *I = commonCastTransforms(CI)) return I; - - // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are - // smaller than the destination type, we can eliminate the truncate by doing - // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well - // as many builtins (sqrt, etc). + // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to + // simpilify this expression to avoid one or more of the trunc/extend + // operations if we can do so without changing the numerical results. + // + // The exact manner in which the widths of the operands interact to limit + // what we can and cannot do safely varies from operation to operation, and + // is explained below in the various case statements. BinaryOperator *OpI = dyn_cast(CI.getOperand(0)); if (OpI && OpI->hasOneUse()) { + Value *LHSOrig = LookThroughFPExtensions(OpI->getOperand(0)); + Value *RHSOrig = LookThroughFPExtensions(OpI->getOperand(1)); + unsigned OpWidth = OpI->getType()->getFPMantissaWidth(); + unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth(); + unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth(); + unsigned SrcWidth = std::max(LHSWidth, RHSWidth); + unsigned DstWidth = CI.getType()->getFPMantissaWidth(); switch (OpI->getOpcode()) { - default: break; - case Instruction::FAdd: - case Instruction::FSub: - case Instruction::FMul: - case Instruction::FDiv: - case Instruction::FRem: - Type *SrcTy = OpI->getType(); - Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); - Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); - if (LHSTrunc->getType() != SrcTy && - RHSTrunc->getType() != SrcTy) { - unsigned DstSize = CI.getType()->getScalarSizeInBits(); - // If the source types were both smaller than the destination type of - // the cast, do this xform. - if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && - RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { - LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); - RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); - return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); + default: break; + case Instruction::FAdd: + case Instruction::FSub: + // For addition and subtraction, the infinitely precise result can + // essentially be arbitrarily wide; proving that double rounding + // will not occur because the result of OpI is exact (as we will for + // FMul, for example) is hopeless. However, we *can* nonetheless + // frequently know that double rounding cannot occur (or that it is + // innocuous) by taking advantage of the specific structure of + // infinitely-precise results that admit double rounding. + // + // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient + // to represent both sources, we can guarantee that the double + // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis, + // "A Rigorous Framework for Fully Supporting the IEEE Standard ..." + // for proof of this fact). + // + // Note: Figueroa does not consider the case where DstFormat != + // SrcFormat. It's possible (likely even!) that this analysis + // could be tightened for those cases, but they are rare (the main + // case of interest here is (float)((double)float + float)). + if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) { + if (LHSOrig->getType() != CI.getType()) + LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); + if (RHSOrig->getType() != CI.getType()) + RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); + Instruction *RI = + BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig); + RI->copyFastMathFlags(OpI); + return RI; } - } - break; + break; + case Instruction::FMul: + // For multiplication, the infinitely precise result has at most + // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient + // that such a value can be exactly represented, then no double + // rounding can possibly occur; we can safely perform the operation + // in the destination format if it can represent both sources. + if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) { + if (LHSOrig->getType() != CI.getType()) + LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); + if (RHSOrig->getType() != CI.getType()) + RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); + Instruction *RI = + BinaryOperator::CreateFMul(LHSOrig, RHSOrig); + RI->copyFastMathFlags(OpI); + return RI; + } + break; + case Instruction::FDiv: + // For division, we use again use the bound from Figueroa's + // dissertation. I am entirely certain that this bound can be + // tightened in the unbalanced operand case by an analysis based on + // the diophantine rational approximation bound, but the well-known + // condition used here is a good conservative first pass. + // TODO: Tighten bound via rigorous analysis of the unbalanced case. + if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) { + if (LHSOrig->getType() != CI.getType()) + LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); + if (RHSOrig->getType() != CI.getType()) + RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); + Instruction *RI = + BinaryOperator::CreateFDiv(LHSOrig, RHSOrig); + RI->copyFastMathFlags(OpI); + return RI; + } + break; + case Instruction::FRem: + // Remainder is straightforward. Remainder is always exact, so the + // type of OpI doesn't enter into things at all. We simply evaluate + // in whichever source type is larger, then convert to the + // destination type. + if (LHSWidth < SrcWidth) + LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType()); + else if (RHSWidth <= SrcWidth) + RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType()); + Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig); + if (Instruction *RI = dyn_cast(ExactResult)) + RI->copyFastMathFlags(OpI); + return CastInst::CreateFPCast(ExactResult, CI.getType()); + } + + // (fptrunc (fneg x)) -> (fneg (fptrunc x)) + if (BinaryOperator::isFNeg(OpI)) { + Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1), + CI.getType()); + Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc); + RI->copyFastMathFlags(OpI); + return RI; } } - - // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) - // NOTE: This should be disabled by -fno-builtin-sqrt if we ever support it. - CallInst *Call = dyn_cast(CI.getOperand(0)); - if (Call && Call->getCalledFunction() && - Call->getCalledFunction()->getName() == "sqrt" && - Call->getNumArgOperands() == 1 && - Call->hasOneUse()) { - CastInst *Arg = dyn_cast(Call->getArgOperand(0)); - if (Arg && Arg->getOpcode() == Instruction::FPExt && - CI.getType()->isFloatTy() && - Call->getType()->isDoubleTy() && - Arg->getType()->isDoubleTy() && - Arg->getOperand(0)->getType()->isFloatTy()) { - Function *Callee = Call->getCalledFunction(); - Module *M = CI.getParent()->getParent()->getParent(); - Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf", - Callee->getAttributes(), - Builder->getFloatTy(), - Builder->getFloatTy(), - NULL); - CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), - "sqrtfcall"); - ret->setAttributes(Callee->getAttributes()); - - - // Remove the old Call. With -fmath-errno, it won't get marked readnone. - ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType())); - EraseInstFromFunction(*Call); - return ret; + + // (fptrunc (select cond, R1, Cst)) --> + // (select cond, (fptrunc R1), (fptrunc Cst)) + SelectInst *SI = dyn_cast(CI.getOperand(0)); + if (SI && + (isa(SI->getOperand(1)) || + isa(SI->getOperand(2)))) { + Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1), + CI.getType()); + Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2), + CI.getType()); + return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc); + } + + IntrinsicInst *II = dyn_cast(CI.getOperand(0)); + if (II) { + switch (II->getIntrinsicID()) { + default: break; + case Intrinsic::fabs: { + // (fptrunc (fabs x)) -> (fabs (fptrunc x)) + Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0), + CI.getType()); + Type *IntrinsicType[] = { CI.getType() }; + Function *Overload = + Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(), + II->getIntrinsicID(), IntrinsicType); + + Value *Args[] = { InnerTrunc }; + return CallInst::Create(Overload, Args, II->getName()); + } } } - - return 0; + + return nullptr; } Instruction *InstCombiner::visitFPExt(CastInst &CI) { @@ -1253,7 +1330,7 @@ Instruction *InstCombiner::visitFPExt(CastInst &CI) { Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { Instruction *OpI = dyn_cast(FI.getOperand(0)); - if (OpI == 0) + if (!OpI) return commonCastTransforms(FI); // fptoui(uitofp(X)) --> X @@ -1261,7 +1338,7 @@ Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { // This is safe if the intermediate type has enough bits in its mantissa to // accurately represent all values of X. For example, do not do this with // i64->float->i64. This is also safe for sitofp case, because any negative - // 'X' value would cause an undefined result for the fptoui. + // 'X' value would cause an undefined result for the fptoui. if ((isa(OpI) || isa(OpI)) && OpI->getOperand(0)->getType() == FI.getType() && (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ @@ -1273,21 +1350,21 @@ Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { Instruction *OpI = dyn_cast(FI.getOperand(0)); - if (OpI == 0) + if (!OpI) return commonCastTransforms(FI); - + // fptosi(sitofp(X)) --> X // fptosi(uitofp(X)) --> X // This is safe if the intermediate type has enough bits in its mantissa to // accurately represent all values of X. For example, do not do this with // i64->float->i64. This is also safe for sitofp case, because any negative - // 'X' value would cause an undefined result for the fptoui. + // 'X' value would cause an undefined result for the fptoui. if ((isa(OpI) || isa(OpI)) && OpI->getOperand(0)->getType() == FI.getType() && (int)FI.getType()->getScalarSizeInBits() <= OpI->getType()->getFPMantissaWidth()) return ReplaceInstUsesWith(FI, OpI->getOperand(0)); - + return commonCastTransforms(FI); } @@ -1303,75 +1380,83 @@ Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { // If the source integer type is not the intptr_t type for this target, do a // trunc or zext to the intptr_t type, then inttoptr of it. This allows the // cast to be exposed to other transforms. - if (TD) { - if (CI.getOperand(0)->getType()->getScalarSizeInBits() > - TD->getPointerSizeInBits()) { - Value *P = Builder->CreateTrunc(CI.getOperand(0), - TD->getIntPtrType(CI.getContext()), "tmp"); - return new IntToPtrInst(P, CI.getType()); - } - if (CI.getOperand(0)->getType()->getScalarSizeInBits() < - TD->getPointerSizeInBits()) { - Value *P = Builder->CreateZExt(CI.getOperand(0), - TD->getIntPtrType(CI.getContext()), "tmp"); + + if (DL) { + unsigned AS = CI.getAddressSpace(); + if (CI.getOperand(0)->getType()->getScalarSizeInBits() != + DL->getPointerSizeInBits(AS)) { + Type *Ty = DL->getIntPtrType(CI.getContext(), AS); + if (CI.getType()->isVectorTy()) // Handle vectors of pointers. + Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements()); + + Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty); return new IntToPtrInst(P, CI.getType()); } } - + if (Instruction *I = commonCastTransforms(CI)) return I; - return 0; + return nullptr; } /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { Value *Src = CI.getOperand(0); - + if (GetElementPtrInst *GEP = dyn_cast(Src)) { // If casting the result of a getelementptr instruction with no offset, turn // this into a cast of the original pointer! - if (GEP->hasAllZeroIndices()) { + if (GEP->hasAllZeroIndices() && + // If CI is an addrspacecast and GEP changes the poiner type, merging + // GEP into CI would undo canonicalizing addrspacecast with different + // pointer types, causing infinite loops. + (!isa(CI) || + GEP->getType() == GEP->getPointerOperand()->getType())) { // Changing the cast operand is usually not a good idea but it is safe - // here because the pointer operand is being replaced with another + // here because the pointer operand is being replaced with another // pointer operand so the opcode doesn't need to change. Worklist.Add(GEP); CI.setOperand(0, GEP->getOperand(0)); return &CI; } - + + if (!DL) + return commonCastTransforms(CI); + // If the GEP has a single use, and the base pointer is a bitcast, and the // GEP computes a constant offset, see if we can convert these three // instructions into fewer. This typically happens with unions and other // non-type-safe code. - if (TD && GEP->hasOneUse() && isa(GEP->getOperand(0)) && - GEP->hasAllConstantIndices()) { - // We are guaranteed to get a constant from EmitGEPOffset. - ConstantInt *OffsetV = cast(EmitGEPOffset(GEP)); - int64_t Offset = OffsetV->getSExtValue(); - + unsigned AS = GEP->getPointerAddressSpace(); + unsigned OffsetBits = DL->getPointerSizeInBits(AS); + APInt Offset(OffsetBits, 0); + BitCastInst *BCI = dyn_cast(GEP->getOperand(0)); + if (GEP->hasOneUse() && + BCI && + GEP->accumulateConstantOffset(*DL, Offset)) { // Get the base pointer input of the bitcast, and the type it points to. - Value *OrigBase = cast(GEP->getOperand(0))->getOperand(0); - Type *GEPIdxTy = - cast(OrigBase->getType())->getElementType(); + Value *OrigBase = BCI->getOperand(0); SmallVector NewIndices; - if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { + if (FindElementAtOffset(OrigBase->getType(), + Offset.getSExtValue(), + NewIndices)) { // If we were able to index down into an element, create the GEP // and bitcast the result. This eliminates one bitcast, potentially // two. Value *NGEP = cast(GEP)->isInBounds() ? - Builder->CreateInBoundsGEP(OrigBase, NewIndices) : - Builder->CreateGEP(OrigBase, NewIndices); + Builder->CreateInBoundsGEP(OrigBase, NewIndices) : + Builder->CreateGEP(OrigBase, NewIndices); NGEP->takeName(GEP); - + if (isa(CI)) return new BitCastInst(NGEP, CI.getType()); assert(isa(CI)); return new PtrToIntInst(NGEP, CI.getType()); - } + } } } - + return commonCastTransforms(CI); } @@ -1379,22 +1464,22 @@ Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { // If the destination integer type is not the intptr_t type for this target, // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast // to be exposed to other transforms. - if (TD) { - if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { - Value *P = Builder->CreatePtrToInt(CI.getOperand(0), - TD->getIntPtrType(CI.getContext()), - "tmp"); - return new TruncInst(P, CI.getType()); - } - if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) { - Value *P = Builder->CreatePtrToInt(CI.getOperand(0), - TD->getIntPtrType(CI.getContext()), - "tmp"); - return new ZExtInst(P, CI.getType()); - } - } - - return commonPointerCastTransforms(CI); + + if (!DL) + return commonPointerCastTransforms(CI); + + Type *Ty = CI.getType(); + unsigned AS = CI.getPointerAddressSpace(); + + if (Ty->getScalarSizeInBits() == DL->getPointerSizeInBits(AS)) + return commonPointerCastTransforms(CI); + + Type *PtrTy = DL->getIntPtrType(CI.getContext(), AS); + if (Ty->isVectorTy()) // Handle vectors of pointers. + PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements()); + + Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy); + return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false); } /// OptimizeVectorResize - This input value (which is known to have vector type) @@ -1408,34 +1493,33 @@ static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy, // element size, or the input is a multiple of the output element size. // Convert the input type to have the same element type as the output. VectorType *SrcTy = cast(InVal->getType()); - + if (SrcTy->getElementType() != DestTy->getElementType()) { // The input types don't need to be identical, but for now they must be the // same size. There is no specific reason we couldn't handle things like // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten - // there yet. + // there yet. if (SrcTy->getElementType()->getPrimitiveSizeInBits() != DestTy->getElementType()->getPrimitiveSizeInBits()) - return 0; - + return nullptr; + SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); InVal = IC.Builder->CreateBitCast(InVal, SrcTy); } - + // Now that the element types match, get the shuffle mask and RHS of the // shuffle to use, which depends on whether we're increasing or decreasing the // size of the input. - SmallVector ShuffleMask; + SmallVector ShuffleMask; Value *V2; - IntegerType *Int32Ty = Type::getInt32Ty(SrcTy->getContext()); - + if (SrcTy->getNumElements() > DestTy->getNumElements()) { // If we're shrinking the number of elements, just shuffle in the low // elements from the input and use undef as the second shuffle input. V2 = UndefValue::get(SrcTy); for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) - ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); - + ShuffleMask.push_back(i); + } else { // If we're increasing the number of elements, shuffle in all of the // elements from InVal and fill the rest of the result elements with zeros @@ -1443,14 +1527,16 @@ static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy, V2 = Constant::getNullValue(SrcTy); unsigned SrcElts = SrcTy->getNumElements(); for (unsigned i = 0, e = SrcElts; i != e; ++i) - ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); + ShuffleMask.push_back(i); // The excess elements reference the first element of the zero input. - ShuffleMask.append(DestTy->getNumElements()-SrcElts, - ConstantInt::get(Int32Ty, SrcElts)); + for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i) + ShuffleMask.push_back(SrcElts); } - - return new ShuffleVectorInst(InVal, V2, ConstantVector::get(ShuffleMask)); + + return new ShuffleVectorInst(InVal, V2, + ConstantDataVector::get(V2->getContext(), + ShuffleMask)); } static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) { @@ -1466,15 +1552,20 @@ static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) { /// insertions into the vector. See the example in the comment for /// OptimizeIntegerToVectorInsertions for the pattern this handles. /// The type of V is always a non-zero multiple of VecEltTy's size. +/// Shift is the number of bits between the lsb of V and the lsb of +/// the vector. /// /// This returns false if the pattern can't be matched or true if it can, /// filling in Elements with the elements found here. -static bool CollectInsertionElements(Value *V, unsigned ElementIndex, +static bool CollectInsertionElements(Value *V, unsigned Shift, SmallVectorImpl &Elements, - Type *VecEltTy) { + Type *VecEltTy, InstCombiner &IC) { + assert(isMultipleOfTypeSize(Shift, VecEltTy) && + "Shift should be a multiple of the element type size"); + // Undef values never contribute useful bits to the result. if (isa(V)) return true; - + // If we got down to a value of the right type, we win, try inserting into the // right element. if (V->getType() == VecEltTy) { @@ -1482,15 +1573,19 @@ static bool CollectInsertionElements(Value *V, unsigned ElementIndex, if (Constant *C = dyn_cast(V)) if (C->isNullValue()) return true; - + + unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy); + if (IC.getDataLayout()->isBigEndian()) + ElementIndex = Elements.size() - ElementIndex - 1; + // Fail if multiple elements are inserted into this slot. - if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0) + if (Elements[ElementIndex]) return false; - + Elements[ElementIndex] = V; return true; } - + if (Constant *C = dyn_cast(V)) { // Figure out the # elements this provides, and bitcast it or slice it up // as required. @@ -1500,8 +1595,8 @@ static bool CollectInsertionElements(Value *V, unsigned ElementIndex, // it to the right type so it gets properly inserted. if (NumElts == 1) return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy), - ElementIndex, Elements, VecEltTy); - + Shift, Elements, VecEltTy, IC); + // Okay, this is a constant that covers multiple elements. Slice it up into // pieces and insert each element-sized piece into the vector. if (!isa(C->getType())) @@ -1509,49 +1604,49 @@ static bool CollectInsertionElements(Value *V, unsigned ElementIndex, C->getType()->getPrimitiveSizeInBits())); unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits(); Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize); - + for (unsigned i = 0; i != NumElts; ++i) { + unsigned ShiftI = Shift+i*ElementSize; Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(), - i*ElementSize)); + ShiftI)); Piece = ConstantExpr::getTrunc(Piece, ElementIntTy); - if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy)) + if (!CollectInsertionElements(Piece, ShiftI, Elements, VecEltTy, IC)) return false; } return true; } - + if (!V->hasOneUse()) return false; - + Instruction *I = dyn_cast(V); - if (I == 0) return false; + if (!I) return false; switch (I->getOpcode()) { default: return false; // Unhandled case. case Instruction::BitCast: - return CollectInsertionElements(I->getOperand(0), ElementIndex, - Elements, VecEltTy); + return CollectInsertionElements(I->getOperand(0), Shift, + Elements, VecEltTy, IC); case Instruction::ZExt: if (!isMultipleOfTypeSize( I->getOperand(0)->getType()->getPrimitiveSizeInBits(), VecEltTy)) return false; - return CollectInsertionElements(I->getOperand(0), ElementIndex, - Elements, VecEltTy); + return CollectInsertionElements(I->getOperand(0), Shift, + Elements, VecEltTy, IC); case Instruction::Or: - return CollectInsertionElements(I->getOperand(0), ElementIndex, - Elements, VecEltTy) && - CollectInsertionElements(I->getOperand(1), ElementIndex, - Elements, VecEltTy); + return CollectInsertionElements(I->getOperand(0), Shift, + Elements, VecEltTy, IC) && + CollectInsertionElements(I->getOperand(1), Shift, + Elements, VecEltTy, IC); case Instruction::Shl: { // Must be shifting by a constant that is a multiple of the element size. ConstantInt *CI = dyn_cast(I->getOperand(1)); - if (CI == 0) return false; - if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false; - unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy); - - return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift, - Elements, VecEltTy); + if (!CI) return false; + Shift += CI->getZExtValue(); + if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false; + return CollectInsertionElements(I->getOperand(0), Shift, + Elements, VecEltTy, IC); } - + } } @@ -1572,25 +1667,28 @@ static bool CollectInsertionElements(Value *V, unsigned ElementIndex, /// Into two insertelements that do "buildvector{%inc, %inc5}". static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI, InstCombiner &IC) { + // We need to know the target byte order to perform this optimization. + if (!IC.getDataLayout()) return nullptr; + VectorType *DestVecTy = cast(CI.getType()); Value *IntInput = CI.getOperand(0); SmallVector Elements(DestVecTy->getNumElements()); if (!CollectInsertionElements(IntInput, 0, Elements, - DestVecTy->getElementType())) - return 0; + DestVecTy->getElementType(), IC)) + return nullptr; // If we succeeded, we know that all of the element are specified by Elements // or are zero if Elements has a null entry. Recast this as a set of // insertions. Value *Result = Constant::getNullValue(CI.getType()); for (unsigned i = 0, e = Elements.size(); i != e; ++i) { - if (Elements[i] == 0) continue; // Unset element. - + if (!Elements[i]) continue; // Unset element. + Result = IC.Builder->CreateInsertElement(Result, Elements[i], IC.Builder->getInt32(i)); } - + return Result; } @@ -1598,12 +1696,15 @@ static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI, /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double /// bitcast. The various long double bitcasts can't get in here. static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ + // We need to know the target byte order to perform this optimization. + if (!IC.getDataLayout()) return nullptr; + Value *Src = CI.getOperand(0); Type *DestTy = CI.getType(); // If this is a bitcast from int to float, check to see if the int is an // extraction from a vector. - Value *VecInput = 0; + Value *VecInput = nullptr; // bitcast(trunc(bitcast(somevector))) if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) && isa(VecInput->getType())) { @@ -1618,13 +1719,16 @@ static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ VecTy->getPrimitiveSizeInBits() / DestWidth); VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); } - - return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0)); + + unsigned Elt = 0; + if (IC.getDataLayout()->isBigEndian()) + Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1; + return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); } } - + // bitcast(trunc(lshr(bitcast(somevector), cst)) - ConstantInt *ShAmt = 0; + ConstantInt *ShAmt = nullptr; if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)), m_ConstantInt(ShAmt)))) && isa(VecInput->getType())) { @@ -1639,12 +1743,14 @@ static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ VecTy->getPrimitiveSizeInBits() / DestWidth); VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); } - + unsigned Elt = ShAmt->getZExtValue() / DestWidth; + if (IC.getDataLayout()->isBigEndian()) + Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt; return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); } } - return 0; + return nullptr; } Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { @@ -1659,21 +1765,11 @@ Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { if (DestTy == Src->getType()) return ReplaceInstUsesWith(CI, Src); - // Bitcasts are transitive. - if (BitCastInst* BSrc = dyn_cast(Src)) { - return CastInst::Create(Instruction::BitCast, BSrc->getOperand(0), DestTy); - } - if (PointerType *DstPTy = dyn_cast(DestTy)) { PointerType *SrcPTy = cast(SrcTy); Type *DstElTy = DstPTy->getElementType(); Type *SrcElTy = SrcPTy->getElementType(); - - // If the address spaces don't match, don't eliminate the bitcast, which is - // required for changing types. - if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) - return 0; - + // If we are casting a alloca to a pointer to a type of the same // size, rewrite the allocation instruction to allocate the "right" type. // There is no need to modify malloc calls because it is their bitcast that @@ -1681,14 +1777,14 @@ Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { if (AllocaInst *AI = dyn_cast(Src)) if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) return V; - + // If the source and destination are pointers, and this cast is equivalent // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. // This can enhance SROA and other transforms that want type-safe pointers. Constant *ZeroUInt = Constant::getNullValue(Type::getInt32Ty(CI.getContext())); unsigned NumZeros = 0; - while (SrcElTy != DstElTy && + while (SrcElTy != DstElTy && isa(SrcElTy) && !SrcElTy->isPointerTy() && SrcElTy->getNumContainedTypes() /* not "{}" */) { SrcElTy = cast(SrcElTy)->getTypeAtIndex(ZeroUInt); @@ -1701,7 +1797,7 @@ Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { return GetElementPtrInst::CreateInBounds(Src, Idxs); } } - + // Try to optimize int -> float bitcasts. if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa(SrcTy)) if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this)) @@ -1714,7 +1810,7 @@ Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) } - + if (isa(SrcTy)) { // If this is a cast from an integer to vector, check to see if the input // is a trunc or zext of a bitcast from vector. If so, we can replace all @@ -1727,7 +1823,7 @@ Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { cast(DestTy), *this)) return I; } - + // If the input is an 'or' instruction, we may be doing shifts and ors to // assemble the elements of the vector manually. Try to rip the code out // and replace it with insertelements. @@ -1737,29 +1833,39 @@ Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { } if (VectorType *SrcVTy = dyn_cast(SrcTy)) { - if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) { - Value *Elem = - Builder->CreateExtractElement(Src, - Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); - return CastInst::Create(Instruction::BitCast, Elem, DestTy); + if (SrcVTy->getNumElements() == 1) { + // If our destination is not a vector, then make this a straight + // scalar-scalar cast. + if (!DestTy->isVectorTy()) { + Value *Elem = + Builder->CreateExtractElement(Src, + Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); + return CastInst::Create(Instruction::BitCast, Elem, DestTy); + } + + // Otherwise, see if our source is an insert. If so, then use the scalar + // component directly. + if (InsertElementInst *IEI = + dyn_cast(CI.getOperand(0))) + return CastInst::Create(Instruction::BitCast, IEI->getOperand(1), + DestTy); } } if (ShuffleVectorInst *SVI = dyn_cast(Src)) { // Okay, we have (bitcast (shuffle ..)). Check to see if this is // a bitcast to a vector with the same # elts. - if (SVI->hasOneUse() && DestTy->isVectorTy() && - cast(DestTy)->getNumElements() == - SVI->getType()->getNumElements() && + if (SVI->hasOneUse() && DestTy->isVectorTy() && + DestTy->getVectorNumElements() == SVI->getType()->getNumElements() && SVI->getType()->getNumElements() == - cast(SVI->getOperand(0)->getType())->getNumElements()) { + SVI->getOperand(0)->getType()->getVectorNumElements()) { BitCastInst *Tmp; // If either of the operands is a cast from CI.getType(), then // evaluating the shuffle in the casted destination's type will allow // us to eliminate at least one cast. - if (((Tmp = dyn_cast(SVI->getOperand(0))) && + if (((Tmp = dyn_cast(SVI->getOperand(0))) && Tmp->getOperand(0)->getType() == DestTy) || - ((Tmp = dyn_cast(SVI->getOperand(1))) && + ((Tmp = dyn_cast(SVI->getOperand(1))) && Tmp->getOperand(0)->getType() == DestTy)) { Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); @@ -1769,8 +1875,31 @@ Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { } } } - + if (SrcTy->isPointerTy()) return commonPointerCastTransforms(CI); return commonCastTransforms(CI); } + +Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) { + // If the destination pointer element type is not the same as the source's + // first do a bitcast to the destination type, and then the addrspacecast. + // This allows the cast to be exposed to other transforms. + Value *Src = CI.getOperand(0); + PointerType *SrcTy = cast(Src->getType()->getScalarType()); + PointerType *DestTy = cast(CI.getType()->getScalarType()); + + Type *DestElemTy = DestTy->getElementType(); + if (SrcTy->getElementType() != DestElemTy) { + Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace()); + if (VectorType *VT = dyn_cast(CI.getType())) { + // Handle vectors of pointers. + MidTy = VectorType::get(MidTy, VT->getNumElements()); + } + + Value *NewBitCast = Builder->CreateBitCast(Src, MidTy); + return new AddrSpaceCastInst(NewBitCast, CI.getType()); + } + + return commonPointerCastTransforms(CI); +}