X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FInstCombine%2FInstructionCombining.cpp;h=bc9a43fa7ecbe2f4831b7cef6ffca9047157ed7e;hb=96363d50018da6e9cfa549695f01aa0d4d2d6a31;hp=64ed817e75c862b61f308218c5ced9d43b31191d;hpb=a78fa8cc2dd6d2ffe5e4fe605f38aae7b3d2fb7a;p=oota-llvm.git diff --git a/lib/Transforms/InstCombine/InstructionCombining.cpp b/lib/Transforms/InstCombine/InstructionCombining.cpp index 64ed817e75c..bc9a43fa7ec 100644 --- a/lib/Transforms/InstCombine/InstructionCombining.cpp +++ b/lib/Transforms/InstCombine/InstructionCombining.cpp @@ -33,30 +33,32 @@ // //===----------------------------------------------------------------------===// -#define DEBUG_TYPE "instcombine" #include "llvm/Transforms/Scalar.h" #include "InstCombine.h" -#include "llvm/IntrinsicInst.h" +#include "llvm-c/Initialization.h" +#include "llvm/ADT/SmallPtrSet.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/ADT/StringSwitch.h" #include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/MemoryBuiltins.h" -#include "llvm/Target/TargetData.h" +#include "llvm/IR/CFG.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/GetElementPtrTypeIterator.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/PatternMatch.h" +#include "llvm/IR/ValueHandle.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/Debug.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Utils/Local.h" -#include "llvm/Support/CFG.h" -#include "llvm/Support/Debug.h" -#include "llvm/Support/GetElementPtrTypeIterator.h" -#include "llvm/Support/PatternMatch.h" -#include "llvm/Support/ValueHandle.h" -#include "llvm/ADT/SmallPtrSet.h" -#include "llvm/ADT/Statistic.h" -#include "llvm/ADT/StringSwitch.h" -#include "llvm-c/Initialization.h" #include #include using namespace llvm; using namespace llvm::PatternMatch; +#define DEBUG_TYPE "instcombine" + STATISTIC(NumCombined , "Number of insts combined"); STATISTIC(NumConstProp, "Number of constant folds"); STATISTIC(NumDeadInst , "Number of dead inst eliminated"); @@ -65,6 +67,11 @@ STATISTIC(NumExpand, "Number of expansions"); STATISTIC(NumFactor , "Number of factorizations"); STATISTIC(NumReassoc , "Number of reassociations"); +static cl::opt UnsafeFPShrink("enable-double-float-shrink", cl::Hidden, + cl::init(false), + cl::desc("Enable unsafe double to float " + "shrinking for math lib calls")); + // Initialization Routines void llvm::initializeInstCombine(PassRegistry &Registry) { initializeInstCombinerPass(Registry); @@ -87,30 +94,34 @@ void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { } +Value *InstCombiner::EmitGEPOffset(User *GEP) { + return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP); +} + /// ShouldChangeType - Return true if it is desirable to convert a computation /// from 'From' to 'To'. We don't want to convert from a legal to an illegal /// type for example, or from a smaller to a larger illegal type. bool InstCombiner::ShouldChangeType(Type *From, Type *To) const { assert(From->isIntegerTy() && To->isIntegerTy()); - - // If we don't have TD, we don't know if the source/dest are legal. - if (!TD) return false; - + + // If we don't have DL, we don't know if the source/dest are legal. + if (!DL) return false; + unsigned FromWidth = From->getPrimitiveSizeInBits(); unsigned ToWidth = To->getPrimitiveSizeInBits(); - bool FromLegal = TD->isLegalInteger(FromWidth); - bool ToLegal = TD->isLegalInteger(ToWidth); - + bool FromLegal = DL->isLegalInteger(FromWidth); + bool ToLegal = DL->isLegalInteger(ToWidth); + // If this is a legal integer from type, and the result would be an illegal // type, don't do the transformation. if (FromLegal && !ToLegal) return false; - + // Otherwise, if both are illegal, do not increase the size of the result. We // do allow things like i160 -> i64, but not i64 -> i160. if (!FromLegal && !ToLegal && ToWidth > FromWidth) return false; - + return true; } @@ -127,7 +138,7 @@ static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) { // We reason about Add and Sub Only. Instruction::BinaryOps Opcode = I.getOpcode(); - if (Opcode != Instruction::Add && + if (Opcode != Instruction::Add && Opcode != Instruction::Sub) { return false; } @@ -152,6 +163,21 @@ static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) { return !Overflow; } +/// Conservatively clears subclassOptionalData after a reassociation or +/// commutation. We preserve fast-math flags when applicable as they can be +/// preserved. +static void ClearSubclassDataAfterReassociation(BinaryOperator &I) { + FPMathOperator *FPMO = dyn_cast(&I); + if (!FPMO) { + I.clearSubclassOptionalData(); + return; + } + + FastMathFlags FMF = I.getFastMathFlags(); + I.clearSubclassOptionalData(); + I.setFastMathFlags(FMF); +} + /// SimplifyAssociativeOrCommutative - This performs a few simplifications for /// operators which are associative or commutative: // @@ -196,22 +222,22 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { Value *C = I.getOperand(1); // Does "B op C" simplify? - if (Value *V = SimplifyBinOp(Opcode, B, C, TD)) { + if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) { // It simplifies to V. Form "A op V". I.setOperand(0, A); I.setOperand(1, V); // Conservatively clear the optional flags, since they may not be // preserved by the reassociation. if (MaintainNoSignedWrap(I, B, C) && - (!Op0 || (isa(Op0) && Op0->hasNoSignedWrap()))) { + (!Op0 || (isa(Op0) && Op0->hasNoSignedWrap()))) { // Note: this is only valid because SimplifyBinOp doesn't look at // the operands to Op0. I.clearSubclassOptionalData(); I.setHasNoSignedWrap(true); } else { - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); } - + Changed = true; ++NumReassoc; continue; @@ -225,13 +251,13 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { Value *C = Op1->getOperand(1); // Does "A op B" simplify? - if (Value *V = SimplifyBinOp(Opcode, A, B, TD)) { + if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) { // It simplifies to V. Form "V op C". I.setOperand(0, V); I.setOperand(1, C); // Conservatively clear the optional flags, since they may not be // preserved by the reassociation. - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); Changed = true; ++NumReassoc; continue; @@ -247,13 +273,13 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { Value *C = I.getOperand(1); // Does "C op A" simplify? - if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) { + if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { // It simplifies to V. Form "V op B". I.setOperand(0, V); I.setOperand(1, B); // Conservatively clear the optional flags, since they may not be // preserved by the reassociation. - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); Changed = true; ++NumReassoc; continue; @@ -267,13 +293,13 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { Value *C = Op1->getOperand(1); // Does "C op A" simplify? - if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) { + if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) { // It simplifies to V. Form "B op V". I.setOperand(0, B); I.setOperand(1, V); // Conservatively clear the optional flags, since they may not be // preserved by the reassociation. - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); Changed = true; ++NumReassoc; continue; @@ -294,13 +320,19 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { Constant *Folded = ConstantExpr::get(Opcode, C1, C2); BinaryOperator *New = BinaryOperator::Create(Opcode, A, B); + if (isa(New)) { + FastMathFlags Flags = I.getFastMathFlags(); + Flags &= Op0->getFastMathFlags(); + Flags &= Op1->getFastMathFlags(); + New->setFastMathFlags(Flags); + } InsertNewInstWith(New, I); New->takeName(Op1); I.setOperand(0, New); I.setOperand(1, Folded); // Conservatively clear the optional flags, since they may not be // preserved by the reassociation. - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); Changed = true; continue; @@ -394,7 +426,7 @@ Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { std::swap(C, D); // Consider forming "A op' (B op D)". // If "B op D" simplifies then it can be formed with no cost. - Value *V = SimplifyBinOp(TopLevelOpcode, B, D, TD); + Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL); // If "B op D" doesn't simplify then only go on if both of the existing // operations "A op' B" and "C op' D" will be zapped as no longer used. if (!V && Op0->hasOneUse() && Op1->hasOneUse()) @@ -416,7 +448,7 @@ Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { std::swap(C, D); // Consider forming "(A op C) op' B". // If "A op C" simplifies then it can be formed with no cost. - Value *V = SimplifyBinOp(TopLevelOpcode, A, C, TD); + Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL); // If "A op C" doesn't simplify then only go on if both of the existing // operations "A op' B" and "C op' D" will be zapped as no longer used. if (!V && Op0->hasOneUse() && Op1->hasOneUse()) @@ -438,8 +470,8 @@ Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' // Do "A op C" and "B op C" both simplify? - if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, TD)) - if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, TD)) { + if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL)) + if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) { // They do! Return "L op' R". ++NumExpand; // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. @@ -447,7 +479,7 @@ Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { (Instruction::isCommutative(InnerOpcode) && L == B && R == A)) return Op0; // Otherwise return "L op' R" if it simplifies. - if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD)) + if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) return V; // Otherwise, create a new instruction. C = Builder->CreateBinOp(InnerOpcode, L, R); @@ -463,8 +495,8 @@ Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op' // Do "A op B" and "A op C" both simplify? - if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, TD)) - if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, TD)) { + if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL)) + if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) { // They do! Return "L op' R". ++NumExpand; // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. @@ -472,7 +504,7 @@ Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { (Instruction::isCommutative(InnerOpcode) && L == C && R == B)) return Op1; // Otherwise return "L op' R" if it simplifies. - if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD)) + if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL)) return V; // Otherwise, create a new instruction. A = Builder->CreateBinOp(InnerOpcode, L, R); @@ -495,10 +527,8 @@ Value *InstCombiner::dyn_castNegVal(Value *V) const { if (ConstantInt *C = dyn_cast(V)) return ConstantExpr::getNeg(C); - if (Constant *C = dyn_cast(V)) - // FIXME: Remove ConstantVector - if ((isa(C) || isa(C)) && - C->getType()->getVectorElementType()->isIntegerTy()) + if (ConstantDataVector *C = dyn_cast(V)) + if (C->getType()->getElementType()->isIntegerTy()) return ConstantExpr::getNeg(C); return 0; @@ -508,18 +538,16 @@ Value *InstCombiner::dyn_castNegVal(Value *V) const { // instruction if the LHS is a constant negative zero (which is the 'negate' // form). // -Value *InstCombiner::dyn_castFNegVal(Value *V) const { - if (BinaryOperator::isFNeg(V)) +Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const { + if (BinaryOperator::isFNeg(V, IgnoreZeroSign)) return BinaryOperator::getFNegArgument(V); // Constants can be considered to be negated values if they can be folded. if (ConstantFP *C = dyn_cast(V)) return ConstantExpr::getFNeg(C); - if (Constant *C = dyn_cast(V)) - // FIXME: Remove ConstantVector - if ((isa(C) || isa(C)) && - C->getType()->getVectorElementType()->isFloatingPointTy()) + if (ConstantDataVector *C = dyn_cast(V)) + if (C->getType()->getElementType()->isFloatingPointTy()) return ConstantExpr::getFNeg(C); return 0; @@ -544,10 +572,15 @@ static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, Value *Op0 = SO, *Op1 = ConstOperand; if (!ConstIsRHS) std::swap(Op0, Op1); - - if (BinaryOperator *BO = dyn_cast(&I)) - return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, + + if (BinaryOperator *BO = dyn_cast(&I)) { + Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1, SO->getName()+".op"); + Instruction *FPInst = dyn_cast(RI); + if (FPInst && isa(FPInst)) + FPInst->copyFastMathFlags(BO); + return RI; + } if (ICmpInst *CI = dyn_cast(&I)) return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1, SO->getName()+".cmp"); @@ -583,7 +616,7 @@ Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements()) return 0; } - + Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); @@ -603,21 +636,20 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { unsigned NumPHIValues = PN->getNumIncomingValues(); if (NumPHIValues == 0) return 0; - + // We normally only transform phis with a single use. However, if a PHI has // multiple uses and they are all the same operation, we can fold *all* of the // uses into the PHI. if (!PN->hasOneUse()) { // Walk the use list for the instruction, comparing them to I. - for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); - UI != E; ++UI) { - Instruction *User = cast(*UI); - if (User != &I && !I.isIdenticalTo(User)) + for (User *U : PN->users()) { + Instruction *UI = cast(U); + if (UI != &I && !I.isIdenticalTo(UI)) return 0; } // Otherwise, we can replace *all* users with the new PHI we form. } - + // Check to see if all of the operands of the PHI are simple constants // (constantint/constantfp/undef). If there is one non-constant value, // remember the BB it is in. If there is more than one or if *it* is a PHI, @@ -631,7 +663,7 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { if (isa(InVal)) return 0; // Itself a phi. if (NonConstBB) return 0; // More than one non-const value. - + NonConstBB = PN->getIncomingBlock(i); // If the InVal is an invoke at the end of the pred block, then we can't @@ -639,14 +671,14 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { if (InvokeInst *II = dyn_cast(InVal)) if (II->getParent() == NonConstBB) return 0; - + // If the incoming non-constant value is in I's block, we will remove one // instruction, but insert another equivalent one, leading to infinite // instcombine. if (NonConstBB == I.getParent()) return 0; } - + // If there is exactly one non-constant value, we can insert a copy of the // operation in that block. However, if this is a critical edge, we would be // inserting the computation one some other paths (e.g. inside a loop). Only @@ -660,12 +692,12 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues()); InsertNewInstBefore(NewPN, *PN); NewPN->takeName(PN); - + // If we are going to have to insert a new computation, do so right before the // predecessors terminator. if (NonConstBB) Builder->SetInsertPoint(NonConstBB->getTerminator()); - + // Next, add all of the operands to the PHI. if (SelectInst *SI = dyn_cast(&I)) { // We only currently try to fold the condition of a select when it is a phi, @@ -678,7 +710,10 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); Value *InV = 0; - if (Constant *InC = dyn_cast(PN->getIncomingValue(i))) + // Beware of ConstantExpr: it may eventually evaluate to getNullValue, + // even if currently isNullValue gives false. + Constant *InC = dyn_cast(PN->getIncomingValue(i)); + if (InC && !isa(InC)) InV = InC->isNullValue() ? FalseVInPred : TrueVInPred; else InV = Builder->CreateSelect(PN->getIncomingValue(i), @@ -710,22 +745,21 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { PN->getIncomingValue(i), C, "phitmp"); NewPN->addIncoming(InV, PN->getIncomingBlock(i)); } - } else { + } else { CastInst *CI = cast(&I); Type *RetTy = CI->getType(); for (unsigned i = 0; i != NumPHIValues; ++i) { Value *InV; if (Constant *InC = dyn_cast(PN->getIncomingValue(i))) InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy); - else + else InV = Builder->CreateCast(CI->getOpcode(), PN->getIncomingValue(i), I.getType(), "phitmp"); NewPN->addIncoming(InV, PN->getIncomingBlock(i)); } } - - for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); - UI != E; ) { + + for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) { Instruction *User = cast(*UI++); if (User == &I) continue; ReplaceInstUsesWith(*User, NewPN); @@ -734,24 +768,30 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { return ReplaceInstUsesWith(I, NewPN); } -/// FindElementAtOffset - Given a type and a constant offset, determine whether -/// or not there is a sequence of GEP indices into the type that will land us at -/// the specified offset. If so, fill them into NewIndices and return the -/// resultant element type, otherwise return null. -Type *InstCombiner::FindElementAtOffset(Type *Ty, int64_t Offset, - SmallVectorImpl &NewIndices) { - if (!TD) return 0; - if (!Ty->isSized()) return 0; - +/// FindElementAtOffset - Given a pointer type and a constant offset, determine +/// whether or not there is a sequence of GEP indices into the pointed type that +/// will land us at the specified offset. If so, fill them into NewIndices and +/// return the resultant element type, otherwise return null. +Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset, + SmallVectorImpl &NewIndices) { + assert(PtrTy->isPtrOrPtrVectorTy()); + + if (!DL) + return 0; + + Type *Ty = PtrTy->getPointerElementType(); + if (!Ty->isSized()) + return 0; + // Start with the index over the outer type. Note that the type size // might be zero (even if the offset isn't zero) if the indexed type // is something like [0 x {int, int}] - Type *IntPtrTy = TD->getIntPtrType(Ty->getContext()); + Type *IntPtrTy = DL->getIntPtrType(PtrTy); int64_t FirstIdx = 0; - if (int64_t TySize = TD->getTypeAllocSize(Ty)) { + if (int64_t TySize = DL->getTypeAllocSize(Ty)) { FirstIdx = Offset/TySize; Offset -= FirstIdx*TySize; - + // Handle hosts where % returns negative instead of values [0..TySize). if (Offset < 0) { --FirstIdx; @@ -760,28 +800,28 @@ Type *InstCombiner::FindElementAtOffset(Type *Ty, int64_t Offset, } assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset"); } - + NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx)); - + // Index into the types. If we fail, set OrigBase to null. while (Offset) { // Indexing into tail padding between struct/array elements. - if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty)) + if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty)) return 0; - + if (StructType *STy = dyn_cast(Ty)) { - const StructLayout *SL = TD->getStructLayout(STy); + const StructLayout *SL = DL->getStructLayout(STy); assert(Offset < (int64_t)SL->getSizeInBytes() && "Offset must stay within the indexed type"); - + unsigned Elt = SL->getElementContainingOffset(Offset); NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), Elt)); - + Offset -= SL->getElementOffset(Elt); Ty = STy->getElementType(Elt); } else if (ArrayType *AT = dyn_cast(Ty)) { - uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType()); + uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType()); assert(EltSize && "Cannot index into a zero-sized array"); NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize)); Offset %= EltSize; @@ -791,7 +831,7 @@ Type *InstCombiner::FindElementAtOffset(Type *Ty, int64_t Offset, return 0; } } - + return Ty; } @@ -805,19 +845,257 @@ static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) { return true; } +/// Descale - Return a value X such that Val = X * Scale, or null if none. If +/// the multiplication is known not to overflow then NoSignedWrap is set. +Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) { + assert(isa(Val->getType()) && "Can only descale integers!"); + assert(cast(Val->getType())->getBitWidth() == + Scale.getBitWidth() && "Scale not compatible with value!"); + + // If Val is zero or Scale is one then Val = Val * Scale. + if (match(Val, m_Zero()) || Scale == 1) { + NoSignedWrap = true; + return Val; + } + + // If Scale is zero then it does not divide Val. + if (Scale.isMinValue()) + return 0; + + // Look through chains of multiplications, searching for a constant that is + // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4 + // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by + // a factor of 4 will produce X*(Y*2). The principle of operation is to bore + // down from Val: + // + // Val = M1 * X || Analysis starts here and works down + // M1 = M2 * Y || Doesn't descend into terms with more + // M2 = Z * 4 \/ than one use + // + // Then to modify a term at the bottom: + // + // Val = M1 * X + // M1 = Z * Y || Replaced M2 with Z + // + // Then to work back up correcting nsw flags. + + // Op - the term we are currently analyzing. Starts at Val then drills down. + // Replaced with its descaled value before exiting from the drill down loop. + Value *Op = Val; + + // Parent - initially null, but after drilling down notes where Op came from. + // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the + // 0'th operand of Val. + std::pair Parent; + + // RequireNoSignedWrap - Set if the transform requires a descaling at deeper + // levels that doesn't overflow. + bool RequireNoSignedWrap = false; + + // logScale - log base 2 of the scale. Negative if not a power of 2. + int32_t logScale = Scale.exactLogBase2(); + + for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down + + if (ConstantInt *CI = dyn_cast(Op)) { + // If Op is a constant divisible by Scale then descale to the quotient. + APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth. + APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder); + if (!Remainder.isMinValue()) + // Not divisible by Scale. + return 0; + // Replace with the quotient in the parent. + Op = ConstantInt::get(CI->getType(), Quotient); + NoSignedWrap = true; + break; + } + + if (BinaryOperator *BO = dyn_cast(Op)) { + + if (BO->getOpcode() == Instruction::Mul) { + // Multiplication. + NoSignedWrap = BO->hasNoSignedWrap(); + if (RequireNoSignedWrap && !NoSignedWrap) + return 0; + + // There are three cases for multiplication: multiplication by exactly + // the scale, multiplication by a constant different to the scale, and + // multiplication by something else. + Value *LHS = BO->getOperand(0); + Value *RHS = BO->getOperand(1); + + if (ConstantInt *CI = dyn_cast(RHS)) { + // Multiplication by a constant. + if (CI->getValue() == Scale) { + // Multiplication by exactly the scale, replace the multiplication + // by its left-hand side in the parent. + Op = LHS; + break; + } + + // Otherwise drill down into the constant. + if (!Op->hasOneUse()) + return 0; + + Parent = std::make_pair(BO, 1); + continue; + } + + // Multiplication by something else. Drill down into the left-hand side + // since that's where the reassociate pass puts the good stuff. + if (!Op->hasOneUse()) + return 0; + + Parent = std::make_pair(BO, 0); + continue; + } + + if (logScale > 0 && BO->getOpcode() == Instruction::Shl && + isa(BO->getOperand(1))) { + // Multiplication by a power of 2. + NoSignedWrap = BO->hasNoSignedWrap(); + if (RequireNoSignedWrap && !NoSignedWrap) + return 0; + + Value *LHS = BO->getOperand(0); + int32_t Amt = cast(BO->getOperand(1))-> + getLimitedValue(Scale.getBitWidth()); + // Op = LHS << Amt. + + if (Amt == logScale) { + // Multiplication by exactly the scale, replace the multiplication + // by its left-hand side in the parent. + Op = LHS; + break; + } + if (Amt < logScale || !Op->hasOneUse()) + return 0; + + // Multiplication by more than the scale. Reduce the multiplying amount + // by the scale in the parent. + Parent = std::make_pair(BO, 1); + Op = ConstantInt::get(BO->getType(), Amt - logScale); + break; + } + } + + if (!Op->hasOneUse()) + return 0; + + if (CastInst *Cast = dyn_cast(Op)) { + if (Cast->getOpcode() == Instruction::SExt) { + // Op is sign-extended from a smaller type, descale in the smaller type. + unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); + APInt SmallScale = Scale.trunc(SmallSize); + // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to + // descale Op as (sext Y) * Scale. In order to have + // sext (Y * SmallScale) = (sext Y) * Scale + // some conditions need to hold however: SmallScale must sign-extend to + // Scale and the multiplication Y * SmallScale should not overflow. + if (SmallScale.sext(Scale.getBitWidth()) != Scale) + // SmallScale does not sign-extend to Scale. + return 0; + assert(SmallScale.exactLogBase2() == logScale); + // Require that Y * SmallScale must not overflow. + RequireNoSignedWrap = true; + + // Drill down through the cast. + Parent = std::make_pair(Cast, 0); + Scale = SmallScale; + continue; + } + + if (Cast->getOpcode() == Instruction::Trunc) { + // Op is truncated from a larger type, descale in the larger type. + // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then + // trunc (Y * sext Scale) = (trunc Y) * Scale + // always holds. However (trunc Y) * Scale may overflow even if + // trunc (Y * sext Scale) does not, so nsw flags need to be cleared + // from this point up in the expression (see later). + if (RequireNoSignedWrap) + return 0; + + // Drill down through the cast. + unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); + Parent = std::make_pair(Cast, 0); + Scale = Scale.sext(LargeSize); + if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits()) + logScale = -1; + assert(Scale.exactLogBase2() == logScale); + continue; + } + } + + // Unsupported expression, bail out. + return 0; + } + + // We know that we can successfully descale, so from here on we can safely + // modify the IR. Op holds the descaled version of the deepest term in the + // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known + // not to overflow. + + if (!Parent.first) + // The expression only had one term. + return Op; + + // Rewrite the parent using the descaled version of its operand. + assert(Parent.first->hasOneUse() && "Drilled down when more than one use!"); + assert(Op != Parent.first->getOperand(Parent.second) && + "Descaling was a no-op?"); + Parent.first->setOperand(Parent.second, Op); + Worklist.Add(Parent.first); + + // Now work back up the expression correcting nsw flags. The logic is based + // on the following observation: if X * Y is known not to overflow as a signed + // multiplication, and Y is replaced by a value Z with smaller absolute value, + // then X * Z will not overflow as a signed multiplication either. As we work + // our way up, having NoSignedWrap 'true' means that the descaled value at the + // current level has strictly smaller absolute value than the original. + Instruction *Ancestor = Parent.first; + do { + if (BinaryOperator *BO = dyn_cast(Ancestor)) { + // If the multiplication wasn't nsw then we can't say anything about the + // value of the descaled multiplication, and we have to clear nsw flags + // from this point on up. + bool OpNoSignedWrap = BO->hasNoSignedWrap(); + NoSignedWrap &= OpNoSignedWrap; + if (NoSignedWrap != OpNoSignedWrap) { + BO->setHasNoSignedWrap(NoSignedWrap); + Worklist.Add(Ancestor); + } + } else if (Ancestor->getOpcode() == Instruction::Trunc) { + // The fact that the descaled input to the trunc has smaller absolute + // value than the original input doesn't tell us anything useful about + // the absolute values of the truncations. + NoSignedWrap = false; + } + assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) && + "Failed to keep proper track of nsw flags while drilling down?"); + + if (Ancestor == Val) + // Got to the top, all done! + return Val; + + // Move up one level in the expression. + assert(Ancestor->hasOneUse() && "Drilled down when more than one use!"); + Ancestor = Ancestor->user_back(); + } while (1); +} + Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { SmallVector Ops(GEP.op_begin(), GEP.op_end()); - if (Value *V = SimplifyGEPInst(Ops, TD)) + if (Value *V = SimplifyGEPInst(Ops, DL)) return ReplaceInstUsesWith(GEP, V); Value *PtrOp = GEP.getOperand(0); // Eliminate unneeded casts for indices, and replace indices which displace // by multiples of a zero size type with zero. - if (TD) { + if (DL) { bool MadeChange = false; - Type *IntPtrTy = TD->getIntPtrType(GEP.getContext()); + Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType()); gep_type_iterator GTI = gep_type_begin(GEP); for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); @@ -829,14 +1107,14 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { // If the element type has zero size then any index over it is equivalent // to an index of zero, so replace it with zero if it is not zero already. if (SeqTy->getElementType()->isSized() && - TD->getTypeAllocSize(SeqTy->getElementType()) == 0) + DL->getTypeAllocSize(SeqTy->getElementType()) == 0) if (!isa(*I) || !cast(*I)->isNullValue()) { *I = Constant::getNullValue(IntPtrTy); MadeChange = true; } Type *IndexTy = (*I)->getType(); - if (IndexTy != IntPtrTy && !IndexTy->isVectorTy()) { + if (IndexTy != IntPtrTy) { // If we are using a wider index than needed for this platform, shrink // it to what we need. If narrower, sign-extend it to what we need. // This explicit cast can make subsequent optimizations more obvious. @@ -855,7 +1133,7 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { if (!shouldMergeGEPs(*cast(&GEP), *Src)) return 0; - // Note that if our source is a gep chain itself that we wait for that + // Note that if our source is a gep chain itself then we wait for that // chain to be resolved before we perform this transformation. This // avoids us creating a TON of code in some cases. if (GEPOperator *SrcGEP = @@ -917,16 +1195,31 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName()); } + // Canonicalize (gep i8* X, -(ptrtoint Y)) to (sub (ptrtoint X), (ptrtoint Y)) + // The GEP pattern is emitted by the SCEV expander for certain kinds of + // pointer arithmetic. + if (DL && GEP.getNumIndices() == 1 && + match(GEP.getOperand(1), m_Neg(m_PtrToInt(m_Value())))) { + unsigned AS = GEP.getPointerAddressSpace(); + if (GEP.getType() == Builder->getInt8PtrTy(AS) && + GEP.getOperand(1)->getType()->getScalarSizeInBits() == + DL->getPointerSizeInBits(AS)) { + Operator *Index = cast(GEP.getOperand(1)); + Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType()); + Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1)); + return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType()); + } + } + // Handle gep(bitcast x) and gep(gep x, 0, 0, 0). Value *StrippedPtr = PtrOp->stripPointerCasts(); PointerType *StrippedPtrTy = dyn_cast(StrippedPtr->getType()); - // We do not handle pointer-vector geps here - if (!StrippedPtr) - return 0; - if (StrippedPtr != PtrOp && - StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { + // We do not handle pointer-vector geps here. + if (!StrippedPtrTy) + return 0; + if (StrippedPtr != PtrOp) { bool HasZeroPointerIndex = false; if (ConstantInt *C = dyn_cast(GEP.getOperand(1))) HasZeroPointerIndex = C->isZero(); @@ -949,9 +1242,17 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { GetElementPtrInst *Res = GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName()); Res->setIsInBounds(GEP.isInBounds()); - return Res; + if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) + return Res; + // Insert Res, and create an addrspacecast. + // e.g., + // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ... + // -> + // %0 = GEP i8 addrspace(1)* X, ... + // addrspacecast i8 addrspace(1)* %0 to i8* + return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType()); } - + if (ArrayType *XATy = dyn_cast(StrippedPtrTy->getElementType())){ // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ? @@ -961,8 +1262,24 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { // to an array of the same type as the destination pointer // array. Because the array type is never stepped over (there // is a leading zero) we can fold the cast into this GEP. - GEP.setOperand(0, StrippedPtr); - return &GEP; + if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) { + GEP.setOperand(0, StrippedPtr); + return &GEP; + } + // Cannot replace the base pointer directly because StrippedPtr's + // address space is different. Instead, create a new GEP followed by + // an addrspacecast. + // e.g., + // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*), + // i32 0, ... + // -> + // %0 = GEP [10 x i8] addrspace(1)* X, ... + // addrspacecast i8 addrspace(1)* %0 to i8* + SmallVector Idx(GEP.idx_begin(), GEP.idx_end()); + Value *NewGEP = GEP.isInBounds() ? + Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) : + Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); + return new AddrSpaceCastInst(NewGEP, GEP.getType()); } } } @@ -971,104 +1288,125 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast Type *SrcElTy = StrippedPtrTy->getElementType(); - Type *ResElTy=cast(PtrOp->getType())->getElementType(); - if (TD && SrcElTy->isArrayTy() && - TD->getTypeAllocSize(cast(SrcElTy)->getElementType()) == - TD->getTypeAllocSize(ResElTy)) { - Value *Idx[2]; - Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); - Idx[1] = GEP.getOperand(1); + Type *ResElTy = PtrOp->getType()->getPointerElementType(); + if (DL && SrcElTy->isArrayTy() && + DL->getTypeAllocSize(SrcElTy->getArrayElementType()) == + DL->getTypeAllocSize(ResElTy)) { + Type *IdxType = DL->getIntPtrType(GEP.getType()); + Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) }; Value *NewGEP = GEP.isInBounds() ? Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) : Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); + // V and GEP are both pointer types --> BitCast - return new BitCastInst(NewGEP, GEP.getType()); + if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) + return new BitCastInst(NewGEP, GEP.getType()); + return new AddrSpaceCastInst(NewGEP, GEP.getType()); } - + // Transform things like: + // %V = mul i64 %N, 4 + // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V + // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast + if (DL && ResElTy->isSized() && SrcElTy->isSized()) { + // Check that changing the type amounts to dividing the index by a scale + // factor. + uint64_t ResSize = DL->getTypeAllocSize(ResElTy); + uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy); + if (ResSize && SrcSize % ResSize == 0) { + Value *Idx = GEP.getOperand(1); + unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); + uint64_t Scale = SrcSize / ResSize; + + // Earlier transforms ensure that the index has type IntPtrType, which + // considerably simplifies the logic by eliminating implicit casts. + assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) && + "Index not cast to pointer width?"); + + bool NSW; + if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { + // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. + // If the multiplication NewIdx * Scale may overflow then the new + // GEP may not be "inbounds". + Value *NewGEP = GEP.isInBounds() && NSW ? + Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) : + Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName()); + + // The NewGEP must be pointer typed, so must the old one -> BitCast + if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) + return new BitCastInst(NewGEP, GEP.getType()); + return new AddrSpaceCastInst(NewGEP, GEP.getType()); + } + } + } + + // Similarly, transform things like: // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp // (where tmp = 8*tmp2) into: // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast - - if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) { - uint64_t ArrayEltSize = - TD->getTypeAllocSize(cast(SrcElTy)->getElementType()); - - // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We - // allow either a mul, shift, or constant here. - Value *NewIdx = 0; - ConstantInt *Scale = 0; - if (ArrayEltSize == 1) { - NewIdx = GEP.getOperand(1); - Scale = ConstantInt::get(cast(NewIdx->getType()), 1); - } else if (ConstantInt *CI = dyn_cast(GEP.getOperand(1))) { - NewIdx = ConstantInt::get(CI->getType(), 1); - Scale = CI; - } else if (Instruction *Inst =dyn_cast(GEP.getOperand(1))){ - if (Inst->getOpcode() == Instruction::Shl && - isa(Inst->getOperand(1))) { - ConstantInt *ShAmt = cast(Inst->getOperand(1)); - uint32_t ShAmtVal = ShAmt->getLimitedValue(64); - Scale = ConstantInt::get(cast(Inst->getType()), - 1ULL << ShAmtVal); - NewIdx = Inst->getOperand(0); - } else if (Inst->getOpcode() == Instruction::Mul && - isa(Inst->getOperand(1))) { - Scale = cast(Inst->getOperand(1)); - NewIdx = Inst->getOperand(0); + if (DL && ResElTy->isSized() && SrcElTy->isSized() && + SrcElTy->isArrayTy()) { + // Check that changing to the array element type amounts to dividing the + // index by a scale factor. + uint64_t ResSize = DL->getTypeAllocSize(ResElTy); + uint64_t ArrayEltSize + = DL->getTypeAllocSize(SrcElTy->getArrayElementType()); + if (ResSize && ArrayEltSize % ResSize == 0) { + Value *Idx = GEP.getOperand(1); + unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); + uint64_t Scale = ArrayEltSize / ResSize; + + // Earlier transforms ensure that the index has type IntPtrType, which + // considerably simplifies the logic by eliminating implicit casts. + assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) && + "Index not cast to pointer width?"); + + bool NSW; + if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { + // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. + // If the multiplication NewIdx * Scale may overflow then the new + // GEP may not be "inbounds". + Value *Off[2] = { + Constant::getNullValue(DL->getIntPtrType(GEP.getType())), + NewIdx + }; + + Value *NewGEP = GEP.isInBounds() && NSW ? + Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) : + Builder->CreateGEP(StrippedPtr, Off, GEP.getName()); + // The NewGEP must be pointer typed, so must the old one -> BitCast + if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) + return new BitCastInst(NewGEP, GEP.getType()); + return new AddrSpaceCastInst(NewGEP, GEP.getType()); } } - - // If the index will be to exactly the right offset with the scale taken - // out, perform the transformation. Note, we don't know whether Scale is - // signed or not. We'll use unsigned version of division/modulo - // operation after making sure Scale doesn't have the sign bit set. - if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL && - Scale->getZExtValue() % ArrayEltSize == 0) { - Scale = ConstantInt::get(Scale->getType(), - Scale->getZExtValue() / ArrayEltSize); - if (Scale->getZExtValue() != 1) { - Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(), - false /*ZExt*/); - NewIdx = Builder->CreateMul(NewIdx, C, "idxscale"); - } - - // Insert the new GEP instruction. - Value *Idx[2]; - Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); - Idx[1] = NewIdx; - Value *NewGEP = GEP.isInBounds() ? - Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()): - Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); - // The NewGEP must be pointer typed, so must the old one -> BitCast - return new BitCastInst(NewGEP, GEP.getType()); - } } } } + if (!DL) + return 0; + /// See if we can simplify: /// X = bitcast A* to B* /// Y = gep X, <...constant indices...> /// into a gep of the original struct. This is important for SROA and alias /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged. if (BitCastInst *BCI = dyn_cast(PtrOp)) { - if (TD && - !isa(BCI->getOperand(0)) && GEP.hasAllConstantIndices() && + Value *Operand = BCI->getOperand(0); + PointerType *OpType = cast(Operand->getType()); + unsigned OffsetBits = DL->getPointerTypeSizeInBits(OpType); + APInt Offset(OffsetBits, 0); + if (!isa(Operand) && + GEP.accumulateConstantOffset(*DL, Offset) && StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { - // Determine how much the GEP moves the pointer. We are guaranteed to get - // a constant back from EmitGEPOffset. - ConstantInt *OffsetV = cast(EmitGEPOffset(&GEP)); - int64_t Offset = OffsetV->getSExtValue(); - // If this GEP instruction doesn't move the pointer, just replace the GEP // with a bitcast of the real input to the dest type. - if (Offset == 0) { + if (!Offset) { // If the bitcast is of an allocation, and the allocation will be // converted to match the type of the cast, don't touch this. - if (isa(BCI->getOperand(0)) || - isMalloc(BCI->getOperand(0))) { + if (isa(Operand) || isAllocationFn(Operand, TLI)) { // See if the bitcast simplifies, if so, don't nuke this GEP yet. if (Instruction *I = visitBitCast(*BCI)) { if (I != BCI) { @@ -1079,81 +1417,112 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { return &GEP; } } - return new BitCastInst(BCI->getOperand(0), GEP.getType()); + return new BitCastInst(Operand, GEP.getType()); } - + // Otherwise, if the offset is non-zero, we need to find out if there is a // field at Offset in 'A's type. If so, we can pull the cast through the // GEP. SmallVector NewIndices; - Type *InTy = - cast(BCI->getOperand(0)->getType())->getElementType(); - if (FindElementAtOffset(InTy, Offset, NewIndices)) { + if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) { Value *NGEP = GEP.isInBounds() ? - Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices) : - Builder->CreateGEP(BCI->getOperand(0), NewIndices); - + Builder->CreateInBoundsGEP(Operand, NewIndices) : + Builder->CreateGEP(Operand, NewIndices); + if (NGEP->getType() == GEP.getType()) return ReplaceInstUsesWith(GEP, NGEP); NGEP->takeName(&GEP); return new BitCastInst(NGEP, GEP.getType()); } } - } - + } + return 0; } +static bool +isAllocSiteRemovable(Instruction *AI, SmallVectorImpl &Users, + const TargetLibraryInfo *TLI) { + SmallVector Worklist; + Worklist.push_back(AI); + do { + Instruction *PI = Worklist.pop_back_val(); + for (User *U : PI->users()) { + Instruction *I = cast(U); + switch (I->getOpcode()) { + default: + // Give up the moment we see something we can't handle. + return false; -static bool IsOnlyNullComparedAndFreed(Value *V, SmallVectorImpl &Users, - int Depth = 0) { - if (Depth == 8) - return false; - - for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); - UI != UE; ++UI) { - User *U = *UI; - if (isFreeCall(U)) { - Users.push_back(U); - continue; - } - if (ICmpInst *ICI = dyn_cast(U)) { - if (ICI->isEquality() && isa(ICI->getOperand(1))) { - Users.push_back(ICI); + case Instruction::BitCast: + case Instruction::GetElementPtr: + Users.push_back(I); + Worklist.push_back(I); continue; - } - } - if (BitCastInst *BCI = dyn_cast(U)) { - if (IsOnlyNullComparedAndFreed(BCI, Users, Depth+1)) { - Users.push_back(BCI); + + case Instruction::ICmp: { + ICmpInst *ICI = cast(I); + // We can fold eq/ne comparisons with null to false/true, respectively. + if (!ICI->isEquality() || !isa(ICI->getOperand(1))) + return false; + Users.push_back(I); continue; } - } - if (GetElementPtrInst *GEPI = dyn_cast(U)) { - if (IsOnlyNullComparedAndFreed(GEPI, Users, Depth+1)) { - Users.push_back(GEPI); + + case Instruction::Call: + // Ignore no-op and store intrinsics. + if (IntrinsicInst *II = dyn_cast(I)) { + switch (II->getIntrinsicID()) { + default: + return false; + + case Intrinsic::memmove: + case Intrinsic::memcpy: + case Intrinsic::memset: { + MemIntrinsic *MI = cast(II); + if (MI->isVolatile() || MI->getRawDest() != PI) + return false; + } + // fall through + case Intrinsic::dbg_declare: + case Intrinsic::dbg_value: + case Intrinsic::invariant_start: + case Intrinsic::invariant_end: + case Intrinsic::lifetime_start: + case Intrinsic::lifetime_end: + case Intrinsic::objectsize: + Users.push_back(I); + continue; + } + } + + if (isFreeCall(I, TLI)) { + Users.push_back(I); + continue; + } + return false; + + case Instruction::Store: { + StoreInst *SI = cast(I); + if (SI->isVolatile() || SI->getPointerOperand() != PI) + return false; + Users.push_back(I); continue; } - } - if (IntrinsicInst *II = dyn_cast(U)) { - if (II->getIntrinsicID() == Intrinsic::lifetime_start || - II->getIntrinsicID() == Intrinsic::lifetime_end) { - Users.push_back(II); - continue; } + llvm_unreachable("missing a return?"); } - return false; - } + } while (!Worklist.empty()); return true; } -Instruction *InstCombiner::visitMalloc(Instruction &MI) { +Instruction *InstCombiner::visitAllocSite(Instruction &MI) { // If we have a malloc call which is only used in any amount of comparisons // to null and free calls, delete the calls and replace the comparisons with // true or false as appropriate. SmallVector Users; - if (IsOnlyNullComparedAndFreed(&MI, Users)) { + if (isAllocSiteRemovable(&MI, Users, TLI)) { for (unsigned i = 0, e = Users.size(); i != e; ++i) { Instruction *I = cast_or_null(&*Users[i]); if (!I) continue; @@ -1164,14 +1533,84 @@ Instruction *InstCombiner::visitMalloc(Instruction &MI) { C->isFalseWhenEqual())); } else if (isa(I) || isa(I)) { ReplaceInstUsesWith(*I, UndefValue::get(I->getType())); + } else if (IntrinsicInst *II = dyn_cast(I)) { + if (II->getIntrinsicID() == Intrinsic::objectsize) { + ConstantInt *CI = cast(II->getArgOperand(1)); + uint64_t DontKnow = CI->isZero() ? -1ULL : 0; + ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow)); + } } EraseInstFromFunction(*I); } + + if (InvokeInst *II = dyn_cast(&MI)) { + // Replace invoke with a NOP intrinsic to maintain the original CFG + Module *M = II->getParent()->getParent()->getParent(); + Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing); + InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(), + None, "", II->getParent()); + } return EraseInstFromFunction(MI); } return 0; } +/// \brief Move the call to free before a NULL test. +/// +/// Check if this free is accessed after its argument has been test +/// against NULL (property 0). +/// If yes, it is legal to move this call in its predecessor block. +/// +/// The move is performed only if the block containing the call to free +/// will be removed, i.e.: +/// 1. it has only one predecessor P, and P has two successors +/// 2. it contains the call and an unconditional branch +/// 3. its successor is the same as its predecessor's successor +/// +/// The profitability is out-of concern here and this function should +/// be called only if the caller knows this transformation would be +/// profitable (e.g., for code size). +static Instruction * +tryToMoveFreeBeforeNullTest(CallInst &FI) { + Value *Op = FI.getArgOperand(0); + BasicBlock *FreeInstrBB = FI.getParent(); + BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor(); + + // Validate part of constraint #1: Only one predecessor + // FIXME: We can extend the number of predecessor, but in that case, we + // would duplicate the call to free in each predecessor and it may + // not be profitable even for code size. + if (!PredBB) + return 0; + + // Validate constraint #2: Does this block contains only the call to + // free and an unconditional branch? + // FIXME: We could check if we can speculate everything in the + // predecessor block + if (FreeInstrBB->size() != 2) + return 0; + BasicBlock *SuccBB; + if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB))) + return 0; + + // Validate the rest of constraint #1 by matching on the pred branch. + TerminatorInst *TI = PredBB->getTerminator(); + BasicBlock *TrueBB, *FalseBB; + ICmpInst::Predicate Pred; + if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB))) + return 0; + if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE) + return 0; + + // Validate constraint #3: Ensure the null case just falls through. + if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB)) + return 0; + assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) && + "Broken CFG: missing edge from predecessor to successor"); + + FI.moveBefore(TI); + return &FI; +} Instruction *InstCombiner::visitFree(CallInst &FI) { @@ -1184,12 +1623,22 @@ Instruction *InstCombiner::visitFree(CallInst &FI) { UndefValue::get(Type::getInt1PtrTy(FI.getContext()))); return EraseInstFromFunction(FI); } - + // If we have 'free null' delete the instruction. This can happen in stl code // when lots of inlining happens. if (isa(Op)) return EraseInstFromFunction(FI); + // If we optimize for code size, try to move the call to free before the null + // test so that simplify cfg can remove the empty block and dead code + // elimination the branch. I.e., helps to turn something like: + // if (foo) free(foo); + // into + // free(foo); + if (MinimizeSize) + if (Instruction *I = tryToMoveFreeBeforeNullTest(FI)) + return I; + return 0; } @@ -1208,23 +1657,23 @@ Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { return &BI; } - // Cannonicalize fcmp_one -> fcmp_oeq + // Canonicalize fcmp_one -> fcmp_oeq FCmpInst::Predicate FPred; Value *Y; - if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), + if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)), TrueDest, FalseDest)) && BI.getCondition()->hasOneUse()) if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE || FPred == FCmpInst::FCMP_OGE) { FCmpInst *Cond = cast(BI.getCondition()); Cond->setPredicate(FCmpInst::getInversePredicate(FPred)); - + // Swap Destinations and condition. BI.swapSuccessors(); Worklist.Add(Cond); return &BI; } - // Cannonicalize icmp_ne -> icmp_eq + // Canonicalize icmp_ne -> icmp_eq ICmpInst::Predicate IPred; if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)), TrueDest, FalseDest)) && @@ -1249,15 +1698,15 @@ Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { if (I->getOpcode() == Instruction::Add) if (ConstantInt *AddRHS = dyn_cast(I->getOperand(1))) { // change 'switch (X+4) case 1:' into 'switch (X) case -3' - unsigned NumCases = SI.getNumCases(); // Skip the first item since that's the default case. - for (unsigned i = 1; i < NumCases; ++i) { - ConstantInt* CaseVal = SI.getCaseValue(i); + for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end(); + i != e; ++i) { + ConstantInt* CaseVal = i.getCaseValue(); Constant* NewCaseVal = ConstantExpr::getSub(cast(CaseVal), AddRHS); assert(isa(NewCaseVal) && "Result of expression should be constant"); - SI.setSuccessorValue(i, cast(NewCaseVal)); + i.setValue(cast(NewCaseVal)); } SI.setCondition(I->getOperand(0)); Worklist.Add(I); @@ -1283,7 +1732,7 @@ Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { } return 0; // Can't handle other constants } - + if (InsertValueInst *IV = dyn_cast(Agg)) { // We're extracting from an insertvalue instruction, compare the indices const unsigned *exti, *exte, *insi, *inse; @@ -1332,7 +1781,7 @@ Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { // %E = extractvalue { i32, { i32 } } %I, 1, 0 // with // %E extractvalue { i32 } { i32 42 }, 0 - return ExtractValueInst::Create(IV->getInsertedValueOperand(), + return ExtractValueInst::Create(IV->getInsertedValueOperand(), makeArrayRef(exti, exte)); } if (IntrinsicInst *II = dyn_cast(Agg)) { @@ -1352,7 +1801,7 @@ Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { EraseInstFromFunction(*II); return BinaryOperator::CreateAdd(LHS, RHS); } - + // If the normal result of the add is dead, and the RHS is a constant, // we can transform this into a range comparison. // overflow = uadd a, -4 --> overflow = icmp ugt a, 3 @@ -1658,7 +2107,7 @@ Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { continue; // If Filter is a subset of LFilter, i.e. every element of Filter is also // an element of LFilter, then discard LFilter. - SmallVector::iterator J = NewClauses.begin() + j; + SmallVectorImpl::iterator J = NewClauses.begin() + j; // If Filter is empty then it is a subset of LFilter. if (!FElts) { // Discard LFilter. @@ -1801,10 +2250,10 @@ static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { /// many instructions are dead or constant). Additionally, if we find a branch /// whose condition is a known constant, we only visit the reachable successors. /// -static bool AddReachableCodeToWorklist(BasicBlock *BB, +static bool AddReachableCodeToWorklist(BasicBlock *BB, SmallPtrSet &Visited, InstCombiner &IC, - const TargetData *TD, + const DataLayout *DL, const TargetLibraryInfo *TLI) { bool MadeIRChange = false; SmallVector Worklist; @@ -1815,33 +2264,33 @@ static bool AddReachableCodeToWorklist(BasicBlock *BB, do { BB = Worklist.pop_back_val(); - + // We have now visited this block! If we've already been here, ignore it. if (!Visited.insert(BB)) continue; for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { Instruction *Inst = BBI++; - + // DCE instruction if trivially dead. - if (isInstructionTriviallyDead(Inst)) { + if (isInstructionTriviallyDead(Inst, TLI)) { ++NumDeadInst; - DEBUG(errs() << "IC: DCE: " << *Inst << '\n'); + DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n'); Inst->eraseFromParent(); continue; } - + // ConstantProp instruction if trivially constant. if (!Inst->use_empty() && isa(Inst->getOperand(0))) - if (Constant *C = ConstantFoldInstruction(Inst, TD, TLI)) { - DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " + if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) { + DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *Inst << '\n'); Inst->replaceAllUsesWith(C); ++NumConstProp; Inst->eraseFromParent(); continue; } - - if (TD) { + + if (DL) { // See if we can constant fold its operands. for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); i != e; ++i) { @@ -1850,7 +2299,7 @@ static bool AddReachableCodeToWorklist(BasicBlock *BB, Constant*& FoldRes = FoldedConstants[CE]; if (!FoldRes) - FoldRes = ConstantFoldConstantExpression(CE, TD, TLI); + FoldRes = ConstantFoldConstantExpression(CE, DL, TLI); if (!FoldRes) FoldRes = CE; @@ -1877,23 +2326,24 @@ static bool AddReachableCodeToWorklist(BasicBlock *BB, } else if (SwitchInst *SI = dyn_cast(TI)) { if (ConstantInt *Cond = dyn_cast(SI->getCondition())) { // See if this is an explicit destination. - for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) - if (SI->getCaseValue(i) == Cond) { - BasicBlock *ReachableBB = SI->getSuccessor(i); + for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); + i != e; ++i) + if (i.getCaseValue() == Cond) { + BasicBlock *ReachableBB = i.getCaseSuccessor(); Worklist.push_back(ReachableBB); continue; } - + // Otherwise it is the default destination. - Worklist.push_back(SI->getSuccessor(0)); + Worklist.push_back(SI->getDefaultDest()); continue; } } - + for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) Worklist.push_back(TI->getSuccessor(i)); } while (!Worklist.empty()); - + // Once we've found all of the instructions to add to instcombine's worklist, // add them in reverse order. This way instcombine will visit from the top // of the function down. This jives well with the way that it adds all uses @@ -1901,14 +2351,14 @@ static bool AddReachableCodeToWorklist(BasicBlock *BB, // some N^2 behavior in pathological cases. IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0], InstrsForInstCombineWorklist.size()); - + return MadeIRChange; } bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { MadeIRChange = false; - - DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " + + DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on " << F.getName() << "\n"); { @@ -1916,7 +2366,7 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { // the reachable instructions. Ignore blocks that are not reachable. Keep // track of which blocks we visit. SmallPtrSet Visited; - MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD, + MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL, TLI); // Do a quick scan over the function. If we find any blocks that are @@ -1952,8 +2402,8 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { if (I == 0) continue; // skip null values. // Check to see if we can DCE the instruction. - if (isInstructionTriviallyDead(I)) { - DEBUG(errs() << "IC: DCE: " << *I << '\n'); + if (isInstructionTriviallyDead(I, TLI)) { + DEBUG(dbgs() << "IC: DCE: " << *I << '\n'); EraseInstFromFunction(*I); ++NumDeadInst; MadeIRChange = true; @@ -1962,8 +2412,8 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { // Instruction isn't dead, see if we can constant propagate it. if (!I->use_empty() && isa(I->getOperand(0))) - if (Constant *C = ConstantFoldInstruction(I, TD, TLI)) { - DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); + if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) { + DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n'); // Add operands to the worklist. ReplaceInstUsesWith(*I, C); @@ -1976,15 +2426,15 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { // See if we can trivially sink this instruction to a successor basic block. if (I->hasOneUse()) { BasicBlock *BB = I->getParent(); - Instruction *UserInst = cast(I->use_back()); + Instruction *UserInst = cast(*I->user_begin()); BasicBlock *UserParent; - + // Get the block the use occurs in. if (PHINode *PN = dyn_cast(UserInst)) - UserParent = PN->getIncomingBlock(I->use_begin().getUse()); + UserParent = PN->getIncomingBlock(*I->use_begin()); else UserParent = UserInst->getParent(); - + if (UserParent != BB) { bool UserIsSuccessor = false; // See if the user is one of our successors. @@ -2006,18 +2456,18 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { // Now that we have an instruction, try combining it to simplify it. Builder->SetInsertPoint(I->getParent(), I); Builder->SetCurrentDebugLocation(I->getDebugLoc()); - + #ifndef NDEBUG std::string OrigI; #endif DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str();); - DEBUG(errs() << "IC: Visiting: " << OrigI << '\n'); + DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n'); if (Instruction *Result = visit(*I)) { ++NumCombined; // Should we replace the old instruction with a new one? if (Result != I) { - DEBUG(errs() << "IC: Old = " << *I << '\n' + DEBUG(dbgs() << "IC: Old = " << *I << '\n' << " New = " << *Result << '\n'); if (!I->getDebugLoc().isUnknown()) @@ -2046,13 +2496,13 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { EraseInstFromFunction(*I); } else { #ifndef NDEBUG - DEBUG(errs() << "IC: Mod = " << OrigI << '\n' + DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n' << " New = " << *I << '\n'); #endif // If the instruction was modified, it's possible that it is now dead. // if so, remove it. - if (isInstructionTriviallyDead(I)) { + if (isInstructionTriviallyDead(I, TLI)) { EraseInstFromFunction(*I); } else { Worklist.Add(I); @@ -2067,18 +2517,46 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { return MadeIRChange; } +namespace { +class InstCombinerLibCallSimplifier : public LibCallSimplifier { + InstCombiner *IC; +public: + InstCombinerLibCallSimplifier(const DataLayout *DL, + const TargetLibraryInfo *TLI, + InstCombiner *IC) + : LibCallSimplifier(DL, TLI, UnsafeFPShrink) { + this->IC = IC; + } + + /// replaceAllUsesWith - override so that instruction replacement + /// can be defined in terms of the instruction combiner framework. + void replaceAllUsesWith(Instruction *I, Value *With) const override { + IC->ReplaceInstUsesWith(*I, With); + } +}; +} bool InstCombiner::runOnFunction(Function &F) { - TD = getAnalysisIfAvailable(); + if (skipOptnoneFunction(F)) + return false; + + DataLayoutPass *DLP = getAnalysisIfAvailable(); + DL = DLP ? &DLP->getDataLayout() : 0; TLI = &getAnalysis(); - + // Minimizing size? + MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, + Attribute::MinSize); + /// Builder - This is an IRBuilder that automatically inserts new /// instructions into the worklist when they are created. - IRBuilder - TheBuilder(F.getContext(), TargetFolder(TD), + IRBuilder + TheBuilder(F.getContext(), TargetFolder(DL), InstCombineIRInserter(Worklist)); Builder = &TheBuilder; - + + InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this); + Simplifier = &TheSimplifier; + bool EverMadeChange = false; // Lower dbg.declare intrinsics otherwise their value may be clobbered @@ -2089,7 +2567,7 @@ bool InstCombiner::runOnFunction(Function &F) { unsigned Iteration = 0; while (DoOneIteration(F, Iteration++)) EverMadeChange = true; - + Builder = 0; return EverMadeChange; }