X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTransforms%2FInstCombine%2FInstructionCombining.cpp;h=8d74976cb18b33f4407fad9f86b80933623edfaa;hp=88405c9d4d8dec35ba298bd9ee87a882a70cc85c;hb=4eb03123dfda2de88a84852834845678833c8c36;hpb=78f8ef42173a3a9867ed789073d4ddc652fb7ff2 diff --git a/lib/Transforms/InstCombine/InstructionCombining.cpp b/lib/Transforms/InstCombine/InstructionCombining.cpp index 88405c9d4d8..8d74976cb18 100644 --- a/lib/Transforms/InstCombine/InstructionCombining.cpp +++ b/lib/Transforms/InstCombine/InstructionCombining.cpp @@ -33,30 +33,35 @@ // //===----------------------------------------------------------------------===// -#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/AssumptionTracker.h" #include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/MemoryBuiltins.h" -#include "llvm/Target/TargetData.h" +#include "llvm/Analysis/ValueTracking.h" +#include "llvm/IR/CFG.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/Dominators.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 +70,12 @@ STATISTIC(NumExpand, "Number of expansions"); STATISTIC(NumFactor , "Number of factorizations"); STATISTIC(NumReassoc , "Number of reassociations"); +static cl::opt + EnableUnsafeFPShrink("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); @@ -77,18 +88,20 @@ void LLVMInitializeInstCombine(LLVMPassRegistryRef R) { char InstCombiner::ID = 0; INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine", "Combine redundant instructions", false, false) +INITIALIZE_PASS_DEPENDENCY(AssumptionTracker) INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) INITIALIZE_PASS_END(InstCombiner, "instcombine", "Combine redundant instructions", false, false) void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { AU.setPreservesCFG(); + AU.addRequired(); AU.addRequired(); } Value *InstCombiner::EmitGEPOffset(User *GEP) { - return llvm::EmitGEPOffset(Builder, *getTargetData(), GEP); + return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP); } /// ShouldChangeType - Return true if it is desirable to convert a computation @@ -97,13 +110,13 @@ Value *InstCombiner::EmitGEPOffset(User *GEP) { 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. @@ -156,6 +169,21 @@ 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: // @@ -200,20 +228,20 @@ 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; @@ -229,13 +257,13 @@ 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; @@ -251,13 +279,13 @@ 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; @@ -271,13 +299,13 @@ 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; @@ -298,13 +326,19 @@ 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; @@ -361,12 +395,163 @@ static bool RightDistributesOverLeft(Instruction::BinaryOps LOp, Instruction::BinaryOps ROp) { if (Instruction::isCommutative(ROp)) return LeftDistributesOverRight(ROp, LOp); + + switch (LOp) { + default: + return false; + // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts. + // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts. + // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts. + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + switch (ROp) { + default: + return false; + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + return true; + } + } // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z", // but this requires knowing that the addition does not overflow and other // such subtleties. return false; } +/// This function returns identity value for given opcode, which can be used to +/// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1). +static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) { + if (isa(V)) + return nullptr; + + if (OpCode == Instruction::Mul) + return ConstantInt::get(V->getType(), 1); + + // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc. + + return nullptr; +} + +/// This function factors binary ops which can be combined using distributive +/// laws. This function tries to transform 'Op' based TopLevelOpcode to enable +/// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called +/// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms +/// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and +/// RHS to 4. +static Instruction::BinaryOps +getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode, + BinaryOperator *Op, Value *&LHS, Value *&RHS) { + if (!Op) + return Instruction::BinaryOpsEnd; + + LHS = Op->getOperand(0); + RHS = Op->getOperand(1); + + switch (TopLevelOpcode) { + default: + return Op->getOpcode(); + + case Instruction::Add: + case Instruction::Sub: + if (Op->getOpcode() == Instruction::Shl) { + if (Constant *CST = dyn_cast(Op->getOperand(1))) { + // The multiplier is really 1 << CST. + RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST); + return Instruction::Mul; + } + } + return Op->getOpcode(); + } + + // TODO: We can add other conversions e.g. shr => div etc. +} + +/// This tries to simplify binary operations by factorizing out common terms +/// (e. g. "(A*B)+(A*C)" -> "A*(B+C)"). +static Value *tryFactorization(InstCombiner::BuilderTy *Builder, + const DataLayout *DL, BinaryOperator &I, + Instruction::BinaryOps InnerOpcode, Value *A, + Value *B, Value *C, Value *D) { + + // If any of A, B, C, D are null, we can not factor I, return early. + // Checking A and C should be enough. + if (!A || !C || !B || !D) + return nullptr; + + Value *SimplifiedInst = nullptr; + Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); + Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); + + // Does "X op' Y" always equal "Y op' X"? + bool InnerCommutative = Instruction::isCommutative(InnerOpcode); + + // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"? + if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode)) + // Does the instruction have the form "(A op' B) op (A op' D)" or, in the + // commutative case, "(A op' B) op (C op' A)"? + if (A == C || (InnerCommutative && A == D)) { + if (A != C) + 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, 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 && LHS->hasOneUse() && RHS->hasOneUse()) + V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName()); + if (V) { + SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V); + } + } + + // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"? + if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode)) + // Does the instruction have the form "(A op' B) op (C op' B)" or, in the + // commutative case, "(A op' B) op (B op' D)"? + if (B == D || (InnerCommutative && B == C)) { + if (B != D) + 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, 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 && LHS->hasOneUse() && RHS->hasOneUse()) + V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName()); + if (V) { + SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B); + } + } + + if (SimplifiedInst) { + ++NumFactor; + SimplifiedInst->takeName(&I); + + // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag. + // TODO: Check for NUW. + if (BinaryOperator *BO = dyn_cast(SimplifiedInst)) { + if (isa(SimplifiedInst)) { + bool HasNSW = false; + if (isa(&I)) + HasNSW = I.hasNoSignedWrap(); + + if (BinaryOperator *Op0 = dyn_cast(LHS)) + if (isa(Op0)) + HasNSW &= Op0->hasNoSignedWrap(); + + if (BinaryOperator *Op1 = dyn_cast(RHS)) + if (isa(Op1)) + HasNSW &= Op1->hasNoSignedWrap(); + BO->setHasNoSignedWrap(HasNSW); + } + } + } + return SimplifiedInst; +} + /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations /// which some other binary operation distributes over either by factorizing /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this @@ -376,63 +561,31 @@ Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); BinaryOperator *Op0 = dyn_cast(LHS); BinaryOperator *Op1 = dyn_cast(RHS); - Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); // op // Factorization. - if (Op0 && Op1 && Op0->getOpcode() == Op1->getOpcode()) { - // The instruction has the form "(A op' B) op (C op' D)". Try to factorize - // a common term. - Value *A = Op0->getOperand(0), *B = Op0->getOperand(1); - Value *C = Op1->getOperand(0), *D = Op1->getOperand(1); - Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op' + Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; + auto TopLevelOpcode = I.getOpcode(); + auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B); + auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D); + + // The instruction has the form "(A op' B) op (C op' D)". Try to factorize + // a common term. + if (LHSOpcode == RHSOpcode) { + if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D)) + return V; + } - // Does "X op' Y" always equal "Y op' X"? - bool InnerCommutative = Instruction::isCommutative(InnerOpcode); - - // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"? - if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode)) - // Does the instruction have the form "(A op' B) op (A op' D)" or, in the - // commutative case, "(A op' B) op (C op' A)"? - if (A == C || (InnerCommutative && A == D)) { - if (A != C) - 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); - // 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()) - V = Builder->CreateBinOp(TopLevelOpcode, B, D, Op1->getName()); - if (V) { - ++NumFactor; - V = Builder->CreateBinOp(InnerOpcode, A, V); - V->takeName(&I); - return V; - } - } + // The instruction has the form "(A op' B) op (C)". Try to factorize common + // term. + if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS, + getIdentityValue(LHSOpcode, RHS))) + return V; - // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"? - if (RightDistributesOverLeft(TopLevelOpcode, InnerOpcode)) - // Does the instruction have the form "(A op' B) op (C op' B)" or, in the - // commutative case, "(A op' B) op (B op' D)"? - if (B == D || (InnerCommutative && B == C)) { - if (B != D) - 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); - // 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()) - V = Builder->CreateBinOp(TopLevelOpcode, A, C, Op0->getName()); - if (V) { - ++NumFactor; - V = Builder->CreateBinOp(InnerOpcode, V, B); - V->takeName(&I); - return V; - } - } - } + // The instruction has the form "(B) op (C op' D)". Try to factorize common + // term. + if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS, + getIdentityValue(RHSOpcode, LHS), C, D)) + return V; // Expansion. if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) { @@ -442,8 +595,8 @@ 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. @@ -451,7 +604,7 @@ 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); @@ -467,8 +620,8 @@ 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. @@ -476,7 +629,7 @@ 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); @@ -485,7 +638,7 @@ Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) { } } - return 0; + return nullptr; } // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction @@ -503,15 +656,15 @@ Value *InstCombiner::dyn_castNegVal(Value *V) const { if (C->getType()->getElementType()->isIntegerTy()) return ConstantExpr::getNeg(C); - return 0; + return nullptr; } // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the // 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. @@ -522,7 +675,7 @@ Value *InstCombiner::dyn_castFNegVal(Value *V) const { if (C->getType()->getElementType()->isFloatingPointTy()) return ConstantExpr::getFNeg(C); - return 0; + return nullptr; } static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, @@ -545,9 +698,14 @@ static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, 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"); @@ -563,13 +721,13 @@ static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO, // not have a second operand. Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { // Don't modify shared select instructions - if (!SI->hasOneUse()) return 0; + if (!SI->hasOneUse()) return nullptr; Value *TV = SI->getOperand(1); Value *FV = SI->getOperand(2); if (isa(TV) || isa(FV)) { // Bool selects with constant operands can be folded to logical ops. - if (SI->getType()->isIntegerTy(1)) return 0; + if (SI->getType()->isIntegerTy(1)) return nullptr; // If it's a bitcast involving vectors, make sure it has the same number of // elements on both sides. @@ -578,10 +736,10 @@ Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { VectorType *SrcTy = dyn_cast(BC->getSrcTy()); // Verify that either both or neither are vectors. - if ((SrcTy == NULL) != (DestTy == NULL)) return 0; + if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr; // If vectors, verify that they have the same number of elements. if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements()) - return 0; + return nullptr; } Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); @@ -590,7 +748,7 @@ Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { return SelectInst::Create(SI->getCondition(), SelectTrueVal, SelectFalseVal); } - return 0; + return nullptr; } @@ -602,18 +760,17 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { PHINode *PN = cast(I.getOperand(0)); unsigned NumPHIValues = PN->getNumIncomingValues(); if (NumPHIValues == 0) - return 0; + return nullptr; // 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)) - return 0; + for (User *U : PN->users()) { + Instruction *UI = cast(U); + if (UI != &I && !I.isIdenticalTo(UI)) + return nullptr; } // Otherwise, we can replace *all* users with the new PHI we form. } @@ -623,14 +780,14 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { // remember the BB it is in. If there is more than one or if *it* is a PHI, // bail out. We don't do arbitrary constant expressions here because moving // their computation can be expensive without a cost model. - BasicBlock *NonConstBB = 0; + BasicBlock *NonConstBB = nullptr; for (unsigned i = 0; i != NumPHIValues; ++i) { Value *InVal = PN->getIncomingValue(i); if (isa(InVal) && !isa(InVal)) continue; - if (isa(InVal)) return 0; // Itself a phi. - if (NonConstBB) return 0; // More than one non-const value. + if (isa(InVal)) return nullptr; // Itself a phi. + if (NonConstBB) return nullptr; // More than one non-const value. NonConstBB = PN->getIncomingBlock(i); @@ -638,22 +795,22 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { // insert a computation after it without breaking the edge. if (InvokeInst *II = dyn_cast(InVal)) if (II->getParent() == NonConstBB) - return 0; + return nullptr; // 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; + return nullptr; } // 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 // do this if the pred block is unconditionally branching into the phi block. - if (NonConstBB != 0) { + if (NonConstBB != nullptr) { BranchInst *BI = dyn_cast(NonConstBB->getTerminator()); - if (!BI || !BI->isUnconditional()) return 0; + if (!BI || !BI->isUnconditional()) return nullptr; } // Okay, we can do the transformation: create the new PHI node. @@ -677,8 +834,11 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { BasicBlock *ThisBB = PN->getIncomingBlock(i); Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB); Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB); - Value *InV = 0; - if (Constant *InC = dyn_cast(PN->getIncomingValue(i))) + Value *InV = nullptr; + // 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), @@ -688,7 +848,7 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { } else if (CmpInst *CI = dyn_cast(&I)) { Constant *C = cast(I.getOperand(1)); for (unsigned i = 0; i != NumPHIValues; ++i) { - Value *InV = 0; + Value *InV = nullptr; if (Constant *InC = dyn_cast(PN->getIncomingValue(i))) InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C); else if (isa(CI)) @@ -702,7 +862,7 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) { } else if (I.getNumOperands() == 2) { Constant *C = cast(I.getOperand(1)); for (unsigned i = 0; i != NumPHIValues; ++i) { - Value *InV = 0; + Value *InV = nullptr; if (Constant *InC = dyn_cast(PN->getIncomingValue(i))) InV = ConstantExpr::get(I.getOpcode(), InC, C); else @@ -724,8 +884,7 @@ Instruction *InstCombiner::FoldOpIntoPhi(Instruction &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); @@ -734,21 +893,27 @@ 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 nullptr; + + Type *Ty = PtrTy->getPointerElementType(); + if (!Ty->isSized()) + return nullptr; // 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; @@ -766,11 +931,11 @@ Type *InstCombiner::FindElementAtOffset(Type *Ty, int64_t Offset, // 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)) - return 0; + if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty)) + return nullptr; 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"); @@ -781,14 +946,14 @@ Type *InstCombiner::FindElementAtOffset(Type *Ty, int64_t Offset, 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; Ty = AT->getElementType(); } else { // Otherwise, we can't index into the middle of this atomic type, bail. - return 0; + return nullptr; } } @@ -805,19 +970,365 @@ 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 nullptr; + + // 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 nullptr; + // 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 nullptr; + + // 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 nullptr; + + 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 nullptr; + + 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 nullptr; + + 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 nullptr; + + // 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 nullptr; + + 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 nullptr; + 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 nullptr; + + // 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 nullptr; + } + + // If Op is zero then Val = Op * Scale. + if (match(Op, m_Zero())) { + NoSignedWrap = true; + return Op; + } + + // 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); +} + +/// \brief Creates node of binary operation with the same attributes as the +/// specified one but with other operands. +static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS, + InstCombiner::BuilderTy *B) { + Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS); + if (BinaryOperator *NewBO = dyn_cast(BORes)) { + if (isa(NewBO)) { + NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap()); + NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap()); + } + if (isa(NewBO)) + NewBO->setIsExact(Inst.isExact()); + } + return BORes; +} + +/// \brief Makes transformation of binary operation specific for vector types. +/// \param Inst Binary operator to transform. +/// \return Pointer to node that must replace the original binary operator, or +/// null pointer if no transformation was made. +Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) { + if (!Inst.getType()->isVectorTy()) return nullptr; + + // It may not be safe to reorder shuffles and things like div, urem, etc. + // because we may trap when executing those ops on unknown vector elements. + // See PR20059. + if (!isSafeToSpeculativelyExecute(&Inst, DL)) return nullptr; + + unsigned VWidth = cast(Inst.getType())->getNumElements(); + Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1); + assert(cast(LHS->getType())->getNumElements() == VWidth); + assert(cast(RHS->getType())->getNumElements() == VWidth); + + // If both arguments of binary operation are shuffles, which use the same + // mask and shuffle within a single vector, it is worthwhile to move the + // shuffle after binary operation: + // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m) + if (isa(LHS) && isa(RHS)) { + ShuffleVectorInst *LShuf = cast(LHS); + ShuffleVectorInst *RShuf = cast(RHS); + if (isa(LShuf->getOperand(1)) && + isa(RShuf->getOperand(1)) && + LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() && + LShuf->getMask() == RShuf->getMask()) { + Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0), + RShuf->getOperand(0), Builder); + Value *Res = Builder->CreateShuffleVector(NewBO, + UndefValue::get(NewBO->getType()), LShuf->getMask()); + return Res; + } + } + + // If one argument is a shuffle within one vector, the other is a constant, + // try moving the shuffle after the binary operation. + ShuffleVectorInst *Shuffle = nullptr; + Constant *C1 = nullptr; + if (isa(LHS)) Shuffle = cast(LHS); + if (isa(RHS)) Shuffle = cast(RHS); + if (isa(LHS)) C1 = cast(LHS); + if (isa(RHS)) C1 = cast(RHS); + if (Shuffle && C1 && + (isa(C1) || isa(C1)) && + isa(Shuffle->getOperand(1)) && + Shuffle->getType() == Shuffle->getOperand(0)->getType()) { + SmallVector ShMask = Shuffle->getShuffleMask(); + // Find constant C2 that has property: + // shuffle(C2, ShMask) = C1 + // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>) + // reorder is not possible. + SmallVector C2M(VWidth, + UndefValue::get(C1->getType()->getScalarType())); + bool MayChange = true; + for (unsigned I = 0; I < VWidth; ++I) { + if (ShMask[I] >= 0) { + assert(ShMask[I] < (int)VWidth); + if (!isa(C2M[ShMask[I]])) { + MayChange = false; + break; + } + C2M[ShMask[I]] = C1->getAggregateElement(I); + } + } + if (MayChange) { + Constant *C2 = ConstantVector::get(C2M); + Value *NewLHS, *NewRHS; + if (isa(LHS)) { + NewLHS = C2; + NewRHS = Shuffle->getOperand(0); + } else { + NewLHS = Shuffle->getOperand(0); + NewRHS = C2; + } + Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder); + Value *Res = Builder->CreateShuffleVector(NewBO, + UndefValue::get(Inst.getType()), Shuffle->getMask()); + return Res; + } + } + + return nullptr; +} + 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, TLI, DT, AT)) 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 +1340,14 @@ 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. @@ -847,21 +1358,106 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { if (MadeChange) return &GEP; } + // Check to see if the inputs to the PHI node are getelementptr instructions. + if (PHINode *PN = dyn_cast(PtrOp)) { + GetElementPtrInst *Op1 = dyn_cast(PN->getOperand(0)); + if (!Op1) + return nullptr; + + signed DI = -1; + + for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) { + GetElementPtrInst *Op2 = dyn_cast(*I); + if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands()) + return nullptr; + + // Keep track of the type as we walk the GEP. + Type *CurTy = Op1->getOperand(0)->getType()->getScalarType(); + + for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) { + if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType()) + return nullptr; + + if (Op1->getOperand(J) != Op2->getOperand(J)) { + if (DI == -1) { + // We have not seen any differences yet in the GEPs feeding the + // PHI yet, so we record this one if it is allowed to be a + // variable. + + // The first two arguments can vary for any GEP, the rest have to be + // static for struct slots + if (J > 1 && CurTy->isStructTy()) + return nullptr; + + DI = J; + } else { + // The GEP is different by more than one input. While this could be + // extended to support GEPs that vary by more than one variable it + // doesn't make sense since it greatly increases the complexity and + // would result in an R+R+R addressing mode which no backend + // directly supports and would need to be broken into several + // simpler instructions anyway. + return nullptr; + } + } + + // Sink down a layer of the type for the next iteration. + if (J > 0) { + if (CompositeType *CT = dyn_cast(CurTy)) { + CurTy = CT->getTypeAtIndex(Op1->getOperand(J)); + } else { + CurTy = nullptr; + } + } + } + } + + GetElementPtrInst *NewGEP = cast(Op1->clone()); + + if (DI == -1) { + // All the GEPs feeding the PHI are identical. Clone one down into our + // BB so that it can be merged with the current GEP. + GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(), + NewGEP); + } else { + // All the GEPs feeding the PHI differ at a single offset. Clone a GEP + // into the current block so it can be merged, and create a new PHI to + // set that index. + Instruction *InsertPt = Builder->GetInsertPoint(); + Builder->SetInsertPoint(PN); + PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(), + PN->getNumOperands()); + Builder->SetInsertPoint(InsertPt); + + for (auto &I : PN->operands()) + NewPN->addIncoming(cast(I)->getOperand(DI), + PN->getIncomingBlock(I)); + + NewGEP->setOperand(DI, NewPN); + GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(), + NewGEP); + NewGEP->setOperand(DI, NewPN); + } + + GEP.setOperand(0, NewGEP); + PtrOp = NewGEP; + } + // Combine Indices - If the source pointer to this getelementptr instruction // is a getelementptr instruction, combine the indices of the two // getelementptr instructions into a single instruction. // if (GEPOperator *Src = dyn_cast(PtrOp)) { if (!shouldMergeGEPs(*cast(&GEP), *Src)) - return 0; + return nullptr; - // 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 = dyn_cast(Src->getOperand(0))) if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP)) - return 0; // Wait until our source is folded to completion. + return nullptr; // Wait until our source is folded to completion. SmallVector Indices; @@ -889,7 +1485,7 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { // intptr_t). Just avoid transforming this until the input has been // normalized. if (SO1->getType() != GO1->getType()) - return 0; + return nullptr; Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum"); } @@ -917,17 +1513,62 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName()); } + if (DL && GEP.getNumIndices() == 1) { + unsigned AS = GEP.getPointerAddressSpace(); + if (GEP.getOperand(1)->getType()->getScalarSizeInBits() == + DL->getPointerSizeInBits(AS)) { + Type *PtrTy = GEP.getPointerOperandType(); + Type *Ty = PtrTy->getPointerElementType(); + uint64_t TyAllocSize = DL->getTypeAllocSize(Ty); + + bool Matched = false; + uint64_t C; + Value *V = nullptr; + if (TyAllocSize == 1) { + V = GEP.getOperand(1); + Matched = true; + } else if (match(GEP.getOperand(1), + m_AShr(m_Value(V), m_ConstantInt(C)))) { + if (TyAllocSize == 1ULL << C) + Matched = true; + } else if (match(GEP.getOperand(1), + m_SDiv(m_Value(V), m_ConstantInt(C)))) { + if (TyAllocSize == C) + Matched = true; + } + + if (Matched) { + // Canonicalize (gep i8* X, -(ptrtoint Y)) + // to (inttoptr (sub (ptrtoint X), (ptrtoint Y))) + // The GEP pattern is emitted by the SCEV expander for certain kinds of + // pointer arithmetic. + if (match(V, m_Neg(m_PtrToInt(m_Value())))) { + Operator *Index = cast(V); + Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType()); + Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1)); + return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType()); + } + // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X)) + // to (bitcast Y) + Value *Y; + if (match(V, m_Sub(m_PtrToInt(m_Value(Y)), + m_PtrToInt(m_Specific(GEP.getOperand(0)))))) { + return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, + 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 (!StrippedPtrTy) - return 0; - - if (StrippedPtr != PtrOp && - StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { + return nullptr; + if (StrippedPtr != PtrOp) { bool HasZeroPointerIndex = false; if (ConstantInt *C = dyn_cast(GEP.getOperand(1))) HasZeroPointerIndex = C->isZero(); @@ -950,7 +1591,15 @@ 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 = @@ -962,8 +1611,24 @@ 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()); } } } @@ -972,103 +1637,133 @@ 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()); + return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, + GEP.getType()); } // 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); + // %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 + return CastInst::CreatePointerBitCastOrAddrSpaceCast(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"); + // 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 (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 + return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, + GEP.getType()); } - - // 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 nullptr; + + // addrspacecast between types is canonicalized as a bitcast, then an + // addrspacecast. To take advantage of the below bitcast + struct GEP, look + // through the addrspacecast. + if (AddrSpaceCastInst *ASC = dyn_cast(PtrOp)) { + // X = bitcast A addrspace(1)* to B addrspace(1)* + // Y = addrspacecast A addrspace(1)* to B addrspace(2)* + // Z = gep Y, <...constant indices...> + // Into an addrspacecasted GEP of the struct. + if (BitCastInst *BC = dyn_cast(ASC->getOperand(0))) + PtrOp = BC; + } + /// 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() && - StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) { - - // Determine how much the GEP moves the pointer. - SmallVector Ops(GEP.idx_begin(), GEP.idx_end()); - int64_t Offset = TD->getIndexedOffset(GEP.getPointerOperandType(), Ops); + Value *Operand = BCI->getOperand(0); + PointerType *OpType = cast(Operand->getType()); + unsigned OffsetBits = DL->getPointerTypeSizeInBits(GEP.getType()); + APInt Offset(OffsetBits, 0); + if (!isa(Operand) && + GEP.accumulateConstantOffset(*DL, Offset)) { // 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)) || - isAllocationFn(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) { @@ -1079,43 +1774,45 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { return &GEP; } } - return new BitCastInst(BCI->getOperand(0), GEP.getType()); + + if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace()) + return new AddrSpaceCastInst(Operand, 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); + + if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace()) + return new AddrSpaceCastInst(NGEP, GEP.getType()); return new BitCastInst(NGEP, GEP.getType()); } } } - return 0; + return nullptr; } - - static bool -isAllocSiteRemovable(Instruction *AI, SmallVectorImpl &Users) { +isAllocSiteRemovable(Instruction *AI, SmallVectorImpl &Users, + const TargetLibraryInfo *TLI) { SmallVector Worklist; Worklist.push_back(AI); do { Instruction *PI = Worklist.pop_back_val(); - for (Value::use_iterator UI = PI->use_begin(), UE = PI->use_end(); UI != UE; - ++UI) { - Instruction *I = cast(*UI); + for (User *U : PI->users()) { + Instruction *I = cast(U); switch (I->getOpcode()) { default: // Give up the moment we see something we can't handle. @@ -1163,7 +1860,7 @@ isAllocSiteRemovable(Instruction *AI, SmallVectorImpl &Users) { } } - if (isFreeCall(I)) { + if (isFreeCall(I, TLI)) { Users.push_back(I); continue; } @@ -1188,7 +1885,7 @@ Instruction *InstCombiner::visitAllocSite(Instruction &MI) { // to null and free calls, delete the calls and replace the comparisons with // true or false as appropriate. SmallVector Users; - if (isAllocSiteRemovable(&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; @@ -1214,13 +1911,69 @@ Instruction *InstCombiner::visitAllocSite(Instruction &MI) { Module *M = II->getParent()->getParent()->getParent(); Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing); InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(), - ArrayRef(), "", II->getParent()); + None, "", II->getParent()); } return EraseInstFromFunction(MI); } - return 0; + return nullptr; } +/// \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 nullptr; + + // 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 nullptr; + BasicBlock *SuccBB; + if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB))) + return nullptr; + + // 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 nullptr; + if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE) + return nullptr; + + // Validate constraint #3: Ensure the null case just falls through. + if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB)) + return nullptr; + 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) { @@ -1239,14 +1992,42 @@ Instruction *InstCombiner::visitFree(CallInst &FI) { if (isa(Op)) return EraseInstFromFunction(FI); - return 0; + // 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 nullptr; } +Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) { + if (RI.getNumOperands() == 0) // ret void + return nullptr; + + Value *ResultOp = RI.getOperand(0); + Type *VTy = ResultOp->getType(); + if (!VTy->isIntegerTy()) + return nullptr; + // There might be assume intrinsics dominating this return that completely + // determine the value. If so, constant fold it. + unsigned BitWidth = VTy->getPrimitiveSizeInBits(); + APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); + computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI); + if ((KnownZero|KnownOne).isAllOnesValue()) + RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne)); + + return nullptr; +} Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { // Change br (not X), label True, label False to: br X, label False, True - Value *X = 0; + Value *X = nullptr; BasicBlock *TrueDest; BasicBlock *FalseDest; if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) && @@ -1257,7 +2038,7 @@ 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)), TrueDest, FalseDest)) && @@ -1273,7 +2054,7 @@ Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { 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)) && @@ -1289,11 +2070,42 @@ Instruction *InstCombiner::visitBranchInst(BranchInst &BI) { return &BI; } - return 0; + return nullptr; } Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { Value *Cond = SI.getCondition(); + unsigned BitWidth = cast(Cond->getType())->getBitWidth(); + APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); + computeKnownBits(Cond, KnownZero, KnownOne); + unsigned LeadingKnownZeros = KnownZero.countLeadingOnes(); + unsigned LeadingKnownOnes = KnownOne.countLeadingOnes(); + + // Compute the number of leading bits we can ignore. + for (auto &C : SI.cases()) { + LeadingKnownZeros = std::min( + LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros()); + LeadingKnownOnes = std::min( + LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes()); + } + + unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes); + + // Truncate the condition operand if the new type is equal to or larger than + // the largest legal integer type. We need to be conservative here since + // x86 generates redundant zero-extenstion instructions if the operand is + // truncated to i8 or i16. + if (BitWidth > NewWidth && NewWidth >= DL->getLargestLegalIntTypeSize()) { + IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth); + Builder->SetInsertPoint(&SI); + Value *NewCond = Builder->CreateTrunc(SI.getCondition(), Ty, "trunc"); + SI.setCondition(NewCond); + + for (auto &C : SI.cases()) + static_cast(&C)->setValue(ConstantInt::get( + SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth))); + } + if (Instruction *I = dyn_cast(Cond)) { if (I->getOpcode() == Instruction::Add) if (ConstantInt *AddRHS = dyn_cast(I->getOperand(1))) { @@ -1313,7 +2125,7 @@ Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) { return &SI; } } - return 0; + return nullptr; } Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { @@ -1330,7 +2142,7 @@ Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { // first index return ExtractValueInst::Create(C2, EV.getIndices().slice(1)); } - return 0; // Can't handle other constants + return nullptr; // Can't handle other constants } if (InsertValueInst *IV = dyn_cast(Agg)) { @@ -1463,7 +2275,7 @@ Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) { // and if again single-use then via load (gep (gep)) to load (gep). // However, double extracts from e.g. function arguments or return values // aren't handled yet. - return 0; + return nullptr; } enum Personality_Type { @@ -1519,7 +2331,7 @@ Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { // Simplify the list of clauses, eg by removing repeated catch clauses // (these are often created by inlining). bool MakeNewInstruction = false; // If true, recreate using the following: - SmallVector NewClauses; // - Clauses for the new instruction; + SmallVector NewClauses; // - Clauses for the new instruction; bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup. SmallPtrSet AlreadyCaught; // Typeinfos known caught already. @@ -1527,8 +2339,8 @@ Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { bool isLastClause = i + 1 == e; if (LI.isCatch(i)) { // A catch clause. - Value *CatchClause = LI.getClause(i); - Constant *TypeInfo = cast(CatchClause->stripPointerCasts()); + Constant *CatchClause = LI.getClause(i); + Constant *TypeInfo = CatchClause->stripPointerCasts(); // If we already saw this clause, there is no point in having a second // copy of it. @@ -1557,7 +2369,7 @@ Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { // equal (for example if one represents a C++ class, and the other some // class derived from it). assert(LI.isFilter(i) && "Unsupported landingpad clause!"); - Value *FilterClause = LI.getClause(i); + Constant *FilterClause = LI.getClause(i); ArrayType *FilterType = cast(FilterClause->getType()); unsigned NumTypeInfos = FilterType->getNumElements(); @@ -1601,8 +2413,8 @@ Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { // catch-alls. If so, the filter can be discarded. bool SawCatchAll = false; for (unsigned j = 0; j != NumTypeInfos; ++j) { - Value *Elt = Filter->getOperand(j); - Constant *TypeInfo = cast(Elt->stripPointerCasts()); + Constant *Elt = Filter->getOperand(j); + Constant *TypeInfo = Elt->stripPointerCasts(); if (isCatchAll(Personality, TypeInfo)) { // This element is a catch-all. Bail out, noting this fact. SawCatchAll = true; @@ -1707,7 +2519,7 @@ 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. @@ -1802,7 +2614,7 @@ Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) { return &LI; } - return 0; + return nullptr; } @@ -1851,9 +2663,9 @@ static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { /// whose condition is a known constant, we only visit the reachable successors. /// static bool AddReachableCodeToWorklist(BasicBlock *BB, - SmallPtrSet &Visited, + SmallPtrSetImpl &Visited, InstCombiner &IC, - const TargetData *TD, + const DataLayout *DL, const TargetLibraryInfo *TLI) { bool MadeIRChange = false; SmallVector Worklist; @@ -1872,17 +2684,17 @@ static bool AddReachableCodeToWorklist(BasicBlock *BB, 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; @@ -1890,16 +2702,16 @@ static bool AddReachableCodeToWorklist(BasicBlock *BB, 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) { ConstantExpr *CE = dyn_cast(i); - if (CE == 0) continue; + if (CE == nullptr) continue; Constant*& FoldRes = FoldedConstants[CE]; if (!FoldRes) - FoldRes = ConstantFoldConstantExpression(CE, TD, TLI); + FoldRes = ConstantFoldConstantExpression(CE, DL, TLI); if (!FoldRes) FoldRes = CE; @@ -1958,7 +2770,7 @@ static bool AddReachableCodeToWorklist(BasicBlock *BB, 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"); { @@ -1966,7 +2778,7 @@ 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 @@ -1999,11 +2811,11 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { while (!Worklist.isEmpty()) { Instruction *I = Worklist.RemoveOne(); - if (I == 0) continue; // skip null values. + if (I == nullptr) 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; @@ -2012,8 +2824,8 @@ 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); @@ -2026,12 +2838,12 @@ 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(); @@ -2047,9 +2859,18 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { // If the user is one of our immediate successors, and if that successor // only has us as a predecessors (we'd have to split the critical edge // otherwise), we can keep going. - if (UserIsSuccessor && UserParent->getSinglePredecessor()) + if (UserIsSuccessor && UserParent->getSinglePredecessor()) { // Okay, the CFG is simple enough, try to sink this instruction. - MadeIRChange |= TryToSinkInstruction(I, UserParent); + if (TryToSinkInstruction(I, UserParent)) { + MadeIRChange = true; + // We'll add uses of the sunk instruction below, but since sinking + // can expose opportunities for it's *operands* add them to the + // worklist + for (Use &U : I->operands()) + if (Instruction *OpI = dyn_cast(U.get())) + Worklist.Add(OpI); + } + } } } @@ -2061,13 +2882,13 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { 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()) @@ -2096,13 +2917,13 @@ 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); @@ -2117,18 +2938,52 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { return MadeIRChange; } +namespace { +class InstCombinerLibCallSimplifier final : public LibCallSimplifier { + InstCombiner *IC; +public: + InstCombinerLibCallSimplifier(const DataLayout *DL, + const TargetLibraryInfo *TLI, + InstCombiner *IC) + : LibCallSimplifier(DL, TLI, EnableUnsafeFPShrink) { + 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; + + AT = &getAnalysis(); + DataLayoutPass *DLP = getAnalysisIfAvailable(); + DL = DLP ? &DLP->getDataLayout() : nullptr; TLI = &getAnalysis(); + DominatorTreeWrapperPass *DTWP = + getAnalysisIfAvailable(); + DT = DTWP ? &DTWP->getDomTree() : nullptr; + + // 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), - InstCombineIRInserter(Worklist)); + TheBuilder(F.getContext(), TargetFolder(DL), + InstCombineIRInserter(Worklist, AT)); Builder = &TheBuilder; + InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this); + Simplifier = &TheSimplifier; + bool EverMadeChange = false; // Lower dbg.declare intrinsics otherwise their value may be clobbered @@ -2140,7 +2995,7 @@ bool InstCombiner::runOnFunction(Function &F) { while (DoOneIteration(F, Iteration++)) EverMadeChange = true; - Builder = 0; + Builder = nullptr; return EverMadeChange; }