X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FBasicAliasAnalysis.cpp;h=22b73b28400c6f2f9e4f456d68575e3f512c7709;hb=4ed4329c37f5a64c16ad4dc1960cbcb66b7118d4;hp=71bcbd2cf09a02d40b1542f521d114f6da6a15b1;hpb=6fb3bd6a658940287789198d3207b0da04c0a4e6;p=oota-llvm.git diff --git a/lib/Analysis/BasicAliasAnalysis.cpp b/lib/Analysis/BasicAliasAnalysis.cpp index 71bcbd2cf09..22b73b28400 100644 --- a/lib/Analysis/BasicAliasAnalysis.cpp +++ b/lib/Analysis/BasicAliasAnalysis.cpp @@ -2,8 +2,8 @@ // // The LLVM Compiler Infrastructure // -// This file was developed by the LLVM research group and is distributed under -// the University of Illinois Open Source License. See LICENSE.TXT for details. +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // @@ -20,14 +20,160 @@ #include "llvm/Function.h" #include "llvm/GlobalVariable.h" #include "llvm/Instructions.h" +#include "llvm/IntrinsicInst.h" #include "llvm/Pass.h" #include "llvm/Target/TargetData.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/Support/Compiler.h" #include "llvm/Support/GetElementPtrTypeIterator.h" +#include "llvm/Support/ManagedStatic.h" #include using namespace llvm; -// Make sure that anything that uses AliasAnalysis pulls in this file... -void llvm::BasicAAStub() {} +//===----------------------------------------------------------------------===// +// Useful predicates +//===----------------------------------------------------------------------===// + +// Determine if an AllocationInst instruction escapes from the function it is +// contained in. If it does not escape, there is no way for another function to +// mod/ref it. We do this by looking at its uses and determining if the uses +// can escape (recursively). +static bool AddressMightEscape(const Value *V) { + for (Value::use_const_iterator UI = V->use_begin(), E = V->use_end(); + UI != E; ++UI) { + const Instruction *I = cast(*UI); + switch (I->getOpcode()) { + case Instruction::Load: + break; //next use. + case Instruction::Store: + if (I->getOperand(0) == V) + return true; // Escapes if the pointer is stored. + break; // next use. + case Instruction::GetElementPtr: + if (AddressMightEscape(I)) + return true; + break; // next use. + case Instruction::BitCast: + if (AddressMightEscape(I)) + return true; + break; // next use + case Instruction::Ret: + // If returned, the address will escape to calling functions, but no + // callees could modify it. + break; // next use + case Instruction::Call: + // If the call is to a few known safe intrinsics, we know that it does + // not escape. + // TODO: Eventually just check the 'nocapture' attribute. + if (!isa(I)) + return true; + break; // next use + default: + return true; + } + } + return false; +} + +static const User *isGEP(const Value *V) { + if (isa(V) || + (isa(V) && + cast(V)->getOpcode() == Instruction::GetElementPtr)) + return cast(V); + return 0; +} + +static const Value *GetGEPOperands(const Value *V, + SmallVector &GEPOps){ + assert(GEPOps.empty() && "Expect empty list to populate!"); + GEPOps.insert(GEPOps.end(), cast(V)->op_begin()+1, + cast(V)->op_end()); + + // Accumulate all of the chained indexes into the operand array + V = cast(V)->getOperand(0); + + while (const User *G = isGEP(V)) { + if (!isa(GEPOps[0]) || isa(GEPOps[0]) || + !cast(GEPOps[0])->isNullValue()) + break; // Don't handle folding arbitrary pointer offsets yet... + GEPOps.erase(GEPOps.begin()); // Drop the zero index + GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end()); + V = G->getOperand(0); + } + return V; +} + +/// isIdentifiedObject - Return true if this pointer refers to a distinct and +/// identifiable object. This returns true for: +/// Global Variables and Functions +/// Allocas and Mallocs +/// ByVal and NoAlias Arguments +/// +static bool isIdentifiedObject(const Value *V) { + if (isa(V) || isa(V)) + return true; + if (const Argument *A = dyn_cast(V)) + return A->hasNoAliasAttr() || A->hasByValAttr(); + return false; +} + +/// isKnownNonNull - Return true if we know that the specified value is never +/// null. +static bool isKnownNonNull(const Value *V) { + // Alloca never returns null, malloc might. + if (isa(V)) return true; + + // A byval argument is never null. + if (const Argument *A = dyn_cast(V)) + return A->hasByValAttr(); + + // Global values are not null unless extern weak. + if (const GlobalValue *GV = dyn_cast(V)) + return !GV->hasExternalWeakLinkage(); + return false; +} + +/// isNonEscapingLocalObject - Return true if the pointer is to a function-local +/// object that never escapes from the function. +static bool isNonEscapingLocalObject(const Value *V) { + // If this is a local allocation, check to see if it escapes. + if (isa(V)) + return !AddressMightEscape(V); + + // If this is an argument that corresponds to a byval or noalias argument, + // it can't escape either. + if (const Argument *A = dyn_cast(V)) + if (A->hasByValAttr() || A->hasNoAliasAttr()) + return !AddressMightEscape(V); + return false; +} + + +/// isObjectSmallerThan - Return true if we can prove that the object specified +/// by V is smaller than Size. +static bool isObjectSmallerThan(const Value *V, unsigned Size, + const TargetData &TD) { + const Type *AccessTy = 0; + if (const GlobalVariable *GV = dyn_cast(V)) + AccessTy = GV->getType()->getElementType(); + + if (const AllocationInst *AI = dyn_cast(V)) + if (!AI->isArrayAllocation()) + AccessTy = AI->getType()->getElementType(); + + if (const Argument *A = dyn_cast(V)) + if (A->hasByValAttr()) + AccessTy = cast(A->getType())->getElementType(); + + if (AccessTy && AccessTy->isSized()) + return TD.getABITypeSize(AccessTy) < Size; + return false; +} + +//===----------------------------------------------------------------------===// +// NoAA Pass +//===----------------------------------------------------------------------===// namespace { /// NoAA - This class implements the -no-aa pass, which always returns "I @@ -35,7 +181,11 @@ namespace { /// implementations, in that it does not chain to a previous analysis. As /// such it doesn't follow many of the rules that other alias analyses must. /// - struct NoAA : public ImmutablePass, public AliasAnalysis { + struct VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis { + static char ID; // Class identification, replacement for typeinfo + NoAA() : ImmutablePass(&ID) {} + explicit NoAA(void *PID) : ImmutablePass(PID) { } + virtual void getAnalysisUsage(AnalysisUsage &AU) const { AU.addRequired(); } @@ -72,22 +222,29 @@ namespace { virtual void deleteValue(Value *V) {} virtual void copyValue(Value *From, Value *To) {} }; +} // End of anonymous namespace - // Register this pass... - RegisterOpt - U("no-aa", "No Alias Analysis (always returns 'may' alias)"); +// Register this pass... +char NoAA::ID = 0; +static RegisterPass +U("no-aa", "No Alias Analysis (always returns 'may' alias)", true, true); - // Declare that we implement the AliasAnalysis interface - RegisterAnalysisGroup V; -} // End of anonymous namespace +// Declare that we implement the AliasAnalysis interface +static RegisterAnalysisGroup V(U); ImmutablePass *llvm::createNoAAPass() { return new NoAA(); } +//===----------------------------------------------------------------------===// +// BasicAA Pass +//===----------------------------------------------------------------------===// + namespace { /// BasicAliasAnalysis - This is the default alias analysis implementation. /// Because it doesn't chain to a previous alias analysis (like -no-aa), it /// derives from the NoAA class. - struct BasicAliasAnalysis : public NoAA { + struct VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA { + static char ID; // Class identification, replacement for typeinfo + BasicAliasAnalysis() : NoAA(&ID) {} AliasResult alias(const Value *V1, unsigned V1Size, const Value *V2, unsigned V2Size); @@ -104,123 +261,37 @@ namespace { /// global) or not. bool pointsToConstantMemory(const Value *P); - virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS, - std::vector *Info); - private: // CheckGEPInstructions - Check two GEP instructions with known // must-aliasing base pointers. This checks to see if the index expressions // preclude the pointers from aliasing... AliasResult - CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, - unsigned G1Size, - const Type *BasePtr2Ty, std::vector &GEP2Ops, - unsigned G2Size); + CheckGEPInstructions(const Type* BasePtr1Ty, + Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size, + const Type *BasePtr2Ty, + Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size); }; +} // End of anonymous namespace - // Register this pass... - RegisterOpt - X("basicaa", "Basic Alias Analysis (default AA impl)"); +// Register this pass... +char BasicAliasAnalysis::ID = 0; +static RegisterPass +X("basicaa", "Basic Alias Analysis (default AA impl)", false, true); - // Declare that we implement the AliasAnalysis interface - RegisterAnalysisGroup Y; -} // End of anonymous namespace +// Declare that we implement the AliasAnalysis interface +static RegisterAnalysisGroup Y(X); ImmutablePass *llvm::createBasicAliasAnalysisPass() { return new BasicAliasAnalysis(); } -// hasUniqueAddress - Return true if the specified value points to something -// with a unique, discernable, address. -static inline bool hasUniqueAddress(const Value *V) { - return isa(V) || isa(V); -} - -// getUnderlyingObject - This traverses the use chain to figure out what object -// the specified value points to. If the value points to, or is derived from, a -// unique object or an argument, return it. -static const Value *getUnderlyingObject(const Value *V) { - if (!isa(V->getType())) return 0; - - // If we are at some type of object... return it. - if (hasUniqueAddress(V) || isa(V)) return V; - - // Traverse through different addressing mechanisms... - if (const Instruction *I = dyn_cast(V)) { - if (isa(I) || isa(I)) - return getUnderlyingObject(I->getOperand(0)); - } else if (const ConstantExpr *CE = dyn_cast(V)) { - if (CE->getOpcode() == Instruction::Cast || - CE->getOpcode() == Instruction::GetElementPtr) - return getUnderlyingObject(CE->getOperand(0)); - } else if (const GlobalValue *GV = dyn_cast(V)) { - return GV; - } - return 0; -} - -static const User *isGEP(const Value *V) { - if (isa(V) || - (isa(V) && - cast(V)->getOpcode() == Instruction::GetElementPtr)) - return cast(V); - return 0; -} - -static const Value *GetGEPOperands(const Value *V, std::vector &GEPOps){ - assert(GEPOps.empty() && "Expect empty list to populate!"); - GEPOps.insert(GEPOps.end(), cast(V)->op_begin()+1, - cast(V)->op_end()); - - // Accumulate all of the chained indexes into the operand array - V = cast(V)->getOperand(0); - - while (const User *G = isGEP(V)) { - if (!isa(GEPOps[0]) || isa(GEPOps[0]) || - !cast(GEPOps[0])->isNullValue()) - break; // Don't handle folding arbitrary pointer offsets yet... - GEPOps.erase(GEPOps.begin()); // Drop the zero index - GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end()); - V = G->getOperand(0); - } - return V; -} /// pointsToConstantMemory - Chase pointers until we find a (constant /// global) or not. bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) { - if (const Value *V = getUnderlyingObject(P)) - if (const GlobalVariable *GV = dyn_cast(V)) - return GV->isConstant(); - return false; -} - -static bool AddressMightEscape(const Value *V) { - for (Value::use_const_iterator UI = V->use_begin(), E = V->use_end(); - UI != E; ++UI) { - const Instruction *I = cast(*UI); - switch (I->getOpcode()) { - case Instruction::Load: break; - case Instruction::Store: - if (I->getOperand(0) == V) - return true; // Escapes if the pointer is stored. - break; - case Instruction::GetElementPtr: - if (AddressMightEscape(I)) return true; - break; - case Instruction::Cast: - if (!isa(I->getType())) - return true; - if (AddressMightEscape(I)) return true; - break; - case Instruction::Ret: - // If returned, the address will escape to calling functions, but no - // callees could modify it. - break; - default: - return true; - } - } + if (const GlobalVariable *GV = + dyn_cast(P->getUnderlyingObject())) + return GV->isConstant(); return false; } @@ -231,26 +302,41 @@ static bool AddressMightEscape(const Value *V) { // AliasAnalysis::ModRefResult BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) { - if (!isa(P)) - if (const AllocationInst *AI = - dyn_cast_or_null(getUnderlyingObject(P))) { - // Okay, the pointer is to a stack allocated object. If we can prove that - // the pointer never "escapes", then we know the call cannot clobber it, - // because it simply can't get its address. - if (!AddressMightEscape(AI)) - return NoModRef; - - // If this is a tail call and P points to a stack location, we know that - // the tail call cannot access or modify the local stack. + if (!isa(P)) { + const Value *Object = P->getUnderlyingObject(); + + // If this is a tail call and P points to a stack location, we know that + // the tail call cannot access or modify the local stack. + // We cannot exclude byval arguments here; these belong to the caller of + // the current function not to the current function, and a tail callee + // may reference them. + if (isa(Object)) if (CallInst *CI = dyn_cast(CS.getInstruction())) - if (CI->isTailCall() && isa(AI)) + if (CI->isTailCall()) return NoModRef; + + // If the pointer is to a locally allocated object that does not escape, + // then the call can not mod/ref the pointer unless the call takes the + // argument without capturing it. + if (isNonEscapingLocalObject(Object)) { + bool passedAsArg = false; + // TODO: Eventually only check 'nocapture' arguments. + for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end(); + CI != CE; ++CI) + if (isa((*CI)->getType()) && + alias(cast(CI), ~0U, P, ~0U) != NoAlias) + passedAsArg = true; + + if (!passedAsArg) + return NoModRef; } + } // The AliasAnalysis base class has some smarts, lets use them. return AliasAnalysis::getModRefInfo(CS, P, Size); } + // alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such // as array references. Note that this function is heavily tail recursive. // Hopefully we have a smart C++ compiler. :) @@ -260,91 +346,62 @@ BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, const Value *V2, unsigned V2Size) { // Strip off any constant expression casts if they exist if (const ConstantExpr *CE = dyn_cast(V1)) - if (CE->getOpcode() == Instruction::Cast && - isa(CE->getOperand(0)->getType())) + if (CE->isCast() && isa(CE->getOperand(0)->getType())) V1 = CE->getOperand(0); if (const ConstantExpr *CE = dyn_cast(V2)) - if (CE->getOpcode() == Instruction::Cast && - isa(CE->getOperand(0)->getType())) + if (CE->isCast() && isa(CE->getOperand(0)->getType())) V2 = CE->getOperand(0); // Are we checking for alias of the same value? if (V1 == V2) return MustAlias; if ((!isa(V1->getType()) || !isa(V2->getType())) && - V1->getType() != Type::LongTy && V2->getType() != Type::LongTy) + V1->getType() != Type::Int64Ty && V2->getType() != Type::Int64Ty) return NoAlias; // Scalars cannot alias each other // Strip off cast instructions... - if (const Instruction *I = dyn_cast(V1)) - if (isa(I->getOperand(0)->getType())) - return alias(I->getOperand(0), V1Size, V2, V2Size); - if (const Instruction *I = dyn_cast(V2)) - if (isa(I->getOperand(0)->getType())) - return alias(V1, V1Size, I->getOperand(0), V2Size); + if (const BitCastInst *I = dyn_cast(V1)) + return alias(I->getOperand(0), V1Size, V2, V2Size); + if (const BitCastInst *I = dyn_cast(V2)) + return alias(V1, V1Size, I->getOperand(0), V2Size); // Figure out what objects these things are pointing to if we can... - const Value *O1 = getUnderlyingObject(V1); - const Value *O2 = getUnderlyingObject(V2); - - // Pointing at a discernible object? - if (O1) { - if (O2) { - if (isa(O1)) { - // Incoming argument cannot alias locally allocated object! - if (isa(O2)) return NoAlias; - // Otherwise, nothing is known... - } else if (isa(O2)) { - // Incoming argument cannot alias locally allocated object! - if (isa(O1)) return NoAlias; - // Otherwise, nothing is known... - } else if (O1 != O2) { - // If they are two different objects, we know that we have no alias... - return NoAlias; - } + const Value *O1 = V1->getUnderlyingObject(); + const Value *O2 = V2->getUnderlyingObject(); - // If they are the same object, they we can look at the indexes. If they - // index off of the object is the same for both pointers, they must alias. - // If they are provably different, they must not alias. Otherwise, we - // can't tell anything. - } - - - if (!isa(O1) && isa(V2)) - return NoAlias; // Unique values don't alias null - - if (isa(O1) || - (isa(O1) && - !cast(O1)->isArrayAllocation())) - if (cast(O1->getType())->getElementType()->isSized()) { - // If the size of the other access is larger than the total size of the - // global/alloca/malloc, it cannot be accessing the global (it's - // undefined to load or store bytes before or after an object). - const Type *ElTy = cast(O1->getType())->getElementType(); - unsigned GlobalSize = getTargetData().getTypeSize(ElTy); - if (GlobalSize < V2Size && V2Size != ~0U) - return NoAlias; - } - } - - if (O2) { - if (!isa(O2) && isa(V1)) - return NoAlias; // Unique values don't alias null - - if (isa(O2) || - (isa(O2) && - !cast(O2)->isArrayAllocation())) - if (cast(O2->getType())->getElementType()->isSized()) { - // If the size of the other access is larger than the total size of the - // global/alloca/malloc, it cannot be accessing the object (it's - // undefined to load or store bytes before or after an object). - const Type *ElTy = cast(O2->getType())->getElementType(); - unsigned GlobalSize = getTargetData().getTypeSize(ElTy); - if (GlobalSize < V1Size && V1Size != ~0U) - return NoAlias; - } + if (O1 != O2) { + // If V1/V2 point to two different objects we know that we have no alias. + if (isIdentifiedObject(O1) && isIdentifiedObject(O2)) + return NoAlias; + + // Incoming argument cannot alias locally allocated object! + if ((isa(O1) && isa(O2)) || + (isa(O2) && isa(O1))) + return NoAlias; + + // Most objects can't alias null. + if ((isa(V2) && isKnownNonNull(O1)) || + (isa(V1) && isKnownNonNull(O2))) + return NoAlias; } - + + // If the size of one access is larger than the entire object on the other + // side, then we know such behavior is undefined and can assume no alias. + const TargetData &TD = getTargetData(); + if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, TD)) || + (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, TD))) + return NoAlias; + + // If one pointer is the result of a call/invoke and the other is a + // non-escaping local object, then we know the object couldn't escape to a + // point where the call could return it. + if ((isa(O1) || isa(O1)) && + isNonEscapingLocalObject(O2)) + return NoAlias; + if ((isa(O2) || isa(O2)) && + isNonEscapingLocalObject(O1)) + return NoAlias; + // If we have two gep instructions with must-alias'ing base pointers, figure // out if the indexes to the GEP tell us anything about the derived pointer. // Note that we also handle chains of getelementptr instructions as well as @@ -353,20 +410,22 @@ BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, if (isGEP(V1) && isGEP(V2)) { // Drill down into the first non-gep value, to test for must-aliasing of // the base pointers. - const Value *BasePtr1 = V1, *BasePtr2 = V2; - do { - BasePtr1 = cast(BasePtr1)->getOperand(0); - } while (isGEP(BasePtr1) && - cast(BasePtr1)->getOperand(1) == - Constant::getNullValue(cast(BasePtr1)->getOperand(1)->getType())); - do { - BasePtr2 = cast(BasePtr2)->getOperand(0); - } while (isGEP(BasePtr2) && - cast(BasePtr2)->getOperand(1) == - Constant::getNullValue(cast(BasePtr2)->getOperand(1)->getType())); + const User *G = cast(V1); + while (isGEP(G->getOperand(0)) && + G->getOperand(1) == + Constant::getNullValue(G->getOperand(1)->getType())) + G = cast(G->getOperand(0)); + const Value *BasePtr1 = G->getOperand(0); + + G = cast(V2); + while (isGEP(G->getOperand(0)) && + G->getOperand(1) == + Constant::getNullValue(G->getOperand(1)->getType())) + G = cast(G->getOperand(0)); + const Value *BasePtr2 = G->getOperand(0); // Do the base pointers alias? - AliasResult BaseAlias = alias(BasePtr1, V1Size, BasePtr2, V2Size); + AliasResult BaseAlias = alias(BasePtr1, ~0U, BasePtr2, ~0U); if (BaseAlias == NoAlias) return NoAlias; if (BaseAlias == MustAlias) { // If the base pointers alias each other exactly, check to see if we can @@ -374,7 +433,7 @@ BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, // non-aliasing. // Collect all of the chained GEP operands together into one simple place - std::vector GEP1Ops, GEP2Ops; + SmallVector GEP1Ops, GEP2Ops; BasePtr1 = GetGEPOperands(V1, GEP1Ops); BasePtr2 = GetGEPOperands(V2, GEP2Ops); @@ -382,8 +441,10 @@ BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, // do the comparison. if (BasePtr1 == BasePtr2) { AliasResult GAlias = - CheckGEPInstructions(BasePtr1->getType(), GEP1Ops, V1Size, - BasePtr2->getType(), GEP2Ops, V2Size); + CheckGEPInstructions(BasePtr1->getType(), + &GEP1Ops[0], GEP1Ops.size(), V1Size, + BasePtr2->getType(), + &GEP2Ops[0], GEP2Ops.size(), V2Size); if (GAlias != MayAlias) return GAlias; } @@ -400,8 +461,8 @@ BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, } if (V1Size != ~0U && V2Size != ~0U) - if (const User *GEP = isGEP(V1)) { - std::vector GEPOperands; + if (isGEP(V1)) { + SmallVector GEPOperands; const Value *BasePtr = GetGEPOperands(V1, GEPOperands); AliasResult R = alias(BasePtr, V1Size, V2, V2Size); @@ -441,7 +502,9 @@ BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, GEPOperands[i] = Constant::getNullValue(GEPOperands[i]->getType()); int64_t Offset = - getTargetData().getIndexedOffset(BasePtr->getType(), GEPOperands); + getTargetData().getIndexedOffset(BasePtr->getType(), + &GEPOperands[0], + GEPOperands.size()); if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size) return NoAlias; @@ -453,14 +516,18 @@ BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, return MayAlias; } -static bool ValuesEqual(Value *V1, Value *V2) { +// This function is used to determin if the indices of two GEP instructions are +// equal. V1 and V2 are the indices. +static bool IndexOperandsEqual(Value *V1, Value *V2) { if (V1->getType() == V2->getType()) return V1 == V2; if (Constant *C1 = dyn_cast(V1)) if (Constant *C2 = dyn_cast(V2)) { - // Sign extend the constants to long types. - C1 = ConstantExpr::getSignExtend(C1, Type::LongTy); - C2 = ConstantExpr::getSignExtend(C2, Type::LongTy); + // Sign extend the constants to long types, if necessary + if (C1->getType() != Type::Int64Ty) + C1 = ConstantExpr::getSExt(C1, Type::Int64Ty); + if (C2->getType() != Type::Int64Ty) + C2 = ConstantExpr::getSExt(C2, Type::Int64Ty); return C1 == C2; } return false; @@ -469,11 +536,10 @@ static bool ValuesEqual(Value *V1, Value *V2) { /// CheckGEPInstructions - Check two GEP instructions with known must-aliasing /// base pointers. This checks to see if the index expressions preclude the /// pointers from aliasing... -AliasAnalysis::AliasResult BasicAliasAnalysis:: -CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, - unsigned G1S, - const Type *BasePtr2Ty, std::vector &GEP2Ops, - unsigned G2S) { +AliasAnalysis::AliasResult +BasicAliasAnalysis::CheckGEPInstructions( + const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S, + const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) { // We currently can't handle the case when the base pointers have different // primitive types. Since this is uncommon anyway, we are happy being // extremely conservative. @@ -484,12 +550,12 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, // Find the (possibly empty) initial sequence of equal values... which are not // necessarily constants. - unsigned NumGEP1Operands = GEP1Ops.size(), NumGEP2Operands = GEP2Ops.size(); + unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops; unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands); unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands); unsigned UnequalOper = 0; while (UnequalOper != MinOperands && - ValuesEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) { + IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) { // Advance through the type as we go... ++UnequalOper; if (const CompositeType *CT = dyn_cast(BasePtr1Ty)) @@ -508,7 +574,10 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, // getelementptrs, check to see if the tail of the leftover one is all zeros. // If so, return mustalias. if (UnequalOper == MinOperands) { - if (GEP1Ops.size() < GEP2Ops.size()) std::swap(GEP1Ops, GEP2Ops); + if (NumGEP1Ops < NumGEP2Ops) { + std::swap(GEP1Ops, GEP2Ops); + std::swap(NumGEP1Ops, NumGEP2Ops); + } bool AllAreZeros = true; for (unsigned i = UnequalOper; i != MaxOperands; ++i) @@ -527,6 +596,14 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, // chain. For example: // A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S)) // + // We have to be careful here about array accesses. In particular, consider: + // A[1][0] vs A[0][i] + // In this case, we don't *know* that the array will be accessed in bounds: + // the index could even be negative. Because of this, we have to + // conservatively *give up* and return may alias. We disregard differing + // array subscripts that are followed by a variable index without going + // through a struct. + // unsigned SizeMax = std::max(G1S, G2S); if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work. @@ -542,20 +619,47 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, if (Constant *G2OC = dyn_cast(const_cast(G2Oper))){ if (G1OC->getType() != G2OC->getType()) { // Sign extend both operands to long. - G1OC = ConstantExpr::getSignExtend(G1OC, Type::LongTy); - G2OC = ConstantExpr::getSignExtend(G2OC, Type::LongTy); + if (G1OC->getType() != Type::Int64Ty) + G1OC = ConstantExpr::getSExt(G1OC, Type::Int64Ty); + if (G2OC->getType() != Type::Int64Ty) + G2OC = ConstantExpr::getSExt(G2OC, Type::Int64Ty); GEP1Ops[FirstConstantOper] = G1OC; GEP2Ops[FirstConstantOper] = G2OC; } - + if (G1OC != G2OC) { + // Handle the "be careful" case above: if this is an array/vector + // subscript, scan for a subsequent variable array index. + if (isa(BasePtr1Ty)) { + const Type *NextTy = + cast(BasePtr1Ty)->getElementType(); + bool isBadCase = false; + + for (unsigned Idx = FirstConstantOper+1; + Idx != MinOperands && isa(NextTy); ++Idx) { + const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx]; + if (!isa(V1) || !isa(V2)) { + isBadCase = true; + break; + } + NextTy = cast(NextTy)->getElementType(); + } + + if (isBadCase) G1OC = 0; + } + // Make sure they are comparable (ie, not constant expressions), and // make sure the GEP with the smaller leading constant is GEP1. - Constant *Compare = ConstantExpr::getSetGT(G1OC, G2OC); - if (ConstantBool *CV = dyn_cast(Compare)) { - if (CV->getValue()) // If they are comparable and G2 > G1 - std::swap(GEP1Ops, GEP2Ops); // Make GEP1 < GEP2 - break; + if (G1OC) { + Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT, + G1OC, G2OC); + if (ConstantInt *CV = dyn_cast(Compare)) { + if (CV->getZExtValue()) { // If they are comparable and G2 > G1 + std::swap(GEP1Ops, GEP2Ops); // Make GEP1 < GEP2 + std::swap(NumGEP1Ops, NumGEP2Ops); + } + break; + } } } } @@ -569,16 +673,18 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, // case, there may still be hope. Check this now. if (FirstConstantOper == MinOperands) { // Make GEP1Ops be the longer one if there is a longer one. - if (GEP1Ops.size() < GEP2Ops.size()) + if (NumGEP1Ops < NumGEP2Ops) { std::swap(GEP1Ops, GEP2Ops); + std::swap(NumGEP1Ops, NumGEP2Ops); + } // Is there anything to check? - if (GEP1Ops.size() > MinOperands) { + if (NumGEP1Ops > MinOperands) { for (unsigned i = FirstConstantOper; i != MaxOperands; ++i) - if (isa(GEP1Ops[i]) && - !cast(GEP1Ops[i])->isNullValue()) { + if (isa(GEP1Ops[i]) && + !cast(GEP1Ops[i])->isZero()) { // Yup, there's a constant in the tail. Set all variables to - // constants in the GEP instruction to make it suiteable for + // constants in the GEP instruction to make it suitable for // TargetData::getIndexedOffset. for (i = 0; i != MaxOperands; ++i) if (!isa(GEP1Ops[i])) @@ -586,15 +692,22 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, // Okay, now get the offset. This is the relative offset for the full // instruction. const TargetData &TD = getTargetData(); - int64_t Offset1 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops); + int64_t Offset1 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops, + NumGEP1Ops); + + // Now check without any constants at the end. + int64_t Offset2 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops, + MinOperands); - // Now crop off any constants from the end... - GEP1Ops.resize(MinOperands); - int64_t Offset2 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops); + // Make sure we compare the absolute difference. + if (Offset1 > Offset2) + std::swap(Offset1, Offset2); // If the tail provided a bit enough offset, return noalias! if ((uint64_t)(Offset2-Offset1) >= SizeMax) return NoAlias; + // Otherwise break - we don't look for another constant in the tail. + break; } } @@ -608,23 +721,30 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, // than the first constant index of GEP2. // Advance BasePtr[12]Ty over this first differing constant operand. - BasePtr2Ty = cast(BasePtr1Ty)->getTypeAtIndex(GEP2Ops[FirstConstantOper]); - BasePtr1Ty = cast(BasePtr1Ty)->getTypeAtIndex(GEP1Ops[FirstConstantOper]); + BasePtr2Ty = cast(BasePtr1Ty)-> + getTypeAtIndex(GEP2Ops[FirstConstantOper]); + BasePtr1Ty = cast(BasePtr1Ty)-> + getTypeAtIndex(GEP1Ops[FirstConstantOper]); // We are going to be using TargetData::getIndexedOffset to determine the // offset that each of the GEP's is reaching. To do this, we have to convert // all variable references to constant references. To do this, we convert the - // initial equal sequence of variables into constant zeros to start with. - for (unsigned i = 0; i != FirstConstantOper; ++i) - if (!isa(GEP1Ops[i]) || !isa(GEP2Ops[i])) - GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Type::UIntTy); + // initial sequence of array subscripts into constant zeros to start with. + const Type *ZeroIdxTy = GEPPointerTy; + for (unsigned i = 0; i != FirstConstantOper; ++i) { + if (!isa(ZeroIdxTy)) + GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Type::Int32Ty); + + if (const CompositeType *CT = dyn_cast(ZeroIdxTy)) + ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]); + } // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok // Loop over the rest of the operands... for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) { - const Value *Op1 = i < GEP1Ops.size() ? GEP1Ops[i] : 0; - const Value *Op2 = i < GEP2Ops.size() ? GEP2Ops[i] : 0; + const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0; + const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0; // If they are equal, use a zero index... if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) { if (!isa(Op1)) @@ -634,10 +754,14 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, if (Op1) { if (const ConstantInt *Op1C = dyn_cast(Op1)) { // If this is an array index, make sure the array element is in range. - if (const ArrayType *AT = dyn_cast(BasePtr1Ty)) - if (Op1C->getRawValue() >= AT->getNumElements()) + if (const ArrayType *AT = dyn_cast(BasePtr1Ty)) { + if (Op1C->getZExtValue() >= AT->getNumElements()) return MayAlias; // Be conservative with out-of-range accesses - + } else if (const VectorType *VT = dyn_cast(BasePtr1Ty)) { + if (Op1C->getZExtValue() >= VT->getNumElements()) + return MayAlias; // Be conservative with out-of-range accesses + } + } else { // GEP1 is known to produce a value less than GEP2. To be // conservatively correct, we must assume the largest possible @@ -650,16 +774,22 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, // value possible. // if (const ArrayType *AT = dyn_cast(BasePtr1Ty)) - GEP1Ops[i] = ConstantSInt::get(Type::LongTy,AT->getNumElements()-1); + GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,AT->getNumElements()-1); + else if (const VectorType *VT = dyn_cast(BasePtr1Ty)) + GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,VT->getNumElements()-1); } } if (Op2) { if (const ConstantInt *Op2C = dyn_cast(Op2)) { // If this is an array index, make sure the array element is in range. - if (const ArrayType *AT = dyn_cast(BasePtr1Ty)) - if (Op2C->getRawValue() >= AT->getNumElements()) + if (const ArrayType *AT = dyn_cast(BasePtr2Ty)) { + if (Op2C->getZExtValue() >= AT->getNumElements()) + return MayAlias; // Be conservative with out-of-range accesses + } else if (const VectorType *VT = dyn_cast(BasePtr2Ty)) { + if (Op2C->getZExtValue() >= VT->getNumElements()) return MayAlias; // Be conservative with out-of-range accesses + } } else { // Conservatively assume the minimum value for this index GEP2Ops[i] = Constant::getNullValue(Op2->getType()); } @@ -682,127 +812,25 @@ CheckGEPInstructions(const Type* BasePtr1Ty, std::vector &GEP1Ops, } if (GEPPointerTy->getElementType()->isSized()) { - int64_t Offset1 = getTargetData().getIndexedOffset(GEPPointerTy, GEP1Ops); - int64_t Offset2 = getTargetData().getIndexedOffset(GEPPointerTy, GEP2Ops); - assert(Offset1 Offset2) + std::swap(Offset1, Offset2); + if ((uint64_t)(Offset2-Offset1) >= SizeMax) { - //std::cerr << "Determined that these two GEP's don't alias [" - // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2; + //cerr << "Determined that these two GEP's don't alias [" + // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2; return NoAlias; } } return MayAlias; } -namespace { - struct StringCompare { - bool operator()(const char *LHS, const char *RHS) { - return strcmp(LHS, RHS) < 0; - } - }; -} - -// Note that this list cannot contain libm functions (such as acos and sqrt) -// that set errno on a domain or other error. -static const char *DoesntAccessMemoryTable[] = { - // LLVM intrinsics: - "llvm.frameaddress", "llvm.returnaddress", "llvm.readport", - "llvm.isunordered", "llvm.sqrt", "llvm.bswap.i16", "llvm.bswap.i32", - "llvm.bswap.i64", "llvm.ctpop", "llvm.ctlz", "llvm.cttz", - - "abs", "labs", "llabs", "imaxabs", "fabs", "fabsf", "fabsl", - "trunc", "truncf", "truncl", "ldexp", - - "atan", "atanf", "atanl", "atan2", "atan2f", "atan2l", - "cbrt", - "cos", "cosf", "cosl", - "exp", "expf", "expl", - "hypot", - "sin", "sinf", "sinl", - "tan", "tanf", "tanl", "tanh", "tanhf", "tanhl", - - "floor", "floorf", "floorl", "ceil", "ceilf", "ceill", - - // ctype.h - "isalnum", "isalpha", "iscntrl", "isdigit", "isgraph", "islower", "isprint" - "ispunct", "isspace", "isupper", "isxdigit", "tolower", "toupper", - - // wctype.h" - "iswalnum", "iswalpha", "iswcntrl", "iswdigit", "iswgraph", "iswlower", - "iswprint", "iswpunct", "iswspace", "iswupper", "iswxdigit", - - "iswctype", "towctrans", "towlower", "towupper", - - "btowc", "wctob", - - "isinf", "isnan", "finite", - - // C99 math functions - "copysign", "copysignf", "copysignd", - "nexttoward", "nexttowardf", "nexttowardd", - "nextafter", "nextafterf", "nextafterd", - - // ISO C99: - "__signbit", "__signbitf", "__signbitl", -}; - -static const unsigned DAMTableSize = - sizeof(DoesntAccessMemoryTable)/sizeof(DoesntAccessMemoryTable[0]); - -static const char *OnlyReadsMemoryTable[] = { - "atoi", "atol", "atof", "atoll", "atoq", "a64l", - "bcmp", "memcmp", "memchr", "memrchr", "wmemcmp", "wmemchr", - - // Strings - "strcmp", "strcasecmp", "strcoll", "strncmp", "strncasecmp", - "strchr", "strcspn", "strlen", "strpbrk", "strrchr", "strspn", "strstr", - "index", "rindex", - - // Wide char strings - "wcschr", "wcscmp", "wcscoll", "wcscspn", "wcslen", "wcsncmp", "wcspbrk", - "wcsrchr", "wcsspn", "wcsstr", - - // glibc - "alphasort", "alphasort64", "versionsort", "versionsort64", - - // C99 - "nan", "nanf", "nand", - - // File I/O - "feof", "ferror", "fileno", - "feof_unlocked", "ferror_unlocked", "fileno_unlocked" -}; - -static const unsigned ORMTableSize = - sizeof(OnlyReadsMemoryTable)/sizeof(OnlyReadsMemoryTable[0]); - -AliasAnalysis::ModRefBehavior -BasicAliasAnalysis::getModRefBehavior(Function *F, CallSite CS, - std::vector *Info) { - if (!F->isExternal()) return UnknownModRefBehavior; - - static bool Initialized = false; - if (!Initialized) { - // Sort the table the first time through. - std::sort(DoesntAccessMemoryTable, DoesntAccessMemoryTable+DAMTableSize, - StringCompare()); - std::sort(OnlyReadsMemoryTable, OnlyReadsMemoryTable+ORMTableSize, - StringCompare()); - Initialized = true; - } - - const char **Ptr = std::lower_bound(DoesntAccessMemoryTable, - DoesntAccessMemoryTable+DAMTableSize, - F->getName().c_str(), StringCompare()); - if (Ptr != DoesntAccessMemoryTable+DAMTableSize && *Ptr == F->getName()) - return DoesNotAccessMemory; - - Ptr = std::lower_bound(OnlyReadsMemoryTable, - OnlyReadsMemoryTable+ORMTableSize, - F->getName().c_str(), StringCompare()); - if (Ptr != OnlyReadsMemoryTable+ORMTableSize && *Ptr == F->getName()) - return OnlyReadsMemory; - - return UnknownModRefBehavior; -} +// Make sure that anything that uses AliasAnalysis pulls in this file... +DEFINING_FILE_FOR(BasicAliasAnalysis)