X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FBasicAliasAnalysis.cpp;h=4f53a6d62559bea022d1ae547c1530481f759ec4;hb=8be3291f5942e3ae4a5d66c480e7aabe2f771031;hp=57d320971dbef17b0b9e3dba382f67bc4cb825c4;hpb=f23d0d392ab47e191e1bbd17dc7487b9d461485c;p=oota-llvm.git diff --git a/lib/Analysis/BasicAliasAnalysis.cpp b/lib/Analysis/BasicAliasAnalysis.cpp index 57d320971db..4f53a6d6255 100644 --- a/lib/Analysis/BasicAliasAnalysis.cpp +++ b/lib/Analysis/BasicAliasAnalysis.cpp @@ -21,13 +21,15 @@ #include "llvm/GlobalVariable.h" #include "llvm/Instructions.h" #include "llvm/IntrinsicInst.h" +#include "llvm/Operator.h" #include "llvm/Pass.h" +#include "llvm/Analysis/CaptureTracking.h" +#include "llvm/Analysis/MemoryBuiltins.h" +#include "llvm/Analysis/ValueTracking.h" #include "llvm/Target/TargetData.h" +#include "llvm/ADT/SmallPtrSet.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 "llvm/Support/ErrorHandling.h" #include using namespace llvm; @@ -35,110 +37,6 @@ using namespace llvm; // 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 argument to the call has the nocapture attribute, then the call - // may store or load to the pointer, but it cannot escape. - if (cast(I)->paramHasAttr(UI.getOperandNo(), - Attribute::NoCapture)) - continue; - - // FIXME: MemIntrinsics should have their operands marked nocapture! - if (isa(I)) - continue; // next use - return true; - case Instruction::Invoke: - // If the argument to the call has the nocapture attribute, then the call - // may store or load to the pointer, but it cannot escape. - if (cast(I)->paramHasAttr(UI.getOperandNo()-2, - Attribute::NoCapture)) - continue; - return true; - 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; -} - -/// isNoAliasCall - Return true if this pointer is returned by a noalias -/// function. -static bool isNoAliasCall(const Value *V) { - if (isa(V) || isa(V)) - return CallSite(const_cast(cast(V))) - .paramHasAttr(0, Attribute::NoAlias); - return false; -} - -/// 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 -/// NoAlias returns -/// -static bool isIdentifiedObject(const Value *V) { - if (isa(V) || isa(V) || isNoAliasCall(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) { @@ -159,17 +57,41 @@ static bool isKnownNonNull(const Value *V) { /// 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) || isNoAliasCall(V)) - return !AddressMightEscape(V); - + if (isa(V) || isNoAliasCall(V)) + // Set StoreCaptures to True so that we can assume in our callers that the + // pointer is not the result of a load instruction. Currently + // PointerMayBeCaptured doesn't have any special analysis for the + // StoreCaptures=false case; if it did, our callers could be refined to be + // more precise. + return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true); + // If this is an argument that corresponds to a byval or noalias argument, - // it can't escape either. + // then it has not escaped before entering the function. Check if it escapes + // inside the function. if (const Argument *A = dyn_cast(V)) - if (A->hasByValAttr() || A->hasNoAliasAttr()) - return !AddressMightEscape(V); + if (A->hasByValAttr() || A->hasNoAliasAttr()) { + // Don't bother analyzing arguments already known not to escape. + if (A->hasNoCaptureAttr()) + return true; + return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true); + } return false; } +/// isEscapeSource - Return true if the pointer is one which would have +/// been considered an escape by isNonEscapingLocalObject. +static bool isEscapeSource(const Value *V) { + if (isa(V) || isa(V) || isa(V)) + return true; + + // The load case works because isNonEscapingLocalObject considers all + // stores to be escapes (it passes true for the StoreCaptures argument + // to PointerMayBeCaptured). + if (isa(V)) + return true; + + return false; +} /// isObjectSmallerThan - Return true if we can prove that the object specified /// by V is smaller than Size. @@ -178,11 +100,17 @@ static bool isObjectSmallerThan(const Value *V, unsigned Size, const Type *AccessTy; if (const GlobalVariable *GV = dyn_cast(V)) { AccessTy = GV->getType()->getElementType(); - } else if (const AllocationInst *AI = dyn_cast(V)) { + } else if (const AllocaInst *AI = dyn_cast(V)) { if (!AI->isArrayAllocation()) AccessTy = AI->getType()->getElementType(); else return false; + } else if (const CallInst* CI = extractMallocCall(V)) { + if (!isArrayMalloc(V, &TD)) + // The size is the argument to the malloc call. + if (const ConstantInt* C = dyn_cast(CI->getArgOperand(0))) + return (C->getZExtValue() < Size); + return false; } else if (const Argument *A = dyn_cast(V)) { if (A->hasByValAttr()) AccessTy = cast(A->getType())->getElementType(); @@ -193,7 +121,7 @@ static bool isObjectSmallerThan(const Value *V, unsigned Size, } if (AccessTy->isSized()) - return TD.getABITypeSize(AccessTy) < Size; + return TD.getTypeAllocSize(AccessTy) < Size; return false; } @@ -207,17 +135,16 @@ 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 VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis { + struct 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(); } virtual void initializePass() { - TD = &getAnalysis(); + TD = getAnalysisIfAvailable(); } virtual AliasResult alias(const Value *V1, unsigned V1Size, @@ -225,17 +152,11 @@ namespace { return MayAlias; } - virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS, - std::vector *Info) { - return UnknownModRefBehavior; - } - virtual void getArgumentAccesses(Function *F, CallSite CS, std::vector &Info) { - assert(0 && "This method may not be called on this function!"); + llvm_unreachable("This method may not be called on this function!"); } - virtual void getMustAliases(Value *P, std::vector &RetVals) { } virtual bool pointsToConstantMemory(const Value *P) { return false; } virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) { return ModRef; @@ -243,10 +164,19 @@ namespace { virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) { return ModRef; } - virtual bool hasNoModRefInfoForCalls() const { return true; } virtual void deleteValue(Value *V) {} virtual void copyValue(Value *From, Value *To) {} + + /// getAdjustedAnalysisPointer - This method is used when a pass implements + /// an analysis interface through multiple inheritance. If needed, it should + /// override this to adjust the this pointer as needed for the specified pass + /// info. + virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) { + if (PI->isPassID(&AliasAnalysis::ID)) + return (AliasAnalysis*)this; + return this; + } }; } // End of anonymous namespace @@ -261,39 +191,85 @@ static RegisterAnalysisGroup V(U); ImmutablePass *llvm::createNoAAPass() { return new NoAA(); } //===----------------------------------------------------------------------===// -// BasicAA Pass +// BasicAliasAnalysis Pass //===----------------------------------------------------------------------===// +#ifndef NDEBUG +static const Function *getParent(const Value *V) { + if (const Instruction *inst = dyn_cast(V)) + return inst->getParent()->getParent(); + + if (const Argument *arg = dyn_cast(V)) + return arg->getParent(); + + return NULL; +} + +static bool notDifferentParent(const Value *O1, const Value *O2) { + + const Function *F1 = getParent(O1); + const Function *F2 = getParent(O2); + + return !F1 || !F2 || F1 == F2; +} +#endif + 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 VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA { + struct 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); + const Value *V2, unsigned V2Size) { + assert(Visited.empty() && "Visited must be cleared after use!"); + assert(notDifferentParent(V1, V2) && + "BasicAliasAnalysis doesn't support interprocedural queries."); + AliasResult Alias = aliasCheck(V1, V1Size, V2, V2Size); + Visited.clear(); + return Alias; + } ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size); ModRefResult getModRefInfo(CallSite CS1, CallSite CS2); - - /// hasNoModRefInfoForCalls - We can provide mod/ref information against - /// non-escaping allocations. - virtual bool hasNoModRefInfoForCalls() const { return false; } /// pointsToConstantMemory - Chase pointers until we find a (constant /// global) or not. bool pointsToConstantMemory(const Value *P); + /// getAdjustedAnalysisPointer - This method is used when a pass implements + /// an analysis interface through multiple inheritance. If needed, it should + /// override this to adjust the this pointer as needed for the specified pass + /// info. + virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) { + if (PI->isPassID(&AliasAnalysis::ID)) + return (AliasAnalysis*)this; + return this; + } + 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, - Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size, - const Type *BasePtr2Ty, - Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size); + // Visited - Track instructions visited by a aliasPHI, aliasSelect(), and aliasGEP(). + SmallPtrSet Visited; + + // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP + // instruction against another. + AliasResult aliasGEP(const GEPOperator *V1, unsigned V1Size, + const Value *V2, unsigned V2Size, + const Value *UnderlyingV1, const Value *UnderlyingV2); + + // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI + // instruction against another. + AliasResult aliasPHI(const PHINode *PN, unsigned PNSize, + const Value *V2, unsigned V2Size); + + /// aliasSelect - Disambiguate a Select instruction against another value. + AliasResult aliasSelect(const SelectInst *SI, unsigned SISize, + const Value *V2, unsigned V2Size); + + AliasResult aliasCheck(const Value *V1, unsigned V1Size, + const Value *V2, unsigned V2Size); }; } // End of anonymous namespace @@ -315,45 +291,127 @@ ImmutablePass *llvm::createBasicAliasAnalysisPass() { bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) { if (const GlobalVariable *GV = dyn_cast(P->getUnderlyingObject())) + // Note: this doesn't require GV to be "ODR" because it isn't legal for a + // global to be marked constant in some modules and non-constant in others. + // GV may even be a declaration, not a definition. return GV->isConstant(); return false; } -// getModRefInfo - Check to see if the specified callsite can clobber the -// specified memory object. Since we only look at local properties of this -// function, we really can't say much about this query. We do, however, use -// simple "address taken" analysis on local objects. -// + +/// getModRefInfo - Check to see if the specified callsite can clobber the +/// specified memory object. Since we only look at local properties of this +/// function, we really can't say much about this query. We do, however, use +/// simple "address taken" analysis on local objects. AliasAnalysis::ModRefResult BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) { - 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()) - 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; + assert(notDifferentParent(CS.getInstruction(), P) && + "AliasAnalysis query involving multiple functions!"); + + 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()) + 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 pointer + // as an argument, and itself doesn't capture it. + if (!isa(Object) && CS.getInstruction() != Object && + isNonEscapingLocalObject(Object)) { + bool PassedAsArg = false; + unsigned ArgNo = 0; + for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end(); + CI != CE; ++CI, ++ArgNo) { + // Only look at the no-capture pointer arguments. + if (!(*CI)->getType()->isPointerTy() || + !CS.paramHasAttr(ArgNo+1, Attribute::NoCapture)) + continue; - if (!passedAsArg) + // If this is a no-capture pointer argument, see if we can tell that it + // is impossible to alias the pointer we're checking. If not, we have to + // assume that the call could touch the pointer, even though it doesn't + // escape. + if (!isNoAlias(cast(CI), ~0U, P, ~0U)) { + PassedAsArg = true; + break; + } + } + + if (!PassedAsArg) + return NoModRef; + } + + // Finally, handle specific knowledge of intrinsics. + IntrinsicInst *II = dyn_cast(CS.getInstruction()); + if (II == 0) + return AliasAnalysis::getModRefInfo(CS, P, Size); + + switch (II->getIntrinsicID()) { + default: break; + case Intrinsic::memcpy: + case Intrinsic::memmove: { + unsigned Len = ~0U; + if (ConstantInt *LenCI = dyn_cast(II->getArgOperand(2))) + Len = LenCI->getZExtValue(); + Value *Dest = II->getArgOperand(0); + Value *Src = II->getArgOperand(1); + if (isNoAlias(Dest, Len, P, Size)) { + if (isNoAlias(Src, Len, P, Size)) return NoModRef; + return Ref; } + break; + } + case Intrinsic::memset: + // Since memset is 'accesses arguments' only, the AliasAnalysis base class + // will handle it for the variable length case. + if (ConstantInt *LenCI = dyn_cast(II->getArgOperand(2))) { + unsigned Len = LenCI->getZExtValue(); + Value *Dest = II->getArgOperand(0); + if (isNoAlias(Dest, Len, P, Size)) + return NoModRef; + } + break; + case Intrinsic::atomic_cmp_swap: + case Intrinsic::atomic_swap: + case Intrinsic::atomic_load_add: + case Intrinsic::atomic_load_sub: + case Intrinsic::atomic_load_and: + case Intrinsic::atomic_load_nand: + case Intrinsic::atomic_load_or: + case Intrinsic::atomic_load_xor: + case Intrinsic::atomic_load_max: + case Intrinsic::atomic_load_min: + case Intrinsic::atomic_load_umax: + case Intrinsic::atomic_load_umin: + if (TD) { + Value *Op1 = II->getArgOperand(0); + unsigned Op1Size = TD->getTypeStoreSize(Op1->getType()); + if (isNoAlias(Op1, Op1Size, P, Size)) + return NoModRef; + } + break; + case Intrinsic::lifetime_start: + case Intrinsic::lifetime_end: + case Intrinsic::invariant_start: { + unsigned PtrSize = cast(II->getArgOperand(0))->getZExtValue(); + if (isNoAlias(II->getArgOperand(1), PtrSize, P, Size)) + return NoModRef; + break; + } + case Intrinsic::invariant_end: { + unsigned PtrSize = cast(II->getArgOperand(1))->getZExtValue(); + if (isNoAlias(II->getArgOperand(2), PtrSize, P, Size)) + return NoModRef; + break; + } } // The AliasAnalysis base class has some smarts, lets use them. @@ -378,514 +436,382 @@ BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) { return NoAA::getModRefInfo(CS1, CS2); } +/// GetIndiceDifference - Dest and Src are the variable indices from two +/// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base +/// pointers. Subtract the GEP2 indices from GEP1 to find the symbolic +/// difference between the two pointers. +static void GetIndiceDifference( + SmallVectorImpl > &Dest, + const SmallVectorImpl > &Src) { + if (Src.empty()) return; + + for (unsigned i = 0, e = Src.size(); i != e; ++i) { + const Value *V = Src[i].first; + int64_t Scale = Src[i].second; + + // Find V in Dest. This is N^2, but pointer indices almost never have more + // than a few variable indexes. + for (unsigned j = 0, e = Dest.size(); j != e; ++j) { + if (Dest[j].first != V) continue; + + // If we found it, subtract off Scale V's from the entry in Dest. If it + // goes to zero, remove the entry. + if (Dest[j].second != Scale) + Dest[j].second -= Scale; + else + Dest.erase(Dest.begin()+j); + Scale = 0; + break; + } + + // If we didn't consume this entry, add it to the end of the Dest list. + if (Scale) + Dest.push_back(std::make_pair(V, -Scale)); + } +} -// alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such -// as array references. -// +/// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction +/// against another pointer. We know that V1 is a GEP, but we don't know +/// anything about V2. UnderlyingV1 is GEP1->getUnderlyingObject(), +/// UnderlyingV2 is the same for V2. +/// AliasAnalysis::AliasResult -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->isCast() && isa(CE->getOperand(0)->getType())) - V1 = CE->getOperand(0); - if (const ConstantExpr *CE = dyn_cast(V2)) - 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())) - return NoAlias; // Scalars cannot alias each other - - // Strip off cast instructions. Since V1 and V2 are pointers, they must be - // pointer<->pointer bitcasts. - 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 = V1->getUnderlyingObject(); - const Value *O2 = V2->getUnderlyingObject(); +BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, unsigned V1Size, + const Value *V2, unsigned V2Size, + const Value *UnderlyingV1, + const Value *UnderlyingV2) { + // If this GEP has been visited before, we're on a use-def cycle. + // Such cycles are only valid when PHI nodes are involved or in unreachable + // code. The visitPHI function catches cycles containing PHIs, but there + // could still be a cycle without PHIs in unreachable code. + if (!Visited.insert(GEP1)) + return MayAlias; - 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; - - // Arguments can't alias with local allocations or noalias calls. - if ((isa(O1) && (isa(O2) || isNoAliasCall(O2))) || - (isa(O2) && (isa(O1) || isNoAliasCall(O1)))) - return NoAlias; + int64_t GEP1BaseOffset; + SmallVector, 4> GEP1VariableIndices; - // 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 - // constant expression getelementptrs here. - // - if (isGEP(V1) && isGEP(V2)) { - const User *GEP1 = cast(V1); - const User *GEP2 = cast(V2); + if (const GEPOperator *GEP2 = dyn_cast(V2)) { + // Do the base pointers alias? + AliasResult BaseAlias = aliasCheck(UnderlyingV1, ~0U, UnderlyingV2, ~0U); + + // If we get a No or May, then return it immediately, no amount of analysis + // will improve this situation. + if (BaseAlias != MustAlias) return BaseAlias; - // If V1 and V2 are identical GEPs, just recurse down on both of them. - // This allows us to analyze things like: - // P = gep A, 0, i, 1 - // Q = gep B, 0, i, 1 - // by just analyzing A and B. This is even safe for variable indices. - if (GEP1->getType() == GEP2->getType() && - GEP1->getNumOperands() == GEP2->getNumOperands() && - GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() && - // All operands are the same, ignoring the base. - std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1)) - return alias(GEP1->getOperand(0), V1Size, GEP2->getOperand(0), V2Size); + // Otherwise, we have a MustAlias. Since the base pointers alias each other + // exactly, see if the computed offset from the common pointer tells us + // about the relation of the resulting pointer. + const Value *GEP1BasePtr = + DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD); + int64_t GEP2BaseOffset; + SmallVector, 4> GEP2VariableIndices; + const Value *GEP2BasePtr = + DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD); - // Drill down into the first non-gep value, to test for must-aliasing of - // the base pointers. - while (isGEP(GEP1->getOperand(0)) && - GEP1->getOperand(1) == - Constant::getNullValue(GEP1->getOperand(1)->getType())) - GEP1 = cast(GEP1->getOperand(0)); - const Value *BasePtr1 = GEP1->getOperand(0); - - while (isGEP(GEP2->getOperand(0)) && - GEP2->getOperand(1) == - Constant::getNullValue(GEP2->getOperand(1)->getType())) - GEP2 = cast(GEP2->getOperand(0)); - const Value *BasePtr2 = GEP2->getOperand(0); + // If DecomposeGEPExpression isn't able to look all the way through the + // addressing operation, we must not have TD and this is too complex for us + // to handle without it. + if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) { + assert(TD == 0 && + "DecomposeGEPExpression and getUnderlyingObject disagree!"); + return MayAlias; + } + + // Subtract the GEP2 pointer from the GEP1 pointer to find out their + // symbolic difference. + GEP1BaseOffset -= GEP2BaseOffset; + GetIndiceDifference(GEP1VariableIndices, GEP2VariableIndices); + + } else { + // Check to see if these two pointers are related by the getelementptr + // instruction. If one pointer is a GEP with a non-zero index of the other + // pointer, we know they cannot alias. - // Do the base pointers alias? - 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 - // figure out anything about the resultant pointers, to try to prove - // non-aliasing. - - // Collect all of the chained GEP operands together into one simple place - SmallVector GEP1Ops, GEP2Ops; - BasePtr1 = GetGEPOperands(V1, GEP1Ops); - BasePtr2 = GetGEPOperands(V2, GEP2Ops); - - // If GetGEPOperands were able to fold to the same must-aliased pointer, - // do the comparison. - if (BasePtr1 == BasePtr2) { - AliasResult GAlias = - CheckGEPInstructions(BasePtr1->getType(), - &GEP1Ops[0], GEP1Ops.size(), V1Size, - BasePtr2->getType(), - &GEP2Ops[0], GEP2Ops.size(), V2Size); - if (GAlias != MayAlias) - return GAlias; - } + // If both accesses are unknown size, we can't do anything useful here. + if (V1Size == ~0U && V2Size == ~0U) + return MayAlias; + + AliasResult R = aliasCheck(UnderlyingV1, ~0U, V2, V2Size); + if (R != MustAlias) + // If V2 may alias GEP base pointer, conservatively returns MayAlias. + // If V2 is known not to alias GEP base pointer, then the two values + // cannot alias per GEP semantics: "A pointer value formed from a + // getelementptr instruction is associated with the addresses associated + // with the first operand of the getelementptr". + return R; + + const Value *GEP1BasePtr = + DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD); + + // If DecomposeGEPExpression isn't able to look all the way through the + // addressing operation, we must not have TD and this is too complex for us + // to handle without it. + if (GEP1BasePtr != UnderlyingV1) { + assert(TD == 0 && + "DecomposeGEPExpression and getUnderlyingObject disagree!"); + return MayAlias; } } - - // Check to see if these two pointers are related by a getelementptr - // instruction. If one pointer is a GEP with a non-zero index of the other - // pointer, we know they cannot alias. + + // In the two GEP Case, if there is no difference in the offsets of the + // computed pointers, the resultant pointers are a must alias. This + // hapens when we have two lexically identical GEP's (for example). // - if (isGEP(V2)) { - std::swap(V1, V2); - std::swap(V1Size, V2Size); + // In the other case, if we have getelementptr , 0, 0, 0, 0, ... and V2 + // must aliases the GEP, the end result is a must alias also. + if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty()) + return MustAlias; + + // If we have a known constant offset, see if this offset is larger than the + // access size being queried. If so, and if no variable indices can remove + // pieces of this constant, then we know we have a no-alias. For example, + // &A[100] != &A. + + // In order to handle cases like &A[100][i] where i is an out of range + // subscript, we have to ignore all constant offset pieces that are a multiple + // of a scaled index. Do this by removing constant offsets that are a + // multiple of any of our variable indices. This allows us to transform + // things like &A[i][1] because i has a stride of (e.g.) 8 bytes but the 1 + // provides an offset of 4 bytes (assuming a <= 4 byte access). + for (unsigned i = 0, e = GEP1VariableIndices.size(); + i != e && GEP1BaseOffset;++i) + if (int64_t RemovedOffset = GEP1BaseOffset/GEP1VariableIndices[i].second) + GEP1BaseOffset -= RemovedOffset*GEP1VariableIndices[i].second; + + // If our known offset is bigger than the access size, we know we don't have + // an alias. + if (GEP1BaseOffset) { + if (GEP1BaseOffset >= (int64_t)V2Size || + GEP1BaseOffset <= -(int64_t)V1Size) + return NoAlias; } - - if (V1Size != ~0U && V2Size != ~0U) - if (isGEP(V1)) { - SmallVector GEPOperands; - const Value *BasePtr = GetGEPOperands(V1, GEPOperands); - - AliasResult R = alias(BasePtr, V1Size, V2, V2Size); - if (R == MustAlias) { - // If there is at least one non-zero constant index, we know they cannot - // alias. - bool ConstantFound = false; - bool AllZerosFound = true; - for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i) - if (const Constant *C = dyn_cast(GEPOperands[i])) { - if (!C->isNullValue()) { - ConstantFound = true; - AllZerosFound = false; - break; - } - } else { - AllZerosFound = false; - } - - // If we have getelementptr , 0, 0, 0, 0, ... and V2 must aliases - // the ptr, the end result is a must alias also. - if (AllZerosFound) - return MustAlias; - - if (ConstantFound) { - if (V2Size <= 1 && V1Size <= 1) // Just pointer check? - return NoAlias; - - // Otherwise we have to check to see that the distance is more than - // the size of the argument... build an index vector that is equal to - // the arguments provided, except substitute 0's for any variable - // indexes we find... - if (cast( - BasePtr->getType())->getElementType()->isSized()) { - for (unsigned i = 0; i != GEPOperands.size(); ++i) - if (!isa(GEPOperands[i])) - GEPOperands[i] = - Constant::getNullValue(GEPOperands[i]->getType()); - int64_t Offset = - getTargetData().getIndexedOffset(BasePtr->getType(), - &GEPOperands[0], - GEPOperands.size()); - - if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size) - return NoAlias; - } - } - } - } - + return MayAlias; } -// This function is used to determine 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, 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; +/// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select +/// instruction against another. +AliasAnalysis::AliasResult +BasicAliasAnalysis::aliasSelect(const SelectInst *SI, unsigned SISize, + const Value *V2, unsigned V2Size) { + // If this select has been visited before, we're on a use-def cycle. + // Such cycles are only valid when PHI nodes are involved or in unreachable + // code. The visitPHI function catches cycles containing PHIs, but there + // could still be a cycle without PHIs in unreachable code. + if (!Visited.insert(SI)) + return MayAlias; + + // If the values are Selects with the same condition, we can do a more precise + // check: just check for aliases between the values on corresponding arms. + if (const SelectInst *SI2 = dyn_cast(V2)) + if (SI->getCondition() == SI2->getCondition()) { + AliasResult Alias = + aliasCheck(SI->getTrueValue(), SISize, + SI2->getTrueValue(), V2Size); + if (Alias == MayAlias) + return MayAlias; + AliasResult ThisAlias = + aliasCheck(SI->getFalseValue(), SISize, + SI2->getFalseValue(), V2Size); + if (ThisAlias != Alias) + return MayAlias; + return Alias; } - return false; + + // If both arms of the Select node NoAlias or MustAlias V2, then returns + // NoAlias / MustAlias. Otherwise, returns MayAlias. + AliasResult Alias = + aliasCheck(V2, V2Size, SI->getTrueValue(), SISize); + if (Alias == MayAlias) + return MayAlias; + + // If V2 is visited, the recursive case will have been caught in the + // above aliasCheck call, so these subsequent calls to aliasCheck + // don't need to assume that V2 is being visited recursively. + Visited.erase(V2); + + AliasResult ThisAlias = + aliasCheck(V2, V2Size, SI->getFalseValue(), SISize); + if (ThisAlias != Alias) + return MayAlias; + return Alias; } -/// 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, 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. - if (BasePtr1Ty != BasePtr2Ty) +// aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction +// against another. +AliasAnalysis::AliasResult +BasicAliasAnalysis::aliasPHI(const PHINode *PN, unsigned PNSize, + const Value *V2, unsigned V2Size) { + // The PHI node has already been visited, avoid recursion any further. + if (!Visited.insert(PN)) return MayAlias; - const PointerType *GEPPointerTy = cast(BasePtr1Ty); - - // Find the (possibly empty) initial sequence of equal values... which are not - // necessarily constants. - unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops; - unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands); - unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands); - unsigned UnequalOper = 0; - while (UnequalOper != MinOperands && - IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) { - // Advance through the type as we go... - ++UnequalOper; - if (const CompositeType *CT = dyn_cast(BasePtr1Ty)) - BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]); - else { - // If all operands equal each other, then the derived pointers must - // alias each other... - BasePtr1Ty = 0; - assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands && - "Ran out of type nesting, but not out of operands?"); - return MustAlias; + // If the values are PHIs in the same block, we can do a more precise + // as well as efficient check: just check for aliases between the values + // on corresponding edges. + if (const PHINode *PN2 = dyn_cast(V2)) + if (PN2->getParent() == PN->getParent()) { + AliasResult Alias = + aliasCheck(PN->getIncomingValue(0), PNSize, + PN2->getIncomingValueForBlock(PN->getIncomingBlock(0)), + V2Size); + if (Alias == MayAlias) + return MayAlias; + for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) { + AliasResult ThisAlias = + aliasCheck(PN->getIncomingValue(i), PNSize, + PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)), + V2Size); + if (ThisAlias != Alias) + return MayAlias; + } + return Alias; } + + SmallPtrSet UniqueSrc; + SmallVector V1Srcs; + for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { + Value *PV1 = PN->getIncomingValue(i); + if (isa(PV1)) + // If any of the source itself is a PHI, return MayAlias conservatively + // to avoid compile time explosion. The worst possible case is if both + // sides are PHI nodes. In which case, this is O(m x n) time where 'm' + // and 'n' are the number of PHI sources. + return MayAlias; + if (UniqueSrc.insert(PV1)) + V1Srcs.push_back(PV1); } - // If we have seen all constant operands, and run out of indexes on one of the - // getelementptrs, check to see if the tail of the leftover one is all zeros. - // If so, return mustalias. - if (UnequalOper == MinOperands) { - if (NumGEP1Ops < NumGEP2Ops) { - std::swap(GEP1Ops, GEP2Ops); - std::swap(NumGEP1Ops, NumGEP2Ops); - } + AliasResult Alias = aliasCheck(V2, V2Size, V1Srcs[0], PNSize); + // Early exit if the check of the first PHI source against V2 is MayAlias. + // Other results are not possible. + if (Alias == MayAlias) + return MayAlias; - bool AllAreZeros = true; - for (unsigned i = UnequalOper; i != MaxOperands; ++i) - if (!isa(GEP1Ops[i]) || - !cast(GEP1Ops[i])->isNullValue()) { - AllAreZeros = false; - break; - } - if (AllAreZeros) return MustAlias; - } + // If all sources of the PHI node NoAlias or MustAlias V2, then returns + // NoAlias / MustAlias. Otherwise, returns MayAlias. + for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) { + Value *V = V1Srcs[i]; + // If V2 is visited, the recursive case will have been caught in the + // above aliasCheck call, so these subsequent calls to aliasCheck + // don't need to assume that V2 is being visited recursively. + Visited.erase(V2); - // So now we know that the indexes derived from the base pointers, - // which are known to alias, are different. We can still determine a - // no-alias result if there are differing constant pairs in the index - // 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. - - // Scan for the first operand that is constant and unequal in the - // two getelementptrs... - unsigned FirstConstantOper = UnequalOper; - for (; FirstConstantOper != MinOperands; ++FirstConstantOper) { - const Value *G1Oper = GEP1Ops[FirstConstantOper]; - const Value *G2Oper = GEP2Ops[FirstConstantOper]; - - if (G1Oper != G2Oper) // Found non-equal constant indexes... - if (Constant *G1OC = dyn_cast(const_cast(G1Oper))) - if (Constant *G2OC = dyn_cast(const_cast(G2Oper))){ - if (G1OC->getType() != G2OC->getType()) { - // Sign extend both operands to long. - 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. - 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; - } - } - } - } - BasePtr1Ty = cast(BasePtr1Ty)->getTypeAtIndex(G1Oper); + AliasResult ThisAlias = aliasCheck(V2, V2Size, V, PNSize); + if (ThisAlias != Alias || ThisAlias == MayAlias) + return MayAlias; } - // No shared constant operands, and we ran out of common operands. At this - // point, the GEP instructions have run through all of their operands, and we - // haven't found evidence that there are any deltas between the GEP's. - // However, one GEP may have more operands than the other. If this is the - // 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 (NumGEP1Ops < NumGEP2Ops) { - std::swap(GEP1Ops, GEP2Ops); - std::swap(NumGEP1Ops, NumGEP2Ops); - } + return Alias; +} - // Is there anything to check? - if (NumGEP1Ops > MinOperands) { - for (unsigned i = FirstConstantOper; i != MaxOperands; ++i) - 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 suitable for - // TargetData::getIndexedOffset. - for (i = 0; i != MaxOperands; ++i) - if (!isa(GEP1Ops[i])) - GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType()); - // 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, - NumGEP1Ops); - - // Now check without any constants at the end. - int64_t Offset2 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops, - MinOperands); - - // 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; - } - } +// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases, +// such as array references. +// +AliasAnalysis::AliasResult +BasicAliasAnalysis::aliasCheck(const Value *V1, unsigned V1Size, + const Value *V2, unsigned V2Size) { + // If either of the memory references is empty, it doesn't matter what the + // pointer values are. + if (V1Size == 0 || V2Size == 0) + return NoAlias; - // Couldn't find anything useful. - return MayAlias; - } + // Strip off any casts if they exist. + V1 = V1->stripPointerCasts(); + V2 = V2->stripPointerCasts(); - // If there are non-equal constants arguments, then we can figure - // out a minimum known delta between the two index expressions... at - // this point we know that the first constant index of GEP1 is less - // 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]); - - // 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 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]); - } + // Are we checking for alias of the same value? + if (V1 == V2) return MustAlias; - // 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 < 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)) - GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType()); - // Otherwise, just keep the constants we have. - } else { - 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->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 - // constant is used in this position. This cannot be the initial - // index to the GEP instructions (because we know we have at least one - // element before this one with the different constant arguments), so - // we know that the current index must be into either a struct or - // array. Because we know it's not constant, this cannot be a - // structure index. Because of this, we can calculate the maximum - // value possible. - // - if (const ArrayType *AT = dyn_cast(BasePtr1Ty)) - 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 (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy()) + return NoAlias; // Scalars cannot alias each other - 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(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()); - } - } - } + // Figure out what objects these things are pointing to if we can. + const Value *O1 = V1->getUnderlyingObject(); + const Value *O2 = V2->getUnderlyingObject(); - if (BasePtr1Ty && Op1) { - if (const CompositeType *CT = dyn_cast(BasePtr1Ty)) - BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]); - else - BasePtr1Ty = 0; - } + // Null values in the default address space don't point to any object, so they + // don't alias any other pointer. + if (const ConstantPointerNull *CPN = dyn_cast(O1)) + if (CPN->getType()->getAddressSpace() == 0) + return NoAlias; + if (const ConstantPointerNull *CPN = dyn_cast(O2)) + if (CPN->getType()->getAddressSpace() == 0) + return NoAlias; - if (BasePtr2Ty && Op2) { - if (const CompositeType *CT = dyn_cast(BasePtr2Ty)) - BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]); - else - BasePtr2Ty = 0; - } + 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; + + // Constant pointers can't alias with non-const isIdentifiedObject objects. + if ((isa(O1) && isIdentifiedObject(O2) && !isa(O2)) || + (isa(O2) && isIdentifiedObject(O1) && !isa(O1))) + return NoAlias; + + // Arguments can't alias with local allocations or noalias calls + // in the same function. + if (((isa(O1) && (isa(O2) || isNoAliasCall(O2))) || + (isa(O2) && (isa(O1) || isNoAliasCall(O1))))) + return NoAlias; + + // Most objects can't alias null. + if ((isa(O2) && isKnownNonNull(O1)) || + (isa(O1) && isKnownNonNull(O2))) + return NoAlias; + + // If one pointer is the result of a call/invoke or load and the other is a + // non-escaping local object within the same function, then we know the + // object couldn't escape to a point where the call could return it. + // + // Note that if the pointers are in different functions, there are a + // variety of complications. A call with a nocapture argument may still + // temporary store the nocapture argument's value in a temporary memory + // location if that memory location doesn't escape. Or it may pass a + // nocapture value to other functions as long as they don't capture it. + if (isEscapeSource(O1) && isNonEscapingLocalObject(O2)) + return NoAlias; + if (isEscapeSource(O2) && isNonEscapingLocalObject(O1)) + return NoAlias; } - if (GEPPointerTy->getElementType()->isSized()) { - int64_t Offset1 = - getTargetData().getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops); - int64_t Offset2 = - getTargetData().getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops); - assert(Offset1 != Offset2 && - "There is at least one different constant here!"); - - // Make sure we compare the absolute difference. - if (Offset1 > Offset2) - std::swap(Offset1, Offset2); - - if ((uint64_t)(Offset2-Offset1) >= SizeMax) { - //cerr << "Determined that these two GEP's don't alias [" - // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2; + // 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. + if (TD) + if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, *TD)) || + (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, *TD))) return NoAlias; - } + + // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the + // GEP can't simplify, we don't even look at the PHI cases. + if (!isa(V1) && isa(V2)) { + std::swap(V1, V2); + std::swap(V1Size, V2Size); + std::swap(O1, O2); } + if (const GEPOperator *GV1 = dyn_cast(V1)) + return aliasGEP(GV1, V1Size, V2, V2Size, O1, O2); + + if (isa(V2) && !isa(V1)) { + std::swap(V1, V2); + std::swap(V1Size, V2Size); + } + if (const PHINode *PN = dyn_cast(V1)) + return aliasPHI(PN, V1Size, V2, V2Size); + + if (isa(V2) && !isa(V1)) { + std::swap(V1, V2); + std::swap(V1Size, V2Size); + } + if (const SelectInst *S1 = dyn_cast(V1)) + return aliasSelect(S1, V1Size, V2, V2Size); + return MayAlias; } -// Make sure that anything that uses AliasAnalysis pulls in this file... +// Make sure that anything that uses AliasAnalysis pulls in this file. DEFINING_FILE_FOR(BasicAliasAnalysis)