X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FBasicAliasAnalysis.cpp;h=fe71f04098b13061a0b04c4dcf2a99bc3fd7f9dd;hb=002ec1482c83003c6aef841395271aa5993d916a;hp=be8fe07f2c97c245c5c8a79ab6b6e7f671d62477;hpb=d087480166d8f0ec2a44732c7ec087f4eb1ba527;p=oota-llvm.git diff --git a/lib/Analysis/BasicAliasAnalysis.cpp b/lib/Analysis/BasicAliasAnalysis.cpp index be8fe07f2c9..fe71f04098b 100644 --- a/lib/Analysis/BasicAliasAnalysis.cpp +++ b/lib/Analysis/BasicAliasAnalysis.cpp @@ -14,14 +14,14 @@ //===----------------------------------------------------------------------===// #include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Analysis/CaptureTracking.h" #include "llvm/Analysis/Passes.h" #include "llvm/Constants.h" #include "llvm/DerivedTypes.h" #include "llvm/Function.h" -#include "llvm/ParameterAttributes.h" #include "llvm/GlobalVariable.h" #include "llvm/Instructions.h" -#include "llvm/Intrinsics.h" +#include "llvm/IntrinsicInst.h" #include "llvm/Pass.h" #include "llvm/Target/TargetData.h" #include "llvm/ADT/SmallVector.h" @@ -32,6 +32,105 @@ #include using namespace llvm; +//===----------------------------------------------------------------------===// +// Useful predicates +//===----------------------------------------------------------------------===// + +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; +} + +/// 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) || isNoAliasCall(V)) + return !PointerMayBeCaptured(V, false); + + // If this is an argument that corresponds to a byval or noalias argument, + // 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()) { + // Don't bother analyzing arguments already known not to escape. + if (A->hasNoCaptureAttr()) + return true; + return !PointerMayBeCaptured(V, false); + } + 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; + if (const GlobalVariable *GV = dyn_cast(V)) { + AccessTy = GV->getType()->getElementType(); + } else if (const AllocationInst *AI = dyn_cast(V)) { + if (!AI->isArrayAllocation()) + AccessTy = AI->getType()->getElementType(); + else + return false; + } else if (const Argument *A = dyn_cast(V)) { + if (A->hasByValAttr()) + AccessTy = cast(A->getType())->getElementType(); + else + return false; + } else { + return false; + } + + if (AccessTy->isSized()) + return TD.getTypePaddedSize(AccessTy) < Size; + return false; +} + +//===----------------------------------------------------------------------===// +// NoAA Pass +//===----------------------------------------------------------------------===// + namespace { /// NoAA - This class implements the -no-aa pass, which always returns "I /// don't know" for alias queries. NoAA is unlike other alias analysis @@ -40,8 +139,8 @@ namespace { /// struct VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis { static char ID; // Class identification, replacement for typeinfo - NoAA() : ImmutablePass((intptr_t)&ID) {} - explicit NoAA(intptr_t PID) : ImmutablePass(PID) { } + NoAA() : ImmutablePass(&ID) {} + explicit NoAA(void *PID) : ImmutablePass(PID) { } virtual void getAnalysisUsage(AnalysisUsage &AU) const { AU.addRequired(); @@ -56,11 +155,6 @@ 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!"); @@ -79,32 +173,34 @@ namespace { virtual void deleteValue(Value *V) {} virtual void copyValue(Value *From, Value *To) {} }; +} // End of anonymous namespace - // Register this pass... - char NoAA::ID = 0; - RegisterPass - 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(U); -} // 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 VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA { static char ID; // Class identification, replacement for typeinfo - BasicAliasAnalysis() : NoAA((intptr_t)&ID) { } + BasicAliasAnalysis() : NoAA(&ID) {} AliasResult alias(const Value *V1, unsigned V1Size, const Value *V2, unsigned V2Size); ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size); - ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) { - return NoAA::getModRefInfo(CS1,CS2); - } + ModRefResult getModRefInfo(CallSite CS1, CallSite CS2); /// hasNoModRefInfoForCalls - We can provide mod/ref information against /// non-escaping allocations. @@ -124,116 +220,30 @@ namespace { const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size); }; +} // End of anonymous namespace - // Register this pass... - char BasicAliasAnalysis::ID = 0; - RegisterPass - 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(X); -} // End of anonymous namespace +// Declare that we implement the AliasAnalysis interface +static RegisterAnalysisGroup Y(X); ImmutablePass *llvm::createBasicAliasAnalysisPass() { return new BasicAliasAnalysis(); } -/// 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. This returns: -/// Arguments, GlobalVariables, Functions, Allocas, Mallocs. -static const Value *getUnderlyingObject(const Value *V) { - if (!isa(V->getType())) return 0; - - // If we are at some type of object, return it. GlobalValues and Allocations - // have unique addresses. - if (isa(V) || isa(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::BitCast || - CE->getOpcode() == Instruction::GetElementPtr) - return getUnderlyingObject(CE->getOperand(0)); - } - 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, - 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; -} /// 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(); + if (const GlobalVariable *GV = + dyn_cast(P->getUnderlyingObject())) + return GV->isConstant(); return false; } -// 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 (!isa(I->getType())) - return true; - 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 - default: - return true; - } - } - 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 @@ -243,22 +253,32 @@ static bool AddressMightEscape(const Value *V) { AliasAnalysis::ModRefResult BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) { if (!isa(P)) { - const Value *Object = getUnderlyingObject(P); - // Allocations and byval arguments are "new" objects. - if (Object && - (isa(Object) || - (isa(Object) && cast(Object)->hasByValAttr()))) { - // 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(Object)) - 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. + 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(Object)) + 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) && CS.getInstruction() != 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; } } @@ -266,9 +286,27 @@ BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) { return AliasAnalysis::getModRefInfo(CS, P, Size); } + +AliasAnalysis::ModRefResult +BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) { + // If CS1 or CS2 are readnone, they don't interact. + ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1); + if (CS1B == DoesNotAccessMemory) return NoModRef; + + ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2); + if (CS2B == DoesNotAccessMemory) return NoModRef; + + // If they both only read from memory, just return ref. + if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory) + return Ref; + + // Otherwise, fall back to NoAA (mod+ref). + return NoAA::getModRefInfo(CS1, CS2); +} + + // 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. :) +// as array references. // AliasAnalysis::AliasResult BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, @@ -284,127 +322,88 @@ BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, // Are we checking for alias of the same value? if (V1 == V2) return MustAlias; - if ((!isa(V1->getType()) || !isa(V2->getType())) && - V1->getType() != Type::Int64Ty && V2->getType() != Type::Int64Ty) + if (!isa(V1->getType()) || !isa(V2->getType())) return NoAlias; // Scalars cannot alias each other - // Strip off cast instructions... + // 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 = getUnderlyingObject(V1); - const Value *O2 = getUnderlyingObject(V2); - - // Pointing at a discernible object? - if (O1) { - if (O2) { - if (const Argument *O1Arg = dyn_cast(O1)) { - // Incoming argument cannot alias locally allocated object! - if (isa(O2)) return NoAlias; - - // If they are two different objects, and one is a noalias argument - // then they do not alias. - if (O1 != O2 && O1Arg->hasNoAliasAttr()) - return NoAlias; - - // Byval arguments can't alias globals or other arguments. - if (O1 != O2 && O1Arg->hasByValAttr()) return NoAlias; - - // Otherwise, nothing is known... - } - - if (const Argument *O2Arg = dyn_cast(O2)) { - // Incoming argument cannot alias locally allocated object! - if (isa(O1)) return NoAlias; - - // If they are two different objects, and one is a noalias argument - // then they do not alias. - if (O1 != O2 && O2Arg->hasNoAliasAttr()) - return NoAlias; - - // Byval arguments can't alias globals or other arguments. - if (O1 != O2 && O2Arg->hasByValAttr()) return NoAlias; - - // Otherwise, nothing is known... - - } else if (O1 != O2 && !isa(O1)) { - // If they are two different objects, and neither is an argument, - // we know that we have no alias. - return NoAlias; - } - - // 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. - } - - // Unique values don't alias null, except non-byval arguments. - if (isa(V2)) { - if (const Argument *O1Arg = dyn_cast(O1)) { - if (O1Arg->hasByValAttr()) - return NoAlias; - } else { - return NoAlias; - } - } + // Figure out what objects these things are pointing to if we can. + const Value *O1 = V1->getUnderlyingObject(); + const Value *O2 = V2->getUnderlyingObject(); - 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().getABITypeSize(ElTy); - if (GlobalSize < V2Size && V2Size != ~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; + + // 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; - 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().getABITypeSize(ElTy); - if (GlobalSize < V1Size && V1Size != ~0U) - 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) && O1 != O2) + return NoAlias; + if ((isa(O2) || isa(O2)) && + isNonEscapingLocalObject(O1) && O1 != O2) + 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 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); + + // Drill down into the first non-gep value, to test for must-aliasing of // the base pointers. - 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); + 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); // Do the base pointers alias? AliasResult BaseAlias = alias(BasePtr1, ~0U, BasePtr2, ~0U); @@ -498,7 +497,7 @@ BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, return MayAlias; } -// This function is used to determin if the indices of two GEP instructions are +// 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()) @@ -666,7 +665,7 @@ BasicAliasAnalysis::CheckGEPInstructions( 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])) @@ -681,9 +680,15 @@ BasicAliasAnalysis::CheckGEPInstructions( 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; } }