1 //===- BasicAliasAnalysis.cpp - Stateless Alias Analysis Impl -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the primary stateless implementation of the
11 // Alias Analysis interface that implements identities (two different
12 // globals cannot alias, etc), but does no stateful analysis.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Analysis/AliasAnalysis.h"
17 #include "llvm/Analysis/Passes.h"
18 #include "llvm/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Function.h"
21 #include "llvm/GlobalAlias.h"
22 #include "llvm/GlobalVariable.h"
23 #include "llvm/Instructions.h"
24 #include "llvm/IntrinsicInst.h"
25 #include "llvm/LLVMContext.h"
26 #include "llvm/Operator.h"
27 #include "llvm/Pass.h"
28 #include "llvm/Analysis/CaptureTracking.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/InstructionSimplify.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/SmallVector.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/GetElementPtrTypeIterator.h"
40 //===----------------------------------------------------------------------===//
42 //===----------------------------------------------------------------------===//
44 /// isKnownNonNull - Return true if we know that the specified value is never
46 static bool isKnownNonNull(const Value *V) {
47 // Alloca never returns null, malloc might.
48 if (isa<AllocaInst>(V)) return true;
50 // A byval argument is never null.
51 if (const Argument *A = dyn_cast<Argument>(V))
52 return A->hasByValAttr();
54 // Global values are not null unless extern weak.
55 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
56 return !GV->hasExternalWeakLinkage();
60 /// isNonEscapingLocalObject - Return true if the pointer is to a function-local
61 /// object that never escapes from the function.
62 static bool isNonEscapingLocalObject(const Value *V) {
63 // If this is a local allocation, check to see if it escapes.
64 if (isa<AllocaInst>(V) || isNoAliasCall(V))
65 // Set StoreCaptures to True so that we can assume in our callers that the
66 // pointer is not the result of a load instruction. Currently
67 // PointerMayBeCaptured doesn't have any special analysis for the
68 // StoreCaptures=false case; if it did, our callers could be refined to be
70 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
72 // If this is an argument that corresponds to a byval or noalias argument,
73 // then it has not escaped before entering the function. Check if it escapes
74 // inside the function.
75 if (const Argument *A = dyn_cast<Argument>(V))
76 if (A->hasByValAttr() || A->hasNoAliasAttr()) {
77 // Don't bother analyzing arguments already known not to escape.
78 if (A->hasNoCaptureAttr())
80 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
85 /// isEscapeSource - Return true if the pointer is one which would have
86 /// been considered an escape by isNonEscapingLocalObject.
87 static bool isEscapeSource(const Value *V) {
88 if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
91 // The load case works because isNonEscapingLocalObject considers all
92 // stores to be escapes (it passes true for the StoreCaptures argument
93 // to PointerMayBeCaptured).
100 /// getObjectSize - Return the size of the object specified by V, or
101 /// UnknownSize if unknown.
102 static uint64_t getObjectSize(const Value *V, const TargetData &TD) {
103 const Type *AccessTy;
104 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
105 if (!GV->hasDefinitiveInitializer())
106 return AliasAnalysis::UnknownSize;
107 AccessTy = GV->getType()->getElementType();
108 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
109 if (!AI->isArrayAllocation())
110 AccessTy = AI->getType()->getElementType();
112 return AliasAnalysis::UnknownSize;
113 } else if (const CallInst* CI = extractMallocCall(V)) {
114 if (!isArrayMalloc(V, &TD))
115 // The size is the argument to the malloc call.
116 if (const ConstantInt* C = dyn_cast<ConstantInt>(CI->getArgOperand(0)))
117 return C->getZExtValue();
118 return AliasAnalysis::UnknownSize;
119 } else if (const Argument *A = dyn_cast<Argument>(V)) {
120 if (A->hasByValAttr())
121 AccessTy = cast<PointerType>(A->getType())->getElementType();
123 return AliasAnalysis::UnknownSize;
125 return AliasAnalysis::UnknownSize;
128 if (AccessTy->isSized())
129 return TD.getTypeAllocSize(AccessTy);
130 return AliasAnalysis::UnknownSize;
133 /// isObjectSmallerThan - Return true if we can prove that the object specified
134 /// by V is smaller than Size.
135 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
136 const TargetData &TD) {
137 uint64_t ObjectSize = getObjectSize(V, TD);
138 return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize < Size;
141 /// isObjectSize - Return true if we can prove that the object specified
142 /// by V has size Size.
143 static bool isObjectSize(const Value *V, uint64_t Size,
144 const TargetData &TD) {
145 uint64_t ObjectSize = getObjectSize(V, TD);
146 return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize == Size;
149 //===----------------------------------------------------------------------===//
150 // GetElementPtr Instruction Decomposition and Analysis
151 //===----------------------------------------------------------------------===//
160 struct VariableGEPIndex {
162 ExtensionKind Extension;
168 /// GetLinearExpression - Analyze the specified value as a linear expression:
169 /// "A*V + B", where A and B are constant integers. Return the scale and offset
170 /// values as APInts and return V as a Value*, and return whether we looked
171 /// through any sign or zero extends. The incoming Value is known to have
172 /// IntegerType and it may already be sign or zero extended.
174 /// Note that this looks through extends, so the high bits may not be
175 /// represented in the result.
176 static Value *GetLinearExpression(Value *V, APInt &Scale, APInt &Offset,
177 ExtensionKind &Extension,
178 const TargetData &TD, unsigned Depth) {
179 assert(V->getType()->isIntegerTy() && "Not an integer value");
181 // Limit our recursion depth.
188 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
189 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
190 switch (BOp->getOpcode()) {
192 case Instruction::Or:
193 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
195 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), &TD))
198 case Instruction::Add:
199 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
201 Offset += RHSC->getValue();
203 case Instruction::Mul:
204 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
206 Offset *= RHSC->getValue();
207 Scale *= RHSC->getValue();
209 case Instruction::Shl:
210 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
212 Offset <<= RHSC->getValue().getLimitedValue();
213 Scale <<= RHSC->getValue().getLimitedValue();
219 // Since GEP indices are sign extended anyway, we don't care about the high
220 // bits of a sign or zero extended value - just scales and offsets. The
221 // extensions have to be consistent though.
222 if ((isa<SExtInst>(V) && Extension != EK_ZeroExt) ||
223 (isa<ZExtInst>(V) && Extension != EK_SignExt)) {
224 Value *CastOp = cast<CastInst>(V)->getOperand(0);
225 unsigned OldWidth = Scale.getBitWidth();
226 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
227 Scale = Scale.trunc(SmallWidth);
228 Offset = Offset.trunc(SmallWidth);
229 Extension = isa<SExtInst>(V) ? EK_SignExt : EK_ZeroExt;
231 Value *Result = GetLinearExpression(CastOp, Scale, Offset, Extension,
233 Scale = Scale.zext(OldWidth);
234 Offset = Offset.zext(OldWidth);
244 /// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it
245 /// into a base pointer with a constant offset and a number of scaled symbolic
248 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale in
249 /// the VarIndices vector) are Value*'s that are known to be scaled by the
250 /// specified amount, but which may have other unrepresented high bits. As such,
251 /// the gep cannot necessarily be reconstructed from its decomposed form.
253 /// When TargetData is around, this function is capable of analyzing everything
254 /// that GetUnderlyingObject can look through. When not, it just looks
255 /// through pointer casts.
258 DecomposeGEPExpression(const Value *V, int64_t &BaseOffs,
259 SmallVectorImpl<VariableGEPIndex> &VarIndices,
260 const TargetData *TD) {
261 // Limit recursion depth to limit compile time in crazy cases.
262 unsigned MaxLookup = 6;
266 // See if this is a bitcast or GEP.
267 const Operator *Op = dyn_cast<Operator>(V);
269 // The only non-operator case we can handle are GlobalAliases.
270 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
271 if (!GA->mayBeOverridden()) {
272 V = GA->getAliasee();
279 if (Op->getOpcode() == Instruction::BitCast) {
280 V = Op->getOperand(0);
284 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
286 // If it's not a GEP, hand it off to SimplifyInstruction to see if it
287 // can come up with something. This matches what GetUnderlyingObject does.
288 if (const Instruction *I = dyn_cast<Instruction>(V))
289 // TODO: Get a DominatorTree and use it here.
290 if (const Value *Simplified =
291 SimplifyInstruction(const_cast<Instruction *>(I), TD)) {
299 // Don't attempt to analyze GEPs over unsized objects.
300 if (!cast<PointerType>(GEPOp->getOperand(0)->getType())
301 ->getElementType()->isSized())
304 // If we are lacking TargetData information, we can't compute the offets of
305 // elements computed by GEPs. However, we can handle bitcast equivalent
308 if (!GEPOp->hasAllZeroIndices())
310 V = GEPOp->getOperand(0);
314 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
315 gep_type_iterator GTI = gep_type_begin(GEPOp);
316 for (User::const_op_iterator I = GEPOp->op_begin()+1,
317 E = GEPOp->op_end(); I != E; ++I) {
319 // Compute the (potentially symbolic) offset in bytes for this index.
320 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
321 // For a struct, add the member offset.
322 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
323 if (FieldNo == 0) continue;
325 BaseOffs += TD->getStructLayout(STy)->getElementOffset(FieldNo);
329 // For an array/pointer, add the element offset, explicitly scaled.
330 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
331 if (CIdx->isZero()) continue;
332 BaseOffs += TD->getTypeAllocSize(*GTI)*CIdx->getSExtValue();
336 uint64_t Scale = TD->getTypeAllocSize(*GTI);
337 ExtensionKind Extension = EK_NotExtended;
339 // If the integer type is smaller than the pointer size, it is implicitly
340 // sign extended to pointer size.
341 unsigned Width = cast<IntegerType>(Index->getType())->getBitWidth();
342 if (TD->getPointerSizeInBits() > Width)
343 Extension = EK_SignExt;
345 // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
346 APInt IndexScale(Width, 0), IndexOffset(Width, 0);
347 Index = GetLinearExpression(Index, IndexScale, IndexOffset, Extension,
350 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
351 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
352 BaseOffs += IndexOffset.getSExtValue()*Scale;
353 Scale *= IndexScale.getSExtValue();
356 // If we already had an occurrence of this index variable, merge this
357 // scale into it. For example, we want to handle:
358 // A[x][x] -> x*16 + x*4 -> x*20
359 // This also ensures that 'x' only appears in the index list once.
360 for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
361 if (VarIndices[i].V == Index &&
362 VarIndices[i].Extension == Extension) {
363 Scale += VarIndices[i].Scale;
364 VarIndices.erase(VarIndices.begin()+i);
369 // Make sure that we have a scale that makes sense for this target's
371 if (unsigned ShiftBits = 64-TD->getPointerSizeInBits()) {
373 Scale = (int64_t)Scale >> ShiftBits;
377 VariableGEPIndex Entry = {Index, Extension, Scale};
378 VarIndices.push_back(Entry);
382 // Analyze the base pointer next.
383 V = GEPOp->getOperand(0);
384 } while (--MaxLookup);
386 // If the chain of expressions is too deep, just return early.
390 /// GetIndexDifference - Dest and Src are the variable indices from two
391 /// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base
392 /// pointers. Subtract the GEP2 indices from GEP1 to find the symbolic
393 /// difference between the two pointers.
394 static void GetIndexDifference(SmallVectorImpl<VariableGEPIndex> &Dest,
395 const SmallVectorImpl<VariableGEPIndex> &Src) {
396 if (Src.empty()) return;
398 for (unsigned i = 0, e = Src.size(); i != e; ++i) {
399 const Value *V = Src[i].V;
400 ExtensionKind Extension = Src[i].Extension;
401 int64_t Scale = Src[i].Scale;
403 // Find V in Dest. This is N^2, but pointer indices almost never have more
404 // than a few variable indexes.
405 for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
406 if (Dest[j].V != V || Dest[j].Extension != Extension) continue;
408 // If we found it, subtract off Scale V's from the entry in Dest. If it
409 // goes to zero, remove the entry.
410 if (Dest[j].Scale != Scale)
411 Dest[j].Scale -= Scale;
413 Dest.erase(Dest.begin()+j);
418 // If we didn't consume this entry, add it to the end of the Dest list.
420 VariableGEPIndex Entry = { V, Extension, -Scale };
421 Dest.push_back(Entry);
426 //===----------------------------------------------------------------------===//
427 // BasicAliasAnalysis Pass
428 //===----------------------------------------------------------------------===//
431 static const Function *getParent(const Value *V) {
432 if (const Instruction *inst = dyn_cast<Instruction>(V))
433 return inst->getParent()->getParent();
435 if (const Argument *arg = dyn_cast<Argument>(V))
436 return arg->getParent();
441 static bool notDifferentParent(const Value *O1, const Value *O2) {
443 const Function *F1 = getParent(O1);
444 const Function *F2 = getParent(O2);
446 return !F1 || !F2 || F1 == F2;
451 /// BasicAliasAnalysis - This is the primary alias analysis implementation.
452 struct BasicAliasAnalysis : public ImmutablePass, public AliasAnalysis {
453 static char ID; // Class identification, replacement for typeinfo
454 BasicAliasAnalysis() : ImmutablePass(ID) {
455 initializeBasicAliasAnalysisPass(*PassRegistry::getPassRegistry());
458 virtual void initializePass() {
459 InitializeAliasAnalysis(this);
462 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
463 AU.addRequired<AliasAnalysis>();
466 virtual AliasResult alias(const Location &LocA,
467 const Location &LocB) {
468 assert(AliasCache.empty() && "AliasCache must be cleared after use!");
469 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
470 "BasicAliasAnalysis doesn't support interprocedural queries.");
471 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.TBAATag,
472 LocB.Ptr, LocB.Size, LocB.TBAATag);
477 virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
478 const Location &Loc);
480 virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
481 ImmutableCallSite CS2) {
482 // The AliasAnalysis base class has some smarts, lets use them.
483 return AliasAnalysis::getModRefInfo(CS1, CS2);
486 /// pointsToConstantMemory - Chase pointers until we find a (constant
488 virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
490 /// getModRefBehavior - Return the behavior when calling the given
492 virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
494 /// getModRefBehavior - Return the behavior when calling the given function.
495 /// For use when the call site is not known.
496 virtual ModRefBehavior getModRefBehavior(const Function *F);
498 /// getAdjustedAnalysisPointer - This method is used when a pass implements
499 /// an analysis interface through multiple inheritance. If needed, it
500 /// should override this to adjust the this pointer as needed for the
501 /// specified pass info.
502 virtual void *getAdjustedAnalysisPointer(const void *ID) {
503 if (ID == &AliasAnalysis::ID)
504 return (AliasAnalysis*)this;
509 // AliasCache - Track alias queries to guard against recursion.
510 typedef std::pair<Location, Location> LocPair;
511 typedef DenseMap<LocPair, AliasResult> AliasCacheTy;
512 AliasCacheTy AliasCache;
514 // Visited - Track instructions visited by pointsToConstantMemory.
515 SmallPtrSet<const Value*, 16> Visited;
517 // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP
518 // instruction against another.
519 AliasResult aliasGEP(const GEPOperator *V1, uint64_t V1Size,
520 const Value *V2, uint64_t V2Size,
521 const MDNode *V2TBAAInfo,
522 const Value *UnderlyingV1, const Value *UnderlyingV2);
524 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI
525 // instruction against another.
526 AliasResult aliasPHI(const PHINode *PN, uint64_t PNSize,
527 const MDNode *PNTBAAInfo,
528 const Value *V2, uint64_t V2Size,
529 const MDNode *V2TBAAInfo);
531 /// aliasSelect - Disambiguate a Select instruction against another value.
532 AliasResult aliasSelect(const SelectInst *SI, uint64_t SISize,
533 const MDNode *SITBAAInfo,
534 const Value *V2, uint64_t V2Size,
535 const MDNode *V2TBAAInfo);
537 AliasResult aliasCheck(const Value *V1, uint64_t V1Size,
538 const MDNode *V1TBAATag,
539 const Value *V2, uint64_t V2Size,
540 const MDNode *V2TBAATag);
542 } // End of anonymous namespace
544 // Register this pass...
545 char BasicAliasAnalysis::ID = 0;
546 INITIALIZE_AG_PASS(BasicAliasAnalysis, AliasAnalysis, "basicaa",
547 "Basic Alias Analysis (stateless AA impl)",
550 ImmutablePass *llvm::createBasicAliasAnalysisPass() {
551 return new BasicAliasAnalysis();
554 /// pointsToConstantMemory - Returns whether the given pointer value
555 /// points to memory that is local to the function, with global constants being
556 /// considered local to all functions.
558 BasicAliasAnalysis::pointsToConstantMemory(const Location &Loc, bool OrLocal) {
559 assert(Visited.empty() && "Visited must be cleared after use!");
561 unsigned MaxLookup = 8;
562 SmallVector<const Value *, 16> Worklist;
563 Worklist.push_back(Loc.Ptr);
565 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), TD);
566 if (!Visited.insert(V)) {
568 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
571 // An alloca instruction defines local memory.
572 if (OrLocal && isa<AllocaInst>(V))
575 // A global constant counts as local memory for our purposes.
576 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
577 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
578 // global to be marked constant in some modules and non-constant in
579 // others. GV may even be a declaration, not a definition.
580 if (!GV->isConstant()) {
582 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
587 // If both select values point to local memory, then so does the select.
588 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
589 Worklist.push_back(SI->getTrueValue());
590 Worklist.push_back(SI->getFalseValue());
594 // If all values incoming to a phi node point to local memory, then so does
596 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
597 // Don't bother inspecting phi nodes with many operands.
598 if (PN->getNumIncomingValues() > MaxLookup) {
600 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
602 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
603 Worklist.push_back(PN->getIncomingValue(i));
607 // Otherwise be conservative.
609 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
611 } while (!Worklist.empty() && --MaxLookup);
614 return Worklist.empty();
617 /// getModRefBehavior - Return the behavior when calling the given call site.
618 AliasAnalysis::ModRefBehavior
619 BasicAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
620 if (CS.doesNotAccessMemory())
621 // Can't do better than this.
622 return DoesNotAccessMemory;
624 ModRefBehavior Min = UnknownModRefBehavior;
626 // If the callsite knows it only reads memory, don't return worse
628 if (CS.onlyReadsMemory())
629 Min = OnlyReadsMemory;
631 // The AliasAnalysis base class has some smarts, lets use them.
632 return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
635 /// getModRefBehavior - Return the behavior when calling the given function.
636 /// For use when the call site is not known.
637 AliasAnalysis::ModRefBehavior
638 BasicAliasAnalysis::getModRefBehavior(const Function *F) {
639 // If the function declares it doesn't access memory, we can't do better.
640 if (F->doesNotAccessMemory())
641 return DoesNotAccessMemory;
643 // For intrinsics, we can check the table.
644 if (unsigned iid = F->getIntrinsicID()) {
645 #define GET_INTRINSIC_MODREF_BEHAVIOR
646 #include "llvm/Intrinsics.gen"
647 #undef GET_INTRINSIC_MODREF_BEHAVIOR
650 ModRefBehavior Min = UnknownModRefBehavior;
652 // If the function declares it only reads memory, go with that.
653 if (F->onlyReadsMemory())
654 Min = OnlyReadsMemory;
656 // Otherwise be conservative.
657 return ModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min);
660 /// getModRefInfo - Check to see if the specified callsite can clobber the
661 /// specified memory object. Since we only look at local properties of this
662 /// function, we really can't say much about this query. We do, however, use
663 /// simple "address taken" analysis on local objects.
664 AliasAnalysis::ModRefResult
665 BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS,
666 const Location &Loc) {
667 assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
668 "AliasAnalysis query involving multiple functions!");
670 const Value *Object = GetUnderlyingObject(Loc.Ptr, TD);
672 // If this is a tail call and Loc.Ptr points to a stack location, we know that
673 // the tail call cannot access or modify the local stack.
674 // We cannot exclude byval arguments here; these belong to the caller of
675 // the current function not to the current function, and a tail callee
676 // may reference them.
677 if (isa<AllocaInst>(Object))
678 if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
679 if (CI->isTailCall())
682 // If the pointer is to a locally allocated object that does not escape,
683 // then the call can not mod/ref the pointer unless the call takes the pointer
684 // as an argument, and itself doesn't capture it.
685 if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
686 isNonEscapingLocalObject(Object)) {
687 bool PassedAsArg = false;
689 for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
690 CI != CE; ++CI, ++ArgNo) {
691 // Only look at the no-capture or byval pointer arguments. If this
692 // pointer were passed to arguments that were neither of these, then it
693 // couldn't be no-capture.
694 if (!(*CI)->getType()->isPointerTy() ||
695 (!CS.paramHasAttr(ArgNo+1, Attribute::NoCapture) &&
696 !CS.paramHasAttr(ArgNo+1, Attribute::ByVal)))
699 // If this is a no-capture pointer argument, see if we can tell that it
700 // is impossible to alias the pointer we're checking. If not, we have to
701 // assume that the call could touch the pointer, even though it doesn't
703 if (!isNoAlias(Location(cast<Value>(CI)), Loc)) {
713 ModRefResult Min = ModRef;
715 // Finally, handle specific knowledge of intrinsics.
716 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
718 switch (II->getIntrinsicID()) {
720 case Intrinsic::memcpy:
721 case Intrinsic::memmove: {
722 uint64_t Len = UnknownSize;
723 if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2)))
724 Len = LenCI->getZExtValue();
725 Value *Dest = II->getArgOperand(0);
726 Value *Src = II->getArgOperand(1);
727 // If it can't overlap the source dest, then it doesn't modref the loc.
728 if (isNoAlias(Location(Dest, Len), Loc)) {
729 if (isNoAlias(Location(Src, Len), Loc))
731 // If it can't overlap the dest, then worst case it reads the loc.
733 } else if (isNoAlias(Location(Src, Len), Loc)) {
734 // If it can't overlap the source, then worst case it mutates the loc.
739 case Intrinsic::memset:
740 // Since memset is 'accesses arguments' only, the AliasAnalysis base class
741 // will handle it for the variable length case.
742 if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
743 uint64_t Len = LenCI->getZExtValue();
744 Value *Dest = II->getArgOperand(0);
745 if (isNoAlias(Location(Dest, Len), Loc))
748 // We know that memset doesn't load anything.
751 case Intrinsic::atomic_cmp_swap:
752 case Intrinsic::atomic_swap:
753 case Intrinsic::atomic_load_add:
754 case Intrinsic::atomic_load_sub:
755 case Intrinsic::atomic_load_and:
756 case Intrinsic::atomic_load_nand:
757 case Intrinsic::atomic_load_or:
758 case Intrinsic::atomic_load_xor:
759 case Intrinsic::atomic_load_max:
760 case Intrinsic::atomic_load_min:
761 case Intrinsic::atomic_load_umax:
762 case Intrinsic::atomic_load_umin:
764 Value *Op1 = II->getArgOperand(0);
765 uint64_t Op1Size = TD->getTypeStoreSize(Op1->getType());
766 MDNode *Tag = II->getMetadata(LLVMContext::MD_tbaa);
767 if (isNoAlias(Location(Op1, Op1Size, Tag), Loc))
771 case Intrinsic::lifetime_start:
772 case Intrinsic::lifetime_end:
773 case Intrinsic::invariant_start: {
775 cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
776 if (isNoAlias(Location(II->getArgOperand(1),
778 II->getMetadata(LLVMContext::MD_tbaa)),
783 case Intrinsic::invariant_end: {
785 cast<ConstantInt>(II->getArgOperand(1))->getZExtValue();
786 if (isNoAlias(Location(II->getArgOperand(2),
788 II->getMetadata(LLVMContext::MD_tbaa)),
793 case Intrinsic::arm_neon_vld1: {
794 // LLVM's vld1 and vst1 intrinsics currently only support a single
797 TD ? TD->getTypeStoreSize(II->getType()) : UnknownSize;
798 if (isNoAlias(Location(II->getArgOperand(0), Size,
799 II->getMetadata(LLVMContext::MD_tbaa)),
804 case Intrinsic::arm_neon_vst1: {
806 TD ? TD->getTypeStoreSize(II->getArgOperand(1)->getType()) : UnknownSize;
807 if (isNoAlias(Location(II->getArgOperand(0), Size,
808 II->getMetadata(LLVMContext::MD_tbaa)),
815 // The AliasAnalysis base class has some smarts, lets use them.
816 return ModRefResult(AliasAnalysis::getModRefInfo(CS, Loc) & Min);
819 /// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
820 /// against another pointer. We know that V1 is a GEP, but we don't know
821 /// anything about V2. UnderlyingV1 is GetUnderlyingObject(GEP1, TD),
822 /// UnderlyingV2 is the same for V2.
824 AliasAnalysis::AliasResult
825 BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
826 const Value *V2, uint64_t V2Size,
827 const MDNode *V2TBAAInfo,
828 const Value *UnderlyingV1,
829 const Value *UnderlyingV2) {
830 int64_t GEP1BaseOffset;
831 SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
833 // If we have two gep instructions with must-alias'ing base pointers, figure
834 // out if the indexes to the GEP tell us anything about the derived pointer.
835 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
836 // Do the base pointers alias?
837 AliasResult BaseAlias = aliasCheck(UnderlyingV1, UnknownSize, 0,
838 UnderlyingV2, UnknownSize, 0);
840 // If we get a No or May, then return it immediately, no amount of analysis
841 // will improve this situation.
842 if (BaseAlias != MustAlias) return BaseAlias;
844 // Otherwise, we have a MustAlias. Since the base pointers alias each other
845 // exactly, see if the computed offset from the common pointer tells us
846 // about the relation of the resulting pointer.
847 const Value *GEP1BasePtr =
848 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
850 int64_t GEP2BaseOffset;
851 SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
852 const Value *GEP2BasePtr =
853 DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD);
855 // If DecomposeGEPExpression isn't able to look all the way through the
856 // addressing operation, we must not have TD and this is too complex for us
857 // to handle without it.
858 if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
860 "DecomposeGEPExpression and GetUnderlyingObject disagree!");
864 // Subtract the GEP2 pointer from the GEP1 pointer to find out their
865 // symbolic difference.
866 GEP1BaseOffset -= GEP2BaseOffset;
867 GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
870 // Check to see if these two pointers are related by the getelementptr
871 // instruction. If one pointer is a GEP with a non-zero index of the other
872 // pointer, we know they cannot alias.
874 // If both accesses are unknown size, we can't do anything useful here.
875 if (V1Size == UnknownSize && V2Size == UnknownSize)
878 AliasResult R = aliasCheck(UnderlyingV1, UnknownSize, 0,
879 V2, V2Size, V2TBAAInfo);
881 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
882 // If V2 is known not to alias GEP base pointer, then the two values
883 // cannot alias per GEP semantics: "A pointer value formed from a
884 // getelementptr instruction is associated with the addresses associated
885 // with the first operand of the getelementptr".
888 const Value *GEP1BasePtr =
889 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
891 // If DecomposeGEPExpression isn't able to look all the way through the
892 // addressing operation, we must not have TD and this is too complex for us
893 // to handle without it.
894 if (GEP1BasePtr != UnderlyingV1) {
896 "DecomposeGEPExpression and GetUnderlyingObject disagree!");
901 // In the two GEP Case, if there is no difference in the offsets of the
902 // computed pointers, the resultant pointers are a must alias. This
903 // hapens when we have two lexically identical GEP's (for example).
905 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
906 // must aliases the GEP, the end result is a must alias also.
907 if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
910 // If there is a difference between the pointers, but the difference is
911 // less than the size of the associated memory object, then we know
912 // that the objects are partially overlapping.
913 if (GEP1BaseOffset != 0 && GEP1VariableIndices.empty()) {
914 if (GEP1BaseOffset >= 0 ?
915 (V2Size != UnknownSize && (uint64_t)GEP1BaseOffset < V2Size) :
916 (V1Size != UnknownSize && -(uint64_t)GEP1BaseOffset < V1Size &&
917 GEP1BaseOffset != INT64_MIN))
921 // If we have a known constant offset, see if this offset is larger than the
922 // access size being queried. If so, and if no variable indices can remove
923 // pieces of this constant, then we know we have a no-alias. For example,
926 // In order to handle cases like &A[100][i] where i is an out of range
927 // subscript, we have to ignore all constant offset pieces that are a multiple
928 // of a scaled index. Do this by removing constant offsets that are a
929 // multiple of any of our variable indices. This allows us to transform
930 // things like &A[i][1] because i has a stride of (e.g.) 8 bytes but the 1
931 // provides an offset of 4 bytes (assuming a <= 4 byte access).
932 for (unsigned i = 0, e = GEP1VariableIndices.size();
933 i != e && GEP1BaseOffset;++i)
934 if (int64_t RemovedOffset = GEP1BaseOffset/GEP1VariableIndices[i].Scale)
935 GEP1BaseOffset -= RemovedOffset*GEP1VariableIndices[i].Scale;
937 // If our known offset is bigger than the access size, we know we don't have
939 if (GEP1BaseOffset) {
940 if (GEP1BaseOffset >= 0 ?
941 (V2Size != UnknownSize && (uint64_t)GEP1BaseOffset >= V2Size) :
942 (V1Size != UnknownSize && -(uint64_t)GEP1BaseOffset >= V1Size &&
943 GEP1BaseOffset != INT64_MIN))
947 // Statically, we can see that the base objects are the same, but the
948 // pointers have dynamic offsets which we can't resolve. And none of our
949 // little tricks above worked.
951 // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the
952 // practical effect of this is protecting TBAA in the case of dynamic
953 // indices into arrays of unions. An alternative way to solve this would
954 // be to have clang emit extra metadata for unions and/or union accesses.
955 // A union-specific solution wouldn't handle the problem for malloc'd
960 static AliasAnalysis::AliasResult
961 MergeAliasResults(AliasAnalysis::AliasResult A, AliasAnalysis::AliasResult B) {
962 // If the results agree, take it.
965 // A mix of PartialAlias and MustAlias is PartialAlias.
966 if ((A == AliasAnalysis::PartialAlias && B == AliasAnalysis::MustAlias) ||
967 (B == AliasAnalysis::PartialAlias && A == AliasAnalysis::MustAlias))
968 return AliasAnalysis::PartialAlias;
969 // Otherwise, we don't know anything.
970 return AliasAnalysis::MayAlias;
973 /// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select
974 /// instruction against another.
975 AliasAnalysis::AliasResult
976 BasicAliasAnalysis::aliasSelect(const SelectInst *SI, uint64_t SISize,
977 const MDNode *SITBAAInfo,
978 const Value *V2, uint64_t V2Size,
979 const MDNode *V2TBAAInfo) {
980 // If the values are Selects with the same condition, we can do a more precise
981 // check: just check for aliases between the values on corresponding arms.
982 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
983 if (SI->getCondition() == SI2->getCondition()) {
985 aliasCheck(SI->getTrueValue(), SISize, SITBAAInfo,
986 SI2->getTrueValue(), V2Size, V2TBAAInfo);
987 if (Alias == MayAlias)
989 AliasResult ThisAlias =
990 aliasCheck(SI->getFalseValue(), SISize, SITBAAInfo,
991 SI2->getFalseValue(), V2Size, V2TBAAInfo);
992 return MergeAliasResults(ThisAlias, Alias);
995 // If both arms of the Select node NoAlias or MustAlias V2, then returns
996 // NoAlias / MustAlias. Otherwise, returns MayAlias.
998 aliasCheck(V2, V2Size, V2TBAAInfo, SI->getTrueValue(), SISize, SITBAAInfo);
999 if (Alias == MayAlias)
1002 AliasResult ThisAlias =
1003 aliasCheck(V2, V2Size, V2TBAAInfo, SI->getFalseValue(), SISize, SITBAAInfo);
1004 return MergeAliasResults(ThisAlias, Alias);
1007 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
1009 AliasAnalysis::AliasResult
1010 BasicAliasAnalysis::aliasPHI(const PHINode *PN, uint64_t PNSize,
1011 const MDNode *PNTBAAInfo,
1012 const Value *V2, uint64_t V2Size,
1013 const MDNode *V2TBAAInfo) {
1014 // If the values are PHIs in the same block, we can do a more precise
1015 // as well as efficient check: just check for aliases between the values
1016 // on corresponding edges.
1017 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1018 if (PN2->getParent() == PN->getParent()) {
1020 aliasCheck(PN->getIncomingValue(0), PNSize, PNTBAAInfo,
1021 PN2->getIncomingValueForBlock(PN->getIncomingBlock(0)),
1022 V2Size, V2TBAAInfo);
1023 if (Alias == MayAlias)
1025 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1026 AliasResult ThisAlias =
1027 aliasCheck(PN->getIncomingValue(i), PNSize, PNTBAAInfo,
1028 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1029 V2Size, V2TBAAInfo);
1030 Alias = MergeAliasResults(ThisAlias, Alias);
1031 if (Alias == MayAlias)
1037 SmallPtrSet<Value*, 4> UniqueSrc;
1038 SmallVector<Value*, 4> V1Srcs;
1039 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1040 Value *PV1 = PN->getIncomingValue(i);
1041 if (isa<PHINode>(PV1))
1042 // If any of the source itself is a PHI, return MayAlias conservatively
1043 // to avoid compile time explosion. The worst possible case is if both
1044 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1045 // and 'n' are the number of PHI sources.
1047 if (UniqueSrc.insert(PV1))
1048 V1Srcs.push_back(PV1);
1051 AliasResult Alias = aliasCheck(V2, V2Size, V2TBAAInfo,
1052 V1Srcs[0], PNSize, PNTBAAInfo);
1053 // Early exit if the check of the first PHI source against V2 is MayAlias.
1054 // Other results are not possible.
1055 if (Alias == MayAlias)
1058 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1059 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1060 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1061 Value *V = V1Srcs[i];
1063 AliasResult ThisAlias = aliasCheck(V2, V2Size, V2TBAAInfo,
1064 V, PNSize, PNTBAAInfo);
1065 Alias = MergeAliasResults(ThisAlias, Alias);
1066 if (Alias == MayAlias)
1073 // aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
1074 // such as array references.
1076 AliasAnalysis::AliasResult
1077 BasicAliasAnalysis::aliasCheck(const Value *V1, uint64_t V1Size,
1078 const MDNode *V1TBAAInfo,
1079 const Value *V2, uint64_t V2Size,
1080 const MDNode *V2TBAAInfo) {
1081 // If either of the memory references is empty, it doesn't matter what the
1082 // pointer values are.
1083 if (V1Size == 0 || V2Size == 0)
1086 // Strip off any casts if they exist.
1087 V1 = V1->stripPointerCasts();
1088 V2 = V2->stripPointerCasts();
1090 // Are we checking for alias of the same value?
1091 if (V1 == V2) return MustAlias;
1093 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1094 return NoAlias; // Scalars cannot alias each other
1096 // Figure out what objects these things are pointing to if we can.
1097 const Value *O1 = GetUnderlyingObject(V1, TD);
1098 const Value *O2 = GetUnderlyingObject(V2, TD);
1100 // Null values in the default address space don't point to any object, so they
1101 // don't alias any other pointer.
1102 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1103 if (CPN->getType()->getAddressSpace() == 0)
1105 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1106 if (CPN->getType()->getAddressSpace() == 0)
1110 // If V1/V2 point to two different objects we know that we have no alias.
1111 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1114 // Constant pointers can't alias with non-const isIdentifiedObject objects.
1115 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1116 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1119 // Arguments can't alias with local allocations or noalias calls
1120 // in the same function.
1121 if (((isa<Argument>(O1) && (isa<AllocaInst>(O2) || isNoAliasCall(O2))) ||
1122 (isa<Argument>(O2) && (isa<AllocaInst>(O1) || isNoAliasCall(O1)))))
1125 // Most objects can't alias null.
1126 if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
1127 (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
1130 // If one pointer is the result of a call/invoke or load and the other is a
1131 // non-escaping local object within the same function, then we know the
1132 // object couldn't escape to a point where the call could return it.
1134 // Note that if the pointers are in different functions, there are a
1135 // variety of complications. A call with a nocapture argument may still
1136 // temporary store the nocapture argument's value in a temporary memory
1137 // location if that memory location doesn't escape. Or it may pass a
1138 // nocapture value to other functions as long as they don't capture it.
1139 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1141 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1145 // If the size of one access is larger than the entire object on the other
1146 // side, then we know such behavior is undefined and can assume no alias.
1148 if ((V1Size != UnknownSize && isObjectSmallerThan(O2, V1Size, *TD)) ||
1149 (V2Size != UnknownSize && isObjectSmallerThan(O1, V2Size, *TD)))
1152 // Check the cache before climbing up use-def chains. This also terminates
1153 // otherwise infinitely recursive queries.
1154 LocPair Locs(Location(V1, V1Size, V1TBAAInfo),
1155 Location(V2, V2Size, V2TBAAInfo));
1157 std::swap(Locs.first, Locs.second);
1158 std::pair<AliasCacheTy::iterator, bool> Pair =
1159 AliasCache.insert(std::make_pair(Locs, MayAlias));
1161 return Pair.first->second;
1163 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1164 // GEP can't simplify, we don't even look at the PHI cases.
1165 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1167 std::swap(V1Size, V2Size);
1170 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1171 AliasResult Result = aliasGEP(GV1, V1Size, V2, V2Size, V2TBAAInfo, O1, O2);
1172 if (Result != MayAlias) return AliasCache[Locs] = Result;
1175 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1177 std::swap(V1Size, V2Size);
1179 if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1180 AliasResult Result = aliasPHI(PN, V1Size, V1TBAAInfo,
1181 V2, V2Size, V2TBAAInfo);
1182 if (Result != MayAlias) return AliasCache[Locs] = Result;
1185 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1187 std::swap(V1Size, V2Size);
1189 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1190 AliasResult Result = aliasSelect(S1, V1Size, V1TBAAInfo,
1191 V2, V2Size, V2TBAAInfo);
1192 if (Result != MayAlias) return AliasCache[Locs] = Result;
1195 // If both pointers are pointing into the same object and one of them
1196 // accesses is accessing the entire object, then the accesses must
1197 // overlap in some way.
1199 if ((V1Size != UnknownSize && isObjectSize(O1, V1Size, *TD)) ||
1200 (V2Size != UnknownSize && isObjectSize(O2, V2Size, *TD)))
1201 return AliasCache[Locs] = PartialAlias;
1203 AliasResult Result =
1204 AliasAnalysis::alias(Location(V1, V1Size, V1TBAAInfo),
1205 Location(V2, V2Size, V2TBAAInfo));
1206 return AliasCache[Locs] = Result;