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/Target/TargetLibraryInfo.h"
34 #include "llvm/ADT/SmallPtrSet.h"
35 #include "llvm/ADT/SmallVector.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/GetElementPtrTypeIterator.h"
41 //===----------------------------------------------------------------------===//
43 //===----------------------------------------------------------------------===//
45 /// isKnownNonNull - Return true if we know that the specified value is never
47 static bool isKnownNonNull(const Value *V) {
48 // Alloca never returns null, malloc might.
49 if (isa<AllocaInst>(V)) return true;
51 // A byval argument is never null.
52 if (const Argument *A = dyn_cast<Argument>(V))
53 return A->hasByValAttr();
55 // Global values are not null unless extern weak.
56 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
57 return !GV->hasExternalWeakLinkage();
61 /// isNonEscapingLocalObject - Return true if the pointer is to a function-local
62 /// object that never escapes from the function.
63 static bool isNonEscapingLocalObject(const Value *V) {
64 // If this is a local allocation, check to see if it escapes.
65 if (isa<AllocaInst>(V) || isNoAliasCall(V))
66 // Set StoreCaptures to True so that we can assume in our callers that the
67 // pointer is not the result of a load instruction. Currently
68 // PointerMayBeCaptured doesn't have any special analysis for the
69 // StoreCaptures=false case; if it did, our callers could be refined to be
71 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
73 // If this is an argument that corresponds to a byval or noalias argument,
74 // then it has not escaped before entering the function. Check if it escapes
75 // inside the function.
76 if (const Argument *A = dyn_cast<Argument>(V))
77 if (A->hasByValAttr() || A->hasNoAliasAttr()) {
78 // Don't bother analyzing arguments already known not to escape.
79 if (A->hasNoCaptureAttr())
81 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
86 /// isEscapeSource - Return true if the pointer is one which would have
87 /// been considered an escape by isNonEscapingLocalObject.
88 static bool isEscapeSource(const Value *V) {
89 if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
92 // The load case works because isNonEscapingLocalObject considers all
93 // stores to be escapes (it passes true for the StoreCaptures argument
94 // to PointerMayBeCaptured).
101 /// getObjectSize - Return the size of the object specified by V, or
102 /// UnknownSize if unknown.
103 static uint64_t getObjectSize(const Value *V, const TargetData &TD) {
105 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
106 if (!GV->hasDefinitiveInitializer())
107 return AliasAnalysis::UnknownSize;
108 AccessTy = GV->getType()->getElementType();
109 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
110 if (!AI->isArrayAllocation())
111 AccessTy = AI->getType()->getElementType();
113 return AliasAnalysis::UnknownSize;
114 } else if (const CallInst* CI = extractMallocCall(V)) {
115 if (!isArrayMalloc(V, &TD))
116 // The size is the argument to the malloc call.
117 if (const ConstantInt* C = dyn_cast<ConstantInt>(CI->getArgOperand(0)))
118 return C->getZExtValue();
119 return AliasAnalysis::UnknownSize;
120 } else if (const Argument *A = dyn_cast<Argument>(V)) {
121 if (A->hasByValAttr())
122 AccessTy = cast<PointerType>(A->getType())->getElementType();
124 return AliasAnalysis::UnknownSize;
126 return AliasAnalysis::UnknownSize;
129 if (AccessTy->isSized())
130 return TD.getTypeAllocSize(AccessTy);
131 return AliasAnalysis::UnknownSize;
134 /// isObjectSmallerThan - Return true if we can prove that the object specified
135 /// by V is smaller than Size.
136 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
137 const TargetData &TD) {
138 uint64_t ObjectSize = getObjectSize(V, TD);
139 return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize < Size;
142 /// isObjectSize - Return true if we can prove that the object specified
143 /// by V has size Size.
144 static bool isObjectSize(const Value *V, uint64_t Size,
145 const TargetData &TD) {
146 uint64_t ObjectSize = getObjectSize(V, TD);
147 return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize == Size;
150 //===----------------------------------------------------------------------===//
151 // GetElementPtr Instruction Decomposition and Analysis
152 //===----------------------------------------------------------------------===//
161 struct VariableGEPIndex {
163 ExtensionKind Extension;
169 /// GetLinearExpression - Analyze the specified value as a linear expression:
170 /// "A*V + B", where A and B are constant integers. Return the scale and offset
171 /// values as APInts and return V as a Value*, and return whether we looked
172 /// through any sign or zero extends. The incoming Value is known to have
173 /// IntegerType and it may already be sign or zero extended.
175 /// Note that this looks through extends, so the high bits may not be
176 /// represented in the result.
177 static Value *GetLinearExpression(Value *V, APInt &Scale, APInt &Offset,
178 ExtensionKind &Extension,
179 const TargetData &TD, unsigned Depth) {
180 assert(V->getType()->isIntegerTy() && "Not an integer value");
182 // Limit our recursion depth.
189 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
190 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
191 switch (BOp->getOpcode()) {
193 case Instruction::Or:
194 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
196 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), &TD))
199 case Instruction::Add:
200 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
202 Offset += RHSC->getValue();
204 case Instruction::Mul:
205 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
207 Offset *= RHSC->getValue();
208 Scale *= RHSC->getValue();
210 case Instruction::Shl:
211 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
213 Offset <<= RHSC->getValue().getLimitedValue();
214 Scale <<= RHSC->getValue().getLimitedValue();
220 // Since GEP indices are sign extended anyway, we don't care about the high
221 // bits of a sign or zero extended value - just scales and offsets. The
222 // extensions have to be consistent though.
223 if ((isa<SExtInst>(V) && Extension != EK_ZeroExt) ||
224 (isa<ZExtInst>(V) && Extension != EK_SignExt)) {
225 Value *CastOp = cast<CastInst>(V)->getOperand(0);
226 unsigned OldWidth = Scale.getBitWidth();
227 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
228 Scale = Scale.trunc(SmallWidth);
229 Offset = Offset.trunc(SmallWidth);
230 Extension = isa<SExtInst>(V) ? EK_SignExt : EK_ZeroExt;
232 Value *Result = GetLinearExpression(CastOp, Scale, Offset, Extension,
234 Scale = Scale.zext(OldWidth);
235 Offset = Offset.zext(OldWidth);
245 /// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it
246 /// into a base pointer with a constant offset and a number of scaled symbolic
249 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale in
250 /// the VarIndices vector) are Value*'s that are known to be scaled by the
251 /// specified amount, but which may have other unrepresented high bits. As such,
252 /// the gep cannot necessarily be reconstructed from its decomposed form.
254 /// When TargetData is around, this function is capable of analyzing everything
255 /// that GetUnderlyingObject can look through. When not, it just looks
256 /// through pointer casts.
259 DecomposeGEPExpression(const Value *V, int64_t &BaseOffs,
260 SmallVectorImpl<VariableGEPIndex> &VarIndices,
261 const TargetData *TD) {
262 // Limit recursion depth to limit compile time in crazy cases.
263 unsigned MaxLookup = 6;
267 // See if this is a bitcast or GEP.
268 const Operator *Op = dyn_cast<Operator>(V);
270 // The only non-operator case we can handle are GlobalAliases.
271 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
272 if (!GA->mayBeOverridden()) {
273 V = GA->getAliasee();
280 if (Op->getOpcode() == Instruction::BitCast) {
281 V = Op->getOperand(0);
285 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
287 // If it's not a GEP, hand it off to SimplifyInstruction to see if it
288 // can come up with something. This matches what GetUnderlyingObject does.
289 if (const Instruction *I = dyn_cast<Instruction>(V))
290 // TODO: Get a DominatorTree and use it here.
291 if (const Value *Simplified =
292 SimplifyInstruction(const_cast<Instruction *>(I), TD)) {
300 // Don't attempt to analyze GEPs over unsized objects.
301 if (!cast<PointerType>(GEPOp->getOperand(0)->getType())
302 ->getElementType()->isSized())
305 // If we are lacking TargetData information, we can't compute the offets of
306 // elements computed by GEPs. However, we can handle bitcast equivalent
309 if (!GEPOp->hasAllZeroIndices())
311 V = GEPOp->getOperand(0);
315 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
316 gep_type_iterator GTI = gep_type_begin(GEPOp);
317 for (User::const_op_iterator I = GEPOp->op_begin()+1,
318 E = GEPOp->op_end(); I != E; ++I) {
320 // Compute the (potentially symbolic) offset in bytes for this index.
321 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
322 // For a struct, add the member offset.
323 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
324 if (FieldNo == 0) continue;
326 BaseOffs += TD->getStructLayout(STy)->getElementOffset(FieldNo);
330 // For an array/pointer, add the element offset, explicitly scaled.
331 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
332 if (CIdx->isZero()) continue;
333 BaseOffs += TD->getTypeAllocSize(*GTI)*CIdx->getSExtValue();
337 uint64_t Scale = TD->getTypeAllocSize(*GTI);
338 ExtensionKind Extension = EK_NotExtended;
340 // If the integer type is smaller than the pointer size, it is implicitly
341 // sign extended to pointer size.
342 unsigned Width = cast<IntegerType>(Index->getType())->getBitWidth();
343 if (TD->getPointerSizeInBits() > Width)
344 Extension = EK_SignExt;
346 // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
347 APInt IndexScale(Width, 0), IndexOffset(Width, 0);
348 Index = GetLinearExpression(Index, IndexScale, IndexOffset, Extension,
351 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
352 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
353 BaseOffs += IndexOffset.getSExtValue()*Scale;
354 Scale *= IndexScale.getSExtValue();
357 // If we already had an occurrence of this index variable, merge this
358 // scale into it. For example, we want to handle:
359 // A[x][x] -> x*16 + x*4 -> x*20
360 // This also ensures that 'x' only appears in the index list once.
361 for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
362 if (VarIndices[i].V == Index &&
363 VarIndices[i].Extension == Extension) {
364 Scale += VarIndices[i].Scale;
365 VarIndices.erase(VarIndices.begin()+i);
370 // Make sure that we have a scale that makes sense for this target's
372 if (unsigned ShiftBits = 64-TD->getPointerSizeInBits()) {
374 Scale = (int64_t)Scale >> ShiftBits;
378 VariableGEPIndex Entry = {Index, Extension,
379 static_cast<int64_t>(Scale)};
380 VarIndices.push_back(Entry);
384 // Analyze the base pointer next.
385 V = GEPOp->getOperand(0);
386 } while (--MaxLookup);
388 // If the chain of expressions is too deep, just return early.
392 /// GetIndexDifference - Dest and Src are the variable indices from two
393 /// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base
394 /// pointers. Subtract the GEP2 indices from GEP1 to find the symbolic
395 /// difference between the two pointers.
396 static void GetIndexDifference(SmallVectorImpl<VariableGEPIndex> &Dest,
397 const SmallVectorImpl<VariableGEPIndex> &Src) {
398 if (Src.empty()) return;
400 for (unsigned i = 0, e = Src.size(); i != e; ++i) {
401 const Value *V = Src[i].V;
402 ExtensionKind Extension = Src[i].Extension;
403 int64_t Scale = Src[i].Scale;
405 // Find V in Dest. This is N^2, but pointer indices almost never have more
406 // than a few variable indexes.
407 for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
408 if (Dest[j].V != V || Dest[j].Extension != Extension) continue;
410 // If we found it, subtract off Scale V's from the entry in Dest. If it
411 // goes to zero, remove the entry.
412 if (Dest[j].Scale != Scale)
413 Dest[j].Scale -= Scale;
415 Dest.erase(Dest.begin()+j);
420 // If we didn't consume this entry, add it to the end of the Dest list.
422 VariableGEPIndex Entry = { V, Extension, -Scale };
423 Dest.push_back(Entry);
428 //===----------------------------------------------------------------------===//
429 // BasicAliasAnalysis Pass
430 //===----------------------------------------------------------------------===//
433 static const Function *getParent(const Value *V) {
434 if (const Instruction *inst = dyn_cast<Instruction>(V))
435 return inst->getParent()->getParent();
437 if (const Argument *arg = dyn_cast<Argument>(V))
438 return arg->getParent();
443 static bool notDifferentParent(const Value *O1, const Value *O2) {
445 const Function *F1 = getParent(O1);
446 const Function *F2 = getParent(O2);
448 return !F1 || !F2 || F1 == F2;
453 /// BasicAliasAnalysis - This is the primary alias analysis implementation.
454 struct BasicAliasAnalysis : public ImmutablePass, public AliasAnalysis {
455 static char ID; // Class identification, replacement for typeinfo
456 BasicAliasAnalysis() : ImmutablePass(ID),
457 // AliasCache rarely has more than 1 or 2 elements,
458 // so start it off fairly small so that clear()
459 // doesn't have to tromp through 64 (the default)
460 // elements on each alias query. This really wants
461 // something like a SmallDenseMap.
463 initializeBasicAliasAnalysisPass(*PassRegistry::getPassRegistry());
466 virtual void initializePass() {
467 InitializeAliasAnalysis(this);
470 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
471 AU.addRequired<AliasAnalysis>();
472 AU.addRequired<TargetLibraryInfo>();
475 virtual AliasResult alias(const Location &LocA,
476 const Location &LocB) {
477 assert(AliasCache.empty() && "AliasCache must be cleared after use!");
478 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
479 "BasicAliasAnalysis doesn't support interprocedural queries.");
480 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.TBAATag,
481 LocB.Ptr, LocB.Size, LocB.TBAATag);
486 virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
487 const Location &Loc);
489 virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
490 ImmutableCallSite CS2) {
491 // The AliasAnalysis base class has some smarts, lets use them.
492 return AliasAnalysis::getModRefInfo(CS1, CS2);
495 /// pointsToConstantMemory - Chase pointers until we find a (constant
497 virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
499 /// getModRefBehavior - Return the behavior when calling the given
501 virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
503 /// getModRefBehavior - Return the behavior when calling the given function.
504 /// For use when the call site is not known.
505 virtual ModRefBehavior getModRefBehavior(const Function *F);
507 /// getAdjustedAnalysisPointer - This method is used when a pass implements
508 /// an analysis interface through multiple inheritance. If needed, it
509 /// should override this to adjust the this pointer as needed for the
510 /// specified pass info.
511 virtual void *getAdjustedAnalysisPointer(const void *ID) {
512 if (ID == &AliasAnalysis::ID)
513 return (AliasAnalysis*)this;
518 // AliasCache - Track alias queries to guard against recursion.
519 typedef std::pair<Location, Location> LocPair;
520 typedef DenseMap<LocPair, AliasResult> AliasCacheTy;
521 AliasCacheTy AliasCache;
523 // Visited - Track instructions visited by pointsToConstantMemory.
524 SmallPtrSet<const Value*, 16> Visited;
526 // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP
527 // instruction against another.
528 AliasResult aliasGEP(const GEPOperator *V1, uint64_t V1Size,
529 const Value *V2, uint64_t V2Size,
530 const MDNode *V2TBAAInfo,
531 const Value *UnderlyingV1, const Value *UnderlyingV2);
533 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI
534 // instruction against another.
535 AliasResult aliasPHI(const PHINode *PN, uint64_t PNSize,
536 const MDNode *PNTBAAInfo,
537 const Value *V2, uint64_t V2Size,
538 const MDNode *V2TBAAInfo);
540 /// aliasSelect - Disambiguate a Select instruction against another value.
541 AliasResult aliasSelect(const SelectInst *SI, uint64_t SISize,
542 const MDNode *SITBAAInfo,
543 const Value *V2, uint64_t V2Size,
544 const MDNode *V2TBAAInfo);
546 AliasResult aliasCheck(const Value *V1, uint64_t V1Size,
547 const MDNode *V1TBAATag,
548 const Value *V2, uint64_t V2Size,
549 const MDNode *V2TBAATag);
551 } // End of anonymous namespace
553 // Register this pass...
554 char BasicAliasAnalysis::ID = 0;
555 INITIALIZE_AG_PASS_BEGIN(BasicAliasAnalysis, AliasAnalysis, "basicaa",
556 "Basic Alias Analysis (stateless AA impl)",
558 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
559 INITIALIZE_AG_PASS_END(BasicAliasAnalysis, AliasAnalysis, "basicaa",
560 "Basic Alias Analysis (stateless AA impl)",
564 ImmutablePass *llvm::createBasicAliasAnalysisPass() {
565 return new BasicAliasAnalysis();
568 /// pointsToConstantMemory - Returns whether the given pointer value
569 /// points to memory that is local to the function, with global constants being
570 /// considered local to all functions.
572 BasicAliasAnalysis::pointsToConstantMemory(const Location &Loc, bool OrLocal) {
573 assert(Visited.empty() && "Visited must be cleared after use!");
575 unsigned MaxLookup = 8;
576 SmallVector<const Value *, 16> Worklist;
577 Worklist.push_back(Loc.Ptr);
579 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), TD);
580 if (!Visited.insert(V)) {
582 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
585 // An alloca instruction defines local memory.
586 if (OrLocal && isa<AllocaInst>(V))
589 // A global constant counts as local memory for our purposes.
590 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
591 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
592 // global to be marked constant in some modules and non-constant in
593 // others. GV may even be a declaration, not a definition.
594 if (!GV->isConstant()) {
596 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
601 // If both select values point to local memory, then so does the select.
602 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
603 Worklist.push_back(SI->getTrueValue());
604 Worklist.push_back(SI->getFalseValue());
608 // If all values incoming to a phi node point to local memory, then so does
610 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
611 // Don't bother inspecting phi nodes with many operands.
612 if (PN->getNumIncomingValues() > MaxLookup) {
614 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
616 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
617 Worklist.push_back(PN->getIncomingValue(i));
621 // Otherwise be conservative.
623 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
625 } while (!Worklist.empty() && --MaxLookup);
628 return Worklist.empty();
631 /// getModRefBehavior - Return the behavior when calling the given call site.
632 AliasAnalysis::ModRefBehavior
633 BasicAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
634 if (CS.doesNotAccessMemory())
635 // Can't do better than this.
636 return DoesNotAccessMemory;
638 ModRefBehavior Min = UnknownModRefBehavior;
640 // If the callsite knows it only reads memory, don't return worse
642 if (CS.onlyReadsMemory())
643 Min = OnlyReadsMemory;
645 // The AliasAnalysis base class has some smarts, lets use them.
646 return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
649 /// getModRefBehavior - Return the behavior when calling the given function.
650 /// For use when the call site is not known.
651 AliasAnalysis::ModRefBehavior
652 BasicAliasAnalysis::getModRefBehavior(const Function *F) {
653 // If the function declares it doesn't access memory, we can't do better.
654 if (F->doesNotAccessMemory())
655 return DoesNotAccessMemory;
657 // For intrinsics, we can check the table.
658 if (unsigned iid = F->getIntrinsicID()) {
659 #define GET_INTRINSIC_MODREF_BEHAVIOR
660 #include "llvm/Intrinsics.gen"
661 #undef GET_INTRINSIC_MODREF_BEHAVIOR
664 ModRefBehavior Min = UnknownModRefBehavior;
666 // If the function declares it only reads memory, go with that.
667 if (F->onlyReadsMemory())
668 Min = OnlyReadsMemory;
670 // Otherwise be conservative.
671 return ModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min);
674 /// getModRefInfo - Check to see if the specified callsite can clobber the
675 /// specified memory object. Since we only look at local properties of this
676 /// function, we really can't say much about this query. We do, however, use
677 /// simple "address taken" analysis on local objects.
678 AliasAnalysis::ModRefResult
679 BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS,
680 const Location &Loc) {
681 assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
682 "AliasAnalysis query involving multiple functions!");
684 const Value *Object = GetUnderlyingObject(Loc.Ptr, TD);
686 // If this is a tail call and Loc.Ptr points to a stack location, we know that
687 // the tail call cannot access or modify the local stack.
688 // We cannot exclude byval arguments here; these belong to the caller of
689 // the current function not to the current function, and a tail callee
690 // may reference them.
691 if (isa<AllocaInst>(Object))
692 if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
693 if (CI->isTailCall())
696 // If the pointer is to a locally allocated object that does not escape,
697 // then the call can not mod/ref the pointer unless the call takes the pointer
698 // as an argument, and itself doesn't capture it.
699 if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
700 isNonEscapingLocalObject(Object)) {
701 bool PassedAsArg = false;
703 for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
704 CI != CE; ++CI, ++ArgNo) {
705 // Only look at the no-capture or byval pointer arguments. If this
706 // pointer were passed to arguments that were neither of these, then it
707 // couldn't be no-capture.
708 if (!(*CI)->getType()->isPointerTy() ||
709 (!CS.paramHasAttr(ArgNo+1, Attribute::NoCapture) &&
710 !CS.paramHasAttr(ArgNo+1, Attribute::ByVal)))
713 // If this is a no-capture pointer argument, see if we can tell that it
714 // is impossible to alias the pointer we're checking. If not, we have to
715 // assume that the call could touch the pointer, even though it doesn't
717 if (!isNoAlias(Location(cast<Value>(CI)), Loc)) {
727 const TargetLibraryInfo &TLI = getAnalysis<TargetLibraryInfo>();
728 ModRefResult Min = ModRef;
730 // Finally, handle specific knowledge of intrinsics.
731 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
733 switch (II->getIntrinsicID()) {
735 case Intrinsic::memcpy:
736 case Intrinsic::memmove: {
737 uint64_t Len = UnknownSize;
738 if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2)))
739 Len = LenCI->getZExtValue();
740 Value *Dest = II->getArgOperand(0);
741 Value *Src = II->getArgOperand(1);
742 // If it can't overlap the source dest, then it doesn't modref the loc.
743 if (isNoAlias(Location(Dest, Len), Loc)) {
744 if (isNoAlias(Location(Src, Len), Loc))
746 // If it can't overlap the dest, then worst case it reads the loc.
748 } else if (isNoAlias(Location(Src, Len), Loc)) {
749 // If it can't overlap the source, then worst case it mutates the loc.
754 case Intrinsic::memset:
755 // Since memset is 'accesses arguments' only, the AliasAnalysis base class
756 // will handle it for the variable length case.
757 if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
758 uint64_t Len = LenCI->getZExtValue();
759 Value *Dest = II->getArgOperand(0);
760 if (isNoAlias(Location(Dest, Len), Loc))
763 // We know that memset doesn't load anything.
766 case Intrinsic::atomic_cmp_swap:
767 case Intrinsic::atomic_swap:
768 case Intrinsic::atomic_load_add:
769 case Intrinsic::atomic_load_sub:
770 case Intrinsic::atomic_load_and:
771 case Intrinsic::atomic_load_nand:
772 case Intrinsic::atomic_load_or:
773 case Intrinsic::atomic_load_xor:
774 case Intrinsic::atomic_load_max:
775 case Intrinsic::atomic_load_min:
776 case Intrinsic::atomic_load_umax:
777 case Intrinsic::atomic_load_umin:
779 Value *Op1 = II->getArgOperand(0);
780 uint64_t Op1Size = TD->getTypeStoreSize(Op1->getType());
781 MDNode *Tag = II->getMetadata(LLVMContext::MD_tbaa);
782 if (isNoAlias(Location(Op1, Op1Size, Tag), Loc))
786 case Intrinsic::lifetime_start:
787 case Intrinsic::lifetime_end:
788 case Intrinsic::invariant_start: {
790 cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
791 if (isNoAlias(Location(II->getArgOperand(1),
793 II->getMetadata(LLVMContext::MD_tbaa)),
798 case Intrinsic::invariant_end: {
800 cast<ConstantInt>(II->getArgOperand(1))->getZExtValue();
801 if (isNoAlias(Location(II->getArgOperand(2),
803 II->getMetadata(LLVMContext::MD_tbaa)),
808 case Intrinsic::arm_neon_vld1: {
809 // LLVM's vld1 and vst1 intrinsics currently only support a single
812 TD ? TD->getTypeStoreSize(II->getType()) : UnknownSize;
813 if (isNoAlias(Location(II->getArgOperand(0), Size,
814 II->getMetadata(LLVMContext::MD_tbaa)),
819 case Intrinsic::arm_neon_vst1: {
821 TD ? TD->getTypeStoreSize(II->getArgOperand(1)->getType()) : UnknownSize;
822 if (isNoAlias(Location(II->getArgOperand(0), Size,
823 II->getMetadata(LLVMContext::MD_tbaa)),
830 // We can bound the aliasing properties of memset_pattern16 just as we can
831 // for memcpy/memset. This is particularly important because the
832 // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
833 // whenever possible.
834 else if (TLI.has(LibFunc::memset_pattern16) &&
835 CS.getCalledFunction() &&
836 CS.getCalledFunction()->getName() == "memset_pattern16") {
837 const Function *MS = CS.getCalledFunction();
838 FunctionType *MemsetType = MS->getFunctionType();
839 if (!MemsetType->isVarArg() && MemsetType->getNumParams() == 3 &&
840 isa<PointerType>(MemsetType->getParamType(0)) &&
841 isa<PointerType>(MemsetType->getParamType(1)) &&
842 isa<IntegerType>(MemsetType->getParamType(2))) {
843 uint64_t Len = UnknownSize;
844 if (const ConstantInt *LenCI = dyn_cast<ConstantInt>(CS.getArgument(2)))
845 Len = LenCI->getZExtValue();
846 const Value *Dest = CS.getArgument(0);
847 const Value *Src = CS.getArgument(1);
848 // If it can't overlap the source dest, then it doesn't modref the loc.
849 if (isNoAlias(Location(Dest, Len), Loc)) {
850 if (isNoAlias(Location(Src, 2), Loc))
852 // If it can't overlap the dest, then worst case it reads the loc.
854 } else if (isNoAlias(Location(Src, 2), Loc)) {
855 // If it can't overlap the source, then worst case it mutates the loc.
861 // The AliasAnalysis base class has some smarts, lets use them.
862 return ModRefResult(AliasAnalysis::getModRefInfo(CS, Loc) & Min);
865 /// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
866 /// against another pointer. We know that V1 is a GEP, but we don't know
867 /// anything about V2. UnderlyingV1 is GetUnderlyingObject(GEP1, TD),
868 /// UnderlyingV2 is the same for V2.
870 AliasAnalysis::AliasResult
871 BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
872 const Value *V2, uint64_t V2Size,
873 const MDNode *V2TBAAInfo,
874 const Value *UnderlyingV1,
875 const Value *UnderlyingV2) {
876 int64_t GEP1BaseOffset;
877 SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
879 // If we have two gep instructions with must-alias'ing base pointers, figure
880 // out if the indexes to the GEP tell us anything about the derived pointer.
881 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
882 // Do the base pointers alias?
883 AliasResult BaseAlias = aliasCheck(UnderlyingV1, UnknownSize, 0,
884 UnderlyingV2, UnknownSize, 0);
886 // If we get a No or May, then return it immediately, no amount of analysis
887 // will improve this situation.
888 if (BaseAlias != MustAlias) return BaseAlias;
890 // Otherwise, we have a MustAlias. Since the base pointers alias each other
891 // exactly, see if the computed offset from the common pointer tells us
892 // about the relation of the resulting pointer.
893 const Value *GEP1BasePtr =
894 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
896 int64_t GEP2BaseOffset;
897 SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
898 const Value *GEP2BasePtr =
899 DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD);
901 // If DecomposeGEPExpression isn't able to look all the way through the
902 // addressing operation, we must not have TD and this is too complex for us
903 // to handle without it.
904 if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
906 "DecomposeGEPExpression and GetUnderlyingObject disagree!");
910 // Subtract the GEP2 pointer from the GEP1 pointer to find out their
911 // symbolic difference.
912 GEP1BaseOffset -= GEP2BaseOffset;
913 GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
916 // Check to see if these two pointers are related by the getelementptr
917 // instruction. If one pointer is a GEP with a non-zero index of the other
918 // pointer, we know they cannot alias.
920 // If both accesses are unknown size, we can't do anything useful here.
921 if (V1Size == UnknownSize && V2Size == UnknownSize)
924 AliasResult R = aliasCheck(UnderlyingV1, UnknownSize, 0,
925 V2, V2Size, V2TBAAInfo);
927 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
928 // If V2 is known not to alias GEP base pointer, then the two values
929 // cannot alias per GEP semantics: "A pointer value formed from a
930 // getelementptr instruction is associated with the addresses associated
931 // with the first operand of the getelementptr".
934 const Value *GEP1BasePtr =
935 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
937 // If DecomposeGEPExpression isn't able to look all the way through the
938 // addressing operation, we must not have TD and this is too complex for us
939 // to handle without it.
940 if (GEP1BasePtr != UnderlyingV1) {
942 "DecomposeGEPExpression and GetUnderlyingObject disagree!");
947 // In the two GEP Case, if there is no difference in the offsets of the
948 // computed pointers, the resultant pointers are a must alias. This
949 // hapens when we have two lexically identical GEP's (for example).
951 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
952 // must aliases the GEP, the end result is a must alias also.
953 if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
956 // If there is a difference between the pointers, but the difference is
957 // less than the size of the associated memory object, then we know
958 // that the objects are partially overlapping.
959 if (GEP1BaseOffset != 0 && GEP1VariableIndices.empty()) {
960 if (GEP1BaseOffset >= 0 ?
961 (V2Size != UnknownSize && (uint64_t)GEP1BaseOffset < V2Size) :
962 (V1Size != UnknownSize && -(uint64_t)GEP1BaseOffset < V1Size &&
963 GEP1BaseOffset != INT64_MIN))
967 // If we have a known constant offset, see if this offset is larger than the
968 // access size being queried. If so, and if no variable indices can remove
969 // pieces of this constant, then we know we have a no-alias. For example,
972 // In order to handle cases like &A[100][i] where i is an out of range
973 // subscript, we have to ignore all constant offset pieces that are a multiple
974 // of a scaled index. Do this by removing constant offsets that are a
975 // multiple of any of our variable indices. This allows us to transform
976 // things like &A[i][1] because i has a stride of (e.g.) 8 bytes but the 1
977 // provides an offset of 4 bytes (assuming a <= 4 byte access).
978 for (unsigned i = 0, e = GEP1VariableIndices.size();
979 i != e && GEP1BaseOffset;++i)
980 if (int64_t RemovedOffset = GEP1BaseOffset/GEP1VariableIndices[i].Scale)
981 GEP1BaseOffset -= RemovedOffset*GEP1VariableIndices[i].Scale;
983 // If our known offset is bigger than the access size, we know we don't have
985 if (GEP1BaseOffset) {
986 if (GEP1BaseOffset >= 0 ?
987 (V2Size != UnknownSize && (uint64_t)GEP1BaseOffset >= V2Size) :
988 (V1Size != UnknownSize && -(uint64_t)GEP1BaseOffset >= V1Size &&
989 GEP1BaseOffset != INT64_MIN))
993 // Statically, we can see that the base objects are the same, but the
994 // pointers have dynamic offsets which we can't resolve. And none of our
995 // little tricks above worked.
997 // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the
998 // practical effect of this is protecting TBAA in the case of dynamic
999 // indices into arrays of unions. An alternative way to solve this would
1000 // be to have clang emit extra metadata for unions and/or union accesses.
1001 // A union-specific solution wouldn't handle the problem for malloc'd
1003 return PartialAlias;
1006 static AliasAnalysis::AliasResult
1007 MergeAliasResults(AliasAnalysis::AliasResult A, AliasAnalysis::AliasResult B) {
1008 // If the results agree, take it.
1011 // A mix of PartialAlias and MustAlias is PartialAlias.
1012 if ((A == AliasAnalysis::PartialAlias && B == AliasAnalysis::MustAlias) ||
1013 (B == AliasAnalysis::PartialAlias && A == AliasAnalysis::MustAlias))
1014 return AliasAnalysis::PartialAlias;
1015 // Otherwise, we don't know anything.
1016 return AliasAnalysis::MayAlias;
1019 /// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select
1020 /// instruction against another.
1021 AliasAnalysis::AliasResult
1022 BasicAliasAnalysis::aliasSelect(const SelectInst *SI, uint64_t SISize,
1023 const MDNode *SITBAAInfo,
1024 const Value *V2, uint64_t V2Size,
1025 const MDNode *V2TBAAInfo) {
1026 // If the values are Selects with the same condition, we can do a more precise
1027 // check: just check for aliases between the values on corresponding arms.
1028 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1029 if (SI->getCondition() == SI2->getCondition()) {
1031 aliasCheck(SI->getTrueValue(), SISize, SITBAAInfo,
1032 SI2->getTrueValue(), V2Size, V2TBAAInfo);
1033 if (Alias == MayAlias)
1035 AliasResult ThisAlias =
1036 aliasCheck(SI->getFalseValue(), SISize, SITBAAInfo,
1037 SI2->getFalseValue(), V2Size, V2TBAAInfo);
1038 return MergeAliasResults(ThisAlias, Alias);
1041 // If both arms of the Select node NoAlias or MustAlias V2, then returns
1042 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1044 aliasCheck(V2, V2Size, V2TBAAInfo, SI->getTrueValue(), SISize, SITBAAInfo);
1045 if (Alias == MayAlias)
1048 AliasResult ThisAlias =
1049 aliasCheck(V2, V2Size, V2TBAAInfo, SI->getFalseValue(), SISize, SITBAAInfo);
1050 return MergeAliasResults(ThisAlias, Alias);
1053 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
1055 AliasAnalysis::AliasResult
1056 BasicAliasAnalysis::aliasPHI(const PHINode *PN, uint64_t PNSize,
1057 const MDNode *PNTBAAInfo,
1058 const Value *V2, uint64_t V2Size,
1059 const MDNode *V2TBAAInfo) {
1060 // If the values are PHIs in the same block, we can do a more precise
1061 // as well as efficient check: just check for aliases between the values
1062 // on corresponding edges.
1063 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1064 if (PN2->getParent() == PN->getParent()) {
1066 aliasCheck(PN->getIncomingValue(0), PNSize, PNTBAAInfo,
1067 PN2->getIncomingValueForBlock(PN->getIncomingBlock(0)),
1068 V2Size, V2TBAAInfo);
1069 if (Alias == MayAlias)
1071 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1072 AliasResult ThisAlias =
1073 aliasCheck(PN->getIncomingValue(i), PNSize, PNTBAAInfo,
1074 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1075 V2Size, V2TBAAInfo);
1076 Alias = MergeAliasResults(ThisAlias, Alias);
1077 if (Alias == MayAlias)
1083 SmallPtrSet<Value*, 4> UniqueSrc;
1084 SmallVector<Value*, 4> V1Srcs;
1085 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1086 Value *PV1 = PN->getIncomingValue(i);
1087 if (isa<PHINode>(PV1))
1088 // If any of the source itself is a PHI, return MayAlias conservatively
1089 // to avoid compile time explosion. The worst possible case is if both
1090 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1091 // and 'n' are the number of PHI sources.
1093 if (UniqueSrc.insert(PV1))
1094 V1Srcs.push_back(PV1);
1097 AliasResult Alias = aliasCheck(V2, V2Size, V2TBAAInfo,
1098 V1Srcs[0], PNSize, PNTBAAInfo);
1099 // Early exit if the check of the first PHI source against V2 is MayAlias.
1100 // Other results are not possible.
1101 if (Alias == MayAlias)
1104 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1105 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1106 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1107 Value *V = V1Srcs[i];
1109 AliasResult ThisAlias = aliasCheck(V2, V2Size, V2TBAAInfo,
1110 V, PNSize, PNTBAAInfo);
1111 Alias = MergeAliasResults(ThisAlias, Alias);
1112 if (Alias == MayAlias)
1119 // aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
1120 // such as array references.
1122 AliasAnalysis::AliasResult
1123 BasicAliasAnalysis::aliasCheck(const Value *V1, uint64_t V1Size,
1124 const MDNode *V1TBAAInfo,
1125 const Value *V2, uint64_t V2Size,
1126 const MDNode *V2TBAAInfo) {
1127 // If either of the memory references is empty, it doesn't matter what the
1128 // pointer values are.
1129 if (V1Size == 0 || V2Size == 0)
1132 // Strip off any casts if they exist.
1133 V1 = V1->stripPointerCasts();
1134 V2 = V2->stripPointerCasts();
1136 // Are we checking for alias of the same value?
1137 if (V1 == V2) return MustAlias;
1139 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1140 return NoAlias; // Scalars cannot alias each other
1142 // Figure out what objects these things are pointing to if we can.
1143 const Value *O1 = GetUnderlyingObject(V1, TD);
1144 const Value *O2 = GetUnderlyingObject(V2, TD);
1146 // Null values in the default address space don't point to any object, so they
1147 // don't alias any other pointer.
1148 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1149 if (CPN->getType()->getAddressSpace() == 0)
1151 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1152 if (CPN->getType()->getAddressSpace() == 0)
1156 // If V1/V2 point to two different objects we know that we have no alias.
1157 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1160 // Constant pointers can't alias with non-const isIdentifiedObject objects.
1161 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1162 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1165 // Arguments can't alias with local allocations or noalias calls
1166 // in the same function.
1167 if (((isa<Argument>(O1) && (isa<AllocaInst>(O2) || isNoAliasCall(O2))) ||
1168 (isa<Argument>(O2) && (isa<AllocaInst>(O1) || isNoAliasCall(O1)))))
1171 // Most objects can't alias null.
1172 if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
1173 (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
1176 // If one pointer is the result of a call/invoke or load and the other is a
1177 // non-escaping local object within the same function, then we know the
1178 // object couldn't escape to a point where the call could return it.
1180 // Note that if the pointers are in different functions, there are a
1181 // variety of complications. A call with a nocapture argument may still
1182 // temporary store the nocapture argument's value in a temporary memory
1183 // location if that memory location doesn't escape. Or it may pass a
1184 // nocapture value to other functions as long as they don't capture it.
1185 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1187 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1191 // If the size of one access is larger than the entire object on the other
1192 // side, then we know such behavior is undefined and can assume no alias.
1194 if ((V1Size != UnknownSize && isObjectSmallerThan(O2, V1Size, *TD)) ||
1195 (V2Size != UnknownSize && isObjectSmallerThan(O1, V2Size, *TD)))
1198 // Check the cache before climbing up use-def chains. This also terminates
1199 // otherwise infinitely recursive queries.
1200 LocPair Locs(Location(V1, V1Size, V1TBAAInfo),
1201 Location(V2, V2Size, V2TBAAInfo));
1203 std::swap(Locs.first, Locs.second);
1204 std::pair<AliasCacheTy::iterator, bool> Pair =
1205 AliasCache.insert(std::make_pair(Locs, MayAlias));
1207 return Pair.first->second;
1209 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1210 // GEP can't simplify, we don't even look at the PHI cases.
1211 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1213 std::swap(V1Size, V2Size);
1216 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1217 AliasResult Result = aliasGEP(GV1, V1Size, V2, V2Size, V2TBAAInfo, O1, O2);
1218 if (Result != MayAlias) return AliasCache[Locs] = Result;
1221 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1223 std::swap(V1Size, V2Size);
1225 if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1226 AliasResult Result = aliasPHI(PN, V1Size, V1TBAAInfo,
1227 V2, V2Size, V2TBAAInfo);
1228 if (Result != MayAlias) return AliasCache[Locs] = Result;
1231 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1233 std::swap(V1Size, V2Size);
1235 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1236 AliasResult Result = aliasSelect(S1, V1Size, V1TBAAInfo,
1237 V2, V2Size, V2TBAAInfo);
1238 if (Result != MayAlias) return AliasCache[Locs] = Result;
1241 // If both pointers are pointing into the same object and one of them
1242 // accesses is accessing the entire object, then the accesses must
1243 // overlap in some way.
1245 if ((V1Size != UnknownSize && isObjectSize(O1, V1Size, *TD)) ||
1246 (V2Size != UnknownSize && isObjectSize(O2, V2Size, *TD)))
1247 return AliasCache[Locs] = PartialAlias;
1249 AliasResult Result =
1250 AliasAnalysis::alias(Location(V1, V1Size, V1TBAAInfo),
1251 Location(V2, V2Size, V2TBAAInfo));
1252 return AliasCache[Locs] = Result;