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/Passes.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/AssumptionCache.h"
21 #include "llvm/Analysis/CFG.h"
22 #include "llvm/Analysis/CaptureTracking.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/LoopInfo.h"
25 #include "llvm/Analysis/MemoryBuiltins.h"
26 #include "llvm/Analysis/TargetLibraryInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/GetElementPtrTypeIterator.h"
34 #include "llvm/IR/GlobalAlias.h"
35 #include "llvm/IR/GlobalVariable.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/LLVMContext.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/ErrorHandling.h"
45 /// Cutoff after which to stop analysing a set of phi nodes potentially involved
46 /// in a cycle. Because we are analysing 'through' phi nodes we need to be
47 /// careful with value equivalence. We use reachability to make sure a value
48 /// cannot be involved in a cycle.
49 const unsigned MaxNumPhiBBsValueReachabilityCheck = 20;
51 // The max limit of the search depth in DecomposeGEPExpression() and
52 // GetUnderlyingObject(), both functions need to use the same search
53 // depth otherwise the algorithm in aliasGEP will assert.
54 static const unsigned MaxLookupSearchDepth = 6;
56 //===----------------------------------------------------------------------===//
58 //===----------------------------------------------------------------------===//
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 // Note even if the argument is marked nocapture we still need to check
78 // for copies made inside the function. The nocapture attribute only
79 // specifies that there are no copies made that outlive the function.
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 DataLayout &DL,
103 const TargetLibraryInfo &TLI,
104 bool RoundToAlign = false) {
106 if (getObjectSize(V, Size, DL, &TLI, RoundToAlign))
108 return MemoryLocation::UnknownSize;
111 /// isObjectSmallerThan - Return true if we can prove that the object specified
112 /// by V is smaller than Size.
113 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
114 const DataLayout &DL,
115 const TargetLibraryInfo &TLI) {
116 // Note that the meanings of the "object" are slightly different in the
117 // following contexts:
118 // c1: llvm::getObjectSize()
119 // c2: llvm.objectsize() intrinsic
120 // c3: isObjectSmallerThan()
121 // c1 and c2 share the same meaning; however, the meaning of "object" in c3
122 // refers to the "entire object".
124 // Consider this example:
125 // char *p = (char*)malloc(100)
128 // In the context of c1 and c2, the "object" pointed by q refers to the
129 // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
131 // However, in the context of c3, the "object" refers to the chunk of memory
132 // being allocated. So, the "object" has 100 bytes, and q points to the middle
133 // the "object". In case q is passed to isObjectSmallerThan() as the 1st
134 // parameter, before the llvm::getObjectSize() is called to get the size of
135 // entire object, we should:
136 // - either rewind the pointer q to the base-address of the object in
137 // question (in this case rewind to p), or
138 // - just give up. It is up to caller to make sure the pointer is pointing
139 // to the base address the object.
141 // We go for 2nd option for simplicity.
142 if (!isIdentifiedObject(V))
145 // This function needs to use the aligned object size because we allow
146 // reads a bit past the end given sufficient alignment.
147 uint64_t ObjectSize = getObjectSize(V, DL, TLI, /*RoundToAlign*/true);
149 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size;
152 /// isObjectSize - Return true if we can prove that the object specified
153 /// by V has size Size.
154 static bool isObjectSize(const Value *V, uint64_t Size,
155 const DataLayout &DL, const TargetLibraryInfo &TLI) {
156 uint64_t ObjectSize = getObjectSize(V, DL, TLI);
157 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size;
160 //===----------------------------------------------------------------------===//
161 // GetElementPtr Instruction Decomposition and Analysis
162 //===----------------------------------------------------------------------===//
171 struct VariableGEPIndex {
173 ExtensionKind Extension;
176 bool operator==(const VariableGEPIndex &Other) const {
177 return V == Other.V && Extension == Other.Extension &&
178 Scale == Other.Scale;
181 bool operator!=(const VariableGEPIndex &Other) const {
182 return !operator==(Other);
188 /// GetLinearExpression - Analyze the specified value as a linear expression:
189 /// "A*V + B", where A and B are constant integers. Return the scale and offset
190 /// values as APInts and return V as a Value*, and return whether we looked
191 /// through any sign or zero extends. The incoming Value is known to have
192 /// IntegerType and it may already be sign or zero extended.
194 /// Note that this looks through extends, so the high bits may not be
195 /// represented in the result.
196 static Value *GetLinearExpression(Value *V, APInt &Scale, APInt &Offset,
197 ExtensionKind &Extension,
198 const DataLayout &DL, unsigned Depth,
199 AssumptionCache *AC, DominatorTree *DT) {
200 assert(V->getType()->isIntegerTy() && "Not an integer value");
202 // Limit our recursion depth.
209 if (ConstantInt *Const = dyn_cast<ConstantInt>(V)) {
210 // if it's a constant, just convert it to an offset
211 // and remove the variable.
212 Offset += Const->getValue();
213 assert(Scale == 0 && "Constant values don't have a scale");
217 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
218 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
219 switch (BOp->getOpcode()) {
221 case Instruction::Or:
222 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
224 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC,
228 case Instruction::Add:
229 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
230 DL, Depth + 1, AC, DT);
231 Offset += RHSC->getValue();
233 case Instruction::Mul:
234 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
235 DL, Depth + 1, AC, DT);
236 Offset *= RHSC->getValue();
237 Scale *= RHSC->getValue();
239 case Instruction::Shl:
240 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
241 DL, Depth + 1, AC, DT);
242 Offset <<= RHSC->getValue().getLimitedValue();
243 Scale <<= RHSC->getValue().getLimitedValue();
249 // Since GEP indices are sign extended anyway, we don't care about the high
250 // bits of a sign or zero extended value - just scales and offsets. The
251 // extensions have to be consistent though.
252 if ((isa<SExtInst>(V) && Extension != EK_ZeroExt) ||
253 (isa<ZExtInst>(V) && Extension != EK_SignExt)) {
254 Value *CastOp = cast<CastInst>(V)->getOperand(0);
255 unsigned OldWidth = Scale.getBitWidth();
256 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
257 Scale = Scale.trunc(SmallWidth);
258 Offset = Offset.trunc(SmallWidth);
259 Extension = isa<SExtInst>(V) ? EK_SignExt : EK_ZeroExt;
261 Value *Result = GetLinearExpression(CastOp, Scale, Offset, Extension, DL,
263 Scale = Scale.zext(OldWidth);
265 // We have to sign-extend even if Extension == EK_ZeroExt as we can't
266 // decompose a sign extension (i.e. zext(x - 1) != zext(x) - zext(-1)).
267 Offset = Offset.sext(OldWidth);
277 /// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it
278 /// into a base pointer with a constant offset and a number of scaled symbolic
281 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale in
282 /// the VarIndices vector) are Value*'s that are known to be scaled by the
283 /// specified amount, but which may have other unrepresented high bits. As such,
284 /// the gep cannot necessarily be reconstructed from its decomposed form.
286 /// When DataLayout is around, this function is capable of analyzing everything
287 /// that GetUnderlyingObject can look through. To be able to do that
288 /// GetUnderlyingObject and DecomposeGEPExpression must use the same search
289 /// depth (MaxLookupSearchDepth).
290 /// When DataLayout not is around, it just looks through pointer casts.
293 DecomposeGEPExpression(const Value *V, int64_t &BaseOffs,
294 SmallVectorImpl<VariableGEPIndex> &VarIndices,
295 bool &MaxLookupReached, const DataLayout &DL,
296 AssumptionCache *AC, DominatorTree *DT) {
297 // Limit recursion depth to limit compile time in crazy cases.
298 unsigned MaxLookup = MaxLookupSearchDepth;
299 MaxLookupReached = false;
303 // See if this is a bitcast or GEP.
304 const Operator *Op = dyn_cast<Operator>(V);
306 // The only non-operator case we can handle are GlobalAliases.
307 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
308 if (!GA->mayBeOverridden()) {
309 V = GA->getAliasee();
316 if (Op->getOpcode() == Instruction::BitCast ||
317 Op->getOpcode() == Instruction::AddrSpaceCast) {
318 V = Op->getOperand(0);
322 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
324 // If it's not a GEP, hand it off to SimplifyInstruction to see if it
325 // can come up with something. This matches what GetUnderlyingObject does.
326 if (const Instruction *I = dyn_cast<Instruction>(V))
327 // TODO: Get a DominatorTree and AssumptionCache and use them here
328 // (these are both now available in this function, but this should be
329 // updated when GetUnderlyingObject is updated). TLI should be
331 if (const Value *Simplified =
332 SimplifyInstruction(const_cast<Instruction *>(I), DL)) {
340 // Don't attempt to analyze GEPs over unsized objects.
341 if (!GEPOp->getOperand(0)->getType()->getPointerElementType()->isSized())
344 unsigned AS = GEPOp->getPointerAddressSpace();
345 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
346 gep_type_iterator GTI = gep_type_begin(GEPOp);
347 for (User::const_op_iterator I = GEPOp->op_begin()+1,
348 E = GEPOp->op_end(); I != E; ++I) {
350 // Compute the (potentially symbolic) offset in bytes for this index.
351 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
352 // For a struct, add the member offset.
353 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
354 if (FieldNo == 0) continue;
356 BaseOffs += DL.getStructLayout(STy)->getElementOffset(FieldNo);
360 // For an array/pointer, add the element offset, explicitly scaled.
361 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
362 if (CIdx->isZero()) continue;
363 BaseOffs += DL.getTypeAllocSize(*GTI) * CIdx->getSExtValue();
367 uint64_t Scale = DL.getTypeAllocSize(*GTI);
368 ExtensionKind Extension = EK_NotExtended;
370 // If the integer type is smaller than the pointer size, it is implicitly
371 // sign extended to pointer size.
372 unsigned Width = Index->getType()->getIntegerBitWidth();
373 if (DL.getPointerSizeInBits(AS) > Width)
374 Extension = EK_SignExt;
376 // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
377 APInt IndexScale(Width, 0), IndexOffset(Width, 0);
378 Index = GetLinearExpression(Index, IndexScale, IndexOffset, Extension, DL,
381 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
382 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
383 BaseOffs += IndexOffset.getSExtValue()*Scale;
384 Scale *= IndexScale.getSExtValue();
386 // If we already had an occurrence of this index variable, merge this
387 // scale into it. For example, we want to handle:
388 // A[x][x] -> x*16 + x*4 -> x*20
389 // This also ensures that 'x' only appears in the index list once.
390 for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
391 if (VarIndices[i].V == Index &&
392 VarIndices[i].Extension == Extension) {
393 Scale += VarIndices[i].Scale;
394 VarIndices.erase(VarIndices.begin()+i);
399 // Make sure that we have a scale that makes sense for this target's
401 if (unsigned ShiftBits = 64 - DL.getPointerSizeInBits(AS)) {
403 Scale = (int64_t)Scale >> ShiftBits;
407 VariableGEPIndex Entry = {Index, Extension,
408 static_cast<int64_t>(Scale)};
409 VarIndices.push_back(Entry);
413 // Analyze the base pointer next.
414 V = GEPOp->getOperand(0);
415 } while (--MaxLookup);
417 // If the chain of expressions is too deep, just return early.
418 MaxLookupReached = true;
422 //===----------------------------------------------------------------------===//
423 // BasicAliasAnalysis Pass
424 //===----------------------------------------------------------------------===//
427 static const Function *getParent(const Value *V) {
428 if (const Instruction *inst = dyn_cast<Instruction>(V))
429 return inst->getParent()->getParent();
431 if (const Argument *arg = dyn_cast<Argument>(V))
432 return arg->getParent();
437 static bool notDifferentParent(const Value *O1, const Value *O2) {
439 const Function *F1 = getParent(O1);
440 const Function *F2 = getParent(O2);
442 return !F1 || !F2 || F1 == F2;
447 /// BasicAliasAnalysis - This is the primary alias analysis implementation.
448 struct BasicAliasAnalysis : public ImmutablePass, public AliasAnalysis {
449 static char ID; // Class identification, replacement for typeinfo
450 BasicAliasAnalysis() : ImmutablePass(ID) {
451 initializeBasicAliasAnalysisPass(*PassRegistry::getPassRegistry());
454 bool doInitialization(Module &M) override;
456 void getAnalysisUsage(AnalysisUsage &AU) const override {
457 AU.addRequired<AliasAnalysis>();
458 AU.addRequired<AssumptionCacheTracker>();
459 AU.addRequired<TargetLibraryInfoWrapperPass>();
462 AliasResult alias(const MemoryLocation &LocA,
463 const MemoryLocation &LocB) override {
464 assert(AliasCache.empty() && "AliasCache must be cleared after use!");
465 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
466 "BasicAliasAnalysis doesn't support interprocedural queries.");
467 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.AATags,
468 LocB.Ptr, LocB.Size, LocB.AATags);
469 // AliasCache rarely has more than 1 or 2 elements, always use
470 // shrink_and_clear so it quickly returns to the inline capacity of the
471 // SmallDenseMap if it ever grows larger.
472 // FIXME: This should really be shrink_to_inline_capacity_and_clear().
473 AliasCache.shrink_and_clear();
474 VisitedPhiBBs.clear();
478 ModRefResult getModRefInfo(ImmutableCallSite CS,
479 const MemoryLocation &Loc) override;
481 ModRefResult getModRefInfo(ImmutableCallSite CS1,
482 ImmutableCallSite CS2) override;
484 /// pointsToConstantMemory - Chase pointers until we find a (constant
486 bool pointsToConstantMemory(const MemoryLocation &Loc,
487 bool OrLocal) override;
489 /// Get the location associated with a pointer argument of a callsite.
490 ModRefResult getArgModRefInfo(ImmutableCallSite CS,
491 unsigned ArgIdx) override;
493 /// getModRefBehavior - Return the behavior when calling the given
495 ModRefBehavior getModRefBehavior(ImmutableCallSite CS) override;
497 /// getModRefBehavior - Return the behavior when calling the given function.
498 /// For use when the call site is not known.
499 ModRefBehavior getModRefBehavior(const Function *F) override;
501 /// getAdjustedAnalysisPointer - This method is used when a pass implements
502 /// an analysis interface through multiple inheritance. If needed, it
503 /// should override this to adjust the this pointer as needed for the
504 /// specified pass info.
505 void *getAdjustedAnalysisPointer(const void *ID) override {
506 if (ID == &AliasAnalysis::ID)
507 return (AliasAnalysis*)this;
512 // AliasCache - Track alias queries to guard against recursion.
513 typedef std::pair<MemoryLocation, MemoryLocation> LocPair;
514 typedef SmallDenseMap<LocPair, AliasResult, 8> AliasCacheTy;
515 AliasCacheTy AliasCache;
517 /// \brief Track phi nodes we have visited. When interpret "Value" pointer
518 /// equality as value equality we need to make sure that the "Value" is not
519 /// part of a cycle. Otherwise, two uses could come from different
520 /// "iterations" of a cycle and see different values for the same "Value"
522 /// The following example shows the problem:
523 /// %p = phi(%alloca1, %addr2)
525 /// %addr1 = gep, %alloca2, 0, %l
526 /// %addr2 = gep %alloca2, 0, (%l + 1)
527 /// alias(%p, %addr1) -> MayAlias !
529 SmallPtrSet<const BasicBlock*, 8> VisitedPhiBBs;
531 // Visited - Track instructions visited by pointsToConstantMemory.
532 SmallPtrSet<const Value*, 16> Visited;
534 /// \brief Check whether two Values can be considered equivalent.
536 /// In addition to pointer equivalence of \p V1 and \p V2 this checks
537 /// whether they can not be part of a cycle in the value graph by looking at
538 /// all visited phi nodes an making sure that the phis cannot reach the
539 /// value. We have to do this because we are looking through phi nodes (That
540 /// is we say noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB).
541 bool isValueEqualInPotentialCycles(const Value *V1, const Value *V2);
543 /// \brief Dest and Src are the variable indices from two decomposed
544 /// GetElementPtr instructions GEP1 and GEP2 which have common base
545 /// pointers. Subtract the GEP2 indices from GEP1 to find the symbolic
546 /// difference between the two pointers.
547 void GetIndexDifference(SmallVectorImpl<VariableGEPIndex> &Dest,
548 const SmallVectorImpl<VariableGEPIndex> &Src);
550 // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP
551 // instruction against another.
552 AliasResult aliasGEP(const GEPOperator *V1, uint64_t V1Size,
553 const AAMDNodes &V1AAInfo,
554 const Value *V2, uint64_t V2Size,
555 const AAMDNodes &V2AAInfo,
556 const Value *UnderlyingV1, const Value *UnderlyingV2);
558 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI
559 // instruction against another.
560 AliasResult aliasPHI(const PHINode *PN, uint64_t PNSize,
561 const AAMDNodes &PNAAInfo,
562 const Value *V2, uint64_t V2Size,
563 const AAMDNodes &V2AAInfo);
565 /// aliasSelect - Disambiguate a Select instruction against another value.
566 AliasResult aliasSelect(const SelectInst *SI, uint64_t SISize,
567 const AAMDNodes &SIAAInfo,
568 const Value *V2, uint64_t V2Size,
569 const AAMDNodes &V2AAInfo);
571 AliasResult aliasCheck(const Value *V1, uint64_t V1Size,
573 const Value *V2, uint64_t V2Size,
576 } // End of anonymous namespace
578 // Register this pass...
579 char BasicAliasAnalysis::ID = 0;
580 INITIALIZE_AG_PASS_BEGIN(BasicAliasAnalysis, AliasAnalysis, "basicaa",
581 "Basic Alias Analysis (stateless AA impl)",
583 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
584 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
585 INITIALIZE_AG_PASS_END(BasicAliasAnalysis, AliasAnalysis, "basicaa",
586 "Basic Alias Analysis (stateless AA impl)",
590 ImmutablePass *llvm::createBasicAliasAnalysisPass() {
591 return new BasicAliasAnalysis();
594 /// pointsToConstantMemory - Returns whether the given pointer value
595 /// points to memory that is local to the function, with global constants being
596 /// considered local to all functions.
597 bool BasicAliasAnalysis::pointsToConstantMemory(const MemoryLocation &Loc,
599 assert(Visited.empty() && "Visited must be cleared after use!");
601 unsigned MaxLookup = 8;
602 SmallVector<const Value *, 16> Worklist;
603 Worklist.push_back(Loc.Ptr);
605 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), *DL);
606 if (!Visited.insert(V).second) {
608 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
611 // An alloca instruction defines local memory.
612 if (OrLocal && isa<AllocaInst>(V))
615 // A global constant counts as local memory for our purposes.
616 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
617 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
618 // global to be marked constant in some modules and non-constant in
619 // others. GV may even be a declaration, not a definition.
620 if (!GV->isConstant()) {
622 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
627 // If both select values point to local memory, then so does the select.
628 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
629 Worklist.push_back(SI->getTrueValue());
630 Worklist.push_back(SI->getFalseValue());
634 // If all values incoming to a phi node point to local memory, then so does
636 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
637 // Don't bother inspecting phi nodes with many operands.
638 if (PN->getNumIncomingValues() > MaxLookup) {
640 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
642 for (Value *IncValue : PN->incoming_values())
643 Worklist.push_back(IncValue);
647 // Otherwise be conservative.
649 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
651 } while (!Worklist.empty() && --MaxLookup);
654 return Worklist.empty();
657 // FIXME: This code is duplicated with MemoryLocation and should be hoisted to
658 // some common utility location.
659 static bool isMemsetPattern16(const Function *MS,
660 const TargetLibraryInfo &TLI) {
661 if (TLI.has(LibFunc::memset_pattern16) &&
662 MS->getName() == "memset_pattern16") {
663 FunctionType *MemsetType = MS->getFunctionType();
664 if (!MemsetType->isVarArg() && MemsetType->getNumParams() == 3 &&
665 isa<PointerType>(MemsetType->getParamType(0)) &&
666 isa<PointerType>(MemsetType->getParamType(1)) &&
667 isa<IntegerType>(MemsetType->getParamType(2)))
674 /// getModRefBehavior - Return the behavior when calling the given call site.
675 AliasAnalysis::ModRefBehavior
676 BasicAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
677 if (CS.doesNotAccessMemory())
678 // Can't do better than this.
679 return DoesNotAccessMemory;
681 ModRefBehavior Min = UnknownModRefBehavior;
683 // If the callsite knows it only reads memory, don't return worse
685 if (CS.onlyReadsMemory())
686 Min = OnlyReadsMemory;
688 // The AliasAnalysis base class has some smarts, lets use them.
689 return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
692 /// getModRefBehavior - Return the behavior when calling the given function.
693 /// For use when the call site is not known.
694 AliasAnalysis::ModRefBehavior
695 BasicAliasAnalysis::getModRefBehavior(const Function *F) {
696 // If the function declares it doesn't access memory, we can't do better.
697 if (F->doesNotAccessMemory())
698 return DoesNotAccessMemory;
700 // For intrinsics, we can check the table.
701 if (Intrinsic::ID iid = F->getIntrinsicID()) {
702 #define GET_INTRINSIC_MODREF_BEHAVIOR
703 #include "llvm/IR/Intrinsics.gen"
704 #undef GET_INTRINSIC_MODREF_BEHAVIOR
707 ModRefBehavior Min = UnknownModRefBehavior;
709 // If the function declares it only reads memory, go with that.
710 if (F->onlyReadsMemory())
711 Min = OnlyReadsMemory;
713 const TargetLibraryInfo &TLI =
714 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
715 if (isMemsetPattern16(F, TLI))
716 Min = OnlyAccessesArgumentPointees;
718 // Otherwise be conservative.
719 return ModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min);
722 AliasAnalysis::ModRefResult
723 BasicAliasAnalysis::getArgModRefInfo(ImmutableCallSite CS, unsigned ArgIdx) {
724 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction()))
725 switch (II->getIntrinsicID()) {
728 case Intrinsic::memset:
729 case Intrinsic::memcpy:
730 case Intrinsic::memmove:
731 assert((ArgIdx == 0 || ArgIdx == 1) &&
732 "Invalid argument index for memory intrinsic");
733 return ArgIdx ? Ref : Mod;
736 // We can bound the aliasing properties of memset_pattern16 just as we can
737 // for memcpy/memset. This is particularly important because the
738 // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
739 // whenever possible.
740 if (CS.getCalledFunction() &&
741 isMemsetPattern16(CS.getCalledFunction(), *TLI)) {
742 assert((ArgIdx == 0 || ArgIdx == 1) &&
743 "Invalid argument index for memset_pattern16");
744 return ArgIdx ? Ref : Mod;
746 // FIXME: Handle memset_pattern4 and memset_pattern8 also.
748 return AliasAnalysis::getArgModRefInfo(CS, ArgIdx);
751 static bool isAssumeIntrinsic(ImmutableCallSite CS) {
752 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
753 if (II && II->getIntrinsicID() == Intrinsic::assume)
759 bool BasicAliasAnalysis::doInitialization(Module &M) {
760 InitializeAliasAnalysis(this, &M.getDataLayout());
764 /// getModRefInfo - Check to see if the specified callsite can clobber the
765 /// specified memory object. Since we only look at local properties of this
766 /// function, we really can't say much about this query. We do, however, use
767 /// simple "address taken" analysis on local objects.
768 AliasAnalysis::ModRefResult
769 BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS,
770 const MemoryLocation &Loc) {
771 assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
772 "AliasAnalysis query involving multiple functions!");
774 const Value *Object = GetUnderlyingObject(Loc.Ptr, *DL);
776 // If this is a tail call and Loc.Ptr points to a stack location, we know that
777 // the tail call cannot access or modify the local stack.
778 // We cannot exclude byval arguments here; these belong to the caller of
779 // the current function not to the current function, and a tail callee
780 // may reference them.
781 if (isa<AllocaInst>(Object))
782 if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
783 if (CI->isTailCall())
786 // If the pointer is to a locally allocated object that does not escape,
787 // then the call can not mod/ref the pointer unless the call takes the pointer
788 // as an argument, and itself doesn't capture it.
789 if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
790 isNonEscapingLocalObject(Object)) {
791 bool PassedAsArg = false;
793 for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
794 CI != CE; ++CI, ++ArgNo) {
795 // Only look at the no-capture or byval pointer arguments. If this
796 // pointer were passed to arguments that were neither of these, then it
797 // couldn't be no-capture.
798 if (!(*CI)->getType()->isPointerTy() ||
799 (!CS.doesNotCapture(ArgNo) && !CS.isByValArgument(ArgNo)))
802 // If this is a no-capture pointer argument, see if we can tell that it
803 // is impossible to alias the pointer we're checking. If not, we have to
804 // assume that the call could touch the pointer, even though it doesn't
806 if (!isNoAlias(MemoryLocation(*CI), MemoryLocation(Object))) {
816 // While the assume intrinsic is marked as arbitrarily writing so that
817 // proper control dependencies will be maintained, it never aliases any
818 // particular memory location.
819 if (isAssumeIntrinsic(CS))
822 // The AliasAnalysis base class has some smarts, lets use them.
823 return AliasAnalysis::getModRefInfo(CS, Loc);
826 AliasAnalysis::ModRefResult
827 BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS1,
828 ImmutableCallSite CS2) {
829 // While the assume intrinsic is marked as arbitrarily writing so that
830 // proper control dependencies will be maintained, it never aliases any
831 // particular memory location.
832 if (isAssumeIntrinsic(CS1) || isAssumeIntrinsic(CS2))
835 // The AliasAnalysis base class has some smarts, lets use them.
836 return AliasAnalysis::getModRefInfo(CS1, CS2);
839 /// \brief Provide ad-hoc rules to disambiguate accesses through two GEP
840 /// operators, both having the exact same pointer operand.
841 static AliasAnalysis::AliasResult
842 aliasSameBasePointerGEPs(const GEPOperator *GEP1, uint64_t V1Size,
843 const GEPOperator *GEP2, uint64_t V2Size,
844 const DataLayout &DL) {
846 assert(GEP1->getPointerOperand() == GEP2->getPointerOperand() &&
847 "Expected GEPs with the same pointer operand");
849 // Try to determine whether GEP1 and GEP2 index through arrays, into structs,
850 // such that the struct field accesses provably cannot alias.
851 // We also need at least two indices (the pointer, and the struct field).
852 if (GEP1->getNumIndices() != GEP2->getNumIndices() ||
853 GEP1->getNumIndices() < 2)
854 return AliasAnalysis::MayAlias;
856 // If we don't know the size of the accesses through both GEPs, we can't
857 // determine whether the struct fields accessed can't alias.
858 if (V1Size == MemoryLocation::UnknownSize ||
859 V2Size == MemoryLocation::UnknownSize)
860 return AliasAnalysis::MayAlias;
863 dyn_cast<ConstantInt>(GEP1->getOperand(GEP1->getNumOperands() - 1));
865 dyn_cast<ConstantInt>(GEP2->getOperand(GEP2->getNumOperands() - 1));
867 // If the last (struct) indices aren't constants, we can't say anything.
868 // If they're identical, the other indices might be also be dynamically
869 // equal, so the GEPs can alias.
870 if (!C1 || !C2 || C1 == C2)
871 return AliasAnalysis::MayAlias;
873 // Find the last-indexed type of the GEP, i.e., the type you'd get if
874 // you stripped the last index.
875 // On the way, look at each indexed type. If there's something other
876 // than an array, different indices can lead to different final types.
877 SmallVector<Value *, 8> IntermediateIndices;
879 // Insert the first index; we don't need to check the type indexed
880 // through it as it only drops the pointer indirection.
881 assert(GEP1->getNumIndices() > 1 && "Not enough GEP indices to examine");
882 IntermediateIndices.push_back(GEP1->getOperand(1));
884 // Insert all the remaining indices but the last one.
885 // Also, check that they all index through arrays.
886 for (unsigned i = 1, e = GEP1->getNumIndices() - 1; i != e; ++i) {
887 if (!isa<ArrayType>(GetElementPtrInst::getIndexedType(
888 GEP1->getSourceElementType(), IntermediateIndices)))
889 return AliasAnalysis::MayAlias;
890 IntermediateIndices.push_back(GEP1->getOperand(i + 1));
893 StructType *LastIndexedStruct =
894 dyn_cast<StructType>(GetElementPtrInst::getIndexedType(
895 GEP1->getSourceElementType(), IntermediateIndices));
897 if (!LastIndexedStruct)
898 return AliasAnalysis::MayAlias;
901 // - both GEPs begin indexing from the exact same pointer;
902 // - the last indices in both GEPs are constants, indexing into a struct;
903 // - said indices are different, hence, the pointed-to fields are different;
904 // - both GEPs only index through arrays prior to that.
906 // This lets us determine that the struct that GEP1 indexes into and the
907 // struct that GEP2 indexes into must either precisely overlap or be
908 // completely disjoint. Because they cannot partially overlap, indexing into
909 // different non-overlapping fields of the struct will never alias.
911 // Therefore, the only remaining thing needed to show that both GEPs can't
912 // alias is that the fields are not overlapping.
913 const StructLayout *SL = DL.getStructLayout(LastIndexedStruct);
914 const uint64_t StructSize = SL->getSizeInBytes();
915 const uint64_t V1Off = SL->getElementOffset(C1->getZExtValue());
916 const uint64_t V2Off = SL->getElementOffset(C2->getZExtValue());
918 auto EltsDontOverlap = [StructSize](uint64_t V1Off, uint64_t V1Size,
919 uint64_t V2Off, uint64_t V2Size) {
920 return V1Off < V2Off && V1Off + V1Size <= V2Off &&
921 ((V2Off + V2Size <= StructSize) ||
922 (V2Off + V2Size - StructSize <= V1Off));
925 if (EltsDontOverlap(V1Off, V1Size, V2Off, V2Size) ||
926 EltsDontOverlap(V2Off, V2Size, V1Off, V1Size))
927 return AliasAnalysis::NoAlias;
929 return AliasAnalysis::MayAlias;
932 /// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
933 /// against another pointer. We know that V1 is a GEP, but we don't know
934 /// anything about V2. UnderlyingV1 is GetUnderlyingObject(GEP1, DL),
935 /// UnderlyingV2 is the same for V2.
937 AliasAnalysis::AliasResult
938 BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
939 const AAMDNodes &V1AAInfo,
940 const Value *V2, uint64_t V2Size,
941 const AAMDNodes &V2AAInfo,
942 const Value *UnderlyingV1,
943 const Value *UnderlyingV2) {
944 int64_t GEP1BaseOffset;
945 bool GEP1MaxLookupReached;
946 SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
948 // We have to get two AssumptionCaches here because GEP1 and V2 may be from
949 // different functions.
950 // FIXME: This really doesn't make any sense. We get a dominator tree below
951 // that can only refer to a single function. But this function (aliasGEP) is
952 // a method on an immutable pass that can be called when there *isn't*
953 // a single function. The old pass management layer makes this "work", but
954 // this isn't really a clean solution.
955 AssumptionCacheTracker &ACT = getAnalysis<AssumptionCacheTracker>();
956 AssumptionCache *AC1 = nullptr, *AC2 = nullptr;
957 if (auto *GEP1I = dyn_cast<Instruction>(GEP1))
958 AC1 = &ACT.getAssumptionCache(
959 const_cast<Function &>(*GEP1I->getParent()->getParent()));
960 if (auto *I2 = dyn_cast<Instruction>(V2))
961 AC2 = &ACT.getAssumptionCache(
962 const_cast<Function &>(*I2->getParent()->getParent()));
964 DominatorTreeWrapperPass *DTWP =
965 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
966 DominatorTree *DT = DTWP ? &DTWP->getDomTree() : nullptr;
968 // If we have two gep instructions with must-alias or not-alias'ing base
969 // pointers, figure out if the indexes to the GEP tell us anything about the
971 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
972 // Do the base pointers alias?
973 AliasResult BaseAlias =
974 aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize, AAMDNodes(),
975 UnderlyingV2, MemoryLocation::UnknownSize, AAMDNodes());
977 // Check for geps of non-aliasing underlying pointers where the offsets are
979 if ((BaseAlias == MayAlias) && V1Size == V2Size) {
980 // Do the base pointers alias assuming type and size.
981 AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size,
982 V1AAInfo, UnderlyingV2,
984 if (PreciseBaseAlias == NoAlias) {
985 // See if the computed offset from the common pointer tells us about the
986 // relation of the resulting pointer.
987 int64_t GEP2BaseOffset;
988 bool GEP2MaxLookupReached;
989 SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
990 const Value *GEP2BasePtr =
991 DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices,
992 GEP2MaxLookupReached, *DL, AC2, DT);
993 const Value *GEP1BasePtr =
994 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
995 GEP1MaxLookupReached, *DL, AC1, DT);
996 // DecomposeGEPExpression and GetUnderlyingObject should return the
997 // same result except when DecomposeGEPExpression has no DataLayout.
998 if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
1000 "DecomposeGEPExpression and GetUnderlyingObject disagree!");
1003 // If the max search depth is reached the result is undefined
1004 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1008 if (GEP1BaseOffset == GEP2BaseOffset &&
1009 GEP1VariableIndices == GEP2VariableIndices)
1011 GEP1VariableIndices.clear();
1015 // If we get a No or May, then return it immediately, no amount of analysis
1016 // will improve this situation.
1017 if (BaseAlias != MustAlias) return BaseAlias;
1019 // Otherwise, we have a MustAlias. Since the base pointers alias each other
1020 // exactly, see if the computed offset from the common pointer tells us
1021 // about the relation of the resulting pointer.
1022 const Value *GEP1BasePtr =
1023 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
1024 GEP1MaxLookupReached, *DL, AC1, DT);
1026 int64_t GEP2BaseOffset;
1027 bool GEP2MaxLookupReached;
1028 SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
1029 const Value *GEP2BasePtr =
1030 DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices,
1031 GEP2MaxLookupReached, *DL, AC2, DT);
1033 // DecomposeGEPExpression and GetUnderlyingObject should return the
1034 // same result except when DecomposeGEPExpression has no DataLayout.
1035 if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
1037 "DecomposeGEPExpression and GetUnderlyingObject disagree!");
1041 // If we know the two GEPs are based off of the exact same pointer (and not
1042 // just the same underlying object), see if that tells us anything about
1043 // the resulting pointers.
1044 if (DL && GEP1->getPointerOperand() == GEP2->getPointerOperand()) {
1045 AliasResult R = aliasSameBasePointerGEPs(GEP1, V1Size, GEP2, V2Size, *DL);
1046 // If we couldn't find anything interesting, don't abandon just yet.
1051 // If the max search depth is reached the result is undefined
1052 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1055 // Subtract the GEP2 pointer from the GEP1 pointer to find out their
1056 // symbolic difference.
1057 GEP1BaseOffset -= GEP2BaseOffset;
1058 GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
1061 // Check to see if these two pointers are related by the getelementptr
1062 // instruction. If one pointer is a GEP with a non-zero index of the other
1063 // pointer, we know they cannot alias.
1065 // If both accesses are unknown size, we can't do anything useful here.
1066 if (V1Size == MemoryLocation::UnknownSize &&
1067 V2Size == MemoryLocation::UnknownSize)
1070 AliasResult R = aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize,
1071 AAMDNodes(), V2, V2Size, V2AAInfo);
1073 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
1074 // If V2 is known not to alias GEP base pointer, then the two values
1075 // cannot alias per GEP semantics: "A pointer value formed from a
1076 // getelementptr instruction is associated with the addresses associated
1077 // with the first operand of the getelementptr".
1080 const Value *GEP1BasePtr =
1081 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
1082 GEP1MaxLookupReached, *DL, AC1, DT);
1084 // DecomposeGEPExpression and GetUnderlyingObject should return the
1085 // same result except when DecomposeGEPExpression has no DataLayout.
1086 if (GEP1BasePtr != UnderlyingV1) {
1088 "DecomposeGEPExpression and GetUnderlyingObject disagree!");
1091 // If the max search depth is reached the result is undefined
1092 if (GEP1MaxLookupReached)
1096 // In the two GEP Case, if there is no difference in the offsets of the
1097 // computed pointers, the resultant pointers are a must alias. This
1098 // hapens when we have two lexically identical GEP's (for example).
1100 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
1101 // must aliases the GEP, the end result is a must alias also.
1102 if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
1105 // If there is a constant difference between the pointers, but the difference
1106 // is less than the size of the associated memory object, then we know
1107 // that the objects are partially overlapping. If the difference is
1108 // greater, we know they do not overlap.
1109 if (GEP1BaseOffset != 0 && GEP1VariableIndices.empty()) {
1110 if (GEP1BaseOffset >= 0) {
1111 if (V2Size != MemoryLocation::UnknownSize) {
1112 if ((uint64_t)GEP1BaseOffset < V2Size)
1113 return PartialAlias;
1117 // We have the situation where:
1120 // ---------------->|
1121 // |-->V1Size |-------> V2Size
1123 // We need to know that V2Size is not unknown, otherwise we might have
1124 // stripped a gep with negative index ('gep <ptr>, -1, ...).
1125 if (V1Size != MemoryLocation::UnknownSize &&
1126 V2Size != MemoryLocation::UnknownSize) {
1127 if (-(uint64_t)GEP1BaseOffset < V1Size)
1128 return PartialAlias;
1134 if (!GEP1VariableIndices.empty()) {
1135 uint64_t Modulo = 0;
1136 bool AllPositive = true;
1137 for (unsigned i = 0, e = GEP1VariableIndices.size(); i != e; ++i) {
1139 // Try to distinguish something like &A[i][1] against &A[42][0].
1140 // Grab the least significant bit set in any of the scales. We
1141 // don't need std::abs here (even if the scale's negative) as we'll
1142 // be ^'ing Modulo with itself later.
1143 Modulo |= (uint64_t) GEP1VariableIndices[i].Scale;
1146 // If the Value could change between cycles, then any reasoning about
1147 // the Value this cycle may not hold in the next cycle. We'll just
1148 // give up if we can't determine conditions that hold for every cycle:
1149 const Value *V = GEP1VariableIndices[i].V;
1151 bool SignKnownZero, SignKnownOne;
1152 ComputeSignBit(const_cast<Value *>(V), SignKnownZero, SignKnownOne, *DL,
1153 0, AC1, nullptr, DT);
1155 // Zero-extension widens the variable, and so forces the sign
1157 bool IsZExt = GEP1VariableIndices[i].Extension == EK_ZeroExt;
1158 SignKnownZero |= IsZExt;
1159 SignKnownOne &= !IsZExt;
1161 // If the variable begins with a zero then we know it's
1162 // positive, regardless of whether the value is signed or
1164 int64_t Scale = GEP1VariableIndices[i].Scale;
1166 (SignKnownZero && Scale >= 0) ||
1167 (SignKnownOne && Scale < 0);
1171 Modulo = Modulo ^ (Modulo & (Modulo - 1));
1173 // We can compute the difference between the two addresses
1174 // mod Modulo. Check whether that difference guarantees that the
1175 // two locations do not alias.
1176 uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1);
1177 if (V1Size != MemoryLocation::UnknownSize &&
1178 V2Size != MemoryLocation::UnknownSize && ModOffset >= V2Size &&
1179 V1Size <= Modulo - ModOffset)
1182 // If we know all the variables are positive, then GEP1 >= GEP1BasePtr.
1183 // If GEP1BasePtr > V2 (GEP1BaseOffset > 0) then we know the pointers
1184 // don't alias if V2Size can fit in the gap between V2 and GEP1BasePtr.
1185 if (AllPositive && GEP1BaseOffset > 0 && V2Size <= (uint64_t) GEP1BaseOffset)
1189 // Statically, we can see that the base objects are the same, but the
1190 // pointers have dynamic offsets which we can't resolve. And none of our
1191 // little tricks above worked.
1193 // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the
1194 // practical effect of this is protecting TBAA in the case of dynamic
1195 // indices into arrays of unions or malloc'd memory.
1196 return PartialAlias;
1199 static AliasAnalysis::AliasResult
1200 MergeAliasResults(AliasAnalysis::AliasResult A, AliasAnalysis::AliasResult B) {
1201 // If the results agree, take it.
1204 // A mix of PartialAlias and MustAlias is PartialAlias.
1205 if ((A == AliasAnalysis::PartialAlias && B == AliasAnalysis::MustAlias) ||
1206 (B == AliasAnalysis::PartialAlias && A == AliasAnalysis::MustAlias))
1207 return AliasAnalysis::PartialAlias;
1208 // Otherwise, we don't know anything.
1209 return AliasAnalysis::MayAlias;
1212 /// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select
1213 /// instruction against another.
1214 AliasAnalysis::AliasResult
1215 BasicAliasAnalysis::aliasSelect(const SelectInst *SI, uint64_t SISize,
1216 const AAMDNodes &SIAAInfo,
1217 const Value *V2, uint64_t V2Size,
1218 const AAMDNodes &V2AAInfo) {
1219 // If the values are Selects with the same condition, we can do a more precise
1220 // check: just check for aliases between the values on corresponding arms.
1221 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1222 if (SI->getCondition() == SI2->getCondition()) {
1224 aliasCheck(SI->getTrueValue(), SISize, SIAAInfo,
1225 SI2->getTrueValue(), V2Size, V2AAInfo);
1226 if (Alias == MayAlias)
1228 AliasResult ThisAlias =
1229 aliasCheck(SI->getFalseValue(), SISize, SIAAInfo,
1230 SI2->getFalseValue(), V2Size, V2AAInfo);
1231 return MergeAliasResults(ThisAlias, Alias);
1234 // If both arms of the Select node NoAlias or MustAlias V2, then returns
1235 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1237 aliasCheck(V2, V2Size, V2AAInfo, SI->getTrueValue(), SISize, SIAAInfo);
1238 if (Alias == MayAlias)
1241 AliasResult ThisAlias =
1242 aliasCheck(V2, V2Size, V2AAInfo, SI->getFalseValue(), SISize, SIAAInfo);
1243 return MergeAliasResults(ThisAlias, Alias);
1246 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
1248 AliasAnalysis::AliasResult
1249 BasicAliasAnalysis::aliasPHI(const PHINode *PN, uint64_t PNSize,
1250 const AAMDNodes &PNAAInfo,
1251 const Value *V2, uint64_t V2Size,
1252 const AAMDNodes &V2AAInfo) {
1253 // Track phi nodes we have visited. We use this information when we determine
1254 // value equivalence.
1255 VisitedPhiBBs.insert(PN->getParent());
1257 // If the values are PHIs in the same block, we can do a more precise
1258 // as well as efficient check: just check for aliases between the values
1259 // on corresponding edges.
1260 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1261 if (PN2->getParent() == PN->getParent()) {
1262 LocPair Locs(MemoryLocation(PN, PNSize, PNAAInfo),
1263 MemoryLocation(V2, V2Size, V2AAInfo));
1265 std::swap(Locs.first, Locs.second);
1266 // Analyse the PHIs' inputs under the assumption that the PHIs are
1268 // If the PHIs are May/MustAlias there must be (recursively) an input
1269 // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
1270 // there must be an operation on the PHIs within the PHIs' value cycle
1271 // that causes a MayAlias.
1272 // Pretend the phis do not alias.
1273 AliasResult Alias = NoAlias;
1274 assert(AliasCache.count(Locs) &&
1275 "There must exist an entry for the phi node");
1276 AliasResult OrigAliasResult = AliasCache[Locs];
1277 AliasCache[Locs] = NoAlias;
1279 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1280 AliasResult ThisAlias =
1281 aliasCheck(PN->getIncomingValue(i), PNSize, PNAAInfo,
1282 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1284 Alias = MergeAliasResults(ThisAlias, Alias);
1285 if (Alias == MayAlias)
1289 // Reset if speculation failed.
1290 if (Alias != NoAlias)
1291 AliasCache[Locs] = OrigAliasResult;
1296 SmallPtrSet<Value*, 4> UniqueSrc;
1297 SmallVector<Value*, 4> V1Srcs;
1298 for (Value *PV1 : PN->incoming_values()) {
1299 if (isa<PHINode>(PV1))
1300 // If any of the source itself is a PHI, return MayAlias conservatively
1301 // to avoid compile time explosion. The worst possible case is if both
1302 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1303 // and 'n' are the number of PHI sources.
1305 if (UniqueSrc.insert(PV1).second)
1306 V1Srcs.push_back(PV1);
1309 AliasResult Alias = aliasCheck(V2, V2Size, V2AAInfo,
1310 V1Srcs[0], PNSize, PNAAInfo);
1311 // Early exit if the check of the first PHI source against V2 is MayAlias.
1312 // Other results are not possible.
1313 if (Alias == MayAlias)
1316 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1317 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1318 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1319 Value *V = V1Srcs[i];
1321 AliasResult ThisAlias = aliasCheck(V2, V2Size, V2AAInfo,
1322 V, PNSize, PNAAInfo);
1323 Alias = MergeAliasResults(ThisAlias, Alias);
1324 if (Alias == MayAlias)
1331 // aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
1332 // such as array references.
1334 AliasAnalysis::AliasResult
1335 BasicAliasAnalysis::aliasCheck(const Value *V1, uint64_t V1Size,
1337 const Value *V2, uint64_t V2Size,
1338 AAMDNodes V2AAInfo) {
1339 // If either of the memory references is empty, it doesn't matter what the
1340 // pointer values are.
1341 if (V1Size == 0 || V2Size == 0)
1344 // Strip off any casts if they exist.
1345 V1 = V1->stripPointerCasts();
1346 V2 = V2->stripPointerCasts();
1348 // If V1 or V2 is undef, the result is NoAlias because we can always pick a
1349 // value for undef that aliases nothing in the program.
1350 if (isa<UndefValue>(V1) || isa<UndefValue>(V2))
1353 // Are we checking for alias of the same value?
1354 // Because we look 'through' phi nodes we could look at "Value" pointers from
1355 // different iterations. We must therefore make sure that this is not the
1356 // case. The function isValueEqualInPotentialCycles ensures that this cannot
1357 // happen by looking at the visited phi nodes and making sure they cannot
1359 if (isValueEqualInPotentialCycles(V1, V2))
1362 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1363 return NoAlias; // Scalars cannot alias each other
1365 // Figure out what objects these things are pointing to if we can.
1366 const Value *O1 = GetUnderlyingObject(V1, *DL, MaxLookupSearchDepth);
1367 const Value *O2 = GetUnderlyingObject(V2, *DL, MaxLookupSearchDepth);
1369 // Null values in the default address space don't point to any object, so they
1370 // don't alias any other pointer.
1371 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1372 if (CPN->getType()->getAddressSpace() == 0)
1374 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1375 if (CPN->getType()->getAddressSpace() == 0)
1379 // If V1/V2 point to two different objects we know that we have no alias.
1380 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1383 // Constant pointers can't alias with non-const isIdentifiedObject objects.
1384 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1385 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1388 // Function arguments can't alias with things that are known to be
1389 // unambigously identified at the function level.
1390 if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1391 (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1394 // Most objects can't alias null.
1395 if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
1396 (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
1399 // If one pointer is the result of a call/invoke or load and the other is a
1400 // non-escaping local object within the same function, then we know the
1401 // object couldn't escape to a point where the call could return it.
1403 // Note that if the pointers are in different functions, there are a
1404 // variety of complications. A call with a nocapture argument may still
1405 // temporary store the nocapture argument's value in a temporary memory
1406 // location if that memory location doesn't escape. Or it may pass a
1407 // nocapture value to other functions as long as they don't capture it.
1408 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1410 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1414 // If the size of one access is larger than the entire object on the other
1415 // side, then we know such behavior is undefined and can assume no alias.
1417 if ((V1Size != MemoryLocation::UnknownSize &&
1418 isObjectSmallerThan(O2, V1Size, *DL, *TLI)) ||
1419 (V2Size != MemoryLocation::UnknownSize &&
1420 isObjectSmallerThan(O1, V2Size, *DL, *TLI)))
1423 // Check the cache before climbing up use-def chains. This also terminates
1424 // otherwise infinitely recursive queries.
1425 LocPair Locs(MemoryLocation(V1, V1Size, V1AAInfo),
1426 MemoryLocation(V2, V2Size, V2AAInfo));
1428 std::swap(Locs.first, Locs.second);
1429 std::pair<AliasCacheTy::iterator, bool> Pair =
1430 AliasCache.insert(std::make_pair(Locs, MayAlias));
1432 return Pair.first->second;
1434 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1435 // GEP can't simplify, we don't even look at the PHI cases.
1436 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1438 std::swap(V1Size, V2Size);
1440 std::swap(V1AAInfo, V2AAInfo);
1442 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1443 AliasResult Result = aliasGEP(GV1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O1, O2);
1444 if (Result != MayAlias) return AliasCache[Locs] = Result;
1447 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1449 std::swap(V1Size, V2Size);
1450 std::swap(V1AAInfo, V2AAInfo);
1452 if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1453 AliasResult Result = aliasPHI(PN, V1Size, V1AAInfo,
1454 V2, V2Size, V2AAInfo);
1455 if (Result != MayAlias) return AliasCache[Locs] = Result;
1458 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1460 std::swap(V1Size, V2Size);
1461 std::swap(V1AAInfo, V2AAInfo);
1463 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1464 AliasResult Result = aliasSelect(S1, V1Size, V1AAInfo,
1465 V2, V2Size, V2AAInfo);
1466 if (Result != MayAlias) return AliasCache[Locs] = Result;
1469 // If both pointers are pointing into the same object and one of them
1470 // accesses is accessing the entire object, then the accesses must
1471 // overlap in some way.
1473 if ((V1Size != MemoryLocation::UnknownSize &&
1474 isObjectSize(O1, V1Size, *DL, *TLI)) ||
1475 (V2Size != MemoryLocation::UnknownSize &&
1476 isObjectSize(O2, V2Size, *DL, *TLI)))
1477 return AliasCache[Locs] = PartialAlias;
1479 AliasResult Result =
1480 AliasAnalysis::alias(MemoryLocation(V1, V1Size, V1AAInfo),
1481 MemoryLocation(V2, V2Size, V2AAInfo));
1482 return AliasCache[Locs] = Result;
1485 bool BasicAliasAnalysis::isValueEqualInPotentialCycles(const Value *V,
1490 const Instruction *Inst = dyn_cast<Instruction>(V);
1494 if (VisitedPhiBBs.empty())
1497 if (VisitedPhiBBs.size() > MaxNumPhiBBsValueReachabilityCheck)
1500 // Use dominance or loop info if available.
1501 DominatorTreeWrapperPass *DTWP =
1502 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
1503 DominatorTree *DT = DTWP ? &DTWP->getDomTree() : nullptr;
1504 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
1505 LoopInfo *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
1507 // Make sure that the visited phis cannot reach the Value. This ensures that
1508 // the Values cannot come from different iterations of a potential cycle the
1509 // phi nodes could be involved in.
1510 for (auto *P : VisitedPhiBBs)
1511 if (isPotentiallyReachable(P->begin(), Inst, DT, LI))
1517 /// GetIndexDifference - Dest and Src are the variable indices from two
1518 /// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base
1519 /// pointers. Subtract the GEP2 indices from GEP1 to find the symbolic
1520 /// difference between the two pointers.
1521 void BasicAliasAnalysis::GetIndexDifference(
1522 SmallVectorImpl<VariableGEPIndex> &Dest,
1523 const SmallVectorImpl<VariableGEPIndex> &Src) {
1527 for (unsigned i = 0, e = Src.size(); i != e; ++i) {
1528 const Value *V = Src[i].V;
1529 ExtensionKind Extension = Src[i].Extension;
1530 int64_t Scale = Src[i].Scale;
1532 // Find V in Dest. This is N^2, but pointer indices almost never have more
1533 // than a few variable indexes.
1534 for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
1535 if (!isValueEqualInPotentialCycles(Dest[j].V, V) ||
1536 Dest[j].Extension != Extension)
1539 // If we found it, subtract off Scale V's from the entry in Dest. If it
1540 // goes to zero, remove the entry.
1541 if (Dest[j].Scale != Scale)
1542 Dest[j].Scale -= Scale;
1544 Dest.erase(Dest.begin() + j);
1549 // If we didn't consume this entry, add it to the end of the Dest list.
1551 VariableGEPIndex Entry = { V, Extension, -Scale };
1552 Dest.push_back(Entry);