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/BasicAliasAnalysis.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/CFG.h"
21 #include "llvm/Analysis/CaptureTracking.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/Analysis/AssumptionCache.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/GlobalAlias.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/ErrorHandling.h"
42 /// Enable analysis of recursive PHI nodes.
43 static cl::opt<bool> EnableRecPhiAnalysis("basicaa-recphi", cl::Hidden,
46 /// SearchLimitReached / SearchTimes shows how often the limit of
47 /// to decompose GEPs is reached. It will affect the precision
48 /// of basic alias analysis.
49 #define DEBUG_TYPE "basicaa"
50 STATISTIC(SearchLimitReached, "Number of times the limit to "
51 "decompose GEPs is reached");
52 STATISTIC(SearchTimes, "Number of times a GEP is decomposed");
54 /// Cutoff after which to stop analysing a set of phi nodes potentially involved
55 /// in a cycle. Because we are analysing 'through' phi nodes we need to be
56 /// careful with value equivalence. We use reachability to make sure a value
57 /// cannot be involved in a cycle.
58 const unsigned MaxNumPhiBBsValueReachabilityCheck = 20;
60 // The max limit of the search depth in DecomposeGEPExpression() and
61 // GetUnderlyingObject(), both functions need to use the same search
62 // depth otherwise the algorithm in aliasGEP will assert.
63 static const unsigned MaxLookupSearchDepth = 6;
65 //===----------------------------------------------------------------------===//
67 //===----------------------------------------------------------------------===//
69 /// Returns true if the pointer is to a function-local object that never
70 /// escapes from the function.
71 static bool isNonEscapingLocalObject(const Value *V) {
72 // If this is a local allocation, check to see if it escapes.
73 if (isa<AllocaInst>(V) || isNoAliasCall(V))
74 // Set StoreCaptures to True so that we can assume in our callers that the
75 // pointer is not the result of a load instruction. Currently
76 // PointerMayBeCaptured doesn't have any special analysis for the
77 // StoreCaptures=false case; if it did, our callers could be refined to be
79 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
81 // If this is an argument that corresponds to a byval or noalias argument,
82 // then it has not escaped before entering the function. Check if it escapes
83 // inside the function.
84 if (const Argument *A = dyn_cast<Argument>(V))
85 if (A->hasByValAttr() || A->hasNoAliasAttr())
86 // Note even if the argument is marked nocapture we still need to check
87 // for copies made inside the function. The nocapture attribute only
88 // specifies that there are no copies made that outlive the function.
89 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
94 /// Returns true if the pointer is one which would have been considered an
95 /// escape by isNonEscapingLocalObject.
96 static bool isEscapeSource(const Value *V) {
97 if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
100 // The load case works because isNonEscapingLocalObject considers all
101 // stores to be escapes (it passes true for the StoreCaptures argument
102 // to PointerMayBeCaptured).
103 if (isa<LoadInst>(V))
109 /// Returns the size of the object specified by V, or UnknownSize if unknown.
110 static uint64_t getObjectSize(const Value *V, const DataLayout &DL,
111 const TargetLibraryInfo &TLI,
112 bool RoundToAlign = false) {
114 if (getObjectSize(V, Size, DL, &TLI, RoundToAlign))
116 return MemoryLocation::UnknownSize;
119 /// Returns true if we can prove that the object specified by V is smaller than
121 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
122 const DataLayout &DL,
123 const TargetLibraryInfo &TLI) {
124 // Note that the meanings of the "object" are slightly different in the
125 // following contexts:
126 // c1: llvm::getObjectSize()
127 // c2: llvm.objectsize() intrinsic
128 // c3: isObjectSmallerThan()
129 // c1 and c2 share the same meaning; however, the meaning of "object" in c3
130 // refers to the "entire object".
132 // Consider this example:
133 // char *p = (char*)malloc(100)
136 // In the context of c1 and c2, the "object" pointed by q refers to the
137 // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
139 // However, in the context of c3, the "object" refers to the chunk of memory
140 // being allocated. So, the "object" has 100 bytes, and q points to the middle
141 // the "object". In case q is passed to isObjectSmallerThan() as the 1st
142 // parameter, before the llvm::getObjectSize() is called to get the size of
143 // entire object, we should:
144 // - either rewind the pointer q to the base-address of the object in
145 // question (in this case rewind to p), or
146 // - just give up. It is up to caller to make sure the pointer is pointing
147 // to the base address the object.
149 // We go for 2nd option for simplicity.
150 if (!isIdentifiedObject(V))
153 // This function needs to use the aligned object size because we allow
154 // reads a bit past the end given sufficient alignment.
155 uint64_t ObjectSize = getObjectSize(V, DL, TLI, /*RoundToAlign*/ true);
157 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size;
160 /// Returns true if we can prove that the object specified by V has size Size.
161 static bool isObjectSize(const Value *V, uint64_t Size, const DataLayout &DL,
162 const TargetLibraryInfo &TLI) {
163 uint64_t ObjectSize = getObjectSize(V, DL, TLI);
164 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size;
167 //===----------------------------------------------------------------------===//
168 // GetElementPtr Instruction Decomposition and Analysis
169 //===----------------------------------------------------------------------===//
171 /// Analyzes the specified value as a linear expression: "A*V + B", where A and
172 /// B are constant integers.
174 /// Returns the scale and offset values as APInts and return V as a Value*, and
175 /// return whether we looked through any sign or zero extends. The incoming
176 /// Value is known to have IntegerType and it may already be sign or zero
179 /// Note that this looks through extends, so the high bits may not be
180 /// represented in the result.
181 /*static*/ const Value *BasicAAResult::GetLinearExpression(
182 const Value *V, APInt &Scale, APInt &Offset, unsigned &ZExtBits,
183 unsigned &SExtBits, const DataLayout &DL, unsigned Depth,
184 AssumptionCache *AC, DominatorTree *DT, bool &NSW, bool &NUW) {
185 assert(V->getType()->isIntegerTy() && "Not an integer value");
187 // Limit our recursion depth.
194 if (const ConstantInt *Const = dyn_cast<ConstantInt>(V)) {
195 // if it's a constant, just convert it to an offset and remove the variable.
196 // If we've been called recursively the Offset bit width will be greater
197 // than the constant's (the Offset's always as wide as the outermost call),
198 // so we'll zext here and process any extension in the isa<SExtInst> &
199 // isa<ZExtInst> cases below.
200 Offset += Const->getValue().zextOrSelf(Offset.getBitWidth());
201 assert(Scale == 0 && "Constant values don't have a scale");
205 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
206 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
208 // If we've been called recursively then Offset and Scale will be wider
209 // that the BOp operands. We'll always zext it here as we'll process sign
210 // extensions below (see the isa<SExtInst> / isa<ZExtInst> cases).
211 APInt RHS = RHSC->getValue().zextOrSelf(Offset.getBitWidth());
213 switch (BOp->getOpcode()) {
215 // We don't understand this instruction, so we can't decompose it any
220 case Instruction::Or:
221 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
223 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC,
230 case Instruction::Add:
231 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
232 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
235 case Instruction::Sub:
236 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
237 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
240 case Instruction::Mul:
241 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
242 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
246 case Instruction::Shl:
247 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
248 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
249 Offset <<= RHS.getLimitedValue();
250 Scale <<= RHS.getLimitedValue();
251 // the semantics of nsw and nuw for left shifts don't match those of
252 // multiplications, so we won't propagate them.
257 if (isa<OverflowingBinaryOperator>(BOp)) {
258 NUW &= BOp->hasNoUnsignedWrap();
259 NSW &= BOp->hasNoSignedWrap();
265 // Since GEP indices are sign extended anyway, we don't care about the high
266 // bits of a sign or zero extended value - just scales and offsets. The
267 // extensions have to be consistent though.
268 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) {
269 Value *CastOp = cast<CastInst>(V)->getOperand(0);
270 unsigned NewWidth = V->getType()->getPrimitiveSizeInBits();
271 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
272 unsigned OldZExtBits = ZExtBits, OldSExtBits = SExtBits;
273 const Value *Result =
274 GetLinearExpression(CastOp, Scale, Offset, ZExtBits, SExtBits, DL,
275 Depth + 1, AC, DT, NSW, NUW);
277 // zext(zext(%x)) == zext(%x), and similiarly for sext; we'll handle this
278 // by just incrementing the number of bits we've extended by.
279 unsigned ExtendedBy = NewWidth - SmallWidth;
281 if (isa<SExtInst>(V) && ZExtBits == 0) {
282 // sext(sext(%x, a), b) == sext(%x, a + b)
285 // We haven't sign-wrapped, so it's valid to decompose sext(%x + c)
286 // into sext(%x) + sext(c). We'll sext the Offset ourselves:
287 unsigned OldWidth = Offset.getBitWidth();
288 Offset = Offset.trunc(SmallWidth).sext(NewWidth).zextOrSelf(OldWidth);
290 // We may have signed-wrapped, so don't decompose sext(%x + c) into
291 // sext(%x) + sext(c)
295 ZExtBits = OldZExtBits;
296 SExtBits = OldSExtBits;
298 SExtBits += ExtendedBy;
300 // sext(zext(%x, a), b) = zext(zext(%x, a), b) = zext(%x, a + b)
303 // We may have unsigned-wrapped, so don't decompose zext(%x + c) into
304 // zext(%x) + zext(c)
308 ZExtBits = OldZExtBits;
309 SExtBits = OldSExtBits;
311 ZExtBits += ExtendedBy;
322 /// If V is a symbolic pointer expression, decompose it into a base pointer
323 /// with a constant offset and a number of scaled symbolic offsets.
325 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale
326 /// in the VarIndices vector) are Value*'s that are known to be scaled by the
327 /// specified amount, but which may have other unrepresented high bits. As
328 /// such, the gep cannot necessarily be reconstructed from its decomposed form.
330 /// When DataLayout is around, this function is capable of analyzing everything
331 /// that GetUnderlyingObject can look through. To be able to do that
332 /// GetUnderlyingObject and DecomposeGEPExpression must use the same search
333 /// depth (MaxLookupSearchDepth). When DataLayout not is around, it just looks
334 /// through pointer casts.
335 /*static*/ const Value *BasicAAResult::DecomposeGEPExpression(
336 const Value *V, int64_t &BaseOffs,
337 SmallVectorImpl<VariableGEPIndex> &VarIndices, bool &MaxLookupReached,
338 const DataLayout &DL, AssumptionCache *AC, DominatorTree *DT) {
339 // Limit recursion depth to limit compile time in crazy cases.
340 unsigned MaxLookup = MaxLookupSearchDepth;
341 MaxLookupReached = false;
346 // See if this is a bitcast or GEP.
347 const Operator *Op = dyn_cast<Operator>(V);
349 // The only non-operator case we can handle are GlobalAliases.
350 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
351 if (!GA->mayBeOverridden()) {
352 V = GA->getAliasee();
359 if (Op->getOpcode() == Instruction::BitCast ||
360 Op->getOpcode() == Instruction::AddrSpaceCast) {
361 V = Op->getOperand(0);
365 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
367 // If it's not a GEP, hand it off to SimplifyInstruction to see if it
368 // can come up with something. This matches what GetUnderlyingObject does.
369 if (const Instruction *I = dyn_cast<Instruction>(V))
370 // TODO: Get a DominatorTree and AssumptionCache and use them here
371 // (these are both now available in this function, but this should be
372 // updated when GetUnderlyingObject is updated). TLI should be
374 if (const Value *Simplified =
375 SimplifyInstruction(const_cast<Instruction *>(I), DL)) {
383 // Don't attempt to analyze GEPs over unsized objects.
384 if (!GEPOp->getOperand(0)->getType()->getPointerElementType()->isSized())
387 unsigned AS = GEPOp->getPointerAddressSpace();
388 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
389 gep_type_iterator GTI = gep_type_begin(GEPOp);
390 for (User::const_op_iterator I = GEPOp->op_begin() + 1, E = GEPOp->op_end();
392 const Value *Index = *I;
393 // Compute the (potentially symbolic) offset in bytes for this index.
394 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
395 // For a struct, add the member offset.
396 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
400 BaseOffs += DL.getStructLayout(STy)->getElementOffset(FieldNo);
404 // For an array/pointer, add the element offset, explicitly scaled.
405 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
408 BaseOffs += DL.getTypeAllocSize(*GTI) * CIdx->getSExtValue();
412 uint64_t Scale = DL.getTypeAllocSize(*GTI);
413 unsigned ZExtBits = 0, SExtBits = 0;
415 // If the integer type is smaller than the pointer size, it is implicitly
416 // sign extended to pointer size.
417 unsigned Width = Index->getType()->getIntegerBitWidth();
418 unsigned PointerSize = DL.getPointerSizeInBits(AS);
419 if (PointerSize > Width)
420 SExtBits += PointerSize - Width;
422 // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
423 APInt IndexScale(Width, 0), IndexOffset(Width, 0);
424 bool NSW = true, NUW = true;
425 Index = GetLinearExpression(Index, IndexScale, IndexOffset, ZExtBits,
426 SExtBits, DL, 0, AC, DT, NSW, NUW);
428 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
429 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
430 BaseOffs += IndexOffset.getSExtValue() * Scale;
431 Scale *= IndexScale.getSExtValue();
433 // If we already had an occurrence of this index variable, merge this
434 // scale into it. For example, we want to handle:
435 // A[x][x] -> x*16 + x*4 -> x*20
436 // This also ensures that 'x' only appears in the index list once.
437 for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
438 if (VarIndices[i].V == Index && VarIndices[i].ZExtBits == ZExtBits &&
439 VarIndices[i].SExtBits == SExtBits) {
440 Scale += VarIndices[i].Scale;
441 VarIndices.erase(VarIndices.begin() + i);
446 // Make sure that we have a scale that makes sense for this target's
448 if (unsigned ShiftBits = 64 - PointerSize) {
450 Scale = (int64_t)Scale >> ShiftBits;
454 VariableGEPIndex Entry = {Index, ZExtBits, SExtBits,
455 static_cast<int64_t>(Scale)};
456 VarIndices.push_back(Entry);
460 // Analyze the base pointer next.
461 V = GEPOp->getOperand(0);
462 } while (--MaxLookup);
464 // If the chain of expressions is too deep, just return early.
465 MaxLookupReached = true;
466 SearchLimitReached++;
470 /// Returns whether the given pointer value points to memory that is local to
471 /// the function, with global constants being considered local to all
473 bool BasicAAResult::pointsToConstantMemory(const MemoryLocation &Loc,
475 assert(Visited.empty() && "Visited must be cleared after use!");
477 unsigned MaxLookup = 8;
478 SmallVector<const Value *, 16> Worklist;
479 Worklist.push_back(Loc.Ptr);
481 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), DL);
482 if (!Visited.insert(V).second) {
484 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
487 // An alloca instruction defines local memory.
488 if (OrLocal && isa<AllocaInst>(V))
491 // A global constant counts as local memory for our purposes.
492 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
493 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
494 // global to be marked constant in some modules and non-constant in
495 // others. GV may even be a declaration, not a definition.
496 if (!GV->isConstant()) {
498 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
503 // If both select values point to local memory, then so does the select.
504 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
505 Worklist.push_back(SI->getTrueValue());
506 Worklist.push_back(SI->getFalseValue());
510 // If all values incoming to a phi node point to local memory, then so does
512 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
513 // Don't bother inspecting phi nodes with many operands.
514 if (PN->getNumIncomingValues() > MaxLookup) {
516 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
518 for (Value *IncValue : PN->incoming_values())
519 Worklist.push_back(IncValue);
523 // Otherwise be conservative.
525 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
527 } while (!Worklist.empty() && --MaxLookup);
530 return Worklist.empty();
533 // FIXME: This code is duplicated with MemoryLocation and should be hoisted to
534 // some common utility location.
535 static bool isMemsetPattern16(const Function *MS,
536 const TargetLibraryInfo &TLI) {
537 if (TLI.has(LibFunc::memset_pattern16) &&
538 MS->getName() == "memset_pattern16") {
539 FunctionType *MemsetType = MS->getFunctionType();
540 if (!MemsetType->isVarArg() && MemsetType->getNumParams() == 3 &&
541 isa<PointerType>(MemsetType->getParamType(0)) &&
542 isa<PointerType>(MemsetType->getParamType(1)) &&
543 isa<IntegerType>(MemsetType->getParamType(2)))
550 /// Returns the behavior when calling the given call site.
551 FunctionModRefBehavior BasicAAResult::getModRefBehavior(ImmutableCallSite CS) {
552 if (CS.doesNotAccessMemory())
553 // Can't do better than this.
554 return FMRB_DoesNotAccessMemory;
556 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
558 // If the callsite knows it only reads memory, don't return worse
560 if (CS.onlyReadsMemory())
561 Min = FMRB_OnlyReadsMemory;
563 if (CS.onlyAccessesArgMemory())
564 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
566 // The AAResultBase base class has some smarts, lets use them.
567 return FunctionModRefBehavior(AAResultBase::getModRefBehavior(CS) & Min);
570 /// Returns the behavior when calling the given function. For use when the call
571 /// site is not known.
572 FunctionModRefBehavior BasicAAResult::getModRefBehavior(const Function *F) {
573 // If the function declares it doesn't access memory, we can't do better.
574 if (F->doesNotAccessMemory())
575 return FMRB_DoesNotAccessMemory;
577 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
579 // If the function declares it only reads memory, go with that.
580 if (F->onlyReadsMemory())
581 Min = FMRB_OnlyReadsMemory;
583 if (F->onlyAccessesArgMemory())
584 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
586 if (isMemsetPattern16(F, TLI))
587 Min = FMRB_OnlyAccessesArgumentPointees;
589 // Otherwise be conservative.
590 return FunctionModRefBehavior(AAResultBase::getModRefBehavior(F) & Min);
593 ModRefInfo BasicAAResult::getArgModRefInfo(ImmutableCallSite CS,
595 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction()))
596 switch (II->getIntrinsicID()) {
599 case Intrinsic::memset:
600 case Intrinsic::memcpy:
601 case Intrinsic::memmove:
602 // We don't currently have a writeonly attribute. All other properties
603 // of these intrinsics are nicely described via attributes in
604 // Intrinsics.td and handled generically below.
609 // We can bound the aliasing properties of memset_pattern16 just as we can
610 // for memcpy/memset. This is particularly important because the
611 // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
612 // whenever possible.
613 if (CS.getCalledFunction() &&
614 isMemsetPattern16(CS.getCalledFunction(), TLI)) {
615 assert((ArgIdx == 0 || ArgIdx == 1) &&
616 "Invalid argument index for memset_pattern16");
617 return ArgIdx ? MRI_Ref : MRI_Mod;
619 // FIXME: Handle memset_pattern4 and memset_pattern8 also.
621 if (CS.paramHasAttr(ArgIdx + 1, Attribute::ReadOnly))
624 if (CS.paramHasAttr(ArgIdx + 1, Attribute::ReadNone))
627 return AAResultBase::getArgModRefInfo(CS, ArgIdx);
630 static bool isAssumeIntrinsic(ImmutableCallSite CS) {
631 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
632 return II && II->getIntrinsicID() == Intrinsic::assume;
636 static const Function *getParent(const Value *V) {
637 if (const Instruction *inst = dyn_cast<Instruction>(V))
638 return inst->getParent()->getParent();
640 if (const Argument *arg = dyn_cast<Argument>(V))
641 return arg->getParent();
646 static bool notDifferentParent(const Value *O1, const Value *O2) {
648 const Function *F1 = getParent(O1);
649 const Function *F2 = getParent(O2);
651 return !F1 || !F2 || F1 == F2;
655 AliasResult BasicAAResult::alias(const MemoryLocation &LocA,
656 const MemoryLocation &LocB) {
657 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
658 "BasicAliasAnalysis doesn't support interprocedural queries.");
660 // If we have a directly cached entry for these locations, we have recursed
661 // through this once, so just return the cached results. Notably, when this
662 // happens, we don't clear the cache.
663 auto CacheIt = AliasCache.find(LocPair(LocA, LocB));
664 if (CacheIt != AliasCache.end())
665 return CacheIt->second;
667 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.AATags, LocB.Ptr,
668 LocB.Size, LocB.AATags);
669 // AliasCache rarely has more than 1 or 2 elements, always use
670 // shrink_and_clear so it quickly returns to the inline capacity of the
671 // SmallDenseMap if it ever grows larger.
672 // FIXME: This should really be shrink_to_inline_capacity_and_clear().
673 AliasCache.shrink_and_clear();
674 VisitedPhiBBs.clear();
678 /// Checks to see if the specified callsite can clobber the specified memory
681 /// Since we only look at local properties of this function, we really can't
682 /// say much about this query. We do, however, use simple "address taken"
683 /// analysis on local objects.
684 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS,
685 const MemoryLocation &Loc) {
686 assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
687 "AliasAnalysis query involving multiple functions!");
689 const Value *Object = GetUnderlyingObject(Loc.Ptr, DL);
691 // If this is a tail call and Loc.Ptr points to a stack location, we know that
692 // the tail call cannot access or modify the local stack.
693 // We cannot exclude byval arguments here; these belong to the caller of
694 // the current function not to the current function, and a tail callee
695 // may reference them.
696 if (isa<AllocaInst>(Object))
697 if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
698 if (CI->isTailCall())
701 // If the pointer is to a locally allocated object that does not escape,
702 // then the call can not mod/ref the pointer unless the call takes the pointer
703 // as an argument, and itself doesn't capture it.
704 if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
705 isNonEscapingLocalObject(Object)) {
706 bool PassedAsArg = false;
708 for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
709 CI != CE; ++CI, ++ArgNo) {
710 // Only look at the no-capture or byval pointer arguments. If this
711 // pointer were passed to arguments that were neither of these, then it
712 // couldn't be no-capture.
713 if (!(*CI)->getType()->isPointerTy() ||
714 (!CS.doesNotCapture(ArgNo) && !CS.isByValArgument(ArgNo)))
717 // If this is a no-capture pointer argument, see if we can tell that it
718 // is impossible to alias the pointer we're checking. If not, we have to
719 // assume that the call could touch the pointer, even though it doesn't
722 getBestAAResults().alias(MemoryLocation(*CI), MemoryLocation(Object));
733 // While the assume intrinsic is marked as arbitrarily writing so that
734 // proper control dependencies will be maintained, it never aliases any
735 // particular memory location.
736 if (isAssumeIntrinsic(CS))
739 // The AAResultBase base class has some smarts, lets use them.
740 return AAResultBase::getModRefInfo(CS, Loc);
743 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS1,
744 ImmutableCallSite CS2) {
745 // While the assume intrinsic is marked as arbitrarily writing so that
746 // proper control dependencies will be maintained, it never aliases any
747 // particular memory location.
748 if (isAssumeIntrinsic(CS1) || isAssumeIntrinsic(CS2))
751 // The AAResultBase base class has some smarts, lets use them.
752 return AAResultBase::getModRefInfo(CS1, CS2);
755 /// Provide ad-hoc rules to disambiguate accesses through two GEP operators,
756 /// both having the exact same pointer operand.
757 static AliasResult aliasSameBasePointerGEPs(const GEPOperator *GEP1,
759 const GEPOperator *GEP2,
761 const DataLayout &DL) {
763 assert(GEP1->getPointerOperand() == GEP2->getPointerOperand() &&
764 "Expected GEPs with the same pointer operand");
766 // Try to determine whether GEP1 and GEP2 index through arrays, into structs,
767 // such that the struct field accesses provably cannot alias.
768 // We also need at least two indices (the pointer, and the struct field).
769 if (GEP1->getNumIndices() != GEP2->getNumIndices() ||
770 GEP1->getNumIndices() < 2)
773 // If we don't know the size of the accesses through both GEPs, we can't
774 // determine whether the struct fields accessed can't alias.
775 if (V1Size == MemoryLocation::UnknownSize ||
776 V2Size == MemoryLocation::UnknownSize)
780 dyn_cast<ConstantInt>(GEP1->getOperand(GEP1->getNumOperands() - 1));
782 dyn_cast<ConstantInt>(GEP2->getOperand(GEP2->getNumOperands() - 1));
784 // If the last (struct) indices are constants and are equal, the other indices
785 // might be also be dynamically equal, so the GEPs can alias.
786 if (C1 && C2 && C1 == C2)
789 // Find the last-indexed type of the GEP, i.e., the type you'd get if
790 // you stripped the last index.
791 // On the way, look at each indexed type. If there's something other
792 // than an array, different indices can lead to different final types.
793 SmallVector<Value *, 8> IntermediateIndices;
795 // Insert the first index; we don't need to check the type indexed
796 // through it as it only drops the pointer indirection.
797 assert(GEP1->getNumIndices() > 1 && "Not enough GEP indices to examine");
798 IntermediateIndices.push_back(GEP1->getOperand(1));
800 // Insert all the remaining indices but the last one.
801 // Also, check that they all index through arrays.
802 for (unsigned i = 1, e = GEP1->getNumIndices() - 1; i != e; ++i) {
803 if (!isa<ArrayType>(GetElementPtrInst::getIndexedType(
804 GEP1->getSourceElementType(), IntermediateIndices)))
806 IntermediateIndices.push_back(GEP1->getOperand(i + 1));
809 auto *Ty = GetElementPtrInst::getIndexedType(
810 GEP1->getSourceElementType(), IntermediateIndices);
811 StructType *LastIndexedStruct = dyn_cast<StructType>(Ty);
813 if (isa<SequentialType>(Ty)) {
815 // - both GEPs begin indexing from the exact same pointer;
816 // - the last indices in both GEPs are constants, indexing into a sequential
817 // type (array or pointer);
818 // - both GEPs only index through arrays prior to that.
820 // Because array indices greater than the number of elements are valid in
821 // GEPs, unless we know the intermediate indices are identical between
822 // GEP1 and GEP2 we cannot guarantee that the last indexed arrays don't
823 // partially overlap. We also need to check that the loaded size matches
824 // the element size, otherwise we could still have overlap.
825 const uint64_t ElementSize =
826 DL.getTypeStoreSize(cast<SequentialType>(Ty)->getElementType());
827 if (V1Size != ElementSize || V2Size != ElementSize)
830 for (unsigned i = 0, e = GEP1->getNumIndices() - 1; i != e; ++i)
831 if (GEP1->getOperand(i + 1) != GEP2->getOperand(i + 1))
834 // Now we know that the array/pointer that GEP1 indexes into and that
835 // that GEP2 indexes into must either precisely overlap or be disjoint.
836 // Because they cannot partially overlap and because fields in an array
837 // cannot overlap, if we can prove the final indices are different between
838 // GEP1 and GEP2, we can conclude GEP1 and GEP2 don't alias.
840 // If the last indices are constants, we've already checked they don't
841 // equal each other so we can exit early.
844 if (isKnownNonEqual(GEP1->getOperand(GEP1->getNumOperands() - 1),
845 GEP2->getOperand(GEP2->getNumOperands() - 1),
849 } else if (!LastIndexedStruct || !C1 || !C2) {
854 // - both GEPs begin indexing from the exact same pointer;
855 // - the last indices in both GEPs are constants, indexing into a struct;
856 // - said indices are different, hence, the pointed-to fields are different;
857 // - both GEPs only index through arrays prior to that.
859 // This lets us determine that the struct that GEP1 indexes into and the
860 // struct that GEP2 indexes into must either precisely overlap or be
861 // completely disjoint. Because they cannot partially overlap, indexing into
862 // different non-overlapping fields of the struct will never alias.
864 // Therefore, the only remaining thing needed to show that both GEPs can't
865 // alias is that the fields are not overlapping.
866 const StructLayout *SL = DL.getStructLayout(LastIndexedStruct);
867 const uint64_t StructSize = SL->getSizeInBytes();
868 const uint64_t V1Off = SL->getElementOffset(C1->getZExtValue());
869 const uint64_t V2Off = SL->getElementOffset(C2->getZExtValue());
871 auto EltsDontOverlap = [StructSize](uint64_t V1Off, uint64_t V1Size,
872 uint64_t V2Off, uint64_t V2Size) {
873 return V1Off < V2Off && V1Off + V1Size <= V2Off &&
874 ((V2Off + V2Size <= StructSize) ||
875 (V2Off + V2Size - StructSize <= V1Off));
878 if (EltsDontOverlap(V1Off, V1Size, V2Off, V2Size) ||
879 EltsDontOverlap(V2Off, V2Size, V1Off, V1Size))
885 /// Provides a bunch of ad-hoc rules to disambiguate a GEP instruction against
888 /// We know that V1 is a GEP, but we don't know anything about V2.
889 /// UnderlyingV1 is GetUnderlyingObject(GEP1, DL), UnderlyingV2 is the same for
891 AliasResult BasicAAResult::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
892 const AAMDNodes &V1AAInfo, const Value *V2,
893 uint64_t V2Size, const AAMDNodes &V2AAInfo,
894 const Value *UnderlyingV1,
895 const Value *UnderlyingV2) {
896 int64_t GEP1BaseOffset;
897 bool GEP1MaxLookupReached;
898 SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
900 // If we have two gep instructions with must-alias or not-alias'ing base
901 // pointers, figure out if the indexes to the GEP tell us anything about the
903 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
904 // Do the base pointers alias?
905 AliasResult BaseAlias =
906 aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize, AAMDNodes(),
907 UnderlyingV2, MemoryLocation::UnknownSize, AAMDNodes());
909 // Check for geps of non-aliasing underlying pointers where the offsets are
911 if ((BaseAlias == MayAlias) && V1Size == V2Size) {
912 // Do the base pointers alias assuming type and size.
913 AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size, V1AAInfo,
914 UnderlyingV2, V2Size, V2AAInfo);
915 if (PreciseBaseAlias == NoAlias) {
916 // See if the computed offset from the common pointer tells us about the
917 // relation of the resulting pointer.
918 int64_t GEP2BaseOffset;
919 bool GEP2MaxLookupReached;
920 SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
921 const Value *GEP2BasePtr =
922 DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices,
923 GEP2MaxLookupReached, DL, &AC, DT);
924 const Value *GEP1BasePtr =
925 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
926 GEP1MaxLookupReached, DL, &AC, DT);
927 // DecomposeGEPExpression and GetUnderlyingObject should return the
928 // same result except when DecomposeGEPExpression has no DataLayout.
929 // FIXME: They always have a DataLayout so this should become an
931 if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
934 // If the max search depth is reached the result is undefined
935 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
939 if (GEP1BaseOffset == GEP2BaseOffset &&
940 GEP1VariableIndices == GEP2VariableIndices)
942 GEP1VariableIndices.clear();
946 // If we get a No or May, then return it immediately, no amount of analysis
947 // will improve this situation.
948 if (BaseAlias != MustAlias)
951 // Otherwise, we have a MustAlias. Since the base pointers alias each other
952 // exactly, see if the computed offset from the common pointer tells us
953 // about the relation of the resulting pointer.
954 const Value *GEP1BasePtr =
955 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
956 GEP1MaxLookupReached, DL, &AC, DT);
958 int64_t GEP2BaseOffset;
959 bool GEP2MaxLookupReached;
960 SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
961 const Value *GEP2BasePtr =
962 DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices,
963 GEP2MaxLookupReached, DL, &AC, DT);
965 // DecomposeGEPExpression and GetUnderlyingObject should return the
966 // same result except when DecomposeGEPExpression has no DataLayout.
967 // FIXME: They always have a DataLayout so this should become an assert.
968 if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
972 // If we know the two GEPs are based off of the exact same pointer (and not
973 // just the same underlying object), see if that tells us anything about
974 // the resulting pointers.
975 if (GEP1->getPointerOperand() == GEP2->getPointerOperand()) {
976 AliasResult R = aliasSameBasePointerGEPs(GEP1, V1Size, GEP2, V2Size, DL);
977 // If we couldn't find anything interesting, don't abandon just yet.
982 // If the max search depth is reached the result is undefined
983 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
986 // Subtract the GEP2 pointer from the GEP1 pointer to find out their
987 // symbolic difference.
988 GEP1BaseOffset -= GEP2BaseOffset;
989 GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
992 // Check to see if these two pointers are related by the getelementptr
993 // instruction. If one pointer is a GEP with a non-zero index of the other
994 // pointer, we know they cannot alias.
996 // If both accesses are unknown size, we can't do anything useful here.
997 if (V1Size == MemoryLocation::UnknownSize &&
998 V2Size == MemoryLocation::UnknownSize)
1001 AliasResult R = aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize,
1002 AAMDNodes(), V2, V2Size, V2AAInfo);
1004 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
1005 // If V2 is known not to alias GEP base pointer, then the two values
1006 // cannot alias per GEP semantics: "A pointer value formed from a
1007 // getelementptr instruction is associated with the addresses associated
1008 // with the first operand of the getelementptr".
1011 const Value *GEP1BasePtr =
1012 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
1013 GEP1MaxLookupReached, DL, &AC, DT);
1015 // DecomposeGEPExpression and GetUnderlyingObject should return the
1016 // same result except when DecomposeGEPExpression has no DataLayout.
1017 // FIXME: They always have a DataLayout so this should become an assert.
1018 if (GEP1BasePtr != UnderlyingV1) {
1021 // If the max search depth is reached the result is undefined
1022 if (GEP1MaxLookupReached)
1026 // In the two GEP Case, if there is no difference in the offsets of the
1027 // computed pointers, the resultant pointers are a must alias. This
1028 // hapens when we have two lexically identical GEP's (for example).
1030 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
1031 // must aliases the GEP, the end result is a must alias also.
1032 if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
1035 // If there is a constant difference between the pointers, but the difference
1036 // is less than the size of the associated memory object, then we know
1037 // that the objects are partially overlapping. If the difference is
1038 // greater, we know they do not overlap.
1039 if (GEP1BaseOffset != 0 && GEP1VariableIndices.empty()) {
1040 if (GEP1BaseOffset >= 0) {
1041 if (V2Size != MemoryLocation::UnknownSize) {
1042 if ((uint64_t)GEP1BaseOffset < V2Size)
1043 return PartialAlias;
1047 // We have the situation where:
1050 // ---------------->|
1051 // |-->V1Size |-------> V2Size
1053 // We need to know that V2Size is not unknown, otherwise we might have
1054 // stripped a gep with negative index ('gep <ptr>, -1, ...).
1055 if (V1Size != MemoryLocation::UnknownSize &&
1056 V2Size != MemoryLocation::UnknownSize) {
1057 if (-(uint64_t)GEP1BaseOffset < V1Size)
1058 return PartialAlias;
1064 if (!GEP1VariableIndices.empty()) {
1065 uint64_t Modulo = 0;
1066 bool AllPositive = true;
1067 for (unsigned i = 0, e = GEP1VariableIndices.size(); i != e; ++i) {
1069 // Try to distinguish something like &A[i][1] against &A[42][0].
1070 // Grab the least significant bit set in any of the scales. We
1071 // don't need std::abs here (even if the scale's negative) as we'll
1072 // be ^'ing Modulo with itself later.
1073 Modulo |= (uint64_t)GEP1VariableIndices[i].Scale;
1076 // If the Value could change between cycles, then any reasoning about
1077 // the Value this cycle may not hold in the next cycle. We'll just
1078 // give up if we can't determine conditions that hold for every cycle:
1079 const Value *V = GEP1VariableIndices[i].V;
1081 bool SignKnownZero, SignKnownOne;
1082 ComputeSignBit(const_cast<Value *>(V), SignKnownZero, SignKnownOne, DL,
1083 0, &AC, nullptr, DT);
1085 // Zero-extension widens the variable, and so forces the sign
1087 bool IsZExt = GEP1VariableIndices[i].ZExtBits > 0 || isa<ZExtInst>(V);
1088 SignKnownZero |= IsZExt;
1089 SignKnownOne &= !IsZExt;
1091 // If the variable begins with a zero then we know it's
1092 // positive, regardless of whether the value is signed or
1094 int64_t Scale = GEP1VariableIndices[i].Scale;
1096 (SignKnownZero && Scale >= 0) || (SignKnownOne && Scale < 0);
1100 Modulo = Modulo ^ (Modulo & (Modulo - 1));
1102 // We can compute the difference between the two addresses
1103 // mod Modulo. Check whether that difference guarantees that the
1104 // two locations do not alias.
1105 uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1);
1106 if (V1Size != MemoryLocation::UnknownSize &&
1107 V2Size != MemoryLocation::UnknownSize && ModOffset >= V2Size &&
1108 V1Size <= Modulo - ModOffset)
1111 // If we know all the variables are positive, then GEP1 >= GEP1BasePtr.
1112 // If GEP1BasePtr > V2 (GEP1BaseOffset > 0) then we know the pointers
1113 // don't alias if V2Size can fit in the gap between V2 and GEP1BasePtr.
1114 if (AllPositive && GEP1BaseOffset > 0 && V2Size <= (uint64_t)GEP1BaseOffset)
1117 if (constantOffsetHeuristic(GEP1VariableIndices, V1Size, V2Size,
1118 GEP1BaseOffset, &AC, DT))
1122 // Statically, we can see that the base objects are the same, but the
1123 // pointers have dynamic offsets which we can't resolve. And none of our
1124 // little tricks above worked.
1126 // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the
1127 // practical effect of this is protecting TBAA in the case of dynamic
1128 // indices into arrays of unions or malloc'd memory.
1129 return PartialAlias;
1132 static AliasResult MergeAliasResults(AliasResult A, AliasResult B) {
1133 // If the results agree, take it.
1136 // A mix of PartialAlias and MustAlias is PartialAlias.
1137 if ((A == PartialAlias && B == MustAlias) ||
1138 (B == PartialAlias && A == MustAlias))
1139 return PartialAlias;
1140 // Otherwise, we don't know anything.
1144 /// Provides a bunch of ad-hoc rules to disambiguate a Select instruction
1145 /// against another.
1146 AliasResult BasicAAResult::aliasSelect(const SelectInst *SI, uint64_t SISize,
1147 const AAMDNodes &SIAAInfo,
1148 const Value *V2, uint64_t V2Size,
1149 const AAMDNodes &V2AAInfo) {
1150 // If the values are Selects with the same condition, we can do a more precise
1151 // check: just check for aliases between the values on corresponding arms.
1152 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1153 if (SI->getCondition() == SI2->getCondition()) {
1154 AliasResult Alias = aliasCheck(SI->getTrueValue(), SISize, SIAAInfo,
1155 SI2->getTrueValue(), V2Size, V2AAInfo);
1156 if (Alias == MayAlias)
1158 AliasResult ThisAlias =
1159 aliasCheck(SI->getFalseValue(), SISize, SIAAInfo,
1160 SI2->getFalseValue(), V2Size, V2AAInfo);
1161 return MergeAliasResults(ThisAlias, Alias);
1164 // If both arms of the Select node NoAlias or MustAlias V2, then returns
1165 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1167 aliasCheck(V2, V2Size, V2AAInfo, SI->getTrueValue(), SISize, SIAAInfo);
1168 if (Alias == MayAlias)
1171 AliasResult ThisAlias =
1172 aliasCheck(V2, V2Size, V2AAInfo, SI->getFalseValue(), SISize, SIAAInfo);
1173 return MergeAliasResults(ThisAlias, Alias);
1176 /// Provide a bunch of ad-hoc rules to disambiguate a PHI instruction against
1178 AliasResult BasicAAResult::aliasPHI(const PHINode *PN, uint64_t PNSize,
1179 const AAMDNodes &PNAAInfo, const Value *V2,
1181 const AAMDNodes &V2AAInfo) {
1182 // Track phi nodes we have visited. We use this information when we determine
1183 // value equivalence.
1184 VisitedPhiBBs.insert(PN->getParent());
1186 // If the values are PHIs in the same block, we can do a more precise
1187 // as well as efficient check: just check for aliases between the values
1188 // on corresponding edges.
1189 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1190 if (PN2->getParent() == PN->getParent()) {
1191 LocPair Locs(MemoryLocation(PN, PNSize, PNAAInfo),
1192 MemoryLocation(V2, V2Size, V2AAInfo));
1194 std::swap(Locs.first, Locs.second);
1195 // Analyse the PHIs' inputs under the assumption that the PHIs are
1197 // If the PHIs are May/MustAlias there must be (recursively) an input
1198 // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
1199 // there must be an operation on the PHIs within the PHIs' value cycle
1200 // that causes a MayAlias.
1201 // Pretend the phis do not alias.
1202 AliasResult Alias = NoAlias;
1203 assert(AliasCache.count(Locs) &&
1204 "There must exist an entry for the phi node");
1205 AliasResult OrigAliasResult = AliasCache[Locs];
1206 AliasCache[Locs] = NoAlias;
1208 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1209 AliasResult ThisAlias =
1210 aliasCheck(PN->getIncomingValue(i), PNSize, PNAAInfo,
1211 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1213 Alias = MergeAliasResults(ThisAlias, Alias);
1214 if (Alias == MayAlias)
1218 // Reset if speculation failed.
1219 if (Alias != NoAlias)
1220 AliasCache[Locs] = OrigAliasResult;
1225 SmallPtrSet<Value *, 4> UniqueSrc;
1226 SmallVector<Value *, 4> V1Srcs;
1227 bool isRecursive = false;
1228 for (Value *PV1 : PN->incoming_values()) {
1229 if (isa<PHINode>(PV1))
1230 // If any of the source itself is a PHI, return MayAlias conservatively
1231 // to avoid compile time explosion. The worst possible case is if both
1232 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1233 // and 'n' are the number of PHI sources.
1236 if (EnableRecPhiAnalysis)
1237 if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) {
1238 // Check whether the incoming value is a GEP that advances the pointer
1239 // result of this PHI node (e.g. in a loop). If this is the case, we
1240 // would recurse and always get a MayAlias. Handle this case specially
1242 if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 &&
1243 isa<ConstantInt>(PV1GEP->idx_begin())) {
1249 if (UniqueSrc.insert(PV1).second)
1250 V1Srcs.push_back(PV1);
1253 // If this PHI node is recursive, set the size of the accessed memory to
1254 // unknown to represent all the possible values the GEP could advance the
1257 PNSize = MemoryLocation::UnknownSize;
1260 aliasCheck(V2, V2Size, V2AAInfo, V1Srcs[0], PNSize, PNAAInfo);
1262 // Early exit if the check of the first PHI source against V2 is MayAlias.
1263 // Other results are not possible.
1264 if (Alias == MayAlias)
1267 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1268 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1269 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1270 Value *V = V1Srcs[i];
1272 AliasResult ThisAlias =
1273 aliasCheck(V2, V2Size, V2AAInfo, V, PNSize, PNAAInfo);
1274 Alias = MergeAliasResults(ThisAlias, Alias);
1275 if (Alias == MayAlias)
1282 /// Provides a bunch of ad-hoc rules to disambiguate in common cases, such as
1283 /// array references.
1284 AliasResult BasicAAResult::aliasCheck(const Value *V1, uint64_t V1Size,
1285 AAMDNodes V1AAInfo, const Value *V2,
1286 uint64_t V2Size, AAMDNodes V2AAInfo) {
1287 // If either of the memory references is empty, it doesn't matter what the
1288 // pointer values are.
1289 if (V1Size == 0 || V2Size == 0)
1292 // Strip off any casts if they exist.
1293 V1 = V1->stripPointerCasts();
1294 V2 = V2->stripPointerCasts();
1296 // If V1 or V2 is undef, the result is NoAlias because we can always pick a
1297 // value for undef that aliases nothing in the program.
1298 if (isa<UndefValue>(V1) || isa<UndefValue>(V2))
1301 // Are we checking for alias of the same value?
1302 // Because we look 'through' phi nodes we could look at "Value" pointers from
1303 // different iterations. We must therefore make sure that this is not the
1304 // case. The function isValueEqualInPotentialCycles ensures that this cannot
1305 // happen by looking at the visited phi nodes and making sure they cannot
1307 if (isValueEqualInPotentialCycles(V1, V2))
1310 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1311 return NoAlias; // Scalars cannot alias each other
1313 // Figure out what objects these things are pointing to if we can.
1314 const Value *O1 = GetUnderlyingObject(V1, DL, MaxLookupSearchDepth);
1315 const Value *O2 = GetUnderlyingObject(V2, DL, MaxLookupSearchDepth);
1317 // Null values in the default address space don't point to any object, so they
1318 // don't alias any other pointer.
1319 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1320 if (CPN->getType()->getAddressSpace() == 0)
1322 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1323 if (CPN->getType()->getAddressSpace() == 0)
1327 // If V1/V2 point to two different objects we know that we have no alias.
1328 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1331 // Constant pointers can't alias with non-const isIdentifiedObject objects.
1332 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1333 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1336 // Function arguments can't alias with things that are known to be
1337 // unambigously identified at the function level.
1338 if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1339 (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1342 // Most objects can't alias null.
1343 if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
1344 (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
1347 // If one pointer is the result of a call/invoke or load and the other is a
1348 // non-escaping local object within the same function, then we know the
1349 // object couldn't escape to a point where the call could return it.
1351 // Note that if the pointers are in different functions, there are a
1352 // variety of complications. A call with a nocapture argument may still
1353 // temporary store the nocapture argument's value in a temporary memory
1354 // location if that memory location doesn't escape. Or it may pass a
1355 // nocapture value to other functions as long as they don't capture it.
1356 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1358 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1362 // If the size of one access is larger than the entire object on the other
1363 // side, then we know such behavior is undefined and can assume no alias.
1364 if ((V1Size != MemoryLocation::UnknownSize &&
1365 isObjectSmallerThan(O2, V1Size, DL, TLI)) ||
1366 (V2Size != MemoryLocation::UnknownSize &&
1367 isObjectSmallerThan(O1, V2Size, DL, TLI)))
1370 // Check the cache before climbing up use-def chains. This also terminates
1371 // otherwise infinitely recursive queries.
1372 LocPair Locs(MemoryLocation(V1, V1Size, V1AAInfo),
1373 MemoryLocation(V2, V2Size, V2AAInfo));
1375 std::swap(Locs.first, Locs.second);
1376 std::pair<AliasCacheTy::iterator, bool> Pair =
1377 AliasCache.insert(std::make_pair(Locs, MayAlias));
1379 return Pair.first->second;
1381 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1382 // GEP can't simplify, we don't even look at the PHI cases.
1383 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1385 std::swap(V1Size, V2Size);
1387 std::swap(V1AAInfo, V2AAInfo);
1389 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1390 AliasResult Result =
1391 aliasGEP(GV1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O1, O2);
1392 if (Result != MayAlias)
1393 return AliasCache[Locs] = Result;
1396 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1398 std::swap(V1Size, V2Size);
1399 std::swap(V1AAInfo, V2AAInfo);
1401 if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1402 AliasResult Result = aliasPHI(PN, V1Size, V1AAInfo, V2, V2Size, V2AAInfo);
1403 if (Result != MayAlias)
1404 return AliasCache[Locs] = Result;
1407 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1409 std::swap(V1Size, V2Size);
1410 std::swap(V1AAInfo, V2AAInfo);
1412 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1413 AliasResult Result =
1414 aliasSelect(S1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo);
1415 if (Result != MayAlias)
1416 return AliasCache[Locs] = Result;
1419 // If both pointers are pointing into the same object and one of them
1420 // accesses is accessing the entire object, then the accesses must
1421 // overlap in some way.
1423 if ((V1Size != MemoryLocation::UnknownSize &&
1424 isObjectSize(O1, V1Size, DL, TLI)) ||
1425 (V2Size != MemoryLocation::UnknownSize &&
1426 isObjectSize(O2, V2Size, DL, TLI)))
1427 return AliasCache[Locs] = PartialAlias;
1429 // Recurse back into the best AA results we have, potentially with refined
1430 // memory locations. We have already ensured that BasicAA has a MayAlias
1431 // cache result for these, so any recursion back into BasicAA won't loop.
1432 AliasResult Result = getBestAAResults().alias(Locs.first, Locs.second);
1433 return AliasCache[Locs] = Result;
1436 /// Check whether two Values can be considered equivalent.
1438 /// In addition to pointer equivalence of \p V1 and \p V2 this checks whether
1439 /// they can not be part of a cycle in the value graph by looking at all
1440 /// visited phi nodes an making sure that the phis cannot reach the value. We
1441 /// have to do this because we are looking through phi nodes (That is we say
1442 /// noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB).
1443 bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V,
1448 const Instruction *Inst = dyn_cast<Instruction>(V);
1452 if (VisitedPhiBBs.empty())
1455 if (VisitedPhiBBs.size() > MaxNumPhiBBsValueReachabilityCheck)
1458 // Make sure that the visited phis cannot reach the Value. This ensures that
1459 // the Values cannot come from different iterations of a potential cycle the
1460 // phi nodes could be involved in.
1461 for (auto *P : VisitedPhiBBs)
1462 if (isPotentiallyReachable(&P->front(), Inst, DT, LI))
1468 /// Computes the symbolic difference between two de-composed GEPs.
1470 /// Dest and Src are the variable indices from two decomposed GetElementPtr
1471 /// instructions GEP1 and GEP2 which have common base pointers.
1472 void BasicAAResult::GetIndexDifference(
1473 SmallVectorImpl<VariableGEPIndex> &Dest,
1474 const SmallVectorImpl<VariableGEPIndex> &Src) {
1478 for (unsigned i = 0, e = Src.size(); i != e; ++i) {
1479 const Value *V = Src[i].V;
1480 unsigned ZExtBits = Src[i].ZExtBits, SExtBits = Src[i].SExtBits;
1481 int64_t Scale = Src[i].Scale;
1483 // Find V in Dest. This is N^2, but pointer indices almost never have more
1484 // than a few variable indexes.
1485 for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
1486 if (!isValueEqualInPotentialCycles(Dest[j].V, V) ||
1487 Dest[j].ZExtBits != ZExtBits || Dest[j].SExtBits != SExtBits)
1490 // If we found it, subtract off Scale V's from the entry in Dest. If it
1491 // goes to zero, remove the entry.
1492 if (Dest[j].Scale != Scale)
1493 Dest[j].Scale -= Scale;
1495 Dest.erase(Dest.begin() + j);
1500 // If we didn't consume this entry, add it to the end of the Dest list.
1502 VariableGEPIndex Entry = {V, ZExtBits, SExtBits, -Scale};
1503 Dest.push_back(Entry);
1508 bool BasicAAResult::constantOffsetHeuristic(
1509 const SmallVectorImpl<VariableGEPIndex> &VarIndices, uint64_t V1Size,
1510 uint64_t V2Size, int64_t BaseOffset, AssumptionCache *AC,
1511 DominatorTree *DT) {
1512 if (VarIndices.size() != 2 || V1Size == MemoryLocation::UnknownSize ||
1513 V2Size == MemoryLocation::UnknownSize)
1516 const VariableGEPIndex &Var0 = VarIndices[0], &Var1 = VarIndices[1];
1518 if (Var0.ZExtBits != Var1.ZExtBits || Var0.SExtBits != Var1.SExtBits ||
1519 Var0.Scale != -Var1.Scale)
1522 unsigned Width = Var1.V->getType()->getIntegerBitWidth();
1524 // We'll strip off the Extensions of Var0 and Var1 and do another round
1525 // of GetLinearExpression decomposition. In the example above, if Var0
1526 // is zext(%x + 1) we should get V1 == %x and V1Offset == 1.
1528 APInt V0Scale(Width, 0), V0Offset(Width, 0), V1Scale(Width, 0),
1530 bool NSW = true, NUW = true;
1531 unsigned V0ZExtBits = 0, V0SExtBits = 0, V1ZExtBits = 0, V1SExtBits = 0;
1532 const Value *V0 = GetLinearExpression(Var0.V, V0Scale, V0Offset, V0ZExtBits,
1533 V0SExtBits, DL, 0, AC, DT, NSW, NUW);
1534 NSW = true, NUW = true;
1535 const Value *V1 = GetLinearExpression(Var1.V, V1Scale, V1Offset, V1ZExtBits,
1536 V1SExtBits, DL, 0, AC, DT, NSW, NUW);
1538 if (V0Scale != V1Scale || V0ZExtBits != V1ZExtBits ||
1539 V0SExtBits != V1SExtBits || !isValueEqualInPotentialCycles(V0, V1))
1542 // We have a hit - Var0 and Var1 only differ by a constant offset!
1544 // If we've been sext'ed then zext'd the maximum difference between Var0 and
1545 // Var1 is possible to calculate, but we're just interested in the absolute
1546 // minimum difference between the two. The minimum distance may occur due to
1547 // wrapping; consider "add i3 %i, 5": if %i == 7 then 7 + 5 mod 8 == 4, and so
1548 // the minimum distance between %i and %i + 5 is 3.
1549 APInt MinDiff = V0Offset - V1Offset, Wrapped = -MinDiff;
1550 MinDiff = APIntOps::umin(MinDiff, Wrapped);
1551 uint64_t MinDiffBytes = MinDiff.getZExtValue() * std::abs(Var0.Scale);
1553 // We can't definitely say whether GEP1 is before or after V2 due to wrapping
1554 // arithmetic (i.e. for some values of GEP1 and V2 GEP1 < V2, and for other
1555 // values GEP1 > V2). We'll therefore only declare NoAlias if both V1Size and
1556 // V2Size can fit in the MinDiffBytes gap.
1557 return V1Size + std::abs(BaseOffset) <= MinDiffBytes &&
1558 V2Size + std::abs(BaseOffset) <= MinDiffBytes;
1561 //===----------------------------------------------------------------------===//
1562 // BasicAliasAnalysis Pass
1563 //===----------------------------------------------------------------------===//
1565 char BasicAA::PassID;
1567 BasicAAResult BasicAA::run(Function &F, AnalysisManager<Function> *AM) {
1568 return BasicAAResult(F.getParent()->getDataLayout(),
1569 AM->getResult<TargetLibraryAnalysis>(F),
1570 AM->getResult<AssumptionAnalysis>(F),
1571 AM->getCachedResult<DominatorTreeAnalysis>(F),
1572 AM->getCachedResult<LoopAnalysis>(F));
1575 BasicAAWrapperPass::BasicAAWrapperPass() : FunctionPass(ID) {
1576 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
1579 char BasicAAWrapperPass::ID = 0;
1580 void BasicAAWrapperPass::anchor() {}
1582 INITIALIZE_PASS_BEGIN(BasicAAWrapperPass, "basicaa",
1583 "Basic Alias Analysis (stateless AA impl)", true, true)
1584 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1585 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1586 INITIALIZE_PASS_END(BasicAAWrapperPass, "basicaa",
1587 "Basic Alias Analysis (stateless AA impl)", true, true)
1589 FunctionPass *llvm::createBasicAAWrapperPass() {
1590 return new BasicAAWrapperPass();
1593 bool BasicAAWrapperPass::runOnFunction(Function &F) {
1594 auto &ACT = getAnalysis<AssumptionCacheTracker>();
1595 auto &TLIWP = getAnalysis<TargetLibraryInfoWrapperPass>();
1596 auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
1597 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
1599 Result.reset(new BasicAAResult(F.getParent()->getDataLayout(), TLIWP.getTLI(),
1600 ACT.getAssumptionCache(F),
1601 DTWP ? &DTWP->getDomTree() : nullptr,
1602 LIWP ? &LIWP->getLoopInfo() : nullptr));
1607 void BasicAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1608 AU.setPreservesAll();
1609 AU.addRequired<AssumptionCacheTracker>();
1610 AU.addRequired<TargetLibraryInfoWrapperPass>();
1613 BasicAAResult llvm::createLegacyPMBasicAAResult(Pass &P, Function &F) {
1614 return BasicAAResult(
1615 F.getParent()->getDataLayout(),
1616 P.getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
1617 P.getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));