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/IR/Constants.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/GlobalAlias.h"
31 #include "llvm/IR/GlobalVariable.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/Operator.h"
36 #include "llvm/Pass.h"
37 #include "llvm/Support/ErrorHandling.h"
41 /// Enable analysis of recursive PHI nodes.
42 static cl::opt<bool> EnableRecPhiAnalysis("basicaa-recphi", cl::Hidden,
45 /// SearchLimitReached / SearchTimes shows how often the limit of
46 /// to decompose GEPs is reached. It will affect the precision
47 /// of basic alias analysis.
48 #define DEBUG_TYPE "basicaa"
49 STATISTIC(SearchLimitReached, "Number of times the limit to "
50 "decompose GEPs is reached");
51 STATISTIC(SearchTimes, "Number of times a GEP is decomposed");
53 /// Cutoff after which to stop analysing a set of phi nodes potentially involved
54 /// in a cycle. Because we are analysing 'through' phi nodes we need to be
55 /// careful with value equivalence. We use reachability to make sure a value
56 /// cannot be involved in a cycle.
57 const unsigned MaxNumPhiBBsValueReachabilityCheck = 20;
59 // The max limit of the search depth in DecomposeGEPExpression() and
60 // GetUnderlyingObject(), both functions need to use the same search
61 // depth otherwise the algorithm in aliasGEP will assert.
62 static const unsigned MaxLookupSearchDepth = 6;
64 //===----------------------------------------------------------------------===//
66 //===----------------------------------------------------------------------===//
68 /// Returns true if the pointer is to a function-local object that never
69 /// escapes from the function.
70 static bool isNonEscapingLocalObject(const Value *V) {
71 // If this is a local allocation, check to see if it escapes.
72 if (isa<AllocaInst>(V) || isNoAliasCall(V))
73 // Set StoreCaptures to True so that we can assume in our callers that the
74 // pointer is not the result of a load instruction. Currently
75 // PointerMayBeCaptured doesn't have any special analysis for the
76 // StoreCaptures=false case; if it did, our callers could be refined to be
78 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
80 // If this is an argument that corresponds to a byval or noalias argument,
81 // then it has not escaped before entering the function. Check if it escapes
82 // inside the function.
83 if (const Argument *A = dyn_cast<Argument>(V))
84 if (A->hasByValAttr() || A->hasNoAliasAttr())
85 // Note even if the argument is marked nocapture we still need to check
86 // for copies made inside the function. The nocapture attribute only
87 // specifies that there are no copies made that outlive the function.
88 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
93 /// Returns true if the pointer is one which would have been considered an
94 /// escape by isNonEscapingLocalObject.
95 static bool isEscapeSource(const Value *V) {
96 if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
99 // The load case works because isNonEscapingLocalObject considers all
100 // stores to be escapes (it passes true for the StoreCaptures argument
101 // to PointerMayBeCaptured).
102 if (isa<LoadInst>(V))
108 /// Returns the size of the object specified by V, or UnknownSize if unknown.
109 static uint64_t getObjectSize(const Value *V, const DataLayout &DL,
110 const TargetLibraryInfo &TLI,
111 bool RoundToAlign = false) {
113 if (getObjectSize(V, Size, DL, &TLI, RoundToAlign))
115 return MemoryLocation::UnknownSize;
118 /// Returns true if we can prove that the object specified by V is smaller than
120 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
121 const DataLayout &DL,
122 const TargetLibraryInfo &TLI) {
123 // Note that the meanings of the "object" are slightly different in the
124 // following contexts:
125 // c1: llvm::getObjectSize()
126 // c2: llvm.objectsize() intrinsic
127 // c3: isObjectSmallerThan()
128 // c1 and c2 share the same meaning; however, the meaning of "object" in c3
129 // refers to the "entire object".
131 // Consider this example:
132 // char *p = (char*)malloc(100)
135 // In the context of c1 and c2, the "object" pointed by q refers to the
136 // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
138 // However, in the context of c3, the "object" refers to the chunk of memory
139 // being allocated. So, the "object" has 100 bytes, and q points to the middle
140 // the "object". In case q is passed to isObjectSmallerThan() as the 1st
141 // parameter, before the llvm::getObjectSize() is called to get the size of
142 // entire object, we should:
143 // - either rewind the pointer q to the base-address of the object in
144 // question (in this case rewind to p), or
145 // - just give up. It is up to caller to make sure the pointer is pointing
146 // to the base address the object.
148 // We go for 2nd option for simplicity.
149 if (!isIdentifiedObject(V))
152 // This function needs to use the aligned object size because we allow
153 // reads a bit past the end given sufficient alignment.
154 uint64_t ObjectSize = getObjectSize(V, DL, TLI, /*RoundToAlign*/ true);
156 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size;
159 /// Returns true if we can prove that the object specified by V has size Size.
160 static bool isObjectSize(const Value *V, uint64_t Size, const DataLayout &DL,
161 const TargetLibraryInfo &TLI) {
162 uint64_t ObjectSize = getObjectSize(V, DL, TLI);
163 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size;
166 //===----------------------------------------------------------------------===//
167 // GetElementPtr Instruction Decomposition and Analysis
168 //===----------------------------------------------------------------------===//
170 /// Analyzes the specified value as a linear expression: "A*V + B", where A and
171 /// B are constant integers.
173 /// Returns the scale and offset values as APInts and return V as a Value*, and
174 /// return whether we looked through any sign or zero extends. The incoming
175 /// Value is known to have IntegerType and it may already be sign or zero
178 /// Note that this looks through extends, so the high bits may not be
179 /// represented in the result.
180 /*static*/ const Value *BasicAliasAnalysis::GetLinearExpression(
181 const Value *V, APInt &Scale, APInt &Offset, unsigned &ZExtBits,
182 unsigned &SExtBits, const DataLayout &DL, unsigned Depth,
183 AssumptionCache *AC, DominatorTree *DT, bool &NSW, bool &NUW) {
184 assert(V->getType()->isIntegerTy() && "Not an integer value");
186 // Limit our recursion depth.
193 if (const ConstantInt *Const = dyn_cast<ConstantInt>(V)) {
194 // if it's a constant, just convert it to an offset and remove the variable.
195 // If we've been called recursively the Offset bit width will be greater
196 // than the constant's (the Offset's always as wide as the outermost call),
197 // so we'll zext here and process any extension in the isa<SExtInst> &
198 // isa<ZExtInst> cases below.
199 Offset += Const->getValue().zextOrSelf(Offset.getBitWidth());
200 assert(Scale == 0 && "Constant values don't have a scale");
204 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
205 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
207 // If we've been called recursively then Offset and Scale will be wider
208 // that the BOp operands. We'll always zext it here as we'll process sign
209 // extensions below (see the isa<SExtInst> / isa<ZExtInst> cases).
210 APInt RHS = RHSC->getValue().zextOrSelf(Offset.getBitWidth());
212 switch (BOp->getOpcode()) {
214 // We don't understand this instruction, so we can't decompose it any
219 case Instruction::Or:
220 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
222 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC,
229 case Instruction::Add:
230 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
231 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
234 case Instruction::Sub:
235 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
236 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
239 case Instruction::Mul:
240 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
241 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
245 case Instruction::Shl:
246 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
247 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
248 Offset <<= RHS.getLimitedValue();
249 Scale <<= RHS.getLimitedValue();
250 // the semantics of nsw and nuw for left shifts don't match those of
251 // multiplications, so we won't propagate them.
256 if (isa<OverflowingBinaryOperator>(BOp)) {
257 NUW &= BOp->hasNoUnsignedWrap();
258 NSW &= BOp->hasNoSignedWrap();
264 // Since GEP indices are sign extended anyway, we don't care about the high
265 // bits of a sign or zero extended value - just scales and offsets. The
266 // extensions have to be consistent though.
267 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) {
268 Value *CastOp = cast<CastInst>(V)->getOperand(0);
269 unsigned NewWidth = V->getType()->getPrimitiveSizeInBits();
270 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
271 unsigned OldZExtBits = ZExtBits, OldSExtBits = SExtBits;
272 const Value *Result =
273 GetLinearExpression(CastOp, Scale, Offset, ZExtBits, SExtBits, DL,
274 Depth + 1, AC, DT, NSW, NUW);
276 // zext(zext(%x)) == zext(%x), and similiarly for sext; we'll handle this
277 // by just incrementing the number of bits we've extended by.
278 unsigned ExtendedBy = NewWidth - SmallWidth;
280 if (isa<SExtInst>(V) && ZExtBits == 0) {
281 // sext(sext(%x, a), b) == sext(%x, a + b)
284 // We haven't sign-wrapped, so it's valid to decompose sext(%x + c)
285 // into sext(%x) + sext(c). We'll sext the Offset ourselves:
286 unsigned OldWidth = Offset.getBitWidth();
287 Offset = Offset.trunc(SmallWidth).sext(NewWidth).zextOrSelf(OldWidth);
289 // We may have signed-wrapped, so don't decompose sext(%x + c) into
290 // sext(%x) + sext(c)
294 ZExtBits = OldZExtBits;
295 SExtBits = OldSExtBits;
297 SExtBits += ExtendedBy;
299 // sext(zext(%x, a), b) = zext(zext(%x, a), b) = zext(%x, a + b)
302 // We may have unsigned-wrapped, so don't decompose zext(%x + c) into
303 // zext(%x) + zext(c)
307 ZExtBits = OldZExtBits;
308 SExtBits = OldSExtBits;
310 ZExtBits += ExtendedBy;
321 /// If V is a symbolic pointer expression, decompose it into a base pointer
322 /// with a constant offset and a number of scaled symbolic offsets.
324 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale
325 /// in the VarIndices vector) are Value*'s that are known to be scaled by the
326 /// specified amount, but which may have other unrepresented high bits. As
327 /// such, the gep cannot necessarily be reconstructed from its decomposed form.
329 /// When DataLayout is around, this function is capable of analyzing everything
330 /// that GetUnderlyingObject can look through. To be able to do that
331 /// GetUnderlyingObject and DecomposeGEPExpression must use the same search
332 /// depth (MaxLookupSearchDepth). When DataLayout not is around, it just looks
333 /// through pointer casts.
334 /*static*/ const Value *BasicAliasAnalysis::DecomposeGEPExpression(
335 const Value *V, int64_t &BaseOffs,
336 SmallVectorImpl<VariableGEPIndex> &VarIndices, bool &MaxLookupReached,
337 const DataLayout &DL, AssumptionCache *AC, DominatorTree *DT) {
338 // Limit recursion depth to limit compile time in crazy cases.
339 unsigned MaxLookup = MaxLookupSearchDepth;
340 MaxLookupReached = false;
345 // See if this is a bitcast or GEP.
346 const Operator *Op = dyn_cast<Operator>(V);
348 // The only non-operator case we can handle are GlobalAliases.
349 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
350 if (!GA->mayBeOverridden()) {
351 V = GA->getAliasee();
358 if (Op->getOpcode() == Instruction::BitCast ||
359 Op->getOpcode() == Instruction::AddrSpaceCast) {
360 V = Op->getOperand(0);
364 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
366 // If it's not a GEP, hand it off to SimplifyInstruction to see if it
367 // can come up with something. This matches what GetUnderlyingObject does.
368 if (const Instruction *I = dyn_cast<Instruction>(V))
369 // TODO: Get a DominatorTree and AssumptionCache and use them here
370 // (these are both now available in this function, but this should be
371 // updated when GetUnderlyingObject is updated). TLI should be
373 if (const Value *Simplified =
374 SimplifyInstruction(const_cast<Instruction *>(I), DL)) {
382 // Don't attempt to analyze GEPs over unsized objects.
383 if (!GEPOp->getOperand(0)->getType()->getPointerElementType()->isSized())
386 unsigned AS = GEPOp->getPointerAddressSpace();
387 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
388 gep_type_iterator GTI = gep_type_begin(GEPOp);
389 for (User::const_op_iterator I = GEPOp->op_begin() + 1, E = GEPOp->op_end();
391 const Value *Index = *I;
392 // Compute the (potentially symbolic) offset in bytes for this index.
393 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
394 // For a struct, add the member offset.
395 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
399 BaseOffs += DL.getStructLayout(STy)->getElementOffset(FieldNo);
403 // For an array/pointer, add the element offset, explicitly scaled.
404 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
407 BaseOffs += DL.getTypeAllocSize(*GTI) * CIdx->getSExtValue();
411 uint64_t Scale = DL.getTypeAllocSize(*GTI);
412 unsigned ZExtBits = 0, SExtBits = 0;
414 // If the integer type is smaller than the pointer size, it is implicitly
415 // sign extended to pointer size.
416 unsigned Width = Index->getType()->getIntegerBitWidth();
417 unsigned PointerSize = DL.getPointerSizeInBits(AS);
418 if (PointerSize > Width)
419 SExtBits += PointerSize - Width;
421 // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
422 APInt IndexScale(Width, 0), IndexOffset(Width, 0);
423 bool NSW = true, NUW = true;
424 Index = GetLinearExpression(Index, IndexScale, IndexOffset, ZExtBits,
425 SExtBits, DL, 0, AC, DT, NSW, NUW);
427 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
428 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
429 BaseOffs += IndexOffset.getSExtValue() * Scale;
430 Scale *= IndexScale.getSExtValue();
432 // If we already had an occurrence of this index variable, merge this
433 // scale into it. For example, we want to handle:
434 // A[x][x] -> x*16 + x*4 -> x*20
435 // This also ensures that 'x' only appears in the index list once.
436 for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
437 if (VarIndices[i].V == Index && VarIndices[i].ZExtBits == ZExtBits &&
438 VarIndices[i].SExtBits == SExtBits) {
439 Scale += VarIndices[i].Scale;
440 VarIndices.erase(VarIndices.begin() + i);
445 // Make sure that we have a scale that makes sense for this target's
447 if (unsigned ShiftBits = 64 - PointerSize) {
449 Scale = (int64_t)Scale >> ShiftBits;
453 VariableGEPIndex Entry = {Index, ZExtBits, SExtBits,
454 static_cast<int64_t>(Scale)};
455 VarIndices.push_back(Entry);
459 // Analyze the base pointer next.
460 V = GEPOp->getOperand(0);
461 } while (--MaxLookup);
463 // If the chain of expressions is too deep, just return early.
464 MaxLookupReached = true;
465 SearchLimitReached++;
469 //===----------------------------------------------------------------------===//
470 // BasicAliasAnalysis Pass
471 //===----------------------------------------------------------------------===//
473 // Register the pass...
474 char BasicAliasAnalysis::ID = 0;
475 INITIALIZE_AG_PASS_BEGIN(BasicAliasAnalysis, AliasAnalysis, "basicaa",
476 "Basic Alias Analysis (stateless AA impl)", false,
478 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
479 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
480 INITIALIZE_AG_PASS_END(BasicAliasAnalysis, AliasAnalysis, "basicaa",
481 "Basic Alias Analysis (stateless AA impl)", false, true,
484 ImmutablePass *llvm::createBasicAliasAnalysisPass() {
485 return new BasicAliasAnalysis();
488 /// Returns whether the given pointer value points to memory that is local to
489 /// the function, with global constants being considered local to all
491 bool BasicAliasAnalysis::pointsToConstantMemory(const MemoryLocation &Loc,
493 assert(Visited.empty() && "Visited must be cleared after use!");
495 unsigned MaxLookup = 8;
496 SmallVector<const Value *, 16> Worklist;
497 Worklist.push_back(Loc.Ptr);
499 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), *DL);
500 if (!Visited.insert(V).second) {
502 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
505 // An alloca instruction defines local memory.
506 if (OrLocal && isa<AllocaInst>(V))
509 // A global constant counts as local memory for our purposes.
510 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
511 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
512 // global to be marked constant in some modules and non-constant in
513 // others. GV may even be a declaration, not a definition.
514 if (!GV->isConstant()) {
516 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
521 // If both select values point to local memory, then so does the select.
522 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
523 Worklist.push_back(SI->getTrueValue());
524 Worklist.push_back(SI->getFalseValue());
528 // If all values incoming to a phi node point to local memory, then so does
530 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
531 // Don't bother inspecting phi nodes with many operands.
532 if (PN->getNumIncomingValues() > MaxLookup) {
534 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
536 for (Value *IncValue : PN->incoming_values())
537 Worklist.push_back(IncValue);
541 // Otherwise be conservative.
543 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
545 } while (!Worklist.empty() && --MaxLookup);
548 return Worklist.empty();
551 // FIXME: This code is duplicated with MemoryLocation and should be hoisted to
552 // some common utility location.
553 static bool isMemsetPattern16(const Function *MS,
554 const TargetLibraryInfo &TLI) {
555 if (TLI.has(LibFunc::memset_pattern16) &&
556 MS->getName() == "memset_pattern16") {
557 FunctionType *MemsetType = MS->getFunctionType();
558 if (!MemsetType->isVarArg() && MemsetType->getNumParams() == 3 &&
559 isa<PointerType>(MemsetType->getParamType(0)) &&
560 isa<PointerType>(MemsetType->getParamType(1)) &&
561 isa<IntegerType>(MemsetType->getParamType(2)))
568 /// Returns the behavior when calling the given call site.
569 FunctionModRefBehavior
570 BasicAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
571 if (CS.doesNotAccessMemory())
572 // Can't do better than this.
573 return FMRB_DoesNotAccessMemory;
575 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
577 // If the callsite knows it only reads memory, don't return worse
579 if (CS.onlyReadsMemory())
580 Min = FMRB_OnlyReadsMemory;
582 if (CS.onlyAccessesArgMemory())
583 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
585 // The AliasAnalysis base class has some smarts, lets use them.
586 return FunctionModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
589 /// Returns the behavior when calling the given function. For use when the call
590 /// site is not known.
591 FunctionModRefBehavior
592 BasicAliasAnalysis::getModRefBehavior(const Function *F) {
593 // If the function declares it doesn't access memory, we can't do better.
594 if (F->doesNotAccessMemory())
595 return FMRB_DoesNotAccessMemory;
597 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
599 // If the function declares it only reads memory, go with that.
600 if (F->onlyReadsMemory())
601 Min = FMRB_OnlyReadsMemory;
603 if (F->onlyAccessesArgMemory())
604 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
606 const TargetLibraryInfo &TLI =
607 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
608 if (isMemsetPattern16(F, TLI))
609 Min = FMRB_OnlyAccessesArgumentPointees;
611 // Otherwise be conservative.
612 return FunctionModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min);
615 ModRefInfo BasicAliasAnalysis::getArgModRefInfo(ImmutableCallSite CS,
617 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction()))
618 switch (II->getIntrinsicID()) {
621 case Intrinsic::memset:
622 case Intrinsic::memcpy:
623 case Intrinsic::memmove:
624 assert((ArgIdx == 0 || ArgIdx == 1) &&
625 "Invalid argument index for memory intrinsic");
626 return ArgIdx ? MRI_Ref : MRI_Mod;
629 // We can bound the aliasing properties of memset_pattern16 just as we can
630 // for memcpy/memset. This is particularly important because the
631 // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
632 // whenever possible.
633 if (CS.getCalledFunction() &&
634 isMemsetPattern16(CS.getCalledFunction(), *TLI)) {
635 assert((ArgIdx == 0 || ArgIdx == 1) &&
636 "Invalid argument index for memset_pattern16");
637 return ArgIdx ? MRI_Ref : MRI_Mod;
639 // FIXME: Handle memset_pattern4 and memset_pattern8 also.
641 return AliasAnalysis::getArgModRefInfo(CS, ArgIdx);
644 static bool isAssumeIntrinsic(ImmutableCallSite CS) {
645 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
646 if (II && II->getIntrinsicID() == Intrinsic::assume)
652 bool BasicAliasAnalysis::doInitialization(Module &M) {
653 InitializeAliasAnalysis(this, &M.getDataLayout());
657 /// Checks to see if the specified callsite can clobber the specified memory
660 /// Since we only look at local properties of this function, we really can't
661 /// say much about this query. We do, however, use simple "address taken"
662 /// analysis on local objects.
663 ModRefInfo BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS,
664 const MemoryLocation &Loc) {
665 assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
666 "AliasAnalysis query involving multiple functions!");
668 const Value *Object = GetUnderlyingObject(Loc.Ptr, *DL);
670 // If this is a tail call and Loc.Ptr points to a stack location, we know that
671 // the tail call cannot access or modify the local stack.
672 // We cannot exclude byval arguments here; these belong to the caller of
673 // the current function not to the current function, and a tail callee
674 // may reference them.
675 if (isa<AllocaInst>(Object))
676 if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
677 if (CI->isTailCall())
680 // If the pointer is to a locally allocated object that does not escape,
681 // then the call can not mod/ref the pointer unless the call takes the pointer
682 // as an argument, and itself doesn't capture it.
683 if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
684 isNonEscapingLocalObject(Object)) {
685 bool PassedAsArg = false;
687 for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
688 CI != CE; ++CI, ++ArgNo) {
689 // Only look at the no-capture or byval pointer arguments. If this
690 // pointer were passed to arguments that were neither of these, then it
691 // couldn't be no-capture.
692 if (!(*CI)->getType()->isPointerTy() ||
693 (!CS.doesNotCapture(ArgNo) && !CS.isByValArgument(ArgNo)))
696 // If this is a no-capture pointer argument, see if we can tell that it
697 // is impossible to alias the pointer we're checking. If not, we have to
698 // assume that the call could touch the pointer, even though it doesn't
700 if (!isNoAlias(MemoryLocation(*CI), MemoryLocation(Object))) {
710 // While the assume intrinsic is marked as arbitrarily writing so that
711 // proper control dependencies will be maintained, it never aliases any
712 // particular memory location.
713 if (isAssumeIntrinsic(CS))
716 // The AliasAnalysis base class has some smarts, lets use them.
717 return AliasAnalysis::getModRefInfo(CS, Loc);
720 ModRefInfo BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS1,
721 ImmutableCallSite CS2) {
722 // While the assume intrinsic is marked as arbitrarily writing so that
723 // proper control dependencies will be maintained, it never aliases any
724 // particular memory location.
725 if (isAssumeIntrinsic(CS1) || isAssumeIntrinsic(CS2))
728 // The AliasAnalysis base class has some smarts, lets use them.
729 return AliasAnalysis::getModRefInfo(CS1, CS2);
732 /// Provide ad-hoc rules to disambiguate accesses through two GEP operators,
733 /// both having the exact same pointer operand.
734 static AliasResult aliasSameBasePointerGEPs(const GEPOperator *GEP1,
736 const GEPOperator *GEP2,
738 const DataLayout &DL) {
740 assert(GEP1->getPointerOperand() == GEP2->getPointerOperand() &&
741 "Expected GEPs with the same pointer operand");
743 // Try to determine whether GEP1 and GEP2 index through arrays, into structs,
744 // such that the struct field accesses provably cannot alias.
745 // We also need at least two indices (the pointer, and the struct field).
746 if (GEP1->getNumIndices() != GEP2->getNumIndices() ||
747 GEP1->getNumIndices() < 2)
750 // If we don't know the size of the accesses through both GEPs, we can't
751 // determine whether the struct fields accessed can't alias.
752 if (V1Size == MemoryLocation::UnknownSize ||
753 V2Size == MemoryLocation::UnknownSize)
757 dyn_cast<ConstantInt>(GEP1->getOperand(GEP1->getNumOperands() - 1));
759 dyn_cast<ConstantInt>(GEP2->getOperand(GEP2->getNumOperands() - 1));
761 // If the last (struct) indices aren't constants, we can't say anything.
762 // If they're identical, the other indices might be also be dynamically
763 // equal, so the GEPs can alias.
764 if (!C1 || !C2 || C1 == C2)
767 // Find the last-indexed type of the GEP, i.e., the type you'd get if
768 // you stripped the last index.
769 // On the way, look at each indexed type. If there's something other
770 // than an array, different indices can lead to different final types.
771 SmallVector<Value *, 8> IntermediateIndices;
773 // Insert the first index; we don't need to check the type indexed
774 // through it as it only drops the pointer indirection.
775 assert(GEP1->getNumIndices() > 1 && "Not enough GEP indices to examine");
776 IntermediateIndices.push_back(GEP1->getOperand(1));
778 // Insert all the remaining indices but the last one.
779 // Also, check that they all index through arrays.
780 for (unsigned i = 1, e = GEP1->getNumIndices() - 1; i != e; ++i) {
781 if (!isa<ArrayType>(GetElementPtrInst::getIndexedType(
782 GEP1->getSourceElementType(), IntermediateIndices)))
784 IntermediateIndices.push_back(GEP1->getOperand(i + 1));
787 StructType *LastIndexedStruct =
788 dyn_cast<StructType>(GetElementPtrInst::getIndexedType(
789 GEP1->getSourceElementType(), IntermediateIndices));
791 if (!LastIndexedStruct)
795 // - both GEPs begin indexing from the exact same pointer;
796 // - the last indices in both GEPs are constants, indexing into a struct;
797 // - said indices are different, hence, the pointed-to fields are different;
798 // - both GEPs only index through arrays prior to that.
800 // This lets us determine that the struct that GEP1 indexes into and the
801 // struct that GEP2 indexes into must either precisely overlap or be
802 // completely disjoint. Because they cannot partially overlap, indexing into
803 // different non-overlapping fields of the struct will never alias.
805 // Therefore, the only remaining thing needed to show that both GEPs can't
806 // alias is that the fields are not overlapping.
807 const StructLayout *SL = DL.getStructLayout(LastIndexedStruct);
808 const uint64_t StructSize = SL->getSizeInBytes();
809 const uint64_t V1Off = SL->getElementOffset(C1->getZExtValue());
810 const uint64_t V2Off = SL->getElementOffset(C2->getZExtValue());
812 auto EltsDontOverlap = [StructSize](uint64_t V1Off, uint64_t V1Size,
813 uint64_t V2Off, uint64_t V2Size) {
814 return V1Off < V2Off && V1Off + V1Size <= V2Off &&
815 ((V2Off + V2Size <= StructSize) ||
816 (V2Off + V2Size - StructSize <= V1Off));
819 if (EltsDontOverlap(V1Off, V1Size, V2Off, V2Size) ||
820 EltsDontOverlap(V2Off, V2Size, V1Off, V1Size))
826 /// Provides a bunch of ad-hoc rules to disambiguate a GEP instruction against
829 /// We know that V1 is a GEP, but we don't know anything about V2.
830 /// UnderlyingV1 is GetUnderlyingObject(GEP1, DL), UnderlyingV2 is the same for
832 AliasResult BasicAliasAnalysis::aliasGEP(
833 const GEPOperator *GEP1, uint64_t V1Size, const AAMDNodes &V1AAInfo,
834 const Value *V2, uint64_t V2Size, const AAMDNodes &V2AAInfo,
835 const Value *UnderlyingV1, const Value *UnderlyingV2) {
836 int64_t GEP1BaseOffset;
837 bool GEP1MaxLookupReached;
838 SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
840 // We have to get two AssumptionCaches here because GEP1 and V2 may be from
841 // different functions.
842 // FIXME: This really doesn't make any sense. We get a dominator tree below
843 // that can only refer to a single function. But this function (aliasGEP) is
844 // a method on an immutable pass that can be called when there *isn't*
845 // a single function. The old pass management layer makes this "work", but
846 // this isn't really a clean solution.
847 AssumptionCacheTracker &ACT = getAnalysis<AssumptionCacheTracker>();
848 AssumptionCache *AC1 = nullptr, *AC2 = nullptr;
849 if (auto *GEP1I = dyn_cast<Instruction>(GEP1))
850 AC1 = &ACT.getAssumptionCache(
851 const_cast<Function &>(*GEP1I->getParent()->getParent()));
852 if (auto *I2 = dyn_cast<Instruction>(V2))
853 AC2 = &ACT.getAssumptionCache(
854 const_cast<Function &>(*I2->getParent()->getParent()));
856 DominatorTreeWrapperPass *DTWP =
857 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
858 DominatorTree *DT = DTWP ? &DTWP->getDomTree() : nullptr;
860 // If we have two gep instructions with must-alias or not-alias'ing base
861 // pointers, figure out if the indexes to the GEP tell us anything about the
863 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
864 // Do the base pointers alias?
865 AliasResult BaseAlias =
866 aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize, AAMDNodes(),
867 UnderlyingV2, MemoryLocation::UnknownSize, AAMDNodes());
869 // Check for geps of non-aliasing underlying pointers where the offsets are
871 if ((BaseAlias == MayAlias) && V1Size == V2Size) {
872 // Do the base pointers alias assuming type and size.
873 AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size, V1AAInfo,
874 UnderlyingV2, V2Size, V2AAInfo);
875 if (PreciseBaseAlias == NoAlias) {
876 // See if the computed offset from the common pointer tells us about the
877 // relation of the resulting pointer.
878 int64_t GEP2BaseOffset;
879 bool GEP2MaxLookupReached;
880 SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
881 const Value *GEP2BasePtr =
882 DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices,
883 GEP2MaxLookupReached, *DL, AC2, DT);
884 const Value *GEP1BasePtr =
885 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
886 GEP1MaxLookupReached, *DL, AC1, DT);
887 // DecomposeGEPExpression and GetUnderlyingObject should return the
888 // same result except when DecomposeGEPExpression has no DataLayout.
889 if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
891 "DecomposeGEPExpression and GetUnderlyingObject disagree!");
894 // If the max search depth is reached the result is undefined
895 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
899 if (GEP1BaseOffset == GEP2BaseOffset &&
900 GEP1VariableIndices == GEP2VariableIndices)
902 GEP1VariableIndices.clear();
906 // If we get a No or May, then return it immediately, no amount of analysis
907 // will improve this situation.
908 if (BaseAlias != MustAlias)
911 // Otherwise, we have a MustAlias. Since the base pointers alias each other
912 // exactly, see if the computed offset from the common pointer tells us
913 // about the relation of the resulting pointer.
914 const Value *GEP1BasePtr =
915 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
916 GEP1MaxLookupReached, *DL, AC1, DT);
918 int64_t GEP2BaseOffset;
919 bool GEP2MaxLookupReached;
920 SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
921 const Value *GEP2BasePtr =
922 DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices,
923 GEP2MaxLookupReached, *DL, AC2, DT);
925 // DecomposeGEPExpression and GetUnderlyingObject should return the
926 // same result except when DecomposeGEPExpression has no DataLayout.
927 if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
928 assert(!DL && "DecomposeGEPExpression and GetUnderlyingObject disagree!");
932 // If we know the two GEPs are based off of the exact same pointer (and not
933 // just the same underlying object), see if that tells us anything about
934 // the resulting pointers.
935 if (DL && GEP1->getPointerOperand() == GEP2->getPointerOperand()) {
936 AliasResult R = aliasSameBasePointerGEPs(GEP1, V1Size, GEP2, V2Size, *DL);
937 // If we couldn't find anything interesting, don't abandon just yet.
942 // If the max search depth is reached the result is undefined
943 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
946 // Subtract the GEP2 pointer from the GEP1 pointer to find out their
947 // symbolic difference.
948 GEP1BaseOffset -= GEP2BaseOffset;
949 GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
952 // Check to see if these two pointers are related by the getelementptr
953 // instruction. If one pointer is a GEP with a non-zero index of the other
954 // pointer, we know they cannot alias.
956 // If both accesses are unknown size, we can't do anything useful here.
957 if (V1Size == MemoryLocation::UnknownSize &&
958 V2Size == MemoryLocation::UnknownSize)
961 AliasResult R = aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize,
962 AAMDNodes(), V2, V2Size, V2AAInfo);
964 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
965 // If V2 is known not to alias GEP base pointer, then the two values
966 // cannot alias per GEP semantics: "A pointer value formed from a
967 // getelementptr instruction is associated with the addresses associated
968 // with the first operand of the getelementptr".
971 const Value *GEP1BasePtr =
972 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
973 GEP1MaxLookupReached, *DL, AC1, DT);
975 // DecomposeGEPExpression and GetUnderlyingObject should return the
976 // same result except when DecomposeGEPExpression has no DataLayout.
977 if (GEP1BasePtr != UnderlyingV1) {
978 assert(!DL && "DecomposeGEPExpression and GetUnderlyingObject disagree!");
981 // If the max search depth is reached the result is undefined
982 if (GEP1MaxLookupReached)
986 // In the two GEP Case, if there is no difference in the offsets of the
987 // computed pointers, the resultant pointers are a must alias. This
988 // hapens when we have two lexically identical GEP's (for example).
990 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
991 // must aliases the GEP, the end result is a must alias also.
992 if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
995 // If there is a constant difference between the pointers, but the difference
996 // is less than the size of the associated memory object, then we know
997 // that the objects are partially overlapping. If the difference is
998 // greater, we know they do not overlap.
999 if (GEP1BaseOffset != 0 && GEP1VariableIndices.empty()) {
1000 if (GEP1BaseOffset >= 0) {
1001 if (V2Size != MemoryLocation::UnknownSize) {
1002 if ((uint64_t)GEP1BaseOffset < V2Size)
1003 return PartialAlias;
1007 // We have the situation where:
1010 // ---------------->|
1011 // |-->V1Size |-------> V2Size
1013 // We need to know that V2Size is not unknown, otherwise we might have
1014 // stripped a gep with negative index ('gep <ptr>, -1, ...).
1015 if (V1Size != MemoryLocation::UnknownSize &&
1016 V2Size != MemoryLocation::UnknownSize) {
1017 if (-(uint64_t)GEP1BaseOffset < V1Size)
1018 return PartialAlias;
1024 if (!GEP1VariableIndices.empty()) {
1025 uint64_t Modulo = 0;
1026 bool AllPositive = true;
1027 for (unsigned i = 0, e = GEP1VariableIndices.size(); i != e; ++i) {
1029 // Try to distinguish something like &A[i][1] against &A[42][0].
1030 // Grab the least significant bit set in any of the scales. We
1031 // don't need std::abs here (even if the scale's negative) as we'll
1032 // be ^'ing Modulo with itself later.
1033 Modulo |= (uint64_t)GEP1VariableIndices[i].Scale;
1036 // If the Value could change between cycles, then any reasoning about
1037 // the Value this cycle may not hold in the next cycle. We'll just
1038 // give up if we can't determine conditions that hold for every cycle:
1039 const Value *V = GEP1VariableIndices[i].V;
1041 bool SignKnownZero, SignKnownOne;
1042 ComputeSignBit(const_cast<Value *>(V), SignKnownZero, SignKnownOne, *DL,
1043 0, AC1, nullptr, DT);
1045 // Zero-extension widens the variable, and so forces the sign
1047 bool IsZExt = GEP1VariableIndices[i].ZExtBits > 0 || isa<ZExtInst>(V);
1048 SignKnownZero |= IsZExt;
1049 SignKnownOne &= !IsZExt;
1051 // If the variable begins with a zero then we know it's
1052 // positive, regardless of whether the value is signed or
1054 int64_t Scale = GEP1VariableIndices[i].Scale;
1056 (SignKnownZero && Scale >= 0) || (SignKnownOne && Scale < 0);
1060 Modulo = Modulo ^ (Modulo & (Modulo - 1));
1062 // We can compute the difference between the two addresses
1063 // mod Modulo. Check whether that difference guarantees that the
1064 // two locations do not alias.
1065 uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1);
1066 if (V1Size != MemoryLocation::UnknownSize &&
1067 V2Size != MemoryLocation::UnknownSize && ModOffset >= V2Size &&
1068 V1Size <= Modulo - ModOffset)
1071 // If we know all the variables are positive, then GEP1 >= GEP1BasePtr.
1072 // If GEP1BasePtr > V2 (GEP1BaseOffset > 0) then we know the pointers
1073 // don't alias if V2Size can fit in the gap between V2 and GEP1BasePtr.
1074 if (AllPositive && GEP1BaseOffset > 0 && V2Size <= (uint64_t)GEP1BaseOffset)
1077 if (constantOffsetHeuristic(GEP1VariableIndices, V1Size, V2Size,
1078 GEP1BaseOffset, DL, AC1, DT))
1082 // Statically, we can see that the base objects are the same, but the
1083 // pointers have dynamic offsets which we can't resolve. And none of our
1084 // little tricks above worked.
1086 // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the
1087 // practical effect of this is protecting TBAA in the case of dynamic
1088 // indices into arrays of unions or malloc'd memory.
1089 return PartialAlias;
1092 static AliasResult MergeAliasResults(AliasResult A, AliasResult B) {
1093 // If the results agree, take it.
1096 // A mix of PartialAlias and MustAlias is PartialAlias.
1097 if ((A == PartialAlias && B == MustAlias) ||
1098 (B == PartialAlias && A == MustAlias))
1099 return PartialAlias;
1100 // Otherwise, we don't know anything.
1104 /// Provides a bunch of ad-hoc rules to disambiguate a Select instruction
1105 /// against another.
1106 AliasResult BasicAliasAnalysis::aliasSelect(const SelectInst *SI,
1108 const AAMDNodes &SIAAInfo,
1109 const Value *V2, uint64_t V2Size,
1110 const AAMDNodes &V2AAInfo) {
1111 // If the values are Selects with the same condition, we can do a more precise
1112 // check: just check for aliases between the values on corresponding arms.
1113 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1114 if (SI->getCondition() == SI2->getCondition()) {
1115 AliasResult Alias = aliasCheck(SI->getTrueValue(), SISize, SIAAInfo,
1116 SI2->getTrueValue(), V2Size, V2AAInfo);
1117 if (Alias == MayAlias)
1119 AliasResult ThisAlias =
1120 aliasCheck(SI->getFalseValue(), SISize, SIAAInfo,
1121 SI2->getFalseValue(), V2Size, V2AAInfo);
1122 return MergeAliasResults(ThisAlias, Alias);
1125 // If both arms of the Select node NoAlias or MustAlias V2, then returns
1126 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1128 aliasCheck(V2, V2Size, V2AAInfo, SI->getTrueValue(), SISize, SIAAInfo);
1129 if (Alias == MayAlias)
1132 AliasResult ThisAlias =
1133 aliasCheck(V2, V2Size, V2AAInfo, SI->getFalseValue(), SISize, SIAAInfo);
1134 return MergeAliasResults(ThisAlias, Alias);
1137 /// Provide a bunch of ad-hoc rules to disambiguate a PHI instruction against
1139 AliasResult BasicAliasAnalysis::aliasPHI(const PHINode *PN, uint64_t PNSize,
1140 const AAMDNodes &PNAAInfo,
1141 const Value *V2, uint64_t V2Size,
1142 const AAMDNodes &V2AAInfo) {
1143 // Track phi nodes we have visited. We use this information when we determine
1144 // value equivalence.
1145 VisitedPhiBBs.insert(PN->getParent());
1147 // If the values are PHIs in the same block, we can do a more precise
1148 // as well as efficient check: just check for aliases between the values
1149 // on corresponding edges.
1150 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1151 if (PN2->getParent() == PN->getParent()) {
1152 LocPair Locs(MemoryLocation(PN, PNSize, PNAAInfo),
1153 MemoryLocation(V2, V2Size, V2AAInfo));
1155 std::swap(Locs.first, Locs.second);
1156 // Analyse the PHIs' inputs under the assumption that the PHIs are
1158 // If the PHIs are May/MustAlias there must be (recursively) an input
1159 // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
1160 // there must be an operation on the PHIs within the PHIs' value cycle
1161 // that causes a MayAlias.
1162 // Pretend the phis do not alias.
1163 AliasResult Alias = NoAlias;
1164 assert(AliasCache.count(Locs) &&
1165 "There must exist an entry for the phi node");
1166 AliasResult OrigAliasResult = AliasCache[Locs];
1167 AliasCache[Locs] = NoAlias;
1169 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1170 AliasResult ThisAlias =
1171 aliasCheck(PN->getIncomingValue(i), PNSize, PNAAInfo,
1172 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1174 Alias = MergeAliasResults(ThisAlias, Alias);
1175 if (Alias == MayAlias)
1179 // Reset if speculation failed.
1180 if (Alias != NoAlias)
1181 AliasCache[Locs] = OrigAliasResult;
1186 SmallPtrSet<Value *, 4> UniqueSrc;
1187 SmallVector<Value *, 4> V1Srcs;
1188 bool isRecursive = false;
1189 for (Value *PV1 : PN->incoming_values()) {
1190 if (isa<PHINode>(PV1))
1191 // If any of the source itself is a PHI, return MayAlias conservatively
1192 // to avoid compile time explosion. The worst possible case is if both
1193 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1194 // and 'n' are the number of PHI sources.
1197 if (EnableRecPhiAnalysis)
1198 if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) {
1199 // Check whether the incoming value is a GEP that advances the pointer
1200 // result of this PHI node (e.g. in a loop). If this is the case, we
1201 // would recurse and always get a MayAlias. Handle this case specially
1203 if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 &&
1204 isa<ConstantInt>(PV1GEP->idx_begin())) {
1210 if (UniqueSrc.insert(PV1).second)
1211 V1Srcs.push_back(PV1);
1214 // If this PHI node is recursive, set the size of the accessed memory to
1215 // unknown to represent all the possible values the GEP could advance the
1218 PNSize = MemoryLocation::UnknownSize;
1221 aliasCheck(V2, V2Size, V2AAInfo, V1Srcs[0], PNSize, PNAAInfo);
1223 // Early exit if the check of the first PHI source against V2 is MayAlias.
1224 // Other results are not possible.
1225 if (Alias == MayAlias)
1228 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1229 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1230 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1231 Value *V = V1Srcs[i];
1233 AliasResult ThisAlias =
1234 aliasCheck(V2, V2Size, V2AAInfo, V, PNSize, PNAAInfo);
1235 Alias = MergeAliasResults(ThisAlias, Alias);
1236 if (Alias == MayAlias)
1243 /// Provideis a bunch of ad-hoc rules to disambiguate in common cases, such as
1244 /// array references.
1245 AliasResult BasicAliasAnalysis::aliasCheck(const Value *V1, uint64_t V1Size,
1246 AAMDNodes V1AAInfo, const Value *V2,
1248 AAMDNodes V2AAInfo) {
1249 // If either of the memory references is empty, it doesn't matter what the
1250 // pointer values are.
1251 if (V1Size == 0 || V2Size == 0)
1254 // Strip off any casts if they exist.
1255 V1 = V1->stripPointerCasts();
1256 V2 = V2->stripPointerCasts();
1258 // If V1 or V2 is undef, the result is NoAlias because we can always pick a
1259 // value for undef that aliases nothing in the program.
1260 if (isa<UndefValue>(V1) || isa<UndefValue>(V2))
1263 // Are we checking for alias of the same value?
1264 // Because we look 'through' phi nodes we could look at "Value" pointers from
1265 // different iterations. We must therefore make sure that this is not the
1266 // case. The function isValueEqualInPotentialCycles ensures that this cannot
1267 // happen by looking at the visited phi nodes and making sure they cannot
1269 if (isValueEqualInPotentialCycles(V1, V2))
1272 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1273 return NoAlias; // Scalars cannot alias each other
1275 // Figure out what objects these things are pointing to if we can.
1276 const Value *O1 = GetUnderlyingObject(V1, *DL, MaxLookupSearchDepth);
1277 const Value *O2 = GetUnderlyingObject(V2, *DL, MaxLookupSearchDepth);
1279 // Null values in the default address space don't point to any object, so they
1280 // don't alias any other pointer.
1281 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1282 if (CPN->getType()->getAddressSpace() == 0)
1284 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1285 if (CPN->getType()->getAddressSpace() == 0)
1289 // If V1/V2 point to two different objects we know that we have no alias.
1290 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1293 // Constant pointers can't alias with non-const isIdentifiedObject objects.
1294 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1295 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1298 // Function arguments can't alias with things that are known to be
1299 // unambigously identified at the function level.
1300 if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1301 (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1304 // Most objects can't alias null.
1305 if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
1306 (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
1309 // If one pointer is the result of a call/invoke or load and the other is a
1310 // non-escaping local object within the same function, then we know the
1311 // object couldn't escape to a point where the call could return it.
1313 // Note that if the pointers are in different functions, there are a
1314 // variety of complications. A call with a nocapture argument may still
1315 // temporary store the nocapture argument's value in a temporary memory
1316 // location if that memory location doesn't escape. Or it may pass a
1317 // nocapture value to other functions as long as they don't capture it.
1318 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1320 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1324 // If the size of one access is larger than the entire object on the other
1325 // side, then we know such behavior is undefined and can assume no alias.
1327 if ((V1Size != MemoryLocation::UnknownSize &&
1328 isObjectSmallerThan(O2, V1Size, *DL, *TLI)) ||
1329 (V2Size != MemoryLocation::UnknownSize &&
1330 isObjectSmallerThan(O1, V2Size, *DL, *TLI)))
1333 // Check the cache before climbing up use-def chains. This also terminates
1334 // otherwise infinitely recursive queries.
1335 LocPair Locs(MemoryLocation(V1, V1Size, V1AAInfo),
1336 MemoryLocation(V2, V2Size, V2AAInfo));
1338 std::swap(Locs.first, Locs.second);
1339 std::pair<AliasCacheTy::iterator, bool> Pair =
1340 AliasCache.insert(std::make_pair(Locs, MayAlias));
1342 return Pair.first->second;
1344 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1345 // GEP can't simplify, we don't even look at the PHI cases.
1346 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1348 std::swap(V1Size, V2Size);
1350 std::swap(V1AAInfo, V2AAInfo);
1352 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1353 AliasResult Result =
1354 aliasGEP(GV1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O1, O2);
1355 if (Result != MayAlias)
1356 return AliasCache[Locs] = Result;
1359 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1361 std::swap(V1Size, V2Size);
1362 std::swap(V1AAInfo, V2AAInfo);
1364 if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1365 AliasResult Result = aliasPHI(PN, V1Size, V1AAInfo, V2, V2Size, V2AAInfo);
1366 if (Result != MayAlias)
1367 return AliasCache[Locs] = Result;
1370 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1372 std::swap(V1Size, V2Size);
1373 std::swap(V1AAInfo, V2AAInfo);
1375 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1376 AliasResult Result =
1377 aliasSelect(S1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo);
1378 if (Result != MayAlias)
1379 return AliasCache[Locs] = Result;
1382 // If both pointers are pointing into the same object and one of them
1383 // accesses is accessing the entire object, then the accesses must
1384 // overlap in some way.
1386 if ((V1Size != MemoryLocation::UnknownSize &&
1387 isObjectSize(O1, V1Size, *DL, *TLI)) ||
1388 (V2Size != MemoryLocation::UnknownSize &&
1389 isObjectSize(O2, V2Size, *DL, *TLI)))
1390 return AliasCache[Locs] = PartialAlias;
1392 AliasResult Result =
1393 AliasAnalysis::alias(MemoryLocation(V1, V1Size, V1AAInfo),
1394 MemoryLocation(V2, V2Size, V2AAInfo));
1395 return AliasCache[Locs] = Result;
1398 /// Check whether two Values can be considered equivalent.
1400 /// In addition to pointer equivalence of \p V1 and \p V2 this checks whether
1401 /// they can not be part of a cycle in the value graph by looking at all
1402 /// visited phi nodes an making sure that the phis cannot reach the value. We
1403 /// have to do this because we are looking through phi nodes (That is we say
1404 /// noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB).
1405 bool BasicAliasAnalysis::isValueEqualInPotentialCycles(const Value *V,
1410 const Instruction *Inst = dyn_cast<Instruction>(V);
1414 if (VisitedPhiBBs.empty())
1417 if (VisitedPhiBBs.size() > MaxNumPhiBBsValueReachabilityCheck)
1420 // Use dominance or loop info if available.
1421 DominatorTreeWrapperPass *DTWP =
1422 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
1423 DominatorTree *DT = DTWP ? &DTWP->getDomTree() : nullptr;
1424 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
1425 LoopInfo *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
1427 // Make sure that the visited phis cannot reach the Value. This ensures that
1428 // the Values cannot come from different iterations of a potential cycle the
1429 // phi nodes could be involved in.
1430 for (auto *P : VisitedPhiBBs)
1431 if (isPotentiallyReachable(P->begin(), Inst, DT, LI))
1437 /// Computes the symbolic difference between two de-composed GEPs.
1439 /// Dest and Src are the variable indices from two decomposed GetElementPtr
1440 /// instructions GEP1 and GEP2 which have common base pointers.
1441 void BasicAliasAnalysis::GetIndexDifference(
1442 SmallVectorImpl<VariableGEPIndex> &Dest,
1443 const SmallVectorImpl<VariableGEPIndex> &Src) {
1447 for (unsigned i = 0, e = Src.size(); i != e; ++i) {
1448 const Value *V = Src[i].V;
1449 unsigned ZExtBits = Src[i].ZExtBits, SExtBits = Src[i].SExtBits;
1450 int64_t Scale = Src[i].Scale;
1452 // Find V in Dest. This is N^2, but pointer indices almost never have more
1453 // than a few variable indexes.
1454 for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
1455 if (!isValueEqualInPotentialCycles(Dest[j].V, V) ||
1456 Dest[j].ZExtBits != ZExtBits || Dest[j].SExtBits != SExtBits)
1459 // If we found it, subtract off Scale V's from the entry in Dest. If it
1460 // goes to zero, remove the entry.
1461 if (Dest[j].Scale != Scale)
1462 Dest[j].Scale -= Scale;
1464 Dest.erase(Dest.begin() + j);
1469 // If we didn't consume this entry, add it to the end of the Dest list.
1471 VariableGEPIndex Entry = {V, ZExtBits, SExtBits, -Scale};
1472 Dest.push_back(Entry);
1477 bool BasicAliasAnalysis::constantOffsetHeuristic(
1478 const SmallVectorImpl<VariableGEPIndex> &VarIndices, uint64_t V1Size,
1479 uint64_t V2Size, int64_t BaseOffset, const DataLayout *DL,
1480 AssumptionCache *AC, DominatorTree *DT) {
1481 if (VarIndices.size() != 2 || V1Size == MemoryLocation::UnknownSize ||
1482 V2Size == MemoryLocation::UnknownSize || !DL)
1485 const VariableGEPIndex &Var0 = VarIndices[0], &Var1 = VarIndices[1];
1487 if (Var0.ZExtBits != Var1.ZExtBits || Var0.SExtBits != Var1.SExtBits ||
1488 Var0.Scale != -Var1.Scale)
1491 unsigned Width = Var1.V->getType()->getIntegerBitWidth();
1493 // We'll strip off the Extensions of Var0 and Var1 and do another round
1494 // of GetLinearExpression decomposition. In the example above, if Var0
1495 // is zext(%x + 1) we should get V1 == %x and V1Offset == 1.
1497 APInt V0Scale(Width, 0), V0Offset(Width, 0), V1Scale(Width, 0),
1499 bool NSW = true, NUW = true;
1500 unsigned V0ZExtBits = 0, V0SExtBits = 0, V1ZExtBits = 0, V1SExtBits = 0;
1501 const Value *V0 = GetLinearExpression(Var0.V, V0Scale, V0Offset, V0ZExtBits,
1502 V0SExtBits, *DL, 0, AC, DT, NSW, NUW);
1503 NSW = true, NUW = true;
1504 const Value *V1 = GetLinearExpression(Var1.V, V1Scale, V1Offset, V1ZExtBits,
1505 V1SExtBits, *DL, 0, AC, DT, NSW, NUW);
1507 if (V0Scale != V1Scale || V0ZExtBits != V1ZExtBits ||
1508 V0SExtBits != V1SExtBits || !isValueEqualInPotentialCycles(V0, V1))
1511 // We have a hit - Var0 and Var1 only differ by a constant offset!
1513 // If we've been sext'ed then zext'd the maximum difference between Var0 and
1514 // Var1 is possible to calculate, but we're just interested in the absolute
1515 // minumum difference between the two. The minimum distance may occur due to
1516 // wrapping; consider "add i3 %i, 5": if %i == 7 then 7 + 5 mod 8 == 4, and so
1517 // the minimum distance between %i and %i + 5 is 3.
1518 APInt MinDiff = V0Offset - V1Offset,
1519 Wrapped = APInt::getMaxValue(Width) - MinDiff + APInt(Width, 1);
1520 MinDiff = APIntOps::umin(MinDiff, Wrapped);
1521 uint64_t MinDiffBytes = MinDiff.getZExtValue() * std::abs(Var0.Scale);
1523 // We can't definitely say whether GEP1 is before or after V2 due to wrapping
1524 // arithmetic (i.e. for some values of GEP1 and V2 GEP1 < V2, and for other
1525 // values GEP1 > V2). We'll therefore only declare NoAlias if both V1Size and
1526 // V2Size can fit in the MinDiffBytes gap.
1527 return V1Size + std::abs(BaseOffset) <= MinDiffBytes &&
1528 V2Size + std::abs(BaseOffset) <= MinDiffBytes;