1 //===- BasicAliasAnalysis.cpp - Stateless Alias Analysis Impl -------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file defines the primary stateless implementation of the
10 // Alias Analysis interface that implements identities (two different
11 // globals cannot alias, etc), but does no stateful analysis.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/BasicAliasAnalysis.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/CFG.h"
23 #include "llvm/Analysis/CaptureTracking.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/MemoryBuiltins.h"
27 #include "llvm/Analysis/MemoryLocation.h"
28 #include "llvm/Analysis/PhiValues.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/Argument.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/Constant.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Dominators.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/GetElementPtrTypeIterator.h"
40 #include "llvm/IR/GlobalAlias.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/InstrTypes.h"
43 #include "llvm/IR/Instruction.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/Intrinsics.h"
47 #include "llvm/IR/Metadata.h"
48 #include "llvm/IR/Operator.h"
49 #include "llvm/IR/Type.h"
50 #include "llvm/IR/User.h"
51 #include "llvm/IR/Value.h"
52 #include "llvm/InitializePasses.h"
53 #include "llvm/Pass.h"
54 #include "llvm/Support/Casting.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Compiler.h"
57 #include "llvm/Support/KnownBits.h"
63 #define DEBUG_TYPE "basicaa"
67 /// Enable analysis of recursive PHI nodes.
68 static cl::opt<bool> EnableRecPhiAnalysis("basicaa-recphi", cl::Hidden,
71 /// By default, even on 32-bit architectures we use 64-bit integers for
72 /// calculations. This will allow us to more-aggressively decompose indexing
73 /// expressions calculated using i64 values (e.g., long long in C) which is
74 /// common enough to worry about.
75 static cl::opt<bool> ForceAtLeast64Bits("basicaa-force-at-least-64b",
76 cl::Hidden, cl::init(true));
77 static cl::opt<bool> DoubleCalcBits("basicaa-double-calc-bits",
78 cl::Hidden, cl::init(false));
80 /// SearchLimitReached / SearchTimes shows how often the limit of
81 /// to decompose GEPs is reached. It will affect the precision
82 /// of basic alias analysis.
83 STATISTIC(SearchLimitReached, "Number of times the limit to "
84 "decompose GEPs is reached");
85 STATISTIC(SearchTimes, "Number of times a GEP is decomposed");
87 /// Cutoff after which to stop analysing a set of phi nodes potentially involved
88 /// in a cycle. Because we are analysing 'through' phi nodes, we need to be
89 /// careful with value equivalence. We use reachability to make sure a value
90 /// cannot be involved in a cycle.
91 const unsigned MaxNumPhiBBsValueReachabilityCheck = 20;
93 // The max limit of the search depth in DecomposeGEPExpression() and
94 // GetUnderlyingObject(), both functions need to use the same search
95 // depth otherwise the algorithm in aliasGEP will assert.
96 static const unsigned MaxLookupSearchDepth = 6;
98 bool BasicAAResult::invalidate(Function &Fn, const PreservedAnalyses &PA,
99 FunctionAnalysisManager::Invalidator &Inv) {
100 // We don't care if this analysis itself is preserved, it has no state. But
101 // we need to check that the analyses it depends on have been. Note that we
102 // may be created without handles to some analyses and in that case don't
104 if (Inv.invalidate<AssumptionAnalysis>(Fn, PA) ||
105 (DT && Inv.invalidate<DominatorTreeAnalysis>(Fn, PA)) ||
106 (LI && Inv.invalidate<LoopAnalysis>(Fn, PA)) ||
107 (PV && Inv.invalidate<PhiValuesAnalysis>(Fn, PA)))
110 // Otherwise this analysis result remains valid.
114 //===----------------------------------------------------------------------===//
116 //===----------------------------------------------------------------------===//
118 /// Returns true if the pointer is to a function-local object that never
119 /// escapes from the function.
120 static bool isNonEscapingLocalObject(
122 SmallDenseMap<const Value *, bool, 8> *IsCapturedCache = nullptr) {
123 SmallDenseMap<const Value *, bool, 8>::iterator CacheIt;
124 if (IsCapturedCache) {
126 std::tie(CacheIt, Inserted) = IsCapturedCache->insert({V, false});
128 // Found cached result, return it!
129 return CacheIt->second;
132 // If this is a local allocation, check to see if it escapes.
133 if (isa<AllocaInst>(V) || isNoAliasCall(V)) {
134 // Set StoreCaptures to True so that we can assume in our callers that the
135 // pointer is not the result of a load instruction. Currently
136 // PointerMayBeCaptured doesn't have any special analysis for the
137 // StoreCaptures=false case; if it did, our callers could be refined to be
139 auto Ret = !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
141 CacheIt->second = Ret;
145 // If this is an argument that corresponds to a byval or noalias argument,
146 // then it has not escaped before entering the function. Check if it escapes
147 // inside the function.
148 if (const Argument *A = dyn_cast<Argument>(V))
149 if (A->hasByValAttr() || A->hasNoAliasAttr()) {
150 // Note even if the argument is marked nocapture, we still need to check
151 // for copies made inside the function. The nocapture attribute only
152 // specifies that there are no copies made that outlive the function.
153 auto Ret = !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
155 CacheIt->second = Ret;
162 /// Returns true if the pointer is one which would have been considered an
163 /// escape by isNonEscapingLocalObject.
164 static bool isEscapeSource(const Value *V) {
165 if (isa<CallBase>(V))
168 if (isa<Argument>(V))
171 // The load case works because isNonEscapingLocalObject considers all
172 // stores to be escapes (it passes true for the StoreCaptures argument
173 // to PointerMayBeCaptured).
174 if (isa<LoadInst>(V))
180 /// Returns the size of the object specified by V or UnknownSize if unknown.
181 static uint64_t getObjectSize(const Value *V, const DataLayout &DL,
182 const TargetLibraryInfo &TLI,
184 bool RoundToAlign = false) {
187 Opts.RoundToAlign = RoundToAlign;
188 Opts.NullIsUnknownSize = NullIsValidLoc;
189 if (getObjectSize(V, Size, DL, &TLI, Opts))
191 return MemoryLocation::UnknownSize;
194 /// Returns true if we can prove that the object specified by V is smaller than
196 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
197 const DataLayout &DL,
198 const TargetLibraryInfo &TLI,
199 bool NullIsValidLoc) {
200 // Note that the meanings of the "object" are slightly different in the
201 // following contexts:
202 // c1: llvm::getObjectSize()
203 // c2: llvm.objectsize() intrinsic
204 // c3: isObjectSmallerThan()
205 // c1 and c2 share the same meaning; however, the meaning of "object" in c3
206 // refers to the "entire object".
208 // Consider this example:
209 // char *p = (char*)malloc(100)
212 // In the context of c1 and c2, the "object" pointed by q refers to the
213 // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
215 // However, in the context of c3, the "object" refers to the chunk of memory
216 // being allocated. So, the "object" has 100 bytes, and q points to the middle
217 // the "object". In case q is passed to isObjectSmallerThan() as the 1st
218 // parameter, before the llvm::getObjectSize() is called to get the size of
219 // entire object, we should:
220 // - either rewind the pointer q to the base-address of the object in
221 // question (in this case rewind to p), or
222 // - just give up. It is up to caller to make sure the pointer is pointing
223 // to the base address the object.
225 // We go for 2nd option for simplicity.
226 if (!isIdentifiedObject(V))
229 // This function needs to use the aligned object size because we allow
230 // reads a bit past the end given sufficient alignment.
231 uint64_t ObjectSize = getObjectSize(V, DL, TLI, NullIsValidLoc,
232 /*RoundToAlign*/ true);
234 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size;
237 /// Return the minimal extent from \p V to the end of the underlying object,
238 /// assuming the result is used in an aliasing query. E.g., we do use the query
239 /// location size and the fact that null pointers cannot alias here.
240 static uint64_t getMinimalExtentFrom(const Value &V,
241 const LocationSize &LocSize,
242 const DataLayout &DL,
243 bool NullIsValidLoc) {
244 // If we have dereferenceability information we know a lower bound for the
245 // extent as accesses for a lower offset would be valid. We need to exclude
246 // the "or null" part if null is a valid pointer.
248 uint64_t DerefBytes = V.getPointerDereferenceableBytes(DL, CanBeNull);
249 DerefBytes = (CanBeNull && NullIsValidLoc) ? 0 : DerefBytes;
250 // If queried with a precise location size, we assume that location size to be
251 // accessed, thus valid.
252 if (LocSize.isPrecise())
253 DerefBytes = std::max(DerefBytes, LocSize.getValue());
257 /// Returns true if we can prove that the object specified by V has size Size.
258 static bool isObjectSize(const Value *V, uint64_t Size, const DataLayout &DL,
259 const TargetLibraryInfo &TLI, bool NullIsValidLoc) {
260 uint64_t ObjectSize = getObjectSize(V, DL, TLI, NullIsValidLoc);
261 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size;
264 //===----------------------------------------------------------------------===//
265 // GetElementPtr Instruction Decomposition and Analysis
266 //===----------------------------------------------------------------------===//
268 /// Analyzes the specified value as a linear expression: "A*V + B", where A and
269 /// B are constant integers.
271 /// Returns the scale and offset values as APInts and return V as a Value*, and
272 /// return whether we looked through any sign or zero extends. The incoming
273 /// Value is known to have IntegerType, and it may already be sign or zero
276 /// Note that this looks through extends, so the high bits may not be
277 /// represented in the result.
278 /*static*/ const Value *BasicAAResult::GetLinearExpression(
279 const Value *V, APInt &Scale, APInt &Offset, unsigned &ZExtBits,
280 unsigned &SExtBits, const DataLayout &DL, unsigned Depth,
281 AssumptionCache *AC, DominatorTree *DT, bool &NSW, bool &NUW) {
282 assert(V->getType()->isIntegerTy() && "Not an integer value");
284 // Limit our recursion depth.
291 if (const ConstantInt *Const = dyn_cast<ConstantInt>(V)) {
292 // If it's a constant, just convert it to an offset and remove the variable.
293 // If we've been called recursively, the Offset bit width will be greater
294 // than the constant's (the Offset's always as wide as the outermost call),
295 // so we'll zext here and process any extension in the isa<SExtInst> &
296 // isa<ZExtInst> cases below.
297 Offset += Const->getValue().zextOrSelf(Offset.getBitWidth());
298 assert(Scale == 0 && "Constant values don't have a scale");
302 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
303 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
304 // If we've been called recursively, then Offset and Scale will be wider
305 // than the BOp operands. We'll always zext it here as we'll process sign
306 // extensions below (see the isa<SExtInst> / isa<ZExtInst> cases).
307 APInt RHS = RHSC->getValue().zextOrSelf(Offset.getBitWidth());
309 switch (BOp->getOpcode()) {
311 // We don't understand this instruction, so we can't decompose it any
316 case Instruction::Or:
317 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
319 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC,
326 case Instruction::Add:
327 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
328 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
331 case Instruction::Sub:
332 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
333 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
336 case Instruction::Mul:
337 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
338 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
342 case Instruction::Shl:
343 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
344 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
346 // We're trying to linearize an expression of the kind:
348 // where the shift count exceeds the bitwidth of the type.
349 // We can't decompose this further (the expression would return
351 if (Offset.getBitWidth() < RHS.getLimitedValue() ||
352 Scale.getBitWidth() < RHS.getLimitedValue()) {
358 Offset <<= RHS.getLimitedValue();
359 Scale <<= RHS.getLimitedValue();
360 // the semantics of nsw and nuw for left shifts don't match those of
361 // multiplications, so we won't propagate them.
366 if (isa<OverflowingBinaryOperator>(BOp)) {
367 NUW &= BOp->hasNoUnsignedWrap();
368 NSW &= BOp->hasNoSignedWrap();
374 // Since GEP indices are sign extended anyway, we don't care about the high
375 // bits of a sign or zero extended value - just scales and offsets. The
376 // extensions have to be consistent though.
377 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) {
378 Value *CastOp = cast<CastInst>(V)->getOperand(0);
379 unsigned NewWidth = V->getType()->getPrimitiveSizeInBits();
380 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
381 unsigned OldZExtBits = ZExtBits, OldSExtBits = SExtBits;
382 const Value *Result =
383 GetLinearExpression(CastOp, Scale, Offset, ZExtBits, SExtBits, DL,
384 Depth + 1, AC, DT, NSW, NUW);
386 // zext(zext(%x)) == zext(%x), and similarly for sext; we'll handle this
387 // by just incrementing the number of bits we've extended by.
388 unsigned ExtendedBy = NewWidth - SmallWidth;
390 if (isa<SExtInst>(V) && ZExtBits == 0) {
391 // sext(sext(%x, a), b) == sext(%x, a + b)
394 // We haven't sign-wrapped, so it's valid to decompose sext(%x + c)
395 // into sext(%x) + sext(c). We'll sext the Offset ourselves:
396 unsigned OldWidth = Offset.getBitWidth();
397 Offset = Offset.trunc(SmallWidth).sext(NewWidth).zextOrSelf(OldWidth);
399 // We may have signed-wrapped, so don't decompose sext(%x + c) into
400 // sext(%x) + sext(c)
404 ZExtBits = OldZExtBits;
405 SExtBits = OldSExtBits;
407 SExtBits += ExtendedBy;
409 // sext(zext(%x, a), b) = zext(zext(%x, a), b) = zext(%x, a + b)
412 // We may have unsigned-wrapped, so don't decompose zext(%x + c) into
413 // zext(%x) + zext(c)
417 ZExtBits = OldZExtBits;
418 SExtBits = OldSExtBits;
420 ZExtBits += ExtendedBy;
431 /// To ensure a pointer offset fits in an integer of size PointerSize
432 /// (in bits) when that size is smaller than the maximum pointer size. This is
433 /// an issue, for example, in particular for 32b pointers with negative indices
434 /// that rely on two's complement wrap-arounds for precise alias information
435 /// where the maximum pointer size is 64b.
436 static APInt adjustToPointerSize(APInt Offset, unsigned PointerSize) {
437 assert(PointerSize <= Offset.getBitWidth() && "Invalid PointerSize!");
438 unsigned ShiftBits = Offset.getBitWidth() - PointerSize;
439 return (Offset << ShiftBits).ashr(ShiftBits);
442 static unsigned getMaxPointerSize(const DataLayout &DL) {
443 unsigned MaxPointerSize = DL.getMaxPointerSizeInBits();
444 if (MaxPointerSize < 64 && ForceAtLeast64Bits) MaxPointerSize = 64;
445 if (DoubleCalcBits) MaxPointerSize *= 2;
447 return MaxPointerSize;
450 /// If V is a symbolic pointer expression, decompose it into a base pointer
451 /// with a constant offset and a number of scaled symbolic offsets.
453 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale
454 /// in the VarIndices vector) are Value*'s that are known to be scaled by the
455 /// specified amount, but which may have other unrepresented high bits. As
456 /// such, the gep cannot necessarily be reconstructed from its decomposed form.
458 /// When DataLayout is around, this function is capable of analyzing everything
459 /// that GetUnderlyingObject can look through. To be able to do that
460 /// GetUnderlyingObject and DecomposeGEPExpression must use the same search
461 /// depth (MaxLookupSearchDepth). When DataLayout not is around, it just looks
462 /// through pointer casts.
463 bool BasicAAResult::DecomposeGEPExpression(const Value *V,
464 DecomposedGEP &Decomposed, const DataLayout &DL, AssumptionCache *AC,
466 // Limit recursion depth to limit compile time in crazy cases.
467 unsigned MaxLookup = MaxLookupSearchDepth;
470 unsigned MaxPointerSize = getMaxPointerSize(DL);
471 Decomposed.VarIndices.clear();
473 // See if this is a bitcast or GEP.
474 const Operator *Op = dyn_cast<Operator>(V);
476 // The only non-operator case we can handle are GlobalAliases.
477 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
478 if (!GA->isInterposable()) {
479 V = GA->getAliasee();
487 if (Op->getOpcode() == Instruction::BitCast ||
488 Op->getOpcode() == Instruction::AddrSpaceCast) {
489 V = Op->getOperand(0);
493 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
495 if (const auto *Call = dyn_cast<CallBase>(V)) {
496 // CaptureTracking can know about special capturing properties of some
497 // intrinsics like launder.invariant.group, that can't be expressed with
498 // the attributes, but have properties like returning aliasing pointer.
499 // Because some analysis may assume that nocaptured pointer is not
500 // returned from some special intrinsic (because function would have to
501 // be marked with returns attribute), it is crucial to use this function
502 // because it should be in sync with CaptureTracking. Not using it may
503 // cause weird miscompilations where 2 aliasing pointers are assumed to
505 if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) {
511 // If it's not a GEP, hand it off to SimplifyInstruction to see if it
512 // can come up with something. This matches what GetUnderlyingObject does.
513 if (const Instruction *I = dyn_cast<Instruction>(V))
514 // TODO: Get a DominatorTree and AssumptionCache and use them here
515 // (these are both now available in this function, but this should be
516 // updated when GetUnderlyingObject is updated). TLI should be
518 if (const Value *Simplified =
519 SimplifyInstruction(const_cast<Instruction *>(I), DL)) {
528 // Don't attempt to analyze GEPs over unsized objects.
529 if (!GEPOp->getSourceElementType()->isSized()) {
534 unsigned AS = GEPOp->getPointerAddressSpace();
535 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
536 gep_type_iterator GTI = gep_type_begin(GEPOp);
537 unsigned PointerSize = DL.getPointerSizeInBits(AS);
538 // Assume all GEP operands are constants until proven otherwise.
539 bool GepHasConstantOffset = true;
540 for (User::const_op_iterator I = GEPOp->op_begin() + 1, E = GEPOp->op_end();
541 I != E; ++I, ++GTI) {
542 const Value *Index = *I;
543 // Compute the (potentially symbolic) offset in bytes for this index.
544 if (StructType *STy = GTI.getStructTypeOrNull()) {
545 // For a struct, add the member offset.
546 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
550 Decomposed.StructOffset +=
551 DL.getStructLayout(STy)->getElementOffset(FieldNo);
555 // For an array/pointer, add the element offset, explicitly scaled.
556 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
559 Decomposed.OtherOffset +=
560 (DL.getTypeAllocSize(GTI.getIndexedType()) *
561 CIdx->getValue().sextOrSelf(MaxPointerSize))
562 .sextOrTrunc(MaxPointerSize);
566 GepHasConstantOffset = false;
568 APInt Scale(MaxPointerSize, DL.getTypeAllocSize(GTI.getIndexedType()));
569 unsigned ZExtBits = 0, SExtBits = 0;
571 // If the integer type is smaller than the pointer size, it is implicitly
572 // sign extended to pointer size.
573 unsigned Width = Index->getType()->getIntegerBitWidth();
574 if (PointerSize > Width)
575 SExtBits += PointerSize - Width;
577 // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
578 APInt IndexScale(Width, 0), IndexOffset(Width, 0);
579 bool NSW = true, NUW = true;
580 const Value *OrigIndex = Index;
581 Index = GetLinearExpression(Index, IndexScale, IndexOffset, ZExtBits,
582 SExtBits, DL, 0, AC, DT, NSW, NUW);
584 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
585 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
587 // It can be the case that, even through C1*V+C2 does not overflow for
588 // relevant values of V, (C2*Scale) can overflow. In that case, we cannot
589 // decompose the expression in this way.
591 // FIXME: C1*Scale and the other operations in the decomposed
592 // (C1*Scale)*V+C2*Scale can also overflow. We should check for this
594 APInt WideScaledOffset = IndexOffset.sextOrTrunc(MaxPointerSize*2) *
595 Scale.sext(MaxPointerSize*2);
596 if (WideScaledOffset.getMinSignedBits() > MaxPointerSize) {
601 ZExtBits = SExtBits = 0;
602 if (PointerSize > Width)
603 SExtBits += PointerSize - Width;
605 Decomposed.OtherOffset += IndexOffset.sextOrTrunc(MaxPointerSize) * Scale;
606 Scale *= IndexScale.sextOrTrunc(MaxPointerSize);
609 // If we already had an occurrence of this index variable, merge this
610 // scale into it. For example, we want to handle:
611 // A[x][x] -> x*16 + x*4 -> x*20
612 // This also ensures that 'x' only appears in the index list once.
613 for (unsigned i = 0, e = Decomposed.VarIndices.size(); i != e; ++i) {
614 if (Decomposed.VarIndices[i].V == Index &&
615 Decomposed.VarIndices[i].ZExtBits == ZExtBits &&
616 Decomposed.VarIndices[i].SExtBits == SExtBits) {
617 Scale += Decomposed.VarIndices[i].Scale;
618 Decomposed.VarIndices.erase(Decomposed.VarIndices.begin() + i);
623 // Make sure that we have a scale that makes sense for this target's
625 Scale = adjustToPointerSize(Scale, PointerSize);
628 VariableGEPIndex Entry = {Index, ZExtBits, SExtBits, Scale};
629 Decomposed.VarIndices.push_back(Entry);
633 // Take care of wrap-arounds
634 if (GepHasConstantOffset) {
635 Decomposed.StructOffset =
636 adjustToPointerSize(Decomposed.StructOffset, PointerSize);
637 Decomposed.OtherOffset =
638 adjustToPointerSize(Decomposed.OtherOffset, PointerSize);
641 // Analyze the base pointer next.
642 V = GEPOp->getOperand(0);
643 } while (--MaxLookup);
645 // If the chain of expressions is too deep, just return early.
647 SearchLimitReached++;
651 /// Returns whether the given pointer value points to memory that is local to
652 /// the function, with global constants being considered local to all
654 bool BasicAAResult::pointsToConstantMemory(const MemoryLocation &Loc,
655 AAQueryInfo &AAQI, bool OrLocal) {
656 assert(Visited.empty() && "Visited must be cleared after use!");
658 unsigned MaxLookup = 8;
659 SmallVector<const Value *, 16> Worklist;
660 Worklist.push_back(Loc.Ptr);
662 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), DL);
663 if (!Visited.insert(V).second) {
665 return AAResultBase::pointsToConstantMemory(Loc, AAQI, OrLocal);
668 // An alloca instruction defines local memory.
669 if (OrLocal && isa<AllocaInst>(V))
672 // A global constant counts as local memory for our purposes.
673 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
674 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
675 // global to be marked constant in some modules and non-constant in
676 // others. GV may even be a declaration, not a definition.
677 if (!GV->isConstant()) {
679 return AAResultBase::pointsToConstantMemory(Loc, AAQI, OrLocal);
684 // If both select values point to local memory, then so does the select.
685 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
686 Worklist.push_back(SI->getTrueValue());
687 Worklist.push_back(SI->getFalseValue());
691 // If all values incoming to a phi node point to local memory, then so does
693 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
694 // Don't bother inspecting phi nodes with many operands.
695 if (PN->getNumIncomingValues() > MaxLookup) {
697 return AAResultBase::pointsToConstantMemory(Loc, AAQI, OrLocal);
699 for (Value *IncValue : PN->incoming_values())
700 Worklist.push_back(IncValue);
704 // Otherwise be conservative.
706 return AAResultBase::pointsToConstantMemory(Loc, AAQI, OrLocal);
707 } while (!Worklist.empty() && --MaxLookup);
710 return Worklist.empty();
713 /// Returns the behavior when calling the given call site.
714 FunctionModRefBehavior BasicAAResult::getModRefBehavior(const CallBase *Call) {
715 if (Call->doesNotAccessMemory())
716 // Can't do better than this.
717 return FMRB_DoesNotAccessMemory;
719 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
721 // If the callsite knows it only reads memory, don't return worse
723 if (Call->onlyReadsMemory())
724 Min = FMRB_OnlyReadsMemory;
725 else if (Call->doesNotReadMemory())
726 Min = FMRB_DoesNotReadMemory;
728 if (Call->onlyAccessesArgMemory())
729 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
730 else if (Call->onlyAccessesInaccessibleMemory())
731 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleMem);
732 else if (Call->onlyAccessesInaccessibleMemOrArgMem())
733 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleOrArgMem);
735 // If the call has operand bundles then aliasing attributes from the function
736 // it calls do not directly apply to the call. This can be made more precise
738 if (!Call->hasOperandBundles())
739 if (const Function *F = Call->getCalledFunction())
741 FunctionModRefBehavior(Min & getBestAAResults().getModRefBehavior(F));
746 /// Returns the behavior when calling the given function. For use when the call
747 /// site is not known.
748 FunctionModRefBehavior BasicAAResult::getModRefBehavior(const Function *F) {
749 // If the function declares it doesn't access memory, we can't do better.
750 if (F->doesNotAccessMemory())
751 return FMRB_DoesNotAccessMemory;
753 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
755 // If the function declares it only reads memory, go with that.
756 if (F->onlyReadsMemory())
757 Min = FMRB_OnlyReadsMemory;
758 else if (F->doesNotReadMemory())
759 Min = FMRB_DoesNotReadMemory;
761 if (F->onlyAccessesArgMemory())
762 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
763 else if (F->onlyAccessesInaccessibleMemory())
764 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleMem);
765 else if (F->onlyAccessesInaccessibleMemOrArgMem())
766 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleOrArgMem);
771 /// Returns true if this is a writeonly (i.e Mod only) parameter.
772 static bool isWriteOnlyParam(const CallBase *Call, unsigned ArgIdx,
773 const TargetLibraryInfo &TLI) {
774 if (Call->paramHasAttr(ArgIdx, Attribute::WriteOnly))
777 // We can bound the aliasing properties of memset_pattern16 just as we can
778 // for memcpy/memset. This is particularly important because the
779 // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
780 // whenever possible.
781 // FIXME Consider handling this in InferFunctionAttr.cpp together with other
784 if (Call->getCalledFunction() &&
785 TLI.getLibFunc(*Call->getCalledFunction(), F) &&
786 F == LibFunc_memset_pattern16 && TLI.has(F))
790 // TODO: memset_pattern4, memset_pattern8
791 // TODO: _chk variants
792 // TODO: strcmp, strcpy
797 ModRefInfo BasicAAResult::getArgModRefInfo(const CallBase *Call,
799 // Checking for known builtin intrinsics and target library functions.
800 if (isWriteOnlyParam(Call, ArgIdx, TLI))
801 return ModRefInfo::Mod;
803 if (Call->paramHasAttr(ArgIdx, Attribute::ReadOnly))
804 return ModRefInfo::Ref;
806 if (Call->paramHasAttr(ArgIdx, Attribute::ReadNone))
807 return ModRefInfo::NoModRef;
809 return AAResultBase::getArgModRefInfo(Call, ArgIdx);
812 static bool isIntrinsicCall(const CallBase *Call, Intrinsic::ID IID) {
813 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Call);
814 return II && II->getIntrinsicID() == IID;
818 static const Function *getParent(const Value *V) {
819 if (const Instruction *inst = dyn_cast<Instruction>(V)) {
820 if (!inst->getParent())
822 return inst->getParent()->getParent();
825 if (const Argument *arg = dyn_cast<Argument>(V))
826 return arg->getParent();
831 static bool notDifferentParent(const Value *O1, const Value *O2) {
833 const Function *F1 = getParent(O1);
834 const Function *F2 = getParent(O2);
836 return !F1 || !F2 || F1 == F2;
840 AliasResult BasicAAResult::alias(const MemoryLocation &LocA,
841 const MemoryLocation &LocB,
843 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
844 "BasicAliasAnalysis doesn't support interprocedural queries.");
846 // If we have a directly cached entry for these locations, we have recursed
847 // through this once, so just return the cached results. Notably, when this
848 // happens, we don't clear the cache.
849 auto CacheIt = AAQI.AliasCache.find(AAQueryInfo::LocPair(LocA, LocB));
850 if (CacheIt != AAQI.AliasCache.end())
851 return CacheIt->second;
853 CacheIt = AAQI.AliasCache.find(AAQueryInfo::LocPair(LocB, LocA));
854 if (CacheIt != AAQI.AliasCache.end())
855 return CacheIt->second;
857 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.AATags, LocB.Ptr,
858 LocB.Size, LocB.AATags, AAQI);
860 VisitedPhiBBs.clear();
864 /// Checks to see if the specified callsite can clobber the specified memory
867 /// Since we only look at local properties of this function, we really can't
868 /// say much about this query. We do, however, use simple "address taken"
869 /// analysis on local objects.
870 ModRefInfo BasicAAResult::getModRefInfo(const CallBase *Call,
871 const MemoryLocation &Loc,
873 assert(notDifferentParent(Call, Loc.Ptr) &&
874 "AliasAnalysis query involving multiple functions!");
876 const Value *Object = GetUnderlyingObject(Loc.Ptr, DL);
878 // Calls marked 'tail' cannot read or write allocas from the current frame
879 // because the current frame might be destroyed by the time they run. However,
880 // a tail call may use an alloca with byval. Calling with byval copies the
881 // contents of the alloca into argument registers or stack slots, so there is
882 // no lifetime issue.
883 if (isa<AllocaInst>(Object))
884 if (const CallInst *CI = dyn_cast<CallInst>(Call))
885 if (CI->isTailCall() &&
886 !CI->getAttributes().hasAttrSomewhere(Attribute::ByVal))
887 return ModRefInfo::NoModRef;
889 // Stack restore is able to modify unescaped dynamic allocas. Assume it may
890 // modify them even though the alloca is not escaped.
891 if (auto *AI = dyn_cast<AllocaInst>(Object))
892 if (!AI->isStaticAlloca() && isIntrinsicCall(Call, Intrinsic::stackrestore))
893 return ModRefInfo::Mod;
895 // If the pointer is to a locally allocated object that does not escape,
896 // then the call can not mod/ref the pointer unless the call takes the pointer
897 // as an argument, and itself doesn't capture it.
898 if (!isa<Constant>(Object) && Call != Object &&
899 isNonEscapingLocalObject(Object, &AAQI.IsCapturedCache)) {
901 // Optimistically assume that call doesn't touch Object and check this
902 // assumption in the following loop.
903 ModRefInfo Result = ModRefInfo::NoModRef;
904 bool IsMustAlias = true;
906 unsigned OperandNo = 0;
907 for (auto CI = Call->data_operands_begin(), CE = Call->data_operands_end();
908 CI != CE; ++CI, ++OperandNo) {
909 // Only look at the no-capture or byval pointer arguments. If this
910 // pointer were passed to arguments that were neither of these, then it
911 // couldn't be no-capture.
912 if (!(*CI)->getType()->isPointerTy() ||
913 (!Call->doesNotCapture(OperandNo) &&
914 OperandNo < Call->getNumArgOperands() &&
915 !Call->isByValArgument(OperandNo)))
918 // Call doesn't access memory through this operand, so we don't care
919 // if it aliases with Object.
920 if (Call->doesNotAccessMemory(OperandNo))
923 // If this is a no-capture pointer argument, see if we can tell that it
924 // is impossible to alias the pointer we're checking.
925 AliasResult AR = getBestAAResults().alias(MemoryLocation(*CI),
926 MemoryLocation(Object), AAQI);
929 // Operand doesn't alias 'Object', continue looking for other aliases
932 // Operand aliases 'Object', but call doesn't modify it. Strengthen
933 // initial assumption and keep looking in case if there are more aliases.
934 if (Call->onlyReadsMemory(OperandNo)) {
935 Result = setRef(Result);
938 // Operand aliases 'Object' but call only writes into it.
939 if (Call->doesNotReadMemory(OperandNo)) {
940 Result = setMod(Result);
943 // This operand aliases 'Object' and call reads and writes into it.
944 // Setting ModRef will not yield an early return below, MustAlias is not
946 Result = ModRefInfo::ModRef;
950 // No operand aliases, reset Must bit. Add below if at least one aliases
951 // and all aliases found are MustAlias.
952 if (isNoModRef(Result))
955 // Early return if we improved mod ref information
956 if (!isModAndRefSet(Result)) {
957 if (isNoModRef(Result))
958 return ModRefInfo::NoModRef;
959 return IsMustAlias ? setMust(Result) : clearMust(Result);
963 // If the call is to malloc or calloc, we can assume that it doesn't
964 // modify any IR visible value. This is only valid because we assume these
965 // routines do not read values visible in the IR. TODO: Consider special
966 // casing realloc and strdup routines which access only their arguments as
967 // well. Or alternatively, replace all of this with inaccessiblememonly once
968 // that's implemented fully.
969 if (isMallocOrCallocLikeFn(Call, &TLI)) {
970 // Be conservative if the accessed pointer may alias the allocation -
971 // fallback to the generic handling below.
972 if (getBestAAResults().alias(MemoryLocation(Call), Loc, AAQI) == NoAlias)
973 return ModRefInfo::NoModRef;
976 // The semantics of memcpy intrinsics forbid overlap between their respective
977 // operands, i.e., source and destination of any given memcpy must no-alias.
978 // If Loc must-aliases either one of these two locations, then it necessarily
979 // no-aliases the other.
980 if (auto *Inst = dyn_cast<AnyMemCpyInst>(Call)) {
981 AliasResult SrcAA, DestAA;
983 if ((SrcAA = getBestAAResults().alias(MemoryLocation::getForSource(Inst),
984 Loc, AAQI)) == MustAlias)
985 // Loc is exactly the memcpy source thus disjoint from memcpy dest.
986 return ModRefInfo::Ref;
987 if ((DestAA = getBestAAResults().alias(MemoryLocation::getForDest(Inst),
988 Loc, AAQI)) == MustAlias)
989 // The converse case.
990 return ModRefInfo::Mod;
992 // It's also possible for Loc to alias both src and dest, or neither.
993 ModRefInfo rv = ModRefInfo::NoModRef;
994 if (SrcAA != NoAlias)
996 if (DestAA != NoAlias)
1001 // While the assume intrinsic is marked as arbitrarily writing so that
1002 // proper control dependencies will be maintained, it never aliases any
1003 // particular memory location.
1004 if (isIntrinsicCall(Call, Intrinsic::assume))
1005 return ModRefInfo::NoModRef;
1007 // Like assumes, guard intrinsics are also marked as arbitrarily writing so
1008 // that proper control dependencies are maintained but they never mods any
1009 // particular memory location.
1011 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the
1012 // heap state at the point the guard is issued needs to be consistent in case
1013 // the guard invokes the "deopt" continuation.
1014 if (isIntrinsicCall(Call, Intrinsic::experimental_guard))
1015 return ModRefInfo::Ref;
1017 // Like assumes, invariant.start intrinsics were also marked as arbitrarily
1018 // writing so that proper control dependencies are maintained but they never
1019 // mod any particular memory location visible to the IR.
1020 // *Unlike* assumes (which are now modeled as NoModRef), invariant.start
1021 // intrinsic is now modeled as reading memory. This prevents hoisting the
1022 // invariant.start intrinsic over stores. Consider:
1025 // invariant_start(ptr)
1029 // This cannot be transformed to:
1032 // invariant_start(ptr)
1037 // The transformation will cause the second store to be ignored (based on
1038 // rules of invariant.start) and print 40, while the first program always
1040 if (isIntrinsicCall(Call, Intrinsic::invariant_start))
1041 return ModRefInfo::Ref;
1043 // The AAResultBase base class has some smarts, lets use them.
1044 return AAResultBase::getModRefInfo(Call, Loc, AAQI);
1047 ModRefInfo BasicAAResult::getModRefInfo(const CallBase *Call1,
1048 const CallBase *Call2,
1049 AAQueryInfo &AAQI) {
1050 // While the assume intrinsic is marked as arbitrarily writing so that
1051 // proper control dependencies will be maintained, it never aliases any
1052 // particular memory location.
1053 if (isIntrinsicCall(Call1, Intrinsic::assume) ||
1054 isIntrinsicCall(Call2, Intrinsic::assume))
1055 return ModRefInfo::NoModRef;
1057 // Like assumes, guard intrinsics are also marked as arbitrarily writing so
1058 // that proper control dependencies are maintained but they never mod any
1059 // particular memory location.
1061 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the
1062 // heap state at the point the guard is issued needs to be consistent in case
1063 // the guard invokes the "deopt" continuation.
1065 // NB! This function is *not* commutative, so we special case two
1066 // possibilities for guard intrinsics.
1068 if (isIntrinsicCall(Call1, Intrinsic::experimental_guard))
1069 return isModSet(createModRefInfo(getModRefBehavior(Call2)))
1071 : ModRefInfo::NoModRef;
1073 if (isIntrinsicCall(Call2, Intrinsic::experimental_guard))
1074 return isModSet(createModRefInfo(getModRefBehavior(Call1)))
1076 : ModRefInfo::NoModRef;
1078 // The AAResultBase base class has some smarts, lets use them.
1079 return AAResultBase::getModRefInfo(Call1, Call2, AAQI);
1082 /// Provide ad-hoc rules to disambiguate accesses through two GEP operators,
1083 /// both having the exact same pointer operand.
1084 static AliasResult aliasSameBasePointerGEPs(const GEPOperator *GEP1,
1085 LocationSize MaybeV1Size,
1086 const GEPOperator *GEP2,
1087 LocationSize MaybeV2Size,
1088 const DataLayout &DL) {
1089 assert(GEP1->getPointerOperand()->stripPointerCastsAndInvariantGroups() ==
1090 GEP2->getPointerOperand()->stripPointerCastsAndInvariantGroups() &&
1091 GEP1->getPointerOperandType() == GEP2->getPointerOperandType() &&
1092 "Expected GEPs with the same pointer operand");
1094 // Try to determine whether GEP1 and GEP2 index through arrays, into structs,
1095 // such that the struct field accesses provably cannot alias.
1096 // We also need at least two indices (the pointer, and the struct field).
1097 if (GEP1->getNumIndices() != GEP2->getNumIndices() ||
1098 GEP1->getNumIndices() < 2)
1101 // If we don't know the size of the accesses through both GEPs, we can't
1102 // determine whether the struct fields accessed can't alias.
1103 if (MaybeV1Size == LocationSize::unknown() ||
1104 MaybeV2Size == LocationSize::unknown())
1107 const uint64_t V1Size = MaybeV1Size.getValue();
1108 const uint64_t V2Size = MaybeV2Size.getValue();
1111 dyn_cast<ConstantInt>(GEP1->getOperand(GEP1->getNumOperands() - 1));
1113 dyn_cast<ConstantInt>(GEP2->getOperand(GEP2->getNumOperands() - 1));
1115 // If the last (struct) indices are constants and are equal, the other indices
1116 // might be also be dynamically equal, so the GEPs can alias.
1118 unsigned BitWidth = std::max(C1->getBitWidth(), C2->getBitWidth());
1119 if (C1->getValue().sextOrSelf(BitWidth) ==
1120 C2->getValue().sextOrSelf(BitWidth))
1124 // Find the last-indexed type of the GEP, i.e., the type you'd get if
1125 // you stripped the last index.
1126 // On the way, look at each indexed type. If there's something other
1127 // than an array, different indices can lead to different final types.
1128 SmallVector<Value *, 8> IntermediateIndices;
1130 // Insert the first index; we don't need to check the type indexed
1131 // through it as it only drops the pointer indirection.
1132 assert(GEP1->getNumIndices() > 1 && "Not enough GEP indices to examine");
1133 IntermediateIndices.push_back(GEP1->getOperand(1));
1135 // Insert all the remaining indices but the last one.
1136 // Also, check that they all index through arrays.
1137 for (unsigned i = 1, e = GEP1->getNumIndices() - 1; i != e; ++i) {
1138 if (!isa<ArrayType>(GetElementPtrInst::getIndexedType(
1139 GEP1->getSourceElementType(), IntermediateIndices)))
1141 IntermediateIndices.push_back(GEP1->getOperand(i + 1));
1144 auto *Ty = GetElementPtrInst::getIndexedType(
1145 GEP1->getSourceElementType(), IntermediateIndices);
1146 StructType *LastIndexedStruct = dyn_cast<StructType>(Ty);
1148 if (isa<SequentialType>(Ty)) {
1150 // - both GEPs begin indexing from the exact same pointer;
1151 // - the last indices in both GEPs are constants, indexing into a sequential
1152 // type (array or pointer);
1153 // - both GEPs only index through arrays prior to that.
1155 // Because array indices greater than the number of elements are valid in
1156 // GEPs, unless we know the intermediate indices are identical between
1157 // GEP1 and GEP2 we cannot guarantee that the last indexed arrays don't
1158 // partially overlap. We also need to check that the loaded size matches
1159 // the element size, otherwise we could still have overlap.
1160 const uint64_t ElementSize =
1161 DL.getTypeStoreSize(cast<SequentialType>(Ty)->getElementType());
1162 if (V1Size != ElementSize || V2Size != ElementSize)
1165 for (unsigned i = 0, e = GEP1->getNumIndices() - 1; i != e; ++i)
1166 if (GEP1->getOperand(i + 1) != GEP2->getOperand(i + 1))
1169 // Now we know that the array/pointer that GEP1 indexes into and that
1170 // that GEP2 indexes into must either precisely overlap or be disjoint.
1171 // Because they cannot partially overlap and because fields in an array
1172 // cannot overlap, if we can prove the final indices are different between
1173 // GEP1 and GEP2, we can conclude GEP1 and GEP2 don't alias.
1175 // If the last indices are constants, we've already checked they don't
1176 // equal each other so we can exit early.
1180 Value *GEP1LastIdx = GEP1->getOperand(GEP1->getNumOperands() - 1);
1181 Value *GEP2LastIdx = GEP2->getOperand(GEP2->getNumOperands() - 1);
1182 if (isa<PHINode>(GEP1LastIdx) || isa<PHINode>(GEP2LastIdx)) {
1183 // If one of the indices is a PHI node, be safe and only use
1184 // computeKnownBits so we don't make any assumptions about the
1185 // relationships between the two indices. This is important if we're
1186 // asking about values from different loop iterations. See PR32314.
1187 // TODO: We may be able to change the check so we only do this when
1188 // we definitely looked through a PHINode.
1189 if (GEP1LastIdx != GEP2LastIdx &&
1190 GEP1LastIdx->getType() == GEP2LastIdx->getType()) {
1191 KnownBits Known1 = computeKnownBits(GEP1LastIdx, DL);
1192 KnownBits Known2 = computeKnownBits(GEP2LastIdx, DL);
1193 if (Known1.Zero.intersects(Known2.One) ||
1194 Known1.One.intersects(Known2.Zero))
1197 } else if (isKnownNonEqual(GEP1LastIdx, GEP2LastIdx, DL))
1201 } else if (!LastIndexedStruct || !C1 || !C2) {
1205 if (C1->getValue().getActiveBits() > 64 ||
1206 C2->getValue().getActiveBits() > 64)
1210 // - both GEPs begin indexing from the exact same pointer;
1211 // - the last indices in both GEPs are constants, indexing into a struct;
1212 // - said indices are different, hence, the pointed-to fields are different;
1213 // - both GEPs only index through arrays prior to that.
1215 // This lets us determine that the struct that GEP1 indexes into and the
1216 // struct that GEP2 indexes into must either precisely overlap or be
1217 // completely disjoint. Because they cannot partially overlap, indexing into
1218 // different non-overlapping fields of the struct will never alias.
1220 // Therefore, the only remaining thing needed to show that both GEPs can't
1221 // alias is that the fields are not overlapping.
1222 const StructLayout *SL = DL.getStructLayout(LastIndexedStruct);
1223 const uint64_t StructSize = SL->getSizeInBytes();
1224 const uint64_t V1Off = SL->getElementOffset(C1->getZExtValue());
1225 const uint64_t V2Off = SL->getElementOffset(C2->getZExtValue());
1227 auto EltsDontOverlap = [StructSize](uint64_t V1Off, uint64_t V1Size,
1228 uint64_t V2Off, uint64_t V2Size) {
1229 return V1Off < V2Off && V1Off + V1Size <= V2Off &&
1230 ((V2Off + V2Size <= StructSize) ||
1231 (V2Off + V2Size - StructSize <= V1Off));
1234 if (EltsDontOverlap(V1Off, V1Size, V2Off, V2Size) ||
1235 EltsDontOverlap(V2Off, V2Size, V1Off, V1Size))
1241 // If a we have (a) a GEP and (b) a pointer based on an alloca, and the
1242 // beginning of the object the GEP points would have a negative offset with
1243 // repsect to the alloca, that means the GEP can not alias pointer (b).
1244 // Note that the pointer based on the alloca may not be a GEP. For
1245 // example, it may be the alloca itself.
1246 // The same applies if (b) is based on a GlobalVariable. Note that just being
1247 // based on isIdentifiedObject() is not enough - we need an identified object
1248 // that does not permit access to negative offsets. For example, a negative
1249 // offset from a noalias argument or call can be inbounds w.r.t the actual
1250 // underlying object.
1252 // For example, consider:
1254 // struct { int f0, int f1, ...} foo;
1256 // foo* random = bar(alloca);
1257 // int *f0 = &alloca.f0
1258 // int *f1 = &random->f1;
1260 // Which is lowered, approximately, to:
1262 // %alloca = alloca %struct.foo
1263 // %random = call %struct.foo* @random(%struct.foo* %alloca)
1264 // %f0 = getelementptr inbounds %struct, %struct.foo* %alloca, i32 0, i32 0
1265 // %f1 = getelementptr inbounds %struct, %struct.foo* %random, i32 0, i32 1
1267 // Assume %f1 and %f0 alias. Then %f1 would point into the object allocated
1268 // by %alloca. Since the %f1 GEP is inbounds, that means %random must also
1269 // point into the same object. But since %f0 points to the beginning of %alloca,
1270 // the highest %f1 can be is (%alloca + 3). This means %random can not be higher
1271 // than (%alloca - 1), and so is not inbounds, a contradiction.
1272 bool BasicAAResult::isGEPBaseAtNegativeOffset(const GEPOperator *GEPOp,
1273 const DecomposedGEP &DecompGEP, const DecomposedGEP &DecompObject,
1274 LocationSize MaybeObjectAccessSize) {
1275 // If the object access size is unknown, or the GEP isn't inbounds, bail.
1276 if (MaybeObjectAccessSize == LocationSize::unknown() || !GEPOp->isInBounds())
1279 const uint64_t ObjectAccessSize = MaybeObjectAccessSize.getValue();
1281 // We need the object to be an alloca or a globalvariable, and want to know
1282 // the offset of the pointer from the object precisely, so no variable
1283 // indices are allowed.
1284 if (!(isa<AllocaInst>(DecompObject.Base) ||
1285 isa<GlobalVariable>(DecompObject.Base)) ||
1286 !DecompObject.VarIndices.empty())
1289 APInt ObjectBaseOffset = DecompObject.StructOffset +
1290 DecompObject.OtherOffset;
1292 // If the GEP has no variable indices, we know the precise offset
1293 // from the base, then use it. If the GEP has variable indices,
1294 // we can't get exact GEP offset to identify pointer alias. So return
1295 // false in that case.
1296 if (!DecompGEP.VarIndices.empty())
1299 APInt GEPBaseOffset = DecompGEP.StructOffset;
1300 GEPBaseOffset += DecompGEP.OtherOffset;
1302 return GEPBaseOffset.sge(ObjectBaseOffset + (int64_t)ObjectAccessSize);
1305 /// Provides a bunch of ad-hoc rules to disambiguate a GEP instruction against
1306 /// another pointer.
1308 /// We know that V1 is a GEP, but we don't know anything about V2.
1309 /// UnderlyingV1 is GetUnderlyingObject(GEP1, DL), UnderlyingV2 is the same for
1311 AliasResult BasicAAResult::aliasGEP(
1312 const GEPOperator *GEP1, LocationSize V1Size, const AAMDNodes &V1AAInfo,
1313 const Value *V2, LocationSize V2Size, const AAMDNodes &V2AAInfo,
1314 const Value *UnderlyingV1, const Value *UnderlyingV2, AAQueryInfo &AAQI) {
1315 DecomposedGEP DecompGEP1, DecompGEP2;
1316 unsigned MaxPointerSize = getMaxPointerSize(DL);
1317 DecompGEP1.StructOffset = DecompGEP1.OtherOffset = APInt(MaxPointerSize, 0);
1318 DecompGEP2.StructOffset = DecompGEP2.OtherOffset = APInt(MaxPointerSize, 0);
1320 bool GEP1MaxLookupReached =
1321 DecomposeGEPExpression(GEP1, DecompGEP1, DL, &AC, DT);
1322 bool GEP2MaxLookupReached =
1323 DecomposeGEPExpression(V2, DecompGEP2, DL, &AC, DT);
1325 APInt GEP1BaseOffset = DecompGEP1.StructOffset + DecompGEP1.OtherOffset;
1326 APInt GEP2BaseOffset = DecompGEP2.StructOffset + DecompGEP2.OtherOffset;
1328 assert(DecompGEP1.Base == UnderlyingV1 && DecompGEP2.Base == UnderlyingV2 &&
1329 "DecomposeGEPExpression returned a result different from "
1330 "GetUnderlyingObject");
1332 // If the GEP's offset relative to its base is such that the base would
1333 // fall below the start of the object underlying V2, then the GEP and V2
1335 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached &&
1336 isGEPBaseAtNegativeOffset(GEP1, DecompGEP1, DecompGEP2, V2Size))
1338 // If we have two gep instructions with must-alias or not-alias'ing base
1339 // pointers, figure out if the indexes to the GEP tell us anything about the
1341 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
1342 // Check for the GEP base being at a negative offset, this time in the other
1344 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached &&
1345 isGEPBaseAtNegativeOffset(GEP2, DecompGEP2, DecompGEP1, V1Size))
1347 // Do the base pointers alias?
1348 AliasResult BaseAlias =
1349 aliasCheck(UnderlyingV1, LocationSize::unknown(), AAMDNodes(),
1350 UnderlyingV2, LocationSize::unknown(), AAMDNodes(), AAQI);
1352 // Check for geps of non-aliasing underlying pointers where the offsets are
1354 if ((BaseAlias == MayAlias) && V1Size == V2Size) {
1355 // Do the base pointers alias assuming type and size.
1356 AliasResult PreciseBaseAlias = aliasCheck(
1357 UnderlyingV1, V1Size, V1AAInfo, UnderlyingV2, V2Size, V2AAInfo, AAQI);
1358 if (PreciseBaseAlias == NoAlias) {
1359 // See if the computed offset from the common pointer tells us about the
1360 // relation of the resulting pointer.
1361 // If the max search depth is reached the result is undefined
1362 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1366 if (GEP1BaseOffset == GEP2BaseOffset &&
1367 DecompGEP1.VarIndices == DecompGEP2.VarIndices)
1372 // If we get a No or May, then return it immediately, no amount of analysis
1373 // will improve this situation.
1374 if (BaseAlias != MustAlias) {
1375 assert(BaseAlias == NoAlias || BaseAlias == MayAlias);
1379 // Otherwise, we have a MustAlias. Since the base pointers alias each other
1380 // exactly, see if the computed offset from the common pointer tells us
1381 // about the relation of the resulting pointer.
1382 // If we know the two GEPs are based off of the exact same pointer (and not
1383 // just the same underlying object), see if that tells us anything about
1384 // the resulting pointers.
1385 if (GEP1->getPointerOperand()->stripPointerCastsAndInvariantGroups() ==
1386 GEP2->getPointerOperand()->stripPointerCastsAndInvariantGroups() &&
1387 GEP1->getPointerOperandType() == GEP2->getPointerOperandType()) {
1388 AliasResult R = aliasSameBasePointerGEPs(GEP1, V1Size, GEP2, V2Size, DL);
1389 // If we couldn't find anything interesting, don't abandon just yet.
1394 // If the max search depth is reached, the result is undefined
1395 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1398 // Subtract the GEP2 pointer from the GEP1 pointer to find out their
1399 // symbolic difference.
1400 GEP1BaseOffset -= GEP2BaseOffset;
1401 GetIndexDifference(DecompGEP1.VarIndices, DecompGEP2.VarIndices);
1404 // Check to see if these two pointers are related by the getelementptr
1405 // instruction. If one pointer is a GEP with a non-zero index of the other
1406 // pointer, we know they cannot alias.
1408 // If both accesses are unknown size, we can't do anything useful here.
1409 if (V1Size == LocationSize::unknown() && V2Size == LocationSize::unknown())
1412 AliasResult R = aliasCheck(UnderlyingV1, LocationSize::unknown(),
1413 AAMDNodes(), V2, LocationSize::unknown(),
1414 V2AAInfo, AAQI, nullptr, UnderlyingV2);
1415 if (R != MustAlias) {
1416 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
1417 // If V2 is known not to alias GEP base pointer, then the two values
1418 // cannot alias per GEP semantics: "Any memory access must be done through
1419 // a pointer value associated with an address range of the memory access,
1420 // otherwise the behavior is undefined.".
1421 assert(R == NoAlias || R == MayAlias);
1425 // If the max search depth is reached the result is undefined
1426 if (GEP1MaxLookupReached)
1430 // In the two GEP Case, if there is no difference in the offsets of the
1431 // computed pointers, the resultant pointers are a must alias. This
1432 // happens when we have two lexically identical GEP's (for example).
1434 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
1435 // must aliases the GEP, the end result is a must alias also.
1436 if (GEP1BaseOffset == 0 && DecompGEP1.VarIndices.empty())
1439 // If there is a constant difference between the pointers, but the difference
1440 // is less than the size of the associated memory object, then we know
1441 // that the objects are partially overlapping. If the difference is
1442 // greater, we know they do not overlap.
1443 if (GEP1BaseOffset != 0 && DecompGEP1.VarIndices.empty()) {
1444 if (GEP1BaseOffset.sge(0)) {
1445 if (V2Size != LocationSize::unknown()) {
1446 if (GEP1BaseOffset.ult(V2Size.getValue()))
1447 return PartialAlias;
1451 // We have the situation where:
1454 // ---------------->|
1455 // |-->V1Size |-------> V2Size
1457 // We need to know that V2Size is not unknown, otherwise we might have
1458 // stripped a gep with negative index ('gep <ptr>, -1, ...).
1459 if (V1Size != LocationSize::unknown() &&
1460 V2Size != LocationSize::unknown()) {
1461 if ((-GEP1BaseOffset).ult(V1Size.getValue()))
1462 return PartialAlias;
1468 if (!DecompGEP1.VarIndices.empty()) {
1469 APInt Modulo(MaxPointerSize, 0);
1470 bool AllPositive = true;
1471 for (unsigned i = 0, e = DecompGEP1.VarIndices.size(); i != e; ++i) {
1473 // Try to distinguish something like &A[i][1] against &A[42][0].
1474 // Grab the least significant bit set in any of the scales. We
1475 // don't need std::abs here (even if the scale's negative) as we'll
1476 // be ^'ing Modulo with itself later.
1477 Modulo |= DecompGEP1.VarIndices[i].Scale;
1480 // If the Value could change between cycles, then any reasoning about
1481 // the Value this cycle may not hold in the next cycle. We'll just
1482 // give up if we can't determine conditions that hold for every cycle:
1483 const Value *V = DecompGEP1.VarIndices[i].V;
1486 computeKnownBits(V, DL, 0, &AC, dyn_cast<Instruction>(GEP1), DT);
1487 bool SignKnownZero = Known.isNonNegative();
1488 bool SignKnownOne = Known.isNegative();
1490 // Zero-extension widens the variable, and so forces the sign
1492 bool IsZExt = DecompGEP1.VarIndices[i].ZExtBits > 0 || isa<ZExtInst>(V);
1493 SignKnownZero |= IsZExt;
1494 SignKnownOne &= !IsZExt;
1496 // If the variable begins with a zero then we know it's
1497 // positive, regardless of whether the value is signed or
1499 APInt Scale = DecompGEP1.VarIndices[i].Scale;
1501 (SignKnownZero && Scale.sge(0)) || (SignKnownOne && Scale.slt(0));
1505 Modulo = Modulo ^ (Modulo & (Modulo - 1));
1507 // We can compute the difference between the two addresses
1508 // mod Modulo. Check whether that difference guarantees that the
1509 // two locations do not alias.
1510 APInt ModOffset = GEP1BaseOffset & (Modulo - 1);
1511 if (V1Size != LocationSize::unknown() &&
1512 V2Size != LocationSize::unknown() && ModOffset.uge(V2Size.getValue()) &&
1513 (Modulo - ModOffset).uge(V1Size.getValue()))
1516 // If we know all the variables are positive, then GEP1 >= GEP1BasePtr.
1517 // If GEP1BasePtr > V2 (GEP1BaseOffset > 0) then we know the pointers
1518 // don't alias if V2Size can fit in the gap between V2 and GEP1BasePtr.
1519 if (AllPositive && GEP1BaseOffset.sgt(0) &&
1520 V2Size != LocationSize::unknown() &&
1521 GEP1BaseOffset.uge(V2Size.getValue()))
1524 if (constantOffsetHeuristic(DecompGEP1.VarIndices, V1Size, V2Size,
1525 GEP1BaseOffset, &AC, DT))
1529 // Statically, we can see that the base objects are the same, but the
1530 // pointers have dynamic offsets which we can't resolve. And none of our
1531 // little tricks above worked.
1535 static AliasResult MergeAliasResults(AliasResult A, AliasResult B) {
1536 // If the results agree, take it.
1539 // A mix of PartialAlias and MustAlias is PartialAlias.
1540 if ((A == PartialAlias && B == MustAlias) ||
1541 (B == PartialAlias && A == MustAlias))
1542 return PartialAlias;
1543 // Otherwise, we don't know anything.
1547 /// Provides a bunch of ad-hoc rules to disambiguate a Select instruction
1548 /// against another.
1550 BasicAAResult::aliasSelect(const SelectInst *SI, LocationSize SISize,
1551 const AAMDNodes &SIAAInfo, const Value *V2,
1552 LocationSize V2Size, const AAMDNodes &V2AAInfo,
1553 const Value *UnderV2, AAQueryInfo &AAQI) {
1554 // If the values are Selects with the same condition, we can do a more precise
1555 // check: just check for aliases between the values on corresponding arms.
1556 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1557 if (SI->getCondition() == SI2->getCondition()) {
1559 aliasCheck(SI->getTrueValue(), SISize, SIAAInfo, SI2->getTrueValue(),
1560 V2Size, V2AAInfo, AAQI);
1561 if (Alias == MayAlias)
1563 AliasResult ThisAlias =
1564 aliasCheck(SI->getFalseValue(), SISize, SIAAInfo,
1565 SI2->getFalseValue(), V2Size, V2AAInfo, AAQI);
1566 return MergeAliasResults(ThisAlias, Alias);
1569 // If both arms of the Select node NoAlias or MustAlias V2, then returns
1570 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1571 AliasResult Alias = aliasCheck(V2, V2Size, V2AAInfo, SI->getTrueValue(),
1572 SISize, SIAAInfo, AAQI, UnderV2);
1573 if (Alias == MayAlias)
1576 AliasResult ThisAlias = aliasCheck(V2, V2Size, V2AAInfo, SI->getFalseValue(),
1577 SISize, SIAAInfo, AAQI, UnderV2);
1578 return MergeAliasResults(ThisAlias, Alias);
1581 /// Provide a bunch of ad-hoc rules to disambiguate a PHI instruction against
1583 AliasResult BasicAAResult::aliasPHI(const PHINode *PN, LocationSize PNSize,
1584 const AAMDNodes &PNAAInfo, const Value *V2,
1585 LocationSize V2Size,
1586 const AAMDNodes &V2AAInfo,
1587 const Value *UnderV2, AAQueryInfo &AAQI) {
1588 // Track phi nodes we have visited. We use this information when we determine
1589 // value equivalence.
1590 VisitedPhiBBs.insert(PN->getParent());
1592 // If the values are PHIs in the same block, we can do a more precise
1593 // as well as efficient check: just check for aliases between the values
1594 // on corresponding edges.
1595 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1596 if (PN2->getParent() == PN->getParent()) {
1597 AAQueryInfo::LocPair Locs(MemoryLocation(PN, PNSize, PNAAInfo),
1598 MemoryLocation(V2, V2Size, V2AAInfo));
1600 std::swap(Locs.first, Locs.second);
1601 // Analyse the PHIs' inputs under the assumption that the PHIs are
1603 // If the PHIs are May/MustAlias there must be (recursively) an input
1604 // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
1605 // there must be an operation on the PHIs within the PHIs' value cycle
1606 // that causes a MayAlias.
1607 // Pretend the phis do not alias.
1608 AliasResult Alias = NoAlias;
1609 AliasResult OrigAliasResult;
1611 // Limited lifetime iterator invalidated by the aliasCheck call below.
1612 auto CacheIt = AAQI.AliasCache.find(Locs);
1613 assert((CacheIt != AAQI.AliasCache.end()) &&
1614 "There must exist an entry for the phi node");
1615 OrigAliasResult = CacheIt->second;
1616 CacheIt->second = NoAlias;
1619 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1620 AliasResult ThisAlias =
1621 aliasCheck(PN->getIncomingValue(i), PNSize, PNAAInfo,
1622 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1623 V2Size, V2AAInfo, AAQI);
1624 Alias = MergeAliasResults(ThisAlias, Alias);
1625 if (Alias == MayAlias)
1629 // Reset if speculation failed.
1630 if (Alias != NoAlias) {
1632 AAQI.AliasCache.insert(std::make_pair(Locs, OrigAliasResult));
1633 assert(!Pair.second && "Entry must have existed");
1634 Pair.first->second = OrigAliasResult;
1639 SmallVector<Value *, 4> V1Srcs;
1640 bool isRecursive = false;
1642 // If we have PhiValues then use it to get the underlying phi values.
1643 const PhiValues::ValueSet &PhiValueSet = PV->getValuesForPhi(PN);
1644 // If we have more phi values than the search depth then return MayAlias
1645 // conservatively to avoid compile time explosion. The worst possible case
1646 // is if both sides are PHI nodes. In which case, this is O(m x n) time
1647 // where 'm' and 'n' are the number of PHI sources.
1648 if (PhiValueSet.size() > MaxLookupSearchDepth)
1650 // Add the values to V1Srcs
1651 for (Value *PV1 : PhiValueSet) {
1652 if (EnableRecPhiAnalysis) {
1653 if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) {
1654 // Check whether the incoming value is a GEP that advances the pointer
1655 // result of this PHI node (e.g. in a loop). If this is the case, we
1656 // would recurse and always get a MayAlias. Handle this case specially
1658 if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 &&
1659 isa<ConstantInt>(PV1GEP->idx_begin())) {
1665 V1Srcs.push_back(PV1);
1668 // If we don't have PhiInfo then just look at the operands of the phi itself
1669 // FIXME: Remove this once we can guarantee that we have PhiInfo always
1670 SmallPtrSet<Value *, 4> UniqueSrc;
1671 for (Value *PV1 : PN->incoming_values()) {
1672 if (isa<PHINode>(PV1))
1673 // If any of the source itself is a PHI, return MayAlias conservatively
1674 // to avoid compile time explosion. The worst possible case is if both
1675 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1676 // and 'n' are the number of PHI sources.
1679 if (EnableRecPhiAnalysis)
1680 if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) {
1681 // Check whether the incoming value is a GEP that advances the pointer
1682 // result of this PHI node (e.g. in a loop). If this is the case, we
1683 // would recurse and always get a MayAlias. Handle this case specially
1685 if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 &&
1686 isa<ConstantInt>(PV1GEP->idx_begin())) {
1692 if (UniqueSrc.insert(PV1).second)
1693 V1Srcs.push_back(PV1);
1697 // If V1Srcs is empty then that means that the phi has no underlying non-phi
1698 // value. This should only be possible in blocks unreachable from the entry
1699 // block, but return MayAlias just in case.
1703 // If this PHI node is recursive, set the size of the accessed memory to
1704 // unknown to represent all the possible values the GEP could advance the
1707 PNSize = LocationSize::unknown();
1709 AliasResult Alias = aliasCheck(V2, V2Size, V2AAInfo, V1Srcs[0], PNSize,
1710 PNAAInfo, AAQI, UnderV2);
1712 // Early exit if the check of the first PHI source against V2 is MayAlias.
1713 // Other results are not possible.
1714 if (Alias == MayAlias)
1717 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1718 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1719 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1720 Value *V = V1Srcs[i];
1722 AliasResult ThisAlias =
1723 aliasCheck(V2, V2Size, V2AAInfo, V, PNSize, PNAAInfo, AAQI, UnderV2);
1724 Alias = MergeAliasResults(ThisAlias, Alias);
1725 if (Alias == MayAlias)
1732 /// Provides a bunch of ad-hoc rules to disambiguate in common cases, such as
1733 /// array references.
1734 AliasResult BasicAAResult::aliasCheck(const Value *V1, LocationSize V1Size,
1735 AAMDNodes V1AAInfo, const Value *V2,
1736 LocationSize V2Size, AAMDNodes V2AAInfo,
1737 AAQueryInfo &AAQI, const Value *O1,
1739 // If either of the memory references is empty, it doesn't matter what the
1740 // pointer values are.
1741 if (V1Size.isZero() || V2Size.isZero())
1744 // Strip off any casts if they exist.
1745 V1 = V1->stripPointerCastsAndInvariantGroups();
1746 V2 = V2->stripPointerCastsAndInvariantGroups();
1748 // If V1 or V2 is undef, the result is NoAlias because we can always pick a
1749 // value for undef that aliases nothing in the program.
1750 if (isa<UndefValue>(V1) || isa<UndefValue>(V2))
1753 // Are we checking for alias of the same value?
1754 // Because we look 'through' phi nodes, we could look at "Value" pointers from
1755 // different iterations. We must therefore make sure that this is not the
1756 // case. The function isValueEqualInPotentialCycles ensures that this cannot
1757 // happen by looking at the visited phi nodes and making sure they cannot
1759 if (isValueEqualInPotentialCycles(V1, V2))
1762 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1763 return NoAlias; // Scalars cannot alias each other
1765 // Figure out what objects these things are pointing to if we can.
1767 O1 = GetUnderlyingObject(V1, DL, MaxLookupSearchDepth);
1770 O2 = GetUnderlyingObject(V2, DL, MaxLookupSearchDepth);
1772 // Null values in the default address space don't point to any object, so they
1773 // don't alias any other pointer.
1774 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1775 if (!NullPointerIsDefined(&F, CPN->getType()->getAddressSpace()))
1777 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1778 if (!NullPointerIsDefined(&F, CPN->getType()->getAddressSpace()))
1782 // If V1/V2 point to two different objects, we know that we have no alias.
1783 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1786 // Constant pointers can't alias with non-const isIdentifiedObject objects.
1787 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1788 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1791 // Function arguments can't alias with things that are known to be
1792 // unambigously identified at the function level.
1793 if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1794 (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1797 // If one pointer is the result of a call/invoke or load and the other is a
1798 // non-escaping local object within the same function, then we know the
1799 // object couldn't escape to a point where the call could return it.
1801 // Note that if the pointers are in different functions, there are a
1802 // variety of complications. A call with a nocapture argument may still
1803 // temporary store the nocapture argument's value in a temporary memory
1804 // location if that memory location doesn't escape. Or it may pass a
1805 // nocapture value to other functions as long as they don't capture it.
1806 if (isEscapeSource(O1) &&
1807 isNonEscapingLocalObject(O2, &AAQI.IsCapturedCache))
1809 if (isEscapeSource(O2) &&
1810 isNonEscapingLocalObject(O1, &AAQI.IsCapturedCache))
1814 // If the size of one access is larger than the entire object on the other
1815 // side, then we know such behavior is undefined and can assume no alias.
1816 bool NullIsValidLocation = NullPointerIsDefined(&F);
1817 if ((isObjectSmallerThan(
1818 O2, getMinimalExtentFrom(*V1, V1Size, DL, NullIsValidLocation), DL,
1819 TLI, NullIsValidLocation)) ||
1820 (isObjectSmallerThan(
1821 O1, getMinimalExtentFrom(*V2, V2Size, DL, NullIsValidLocation), DL,
1822 TLI, NullIsValidLocation)))
1825 // Check the cache before climbing up use-def chains. This also terminates
1826 // otherwise infinitely recursive queries.
1827 AAQueryInfo::LocPair Locs(MemoryLocation(V1, V1Size, V1AAInfo),
1828 MemoryLocation(V2, V2Size, V2AAInfo));
1830 std::swap(Locs.first, Locs.second);
1831 std::pair<AAQueryInfo::AliasCacheT::iterator, bool> Pair =
1832 AAQI.AliasCache.try_emplace(Locs, MayAlias);
1834 return Pair.first->second;
1836 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1837 // GEP can't simplify, we don't even look at the PHI cases.
1838 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1840 std::swap(V1Size, V2Size);
1842 std::swap(V1AAInfo, V2AAInfo);
1844 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1845 AliasResult Result =
1846 aliasGEP(GV1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O1, O2, AAQI);
1847 if (Result != MayAlias) {
1848 auto ItInsPair = AAQI.AliasCache.insert(std::make_pair(Locs, Result));
1849 assert(!ItInsPair.second && "Entry must have existed");
1850 ItInsPair.first->second = Result;
1855 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1858 std::swap(V1Size, V2Size);
1859 std::swap(V1AAInfo, V2AAInfo);
1861 if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1862 AliasResult Result =
1863 aliasPHI(PN, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O2, AAQI);
1864 if (Result != MayAlias) {
1865 Pair = AAQI.AliasCache.try_emplace(Locs, Result);
1866 assert(!Pair.second && "Entry must have existed");
1867 return Pair.first->second = Result;
1871 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1874 std::swap(V1Size, V2Size);
1875 std::swap(V1AAInfo, V2AAInfo);
1877 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1878 AliasResult Result =
1879 aliasSelect(S1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O2, AAQI);
1880 if (Result != MayAlias) {
1881 Pair = AAQI.AliasCache.try_emplace(Locs, Result);
1882 assert(!Pair.second && "Entry must have existed");
1883 return Pair.first->second = Result;
1887 // If both pointers are pointing into the same object and one of them
1888 // accesses the entire object, then the accesses must overlap in some way.
1890 if (V1Size.isPrecise() && V2Size.isPrecise() &&
1891 (isObjectSize(O1, V1Size.getValue(), DL, TLI, NullIsValidLocation) ||
1892 isObjectSize(O2, V2Size.getValue(), DL, TLI, NullIsValidLocation))) {
1893 Pair = AAQI.AliasCache.try_emplace(Locs, PartialAlias);
1894 assert(!Pair.second && "Entry must have existed");
1895 return Pair.first->second = PartialAlias;
1898 // Recurse back into the best AA results we have, potentially with refined
1899 // memory locations. We have already ensured that BasicAA has a MayAlias
1900 // cache result for these, so any recursion back into BasicAA won't loop.
1901 AliasResult Result = getBestAAResults().alias(Locs.first, Locs.second, AAQI);
1902 Pair = AAQI.AliasCache.try_emplace(Locs, Result);
1903 assert(!Pair.second && "Entry must have existed");
1904 return Pair.first->second = Result;
1907 /// Check whether two Values can be considered equivalent.
1909 /// In addition to pointer equivalence of \p V1 and \p V2 this checks whether
1910 /// they can not be part of a cycle in the value graph by looking at all
1911 /// visited phi nodes an making sure that the phis cannot reach the value. We
1912 /// have to do this because we are looking through phi nodes (That is we say
1913 /// noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB).
1914 bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V,
1919 const Instruction *Inst = dyn_cast<Instruction>(V);
1923 if (VisitedPhiBBs.empty())
1926 if (VisitedPhiBBs.size() > MaxNumPhiBBsValueReachabilityCheck)
1929 // Make sure that the visited phis cannot reach the Value. This ensures that
1930 // the Values cannot come from different iterations of a potential cycle the
1931 // phi nodes could be involved in.
1932 for (auto *P : VisitedPhiBBs)
1933 if (isPotentiallyReachable(&P->front(), Inst, nullptr, DT, LI))
1939 /// Computes the symbolic difference between two de-composed GEPs.
1941 /// Dest and Src are the variable indices from two decomposed GetElementPtr
1942 /// instructions GEP1 and GEP2 which have common base pointers.
1943 void BasicAAResult::GetIndexDifference(
1944 SmallVectorImpl<VariableGEPIndex> &Dest,
1945 const SmallVectorImpl<VariableGEPIndex> &Src) {
1949 for (unsigned i = 0, e = Src.size(); i != e; ++i) {
1950 const Value *V = Src[i].V;
1951 unsigned ZExtBits = Src[i].ZExtBits, SExtBits = Src[i].SExtBits;
1952 APInt Scale = Src[i].Scale;
1954 // Find V in Dest. This is N^2, but pointer indices almost never have more
1955 // than a few variable indexes.
1956 for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
1957 if (!isValueEqualInPotentialCycles(Dest[j].V, V) ||
1958 Dest[j].ZExtBits != ZExtBits || Dest[j].SExtBits != SExtBits)
1961 // If we found it, subtract off Scale V's from the entry in Dest. If it
1962 // goes to zero, remove the entry.
1963 if (Dest[j].Scale != Scale)
1964 Dest[j].Scale -= Scale;
1966 Dest.erase(Dest.begin() + j);
1971 // If we didn't consume this entry, add it to the end of the Dest list.
1973 VariableGEPIndex Entry = {V, ZExtBits, SExtBits, -Scale};
1974 Dest.push_back(Entry);
1979 bool BasicAAResult::constantOffsetHeuristic(
1980 const SmallVectorImpl<VariableGEPIndex> &VarIndices,
1981 LocationSize MaybeV1Size, LocationSize MaybeV2Size, APInt BaseOffset,
1982 AssumptionCache *AC, DominatorTree *DT) {
1983 if (VarIndices.size() != 2 || MaybeV1Size == LocationSize::unknown() ||
1984 MaybeV2Size == LocationSize::unknown())
1987 const uint64_t V1Size = MaybeV1Size.getValue();
1988 const uint64_t V2Size = MaybeV2Size.getValue();
1990 const VariableGEPIndex &Var0 = VarIndices[0], &Var1 = VarIndices[1];
1992 if (Var0.ZExtBits != Var1.ZExtBits || Var0.SExtBits != Var1.SExtBits ||
1993 Var0.Scale != -Var1.Scale)
1996 unsigned Width = Var1.V->getType()->getIntegerBitWidth();
1998 // We'll strip off the Extensions of Var0 and Var1 and do another round
1999 // of GetLinearExpression decomposition. In the example above, if Var0
2000 // is zext(%x + 1) we should get V1 == %x and V1Offset == 1.
2002 APInt V0Scale(Width, 0), V0Offset(Width, 0), V1Scale(Width, 0),
2004 bool NSW = true, NUW = true;
2005 unsigned V0ZExtBits = 0, V0SExtBits = 0, V1ZExtBits = 0, V1SExtBits = 0;
2006 const Value *V0 = GetLinearExpression(Var0.V, V0Scale, V0Offset, V0ZExtBits,
2007 V0SExtBits, DL, 0, AC, DT, NSW, NUW);
2010 const Value *V1 = GetLinearExpression(Var1.V, V1Scale, V1Offset, V1ZExtBits,
2011 V1SExtBits, DL, 0, AC, DT, NSW, NUW);
2013 if (V0Scale != V1Scale || V0ZExtBits != V1ZExtBits ||
2014 V0SExtBits != V1SExtBits || !isValueEqualInPotentialCycles(V0, V1))
2017 // We have a hit - Var0 and Var1 only differ by a constant offset!
2019 // If we've been sext'ed then zext'd the maximum difference between Var0 and
2020 // Var1 is possible to calculate, but we're just interested in the absolute
2021 // minimum difference between the two. The minimum distance may occur due to
2022 // wrapping; consider "add i3 %i, 5": if %i == 7 then 7 + 5 mod 8 == 4, and so
2023 // the minimum distance between %i and %i + 5 is 3.
2024 APInt MinDiff = V0Offset - V1Offset, Wrapped = -MinDiff;
2025 MinDiff = APIntOps::umin(MinDiff, Wrapped);
2026 APInt MinDiffBytes =
2027 MinDiff.zextOrTrunc(Var0.Scale.getBitWidth()) * Var0.Scale.abs();
2029 // We can't definitely say whether GEP1 is before or after V2 due to wrapping
2030 // arithmetic (i.e. for some values of GEP1 and V2 GEP1 < V2, and for other
2031 // values GEP1 > V2). We'll therefore only declare NoAlias if both V1Size and
2032 // V2Size can fit in the MinDiffBytes gap.
2033 return MinDiffBytes.uge(V1Size + BaseOffset.abs()) &&
2034 MinDiffBytes.uge(V2Size + BaseOffset.abs());
2037 //===----------------------------------------------------------------------===//
2038 // BasicAliasAnalysis Pass
2039 //===----------------------------------------------------------------------===//
2041 AnalysisKey BasicAA::Key;
2043 BasicAAResult BasicAA::run(Function &F, FunctionAnalysisManager &AM) {
2044 return BasicAAResult(F.getParent()->getDataLayout(),
2046 AM.getResult<TargetLibraryAnalysis>(F),
2047 AM.getResult<AssumptionAnalysis>(F),
2048 &AM.getResult<DominatorTreeAnalysis>(F),
2049 AM.getCachedResult<LoopAnalysis>(F),
2050 AM.getCachedResult<PhiValuesAnalysis>(F));
2053 BasicAAWrapperPass::BasicAAWrapperPass() : FunctionPass(ID) {
2054 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
2057 char BasicAAWrapperPass::ID = 0;
2059 void BasicAAWrapperPass::anchor() {}
2061 INITIALIZE_PASS_BEGIN(BasicAAWrapperPass, "basicaa",
2062 "Basic Alias Analysis (stateless AA impl)", false, true)
2063 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2064 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2065 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
2066 INITIALIZE_PASS_END(BasicAAWrapperPass, "basicaa",
2067 "Basic Alias Analysis (stateless AA impl)", false, true)
2069 FunctionPass *llvm::createBasicAAWrapperPass() {
2070 return new BasicAAWrapperPass();
2073 bool BasicAAWrapperPass::runOnFunction(Function &F) {
2074 auto &ACT = getAnalysis<AssumptionCacheTracker>();
2075 auto &TLIWP = getAnalysis<TargetLibraryInfoWrapperPass>();
2076 auto &DTWP = getAnalysis<DominatorTreeWrapperPass>();
2077 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2078 auto *PVWP = getAnalysisIfAvailable<PhiValuesWrapperPass>();
2080 Result.reset(new BasicAAResult(F.getParent()->getDataLayout(), F,
2081 TLIWP.getTLI(F), ACT.getAssumptionCache(F),
2083 LIWP ? &LIWP->getLoopInfo() : nullptr,
2084 PVWP ? &PVWP->getResult() : nullptr));
2089 void BasicAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
2090 AU.setPreservesAll();
2091 AU.addRequired<AssumptionCacheTracker>();
2092 AU.addRequired<DominatorTreeWrapperPass>();
2093 AU.addRequired<TargetLibraryInfoWrapperPass>();
2094 AU.addUsedIfAvailable<PhiValuesWrapperPass>();
2097 BasicAAResult llvm::createLegacyPMBasicAAResult(Pass &P, Function &F) {
2098 return BasicAAResult(
2099 F.getParent()->getDataLayout(), F,
2100 P.getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F),
2101 P.getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));