xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/BasicAliasAnalysis.cpp (revision 9f23cbd6cae82fd77edfad7173432fa8dccd0a95)
1 //===- BasicAliasAnalysis.cpp - Stateless Alias Analysis Impl -------------===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
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.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Analysis/BasicAliasAnalysis.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/ScopeExit.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/AssumptionCache.h"
23 #include "llvm/Analysis/CFG.h"
24 #include "llvm/Analysis/CaptureTracking.h"
25 #include "llvm/Analysis/MemoryBuiltins.h"
26 #include "llvm/Analysis/MemoryLocation.h"
27 #include "llvm/Analysis/TargetLibraryInfo.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/IR/Argument.h"
30 #include "llvm/IR/Attributes.h"
31 #include "llvm/IR/Constant.h"
32 #include "llvm/IR/ConstantRange.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/Dominators.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/GetElementPtrTypeIterator.h"
39 #include "llvm/IR/GlobalAlias.h"
40 #include "llvm/IR/GlobalVariable.h"
41 #include "llvm/IR/InstrTypes.h"
42 #include "llvm/IR/Instruction.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/Intrinsics.h"
46 #include "llvm/IR/Operator.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/IR/User.h"
49 #include "llvm/IR/Value.h"
50 #include "llvm/InitializePasses.h"
51 #include "llvm/Pass.h"
52 #include "llvm/Support/Casting.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/Support/KnownBits.h"
56 #include "llvm/Support/SaveAndRestore.h"
57 #include <cassert>
58 #include <cstdint>
59 #include <cstdlib>
60 #include <optional>
61 #include <utility>
62 
63 #define DEBUG_TYPE "basicaa"
64 
65 using namespace llvm;
66 
67 /// Enable analysis of recursive PHI nodes.
68 static cl::opt<bool> EnableRecPhiAnalysis("basic-aa-recphi", cl::Hidden,
69                                           cl::init(true));
70 
71 static cl::opt<bool> EnableSeparateStorageAnalysis("basic-aa-separate-storage",
72                                                    cl::Hidden, cl::init(false));
73 
74 /// SearchLimitReached / SearchTimes shows how often the limit of
75 /// to decompose GEPs is reached. It will affect the precision
76 /// of basic alias analysis.
77 STATISTIC(SearchLimitReached, "Number of times the limit to "
78                               "decompose GEPs is reached");
79 STATISTIC(SearchTimes, "Number of times a GEP is decomposed");
80 
81 // The max limit of the search depth in DecomposeGEPExpression() and
82 // getUnderlyingObject().
83 static const unsigned MaxLookupSearchDepth = 6;
84 
85 bool BasicAAResult::invalidate(Function &Fn, const PreservedAnalyses &PA,
86                                FunctionAnalysisManager::Invalidator &Inv) {
87   // We don't care if this analysis itself is preserved, it has no state. But
88   // we need to check that the analyses it depends on have been. Note that we
89   // may be created without handles to some analyses and in that case don't
90   // depend on them.
91   if (Inv.invalidate<AssumptionAnalysis>(Fn, PA) ||
92       (DT && Inv.invalidate<DominatorTreeAnalysis>(Fn, PA)))
93     return true;
94 
95   // Otherwise this analysis result remains valid.
96   return false;
97 }
98 
99 //===----------------------------------------------------------------------===//
100 // Useful predicates
101 //===----------------------------------------------------------------------===//
102 
103 /// Returns the size of the object specified by V or UnknownSize if unknown.
104 static uint64_t getObjectSize(const Value *V, const DataLayout &DL,
105                               const TargetLibraryInfo &TLI,
106                               bool NullIsValidLoc,
107                               bool RoundToAlign = false) {
108   uint64_t Size;
109   ObjectSizeOpts Opts;
110   Opts.RoundToAlign = RoundToAlign;
111   Opts.NullIsUnknownSize = NullIsValidLoc;
112   if (getObjectSize(V, Size, DL, &TLI, Opts))
113     return Size;
114   return MemoryLocation::UnknownSize;
115 }
116 
117 /// Returns true if we can prove that the object specified by V is smaller than
118 /// Size.
119 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
120                                 const DataLayout &DL,
121                                 const TargetLibraryInfo &TLI,
122                                 bool NullIsValidLoc) {
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".
130   //
131   //  Consider this example:
132   //     char *p = (char*)malloc(100)
133   //     char *q = p+80;
134   //
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.
137   //
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.
147   //
148   // We go for 2nd option for simplicity.
149   if (!isIdentifiedObject(V))
150     return false;
151 
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, NullIsValidLoc,
155                                       /*RoundToAlign*/ true);
156 
157   return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size;
158 }
159 
160 /// Return the minimal extent from \p V to the end of the underlying object,
161 /// assuming the result is used in an aliasing query. E.g., we do use the query
162 /// location size and the fact that null pointers cannot alias here.
163 static uint64_t getMinimalExtentFrom(const Value &V,
164                                      const LocationSize &LocSize,
165                                      const DataLayout &DL,
166                                      bool NullIsValidLoc) {
167   // If we have dereferenceability information we know a lower bound for the
168   // extent as accesses for a lower offset would be valid. We need to exclude
169   // the "or null" part if null is a valid pointer. We can ignore frees, as an
170   // access after free would be undefined behavior.
171   bool CanBeNull, CanBeFreed;
172   uint64_t DerefBytes =
173     V.getPointerDereferenceableBytes(DL, CanBeNull, CanBeFreed);
174   DerefBytes = (CanBeNull && NullIsValidLoc) ? 0 : DerefBytes;
175   // If queried with a precise location size, we assume that location size to be
176   // accessed, thus valid.
177   if (LocSize.isPrecise())
178     DerefBytes = std::max(DerefBytes, LocSize.getValue());
179   return DerefBytes;
180 }
181 
182 /// Returns true if we can prove that the object specified by V has size Size.
183 static bool isObjectSize(const Value *V, uint64_t Size, const DataLayout &DL,
184                          const TargetLibraryInfo &TLI, bool NullIsValidLoc) {
185   uint64_t ObjectSize = getObjectSize(V, DL, TLI, NullIsValidLoc);
186   return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size;
187 }
188 
189 //===----------------------------------------------------------------------===//
190 // CaptureInfo implementations
191 //===----------------------------------------------------------------------===//
192 
193 CaptureInfo::~CaptureInfo() = default;
194 
195 bool SimpleCaptureInfo::isNotCapturedBeforeOrAt(const Value *Object,
196                                                 const Instruction *I) {
197   return isNonEscapingLocalObject(Object, &IsCapturedCache);
198 }
199 
200 bool EarliestEscapeInfo::isNotCapturedBeforeOrAt(const Value *Object,
201                                                  const Instruction *I) {
202   if (!isIdentifiedFunctionLocal(Object))
203     return false;
204 
205   auto Iter = EarliestEscapes.insert({Object, nullptr});
206   if (Iter.second) {
207     Instruction *EarliestCapture = FindEarliestCapture(
208         Object, *const_cast<Function *>(I->getFunction()),
209         /*ReturnCaptures=*/false, /*StoreCaptures=*/true, DT, EphValues);
210     if (EarliestCapture) {
211       auto Ins = Inst2Obj.insert({EarliestCapture, {}});
212       Ins.first->second.push_back(Object);
213     }
214     Iter.first->second = EarliestCapture;
215   }
216 
217   // No capturing instruction.
218   if (!Iter.first->second)
219     return true;
220 
221   return I != Iter.first->second &&
222          !isPotentiallyReachable(Iter.first->second, I, nullptr, &DT, &LI);
223 }
224 
225 void EarliestEscapeInfo::removeInstruction(Instruction *I) {
226   auto Iter = Inst2Obj.find(I);
227   if (Iter != Inst2Obj.end()) {
228     for (const Value *Obj : Iter->second)
229       EarliestEscapes.erase(Obj);
230     Inst2Obj.erase(I);
231   }
232 }
233 
234 //===----------------------------------------------------------------------===//
235 // GetElementPtr Instruction Decomposition and Analysis
236 //===----------------------------------------------------------------------===//
237 
238 namespace {
239 /// Represents zext(sext(trunc(V))).
240 struct CastedValue {
241   const Value *V;
242   unsigned ZExtBits = 0;
243   unsigned SExtBits = 0;
244   unsigned TruncBits = 0;
245 
246   explicit CastedValue(const Value *V) : V(V) {}
247   explicit CastedValue(const Value *V, unsigned ZExtBits, unsigned SExtBits,
248                        unsigned TruncBits)
249       : V(V), ZExtBits(ZExtBits), SExtBits(SExtBits), TruncBits(TruncBits) {}
250 
251   unsigned getBitWidth() const {
252     return V->getType()->getPrimitiveSizeInBits() - TruncBits + ZExtBits +
253            SExtBits;
254   }
255 
256   CastedValue withValue(const Value *NewV) const {
257     return CastedValue(NewV, ZExtBits, SExtBits, TruncBits);
258   }
259 
260   /// Replace V with zext(NewV)
261   CastedValue withZExtOfValue(const Value *NewV) const {
262     unsigned ExtendBy = V->getType()->getPrimitiveSizeInBits() -
263                         NewV->getType()->getPrimitiveSizeInBits();
264     if (ExtendBy <= TruncBits)
265       return CastedValue(NewV, ZExtBits, SExtBits, TruncBits - ExtendBy);
266 
267     // zext(sext(zext(NewV))) == zext(zext(zext(NewV)))
268     ExtendBy -= TruncBits;
269     return CastedValue(NewV, ZExtBits + SExtBits + ExtendBy, 0, 0);
270   }
271 
272   /// Replace V with sext(NewV)
273   CastedValue withSExtOfValue(const Value *NewV) const {
274     unsigned ExtendBy = V->getType()->getPrimitiveSizeInBits() -
275                         NewV->getType()->getPrimitiveSizeInBits();
276     if (ExtendBy <= TruncBits)
277       return CastedValue(NewV, ZExtBits, SExtBits, TruncBits - ExtendBy);
278 
279     // zext(sext(sext(NewV)))
280     ExtendBy -= TruncBits;
281     return CastedValue(NewV, ZExtBits, SExtBits + ExtendBy, 0);
282   }
283 
284   APInt evaluateWith(APInt N) const {
285     assert(N.getBitWidth() == V->getType()->getPrimitiveSizeInBits() &&
286            "Incompatible bit width");
287     if (TruncBits) N = N.trunc(N.getBitWidth() - TruncBits);
288     if (SExtBits) N = N.sext(N.getBitWidth() + SExtBits);
289     if (ZExtBits) N = N.zext(N.getBitWidth() + ZExtBits);
290     return N;
291   }
292 
293   ConstantRange evaluateWith(ConstantRange N) const {
294     assert(N.getBitWidth() == V->getType()->getPrimitiveSizeInBits() &&
295            "Incompatible bit width");
296     if (TruncBits) N = N.truncate(N.getBitWidth() - TruncBits);
297     if (SExtBits) N = N.signExtend(N.getBitWidth() + SExtBits);
298     if (ZExtBits) N = N.zeroExtend(N.getBitWidth() + ZExtBits);
299     return N;
300   }
301 
302   bool canDistributeOver(bool NUW, bool NSW) const {
303     // zext(x op<nuw> y) == zext(x) op<nuw> zext(y)
304     // sext(x op<nsw> y) == sext(x) op<nsw> sext(y)
305     // trunc(x op y) == trunc(x) op trunc(y)
306     return (!ZExtBits || NUW) && (!SExtBits || NSW);
307   }
308 
309   bool hasSameCastsAs(const CastedValue &Other) const {
310     return ZExtBits == Other.ZExtBits && SExtBits == Other.SExtBits &&
311            TruncBits == Other.TruncBits;
312   }
313 };
314 
315 /// Represents zext(sext(trunc(V))) * Scale + Offset.
316 struct LinearExpression {
317   CastedValue Val;
318   APInt Scale;
319   APInt Offset;
320 
321   /// True if all operations in this expression are NSW.
322   bool IsNSW;
323 
324   LinearExpression(const CastedValue &Val, const APInt &Scale,
325                    const APInt &Offset, bool IsNSW)
326       : Val(Val), Scale(Scale), Offset(Offset), IsNSW(IsNSW) {}
327 
328   LinearExpression(const CastedValue &Val) : Val(Val), IsNSW(true) {
329     unsigned BitWidth = Val.getBitWidth();
330     Scale = APInt(BitWidth, 1);
331     Offset = APInt(BitWidth, 0);
332   }
333 
334   LinearExpression mul(const APInt &Other, bool MulIsNSW) const {
335     // The check for zero offset is necessary, because generally
336     // (X +nsw Y) *nsw Z does not imply (X *nsw Z) +nsw (Y *nsw Z).
337     bool NSW = IsNSW && (Other.isOne() || (MulIsNSW && Offset.isZero()));
338     return LinearExpression(Val, Scale * Other, Offset * Other, NSW);
339   }
340 };
341 }
342 
343 /// Analyzes the specified value as a linear expression: "A*V + B", where A and
344 /// B are constant integers.
345 static LinearExpression GetLinearExpression(
346     const CastedValue &Val,  const DataLayout &DL, unsigned Depth,
347     AssumptionCache *AC, DominatorTree *DT) {
348   // Limit our recursion depth.
349   if (Depth == 6)
350     return Val;
351 
352   if (const ConstantInt *Const = dyn_cast<ConstantInt>(Val.V))
353     return LinearExpression(Val, APInt(Val.getBitWidth(), 0),
354                             Val.evaluateWith(Const->getValue()), true);
355 
356   if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(Val.V)) {
357     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
358       APInt RHS = Val.evaluateWith(RHSC->getValue());
359       // The only non-OBO case we deal with is or, and only limited to the
360       // case where it is both nuw and nsw.
361       bool NUW = true, NSW = true;
362       if (isa<OverflowingBinaryOperator>(BOp)) {
363         NUW &= BOp->hasNoUnsignedWrap();
364         NSW &= BOp->hasNoSignedWrap();
365       }
366       if (!Val.canDistributeOver(NUW, NSW))
367         return Val;
368 
369       // While we can distribute over trunc, we cannot preserve nowrap flags
370       // in that case.
371       if (Val.TruncBits)
372         NUW = NSW = false;
373 
374       LinearExpression E(Val);
375       switch (BOp->getOpcode()) {
376       default:
377         // We don't understand this instruction, so we can't decompose it any
378         // further.
379         return Val;
380       case Instruction::Or:
381         // X|C == X+C if all the bits in C are unset in X.  Otherwise we can't
382         // analyze it.
383         if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC,
384                                BOp, DT))
385           return Val;
386 
387         [[fallthrough]];
388       case Instruction::Add: {
389         E = GetLinearExpression(Val.withValue(BOp->getOperand(0)), DL,
390                                 Depth + 1, AC, DT);
391         E.Offset += RHS;
392         E.IsNSW &= NSW;
393         break;
394       }
395       case Instruction::Sub: {
396         E = GetLinearExpression(Val.withValue(BOp->getOperand(0)), DL,
397                                 Depth + 1, AC, DT);
398         E.Offset -= RHS;
399         E.IsNSW &= NSW;
400         break;
401       }
402       case Instruction::Mul:
403         E = GetLinearExpression(Val.withValue(BOp->getOperand(0)), DL,
404                                 Depth + 1, AC, DT)
405                 .mul(RHS, NSW);
406         break;
407       case Instruction::Shl:
408         // We're trying to linearize an expression of the kind:
409         //   shl i8 -128, 36
410         // where the shift count exceeds the bitwidth of the type.
411         // We can't decompose this further (the expression would return
412         // a poison value).
413         if (RHS.getLimitedValue() > Val.getBitWidth())
414           return Val;
415 
416         E = GetLinearExpression(Val.withValue(BOp->getOperand(0)), DL,
417                                 Depth + 1, AC, DT);
418         E.Offset <<= RHS.getLimitedValue();
419         E.Scale <<= RHS.getLimitedValue();
420         E.IsNSW &= NSW;
421         break;
422       }
423       return E;
424     }
425   }
426 
427   if (isa<ZExtInst>(Val.V))
428     return GetLinearExpression(
429         Val.withZExtOfValue(cast<CastInst>(Val.V)->getOperand(0)),
430         DL, Depth + 1, AC, DT);
431 
432   if (isa<SExtInst>(Val.V))
433     return GetLinearExpression(
434         Val.withSExtOfValue(cast<CastInst>(Val.V)->getOperand(0)),
435         DL, Depth + 1, AC, DT);
436 
437   return Val;
438 }
439 
440 /// To ensure a pointer offset fits in an integer of size IndexSize
441 /// (in bits) when that size is smaller than the maximum index size. This is
442 /// an issue, for example, in particular for 32b pointers with negative indices
443 /// that rely on two's complement wrap-arounds for precise alias information
444 /// where the maximum index size is 64b.
445 static APInt adjustToIndexSize(const APInt &Offset, unsigned IndexSize) {
446   assert(IndexSize <= Offset.getBitWidth() && "Invalid IndexSize!");
447   unsigned ShiftBits = Offset.getBitWidth() - IndexSize;
448   return (Offset << ShiftBits).ashr(ShiftBits);
449 }
450 
451 namespace {
452 // A linear transformation of a Value; this class represents
453 // ZExt(SExt(Trunc(V, TruncBits), SExtBits), ZExtBits) * Scale.
454 struct VariableGEPIndex {
455   CastedValue Val;
456   APInt Scale;
457 
458   // Context instruction to use when querying information about this index.
459   const Instruction *CxtI;
460 
461   /// True if all operations in this expression are NSW.
462   bool IsNSW;
463 
464   void dump() const {
465     print(dbgs());
466     dbgs() << "\n";
467   }
468   void print(raw_ostream &OS) const {
469     OS << "(V=" << Val.V->getName()
470        << ", zextbits=" << Val.ZExtBits
471        << ", sextbits=" << Val.SExtBits
472        << ", truncbits=" << Val.TruncBits
473        << ", scale=" << Scale << ")";
474   }
475 };
476 }
477 
478 // Represents the internal structure of a GEP, decomposed into a base pointer,
479 // constant offsets, and variable scaled indices.
480 struct BasicAAResult::DecomposedGEP {
481   // Base pointer of the GEP
482   const Value *Base;
483   // Total constant offset from base.
484   APInt Offset;
485   // Scaled variable (non-constant) indices.
486   SmallVector<VariableGEPIndex, 4> VarIndices;
487   // Are all operations inbounds GEPs or non-indexing operations?
488   // (std::nullopt iff expression doesn't involve any geps)
489   std::optional<bool> InBounds;
490 
491   void dump() const {
492     print(dbgs());
493     dbgs() << "\n";
494   }
495   void print(raw_ostream &OS) const {
496     OS << "(DecomposedGEP Base=" << Base->getName()
497        << ", Offset=" << Offset
498        << ", VarIndices=[";
499     for (size_t i = 0; i < VarIndices.size(); i++) {
500       if (i != 0)
501         OS << ", ";
502       VarIndices[i].print(OS);
503     }
504     OS << "])";
505   }
506 };
507 
508 
509 /// If V is a symbolic pointer expression, decompose it into a base pointer
510 /// with a constant offset and a number of scaled symbolic offsets.
511 ///
512 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale
513 /// in the VarIndices vector) are Value*'s that are known to be scaled by the
514 /// specified amount, but which may have other unrepresented high bits. As
515 /// such, the gep cannot necessarily be reconstructed from its decomposed form.
516 BasicAAResult::DecomposedGEP
517 BasicAAResult::DecomposeGEPExpression(const Value *V, const DataLayout &DL,
518                                       AssumptionCache *AC, DominatorTree *DT) {
519   // Limit recursion depth to limit compile time in crazy cases.
520   unsigned MaxLookup = MaxLookupSearchDepth;
521   SearchTimes++;
522   const Instruction *CxtI = dyn_cast<Instruction>(V);
523 
524   unsigned MaxIndexSize = DL.getMaxIndexSizeInBits();
525   DecomposedGEP Decomposed;
526   Decomposed.Offset = APInt(MaxIndexSize, 0);
527   do {
528     // See if this is a bitcast or GEP.
529     const Operator *Op = dyn_cast<Operator>(V);
530     if (!Op) {
531       // The only non-operator case we can handle are GlobalAliases.
532       if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
533         if (!GA->isInterposable()) {
534           V = GA->getAliasee();
535           continue;
536         }
537       }
538       Decomposed.Base = V;
539       return Decomposed;
540     }
541 
542     if (Op->getOpcode() == Instruction::BitCast ||
543         Op->getOpcode() == Instruction::AddrSpaceCast) {
544       V = Op->getOperand(0);
545       continue;
546     }
547 
548     const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
549     if (!GEPOp) {
550       if (const auto *PHI = dyn_cast<PHINode>(V)) {
551         // Look through single-arg phi nodes created by LCSSA.
552         if (PHI->getNumIncomingValues() == 1) {
553           V = PHI->getIncomingValue(0);
554           continue;
555         }
556       } else if (const auto *Call = dyn_cast<CallBase>(V)) {
557         // CaptureTracking can know about special capturing properties of some
558         // intrinsics like launder.invariant.group, that can't be expressed with
559         // the attributes, but have properties like returning aliasing pointer.
560         // Because some analysis may assume that nocaptured pointer is not
561         // returned from some special intrinsic (because function would have to
562         // be marked with returns attribute), it is crucial to use this function
563         // because it should be in sync with CaptureTracking. Not using it may
564         // cause weird miscompilations where 2 aliasing pointers are assumed to
565         // noalias.
566         if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) {
567           V = RP;
568           continue;
569         }
570       }
571 
572       Decomposed.Base = V;
573       return Decomposed;
574     }
575 
576     // Track whether we've seen at least one in bounds gep, and if so, whether
577     // all geps parsed were in bounds.
578     if (Decomposed.InBounds == std::nullopt)
579       Decomposed.InBounds = GEPOp->isInBounds();
580     else if (!GEPOp->isInBounds())
581       Decomposed.InBounds = false;
582 
583     assert(GEPOp->getSourceElementType()->isSized() && "GEP must be sized");
584 
585     unsigned AS = GEPOp->getPointerAddressSpace();
586     // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
587     gep_type_iterator GTI = gep_type_begin(GEPOp);
588     unsigned IndexSize = DL.getIndexSizeInBits(AS);
589     // Assume all GEP operands are constants until proven otherwise.
590     bool GepHasConstantOffset = true;
591     for (User::const_op_iterator I = GEPOp->op_begin() + 1, E = GEPOp->op_end();
592          I != E; ++I, ++GTI) {
593       const Value *Index = *I;
594       // Compute the (potentially symbolic) offset in bytes for this index.
595       if (StructType *STy = GTI.getStructTypeOrNull()) {
596         // For a struct, add the member offset.
597         unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
598         if (FieldNo == 0)
599           continue;
600 
601         Decomposed.Offset += DL.getStructLayout(STy)->getElementOffset(FieldNo);
602         continue;
603       }
604 
605       // For an array/pointer, add the element offset, explicitly scaled.
606       if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
607         if (CIdx->isZero())
608           continue;
609 
610         // Don't attempt to analyze GEPs if the scalable index is not zero.
611         TypeSize AllocTypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
612         if (AllocTypeSize.isScalable()) {
613           Decomposed.Base = V;
614           return Decomposed;
615         }
616 
617         Decomposed.Offset += AllocTypeSize.getFixedValue() *
618                              CIdx->getValue().sextOrTrunc(MaxIndexSize);
619         continue;
620       }
621 
622       TypeSize AllocTypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
623       if (AllocTypeSize.isScalable()) {
624         Decomposed.Base = V;
625         return Decomposed;
626       }
627 
628       GepHasConstantOffset = false;
629 
630       // If the integer type is smaller than the index size, it is implicitly
631       // sign extended or truncated to index size.
632       unsigned Width = Index->getType()->getIntegerBitWidth();
633       unsigned SExtBits = IndexSize > Width ? IndexSize - Width : 0;
634       unsigned TruncBits = IndexSize < Width ? Width - IndexSize : 0;
635       LinearExpression LE = GetLinearExpression(
636           CastedValue(Index, 0, SExtBits, TruncBits), DL, 0, AC, DT);
637 
638       // Scale by the type size.
639       unsigned TypeSize = AllocTypeSize.getFixedValue();
640       LE = LE.mul(APInt(IndexSize, TypeSize), GEPOp->isInBounds());
641       Decomposed.Offset += LE.Offset.sext(MaxIndexSize);
642       APInt Scale = LE.Scale.sext(MaxIndexSize);
643 
644       // If we already had an occurrence of this index variable, merge this
645       // scale into it.  For example, we want to handle:
646       //   A[x][x] -> x*16 + x*4 -> x*20
647       // This also ensures that 'x' only appears in the index list once.
648       for (unsigned i = 0, e = Decomposed.VarIndices.size(); i != e; ++i) {
649         if (Decomposed.VarIndices[i].Val.V == LE.Val.V &&
650             Decomposed.VarIndices[i].Val.hasSameCastsAs(LE.Val)) {
651           Scale += Decomposed.VarIndices[i].Scale;
652           Decomposed.VarIndices.erase(Decomposed.VarIndices.begin() + i);
653           break;
654         }
655       }
656 
657       // Make sure that we have a scale that makes sense for this target's
658       // index size.
659       Scale = adjustToIndexSize(Scale, IndexSize);
660 
661       if (!!Scale) {
662         VariableGEPIndex Entry = {LE.Val, Scale, CxtI, LE.IsNSW};
663         Decomposed.VarIndices.push_back(Entry);
664       }
665     }
666 
667     // Take care of wrap-arounds
668     if (GepHasConstantOffset)
669       Decomposed.Offset = adjustToIndexSize(Decomposed.Offset, IndexSize);
670 
671     // Analyze the base pointer next.
672     V = GEPOp->getOperand(0);
673   } while (--MaxLookup);
674 
675   // If the chain of expressions is too deep, just return early.
676   Decomposed.Base = V;
677   SearchLimitReached++;
678   return Decomposed;
679 }
680 
681 ModRefInfo BasicAAResult::getModRefInfoMask(const MemoryLocation &Loc,
682                                             AAQueryInfo &AAQI,
683                                             bool IgnoreLocals) {
684   assert(Visited.empty() && "Visited must be cleared after use!");
685   auto _ = make_scope_exit([&] { Visited.clear(); });
686 
687   unsigned MaxLookup = 8;
688   SmallVector<const Value *, 16> Worklist;
689   Worklist.push_back(Loc.Ptr);
690   ModRefInfo Result = ModRefInfo::NoModRef;
691 
692   do {
693     const Value *V = getUnderlyingObject(Worklist.pop_back_val());
694     if (!Visited.insert(V).second)
695       continue;
696 
697     // Ignore allocas if we were instructed to do so.
698     if (IgnoreLocals && isa<AllocaInst>(V))
699       continue;
700 
701     // If the location points to memory that is known to be invariant for
702     // the life of the underlying SSA value, then we can exclude Mod from
703     // the set of valid memory effects.
704     //
705     // An argument that is marked readonly and noalias is known to be
706     // invariant while that function is executing.
707     if (const Argument *Arg = dyn_cast<Argument>(V)) {
708       if (Arg->hasNoAliasAttr() && Arg->onlyReadsMemory()) {
709         Result |= ModRefInfo::Ref;
710         continue;
711       }
712     }
713 
714     // A global constant can't be mutated.
715     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
716       // Note: this doesn't require GV to be "ODR" because it isn't legal for a
717       // global to be marked constant in some modules and non-constant in
718       // others.  GV may even be a declaration, not a definition.
719       if (!GV->isConstant())
720         return AAResultBase::getModRefInfoMask(Loc, AAQI, IgnoreLocals);
721       continue;
722     }
723 
724     // If both select values point to local memory, then so does the select.
725     if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
726       Worklist.push_back(SI->getTrueValue());
727       Worklist.push_back(SI->getFalseValue());
728       continue;
729     }
730 
731     // If all values incoming to a phi node point to local memory, then so does
732     // the phi.
733     if (const PHINode *PN = dyn_cast<PHINode>(V)) {
734       // Don't bother inspecting phi nodes with many operands.
735       if (PN->getNumIncomingValues() > MaxLookup)
736         return AAResultBase::getModRefInfoMask(Loc, AAQI, IgnoreLocals);
737       append_range(Worklist, PN->incoming_values());
738       continue;
739     }
740 
741     // Otherwise be conservative.
742     return AAResultBase::getModRefInfoMask(Loc, AAQI, IgnoreLocals);
743   } while (!Worklist.empty() && --MaxLookup);
744 
745   // If we hit the maximum number of instructions to examine, be conservative.
746   if (!Worklist.empty())
747     return AAResultBase::getModRefInfoMask(Loc, AAQI, IgnoreLocals);
748 
749   return Result;
750 }
751 
752 static bool isIntrinsicCall(const CallBase *Call, Intrinsic::ID IID) {
753   const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Call);
754   return II && II->getIntrinsicID() == IID;
755 }
756 
757 /// Returns the behavior when calling the given call site.
758 MemoryEffects BasicAAResult::getMemoryEffects(const CallBase *Call,
759                                               AAQueryInfo &AAQI) {
760   MemoryEffects Min = Call->getAttributes().getMemoryEffects();
761 
762   if (const Function *F = dyn_cast<Function>(Call->getCalledOperand())) {
763     MemoryEffects FuncME = AAQI.AAR.getMemoryEffects(F);
764     // Operand bundles on the call may also read or write memory, in addition
765     // to the behavior of the called function.
766     if (Call->hasReadingOperandBundles())
767       FuncME |= MemoryEffects::readOnly();
768     if (Call->hasClobberingOperandBundles())
769       FuncME |= MemoryEffects::writeOnly();
770     Min &= FuncME;
771   }
772 
773   return Min;
774 }
775 
776 /// Returns the behavior when calling the given function. For use when the call
777 /// site is not known.
778 MemoryEffects BasicAAResult::getMemoryEffects(const Function *F) {
779   switch (F->getIntrinsicID()) {
780   case Intrinsic::experimental_guard:
781   case Intrinsic::experimental_deoptimize:
782     // These intrinsics can read arbitrary memory, and additionally modref
783     // inaccessible memory to model control dependence.
784     return MemoryEffects::readOnly() |
785            MemoryEffects::inaccessibleMemOnly(ModRefInfo::ModRef);
786   }
787 
788   return F->getMemoryEffects();
789 }
790 
791 ModRefInfo BasicAAResult::getArgModRefInfo(const CallBase *Call,
792                                            unsigned ArgIdx) {
793   if (Call->paramHasAttr(ArgIdx, Attribute::WriteOnly))
794     return ModRefInfo::Mod;
795 
796   if (Call->paramHasAttr(ArgIdx, Attribute::ReadOnly))
797     return ModRefInfo::Ref;
798 
799   if (Call->paramHasAttr(ArgIdx, Attribute::ReadNone))
800     return ModRefInfo::NoModRef;
801 
802   return AAResultBase::getArgModRefInfo(Call, ArgIdx);
803 }
804 
805 #ifndef NDEBUG
806 static const Function *getParent(const Value *V) {
807   if (const Instruction *inst = dyn_cast<Instruction>(V)) {
808     if (!inst->getParent())
809       return nullptr;
810     return inst->getParent()->getParent();
811   }
812 
813   if (const Argument *arg = dyn_cast<Argument>(V))
814     return arg->getParent();
815 
816   return nullptr;
817 }
818 
819 static bool notDifferentParent(const Value *O1, const Value *O2) {
820 
821   const Function *F1 = getParent(O1);
822   const Function *F2 = getParent(O2);
823 
824   return !F1 || !F2 || F1 == F2;
825 }
826 #endif
827 
828 AliasResult BasicAAResult::alias(const MemoryLocation &LocA,
829                                  const MemoryLocation &LocB, AAQueryInfo &AAQI,
830                                  const Instruction *CtxI) {
831   assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
832          "BasicAliasAnalysis doesn't support interprocedural queries.");
833   return aliasCheck(LocA.Ptr, LocA.Size, LocB.Ptr, LocB.Size, AAQI, CtxI);
834 }
835 
836 /// Checks to see if the specified callsite can clobber the specified memory
837 /// object.
838 ///
839 /// Since we only look at local properties of this function, we really can't
840 /// say much about this query.  We do, however, use simple "address taken"
841 /// analysis on local objects.
842 ModRefInfo BasicAAResult::getModRefInfo(const CallBase *Call,
843                                         const MemoryLocation &Loc,
844                                         AAQueryInfo &AAQI) {
845   assert(notDifferentParent(Call, Loc.Ptr) &&
846          "AliasAnalysis query involving multiple functions!");
847 
848   const Value *Object = getUnderlyingObject(Loc.Ptr);
849 
850   // Calls marked 'tail' cannot read or write allocas from the current frame
851   // because the current frame might be destroyed by the time they run. However,
852   // a tail call may use an alloca with byval. Calling with byval copies the
853   // contents of the alloca into argument registers or stack slots, so there is
854   // no lifetime issue.
855   if (isa<AllocaInst>(Object))
856     if (const CallInst *CI = dyn_cast<CallInst>(Call))
857       if (CI->isTailCall() &&
858           !CI->getAttributes().hasAttrSomewhere(Attribute::ByVal))
859         return ModRefInfo::NoModRef;
860 
861   // Stack restore is able to modify unescaped dynamic allocas. Assume it may
862   // modify them even though the alloca is not escaped.
863   if (auto *AI = dyn_cast<AllocaInst>(Object))
864     if (!AI->isStaticAlloca() && isIntrinsicCall(Call, Intrinsic::stackrestore))
865       return ModRefInfo::Mod;
866 
867   // If the pointer is to a locally allocated object that does not escape,
868   // then the call can not mod/ref the pointer unless the call takes the pointer
869   // as an argument, and itself doesn't capture it.
870   if (!isa<Constant>(Object) && Call != Object &&
871       AAQI.CI->isNotCapturedBeforeOrAt(Object, Call)) {
872 
873     // Optimistically assume that call doesn't touch Object and check this
874     // assumption in the following loop.
875     ModRefInfo Result = ModRefInfo::NoModRef;
876 
877     unsigned OperandNo = 0;
878     for (auto CI = Call->data_operands_begin(), CE = Call->data_operands_end();
879          CI != CE; ++CI, ++OperandNo) {
880       // Only look at the no-capture or byval pointer arguments.  If this
881       // pointer were passed to arguments that were neither of these, then it
882       // couldn't be no-capture.
883       if (!(*CI)->getType()->isPointerTy() ||
884           (!Call->doesNotCapture(OperandNo) && OperandNo < Call->arg_size() &&
885            !Call->isByValArgument(OperandNo)))
886         continue;
887 
888       // Call doesn't access memory through this operand, so we don't care
889       // if it aliases with Object.
890       if (Call->doesNotAccessMemory(OperandNo))
891         continue;
892 
893       // If this is a no-capture pointer argument, see if we can tell that it
894       // is impossible to alias the pointer we're checking.
895       AliasResult AR =
896           AAQI.AAR.alias(MemoryLocation::getBeforeOrAfter(*CI),
897                          MemoryLocation::getBeforeOrAfter(Object), AAQI);
898       // Operand doesn't alias 'Object', continue looking for other aliases
899       if (AR == AliasResult::NoAlias)
900         continue;
901       // Operand aliases 'Object', but call doesn't modify it. Strengthen
902       // initial assumption and keep looking in case if there are more aliases.
903       if (Call->onlyReadsMemory(OperandNo)) {
904         Result |= ModRefInfo::Ref;
905         continue;
906       }
907       // Operand aliases 'Object' but call only writes into it.
908       if (Call->onlyWritesMemory(OperandNo)) {
909         Result |= ModRefInfo::Mod;
910         continue;
911       }
912       // This operand aliases 'Object' and call reads and writes into it.
913       // Setting ModRef will not yield an early return below, MustAlias is not
914       // used further.
915       Result = ModRefInfo::ModRef;
916       break;
917     }
918 
919     // Early return if we improved mod ref information
920     if (!isModAndRefSet(Result))
921       return Result;
922   }
923 
924   // If the call is malloc/calloc like, we can assume that it doesn't
925   // modify any IR visible value.  This is only valid because we assume these
926   // routines do not read values visible in the IR.  TODO: Consider special
927   // casing realloc and strdup routines which access only their arguments as
928   // well.  Or alternatively, replace all of this with inaccessiblememonly once
929   // that's implemented fully.
930   if (isMallocOrCallocLikeFn(Call, &TLI)) {
931     // Be conservative if the accessed pointer may alias the allocation -
932     // fallback to the generic handling below.
933     if (AAQI.AAR.alias(MemoryLocation::getBeforeOrAfter(Call), Loc, AAQI) ==
934         AliasResult::NoAlias)
935       return ModRefInfo::NoModRef;
936   }
937 
938   // Like assumes, invariant.start intrinsics were also marked as arbitrarily
939   // writing so that proper control dependencies are maintained but they never
940   // mod any particular memory location visible to the IR.
941   // *Unlike* assumes (which are now modeled as NoModRef), invariant.start
942   // intrinsic is now modeled as reading memory. This prevents hoisting the
943   // invariant.start intrinsic over stores. Consider:
944   // *ptr = 40;
945   // *ptr = 50;
946   // invariant_start(ptr)
947   // int val = *ptr;
948   // print(val);
949   //
950   // This cannot be transformed to:
951   //
952   // *ptr = 40;
953   // invariant_start(ptr)
954   // *ptr = 50;
955   // int val = *ptr;
956   // print(val);
957   //
958   // The transformation will cause the second store to be ignored (based on
959   // rules of invariant.start)  and print 40, while the first program always
960   // prints 50.
961   if (isIntrinsicCall(Call, Intrinsic::invariant_start))
962     return ModRefInfo::Ref;
963 
964   // The AAResultBase base class has some smarts, lets use them.
965   return AAResultBase::getModRefInfo(Call, Loc, AAQI);
966 }
967 
968 ModRefInfo BasicAAResult::getModRefInfo(const CallBase *Call1,
969                                         const CallBase *Call2,
970                                         AAQueryInfo &AAQI) {
971   // Guard intrinsics are marked as arbitrarily writing so that proper control
972   // dependencies are maintained but they never mods any particular memory
973   // location.
974   //
975   // *Unlike* assumes, guard intrinsics are modeled as reading memory since the
976   // heap state at the point the guard is issued needs to be consistent in case
977   // the guard invokes the "deopt" continuation.
978 
979   // NB! This function is *not* commutative, so we special case two
980   // possibilities for guard intrinsics.
981 
982   if (isIntrinsicCall(Call1, Intrinsic::experimental_guard))
983     return isModSet(getMemoryEffects(Call2, AAQI).getModRef())
984                ? ModRefInfo::Ref
985                : ModRefInfo::NoModRef;
986 
987   if (isIntrinsicCall(Call2, Intrinsic::experimental_guard))
988     return isModSet(getMemoryEffects(Call1, AAQI).getModRef())
989                ? ModRefInfo::Mod
990                : ModRefInfo::NoModRef;
991 
992   // The AAResultBase base class has some smarts, lets use them.
993   return AAResultBase::getModRefInfo(Call1, Call2, AAQI);
994 }
995 
996 /// Return true if we know V to the base address of the corresponding memory
997 /// object.  This implies that any address less than V must be out of bounds
998 /// for the underlying object.  Note that just being isIdentifiedObject() is
999 /// not enough - For example, a negative offset from a noalias argument or call
1000 /// can be inbounds w.r.t the actual underlying object.
1001 static bool isBaseOfObject(const Value *V) {
1002   // TODO: We can handle other cases here
1003   // 1) For GC languages, arguments to functions are often required to be
1004   //    base pointers.
1005   // 2) Result of allocation routines are often base pointers.  Leverage TLI.
1006   return (isa<AllocaInst>(V) || isa<GlobalVariable>(V));
1007 }
1008 
1009 /// Provides a bunch of ad-hoc rules to disambiguate a GEP instruction against
1010 /// another pointer.
1011 ///
1012 /// We know that V1 is a GEP, but we don't know anything about V2.
1013 /// UnderlyingV1 is getUnderlyingObject(GEP1), UnderlyingV2 is the same for
1014 /// V2.
1015 AliasResult BasicAAResult::aliasGEP(
1016     const GEPOperator *GEP1, LocationSize V1Size,
1017     const Value *V2, LocationSize V2Size,
1018     const Value *UnderlyingV1, const Value *UnderlyingV2, AAQueryInfo &AAQI) {
1019   if (!V1Size.hasValue() && !V2Size.hasValue()) {
1020     // TODO: This limitation exists for compile-time reasons. Relax it if we
1021     // can avoid exponential pathological cases.
1022     if (!isa<GEPOperator>(V2))
1023       return AliasResult::MayAlias;
1024 
1025     // If both accesses have unknown size, we can only check whether the base
1026     // objects don't alias.
1027     AliasResult BaseAlias =
1028         AAQI.AAR.alias(MemoryLocation::getBeforeOrAfter(UnderlyingV1),
1029                        MemoryLocation::getBeforeOrAfter(UnderlyingV2), AAQI);
1030     return BaseAlias == AliasResult::NoAlias ? AliasResult::NoAlias
1031                                              : AliasResult::MayAlias;
1032   }
1033 
1034   DecomposedGEP DecompGEP1 = DecomposeGEPExpression(GEP1, DL, &AC, DT);
1035   DecomposedGEP DecompGEP2 = DecomposeGEPExpression(V2, DL, &AC, DT);
1036 
1037   // Bail if we were not able to decompose anything.
1038   if (DecompGEP1.Base == GEP1 && DecompGEP2.Base == V2)
1039     return AliasResult::MayAlias;
1040 
1041   // Subtract the GEP2 pointer from the GEP1 pointer to find out their
1042   // symbolic difference.
1043   subtractDecomposedGEPs(DecompGEP1, DecompGEP2, AAQI);
1044 
1045   // If an inbounds GEP would have to start from an out of bounds address
1046   // for the two to alias, then we can assume noalias.
1047   if (*DecompGEP1.InBounds && DecompGEP1.VarIndices.empty() &&
1048       V2Size.hasValue() && DecompGEP1.Offset.sge(V2Size.getValue()) &&
1049       isBaseOfObject(DecompGEP2.Base))
1050     return AliasResult::NoAlias;
1051 
1052   if (isa<GEPOperator>(V2)) {
1053     // Symmetric case to above.
1054     if (*DecompGEP2.InBounds && DecompGEP1.VarIndices.empty() &&
1055         V1Size.hasValue() && DecompGEP1.Offset.sle(-V1Size.getValue()) &&
1056         isBaseOfObject(DecompGEP1.Base))
1057       return AliasResult::NoAlias;
1058   }
1059 
1060   // For GEPs with identical offsets, we can preserve the size and AAInfo
1061   // when performing the alias check on the underlying objects.
1062   if (DecompGEP1.Offset == 0 && DecompGEP1.VarIndices.empty())
1063     return AAQI.AAR.alias(MemoryLocation(DecompGEP1.Base, V1Size),
1064                           MemoryLocation(DecompGEP2.Base, V2Size), AAQI);
1065 
1066   // Do the base pointers alias?
1067   AliasResult BaseAlias =
1068       AAQI.AAR.alias(MemoryLocation::getBeforeOrAfter(DecompGEP1.Base),
1069                      MemoryLocation::getBeforeOrAfter(DecompGEP2.Base), AAQI);
1070 
1071   // If we get a No or May, then return it immediately, no amount of analysis
1072   // will improve this situation.
1073   if (BaseAlias != AliasResult::MustAlias) {
1074     assert(BaseAlias == AliasResult::NoAlias ||
1075            BaseAlias == AliasResult::MayAlias);
1076     return BaseAlias;
1077   }
1078 
1079   // If there is a constant difference between the pointers, but the difference
1080   // is less than the size of the associated memory object, then we know
1081   // that the objects are partially overlapping.  If the difference is
1082   // greater, we know they do not overlap.
1083   if (DecompGEP1.VarIndices.empty()) {
1084     APInt &Off = DecompGEP1.Offset;
1085 
1086     // Initialize for Off >= 0 (V2 <= GEP1) case.
1087     const Value *LeftPtr = V2;
1088     const Value *RightPtr = GEP1;
1089     LocationSize VLeftSize = V2Size;
1090     LocationSize VRightSize = V1Size;
1091     const bool Swapped = Off.isNegative();
1092 
1093     if (Swapped) {
1094       // Swap if we have the situation where:
1095       // +                +
1096       // | BaseOffset     |
1097       // ---------------->|
1098       // |-->V1Size       |-------> V2Size
1099       // GEP1             V2
1100       std::swap(LeftPtr, RightPtr);
1101       std::swap(VLeftSize, VRightSize);
1102       Off = -Off;
1103     }
1104 
1105     if (!VLeftSize.hasValue())
1106       return AliasResult::MayAlias;
1107 
1108     const uint64_t LSize = VLeftSize.getValue();
1109     if (Off.ult(LSize)) {
1110       // Conservatively drop processing if a phi was visited and/or offset is
1111       // too big.
1112       AliasResult AR = AliasResult::PartialAlias;
1113       if (VRightSize.hasValue() && Off.ule(INT32_MAX) &&
1114           (Off + VRightSize.getValue()).ule(LSize)) {
1115         // Memory referenced by right pointer is nested. Save the offset in
1116         // cache. Note that originally offset estimated as GEP1-V2, but
1117         // AliasResult contains the shift that represents GEP1+Offset=V2.
1118         AR.setOffset(-Off.getSExtValue());
1119         AR.swap(Swapped);
1120       }
1121       return AR;
1122     }
1123     return AliasResult::NoAlias;
1124   }
1125 
1126   // We need to know both acess sizes for all the following heuristics.
1127   if (!V1Size.hasValue() || !V2Size.hasValue())
1128     return AliasResult::MayAlias;
1129 
1130   APInt GCD;
1131   ConstantRange OffsetRange = ConstantRange(DecompGEP1.Offset);
1132   for (unsigned i = 0, e = DecompGEP1.VarIndices.size(); i != e; ++i) {
1133     const VariableGEPIndex &Index = DecompGEP1.VarIndices[i];
1134     const APInt &Scale = Index.Scale;
1135     APInt ScaleForGCD = Scale;
1136     if (!Index.IsNSW)
1137       ScaleForGCD = APInt::getOneBitSet(Scale.getBitWidth(),
1138                                         Scale.countTrailingZeros());
1139 
1140     if (i == 0)
1141       GCD = ScaleForGCD.abs();
1142     else
1143       GCD = APIntOps::GreatestCommonDivisor(GCD, ScaleForGCD.abs());
1144 
1145     ConstantRange CR = computeConstantRange(Index.Val.V, /* ForSigned */ false,
1146                                             true, &AC, Index.CxtI);
1147     KnownBits Known =
1148         computeKnownBits(Index.Val.V, DL, 0, &AC, Index.CxtI, DT);
1149     CR = CR.intersectWith(
1150         ConstantRange::fromKnownBits(Known, /* Signed */ true),
1151         ConstantRange::Signed);
1152     CR = Index.Val.evaluateWith(CR).sextOrTrunc(OffsetRange.getBitWidth());
1153 
1154     assert(OffsetRange.getBitWidth() == Scale.getBitWidth() &&
1155            "Bit widths are normalized to MaxIndexSize");
1156     if (Index.IsNSW)
1157       OffsetRange = OffsetRange.add(CR.smul_sat(ConstantRange(Scale)));
1158     else
1159       OffsetRange = OffsetRange.add(CR.smul_fast(ConstantRange(Scale)));
1160   }
1161 
1162   // We now have accesses at two offsets from the same base:
1163   //  1. (...)*GCD + DecompGEP1.Offset with size V1Size
1164   //  2. 0 with size V2Size
1165   // Using arithmetic modulo GCD, the accesses are at
1166   // [ModOffset..ModOffset+V1Size) and [0..V2Size). If the first access fits
1167   // into the range [V2Size..GCD), then we know they cannot overlap.
1168   APInt ModOffset = DecompGEP1.Offset.srem(GCD);
1169   if (ModOffset.isNegative())
1170     ModOffset += GCD; // We want mod, not rem.
1171   if (ModOffset.uge(V2Size.getValue()) &&
1172       (GCD - ModOffset).uge(V1Size.getValue()))
1173     return AliasResult::NoAlias;
1174 
1175   // Compute ranges of potentially accessed bytes for both accesses. If the
1176   // interseciton is empty, there can be no overlap.
1177   unsigned BW = OffsetRange.getBitWidth();
1178   ConstantRange Range1 = OffsetRange.add(
1179       ConstantRange(APInt(BW, 0), APInt(BW, V1Size.getValue())));
1180   ConstantRange Range2 =
1181       ConstantRange(APInt(BW, 0), APInt(BW, V2Size.getValue()));
1182   if (Range1.intersectWith(Range2).isEmptySet())
1183     return AliasResult::NoAlias;
1184 
1185   // Try to determine the range of values for VarIndex such that
1186   // VarIndex <= -MinAbsVarIndex || MinAbsVarIndex <= VarIndex.
1187   std::optional<APInt> MinAbsVarIndex;
1188   if (DecompGEP1.VarIndices.size() == 1) {
1189     // VarIndex = Scale*V.
1190     const VariableGEPIndex &Var = DecompGEP1.VarIndices[0];
1191     if (Var.Val.TruncBits == 0 &&
1192         isKnownNonZero(Var.Val.V, DL, 0, &AC, Var.CxtI, DT)) {
1193       // If V != 0, then abs(VarIndex) > 0.
1194       MinAbsVarIndex = APInt(Var.Scale.getBitWidth(), 1);
1195 
1196       // Check if abs(V*Scale) >= abs(Scale) holds in the presence of
1197       // potentially wrapping math.
1198       auto MultiplyByScaleNoWrap = [](const VariableGEPIndex &Var) {
1199         if (Var.IsNSW)
1200           return true;
1201 
1202         int ValOrigBW = Var.Val.V->getType()->getPrimitiveSizeInBits();
1203         // If Scale is small enough so that abs(V*Scale) >= abs(Scale) holds.
1204         // The max value of abs(V) is 2^ValOrigBW - 1. Multiplying with a
1205         // constant smaller than 2^(bitwidth(Val) - ValOrigBW) won't wrap.
1206         int MaxScaleValueBW = Var.Val.getBitWidth() - ValOrigBW;
1207         if (MaxScaleValueBW <= 0)
1208           return false;
1209         return Var.Scale.ule(
1210             APInt::getMaxValue(MaxScaleValueBW).zext(Var.Scale.getBitWidth()));
1211       };
1212       // Refine MinAbsVarIndex, if abs(Scale*V) >= abs(Scale) holds in the
1213       // presence of potentially wrapping math.
1214       if (MultiplyByScaleNoWrap(Var)) {
1215         // If V != 0 then abs(VarIndex) >= abs(Scale).
1216         MinAbsVarIndex = Var.Scale.abs();
1217       }
1218     }
1219   } else if (DecompGEP1.VarIndices.size() == 2) {
1220     // VarIndex = Scale*V0 + (-Scale)*V1.
1221     // If V0 != V1 then abs(VarIndex) >= abs(Scale).
1222     // Check that MayBeCrossIteration is false, to avoid reasoning about
1223     // inequality of values across loop iterations.
1224     const VariableGEPIndex &Var0 = DecompGEP1.VarIndices[0];
1225     const VariableGEPIndex &Var1 = DecompGEP1.VarIndices[1];
1226     if (Var0.Scale == -Var1.Scale && Var0.Val.TruncBits == 0 &&
1227         Var0.Val.hasSameCastsAs(Var1.Val) && !AAQI.MayBeCrossIteration &&
1228         isKnownNonEqual(Var0.Val.V, Var1.Val.V, DL, &AC, /* CxtI */ nullptr,
1229                         DT))
1230       MinAbsVarIndex = Var0.Scale.abs();
1231   }
1232 
1233   if (MinAbsVarIndex) {
1234     // The constant offset will have added at least +/-MinAbsVarIndex to it.
1235     APInt OffsetLo = DecompGEP1.Offset - *MinAbsVarIndex;
1236     APInt OffsetHi = DecompGEP1.Offset + *MinAbsVarIndex;
1237     // We know that Offset <= OffsetLo || Offset >= OffsetHi
1238     if (OffsetLo.isNegative() && (-OffsetLo).uge(V1Size.getValue()) &&
1239         OffsetHi.isNonNegative() && OffsetHi.uge(V2Size.getValue()))
1240       return AliasResult::NoAlias;
1241   }
1242 
1243   if (constantOffsetHeuristic(DecompGEP1, V1Size, V2Size, &AC, DT, AAQI))
1244     return AliasResult::NoAlias;
1245 
1246   // Statically, we can see that the base objects are the same, but the
1247   // pointers have dynamic offsets which we can't resolve. And none of our
1248   // little tricks above worked.
1249   return AliasResult::MayAlias;
1250 }
1251 
1252 static AliasResult MergeAliasResults(AliasResult A, AliasResult B) {
1253   // If the results agree, take it.
1254   if (A == B)
1255     return A;
1256   // A mix of PartialAlias and MustAlias is PartialAlias.
1257   if ((A == AliasResult::PartialAlias && B == AliasResult::MustAlias) ||
1258       (B == AliasResult::PartialAlias && A == AliasResult::MustAlias))
1259     return AliasResult::PartialAlias;
1260   // Otherwise, we don't know anything.
1261   return AliasResult::MayAlias;
1262 }
1263 
1264 /// Provides a bunch of ad-hoc rules to disambiguate a Select instruction
1265 /// against another.
1266 AliasResult
1267 BasicAAResult::aliasSelect(const SelectInst *SI, LocationSize SISize,
1268                            const Value *V2, LocationSize V2Size,
1269                            AAQueryInfo &AAQI) {
1270   // If the values are Selects with the same condition, we can do a more precise
1271   // check: just check for aliases between the values on corresponding arms.
1272   if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1273     if (isValueEqualInPotentialCycles(SI->getCondition(), SI2->getCondition(),
1274                                       AAQI)) {
1275       AliasResult Alias =
1276           AAQI.AAR.alias(MemoryLocation(SI->getTrueValue(), SISize),
1277                          MemoryLocation(SI2->getTrueValue(), V2Size), AAQI);
1278       if (Alias == AliasResult::MayAlias)
1279         return AliasResult::MayAlias;
1280       AliasResult ThisAlias =
1281           AAQI.AAR.alias(MemoryLocation(SI->getFalseValue(), SISize),
1282                          MemoryLocation(SI2->getFalseValue(), V2Size), AAQI);
1283       return MergeAliasResults(ThisAlias, Alias);
1284     }
1285 
1286   // If both arms of the Select node NoAlias or MustAlias V2, then returns
1287   // NoAlias / MustAlias. Otherwise, returns MayAlias.
1288   AliasResult Alias = AAQI.AAR.alias(MemoryLocation(SI->getTrueValue(), SISize),
1289                                      MemoryLocation(V2, V2Size), AAQI);
1290   if (Alias == AliasResult::MayAlias)
1291     return AliasResult::MayAlias;
1292 
1293   AliasResult ThisAlias =
1294       AAQI.AAR.alias(MemoryLocation(SI->getFalseValue(), SISize),
1295                      MemoryLocation(V2, V2Size), AAQI);
1296   return MergeAliasResults(ThisAlias, Alias);
1297 }
1298 
1299 /// Provide a bunch of ad-hoc rules to disambiguate a PHI instruction against
1300 /// another.
1301 AliasResult BasicAAResult::aliasPHI(const PHINode *PN, LocationSize PNSize,
1302                                     const Value *V2, LocationSize V2Size,
1303                                     AAQueryInfo &AAQI) {
1304   if (!PN->getNumIncomingValues())
1305     return AliasResult::NoAlias;
1306   // If the values are PHIs in the same block, we can do a more precise
1307   // as well as efficient check: just check for aliases between the values
1308   // on corresponding edges.
1309   if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1310     if (PN2->getParent() == PN->getParent()) {
1311       std::optional<AliasResult> Alias;
1312       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1313         AliasResult ThisAlias = AAQI.AAR.alias(
1314             MemoryLocation(PN->getIncomingValue(i), PNSize),
1315             MemoryLocation(
1316                 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)), V2Size),
1317             AAQI);
1318         if (Alias)
1319           *Alias = MergeAliasResults(*Alias, ThisAlias);
1320         else
1321           Alias = ThisAlias;
1322         if (*Alias == AliasResult::MayAlias)
1323           break;
1324       }
1325       return *Alias;
1326     }
1327 
1328   SmallVector<Value *, 4> V1Srcs;
1329   // If a phi operand recurses back to the phi, we can still determine NoAlias
1330   // if we don't alias the underlying objects of the other phi operands, as we
1331   // know that the recursive phi needs to be based on them in some way.
1332   bool isRecursive = false;
1333   auto CheckForRecPhi = [&](Value *PV) {
1334     if (!EnableRecPhiAnalysis)
1335       return false;
1336     if (getUnderlyingObject(PV) == PN) {
1337       isRecursive = true;
1338       return true;
1339     }
1340     return false;
1341   };
1342 
1343   SmallPtrSet<Value *, 4> UniqueSrc;
1344   Value *OnePhi = nullptr;
1345   for (Value *PV1 : PN->incoming_values()) {
1346     // Skip the phi itself being the incoming value.
1347     if (PV1 == PN)
1348       continue;
1349 
1350     if (isa<PHINode>(PV1)) {
1351       if (OnePhi && OnePhi != PV1) {
1352         // To control potential compile time explosion, we choose to be
1353         // conserviate when we have more than one Phi input.  It is important
1354         // that we handle the single phi case as that lets us handle LCSSA
1355         // phi nodes and (combined with the recursive phi handling) simple
1356         // pointer induction variable patterns.
1357         return AliasResult::MayAlias;
1358       }
1359       OnePhi = PV1;
1360     }
1361 
1362     if (CheckForRecPhi(PV1))
1363       continue;
1364 
1365     if (UniqueSrc.insert(PV1).second)
1366       V1Srcs.push_back(PV1);
1367   }
1368 
1369   if (OnePhi && UniqueSrc.size() > 1)
1370     // Out of an abundance of caution, allow only the trivial lcssa and
1371     // recursive phi cases.
1372     return AliasResult::MayAlias;
1373 
1374   // If V1Srcs is empty then that means that the phi has no underlying non-phi
1375   // value. This should only be possible in blocks unreachable from the entry
1376   // block, but return MayAlias just in case.
1377   if (V1Srcs.empty())
1378     return AliasResult::MayAlias;
1379 
1380   // If this PHI node is recursive, indicate that the pointer may be moved
1381   // across iterations. We can only prove NoAlias if different underlying
1382   // objects are involved.
1383   if (isRecursive)
1384     PNSize = LocationSize::beforeOrAfterPointer();
1385 
1386   // In the recursive alias queries below, we may compare values from two
1387   // different loop iterations.
1388   SaveAndRestore SavedMayBeCrossIteration(AAQI.MayBeCrossIteration, true);
1389 
1390   AliasResult Alias = AAQI.AAR.alias(MemoryLocation(V1Srcs[0], PNSize),
1391                                      MemoryLocation(V2, V2Size), AAQI);
1392 
1393   // Early exit if the check of the first PHI source against V2 is MayAlias.
1394   // Other results are not possible.
1395   if (Alias == AliasResult::MayAlias)
1396     return AliasResult::MayAlias;
1397   // With recursive phis we cannot guarantee that MustAlias/PartialAlias will
1398   // remain valid to all elements and needs to conservatively return MayAlias.
1399   if (isRecursive && Alias != AliasResult::NoAlias)
1400     return AliasResult::MayAlias;
1401 
1402   // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1403   // NoAlias / MustAlias. Otherwise, returns MayAlias.
1404   for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1405     Value *V = V1Srcs[i];
1406 
1407     AliasResult ThisAlias = AAQI.AAR.alias(
1408         MemoryLocation(V, PNSize), MemoryLocation(V2, V2Size), AAQI);
1409     Alias = MergeAliasResults(ThisAlias, Alias);
1410     if (Alias == AliasResult::MayAlias)
1411       break;
1412   }
1413 
1414   return Alias;
1415 }
1416 
1417 /// Provides a bunch of ad-hoc rules to disambiguate in common cases, such as
1418 /// array references.
1419 AliasResult BasicAAResult::aliasCheck(const Value *V1, LocationSize V1Size,
1420                                       const Value *V2, LocationSize V2Size,
1421                                       AAQueryInfo &AAQI,
1422                                       const Instruction *CtxI) {
1423   // If either of the memory references is empty, it doesn't matter what the
1424   // pointer values are.
1425   if (V1Size.isZero() || V2Size.isZero())
1426     return AliasResult::NoAlias;
1427 
1428   // Strip off any casts if they exist.
1429   V1 = V1->stripPointerCastsForAliasAnalysis();
1430   V2 = V2->stripPointerCastsForAliasAnalysis();
1431 
1432   // If V1 or V2 is undef, the result is NoAlias because we can always pick a
1433   // value for undef that aliases nothing in the program.
1434   if (isa<UndefValue>(V1) || isa<UndefValue>(V2))
1435     return AliasResult::NoAlias;
1436 
1437   // Are we checking for alias of the same value?
1438   // Because we look 'through' phi nodes, we could look at "Value" pointers from
1439   // different iterations. We must therefore make sure that this is not the
1440   // case. The function isValueEqualInPotentialCycles ensures that this cannot
1441   // happen by looking at the visited phi nodes and making sure they cannot
1442   // reach the value.
1443   if (isValueEqualInPotentialCycles(V1, V2, AAQI))
1444     return AliasResult::MustAlias;
1445 
1446   if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1447     return AliasResult::NoAlias; // Scalars cannot alias each other
1448 
1449   // Figure out what objects these things are pointing to if we can.
1450   const Value *O1 = getUnderlyingObject(V1, MaxLookupSearchDepth);
1451   const Value *O2 = getUnderlyingObject(V2, MaxLookupSearchDepth);
1452 
1453   // Null values in the default address space don't point to any object, so they
1454   // don't alias any other pointer.
1455   if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1456     if (!NullPointerIsDefined(&F, CPN->getType()->getAddressSpace()))
1457       return AliasResult::NoAlias;
1458   if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1459     if (!NullPointerIsDefined(&F, CPN->getType()->getAddressSpace()))
1460       return AliasResult::NoAlias;
1461 
1462   if (O1 != O2) {
1463     // If V1/V2 point to two different objects, we know that we have no alias.
1464     if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1465       return AliasResult::NoAlias;
1466 
1467     // Constant pointers can't alias with non-const isIdentifiedObject objects.
1468     if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1469         (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1470       return AliasResult::NoAlias;
1471 
1472     // Function arguments can't alias with things that are known to be
1473     // unambigously identified at the function level.
1474     if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1475         (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1476       return AliasResult::NoAlias;
1477 
1478     // If one pointer is the result of a call/invoke or load and the other is a
1479     // non-escaping local object within the same function, then we know the
1480     // object couldn't escape to a point where the call could return it.
1481     //
1482     // Note that if the pointers are in different functions, there are a
1483     // variety of complications. A call with a nocapture argument may still
1484     // temporary store the nocapture argument's value in a temporary memory
1485     // location if that memory location doesn't escape. Or it may pass a
1486     // nocapture value to other functions as long as they don't capture it.
1487     if (isEscapeSource(O1) &&
1488         AAQI.CI->isNotCapturedBeforeOrAt(O2, cast<Instruction>(O1)))
1489       return AliasResult::NoAlias;
1490     if (isEscapeSource(O2) &&
1491         AAQI.CI->isNotCapturedBeforeOrAt(O1, cast<Instruction>(O2)))
1492       return AliasResult::NoAlias;
1493   }
1494 
1495   // If the size of one access is larger than the entire object on the other
1496   // side, then we know such behavior is undefined and can assume no alias.
1497   bool NullIsValidLocation = NullPointerIsDefined(&F);
1498   if ((isObjectSmallerThan(
1499           O2, getMinimalExtentFrom(*V1, V1Size, DL, NullIsValidLocation), DL,
1500           TLI, NullIsValidLocation)) ||
1501       (isObjectSmallerThan(
1502           O1, getMinimalExtentFrom(*V2, V2Size, DL, NullIsValidLocation), DL,
1503           TLI, NullIsValidLocation)))
1504     return AliasResult::NoAlias;
1505 
1506   if (CtxI && EnableSeparateStorageAnalysis) {
1507     for (auto &AssumeVH : AC.assumptions()) {
1508       if (!AssumeVH)
1509         continue;
1510 
1511       AssumeInst *Assume = cast<AssumeInst>(AssumeVH);
1512 
1513       for (unsigned Idx = 0; Idx < Assume->getNumOperandBundles(); Idx++) {
1514         OperandBundleUse OBU = Assume->getOperandBundleAt(Idx);
1515         if (OBU.getTagName() == "separate_storage") {
1516           assert(OBU.Inputs.size() == 2);
1517           const Value *Hint1 = OBU.Inputs[0].get();
1518           const Value *Hint2 = OBU.Inputs[1].get();
1519           const Value *HintO1 = getUnderlyingObject(Hint1);
1520           const Value *HintO2 = getUnderlyingObject(Hint2);
1521 
1522           if (((O1 == HintO1 && O2 == HintO2) ||
1523                (O1 == HintO2 && O2 == HintO1)) &&
1524               isValidAssumeForContext(Assume, CtxI, DT))
1525             return AliasResult::NoAlias;
1526         }
1527       }
1528     }
1529   }
1530 
1531   // If one the accesses may be before the accessed pointer, canonicalize this
1532   // by using unknown after-pointer sizes for both accesses. This is
1533   // equivalent, because regardless of which pointer is lower, one of them
1534   // will always came after the other, as long as the underlying objects aren't
1535   // disjoint. We do this so that the rest of BasicAA does not have to deal
1536   // with accesses before the base pointer, and to improve cache utilization by
1537   // merging equivalent states.
1538   if (V1Size.mayBeBeforePointer() || V2Size.mayBeBeforePointer()) {
1539     V1Size = LocationSize::afterPointer();
1540     V2Size = LocationSize::afterPointer();
1541   }
1542 
1543   // FIXME: If this depth limit is hit, then we may cache sub-optimal results
1544   // for recursive queries. For this reason, this limit is chosen to be large
1545   // enough to be very rarely hit, while still being small enough to avoid
1546   // stack overflows.
1547   if (AAQI.Depth >= 512)
1548     return AliasResult::MayAlias;
1549 
1550   // Check the cache before climbing up use-def chains. This also terminates
1551   // otherwise infinitely recursive queries. Include MayBeCrossIteration in the
1552   // cache key, because some cases where MayBeCrossIteration==false returns
1553   // MustAlias or NoAlias may become MayAlias under MayBeCrossIteration==true.
1554   AAQueryInfo::LocPair Locs({V1, V1Size, AAQI.MayBeCrossIteration},
1555                             {V2, V2Size, AAQI.MayBeCrossIteration});
1556   const bool Swapped = V1 > V2;
1557   if (Swapped)
1558     std::swap(Locs.first, Locs.second);
1559   const auto &Pair = AAQI.AliasCache.try_emplace(
1560       Locs, AAQueryInfo::CacheEntry{AliasResult::NoAlias, 0});
1561   if (!Pair.second) {
1562     auto &Entry = Pair.first->second;
1563     if (!Entry.isDefinitive()) {
1564       // Remember that we used an assumption.
1565       ++Entry.NumAssumptionUses;
1566       ++AAQI.NumAssumptionUses;
1567     }
1568     // Cache contains sorted {V1,V2} pairs but we should return original order.
1569     auto Result = Entry.Result;
1570     Result.swap(Swapped);
1571     return Result;
1572   }
1573 
1574   int OrigNumAssumptionUses = AAQI.NumAssumptionUses;
1575   unsigned OrigNumAssumptionBasedResults = AAQI.AssumptionBasedResults.size();
1576   AliasResult Result =
1577       aliasCheckRecursive(V1, V1Size, V2, V2Size, AAQI, O1, O2);
1578 
1579   auto It = AAQI.AliasCache.find(Locs);
1580   assert(It != AAQI.AliasCache.end() && "Must be in cache");
1581   auto &Entry = It->second;
1582 
1583   // Check whether a NoAlias assumption has been used, but disproven.
1584   bool AssumptionDisproven =
1585       Entry.NumAssumptionUses > 0 && Result != AliasResult::NoAlias;
1586   if (AssumptionDisproven)
1587     Result = AliasResult::MayAlias;
1588 
1589   // This is a definitive result now, when considered as a root query.
1590   AAQI.NumAssumptionUses -= Entry.NumAssumptionUses;
1591   Entry.Result = Result;
1592   // Cache contains sorted {V1,V2} pairs.
1593   Entry.Result.swap(Swapped);
1594   Entry.NumAssumptionUses = -1;
1595 
1596   // If the assumption has been disproven, remove any results that may have
1597   // been based on this assumption. Do this after the Entry updates above to
1598   // avoid iterator invalidation.
1599   if (AssumptionDisproven)
1600     while (AAQI.AssumptionBasedResults.size() > OrigNumAssumptionBasedResults)
1601       AAQI.AliasCache.erase(AAQI.AssumptionBasedResults.pop_back_val());
1602 
1603   // The result may still be based on assumptions higher up in the chain.
1604   // Remember it, so it can be purged from the cache later.
1605   if (OrigNumAssumptionUses != AAQI.NumAssumptionUses &&
1606       Result != AliasResult::MayAlias)
1607     AAQI.AssumptionBasedResults.push_back(Locs);
1608   return Result;
1609 }
1610 
1611 AliasResult BasicAAResult::aliasCheckRecursive(
1612     const Value *V1, LocationSize V1Size,
1613     const Value *V2, LocationSize V2Size,
1614     AAQueryInfo &AAQI, const Value *O1, const Value *O2) {
1615   if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1616     AliasResult Result = aliasGEP(GV1, V1Size, V2, V2Size, O1, O2, AAQI);
1617     if (Result != AliasResult::MayAlias)
1618       return Result;
1619   } else if (const GEPOperator *GV2 = dyn_cast<GEPOperator>(V2)) {
1620     AliasResult Result = aliasGEP(GV2, V2Size, V1, V1Size, O2, O1, AAQI);
1621     Result.swap();
1622     if (Result != AliasResult::MayAlias)
1623       return Result;
1624   }
1625 
1626   if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1627     AliasResult Result = aliasPHI(PN, V1Size, V2, V2Size, AAQI);
1628     if (Result != AliasResult::MayAlias)
1629       return Result;
1630   } else if (const PHINode *PN = dyn_cast<PHINode>(V2)) {
1631     AliasResult Result = aliasPHI(PN, V2Size, V1, V1Size, AAQI);
1632     Result.swap();
1633     if (Result != AliasResult::MayAlias)
1634       return Result;
1635   }
1636 
1637   if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1638     AliasResult Result = aliasSelect(S1, V1Size, V2, V2Size, AAQI);
1639     if (Result != AliasResult::MayAlias)
1640       return Result;
1641   } else if (const SelectInst *S2 = dyn_cast<SelectInst>(V2)) {
1642     AliasResult Result = aliasSelect(S2, V2Size, V1, V1Size, AAQI);
1643     Result.swap();
1644     if (Result != AliasResult::MayAlias)
1645       return Result;
1646   }
1647 
1648   // If both pointers are pointing into the same object and one of them
1649   // accesses the entire object, then the accesses must overlap in some way.
1650   if (O1 == O2) {
1651     bool NullIsValidLocation = NullPointerIsDefined(&F);
1652     if (V1Size.isPrecise() && V2Size.isPrecise() &&
1653         (isObjectSize(O1, V1Size.getValue(), DL, TLI, NullIsValidLocation) ||
1654          isObjectSize(O2, V2Size.getValue(), DL, TLI, NullIsValidLocation)))
1655       return AliasResult::PartialAlias;
1656   }
1657 
1658   return AliasResult::MayAlias;
1659 }
1660 
1661 /// Check whether two Values can be considered equivalent.
1662 ///
1663 /// If the values may come from different cycle iterations, this will also
1664 /// check that the values are not part of cycle. We have to do this because we
1665 /// are looking through phi nodes, that is we say
1666 /// noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB).
1667 bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V,
1668                                                   const Value *V2,
1669                                                   const AAQueryInfo &AAQI) {
1670   if (V != V2)
1671     return false;
1672 
1673   if (!AAQI.MayBeCrossIteration)
1674     return true;
1675 
1676   // Non-instructions and instructions in the entry block cannot be part of
1677   // a loop.
1678   const Instruction *Inst = dyn_cast<Instruction>(V);
1679   if (!Inst || Inst->getParent()->isEntryBlock())
1680     return true;
1681 
1682   // Check whether the instruction is part of a cycle, by checking whether the
1683   // block can (non-trivially) reach itself.
1684   BasicBlock *BB = const_cast<BasicBlock *>(Inst->getParent());
1685   SmallVector<BasicBlock *> Succs(successors(BB));
1686   return !Succs.empty() &&
1687          !isPotentiallyReachableFromMany(Succs, BB, nullptr, DT);
1688 }
1689 
1690 /// Computes the symbolic difference between two de-composed GEPs.
1691 void BasicAAResult::subtractDecomposedGEPs(DecomposedGEP &DestGEP,
1692                                            const DecomposedGEP &SrcGEP,
1693                                            const AAQueryInfo &AAQI) {
1694   DestGEP.Offset -= SrcGEP.Offset;
1695   for (const VariableGEPIndex &Src : SrcGEP.VarIndices) {
1696     // Find V in Dest.  This is N^2, but pointer indices almost never have more
1697     // than a few variable indexes.
1698     bool Found = false;
1699     for (auto I : enumerate(DestGEP.VarIndices)) {
1700       VariableGEPIndex &Dest = I.value();
1701       if (!isValueEqualInPotentialCycles(Dest.Val.V, Src.Val.V, AAQI) ||
1702           !Dest.Val.hasSameCastsAs(Src.Val))
1703         continue;
1704 
1705       // If we found it, subtract off Scale V's from the entry in Dest.  If it
1706       // goes to zero, remove the entry.
1707       if (Dest.Scale != Src.Scale) {
1708         Dest.Scale -= Src.Scale;
1709         Dest.IsNSW = false;
1710       } else {
1711         DestGEP.VarIndices.erase(DestGEP.VarIndices.begin() + I.index());
1712       }
1713       Found = true;
1714       break;
1715     }
1716 
1717     // If we didn't consume this entry, add it to the end of the Dest list.
1718     if (!Found) {
1719       VariableGEPIndex Entry = {Src.Val, -Src.Scale, Src.CxtI, Src.IsNSW};
1720       DestGEP.VarIndices.push_back(Entry);
1721     }
1722   }
1723 }
1724 
1725 bool BasicAAResult::constantOffsetHeuristic(const DecomposedGEP &GEP,
1726                                             LocationSize MaybeV1Size,
1727                                             LocationSize MaybeV2Size,
1728                                             AssumptionCache *AC,
1729                                             DominatorTree *DT,
1730                                             const AAQueryInfo &AAQI) {
1731   if (GEP.VarIndices.size() != 2 || !MaybeV1Size.hasValue() ||
1732       !MaybeV2Size.hasValue())
1733     return false;
1734 
1735   const uint64_t V1Size = MaybeV1Size.getValue();
1736   const uint64_t V2Size = MaybeV2Size.getValue();
1737 
1738   const VariableGEPIndex &Var0 = GEP.VarIndices[0], &Var1 = GEP.VarIndices[1];
1739 
1740   if (Var0.Val.TruncBits != 0 || !Var0.Val.hasSameCastsAs(Var1.Val) ||
1741       Var0.Scale != -Var1.Scale ||
1742       Var0.Val.V->getType() != Var1.Val.V->getType())
1743     return false;
1744 
1745   // We'll strip off the Extensions of Var0 and Var1 and do another round
1746   // of GetLinearExpression decomposition. In the example above, if Var0
1747   // is zext(%x + 1) we should get V1 == %x and V1Offset == 1.
1748 
1749   LinearExpression E0 =
1750       GetLinearExpression(CastedValue(Var0.Val.V), DL, 0, AC, DT);
1751   LinearExpression E1 =
1752       GetLinearExpression(CastedValue(Var1.Val.V), DL, 0, AC, DT);
1753   if (E0.Scale != E1.Scale || !E0.Val.hasSameCastsAs(E1.Val) ||
1754       !isValueEqualInPotentialCycles(E0.Val.V, E1.Val.V, AAQI))
1755     return false;
1756 
1757   // We have a hit - Var0 and Var1 only differ by a constant offset!
1758 
1759   // If we've been sext'ed then zext'd the maximum difference between Var0 and
1760   // Var1 is possible to calculate, but we're just interested in the absolute
1761   // minimum difference between the two. The minimum distance may occur due to
1762   // wrapping; consider "add i3 %i, 5": if %i == 7 then 7 + 5 mod 8 == 4, and so
1763   // the minimum distance between %i and %i + 5 is 3.
1764   APInt MinDiff = E0.Offset - E1.Offset, Wrapped = -MinDiff;
1765   MinDiff = APIntOps::umin(MinDiff, Wrapped);
1766   APInt MinDiffBytes =
1767     MinDiff.zextOrTrunc(Var0.Scale.getBitWidth()) * Var0.Scale.abs();
1768 
1769   // We can't definitely say whether GEP1 is before or after V2 due to wrapping
1770   // arithmetic (i.e. for some values of GEP1 and V2 GEP1 < V2, and for other
1771   // values GEP1 > V2). We'll therefore only declare NoAlias if both V1Size and
1772   // V2Size can fit in the MinDiffBytes gap.
1773   return MinDiffBytes.uge(V1Size + GEP.Offset.abs()) &&
1774          MinDiffBytes.uge(V2Size + GEP.Offset.abs());
1775 }
1776 
1777 //===----------------------------------------------------------------------===//
1778 // BasicAliasAnalysis Pass
1779 //===----------------------------------------------------------------------===//
1780 
1781 AnalysisKey BasicAA::Key;
1782 
1783 BasicAAResult BasicAA::run(Function &F, FunctionAnalysisManager &AM) {
1784   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1785   auto &AC = AM.getResult<AssumptionAnalysis>(F);
1786   auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
1787   return BasicAAResult(F.getParent()->getDataLayout(), F, TLI, AC, DT);
1788 }
1789 
1790 BasicAAWrapperPass::BasicAAWrapperPass() : FunctionPass(ID) {
1791   initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
1792 }
1793 
1794 char BasicAAWrapperPass::ID = 0;
1795 
1796 void BasicAAWrapperPass::anchor() {}
1797 
1798 INITIALIZE_PASS_BEGIN(BasicAAWrapperPass, "basic-aa",
1799                       "Basic Alias Analysis (stateless AA impl)", true, true)
1800 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1801 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1802 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1803 INITIALIZE_PASS_END(BasicAAWrapperPass, "basic-aa",
1804                     "Basic Alias Analysis (stateless AA impl)", true, true)
1805 
1806 FunctionPass *llvm::createBasicAAWrapperPass() {
1807   return new BasicAAWrapperPass();
1808 }
1809 
1810 bool BasicAAWrapperPass::runOnFunction(Function &F) {
1811   auto &ACT = getAnalysis<AssumptionCacheTracker>();
1812   auto &TLIWP = getAnalysis<TargetLibraryInfoWrapperPass>();
1813   auto &DTWP = getAnalysis<DominatorTreeWrapperPass>();
1814 
1815   Result.reset(new BasicAAResult(F.getParent()->getDataLayout(), F,
1816                                  TLIWP.getTLI(F), ACT.getAssumptionCache(F),
1817                                  &DTWP.getDomTree()));
1818 
1819   return false;
1820 }
1821 
1822 void BasicAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1823   AU.setPreservesAll();
1824   AU.addRequiredTransitive<AssumptionCacheTracker>();
1825   AU.addRequiredTransitive<DominatorTreeWrapperPass>();
1826   AU.addRequiredTransitive<TargetLibraryInfoWrapperPass>();
1827 }
1828 
1829 BasicAAResult llvm::createLegacyPMBasicAAResult(Pass &P, Function &F) {
1830   return BasicAAResult(
1831       F.getParent()->getDataLayout(), F,
1832       P.getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F),
1833       P.getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
1834 }
1835