BasicAliasAnalysis.cpp revision 4cccb87b4d766719cd8cdf98bed1d433d245adb0
1//===- BasicAliasAnalysis.cpp - Local Alias Analysis Impl -----------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines the default implementation of the Alias Analysis interface
11// that simply implements a few identities (two different globals cannot alias,
12// etc), but otherwise does no analysis.
13//
14//===----------------------------------------------------------------------===//
15
16#include "llvm/Analysis/AliasAnalysis.h"
17#include "llvm/Analysis/Passes.h"
18#include "llvm/Constants.h"
19#include "llvm/DerivedTypes.h"
20#include "llvm/Function.h"
21#include "llvm/GlobalAlias.h"
22#include "llvm/GlobalVariable.h"
23#include "llvm/Instructions.h"
24#include "llvm/IntrinsicInst.h"
25#include "llvm/LLVMContext.h"
26#include "llvm/Operator.h"
27#include "llvm/Pass.h"
28#include "llvm/Analysis/CaptureTracking.h"
29#include "llvm/Analysis/MemoryBuiltins.h"
30#include "llvm/Analysis/ValueTracking.h"
31#include "llvm/Target/TargetData.h"
32#include "llvm/ADT/SmallPtrSet.h"
33#include "llvm/ADT/SmallVector.h"
34#include "llvm/Support/ErrorHandling.h"
35#include "llvm/Support/GetElementPtrTypeIterator.h"
36#include <algorithm>
37using namespace llvm;
38
39//===----------------------------------------------------------------------===//
40// Useful predicates
41//===----------------------------------------------------------------------===//
42
43/// isKnownNonNull - Return true if we know that the specified value is never
44/// null.
45static bool isKnownNonNull(const Value *V) {
46  // Alloca never returns null, malloc might.
47  if (isa<AllocaInst>(V)) return true;
48
49  // A byval argument is never null.
50  if (const Argument *A = dyn_cast<Argument>(V))
51    return A->hasByValAttr();
52
53  // Global values are not null unless extern weak.
54  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
55    return !GV->hasExternalWeakLinkage();
56  return false;
57}
58
59/// isNonEscapingLocalObject - Return true if the pointer is to a function-local
60/// object that never escapes from the function.
61static bool isNonEscapingLocalObject(const Value *V) {
62  // If this is a local allocation, check to see if it escapes.
63  if (isa<AllocaInst>(V) || isNoAliasCall(V))
64    // Set StoreCaptures to True so that we can assume in our callers that the
65    // pointer is not the result of a load instruction. Currently
66    // PointerMayBeCaptured doesn't have any special analysis for the
67    // StoreCaptures=false case; if it did, our callers could be refined to be
68    // more precise.
69    return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
70
71  // If this is an argument that corresponds to a byval or noalias argument,
72  // then it has not escaped before entering the function.  Check if it escapes
73  // inside the function.
74  if (const Argument *A = dyn_cast<Argument>(V))
75    if (A->hasByValAttr() || A->hasNoAliasAttr()) {
76      // Don't bother analyzing arguments already known not to escape.
77      if (A->hasNoCaptureAttr())
78        return true;
79      return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
80    }
81  return false;
82}
83
84/// isEscapeSource - Return true if the pointer is one which would have
85/// been considered an escape by isNonEscapingLocalObject.
86static bool isEscapeSource(const Value *V) {
87  if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
88    return true;
89
90  // The load case works because isNonEscapingLocalObject considers all
91  // stores to be escapes (it passes true for the StoreCaptures argument
92  // to PointerMayBeCaptured).
93  if (isa<LoadInst>(V))
94    return true;
95
96  return false;
97}
98
99/// isObjectSmallerThan - Return true if we can prove that the object specified
100/// by V is smaller than Size.
101static bool isObjectSmallerThan(const Value *V, uint64_t Size,
102                                const TargetData &TD) {
103  const Type *AccessTy;
104  if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
105    AccessTy = GV->getType()->getElementType();
106  } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
107    if (!AI->isArrayAllocation())
108      AccessTy = AI->getType()->getElementType();
109    else
110      return false;
111  } else if (const CallInst* CI = extractMallocCall(V)) {
112    if (!isArrayMalloc(V, &TD))
113      // The size is the argument to the malloc call.
114      if (const ConstantInt* C = dyn_cast<ConstantInt>(CI->getArgOperand(0)))
115        return (C->getZExtValue() < Size);
116    return false;
117  } else if (const Argument *A = dyn_cast<Argument>(V)) {
118    if (A->hasByValAttr())
119      AccessTy = cast<PointerType>(A->getType())->getElementType();
120    else
121      return false;
122  } else {
123    return false;
124  }
125
126  if (AccessTy->isSized())
127    return TD.getTypeAllocSize(AccessTy) < Size;
128  return false;
129}
130
131//===----------------------------------------------------------------------===//
132// NoAA Pass
133//===----------------------------------------------------------------------===//
134
135namespace {
136  /// NoAA - This class implements the -no-aa pass, which always returns "I
137  /// don't know" for alias queries.  NoAA is unlike other alias analysis
138  /// implementations, in that it does not chain to a previous analysis.  As
139  /// such it doesn't follow many of the rules that other alias analyses must.
140  ///
141  struct NoAA : public ImmutablePass, public AliasAnalysis {
142    static char ID; // Class identification, replacement for typeinfo
143    NoAA() : ImmutablePass(ID) {
144      initializeNoAAPass(*PassRegistry::getPassRegistry());
145    }
146    explicit NoAA(char &PID) : ImmutablePass(PID) {}
147
148    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
149    }
150
151    virtual void initializePass() {
152      TD = getAnalysisIfAvailable<TargetData>();
153    }
154
155    virtual AliasResult alias(const Location &LocA, const Location &LocB) {
156      return MayAlias;
157    }
158
159    virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS) {
160      return UnknownModRefBehavior;
161    }
162    virtual ModRefBehavior getModRefBehavior(const Function *F) {
163      return UnknownModRefBehavior;
164    }
165
166    virtual bool pointsToConstantMemory(const Location &Loc) { return false; }
167    virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
168                                       const Location &Loc) {
169      return ModRef;
170    }
171    virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
172                                       ImmutableCallSite CS2) {
173      return ModRef;
174    }
175
176    virtual void deleteValue(Value *V) {}
177    virtual void copyValue(Value *From, Value *To) {}
178
179    /// getAdjustedAnalysisPointer - This method is used when a pass implements
180    /// an analysis interface through multiple inheritance.  If needed, it
181    /// should override this to adjust the this pointer as needed for the
182    /// specified pass info.
183    virtual void *getAdjustedAnalysisPointer(const void *ID) {
184      if (ID == &AliasAnalysis::ID)
185        return (AliasAnalysis*)this;
186      return this;
187    }
188  };
189}  // End of anonymous namespace
190
191// Register this pass...
192char NoAA::ID = 0;
193INITIALIZE_AG_PASS(NoAA, AliasAnalysis, "no-aa",
194                   "No Alias Analysis (always returns 'may' alias)",
195                   true, true, true)
196
197ImmutablePass *llvm::createNoAAPass() { return new NoAA(); }
198
199//===----------------------------------------------------------------------===//
200// GetElementPtr Instruction Decomposition and Analysis
201//===----------------------------------------------------------------------===//
202
203namespace {
204  enum ExtensionKind {
205    EK_NotExtended,
206    EK_SignExt,
207    EK_ZeroExt
208  };
209
210  struct VariableGEPIndex {
211    const Value *V;
212    ExtensionKind Extension;
213    int64_t Scale;
214  };
215}
216
217
218/// GetLinearExpression - Analyze the specified value as a linear expression:
219/// "A*V + B", where A and B are constant integers.  Return the scale and offset
220/// values as APInts and return V as a Value*, and return whether we looked
221/// through any sign or zero extends.  The incoming Value is known to have
222/// IntegerType and it may already be sign or zero extended.
223///
224/// Note that this looks through extends, so the high bits may not be
225/// represented in the result.
226static Value *GetLinearExpression(Value *V, APInt &Scale, APInt &Offset,
227                                  ExtensionKind &Extension,
228                                  const TargetData &TD, unsigned Depth) {
229  assert(V->getType()->isIntegerTy() && "Not an integer value");
230
231  // Limit our recursion depth.
232  if (Depth == 6) {
233    Scale = 1;
234    Offset = 0;
235    return V;
236  }
237
238  if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
239    if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
240      switch (BOp->getOpcode()) {
241      default: break;
242      case Instruction::Or:
243        // X|C == X+C if all the bits in C are unset in X.  Otherwise we can't
244        // analyze it.
245        if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), &TD))
246          break;
247        // FALL THROUGH.
248      case Instruction::Add:
249        V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
250                                TD, Depth+1);
251        Offset += RHSC->getValue();
252        return V;
253      case Instruction::Mul:
254        V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
255                                TD, Depth+1);
256        Offset *= RHSC->getValue();
257        Scale *= RHSC->getValue();
258        return V;
259      case Instruction::Shl:
260        V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
261                                TD, Depth+1);
262        Offset <<= RHSC->getValue().getLimitedValue();
263        Scale <<= RHSC->getValue().getLimitedValue();
264        return V;
265      }
266    }
267  }
268
269  // Since GEP indices are sign extended anyway, we don't care about the high
270  // bits of a sign or zero extended value - just scales and offsets.  The
271  // extensions have to be consistent though.
272  if ((isa<SExtInst>(V) && Extension != EK_ZeroExt) ||
273      (isa<ZExtInst>(V) && Extension != EK_SignExt)) {
274    Value *CastOp = cast<CastInst>(V)->getOperand(0);
275    unsigned OldWidth = Scale.getBitWidth();
276    unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
277    Scale.trunc(SmallWidth);
278    Offset.trunc(SmallWidth);
279    Extension = isa<SExtInst>(V) ? EK_SignExt : EK_ZeroExt;
280
281    Value *Result = GetLinearExpression(CastOp, Scale, Offset, Extension,
282                                        TD, Depth+1);
283    Scale.zext(OldWidth);
284    Offset.zext(OldWidth);
285
286    return Result;
287  }
288
289  Scale = 1;
290  Offset = 0;
291  return V;
292}
293
294/// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it
295/// into a base pointer with a constant offset and a number of scaled symbolic
296/// offsets.
297///
298/// The scaled symbolic offsets (represented by pairs of a Value* and a scale in
299/// the VarIndices vector) are Value*'s that are known to be scaled by the
300/// specified amount, but which may have other unrepresented high bits. As such,
301/// the gep cannot necessarily be reconstructed from its decomposed form.
302///
303/// When TargetData is around, this function is capable of analyzing everything
304/// that Value::getUnderlyingObject() can look through.  When not, it just looks
305/// through pointer casts.
306///
307static const Value *
308DecomposeGEPExpression(const Value *V, int64_t &BaseOffs,
309                       SmallVectorImpl<VariableGEPIndex> &VarIndices,
310                       const TargetData *TD) {
311  // Limit recursion depth to limit compile time in crazy cases.
312  unsigned MaxLookup = 6;
313
314  BaseOffs = 0;
315  do {
316    // See if this is a bitcast or GEP.
317    const Operator *Op = dyn_cast<Operator>(V);
318    if (Op == 0) {
319      // The only non-operator case we can handle are GlobalAliases.
320      if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
321        if (!GA->mayBeOverridden()) {
322          V = GA->getAliasee();
323          continue;
324        }
325      }
326      return V;
327    }
328
329    if (Op->getOpcode() == Instruction::BitCast) {
330      V = Op->getOperand(0);
331      continue;
332    }
333
334    const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
335    if (GEPOp == 0)
336      return V;
337
338    // Don't attempt to analyze GEPs over unsized objects.
339    if (!cast<PointerType>(GEPOp->getOperand(0)->getType())
340        ->getElementType()->isSized())
341      return V;
342
343    // If we are lacking TargetData information, we can't compute the offets of
344    // elements computed by GEPs.  However, we can handle bitcast equivalent
345    // GEPs.
346    if (TD == 0) {
347      if (!GEPOp->hasAllZeroIndices())
348        return V;
349      V = GEPOp->getOperand(0);
350      continue;
351    }
352
353    // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
354    gep_type_iterator GTI = gep_type_begin(GEPOp);
355    for (User::const_op_iterator I = GEPOp->op_begin()+1,
356         E = GEPOp->op_end(); I != E; ++I) {
357      Value *Index = *I;
358      // Compute the (potentially symbolic) offset in bytes for this index.
359      if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
360        // For a struct, add the member offset.
361        unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
362        if (FieldNo == 0) continue;
363
364        BaseOffs += TD->getStructLayout(STy)->getElementOffset(FieldNo);
365        continue;
366      }
367
368      // For an array/pointer, add the element offset, explicitly scaled.
369      if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
370        if (CIdx->isZero()) continue;
371        BaseOffs += TD->getTypeAllocSize(*GTI)*CIdx->getSExtValue();
372        continue;
373      }
374
375      uint64_t Scale = TD->getTypeAllocSize(*GTI);
376      ExtensionKind Extension = EK_NotExtended;
377
378      // If the integer type is smaller than the pointer size, it is implicitly
379      // sign extended to pointer size.
380      unsigned Width = cast<IntegerType>(Index->getType())->getBitWidth();
381      if (TD->getPointerSizeInBits() > Width)
382        Extension = EK_SignExt;
383
384      // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
385      APInt IndexScale(Width, 0), IndexOffset(Width, 0);
386      Index = GetLinearExpression(Index, IndexScale, IndexOffset, Extension,
387                                  *TD, 0);
388
389      // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
390      // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
391      BaseOffs += IndexOffset.getSExtValue()*Scale;
392      Scale *= IndexScale.getSExtValue();
393
394
395      // If we already had an occurrance of this index variable, merge this
396      // scale into it.  For example, we want to handle:
397      //   A[x][x] -> x*16 + x*4 -> x*20
398      // This also ensures that 'x' only appears in the index list once.
399      for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
400        if (VarIndices[i].V == Index &&
401            VarIndices[i].Extension == Extension) {
402          Scale += VarIndices[i].Scale;
403          VarIndices.erase(VarIndices.begin()+i);
404          break;
405        }
406      }
407
408      // Make sure that we have a scale that makes sense for this target's
409      // pointer size.
410      if (unsigned ShiftBits = 64-TD->getPointerSizeInBits()) {
411        Scale <<= ShiftBits;
412        Scale = (int64_t)Scale >> ShiftBits;
413      }
414
415      if (Scale) {
416        VariableGEPIndex Entry = {Index, Extension, Scale};
417        VarIndices.push_back(Entry);
418      }
419    }
420
421    // Analyze the base pointer next.
422    V = GEPOp->getOperand(0);
423  } while (--MaxLookup);
424
425  // If the chain of expressions is too deep, just return early.
426  return V;
427}
428
429/// GetIndexDifference - Dest and Src are the variable indices from two
430/// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base
431/// pointers.  Subtract the GEP2 indices from GEP1 to find the symbolic
432/// difference between the two pointers.
433static void GetIndexDifference(SmallVectorImpl<VariableGEPIndex> &Dest,
434                               const SmallVectorImpl<VariableGEPIndex> &Src) {
435  if (Src.empty()) return;
436
437  for (unsigned i = 0, e = Src.size(); i != e; ++i) {
438    const Value *V = Src[i].V;
439    ExtensionKind Extension = Src[i].Extension;
440    int64_t Scale = Src[i].Scale;
441
442    // Find V in Dest.  This is N^2, but pointer indices almost never have more
443    // than a few variable indexes.
444    for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
445      if (Dest[j].V != V || Dest[j].Extension != Extension) continue;
446
447      // If we found it, subtract off Scale V's from the entry in Dest.  If it
448      // goes to zero, remove the entry.
449      if (Dest[j].Scale != Scale)
450        Dest[j].Scale -= Scale;
451      else
452        Dest.erase(Dest.begin()+j);
453      Scale = 0;
454      break;
455    }
456
457    // If we didn't consume this entry, add it to the end of the Dest list.
458    if (Scale) {
459      VariableGEPIndex Entry = { V, Extension, -Scale };
460      Dest.push_back(Entry);
461    }
462  }
463}
464
465//===----------------------------------------------------------------------===//
466// BasicAliasAnalysis Pass
467//===----------------------------------------------------------------------===//
468
469#ifndef NDEBUG
470static const Function *getParent(const Value *V) {
471  if (const Instruction *inst = dyn_cast<Instruction>(V))
472    return inst->getParent()->getParent();
473
474  if (const Argument *arg = dyn_cast<Argument>(V))
475    return arg->getParent();
476
477  return NULL;
478}
479
480static bool notDifferentParent(const Value *O1, const Value *O2) {
481
482  const Function *F1 = getParent(O1);
483  const Function *F2 = getParent(O2);
484
485  return !F1 || !F2 || F1 == F2;
486}
487#endif
488
489namespace {
490  /// BasicAliasAnalysis - This is the default alias analysis implementation.
491  /// Because it doesn't chain to a previous alias analysis (like -no-aa), it
492  /// derives from the NoAA class.
493  struct BasicAliasAnalysis : public NoAA {
494    static char ID; // Class identification, replacement for typeinfo
495    BasicAliasAnalysis() : NoAA(ID) {
496      initializeBasicAliasAnalysisPass(*PassRegistry::getPassRegistry());
497    }
498
499    virtual void initializePass() {
500      InitializeAliasAnalysis(this);
501    }
502
503    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
504      AU.addRequired<AliasAnalysis>();
505    }
506
507    virtual AliasResult alias(const Location &LocA,
508                              const Location &LocB) {
509      assert(Visited.empty() && "Visited must be cleared after use!");
510      assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
511             "BasicAliasAnalysis doesn't support interprocedural queries.");
512      AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.TBAATag,
513                                     LocB.Ptr, LocB.Size, LocB.TBAATag);
514      Visited.clear();
515      return Alias;
516    }
517
518    virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
519                                       const Location &Loc);
520
521    virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
522                                       ImmutableCallSite CS2) {
523      // The AliasAnalysis base class has some smarts, lets use them.
524      return AliasAnalysis::getModRefInfo(CS1, CS2);
525    }
526
527    /// pointsToConstantMemory - Chase pointers until we find a (constant
528    /// global) or not.
529    virtual bool pointsToConstantMemory(const Location &Loc);
530
531    /// getModRefBehavior - Return the behavior when calling the given
532    /// call site.
533    virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
534
535    /// getModRefBehavior - Return the behavior when calling the given function.
536    /// For use when the call site is not known.
537    virtual ModRefBehavior getModRefBehavior(const Function *F);
538
539    /// getAdjustedAnalysisPointer - This method is used when a pass implements
540    /// an analysis interface through multiple inheritance.  If needed, it
541    /// should override this to adjust the this pointer as needed for the
542    /// specified pass info.
543    virtual void *getAdjustedAnalysisPointer(const void *ID) {
544      if (ID == &AliasAnalysis::ID)
545        return (AliasAnalysis*)this;
546      return this;
547    }
548
549  private:
550    // Visited - Track instructions visited by a aliasPHI, aliasSelect(), and aliasGEP().
551    SmallPtrSet<const Value*, 16> Visited;
552
553    // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP
554    // instruction against another.
555    AliasResult aliasGEP(const GEPOperator *V1, uint64_t V1Size,
556                         const Value *V2, uint64_t V2Size,
557                         const MDNode *V2TBAAInfo,
558                         const Value *UnderlyingV1, const Value *UnderlyingV2);
559
560    // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI
561    // instruction against another.
562    AliasResult aliasPHI(const PHINode *PN, uint64_t PNSize,
563                         const MDNode *PNTBAAInfo,
564                         const Value *V2, uint64_t V2Size,
565                         const MDNode *V2TBAAInfo);
566
567    /// aliasSelect - Disambiguate a Select instruction against another value.
568    AliasResult aliasSelect(const SelectInst *SI, uint64_t SISize,
569                            const MDNode *SITBAAInfo,
570                            const Value *V2, uint64_t V2Size,
571                            const MDNode *V2TBAAInfo);
572
573    AliasResult aliasCheck(const Value *V1, uint64_t V1Size,
574                           const MDNode *V1TBAATag,
575                           const Value *V2, uint64_t V2Size,
576                           const MDNode *V2TBAATag);
577  };
578}  // End of anonymous namespace
579
580// Register this pass...
581char BasicAliasAnalysis::ID = 0;
582INITIALIZE_AG_PASS(BasicAliasAnalysis, AliasAnalysis, "basicaa",
583                   "Basic Alias Analysis (default AA impl)",
584                   false, true, false)
585
586ImmutablePass *llvm::createBasicAliasAnalysisPass() {
587  return new BasicAliasAnalysis();
588}
589
590
591/// pointsToConstantMemory - Chase pointers until we find a (constant
592/// global) or not.
593bool BasicAliasAnalysis::pointsToConstantMemory(const Location &Loc) {
594  if (const GlobalVariable *GV =
595        dyn_cast<GlobalVariable>(Loc.Ptr->getUnderlyingObject()))
596    // Note: this doesn't require GV to be "ODR" because it isn't legal for a
597    // global to be marked constant in some modules and non-constant in others.
598    // GV may even be a declaration, not a definition.
599    return GV->isConstant();
600
601  return AliasAnalysis::pointsToConstantMemory(Loc);
602}
603
604/// getModRefBehavior - Return the behavior when calling the given call site.
605AliasAnalysis::ModRefBehavior
606BasicAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
607  if (CS.doesNotAccessMemory())
608    // Can't do better than this.
609    return DoesNotAccessMemory;
610
611  ModRefBehavior Min = UnknownModRefBehavior;
612
613  // If the callsite knows it only reads memory, don't return worse
614  // than that.
615  if (CS.onlyReadsMemory())
616    Min = OnlyReadsMemory;
617
618  // The AliasAnalysis base class has some smarts, lets use them.
619  return std::min(AliasAnalysis::getModRefBehavior(CS), Min);
620}
621
622/// getModRefBehavior - Return the behavior when calling the given function.
623/// For use when the call site is not known.
624AliasAnalysis::ModRefBehavior
625BasicAliasAnalysis::getModRefBehavior(const Function *F) {
626  if (F->doesNotAccessMemory())
627    // Can't do better than this.
628    return DoesNotAccessMemory;
629  if (F->onlyReadsMemory())
630    return OnlyReadsMemory;
631  if (unsigned id = F->getIntrinsicID())
632    return getIntrinsicModRefBehavior(id);
633
634  return AliasAnalysis::getModRefBehavior(F);
635}
636
637/// getModRefInfo - Check to see if the specified callsite can clobber the
638/// specified memory object.  Since we only look at local properties of this
639/// function, we really can't say much about this query.  We do, however, use
640/// simple "address taken" analysis on local objects.
641AliasAnalysis::ModRefResult
642BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS,
643                                  const Location &Loc) {
644  assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
645         "AliasAnalysis query involving multiple functions!");
646
647  const Value *Object = Loc.Ptr->getUnderlyingObject();
648
649  // If this is a tail call and Loc.Ptr points to a stack location, we know that
650  // the tail call cannot access or modify the local stack.
651  // We cannot exclude byval arguments here; these belong to the caller of
652  // the current function not to the current function, and a tail callee
653  // may reference them.
654  if (isa<AllocaInst>(Object))
655    if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
656      if (CI->isTailCall())
657        return NoModRef;
658
659  // If the pointer is to a locally allocated object that does not escape,
660  // then the call can not mod/ref the pointer unless the call takes the pointer
661  // as an argument, and itself doesn't capture it.
662  if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
663      isNonEscapingLocalObject(Object)) {
664    bool PassedAsArg = false;
665    unsigned ArgNo = 0;
666    for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
667         CI != CE; ++CI, ++ArgNo) {
668      // Only look at the no-capture pointer arguments.
669      if (!(*CI)->getType()->isPointerTy() ||
670          !CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))
671        continue;
672
673      // If this is a no-capture pointer argument, see if we can tell that it
674      // is impossible to alias the pointer we're checking.  If not, we have to
675      // assume that the call could touch the pointer, even though it doesn't
676      // escape.
677      if (!isNoAlias(Location(cast<Value>(CI)), Loc)) {
678        PassedAsArg = true;
679        break;
680      }
681    }
682
683    if (!PassedAsArg)
684      return NoModRef;
685  }
686
687  // Finally, handle specific knowledge of intrinsics.
688  const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
689  if (II != 0)
690    switch (II->getIntrinsicID()) {
691    default: break;
692    case Intrinsic::memcpy:
693    case Intrinsic::memmove: {
694      uint64_t Len = UnknownSize;
695      if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2)))
696        Len = LenCI->getZExtValue();
697      Value *Dest = II->getArgOperand(0);
698      Value *Src = II->getArgOperand(1);
699      if (isNoAlias(Location(Dest, Len), Loc)) {
700        if (isNoAlias(Location(Src, Len), Loc))
701          return NoModRef;
702        return Ref;
703      }
704      break;
705    }
706    case Intrinsic::memset:
707      // Since memset is 'accesses arguments' only, the AliasAnalysis base class
708      // will handle it for the variable length case.
709      if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
710        uint64_t Len = LenCI->getZExtValue();
711        Value *Dest = II->getArgOperand(0);
712        if (isNoAlias(Location(Dest, Len), Loc))
713          return NoModRef;
714      }
715      break;
716    case Intrinsic::atomic_cmp_swap:
717    case Intrinsic::atomic_swap:
718    case Intrinsic::atomic_load_add:
719    case Intrinsic::atomic_load_sub:
720    case Intrinsic::atomic_load_and:
721    case Intrinsic::atomic_load_nand:
722    case Intrinsic::atomic_load_or:
723    case Intrinsic::atomic_load_xor:
724    case Intrinsic::atomic_load_max:
725    case Intrinsic::atomic_load_min:
726    case Intrinsic::atomic_load_umax:
727    case Intrinsic::atomic_load_umin:
728      if (TD) {
729        Value *Op1 = II->getArgOperand(0);
730        uint64_t Op1Size = TD->getTypeStoreSize(Op1->getType());
731        MDNode *Tag = II->getMetadata(LLVMContext::MD_tbaa);
732        if (isNoAlias(Location(Op1, Op1Size, Tag), Loc))
733          return NoModRef;
734      }
735      break;
736    case Intrinsic::lifetime_start:
737    case Intrinsic::lifetime_end:
738    case Intrinsic::invariant_start: {
739      uint64_t PtrSize =
740        cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
741      if (isNoAlias(Location(II->getArgOperand(1),
742                             PtrSize,
743                             II->getMetadata(LLVMContext::MD_tbaa)),
744                    Loc))
745        return NoModRef;
746      break;
747    }
748    case Intrinsic::invariant_end: {
749      uint64_t PtrSize =
750        cast<ConstantInt>(II->getArgOperand(1))->getZExtValue();
751      if (isNoAlias(Location(II->getArgOperand(2),
752                             PtrSize,
753                             II->getMetadata(LLVMContext::MD_tbaa)),
754                    Loc))
755        return NoModRef;
756      break;
757    }
758    }
759
760  // The AliasAnalysis base class has some smarts, lets use them.
761  return AliasAnalysis::getModRefInfo(CS, Loc);
762}
763
764/// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
765/// against another pointer.  We know that V1 is a GEP, but we don't know
766/// anything about V2.  UnderlyingV1 is GEP1->getUnderlyingObject(),
767/// UnderlyingV2 is the same for V2.
768///
769AliasAnalysis::AliasResult
770BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
771                             const Value *V2, uint64_t V2Size,
772                             const MDNode *V2TBAAInfo,
773                             const Value *UnderlyingV1,
774                             const Value *UnderlyingV2) {
775  // If this GEP has been visited before, we're on a use-def cycle.
776  // Such cycles are only valid when PHI nodes are involved or in unreachable
777  // code. The visitPHI function catches cycles containing PHIs, but there
778  // could still be a cycle without PHIs in unreachable code.
779  if (!Visited.insert(GEP1))
780    return MayAlias;
781
782  int64_t GEP1BaseOffset;
783  SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
784
785  // If we have two gep instructions with must-alias'ing base pointers, figure
786  // out if the indexes to the GEP tell us anything about the derived pointer.
787  if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
788    // Do the base pointers alias?
789    AliasResult BaseAlias = aliasCheck(UnderlyingV1, UnknownSize, 0,
790                                       UnderlyingV2, UnknownSize, 0);
791
792    // If we get a No or May, then return it immediately, no amount of analysis
793    // will improve this situation.
794    if (BaseAlias != MustAlias) return BaseAlias;
795
796    // Otherwise, we have a MustAlias.  Since the base pointers alias each other
797    // exactly, see if the computed offset from the common pointer tells us
798    // about the relation of the resulting pointer.
799    const Value *GEP1BasePtr =
800      DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
801
802    int64_t GEP2BaseOffset;
803    SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
804    const Value *GEP2BasePtr =
805      DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD);
806
807    // If DecomposeGEPExpression isn't able to look all the way through the
808    // addressing operation, we must not have TD and this is too complex for us
809    // to handle without it.
810    if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
811      assert(TD == 0 &&
812             "DecomposeGEPExpression and getUnderlyingObject disagree!");
813      return MayAlias;
814    }
815
816    // Subtract the GEP2 pointer from the GEP1 pointer to find out their
817    // symbolic difference.
818    GEP1BaseOffset -= GEP2BaseOffset;
819    GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
820
821  } else {
822    // Check to see if these two pointers are related by the getelementptr
823    // instruction.  If one pointer is a GEP with a non-zero index of the other
824    // pointer, we know they cannot alias.
825
826    // If both accesses are unknown size, we can't do anything useful here.
827    if (V1Size == UnknownSize && V2Size == UnknownSize)
828      return MayAlias;
829
830    AliasResult R = aliasCheck(UnderlyingV1, UnknownSize, 0,
831                               V2, V2Size, V2TBAAInfo);
832    if (R != MustAlias)
833      // If V2 may alias GEP base pointer, conservatively returns MayAlias.
834      // If V2 is known not to alias GEP base pointer, then the two values
835      // cannot alias per GEP semantics: "A pointer value formed from a
836      // getelementptr instruction is associated with the addresses associated
837      // with the first operand of the getelementptr".
838      return R;
839
840    const Value *GEP1BasePtr =
841      DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
842
843    // If DecomposeGEPExpression isn't able to look all the way through the
844    // addressing operation, we must not have TD and this is too complex for us
845    // to handle without it.
846    if (GEP1BasePtr != UnderlyingV1) {
847      assert(TD == 0 &&
848             "DecomposeGEPExpression and getUnderlyingObject disagree!");
849      return MayAlias;
850    }
851  }
852
853  // In the two GEP Case, if there is no difference in the offsets of the
854  // computed pointers, the resultant pointers are a must alias.  This
855  // hapens when we have two lexically identical GEP's (for example).
856  //
857  // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
858  // must aliases the GEP, the end result is a must alias also.
859  if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
860    return MustAlias;
861
862  // If we have a known constant offset, see if this offset is larger than the
863  // access size being queried.  If so, and if no variable indices can remove
864  // pieces of this constant, then we know we have a no-alias.  For example,
865  //   &A[100] != &A.
866
867  // In order to handle cases like &A[100][i] where i is an out of range
868  // subscript, we have to ignore all constant offset pieces that are a multiple
869  // of a scaled index.  Do this by removing constant offsets that are a
870  // multiple of any of our variable indices.  This allows us to transform
871  // things like &A[i][1] because i has a stride of (e.g.) 8 bytes but the 1
872  // provides an offset of 4 bytes (assuming a <= 4 byte access).
873  for (unsigned i = 0, e = GEP1VariableIndices.size();
874       i != e && GEP1BaseOffset;++i)
875    if (int64_t RemovedOffset = GEP1BaseOffset/GEP1VariableIndices[i].Scale)
876      GEP1BaseOffset -= RemovedOffset*GEP1VariableIndices[i].Scale;
877
878  // If our known offset is bigger than the access size, we know we don't have
879  // an alias.
880  if (GEP1BaseOffset) {
881    if (GEP1BaseOffset >= 0 ?
882        (V2Size != UnknownSize && (uint64_t)GEP1BaseOffset >= V2Size) :
883        (V1Size != UnknownSize && -(uint64_t)GEP1BaseOffset >= V1Size &&
884         GEP1BaseOffset != INT64_MIN))
885      return NoAlias;
886  }
887
888  return MayAlias;
889}
890
891/// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select
892/// instruction against another.
893AliasAnalysis::AliasResult
894BasicAliasAnalysis::aliasSelect(const SelectInst *SI, uint64_t SISize,
895                                const MDNode *SITBAAInfo,
896                                const Value *V2, uint64_t V2Size,
897                                const MDNode *V2TBAAInfo) {
898  // If this select has been visited before, we're on a use-def cycle.
899  // Such cycles are only valid when PHI nodes are involved or in unreachable
900  // code. The visitPHI function catches cycles containing PHIs, but there
901  // could still be a cycle without PHIs in unreachable code.
902  if (!Visited.insert(SI))
903    return MayAlias;
904
905  // If the values are Selects with the same condition, we can do a more precise
906  // check: just check for aliases between the values on corresponding arms.
907  if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
908    if (SI->getCondition() == SI2->getCondition()) {
909      AliasResult Alias =
910        aliasCheck(SI->getTrueValue(), SISize, SITBAAInfo,
911                   SI2->getTrueValue(), V2Size, V2TBAAInfo);
912      if (Alias == MayAlias)
913        return MayAlias;
914      AliasResult ThisAlias =
915        aliasCheck(SI->getFalseValue(), SISize, SITBAAInfo,
916                   SI2->getFalseValue(), V2Size, V2TBAAInfo);
917      if (ThisAlias != Alias)
918        return MayAlias;
919      return Alias;
920    }
921
922  // If both arms of the Select node NoAlias or MustAlias V2, then returns
923  // NoAlias / MustAlias. Otherwise, returns MayAlias.
924  AliasResult Alias =
925    aliasCheck(V2, V2Size, V2TBAAInfo, SI->getTrueValue(), SISize, SITBAAInfo);
926  if (Alias == MayAlias)
927    return MayAlias;
928
929  // If V2 is visited, the recursive case will have been caught in the
930  // above aliasCheck call, so these subsequent calls to aliasCheck
931  // don't need to assume that V2 is being visited recursively.
932  Visited.erase(V2);
933
934  AliasResult ThisAlias =
935    aliasCheck(V2, V2Size, V2TBAAInfo, SI->getFalseValue(), SISize, SITBAAInfo);
936  if (ThisAlias != Alias)
937    return MayAlias;
938  return Alias;
939}
940
941// aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
942// against another.
943AliasAnalysis::AliasResult
944BasicAliasAnalysis::aliasPHI(const PHINode *PN, uint64_t PNSize,
945                             const MDNode *PNTBAAInfo,
946                             const Value *V2, uint64_t V2Size,
947                             const MDNode *V2TBAAInfo) {
948  // The PHI node has already been visited, avoid recursion any further.
949  if (!Visited.insert(PN))
950    return MayAlias;
951
952  // If the values are PHIs in the same block, we can do a more precise
953  // as well as efficient check: just check for aliases between the values
954  // on corresponding edges.
955  if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
956    if (PN2->getParent() == PN->getParent()) {
957      AliasResult Alias =
958        aliasCheck(PN->getIncomingValue(0), PNSize, PNTBAAInfo,
959                   PN2->getIncomingValueForBlock(PN->getIncomingBlock(0)),
960                   V2Size, V2TBAAInfo);
961      if (Alias == MayAlias)
962        return MayAlias;
963      for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
964        AliasResult ThisAlias =
965          aliasCheck(PN->getIncomingValue(i), PNSize, PNTBAAInfo,
966                     PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
967                     V2Size, V2TBAAInfo);
968        if (ThisAlias != Alias)
969          return MayAlias;
970      }
971      return Alias;
972    }
973
974  SmallPtrSet<Value*, 4> UniqueSrc;
975  SmallVector<Value*, 4> V1Srcs;
976  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
977    Value *PV1 = PN->getIncomingValue(i);
978    if (isa<PHINode>(PV1))
979      // If any of the source itself is a PHI, return MayAlias conservatively
980      // to avoid compile time explosion. The worst possible case is if both
981      // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
982      // and 'n' are the number of PHI sources.
983      return MayAlias;
984    if (UniqueSrc.insert(PV1))
985      V1Srcs.push_back(PV1);
986  }
987
988  AliasResult Alias = aliasCheck(V2, V2Size, V2TBAAInfo,
989                                 V1Srcs[0], PNSize, PNTBAAInfo);
990  // Early exit if the check of the first PHI source against V2 is MayAlias.
991  // Other results are not possible.
992  if (Alias == MayAlias)
993    return MayAlias;
994
995  // If all sources of the PHI node NoAlias or MustAlias V2, then returns
996  // NoAlias / MustAlias. Otherwise, returns MayAlias.
997  for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
998    Value *V = V1Srcs[i];
999
1000    // If V2 is visited, the recursive case will have been caught in the
1001    // above aliasCheck call, so these subsequent calls to aliasCheck
1002    // don't need to assume that V2 is being visited recursively.
1003    Visited.erase(V2);
1004
1005    AliasResult ThisAlias = aliasCheck(V2, V2Size, V2TBAAInfo,
1006                                       V, PNSize, PNTBAAInfo);
1007    if (ThisAlias != Alias || ThisAlias == MayAlias)
1008      return MayAlias;
1009  }
1010
1011  return Alias;
1012}
1013
1014// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
1015// such as array references.
1016//
1017AliasAnalysis::AliasResult
1018BasicAliasAnalysis::aliasCheck(const Value *V1, uint64_t V1Size,
1019                               const MDNode *V1TBAAInfo,
1020                               const Value *V2, uint64_t V2Size,
1021                               const MDNode *V2TBAAInfo) {
1022  // If either of the memory references is empty, it doesn't matter what the
1023  // pointer values are.
1024  if (V1Size == 0 || V2Size == 0)
1025    return NoAlias;
1026
1027  // Strip off any casts if they exist.
1028  V1 = V1->stripPointerCasts();
1029  V2 = V2->stripPointerCasts();
1030
1031  // Are we checking for alias of the same value?
1032  if (V1 == V2) return MustAlias;
1033
1034  if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1035    return NoAlias;  // Scalars cannot alias each other
1036
1037  // Figure out what objects these things are pointing to if we can.
1038  const Value *O1 = V1->getUnderlyingObject();
1039  const Value *O2 = V2->getUnderlyingObject();
1040
1041  // Null values in the default address space don't point to any object, so they
1042  // don't alias any other pointer.
1043  if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1044    if (CPN->getType()->getAddressSpace() == 0)
1045      return NoAlias;
1046  if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1047    if (CPN->getType()->getAddressSpace() == 0)
1048      return NoAlias;
1049
1050  if (O1 != O2) {
1051    // If V1/V2 point to two different objects we know that we have no alias.
1052    if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1053      return NoAlias;
1054
1055    // Constant pointers can't alias with non-const isIdentifiedObject objects.
1056    if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1057        (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1058      return NoAlias;
1059
1060    // Arguments can't alias with local allocations or noalias calls
1061    // in the same function.
1062    if (((isa<Argument>(O1) && (isa<AllocaInst>(O2) || isNoAliasCall(O2))) ||
1063         (isa<Argument>(O2) && (isa<AllocaInst>(O1) || isNoAliasCall(O1)))))
1064      return NoAlias;
1065
1066    // Most objects can't alias null.
1067    if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
1068        (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
1069      return NoAlias;
1070
1071    // If one pointer is the result of a call/invoke or load and the other is a
1072    // non-escaping local object within the same function, then we know the
1073    // object couldn't escape to a point where the call could return it.
1074    //
1075    // Note that if the pointers are in different functions, there are a
1076    // variety of complications. A call with a nocapture argument may still
1077    // temporary store the nocapture argument's value in a temporary memory
1078    // location if that memory location doesn't escape. Or it may pass a
1079    // nocapture value to other functions as long as they don't capture it.
1080    if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1081      return NoAlias;
1082    if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1083      return NoAlias;
1084  }
1085
1086  // If the size of one access is larger than the entire object on the other
1087  // side, then we know such behavior is undefined and can assume no alias.
1088  if (TD)
1089    if ((V1Size != UnknownSize && isObjectSmallerThan(O2, V1Size, *TD)) ||
1090        (V2Size != UnknownSize && isObjectSmallerThan(O1, V2Size, *TD)))
1091      return NoAlias;
1092
1093  // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1094  // GEP can't simplify, we don't even look at the PHI cases.
1095  if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1096    std::swap(V1, V2);
1097    std::swap(V1Size, V2Size);
1098    std::swap(O1, O2);
1099  }
1100  if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1101    AliasResult Result = aliasGEP(GV1, V1Size, V2, V2Size, V2TBAAInfo, O1, O2);
1102    if (Result != MayAlias) return Result;
1103  }
1104
1105  if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1106    std::swap(V1, V2);
1107    std::swap(V1Size, V2Size);
1108  }
1109  if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1110    AliasResult Result = aliasPHI(PN, V1Size, V1TBAAInfo,
1111                                  V2, V2Size, V2TBAAInfo);
1112    if (Result != MayAlias) return Result;
1113  }
1114
1115  if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1116    std::swap(V1, V2);
1117    std::swap(V1Size, V2Size);
1118  }
1119  if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1120    AliasResult Result = aliasSelect(S1, V1Size, V1TBAAInfo,
1121                                     V2, V2Size, V2TBAAInfo);
1122    if (Result != MayAlias) return Result;
1123  }
1124
1125  return AliasAnalysis::alias(Location(V1, V1Size, V1TBAAInfo),
1126                              Location(V2, V2Size, V2TBAAInfo));
1127}
1128
1129// Make sure that anything that uses AliasAnalysis pulls in this file.
1130DEFINING_FILE_FOR(BasicAliasAnalysis)
1131