BasicAliasAnalysis.cpp revision 74c996cbd12aa39e75990e3e694549bb1be90e4d
1//===- BasicAliasAnalysis.cpp - Stateless 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 primary stateless implementation of the
11// Alias Analysis interface that implements identities (two different
12// globals cannot alias, etc), but does no stateful analysis.
13//
14//===----------------------------------------------------------------------===//
15
16#include "llvm/Analysis/Passes.h"
17#include "llvm/ADT/SmallPtrSet.h"
18#include "llvm/ADT/SmallVector.h"
19#include "llvm/Analysis/AliasAnalysis.h"
20#include "llvm/Analysis/CaptureTracking.h"
21#include "llvm/Analysis/InstructionSimplify.h"
22#include "llvm/Analysis/MemoryBuiltins.h"
23#include "llvm/Analysis/ValueTracking.h"
24#include "llvm/IR/Constants.h"
25#include "llvm/IR/DataLayout.h"
26#include "llvm/IR/DerivedTypes.h"
27#include "llvm/IR/Function.h"
28#include "llvm/IR/GlobalAlias.h"
29#include "llvm/IR/GlobalVariable.h"
30#include "llvm/IR/Instructions.h"
31#include "llvm/IR/IntrinsicInst.h"
32#include "llvm/IR/LLVMContext.h"
33#include "llvm/IR/Operator.h"
34#include "llvm/Pass.h"
35#include "llvm/Support/ErrorHandling.h"
36#include "llvm/Support/GetElementPtrTypeIterator.h"
37#include "llvm/Target/TargetLibraryInfo.h"
38#include <algorithm>
39using namespace llvm;
40
41//===----------------------------------------------------------------------===//
42// Useful predicates
43//===----------------------------------------------------------------------===//
44
45/// isNonEscapingLocalObject - Return true if the pointer is to a function-local
46/// object that never escapes from the function.
47static bool isNonEscapingLocalObject(const Value *V) {
48  // If this is a local allocation, check to see if it escapes.
49  if (isa<AllocaInst>(V) || isNoAliasCall(V))
50    // Set StoreCaptures to True so that we can assume in our callers that the
51    // pointer is not the result of a load instruction. Currently
52    // PointerMayBeCaptured doesn't have any special analysis for the
53    // StoreCaptures=false case; if it did, our callers could be refined to be
54    // more precise.
55    return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
56
57  // If this is an argument that corresponds to a byval or noalias argument,
58  // then it has not escaped before entering the function.  Check if it escapes
59  // inside the function.
60  if (const Argument *A = dyn_cast<Argument>(V))
61    if (A->hasByValAttr() || A->hasNoAliasAttr())
62      // Note even if the argument is marked nocapture we still need to check
63      // for copies made inside the function. The nocapture attribute only
64      // specifies that there are no copies made that outlive the function.
65      return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
66
67  return false;
68}
69
70/// isEscapeSource - Return true if the pointer is one which would have
71/// been considered an escape by isNonEscapingLocalObject.
72static bool isEscapeSource(const Value *V) {
73  if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
74    return true;
75
76  // The load case works because isNonEscapingLocalObject considers all
77  // stores to be escapes (it passes true for the StoreCaptures argument
78  // to PointerMayBeCaptured).
79  if (isa<LoadInst>(V))
80    return true;
81
82  return false;
83}
84
85/// getObjectSize - Return the size of the object specified by V, or
86/// UnknownSize if unknown.
87static uint64_t getObjectSize(const Value *V, const DataLayout &TD,
88                              const TargetLibraryInfo &TLI,
89                              bool RoundToAlign = false) {
90  uint64_t Size;
91  if (getObjectSize(V, Size, &TD, &TLI, RoundToAlign))
92    return Size;
93  return AliasAnalysis::UnknownSize;
94}
95
96/// isObjectSmallerThan - Return true if we can prove that the object specified
97/// by V is smaller than Size.
98static bool isObjectSmallerThan(const Value *V, uint64_t Size,
99                                const DataLayout &TD,
100                                const TargetLibraryInfo &TLI) {
101  // Note that the meanings of the "object" are slightly different in the
102  // following contexts:
103  //    c1: llvm::getObjectSize()
104  //    c2: llvm.objectsize() intrinsic
105  //    c3: isObjectSmallerThan()
106  // c1 and c2 share the same meaning; however, the meaning of "object" in c3
107  // refers to the "entire object".
108  //
109  //  Consider this example:
110  //     char *p = (char*)malloc(100)
111  //     char *q = p+80;
112  //
113  //  In the context of c1 and c2, the "object" pointed by q refers to the
114  // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
115  //
116  //  However, in the context of c3, the "object" refers to the chunk of memory
117  // being allocated. So, the "object" has 100 bytes, and q points to the middle
118  // the "object". In case q is passed to isObjectSmallerThan() as the 1st
119  // parameter, before the llvm::getObjectSize() is called to get the size of
120  // entire object, we should:
121  //    - either rewind the pointer q to the base-address of the object in
122  //      question (in this case rewind to p), or
123  //    - just give up. It is up to caller to make sure the pointer is pointing
124  //      to the base address the object.
125  //
126  // We go for 2nd option for simplicity.
127  if (!isIdentifiedObject(V))
128    return false;
129
130  // This function needs to use the aligned object size because we allow
131  // reads a bit past the end given sufficient alignment.
132  uint64_t ObjectSize = getObjectSize(V, TD, TLI, /*RoundToAlign*/true);
133
134  return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize < Size;
135}
136
137/// isObjectSize - Return true if we can prove that the object specified
138/// by V has size Size.
139static bool isObjectSize(const Value *V, uint64_t Size,
140                         const DataLayout &TD, const TargetLibraryInfo &TLI) {
141  uint64_t ObjectSize = getObjectSize(V, TD, TLI);
142  return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize == Size;
143}
144
145/// isIdentifiedFunctionLocal - Return true if V is umabigously identified
146/// at the function-level. Different IdentifiedFunctionLocals can't alias.
147/// Further, an IdentifiedFunctionLocal can not alias with any function
148/// arguments other than itself, which is not neccessarily true for
149/// IdentifiedObjects.
150static bool isIdentifiedFunctionLocal(const Value *V)
151{
152  return isa<AllocaInst>(V) || isNoAliasCall(V) || isNoAliasArgument(V);
153}
154
155
156//===----------------------------------------------------------------------===//
157// GetElementPtr Instruction Decomposition and Analysis
158//===----------------------------------------------------------------------===//
159
160namespace {
161  enum ExtensionKind {
162    EK_NotExtended,
163    EK_SignExt,
164    EK_ZeroExt
165  };
166
167  struct VariableGEPIndex {
168    const Value *V;
169    ExtensionKind Extension;
170    int64_t Scale;
171
172    bool operator==(const VariableGEPIndex &Other) const {
173      return V == Other.V && Extension == Other.Extension &&
174        Scale == Other.Scale;
175    }
176
177    bool operator!=(const VariableGEPIndex &Other) const {
178      return !operator==(Other);
179    }
180  };
181}
182
183
184/// GetLinearExpression - Analyze the specified value as a linear expression:
185/// "A*V + B", where A and B are constant integers.  Return the scale and offset
186/// values as APInts and return V as a Value*, and return whether we looked
187/// through any sign or zero extends.  The incoming Value is known to have
188/// IntegerType and it may already be sign or zero extended.
189///
190/// Note that this looks through extends, so the high bits may not be
191/// represented in the result.
192static Value *GetLinearExpression(Value *V, APInt &Scale, APInt &Offset,
193                                  ExtensionKind &Extension,
194                                  const DataLayout &TD, unsigned Depth) {
195  assert(V->getType()->isIntegerTy() && "Not an integer value");
196
197  // Limit our recursion depth.
198  if (Depth == 6) {
199    Scale = 1;
200    Offset = 0;
201    return V;
202  }
203
204  if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
205    if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
206      switch (BOp->getOpcode()) {
207      default: break;
208      case Instruction::Or:
209        // X|C == X+C if all the bits in C are unset in X.  Otherwise we can't
210        // analyze it.
211        if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), &TD))
212          break;
213        // FALL THROUGH.
214      case Instruction::Add:
215        V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
216                                TD, Depth+1);
217        Offset += RHSC->getValue();
218        return V;
219      case Instruction::Mul:
220        V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
221                                TD, Depth+1);
222        Offset *= RHSC->getValue();
223        Scale *= RHSC->getValue();
224        return V;
225      case Instruction::Shl:
226        V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
227                                TD, Depth+1);
228        Offset <<= RHSC->getValue().getLimitedValue();
229        Scale <<= RHSC->getValue().getLimitedValue();
230        return V;
231      }
232    }
233  }
234
235  // Since GEP indices are sign extended anyway, we don't care about the high
236  // bits of a sign or zero extended value - just scales and offsets.  The
237  // extensions have to be consistent though.
238  if ((isa<SExtInst>(V) && Extension != EK_ZeroExt) ||
239      (isa<ZExtInst>(V) && Extension != EK_SignExt)) {
240    Value *CastOp = cast<CastInst>(V)->getOperand(0);
241    unsigned OldWidth = Scale.getBitWidth();
242    unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
243    Scale = Scale.trunc(SmallWidth);
244    Offset = Offset.trunc(SmallWidth);
245    Extension = isa<SExtInst>(V) ? EK_SignExt : EK_ZeroExt;
246
247    Value *Result = GetLinearExpression(CastOp, Scale, Offset, Extension,
248                                        TD, Depth+1);
249    Scale = Scale.zext(OldWidth);
250    Offset = Offset.zext(OldWidth);
251
252    return Result;
253  }
254
255  Scale = 1;
256  Offset = 0;
257  return V;
258}
259
260/// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it
261/// into a base pointer with a constant offset and a number of scaled symbolic
262/// offsets.
263///
264/// The scaled symbolic offsets (represented by pairs of a Value* and a scale in
265/// the VarIndices vector) are Value*'s that are known to be scaled by the
266/// specified amount, but which may have other unrepresented high bits. As such,
267/// the gep cannot necessarily be reconstructed from its decomposed form.
268///
269/// When DataLayout is around, this function is capable of analyzing everything
270/// that GetUnderlyingObject can look through.  When not, it just looks
271/// through pointer casts.
272///
273static const Value *
274DecomposeGEPExpression(const Value *V, int64_t &BaseOffs,
275                       SmallVectorImpl<VariableGEPIndex> &VarIndices,
276                       const DataLayout *TD) {
277  // Limit recursion depth to limit compile time in crazy cases.
278  unsigned MaxLookup = 6;
279
280  BaseOffs = 0;
281  do {
282    // See if this is a bitcast or GEP.
283    const Operator *Op = dyn_cast<Operator>(V);
284    if (Op == 0) {
285      // The only non-operator case we can handle are GlobalAliases.
286      if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
287        if (!GA->mayBeOverridden()) {
288          V = GA->getAliasee();
289          continue;
290        }
291      }
292      return V;
293    }
294
295    if (Op->getOpcode() == Instruction::BitCast) {
296      V = Op->getOperand(0);
297      continue;
298    }
299
300    const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
301    if (GEPOp == 0) {
302      // If it's not a GEP, hand it off to SimplifyInstruction to see if it
303      // can come up with something. This matches what GetUnderlyingObject does.
304      if (const Instruction *I = dyn_cast<Instruction>(V))
305        // TODO: Get a DominatorTree and use it here.
306        if (const Value *Simplified =
307              SimplifyInstruction(const_cast<Instruction *>(I), TD)) {
308          V = Simplified;
309          continue;
310        }
311
312      return V;
313    }
314
315    // Don't attempt to analyze GEPs over unsized objects.
316    if (!GEPOp->getOperand(0)->getType()->getPointerElementType()->isSized())
317      return V;
318
319    // If we are lacking DataLayout information, we can't compute the offets of
320    // elements computed by GEPs.  However, we can handle bitcast equivalent
321    // GEPs.
322    if (TD == 0) {
323      if (!GEPOp->hasAllZeroIndices())
324        return V;
325      V = GEPOp->getOperand(0);
326      continue;
327    }
328
329    unsigned AS = GEPOp->getPointerAddressSpace();
330    // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
331    gep_type_iterator GTI = gep_type_begin(GEPOp);
332    for (User::const_op_iterator I = GEPOp->op_begin()+1,
333         E = GEPOp->op_end(); I != E; ++I) {
334      Value *Index = *I;
335      // Compute the (potentially symbolic) offset in bytes for this index.
336      if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
337        // For a struct, add the member offset.
338        unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
339        if (FieldNo == 0) continue;
340
341        BaseOffs += TD->getStructLayout(STy)->getElementOffset(FieldNo);
342        continue;
343      }
344
345      // For an array/pointer, add the element offset, explicitly scaled.
346      if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
347        if (CIdx->isZero()) continue;
348        BaseOffs += TD->getTypeAllocSize(*GTI)*CIdx->getSExtValue();
349        continue;
350      }
351
352      uint64_t Scale = TD->getTypeAllocSize(*GTI);
353      ExtensionKind Extension = EK_NotExtended;
354
355      // If the integer type is smaller than the pointer size, it is implicitly
356      // sign extended to pointer size.
357      unsigned Width = Index->getType()->getIntegerBitWidth();
358      if (TD->getPointerSizeInBits(AS) > Width)
359        Extension = EK_SignExt;
360
361      // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
362      APInt IndexScale(Width, 0), IndexOffset(Width, 0);
363      Index = GetLinearExpression(Index, IndexScale, IndexOffset, Extension,
364                                  *TD, 0);
365
366      // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
367      // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
368      BaseOffs += IndexOffset.getSExtValue()*Scale;
369      Scale *= IndexScale.getSExtValue();
370
371      // If we already had an occurrence of this index variable, merge this
372      // scale into it.  For example, we want to handle:
373      //   A[x][x] -> x*16 + x*4 -> x*20
374      // This also ensures that 'x' only appears in the index list once.
375      for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
376        if (VarIndices[i].V == Index &&
377            VarIndices[i].Extension == Extension) {
378          Scale += VarIndices[i].Scale;
379          VarIndices.erase(VarIndices.begin()+i);
380          break;
381        }
382      }
383
384      // Make sure that we have a scale that makes sense for this target's
385      // pointer size.
386      if (unsigned ShiftBits = 64 - TD->getPointerSizeInBits(AS)) {
387        Scale <<= ShiftBits;
388        Scale = (int64_t)Scale >> ShiftBits;
389      }
390
391      if (Scale) {
392        VariableGEPIndex Entry = {Index, Extension,
393                                  static_cast<int64_t>(Scale)};
394        VarIndices.push_back(Entry);
395      }
396    }
397
398    // Analyze the base pointer next.
399    V = GEPOp->getOperand(0);
400  } while (--MaxLookup);
401
402  // If the chain of expressions is too deep, just return early.
403  return V;
404}
405
406/// GetIndexDifference - Dest and Src are the variable indices from two
407/// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base
408/// pointers.  Subtract the GEP2 indices from GEP1 to find the symbolic
409/// difference between the two pointers.
410static void GetIndexDifference(SmallVectorImpl<VariableGEPIndex> &Dest,
411                               const SmallVectorImpl<VariableGEPIndex> &Src) {
412  if (Src.empty()) return;
413
414  for (unsigned i = 0, e = Src.size(); i != e; ++i) {
415    const Value *V = Src[i].V;
416    ExtensionKind Extension = Src[i].Extension;
417    int64_t Scale = Src[i].Scale;
418
419    // Find V in Dest.  This is N^2, but pointer indices almost never have more
420    // than a few variable indexes.
421    for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
422      if (Dest[j].V != V || Dest[j].Extension != Extension) continue;
423
424      // If we found it, subtract off Scale V's from the entry in Dest.  If it
425      // goes to zero, remove the entry.
426      if (Dest[j].Scale != Scale)
427        Dest[j].Scale -= Scale;
428      else
429        Dest.erase(Dest.begin()+j);
430      Scale = 0;
431      break;
432    }
433
434    // If we didn't consume this entry, add it to the end of the Dest list.
435    if (Scale) {
436      VariableGEPIndex Entry = { V, Extension, -Scale };
437      Dest.push_back(Entry);
438    }
439  }
440}
441
442//===----------------------------------------------------------------------===//
443// BasicAliasAnalysis Pass
444//===----------------------------------------------------------------------===//
445
446#ifndef NDEBUG
447static const Function *getParent(const Value *V) {
448  if (const Instruction *inst = dyn_cast<Instruction>(V))
449    return inst->getParent()->getParent();
450
451  if (const Argument *arg = dyn_cast<Argument>(V))
452    return arg->getParent();
453
454  return NULL;
455}
456
457static bool notDifferentParent(const Value *O1, const Value *O2) {
458
459  const Function *F1 = getParent(O1);
460  const Function *F2 = getParent(O2);
461
462  return !F1 || !F2 || F1 == F2;
463}
464#endif
465
466namespace {
467  /// BasicAliasAnalysis - This is the primary alias analysis implementation.
468  struct BasicAliasAnalysis : public ImmutablePass, public AliasAnalysis {
469    static char ID; // Class identification, replacement for typeinfo
470    BasicAliasAnalysis() : ImmutablePass(ID) {
471      initializeBasicAliasAnalysisPass(*PassRegistry::getPassRegistry());
472    }
473
474    virtual void initializePass() {
475      InitializeAliasAnalysis(this);
476    }
477
478    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
479      AU.addRequired<AliasAnalysis>();
480      AU.addRequired<TargetLibraryInfo>();
481    }
482
483    virtual AliasResult alias(const Location &LocA,
484                              const Location &LocB) {
485      assert(AliasCache.empty() && "AliasCache must be cleared after use!");
486      assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
487             "BasicAliasAnalysis doesn't support interprocedural queries.");
488      AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.TBAATag,
489                                     LocB.Ptr, LocB.Size, LocB.TBAATag);
490      // AliasCache rarely has more than 1 or 2 elements, always use
491      // shrink_and_clear so it quickly returns to the inline capacity of the
492      // SmallDenseMap if it ever grows larger.
493      // FIXME: This should really be shrink_to_inline_capacity_and_clear().
494      AliasCache.shrink_and_clear();
495      return Alias;
496    }
497
498    virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
499                                       const Location &Loc);
500
501    virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
502                                       ImmutableCallSite CS2) {
503      // The AliasAnalysis base class has some smarts, lets use them.
504      return AliasAnalysis::getModRefInfo(CS1, CS2);
505    }
506
507    /// pointsToConstantMemory - Chase pointers until we find a (constant
508    /// global) or not.
509    virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
510
511    /// getModRefBehavior - Return the behavior when calling the given
512    /// call site.
513    virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
514
515    /// getModRefBehavior - Return the behavior when calling the given function.
516    /// For use when the call site is not known.
517    virtual ModRefBehavior getModRefBehavior(const Function *F);
518
519    /// getAdjustedAnalysisPointer - This method is used when a pass implements
520    /// an analysis interface through multiple inheritance.  If needed, it
521    /// should override this to adjust the this pointer as needed for the
522    /// specified pass info.
523    virtual void *getAdjustedAnalysisPointer(const void *ID) {
524      if (ID == &AliasAnalysis::ID)
525        return (AliasAnalysis*)this;
526      return this;
527    }
528
529  private:
530    // AliasCache - Track alias queries to guard against recursion.
531    typedef std::pair<Location, Location> LocPair;
532    typedef SmallDenseMap<LocPair, AliasResult, 8> AliasCacheTy;
533    AliasCacheTy AliasCache;
534
535    // Visited - Track instructions visited by pointsToConstantMemory.
536    SmallPtrSet<const Value*, 16> Visited;
537
538    // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP
539    // instruction against another.
540    AliasResult aliasGEP(const GEPOperator *V1, uint64_t V1Size,
541                         const MDNode *V1TBAAInfo,
542                         const Value *V2, uint64_t V2Size,
543                         const MDNode *V2TBAAInfo,
544                         const Value *UnderlyingV1, const Value *UnderlyingV2);
545
546    // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI
547    // instruction against another.
548    AliasResult aliasPHI(const PHINode *PN, uint64_t PNSize,
549                         const MDNode *PNTBAAInfo,
550                         const Value *V2, uint64_t V2Size,
551                         const MDNode *V2TBAAInfo);
552
553    /// aliasSelect - Disambiguate a Select instruction against another value.
554    AliasResult aliasSelect(const SelectInst *SI, uint64_t SISize,
555                            const MDNode *SITBAAInfo,
556                            const Value *V2, uint64_t V2Size,
557                            const MDNode *V2TBAAInfo);
558
559    AliasResult aliasCheck(const Value *V1, uint64_t V1Size,
560                           const MDNode *V1TBAATag,
561                           const Value *V2, uint64_t V2Size,
562                           const MDNode *V2TBAATag);
563  };
564}  // End of anonymous namespace
565
566// Register this pass...
567char BasicAliasAnalysis::ID = 0;
568INITIALIZE_AG_PASS_BEGIN(BasicAliasAnalysis, AliasAnalysis, "basicaa",
569                   "Basic Alias Analysis (stateless AA impl)",
570                   false, true, false)
571INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
572INITIALIZE_AG_PASS_END(BasicAliasAnalysis, AliasAnalysis, "basicaa",
573                   "Basic Alias Analysis (stateless AA impl)",
574                   false, true, false)
575
576
577ImmutablePass *llvm::createBasicAliasAnalysisPass() {
578  return new BasicAliasAnalysis();
579}
580
581/// pointsToConstantMemory - Returns whether the given pointer value
582/// points to memory that is local to the function, with global constants being
583/// considered local to all functions.
584bool
585BasicAliasAnalysis::pointsToConstantMemory(const Location &Loc, bool OrLocal) {
586  assert(Visited.empty() && "Visited must be cleared after use!");
587
588  unsigned MaxLookup = 8;
589  SmallVector<const Value *, 16> Worklist;
590  Worklist.push_back(Loc.Ptr);
591  do {
592    const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), TD);
593    if (!Visited.insert(V)) {
594      Visited.clear();
595      return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
596    }
597
598    // An alloca instruction defines local memory.
599    if (OrLocal && isa<AllocaInst>(V))
600      continue;
601
602    // A global constant counts as local memory for our purposes.
603    if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
604      // Note: this doesn't require GV to be "ODR" because it isn't legal for a
605      // global to be marked constant in some modules and non-constant in
606      // others.  GV may even be a declaration, not a definition.
607      if (!GV->isConstant()) {
608        Visited.clear();
609        return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
610      }
611      continue;
612    }
613
614    // If both select values point to local memory, then so does the select.
615    if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
616      Worklist.push_back(SI->getTrueValue());
617      Worklist.push_back(SI->getFalseValue());
618      continue;
619    }
620
621    // If all values incoming to a phi node point to local memory, then so does
622    // the phi.
623    if (const PHINode *PN = dyn_cast<PHINode>(V)) {
624      // Don't bother inspecting phi nodes with many operands.
625      if (PN->getNumIncomingValues() > MaxLookup) {
626        Visited.clear();
627        return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
628      }
629      for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
630        Worklist.push_back(PN->getIncomingValue(i));
631      continue;
632    }
633
634    // Otherwise be conservative.
635    Visited.clear();
636    return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
637
638  } while (!Worklist.empty() && --MaxLookup);
639
640  Visited.clear();
641  return Worklist.empty();
642}
643
644/// getModRefBehavior - Return the behavior when calling the given call site.
645AliasAnalysis::ModRefBehavior
646BasicAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
647  if (CS.doesNotAccessMemory())
648    // Can't do better than this.
649    return DoesNotAccessMemory;
650
651  ModRefBehavior Min = UnknownModRefBehavior;
652
653  // If the callsite knows it only reads memory, don't return worse
654  // than that.
655  if (CS.onlyReadsMemory())
656    Min = OnlyReadsMemory;
657
658  // The AliasAnalysis base class has some smarts, lets use them.
659  return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
660}
661
662/// getModRefBehavior - Return the behavior when calling the given function.
663/// For use when the call site is not known.
664AliasAnalysis::ModRefBehavior
665BasicAliasAnalysis::getModRefBehavior(const Function *F) {
666  // If the function declares it doesn't access memory, we can't do better.
667  if (F->doesNotAccessMemory())
668    return DoesNotAccessMemory;
669
670  // For intrinsics, we can check the table.
671  if (unsigned iid = F->getIntrinsicID()) {
672#define GET_INTRINSIC_MODREF_BEHAVIOR
673#include "llvm/IR/Intrinsics.gen"
674#undef GET_INTRINSIC_MODREF_BEHAVIOR
675  }
676
677  ModRefBehavior Min = UnknownModRefBehavior;
678
679  // If the function declares it only reads memory, go with that.
680  if (F->onlyReadsMemory())
681    Min = OnlyReadsMemory;
682
683  // Otherwise be conservative.
684  return ModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min);
685}
686
687/// getModRefInfo - Check to see if the specified callsite can clobber the
688/// specified memory object.  Since we only look at local properties of this
689/// function, we really can't say much about this query.  We do, however, use
690/// simple "address taken" analysis on local objects.
691AliasAnalysis::ModRefResult
692BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS,
693                                  const Location &Loc) {
694  assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
695         "AliasAnalysis query involving multiple functions!");
696
697  const Value *Object = GetUnderlyingObject(Loc.Ptr, TD);
698
699  // If this is a tail call and Loc.Ptr points to a stack location, we know that
700  // the tail call cannot access or modify the local stack.
701  // We cannot exclude byval arguments here; these belong to the caller of
702  // the current function not to the current function, and a tail callee
703  // may reference them.
704  if (isa<AllocaInst>(Object))
705    if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
706      if (CI->isTailCall())
707        return NoModRef;
708
709  // If the pointer is to a locally allocated object that does not escape,
710  // then the call can not mod/ref the pointer unless the call takes the pointer
711  // as an argument, and itself doesn't capture it.
712  if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
713      isNonEscapingLocalObject(Object)) {
714    bool PassedAsArg = false;
715    unsigned ArgNo = 0;
716    for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
717         CI != CE; ++CI, ++ArgNo) {
718      // Only look at the no-capture or byval pointer arguments.  If this
719      // pointer were passed to arguments that were neither of these, then it
720      // couldn't be no-capture.
721      if (!(*CI)->getType()->isPointerTy() ||
722          (!CS.doesNotCapture(ArgNo) && !CS.isByValArgument(ArgNo)))
723        continue;
724
725      // If this is a no-capture pointer argument, see if we can tell that it
726      // is impossible to alias the pointer we're checking.  If not, we have to
727      // assume that the call could touch the pointer, even though it doesn't
728      // escape.
729      if (!isNoAlias(Location(*CI), Location(Object))) {
730        PassedAsArg = true;
731        break;
732      }
733    }
734
735    if (!PassedAsArg)
736      return NoModRef;
737  }
738
739  const TargetLibraryInfo &TLI = getAnalysis<TargetLibraryInfo>();
740  ModRefResult Min = ModRef;
741
742  // Finally, handle specific knowledge of intrinsics.
743  const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
744  if (II != 0)
745    switch (II->getIntrinsicID()) {
746    default: break;
747    case Intrinsic::memcpy:
748    case Intrinsic::memmove: {
749      uint64_t Len = UnknownSize;
750      if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2)))
751        Len = LenCI->getZExtValue();
752      Value *Dest = II->getArgOperand(0);
753      Value *Src = II->getArgOperand(1);
754      // If it can't overlap the source dest, then it doesn't modref the loc.
755      if (isNoAlias(Location(Dest, Len), Loc)) {
756        if (isNoAlias(Location(Src, Len), Loc))
757          return NoModRef;
758        // If it can't overlap the dest, then worst case it reads the loc.
759        Min = Ref;
760      } else if (isNoAlias(Location(Src, Len), Loc)) {
761        // If it can't overlap the source, then worst case it mutates the loc.
762        Min = Mod;
763      }
764      break;
765    }
766    case Intrinsic::memset:
767      // Since memset is 'accesses arguments' only, the AliasAnalysis base class
768      // will handle it for the variable length case.
769      if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
770        uint64_t Len = LenCI->getZExtValue();
771        Value *Dest = II->getArgOperand(0);
772        if (isNoAlias(Location(Dest, Len), Loc))
773          return NoModRef;
774      }
775      // We know that memset doesn't load anything.
776      Min = Mod;
777      break;
778    case Intrinsic::lifetime_start:
779    case Intrinsic::lifetime_end:
780    case Intrinsic::invariant_start: {
781      uint64_t PtrSize =
782        cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
783      if (isNoAlias(Location(II->getArgOperand(1),
784                             PtrSize,
785                             II->getMetadata(LLVMContext::MD_tbaa)),
786                    Loc))
787        return NoModRef;
788      break;
789    }
790    case Intrinsic::invariant_end: {
791      uint64_t PtrSize =
792        cast<ConstantInt>(II->getArgOperand(1))->getZExtValue();
793      if (isNoAlias(Location(II->getArgOperand(2),
794                             PtrSize,
795                             II->getMetadata(LLVMContext::MD_tbaa)),
796                    Loc))
797        return NoModRef;
798      break;
799    }
800    case Intrinsic::arm_neon_vld1: {
801      // LLVM's vld1 and vst1 intrinsics currently only support a single
802      // vector register.
803      uint64_t Size =
804        TD ? TD->getTypeStoreSize(II->getType()) : UnknownSize;
805      if (isNoAlias(Location(II->getArgOperand(0), Size,
806                             II->getMetadata(LLVMContext::MD_tbaa)),
807                    Loc))
808        return NoModRef;
809      break;
810    }
811    case Intrinsic::arm_neon_vst1: {
812      uint64_t Size =
813        TD ? TD->getTypeStoreSize(II->getArgOperand(1)->getType()) : UnknownSize;
814      if (isNoAlias(Location(II->getArgOperand(0), Size,
815                             II->getMetadata(LLVMContext::MD_tbaa)),
816                    Loc))
817        return NoModRef;
818      break;
819    }
820    }
821
822  // We can bound the aliasing properties of memset_pattern16 just as we can
823  // for memcpy/memset.  This is particularly important because the
824  // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
825  // whenever possible.
826  else if (TLI.has(LibFunc::memset_pattern16) &&
827           CS.getCalledFunction() &&
828           CS.getCalledFunction()->getName() == "memset_pattern16") {
829    const Function *MS = CS.getCalledFunction();
830    FunctionType *MemsetType = MS->getFunctionType();
831    if (!MemsetType->isVarArg() && MemsetType->getNumParams() == 3 &&
832        isa<PointerType>(MemsetType->getParamType(0)) &&
833        isa<PointerType>(MemsetType->getParamType(1)) &&
834        isa<IntegerType>(MemsetType->getParamType(2))) {
835      uint64_t Len = UnknownSize;
836      if (const ConstantInt *LenCI = dyn_cast<ConstantInt>(CS.getArgument(2)))
837        Len = LenCI->getZExtValue();
838      const Value *Dest = CS.getArgument(0);
839      const Value *Src = CS.getArgument(1);
840      // If it can't overlap the source dest, then it doesn't modref the loc.
841      if (isNoAlias(Location(Dest, Len), Loc)) {
842        // Always reads 16 bytes of the source.
843        if (isNoAlias(Location(Src, 16), Loc))
844          return NoModRef;
845        // If it can't overlap the dest, then worst case it reads the loc.
846        Min = Ref;
847      // Always reads 16 bytes of the source.
848      } else if (isNoAlias(Location(Src, 16), Loc)) {
849        // If it can't overlap the source, then worst case it mutates the loc.
850        Min = Mod;
851      }
852    }
853  }
854
855  // The AliasAnalysis base class has some smarts, lets use them.
856  return ModRefResult(AliasAnalysis::getModRefInfo(CS, Loc) & Min);
857}
858
859static bool areVarIndicesEqual(SmallVectorImpl<VariableGEPIndex> &Indices1,
860                               SmallVectorImpl<VariableGEPIndex> &Indices2) {
861  unsigned Size1 = Indices1.size();
862  unsigned Size2 = Indices2.size();
863
864  if (Size1 != Size2)
865    return false;
866
867  for (unsigned I = 0; I != Size1; ++I)
868    if (Indices1[I] != Indices2[I])
869      return false;
870
871  return true;
872}
873
874/// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
875/// against another pointer.  We know that V1 is a GEP, but we don't know
876/// anything about V2.  UnderlyingV1 is GetUnderlyingObject(GEP1, TD),
877/// UnderlyingV2 is the same for V2.
878///
879AliasAnalysis::AliasResult
880BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
881                             const MDNode *V1TBAAInfo,
882                             const Value *V2, uint64_t V2Size,
883                             const MDNode *V2TBAAInfo,
884                             const Value *UnderlyingV1,
885                             const Value *UnderlyingV2) {
886  int64_t GEP1BaseOffset;
887  SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
888
889  // If we have two gep instructions with must-alias or not-alias'ing base
890  // pointers, figure out if the indexes to the GEP tell us anything about the
891  // derived pointer.
892  if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
893    // Do the base pointers alias?
894    AliasResult BaseAlias = aliasCheck(UnderlyingV1, UnknownSize, 0,
895                                       UnderlyingV2, UnknownSize, 0);
896
897    // Check for geps of non-aliasing underlying pointers where the offsets are
898    // identical.
899    if ((BaseAlias == MayAlias) && V1Size == V2Size) {
900      // Do the base pointers alias assuming type and size.
901      AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size,
902                                                V1TBAAInfo, UnderlyingV2,
903                                                V2Size, V2TBAAInfo);
904      if (PreciseBaseAlias == NoAlias) {
905        // See if the computed offset from the common pointer tells us about the
906        // relation of the resulting pointer.
907        int64_t GEP2BaseOffset;
908        SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
909        const Value *GEP2BasePtr =
910          DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD);
911        const Value *GEP1BasePtr =
912          DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
913        // DecomposeGEPExpression and GetUnderlyingObject should return the
914        // same result except when DecomposeGEPExpression has no DataLayout.
915        if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
916          assert(TD == 0 &&
917             "DecomposeGEPExpression and GetUnderlyingObject disagree!");
918          return MayAlias;
919        }
920        // Same offsets.
921        if (GEP1BaseOffset == GEP2BaseOffset &&
922            areVarIndicesEqual(GEP1VariableIndices, GEP2VariableIndices))
923          return NoAlias;
924        GEP1VariableIndices.clear();
925      }
926    }
927
928    // If we get a No or May, then return it immediately, no amount of analysis
929    // will improve this situation.
930    if (BaseAlias != MustAlias) return BaseAlias;
931
932    // Otherwise, we have a MustAlias.  Since the base pointers alias each other
933    // exactly, see if the computed offset from the common pointer tells us
934    // about the relation of the resulting pointer.
935    const Value *GEP1BasePtr =
936      DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
937
938    int64_t GEP2BaseOffset;
939    SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
940    const Value *GEP2BasePtr =
941      DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD);
942
943    // DecomposeGEPExpression and GetUnderlyingObject should return the
944    // same result except when DecomposeGEPExpression has no DataLayout.
945    if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
946      assert(TD == 0 &&
947             "DecomposeGEPExpression and GetUnderlyingObject disagree!");
948      return MayAlias;
949    }
950
951    // Subtract the GEP2 pointer from the GEP1 pointer to find out their
952    // symbolic difference.
953    GEP1BaseOffset -= GEP2BaseOffset;
954    GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
955
956  } else {
957    // Check to see if these two pointers are related by the getelementptr
958    // instruction.  If one pointer is a GEP with a non-zero index of the other
959    // pointer, we know they cannot alias.
960
961    // If both accesses are unknown size, we can't do anything useful here.
962    if (V1Size == UnknownSize && V2Size == UnknownSize)
963      return MayAlias;
964
965    AliasResult R = aliasCheck(UnderlyingV1, UnknownSize, 0,
966                               V2, V2Size, V2TBAAInfo);
967    if (R != MustAlias)
968      // If V2 may alias GEP base pointer, conservatively returns MayAlias.
969      // If V2 is known not to alias GEP base pointer, then the two values
970      // cannot alias per GEP semantics: "A pointer value formed from a
971      // getelementptr instruction is associated with the addresses associated
972      // with the first operand of the getelementptr".
973      return R;
974
975    const Value *GEP1BasePtr =
976      DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
977
978    // DecomposeGEPExpression and GetUnderlyingObject should return the
979    // same result except when DecomposeGEPExpression has no DataLayout.
980    if (GEP1BasePtr != UnderlyingV1) {
981      assert(TD == 0 &&
982             "DecomposeGEPExpression and GetUnderlyingObject disagree!");
983      return MayAlias;
984    }
985  }
986
987  // In the two GEP Case, if there is no difference in the offsets of the
988  // computed pointers, the resultant pointers are a must alias.  This
989  // hapens when we have two lexically identical GEP's (for example).
990  //
991  // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
992  // must aliases the GEP, the end result is a must alias also.
993  if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
994    return MustAlias;
995
996  // If there is a constant difference between the pointers, but the difference
997  // is less than the size of the associated memory object, then we know
998  // that the objects are partially overlapping.  If the difference is
999  // greater, we know they do not overlap.
1000  if (GEP1BaseOffset != 0 && GEP1VariableIndices.empty()) {
1001    if (GEP1BaseOffset >= 0) {
1002      if (V2Size != UnknownSize) {
1003        if ((uint64_t)GEP1BaseOffset < V2Size)
1004          return PartialAlias;
1005        return NoAlias;
1006      }
1007    } else {
1008      if (V1Size != UnknownSize) {
1009        if (-(uint64_t)GEP1BaseOffset < V1Size)
1010          return PartialAlias;
1011        return NoAlias;
1012      }
1013    }
1014  }
1015
1016  // Try to distinguish something like &A[i][1] against &A[42][0].
1017  // Grab the least significant bit set in any of the scales.
1018  if (!GEP1VariableIndices.empty()) {
1019    uint64_t Modulo = 0;
1020    for (unsigned i = 0, e = GEP1VariableIndices.size(); i != e; ++i)
1021      Modulo |= (uint64_t)GEP1VariableIndices[i].Scale;
1022    Modulo = Modulo ^ (Modulo & (Modulo - 1));
1023
1024    // We can compute the difference between the two addresses
1025    // mod Modulo. Check whether that difference guarantees that the
1026    // two locations do not alias.
1027    uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1);
1028    if (V1Size != UnknownSize && V2Size != UnknownSize &&
1029        ModOffset >= V2Size && V1Size <= Modulo - ModOffset)
1030      return NoAlias;
1031  }
1032
1033  // Statically, we can see that the base objects are the same, but the
1034  // pointers have dynamic offsets which we can't resolve. And none of our
1035  // little tricks above worked.
1036  //
1037  // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the
1038  // practical effect of this is protecting TBAA in the case of dynamic
1039  // indices into arrays of unions or malloc'd memory.
1040  return PartialAlias;
1041}
1042
1043static AliasAnalysis::AliasResult
1044MergeAliasResults(AliasAnalysis::AliasResult A, AliasAnalysis::AliasResult B) {
1045  // If the results agree, take it.
1046  if (A == B)
1047    return A;
1048  // A mix of PartialAlias and MustAlias is PartialAlias.
1049  if ((A == AliasAnalysis::PartialAlias && B == AliasAnalysis::MustAlias) ||
1050      (B == AliasAnalysis::PartialAlias && A == AliasAnalysis::MustAlias))
1051    return AliasAnalysis::PartialAlias;
1052  // Otherwise, we don't know anything.
1053  return AliasAnalysis::MayAlias;
1054}
1055
1056/// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select
1057/// instruction against another.
1058AliasAnalysis::AliasResult
1059BasicAliasAnalysis::aliasSelect(const SelectInst *SI, uint64_t SISize,
1060                                const MDNode *SITBAAInfo,
1061                                const Value *V2, uint64_t V2Size,
1062                                const MDNode *V2TBAAInfo) {
1063  // If the values are Selects with the same condition, we can do a more precise
1064  // check: just check for aliases between the values on corresponding arms.
1065  if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1066    if (SI->getCondition() == SI2->getCondition()) {
1067      AliasResult Alias =
1068        aliasCheck(SI->getTrueValue(), SISize, SITBAAInfo,
1069                   SI2->getTrueValue(), V2Size, V2TBAAInfo);
1070      if (Alias == MayAlias)
1071        return MayAlias;
1072      AliasResult ThisAlias =
1073        aliasCheck(SI->getFalseValue(), SISize, SITBAAInfo,
1074                   SI2->getFalseValue(), V2Size, V2TBAAInfo);
1075      return MergeAliasResults(ThisAlias, Alias);
1076    }
1077
1078  // If both arms of the Select node NoAlias or MustAlias V2, then returns
1079  // NoAlias / MustAlias. Otherwise, returns MayAlias.
1080  AliasResult Alias =
1081    aliasCheck(V2, V2Size, V2TBAAInfo, SI->getTrueValue(), SISize, SITBAAInfo);
1082  if (Alias == MayAlias)
1083    return MayAlias;
1084
1085  AliasResult ThisAlias =
1086    aliasCheck(V2, V2Size, V2TBAAInfo, SI->getFalseValue(), SISize, SITBAAInfo);
1087  return MergeAliasResults(ThisAlias, Alias);
1088}
1089
1090// aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
1091// against another.
1092AliasAnalysis::AliasResult
1093BasicAliasAnalysis::aliasPHI(const PHINode *PN, uint64_t PNSize,
1094                             const MDNode *PNTBAAInfo,
1095                             const Value *V2, uint64_t V2Size,
1096                             const MDNode *V2TBAAInfo) {
1097  // If the values are PHIs in the same block, we can do a more precise
1098  // as well as efficient check: just check for aliases between the values
1099  // on corresponding edges.
1100  if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1101    if (PN2->getParent() == PN->getParent()) {
1102      LocPair Locs(Location(PN, PNSize, PNTBAAInfo),
1103                   Location(V2, V2Size, V2TBAAInfo));
1104      if (PN > V2)
1105        std::swap(Locs.first, Locs.second);
1106      // Analyse the PHIs' inputs under the assumption that the PHIs are
1107      // NoAlias.
1108      // If the PHIs are May/MustAlias there must be (recursively) an input
1109      // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
1110      // there must be an operation on the PHIs within the PHIs' value cycle
1111      // that causes a MayAlias.
1112      // Pretend the phis do not alias.
1113      AliasResult Alias = NoAlias;
1114      assert(AliasCache.count(Locs) &&
1115             "There must exist an entry for the phi node");
1116      AliasResult OrigAliasResult = AliasCache[Locs];
1117      AliasCache[Locs] = NoAlias;
1118
1119      for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1120        AliasResult ThisAlias =
1121          aliasCheck(PN->getIncomingValue(i), PNSize, PNTBAAInfo,
1122                     PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1123                     V2Size, V2TBAAInfo);
1124        Alias = MergeAliasResults(ThisAlias, Alias);
1125        if (Alias == MayAlias)
1126          break;
1127      }
1128
1129      // Reset if speculation failed.
1130      if (Alias != NoAlias)
1131        AliasCache[Locs] = OrigAliasResult;
1132
1133      return Alias;
1134    }
1135
1136  SmallPtrSet<Value*, 4> UniqueSrc;
1137  SmallVector<Value*, 4> V1Srcs;
1138  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1139    Value *PV1 = PN->getIncomingValue(i);
1140    if (isa<PHINode>(PV1))
1141      // If any of the source itself is a PHI, return MayAlias conservatively
1142      // to avoid compile time explosion. The worst possible case is if both
1143      // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1144      // and 'n' are the number of PHI sources.
1145      return MayAlias;
1146    if (UniqueSrc.insert(PV1))
1147      V1Srcs.push_back(PV1);
1148  }
1149
1150  AliasResult Alias = aliasCheck(V2, V2Size, V2TBAAInfo,
1151                                 V1Srcs[0], PNSize, PNTBAAInfo);
1152  // Early exit if the check of the first PHI source against V2 is MayAlias.
1153  // Other results are not possible.
1154  if (Alias == MayAlias)
1155    return MayAlias;
1156
1157  // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1158  // NoAlias / MustAlias. Otherwise, returns MayAlias.
1159  for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1160    Value *V = V1Srcs[i];
1161
1162    AliasResult ThisAlias = aliasCheck(V2, V2Size, V2TBAAInfo,
1163                                       V, PNSize, PNTBAAInfo);
1164    Alias = MergeAliasResults(ThisAlias, Alias);
1165    if (Alias == MayAlias)
1166      break;
1167  }
1168
1169  return Alias;
1170}
1171
1172// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
1173// such as array references.
1174//
1175AliasAnalysis::AliasResult
1176BasicAliasAnalysis::aliasCheck(const Value *V1, uint64_t V1Size,
1177                               const MDNode *V1TBAAInfo,
1178                               const Value *V2, uint64_t V2Size,
1179                               const MDNode *V2TBAAInfo) {
1180  // If either of the memory references is empty, it doesn't matter what the
1181  // pointer values are.
1182  if (V1Size == 0 || V2Size == 0)
1183    return NoAlias;
1184
1185  // Strip off any casts if they exist.
1186  V1 = V1->stripPointerCasts();
1187  V2 = V2->stripPointerCasts();
1188
1189  // Are we checking for alias of the same value?
1190  if (V1 == V2) return MustAlias;
1191
1192  if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1193    return NoAlias;  // Scalars cannot alias each other
1194
1195  // Figure out what objects these things are pointing to if we can.
1196  const Value *O1 = GetUnderlyingObject(V1, TD);
1197  const Value *O2 = GetUnderlyingObject(V2, TD);
1198
1199  // Null values in the default address space don't point to any object, so they
1200  // don't alias any other pointer.
1201  if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1202    if (CPN->getType()->getAddressSpace() == 0)
1203      return NoAlias;
1204  if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1205    if (CPN->getType()->getAddressSpace() == 0)
1206      return NoAlias;
1207
1208  if (O1 != O2) {
1209    // If V1/V2 point to two different objects we know that we have no alias.
1210    if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1211      return NoAlias;
1212
1213    // Constant pointers can't alias with non-const isIdentifiedObject objects.
1214    if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1215        (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1216      return NoAlias;
1217
1218    // Function arguments can't alias with things that are known to be
1219    // unambigously identified at the function level.
1220    if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1221        (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1222      return NoAlias;
1223
1224    // Most objects can't alias null.
1225    if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
1226        (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
1227      return NoAlias;
1228
1229    // If one pointer is the result of a call/invoke or load and the other is a
1230    // non-escaping local object within the same function, then we know the
1231    // object couldn't escape to a point where the call could return it.
1232    //
1233    // Note that if the pointers are in different functions, there are a
1234    // variety of complications. A call with a nocapture argument may still
1235    // temporary store the nocapture argument's value in a temporary memory
1236    // location if that memory location doesn't escape. Or it may pass a
1237    // nocapture value to other functions as long as they don't capture it.
1238    if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1239      return NoAlias;
1240    if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1241      return NoAlias;
1242  }
1243
1244  // If the size of one access is larger than the entire object on the other
1245  // side, then we know such behavior is undefined and can assume no alias.
1246  if (TD)
1247    if ((V1Size != UnknownSize && isObjectSmallerThan(O2, V1Size, *TD, *TLI)) ||
1248        (V2Size != UnknownSize && isObjectSmallerThan(O1, V2Size, *TD, *TLI)))
1249      return NoAlias;
1250
1251  // Check the cache before climbing up use-def chains. This also terminates
1252  // otherwise infinitely recursive queries.
1253  LocPair Locs(Location(V1, V1Size, V1TBAAInfo),
1254               Location(V2, V2Size, V2TBAAInfo));
1255  if (V1 > V2)
1256    std::swap(Locs.first, Locs.second);
1257  std::pair<AliasCacheTy::iterator, bool> Pair =
1258    AliasCache.insert(std::make_pair(Locs, MayAlias));
1259  if (!Pair.second)
1260    return Pair.first->second;
1261
1262  // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1263  // GEP can't simplify, we don't even look at the PHI cases.
1264  if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1265    std::swap(V1, V2);
1266    std::swap(V1Size, V2Size);
1267    std::swap(O1, O2);
1268    std::swap(V1TBAAInfo, V2TBAAInfo);
1269  }
1270  if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1271    AliasResult Result = aliasGEP(GV1, V1Size, V1TBAAInfo, V2, V2Size, V2TBAAInfo, O1, O2);
1272    if (Result != MayAlias) return AliasCache[Locs] = Result;
1273  }
1274
1275  if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1276    std::swap(V1, V2);
1277    std::swap(V1Size, V2Size);
1278    std::swap(V1TBAAInfo, V2TBAAInfo);
1279  }
1280  if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1281    AliasResult Result = aliasPHI(PN, V1Size, V1TBAAInfo,
1282                                  V2, V2Size, V2TBAAInfo);
1283    if (Result != MayAlias) return AliasCache[Locs] = Result;
1284  }
1285
1286  if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1287    std::swap(V1, V2);
1288    std::swap(V1Size, V2Size);
1289    std::swap(V1TBAAInfo, V2TBAAInfo);
1290  }
1291  if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1292    AliasResult Result = aliasSelect(S1, V1Size, V1TBAAInfo,
1293                                     V2, V2Size, V2TBAAInfo);
1294    if (Result != MayAlias) return AliasCache[Locs] = Result;
1295  }
1296
1297  // If both pointers are pointing into the same object and one of them
1298  // accesses is accessing the entire object, then the accesses must
1299  // overlap in some way.
1300  if (TD && O1 == O2)
1301    if ((V1Size != UnknownSize && isObjectSize(O1, V1Size, *TD, *TLI)) ||
1302        (V2Size != UnknownSize && isObjectSize(O2, V2Size, *TD, *TLI)))
1303      return AliasCache[Locs] = PartialAlias;
1304
1305  AliasResult Result =
1306    AliasAnalysis::alias(Location(V1, V1Size, V1TBAAInfo),
1307                         Location(V2, V2Size, V2TBAAInfo));
1308  return AliasCache[Locs] = Result;
1309}
1310