BasicAliasAnalysis.cpp revision 6e9e010752e475c27d929147dc07d6261fc0f0f1
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/CaptureTracking.h"
18#include "llvm/Analysis/MallocHelper.h"
19#include "llvm/Analysis/Passes.h"
20#include "llvm/Constants.h"
21#include "llvm/DerivedTypes.h"
22#include "llvm/Function.h"
23#include "llvm/GlobalVariable.h"
24#include "llvm/Instructions.h"
25#include "llvm/IntrinsicInst.h"
26#include "llvm/LLVMContext.h"
27#include "llvm/Operator.h"
28#include "llvm/Pass.h"
29#include "llvm/Target/TargetData.h"
30#include "llvm/ADT/SmallSet.h"
31#include "llvm/ADT/SmallVector.h"
32#include "llvm/ADT/STLExtras.h"
33#include "llvm/Support/Compiler.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
43static const GEPOperator *isGEP(const Value *V) {
44  return dyn_cast<GEPOperator>(V);
45}
46
47static const Value *GetGEPOperands(const Value *V,
48                                   SmallVector<Value*, 16> &GEPOps) {
49  assert(GEPOps.empty() && "Expect empty list to populate!");
50  GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1,
51                cast<User>(V)->op_end());
52
53  // Accumulate all of the chained indexes into the operand array
54  V = cast<User>(V)->getOperand(0);
55
56  while (const User *G = isGEP(V)) {
57    if (!isa<Constant>(GEPOps[0]) || isa<GlobalValue>(GEPOps[0]) ||
58        !cast<Constant>(GEPOps[0])->isNullValue())
59      break;  // Don't handle folding arbitrary pointer offsets yet...
60    GEPOps.erase(GEPOps.begin());   // Drop the zero index
61    GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end());
62    V = G->getOperand(0);
63  }
64  return V;
65}
66
67/// isKnownNonNull - Return true if we know that the specified value is never
68/// null.
69static bool isKnownNonNull(const Value *V) {
70  // Alloca never returns null, malloc might.
71  if (isa<AllocaInst>(V)) return true;
72
73  // A byval argument is never null.
74  if (const Argument *A = dyn_cast<Argument>(V))
75    return A->hasByValAttr();
76
77  // Global values are not null unless extern weak.
78  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
79    return !GV->hasExternalWeakLinkage();
80  return false;
81}
82
83/// isNonEscapingLocalObject - Return true if the pointer is to a function-local
84/// object that never escapes from the function.
85static bool isNonEscapingLocalObject(const Value *V) {
86  // If this is a local allocation, check to see if it escapes.
87  if (isa<AllocationInst>(V) || isNoAliasCall(V))
88    return !PointerMayBeCaptured(V, false);
89
90  // If this is an argument that corresponds to a byval or noalias argument,
91  // then it has not escaped before entering the function.  Check if it escapes
92  // inside the function.
93  if (const Argument *A = dyn_cast<Argument>(V))
94    if (A->hasByValAttr() || A->hasNoAliasAttr()) {
95      // Don't bother analyzing arguments already known not to escape.
96      if (A->hasNoCaptureAttr())
97        return true;
98      return !PointerMayBeCaptured(V, false);
99    }
100  return false;
101}
102
103
104/// isObjectSmallerThan - Return true if we can prove that the object specified
105/// by V is smaller than Size.
106static bool isObjectSmallerThan(const Value *V, unsigned Size,
107                                LLVMContext &Context, const TargetData &TD) {
108  const Type *AccessTy;
109  if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
110    AccessTy = GV->getType()->getElementType();
111  } else if (const AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
112    if (!AI->isArrayAllocation())
113      AccessTy = AI->getType()->getElementType();
114    else
115      return false;
116  } else if (const CallInst* CI = extractMallocCall(V)) {
117    if (!isArrayMalloc(V, Context, &TD))
118      // The size is the argument to the malloc call.
119      if (const ConstantInt* C = dyn_cast<ConstantInt>(CI->getOperand(1)))
120        return (C->getZExtValue() < Size);
121    return false;
122  } else if (const Argument *A = dyn_cast<Argument>(V)) {
123    if (A->hasByValAttr())
124      AccessTy = cast<PointerType>(A->getType())->getElementType();
125    else
126      return false;
127  } else {
128    return false;
129  }
130
131  if (AccessTy->isSized())
132    return TD.getTypeAllocSize(AccessTy) < Size;
133  return false;
134}
135
136//===----------------------------------------------------------------------===//
137// NoAA Pass
138//===----------------------------------------------------------------------===//
139
140namespace {
141  /// NoAA - This class implements the -no-aa pass, which always returns "I
142  /// don't know" for alias queries.  NoAA is unlike other alias analysis
143  /// implementations, in that it does not chain to a previous analysis.  As
144  /// such it doesn't follow many of the rules that other alias analyses must.
145  ///
146  struct VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis {
147    static char ID; // Class identification, replacement for typeinfo
148    NoAA() : ImmutablePass(&ID) {}
149    explicit NoAA(void *PID) : ImmutablePass(PID) { }
150
151    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
152    }
153
154    virtual void initializePass() {
155      TD = getAnalysisIfAvailable<TargetData>();
156    }
157
158    virtual AliasResult alias(const Value *V1, unsigned V1Size,
159                              const Value *V2, unsigned V2Size) {
160      return MayAlias;
161    }
162
163    virtual void getArgumentAccesses(Function *F, CallSite CS,
164                                     std::vector<PointerAccessInfo> &Info) {
165      llvm_unreachable("This method may not be called on this function!");
166    }
167
168    virtual void getMustAliases(Value *P, std::vector<Value*> &RetVals) { }
169    virtual bool pointsToConstantMemory(const Value *P) { return false; }
170    virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) {
171      return ModRef;
172    }
173    virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
174      return ModRef;
175    }
176    virtual bool hasNoModRefInfoForCalls() const { return true; }
177
178    virtual void deleteValue(Value *V) {}
179    virtual void copyValue(Value *From, Value *To) {}
180  };
181}  // End of anonymous namespace
182
183// Register this pass...
184char NoAA::ID = 0;
185static RegisterPass<NoAA>
186U("no-aa", "No Alias Analysis (always returns 'may' alias)", true, true);
187
188// Declare that we implement the AliasAnalysis interface
189static RegisterAnalysisGroup<AliasAnalysis> V(U);
190
191ImmutablePass *llvm::createNoAAPass() { return new NoAA(); }
192
193//===----------------------------------------------------------------------===//
194// BasicAA Pass
195//===----------------------------------------------------------------------===//
196
197namespace {
198  /// BasicAliasAnalysis - This is the default alias analysis implementation.
199  /// Because it doesn't chain to a previous alias analysis (like -no-aa), it
200  /// derives from the NoAA class.
201  struct VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA {
202    static char ID; // Class identification, replacement for typeinfo
203    BasicAliasAnalysis() : NoAA(&ID) {}
204    AliasResult alias(const Value *V1, unsigned V1Size,
205                      const Value *V2, unsigned V2Size) {
206      assert(VisitedPHIs.empty() && "VisitedPHIs must be cleared after use!");
207      AliasResult Alias = aliasCheck(V1, V1Size, V2, V2Size);
208      VisitedPHIs.clear();
209      return Alias;
210    }
211
212    ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
213    ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
214
215    /// hasNoModRefInfoForCalls - We can provide mod/ref information against
216    /// non-escaping allocations.
217    virtual bool hasNoModRefInfoForCalls() const { return false; }
218
219    /// pointsToConstantMemory - Chase pointers until we find a (constant
220    /// global) or not.
221    bool pointsToConstantMemory(const Value *P);
222
223  private:
224    // VisitedPHIs - Track PHI nodes visited by a aliasCheck() call.
225    SmallSet<const PHINode*, 16> VisitedPHIs;
226
227    // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
228    // against another.
229    AliasResult aliasGEP(const Value *V1, unsigned V1Size,
230                         const Value *V2, unsigned V2Size);
231
232    // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
233    // against another.
234    AliasResult aliasPHI(const PHINode *PN, unsigned PNSize,
235                         const Value *V2, unsigned V2Size);
236
237    AliasResult aliasCheck(const Value *V1, unsigned V1Size,
238                           const Value *V2, unsigned V2Size);
239
240    // CheckGEPInstructions - Check two GEP instructions with known
241    // must-aliasing base pointers.  This checks to see if the index expressions
242    // preclude the pointers from aliasing...
243    AliasResult
244    CheckGEPInstructions(const Type* BasePtr1Ty,
245                         Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size,
246                         const Type *BasePtr2Ty,
247                         Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size);
248  };
249}  // End of anonymous namespace
250
251// Register this pass...
252char BasicAliasAnalysis::ID = 0;
253static RegisterPass<BasicAliasAnalysis>
254X("basicaa", "Basic Alias Analysis (default AA impl)", false, true);
255
256// Declare that we implement the AliasAnalysis interface
257static RegisterAnalysisGroup<AliasAnalysis, true> Y(X);
258
259ImmutablePass *llvm::createBasicAliasAnalysisPass() {
260  return new BasicAliasAnalysis();
261}
262
263
264/// pointsToConstantMemory - Chase pointers until we find a (constant
265/// global) or not.
266bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) {
267  if (const GlobalVariable *GV =
268        dyn_cast<GlobalVariable>(P->getUnderlyingObject()))
269    return GV->isConstant();
270  return false;
271}
272
273
274// getModRefInfo - Check to see if the specified callsite can clobber the
275// specified memory object.  Since we only look at local properties of this
276// function, we really can't say much about this query.  We do, however, use
277// simple "address taken" analysis on local objects.
278//
279AliasAnalysis::ModRefResult
280BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
281  if (!isa<Constant>(P)) {
282    const Value *Object = P->getUnderlyingObject();
283
284    // If this is a tail call and P points to a stack location, we know that
285    // the tail call cannot access or modify the local stack.
286    // We cannot exclude byval arguments here; these belong to the caller of
287    // the current function not to the current function, and a tail callee
288    // may reference them.
289    if (isa<AllocaInst>(Object))
290      if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
291        if (CI->isTailCall())
292          return NoModRef;
293
294    // If the pointer is to a locally allocated object that does not escape,
295    // then the call can not mod/ref the pointer unless the call takes the
296    // argument without capturing it.
297    if (isNonEscapingLocalObject(Object) && CS.getInstruction() != Object) {
298      bool passedAsArg = false;
299      // TODO: Eventually only check 'nocapture' arguments.
300      for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
301           CI != CE; ++CI)
302        if (isa<PointerType>((*CI)->getType()) &&
303            alias(cast<Value>(CI), ~0U, P, ~0U) != NoAlias)
304          passedAsArg = true;
305
306      if (!passedAsArg)
307        return NoModRef;
308    }
309
310    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
311      switch (II->getIntrinsicID()) {
312      default: break;
313      case Intrinsic::memcpy:
314      case Intrinsic::memmove: {
315        unsigned Len = ~0U;
316        if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getOperand(3)))
317          Len = LenCI->getZExtValue();
318        Value *Dest = II->getOperand(1);
319        Value *Src = II->getOperand(2);
320        if (alias(Dest, Len, P, Size) == NoAlias) {
321          if (alias(Src, Len, P, Size) == NoAlias)
322            return NoModRef;
323          return Ref;
324        }
325        }
326        break;
327      case Intrinsic::memset:
328        if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getOperand(3))) {
329          unsigned Len = LenCI->getZExtValue();
330          Value *Dest = II->getOperand(1);
331          if (alias(Dest, Len, P, Size) == NoAlias)
332            return NoModRef;
333        }
334        break;
335      case Intrinsic::atomic_cmp_swap:
336      case Intrinsic::atomic_swap:
337      case Intrinsic::atomic_load_add:
338      case Intrinsic::atomic_load_sub:
339      case Intrinsic::atomic_load_and:
340      case Intrinsic::atomic_load_nand:
341      case Intrinsic::atomic_load_or:
342      case Intrinsic::atomic_load_xor:
343      case Intrinsic::atomic_load_max:
344      case Intrinsic::atomic_load_min:
345      case Intrinsic::atomic_load_umax:
346      case Intrinsic::atomic_load_umin:
347        if (TD) {
348          Value *Op1 = II->getOperand(1);
349          unsigned Op1Size = TD->getTypeStoreSize(Op1->getType());
350          if (alias(Op1, Op1Size, P, Size) == NoAlias)
351            return NoModRef;
352        }
353        break;
354      case Intrinsic::lifetime_start:
355      case Intrinsic::lifetime_end:
356      case Intrinsic::invariant_start: {
357        unsigned PtrSize = cast<ConstantInt>(II->getOperand(1))->getZExtValue();
358        if (alias(II->getOperand(2), PtrSize, P, Size) == NoAlias)
359          return NoModRef;
360      }
361      case Intrinsic::invariant_end: {
362        unsigned PtrSize = cast<ConstantInt>(II->getOperand(2))->getZExtValue();
363        if (alias(II->getOperand(3), PtrSize, P, Size) == NoAlias)
364          return NoModRef;
365      }
366      }
367    }
368  }
369
370  // The AliasAnalysis base class has some smarts, lets use them.
371  return AliasAnalysis::getModRefInfo(CS, P, Size);
372}
373
374
375AliasAnalysis::ModRefResult
376BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) {
377  // If CS1 or CS2 are readnone, they don't interact.
378  ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1);
379  if (CS1B == DoesNotAccessMemory) return NoModRef;
380
381  ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2);
382  if (CS2B == DoesNotAccessMemory) return NoModRef;
383
384  // If they both only read from memory, just return ref.
385  if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory)
386    return Ref;
387
388  // Otherwise, fall back to NoAA (mod+ref).
389  return NoAA::getModRefInfo(CS1, CS2);
390}
391
392// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
393// against another.
394//
395AliasAnalysis::AliasResult
396BasicAliasAnalysis::aliasGEP(const Value *V1, unsigned V1Size,
397                             const Value *V2, unsigned V2Size) {
398  // If we have two gep instructions with must-alias'ing base pointers, figure
399  // out if the indexes to the GEP tell us anything about the derived pointer.
400  // Note that we also handle chains of getelementptr instructions as well as
401  // constant expression getelementptrs here.
402  //
403  if (isGEP(V1) && isGEP(V2)) {
404    const User *GEP1 = cast<User>(V1);
405    const User *GEP2 = cast<User>(V2);
406
407    // If V1 and V2 are identical GEPs, just recurse down on both of them.
408    // This allows us to analyze things like:
409    //   P = gep A, 0, i, 1
410    //   Q = gep B, 0, i, 1
411    // by just analyzing A and B.  This is even safe for variable indices.
412    if (GEP1->getType() == GEP2->getType() &&
413        GEP1->getNumOperands() == GEP2->getNumOperands() &&
414        GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() &&
415        // All operands are the same, ignoring the base.
416        std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1))
417      return aliasCheck(GEP1->getOperand(0), V1Size,
418                        GEP2->getOperand(0), V2Size);
419
420    // Drill down into the first non-gep value, to test for must-aliasing of
421    // the base pointers.
422    while (isGEP(GEP1->getOperand(0)) &&
423           GEP1->getOperand(1) ==
424           Constant::getNullValue(GEP1->getOperand(1)->getType()))
425      GEP1 = cast<User>(GEP1->getOperand(0));
426    const Value *BasePtr1 = GEP1->getOperand(0);
427
428    while (isGEP(GEP2->getOperand(0)) &&
429           GEP2->getOperand(1) ==
430           Constant::getNullValue(GEP2->getOperand(1)->getType()))
431      GEP2 = cast<User>(GEP2->getOperand(0));
432    const Value *BasePtr2 = GEP2->getOperand(0);
433
434    // Do the base pointers alias?
435    AliasResult BaseAlias = aliasCheck(BasePtr1, ~0U, BasePtr2, ~0U);
436    if (BaseAlias == NoAlias) return NoAlias;
437    if (BaseAlias == MustAlias) {
438      // If the base pointers alias each other exactly, check to see if we can
439      // figure out anything about the resultant pointers, to try to prove
440      // non-aliasing.
441
442      // Collect all of the chained GEP operands together into one simple place
443      SmallVector<Value*, 16> GEP1Ops, GEP2Ops;
444      BasePtr1 = GetGEPOperands(V1, GEP1Ops);
445      BasePtr2 = GetGEPOperands(V2, GEP2Ops);
446
447      // If GetGEPOperands were able to fold to the same must-aliased pointer,
448      // do the comparison.
449      if (BasePtr1 == BasePtr2) {
450        AliasResult GAlias =
451          CheckGEPInstructions(BasePtr1->getType(),
452                               &GEP1Ops[0], GEP1Ops.size(), V1Size,
453                               BasePtr2->getType(),
454                               &GEP2Ops[0], GEP2Ops.size(), V2Size);
455        if (GAlias != MayAlias)
456          return GAlias;
457      }
458    }
459  }
460
461  // Check to see if these two pointers are related by a getelementptr
462  // instruction.  If one pointer is a GEP with a non-zero index of the other
463  // pointer, we know they cannot alias.
464  //
465  if (V1Size == ~0U || V2Size == ~0U)
466    return MayAlias;
467
468  SmallVector<Value*, 16> GEPOperands;
469  const Value *BasePtr = GetGEPOperands(V1, GEPOperands);
470
471  AliasResult R = aliasCheck(BasePtr, ~0U, V2, V2Size);
472  if (R != MustAlias)
473    // If V2 may alias GEP base pointer, conservatively returns MayAlias.
474    // If V2 is known not to alias GEP base pointer, then the two values
475    // cannot alias per GEP semantics: "A pointer value formed from a
476    // getelementptr instruction is associated with the addresses associated
477    // with the first operand of the getelementptr".
478    return R;
479
480  // If there is at least one non-zero constant index, we know they cannot
481  // alias.
482  bool ConstantFound = false;
483  bool AllZerosFound = true;
484  for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i)
485    if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) {
486      if (!C->isNullValue()) {
487        ConstantFound = true;
488        AllZerosFound = false;
489        break;
490      }
491    } else {
492      AllZerosFound = false;
493    }
494
495  // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases
496  // the ptr, the end result is a must alias also.
497  if (AllZerosFound)
498    return MustAlias;
499
500  if (ConstantFound) {
501    if (V2Size <= 1 && V1Size <= 1)  // Just pointer check?
502      return NoAlias;
503
504    // Otherwise we have to check to see that the distance is more than
505    // the size of the argument... build an index vector that is equal to
506    // the arguments provided, except substitute 0's for any variable
507    // indexes we find...
508    if (TD &&
509        cast<PointerType>(BasePtr->getType())->getElementType()->isSized()) {
510      for (unsigned i = 0; i != GEPOperands.size(); ++i)
511        if (!isa<ConstantInt>(GEPOperands[i]))
512          GEPOperands[i] = Constant::getNullValue(GEPOperands[i]->getType());
513      int64_t Offset = TD->getIndexedOffset(BasePtr->getType(),
514                                            &GEPOperands[0],
515                                            GEPOperands.size());
516
517      if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size)
518        return NoAlias;
519    }
520  }
521
522  return MayAlias;
523}
524
525// aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
526// against another.
527AliasAnalysis::AliasResult
528BasicAliasAnalysis::aliasPHI(const PHINode *PN, unsigned PNSize,
529                             const Value *V2, unsigned V2Size) {
530  // The PHI node has already been visited, avoid recursion any further.
531  if (!VisitedPHIs.insert(PN))
532    return MayAlias;
533
534  SmallSet<Value*, 4> UniqueSrc;
535  SmallVector<Value*, 4> V1Srcs;
536  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
537    Value *PV1 = PN->getIncomingValue(i);
538    if (isa<PHINode>(PV1))
539      // If any of the source itself is a PHI, return MayAlias conservatively
540      // to avoid compile time explosion. The worst possible case is if both
541      // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
542      // and 'n' are the number of PHI sources.
543      return MayAlias;
544    if (UniqueSrc.insert(PV1))
545      V1Srcs.push_back(PV1);
546  }
547
548  AliasResult Alias = aliasCheck(V1Srcs[0], PNSize, V2, V2Size);
549  // Early exit if the check of the first PHI source against V2 is MayAlias.
550  // Other results are not possible.
551  if (Alias == MayAlias)
552    return MayAlias;
553
554  // If all sources of the PHI node NoAlias or MustAlias V2, then returns
555  // NoAlias / MustAlias. Otherwise, returns MayAlias.
556  for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
557    Value *V = V1Srcs[i];
558    AliasResult ThisAlias = aliasCheck(V, PNSize, V2, V2Size);
559    if (ThisAlias != Alias || ThisAlias == MayAlias)
560      return MayAlias;
561  }
562
563  return Alias;
564}
565
566// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
567// such as array references.
568//
569AliasAnalysis::AliasResult
570BasicAliasAnalysis::aliasCheck(const Value *V1, unsigned V1Size,
571                               const Value *V2, unsigned V2Size) {
572  // Strip off any casts if they exist.
573  V1 = V1->stripPointerCasts();
574  V2 = V2->stripPointerCasts();
575
576  // Are we checking for alias of the same value?
577  if (V1 == V2) return MustAlias;
578
579  if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType()))
580    return NoAlias;  // Scalars cannot alias each other
581
582  // Figure out what objects these things are pointing to if we can.
583  const Value *O1 = V1->getUnderlyingObject();
584  const Value *O2 = V2->getUnderlyingObject();
585
586  if (O1 != O2) {
587    // If V1/V2 point to two different objects we know that we have no alias.
588    if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
589      return NoAlias;
590
591    // Arguments can't alias with local allocations or noalias calls.
592    if ((isa<Argument>(O1) && (isa<AllocationInst>(O2) || isNoAliasCall(O2))) ||
593        (isa<Argument>(O2) && (isa<AllocationInst>(O1) || isNoAliasCall(O1))))
594      return NoAlias;
595
596    // Most objects can't alias null.
597    if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) ||
598        (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2)))
599      return NoAlias;
600  }
601
602  // If the size of one access is larger than the entire object on the other
603  // side, then we know such behavior is undefined and can assume no alias.
604  LLVMContext &Context = V1->getContext();
605  if (TD)
606    if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, Context, *TD)) ||
607        (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, Context, *TD)))
608      return NoAlias;
609
610  // If one pointer is the result of a call/invoke and the other is a
611  // non-escaping local object, then we know the object couldn't escape to a
612  // point where the call could return it.
613  if ((isa<CallInst>(O1) || isa<InvokeInst>(O1)) &&
614      isNonEscapingLocalObject(O2) && O1 != O2)
615    return NoAlias;
616  if ((isa<CallInst>(O2) || isa<InvokeInst>(O2)) &&
617      isNonEscapingLocalObject(O1) && O1 != O2)
618    return NoAlias;
619
620  if (!isGEP(V1) && isGEP(V2)) {
621    std::swap(V1, V2);
622    std::swap(V1Size, V2Size);
623  }
624  if (isGEP(V1))
625    return aliasGEP(V1, V1Size, V2, V2Size);
626
627  if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
628    std::swap(V1, V2);
629    std::swap(V1Size, V2Size);
630  }
631  if (const PHINode *PN = dyn_cast<PHINode>(V1))
632    return aliasPHI(PN, V1Size, V2, V2Size);
633
634  return MayAlias;
635}
636
637// This function is used to determine if the indices of two GEP instructions are
638// equal. V1 and V2 are the indices.
639static bool IndexOperandsEqual(Value *V1, Value *V2, LLVMContext &Context) {
640  if (V1->getType() == V2->getType())
641    return V1 == V2;
642  if (Constant *C1 = dyn_cast<Constant>(V1))
643    if (Constant *C2 = dyn_cast<Constant>(V2)) {
644      // Sign extend the constants to long types, if necessary
645      if (C1->getType() != Type::getInt64Ty(Context))
646        C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(Context));
647      if (C2->getType() != Type::getInt64Ty(Context))
648        C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(Context));
649      return C1 == C2;
650    }
651  return false;
652}
653
654/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing
655/// base pointers.  This checks to see if the index expressions preclude the
656/// pointers from aliasing...
657AliasAnalysis::AliasResult
658BasicAliasAnalysis::CheckGEPInstructions(
659  const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S,
660  const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) {
661  // We currently can't handle the case when the base pointers have different
662  // primitive types.  Since this is uncommon anyway, we are happy being
663  // extremely conservative.
664  if (BasePtr1Ty != BasePtr2Ty)
665    return MayAlias;
666
667  const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty);
668
669  LLVMContext &Context = GEPPointerTy->getContext();
670
671  // Find the (possibly empty) initial sequence of equal values... which are not
672  // necessarily constants.
673  unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops;
674  unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands);
675  unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands);
676  unsigned UnequalOper = 0;
677  while (UnequalOper != MinOperands &&
678         IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper],
679         Context)) {
680    // Advance through the type as we go...
681    ++UnequalOper;
682    if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
683      BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]);
684    else {
685      // If all operands equal each other, then the derived pointers must
686      // alias each other...
687      BasePtr1Ty = 0;
688      assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands &&
689             "Ran out of type nesting, but not out of operands?");
690      return MustAlias;
691    }
692  }
693
694  // If we have seen all constant operands, and run out of indexes on one of the
695  // getelementptrs, check to see if the tail of the leftover one is all zeros.
696  // If so, return mustalias.
697  if (UnequalOper == MinOperands) {
698    if (NumGEP1Ops < NumGEP2Ops) {
699      std::swap(GEP1Ops, GEP2Ops);
700      std::swap(NumGEP1Ops, NumGEP2Ops);
701    }
702
703    bool AllAreZeros = true;
704    for (unsigned i = UnequalOper; i != MaxOperands; ++i)
705      if (!isa<Constant>(GEP1Ops[i]) ||
706          !cast<Constant>(GEP1Ops[i])->isNullValue()) {
707        AllAreZeros = false;
708        break;
709      }
710    if (AllAreZeros) return MustAlias;
711  }
712
713
714  // So now we know that the indexes derived from the base pointers,
715  // which are known to alias, are different.  We can still determine a
716  // no-alias result if there are differing constant pairs in the index
717  // chain.  For example:
718  //        A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S))
719  //
720  // We have to be careful here about array accesses.  In particular, consider:
721  //        A[1][0] vs A[0][i]
722  // In this case, we don't *know* that the array will be accessed in bounds:
723  // the index could even be negative.  Because of this, we have to
724  // conservatively *give up* and return may alias.  We disregard differing
725  // array subscripts that are followed by a variable index without going
726  // through a struct.
727  //
728  unsigned SizeMax = std::max(G1S, G2S);
729  if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work.
730
731  // Scan for the first operand that is constant and unequal in the
732  // two getelementptrs...
733  unsigned FirstConstantOper = UnequalOper;
734  for (; FirstConstantOper != MinOperands; ++FirstConstantOper) {
735    const Value *G1Oper = GEP1Ops[FirstConstantOper];
736    const Value *G2Oper = GEP2Ops[FirstConstantOper];
737
738    if (G1Oper != G2Oper)   // Found non-equal constant indexes...
739      if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper)))
740        if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){
741          if (G1OC->getType() != G2OC->getType()) {
742            // Sign extend both operands to long.
743            if (G1OC->getType() != Type::getInt64Ty(Context))
744              G1OC = ConstantExpr::getSExt(G1OC, Type::getInt64Ty(Context));
745            if (G2OC->getType() != Type::getInt64Ty(Context))
746              G2OC = ConstantExpr::getSExt(G2OC, Type::getInt64Ty(Context));
747            GEP1Ops[FirstConstantOper] = G1OC;
748            GEP2Ops[FirstConstantOper] = G2OC;
749          }
750
751          if (G1OC != G2OC) {
752            // Handle the "be careful" case above: if this is an array/vector
753            // subscript, scan for a subsequent variable array index.
754            if (const SequentialType *STy =
755                  dyn_cast<SequentialType>(BasePtr1Ty)) {
756              const Type *NextTy = STy;
757              bool isBadCase = false;
758
759              for (unsigned Idx = FirstConstantOper;
760                   Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) {
761                const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx];
762                if (!isa<Constant>(V1) || !isa<Constant>(V2)) {
763                  isBadCase = true;
764                  break;
765                }
766                // If the array is indexed beyond the bounds of the static type
767                // at this level, it will also fall into the "be careful" case.
768                // It would theoretically be possible to analyze these cases,
769                // but for now just be conservatively correct.
770                if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
771                  if (cast<ConstantInt>(G1OC)->getZExtValue() >=
772                        ATy->getNumElements() ||
773                      cast<ConstantInt>(G2OC)->getZExtValue() >=
774                        ATy->getNumElements()) {
775                    isBadCase = true;
776                    break;
777                  }
778                if (const VectorType *VTy = dyn_cast<VectorType>(STy))
779                  if (cast<ConstantInt>(G1OC)->getZExtValue() >=
780                        VTy->getNumElements() ||
781                      cast<ConstantInt>(G2OC)->getZExtValue() >=
782                        VTy->getNumElements()) {
783                    isBadCase = true;
784                    break;
785                  }
786                STy = cast<SequentialType>(NextTy);
787                NextTy = cast<SequentialType>(NextTy)->getElementType();
788              }
789
790              if (isBadCase) G1OC = 0;
791            }
792
793            // Make sure they are comparable (ie, not constant expressions), and
794            // make sure the GEP with the smaller leading constant is GEP1.
795            if (G1OC) {
796              Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT,
797                                                        G1OC, G2OC);
798              if (ConstantInt *CV = dyn_cast<ConstantInt>(Compare)) {
799                if (CV->getZExtValue()) {  // If they are comparable and G2 > G1
800                  std::swap(GEP1Ops, GEP2Ops);  // Make GEP1 < GEP2
801                  std::swap(NumGEP1Ops, NumGEP2Ops);
802                }
803                break;
804              }
805            }
806          }
807        }
808    BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper);
809  }
810
811  // No shared constant operands, and we ran out of common operands.  At this
812  // point, the GEP instructions have run through all of their operands, and we
813  // haven't found evidence that there are any deltas between the GEP's.
814  // However, one GEP may have more operands than the other.  If this is the
815  // case, there may still be hope.  Check this now.
816  if (FirstConstantOper == MinOperands) {
817    // Without TargetData, we won't know what the offsets are.
818    if (!TD)
819      return MayAlias;
820
821    // Make GEP1Ops be the longer one if there is a longer one.
822    if (NumGEP1Ops < NumGEP2Ops) {
823      std::swap(GEP1Ops, GEP2Ops);
824      std::swap(NumGEP1Ops, NumGEP2Ops);
825    }
826
827    // Is there anything to check?
828    if (NumGEP1Ops > MinOperands) {
829      for (unsigned i = FirstConstantOper; i != MaxOperands; ++i)
830        if (isa<ConstantInt>(GEP1Ops[i]) &&
831            !cast<ConstantInt>(GEP1Ops[i])->isZero()) {
832          // Yup, there's a constant in the tail.  Set all variables to
833          // constants in the GEP instruction to make it suitable for
834          // TargetData::getIndexedOffset.
835          for (i = 0; i != MaxOperands; ++i)
836            if (!isa<ConstantInt>(GEP1Ops[i]))
837              GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType());
838          // Okay, now get the offset.  This is the relative offset for the full
839          // instruction.
840          int64_t Offset1 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops,
841                                                 NumGEP1Ops);
842
843          // Now check without any constants at the end.
844          int64_t Offset2 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops,
845                                                 MinOperands);
846
847          // Make sure we compare the absolute difference.
848          if (Offset1 > Offset2)
849            std::swap(Offset1, Offset2);
850
851          // If the tail provided a bit enough offset, return noalias!
852          if ((uint64_t)(Offset2-Offset1) >= SizeMax)
853            return NoAlias;
854          // Otherwise break - we don't look for another constant in the tail.
855          break;
856        }
857    }
858
859    // Couldn't find anything useful.
860    return MayAlias;
861  }
862
863  // If there are non-equal constants arguments, then we can figure
864  // out a minimum known delta between the two index expressions... at
865  // this point we know that the first constant index of GEP1 is less
866  // than the first constant index of GEP2.
867
868  // Advance BasePtr[12]Ty over this first differing constant operand.
869  BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)->
870      getTypeAtIndex(GEP2Ops[FirstConstantOper]);
871  BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->
872      getTypeAtIndex(GEP1Ops[FirstConstantOper]);
873
874  // We are going to be using TargetData::getIndexedOffset to determine the
875  // offset that each of the GEP's is reaching.  To do this, we have to convert
876  // all variable references to constant references.  To do this, we convert the
877  // initial sequence of array subscripts into constant zeros to start with.
878  const Type *ZeroIdxTy = GEPPointerTy;
879  for (unsigned i = 0; i != FirstConstantOper; ++i) {
880    if (!isa<StructType>(ZeroIdxTy))
881      GEP1Ops[i] = GEP2Ops[i] =
882                              Constant::getNullValue(Type::getInt32Ty(Context));
883
884    if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy))
885      ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]);
886  }
887
888  // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok
889
890  // Loop over the rest of the operands...
891  for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) {
892    const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0;
893    const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0;
894    // If they are equal, use a zero index...
895    if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) {
896      if (!isa<ConstantInt>(Op1))
897        GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType());
898      // Otherwise, just keep the constants we have.
899    } else {
900      if (Op1) {
901        if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
902          // If this is an array index, make sure the array element is in range.
903          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) {
904            if (Op1C->getZExtValue() >= AT->getNumElements())
905              return MayAlias;  // Be conservative with out-of-range accesses
906          } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) {
907            if (Op1C->getZExtValue() >= VT->getNumElements())
908              return MayAlias;  // Be conservative with out-of-range accesses
909          }
910
911        } else {
912          // GEP1 is known to produce a value less than GEP2.  To be
913          // conservatively correct, we must assume the largest possible
914          // constant is used in this position.  This cannot be the initial
915          // index to the GEP instructions (because we know we have at least one
916          // element before this one with the different constant arguments), so
917          // we know that the current index must be into either a struct or
918          // array.  Because we know it's not constant, this cannot be a
919          // structure index.  Because of this, we can calculate the maximum
920          // value possible.
921          //
922          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
923            GEP1Ops[i] =
924                  ConstantInt::get(Type::getInt64Ty(Context),
925                                   AT->getNumElements()-1);
926          else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty))
927            GEP1Ops[i] =
928                  ConstantInt::get(Type::getInt64Ty(Context),
929                                   VT->getNumElements()-1);
930        }
931      }
932
933      if (Op2) {
934        if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) {
935          // If this is an array index, make sure the array element is in range.
936          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) {
937            if (Op2C->getZExtValue() >= AT->getNumElements())
938              return MayAlias;  // Be conservative with out-of-range accesses
939          } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) {
940            if (Op2C->getZExtValue() >= VT->getNumElements())
941              return MayAlias;  // Be conservative with out-of-range accesses
942          }
943        } else {  // Conservatively assume the minimum value for this index
944          GEP2Ops[i] = Constant::getNullValue(Op2->getType());
945        }
946      }
947    }
948
949    if (BasePtr1Ty && Op1) {
950      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
951        BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]);
952      else
953        BasePtr1Ty = 0;
954    }
955
956    if (BasePtr2Ty && Op2) {
957      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty))
958        BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]);
959      else
960        BasePtr2Ty = 0;
961    }
962  }
963
964  if (TD && GEPPointerTy->getElementType()->isSized()) {
965    int64_t Offset1 =
966      TD->getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops);
967    int64_t Offset2 =
968      TD->getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops);
969    assert(Offset1 != Offset2 &&
970           "There is at least one different constant here!");
971
972    // Make sure we compare the absolute difference.
973    if (Offset1 > Offset2)
974      std::swap(Offset1, Offset2);
975
976    if ((uint64_t)(Offset2-Offset1) >= SizeMax) {
977      //cerr << "Determined that these two GEP's don't alias ["
978      //     << SizeMax << " bytes]: \n" << *GEP1 << *GEP2;
979      return NoAlias;
980    }
981  }
982  return MayAlias;
983}
984
985// Make sure that anything that uses AliasAnalysis pulls in this file...
986DEFINING_FILE_FOR(BasicAliasAnalysis)
987