BasicAliasAnalysis.cpp revision 42e3c81c5f23c42ba2415a8f797d430dad387734
1//===- BasicAliasAnalysis.cpp - Local Alias Analysis Impl -----------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by the LLVM research group and is distributed under
6// the University of Illinois Open Source License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines the default implementation of the Alias Analysis interface
11// that simply implements a few identities (two different globals cannot alias,
12// etc), but otherwise does no analysis.
13//
14//===----------------------------------------------------------------------===//
15
16#include "llvm/Analysis/AliasAnalysis.h"
17#include "llvm/Analysis/Passes.h"
18#include "llvm/Constants.h"
19#include "llvm/DerivedTypes.h"
20#include "llvm/Function.h"
21#include "llvm/GlobalVariable.h"
22#include "llvm/Instructions.h"
23#include "llvm/Pass.h"
24#include "llvm/Target/TargetData.h"
25#include "llvm/Support/GetElementPtrTypeIterator.h"
26#include <algorithm>
27using namespace llvm;
28
29// Make sure that anything that uses AliasAnalysis pulls in this file...
30void llvm::BasicAAStub() {}
31
32namespace {
33  /// NoAA - This class implements the -no-aa pass, which always returns "I
34  /// don't know" for alias queries.  NoAA is unlike other alias analysis
35  /// implementations, in that it does not chain to a previous analysis.  As
36  /// such it doesn't follow many of the rules that other alias analyses must.
37  ///
38  struct NoAA : public ImmutablePass, public AliasAnalysis {
39    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
40      AU.addRequired<TargetData>();
41    }
42
43    virtual void initializePass() {
44      TD = &getAnalysis<TargetData>();
45    }
46
47    virtual AliasResult alias(const Value *V1, unsigned V1Size,
48                              const Value *V2, unsigned V2Size) {
49      return MayAlias;
50    }
51
52    virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
53                                         std::vector<PointerAccessInfo> *Info) {
54      return UnknownModRefBehavior;
55    }
56
57    virtual void getArgumentAccesses(Function *F, CallSite CS,
58                                     std::vector<PointerAccessInfo> &Info) {
59      assert(0 && "This method may not be called on this function!");
60    }
61
62    virtual void getMustAliases(Value *P, std::vector<Value*> &RetVals) { }
63    virtual bool pointsToConstantMemory(const Value *P) { return false; }
64    virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) {
65      return ModRef;
66    }
67    virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
68      return ModRef;
69    }
70    virtual bool hasNoModRefInfoForCalls() const { return true; }
71
72    virtual void deleteValue(Value *V) {}
73    virtual void copyValue(Value *From, Value *To) {}
74  };
75
76  // Register this pass...
77  RegisterOpt<NoAA>
78  U("no-aa", "No Alias Analysis (always returns 'may' alias)");
79
80  // Declare that we implement the AliasAnalysis interface
81  RegisterAnalysisGroup<AliasAnalysis, NoAA> V;
82}  // End of anonymous namespace
83
84ImmutablePass *llvm::createNoAAPass() { return new NoAA(); }
85
86namespace {
87  /// BasicAliasAnalysis - This is the default alias analysis implementation.
88  /// Because it doesn't chain to a previous alias analysis (like -no-aa), it
89  /// derives from the NoAA class.
90  struct BasicAliasAnalysis : public NoAA {
91    AliasResult alias(const Value *V1, unsigned V1Size,
92                      const Value *V2, unsigned V2Size);
93
94    ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
95    ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
96      return NoAA::getModRefInfo(CS1,CS2);
97    }
98
99    /// hasNoModRefInfoForCalls - We can provide mod/ref information against
100    /// non-escaping allocations.
101    virtual bool hasNoModRefInfoForCalls() const { return false; }
102
103    /// pointsToConstantMemory - Chase pointers until we find a (constant
104    /// global) or not.
105    bool pointsToConstantMemory(const Value *P);
106
107    virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
108                                             std::vector<PointerAccessInfo> *Info);
109
110  private:
111    // CheckGEPInstructions - Check two GEP instructions with known
112    // must-aliasing base pointers.  This checks to see if the index expressions
113    // preclude the pointers from aliasing...
114    AliasResult
115    CheckGEPInstructions(const Type* BasePtr1Ty, std::vector<Value*> &GEP1Ops,
116                         unsigned G1Size,
117                         const Type *BasePtr2Ty, std::vector<Value*> &GEP2Ops,
118                         unsigned G2Size);
119  };
120
121  // Register this pass...
122  RegisterOpt<BasicAliasAnalysis>
123  X("basicaa", "Basic Alias Analysis (default AA impl)");
124
125  // Declare that we implement the AliasAnalysis interface
126  RegisterAnalysisGroup<AliasAnalysis, BasicAliasAnalysis, true> Y;
127}  // End of anonymous namespace
128
129ImmutablePass *llvm::createBasicAliasAnalysisPass() {
130  return new BasicAliasAnalysis();
131}
132
133// hasUniqueAddress - Return true if the specified value points to something
134// with a unique, discernable, address.
135static inline bool hasUniqueAddress(const Value *V) {
136  return isa<GlobalValue>(V) || isa<AllocationInst>(V);
137}
138
139// getUnderlyingObject - This traverses the use chain to figure out what object
140// the specified value points to.  If the value points to, or is derived from, a
141// unique object or an argument, return it.
142static const Value *getUnderlyingObject(const Value *V) {
143  if (!isa<PointerType>(V->getType())) return 0;
144
145  // If we are at some type of object... return it.
146  if (hasUniqueAddress(V) || isa<Argument>(V)) return V;
147
148  // Traverse through different addressing mechanisms...
149  if (const Instruction *I = dyn_cast<Instruction>(V)) {
150    if (isa<CastInst>(I) || isa<GetElementPtrInst>(I))
151      return getUnderlyingObject(I->getOperand(0));
152  } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
153    if (CE->getOpcode() == Instruction::Cast ||
154        CE->getOpcode() == Instruction::GetElementPtr)
155      return getUnderlyingObject(CE->getOperand(0));
156  } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
157    return GV;
158  }
159  return 0;
160}
161
162static const User *isGEP(const Value *V) {
163  if (isa<GetElementPtrInst>(V) ||
164      (isa<ConstantExpr>(V) &&
165       cast<ConstantExpr>(V)->getOpcode() == Instruction::GetElementPtr))
166    return cast<User>(V);
167  return 0;
168}
169
170static const Value *GetGEPOperands(const Value *V, std::vector<Value*> &GEPOps){
171  assert(GEPOps.empty() && "Expect empty list to populate!");
172  GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1,
173                cast<User>(V)->op_end());
174
175  // Accumulate all of the chained indexes into the operand array
176  V = cast<User>(V)->getOperand(0);
177
178  while (const User *G = isGEP(V)) {
179    if (!isa<Constant>(GEPOps[0]) || isa<GlobalValue>(GEPOps[0]) ||
180        !cast<Constant>(GEPOps[0])->isNullValue())
181      break;  // Don't handle folding arbitrary pointer offsets yet...
182    GEPOps.erase(GEPOps.begin());   // Drop the zero index
183    GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end());
184    V = G->getOperand(0);
185  }
186  return V;
187}
188
189/// pointsToConstantMemory - Chase pointers until we find a (constant
190/// global) or not.
191bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) {
192  if (const Value *V = getUnderlyingObject(P))
193    if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
194      return GV->isConstant();
195  return false;
196}
197
198static bool AddressMightEscape(const Value *V) {
199  for (Value::use_const_iterator UI = V->use_begin(), E = V->use_end();
200       UI != E; ++UI) {
201    const Instruction *I = cast<Instruction>(*UI);
202    switch (I->getOpcode()) {
203    case Instruction::Load: break;
204    case Instruction::Store:
205      if (I->getOperand(0) == V)
206        return true; // Escapes if the pointer is stored.
207      break;
208    case Instruction::GetElementPtr:
209      if (AddressMightEscape(I)) return true;
210      break;
211    case Instruction::Cast:
212      if (!isa<PointerType>(I->getType()))
213        return true;
214      if (AddressMightEscape(I)) return true;
215      break;
216    case Instruction::Ret:
217      // If returned, the address will escape to calling functions, but no
218      // callees could modify it.
219      break;
220    default:
221      return true;
222    }
223  }
224  return false;
225}
226
227// getModRefInfo - Check to see if the specified callsite can clobber the
228// specified memory object.  Since we only look at local properties of this
229// function, we really can't say much about this query.  We do, however, use
230// simple "address taken" analysis on local objects.
231//
232AliasAnalysis::ModRefResult
233BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
234  if (!isa<Constant>(P))
235    if (const AllocationInst *AI =
236                  dyn_cast_or_null<AllocationInst>(getUnderlyingObject(P))) {
237      // Okay, the pointer is to a stack allocated object.  If we can prove that
238      // the pointer never "escapes", then we know the call cannot clobber it,
239      // because it simply can't get its address.
240      if (!AddressMightEscape(AI))
241        return NoModRef;
242
243      // If this is a tail call and P points to a stack location, we know that
244      // the tail call cannot access or modify the local stack.
245      if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
246        if (CI->isTailCall() && isa<AllocaInst>(AI))
247          return NoModRef;
248    }
249
250  // The AliasAnalysis base class has some smarts, lets use them.
251  return AliasAnalysis::getModRefInfo(CS, P, Size);
252}
253
254// alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such
255// as array references.  Note that this function is heavily tail recursive.
256// Hopefully we have a smart C++ compiler.  :)
257//
258AliasAnalysis::AliasResult
259BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size,
260                          const Value *V2, unsigned V2Size) {
261  // Strip off any constant expression casts if they exist
262  if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V1))
263    if (CE->getOpcode() == Instruction::Cast &&
264        isa<PointerType>(CE->getOperand(0)->getType()))
265      V1 = CE->getOperand(0);
266  if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V2))
267    if (CE->getOpcode() == Instruction::Cast &&
268        isa<PointerType>(CE->getOperand(0)->getType()))
269      V2 = CE->getOperand(0);
270
271  // Are we checking for alias of the same value?
272  if (V1 == V2) return MustAlias;
273
274  if ((!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) &&
275      V1->getType() != Type::LongTy && V2->getType() != Type::LongTy)
276    return NoAlias;  // Scalars cannot alias each other
277
278  // Strip off cast instructions...
279  if (const Instruction *I = dyn_cast<CastInst>(V1))
280    if (isa<PointerType>(I->getOperand(0)->getType()))
281      return alias(I->getOperand(0), V1Size, V2, V2Size);
282  if (const Instruction *I = dyn_cast<CastInst>(V2))
283    if (isa<PointerType>(I->getOperand(0)->getType()))
284      return alias(V1, V1Size, I->getOperand(0), V2Size);
285
286  // Figure out what objects these things are pointing to if we can...
287  const Value *O1 = getUnderlyingObject(V1);
288  const Value *O2 = getUnderlyingObject(V2);
289
290  // Pointing at a discernible object?
291  if (O1) {
292    if (O2) {
293      if (isa<Argument>(O1)) {
294        // Incoming argument cannot alias locally allocated object!
295        if (isa<AllocationInst>(O2)) return NoAlias;
296        // Otherwise, nothing is known...
297      } else if (isa<Argument>(O2)) {
298        // Incoming argument cannot alias locally allocated object!
299        if (isa<AllocationInst>(O1)) return NoAlias;
300        // Otherwise, nothing is known...
301      } else if (O1 != O2) {
302        // If they are two different objects, we know that we have no alias...
303        return NoAlias;
304      }
305
306      // If they are the same object, they we can look at the indexes.  If they
307      // index off of the object is the same for both pointers, they must alias.
308      // If they are provably different, they must not alias.  Otherwise, we
309      // can't tell anything.
310    }
311
312
313    if (!isa<Argument>(O1) && isa<ConstantPointerNull>(V2))
314      return NoAlias;                    // Unique values don't alias null
315
316    if (isa<GlobalVariable>(O1) ||
317        (isa<AllocationInst>(O1) &&
318         !cast<AllocationInst>(O1)->isArrayAllocation()))
319      if (cast<PointerType>(O1->getType())->getElementType()->isSized()) {
320        // If the size of the other access is larger than the total size of the
321        // global/alloca/malloc, it cannot be accessing the global (it's
322        // undefined to load or store bytes before or after an object).
323        const Type *ElTy = cast<PointerType>(O1->getType())->getElementType();
324        unsigned GlobalSize = getTargetData().getTypeSize(ElTy);
325        if (GlobalSize < V2Size && V2Size != ~0U)
326          return NoAlias;
327      }
328  }
329
330  if (O2) {
331    if (!isa<Argument>(O2) && isa<ConstantPointerNull>(V1))
332      return NoAlias;                    // Unique values don't alias null
333
334    if (isa<GlobalVariable>(O2) ||
335        (isa<AllocationInst>(O2) &&
336         !cast<AllocationInst>(O2)->isArrayAllocation()))
337      if (cast<PointerType>(O2->getType())->getElementType()->isSized()) {
338        // If the size of the other access is larger than the total size of the
339        // global/alloca/malloc, it cannot be accessing the object (it's
340        // undefined to load or store bytes before or after an object).
341        const Type *ElTy = cast<PointerType>(O2->getType())->getElementType();
342        unsigned GlobalSize = getTargetData().getTypeSize(ElTy);
343        if (GlobalSize < V1Size && V1Size != ~0U)
344          return NoAlias;
345      }
346  }
347
348  // If we have two gep instructions with must-alias'ing base pointers, figure
349  // out if the indexes to the GEP tell us anything about the derived pointer.
350  // Note that we also handle chains of getelementptr instructions as well as
351  // constant expression getelementptrs here.
352  //
353  if (isGEP(V1) && isGEP(V2)) {
354    // Drill down into the first non-gep value, to test for must-aliasing of
355    // the base pointers.
356    const Value *BasePtr1 = V1, *BasePtr2 = V2;
357    do {
358      BasePtr1 = cast<User>(BasePtr1)->getOperand(0);
359    } while (isGEP(BasePtr1) &&
360             cast<User>(BasePtr1)->getOperand(1) ==
361       Constant::getNullValue(cast<User>(BasePtr1)->getOperand(1)->getType()));
362    do {
363      BasePtr2 = cast<User>(BasePtr2)->getOperand(0);
364    } while (isGEP(BasePtr2) &&
365             cast<User>(BasePtr2)->getOperand(1) ==
366       Constant::getNullValue(cast<User>(BasePtr2)->getOperand(1)->getType()));
367
368    // Do the base pointers alias?
369    AliasResult BaseAlias = alias(BasePtr1, V1Size, BasePtr2, V2Size);
370    if (BaseAlias == NoAlias) return NoAlias;
371    if (BaseAlias == MustAlias) {
372      // If the base pointers alias each other exactly, check to see if we can
373      // figure out anything about the resultant pointers, to try to prove
374      // non-aliasing.
375
376      // Collect all of the chained GEP operands together into one simple place
377      std::vector<Value*> GEP1Ops, GEP2Ops;
378      BasePtr1 = GetGEPOperands(V1, GEP1Ops);
379      BasePtr2 = GetGEPOperands(V2, GEP2Ops);
380
381      // If GetGEPOperands were able to fold to the same must-aliased pointer,
382      // do the comparison.
383      if (BasePtr1 == BasePtr2) {
384        AliasResult GAlias =
385          CheckGEPInstructions(BasePtr1->getType(), GEP1Ops, V1Size,
386                               BasePtr2->getType(), GEP2Ops, V2Size);
387        if (GAlias != MayAlias)
388          return GAlias;
389      }
390    }
391  }
392
393  // Check to see if these two pointers are related by a getelementptr
394  // instruction.  If one pointer is a GEP with a non-zero index of the other
395  // pointer, we know they cannot alias.
396  //
397  if (isGEP(V2)) {
398    std::swap(V1, V2);
399    std::swap(V1Size, V2Size);
400  }
401
402  if (V1Size != ~0U && V2Size != ~0U)
403    if (const User *GEP = isGEP(V1)) {
404      std::vector<Value*> GEPOperands;
405      const Value *BasePtr = GetGEPOperands(V1, GEPOperands);
406
407      AliasResult R = alias(BasePtr, V1Size, V2, V2Size);
408      if (R == MustAlias) {
409        // If there is at least one non-zero constant index, we know they cannot
410        // alias.
411        bool ConstantFound = false;
412        bool AllZerosFound = true;
413        for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i)
414          if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) {
415            if (!C->isNullValue()) {
416              ConstantFound = true;
417              AllZerosFound = false;
418              break;
419            }
420          } else {
421            AllZerosFound = false;
422          }
423
424        // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases
425        // the ptr, the end result is a must alias also.
426        if (AllZerosFound)
427          return MustAlias;
428
429        if (ConstantFound) {
430          if (V2Size <= 1 && V1Size <= 1)  // Just pointer check?
431            return NoAlias;
432
433          // Otherwise we have to check to see that the distance is more than
434          // the size of the argument... build an index vector that is equal to
435          // the arguments provided, except substitute 0's for any variable
436          // indexes we find...
437          if (cast<PointerType>(
438                BasePtr->getType())->getElementType()->isSized()) {
439            for (unsigned i = 0; i != GEPOperands.size(); ++i)
440              if (!isa<ConstantInt>(GEPOperands[i]))
441                GEPOperands[i] =
442                  Constant::getNullValue(GEPOperands[i]->getType());
443            int64_t Offset =
444              getTargetData().getIndexedOffset(BasePtr->getType(), GEPOperands);
445
446            if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size)
447              return NoAlias;
448          }
449        }
450      }
451    }
452
453  return MayAlias;
454}
455
456static bool ValuesEqual(Value *V1, Value *V2) {
457  if (V1->getType() == V2->getType())
458    return V1 == V2;
459  if (Constant *C1 = dyn_cast<Constant>(V1))
460    if (Constant *C2 = dyn_cast<Constant>(V2)) {
461      // Sign extend the constants to long types.
462      C1 = ConstantExpr::getSignExtend(C1, Type::LongTy);
463      C2 = ConstantExpr::getSignExtend(C2, Type::LongTy);
464      return C1 == C2;
465    }
466  return false;
467}
468
469/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing
470/// base pointers.  This checks to see if the index expressions preclude the
471/// pointers from aliasing...
472AliasAnalysis::AliasResult BasicAliasAnalysis::
473CheckGEPInstructions(const Type* BasePtr1Ty, std::vector<Value*> &GEP1Ops,
474                     unsigned G1S,
475                     const Type *BasePtr2Ty, std::vector<Value*> &GEP2Ops,
476                     unsigned G2S) {
477  // We currently can't handle the case when the base pointers have different
478  // primitive types.  Since this is uncommon anyway, we are happy being
479  // extremely conservative.
480  if (BasePtr1Ty != BasePtr2Ty)
481    return MayAlias;
482
483  const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty);
484
485  // Find the (possibly empty) initial sequence of equal values... which are not
486  // necessarily constants.
487  unsigned NumGEP1Operands = GEP1Ops.size(), NumGEP2Operands = GEP2Ops.size();
488  unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands);
489  unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands);
490  unsigned UnequalOper = 0;
491  while (UnequalOper != MinOperands &&
492         ValuesEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) {
493    // Advance through the type as we go...
494    ++UnequalOper;
495    if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
496      BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]);
497    else {
498      // If all operands equal each other, then the derived pointers must
499      // alias each other...
500      BasePtr1Ty = 0;
501      assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands &&
502             "Ran out of type nesting, but not out of operands?");
503      return MustAlias;
504    }
505  }
506
507  // If we have seen all constant operands, and run out of indexes on one of the
508  // getelementptrs, check to see if the tail of the leftover one is all zeros.
509  // If so, return mustalias.
510  if (UnequalOper == MinOperands) {
511    if (GEP1Ops.size() < GEP2Ops.size()) std::swap(GEP1Ops, GEP2Ops);
512
513    bool AllAreZeros = true;
514    for (unsigned i = UnequalOper; i != MaxOperands; ++i)
515      if (!isa<Constant>(GEP1Ops[i]) ||
516          !cast<Constant>(GEP1Ops[i])->isNullValue()) {
517        AllAreZeros = false;
518        break;
519      }
520    if (AllAreZeros) return MustAlias;
521  }
522
523
524  // So now we know that the indexes derived from the base pointers,
525  // which are known to alias, are different.  We can still determine a
526  // no-alias result if there are differing constant pairs in the index
527  // chain.  For example:
528  //        A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S))
529  //
530  unsigned SizeMax = std::max(G1S, G2S);
531  if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work.
532
533  // Scan for the first operand that is constant and unequal in the
534  // two getelementptrs...
535  unsigned FirstConstantOper = UnequalOper;
536  for (; FirstConstantOper != MinOperands; ++FirstConstantOper) {
537    const Value *G1Oper = GEP1Ops[FirstConstantOper];
538    const Value *G2Oper = GEP2Ops[FirstConstantOper];
539
540    if (G1Oper != G2Oper)   // Found non-equal constant indexes...
541      if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper)))
542        if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){
543          if (G1OC->getType() != G2OC->getType()) {
544            // Sign extend both operands to long.
545            G1OC = ConstantExpr::getSignExtend(G1OC, Type::LongTy);
546            G2OC = ConstantExpr::getSignExtend(G2OC, Type::LongTy);
547            GEP1Ops[FirstConstantOper] = G1OC;
548            GEP2Ops[FirstConstantOper] = G2OC;
549          }
550
551          if (G1OC != G2OC) {
552            // Make sure they are comparable (ie, not constant expressions), and
553            // make sure the GEP with the smaller leading constant is GEP1.
554            Constant *Compare = ConstantExpr::getSetGT(G1OC, G2OC);
555            if (ConstantBool *CV = dyn_cast<ConstantBool>(Compare)) {
556              if (CV->getValue())   // If they are comparable and G2 > G1
557                std::swap(GEP1Ops, GEP2Ops);  // Make GEP1 < GEP2
558              break;
559            }
560          }
561        }
562    BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper);
563  }
564
565  // No shared constant operands, and we ran out of common operands.  At this
566  // point, the GEP instructions have run through all of their operands, and we
567  // haven't found evidence that there are any deltas between the GEP's.
568  // However, one GEP may have more operands than the other.  If this is the
569  // case, there may still be hope.  Check this now.
570  if (FirstConstantOper == MinOperands) {
571    // Make GEP1Ops be the longer one if there is a longer one.
572    if (GEP1Ops.size() < GEP2Ops.size())
573      std::swap(GEP1Ops, GEP2Ops);
574
575    // Is there anything to check?
576    if (GEP1Ops.size() > MinOperands) {
577      for (unsigned i = FirstConstantOper; i != MaxOperands; ++i)
578        if (isa<ConstantInt>(GEP1Ops[i]) &&
579            !cast<Constant>(GEP1Ops[i])->isNullValue()) {
580          // Yup, there's a constant in the tail.  Set all variables to
581          // constants in the GEP instruction to make it suiteable for
582          // TargetData::getIndexedOffset.
583          for (i = 0; i != MaxOperands; ++i)
584            if (!isa<ConstantInt>(GEP1Ops[i]))
585              GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType());
586          // Okay, now get the offset.  This is the relative offset for the full
587          // instruction.
588          const TargetData &TD = getTargetData();
589          int64_t Offset1 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops);
590
591          // Now crop off any constants from the end...
592          GEP1Ops.resize(MinOperands);
593          int64_t Offset2 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops);
594
595          // If the tail provided a bit enough offset, return noalias!
596          if ((uint64_t)(Offset2-Offset1) >= SizeMax)
597            return NoAlias;
598        }
599    }
600
601    // Couldn't find anything useful.
602    return MayAlias;
603  }
604
605  // If there are non-equal constants arguments, then we can figure
606  // out a minimum known delta between the two index expressions... at
607  // this point we know that the first constant index of GEP1 is less
608  // than the first constant index of GEP2.
609
610  // Advance BasePtr[12]Ty over this first differing constant operand.
611  BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(GEP2Ops[FirstConstantOper]);
612  BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(GEP1Ops[FirstConstantOper]);
613
614  // We are going to be using TargetData::getIndexedOffset to determine the
615  // offset that each of the GEP's is reaching.  To do this, we have to convert
616  // all variable references to constant references.  To do this, we convert the
617  // initial equal sequence of variables into constant zeros to start with.
618  for (unsigned i = 0; i != FirstConstantOper; ++i)
619    if (!isa<ConstantInt>(GEP1Ops[i]) || !isa<ConstantInt>(GEP2Ops[i]))
620      GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Type::UIntTy);
621
622  // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok
623
624  // Loop over the rest of the operands...
625  for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) {
626    const Value *Op1 = i < GEP1Ops.size() ? GEP1Ops[i] : 0;
627    const Value *Op2 = i < GEP2Ops.size() ? GEP2Ops[i] : 0;
628    // If they are equal, use a zero index...
629    if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) {
630      if (!isa<ConstantInt>(Op1))
631        GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType());
632      // Otherwise, just keep the constants we have.
633    } else {
634      if (Op1) {
635        if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
636          // If this is an array index, make sure the array element is in range.
637          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
638            if (Op1C->getRawValue() >= AT->getNumElements())
639              return MayAlias;  // Be conservative with out-of-range accesses
640
641        } else {
642          // GEP1 is known to produce a value less than GEP2.  To be
643          // conservatively correct, we must assume the largest possible
644          // constant is used in this position.  This cannot be the initial
645          // index to the GEP instructions (because we know we have at least one
646          // element before this one with the different constant arguments), so
647          // we know that the current index must be into either a struct or
648          // array.  Because we know it's not constant, this cannot be a
649          // structure index.  Because of this, we can calculate the maximum
650          // value possible.
651          //
652          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
653            GEP1Ops[i] = ConstantSInt::get(Type::LongTy,AT->getNumElements()-1);
654        }
655      }
656
657      if (Op2) {
658        if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) {
659          // If this is an array index, make sure the array element is in range.
660          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
661            if (Op2C->getRawValue() >= AT->getNumElements())
662              return MayAlias;  // Be conservative with out-of-range accesses
663        } else {  // Conservatively assume the minimum value for this index
664          GEP2Ops[i] = Constant::getNullValue(Op2->getType());
665        }
666      }
667    }
668
669    if (BasePtr1Ty && Op1) {
670      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
671        BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]);
672      else
673        BasePtr1Ty = 0;
674    }
675
676    if (BasePtr2Ty && Op2) {
677      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty))
678        BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]);
679      else
680        BasePtr2Ty = 0;
681    }
682  }
683
684  if (GEPPointerTy->getElementType()->isSized()) {
685    int64_t Offset1 = getTargetData().getIndexedOffset(GEPPointerTy, GEP1Ops);
686    int64_t Offset2 = getTargetData().getIndexedOffset(GEPPointerTy, GEP2Ops);
687    assert(Offset1<Offset2 && "There is at least one different constant here!");
688
689    if ((uint64_t)(Offset2-Offset1) >= SizeMax) {
690      //std::cerr << "Determined that these two GEP's don't alias ["
691      //          << SizeMax << " bytes]: \n" << *GEP1 << *GEP2;
692      return NoAlias;
693    }
694  }
695  return MayAlias;
696}
697
698namespace {
699  struct StringCompare {
700    bool operator()(const char *LHS, const char *RHS) {
701      return strcmp(LHS, RHS) < 0;
702    }
703  };
704}
705
706// Note that this list cannot contain libm functions (such as acos and sqrt)
707// that set errno on a domain or other error.
708static const char *DoesntAccessMemoryTable[] = {
709  // LLVM intrinsics:
710  "llvm.frameaddress", "llvm.returnaddress", "llvm.readport",
711  "llvm.isunordered", "llvm.sqrt", "llvm.ctpop", "llvm.ctlz", "llvm.cttz",
712
713  "abs", "labs", "llabs", "imaxabs", "fabs", "fabsf", "fabsl",
714  "trunc", "truncf", "truncl", "ldexp",
715
716  "atan", "atanf", "atanl",   "atan2", "atan2f", "atan2l",
717  "cbrt",
718  "cos", "cosf", "cosl",
719  "exp", "expf", "expl",
720  "hypot",
721  "sin", "sinf", "sinl",
722  "tan", "tanf", "tanl",      "tanh", "tanhf", "tanhl",
723
724  // ctype.h
725  "isalnum", "isalpha", "iscntrl", "isdigit", "isgraph", "islower", "isprint"
726  "ispunct", "isspace", "isupper", "isxdigit", "tolower", "toupper",
727
728  // wctype.h"
729  "iswalnum", "iswalpha", "iswcntrl", "iswdigit", "iswgraph", "iswlower",
730  "iswprint", "iswpunct", "iswspace", "iswupper", "iswxdigit",
731
732  "iswctype", "towctrans", "towlower", "towupper",
733
734  "btowc", "wctob",
735
736  "isinf", "isnan", "finite",
737
738  // C99 math functions
739  "copysign", "copysignf", "copysignd",
740  "nexttoward", "nexttowardf", "nexttowardd",
741  "nextafter", "nextafterf", "nextafterd",
742
743  // glibc functions:
744  "__fpclassify", "__fpclassifyf", "__fpclassifyl",
745  "__signbit", "__signbitf", "__signbitl",
746};
747
748static const unsigned DAMTableSize =
749    sizeof(DoesntAccessMemoryTable)/sizeof(DoesntAccessMemoryTable[0]);
750
751static const char *OnlyReadsMemoryTable[] = {
752  "atoi", "atol", "atof", "atoll", "atoq", "a64l",
753  "bcmp", "memcmp", "memchr", "memrchr", "wmemcmp", "wmemchr",
754
755  // Strings
756  "strcmp", "strcasecmp", "strcoll", "strncmp", "strncasecmp",
757  "strchr", "strcspn", "strlen", "strpbrk", "strrchr", "strspn", "strstr",
758  "index", "rindex",
759
760  // Wide char strings
761  "wcschr", "wcscmp", "wcscoll", "wcscspn", "wcslen", "wcsncmp", "wcspbrk",
762  "wcsrchr", "wcsspn", "wcsstr",
763
764  // glibc
765  "alphasort", "alphasort64", "versionsort", "versionsort64",
766
767  // C99
768  "nan", "nanf", "nand",
769
770  // File I/O
771  "feof", "ferror", "fileno",
772  "feof_unlocked", "ferror_unlocked", "fileno_unlocked"
773};
774
775static const unsigned ORMTableSize =
776    sizeof(OnlyReadsMemoryTable)/sizeof(OnlyReadsMemoryTable[0]);
777
778AliasAnalysis::ModRefBehavior
779BasicAliasAnalysis::getModRefBehavior(Function *F, CallSite CS,
780                                      std::vector<PointerAccessInfo> *Info) {
781  if (!F->isExternal()) return UnknownModRefBehavior;
782
783  static bool Initialized = false;
784  if (!Initialized) {
785    // Sort the table the first time through.
786    std::sort(DoesntAccessMemoryTable, DoesntAccessMemoryTable+DAMTableSize,
787              StringCompare());
788    std::sort(OnlyReadsMemoryTable, OnlyReadsMemoryTable+ORMTableSize,
789              StringCompare());
790    Initialized = true;
791  }
792
793  const char **Ptr = std::lower_bound(DoesntAccessMemoryTable,
794                                      DoesntAccessMemoryTable+DAMTableSize,
795                                      F->getName().c_str(), StringCompare());
796  if (Ptr != DoesntAccessMemoryTable+DAMTableSize && *Ptr == F->getName())
797    return DoesNotAccessMemory;
798
799  Ptr = std::lower_bound(OnlyReadsMemoryTable,
800                         OnlyReadsMemoryTable+ORMTableSize,
801                         F->getName().c_str(), StringCompare());
802  if (Ptr != OnlyReadsMemoryTable+ORMTableSize && *Ptr == F->getName())
803    return OnlyReadsMemory;
804
805  return UnknownModRefBehavior;
806}
807