BasicAliasAnalysis.cpp revision 7765d71304f0f26ab348deb41af9b6ae033aae4d
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/Compiler.h"
26#include "llvm/Support/GetElementPtrTypeIterator.h"
27#include "llvm/Support/ManagedStatic.h"
28#include <algorithm>
29using namespace llvm;
30
31namespace {
32  /// NoAA - This class implements the -no-aa pass, which always returns "I
33  /// don't know" for alias queries.  NoAA is unlike other alias analysis
34  /// implementations, in that it does not chain to a previous analysis.  As
35  /// such it doesn't follow many of the rules that other alias analyses must.
36  ///
37  struct VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis {
38    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
39      AU.addRequired<TargetData>();
40    }
41
42    virtual void initializePass() {
43      TD = &getAnalysis<TargetData>();
44    }
45
46    virtual AliasResult alias(const Value *V1, unsigned V1Size,
47                              const Value *V2, unsigned V2Size) {
48      return MayAlias;
49    }
50
51    virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
52                                         std::vector<PointerAccessInfo> *Info) {
53      return UnknownModRefBehavior;
54    }
55
56    virtual void getArgumentAccesses(Function *F, CallSite CS,
57                                     std::vector<PointerAccessInfo> &Info) {
58      assert(0 && "This method may not be called on this function!");
59    }
60
61    virtual void getMustAliases(Value *P, std::vector<Value*> &RetVals) { }
62    virtual bool pointsToConstantMemory(const Value *P) { return false; }
63    virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) {
64      return ModRef;
65    }
66    virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
67      return ModRef;
68    }
69    virtual bool hasNoModRefInfoForCalls() const { return true; }
70
71    virtual void deleteValue(Value *V) {}
72    virtual void copyValue(Value *From, Value *To) {}
73  };
74
75  // Register this pass...
76  RegisterPass<NoAA>
77  U("no-aa", "No Alias Analysis (always returns 'may' alias)");
78
79  // Declare that we implement the AliasAnalysis interface
80  RegisterAnalysisGroup<AliasAnalysis> V(U);
81}  // End of anonymous namespace
82
83ImmutablePass *llvm::createNoAAPass() { return new NoAA(); }
84
85namespace {
86  /// BasicAliasAnalysis - This is the default alias analysis implementation.
87  /// Because it doesn't chain to a previous alias analysis (like -no-aa), it
88  /// derives from the NoAA class.
89  struct VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA {
90    AliasResult alias(const Value *V1, unsigned V1Size,
91                      const Value *V2, unsigned V2Size);
92
93    ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
94    ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
95      return NoAA::getModRefInfo(CS1,CS2);
96    }
97
98    /// hasNoModRefInfoForCalls - We can provide mod/ref information against
99    /// non-escaping allocations.
100    virtual bool hasNoModRefInfoForCalls() const { return false; }
101
102    /// pointsToConstantMemory - Chase pointers until we find a (constant
103    /// global) or not.
104    bool pointsToConstantMemory(const Value *P);
105
106    virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
107                                             std::vector<PointerAccessInfo> *Info);
108
109  private:
110    // CheckGEPInstructions - Check two GEP instructions with known
111    // must-aliasing base pointers.  This checks to see if the index expressions
112    // preclude the pointers from aliasing...
113    AliasResult
114    CheckGEPInstructions(const Type* BasePtr1Ty, std::vector<Value*> &GEP1Ops,
115                         unsigned G1Size,
116                         const Type *BasePtr2Ty, std::vector<Value*> &GEP2Ops,
117                         unsigned G2Size);
118  };
119
120  // Register this pass...
121  RegisterPass<BasicAliasAnalysis>
122  X("basicaa", "Basic Alias Analysis (default AA impl)");
123
124  // Declare that we implement the AliasAnalysis interface
125  RegisterAnalysisGroup<AliasAnalysis, true> Y(X);
126}  // End of anonymous namespace
127
128ImmutablePass *llvm::createBasicAliasAnalysisPass() {
129  return new BasicAliasAnalysis();
130}
131
132// hasUniqueAddress - Return true if the specified value points to something
133// with a unique, discernable, address.
134static inline bool hasUniqueAddress(const Value *V) {
135  return isa<GlobalValue>(V) || isa<AllocationInst>(V);
136}
137
138// getUnderlyingObject - This traverses the use chain to figure out what object
139// the specified value points to.  If the value points to, or is derived from, a
140// unique object or an argument, return it.
141static const Value *getUnderlyingObject(const Value *V) {
142  if (!isa<PointerType>(V->getType())) return 0;
143
144  // If we are at some type of object... return it.
145  if (hasUniqueAddress(V) || isa<Argument>(V)) return V;
146
147  // Traverse through different addressing mechanisms...
148  if (const Instruction *I = dyn_cast<Instruction>(V)) {
149    if (isa<CastInst>(I) || isa<GetElementPtrInst>(I))
150      return getUnderlyingObject(I->getOperand(0));
151  } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
152    if (CE->getOpcode() == Instruction::Cast ||
153        CE->getOpcode() == Instruction::GetElementPtr)
154      return getUnderlyingObject(CE->getOperand(0));
155  } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
156    return GV;
157  }
158  return 0;
159}
160
161static const User *isGEP(const Value *V) {
162  if (isa<GetElementPtrInst>(V) ||
163      (isa<ConstantExpr>(V) &&
164       cast<ConstantExpr>(V)->getOpcode() == Instruction::GetElementPtr))
165    return cast<User>(V);
166  return 0;
167}
168
169static const Value *GetGEPOperands(const Value *V, std::vector<Value*> &GEPOps){
170  assert(GEPOps.empty() && "Expect empty list to populate!");
171  GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1,
172                cast<User>(V)->op_end());
173
174  // Accumulate all of the chained indexes into the operand array
175  V = cast<User>(V)->getOperand(0);
176
177  while (const User *G = isGEP(V)) {
178    if (!isa<Constant>(GEPOps[0]) || isa<GlobalValue>(GEPOps[0]) ||
179        !cast<Constant>(GEPOps[0])->isNullValue())
180      break;  // Don't handle folding arbitrary pointer offsets yet...
181    GEPOps.erase(GEPOps.begin());   // Drop the zero index
182    GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end());
183    V = G->getOperand(0);
184  }
185  return V;
186}
187
188/// pointsToConstantMemory - Chase pointers until we find a (constant
189/// global) or not.
190bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) {
191  if (const Value *V = getUnderlyingObject(P))
192    if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
193      return GV->isConstant();
194  return false;
195}
196
197static bool AddressMightEscape(const Value *V) {
198  for (Value::use_const_iterator UI = V->use_begin(), E = V->use_end();
199       UI != E; ++UI) {
200    const Instruction *I = cast<Instruction>(*UI);
201    switch (I->getOpcode()) {
202    case Instruction::Load: break;
203    case Instruction::Store:
204      if (I->getOperand(0) == V)
205        return true; // Escapes if the pointer is stored.
206      break;
207    case Instruction::GetElementPtr:
208      if (AddressMightEscape(I)) return true;
209      break;
210    case Instruction::Cast:
211      if (!isa<PointerType>(I->getType()))
212        return true;
213      if (AddressMightEscape(I)) return true;
214      break;
215    case Instruction::Ret:
216      // If returned, the address will escape to calling functions, but no
217      // callees could modify it.
218      break;
219    default:
220      return true;
221    }
222  }
223  return false;
224}
225
226// getModRefInfo - Check to see if the specified callsite can clobber the
227// specified memory object.  Since we only look at local properties of this
228// function, we really can't say much about this query.  We do, however, use
229// simple "address taken" analysis on local objects.
230//
231AliasAnalysis::ModRefResult
232BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
233  if (!isa<Constant>(P))
234    if (const AllocationInst *AI =
235                  dyn_cast_or_null<AllocationInst>(getUnderlyingObject(P))) {
236      // Okay, the pointer is to a stack allocated object.  If we can prove that
237      // the pointer never "escapes", then we know the call cannot clobber it,
238      // because it simply can't get its address.
239      if (!AddressMightEscape(AI))
240        return NoModRef;
241
242      // If this is a tail call and P points to a stack location, we know that
243      // the tail call cannot access or modify the local stack.
244      if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
245        if (CI->isTailCall() && isa<AllocaInst>(AI))
246          return NoModRef;
247    }
248
249  // The AliasAnalysis base class has some smarts, lets use them.
250  return AliasAnalysis::getModRefInfo(CS, P, Size);
251}
252
253// alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such
254// as array references.  Note that this function is heavily tail recursive.
255// Hopefully we have a smart C++ compiler.  :)
256//
257AliasAnalysis::AliasResult
258BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size,
259                          const Value *V2, unsigned V2Size) {
260  // Strip off any constant expression casts if they exist
261  if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V1))
262    if (CE->getOpcode() == Instruction::Cast &&
263        isa<PointerType>(CE->getOperand(0)->getType()))
264      V1 = CE->getOperand(0);
265  if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V2))
266    if (CE->getOpcode() == Instruction::Cast &&
267        isa<PointerType>(CE->getOperand(0)->getType()))
268      V2 = CE->getOperand(0);
269
270  // Are we checking for alias of the same value?
271  if (V1 == V2) return MustAlias;
272
273  if ((!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) &&
274      V1->getType() != Type::LongTy && V2->getType() != Type::LongTy)
275    return NoAlias;  // Scalars cannot alias each other
276
277  // Strip off cast instructions...
278  if (const Instruction *I = dyn_cast<CastInst>(V1))
279    if (isa<PointerType>(I->getOperand(0)->getType()))
280      return alias(I->getOperand(0), V1Size, V2, V2Size);
281  if (const Instruction *I = dyn_cast<CastInst>(V2))
282    if (isa<PointerType>(I->getOperand(0)->getType()))
283      return alias(V1, V1Size, I->getOperand(0), V2Size);
284
285  // Figure out what objects these things are pointing to if we can...
286  const Value *O1 = getUnderlyingObject(V1);
287  const Value *O2 = getUnderlyingObject(V2);
288
289  // Pointing at a discernible object?
290  if (O1) {
291    if (O2) {
292      if (isa<Argument>(O1)) {
293        // Incoming argument cannot alias locally allocated object!
294        if (isa<AllocationInst>(O2)) return NoAlias;
295        // Otherwise, nothing is known...
296      } else if (isa<Argument>(O2)) {
297        // Incoming argument cannot alias locally allocated object!
298        if (isa<AllocationInst>(O1)) return NoAlias;
299        // Otherwise, nothing is known...
300      } else if (O1 != O2) {
301        // If they are two different objects, we know that we have no alias...
302        return NoAlias;
303      }
304
305      // If they are the same object, they we can look at the indexes.  If they
306      // index off of the object is the same for both pointers, they must alias.
307      // If they are provably different, they must not alias.  Otherwise, we
308      // can't tell anything.
309    }
310
311
312    if (!isa<Argument>(O1) && isa<ConstantPointerNull>(V2))
313      return NoAlias;                    // Unique values don't alias null
314
315    if (isa<GlobalVariable>(O1) ||
316        (isa<AllocationInst>(O1) &&
317         !cast<AllocationInst>(O1)->isArrayAllocation()))
318      if (cast<PointerType>(O1->getType())->getElementType()->isSized()) {
319        // If the size of the other access is larger than the total size of the
320        // global/alloca/malloc, it cannot be accessing the global (it's
321        // undefined to load or store bytes before or after an object).
322        const Type *ElTy = cast<PointerType>(O1->getType())->getElementType();
323        unsigned GlobalSize = getTargetData().getTypeSize(ElTy);
324        if (GlobalSize < V2Size && V2Size != ~0U)
325          return NoAlias;
326      }
327  }
328
329  if (O2) {
330    if (!isa<Argument>(O2) && isa<ConstantPointerNull>(V1))
331      return NoAlias;                    // Unique values don't alias null
332
333    if (isa<GlobalVariable>(O2) ||
334        (isa<AllocationInst>(O2) &&
335         !cast<AllocationInst>(O2)->isArrayAllocation()))
336      if (cast<PointerType>(O2->getType())->getElementType()->isSized()) {
337        // If the size of the other access is larger than the total size of the
338        // global/alloca/malloc, it cannot be accessing the object (it's
339        // undefined to load or store bytes before or after an object).
340        const Type *ElTy = cast<PointerType>(O2->getType())->getElementType();
341        unsigned GlobalSize = getTargetData().getTypeSize(ElTy);
342        if (GlobalSize < V1Size && V1Size != ~0U)
343          return NoAlias;
344      }
345  }
346
347  // If we have two gep instructions with must-alias'ing base pointers, figure
348  // out if the indexes to the GEP tell us anything about the derived pointer.
349  // Note that we also handle chains of getelementptr instructions as well as
350  // constant expression getelementptrs here.
351  //
352  if (isGEP(V1) && isGEP(V2)) {
353    // Drill down into the first non-gep value, to test for must-aliasing of
354    // the base pointers.
355    const Value *BasePtr1 = V1, *BasePtr2 = V2;
356    do {
357      BasePtr1 = cast<User>(BasePtr1)->getOperand(0);
358    } while (isGEP(BasePtr1) &&
359             cast<User>(BasePtr1)->getOperand(1) ==
360       Constant::getNullValue(cast<User>(BasePtr1)->getOperand(1)->getType()));
361    do {
362      BasePtr2 = cast<User>(BasePtr2)->getOperand(0);
363    } while (isGEP(BasePtr2) &&
364             cast<User>(BasePtr2)->getOperand(1) ==
365       Constant::getNullValue(cast<User>(BasePtr2)->getOperand(1)->getType()));
366
367    // Do the base pointers alias?
368    AliasResult BaseAlias = alias(BasePtr1, V1Size, BasePtr2, V2Size);
369    if (BaseAlias == NoAlias) return NoAlias;
370    if (BaseAlias == MustAlias) {
371      // If the base pointers alias each other exactly, check to see if we can
372      // figure out anything about the resultant pointers, to try to prove
373      // non-aliasing.
374
375      // Collect all of the chained GEP operands together into one simple place
376      std::vector<Value*> GEP1Ops, GEP2Ops;
377      BasePtr1 = GetGEPOperands(V1, GEP1Ops);
378      BasePtr2 = GetGEPOperands(V2, GEP2Ops);
379
380      // If GetGEPOperands were able to fold to the same must-aliased pointer,
381      // do the comparison.
382      if (BasePtr1 == BasePtr2) {
383        AliasResult GAlias =
384          CheckGEPInstructions(BasePtr1->getType(), GEP1Ops, V1Size,
385                               BasePtr2->getType(), GEP2Ops, V2Size);
386        if (GAlias != MayAlias)
387          return GAlias;
388      }
389    }
390  }
391
392  // Check to see if these two pointers are related by a getelementptr
393  // instruction.  If one pointer is a GEP with a non-zero index of the other
394  // pointer, we know they cannot alias.
395  //
396  if (isGEP(V2)) {
397    std::swap(V1, V2);
398    std::swap(V1Size, V2Size);
399  }
400
401  if (V1Size != ~0U && V2Size != ~0U)
402    if (isGEP(V1)) {
403      std::vector<Value*> GEPOperands;
404      const Value *BasePtr = GetGEPOperands(V1, GEPOperands);
405
406      AliasResult R = alias(BasePtr, V1Size, V2, V2Size);
407      if (R == MustAlias) {
408        // If there is at least one non-zero constant index, we know they cannot
409        // alias.
410        bool ConstantFound = false;
411        bool AllZerosFound = true;
412        for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i)
413          if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) {
414            if (!C->isNullValue()) {
415              ConstantFound = true;
416              AllZerosFound = false;
417              break;
418            }
419          } else {
420            AllZerosFound = false;
421          }
422
423        // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases
424        // the ptr, the end result is a must alias also.
425        if (AllZerosFound)
426          return MustAlias;
427
428        if (ConstantFound) {
429          if (V2Size <= 1 && V1Size <= 1)  // Just pointer check?
430            return NoAlias;
431
432          // Otherwise we have to check to see that the distance is more than
433          // the size of the argument... build an index vector that is equal to
434          // the arguments provided, except substitute 0's for any variable
435          // indexes we find...
436          if (cast<PointerType>(
437                BasePtr->getType())->getElementType()->isSized()) {
438            for (unsigned i = 0; i != GEPOperands.size(); ++i)
439              if (!isa<ConstantInt>(GEPOperands[i]))
440                GEPOperands[i] =
441                  Constant::getNullValue(GEPOperands[i]->getType());
442            int64_t Offset =
443              getTargetData().getIndexedOffset(BasePtr->getType(), GEPOperands);
444
445            if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size)
446              return NoAlias;
447          }
448        }
449      }
450    }
451
452  return MayAlias;
453}
454
455static bool ValuesEqual(Value *V1, Value *V2) {
456  if (V1->getType() == V2->getType())
457    return V1 == V2;
458  if (Constant *C1 = dyn_cast<Constant>(V1))
459    if (Constant *C2 = dyn_cast<Constant>(V2)) {
460      // Sign extend the constants to long types.
461      C1 = ConstantExpr::getSignExtend(C1, Type::LongTy);
462      C2 = ConstantExpr::getSignExtend(C2, Type::LongTy);
463      return C1 == C2;
464    }
465  return false;
466}
467
468/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing
469/// base pointers.  This checks to see if the index expressions preclude the
470/// pointers from aliasing...
471AliasAnalysis::AliasResult
472BasicAliasAnalysis::CheckGEPInstructions(
473  const Type* BasePtr1Ty, std::vector<Value*> &GEP1Ops, unsigned G1S,
474  const Type *BasePtr2Ty, std::vector<Value*> &GEP2Ops, unsigned G2S) {
475  // We currently can't handle the case when the base pointers have different
476  // primitive types.  Since this is uncommon anyway, we are happy being
477  // extremely conservative.
478  if (BasePtr1Ty != BasePtr2Ty)
479    return MayAlias;
480
481  const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty);
482
483  // Find the (possibly empty) initial sequence of equal values... which are not
484  // necessarily constants.
485  unsigned NumGEP1Operands = GEP1Ops.size(), NumGEP2Operands = GEP2Ops.size();
486  unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands);
487  unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands);
488  unsigned UnequalOper = 0;
489  while (UnequalOper != MinOperands &&
490         ValuesEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) {
491    // Advance through the type as we go...
492    ++UnequalOper;
493    if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
494      BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]);
495    else {
496      // If all operands equal each other, then the derived pointers must
497      // alias each other...
498      BasePtr1Ty = 0;
499      assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands &&
500             "Ran out of type nesting, but not out of operands?");
501      return MustAlias;
502    }
503  }
504
505  // If we have seen all constant operands, and run out of indexes on one of the
506  // getelementptrs, check to see if the tail of the leftover one is all zeros.
507  // If so, return mustalias.
508  if (UnequalOper == MinOperands) {
509    if (GEP1Ops.size() < GEP2Ops.size()) std::swap(GEP1Ops, GEP2Ops);
510
511    bool AllAreZeros = true;
512    for (unsigned i = UnequalOper; i != MaxOperands; ++i)
513      if (!isa<Constant>(GEP1Ops[i]) ||
514          !cast<Constant>(GEP1Ops[i])->isNullValue()) {
515        AllAreZeros = false;
516        break;
517      }
518    if (AllAreZeros) return MustAlias;
519  }
520
521
522  // So now we know that the indexes derived from the base pointers,
523  // which are known to alias, are different.  We can still determine a
524  // no-alias result if there are differing constant pairs in the index
525  // chain.  For example:
526  //        A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S))
527  //
528  // We have to be careful here about array accesses.  In particular, consider:
529  //        A[1][0] vs A[0][i]
530  // In this case, we don't *know* that the array will be accessed in bounds:
531  // the index could even be negative.  Because of this, we have to
532  // conservatively *give up* and return may alias.  We disregard differing
533  // array subscripts that are followed by a variable index without going
534  // through a struct.
535  //
536  unsigned SizeMax = std::max(G1S, G2S);
537  if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work.
538
539  // Scan for the first operand that is constant and unequal in the
540  // two getelementptrs...
541  unsigned FirstConstantOper = UnequalOper;
542  for (; FirstConstantOper != MinOperands; ++FirstConstantOper) {
543    const Value *G1Oper = GEP1Ops[FirstConstantOper];
544    const Value *G2Oper = GEP2Ops[FirstConstantOper];
545
546    if (G1Oper != G2Oper)   // Found non-equal constant indexes...
547      if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper)))
548        if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){
549          if (G1OC->getType() != G2OC->getType()) {
550            // Sign extend both operands to long.
551            G1OC = ConstantExpr::getSignExtend(G1OC, Type::LongTy);
552            G2OC = ConstantExpr::getSignExtend(G2OC, Type::LongTy);
553            GEP1Ops[FirstConstantOper] = G1OC;
554            GEP2Ops[FirstConstantOper] = G2OC;
555          }
556
557          if (G1OC != G2OC) {
558            // Handle the "be careful" case above: if this is an array/packed
559            // subscript, scan for a subsequent variable array index.
560            if (isa<SequentialType>(BasePtr1Ty))  {
561              const Type *NextTy =
562                cast<SequentialType>(BasePtr1Ty)->getElementType();
563              bool isBadCase = false;
564
565              for (unsigned Idx = FirstConstantOper+1;
566                   Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) {
567                const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx];
568                if (!isa<Constant>(V1) || !isa<Constant>(V2)) {
569                  isBadCase = true;
570                  break;
571                }
572                NextTy = cast<SequentialType>(NextTy)->getElementType();
573              }
574
575              if (isBadCase) G1OC = 0;
576            }
577
578            // Make sure they are comparable (ie, not constant expressions), and
579            // make sure the GEP with the smaller leading constant is GEP1.
580            if (G1OC) {
581              Constant *Compare = ConstantExpr::getSetGT(G1OC, G2OC);
582              if (ConstantBool *CV = dyn_cast<ConstantBool>(Compare)) {
583                if (CV->getValue())   // If they are comparable and G2 > G1
584                  std::swap(GEP1Ops, GEP2Ops);  // Make GEP1 < GEP2
585                break;
586              }
587            }
588          }
589        }
590    BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper);
591  }
592
593  // No shared constant operands, and we ran out of common operands.  At this
594  // point, the GEP instructions have run through all of their operands, and we
595  // haven't found evidence that there are any deltas between the GEP's.
596  // However, one GEP may have more operands than the other.  If this is the
597  // case, there may still be hope.  Check this now.
598  if (FirstConstantOper == MinOperands) {
599    // Make GEP1Ops be the longer one if there is a longer one.
600    if (GEP1Ops.size() < GEP2Ops.size())
601      std::swap(GEP1Ops, GEP2Ops);
602
603    // Is there anything to check?
604    if (GEP1Ops.size() > MinOperands) {
605      for (unsigned i = FirstConstantOper; i != MaxOperands; ++i)
606        if (isa<ConstantInt>(GEP1Ops[i]) &&
607            !cast<Constant>(GEP1Ops[i])->isNullValue()) {
608          // Yup, there's a constant in the tail.  Set all variables to
609          // constants in the GEP instruction to make it suiteable for
610          // TargetData::getIndexedOffset.
611          for (i = 0; i != MaxOperands; ++i)
612            if (!isa<ConstantInt>(GEP1Ops[i]))
613              GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType());
614          // Okay, now get the offset.  This is the relative offset for the full
615          // instruction.
616          const TargetData &TD = getTargetData();
617          int64_t Offset1 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops);
618
619          // Now crop off any constants from the end...
620          GEP1Ops.resize(MinOperands);
621          int64_t Offset2 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops);
622
623          // If the tail provided a bit enough offset, return noalias!
624          if ((uint64_t)(Offset2-Offset1) >= SizeMax)
625            return NoAlias;
626        }
627    }
628
629    // Couldn't find anything useful.
630    return MayAlias;
631  }
632
633  // If there are non-equal constants arguments, then we can figure
634  // out a minimum known delta between the two index expressions... at
635  // this point we know that the first constant index of GEP1 is less
636  // than the first constant index of GEP2.
637
638  // Advance BasePtr[12]Ty over this first differing constant operand.
639  BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)->
640      getTypeAtIndex(GEP2Ops[FirstConstantOper]);
641  BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->
642      getTypeAtIndex(GEP1Ops[FirstConstantOper]);
643
644  // We are going to be using TargetData::getIndexedOffset to determine the
645  // offset that each of the GEP's is reaching.  To do this, we have to convert
646  // all variable references to constant references.  To do this, we convert the
647  // initial sequence of array subscripts into constant zeros to start with.
648  const Type *ZeroIdxTy = GEPPointerTy;
649  for (unsigned i = 0; i != FirstConstantOper; ++i) {
650    if (!isa<StructType>(ZeroIdxTy))
651      GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Type::UIntTy);
652
653    if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy))
654      ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]);
655  }
656
657  // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok
658
659  // Loop over the rest of the operands...
660  for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) {
661    const Value *Op1 = i < GEP1Ops.size() ? GEP1Ops[i] : 0;
662    const Value *Op2 = i < GEP2Ops.size() ? GEP2Ops[i] : 0;
663    // If they are equal, use a zero index...
664    if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) {
665      if (!isa<ConstantInt>(Op1))
666        GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType());
667      // Otherwise, just keep the constants we have.
668    } else {
669      if (Op1) {
670        if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
671          // If this is an array index, make sure the array element is in range.
672          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) {
673            if (Op1C->getZExtValue() >= AT->getNumElements())
674              return MayAlias;  // Be conservative with out-of-range accesses
675          } else if (const PackedType *PT = dyn_cast<PackedType>(BasePtr1Ty)) {
676            if (Op1C->getZExtValue() >= PT->getNumElements())
677              return MayAlias;  // Be conservative with out-of-range accesses
678          }
679
680        } else {
681          // GEP1 is known to produce a value less than GEP2.  To be
682          // conservatively correct, we must assume the largest possible
683          // constant is used in this position.  This cannot be the initial
684          // index to the GEP instructions (because we know we have at least one
685          // element before this one with the different constant arguments), so
686          // we know that the current index must be into either a struct or
687          // array.  Because we know it's not constant, this cannot be a
688          // structure index.  Because of this, we can calculate the maximum
689          // value possible.
690          //
691          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
692            GEP1Ops[i] = ConstantInt::get(Type::LongTy, AT->getNumElements()-1);
693          else if (const PackedType *PT = dyn_cast<PackedType>(BasePtr1Ty))
694            GEP1Ops[i] = ConstantInt::get(Type::LongTy, PT->getNumElements()-1);
695
696        }
697      }
698
699      if (Op2) {
700        if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) {
701          // If this is an array index, make sure the array element is in range.
702          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) {
703            if (Op2C->getZExtValue() >= AT->getNumElements())
704              return MayAlias;  // Be conservative with out-of-range accesses
705          } else if (const PackedType *PT = dyn_cast<PackedType>(BasePtr1Ty)) {
706            if (Op2C->getZExtValue() >= PT->getNumElements())
707              return MayAlias;  // Be conservative with out-of-range accesses
708          }
709        } else {  // Conservatively assume the minimum value for this index
710          GEP2Ops[i] = Constant::getNullValue(Op2->getType());
711        }
712      }
713    }
714
715    if (BasePtr1Ty && Op1) {
716      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
717        BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]);
718      else
719        BasePtr1Ty = 0;
720    }
721
722    if (BasePtr2Ty && Op2) {
723      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty))
724        BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]);
725      else
726        BasePtr2Ty = 0;
727    }
728  }
729
730  if (GEPPointerTy->getElementType()->isSized()) {
731    int64_t Offset1 = getTargetData().getIndexedOffset(GEPPointerTy, GEP1Ops);
732    int64_t Offset2 = getTargetData().getIndexedOffset(GEPPointerTy, GEP2Ops);
733    assert(Offset1<Offset2 && "There is at least one different constant here!");
734
735    if ((uint64_t)(Offset2-Offset1) >= SizeMax) {
736      //std::cerr << "Determined that these two GEP's don't alias ["
737      //          << SizeMax << " bytes]: \n" << *GEP1 << *GEP2;
738      return NoAlias;
739    }
740  }
741  return MayAlias;
742}
743
744namespace {
745  struct StringCompare {
746    bool operator()(const char *LHS, const char *RHS) {
747      return strcmp(LHS, RHS) < 0;
748    }
749  };
750}
751
752// Note that this list cannot contain libm functions (such as acos and sqrt)
753// that set errno on a domain or other error.
754static const char *DoesntAccessMemoryFns[] = {
755  "abs", "labs", "llabs", "imaxabs", "fabs", "fabsf", "fabsl",
756  "trunc", "truncf", "truncl", "ldexp",
757
758  "atan", "atanf", "atanl",   "atan2", "atan2f", "atan2l",
759  "cbrt",
760  "cos", "cosf", "cosl",
761  "exp", "expf", "expl",
762  "hypot",
763  "sin", "sinf", "sinl",
764  "tan", "tanf", "tanl",      "tanh", "tanhf", "tanhl",
765
766  "floor", "floorf", "floorl", "ceil", "ceilf", "ceill",
767
768  // ctype.h
769  "isalnum", "isalpha", "iscntrl", "isdigit", "isgraph", "islower", "isprint"
770  "ispunct", "isspace", "isupper", "isxdigit", "tolower", "toupper",
771
772  // wctype.h"
773  "iswalnum", "iswalpha", "iswcntrl", "iswdigit", "iswgraph", "iswlower",
774  "iswprint", "iswpunct", "iswspace", "iswupper", "iswxdigit",
775
776  "iswctype", "towctrans", "towlower", "towupper",
777
778  "btowc", "wctob",
779
780  "isinf", "isnan", "finite",
781
782  // C99 math functions
783  "copysign", "copysignf", "copysignd",
784  "nexttoward", "nexttowardf", "nexttowardd",
785  "nextafter", "nextafterf", "nextafterd",
786
787  // ISO C99:
788  "__signbit", "__signbitf", "__signbitl",
789};
790
791
792static const char *OnlyReadsMemoryFns[] = {
793  "atoi", "atol", "atof", "atoll", "atoq", "a64l",
794  "bcmp", "memcmp", "memchr", "memrchr", "wmemcmp", "wmemchr",
795
796  // Strings
797  "strcmp", "strcasecmp", "strcoll", "strncmp", "strncasecmp",
798  "strchr", "strcspn", "strlen", "strpbrk", "strrchr", "strspn", "strstr",
799  "index", "rindex",
800
801  // Wide char strings
802  "wcschr", "wcscmp", "wcscoll", "wcscspn", "wcslen", "wcsncmp", "wcspbrk",
803  "wcsrchr", "wcsspn", "wcsstr",
804
805  // glibc
806  "alphasort", "alphasort64", "versionsort", "versionsort64",
807
808  // C99
809  "nan", "nanf", "nand",
810
811  // File I/O
812  "feof", "ferror", "fileno",
813  "feof_unlocked", "ferror_unlocked", "fileno_unlocked"
814};
815
816static ManagedStatic<std::vector<const char*> > NoMemoryTable;
817static ManagedStatic<std::vector<const char*> > OnlyReadsMemoryTable;
818
819
820AliasAnalysis::ModRefBehavior
821BasicAliasAnalysis::getModRefBehavior(Function *F, CallSite CS,
822                                      std::vector<PointerAccessInfo> *Info) {
823  if (!F->isExternal()) return UnknownModRefBehavior;
824
825  static bool Initialized = false;
826  if (!Initialized) {
827    NoMemoryTable->insert(NoMemoryTable->end(),
828                          DoesntAccessMemoryFns,
829                          DoesntAccessMemoryFns+
830                sizeof(DoesntAccessMemoryFns)/sizeof(DoesntAccessMemoryFns[0]));
831
832    OnlyReadsMemoryTable->insert(OnlyReadsMemoryTable->end(),
833                                OnlyReadsMemoryFns,
834                                OnlyReadsMemoryFns+
835                      sizeof(OnlyReadsMemoryFns)/sizeof(OnlyReadsMemoryFns[0]));
836#define GET_MODREF_BEHAVIOR
837#include "llvm/Intrinsics.gen"
838#undef GET_MODREF_BEHAVIOR
839
840    // Sort the table the first time through.
841    std::sort(NoMemoryTable->begin(), NoMemoryTable->end(), StringCompare());
842    std::sort(OnlyReadsMemoryTable->begin(), OnlyReadsMemoryTable->end(),
843              StringCompare());
844    Initialized = true;
845  }
846
847  std::vector<const char*>::iterator Ptr =
848    std::lower_bound(NoMemoryTable->begin(), NoMemoryTable->end(),
849                     F->getName().c_str(), StringCompare());
850  if (Ptr != NoMemoryTable->end() && *Ptr == F->getName())
851    return DoesNotAccessMemory;
852
853  Ptr = std::lower_bound(OnlyReadsMemoryTable->begin(),
854                         OnlyReadsMemoryTable->end(),
855                         F->getName().c_str(), StringCompare());
856  if (Ptr != OnlyReadsMemoryTable->end() && *Ptr == F->getName())
857    return OnlyReadsMemory;
858
859  return UnknownModRefBehavior;
860}
861
862// Make sure that anything that uses AliasAnalysis pulls in this file...
863DEFINING_FILE_FOR(BasicAliasAnalysis)
864