BasicAliasAnalysis.cpp revision 851bfff50f7251e4d00b5fffbb5e140544cc4a63
1//===- BasicAliasAnalysis.cpp - Local Alias Analysis Impl -----------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines the default implementation of the Alias Analysis interface
11// that simply implements a few identities (two different globals cannot alias,
12// etc), but otherwise does no analysis.
13//
14//===----------------------------------------------------------------------===//
15
16#include "llvm/Analysis/AliasAnalysis.h"
17#include "llvm/Analysis/Passes.h"
18#include "llvm/Constants.h"
19#include "llvm/DerivedTypes.h"
20#include "llvm/Function.h"
21#include "llvm/ParameterAttributes.h"
22#include "llvm/GlobalVariable.h"
23#include "llvm/Instructions.h"
24#include "llvm/IntrinsicInst.h"
25#include "llvm/Pass.h"
26#include "llvm/Target/TargetData.h"
27#include "llvm/ADT/SmallVector.h"
28#include "llvm/ADT/STLExtras.h"
29#include "llvm/Support/Compiler.h"
30#include "llvm/Support/GetElementPtrTypeIterator.h"
31#include "llvm/Support/ManagedStatic.h"
32#include <algorithm>
33using namespace llvm;
34
35namespace {
36  /// NoAA - This class implements the -no-aa pass, which always returns "I
37  /// don't know" for alias queries.  NoAA is unlike other alias analysis
38  /// implementations, in that it does not chain to a previous analysis.  As
39  /// such it doesn't follow many of the rules that other alias analyses must.
40  ///
41  struct VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis {
42    static char ID; // Class identification, replacement for typeinfo
43    NoAA() : ImmutablePass((intptr_t)&ID) {}
44    explicit NoAA(intptr_t PID) : ImmutablePass(PID) { }
45
46    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
47      AU.addRequired<TargetData>();
48    }
49
50    virtual void initializePass() {
51      TD = &getAnalysis<TargetData>();
52    }
53
54    virtual AliasResult alias(const Value *V1, unsigned V1Size,
55                              const Value *V2, unsigned V2Size) {
56      return MayAlias;
57    }
58
59    virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
60                                         std::vector<PointerAccessInfo> *Info) {
61      return UnknownModRefBehavior;
62    }
63
64    virtual void getArgumentAccesses(Function *F, CallSite CS,
65                                     std::vector<PointerAccessInfo> &Info) {
66      assert(0 && "This method may not be called on this function!");
67    }
68
69    virtual void getMustAliases(Value *P, std::vector<Value*> &RetVals) { }
70    virtual bool pointsToConstantMemory(const Value *P) { return false; }
71    virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) {
72      return ModRef;
73    }
74    virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
75      return ModRef;
76    }
77    virtual bool hasNoModRefInfoForCalls() const { return true; }
78
79    virtual void deleteValue(Value *V) {}
80    virtual void copyValue(Value *From, Value *To) {}
81  };
82
83  // Register this pass...
84  char NoAA::ID = 0;
85  RegisterPass<NoAA>
86  U("no-aa", "No Alias Analysis (always returns 'may' alias)");
87
88  // Declare that we implement the AliasAnalysis interface
89  RegisterAnalysisGroup<AliasAnalysis> V(U);
90}  // End of anonymous namespace
91
92ImmutablePass *llvm::createNoAAPass() { return new NoAA(); }
93
94namespace {
95  /// BasicAliasAnalysis - This is the default alias analysis implementation.
96  /// Because it doesn't chain to a previous alias analysis (like -no-aa), it
97  /// derives from the NoAA class.
98  struct VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA {
99    static char ID; // Class identification, replacement for typeinfo
100    BasicAliasAnalysis() : NoAA((intptr_t)&ID) { }
101    AliasResult alias(const Value *V1, unsigned V1Size,
102                      const Value *V2, unsigned V2Size);
103
104    ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
105    ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
106      return NoAA::getModRefInfo(CS1,CS2);
107    }
108
109    /// hasNoModRefInfoForCalls - We can provide mod/ref information against
110    /// non-escaping allocations.
111    virtual bool hasNoModRefInfoForCalls() const { return false; }
112
113    /// pointsToConstantMemory - Chase pointers until we find a (constant
114    /// global) or not.
115    bool pointsToConstantMemory(const Value *P);
116
117  private:
118    // CheckGEPInstructions - Check two GEP instructions with known
119    // must-aliasing base pointers.  This checks to see if the index expressions
120    // preclude the pointers from aliasing...
121    AliasResult
122    CheckGEPInstructions(const Type* BasePtr1Ty,
123                         Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size,
124                         const Type *BasePtr2Ty,
125                         Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size);
126  };
127
128  // Register this pass...
129  char BasicAliasAnalysis::ID = 0;
130  RegisterPass<BasicAliasAnalysis>
131  X("basicaa", "Basic Alias Analysis (default AA impl)");
132
133  // Declare that we implement the AliasAnalysis interface
134  RegisterAnalysisGroup<AliasAnalysis, true> Y(X);
135}  // End of anonymous namespace
136
137ImmutablePass *llvm::createBasicAliasAnalysisPass() {
138  return new BasicAliasAnalysis();
139}
140
141/// getUnderlyingObject - This traverses the use chain to figure out what object
142/// the specified value points to.  If the value points to, or is derived from,
143/// a unique object or an argument, return it.  This returns:
144///    Arguments, GlobalVariables, Functions, Allocas, Mallocs.
145static const Value *getUnderlyingObject(const Value *V) {
146  if (!isa<PointerType>(V->getType())) return 0;
147
148  // If we are at some type of object, return it. GlobalValues and Allocations
149  // have unique addresses.
150  if (isa<GlobalValue>(V) || isa<AllocationInst>(V) || isa<Argument>(V))
151    return V;
152
153  // Traverse through different addressing mechanisms...
154  if (const Instruction *I = dyn_cast<Instruction>(V)) {
155    if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I))
156      return getUnderlyingObject(I->getOperand(0));
157  } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
158    if (CE->getOpcode() == Instruction::BitCast ||
159        CE->getOpcode() == Instruction::GetElementPtr)
160      return getUnderlyingObject(CE->getOperand(0));
161  }
162  return 0;
163}
164
165static const User *isGEP(const Value *V) {
166  if (isa<GetElementPtrInst>(V) ||
167      (isa<ConstantExpr>(V) &&
168       cast<ConstantExpr>(V)->getOpcode() == Instruction::GetElementPtr))
169    return cast<User>(V);
170  return 0;
171}
172
173static const Value *GetGEPOperands(const Value *V,
174                                   SmallVector<Value*, 16> &GEPOps){
175  assert(GEPOps.empty() && "Expect empty list to populate!");
176  GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1,
177                cast<User>(V)->op_end());
178
179  // Accumulate all of the chained indexes into the operand array
180  V = cast<User>(V)->getOperand(0);
181
182  while (const User *G = isGEP(V)) {
183    if (!isa<Constant>(GEPOps[0]) || isa<GlobalValue>(GEPOps[0]) ||
184        !cast<Constant>(GEPOps[0])->isNullValue())
185      break;  // Don't handle folding arbitrary pointer offsets yet...
186    GEPOps.erase(GEPOps.begin());   // Drop the zero index
187    GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end());
188    V = G->getOperand(0);
189  }
190  return V;
191}
192
193/// pointsToConstantMemory - Chase pointers until we find a (constant
194/// global) or not.
195bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) {
196  if (const Value *V = getUnderlyingObject(P))
197    if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
198      return GV->isConstant();
199  return false;
200}
201
202// Determine if an AllocationInst instruction escapes from the function it is
203// contained in. If it does not escape, there is no way for another function to
204// mod/ref it.  We do this by looking at its uses and determining if the uses
205// can escape (recursively).
206static bool AddressMightEscape(const Value *V) {
207  for (Value::use_const_iterator UI = V->use_begin(), E = V->use_end();
208       UI != E; ++UI) {
209    const Instruction *I = cast<Instruction>(*UI);
210    switch (I->getOpcode()) {
211    case Instruction::Load:
212      break; //next use.
213    case Instruction::Store:
214      if (I->getOperand(0) == V)
215        return true; // Escapes if the pointer is stored.
216      break; // next use.
217    case Instruction::GetElementPtr:
218      if (AddressMightEscape(I))
219        return true;
220      break; // next use.
221    case Instruction::BitCast:
222      if (AddressMightEscape(I))
223        return true;
224      break; // next use
225    case Instruction::Ret:
226      // If returned, the address will escape to calling functions, but no
227      // callees could modify it.
228      break; // next use
229    case Instruction::Call:
230      // If the call is to a few known safe intrinsics, we know that it does
231      // not escape
232      if (!isa<MemIntrinsic>(I))
233        return true;
234      break;  // next use
235    default:
236      return true;
237    }
238  }
239  return false;
240}
241
242// getModRefInfo - Check to see if the specified callsite can clobber the
243// specified memory object.  Since we only look at local properties of this
244// function, we really can't say much about this query.  We do, however, use
245// simple "address taken" analysis on local objects.
246//
247AliasAnalysis::ModRefResult
248BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
249  if (!isa<Constant>(P)) {
250    const Value *Object = getUnderlyingObject(P);
251    // Allocations and byval arguments are "new" objects.
252    if (Object &&
253        (isa<AllocationInst>(Object) || isa<Argument>(Object))) {
254      // Okay, the pointer is to a stack allocated (or effectively so, for
255      // for noalias parameters) object.  If the address of this object doesn't
256      // escape from this function body to a callee, then we know that no
257      // callees can mod/ref it unless they are actually passed it.
258      if (isa<AllocationInst>(Object) ||
259          cast<Argument>(Object)->hasByValAttr() ||
260          cast<Argument>(Object)->hasNoAliasAttr())
261        if (!AddressMightEscape(Object)) {
262          bool passedAsArg = false;
263          for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
264              CI != CE; ++CI)
265            if (isa<PointerType>((*CI)->getType()) &&
266                ( getUnderlyingObject(*CI) == P ||
267                  alias(cast<Value>(CI), ~0ULL, P, ~0ULL) != NoAlias) )
268              passedAsArg = true;
269
270          if (!passedAsArg)
271            return NoModRef;
272        }
273
274      // If this is a tail call and P points to a stack location, we know that
275      // the tail call cannot access or modify the local stack.
276      if (isa<AllocaInst>(Object) ||
277          (isa<Argument>(Object) && cast<Argument>(Object)->hasByValAttr()))
278        if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
279          if (CI->isTailCall())
280            return NoModRef;
281    }
282  }
283
284  // The AliasAnalysis base class has some smarts, lets use them.
285  return AliasAnalysis::getModRefInfo(CS, P, Size);
286}
287
288// alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such
289// as array references.  Note that this function is heavily tail recursive.
290// Hopefully we have a smart C++ compiler.  :)
291//
292AliasAnalysis::AliasResult
293BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size,
294                          const Value *V2, unsigned V2Size) {
295  // Strip off any constant expression casts if they exist
296  if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V1))
297    if (CE->isCast() && isa<PointerType>(CE->getOperand(0)->getType()))
298      V1 = CE->getOperand(0);
299  if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V2))
300    if (CE->isCast() && isa<PointerType>(CE->getOperand(0)->getType()))
301      V2 = CE->getOperand(0);
302
303  // Are we checking for alias of the same value?
304  if (V1 == V2) return MustAlias;
305
306  if ((!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) &&
307      V1->getType() != Type::Int64Ty && V2->getType() != Type::Int64Ty)
308    return NoAlias;  // Scalars cannot alias each other
309
310  // Strip off cast instructions...
311  if (const BitCastInst *I = dyn_cast<BitCastInst>(V1))
312    return alias(I->getOperand(0), V1Size, V2, V2Size);
313  if (const BitCastInst *I = dyn_cast<BitCastInst>(V2))
314    return alias(V1, V1Size, I->getOperand(0), V2Size);
315
316  // Figure out what objects these things are pointing to if we can...
317  const Value *O1 = getUnderlyingObject(V1);
318  const Value *O2 = getUnderlyingObject(V2);
319
320  // Pointing at a discernible object?
321  if (O1) {
322    if (O2) {
323      if (const Argument *O1Arg = dyn_cast<Argument>(O1)) {
324        // Incoming argument cannot alias locally allocated object!
325        if (isa<AllocationInst>(O2)) return NoAlias;
326
327        // If they are two different objects, and one is a noalias argument
328        // then they do not alias.
329        if (O1 != O2 && O1Arg->hasNoAliasAttr())
330          return NoAlias;
331
332        // Byval arguments can't alias globals or other arguments.
333        if (O1 != O2 && O1Arg->hasByValAttr()) return NoAlias;
334
335        // Otherwise, nothing is known...
336      }
337
338      if (const Argument *O2Arg = dyn_cast<Argument>(O2)) {
339        // Incoming argument cannot alias locally allocated object!
340        if (isa<AllocationInst>(O1)) return NoAlias;
341
342        // If they are two different objects, and one is a noalias argument
343        // then they do not alias.
344        if (O1 != O2 && O2Arg->hasNoAliasAttr())
345          return NoAlias;
346
347        // Byval arguments can't alias globals or other arguments.
348        if (O1 != O2 && O2Arg->hasByValAttr()) return NoAlias;
349
350        // Otherwise, nothing is known...
351
352      } else if (O1 != O2 && !isa<Argument>(O1)) {
353        // If they are two different objects, and neither is an argument,
354        // we know that we have no alias.
355        return NoAlias;
356      }
357
358      // If they are the same object, they we can look at the indexes.  If they
359      // index off of the object is the same for both pointers, they must alias.
360      // If they are provably different, they must not alias.  Otherwise, we
361      // can't tell anything.
362    }
363
364    // Unique values don't alias null, except non-byval arguments.
365    if (isa<ConstantPointerNull>(V2)) {
366      if (const Argument *O1Arg = dyn_cast<Argument>(O1)) {
367        if (O1Arg->hasByValAttr())
368          return NoAlias;
369      } else {
370        return NoAlias;
371      }
372    }
373
374    if (isa<GlobalVariable>(O1) ||
375        (isa<AllocationInst>(O1) &&
376         !cast<AllocationInst>(O1)->isArrayAllocation()))
377      if (cast<PointerType>(O1->getType())->getElementType()->isSized()) {
378        // If the size of the other access is larger than the total size of the
379        // global/alloca/malloc, it cannot be accessing the global (it's
380        // undefined to load or store bytes before or after an object).
381        const Type *ElTy = cast<PointerType>(O1->getType())->getElementType();
382        unsigned GlobalSize = getTargetData().getABITypeSize(ElTy);
383        if (GlobalSize < V2Size && V2Size != ~0U)
384          return NoAlias;
385      }
386  }
387
388  if (O2) {
389    if (!isa<Argument>(O2) && isa<ConstantPointerNull>(V1))
390      return NoAlias;                    // Unique values don't alias null
391
392    if (isa<GlobalVariable>(O2) ||
393        (isa<AllocationInst>(O2) &&
394         !cast<AllocationInst>(O2)->isArrayAllocation()))
395      if (cast<PointerType>(O2->getType())->getElementType()->isSized()) {
396        // If the size of the other access is larger than the total size of the
397        // global/alloca/malloc, it cannot be accessing the object (it's
398        // undefined to load or store bytes before or after an object).
399        const Type *ElTy = cast<PointerType>(O2->getType())->getElementType();
400        unsigned GlobalSize = getTargetData().getABITypeSize(ElTy);
401        if (GlobalSize < V1Size && V1Size != ~0U)
402          return NoAlias;
403      }
404  }
405
406  // If we have two gep instructions with must-alias'ing base pointers, figure
407  // out if the indexes to the GEP tell us anything about the derived pointer.
408  // Note that we also handle chains of getelementptr instructions as well as
409  // constant expression getelementptrs here.
410  //
411  if (isGEP(V1) && isGEP(V2)) {
412    // Drill down into the first non-gep value, to test for must-aliasing of
413    // the base pointers.
414    const User *G = cast<User>(V1);
415    while (isGEP(G->getOperand(0)) &&
416           G->getOperand(1) ==
417           Constant::getNullValue(G->getOperand(1)->getType()))
418      G = cast<User>(G->getOperand(0));
419    const Value *BasePtr1 = G->getOperand(0);
420
421    G = cast<User>(V2);
422    while (isGEP(G->getOperand(0)) &&
423           G->getOperand(1) ==
424           Constant::getNullValue(G->getOperand(1)->getType()))
425      G = cast<User>(G->getOperand(0));
426    const Value *BasePtr2 = G->getOperand(0);
427
428    // Do the base pointers alias?
429    AliasResult BaseAlias = alias(BasePtr1, ~0U, BasePtr2, ~0U);
430    if (BaseAlias == NoAlias) return NoAlias;
431    if (BaseAlias == MustAlias) {
432      // If the base pointers alias each other exactly, check to see if we can
433      // figure out anything about the resultant pointers, to try to prove
434      // non-aliasing.
435
436      // Collect all of the chained GEP operands together into one simple place
437      SmallVector<Value*, 16> GEP1Ops, GEP2Ops;
438      BasePtr1 = GetGEPOperands(V1, GEP1Ops);
439      BasePtr2 = GetGEPOperands(V2, GEP2Ops);
440
441      // If GetGEPOperands were able to fold to the same must-aliased pointer,
442      // do the comparison.
443      if (BasePtr1 == BasePtr2) {
444        AliasResult GAlias =
445          CheckGEPInstructions(BasePtr1->getType(),
446                               &GEP1Ops[0], GEP1Ops.size(), V1Size,
447                               BasePtr2->getType(),
448                               &GEP2Ops[0], GEP2Ops.size(), V2Size);
449        if (GAlias != MayAlias)
450          return GAlias;
451      }
452    }
453  }
454
455  // Check to see if these two pointers are related by a getelementptr
456  // instruction.  If one pointer is a GEP with a non-zero index of the other
457  // pointer, we know they cannot alias.
458  //
459  if (isGEP(V2)) {
460    std::swap(V1, V2);
461    std::swap(V1Size, V2Size);
462  }
463
464  if (V1Size != ~0U && V2Size != ~0U)
465    if (isGEP(V1)) {
466      SmallVector<Value*, 16> GEPOperands;
467      const Value *BasePtr = GetGEPOperands(V1, GEPOperands);
468
469      AliasResult R = alias(BasePtr, V1Size, V2, V2Size);
470      if (R == MustAlias) {
471        // If there is at least one non-zero constant index, we know they cannot
472        // alias.
473        bool ConstantFound = false;
474        bool AllZerosFound = true;
475        for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i)
476          if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) {
477            if (!C->isNullValue()) {
478              ConstantFound = true;
479              AllZerosFound = false;
480              break;
481            }
482          } else {
483            AllZerosFound = false;
484          }
485
486        // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases
487        // the ptr, the end result is a must alias also.
488        if (AllZerosFound)
489          return MustAlias;
490
491        if (ConstantFound) {
492          if (V2Size <= 1 && V1Size <= 1)  // Just pointer check?
493            return NoAlias;
494
495          // Otherwise we have to check to see that the distance is more than
496          // the size of the argument... build an index vector that is equal to
497          // the arguments provided, except substitute 0's for any variable
498          // indexes we find...
499          if (cast<PointerType>(
500                BasePtr->getType())->getElementType()->isSized()) {
501            for (unsigned i = 0; i != GEPOperands.size(); ++i)
502              if (!isa<ConstantInt>(GEPOperands[i]))
503                GEPOperands[i] =
504                  Constant::getNullValue(GEPOperands[i]->getType());
505            int64_t Offset =
506              getTargetData().getIndexedOffset(BasePtr->getType(),
507                                               &GEPOperands[0],
508                                               GEPOperands.size());
509
510            if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size)
511              return NoAlias;
512          }
513        }
514      }
515    }
516
517  return MayAlias;
518}
519
520// This function is used to determin if the indices of two GEP instructions are
521// equal. V1 and V2 are the indices.
522static bool IndexOperandsEqual(Value *V1, Value *V2) {
523  if (V1->getType() == V2->getType())
524    return V1 == V2;
525  if (Constant *C1 = dyn_cast<Constant>(V1))
526    if (Constant *C2 = dyn_cast<Constant>(V2)) {
527      // Sign extend the constants to long types, if necessary
528      if (C1->getType() != Type::Int64Ty)
529        C1 = ConstantExpr::getSExt(C1, Type::Int64Ty);
530      if (C2->getType() != Type::Int64Ty)
531        C2 = ConstantExpr::getSExt(C2, Type::Int64Ty);
532      return C1 == C2;
533    }
534  return false;
535}
536
537/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing
538/// base pointers.  This checks to see if the index expressions preclude the
539/// pointers from aliasing...
540AliasAnalysis::AliasResult
541BasicAliasAnalysis::CheckGEPInstructions(
542  const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S,
543  const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) {
544  // We currently can't handle the case when the base pointers have different
545  // primitive types.  Since this is uncommon anyway, we are happy being
546  // extremely conservative.
547  if (BasePtr1Ty != BasePtr2Ty)
548    return MayAlias;
549
550  const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty);
551
552  // Find the (possibly empty) initial sequence of equal values... which are not
553  // necessarily constants.
554  unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops;
555  unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands);
556  unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands);
557  unsigned UnequalOper = 0;
558  while (UnequalOper != MinOperands &&
559         IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) {
560    // Advance through the type as we go...
561    ++UnequalOper;
562    if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
563      BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]);
564    else {
565      // If all operands equal each other, then the derived pointers must
566      // alias each other...
567      BasePtr1Ty = 0;
568      assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands &&
569             "Ran out of type nesting, but not out of operands?");
570      return MustAlias;
571    }
572  }
573
574  // If we have seen all constant operands, and run out of indexes on one of the
575  // getelementptrs, check to see if the tail of the leftover one is all zeros.
576  // If so, return mustalias.
577  if (UnequalOper == MinOperands) {
578    if (NumGEP1Ops < NumGEP2Ops) {
579      std::swap(GEP1Ops, GEP2Ops);
580      std::swap(NumGEP1Ops, NumGEP2Ops);
581    }
582
583    bool AllAreZeros = true;
584    for (unsigned i = UnequalOper; i != MaxOperands; ++i)
585      if (!isa<Constant>(GEP1Ops[i]) ||
586          !cast<Constant>(GEP1Ops[i])->isNullValue()) {
587        AllAreZeros = false;
588        break;
589      }
590    if (AllAreZeros) return MustAlias;
591  }
592
593
594  // So now we know that the indexes derived from the base pointers,
595  // which are known to alias, are different.  We can still determine a
596  // no-alias result if there are differing constant pairs in the index
597  // chain.  For example:
598  //        A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S))
599  //
600  // We have to be careful here about array accesses.  In particular, consider:
601  //        A[1][0] vs A[0][i]
602  // In this case, we don't *know* that the array will be accessed in bounds:
603  // the index could even be negative.  Because of this, we have to
604  // conservatively *give up* and return may alias.  We disregard differing
605  // array subscripts that are followed by a variable index without going
606  // through a struct.
607  //
608  unsigned SizeMax = std::max(G1S, G2S);
609  if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work.
610
611  // Scan for the first operand that is constant and unequal in the
612  // two getelementptrs...
613  unsigned FirstConstantOper = UnequalOper;
614  for (; FirstConstantOper != MinOperands; ++FirstConstantOper) {
615    const Value *G1Oper = GEP1Ops[FirstConstantOper];
616    const Value *G2Oper = GEP2Ops[FirstConstantOper];
617
618    if (G1Oper != G2Oper)   // Found non-equal constant indexes...
619      if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper)))
620        if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){
621          if (G1OC->getType() != G2OC->getType()) {
622            // Sign extend both operands to long.
623            if (G1OC->getType() != Type::Int64Ty)
624              G1OC = ConstantExpr::getSExt(G1OC, Type::Int64Ty);
625            if (G2OC->getType() != Type::Int64Ty)
626              G2OC = ConstantExpr::getSExt(G2OC, Type::Int64Ty);
627            GEP1Ops[FirstConstantOper] = G1OC;
628            GEP2Ops[FirstConstantOper] = G2OC;
629          }
630
631          if (G1OC != G2OC) {
632            // Handle the "be careful" case above: if this is an array/vector
633            // subscript, scan for a subsequent variable array index.
634            if (isa<SequentialType>(BasePtr1Ty))  {
635              const Type *NextTy =
636                cast<SequentialType>(BasePtr1Ty)->getElementType();
637              bool isBadCase = false;
638
639              for (unsigned Idx = FirstConstantOper+1;
640                   Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) {
641                const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx];
642                if (!isa<Constant>(V1) || !isa<Constant>(V2)) {
643                  isBadCase = true;
644                  break;
645                }
646                NextTy = cast<SequentialType>(NextTy)->getElementType();
647              }
648
649              if (isBadCase) G1OC = 0;
650            }
651
652            // Make sure they are comparable (ie, not constant expressions), and
653            // make sure the GEP with the smaller leading constant is GEP1.
654            if (G1OC) {
655              Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT,
656                                                        G1OC, G2OC);
657              if (ConstantInt *CV = dyn_cast<ConstantInt>(Compare)) {
658                if (CV->getZExtValue()) {  // If they are comparable and G2 > G1
659                  std::swap(GEP1Ops, GEP2Ops);  // Make GEP1 < GEP2
660                  std::swap(NumGEP1Ops, NumGEP2Ops);
661                }
662                break;
663              }
664            }
665          }
666        }
667    BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper);
668  }
669
670  // No shared constant operands, and we ran out of common operands.  At this
671  // point, the GEP instructions have run through all of their operands, and we
672  // haven't found evidence that there are any deltas between the GEP's.
673  // However, one GEP may have more operands than the other.  If this is the
674  // case, there may still be hope.  Check this now.
675  if (FirstConstantOper == MinOperands) {
676    // Make GEP1Ops be the longer one if there is a longer one.
677    if (NumGEP1Ops < NumGEP2Ops) {
678      std::swap(GEP1Ops, GEP2Ops);
679      std::swap(NumGEP1Ops, NumGEP2Ops);
680    }
681
682    // Is there anything to check?
683    if (NumGEP1Ops > MinOperands) {
684      for (unsigned i = FirstConstantOper; i != MaxOperands; ++i)
685        if (isa<ConstantInt>(GEP1Ops[i]) &&
686            !cast<ConstantInt>(GEP1Ops[i])->isZero()) {
687          // Yup, there's a constant in the tail.  Set all variables to
688          // constants in the GEP instruction to make it suiteable for
689          // TargetData::getIndexedOffset.
690          for (i = 0; i != MaxOperands; ++i)
691            if (!isa<ConstantInt>(GEP1Ops[i]))
692              GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType());
693          // Okay, now get the offset.  This is the relative offset for the full
694          // instruction.
695          const TargetData &TD = getTargetData();
696          int64_t Offset1 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops,
697                                                NumGEP1Ops);
698
699          // Now check without any constants at the end.
700          int64_t Offset2 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops,
701                                                MinOperands);
702
703          // If the tail provided a bit enough offset, return noalias!
704          if ((uint64_t)(Offset2-Offset1) >= SizeMax)
705            return NoAlias;
706        }
707    }
708
709    // Couldn't find anything useful.
710    return MayAlias;
711  }
712
713  // If there are non-equal constants arguments, then we can figure
714  // out a minimum known delta between the two index expressions... at
715  // this point we know that the first constant index of GEP1 is less
716  // than the first constant index of GEP2.
717
718  // Advance BasePtr[12]Ty over this first differing constant operand.
719  BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)->
720      getTypeAtIndex(GEP2Ops[FirstConstantOper]);
721  BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->
722      getTypeAtIndex(GEP1Ops[FirstConstantOper]);
723
724  // We are going to be using TargetData::getIndexedOffset to determine the
725  // offset that each of the GEP's is reaching.  To do this, we have to convert
726  // all variable references to constant references.  To do this, we convert the
727  // initial sequence of array subscripts into constant zeros to start with.
728  const Type *ZeroIdxTy = GEPPointerTy;
729  for (unsigned i = 0; i != FirstConstantOper; ++i) {
730    if (!isa<StructType>(ZeroIdxTy))
731      GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Type::Int32Ty);
732
733    if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy))
734      ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]);
735  }
736
737  // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok
738
739  // Loop over the rest of the operands...
740  for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) {
741    const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0;
742    const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0;
743    // If they are equal, use a zero index...
744    if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) {
745      if (!isa<ConstantInt>(Op1))
746        GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType());
747      // Otherwise, just keep the constants we have.
748    } else {
749      if (Op1) {
750        if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
751          // If this is an array index, make sure the array element is in range.
752          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) {
753            if (Op1C->getZExtValue() >= AT->getNumElements())
754              return MayAlias;  // Be conservative with out-of-range accesses
755          } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) {
756            if (Op1C->getZExtValue() >= VT->getNumElements())
757              return MayAlias;  // Be conservative with out-of-range accesses
758          }
759
760        } else {
761          // GEP1 is known to produce a value less than GEP2.  To be
762          // conservatively correct, we must assume the largest possible
763          // constant is used in this position.  This cannot be the initial
764          // index to the GEP instructions (because we know we have at least one
765          // element before this one with the different constant arguments), so
766          // we know that the current index must be into either a struct or
767          // array.  Because we know it's not constant, this cannot be a
768          // structure index.  Because of this, we can calculate the maximum
769          // value possible.
770          //
771          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
772            GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,AT->getNumElements()-1);
773          else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty))
774            GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,VT->getNumElements()-1);
775        }
776      }
777
778      if (Op2) {
779        if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) {
780          // If this is an array index, make sure the array element is in range.
781          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) {
782            if (Op2C->getZExtValue() >= AT->getNumElements())
783              return MayAlias;  // Be conservative with out-of-range accesses
784          } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) {
785            if (Op2C->getZExtValue() >= VT->getNumElements())
786              return MayAlias;  // Be conservative with out-of-range accesses
787          }
788        } else {  // Conservatively assume the minimum value for this index
789          GEP2Ops[i] = Constant::getNullValue(Op2->getType());
790        }
791      }
792    }
793
794    if (BasePtr1Ty && Op1) {
795      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
796        BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]);
797      else
798        BasePtr1Ty = 0;
799    }
800
801    if (BasePtr2Ty && Op2) {
802      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty))
803        BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]);
804      else
805        BasePtr2Ty = 0;
806    }
807  }
808
809  if (GEPPointerTy->getElementType()->isSized()) {
810    int64_t Offset1 =
811      getTargetData().getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops);
812    int64_t Offset2 =
813      getTargetData().getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops);
814    assert(Offset1 != Offset2 &&
815           "There is at least one different constant here!");
816
817    // Make sure we compare the absolute difference.
818    if (Offset1 > Offset2)
819      std::swap(Offset1, Offset2);
820
821    if ((uint64_t)(Offset2-Offset1) >= SizeMax) {
822      //cerr << "Determined that these two GEP's don't alias ["
823      //     << SizeMax << " bytes]: \n" << *GEP1 << *GEP2;
824      return NoAlias;
825    }
826  }
827  return MayAlias;
828}
829
830// Make sure that anything that uses AliasAnalysis pulls in this file...
831DEFINING_FILE_FOR(BasicAliasAnalysis)
832