BasicAliasAnalysis.cpp revision 194710
1//===- BasicAliasAnalysis.cpp - Local Alias Analysis Impl -----------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines the default implementation of the Alias Analysis interface
11// that simply implements a few identities (two different globals cannot alias,
12// etc), but otherwise does no analysis.
13//
14//===----------------------------------------------------------------------===//
15
16#include "llvm/Analysis/AliasAnalysis.h"
17#include "llvm/Analysis/CaptureTracking.h"
18#include "llvm/Analysis/Passes.h"
19#include "llvm/Constants.h"
20#include "llvm/DerivedTypes.h"
21#include "llvm/Function.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 <algorithm>
32using namespace llvm;
33
34//===----------------------------------------------------------------------===//
35// Useful predicates
36//===----------------------------------------------------------------------===//
37
38static const User *isGEP(const Value *V) {
39  if (isa<GetElementPtrInst>(V) ||
40      (isa<ConstantExpr>(V) &&
41       cast<ConstantExpr>(V)->getOpcode() == Instruction::GetElementPtr))
42    return cast<User>(V);
43  return 0;
44}
45
46static const Value *GetGEPOperands(const Value *V,
47                                   SmallVector<Value*, 16> &GEPOps) {
48  assert(GEPOps.empty() && "Expect empty list to populate!");
49  GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1,
50                cast<User>(V)->op_end());
51
52  // Accumulate all of the chained indexes into the operand array
53  V = cast<User>(V)->getOperand(0);
54
55  while (const User *G = isGEP(V)) {
56    if (!isa<Constant>(GEPOps[0]) || isa<GlobalValue>(GEPOps[0]) ||
57        !cast<Constant>(GEPOps[0])->isNullValue())
58      break;  // Don't handle folding arbitrary pointer offsets yet...
59    GEPOps.erase(GEPOps.begin());   // Drop the zero index
60    GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end());
61    V = G->getOperand(0);
62  }
63  return V;
64}
65
66/// isKnownNonNull - Return true if we know that the specified value is never
67/// null.
68static bool isKnownNonNull(const Value *V) {
69  // Alloca never returns null, malloc might.
70  if (isa<AllocaInst>(V)) return true;
71
72  // A byval argument is never null.
73  if (const Argument *A = dyn_cast<Argument>(V))
74    return A->hasByValAttr();
75
76  // Global values are not null unless extern weak.
77  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
78    return !GV->hasExternalWeakLinkage();
79  return false;
80}
81
82/// isNonEscapingLocalObject - Return true if the pointer is to a function-local
83/// object that never escapes from the function.
84static bool isNonEscapingLocalObject(const Value *V) {
85  // If this is a local allocation, check to see if it escapes.
86  if (isa<AllocationInst>(V) || isNoAliasCall(V))
87    return !PointerMayBeCaptured(V, false);
88
89  // If this is an argument that corresponds to a byval or noalias argument,
90  // then it has not escaped before entering the function.  Check if it escapes
91  // inside the function.
92  if (const Argument *A = dyn_cast<Argument>(V))
93    if (A->hasByValAttr() || A->hasNoAliasAttr()) {
94      // Don't bother analyzing arguments already known not to escape.
95      if (A->hasNoCaptureAttr())
96        return true;
97      return !PointerMayBeCaptured(V, false);
98    }
99  return false;
100}
101
102
103/// isObjectSmallerThan - Return true if we can prove that the object specified
104/// by V is smaller than Size.
105static bool isObjectSmallerThan(const Value *V, unsigned Size,
106                                const TargetData &TD) {
107  const Type *AccessTy;
108  if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
109    AccessTy = GV->getType()->getElementType();
110  } else if (const AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
111    if (!AI->isArrayAllocation())
112      AccessTy = AI->getType()->getElementType();
113    else
114      return false;
115  } else if (const Argument *A = dyn_cast<Argument>(V)) {
116    if (A->hasByValAttr())
117      AccessTy = cast<PointerType>(A->getType())->getElementType();
118    else
119      return false;
120  } else {
121    return false;
122  }
123
124  if (AccessTy->isSized())
125    return TD.getTypeAllocSize(AccessTy) < Size;
126  return false;
127}
128
129//===----------------------------------------------------------------------===//
130// NoAA Pass
131//===----------------------------------------------------------------------===//
132
133namespace {
134  /// NoAA - This class implements the -no-aa pass, which always returns "I
135  /// don't know" for alias queries.  NoAA is unlike other alias analysis
136  /// implementations, in that it does not chain to a previous analysis.  As
137  /// such it doesn't follow many of the rules that other alias analyses must.
138  ///
139  struct VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis {
140    static char ID; // Class identification, replacement for typeinfo
141    NoAA() : ImmutablePass(&ID) {}
142    explicit NoAA(void *PID) : ImmutablePass(PID) { }
143
144    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
145      AU.addRequired<TargetData>();
146    }
147
148    virtual void initializePass() {
149      TD = &getAnalysis<TargetData>();
150    }
151
152    virtual AliasResult alias(const Value *V1, unsigned V1Size,
153                              const Value *V2, unsigned V2Size) {
154      return MayAlias;
155    }
156
157    virtual void getArgumentAccesses(Function *F, CallSite CS,
158                                     std::vector<PointerAccessInfo> &Info) {
159      assert(0 && "This method may not be called on this function!");
160    }
161
162    virtual void getMustAliases(Value *P, std::vector<Value*> &RetVals) { }
163    virtual bool pointsToConstantMemory(const Value *P) { return false; }
164    virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) {
165      return ModRef;
166    }
167    virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
168      return ModRef;
169    }
170    virtual bool hasNoModRefInfoForCalls() const { return true; }
171
172    virtual void deleteValue(Value *V) {}
173    virtual void copyValue(Value *From, Value *To) {}
174  };
175}  // End of anonymous namespace
176
177// Register this pass...
178char NoAA::ID = 0;
179static RegisterPass<NoAA>
180U("no-aa", "No Alias Analysis (always returns 'may' alias)", true, true);
181
182// Declare that we implement the AliasAnalysis interface
183static RegisterAnalysisGroup<AliasAnalysis> V(U);
184
185ImmutablePass *llvm::createNoAAPass() { return new NoAA(); }
186
187//===----------------------------------------------------------------------===//
188// BasicAA Pass
189//===----------------------------------------------------------------------===//
190
191namespace {
192  /// BasicAliasAnalysis - This is the default alias analysis implementation.
193  /// Because it doesn't chain to a previous alias analysis (like -no-aa), it
194  /// derives from the NoAA class.
195  struct VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA {
196    static char ID; // Class identification, replacement for typeinfo
197    BasicAliasAnalysis() : NoAA(&ID) {}
198    AliasResult alias(const Value *V1, unsigned V1Size,
199                      const Value *V2, unsigned V2Size);
200
201    ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
202    ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
203
204    /// hasNoModRefInfoForCalls - We can provide mod/ref information against
205    /// non-escaping allocations.
206    virtual bool hasNoModRefInfoForCalls() const { return false; }
207
208    /// pointsToConstantMemory - Chase pointers until we find a (constant
209    /// global) or not.
210    bool pointsToConstantMemory(const Value *P);
211
212  private:
213    // CheckGEPInstructions - Check two GEP instructions with known
214    // must-aliasing base pointers.  This checks to see if the index expressions
215    // preclude the pointers from aliasing...
216    AliasResult
217    CheckGEPInstructions(const Type* BasePtr1Ty,
218                         Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size,
219                         const Type *BasePtr2Ty,
220                         Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size);
221  };
222}  // End of anonymous namespace
223
224// Register this pass...
225char BasicAliasAnalysis::ID = 0;
226static RegisterPass<BasicAliasAnalysis>
227X("basicaa", "Basic Alias Analysis (default AA impl)", false, true);
228
229// Declare that we implement the AliasAnalysis interface
230static RegisterAnalysisGroup<AliasAnalysis, true> Y(X);
231
232ImmutablePass *llvm::createBasicAliasAnalysisPass() {
233  return new BasicAliasAnalysis();
234}
235
236
237/// pointsToConstantMemory - Chase pointers until we find a (constant
238/// global) or not.
239bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) {
240  if (const GlobalVariable *GV =
241        dyn_cast<GlobalVariable>(P->getUnderlyingObject()))
242    return GV->isConstant();
243  return false;
244}
245
246
247// getModRefInfo - Check to see if the specified callsite can clobber the
248// specified memory object.  Since we only look at local properties of this
249// function, we really can't say much about this query.  We do, however, use
250// simple "address taken" analysis on local objects.
251//
252AliasAnalysis::ModRefResult
253BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
254  if (!isa<Constant>(P)) {
255    const Value *Object = P->getUnderlyingObject();
256
257    // If this is a tail call and P points to a stack location, we know that
258    // the tail call cannot access or modify the local stack.
259    // We cannot exclude byval arguments here; these belong to the caller of
260    // the current function not to the current function, and a tail callee
261    // may reference them.
262    if (isa<AllocaInst>(Object))
263      if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
264        if (CI->isTailCall())
265          return NoModRef;
266
267    // If the pointer is to a locally allocated object that does not escape,
268    // then the call can not mod/ref the pointer unless the call takes the
269    // argument without capturing it.
270    if (isNonEscapingLocalObject(Object) && CS.getInstruction() != Object) {
271      bool passedAsArg = false;
272      // TODO: Eventually only check 'nocapture' arguments.
273      for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
274           CI != CE; ++CI)
275        if (isa<PointerType>((*CI)->getType()) &&
276            alias(cast<Value>(CI), ~0U, P, ~0U) != NoAlias)
277          passedAsArg = true;
278
279      if (!passedAsArg)
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
289AliasAnalysis::ModRefResult
290BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) {
291  // If CS1 or CS2 are readnone, they don't interact.
292  ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1);
293  if (CS1B == DoesNotAccessMemory) return NoModRef;
294
295  ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2);
296  if (CS2B == DoesNotAccessMemory) return NoModRef;
297
298  // If they both only read from memory, just return ref.
299  if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory)
300    return Ref;
301
302  // Otherwise, fall back to NoAA (mod+ref).
303  return NoAA::getModRefInfo(CS1, CS2);
304}
305
306
307// alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such
308// as array references.
309//
310AliasAnalysis::AliasResult
311BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size,
312                          const Value *V2, unsigned V2Size) {
313  // Strip off any constant expression casts if they exist
314  if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V1))
315    if (CE->isCast() && isa<PointerType>(CE->getOperand(0)->getType()))
316      V1 = CE->getOperand(0);
317  if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V2))
318    if (CE->isCast() && isa<PointerType>(CE->getOperand(0)->getType()))
319      V2 = CE->getOperand(0);
320
321  // Are we checking for alias of the same value?
322  if (V1 == V2) return MustAlias;
323
324  if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType()))
325    return NoAlias;  // Scalars cannot alias each other
326
327  // Strip off cast instructions.   Since V1 and V2 are pointers, they must be
328  // pointer<->pointer bitcasts.
329  if (const BitCastInst *I = dyn_cast<BitCastInst>(V1))
330    return alias(I->getOperand(0), V1Size, V2, V2Size);
331  if (const BitCastInst *I = dyn_cast<BitCastInst>(V2))
332    return alias(V1, V1Size, I->getOperand(0), V2Size);
333
334  // Figure out what objects these things are pointing to if we can.
335  const Value *O1 = V1->getUnderlyingObject();
336  const Value *O2 = V2->getUnderlyingObject();
337
338  if (O1 != O2) {
339    // If V1/V2 point to two different objects we know that we have no alias.
340    if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
341      return NoAlias;
342
343    // Arguments can't alias with local allocations or noalias calls.
344    if ((isa<Argument>(O1) && (isa<AllocationInst>(O2) || isNoAliasCall(O2))) ||
345        (isa<Argument>(O2) && (isa<AllocationInst>(O1) || isNoAliasCall(O1))))
346      return NoAlias;
347
348    // Most objects can't alias null.
349    if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) ||
350        (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2)))
351      return NoAlias;
352  }
353
354  // If the size of one access is larger than the entire object on the other
355  // side, then we know such behavior is undefined and can assume no alias.
356  const TargetData &TD = getTargetData();
357  if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, TD)) ||
358      (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, TD)))
359    return NoAlias;
360
361  // If one pointer is the result of a call/invoke and the other is a
362  // non-escaping local object, then we know the object couldn't escape to a
363  // point where the call could return it.
364  if ((isa<CallInst>(O1) || isa<InvokeInst>(O1)) &&
365      isNonEscapingLocalObject(O2) && O1 != O2)
366    return NoAlias;
367  if ((isa<CallInst>(O2) || isa<InvokeInst>(O2)) &&
368      isNonEscapingLocalObject(O1) && O1 != O2)
369    return NoAlias;
370
371  // If we have two gep instructions with must-alias'ing base pointers, figure
372  // out if the indexes to the GEP tell us anything about the derived pointer.
373  // Note that we also handle chains of getelementptr instructions as well as
374  // constant expression getelementptrs here.
375  //
376  if (isGEP(V1) && isGEP(V2)) {
377    const User *GEP1 = cast<User>(V1);
378    const User *GEP2 = cast<User>(V2);
379
380    // If V1 and V2 are identical GEPs, just recurse down on both of them.
381    // This allows us to analyze things like:
382    //   P = gep A, 0, i, 1
383    //   Q = gep B, 0, i, 1
384    // by just analyzing A and B.  This is even safe for variable indices.
385    if (GEP1->getType() == GEP2->getType() &&
386        GEP1->getNumOperands() == GEP2->getNumOperands() &&
387        GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() &&
388        // All operands are the same, ignoring the base.
389        std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1))
390      return alias(GEP1->getOperand(0), V1Size, GEP2->getOperand(0), V2Size);
391
392
393    // Drill down into the first non-gep value, to test for must-aliasing of
394    // the base pointers.
395    while (isGEP(GEP1->getOperand(0)) &&
396           GEP1->getOperand(1) ==
397           Constant::getNullValue(GEP1->getOperand(1)->getType()))
398      GEP1 = cast<User>(GEP1->getOperand(0));
399    const Value *BasePtr1 = GEP1->getOperand(0);
400
401    while (isGEP(GEP2->getOperand(0)) &&
402           GEP2->getOperand(1) ==
403           Constant::getNullValue(GEP2->getOperand(1)->getType()))
404      GEP2 = cast<User>(GEP2->getOperand(0));
405    const Value *BasePtr2 = GEP2->getOperand(0);
406
407    // Do the base pointers alias?
408    AliasResult BaseAlias = alias(BasePtr1, ~0U, BasePtr2, ~0U);
409    if (BaseAlias == NoAlias) return NoAlias;
410    if (BaseAlias == MustAlias) {
411      // If the base pointers alias each other exactly, check to see if we can
412      // figure out anything about the resultant pointers, to try to prove
413      // non-aliasing.
414
415      // Collect all of the chained GEP operands together into one simple place
416      SmallVector<Value*, 16> GEP1Ops, GEP2Ops;
417      BasePtr1 = GetGEPOperands(V1, GEP1Ops);
418      BasePtr2 = GetGEPOperands(V2, GEP2Ops);
419
420      // If GetGEPOperands were able to fold to the same must-aliased pointer,
421      // do the comparison.
422      if (BasePtr1 == BasePtr2) {
423        AliasResult GAlias =
424          CheckGEPInstructions(BasePtr1->getType(),
425                               &GEP1Ops[0], GEP1Ops.size(), V1Size,
426                               BasePtr2->getType(),
427                               &GEP2Ops[0], GEP2Ops.size(), V2Size);
428        if (GAlias != MayAlias)
429          return GAlias;
430      }
431    }
432  }
433
434  // Check to see if these two pointers are related by a getelementptr
435  // instruction.  If one pointer is a GEP with a non-zero index of the other
436  // pointer, we know they cannot alias.
437  //
438  if (isGEP(V2)) {
439    std::swap(V1, V2);
440    std::swap(V1Size, V2Size);
441  }
442
443  if (V1Size != ~0U && V2Size != ~0U)
444    if (isGEP(V1)) {
445      SmallVector<Value*, 16> GEPOperands;
446      const Value *BasePtr = GetGEPOperands(V1, GEPOperands);
447
448      AliasResult R = alias(BasePtr, V1Size, V2, V2Size);
449      if (R == MustAlias) {
450        // If there is at least one non-zero constant index, we know they cannot
451        // alias.
452        bool ConstantFound = false;
453        bool AllZerosFound = true;
454        for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i)
455          if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) {
456            if (!C->isNullValue()) {
457              ConstantFound = true;
458              AllZerosFound = false;
459              break;
460            }
461          } else {
462            AllZerosFound = false;
463          }
464
465        // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases
466        // the ptr, the end result is a must alias also.
467        if (AllZerosFound)
468          return MustAlias;
469
470        if (ConstantFound) {
471          if (V2Size <= 1 && V1Size <= 1)  // Just pointer check?
472            return NoAlias;
473
474          // Otherwise we have to check to see that the distance is more than
475          // the size of the argument... build an index vector that is equal to
476          // the arguments provided, except substitute 0's for any variable
477          // indexes we find...
478          if (cast<PointerType>(
479                BasePtr->getType())->getElementType()->isSized()) {
480            for (unsigned i = 0; i != GEPOperands.size(); ++i)
481              if (!isa<ConstantInt>(GEPOperands[i]))
482                GEPOperands[i] =
483                  Constant::getNullValue(GEPOperands[i]->getType());
484            int64_t Offset =
485              getTargetData().getIndexedOffset(BasePtr->getType(),
486                                               &GEPOperands[0],
487                                               GEPOperands.size());
488
489            if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size)
490              return NoAlias;
491          }
492        }
493      }
494    }
495
496  return MayAlias;
497}
498
499// This function is used to determine if the indices of two GEP instructions are
500// equal. V1 and V2 are the indices.
501static bool IndexOperandsEqual(Value *V1, Value *V2) {
502  if (V1->getType() == V2->getType())
503    return V1 == V2;
504  if (Constant *C1 = dyn_cast<Constant>(V1))
505    if (Constant *C2 = dyn_cast<Constant>(V2)) {
506      // Sign extend the constants to long types, if necessary
507      if (C1->getType() != Type::Int64Ty)
508        C1 = ConstantExpr::getSExt(C1, Type::Int64Ty);
509      if (C2->getType() != Type::Int64Ty)
510        C2 = ConstantExpr::getSExt(C2, Type::Int64Ty);
511      return C1 == C2;
512    }
513  return false;
514}
515
516/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing
517/// base pointers.  This checks to see if the index expressions preclude the
518/// pointers from aliasing...
519AliasAnalysis::AliasResult
520BasicAliasAnalysis::CheckGEPInstructions(
521  const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S,
522  const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) {
523  // We currently can't handle the case when the base pointers have different
524  // primitive types.  Since this is uncommon anyway, we are happy being
525  // extremely conservative.
526  if (BasePtr1Ty != BasePtr2Ty)
527    return MayAlias;
528
529  const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty);
530
531  // Find the (possibly empty) initial sequence of equal values... which are not
532  // necessarily constants.
533  unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops;
534  unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands);
535  unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands);
536  unsigned UnequalOper = 0;
537  while (UnequalOper != MinOperands &&
538         IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) {
539    // Advance through the type as we go...
540    ++UnequalOper;
541    if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
542      BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]);
543    else {
544      // If all operands equal each other, then the derived pointers must
545      // alias each other...
546      BasePtr1Ty = 0;
547      assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands &&
548             "Ran out of type nesting, but not out of operands?");
549      return MustAlias;
550    }
551  }
552
553  // If we have seen all constant operands, and run out of indexes on one of the
554  // getelementptrs, check to see if the tail of the leftover one is all zeros.
555  // If so, return mustalias.
556  if (UnequalOper == MinOperands) {
557    if (NumGEP1Ops < NumGEP2Ops) {
558      std::swap(GEP1Ops, GEP2Ops);
559      std::swap(NumGEP1Ops, NumGEP2Ops);
560    }
561
562    bool AllAreZeros = true;
563    for (unsigned i = UnequalOper; i != MaxOperands; ++i)
564      if (!isa<Constant>(GEP1Ops[i]) ||
565          !cast<Constant>(GEP1Ops[i])->isNullValue()) {
566        AllAreZeros = false;
567        break;
568      }
569    if (AllAreZeros) return MustAlias;
570  }
571
572
573  // So now we know that the indexes derived from the base pointers,
574  // which are known to alias, are different.  We can still determine a
575  // no-alias result if there are differing constant pairs in the index
576  // chain.  For example:
577  //        A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S))
578  //
579  // We have to be careful here about array accesses.  In particular, consider:
580  //        A[1][0] vs A[0][i]
581  // In this case, we don't *know* that the array will be accessed in bounds:
582  // the index could even be negative.  Because of this, we have to
583  // conservatively *give up* and return may alias.  We disregard differing
584  // array subscripts that are followed by a variable index without going
585  // through a struct.
586  //
587  unsigned SizeMax = std::max(G1S, G2S);
588  if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work.
589
590  // Scan for the first operand that is constant and unequal in the
591  // two getelementptrs...
592  unsigned FirstConstantOper = UnequalOper;
593  for (; FirstConstantOper != MinOperands; ++FirstConstantOper) {
594    const Value *G1Oper = GEP1Ops[FirstConstantOper];
595    const Value *G2Oper = GEP2Ops[FirstConstantOper];
596
597    if (G1Oper != G2Oper)   // Found non-equal constant indexes...
598      if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper)))
599        if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){
600          if (G1OC->getType() != G2OC->getType()) {
601            // Sign extend both operands to long.
602            if (G1OC->getType() != Type::Int64Ty)
603              G1OC = ConstantExpr::getSExt(G1OC, Type::Int64Ty);
604            if (G2OC->getType() != Type::Int64Ty)
605              G2OC = ConstantExpr::getSExt(G2OC, Type::Int64Ty);
606            GEP1Ops[FirstConstantOper] = G1OC;
607            GEP2Ops[FirstConstantOper] = G2OC;
608          }
609
610          if (G1OC != G2OC) {
611            // Handle the "be careful" case above: if this is an array/vector
612            // subscript, scan for a subsequent variable array index.
613            if (const SequentialType *STy =
614                  dyn_cast<SequentialType>(BasePtr1Ty)) {
615              const Type *NextTy = STy;
616              bool isBadCase = false;
617
618              for (unsigned Idx = FirstConstantOper;
619                   Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) {
620                const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx];
621                if (!isa<Constant>(V1) || !isa<Constant>(V2)) {
622                  isBadCase = true;
623                  break;
624                }
625                // If the array is indexed beyond the bounds of the static type
626                // at this level, it will also fall into the "be careful" case.
627                // It would theoretically be possible to analyze these cases,
628                // but for now just be conservatively correct.
629                if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
630                  if (cast<ConstantInt>(G1OC)->getZExtValue() >=
631                        ATy->getNumElements() ||
632                      cast<ConstantInt>(G2OC)->getZExtValue() >=
633                        ATy->getNumElements()) {
634                    isBadCase = true;
635                    break;
636                  }
637                if (const VectorType *VTy = dyn_cast<VectorType>(STy))
638                  if (cast<ConstantInt>(G1OC)->getZExtValue() >=
639                        VTy->getNumElements() ||
640                      cast<ConstantInt>(G2OC)->getZExtValue() >=
641                        VTy->getNumElements()) {
642                    isBadCase = true;
643                    break;
644                  }
645                STy = cast<SequentialType>(NextTy);
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 suitable 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          // Make sure we compare the absolute difference.
704          if (Offset1 > Offset2)
705            std::swap(Offset1, Offset2);
706
707          // If the tail provided a bit enough offset, return noalias!
708          if ((uint64_t)(Offset2-Offset1) >= SizeMax)
709            return NoAlias;
710          // Otherwise break - we don't look for another constant in the tail.
711          break;
712        }
713    }
714
715    // Couldn't find anything useful.
716    return MayAlias;
717  }
718
719  // If there are non-equal constants arguments, then we can figure
720  // out a minimum known delta between the two index expressions... at
721  // this point we know that the first constant index of GEP1 is less
722  // than the first constant index of GEP2.
723
724  // Advance BasePtr[12]Ty over this first differing constant operand.
725  BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)->
726      getTypeAtIndex(GEP2Ops[FirstConstantOper]);
727  BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->
728      getTypeAtIndex(GEP1Ops[FirstConstantOper]);
729
730  // We are going to be using TargetData::getIndexedOffset to determine the
731  // offset that each of the GEP's is reaching.  To do this, we have to convert
732  // all variable references to constant references.  To do this, we convert the
733  // initial sequence of array subscripts into constant zeros to start with.
734  const Type *ZeroIdxTy = GEPPointerTy;
735  for (unsigned i = 0; i != FirstConstantOper; ++i) {
736    if (!isa<StructType>(ZeroIdxTy))
737      GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Type::Int32Ty);
738
739    if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy))
740      ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]);
741  }
742
743  // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok
744
745  // Loop over the rest of the operands...
746  for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) {
747    const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0;
748    const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0;
749    // If they are equal, use a zero index...
750    if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) {
751      if (!isa<ConstantInt>(Op1))
752        GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType());
753      // Otherwise, just keep the constants we have.
754    } else {
755      if (Op1) {
756        if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
757          // If this is an array index, make sure the array element is in range.
758          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) {
759            if (Op1C->getZExtValue() >= AT->getNumElements())
760              return MayAlias;  // Be conservative with out-of-range accesses
761          } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) {
762            if (Op1C->getZExtValue() >= VT->getNumElements())
763              return MayAlias;  // Be conservative with out-of-range accesses
764          }
765
766        } else {
767          // GEP1 is known to produce a value less than GEP2.  To be
768          // conservatively correct, we must assume the largest possible
769          // constant is used in this position.  This cannot be the initial
770          // index to the GEP instructions (because we know we have at least one
771          // element before this one with the different constant arguments), so
772          // we know that the current index must be into either a struct or
773          // array.  Because we know it's not constant, this cannot be a
774          // structure index.  Because of this, we can calculate the maximum
775          // value possible.
776          //
777          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
778            GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,AT->getNumElements()-1);
779          else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty))
780            GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,VT->getNumElements()-1);
781        }
782      }
783
784      if (Op2) {
785        if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) {
786          // If this is an array index, make sure the array element is in range.
787          if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) {
788            if (Op2C->getZExtValue() >= AT->getNumElements())
789              return MayAlias;  // Be conservative with out-of-range accesses
790          } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) {
791            if (Op2C->getZExtValue() >= VT->getNumElements())
792              return MayAlias;  // Be conservative with out-of-range accesses
793          }
794        } else {  // Conservatively assume the minimum value for this index
795          GEP2Ops[i] = Constant::getNullValue(Op2->getType());
796        }
797      }
798    }
799
800    if (BasePtr1Ty && Op1) {
801      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
802        BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]);
803      else
804        BasePtr1Ty = 0;
805    }
806
807    if (BasePtr2Ty && Op2) {
808      if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty))
809        BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]);
810      else
811        BasePtr2Ty = 0;
812    }
813  }
814
815  if (GEPPointerTy->getElementType()->isSized()) {
816    int64_t Offset1 =
817      getTargetData().getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops);
818    int64_t Offset2 =
819      getTargetData().getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops);
820    assert(Offset1 != Offset2 &&
821           "There is at least one different constant here!");
822
823    // Make sure we compare the absolute difference.
824    if (Offset1 > Offset2)
825      std::swap(Offset1, Offset2);
826
827    if ((uint64_t)(Offset2-Offset1) >= SizeMax) {
828      //cerr << "Determined that these two GEP's don't alias ["
829      //     << SizeMax << " bytes]: \n" << *GEP1 << *GEP2;
830      return NoAlias;
831    }
832  }
833  return MayAlias;
834}
835
836// Make sure that anything that uses AliasAnalysis pulls in this file...
837DEFINING_FILE_FOR(BasicAliasAnalysis)
838