LLVM API Documentation

BasicAliasAnalysis.cpp

Go to the documentation of this file.
00001 //===- BasicAliasAnalysis.cpp - Local Alias Analysis Impl -----------------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This file defines the default implementation of the Alias Analysis interface
00011 // that simply implements a few identities (two different globals cannot alias,
00012 // etc), but otherwise does no analysis.
00013 //
00014 //===----------------------------------------------------------------------===//
00015 
00016 #include "llvm/Analysis/AliasAnalysis.h"
00017 #include "llvm/Analysis/Passes.h"
00018 #include "llvm/Constants.h"
00019 #include "llvm/DerivedTypes.h"
00020 #include "llvm/Function.h"
00021 #include "llvm/GlobalVariable.h"
00022 #include "llvm/Instructions.h"
00023 #include "llvm/IntrinsicInst.h"
00024 #include "llvm/Pass.h"
00025 #include "llvm/Target/TargetData.h"
00026 #include "llvm/ADT/SmallVector.h"
00027 #include "llvm/ADT/STLExtras.h"
00028 #include "llvm/Support/Compiler.h"
00029 #include "llvm/Support/GetElementPtrTypeIterator.h"
00030 #include "llvm/Support/ManagedStatic.h"
00031 #include <algorithm>
00032 using namespace llvm;
00033 
00034 //===----------------------------------------------------------------------===//
00035 // Useful predicates
00036 //===----------------------------------------------------------------------===//
00037 
00038 // Determine if an AllocationInst instruction escapes from the function it is
00039 // contained in. If it does not escape, there is no way for another function to
00040 // mod/ref it.  We do this by looking at its uses and determining if the uses
00041 // can escape (recursively).
00042 static bool AddressMightEscape(const Value *V) {
00043   for (Value::use_const_iterator UI = V->use_begin(), E = V->use_end();
00044        UI != E; ++UI) {
00045     const Instruction *I = cast<Instruction>(*UI);
00046     switch (I->getOpcode()) {
00047     case Instruction::Load: 
00048       break; //next use.
00049     case Instruction::Store:
00050       if (I->getOperand(0) == V)
00051         return true; // Escapes if the pointer is stored.
00052       break; // next use.
00053     case Instruction::GetElementPtr:
00054       if (AddressMightEscape(I))
00055         return true;
00056       break; // next use.
00057     case Instruction::BitCast:
00058       if (AddressMightEscape(I))
00059         return true;
00060       break; // next use
00061     case Instruction::Ret:
00062       // If returned, the address will escape to calling functions, but no
00063       // callees could modify it.
00064       break; // next use
00065     case Instruction::Call:
00066       // If the call is to a few known safe intrinsics, we know that it does
00067       // not escape.
00068       // TODO: Eventually just check the 'nocapture' attribute.
00069       if (!isa<MemIntrinsic>(I))
00070         return true;
00071       break;  // next use
00072     default:
00073       return true;
00074     }
00075   }
00076   return false;
00077 }
00078 
00079 static const User *isGEP(const Value *V) {
00080   if (isa<GetElementPtrInst>(V) ||
00081       (isa<ConstantExpr>(V) &&
00082        cast<ConstantExpr>(V)->getOpcode() == Instruction::GetElementPtr))
00083     return cast<User>(V);
00084   return 0;
00085 }
00086 
00087 static const Value *GetGEPOperands(const Value *V, 
00088                                    SmallVector<Value*, 16> &GEPOps){
00089   assert(GEPOps.empty() && "Expect empty list to populate!");
00090   GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1,
00091                 cast<User>(V)->op_end());
00092 
00093   // Accumulate all of the chained indexes into the operand array
00094   V = cast<User>(V)->getOperand(0);
00095 
00096   while (const User *G = isGEP(V)) {
00097     if (!isa<Constant>(GEPOps[0]) || isa<GlobalValue>(GEPOps[0]) ||
00098         !cast<Constant>(GEPOps[0])->isNullValue())
00099       break;  // Don't handle folding arbitrary pointer offsets yet...
00100     GEPOps.erase(GEPOps.begin());   // Drop the zero index
00101     GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end());
00102     V = G->getOperand(0);
00103   }
00104   return V;
00105 }
00106 
00107 /// isIdentifiedObject - Return true if this pointer refers to a distinct and
00108 /// identifiable object.  This returns true for:
00109 ///    Global Variables and Functions
00110 ///    Allocas and Mallocs
00111 ///    ByVal and NoAlias Arguments
00112 ///
00113 static bool isIdentifiedObject(const Value *V) {
00114   if (isa<GlobalValue>(V) || isa<AllocationInst>(V))
00115     return true;
00116   if (const Argument *A = dyn_cast<Argument>(V))
00117     return A->hasNoAliasAttr() || A->hasByValAttr();
00118   return false;
00119 }
00120 
00121 /// isKnownNonNull - Return true if we know that the specified value is never
00122 /// null.
00123 static bool isKnownNonNull(const Value *V) {
00124   // Alloca never returns null, malloc might.
00125   if (isa<AllocaInst>(V)) return true;
00126   
00127   // A byval argument is never null.
00128   if (const Argument *A = dyn_cast<Argument>(V))
00129     return A->hasByValAttr();
00130 
00131   // Global values are not null unless extern weak.
00132   if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
00133     return !GV->hasExternalWeakLinkage();
00134   return false;
00135 }
00136 
00137 /// isNonEscapingLocalObject - Return true if the pointer is to a function-local
00138 /// object that never escapes from the function.
00139 static bool isNonEscapingLocalObject(const Value *V) {
00140   // If this is a local allocation, check to see if it escapes.
00141   if (isa<AllocationInst>(V))
00142     return !AddressMightEscape(V);
00143       
00144   // If this is an argument that corresponds to a byval or noalias argument,
00145   // it can't escape either.
00146   if (const Argument *A = dyn_cast<Argument>(V))
00147     if (A->hasByValAttr() || A->hasNoAliasAttr())
00148       return !AddressMightEscape(V);
00149   return false;
00150 }
00151 
00152 
00153 /// isObjectSmallerThan - Return true if we can prove that the object specified
00154 /// by V is smaller than Size.
00155 static bool isObjectSmallerThan(const Value *V, unsigned Size,
00156                                 const TargetData &TD) {
00157   const Type *AccessTy = 0;
00158   if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
00159     AccessTy = GV->getType()->getElementType();
00160   
00161   if (const AllocationInst *AI = dyn_cast<AllocationInst>(V))
00162     if (!AI->isArrayAllocation())
00163       AccessTy = AI->getType()->getElementType();
00164 
00165   if (const Argument *A = dyn_cast<Argument>(V))
00166     if (A->hasByValAttr())
00167       AccessTy = cast<PointerType>(A->getType())->getElementType();
00168   
00169   if (AccessTy && AccessTy->isSized())
00170     return TD.getABITypeSize(AccessTy) < Size;
00171   return false;
00172 }
00173 
00174 //===----------------------------------------------------------------------===//
00175 // NoAA Pass
00176 //===----------------------------------------------------------------------===//
00177 
00178 namespace {
00179   /// NoAA - This class implements the -no-aa pass, which always returns "I
00180   /// don't know" for alias queries.  NoAA is unlike other alias analysis
00181   /// implementations, in that it does not chain to a previous analysis.  As
00182   /// such it doesn't follow many of the rules that other alias analyses must.
00183   ///
00184   struct VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis {
00185     static char ID; // Class identification, replacement for typeinfo
00186     NoAA() : ImmutablePass(&ID) {}
00187     explicit NoAA(void *PID) : ImmutablePass(PID) { }
00188 
00189     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
00190       AU.addRequired<TargetData>();
00191     }
00192 
00193     virtual void initializePass() {
00194       TD = &getAnalysis<TargetData>();
00195     }
00196 
00197     virtual AliasResult alias(const Value *V1, unsigned V1Size,
00198                               const Value *V2, unsigned V2Size) {
00199       return MayAlias;
00200     }
00201 
00202     virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
00203                                          std::vector<PointerAccessInfo> *Info) {
00204       return UnknownModRefBehavior;
00205     }
00206 
00207     virtual void getArgumentAccesses(Function *F, CallSite CS,
00208                                      std::vector<PointerAccessInfo> &Info) {
00209       assert(0 && "This method may not be called on this function!");
00210     }
00211 
00212     virtual void getMustAliases(Value *P, std::vector<Value*> &RetVals) { }
00213     virtual bool pointsToConstantMemory(const Value *P) { return false; }
00214     virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) {
00215       return ModRef;
00216     }
00217     virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
00218       return ModRef;
00219     }
00220     virtual bool hasNoModRefInfoForCalls() const { return true; }
00221 
00222     virtual void deleteValue(Value *V) {}
00223     virtual void copyValue(Value *From, Value *To) {}
00224   };
00225 }  // End of anonymous namespace
00226 
00227 // Register this pass...
00228 char NoAA::ID = 0;
00229 static RegisterPass<NoAA>
00230 U("no-aa", "No Alias Analysis (always returns 'may' alias)", true, true);
00231 
00232 // Declare that we implement the AliasAnalysis interface
00233 static RegisterAnalysisGroup<AliasAnalysis> V(U);
00234 
00235 ImmutablePass *llvm::createNoAAPass() { return new NoAA(); }
00236 
00237 //===----------------------------------------------------------------------===//
00238 // BasicAA Pass
00239 //===----------------------------------------------------------------------===//
00240 
00241 namespace {
00242   /// BasicAliasAnalysis - This is the default alias analysis implementation.
00243   /// Because it doesn't chain to a previous alias analysis (like -no-aa), it
00244   /// derives from the NoAA class.
00245   struct VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA {
00246     static char ID; // Class identification, replacement for typeinfo
00247     BasicAliasAnalysis() : NoAA(&ID) {}
00248     AliasResult alias(const Value *V1, unsigned V1Size,
00249                       const Value *V2, unsigned V2Size);
00250 
00251     ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
00252     ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
00253       return NoAA::getModRefInfo(CS1,CS2);
00254     }
00255 
00256     /// hasNoModRefInfoForCalls - We can provide mod/ref information against
00257     /// non-escaping allocations.
00258     virtual bool hasNoModRefInfoForCalls() const { return false; }
00259 
00260     /// pointsToConstantMemory - Chase pointers until we find a (constant
00261     /// global) or not.
00262     bool pointsToConstantMemory(const Value *P);
00263 
00264   private:
00265     // CheckGEPInstructions - Check two GEP instructions with known
00266     // must-aliasing base pointers.  This checks to see if the index expressions
00267     // preclude the pointers from aliasing...
00268     AliasResult
00269     CheckGEPInstructions(const Type* BasePtr1Ty,
00270                          Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size,
00271                          const Type *BasePtr2Ty,
00272                          Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size);
00273   };
00274 }  // End of anonymous namespace
00275 
00276 // Register this pass...
00277 char BasicAliasAnalysis::ID = 0;
00278 static RegisterPass<BasicAliasAnalysis>
00279 X("basicaa", "Basic Alias Analysis (default AA impl)", false, true);
00280 
00281 // Declare that we implement the AliasAnalysis interface
00282 static RegisterAnalysisGroup<AliasAnalysis, true> Y(X);
00283 
00284 ImmutablePass *llvm::createBasicAliasAnalysisPass() {
00285   return new BasicAliasAnalysis();
00286 }
00287 
00288 
00289 /// pointsToConstantMemory - Chase pointers until we find a (constant
00290 /// global) or not.
00291 bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) {
00292   if (const GlobalVariable *GV = 
00293         dyn_cast<GlobalVariable>(P->getUnderlyingObject()))
00294     return GV->isConstant();
00295   return false;
00296 }
00297 
00298 // getModRefInfo - Check to see if the specified callsite can clobber the
00299 // specified memory object.  Since we only look at local properties of this
00300 // function, we really can't say much about this query.  We do, however, use
00301 // simple "address taken" analysis on local objects.
00302 //
00303 AliasAnalysis::ModRefResult
00304 BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
00305   if (!isa<Constant>(P)) {
00306     const Value *Object = P->getUnderlyingObject();
00307     
00308     // If this is a tail call and P points to a stack location, we know that
00309     // the tail call cannot access or modify the local stack.
00310     // We cannot exclude byval arguments here; these belong to the caller of
00311     // the current function not to the current function, and a tail callee
00312     // may reference them.
00313     if (isa<AllocaInst>(Object))
00314       if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
00315         if (CI->isTailCall())
00316           return NoModRef;
00317     
00318     // If the pointer is to a locally allocated object that does not escape,
00319     // then the call can not mod/ref the pointer unless the call takes the
00320     // argument without capturing it.
00321     if (isNonEscapingLocalObject(Object)) {
00322       bool passedAsArg = false;
00323       // TODO: Eventually only check 'nocapture' arguments.
00324       for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
00325            CI != CE; ++CI)
00326         if (isa<PointerType>((*CI)->getType()) &&
00327             alias(cast<Value>(CI), ~0U, P, ~0U) != NoAlias)
00328           passedAsArg = true;
00329       
00330       if (!passedAsArg)
00331         return NoModRef;
00332     }
00333   }
00334 
00335   // The AliasAnalysis base class has some smarts, lets use them.
00336   return AliasAnalysis::getModRefInfo(CS, P, Size);
00337 }
00338 
00339 
00340 // alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such
00341 // as array references.  Note that this function is heavily tail recursive.
00342 // Hopefully we have a smart C++ compiler.  :)
00343 //
00344 AliasAnalysis::AliasResult
00345 BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size,
00346                           const Value *V2, unsigned V2Size) {
00347   // Strip off any constant expression casts if they exist
00348   if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V1))
00349     if (CE->isCast() && isa<PointerType>(CE->getOperand(0)->getType()))
00350       V1 = CE->getOperand(0);
00351   if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V2))
00352     if (CE->isCast() && isa<PointerType>(CE->getOperand(0)->getType()))
00353       V2 = CE->getOperand(0);
00354 
00355   // Are we checking for alias of the same value?
00356   if (V1 == V2) return MustAlias;
00357 
00358   if ((!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) &&
00359       V1->getType() != Type::Int64Ty && V2->getType() != Type::Int64Ty)
00360     return NoAlias;  // Scalars cannot alias each other
00361 
00362   // Strip off cast instructions...
00363   if (const BitCastInst *I = dyn_cast<BitCastInst>(V1))
00364     return alias(I->getOperand(0), V1Size, V2, V2Size);
00365   if (const BitCastInst *I = dyn_cast<BitCastInst>(V2))
00366     return alias(V1, V1Size, I->getOperand(0), V2Size);
00367 
00368   // Figure out what objects these things are pointing to if we can...
00369   const Value *O1 = V1->getUnderlyingObject();
00370   const Value *O2 = V2->getUnderlyingObject();
00371 
00372   if (O1 != O2) {
00373     // If V1/V2 point to two different objects we know that we have no alias.
00374     if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
00375       return NoAlias;
00376   
00377     // Incoming argument cannot alias locally allocated object!
00378     if ((isa<Argument>(O1) && isa<AllocationInst>(O2)) ||
00379         (isa<Argument>(O2) && isa<AllocationInst>(O1)))
00380       return NoAlias;
00381     
00382     // Most objects can't alias null.
00383     if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) ||
00384         (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2)))
00385       return NoAlias;
00386   }
00387   
00388   // If the size of one access is larger than the entire object on the other
00389   // side, then we know such behavior is undefined and can assume no alias.
00390   const TargetData &TD = getTargetData();
00391   if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, TD)) ||
00392       (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, TD)))
00393     return NoAlias;
00394   
00395   // If one pointer is the result of a call/invoke and the other is a
00396   // non-escaping local object, then we know the object couldn't escape to a
00397   // point where the call could return it.
00398   if ((isa<CallInst>(O1) || isa<InvokeInst>(O1)) &&
00399       isNonEscapingLocalObject(O2))
00400     return NoAlias;
00401   if ((isa<CallInst>(O2) || isa<InvokeInst>(O2)) &&
00402       isNonEscapingLocalObject(O1))
00403     return NoAlias;
00404   
00405   // If we have two gep instructions with must-alias'ing base pointers, figure
00406   // out if the indexes to the GEP tell us anything about the derived pointer.
00407   // Note that we also handle chains of getelementptr instructions as well as
00408   // constant expression getelementptrs here.
00409   //
00410   if (isGEP(V1) && isGEP(V2)) {
00411     // Drill down into the first non-gep value, to test for must-aliasing of
00412     // the base pointers.
00413     const User *G = cast<User>(V1);
00414     while (isGEP(G->getOperand(0)) &&
00415            G->getOperand(1) ==
00416            Constant::getNullValue(G->getOperand(1)->getType()))
00417       G = cast<User>(G->getOperand(0));
00418     const Value *BasePtr1 = G->getOperand(0);
00419 
00420     G = cast<User>(V2);
00421     while (isGEP(G->getOperand(0)) &&
00422            G->getOperand(1) ==
00423            Constant::getNullValue(G->getOperand(1)->getType()))
00424       G = cast<User>(G->getOperand(0));
00425     const Value *BasePtr2 = G->getOperand(0);
00426 
00427     // Do the base pointers alias?
00428     AliasResult BaseAlias = alias(BasePtr1, ~0U, BasePtr2, ~0U);
00429     if (BaseAlias == NoAlias) return NoAlias;
00430     if (BaseAlias == MustAlias) {
00431       // If the base pointers alias each other exactly, check to see if we can
00432       // figure out anything about the resultant pointers, to try to prove
00433       // non-aliasing.
00434 
00435       // Collect all of the chained GEP operands together into one simple place
00436       SmallVector<Value*, 16> GEP1Ops, GEP2Ops;
00437       BasePtr1 = GetGEPOperands(V1, GEP1Ops);
00438       BasePtr2 = GetGEPOperands(V2, GEP2Ops);
00439 
00440       // If GetGEPOperands were able to fold to the same must-aliased pointer,
00441       // do the comparison.
00442       if (BasePtr1 == BasePtr2) {
00443         AliasResult GAlias =
00444           CheckGEPInstructions(BasePtr1->getType(),
00445                                &GEP1Ops[0], GEP1Ops.size(), V1Size,
00446                                BasePtr2->getType(),
00447                                &GEP2Ops[0], GEP2Ops.size(), V2Size);
00448         if (GAlias != MayAlias)
00449           return GAlias;
00450       }
00451     }
00452   }
00453 
00454   // Check to see if these two pointers are related by a getelementptr
00455   // instruction.  If one pointer is a GEP with a non-zero index of the other
00456   // pointer, we know they cannot alias.
00457   //
00458   if (isGEP(V2)) {
00459     std::swap(V1, V2);
00460     std::swap(V1Size, V2Size);
00461   }
00462 
00463   if (V1Size != ~0U && V2Size != ~0U)
00464     if (isGEP(V1)) {
00465       SmallVector<Value*, 16> GEPOperands;
00466       const Value *BasePtr = GetGEPOperands(V1, GEPOperands);
00467 
00468       AliasResult R = alias(BasePtr, V1Size, V2, V2Size);
00469       if (R == MustAlias) {
00470         // If there is at least one non-zero constant index, we know they cannot
00471         // alias.
00472         bool ConstantFound = false;
00473         bool AllZerosFound = true;
00474         for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i)
00475           if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) {
00476             if (!C->isNullValue()) {
00477               ConstantFound = true;
00478               AllZerosFound = false;
00479               break;
00480             }
00481           } else {
00482             AllZerosFound = false;
00483           }
00484 
00485         // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases
00486         // the ptr, the end result is a must alias also.
00487         if (AllZerosFound)
00488           return MustAlias;
00489 
00490         if (ConstantFound) {
00491           if (V2Size <= 1 && V1Size <= 1)  // Just pointer check?
00492             return NoAlias;
00493 
00494           // Otherwise we have to check to see that the distance is more than
00495           // the size of the argument... build an index vector that is equal to
00496           // the arguments provided, except substitute 0's for any variable
00497           // indexes we find...
00498           if (cast<PointerType>(
00499                 BasePtr->getType())->getElementType()->isSized()) {
00500             for (unsigned i = 0; i != GEPOperands.size(); ++i)
00501               if (!isa<ConstantInt>(GEPOperands[i]))
00502                 GEPOperands[i] =
00503                   Constant::getNullValue(GEPOperands[i]->getType());
00504             int64_t Offset =
00505               getTargetData().getIndexedOffset(BasePtr->getType(),
00506                                                &GEPOperands[0],
00507                                                GEPOperands.size());
00508 
00509             if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size)
00510               return NoAlias;
00511           }
00512         }
00513       }
00514     }
00515 
00516   return MayAlias;
00517 }
00518 
00519 // This function is used to determin if the indices of two GEP instructions are
00520 // equal. V1 and V2 are the indices.
00521 static bool IndexOperandsEqual(Value *V1, Value *V2) {
00522   if (V1->getType() == V2->getType())
00523     return V1 == V2;
00524   if (Constant *C1 = dyn_cast<Constant>(V1))
00525     if (Constant *C2 = dyn_cast<Constant>(V2)) {
00526       // Sign extend the constants to long types, if necessary
00527       if (C1->getType() != Type::Int64Ty)
00528         C1 = ConstantExpr::getSExt(C1, Type::Int64Ty);
00529       if (C2->getType() != Type::Int64Ty) 
00530         C2 = ConstantExpr::getSExt(C2, Type::Int64Ty);
00531       return C1 == C2;
00532     }
00533   return false;
00534 }
00535 
00536 /// CheckGEPInstructions - Check two GEP instructions with known must-aliasing
00537 /// base pointers.  This checks to see if the index expressions preclude the
00538 /// pointers from aliasing...
00539 AliasAnalysis::AliasResult 
00540 BasicAliasAnalysis::CheckGEPInstructions(
00541   const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S,
00542   const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) {
00543   // We currently can't handle the case when the base pointers have different
00544   // primitive types.  Since this is uncommon anyway, we are happy being
00545   // extremely conservative.
00546   if (BasePtr1Ty != BasePtr2Ty)
00547     return MayAlias;
00548 
00549   const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty);
00550 
00551   // Find the (possibly empty) initial sequence of equal values... which are not
00552   // necessarily constants.
00553   unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops;
00554   unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands);
00555   unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands);
00556   unsigned UnequalOper = 0;
00557   while (UnequalOper != MinOperands &&
00558          IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) {
00559     // Advance through the type as we go...
00560     ++UnequalOper;
00561     if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
00562       BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]);
00563     else {
00564       // If all operands equal each other, then the derived pointers must
00565       // alias each other...
00566       BasePtr1Ty = 0;
00567       assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands &&
00568              "Ran out of type nesting, but not out of operands?");
00569       return MustAlias;
00570     }
00571   }
00572 
00573   // If we have seen all constant operands, and run out of indexes on one of the
00574   // getelementptrs, check to see if the tail of the leftover one is all zeros.
00575   // If so, return mustalias.
00576   if (UnequalOper == MinOperands) {
00577     if (NumGEP1Ops < NumGEP2Ops) {
00578       std::swap(GEP1Ops, GEP2Ops);
00579       std::swap(NumGEP1Ops, NumGEP2Ops);
00580     }
00581 
00582     bool AllAreZeros = true;
00583     for (unsigned i = UnequalOper; i != MaxOperands; ++i)
00584       if (!isa<Constant>(GEP1Ops[i]) ||
00585           !cast<Constant>(GEP1Ops[i])->isNullValue()) {
00586         AllAreZeros = false;
00587         break;
00588       }
00589     if (AllAreZeros) return MustAlias;
00590   }
00591 
00592 
00593   // So now we know that the indexes derived from the base pointers,
00594   // which are known to alias, are different.  We can still determine a
00595   // no-alias result if there are differing constant pairs in the index
00596   // chain.  For example:
00597   //        A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S))
00598   //
00599   // We have to be careful here about array accesses.  In particular, consider:
00600   //        A[1][0] vs A[0][i]
00601   // In this case, we don't *know* that the array will be accessed in bounds:
00602   // the index could even be negative.  Because of this, we have to
00603   // conservatively *give up* and return may alias.  We disregard differing
00604   // array subscripts that are followed by a variable index without going
00605   // through a struct.
00606   //
00607   unsigned SizeMax = std::max(G1S, G2S);
00608   if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work.
00609 
00610   // Scan for the first operand that is constant and unequal in the
00611   // two getelementptrs...
00612   unsigned FirstConstantOper = UnequalOper;
00613   for (; FirstConstantOper != MinOperands; ++FirstConstantOper) {
00614     const Value *G1Oper = GEP1Ops[FirstConstantOper];
00615     const Value *G2Oper = GEP2Ops[FirstConstantOper];
00616 
00617     if (G1Oper != G2Oper)   // Found non-equal constant indexes...
00618       if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper)))
00619         if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){
00620           if (G1OC->getType() != G2OC->getType()) {
00621             // Sign extend both operands to long.
00622             if (G1OC->getType() != Type::Int64Ty)
00623               G1OC = ConstantExpr::getSExt(G1OC, Type::Int64Ty);
00624             if (G2OC->getType() != Type::Int64Ty) 
00625               G2OC = ConstantExpr::getSExt(G2OC, Type::Int64Ty);
00626             GEP1Ops[FirstConstantOper] = G1OC;
00627             GEP2Ops[FirstConstantOper] = G2OC;
00628           }
00629           
00630           if (G1OC != G2OC) {
00631             // Handle the "be careful" case above: if this is an array/vector
00632             // subscript, scan for a subsequent variable array index.
00633             if (isa<SequentialType>(BasePtr1Ty))  {
00634               const Type *NextTy =
00635                 cast<SequentialType>(BasePtr1Ty)->getElementType();
00636               bool isBadCase = false;
00637               
00638               for (unsigned Idx = FirstConstantOper+1;
00639                    Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) {
00640                 const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx];
00641                 if (!isa<Constant>(V1) || !isa<Constant>(V2)) {
00642                   isBadCase = true;
00643                   break;
00644                 }
00645                 NextTy = cast<SequentialType>(NextTy)->getElementType();
00646               }
00647               
00648               if (isBadCase) G1OC = 0;
00649             }
00650 
00651             // Make sure they are comparable (ie, not constant expressions), and
00652             // make sure the GEP with the smaller leading constant is GEP1.
00653             if (G1OC) {
00654               Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT, 
00655                                                         G1OC, G2OC);
00656               if (ConstantInt *CV = dyn_cast<ConstantInt>(Compare)) {
00657                 if (CV->getZExtValue()) {  // If they are comparable and G2 > G1
00658                   std::swap(GEP1Ops, GEP2Ops);  // Make GEP1 < GEP2
00659                   std::swap(NumGEP1Ops, NumGEP2Ops);
00660                 }
00661                 break;
00662               }
00663             }
00664           }
00665         }
00666     BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper);
00667   }
00668 
00669   // No shared constant operands, and we ran out of common operands.  At this
00670   // point, the GEP instructions have run through all of their operands, and we
00671   // haven't found evidence that there are any deltas between the GEP's.
00672   // However, one GEP may have more operands than the other.  If this is the
00673   // case, there may still be hope.  Check this now.
00674   if (FirstConstantOper == MinOperands) {
00675     // Make GEP1Ops be the longer one if there is a longer one.
00676     if (NumGEP1Ops < NumGEP2Ops) {
00677       std::swap(GEP1Ops, GEP2Ops);
00678       std::swap(NumGEP1Ops, NumGEP2Ops);
00679     }
00680 
00681     // Is there anything to check?
00682     if (NumGEP1Ops > MinOperands) {
00683       for (unsigned i = FirstConstantOper; i != MaxOperands; ++i)
00684         if (isa<ConstantInt>(GEP1Ops[i]) && 
00685             !cast<ConstantInt>(GEP1Ops[i])->isZero()) {
00686           // Yup, there's a constant in the tail.  Set all variables to
00687           // constants in the GEP instruction to make it suitable for
00688           // TargetData::getIndexedOffset.
00689           for (i = 0; i != MaxOperands; ++i)
00690             if (!isa<ConstantInt>(GEP1Ops[i]))
00691               GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType());
00692           // Okay, now get the offset.  This is the relative offset for the full
00693           // instruction.
00694           const TargetData &TD = getTargetData();
00695           int64_t Offset1 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops,
00696                                                 NumGEP1Ops);
00697 
00698           // Now check without any constants at the end.
00699           int64_t Offset2 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops,
00700                                                 MinOperands);
00701 
00702           // Make sure we compare the absolute difference.
00703           if (Offset1 > Offset2)
00704             std::swap(Offset1, Offset2);
00705 
00706           // If the tail provided a bit enough offset, return noalias!
00707           if ((uint64_t)(Offset2-Offset1) >= SizeMax)
00708             return NoAlias;
00709           // Otherwise break - we don't look for another constant in the tail.
00710           break;
00711         }
00712     }
00713 
00714     // Couldn't find anything useful.
00715     return MayAlias;
00716   }
00717 
00718   // If there are non-equal constants arguments, then we can figure
00719   // out a minimum known delta between the two index expressions... at
00720   // this point we know that the first constant index of GEP1 is less
00721   // than the first constant index of GEP2.
00722 
00723   // Advance BasePtr[12]Ty over this first differing constant operand.
00724   BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)->
00725       getTypeAtIndex(GEP2Ops[FirstConstantOper]);
00726   BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->
00727       getTypeAtIndex(GEP1Ops[FirstConstantOper]);
00728 
00729   // We are going to be using TargetData::getIndexedOffset to determine the
00730   // offset that each of the GEP's is reaching.  To do this, we have to convert
00731   // all variable references to constant references.  To do this, we convert the
00732   // initial sequence of array subscripts into constant zeros to start with.
00733   const Type *ZeroIdxTy = GEPPointerTy;
00734   for (unsigned i = 0; i != FirstConstantOper; ++i) {
00735     if (!isa<StructType>(ZeroIdxTy))
00736       GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Type::Int32Ty);
00737 
00738     if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy))
00739       ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]);
00740   }
00741 
00742   // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok
00743 
00744   // Loop over the rest of the operands...
00745   for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) {
00746     const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0;
00747     const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0;
00748     // If they are equal, use a zero index...
00749     if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) {
00750       if (!isa<ConstantInt>(Op1))
00751         GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType());
00752       // Otherwise, just keep the constants we have.
00753     } else {
00754       if (Op1) {
00755         if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
00756           // If this is an array index, make sure the array element is in range.
00757           if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) {
00758             if (Op1C->getZExtValue() >= AT->getNumElements())
00759               return MayAlias;  // Be conservative with out-of-range accesses
00760           } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) {
00761             if (Op1C->getZExtValue() >= VT->getNumElements())
00762               return MayAlias;  // Be conservative with out-of-range accesses
00763           }
00764           
00765         } else {
00766           // GEP1 is known to produce a value less than GEP2.  To be
00767           // conservatively correct, we must assume the largest possible
00768           // constant is used in this position.  This cannot be the initial
00769           // index to the GEP instructions (because we know we have at least one
00770           // element before this one with the different constant arguments), so
00771           // we know that the current index must be into either a struct or
00772           // array.  Because we know it's not constant, this cannot be a
00773           // structure index.  Because of this, we can calculate the maximum
00774           // value possible.
00775           //
00776           if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
00777             GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,AT->getNumElements()-1);
00778           else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty))
00779             GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,VT->getNumElements()-1);
00780         }
00781       }
00782 
00783       if (Op2) {
00784         if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) {
00785           // If this is an array index, make sure the array element is in range.
00786           if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) {
00787             if (Op2C->getZExtValue() >= AT->getNumElements())
00788               return MayAlias;  // Be conservative with out-of-range accesses
00789           } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) {
00790             if (Op2C->getZExtValue() >= VT->getNumElements())
00791               return MayAlias;  // Be conservative with out-of-range accesses
00792           }
00793         } else {  // Conservatively assume the minimum value for this index
00794           GEP2Ops[i] = Constant::getNullValue(Op2->getType());
00795         }
00796       }
00797     }
00798 
00799     if (BasePtr1Ty && Op1) {
00800       if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
00801         BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]);
00802       else
00803         BasePtr1Ty = 0;
00804     }
00805 
00806     if (BasePtr2Ty && Op2) {
00807       if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty))
00808         BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]);
00809       else
00810         BasePtr2Ty = 0;
00811     }
00812   }
00813 
00814   if (GEPPointerTy->getElementType()->isSized()) {
00815     int64_t Offset1 =
00816       getTargetData().getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops);
00817     int64_t Offset2 = 
00818       getTargetData().getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops);
00819     assert(Offset1 != Offset2 &&
00820            "There is at least one different constant here!");
00821     
00822     // Make sure we compare the absolute difference.
00823     if (Offset1 > Offset2)
00824       std::swap(Offset1, Offset2);
00825     
00826     if ((uint64_t)(Offset2-Offset1) >= SizeMax) {
00827       //cerr << "Determined that these two GEP's don't alias ["
00828       //     << SizeMax << " bytes]: \n" << *GEP1 << *GEP2;
00829       return NoAlias;
00830     }
00831   }
00832   return MayAlias;
00833 }
00834 
00835 // Make sure that anything that uses AliasAnalysis pulls in this file...
00836 DEFINING_FILE_FOR(BasicAliasAnalysis)



This web site is hosted by the Computer Science Department at the University of Illinois at Urbana-Champaign.