Home | History | Annotate | Download | only in Analysis
      1 //===- BasicAliasAnalysis.cpp - Stateless 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 primary stateless implementation of the
     11 // Alias Analysis interface that implements identities (two different
     12 // globals cannot alias, etc), but does no stateful 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/GlobalAlias.h"
     22 #include "llvm/GlobalVariable.h"
     23 #include "llvm/Instructions.h"
     24 #include "llvm/IntrinsicInst.h"
     25 #include "llvm/LLVMContext.h"
     26 #include "llvm/Operator.h"
     27 #include "llvm/Pass.h"
     28 #include "llvm/Analysis/CaptureTracking.h"
     29 #include "llvm/Analysis/MemoryBuiltins.h"
     30 #include "llvm/Analysis/InstructionSimplify.h"
     31 #include "llvm/Analysis/ValueTracking.h"
     32 #include "llvm/Target/TargetData.h"
     33 #include "llvm/Target/TargetLibraryInfo.h"
     34 #include "llvm/ADT/SmallPtrSet.h"
     35 #include "llvm/ADT/SmallVector.h"
     36 #include "llvm/Support/ErrorHandling.h"
     37 #include "llvm/Support/GetElementPtrTypeIterator.h"
     38 #include <algorithm>
     39 using namespace llvm;
     40 
     41 //===----------------------------------------------------------------------===//
     42 // Useful predicates
     43 //===----------------------------------------------------------------------===//
     44 
     45 /// isKnownNonNull - Return true if we know that the specified value is never
     46 /// null.
     47 static bool isKnownNonNull(const Value *V) {
     48   // Alloca never returns null, malloc might.
     49   if (isa<AllocaInst>(V)) return true;
     50 
     51   // A byval argument is never null.
     52   if (const Argument *A = dyn_cast<Argument>(V))
     53     return A->hasByValAttr();
     54 
     55   // Global values are not null unless extern weak.
     56   if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
     57     return !GV->hasExternalWeakLinkage();
     58   return false;
     59 }
     60 
     61 /// isNonEscapingLocalObject - Return true if the pointer is to a function-local
     62 /// object that never escapes from the function.
     63 static bool isNonEscapingLocalObject(const Value *V) {
     64   // If this is a local allocation, check to see if it escapes.
     65   if (isa<AllocaInst>(V) || isNoAliasCall(V))
     66     // Set StoreCaptures to True so that we can assume in our callers that the
     67     // pointer is not the result of a load instruction. Currently
     68     // PointerMayBeCaptured doesn't have any special analysis for the
     69     // StoreCaptures=false case; if it did, our callers could be refined to be
     70     // more precise.
     71     return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
     72 
     73   // If this is an argument that corresponds to a byval or noalias argument,
     74   // then it has not escaped before entering the function.  Check if it escapes
     75   // inside the function.
     76   if (const Argument *A = dyn_cast<Argument>(V))
     77     if (A->hasByValAttr() || A->hasNoAliasAttr()) {
     78       // Don't bother analyzing arguments already known not to escape.
     79       if (A->hasNoCaptureAttr())
     80         return true;
     81       return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
     82     }
     83   return false;
     84 }
     85 
     86 /// isEscapeSource - Return true if the pointer is one which would have
     87 /// been considered an escape by isNonEscapingLocalObject.
     88 static bool isEscapeSource(const Value *V) {
     89   if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
     90     return true;
     91 
     92   // The load case works because isNonEscapingLocalObject considers all
     93   // stores to be escapes (it passes true for the StoreCaptures argument
     94   // to PointerMayBeCaptured).
     95   if (isa<LoadInst>(V))
     96     return true;
     97 
     98   return false;
     99 }
    100 
    101 /// getObjectSize - Return the size of the object specified by V, or
    102 /// UnknownSize if unknown.
    103 static uint64_t getObjectSize(const Value *V, const TargetData &TD) {
    104   Type *AccessTy;
    105   if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
    106     if (!GV->hasDefinitiveInitializer())
    107       return AliasAnalysis::UnknownSize;
    108     AccessTy = GV->getType()->getElementType();
    109   } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
    110     if (!AI->isArrayAllocation())
    111       AccessTy = AI->getType()->getElementType();
    112     else
    113       return AliasAnalysis::UnknownSize;
    114   } else if (const CallInst* CI = extractMallocCall(V)) {
    115     if (!isArrayMalloc(V, &TD))
    116       // The size is the argument to the malloc call.
    117       if (const ConstantInt* C = dyn_cast<ConstantInt>(CI->getArgOperand(0)))
    118         return C->getZExtValue();
    119     return AliasAnalysis::UnknownSize;
    120   } else if (const Argument *A = dyn_cast<Argument>(V)) {
    121     if (A->hasByValAttr())
    122       AccessTy = cast<PointerType>(A->getType())->getElementType();
    123     else
    124       return AliasAnalysis::UnknownSize;
    125   } else {
    126     return AliasAnalysis::UnknownSize;
    127   }
    128 
    129   if (AccessTy->isSized())
    130     return TD.getTypeAllocSize(AccessTy);
    131   return AliasAnalysis::UnknownSize;
    132 }
    133 
    134 /// isObjectSmallerThan - Return true if we can prove that the object specified
    135 /// by V is smaller than Size.
    136 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
    137                                 const TargetData &TD) {
    138   uint64_t ObjectSize = getObjectSize(V, TD);
    139   return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize < Size;
    140 }
    141 
    142 /// isObjectSize - Return true if we can prove that the object specified
    143 /// by V has size Size.
    144 static bool isObjectSize(const Value *V, uint64_t Size,
    145                          const TargetData &TD) {
    146   uint64_t ObjectSize = getObjectSize(V, TD);
    147   return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize == Size;
    148 }
    149 
    150 //===----------------------------------------------------------------------===//
    151 // GetElementPtr Instruction Decomposition and Analysis
    152 //===----------------------------------------------------------------------===//
    153 
    154 namespace {
    155   enum ExtensionKind {
    156     EK_NotExtended,
    157     EK_SignExt,
    158     EK_ZeroExt
    159   };
    160 
    161   struct VariableGEPIndex {
    162     const Value *V;
    163     ExtensionKind Extension;
    164     int64_t Scale;
    165   };
    166 }
    167 
    168 
    169 /// GetLinearExpression - Analyze the specified value as a linear expression:
    170 /// "A*V + B", where A and B are constant integers.  Return the scale and offset
    171 /// values as APInts and return V as a Value*, and return whether we looked
    172 /// through any sign or zero extends.  The incoming Value is known to have
    173 /// IntegerType and it may already be sign or zero extended.
    174 ///
    175 /// Note that this looks through extends, so the high bits may not be
    176 /// represented in the result.
    177 static Value *GetLinearExpression(Value *V, APInt &Scale, APInt &Offset,
    178                                   ExtensionKind &Extension,
    179                                   const TargetData &TD, unsigned Depth) {
    180   assert(V->getType()->isIntegerTy() && "Not an integer value");
    181 
    182   // Limit our recursion depth.
    183   if (Depth == 6) {
    184     Scale = 1;
    185     Offset = 0;
    186     return V;
    187   }
    188 
    189   if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
    190     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
    191       switch (BOp->getOpcode()) {
    192       default: break;
    193       case Instruction::Or:
    194         // X|C == X+C if all the bits in C are unset in X.  Otherwise we can't
    195         // analyze it.
    196         if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), &TD))
    197           break;
    198         // FALL THROUGH.
    199       case Instruction::Add:
    200         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
    201                                 TD, Depth+1);
    202         Offset += RHSC->getValue();
    203         return V;
    204       case Instruction::Mul:
    205         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
    206                                 TD, Depth+1);
    207         Offset *= RHSC->getValue();
    208         Scale *= RHSC->getValue();
    209         return V;
    210       case Instruction::Shl:
    211         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
    212                                 TD, Depth+1);
    213         Offset <<= RHSC->getValue().getLimitedValue();
    214         Scale <<= RHSC->getValue().getLimitedValue();
    215         return V;
    216       }
    217     }
    218   }
    219 
    220   // Since GEP indices are sign extended anyway, we don't care about the high
    221   // bits of a sign or zero extended value - just scales and offsets.  The
    222   // extensions have to be consistent though.
    223   if ((isa<SExtInst>(V) && Extension != EK_ZeroExt) ||
    224       (isa<ZExtInst>(V) && Extension != EK_SignExt)) {
    225     Value *CastOp = cast<CastInst>(V)->getOperand(0);
    226     unsigned OldWidth = Scale.getBitWidth();
    227     unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
    228     Scale = Scale.trunc(SmallWidth);
    229     Offset = Offset.trunc(SmallWidth);
    230     Extension = isa<SExtInst>(V) ? EK_SignExt : EK_ZeroExt;
    231 
    232     Value *Result = GetLinearExpression(CastOp, Scale, Offset, Extension,
    233                                         TD, Depth+1);
    234     Scale = Scale.zext(OldWidth);
    235     Offset = Offset.zext(OldWidth);
    236 
    237     return Result;
    238   }
    239 
    240   Scale = 1;
    241   Offset = 0;
    242   return V;
    243 }
    244 
    245 /// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it
    246 /// into a base pointer with a constant offset and a number of scaled symbolic
    247 /// offsets.
    248 ///
    249 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale in
    250 /// the VarIndices vector) are Value*'s that are known to be scaled by the
    251 /// specified amount, but which may have other unrepresented high bits. As such,
    252 /// the gep cannot necessarily be reconstructed from its decomposed form.
    253 ///
    254 /// When TargetData is around, this function is capable of analyzing everything
    255 /// that GetUnderlyingObject can look through.  When not, it just looks
    256 /// through pointer casts.
    257 ///
    258 static const Value *
    259 DecomposeGEPExpression(const Value *V, int64_t &BaseOffs,
    260                        SmallVectorImpl<VariableGEPIndex> &VarIndices,
    261                        const TargetData *TD) {
    262   // Limit recursion depth to limit compile time in crazy cases.
    263   unsigned MaxLookup = 6;
    264 
    265   BaseOffs = 0;
    266   do {
    267     // See if this is a bitcast or GEP.
    268     const Operator *Op = dyn_cast<Operator>(V);
    269     if (Op == 0) {
    270       // The only non-operator case we can handle are GlobalAliases.
    271       if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
    272         if (!GA->mayBeOverridden()) {
    273           V = GA->getAliasee();
    274           continue;
    275         }
    276       }
    277       return V;
    278     }
    279 
    280     if (Op->getOpcode() == Instruction::BitCast) {
    281       V = Op->getOperand(0);
    282       continue;
    283     }
    284 
    285     const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
    286     if (GEPOp == 0) {
    287       // If it's not a GEP, hand it off to SimplifyInstruction to see if it
    288       // can come up with something. This matches what GetUnderlyingObject does.
    289       if (const Instruction *I = dyn_cast<Instruction>(V))
    290         // TODO: Get a DominatorTree and use it here.
    291         if (const Value *Simplified =
    292               SimplifyInstruction(const_cast<Instruction *>(I), TD)) {
    293           V = Simplified;
    294           continue;
    295         }
    296 
    297       return V;
    298     }
    299 
    300     // Don't attempt to analyze GEPs over unsized objects.
    301     if (!cast<PointerType>(GEPOp->getOperand(0)->getType())
    302         ->getElementType()->isSized())
    303       return V;
    304 
    305     // If we are lacking TargetData information, we can't compute the offets of
    306     // elements computed by GEPs.  However, we can handle bitcast equivalent
    307     // GEPs.
    308     if (TD == 0) {
    309       if (!GEPOp->hasAllZeroIndices())
    310         return V;
    311       V = GEPOp->getOperand(0);
    312       continue;
    313     }
    314 
    315     // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
    316     gep_type_iterator GTI = gep_type_begin(GEPOp);
    317     for (User::const_op_iterator I = GEPOp->op_begin()+1,
    318          E = GEPOp->op_end(); I != E; ++I) {
    319       Value *Index = *I;
    320       // Compute the (potentially symbolic) offset in bytes for this index.
    321       if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
    322         // For a struct, add the member offset.
    323         unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
    324         if (FieldNo == 0) continue;
    325 
    326         BaseOffs += TD->getStructLayout(STy)->getElementOffset(FieldNo);
    327         continue;
    328       }
    329 
    330       // For an array/pointer, add the element offset, explicitly scaled.
    331       if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
    332         if (CIdx->isZero()) continue;
    333         BaseOffs += TD->getTypeAllocSize(*GTI)*CIdx->getSExtValue();
    334         continue;
    335       }
    336 
    337       uint64_t Scale = TD->getTypeAllocSize(*GTI);
    338       ExtensionKind Extension = EK_NotExtended;
    339 
    340       // If the integer type is smaller than the pointer size, it is implicitly
    341       // sign extended to pointer size.
    342       unsigned Width = cast<IntegerType>(Index->getType())->getBitWidth();
    343       if (TD->getPointerSizeInBits() > Width)
    344         Extension = EK_SignExt;
    345 
    346       // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
    347       APInt IndexScale(Width, 0), IndexOffset(Width, 0);
    348       Index = GetLinearExpression(Index, IndexScale, IndexOffset, Extension,
    349                                   *TD, 0);
    350 
    351       // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
    352       // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
    353       BaseOffs += IndexOffset.getSExtValue()*Scale;
    354       Scale *= IndexScale.getSExtValue();
    355 
    356 
    357       // If we already had an occurrence of this index variable, merge this
    358       // scale into it.  For example, we want to handle:
    359       //   A[x][x] -> x*16 + x*4 -> x*20
    360       // This also ensures that 'x' only appears in the index list once.
    361       for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
    362         if (VarIndices[i].V == Index &&
    363             VarIndices[i].Extension == Extension) {
    364           Scale += VarIndices[i].Scale;
    365           VarIndices.erase(VarIndices.begin()+i);
    366           break;
    367         }
    368       }
    369 
    370       // Make sure that we have a scale that makes sense for this target's
    371       // pointer size.
    372       if (unsigned ShiftBits = 64-TD->getPointerSizeInBits()) {
    373         Scale <<= ShiftBits;
    374         Scale = (int64_t)Scale >> ShiftBits;
    375       }
    376 
    377       if (Scale) {
    378         VariableGEPIndex Entry = {Index, Extension,
    379                                   static_cast<int64_t>(Scale)};
    380         VarIndices.push_back(Entry);
    381       }
    382     }
    383 
    384     // Analyze the base pointer next.
    385     V = GEPOp->getOperand(0);
    386   } while (--MaxLookup);
    387 
    388   // If the chain of expressions is too deep, just return early.
    389   return V;
    390 }
    391 
    392 /// GetIndexDifference - Dest and Src are the variable indices from two
    393 /// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base
    394 /// pointers.  Subtract the GEP2 indices from GEP1 to find the symbolic
    395 /// difference between the two pointers.
    396 static void GetIndexDifference(SmallVectorImpl<VariableGEPIndex> &Dest,
    397                                const SmallVectorImpl<VariableGEPIndex> &Src) {
    398   if (Src.empty()) return;
    399 
    400   for (unsigned i = 0, e = Src.size(); i != e; ++i) {
    401     const Value *V = Src[i].V;
    402     ExtensionKind Extension = Src[i].Extension;
    403     int64_t Scale = Src[i].Scale;
    404 
    405     // Find V in Dest.  This is N^2, but pointer indices almost never have more
    406     // than a few variable indexes.
    407     for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
    408       if (Dest[j].V != V || Dest[j].Extension != Extension) continue;
    409 
    410       // If we found it, subtract off Scale V's from the entry in Dest.  If it
    411       // goes to zero, remove the entry.
    412       if (Dest[j].Scale != Scale)
    413         Dest[j].Scale -= Scale;
    414       else
    415         Dest.erase(Dest.begin()+j);
    416       Scale = 0;
    417       break;
    418     }
    419 
    420     // If we didn't consume this entry, add it to the end of the Dest list.
    421     if (Scale) {
    422       VariableGEPIndex Entry = { V, Extension, -Scale };
    423       Dest.push_back(Entry);
    424     }
    425   }
    426 }
    427 
    428 //===----------------------------------------------------------------------===//
    429 // BasicAliasAnalysis Pass
    430 //===----------------------------------------------------------------------===//
    431 
    432 #ifndef NDEBUG
    433 static const Function *getParent(const Value *V) {
    434   if (const Instruction *inst = dyn_cast<Instruction>(V))
    435     return inst->getParent()->getParent();
    436 
    437   if (const Argument *arg = dyn_cast<Argument>(V))
    438     return arg->getParent();
    439 
    440   return NULL;
    441 }
    442 
    443 static bool notDifferentParent(const Value *O1, const Value *O2) {
    444 
    445   const Function *F1 = getParent(O1);
    446   const Function *F2 = getParent(O2);
    447 
    448   return !F1 || !F2 || F1 == F2;
    449 }
    450 #endif
    451 
    452 namespace {
    453   /// BasicAliasAnalysis - This is the primary alias analysis implementation.
    454   struct BasicAliasAnalysis : public ImmutablePass, public AliasAnalysis {
    455     static char ID; // Class identification, replacement for typeinfo
    456     BasicAliasAnalysis() : ImmutablePass(ID),
    457                            // AliasCache rarely has more than 1 or 2 elements,
    458                            // so start it off fairly small so that clear()
    459                            // doesn't have to tromp through 64 (the default)
    460                            // elements on each alias query. This really wants
    461                            // something like a SmallDenseMap.
    462                            AliasCache(8) {
    463       initializeBasicAliasAnalysisPass(*PassRegistry::getPassRegistry());
    464     }
    465 
    466     virtual void initializePass() {
    467       InitializeAliasAnalysis(this);
    468     }
    469 
    470     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
    471       AU.addRequired<AliasAnalysis>();
    472       AU.addRequired<TargetLibraryInfo>();
    473     }
    474 
    475     virtual AliasResult alias(const Location &LocA,
    476                               const Location &LocB) {
    477       assert(AliasCache.empty() && "AliasCache must be cleared after use!");
    478       assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
    479              "BasicAliasAnalysis doesn't support interprocedural queries.");
    480       AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.TBAATag,
    481                                      LocB.Ptr, LocB.Size, LocB.TBAATag);
    482       AliasCache.clear();
    483       return Alias;
    484     }
    485 
    486     virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
    487                                        const Location &Loc);
    488 
    489     virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
    490                                        ImmutableCallSite CS2) {
    491       // The AliasAnalysis base class has some smarts, lets use them.
    492       return AliasAnalysis::getModRefInfo(CS1, CS2);
    493     }
    494 
    495     /// pointsToConstantMemory - Chase pointers until we find a (constant
    496     /// global) or not.
    497     virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
    498 
    499     /// getModRefBehavior - Return the behavior when calling the given
    500     /// call site.
    501     virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
    502 
    503     /// getModRefBehavior - Return the behavior when calling the given function.
    504     /// For use when the call site is not known.
    505     virtual ModRefBehavior getModRefBehavior(const Function *F);
    506 
    507     /// getAdjustedAnalysisPointer - This method is used when a pass implements
    508     /// an analysis interface through multiple inheritance.  If needed, it
    509     /// should override this to adjust the this pointer as needed for the
    510     /// specified pass info.
    511     virtual void *getAdjustedAnalysisPointer(const void *ID) {
    512       if (ID == &AliasAnalysis::ID)
    513         return (AliasAnalysis*)this;
    514       return this;
    515     }
    516 
    517   private:
    518     // AliasCache - Track alias queries to guard against recursion.
    519     typedef std::pair<Location, Location> LocPair;
    520     typedef DenseMap<LocPair, AliasResult> AliasCacheTy;
    521     AliasCacheTy AliasCache;
    522 
    523     // Visited - Track instructions visited by pointsToConstantMemory.
    524     SmallPtrSet<const Value*, 16> Visited;
    525 
    526     // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP
    527     // instruction against another.
    528     AliasResult aliasGEP(const GEPOperator *V1, uint64_t V1Size,
    529                          const Value *V2, uint64_t V2Size,
    530                          const MDNode *V2TBAAInfo,
    531                          const Value *UnderlyingV1, const Value *UnderlyingV2);
    532 
    533     // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI
    534     // instruction against another.
    535     AliasResult aliasPHI(const PHINode *PN, uint64_t PNSize,
    536                          const MDNode *PNTBAAInfo,
    537                          const Value *V2, uint64_t V2Size,
    538                          const MDNode *V2TBAAInfo);
    539 
    540     /// aliasSelect - Disambiguate a Select instruction against another value.
    541     AliasResult aliasSelect(const SelectInst *SI, uint64_t SISize,
    542                             const MDNode *SITBAAInfo,
    543                             const Value *V2, uint64_t V2Size,
    544                             const MDNode *V2TBAAInfo);
    545 
    546     AliasResult aliasCheck(const Value *V1, uint64_t V1Size,
    547                            const MDNode *V1TBAATag,
    548                            const Value *V2, uint64_t V2Size,
    549                            const MDNode *V2TBAATag);
    550   };
    551 }  // End of anonymous namespace
    552 
    553 // Register this pass...
    554 char BasicAliasAnalysis::ID = 0;
    555 INITIALIZE_AG_PASS_BEGIN(BasicAliasAnalysis, AliasAnalysis, "basicaa",
    556                    "Basic Alias Analysis (stateless AA impl)",
    557                    false, true, false)
    558 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
    559 INITIALIZE_AG_PASS_END(BasicAliasAnalysis, AliasAnalysis, "basicaa",
    560                    "Basic Alias Analysis (stateless AA impl)",
    561                    false, true, false)
    562 
    563 
    564 ImmutablePass *llvm::createBasicAliasAnalysisPass() {
    565   return new BasicAliasAnalysis();
    566 }
    567 
    568 /// pointsToConstantMemory - Returns whether the given pointer value
    569 /// points to memory that is local to the function, with global constants being
    570 /// considered local to all functions.
    571 bool
    572 BasicAliasAnalysis::pointsToConstantMemory(const Location &Loc, bool OrLocal) {
    573   assert(Visited.empty() && "Visited must be cleared after use!");
    574 
    575   unsigned MaxLookup = 8;
    576   SmallVector<const Value *, 16> Worklist;
    577   Worklist.push_back(Loc.Ptr);
    578   do {
    579     const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), TD);
    580     if (!Visited.insert(V)) {
    581       Visited.clear();
    582       return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
    583     }
    584 
    585     // An alloca instruction defines local memory.
    586     if (OrLocal && isa<AllocaInst>(V))
    587       continue;
    588 
    589     // A global constant counts as local memory for our purposes.
    590     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
    591       // Note: this doesn't require GV to be "ODR" because it isn't legal for a
    592       // global to be marked constant in some modules and non-constant in
    593       // others.  GV may even be a declaration, not a definition.
    594       if (!GV->isConstant()) {
    595         Visited.clear();
    596         return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
    597       }
    598       continue;
    599     }
    600 
    601     // If both select values point to local memory, then so does the select.
    602     if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
    603       Worklist.push_back(SI->getTrueValue());
    604       Worklist.push_back(SI->getFalseValue());
    605       continue;
    606     }
    607 
    608     // If all values incoming to a phi node point to local memory, then so does
    609     // the phi.
    610     if (const PHINode *PN = dyn_cast<PHINode>(V)) {
    611       // Don't bother inspecting phi nodes with many operands.
    612       if (PN->getNumIncomingValues() > MaxLookup) {
    613         Visited.clear();
    614         return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
    615       }
    616       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
    617         Worklist.push_back(PN->getIncomingValue(i));
    618       continue;
    619     }
    620 
    621     // Otherwise be conservative.
    622     Visited.clear();
    623     return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
    624 
    625   } while (!Worklist.empty() && --MaxLookup);
    626 
    627   Visited.clear();
    628   return Worklist.empty();
    629 }
    630 
    631 /// getModRefBehavior - Return the behavior when calling the given call site.
    632 AliasAnalysis::ModRefBehavior
    633 BasicAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
    634   if (CS.doesNotAccessMemory())
    635     // Can't do better than this.
    636     return DoesNotAccessMemory;
    637 
    638   ModRefBehavior Min = UnknownModRefBehavior;
    639 
    640   // If the callsite knows it only reads memory, don't return worse
    641   // than that.
    642   if (CS.onlyReadsMemory())
    643     Min = OnlyReadsMemory;
    644 
    645   // The AliasAnalysis base class has some smarts, lets use them.
    646   return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
    647 }
    648 
    649 /// getModRefBehavior - Return the behavior when calling the given function.
    650 /// For use when the call site is not known.
    651 AliasAnalysis::ModRefBehavior
    652 BasicAliasAnalysis::getModRefBehavior(const Function *F) {
    653   // If the function declares it doesn't access memory, we can't do better.
    654   if (F->doesNotAccessMemory())
    655     return DoesNotAccessMemory;
    656 
    657   // For intrinsics, we can check the table.
    658   if (unsigned iid = F->getIntrinsicID()) {
    659 #define GET_INTRINSIC_MODREF_BEHAVIOR
    660 #include "llvm/Intrinsics.gen"
    661 #undef GET_INTRINSIC_MODREF_BEHAVIOR
    662   }
    663 
    664   ModRefBehavior Min = UnknownModRefBehavior;
    665 
    666   // If the function declares it only reads memory, go with that.
    667   if (F->onlyReadsMemory())
    668     Min = OnlyReadsMemory;
    669 
    670   // Otherwise be conservative.
    671   return ModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min);
    672 }
    673 
    674 /// getModRefInfo - Check to see if the specified callsite can clobber the
    675 /// specified memory object.  Since we only look at local properties of this
    676 /// function, we really can't say much about this query.  We do, however, use
    677 /// simple "address taken" analysis on local objects.
    678 AliasAnalysis::ModRefResult
    679 BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS,
    680                                   const Location &Loc) {
    681   assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
    682          "AliasAnalysis query involving multiple functions!");
    683 
    684   const Value *Object = GetUnderlyingObject(Loc.Ptr, TD);
    685 
    686   // If this is a tail call and Loc.Ptr points to a stack location, we know that
    687   // the tail call cannot access or modify the local stack.
    688   // We cannot exclude byval arguments here; these belong to the caller of
    689   // the current function not to the current function, and a tail callee
    690   // may reference them.
    691   if (isa<AllocaInst>(Object))
    692     if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
    693       if (CI->isTailCall())
    694         return NoModRef;
    695 
    696   // If the pointer is to a locally allocated object that does not escape,
    697   // then the call can not mod/ref the pointer unless the call takes the pointer
    698   // as an argument, and itself doesn't capture it.
    699   if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
    700       isNonEscapingLocalObject(Object)) {
    701     bool PassedAsArg = false;
    702     unsigned ArgNo = 0;
    703     for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
    704          CI != CE; ++CI, ++ArgNo) {
    705       // Only look at the no-capture or byval pointer arguments.  If this
    706       // pointer were passed to arguments that were neither of these, then it
    707       // couldn't be no-capture.
    708       if (!(*CI)->getType()->isPointerTy() ||
    709           (!CS.paramHasAttr(ArgNo+1, Attribute::NoCapture) &&
    710            !CS.paramHasAttr(ArgNo+1, Attribute::ByVal)))
    711         continue;
    712 
    713       // If this is a no-capture pointer argument, see if we can tell that it
    714       // is impossible to alias the pointer we're checking.  If not, we have to
    715       // assume that the call could touch the pointer, even though it doesn't
    716       // escape.
    717       if (!isNoAlias(Location(*CI), Location(Object))) {
    718         PassedAsArg = true;
    719         break;
    720       }
    721     }
    722 
    723     if (!PassedAsArg)
    724       return NoModRef;
    725   }
    726 
    727   const TargetLibraryInfo &TLI = getAnalysis<TargetLibraryInfo>();
    728   ModRefResult Min = ModRef;
    729 
    730   // Finally, handle specific knowledge of intrinsics.
    731   const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
    732   if (II != 0)
    733     switch (II->getIntrinsicID()) {
    734     default: break;
    735     case Intrinsic::memcpy:
    736     case Intrinsic::memmove: {
    737       uint64_t Len = UnknownSize;
    738       if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2)))
    739         Len = LenCI->getZExtValue();
    740       Value *Dest = II->getArgOperand(0);
    741       Value *Src = II->getArgOperand(1);
    742       // If it can't overlap the source dest, then it doesn't modref the loc.
    743       if (isNoAlias(Location(Dest, Len), Loc)) {
    744         if (isNoAlias(Location(Src, Len), Loc))
    745           return NoModRef;
    746         // If it can't overlap the dest, then worst case it reads the loc.
    747         Min = Ref;
    748       } else if (isNoAlias(Location(Src, Len), Loc)) {
    749         // If it can't overlap the source, then worst case it mutates the loc.
    750         Min = Mod;
    751       }
    752       break;
    753     }
    754     case Intrinsic::memset:
    755       // Since memset is 'accesses arguments' only, the AliasAnalysis base class
    756       // will handle it for the variable length case.
    757       if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
    758         uint64_t Len = LenCI->getZExtValue();
    759         Value *Dest = II->getArgOperand(0);
    760         if (isNoAlias(Location(Dest, Len), Loc))
    761           return NoModRef;
    762       }
    763       // We know that memset doesn't load anything.
    764       Min = Mod;
    765       break;
    766     case Intrinsic::lifetime_start:
    767     case Intrinsic::lifetime_end:
    768     case Intrinsic::invariant_start: {
    769       uint64_t PtrSize =
    770         cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
    771       if (isNoAlias(Location(II->getArgOperand(1),
    772                              PtrSize,
    773                              II->getMetadata(LLVMContext::MD_tbaa)),
    774                     Loc))
    775         return NoModRef;
    776       break;
    777     }
    778     case Intrinsic::invariant_end: {
    779       uint64_t PtrSize =
    780         cast<ConstantInt>(II->getArgOperand(1))->getZExtValue();
    781       if (isNoAlias(Location(II->getArgOperand(2),
    782                              PtrSize,
    783                              II->getMetadata(LLVMContext::MD_tbaa)),
    784                     Loc))
    785         return NoModRef;
    786       break;
    787     }
    788     //case Intrinsic::arm_neon_vld1: {
    789     //  // LLVM's vld1 and vst1 intrinsics currently only support a single
    790     //  // vector register.
    791     //  uint64_t Size =
    792     //    TD ? TD->getTypeStoreSize(II->getType()) : UnknownSize;
    793     //  if (isNoAlias(Location(II->getArgOperand(0), Size,
    794     //                         II->getMetadata(LLVMContext::MD_tbaa)),
    795     //                Loc))
    796     //    return NoModRef;
    797     //  break;
    798     //}
    799     //case Intrinsic::arm_neon_vst1: {
    800     //  uint64_t Size =
    801     //    TD ? TD->getTypeStoreSize(II->getArgOperand(1)->getType()) : UnknownSize;
    802     //  if (isNoAlias(Location(II->getArgOperand(0), Size,
    803     //                         II->getMetadata(LLVMContext::MD_tbaa)),
    804     //                Loc))
    805     //    return NoModRef;
    806     //  break;
    807     //}
    808     }
    809 
    810   // We can bound the aliasing properties of memset_pattern16 just as we can
    811   // for memcpy/memset.  This is particularly important because the
    812   // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
    813   // whenever possible.
    814   else if (TLI.has(LibFunc::memset_pattern16) &&
    815            CS.getCalledFunction() &&
    816            CS.getCalledFunction()->getName() == "memset_pattern16") {
    817     const Function *MS = CS.getCalledFunction();
    818     FunctionType *MemsetType = MS->getFunctionType();
    819     if (!MemsetType->isVarArg() && MemsetType->getNumParams() == 3 &&
    820         isa<PointerType>(MemsetType->getParamType(0)) &&
    821         isa<PointerType>(MemsetType->getParamType(1)) &&
    822         isa<IntegerType>(MemsetType->getParamType(2))) {
    823       uint64_t Len = UnknownSize;
    824       if (const ConstantInt *LenCI = dyn_cast<ConstantInt>(CS.getArgument(2)))
    825         Len = LenCI->getZExtValue();
    826       const Value *Dest = CS.getArgument(0);
    827       const Value *Src = CS.getArgument(1);
    828       // If it can't overlap the source dest, then it doesn't modref the loc.
    829       if (isNoAlias(Location(Dest, Len), Loc)) {
    830         // Always reads 16 bytes of the source.
    831         if (isNoAlias(Location(Src, 16), Loc))
    832           return NoModRef;
    833         // If it can't overlap the dest, then worst case it reads the loc.
    834         Min = Ref;
    835       // Always reads 16 bytes of the source.
    836       } else if (isNoAlias(Location(Src, 16), Loc)) {
    837         // If it can't overlap the source, then worst case it mutates the loc.
    838         Min = Mod;
    839       }
    840     }
    841   }
    842 
    843   // The AliasAnalysis base class has some smarts, lets use them.
    844   return ModRefResult(AliasAnalysis::getModRefInfo(CS, Loc) & Min);
    845 }
    846 
    847 /// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
    848 /// against another pointer.  We know that V1 is a GEP, but we don't know
    849 /// anything about V2.  UnderlyingV1 is GetUnderlyingObject(GEP1, TD),
    850 /// UnderlyingV2 is the same for V2.
    851 ///
    852 AliasAnalysis::AliasResult
    853 BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
    854                              const Value *V2, uint64_t V2Size,
    855                              const MDNode *V2TBAAInfo,
    856                              const Value *UnderlyingV1,
    857                              const Value *UnderlyingV2) {
    858   int64_t GEP1BaseOffset;
    859   SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
    860 
    861   // If we have two gep instructions with must-alias'ing base pointers, figure
    862   // out if the indexes to the GEP tell us anything about the derived pointer.
    863   if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
    864     // Do the base pointers alias?
    865     AliasResult BaseAlias = aliasCheck(UnderlyingV1, UnknownSize, 0,
    866                                        UnderlyingV2, UnknownSize, 0);
    867 
    868     // If we get a No or May, then return it immediately, no amount of analysis
    869     // will improve this situation.
    870     if (BaseAlias != MustAlias) return BaseAlias;
    871 
    872     // Otherwise, we have a MustAlias.  Since the base pointers alias each other
    873     // exactly, see if the computed offset from the common pointer tells us
    874     // about the relation of the resulting pointer.
    875     const Value *GEP1BasePtr =
    876       DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
    877 
    878     int64_t GEP2BaseOffset;
    879     SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
    880     const Value *GEP2BasePtr =
    881       DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD);
    882 
    883     // If DecomposeGEPExpression isn't able to look all the way through the
    884     // addressing operation, we must not have TD and this is too complex for us
    885     // to handle without it.
    886     if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
    887       assert(TD == 0 &&
    888              "DecomposeGEPExpression and GetUnderlyingObject disagree!");
    889       return MayAlias;
    890     }
    891 
    892     // Subtract the GEP2 pointer from the GEP1 pointer to find out their
    893     // symbolic difference.
    894     GEP1BaseOffset -= GEP2BaseOffset;
    895     GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
    896 
    897   } else {
    898     // Check to see if these two pointers are related by the getelementptr
    899     // instruction.  If one pointer is a GEP with a non-zero index of the other
    900     // pointer, we know they cannot alias.
    901 
    902     // If both accesses are unknown size, we can't do anything useful here.
    903     if (V1Size == UnknownSize && V2Size == UnknownSize)
    904       return MayAlias;
    905 
    906     AliasResult R = aliasCheck(UnderlyingV1, UnknownSize, 0,
    907                                V2, V2Size, V2TBAAInfo);
    908     if (R != MustAlias)
    909       // If V2 may alias GEP base pointer, conservatively returns MayAlias.
    910       // If V2 is known not to alias GEP base pointer, then the two values
    911       // cannot alias per GEP semantics: "A pointer value formed from a
    912       // getelementptr instruction is associated with the addresses associated
    913       // with the first operand of the getelementptr".
    914       return R;
    915 
    916     const Value *GEP1BasePtr =
    917       DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
    918 
    919     // If DecomposeGEPExpression isn't able to look all the way through the
    920     // addressing operation, we must not have TD and this is too complex for us
    921     // to handle without it.
    922     if (GEP1BasePtr != UnderlyingV1) {
    923       assert(TD == 0 &&
    924              "DecomposeGEPExpression and GetUnderlyingObject disagree!");
    925       return MayAlias;
    926     }
    927   }
    928 
    929   // In the two GEP Case, if there is no difference in the offsets of the
    930   // computed pointers, the resultant pointers are a must alias.  This
    931   // hapens when we have two lexically identical GEP's (for example).
    932   //
    933   // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
    934   // must aliases the GEP, the end result is a must alias also.
    935   if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
    936     return MustAlias;
    937 
    938   // If there is a constant difference between the pointers, but the difference
    939   // is less than the size of the associated memory object, then we know
    940   // that the objects are partially overlapping.  If the difference is
    941   // greater, we know they do not overlap.
    942   if (GEP1BaseOffset != 0 && GEP1VariableIndices.empty()) {
    943     if (GEP1BaseOffset >= 0) {
    944       if (V2Size != UnknownSize) {
    945         if ((uint64_t)GEP1BaseOffset < V2Size)
    946           return PartialAlias;
    947         return NoAlias;
    948       }
    949     } else {
    950       if (V1Size != UnknownSize) {
    951         if (-(uint64_t)GEP1BaseOffset < V1Size)
    952           return PartialAlias;
    953         return NoAlias;
    954       }
    955     }
    956   }
    957 
    958   // Try to distinguish something like &A[i][1] against &A[42][0].
    959   // Grab the least significant bit set in any of the scales.
    960   if (!GEP1VariableIndices.empty()) {
    961     uint64_t Modulo = 0;
    962     for (unsigned i = 0, e = GEP1VariableIndices.size(); i != e; ++i)
    963       Modulo |= (uint64_t)GEP1VariableIndices[i].Scale;
    964     Modulo = Modulo ^ (Modulo & (Modulo - 1));
    965 
    966     // We can compute the difference between the two addresses
    967     // mod Modulo. Check whether that difference guarantees that the
    968     // two locations do not alias.
    969     uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1);
    970     if (V1Size != UnknownSize && V2Size != UnknownSize &&
    971         ModOffset >= V2Size && V1Size <= Modulo - ModOffset)
    972       return NoAlias;
    973   }
    974 
    975   // Statically, we can see that the base objects are the same, but the
    976   // pointers have dynamic offsets which we can't resolve. And none of our
    977   // little tricks above worked.
    978   //
    979   // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the
    980   // practical effect of this is protecting TBAA in the case of dynamic
    981   // indices into arrays of unions. An alternative way to solve this would
    982   // be to have clang emit extra metadata for unions and/or union accesses.
    983   // A union-specific solution wouldn't handle the problem for malloc'd
    984   // memory however.
    985   return PartialAlias;
    986 }
    987 
    988 static AliasAnalysis::AliasResult
    989 MergeAliasResults(AliasAnalysis::AliasResult A, AliasAnalysis::AliasResult B) {
    990   // If the results agree, take it.
    991   if (A == B)
    992     return A;
    993   // A mix of PartialAlias and MustAlias is PartialAlias.
    994   if ((A == AliasAnalysis::PartialAlias && B == AliasAnalysis::MustAlias) ||
    995       (B == AliasAnalysis::PartialAlias && A == AliasAnalysis::MustAlias))
    996     return AliasAnalysis::PartialAlias;
    997   // Otherwise, we don't know anything.
    998   return AliasAnalysis::MayAlias;
    999 }
   1000 
   1001 /// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select
   1002 /// instruction against another.
   1003 AliasAnalysis::AliasResult
   1004 BasicAliasAnalysis::aliasSelect(const SelectInst *SI, uint64_t SISize,
   1005                                 const MDNode *SITBAAInfo,
   1006                                 const Value *V2, uint64_t V2Size,
   1007                                 const MDNode *V2TBAAInfo) {
   1008   // If the values are Selects with the same condition, we can do a more precise
   1009   // check: just check for aliases between the values on corresponding arms.
   1010   if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
   1011     if (SI->getCondition() == SI2->getCondition()) {
   1012       AliasResult Alias =
   1013         aliasCheck(SI->getTrueValue(), SISize, SITBAAInfo,
   1014                    SI2->getTrueValue(), V2Size, V2TBAAInfo);
   1015       if (Alias == MayAlias)
   1016         return MayAlias;
   1017       AliasResult ThisAlias =
   1018         aliasCheck(SI->getFalseValue(), SISize, SITBAAInfo,
   1019                    SI2->getFalseValue(), V2Size, V2TBAAInfo);
   1020       return MergeAliasResults(ThisAlias, Alias);
   1021     }
   1022 
   1023   // If both arms of the Select node NoAlias or MustAlias V2, then returns
   1024   // NoAlias / MustAlias. Otherwise, returns MayAlias.
   1025   AliasResult Alias =
   1026     aliasCheck(V2, V2Size, V2TBAAInfo, SI->getTrueValue(), SISize, SITBAAInfo);
   1027   if (Alias == MayAlias)
   1028     return MayAlias;
   1029 
   1030   AliasResult ThisAlias =
   1031     aliasCheck(V2, V2Size, V2TBAAInfo, SI->getFalseValue(), SISize, SITBAAInfo);
   1032   return MergeAliasResults(ThisAlias, Alias);
   1033 }
   1034 
   1035 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
   1036 // against another.
   1037 AliasAnalysis::AliasResult
   1038 BasicAliasAnalysis::aliasPHI(const PHINode *PN, uint64_t PNSize,
   1039                              const MDNode *PNTBAAInfo,
   1040                              const Value *V2, uint64_t V2Size,
   1041                              const MDNode *V2TBAAInfo) {
   1042   // If the values are PHIs in the same block, we can do a more precise
   1043   // as well as efficient check: just check for aliases between the values
   1044   // on corresponding edges.
   1045   if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
   1046     if (PN2->getParent() == PN->getParent()) {
   1047       AliasResult Alias =
   1048         aliasCheck(PN->getIncomingValue(0), PNSize, PNTBAAInfo,
   1049                    PN2->getIncomingValueForBlock(PN->getIncomingBlock(0)),
   1050                    V2Size, V2TBAAInfo);
   1051       if (Alias == MayAlias)
   1052         return MayAlias;
   1053       for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
   1054         AliasResult ThisAlias =
   1055           aliasCheck(PN->getIncomingValue(i), PNSize, PNTBAAInfo,
   1056                      PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
   1057                      V2Size, V2TBAAInfo);
   1058         Alias = MergeAliasResults(ThisAlias, Alias);
   1059         if (Alias == MayAlias)
   1060           break;
   1061       }
   1062       return Alias;
   1063     }
   1064 
   1065   SmallPtrSet<Value*, 4> UniqueSrc;
   1066   SmallVector<Value*, 4> V1Srcs;
   1067   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
   1068     Value *PV1 = PN->getIncomingValue(i);
   1069     if (isa<PHINode>(PV1))
   1070       // If any of the source itself is a PHI, return MayAlias conservatively
   1071       // to avoid compile time explosion. The worst possible case is if both
   1072       // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
   1073       // and 'n' are the number of PHI sources.
   1074       return MayAlias;
   1075     if (UniqueSrc.insert(PV1))
   1076       V1Srcs.push_back(PV1);
   1077   }
   1078 
   1079   AliasResult Alias = aliasCheck(V2, V2Size, V2TBAAInfo,
   1080                                  V1Srcs[0], PNSize, PNTBAAInfo);
   1081   // Early exit if the check of the first PHI source against V2 is MayAlias.
   1082   // Other results are not possible.
   1083   if (Alias == MayAlias)
   1084     return MayAlias;
   1085 
   1086   // If all sources of the PHI node NoAlias or MustAlias V2, then returns
   1087   // NoAlias / MustAlias. Otherwise, returns MayAlias.
   1088   for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
   1089     Value *V = V1Srcs[i];
   1090 
   1091     AliasResult ThisAlias = aliasCheck(V2, V2Size, V2TBAAInfo,
   1092                                        V, PNSize, PNTBAAInfo);
   1093     Alias = MergeAliasResults(ThisAlias, Alias);
   1094     if (Alias == MayAlias)
   1095       break;
   1096   }
   1097 
   1098   return Alias;
   1099 }
   1100 
   1101 // aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
   1102 // such as array references.
   1103 //
   1104 AliasAnalysis::AliasResult
   1105 BasicAliasAnalysis::aliasCheck(const Value *V1, uint64_t V1Size,
   1106                                const MDNode *V1TBAAInfo,
   1107                                const Value *V2, uint64_t V2Size,
   1108                                const MDNode *V2TBAAInfo) {
   1109   // If either of the memory references is empty, it doesn't matter what the
   1110   // pointer values are.
   1111   if (V1Size == 0 || V2Size == 0)
   1112     return NoAlias;
   1113 
   1114   // Strip off any casts if they exist.
   1115   V1 = V1->stripPointerCasts();
   1116   V2 = V2->stripPointerCasts();
   1117 
   1118   // Are we checking for alias of the same value?
   1119   if (V1 == V2) return MustAlias;
   1120 
   1121   if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
   1122     return NoAlias;  // Scalars cannot alias each other
   1123 
   1124   // Figure out what objects these things are pointing to if we can.
   1125   const Value *O1 = GetUnderlyingObject(V1, TD);
   1126   const Value *O2 = GetUnderlyingObject(V2, TD);
   1127 
   1128   // Null values in the default address space don't point to any object, so they
   1129   // don't alias any other pointer.
   1130   if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
   1131     if (CPN->getType()->getAddressSpace() == 0)
   1132       return NoAlias;
   1133   if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
   1134     if (CPN->getType()->getAddressSpace() == 0)
   1135       return NoAlias;
   1136 
   1137   if (O1 != O2) {
   1138     // If V1/V2 point to two different objects we know that we have no alias.
   1139     if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
   1140       return NoAlias;
   1141 
   1142     // Constant pointers can't alias with non-const isIdentifiedObject objects.
   1143     if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
   1144         (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
   1145       return NoAlias;
   1146 
   1147     // Arguments can't alias with local allocations or noalias calls
   1148     // in the same function.
   1149     if (((isa<Argument>(O1) && (isa<AllocaInst>(O2) || isNoAliasCall(O2))) ||
   1150          (isa<Argument>(O2) && (isa<AllocaInst>(O1) || isNoAliasCall(O1)))))
   1151       return NoAlias;
   1152 
   1153     // Most objects can't alias null.
   1154     if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
   1155         (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
   1156       return NoAlias;
   1157 
   1158     // If one pointer is the result of a call/invoke or load and the other is a
   1159     // non-escaping local object within the same function, then we know the
   1160     // object couldn't escape to a point where the call could return it.
   1161     //
   1162     // Note that if the pointers are in different functions, there are a
   1163     // variety of complications. A call with a nocapture argument may still
   1164     // temporary store the nocapture argument's value in a temporary memory
   1165     // location if that memory location doesn't escape. Or it may pass a
   1166     // nocapture value to other functions as long as they don't capture it.
   1167     if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
   1168       return NoAlias;
   1169     if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
   1170       return NoAlias;
   1171   }
   1172 
   1173   // If the size of one access is larger than the entire object on the other
   1174   // side, then we know such behavior is undefined and can assume no alias.
   1175   if (TD)
   1176     if ((V1Size != UnknownSize && isObjectSmallerThan(O2, V1Size, *TD)) ||
   1177         (V2Size != UnknownSize && isObjectSmallerThan(O1, V2Size, *TD)))
   1178       return NoAlias;
   1179 
   1180   // Check the cache before climbing up use-def chains. This also terminates
   1181   // otherwise infinitely recursive queries.
   1182   LocPair Locs(Location(V1, V1Size, V1TBAAInfo),
   1183                Location(V2, V2Size, V2TBAAInfo));
   1184   if (V1 > V2)
   1185     std::swap(Locs.first, Locs.second);
   1186   std::pair<AliasCacheTy::iterator, bool> Pair =
   1187     AliasCache.insert(std::make_pair(Locs, MayAlias));
   1188   if (!Pair.second)
   1189     return Pair.first->second;
   1190 
   1191   // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
   1192   // GEP can't simplify, we don't even look at the PHI cases.
   1193   if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
   1194     std::swap(V1, V2);
   1195     std::swap(V1Size, V2Size);
   1196     std::swap(O1, O2);
   1197   }
   1198   if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
   1199     AliasResult Result = aliasGEP(GV1, V1Size, V2, V2Size, V2TBAAInfo, O1, O2);
   1200     if (Result != MayAlias) return AliasCache[Locs] = Result;
   1201   }
   1202 
   1203   if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
   1204     std::swap(V1, V2);
   1205     std::swap(V1Size, V2Size);
   1206   }
   1207   if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
   1208     AliasResult Result = aliasPHI(PN, V1Size, V1TBAAInfo,
   1209                                   V2, V2Size, V2TBAAInfo);
   1210     if (Result != MayAlias) return AliasCache[Locs] = Result;
   1211   }
   1212 
   1213   if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
   1214     std::swap(V1, V2);
   1215     std::swap(V1Size, V2Size);
   1216   }
   1217   if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
   1218     AliasResult Result = aliasSelect(S1, V1Size, V1TBAAInfo,
   1219                                      V2, V2Size, V2TBAAInfo);
   1220     if (Result != MayAlias) return AliasCache[Locs] = Result;
   1221   }
   1222 
   1223   // If both pointers are pointing into the same object and one of them
   1224   // accesses is accessing the entire object, then the accesses must
   1225   // overlap in some way.
   1226   if (TD && O1 == O2)
   1227     if ((V1Size != UnknownSize && isObjectSize(O1, V1Size, *TD)) ||
   1228         (V2Size != UnknownSize && isObjectSize(O2, V2Size, *TD)))
   1229       return AliasCache[Locs] = PartialAlias;
   1230 
   1231   AliasResult Result =
   1232     AliasAnalysis::alias(Location(V1, V1Size, V1TBAAInfo),
   1233                          Location(V2, V2Size, V2TBAAInfo));
   1234   return AliasCache[Locs] = Result;
   1235 }
   1236