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/Passes.h"
     17 #include "llvm/ADT/SmallPtrSet.h"
     18 #include "llvm/ADT/SmallVector.h"
     19 #include "llvm/Analysis/AliasAnalysis.h"
     20 #include "llvm/Analysis/CaptureTracking.h"
     21 #include "llvm/Analysis/InstructionSimplify.h"
     22 #include "llvm/Analysis/MemoryBuiltins.h"
     23 #include "llvm/Analysis/ValueTracking.h"
     24 #include "llvm/IR/Constants.h"
     25 #include "llvm/IR/DataLayout.h"
     26 #include "llvm/IR/DerivedTypes.h"
     27 #include "llvm/IR/Function.h"
     28 #include "llvm/IR/GlobalAlias.h"
     29 #include "llvm/IR/GlobalVariable.h"
     30 #include "llvm/IR/Instructions.h"
     31 #include "llvm/IR/IntrinsicInst.h"
     32 #include "llvm/IR/LLVMContext.h"
     33 #include "llvm/IR/Operator.h"
     34 #include "llvm/Pass.h"
     35 #include "llvm/Support/ErrorHandling.h"
     36 #include "llvm/Support/GetElementPtrTypeIterator.h"
     37 #include "llvm/Target/TargetLibraryInfo.h"
     38 #include <algorithm>
     39 using namespace llvm;
     40 
     41 //===----------------------------------------------------------------------===//
     42 // Useful predicates
     43 //===----------------------------------------------------------------------===//
     44 
     45 /// isNonEscapingLocalObject - Return true if the pointer is to a function-local
     46 /// object that never escapes from the function.
     47 static bool isNonEscapingLocalObject(const Value *V) {
     48   // If this is a local allocation, check to see if it escapes.
     49   if (isa<AllocaInst>(V) || isNoAliasCall(V))
     50     // Set StoreCaptures to True so that we can assume in our callers that the
     51     // pointer is not the result of a load instruction. Currently
     52     // PointerMayBeCaptured doesn't have any special analysis for the
     53     // StoreCaptures=false case; if it did, our callers could be refined to be
     54     // more precise.
     55     return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
     56 
     57   // If this is an argument that corresponds to a byval or noalias argument,
     58   // then it has not escaped before entering the function.  Check if it escapes
     59   // inside the function.
     60   if (const Argument *A = dyn_cast<Argument>(V))
     61     if (A->hasByValAttr() || A->hasNoAliasAttr())
     62       // Note even if the argument is marked nocapture we still need to check
     63       // for copies made inside the function. The nocapture attribute only
     64       // specifies that there are no copies made that outlive the function.
     65       return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
     66 
     67   return false;
     68 }
     69 
     70 /// isEscapeSource - Return true if the pointer is one which would have
     71 /// been considered an escape by isNonEscapingLocalObject.
     72 static bool isEscapeSource(const Value *V) {
     73   if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
     74     return true;
     75 
     76   // The load case works because isNonEscapingLocalObject considers all
     77   // stores to be escapes (it passes true for the StoreCaptures argument
     78   // to PointerMayBeCaptured).
     79   if (isa<LoadInst>(V))
     80     return true;
     81 
     82   return false;
     83 }
     84 
     85 /// getObjectSize - Return the size of the object specified by V, or
     86 /// UnknownSize if unknown.
     87 static uint64_t getObjectSize(const Value *V, const DataLayout &TD,
     88                               const TargetLibraryInfo &TLI,
     89                               bool RoundToAlign = false) {
     90   uint64_t Size;
     91   if (getObjectSize(V, Size, &TD, &TLI, RoundToAlign))
     92     return Size;
     93   return AliasAnalysis::UnknownSize;
     94 }
     95 
     96 /// isObjectSmallerThan - Return true if we can prove that the object specified
     97 /// by V is smaller than Size.
     98 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
     99                                 const DataLayout &TD,
    100                                 const TargetLibraryInfo &TLI) {
    101   // Note that the meanings of the "object" are slightly different in the
    102   // following contexts:
    103   //    c1: llvm::getObjectSize()
    104   //    c2: llvm.objectsize() intrinsic
    105   //    c3: isObjectSmallerThan()
    106   // c1 and c2 share the same meaning; however, the meaning of "object" in c3
    107   // refers to the "entire object".
    108   //
    109   //  Consider this example:
    110   //     char *p = (char*)malloc(100)
    111   //     char *q = p+80;
    112   //
    113   //  In the context of c1 and c2, the "object" pointed by q refers to the
    114   // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
    115   //
    116   //  However, in the context of c3, the "object" refers to the chunk of memory
    117   // being allocated. So, the "object" has 100 bytes, and q points to the middle
    118   // the "object". In case q is passed to isObjectSmallerThan() as the 1st
    119   // parameter, before the llvm::getObjectSize() is called to get the size of
    120   // entire object, we should:
    121   //    - either rewind the pointer q to the base-address of the object in
    122   //      question (in this case rewind to p), or
    123   //    - just give up. It is up to caller to make sure the pointer is pointing
    124   //      to the base address the object.
    125   //
    126   // We go for 2nd option for simplicity.
    127   if (!isIdentifiedObject(V))
    128     return false;
    129 
    130   // This function needs to use the aligned object size because we allow
    131   // reads a bit past the end given sufficient alignment.
    132   uint64_t ObjectSize = getObjectSize(V, TD, TLI, /*RoundToAlign*/true);
    133 
    134   return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize < Size;
    135 }
    136 
    137 /// isObjectSize - Return true if we can prove that the object specified
    138 /// by V has size Size.
    139 static bool isObjectSize(const Value *V, uint64_t Size,
    140                          const DataLayout &TD, const TargetLibraryInfo &TLI) {
    141   uint64_t ObjectSize = getObjectSize(V, TD, TLI);
    142   return ObjectSize != AliasAnalysis::UnknownSize && ObjectSize == Size;
    143 }
    144 
    145 /// isIdentifiedFunctionLocal - Return true if V is umabigously identified
    146 /// at the function-level. Different IdentifiedFunctionLocals can't alias.
    147 /// Further, an IdentifiedFunctionLocal can not alias with any function
    148 /// arguments other than itself, which is not neccessarily true for
    149 /// IdentifiedObjects.
    150 static bool isIdentifiedFunctionLocal(const Value *V)
    151 {
    152   return isa<AllocaInst>(V) || isNoAliasCall(V) || isNoAliasArgument(V);
    153 }
    154 
    155 
    156 //===----------------------------------------------------------------------===//
    157 // GetElementPtr Instruction Decomposition and Analysis
    158 //===----------------------------------------------------------------------===//
    159 
    160 namespace {
    161   enum ExtensionKind {
    162     EK_NotExtended,
    163     EK_SignExt,
    164     EK_ZeroExt
    165   };
    166 
    167   struct VariableGEPIndex {
    168     const Value *V;
    169     ExtensionKind Extension;
    170     int64_t Scale;
    171 
    172     bool operator==(const VariableGEPIndex &Other) const {
    173       return V == Other.V && Extension == Other.Extension &&
    174         Scale == Other.Scale;
    175     }
    176 
    177     bool operator!=(const VariableGEPIndex &Other) const {
    178       return !operator==(Other);
    179     }
    180   };
    181 }
    182 
    183 
    184 /// GetLinearExpression - Analyze the specified value as a linear expression:
    185 /// "A*V + B", where A and B are constant integers.  Return the scale and offset
    186 /// values as APInts and return V as a Value*, and return whether we looked
    187 /// through any sign or zero extends.  The incoming Value is known to have
    188 /// IntegerType and it may already be sign or zero extended.
    189 ///
    190 /// Note that this looks through extends, so the high bits may not be
    191 /// represented in the result.
    192 static Value *GetLinearExpression(Value *V, APInt &Scale, APInt &Offset,
    193                                   ExtensionKind &Extension,
    194                                   const DataLayout &TD, unsigned Depth) {
    195   assert(V->getType()->isIntegerTy() && "Not an integer value");
    196 
    197   // Limit our recursion depth.
    198   if (Depth == 6) {
    199     Scale = 1;
    200     Offset = 0;
    201     return V;
    202   }
    203 
    204   if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
    205     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
    206       switch (BOp->getOpcode()) {
    207       default: break;
    208       case Instruction::Or:
    209         // X|C == X+C if all the bits in C are unset in X.  Otherwise we can't
    210         // analyze it.
    211         if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), &TD))
    212           break;
    213         // FALL THROUGH.
    214       case Instruction::Add:
    215         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
    216                                 TD, Depth+1);
    217         Offset += RHSC->getValue();
    218         return V;
    219       case Instruction::Mul:
    220         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
    221                                 TD, Depth+1);
    222         Offset *= RHSC->getValue();
    223         Scale *= RHSC->getValue();
    224         return V;
    225       case Instruction::Shl:
    226         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension,
    227                                 TD, Depth+1);
    228         Offset <<= RHSC->getValue().getLimitedValue();
    229         Scale <<= RHSC->getValue().getLimitedValue();
    230         return V;
    231       }
    232     }
    233   }
    234 
    235   // Since GEP indices are sign extended anyway, we don't care about the high
    236   // bits of a sign or zero extended value - just scales and offsets.  The
    237   // extensions have to be consistent though.
    238   if ((isa<SExtInst>(V) && Extension != EK_ZeroExt) ||
    239       (isa<ZExtInst>(V) && Extension != EK_SignExt)) {
    240     Value *CastOp = cast<CastInst>(V)->getOperand(0);
    241     unsigned OldWidth = Scale.getBitWidth();
    242     unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
    243     Scale = Scale.trunc(SmallWidth);
    244     Offset = Offset.trunc(SmallWidth);
    245     Extension = isa<SExtInst>(V) ? EK_SignExt : EK_ZeroExt;
    246 
    247     Value *Result = GetLinearExpression(CastOp, Scale, Offset, Extension,
    248                                         TD, Depth+1);
    249     Scale = Scale.zext(OldWidth);
    250     Offset = Offset.zext(OldWidth);
    251 
    252     return Result;
    253   }
    254 
    255   Scale = 1;
    256   Offset = 0;
    257   return V;
    258 }
    259 
    260 /// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it
    261 /// into a base pointer with a constant offset and a number of scaled symbolic
    262 /// offsets.
    263 ///
    264 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale in
    265 /// the VarIndices vector) are Value*'s that are known to be scaled by the
    266 /// specified amount, but which may have other unrepresented high bits. As such,
    267 /// the gep cannot necessarily be reconstructed from its decomposed form.
    268 ///
    269 /// When DataLayout is around, this function is capable of analyzing everything
    270 /// that GetUnderlyingObject can look through.  When not, it just looks
    271 /// through pointer casts.
    272 ///
    273 static const Value *
    274 DecomposeGEPExpression(const Value *V, int64_t &BaseOffs,
    275                        SmallVectorImpl<VariableGEPIndex> &VarIndices,
    276                        const DataLayout *TD) {
    277   // Limit recursion depth to limit compile time in crazy cases.
    278   unsigned MaxLookup = 6;
    279 
    280   BaseOffs = 0;
    281   do {
    282     // See if this is a bitcast or GEP.
    283     const Operator *Op = dyn_cast<Operator>(V);
    284     if (Op == 0) {
    285       // The only non-operator case we can handle are GlobalAliases.
    286       if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
    287         if (!GA->mayBeOverridden()) {
    288           V = GA->getAliasee();
    289           continue;
    290         }
    291       }
    292       return V;
    293     }
    294 
    295     if (Op->getOpcode() == Instruction::BitCast) {
    296       V = Op->getOperand(0);
    297       continue;
    298     }
    299 
    300     const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
    301     if (GEPOp == 0) {
    302       // If it's not a GEP, hand it off to SimplifyInstruction to see if it
    303       // can come up with something. This matches what GetUnderlyingObject does.
    304       if (const Instruction *I = dyn_cast<Instruction>(V))
    305         // TODO: Get a DominatorTree and use it here.
    306         if (const Value *Simplified =
    307               SimplifyInstruction(const_cast<Instruction *>(I), TD)) {
    308           V = Simplified;
    309           continue;
    310         }
    311 
    312       return V;
    313     }
    314 
    315     // Don't attempt to analyze GEPs over unsized objects.
    316     if (!cast<PointerType>(GEPOp->getOperand(0)->getType())
    317         ->getElementType()->isSized())
    318       return V;
    319 
    320     // If we are lacking DataLayout information, we can't compute the offets of
    321     // elements computed by GEPs.  However, we can handle bitcast equivalent
    322     // GEPs.
    323     if (TD == 0) {
    324       if (!GEPOp->hasAllZeroIndices())
    325         return V;
    326       V = GEPOp->getOperand(0);
    327       continue;
    328     }
    329 
    330     // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
    331     gep_type_iterator GTI = gep_type_begin(GEPOp);
    332     for (User::const_op_iterator I = GEPOp->op_begin()+1,
    333          E = GEPOp->op_end(); I != E; ++I) {
    334       Value *Index = *I;
    335       // Compute the (potentially symbolic) offset in bytes for this index.
    336       if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
    337         // For a struct, add the member offset.
    338         unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
    339         if (FieldNo == 0) continue;
    340 
    341         BaseOffs += TD->getStructLayout(STy)->getElementOffset(FieldNo);
    342         continue;
    343       }
    344 
    345       // For an array/pointer, add the element offset, explicitly scaled.
    346       if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
    347         if (CIdx->isZero()) continue;
    348         BaseOffs += TD->getTypeAllocSize(*GTI)*CIdx->getSExtValue();
    349         continue;
    350       }
    351 
    352       uint64_t Scale = TD->getTypeAllocSize(*GTI);
    353       ExtensionKind Extension = EK_NotExtended;
    354 
    355       // If the integer type is smaller than the pointer size, it is implicitly
    356       // sign extended to pointer size.
    357       unsigned Width = cast<IntegerType>(Index->getType())->getBitWidth();
    358       if (TD->getPointerSizeInBits() > Width)
    359         Extension = EK_SignExt;
    360 
    361       // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
    362       APInt IndexScale(Width, 0), IndexOffset(Width, 0);
    363       Index = GetLinearExpression(Index, IndexScale, IndexOffset, Extension,
    364                                   *TD, 0);
    365 
    366       // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
    367       // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
    368       BaseOffs += IndexOffset.getSExtValue()*Scale;
    369       Scale *= IndexScale.getSExtValue();
    370 
    371 
    372       // If we already had an occurrence of this index variable, merge this
    373       // scale into it.  For example, we want to handle:
    374       //   A[x][x] -> x*16 + x*4 -> x*20
    375       // This also ensures that 'x' only appears in the index list once.
    376       for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
    377         if (VarIndices[i].V == Index &&
    378             VarIndices[i].Extension == Extension) {
    379           Scale += VarIndices[i].Scale;
    380           VarIndices.erase(VarIndices.begin()+i);
    381           break;
    382         }
    383       }
    384 
    385       // Make sure that we have a scale that makes sense for this target's
    386       // pointer size.
    387       if (unsigned ShiftBits = 64-TD->getPointerSizeInBits()) {
    388         Scale <<= ShiftBits;
    389         Scale = (int64_t)Scale >> ShiftBits;
    390       }
    391 
    392       if (Scale) {
    393         VariableGEPIndex Entry = {Index, Extension,
    394                                   static_cast<int64_t>(Scale)};
    395         VarIndices.push_back(Entry);
    396       }
    397     }
    398 
    399     // Analyze the base pointer next.
    400     V = GEPOp->getOperand(0);
    401   } while (--MaxLookup);
    402 
    403   // If the chain of expressions is too deep, just return early.
    404   return V;
    405 }
    406 
    407 /// GetIndexDifference - Dest and Src are the variable indices from two
    408 /// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base
    409 /// pointers.  Subtract the GEP2 indices from GEP1 to find the symbolic
    410 /// difference between the two pointers.
    411 static void GetIndexDifference(SmallVectorImpl<VariableGEPIndex> &Dest,
    412                                const SmallVectorImpl<VariableGEPIndex> &Src) {
    413   if (Src.empty()) return;
    414 
    415   for (unsigned i = 0, e = Src.size(); i != e; ++i) {
    416     const Value *V = Src[i].V;
    417     ExtensionKind Extension = Src[i].Extension;
    418     int64_t Scale = Src[i].Scale;
    419 
    420     // Find V in Dest.  This is N^2, but pointer indices almost never have more
    421     // than a few variable indexes.
    422     for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
    423       if (Dest[j].V != V || Dest[j].Extension != Extension) continue;
    424 
    425       // If we found it, subtract off Scale V's from the entry in Dest.  If it
    426       // goes to zero, remove the entry.
    427       if (Dest[j].Scale != Scale)
    428         Dest[j].Scale -= Scale;
    429       else
    430         Dest.erase(Dest.begin()+j);
    431       Scale = 0;
    432       break;
    433     }
    434 
    435     // If we didn't consume this entry, add it to the end of the Dest list.
    436     if (Scale) {
    437       VariableGEPIndex Entry = { V, Extension, -Scale };
    438       Dest.push_back(Entry);
    439     }
    440   }
    441 }
    442 
    443 //===----------------------------------------------------------------------===//
    444 // BasicAliasAnalysis Pass
    445 //===----------------------------------------------------------------------===//
    446 
    447 #ifndef NDEBUG
    448 static const Function *getParent(const Value *V) {
    449   if (const Instruction *inst = dyn_cast<Instruction>(V))
    450     return inst->getParent()->getParent();
    451 
    452   if (const Argument *arg = dyn_cast<Argument>(V))
    453     return arg->getParent();
    454 
    455   return NULL;
    456 }
    457 
    458 static bool notDifferentParent(const Value *O1, const Value *O2) {
    459 
    460   const Function *F1 = getParent(O1);
    461   const Function *F2 = getParent(O2);
    462 
    463   return !F1 || !F2 || F1 == F2;
    464 }
    465 #endif
    466 
    467 namespace {
    468   /// BasicAliasAnalysis - This is the primary alias analysis implementation.
    469   struct BasicAliasAnalysis : public ImmutablePass, public AliasAnalysis {
    470     static char ID; // Class identification, replacement for typeinfo
    471     BasicAliasAnalysis() : ImmutablePass(ID) {
    472       initializeBasicAliasAnalysisPass(*PassRegistry::getPassRegistry());
    473     }
    474 
    475     virtual void initializePass() {
    476       InitializeAliasAnalysis(this);
    477     }
    478 
    479     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
    480       AU.addRequired<AliasAnalysis>();
    481       AU.addRequired<TargetLibraryInfo>();
    482     }
    483 
    484     virtual AliasResult alias(const Location &LocA,
    485                               const Location &LocB) {
    486       assert(AliasCache.empty() && "AliasCache must be cleared after use!");
    487       assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
    488              "BasicAliasAnalysis doesn't support interprocedural queries.");
    489       AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.TBAATag,
    490                                      LocB.Ptr, LocB.Size, LocB.TBAATag);
    491       // AliasCache rarely has more than 1 or 2 elements, always use
    492       // shrink_and_clear so it quickly returns to the inline capacity of the
    493       // SmallDenseMap if it ever grows larger.
    494       // FIXME: This should really be shrink_to_inline_capacity_and_clear().
    495       AliasCache.shrink_and_clear();
    496       return Alias;
    497     }
    498 
    499     virtual ModRefResult getModRefInfo(ImmutableCallSite CS,
    500                                        const Location &Loc);
    501 
    502     virtual ModRefResult getModRefInfo(ImmutableCallSite CS1,
    503                                        ImmutableCallSite CS2) {
    504       // The AliasAnalysis base class has some smarts, lets use them.
    505       return AliasAnalysis::getModRefInfo(CS1, CS2);
    506     }
    507 
    508     /// pointsToConstantMemory - Chase pointers until we find a (constant
    509     /// global) or not.
    510     virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal);
    511 
    512     /// getModRefBehavior - Return the behavior when calling the given
    513     /// call site.
    514     virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS);
    515 
    516     /// getModRefBehavior - Return the behavior when calling the given function.
    517     /// For use when the call site is not known.
    518     virtual ModRefBehavior getModRefBehavior(const Function *F);
    519 
    520     /// getAdjustedAnalysisPointer - This method is used when a pass implements
    521     /// an analysis interface through multiple inheritance.  If needed, it
    522     /// should override this to adjust the this pointer as needed for the
    523     /// specified pass info.
    524     virtual void *getAdjustedAnalysisPointer(const void *ID) {
    525       if (ID == &AliasAnalysis::ID)
    526         return (AliasAnalysis*)this;
    527       return this;
    528     }
    529 
    530   private:
    531     // AliasCache - Track alias queries to guard against recursion.
    532     typedef std::pair<Location, Location> LocPair;
    533     typedef SmallDenseMap<LocPair, AliasResult, 8> AliasCacheTy;
    534     AliasCacheTy AliasCache;
    535 
    536     // Visited - Track instructions visited by pointsToConstantMemory.
    537     SmallPtrSet<const Value*, 16> Visited;
    538 
    539     // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP
    540     // instruction against another.
    541     AliasResult aliasGEP(const GEPOperator *V1, uint64_t V1Size,
    542                          const MDNode *V1TBAAInfo,
    543                          const Value *V2, uint64_t V2Size,
    544                          const MDNode *V2TBAAInfo,
    545                          const Value *UnderlyingV1, const Value *UnderlyingV2);
    546 
    547     // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI
    548     // instruction against another.
    549     AliasResult aliasPHI(const PHINode *PN, uint64_t PNSize,
    550                          const MDNode *PNTBAAInfo,
    551                          const Value *V2, uint64_t V2Size,
    552                          const MDNode *V2TBAAInfo);
    553 
    554     /// aliasSelect - Disambiguate a Select instruction against another value.
    555     AliasResult aliasSelect(const SelectInst *SI, uint64_t SISize,
    556                             const MDNode *SITBAAInfo,
    557                             const Value *V2, uint64_t V2Size,
    558                             const MDNode *V2TBAAInfo);
    559 
    560     AliasResult aliasCheck(const Value *V1, uint64_t V1Size,
    561                            const MDNode *V1TBAATag,
    562                            const Value *V2, uint64_t V2Size,
    563                            const MDNode *V2TBAATag);
    564   };
    565 }  // End of anonymous namespace
    566 
    567 // Register this pass...
    568 char BasicAliasAnalysis::ID = 0;
    569 INITIALIZE_AG_PASS_BEGIN(BasicAliasAnalysis, AliasAnalysis, "basicaa",
    570                    "Basic Alias Analysis (stateless AA impl)",
    571                    false, true, false)
    572 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
    573 INITIALIZE_AG_PASS_END(BasicAliasAnalysis, AliasAnalysis, "basicaa",
    574                    "Basic Alias Analysis (stateless AA impl)",
    575                    false, true, false)
    576 
    577 
    578 ImmutablePass *llvm::createBasicAliasAnalysisPass() {
    579   return new BasicAliasAnalysis();
    580 }
    581 
    582 /// pointsToConstantMemory - Returns whether the given pointer value
    583 /// points to memory that is local to the function, with global constants being
    584 /// considered local to all functions.
    585 bool
    586 BasicAliasAnalysis::pointsToConstantMemory(const Location &Loc, bool OrLocal) {
    587   assert(Visited.empty() && "Visited must be cleared after use!");
    588 
    589   unsigned MaxLookup = 8;
    590   SmallVector<const Value *, 16> Worklist;
    591   Worklist.push_back(Loc.Ptr);
    592   do {
    593     const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), TD);
    594     if (!Visited.insert(V)) {
    595       Visited.clear();
    596       return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
    597     }
    598 
    599     // An alloca instruction defines local memory.
    600     if (OrLocal && isa<AllocaInst>(V))
    601       continue;
    602 
    603     // A global constant counts as local memory for our purposes.
    604     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
    605       // Note: this doesn't require GV to be "ODR" because it isn't legal for a
    606       // global to be marked constant in some modules and non-constant in
    607       // others.  GV may even be a declaration, not a definition.
    608       if (!GV->isConstant()) {
    609         Visited.clear();
    610         return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
    611       }
    612       continue;
    613     }
    614 
    615     // If both select values point to local memory, then so does the select.
    616     if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
    617       Worklist.push_back(SI->getTrueValue());
    618       Worklist.push_back(SI->getFalseValue());
    619       continue;
    620     }
    621 
    622     // If all values incoming to a phi node point to local memory, then so does
    623     // the phi.
    624     if (const PHINode *PN = dyn_cast<PHINode>(V)) {
    625       // Don't bother inspecting phi nodes with many operands.
    626       if (PN->getNumIncomingValues() > MaxLookup) {
    627         Visited.clear();
    628         return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
    629       }
    630       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
    631         Worklist.push_back(PN->getIncomingValue(i));
    632       continue;
    633     }
    634 
    635     // Otherwise be conservative.
    636     Visited.clear();
    637     return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
    638 
    639   } while (!Worklist.empty() && --MaxLookup);
    640 
    641   Visited.clear();
    642   return Worklist.empty();
    643 }
    644 
    645 /// getModRefBehavior - Return the behavior when calling the given call site.
    646 AliasAnalysis::ModRefBehavior
    647 BasicAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
    648   if (CS.doesNotAccessMemory())
    649     // Can't do better than this.
    650     return DoesNotAccessMemory;
    651 
    652   ModRefBehavior Min = UnknownModRefBehavior;
    653 
    654   // If the callsite knows it only reads memory, don't return worse
    655   // than that.
    656   if (CS.onlyReadsMemory())
    657     Min = OnlyReadsMemory;
    658 
    659   // The AliasAnalysis base class has some smarts, lets use them.
    660   return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
    661 }
    662 
    663 /// getModRefBehavior - Return the behavior when calling the given function.
    664 /// For use when the call site is not known.
    665 AliasAnalysis::ModRefBehavior
    666 BasicAliasAnalysis::getModRefBehavior(const Function *F) {
    667   // If the function declares it doesn't access memory, we can't do better.
    668   if (F->doesNotAccessMemory())
    669     return DoesNotAccessMemory;
    670 
    671   // For intrinsics, we can check the table.
    672   if (unsigned iid = F->getIntrinsicID()) {
    673 #define GET_INTRINSIC_MODREF_BEHAVIOR
    674 #include "llvm/IR/Intrinsics.gen"
    675 #undef GET_INTRINSIC_MODREF_BEHAVIOR
    676   }
    677 
    678   ModRefBehavior Min = UnknownModRefBehavior;
    679 
    680   // If the function declares it only reads memory, go with that.
    681   if (F->onlyReadsMemory())
    682     Min = OnlyReadsMemory;
    683 
    684   // Otherwise be conservative.
    685   return ModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min);
    686 }
    687 
    688 /// getModRefInfo - Check to see if the specified callsite can clobber the
    689 /// specified memory object.  Since we only look at local properties of this
    690 /// function, we really can't say much about this query.  We do, however, use
    691 /// simple "address taken" analysis on local objects.
    692 AliasAnalysis::ModRefResult
    693 BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS,
    694                                   const Location &Loc) {
    695   assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
    696          "AliasAnalysis query involving multiple functions!");
    697 
    698   const Value *Object = GetUnderlyingObject(Loc.Ptr, TD);
    699 
    700   // If this is a tail call and Loc.Ptr points to a stack location, we know that
    701   // the tail call cannot access or modify the local stack.
    702   // We cannot exclude byval arguments here; these belong to the caller of
    703   // the current function not to the current function, and a tail callee
    704   // may reference them.
    705   if (isa<AllocaInst>(Object))
    706     if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
    707       if (CI->isTailCall())
    708         return NoModRef;
    709 
    710   // If the pointer is to a locally allocated object that does not escape,
    711   // then the call can not mod/ref the pointer unless the call takes the pointer
    712   // as an argument, and itself doesn't capture it.
    713   if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
    714       isNonEscapingLocalObject(Object)) {
    715     bool PassedAsArg = false;
    716     unsigned ArgNo = 0;
    717     for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
    718          CI != CE; ++CI, ++ArgNo) {
    719       // Only look at the no-capture or byval pointer arguments.  If this
    720       // pointer were passed to arguments that were neither of these, then it
    721       // couldn't be no-capture.
    722       if (!(*CI)->getType()->isPointerTy() ||
    723           (!CS.doesNotCapture(ArgNo) && !CS.isByValArgument(ArgNo)))
    724         continue;
    725 
    726       // If this is a no-capture pointer argument, see if we can tell that it
    727       // is impossible to alias the pointer we're checking.  If not, we have to
    728       // assume that the call could touch the pointer, even though it doesn't
    729       // escape.
    730       if (!isNoAlias(Location(*CI), Location(Object))) {
    731         PassedAsArg = true;
    732         break;
    733       }
    734     }
    735 
    736     if (!PassedAsArg)
    737       return NoModRef;
    738   }
    739 
    740   const TargetLibraryInfo &TLI = getAnalysis<TargetLibraryInfo>();
    741   ModRefResult Min = ModRef;
    742 
    743   // Finally, handle specific knowledge of intrinsics.
    744   const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
    745   if (II != 0)
    746     switch (II->getIntrinsicID()) {
    747     default: break;
    748     case Intrinsic::memcpy:
    749     case Intrinsic::memmove: {
    750       uint64_t Len = UnknownSize;
    751       if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2)))
    752         Len = LenCI->getZExtValue();
    753       Value *Dest = II->getArgOperand(0);
    754       Value *Src = II->getArgOperand(1);
    755       // If it can't overlap the source dest, then it doesn't modref the loc.
    756       if (isNoAlias(Location(Dest, Len), Loc)) {
    757         if (isNoAlias(Location(Src, Len), Loc))
    758           return NoModRef;
    759         // If it can't overlap the dest, then worst case it reads the loc.
    760         Min = Ref;
    761       } else if (isNoAlias(Location(Src, Len), Loc)) {
    762         // If it can't overlap the source, then worst case it mutates the loc.
    763         Min = Mod;
    764       }
    765       break;
    766     }
    767     case Intrinsic::memset:
    768       // Since memset is 'accesses arguments' only, the AliasAnalysis base class
    769       // will handle it for the variable length case.
    770       if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
    771         uint64_t Len = LenCI->getZExtValue();
    772         Value *Dest = II->getArgOperand(0);
    773         if (isNoAlias(Location(Dest, Len), Loc))
    774           return NoModRef;
    775       }
    776       // We know that memset doesn't load anything.
    777       Min = Mod;
    778       break;
    779     case Intrinsic::lifetime_start:
    780     case Intrinsic::lifetime_end:
    781     case Intrinsic::invariant_start: {
    782       uint64_t PtrSize =
    783         cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
    784       if (isNoAlias(Location(II->getArgOperand(1),
    785                              PtrSize,
    786                              II->getMetadata(LLVMContext::MD_tbaa)),
    787                     Loc))
    788         return NoModRef;
    789       break;
    790     }
    791     case Intrinsic::invariant_end: {
    792       uint64_t PtrSize =
    793         cast<ConstantInt>(II->getArgOperand(1))->getZExtValue();
    794       if (isNoAlias(Location(II->getArgOperand(2),
    795                              PtrSize,
    796                              II->getMetadata(LLVMContext::MD_tbaa)),
    797                     Loc))
    798         return NoModRef;
    799       break;
    800     }
    801     case Intrinsic::arm_neon_vld1: {
    802       // LLVM's vld1 and vst1 intrinsics currently only support a single
    803       // vector register.
    804       uint64_t Size =
    805         TD ? TD->getTypeStoreSize(II->getType()) : UnknownSize;
    806       if (isNoAlias(Location(II->getArgOperand(0), Size,
    807                              II->getMetadata(LLVMContext::MD_tbaa)),
    808                     Loc))
    809         return NoModRef;
    810       break;
    811     }
    812     case Intrinsic::arm_neon_vst1: {
    813       uint64_t Size =
    814         TD ? TD->getTypeStoreSize(II->getArgOperand(1)->getType()) : UnknownSize;
    815       if (isNoAlias(Location(II->getArgOperand(0), Size,
    816                              II->getMetadata(LLVMContext::MD_tbaa)),
    817                     Loc))
    818         return NoModRef;
    819       break;
    820     }
    821     }
    822 
    823   // We can bound the aliasing properties of memset_pattern16 just as we can
    824   // for memcpy/memset.  This is particularly important because the
    825   // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
    826   // whenever possible.
    827   else if (TLI.has(LibFunc::memset_pattern16) &&
    828            CS.getCalledFunction() &&
    829            CS.getCalledFunction()->getName() == "memset_pattern16") {
    830     const Function *MS = CS.getCalledFunction();
    831     FunctionType *MemsetType = MS->getFunctionType();
    832     if (!MemsetType->isVarArg() && MemsetType->getNumParams() == 3 &&
    833         isa<PointerType>(MemsetType->getParamType(0)) &&
    834         isa<PointerType>(MemsetType->getParamType(1)) &&
    835         isa<IntegerType>(MemsetType->getParamType(2))) {
    836       uint64_t Len = UnknownSize;
    837       if (const ConstantInt *LenCI = dyn_cast<ConstantInt>(CS.getArgument(2)))
    838         Len = LenCI->getZExtValue();
    839       const Value *Dest = CS.getArgument(0);
    840       const Value *Src = CS.getArgument(1);
    841       // If it can't overlap the source dest, then it doesn't modref the loc.
    842       if (isNoAlias(Location(Dest, Len), Loc)) {
    843         // Always reads 16 bytes of the source.
    844         if (isNoAlias(Location(Src, 16), Loc))
    845           return NoModRef;
    846         // If it can't overlap the dest, then worst case it reads the loc.
    847         Min = Ref;
    848       // Always reads 16 bytes of the source.
    849       } else if (isNoAlias(Location(Src, 16), Loc)) {
    850         // If it can't overlap the source, then worst case it mutates the loc.
    851         Min = Mod;
    852       }
    853     }
    854   }
    855 
    856   // The AliasAnalysis base class has some smarts, lets use them.
    857   return ModRefResult(AliasAnalysis::getModRefInfo(CS, Loc) & Min);
    858 }
    859 
    860 static bool areVarIndicesEqual(SmallVectorImpl<VariableGEPIndex> &Indices1,
    861                                SmallVectorImpl<VariableGEPIndex> &Indices2) {
    862   unsigned Size1 = Indices1.size();
    863   unsigned Size2 = Indices2.size();
    864 
    865   if (Size1 != Size2)
    866     return false;
    867 
    868   for (unsigned I = 0; I != Size1; ++I)
    869     if (Indices1[I] != Indices2[I])
    870       return false;
    871 
    872   return true;
    873 }
    874 
    875 /// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
    876 /// against another pointer.  We know that V1 is a GEP, but we don't know
    877 /// anything about V2.  UnderlyingV1 is GetUnderlyingObject(GEP1, TD),
    878 /// UnderlyingV2 is the same for V2.
    879 ///
    880 AliasAnalysis::AliasResult
    881 BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
    882                              const MDNode *V1TBAAInfo,
    883                              const Value *V2, uint64_t V2Size,
    884                              const MDNode *V2TBAAInfo,
    885                              const Value *UnderlyingV1,
    886                              const Value *UnderlyingV2) {
    887   int64_t GEP1BaseOffset;
    888   SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
    889 
    890   // If we have two gep instructions with must-alias or not-alias'ing base
    891   // pointers, figure out if the indexes to the GEP tell us anything about the
    892   // derived pointer.
    893   if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
    894     // Do the base pointers alias?
    895     AliasResult BaseAlias = aliasCheck(UnderlyingV1, UnknownSize, 0,
    896                                        UnderlyingV2, UnknownSize, 0);
    897 
    898     // Check for geps of non-aliasing underlying pointers where the offsets are
    899     // identical.
    900     if ((BaseAlias == MayAlias) && V1Size == V2Size) {
    901       // Do the base pointers alias assuming type and size.
    902       AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size,
    903                                                 V1TBAAInfo, UnderlyingV2,
    904                                                 V2Size, V2TBAAInfo);
    905       if (PreciseBaseAlias == NoAlias) {
    906         // See if the computed offset from the common pointer tells us about the
    907         // relation of the resulting pointer.
    908         int64_t GEP2BaseOffset;
    909         SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
    910         const Value *GEP2BasePtr =
    911           DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD);
    912         const Value *GEP1BasePtr =
    913           DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
    914         // DecomposeGEPExpression and GetUnderlyingObject should return the
    915         // same result except when DecomposeGEPExpression has no DataLayout.
    916         if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
    917           assert(TD == 0 &&
    918              "DecomposeGEPExpression and GetUnderlyingObject disagree!");
    919           return MayAlias;
    920         }
    921         // Same offsets.
    922         if (GEP1BaseOffset == GEP2BaseOffset &&
    923             areVarIndicesEqual(GEP1VariableIndices, GEP2VariableIndices))
    924           return NoAlias;
    925         GEP1VariableIndices.clear();
    926       }
    927     }
    928 
    929     // If we get a No or May, then return it immediately, no amount of analysis
    930     // will improve this situation.
    931     if (BaseAlias != MustAlias) return BaseAlias;
    932 
    933     // Otherwise, we have a MustAlias.  Since the base pointers alias each other
    934     // exactly, see if the computed offset from the common pointer tells us
    935     // about the relation of the resulting pointer.
    936     const Value *GEP1BasePtr =
    937       DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
    938 
    939     int64_t GEP2BaseOffset;
    940     SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
    941     const Value *GEP2BasePtr =
    942       DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD);
    943 
    944     // DecomposeGEPExpression and GetUnderlyingObject should return the
    945     // same result except when DecomposeGEPExpression has no DataLayout.
    946     if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
    947       assert(TD == 0 &&
    948              "DecomposeGEPExpression and GetUnderlyingObject disagree!");
    949       return MayAlias;
    950     }
    951 
    952     // Subtract the GEP2 pointer from the GEP1 pointer to find out their
    953     // symbolic difference.
    954     GEP1BaseOffset -= GEP2BaseOffset;
    955     GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
    956 
    957   } else {
    958     // Check to see if these two pointers are related by the getelementptr
    959     // instruction.  If one pointer is a GEP with a non-zero index of the other
    960     // pointer, we know they cannot alias.
    961 
    962     // If both accesses are unknown size, we can't do anything useful here.
    963     if (V1Size == UnknownSize && V2Size == UnknownSize)
    964       return MayAlias;
    965 
    966     AliasResult R = aliasCheck(UnderlyingV1, UnknownSize, 0,
    967                                V2, V2Size, V2TBAAInfo);
    968     if (R != MustAlias)
    969       // If V2 may alias GEP base pointer, conservatively returns MayAlias.
    970       // If V2 is known not to alias GEP base pointer, then the two values
    971       // cannot alias per GEP semantics: "A pointer value formed from a
    972       // getelementptr instruction is associated with the addresses associated
    973       // with the first operand of the getelementptr".
    974       return R;
    975 
    976     const Value *GEP1BasePtr =
    977       DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
    978 
    979     // DecomposeGEPExpression and GetUnderlyingObject should return the
    980     // same result except when DecomposeGEPExpression has no DataLayout.
    981     if (GEP1BasePtr != UnderlyingV1) {
    982       assert(TD == 0 &&
    983              "DecomposeGEPExpression and GetUnderlyingObject disagree!");
    984       return MayAlias;
    985     }
    986   }
    987 
    988   // In the two GEP Case, if there is no difference in the offsets of the
    989   // computed pointers, the resultant pointers are a must alias.  This
    990   // hapens when we have two lexically identical GEP's (for example).
    991   //
    992   // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
    993   // must aliases the GEP, the end result is a must alias also.
    994   if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
    995     return MustAlias;
    996 
    997   // If there is a constant difference between the pointers, but the difference
    998   // is less than the size of the associated memory object, then we know
    999   // that the objects are partially overlapping.  If the difference is
   1000   // greater, we know they do not overlap.
   1001   if (GEP1BaseOffset != 0 && GEP1VariableIndices.empty()) {
   1002     if (GEP1BaseOffset >= 0) {
   1003       if (V2Size != UnknownSize) {
   1004         if ((uint64_t)GEP1BaseOffset < V2Size)
   1005           return PartialAlias;
   1006         return NoAlias;
   1007       }
   1008     } else {
   1009       if (V1Size != UnknownSize) {
   1010         if (-(uint64_t)GEP1BaseOffset < V1Size)
   1011           return PartialAlias;
   1012         return NoAlias;
   1013       }
   1014     }
   1015   }
   1016 
   1017   // Try to distinguish something like &A[i][1] against &A[42][0].
   1018   // Grab the least significant bit set in any of the scales.
   1019   if (!GEP1VariableIndices.empty()) {
   1020     uint64_t Modulo = 0;
   1021     for (unsigned i = 0, e = GEP1VariableIndices.size(); i != e; ++i)
   1022       Modulo |= (uint64_t)GEP1VariableIndices[i].Scale;
   1023     Modulo = Modulo ^ (Modulo & (Modulo - 1));
   1024 
   1025     // We can compute the difference between the two addresses
   1026     // mod Modulo. Check whether that difference guarantees that the
   1027     // two locations do not alias.
   1028     uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1);
   1029     if (V1Size != UnknownSize && V2Size != UnknownSize &&
   1030         ModOffset >= V2Size && V1Size <= Modulo - ModOffset)
   1031       return NoAlias;
   1032   }
   1033 
   1034   // Statically, we can see that the base objects are the same, but the
   1035   // pointers have dynamic offsets which we can't resolve. And none of our
   1036   // little tricks above worked.
   1037   //
   1038   // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the
   1039   // practical effect of this is protecting TBAA in the case of dynamic
   1040   // indices into arrays of unions or malloc'd memory.
   1041   return PartialAlias;
   1042 }
   1043 
   1044 static AliasAnalysis::AliasResult
   1045 MergeAliasResults(AliasAnalysis::AliasResult A, AliasAnalysis::AliasResult B) {
   1046   // If the results agree, take it.
   1047   if (A == B)
   1048     return A;
   1049   // A mix of PartialAlias and MustAlias is PartialAlias.
   1050   if ((A == AliasAnalysis::PartialAlias && B == AliasAnalysis::MustAlias) ||
   1051       (B == AliasAnalysis::PartialAlias && A == AliasAnalysis::MustAlias))
   1052     return AliasAnalysis::PartialAlias;
   1053   // Otherwise, we don't know anything.
   1054   return AliasAnalysis::MayAlias;
   1055 }
   1056 
   1057 /// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select
   1058 /// instruction against another.
   1059 AliasAnalysis::AliasResult
   1060 BasicAliasAnalysis::aliasSelect(const SelectInst *SI, uint64_t SISize,
   1061                                 const MDNode *SITBAAInfo,
   1062                                 const Value *V2, uint64_t V2Size,
   1063                                 const MDNode *V2TBAAInfo) {
   1064   // If the values are Selects with the same condition, we can do a more precise
   1065   // check: just check for aliases between the values on corresponding arms.
   1066   if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
   1067     if (SI->getCondition() == SI2->getCondition()) {
   1068       AliasResult Alias =
   1069         aliasCheck(SI->getTrueValue(), SISize, SITBAAInfo,
   1070                    SI2->getTrueValue(), V2Size, V2TBAAInfo);
   1071       if (Alias == MayAlias)
   1072         return MayAlias;
   1073       AliasResult ThisAlias =
   1074         aliasCheck(SI->getFalseValue(), SISize, SITBAAInfo,
   1075                    SI2->getFalseValue(), V2Size, V2TBAAInfo);
   1076       return MergeAliasResults(ThisAlias, Alias);
   1077     }
   1078 
   1079   // If both arms of the Select node NoAlias or MustAlias V2, then returns
   1080   // NoAlias / MustAlias. Otherwise, returns MayAlias.
   1081   AliasResult Alias =
   1082     aliasCheck(V2, V2Size, V2TBAAInfo, SI->getTrueValue(), SISize, SITBAAInfo);
   1083   if (Alias == MayAlias)
   1084     return MayAlias;
   1085 
   1086   AliasResult ThisAlias =
   1087     aliasCheck(V2, V2Size, V2TBAAInfo, SI->getFalseValue(), SISize, SITBAAInfo);
   1088   return MergeAliasResults(ThisAlias, Alias);
   1089 }
   1090 
   1091 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
   1092 // against another.
   1093 AliasAnalysis::AliasResult
   1094 BasicAliasAnalysis::aliasPHI(const PHINode *PN, uint64_t PNSize,
   1095                              const MDNode *PNTBAAInfo,
   1096                              const Value *V2, uint64_t V2Size,
   1097                              const MDNode *V2TBAAInfo) {
   1098   // If the values are PHIs in the same block, we can do a more precise
   1099   // as well as efficient check: just check for aliases between the values
   1100   // on corresponding edges.
   1101   if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
   1102     if (PN2->getParent() == PN->getParent()) {
   1103       LocPair Locs(Location(PN, PNSize, PNTBAAInfo),
   1104                    Location(V2, V2Size, V2TBAAInfo));
   1105       if (PN > V2)
   1106         std::swap(Locs.first, Locs.second);
   1107       // Analyse the PHIs' inputs under the assumption that the PHIs are
   1108       // NoAlias.
   1109       // If the PHIs are May/MustAlias there must be (recursively) an input
   1110       // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
   1111       // there must be an operation on the PHIs within the PHIs' value cycle
   1112       // that causes a MayAlias.
   1113       // Pretend the phis do not alias.
   1114       AliasResult Alias = NoAlias;
   1115       assert(AliasCache.count(Locs) &&
   1116              "There must exist an entry for the phi node");
   1117       AliasResult OrigAliasResult = AliasCache[Locs];
   1118       AliasCache[Locs] = NoAlias;
   1119 
   1120       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
   1121         AliasResult ThisAlias =
   1122           aliasCheck(PN->getIncomingValue(i), PNSize, PNTBAAInfo,
   1123                      PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
   1124                      V2Size, V2TBAAInfo);
   1125         Alias = MergeAliasResults(ThisAlias, Alias);
   1126         if (Alias == MayAlias)
   1127           break;
   1128       }
   1129 
   1130       // Reset if speculation failed.
   1131       if (Alias != NoAlias)
   1132         AliasCache[Locs] = OrigAliasResult;
   1133 
   1134       return Alias;
   1135     }
   1136 
   1137   SmallPtrSet<Value*, 4> UniqueSrc;
   1138   SmallVector<Value*, 4> V1Srcs;
   1139   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
   1140     Value *PV1 = PN->getIncomingValue(i);
   1141     if (isa<PHINode>(PV1))
   1142       // If any of the source itself is a PHI, return MayAlias conservatively
   1143       // to avoid compile time explosion. The worst possible case is if both
   1144       // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
   1145       // and 'n' are the number of PHI sources.
   1146       return MayAlias;
   1147     if (UniqueSrc.insert(PV1))
   1148       V1Srcs.push_back(PV1);
   1149   }
   1150 
   1151   AliasResult Alias = aliasCheck(V2, V2Size, V2TBAAInfo,
   1152                                  V1Srcs[0], PNSize, PNTBAAInfo);
   1153   // Early exit if the check of the first PHI source against V2 is MayAlias.
   1154   // Other results are not possible.
   1155   if (Alias == MayAlias)
   1156     return MayAlias;
   1157 
   1158   // If all sources of the PHI node NoAlias or MustAlias V2, then returns
   1159   // NoAlias / MustAlias. Otherwise, returns MayAlias.
   1160   for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
   1161     Value *V = V1Srcs[i];
   1162 
   1163     AliasResult ThisAlias = aliasCheck(V2, V2Size, V2TBAAInfo,
   1164                                        V, PNSize, PNTBAAInfo);
   1165     Alias = MergeAliasResults(ThisAlias, Alias);
   1166     if (Alias == MayAlias)
   1167       break;
   1168   }
   1169 
   1170   return Alias;
   1171 }
   1172 
   1173 // aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
   1174 // such as array references.
   1175 //
   1176 AliasAnalysis::AliasResult
   1177 BasicAliasAnalysis::aliasCheck(const Value *V1, uint64_t V1Size,
   1178                                const MDNode *V1TBAAInfo,
   1179                                const Value *V2, uint64_t V2Size,
   1180                                const MDNode *V2TBAAInfo) {
   1181   // If either of the memory references is empty, it doesn't matter what the
   1182   // pointer values are.
   1183   if (V1Size == 0 || V2Size == 0)
   1184     return NoAlias;
   1185 
   1186   // Strip off any casts if they exist.
   1187   V1 = V1->stripPointerCasts();
   1188   V2 = V2->stripPointerCasts();
   1189 
   1190   // Are we checking for alias of the same value?
   1191   if (V1 == V2) return MustAlias;
   1192 
   1193   if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
   1194     return NoAlias;  // Scalars cannot alias each other
   1195 
   1196   // Figure out what objects these things are pointing to if we can.
   1197   const Value *O1 = GetUnderlyingObject(V1, TD);
   1198   const Value *O2 = GetUnderlyingObject(V2, TD);
   1199 
   1200   // Null values in the default address space don't point to any object, so they
   1201   // don't alias any other pointer.
   1202   if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
   1203     if (CPN->getType()->getAddressSpace() == 0)
   1204       return NoAlias;
   1205   if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
   1206     if (CPN->getType()->getAddressSpace() == 0)
   1207       return NoAlias;
   1208 
   1209   if (O1 != O2) {
   1210     // If V1/V2 point to two different objects we know that we have no alias.
   1211     if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
   1212       return NoAlias;
   1213 
   1214     // Constant pointers can't alias with non-const isIdentifiedObject objects.
   1215     if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
   1216         (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
   1217       return NoAlias;
   1218 
   1219     // Function arguments can't alias with things that are known to be
   1220     // unambigously identified at the function level.
   1221     if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
   1222         (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
   1223       return NoAlias;
   1224 
   1225     // Most objects can't alias null.
   1226     if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
   1227         (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
   1228       return NoAlias;
   1229 
   1230     // If one pointer is the result of a call/invoke or load and the other is a
   1231     // non-escaping local object within the same function, then we know the
   1232     // object couldn't escape to a point where the call could return it.
   1233     //
   1234     // Note that if the pointers are in different functions, there are a
   1235     // variety of complications. A call with a nocapture argument may still
   1236     // temporary store the nocapture argument's value in a temporary memory
   1237     // location if that memory location doesn't escape. Or it may pass a
   1238     // nocapture value to other functions as long as they don't capture it.
   1239     if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
   1240       return NoAlias;
   1241     if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
   1242       return NoAlias;
   1243   }
   1244 
   1245   // If the size of one access is larger than the entire object on the other
   1246   // side, then we know such behavior is undefined and can assume no alias.
   1247   if (TD)
   1248     if ((V1Size != UnknownSize && isObjectSmallerThan(O2, V1Size, *TD, *TLI)) ||
   1249         (V2Size != UnknownSize && isObjectSmallerThan(O1, V2Size, *TD, *TLI)))
   1250       return NoAlias;
   1251 
   1252   // Check the cache before climbing up use-def chains. This also terminates
   1253   // otherwise infinitely recursive queries.
   1254   LocPair Locs(Location(V1, V1Size, V1TBAAInfo),
   1255                Location(V2, V2Size, V2TBAAInfo));
   1256   if (V1 > V2)
   1257     std::swap(Locs.first, Locs.second);
   1258   std::pair<AliasCacheTy::iterator, bool> Pair =
   1259     AliasCache.insert(std::make_pair(Locs, MayAlias));
   1260   if (!Pair.second)
   1261     return Pair.first->second;
   1262 
   1263   // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
   1264   // GEP can't simplify, we don't even look at the PHI cases.
   1265   if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
   1266     std::swap(V1, V2);
   1267     std::swap(V1Size, V2Size);
   1268     std::swap(O1, O2);
   1269     std::swap(V1TBAAInfo, V2TBAAInfo);
   1270   }
   1271   if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
   1272     AliasResult Result = aliasGEP(GV1, V1Size, V1TBAAInfo, V2, V2Size, V2TBAAInfo, O1, O2);
   1273     if (Result != MayAlias) return AliasCache[Locs] = Result;
   1274   }
   1275 
   1276   if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
   1277     std::swap(V1, V2);
   1278     std::swap(V1Size, V2Size);
   1279     std::swap(V1TBAAInfo, V2TBAAInfo);
   1280   }
   1281   if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
   1282     AliasResult Result = aliasPHI(PN, V1Size, V1TBAAInfo,
   1283                                   V2, V2Size, V2TBAAInfo);
   1284     if (Result != MayAlias) return AliasCache[Locs] = Result;
   1285   }
   1286 
   1287   if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
   1288     std::swap(V1, V2);
   1289     std::swap(V1Size, V2Size);
   1290     std::swap(V1TBAAInfo, V2TBAAInfo);
   1291   }
   1292   if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
   1293     AliasResult Result = aliasSelect(S1, V1Size, V1TBAAInfo,
   1294                                      V2, V2Size, V2TBAAInfo);
   1295     if (Result != MayAlias) return AliasCache[Locs] = Result;
   1296   }
   1297 
   1298   // If both pointers are pointing into the same object and one of them
   1299   // accesses is accessing the entire object, then the accesses must
   1300   // overlap in some way.
   1301   if (TD && O1 == O2)
   1302     if ((V1Size != UnknownSize && isObjectSize(O1, V1Size, *TD, *TLI)) ||
   1303         (V2Size != UnknownSize && isObjectSize(O2, V2Size, *TD, *TLI)))
   1304       return AliasCache[Locs] = PartialAlias;
   1305 
   1306   AliasResult Result =
   1307     AliasAnalysis::alias(Location(V1, V1Size, V1TBAAInfo),
   1308                          Location(V2, V2Size, V2TBAAInfo));
   1309   return AliasCache[Locs] = Result;
   1310 }
   1311