MemoryDependenceAnalysis.cpp

//===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation  --*- C++ -*-===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements an analysis that determines, for a given memory
// operation, what preceding memory operations it depends on.  It builds on 
// alias analysis information, and tries to provide a lazy, caching interface to
// a common kind of alias information query.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "memdep"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Function.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/MemoryBuiltins.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/PredIteratorCache.h"
#include "llvm/Support/Debug.h"
using namespace llvm;

STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");

STATISTIC(NumCacheNonLocalPtr,
          "Number of fully cached non-local ptr responses");
STATISTIC(NumCacheDirtyNonLocalPtr,
          "Number of cached, but dirty, non-local ptr responses");
STATISTIC(NumUncacheNonLocalPtr,
          "Number of uncached non-local ptr responses");
STATISTIC(NumCacheCompleteNonLocalPtr,
          "Number of block queries that were completely cached");

char MemoryDependenceAnalysis::ID = 0;
  
// Register this pass...
static RegisterPass<MemoryDependenceAnalysis> X("memdep",
                                     "Memory Dependence Analysis", false, true);

MemoryDependenceAnalysis::MemoryDependenceAnalysis()
: FunctionPass(&ID), PredCache(0) {
}
MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
}

/// Clean up memory in between runs
void MemoryDependenceAnalysis::releaseMemory() {
  LocalDeps.clear();
  NonLocalDeps.clear();
  NonLocalPointerDeps.clear();
  ReverseLocalDeps.clear();
  ReverseNonLocalDeps.clear();
  ReverseNonLocalPtrDeps.clear();
  PredCache->clear();
}



/// getAnalysisUsage - Does not modify anything.  It uses Alias Analysis.
///
void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
  AU.setPreservesAll();
  AU.addRequiredTransitive<AliasAnalysis>();
}

bool MemoryDependenceAnalysis::runOnFunction(Function &) {
  AA = &getAnalysis<AliasAnalysis>();
  if (PredCache == 0)
    PredCache.reset(new PredIteratorCache());
  return false;
}

/// RemoveFromReverseMap - This is a helper function that removes Val from
/// 'Inst's set in ReverseMap.  If the set becomes empty, remove Inst's entry.
template <typename KeyTy>
static void RemoveFromReverseMap(DenseMap<Instruction*, 
                                 SmallPtrSet<KeyTy, 4> > &ReverseMap,
                                 Instruction *Inst, KeyTy Val) {
  typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
  InstIt = ReverseMap.find(Inst);
  assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
  bool Found = InstIt->second.erase(Val);
  assert(Found && "Invalid reverse map!"); Found=Found;
  if (InstIt->second.empty())
    ReverseMap.erase(InstIt);
}


/// getCallSiteDependencyFrom - Private helper for finding the local
/// dependencies of a call site.
MemDepResult MemoryDependenceAnalysis::
getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
                          BasicBlock::iterator ScanIt, BasicBlock *BB) {
  // Walk backwards through the block, looking for dependencies
  while (ScanIt != BB->begin()) {
    Instruction *Inst = --ScanIt;
    
    // If this inst is a memory op, get the pointer it accessed
    Value *Pointer = 0;
    uint64_t PointerSize = 0;
    if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
      Pointer = S->getPointerOperand();
      PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType());
    } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
      Pointer = V->getOperand(0);
      PointerSize = AA->getTypeStoreSize(V->getType());
    } else if (isFreeCall(Inst)) {
      Pointer = Inst->getOperand(1);
      // calls to free() erase the entire structure
      PointerSize = ~0ULL;
    } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
      // Debug intrinsics don't cause dependences.
      if (isa<DbgInfoIntrinsic>(Inst)) continue;
      CallSite InstCS = CallSite::get(Inst);
      // If these two calls do not interfere, look past it.
      switch (AA->getModRefInfo(CS, InstCS)) {
      case AliasAnalysis::NoModRef:
        // If the two calls don't interact (e.g. InstCS is readnone) keep
        // scanning.
        continue;
      case AliasAnalysis::Ref:
        // If the two calls read the same memory locations and CS is a readonly
        // function, then we have two cases: 1) the calls may not interfere with
        // each other at all.  2) the calls may produce the same value.  In case
        // #1 we want to ignore the values, in case #2, we want to return Inst
        // as a Def dependence.  This allows us to CSE in cases like:
        //   X = strlen(P);
        //    memchr(...);
        //   Y = strlen(P);  // Y = X
        if (isReadOnlyCall) {
          if (CS.getCalledFunction() != 0 &&
              CS.getCalledFunction() == InstCS.getCalledFunction())
            return MemDepResult::getDef(Inst);
          // Ignore unrelated read/read call dependences.
          continue;
        }
        // FALL THROUGH
      default:
        return MemDepResult::getClobber(Inst);
      }
    } else {
      // Non-memory instruction.
      continue;
    }
    
    if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef)
      return MemDepResult::getClobber(Inst);
  }
  
  // No dependence found.  If this is the entry block of the function, it is a
  // clobber, otherwise it is non-local.
  if (BB != &BB->getParent()->getEntryBlock())
    return MemDepResult::getNonLocal();
  return MemDepResult::getClobber(ScanIt);
}

/// getPointerDependencyFrom - Return the instruction on which a memory
/// location depends.  If isLoad is true, this routine ignore may-aliases with
/// read-only operations.
MemDepResult MemoryDependenceAnalysis::
getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, 
                         BasicBlock::iterator ScanIt, BasicBlock *BB) {

  Value *invariantTag = 0;

  // Walk backwards through the basic block, looking for dependencies.
  while (ScanIt != BB->begin()) {
    Instruction *Inst = --ScanIt;

    // If we're in an invariant region, no dependencies can be found before
    // we pass an invariant-begin marker.
    if (invariantTag == Inst) {
      invariantTag = 0;
      continue;
    } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
      // If we pass an invariant-end marker, then we've just entered an
      // invariant region and can start ignoring dependencies.
      if (II->getIntrinsicID() == Intrinsic::invariant_end) {
        uint64_t invariantSize = ~0ULL;
        if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getOperand(2)))
          invariantSize = CI->getZExtValue();
        
        AliasAnalysis::AliasResult R =
          AA->alias(II->getOperand(3), invariantSize, MemPtr, MemSize);
        if (R == AliasAnalysis::MustAlias) {
          invariantTag = II->getOperand(1);
          continue;
        }
      
      // If we reach a lifetime begin or end marker, then the query ends here
      // because the value is undefined.
      } else if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
                 II->getIntrinsicID() == Intrinsic::lifetime_end) {
        uint64_t invariantSize = ~0ULL;
        if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getOperand(1)))
          invariantSize = CI->getZExtValue();

        AliasAnalysis::AliasResult R =
          AA->alias(II->getOperand(2), invariantSize, MemPtr, MemSize);
        if (R == AliasAnalysis::MustAlias)
          return MemDepResult::getDef(II);
      }
    }

    // If we're querying on a load and we're in an invariant region, we're done
    // at this point. Nothing a load depends on can live in an invariant region.
    if (isLoad && invariantTag) continue;

    // Debug intrinsics don't cause dependences.
    if (isa<DbgInfoIntrinsic>(Inst)) continue;

    // Values depend on loads if the pointers are must aliased.  This means that
    // a load depends on another must aliased load from the same value.
    if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
      Value *Pointer = LI->getPointerOperand();
      uint64_t PointerSize = AA->getTypeStoreSize(LI->getType());
      
      // If we found a pointer, check if it could be the same as our pointer.
      AliasAnalysis::AliasResult R =
        AA->alias(Pointer, PointerSize, MemPtr, MemSize);
      if (R == AliasAnalysis::NoAlias)
        continue;
      
      // May-alias loads don't depend on each other without a dependence.
      if (isLoad && R == AliasAnalysis::MayAlias)
        continue;
      // Stores depend on may and must aliased loads, loads depend on must-alias
      // loads.
      return MemDepResult::getDef(Inst);
    }
    
    if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
      // There can't be stores to the value we care about inside an 
      // invariant region.
      if (invariantTag) continue;
      
      // If alias analysis can tell that this store is guaranteed to not modify
      // the query pointer, ignore it.  Use getModRefInfo to handle cases where
      // the query pointer points to constant memory etc.
      if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef)
        continue;

      // Ok, this store might clobber the query pointer.  Check to see if it is
      // a must alias: in this case, we want to return this as a def.
      Value *Pointer = SI->getPointerOperand();
      uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
      
      // If we found a pointer, check if it could be the same as our pointer.
      AliasAnalysis::AliasResult R =
        AA->alias(Pointer, PointerSize, MemPtr, MemSize);
      
      if (R == AliasAnalysis::NoAlias)
        continue;
      if (R == AliasAnalysis::MayAlias)
        return MemDepResult::getClobber(Inst);
      return MemDepResult::getDef(Inst);
    }

    // If this is an allocation, and if we know that the accessed pointer is to
    // the allocation, return Def.  This means that there is no dependence and
    // the access can be optimized based on that.  For example, a load could
    // turn into undef.
    // Note: Only determine this to be a malloc if Inst is the malloc call, not
    // a subsequent bitcast of the malloc call result.  There can be stores to
    // the malloced memory between the malloc call and its bitcast uses, and we
    // need to continue scanning until the malloc call.
    if (isa<AllocaInst>(Inst) || extractMallocCall(Inst)) {
      Value *AccessPtr = MemPtr->getUnderlyingObject();
      
      if (AccessPtr == Inst ||
          AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
        return MemDepResult::getDef(Inst);
      continue;
    }

    // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
    switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
    case AliasAnalysis::NoModRef:
      // If the call has no effect on the queried pointer, just ignore it.
      continue;
    case AliasAnalysis::Mod:
      // If we're in an invariant region, we can ignore calls that ONLY
      // modify the pointer.
      if (invariantTag) continue;
      return MemDepResult::getClobber(Inst);
    case AliasAnalysis::Ref:
      // If the call is known to never store to the pointer, and if this is a
      // load query, we can safely ignore it (scan past it).
      if (isLoad)
        continue;
    default:
      // Otherwise, there is a potential dependence.  Return a clobber.
      return MemDepResult::getClobber(Inst);
    }
  }
  
  // No dependence found.  If this is the entry block of the function, it is a
  // clobber, otherwise it is non-local.
  if (BB != &BB->getParent()->getEntryBlock())
    return MemDepResult::getNonLocal();
  return MemDepResult::getClobber(ScanIt);
}

/// getDependency - Return the instruction on which a memory operation
/// depends.
MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
  Instruction *ScanPos = QueryInst;
  
  // Check for a cached result
  MemDepResult &LocalCache = LocalDeps[QueryInst];
  
  // If the cached entry is non-dirty, just return it.  Note that this depends
  // on MemDepResult's default constructing to 'dirty'.
  if (!LocalCache.isDirty())
    return LocalCache;
    
  // Otherwise, if we have a dirty entry, we know we can start the scan at that
  // instruction, which may save us some work.
  if (Instruction *Inst = LocalCache.getInst()) {
    ScanPos = Inst;
   
    RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
  }
  
  BasicBlock *QueryParent = QueryInst->getParent();
  
  Value *MemPtr = 0;
  uint64_t MemSize = 0;
  
  // Do the scan.
  if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
    // No dependence found.  If this is the entry block of the function, it is a
    // clobber, otherwise it is non-local.
    if (QueryParent != &QueryParent->getParent()->getEntryBlock())
      LocalCache = MemDepResult::getNonLocal();
    else
      LocalCache = MemDepResult::getClobber(QueryInst);
  } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
    // If this is a volatile store, don't mess around with it.  Just return the
    // previous instruction as a clobber.
    if (SI->isVolatile())
      LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
    else {
      MemPtr = SI->getPointerOperand();
      MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
    }
  } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
    // If this is a volatile load, don't mess around with it.  Just return the
    // previous instruction as a clobber.
    if (LI->isVolatile())
      LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
    else {
      MemPtr = LI->getPointerOperand();
      MemSize = AA->getTypeStoreSize(LI->getType());
    }
  } else if (isFreeCall(QueryInst)) {
    MemPtr = QueryInst->getOperand(1);
    // calls to free() erase the entire structure, not just a field.
    MemSize = ~0UL;
  } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
    int IntrinsicID = 0;  // Intrinsic IDs start at 1.
    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
      IntrinsicID = II->getIntrinsicID();

    switch (IntrinsicID) {
      case Intrinsic::lifetime_start:
      case Intrinsic::lifetime_end:
      case Intrinsic::invariant_start:
        MemPtr = QueryInst->getOperand(2);
        MemSize = cast<ConstantInt>(QueryInst->getOperand(1))->getZExtValue();
        break;
      case Intrinsic::invariant_end:
        MemPtr = QueryInst->getOperand(3);
        MemSize = cast<ConstantInt>(QueryInst->getOperand(2))->getZExtValue();
        break;
      default:
        CallSite QueryCS = CallSite::get(QueryInst);
        bool isReadOnly = AA->onlyReadsMemory(QueryCS);
        LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
                                               QueryParent);
    }
  } else {
    // Non-memory instruction.
    LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
  }
  
  // If we need to do a pointer scan, make it happen.
  if (MemPtr) {
    bool isLoad = !QueryInst->mayWriteToMemory();
    if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) {
      isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end;
    }
    LocalCache = getPointerDependencyFrom(MemPtr, MemSize, isLoad, ScanPos,
                                          QueryParent);
  }
  
  // Remember the result!
  if (Instruction *I = LocalCache.getInst())
    ReverseLocalDeps[I].insert(QueryInst);
  
  return LocalCache;
}

#ifndef NDEBUG
/// AssertSorted - This method is used when -debug is specified to verify that
/// cache arrays are properly kept sorted.
static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
                         int Count = -1) {
  if (Count == -1) Count = Cache.size();
  if (Count == 0) return;

  for (unsigned i = 1; i != unsigned(Count); ++i)
    assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!");
}
#endif

/// getNonLocalCallDependency - Perform a full dependency query for the
/// specified call, returning the set of blocks that the value is
/// potentially live across.  The returned set of results will include a
/// "NonLocal" result for all blocks where the value is live across.
///
/// This method assumes the instruction returns a "NonLocal" dependency
/// within its own block.
///
/// This returns a reference to an internal data structure that may be
/// invalidated on the next non-local query or when an instruction is
/// removed.  Clients must copy this data if they want it around longer than
/// that.
const MemoryDependenceAnalysis::NonLocalDepInfo &
MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
  assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
 "getNonLocalCallDependency should only be used on calls with non-local deps!");
  PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
  NonLocalDepInfo &Cache = CacheP.first;

  /// DirtyBlocks - This is the set of blocks that need to be recomputed.  In
  /// the cached case, this can happen due to instructions being deleted etc. In
  /// the uncached case, this starts out as the set of predecessors we care
  /// about.
  SmallVector<BasicBlock*, 32> DirtyBlocks;
  
  if (!Cache.empty()) {
    // Okay, we have a cache entry.  If we know it is not dirty, just return it
    // with no computation.
    if (!CacheP.second) {
      NumCacheNonLocal++;
      return Cache;
    }
    
    // If we already have a partially computed set of results, scan them to
    // determine what is dirty, seeding our initial DirtyBlocks worklist.
    for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
       I != E; ++I)
      if (I->second.isDirty())
        DirtyBlocks.push_back(I->first);
    
    // Sort the cache so that we can do fast binary search lookups below.
    std::sort(Cache.begin(), Cache.end());
    
    ++NumCacheDirtyNonLocal;
    //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
    //     << Cache.size() << " cached: " << *QueryInst;
  } else {
    // Seed DirtyBlocks with each of the preds of QueryInst's block.
    BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
    for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
      DirtyBlocks.push_back(*PI);
    NumUncacheNonLocal++;
  }
  
  // isReadonlyCall - If this is a read-only call, we can be more aggressive.
  bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);

  SmallPtrSet<BasicBlock*, 64> Visited;
  
  unsigned NumSortedEntries = Cache.size();
  DEBUG(AssertSorted(Cache));
  
  // Iterate while we still have blocks to update.
  while (!DirtyBlocks.empty()) {
    BasicBlock *DirtyBB = DirtyBlocks.back();
    DirtyBlocks.pop_back();
    
    // Already processed this block?
    if (!Visited.insert(DirtyBB))
      continue;
    
    // Do a binary search to see if we already have an entry for this block in
    // the cache set.  If so, find it.
    DEBUG(AssertSorted(Cache, NumSortedEntries));
    NonLocalDepInfo::iterator Entry = 
      std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
                       std::make_pair(DirtyBB, MemDepResult()));
    if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB)
      --Entry;
    
    MemDepResult *ExistingResult = 0;
    if (Entry != Cache.begin()+NumSortedEntries && 
        Entry->first == DirtyBB) {
      // If we already have an entry, and if it isn't already dirty, the block
      // is done.
      if (!Entry->second.isDirty())
        continue;
      
      // Otherwise, remember this slot so we can update the value.
      ExistingResult = &Entry->second;
    }
    
    // If the dirty entry has a pointer, start scanning from it so we don't have
    // to rescan the entire block.
    BasicBlock::iterator ScanPos = DirtyBB->end();
    if (ExistingResult) {
      if (Instruction *Inst = ExistingResult->getInst()) {
        ScanPos = Inst;
        // We're removing QueryInst's use of Inst.
        RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
                             QueryCS.getInstruction());
      }
    }
    
    // Find out if this block has a local dependency for QueryInst.
    MemDepResult Dep;
    
    if (ScanPos != DirtyBB->begin()) {
      Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
    } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
      // No dependence found.  If this is the entry block of the function, it is
      // a clobber, otherwise it is non-local.
      Dep = MemDepResult::getNonLocal();
    } else {
      Dep = MemDepResult::getClobber(ScanPos);
    }
    
    // If we had a dirty entry for the block, update it.  Otherwise, just add
    // a new entry.
    if (ExistingResult)
      *ExistingResult = Dep;
    else
      Cache.push_back(std::make_pair(DirtyBB, Dep));
    
    // If the block has a dependency (i.e. it isn't completely transparent to
    // the value), remember the association!
    if (!Dep.isNonLocal()) {
      // Keep the ReverseNonLocalDeps map up to date so we can efficiently
      // update this when we remove instructions.
      if (Instruction *Inst = Dep.getInst())
        ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
    } else {
    
      // If the block *is* completely transparent to the load, we need to check
      // the predecessors of this block.  Add them to our worklist.
      for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
        DirtyBlocks.push_back(*PI);
    }
  }
  
  return Cache;
}

/// getNonLocalPointerDependency - Perform a full dependency query for an
/// access to the specified (non-volatile) memory location, returning the
/// set of instructions that either define or clobber the value.
///
/// This method assumes the pointer has a "NonLocal" dependency within its
/// own block.
///
void MemoryDependenceAnalysis::
getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
                             SmallVectorImpl<NonLocalDepEntry> &Result) {
  assert(isa<PointerType>(Pointer->getType()) &&
         "Can't get pointer deps of a non-pointer!");
  Result.clear();
  
  // We know that the pointer value is live into FromBB find the def/clobbers
  // from presecessors.
  const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
  uint64_t PointeeSize = AA->getTypeStoreSize(EltTy);
  
  // This is the set of blocks we've inspected, and the pointer we consider in
  // each block.  Because of critical edges, we currently bail out if querying
  // a block with multiple different pointers.  This can happen during PHI
  // translation.
  DenseMap<BasicBlock*, Value*> Visited;
  if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB,
                                   Result, Visited, true))
    return;
  Result.clear();
  Result.push_back(std::make_pair(FromBB,
                                  MemDepResult::getClobber(FromBB->begin())));
}

/// GetNonLocalInfoForBlock - Compute the memdep value for BB with
/// Pointer/PointeeSize using either cached information in Cache or by doing a
/// lookup (which may use dirty cache info if available).  If we do a lookup,
/// add the result to the cache.
MemDepResult MemoryDependenceAnalysis::
GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
                        bool isLoad, BasicBlock *BB,
                        NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
  
  // Do a binary search to see if we already have an entry for this block in
  // the cache set.  If so, find it.
  NonLocalDepInfo::iterator Entry =
    std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
                     std::make_pair(BB, MemDepResult()));
  if (Entry != Cache->begin() && prior(Entry)->first == BB)
    --Entry;
  
  MemDepResult *ExistingResult = 0;
  if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
    ExistingResult = &Entry->second;
  
  // If we have a cached entry, and it is non-dirty, use it as the value for
  // this dependency.
  if (ExistingResult && !ExistingResult->isDirty()) {
    ++NumCacheNonLocalPtr;
    return *ExistingResult;
  }    
  
  // Otherwise, we have to scan for the value.  If we have a dirty cache
  // entry, start scanning from its position, otherwise we scan from the end
  // of the block.
  BasicBlock::iterator ScanPos = BB->end();
  if (ExistingResult && ExistingResult->getInst()) {
    assert(ExistingResult->getInst()->getParent() == BB &&
           "Instruction invalidated?");
    ++NumCacheDirtyNonLocalPtr;
    ScanPos = ExistingResult->getInst();
    
    // Eliminating the dirty entry from 'Cache', so update the reverse info.
    ValueIsLoadPair CacheKey(Pointer, isLoad);
    RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
  } else {
    ++NumUncacheNonLocalPtr;
  }
  
  // Scan the block for the dependency.
  MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, 
                                              ScanPos, BB);
  
  // If we had a dirty entry for the block, update it.  Otherwise, just add
  // a new entry.
  if (ExistingResult)
    *ExistingResult = Dep;
  else
    Cache->push_back(std::make_pair(BB, Dep));
  
  // If the block has a dependency (i.e. it isn't completely transparent to
  // the value), remember the reverse association because we just added it
  // to Cache!
  if (Dep.isNonLocal())
    return Dep;
  
  // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
  // update MemDep when we remove instructions.
  Instruction *Inst = Dep.getInst();
  assert(Inst && "Didn't depend on anything?");
  ValueIsLoadPair CacheKey(Pointer, isLoad);
  ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
  return Dep;
}

/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
/// number of elements in the array that are already properly ordered.  This is
/// optimized for the case when only a few entries are added.
static void 
SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
                         unsigned NumSortedEntries) {
  switch (Cache.size() - NumSortedEntries) {
  case 0:
    // done, no new entries.
    break;
  case 2: {
    // Two new entries, insert the last one into place.
    MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
    Cache.pop_back();
    MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
      std::upper_bound(Cache.begin(), Cache.end()-1, Val);
    Cache.insert(Entry, Val);
    // FALL THROUGH.
  }
  case 1:
    // One new entry, Just insert the new value at the appropriate position.
    if (Cache.size() != 1) {
      MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back();
      Cache.pop_back();
      MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
        std::upper_bound(Cache.begin(), Cache.end(), Val);
      Cache.insert(Entry, Val);
    }
    break;
  default:
    // Added many values, do a full scale sort.
    std::sort(Cache.begin(), Cache.end());
    break;
  }
}

/// isPHITranslatable - Return true if the specified computation is derived from
/// a PHI node in the current block and if it is simple enough for us to handle.
static bool isPHITranslatable(Instruction *Inst) {
  if (isa<PHINode>(Inst))
    return true;
  
  // We can handle bitcast of a PHI, but the PHI needs to be in the same block
  // as the bitcast.
  if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) {
    Instruction *OpI = dyn_cast<Instruction>(BC->getOperand(0));
    if (OpI == 0 || OpI->getParent() != Inst->getParent())
      return true;
    return isPHITranslatable(OpI);
  }
  
  // We can translate a GEP if all of its operands defined in this block are phi
  // translatable. 
  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
    for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) {
      Instruction *OpI = dyn_cast<Instruction>(GEP->getOperand(i));
      if (OpI == 0 || OpI->getParent() != Inst->getParent())
        continue;
      
      if (!isPHITranslatable(OpI))
        return false;
    }
    return true;
  }
  
  if (Inst->getOpcode() == Instruction::Add &&
      isa<ConstantInt>(Inst->getOperand(1))) {
    Instruction *OpI = dyn_cast<Instruction>(Inst->getOperand(0));
    if (OpI == 0 || OpI->getParent() != Inst->getParent())
      return true;
    return isPHITranslatable(OpI);
  }

  //   cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
  //   if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
  //     cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
  
  return false;
}

/// GetPHITranslatedValue - Given a computation that satisfied the
/// isPHITranslatable predicate, see if we can translate the computation into
/// the specified predecessor block.  If so, return that value.
Value *MemoryDependenceAnalysis::
GetPHITranslatedValue(Value *InVal, BasicBlock *CurBB, BasicBlock *Pred,
                      const TargetData *TD) const {  
  // If the input value is not an instruction, or if it is not defined in CurBB,
  // then we don't need to phi translate it.
  Instruction *Inst = dyn_cast<Instruction>(InVal);
  if (Inst == 0 || Inst->getParent() != CurBB)
    return InVal;
  
  if (PHINode *PN = dyn_cast<PHINode>(Inst))
    return PN->getIncomingValueForBlock(Pred);
  
  // Handle bitcast of PHI.
  if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) {
    // PHI translate the input operand.
    Value *PHIIn = GetPHITranslatedValue(BC->getOperand(0), CurBB, Pred, TD);
    if (PHIIn == 0) return 0;
    
    // Constants are trivial to phi translate.
    if (Constant *C = dyn_cast<Constant>(PHIIn))
      return ConstantExpr::getBitCast(C, BC->getType());
    
    // Otherwise we have to see if a bitcasted version of the incoming pointer
    // is available.  If so, we can use it, otherwise we have to fail.
    for (Value::use_iterator UI = PHIIn->use_begin(), E = PHIIn->use_end();
         UI != E; ++UI) {
      if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI))
        if (BCI->getType() == BC->getType())
          return BCI;
    }
    return 0;
  }

  // Handle getelementptr with at least one PHI translatable operand.
  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
    SmallVector<Value*, 8> GEPOps;
    BasicBlock *CurBB = GEP->getParent();
    for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) {
      Value *GEPOp = GEP->getOperand(i);
      // No PHI translation is needed of operands whose values are live in to
      // the predecessor block.
      if (!isa<Instruction>(GEPOp) ||
          cast<Instruction>(GEPOp)->getParent() != CurBB) {
        GEPOps.push_back(GEPOp);
        continue;
      }
      
      // If the operand is a phi node, do phi translation.
      Value *InOp = GetPHITranslatedValue(GEPOp, CurBB, Pred, TD);
      if (InOp == 0) return 0;
      
      GEPOps.push_back(InOp);
    }
    
    // Simplify the GEP to handle 'gep x, 0' -> x etc.
    if (Value *V = SimplifyGEPInst(&GEPOps[0], GEPOps.size(), TD))
      return V;

    // Scan to see if we have this GEP available.
    Value *APHIOp = GEPOps[0];
    for (Value::use_iterator UI = APHIOp->use_begin(), E = APHIOp->use_end();
         UI != E; ++UI) {
      if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI))
        if (GEPI->getType() == GEP->getType() &&
            GEPI->getNumOperands() == GEPOps.size() &&
            GEPI->getParent()->getParent() == CurBB->getParent()) {
          bool Mismatch = false;
          for (unsigned i = 0, e = GEPOps.size(); i != e; ++i)
            if (GEPI->getOperand(i) != GEPOps[i]) {
              Mismatch = true;
              break;
            }
          if (!Mismatch)
            return GEPI;
        }
    }
    return 0;
  }
  
  // Handle add with a constant RHS.
  if (Inst->getOpcode() == Instruction::Add &&
      isa<ConstantInt>(Inst->getOperand(1))) {
    // PHI translate the LHS.
    Value *LHS;
    Constant *RHS = cast<ConstantInt>(Inst->getOperand(1));
    Instruction *OpI = dyn_cast<Instruction>(Inst->getOperand(0));
    bool isNSW = cast<BinaryOperator>(Inst)->hasNoSignedWrap();
    bool isNUW = cast<BinaryOperator>(Inst)->hasNoUnsignedWrap();
    
    if (OpI == 0 || OpI->getParent() != Inst->getParent())
      LHS = Inst->getOperand(0);
    else {
      LHS = GetPHITranslatedValue(Inst->getOperand(0), CurBB, Pred, TD);
      if (LHS == 0)
        return 0;
    }
    
    // If the PHI translated LHS is an add of a constant, fold the immediates.
    if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(LHS))
      if (BOp->getOpcode() == Instruction::Add)
        if (ConstantInt *CI = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
          LHS = BOp->getOperand(0);
          RHS = ConstantExpr::getAdd(RHS, CI);
          isNSW = isNUW = false;
        }
    
    // See if the add simplifies away.
    if (Value *Res = SimplifyAddInst(LHS, RHS, isNSW, isNUW, TD))
      return Res;
    
    // Otherwise, see if we have this add available somewhere.
    for (Value::use_iterator UI = LHS->use_begin(), E = LHS->use_end();
         UI != E; ++UI) {
      if (BinaryOperator *BO = dyn_cast<BinaryOperator>(*UI))
        if (BO->getOperand(0) == LHS && BO->getOperand(1) == RHS &&
            BO->getParent()->getParent() == CurBB->getParent())
          return BO;
    }
    
    return 0;
  }
  
  return 0;
}

/// GetAvailablePHITranslatePointer - Return the value computed by
/// PHITranslatePointer if it dominates PredBB, otherwise return null.
Value *MemoryDependenceAnalysis::
GetAvailablePHITranslatedValue(Value *V,
                               BasicBlock *CurBB, BasicBlock *PredBB,
                               const TargetData *TD,
                               const DominatorTree &DT) const {
  // See if PHI translation succeeds.
  V = GetPHITranslatedValue(V, CurBB, PredBB, TD);
  if (V == 0) return 0;
  
  // Make sure the value is live in the predecessor.
  if (Instruction *Inst = dyn_cast_or_null<Instruction>(V))
    if (!DT.dominates(Inst->getParent(), PredBB))
      return 0;
  return V;
}


/// InsertPHITranslatedPointer - Insert a computation of the PHI translated
/// version of 'V' for the edge PredBB->CurBB into the end of the PredBB
/// block.  All newly created instructions are added to the NewInsts list.
///
Value *MemoryDependenceAnalysis::
InsertPHITranslatedPointer(Value *InVal, BasicBlock *CurBB,
                           BasicBlock *PredBB, const TargetData *TD,
                           const DominatorTree &DT,
                           SmallVectorImpl<Instruction*> &NewInsts) const {
  // See if we have a version of this value already available and dominating
  // PredBB.  If so, there is no need to insert a new copy.
  if (Value *Res = GetAvailablePHITranslatedValue(InVal, CurBB, PredBB, TD, DT))
    return Res;
  
  // If we don't have an available version of this value, it must be an
  // instruction.
  Instruction *Inst = cast<Instruction>(InVal);
  
  // Handle bitcast of PHI translatable value.
  if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) {
    Value *OpVal = InsertPHITranslatedPointer(BC->getOperand(0),
                                              CurBB, PredBB, TD, DT, NewInsts);
    if (OpVal == 0) return 0;
      
    // Otherwise insert a bitcast at the end of PredBB.
    BitCastInst *New = new BitCastInst(OpVal, InVal->getType(),
                                       InVal->getName()+".phi.trans.insert",
                                       PredBB->getTerminator());
    NewInsts.push_back(New);
    return New;
  }
  
  // Handle getelementptr with at least one PHI operand.
  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
    SmallVector<Value*, 8> GEPOps;
    BasicBlock *CurBB = GEP->getParent();
    for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) {
      Value *OpVal = InsertPHITranslatedPointer(GEP->getOperand(i),
                                                CurBB, PredBB, TD, DT, NewInsts);
      if (OpVal == 0) return 0;
      GEPOps.push_back(OpVal);
    }
    
    GetElementPtrInst *Result = 
      GetElementPtrInst::Create(GEPOps[0], GEPOps.begin()+1, GEPOps.end(),
                                InVal->getName()+".phi.trans.insert",
                                PredBB->getTerminator());
    Result->setIsInBounds(GEP->isInBounds());
    NewInsts.push_back(Result);
    return Result;
  }
  
#if 0
  // FIXME: This code works, but it is unclear that we actually want to insert
  // a big chain of computation in order to make a value available in a block.
  // This needs to be evaluated carefully to consider its cost trade offs.
  
  // Handle add with a constant RHS.
  if (Inst->getOpcode() == Instruction::Add &&
      isa<ConstantInt>(Inst->getOperand(1))) {
    // PHI translate the LHS.
    Value *OpVal = InsertPHITranslatedPointer(Inst->getOperand(0),
                                              CurBB, PredBB, TD, DT, NewInsts);
    if (OpVal == 0) return 0;
    
    BinaryOperator *Res = BinaryOperator::CreateAdd(OpVal, Inst->getOperand(1),
                                           InVal->getName()+".phi.trans.insert",
                                                    PredBB->getTerminator());
    Res->setHasNoSignedWrap(cast<BinaryOperator>(Inst)->hasNoSignedWrap());
    Res->setHasNoUnsignedWrap(cast<BinaryOperator>(Inst)->hasNoUnsignedWrap());
    NewInsts.push_back(Res);
    return Res;
  }
#endif
  
  return 0;
}

/// getNonLocalPointerDepFromBB - Perform a dependency query based on
/// pointer/pointeesize starting at the end of StartBB.  Add any clobber/def
/// results to the results vector and keep track of which blocks are visited in
/// 'Visited'.
///
/// This has special behavior for the first block queries (when SkipFirstBlock
/// is true).  In this special case, it ignores the contents of the specified
/// block and starts returning dependence info for its predecessors.
///
/// This function returns false on success, or true to indicate that it could
/// not compute dependence information for some reason.  This should be treated
/// as a clobber dependence on the first instruction in the predecessor block.
bool MemoryDependenceAnalysis::
getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
                            bool isLoad, BasicBlock *StartBB,
                            SmallVectorImpl<NonLocalDepEntry> &Result,
                            DenseMap<BasicBlock*, Value*> &Visited,
                            bool SkipFirstBlock) {
  
  // Look up the cached info for Pointer.
  ValueIsLoadPair CacheKey(Pointer, isLoad);