GVN.cpp

  // If the load and store are to the exact same address, they should have been
  // a must alias.  AA must have gotten confused.
  // FIXME: Study to see if/when this happens.
  if (LoadOffset == StoreOffset) {
#if 0
    dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
    << "Base       = " << *StoreBase << "\n"
    << "Store Ptr  = " << *WritePtr << "\n"
    << "Store Offs = " << StoreOffset << "\n"
    << "Load Ptr   = " << *LoadPtr << "\n";
    abort();
#endif
    return -1;
  }
  
  // If the load and store don't overlap at all, the store doesn't provide
  // anything to the load.  In this case, they really don't alias at all, AA
  // must have gotten confused.
  // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
  // remove this check, as it is duplicated with what we have below.
  uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
  
  if ((WriteSizeInBits & 7) | (LoadSize & 7))
    return -1;
  uint64_t StoreSize = WriteSizeInBits >> 3;  // Convert to bytes.
  LoadSize >>= 3;
  
  
  bool isAAFailure = false;
  if (StoreOffset < LoadOffset) {
    isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
  } else {
    isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
  }
  if (isAAFailure) {
#if 0
    dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
    << "Base       = " << *StoreBase << "\n"
    << "Store Ptr  = " << *WritePtr << "\n"
    << "Store Offs = " << StoreOffset << "\n"
    << "Load Ptr   = " << *LoadPtr << "\n";
    abort();
#endif
    return -1;
  }
  
  // If the Load isn't completely contained within the stored bits, we don't
  // have all the bits to feed it.  We could do something crazy in the future
  // (issue a smaller load then merge the bits in) but this seems unlikely to be
  // valuable.
  if (StoreOffset > LoadOffset ||
      StoreOffset+StoreSize < LoadOffset+LoadSize)
    return -1;
  
  // Okay, we can do this transformation.  Return the number of bytes into the
  // store that the load is.
  return LoadOffset-StoreOffset;
}  

/// AnalyzeLoadFromClobberingStore - This function is called when we have a
/// memdep query of a load that ends up being a clobbering store.
static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
                                          StoreInst *DepSI,
                                          const TargetData &TD) {
  // Cannot handle reading from store of first-class aggregate yet.
  if (isa<StructType>(DepSI->getOperand(0)->getType()) ||
      isa<ArrayType>(DepSI->getOperand(0)->getType()))
    return -1;

  Value *StorePtr = DepSI->getPointerOperand();
  uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType());
  return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
                                        StorePtr, StoreSize, TD);
}

static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
                                            MemIntrinsic *MI,
                                            const TargetData &TD) {
  // If the mem operation is a non-constant size, we can't handle it.
  ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
  if (SizeCst == 0) return -1;
  uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;

  // If this is memset, we just need to see if the offset is valid in the size
  // of the memset..
  if (MI->getIntrinsicID() == Intrinsic::memset)
    return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
                                          MemSizeInBits, TD);
  
  // If we have a memcpy/memmove, the only case we can handle is if this is a
  // copy from constant memory.  In that case, we can read directly from the
  // constant memory.
  MemTransferInst *MTI = cast<MemTransferInst>(MI);
  
  Constant *Src = dyn_cast<Constant>(MTI->getSource());
  if (Src == 0) return -1;
  
  GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject());
  if (GV == 0 || !GV->isConstant()) return -1;
  
  // See if the access is within the bounds of the transfer.
  int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
                                              MI->getDest(), MemSizeInBits, TD);
  if (Offset == -1)
    return Offset;
  
  // Otherwise, see if we can constant fold a load from the constant with the
  // offset applied as appropriate.
  Src = ConstantExpr::getBitCast(Src,
                                 llvm::Type::getInt8PtrTy(Src->getContext()));
  Constant *OffsetCst = 
    ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
  Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
  Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
  if (ConstantFoldLoadFromConstPtr(Src, &TD))
    return Offset;
  return -1;
}
                                            

/// GetStoreValueForLoad - This function is called when we have a
/// memdep query of a load that ends up being a clobbering store.  This means
/// that the store *may* provide bits used by the load but we can't be sure
/// because the pointers don't mustalias.  Check this case to see if there is
/// anything more we can do before we give up.
static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
                                   const Type *LoadTy,
                                   Instruction *InsertPt, const TargetData &TD){
  LLVMContext &Ctx = SrcVal->getType()->getContext();
  
  uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8;
  uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
  
  IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
  
  // Compute which bits of the stored value are being used by the load.  Convert
  // to an integer type to start with.
  if (isa<PointerType>(SrcVal->getType()))
    SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
  if (!isa<IntegerType>(SrcVal->getType()))
    SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
                                   "tmp");
  
  // Shift the bits to the least significant depending on endianness.
  unsigned ShiftAmt;
  if (TD.isLittleEndian())
    ShiftAmt = Offset*8;
  else
    ShiftAmt = (StoreSize-LoadSize-Offset)*8;
  
  if (ShiftAmt)
    SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
  
  if (LoadSize != StoreSize)
    SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
                                 "tmp");
  
  return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
}

/// GetMemInstValueForLoad - This function is called when we have a
/// memdep query of a load that ends up being a clobbering mem intrinsic.
static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
                                     const Type *LoadTy, Instruction *InsertPt,
                                     const TargetData &TD){
  LLVMContext &Ctx = LoadTy->getContext();
  uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;

  IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
  
  // We know that this method is only called when the mem transfer fully
  // provides the bits for the load.
  if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
    // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
    // independently of what the offset is.
    Value *Val = MSI->getValue();
    if (LoadSize != 1)
      Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
    
    Value *OneElt = Val;
    
    // Splat the value out to the right number of bits.
    for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
      // If we can double the number of bytes set, do it.
      if (NumBytesSet*2 <= LoadSize) {
        Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
        Val = Builder.CreateOr(Val, ShVal);
        NumBytesSet <<= 1;
        continue;
      }
      
      // Otherwise insert one byte at a time.
      Value *ShVal = Builder.CreateShl(Val, 1*8);
      Val = Builder.CreateOr(OneElt, ShVal);
      ++NumBytesSet;
    }
    
    return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
  }
 
  // Otherwise, this is a memcpy/memmove from a constant global.
  MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
  Constant *Src = cast<Constant>(MTI->getSource());

  // Otherwise, see if we can constant fold a load from the constant with the
  // offset applied as appropriate.
  Src = ConstantExpr::getBitCast(Src,
                                 llvm::Type::getInt8PtrTy(Src->getContext()));
  Constant *OffsetCst = 
  ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
  Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
  Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
  return ConstantFoldLoadFromConstPtr(Src, &TD);
}



struct AvailableValueInBlock {
  /// BB - The basic block in question.
  BasicBlock *BB;
  enum ValType {
    SimpleVal,  // A simple offsetted value that is accessed.
    MemIntrin   // A memory intrinsic which is loaded from.
  };
  
  /// V - The value that is live out of the block.
  PointerIntPair<Value *, 1, ValType> Val;
  
  /// Offset - The byte offset in Val that is interesting for the load query.
  unsigned Offset;
  
  static AvailableValueInBlock get(BasicBlock *BB, Value *V,
                                   unsigned Offset = 0) {
    AvailableValueInBlock Res;
    Res.BB = BB;
    Res.Val.setPointer(V);
    Res.Val.setInt(SimpleVal);
    Res.Offset = Offset;
    return Res;
  }

  static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
                                     unsigned Offset = 0) {
    AvailableValueInBlock Res;
    Res.BB = BB;
    Res.Val.setPointer(MI);
    Res.Val.setInt(MemIntrin);
    Res.Offset = Offset;
    return Res;
  }
  
  bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
  Value *getSimpleValue() const {
    assert(isSimpleValue() && "Wrong accessor");
    return Val.getPointer();
  }
  
  MemIntrinsic *getMemIntrinValue() const {
    assert(!isSimpleValue() && "Wrong accessor");
    return cast<MemIntrinsic>(Val.getPointer());
  }
  
  /// MaterializeAdjustedValue - Emit code into this block to adjust the value
  /// defined here to the specified type.  This handles various coercion cases.
  Value *MaterializeAdjustedValue(const Type *LoadTy,
                                  const TargetData *TD) const {
    Value *Res;
    if (isSimpleValue()) {
      Res = getSimpleValue();
      if (Res->getType() != LoadTy) {
        assert(TD && "Need target data to handle type mismatch case");
        Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
                                   *TD);
        
        DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << "  "
                     << *getSimpleValue() << '\n'
                     << *Res << '\n' << "\n\n\n");
      }
    } else {
      Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
                                   LoadTy, BB->getTerminator(), *TD);
      DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
                   << "  " << *getMemIntrinValue() << '\n'
                   << *Res << '\n' << "\n\n\n");
    }
    return Res;
  }
};

/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
/// construct SSA form, allowing us to eliminate LI.  This returns the value
/// that should be used at LI's definition site.
static Value *ConstructSSAForLoadSet(LoadInst *LI, 
                         SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
                                     const TargetData *TD,
                                     const DominatorTree &DT,
                                     AliasAnalysis *AA) {
  // Check for the fully redundant, dominating load case.  In this case, we can
  // just use the dominating value directly.
  if (ValuesPerBlock.size() == 1 && 
      DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent()))
    return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD);

  // Otherwise, we have to construct SSA form.
  SmallVector<PHINode*, 8> NewPHIs;
  SSAUpdater SSAUpdate(&NewPHIs);
  SSAUpdate.Initialize(LI);
  
  const Type *LoadTy = LI->getType();
  
  for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
    const AvailableValueInBlock &AV = ValuesPerBlock[i];
    BasicBlock *BB = AV.BB;
    
    if (SSAUpdate.HasValueForBlock(BB))
      continue;

    SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD));
  }
  
  // Perform PHI construction.
  Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
  
  // If new PHI nodes were created, notify alias analysis.
  if (isa<PointerType>(V->getType()))
    for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
      AA->copyValue(LI, NewPHIs[i]);

  return V;
}

static bool isLifetimeStart(Instruction *Inst) {
  if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
    return II->getIntrinsicID() == Intrinsic::lifetime_start;
  return false;
}

/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
/// non-local by performing PHI construction.
bool GVN::processNonLocalLoad(LoadInst *LI,
                              SmallVectorImpl<Instruction*> &toErase) {
  // Find the non-local dependencies of the load.
  SmallVector<NonLocalDepResult, 64> Deps;
  MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
                                   Deps);
  //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
  //             << Deps.size() << *LI << '\n');

  // If we had to process more than one hundred blocks to find the
  // dependencies, this load isn't worth worrying about.  Optimizing
  // it will be too expensive.
  if (Deps.size() > 100)
    return false;

  // If we had a phi translation failure, we'll have a single entry which is a
  // clobber in the current block.  Reject this early.
  if (Deps.size() == 1 && Deps[0].getResult().isClobber()) {
    DEBUG(
      dbgs() << "GVN: non-local load ";
      WriteAsOperand(dbgs(), LI);
      dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
    );
    return false;
  }

  // Filter out useless results (non-locals, etc).  Keep track of the blocks
  // where we have a value available in repl, also keep track of whether we see
  // dependencies that produce an unknown value for the load (such as a call
  // that could potentially clobber the load).
  SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
  SmallVector<BasicBlock*, 16> UnavailableBlocks;

  const TargetData *TD = 0;
  
  for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
    BasicBlock *DepBB = Deps[i].getBB();
    MemDepResult DepInfo = Deps[i].getResult();

    if (DepInfo.isClobber()) {
      // The address being loaded in this non-local block may not be the same as
      // the pointer operand of the load if PHI translation occurs.  Make sure
      // to consider the right address.
      Value *Address = Deps[i].getAddress();
      
      // If the dependence is to a store that writes to a superset of the bits
      // read by the load, we can extract the bits we need for the load from the
      // stored value.
      if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
        if (TD == 0)
          TD = getAnalysisIfAvailable<TargetData>();
        if (TD && Address) {
          int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
                                                      DepSI, *TD);
          if (Offset != -1) {
            ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
                                                           DepSI->getOperand(0),
                                                                Offset));
            continue;
          }
        }
      }

      // If the clobbering value is a memset/memcpy/memmove, see if we can
      // forward a value on from it.
      if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
        if (TD == 0)
          TD = getAnalysisIfAvailable<TargetData>();
        if (TD && Address) {
          int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
                                                        DepMI, *TD);
          if (Offset != -1) {
            ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
                                                                  Offset));
            continue;
          }            
        }
      }
      
      UnavailableBlocks.push_back(DepBB);
      continue;
    }

    Instruction *DepInst = DepInfo.getInst();

    // Loading the allocation -> undef.
    if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
        // Loading immediately after lifetime begin -> undef.
        isLifetimeStart(DepInst)) {
      ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
                                             UndefValue::get(LI->getType())));
      continue;
    }
    
    if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
      // Reject loads and stores that are to the same address but are of
      // different types if we have to.
      if (S->getOperand(0)->getType() != LI->getType()) {
        if (TD == 0)
          TD = getAnalysisIfAvailable<TargetData>();
        
        // If the stored value is larger or equal to the loaded value, we can
        // reuse it.
        if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0),
                                                        LI->getType(), *TD)) {
          UnavailableBlocks.push_back(DepBB);
          continue;
        }
      }

      ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
                                                          S->getOperand(0)));
      continue;
    }
    
    if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
      // If the types mismatch and we can't handle it, reject reuse of the load.
      if (LD->getType() != LI->getType()) {
        if (TD == 0)
          TD = getAnalysisIfAvailable<TargetData>();
        
        // If the stored value is larger or equal to the loaded value, we can
        // reuse it.
        if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
          UnavailableBlocks.push_back(DepBB);
          continue;
        }          
      }
      ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
      continue;
    }
    
    UnavailableBlocks.push_back(DepBB);
    continue;
  }

  // If we have no predecessors that produce a known value for this load, exit
  // early.
  if (ValuesPerBlock.empty()) return false;

  // If all of the instructions we depend on produce a known value for this
  // load, then it is fully redundant and we can use PHI insertion to compute
  // its value.  Insert PHIs and remove the fully redundant value now.
  if (UnavailableBlocks.empty()) {
    DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
    
    // Perform PHI construction.
    Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
                                      VN.getAliasAnalysis());
    LI->replaceAllUsesWith(V);

    if (isa<PHINode>(V))
      V->takeName(LI);
    if (isa<PointerType>(V->getType()))
      MD->invalidateCachedPointerInfo(V);
    toErase.push_back(LI);
    NumGVNLoad++;
    return true;
  }

  if (!EnablePRE || !EnableLoadPRE)
    return false;

  // Okay, we have *some* definitions of the value.  This means that the value
  // is available in some of our (transitive) predecessors.  Lets think about
  // doing PRE of this load.  This will involve inserting a new load into the
  // predecessor when it's not available.  We could do this in general, but
  // prefer to not increase code size.  As such, we only do this when we know
  // that we only have to insert *one* load (which means we're basically moving
  // the load, not inserting a new one).

  SmallPtrSet<BasicBlock *, 4> Blockers;
  for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
    Blockers.insert(UnavailableBlocks[i]);

  // Lets find first basic block with more than one predecessor.  Walk backwards
  // through predecessors if needed.
  BasicBlock *LoadBB = LI->getParent();
  BasicBlock *TmpBB = LoadBB;

  bool isSinglePred = false;
  bool allSingleSucc = true;
  while (TmpBB->getSinglePredecessor()) {
    isSinglePred = true;
    TmpBB = TmpBB->getSinglePredecessor();
    if (TmpBB == LoadBB) // Infinite (unreachable) loop.
      return false;
    if (Blockers.count(TmpBB))
      return false;
    if (TmpBB->getTerminator()->getNumSuccessors() != 1)
      allSingleSucc = false;
  }

  assert(TmpBB);
  LoadBB = TmpBB;

  // If we have a repl set with LI itself in it, this means we have a loop where
  // at least one of the values is LI.  Since this means that we won't be able
  // to eliminate LI even if we insert uses in the other predecessors, we will
  // end up increasing code size.  Reject this by scanning for LI.
  for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
    if (ValuesPerBlock[i].isSimpleValue() &&
        ValuesPerBlock[i].getSimpleValue() == LI)
      return false;

  // FIXME: It is extremely unclear what this loop is doing, other than
  // artificially restricting loadpre.
  if (isSinglePred) {
    bool isHot = false;
    for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
      const AvailableValueInBlock &AV = ValuesPerBlock[i];
      if (AV.isSimpleValue())
        // "Hot" Instruction is in some loop (because it dominates its dep.
        // instruction).
        if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
          if (DT->dominates(LI, I)) {
            isHot = true;
            break;
          }
    }

    // We are interested only in "hot" instructions. We don't want to do any
    // mis-optimizations here.
    if (!isHot)
      return false;
  }

  // Okay, we have some hope :).  Check to see if the loaded value is fully
  // available in all but one predecessor.
  // FIXME: If we could restructure the CFG, we could make a common pred with
  // all the preds that don't have an available LI and insert a new load into
  // that one block.
  BasicBlock *UnavailablePred = 0;

  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
  for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
    FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
  for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
    FullyAvailableBlocks[UnavailableBlocks[i]] = false;

  for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
       PI != E; ++PI) {
    if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
      continue;

    // If this load is not available in multiple predecessors, reject it.
    if (UnavailablePred && UnavailablePred != *PI)
      return false;
    UnavailablePred = *PI;
  }

  assert(UnavailablePred != 0 &&
         "Fully available value should be eliminated above!");

  // We don't currently handle critical edges :(
  if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
    DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
                 << UnavailablePred->getName() << "': " << *LI << '\n');
    return false;
  }
  
  // Do PHI translation to get its value in the predecessor if necessary.  The
  // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
  //
  SmallVector<Instruction*, 8> NewInsts;
  
  // If all preds have a single successor, then we know it is safe to insert the
  // load on the pred (?!?), so we can insert code to materialize the pointer if
  // it is not available.
  PHITransAddr Address(LI->getOperand(0), TD);
  Value *LoadPtr = 0;
  if (allSingleSucc) {
    LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
                                                *DT, NewInsts);
  } else {
    Address.PHITranslateValue(LoadBB, UnavailablePred);
    LoadPtr = Address.getAddr();
    
    // Make sure the value is live in the predecessor.
    if (Instruction *Inst = dyn_cast_or_null<Instruction>(LoadPtr))
      if (!DT->dominates(Inst->getParent(), UnavailablePred))
        LoadPtr = 0;
  }

  // If we couldn't find or insert a computation of this phi translated value,
  // we fail PRE.
  if (LoadPtr == 0) {
    assert(NewInsts.empty() && "Shouldn't insert insts on failure");
    DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
                 << *LI->getOperand(0) << "\n");
    return false;
  }

  // Assign value numbers to these new instructions.
  for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
    // FIXME: We really _ought_ to insert these value numbers into their 
    // parent's availability map.  However, in doing so, we risk getting into
    // ordering issues.  If a block hasn't been processed yet, we would be
    // marking a value as AVAIL-IN, which isn't what we intend.
    VN.lookup_or_add(NewInsts[i]);
  }
  
  // Make sure it is valid to move this load here.  We have to watch out for:
  //  @1 = getelementptr (i8* p, ...
  //  test p and branch if == 0
  //  load @1
  // It is valid to have the getelementptr before the test, even if p can be 0,
  // as getelementptr only does address arithmetic.
  // If we are not pushing the value through any multiple-successor blocks
  // we do not have this case.  Otherwise, check that the load is safe to
  // put anywhere; this can be improved, but should be conservatively safe.
  if (!allSingleSucc &&
      // FIXME: REEVALUTE THIS.
      !isSafeToLoadUnconditionally(LoadPtr,
                                   UnavailablePred->getTerminator(),
                                   LI->getAlignment(), TD)) {
    assert(NewInsts.empty() && "Should not have inserted instructions");
    return false;
  }

  // Okay, we can eliminate this load by inserting a reload in the predecessor
  // and using PHI construction to get the value in the other predecessors, do
  // it.
  DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
  DEBUG(if (!NewInsts.empty())
          dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
                 << *NewInsts.back() << '\n');
  
  Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
                                LI->getAlignment(),
                                UnavailablePred->getTerminator());

  // Add the newly created load.
  ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,NewLoad));

  // Perform PHI construction.
  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
                                    VN.getAliasAnalysis());
  LI->replaceAllUsesWith(V);
  if (isa<PHINode>(V))
    V->takeName(LI);
  if (isa<PointerType>(V->getType()))
    MD->invalidateCachedPointerInfo(V);
  toErase.push_back(LI);
  NumPRELoad++;
  return true;
}

/// processLoad - Attempt to eliminate a load, first by eliminating it
/// locally, and then attempting non-local elimination if that fails.
bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
  if (!MD)
    return false;

  if (L->isVolatile())
    return false;

  // ... to a pointer that has been loaded from before...
  MemDepResult Dep = MD->getDependency(L);

  // If the value isn't available, don't do anything!
  if (Dep.isClobber()) {
    // Check to see if we have something like this:
    //   store i32 123, i32* %P
    //   %A = bitcast i32* %P to i8*
    //   %B = gep i8* %A, i32 1
    //   %C = load i8* %B
    //
    // We could do that by recognizing if the clobber instructions are obviously
    // a common base + constant offset, and if the previous store (or memset)
    // completely covers this load.  This sort of thing can happen in bitfield
    // access code.
    Value *AvailVal = 0;
    if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
      if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
        int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
                                                    L->getPointerOperand(),
                                                    DepSI, *TD);
        if (Offset != -1)
          AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
                                          L->getType(), L, *TD);
      }
    
    // If the clobbering value is a memset/memcpy/memmove, see if we can forward
    // a value on from it.
    if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
      if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
        int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
                                                      L->getPointerOperand(),
                                                      DepMI, *TD);
        if (Offset != -1)
          AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD);
      }
    }
        
    if (AvailVal) {
      DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
            << *AvailVal << '\n' << *L << "\n\n\n");
      
      // Replace the load!
      L->replaceAllUsesWith(AvailVal);
      if (isa<PointerType>(AvailVal->getType()))
        MD->invalidateCachedPointerInfo(AvailVal);
      toErase.push_back(L);
      NumGVNLoad++;
      return true;
    }
        
    DEBUG(
      // fast print dep, using operator<< on instruction would be too slow
      dbgs() << "GVN: load ";
      WriteAsOperand(dbgs(), L);
      Instruction *I = Dep.getInst();
      dbgs() << " is clobbered by " << *I << '\n';
    );
    return false;
  }

  // If it is defined in another block, try harder.
  if (Dep.isNonLocal())
    return processNonLocalLoad(L, toErase);

  Instruction *DepInst = Dep.getInst();
  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
    Value *StoredVal = DepSI->getOperand(0);
    
    // The store and load are to a must-aliased pointer, but they may not
    // actually have the same type.  See if we know how to reuse the stored
    // value (depending on its type).
    const TargetData *TD = 0;
    if (StoredVal->getType() != L->getType()) {
      if ((TD = getAnalysisIfAvailable<TargetData>())) {
        StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
                                                   L, *TD);
        if (StoredVal == 0)
          return false;
        
        DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
                     << '\n' << *L << "\n\n\n");
      }
      else 
        return false;
    }

    // Remove it!
    L->replaceAllUsesWith(StoredVal);
    if (isa<PointerType>(StoredVal->getType()))
      MD->invalidateCachedPointerInfo(StoredVal);
    toErase.push_back(L);
    NumGVNLoad++;
    return true;
  }

  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
    Value *AvailableVal = DepLI;
    
    // The loads are of a must-aliased pointer, but they may not actually have
    // the same type.  See if we know how to reuse the previously loaded value
    // (depending on its type).
    const TargetData *TD = 0;
    if (DepLI->getType() != L->getType()) {
      if ((TD = getAnalysisIfAvailable<TargetData>())) {
        AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
        if (AvailableVal == 0)
          return false;
      
        DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
                     << "\n" << *L << "\n\n\n");
      }
      else 
        return false;
    }
    
    // Remove it!
    L->replaceAllUsesWith(AvailableVal);
    if (isa<PointerType>(DepLI->getType()))
      MD->invalidateCachedPointerInfo(DepLI);
    toErase.push_back(L);
    NumGVNLoad++;
    return true;
  }

  // If this load really doesn't depend on anything, then we must be loading an
  // undef value.  This can happen when loading for a fresh allocation with no
  // intervening stores, for example.
  if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
    L->replaceAllUsesWith(UndefValue::get(L->getType()));
    toErase.push_back(L);
    NumGVNLoad++;
    return true;
  }
  
  // If this load occurs either right after a lifetime begin,
  // then the loaded value is undefined.
  if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
    if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
      L->replaceAllUsesWith(UndefValue::get(L->getType()));
      toErase.push_back(L);
      NumGVNLoad++;
      return true;
    }
  }

  return false;
}

Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
  DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
  if (I == localAvail.end())
    return 0;

  ValueNumberScope *Locals = I->second;
  while (Locals) {
    DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num);
    if (I != Locals->table.end())
      return I->second;
    Locals = Locals->parent;
  }

  return 0;
}


/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
bool GVN::processInstruction(Instruction *I,
                             SmallVectorImpl<Instruction*> &toErase) {
  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
    bool Changed = processLoad(LI, toErase);

    if (!Changed) {
      unsigned Num = VN.lookup_or_add(LI);
      localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI));
    }

    return Changed;
  }

  uint32_t NextNum = VN.getNextUnusedValueNumber();
  unsigned Num = VN.lookup_or_add(I);

  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
    localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));

    if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
      return false;

    Value *BranchCond = BI->getCondition();
    uint32_t CondVN = VN.lookup_or_add(BranchCond);

    BasicBlock *TrueSucc = BI->getSuccessor(0);
    BasicBlock *FalseSucc = BI->getSuccessor(1);

    if (TrueSucc->getSinglePredecessor())
      localAvail[TrueSucc]->table[CondVN] =
        ConstantInt::getTrue(TrueSucc->getContext());
    if (FalseSucc->getSinglePredecessor())
      localAvail[FalseSucc]->table[CondVN] =
        ConstantInt::getFalse(TrueSucc->getContext());

    return false;

  // Allocations are always uniquely numbered, so we can save time and memory
  // by fast failing them.
  } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) {
    localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
    return false;
  }

  // Collapse PHI nodes
  if (PHINode* p = dyn_cast<PHINode>(I)) {
    Value *constVal = CollapsePhi(p);

    if (constVal) {
      p->replaceAllUsesWith(constVal);
      if (MD && isa<PointerType>(constVal->getType()))
        MD->invalidateCachedPointerInfo(constVal);
      VN.erase(p);

      toErase.push_back(p);
    } else {
      localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
    }

  // If the number we were assigned was a brand new VN, then we don't
  // need to do a lookup to see if the number already exists
  // somewhere in the domtree: it can't!
  } else if (Num == NextNum) {
    localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));

  // Perform fast-path value-number based elimination of values inherited from
  // dominators.
  } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
    // Remove it!
    VN.erase(I);
    I->replaceAllUsesWith(repl);
    if (MD && isa<PointerType>(repl->getType()))
      MD->invalidateCachedPointerInfo(repl);
    toErase.push_back(I);
    return true;

  } else {
    localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
  }

  return false;
}

/// runOnFunction - This is the main transformation entry point for a function.
bool GVN::runOnFunction(Function& F) {
  if (!NoLoads)
    MD = &getAnalysis<MemoryDependenceAnalysis>();
  DT = &getAnalysis<DominatorTree>();
  VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
  VN.setMemDep(MD);
  VN.setDomTree(DT);

  bool Changed = false;
  bool ShouldContinue = true;

  // Merge unconditional branches, allowing PRE to catch more
  // optimization opportunities.
  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
    BasicBlock *BB = FI;
    ++FI;
    bool removedBlock = MergeBlockIntoPredecessor(BB, this);
    if (removedBlock) NumGVNBlocks++;

    Changed |= removedBlock;
  }

  unsigned Iteration = 0;

  while (ShouldContinue) {
    DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
    ShouldContinue = iterateOnFunction(F);
    Changed |= ShouldContinue;
    ++Iteration;
  }

  if (EnablePRE) {
    bool PREChanged = true;
    while (PREChanged) {
      PREChanged = performPRE(F);
      Changed |= PREChanged;
    }
  }
  // FIXME: Should perform GVN again after PRE does something.  PRE can move
  // computations into blocks where they become fully redundant.  Note that
  // we can't do this until PRE's critical edge splitting updates memdep.
  // Actually, when this happens, we should just fully integrate PRE into GVN.

  cleanupGlobalSets();

  return Changed;
}


bool GVN::processBlock(BasicBlock *BB) {
  // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
  // incrementing BI before processing an instruction).