GVN.cpp

      }
      
      // Whether we removed it or not, we can't
      // go any further
      break;
    } else if (!last) {
      // If we don't depend on a store, and we haven't
      // been loaded before, bail.
      break;
    } else if (dep == last) {
      // Remove it!
      MD.removeInstruction(L);
      
      L->replaceAllUsesWith(last);
      toErase.push_back(L);
      deletedLoad = true;
      NumGVNLoad++;
        
      break;
    } else {
      dep = MD.getDependency(L, dep);
    }
  }

  if (dep != MemoryDependenceAnalysis::None &&
      dep != MemoryDependenceAnalysis::NonLocal &&
      isa<AllocationInst>(dep)) {
    // Check that this load is actually from the
    // allocation we found
    Value* v = L->getOperand(0);
    while (true) {
      if (BitCastInst *BC = dyn_cast<BitCastInst>(v))
        v = BC->getOperand(0);
      else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(v))
        v = GEP->getOperand(0);
      else
        break;
    }
    if (v == dep) {
      // If this load depends directly on an allocation, there isn't
      // anything stored there; therefore, we can optimize this load
      // to undef.
      MD.removeInstruction(L);

      L->replaceAllUsesWith(UndefValue::get(L->getType()));
      toErase.push_back(L);
      deletedLoad = true;
      NumGVNLoad++;
    }
  }

  if (!deletedLoad)
    last = L;
  
  return deletedLoad;
}

/// performCallSlotOptzn - takes a memcpy and a call that it depends on,
/// and checks for the possibility of a call slot optimization by having
/// the call write its result directly into the destination of the memcpy.
bool GVN::performCallSlotOptzn(MemCpyInst* cpy, CallInst* C,
                               SmallVector<Instruction*, 4>& toErase) {
  // The general transformation to keep in mind is
  //
  //   call @func(..., src, ...)
  //   memcpy(dest, src, ...)
  //
  // ->
  //
  //   memcpy(dest, src, ...)
  //   call @func(..., dest, ...)
  //
  // Since moving the memcpy is technically awkward, we additionally check that
  // src only holds uninitialized values at the moment of the call, meaning that
  // the memcpy can be discarded rather than moved.

  // Deliberately get the source and destination with bitcasts stripped away,
  // because we'll need to do type comparisons based on the underlying type.
  Value* cpyDest = cpy->getDest();
  Value* cpySrc = cpy->getSource();
  CallSite CS = CallSite::get(C);

  // We need to be able to reason about the size of the memcpy, so we require
  // that it be a constant.
  ConstantInt* cpyLength = dyn_cast<ConstantInt>(cpy->getLength());
  if (!cpyLength)
    return false;

  // Require that src be an alloca.  This simplifies the reasoning considerably.
  AllocaInst* srcAlloca = dyn_cast<AllocaInst>(cpySrc);
  if (!srcAlloca)
    return false;

  // Check that all of src is copied to dest.
  TargetData& TD = getAnalysis<TargetData>();

  ConstantInt* srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
  if (!srcArraySize)
    return false;

  uint64_t srcSize = TD.getABITypeSize(srcAlloca->getAllocatedType()) *
    srcArraySize->getZExtValue();

  if (cpyLength->getZExtValue() < srcSize)
    return false;

  // Check that accessing the first srcSize bytes of dest will not cause a
  // trap.  Otherwise the transform is invalid since it might cause a trap
  // to occur earlier than it otherwise would.
  if (AllocaInst* A = dyn_cast<AllocaInst>(cpyDest)) {
    // The destination is an alloca.  Check it is larger than srcSize.
    ConstantInt* destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
    if (!destArraySize)
      return false;

    uint64_t destSize = TD.getABITypeSize(A->getAllocatedType()) *
      destArraySize->getZExtValue();

    if (destSize < srcSize)
      return false;
  } else if (Argument* A = dyn_cast<Argument>(cpyDest)) {
    // If the destination is an sret parameter then only accesses that are
    // outside of the returned struct type can trap.
    if (!A->hasStructRetAttr())
      return false;

    const Type* StructTy = cast<PointerType>(A->getType())->getElementType();
    uint64_t destSize = TD.getABITypeSize(StructTy);

    if (destSize < srcSize)
      return false;
  } else {
    return false;
  }

  // Check that src is not accessed except via the call and the memcpy.  This
  // guarantees that it holds only undefined values when passed in (so the final
  // memcpy can be dropped), that it is not read or written between the call and
  // the memcpy, and that writing beyond the end of it is undefined.

  SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
                                   srcAlloca->use_end());
  while (!srcUseList.empty()) {
    User* UI = srcUseList.back();
    srcUseList.pop_back();

    if (isa<GetElementPtrInst>(UI) || isa<BitCastInst>(UI)) {
      for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
           I != E; ++I)
        srcUseList.push_back(*I);
    } else if (UI != C && UI != cpy) {
      return false;
    }
  }

  // Since we're changing the parameter to the callsite, we need to make sure
  // that what would be the new parameter dominates the callsite.
  DominatorTree& DT = getAnalysis<DominatorTree>();
  if (Instruction* cpyDestInst = dyn_cast<Instruction>(cpyDest))
    if (!DT.dominates(cpyDestInst, C))
      return false;

  // In addition to knowing that the call does not access src in some
  // unexpected manner, for example via a global, which we deduce from
  // the use analysis, we also need to know that it does not sneakily
  // access dest.  We rely on AA to figure this out for us.
  AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
  if (AA.getModRefInfo(C, cpy->getRawDest(), srcSize) !=
      AliasAnalysis::NoModRef)
    return false;

  // All the checks have passed, so do the transformation.
  for (unsigned i = 0; i < CS.arg_size(); ++i)
    if (CS.getArgument(i) == cpySrc) {
      if (cpySrc->getType() != cpyDest->getType())
        cpyDest = CastInst::createPointerCast(cpyDest, cpySrc->getType(),
                                              cpyDest->getName(), C);
      CS.setArgument(i, cpyDest);
    }

  // Drop any cached information about the call, because we may have changed
  // its dependence information by changing its parameter.
  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
  MD.dropInstruction(C);

  // Remove the memcpy
  MD.removeInstruction(cpy);
  toErase.push_back(cpy);

  return true;
}

/// processMemCpy - perform simplication of memcpy's.  If we have memcpy A which
/// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be
/// a memcpy from X to Z (or potentially a memmove, depending on circumstances).
///  This allows later passes to remove the first memcpy altogether.
bool GVN::processMemCpy(MemCpyInst* M, MemCpyInst* MDep,
                        SmallVector<Instruction*, 4>& toErase) {
  // We can only transforms memcpy's where the dest of one is the source of the
  // other
  if (M->getSource() != MDep->getDest())
    return false;
  
  // Second, the length of the memcpy's must be the same, or the preceeding one
  // must be larger than the following one.
  ConstantInt* C1 = dyn_cast<ConstantInt>(MDep->getLength());
  ConstantInt* C2 = dyn_cast<ConstantInt>(M->getLength());
  if (!C1 || !C2)
    return false;
  
  uint64_t DepSize = C1->getValue().getZExtValue();
  uint64_t CpySize = C2->getValue().getZExtValue();
  
  if (DepSize < CpySize)
    return false;
  
  // Finally, we have to make sure that the dest of the second does not
  // alias the source of the first
  AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
  if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) !=
      AliasAnalysis::NoAlias)
    return false;
  else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) !=
           AliasAnalysis::NoAlias)
    return false;
  else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize)
           != AliasAnalysis::NoAlias)
    return false;
  
  // If all checks passed, then we can transform these memcpy's
  Function* MemCpyFun = Intrinsic::getDeclaration(
                                 M->getParent()->getParent()->getParent(),
                                 M->getIntrinsicID());
    
  std::vector<Value*> args;
  args.push_back(M->getRawDest());
  args.push_back(MDep->getRawSource());
  args.push_back(M->getLength());
  args.push_back(M->getAlignment());
  
  CallInst* C = new CallInst(MemCpyFun, args.begin(), args.end(), "", M);
  
  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
  if (MD.getDependency(C) == MDep) {
    MD.dropInstruction(M);
    toErase.push_back(M);
    return true;
  } else {
    MD.removeInstruction(C);
    toErase.push_back(C);
    return false;
  }
}

/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
bool GVN::processInstruction(Instruction* I,
                                ValueNumberedSet& currAvail,
                                DenseMap<Value*, LoadInst*>& lastSeenLoad,
                                SmallVector<Instruction*, 4>& toErase) {
  if (LoadInst* L = dyn_cast<LoadInst>(I)) {
    return processLoad(L, lastSeenLoad, toErase);
  } else if (MemCpyInst* M = dyn_cast<MemCpyInst>(I)) {
    MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();

    // The are two possible optimizations we can do for memcpy:
    //   a) memcpy-memcpy xform which exposes redundance for DSE
    //   b) call-memcpy xform for return slot optimization
    Instruction* dep = MD.getDependency(M);
    if (dep == MemoryDependenceAnalysis::None ||
        dep == MemoryDependenceAnalysis::NonLocal)
      return false;
    if (MemCpyInst *MemCpy = dyn_cast<MemCpyInst>(dep))
      return processMemCpy(M, MemCpy, toErase);
    if (CallInst* C = dyn_cast<CallInst>(dep))
      return performCallSlotOptzn(M, C, toErase);
    return false;
  }
  
  unsigned num = VN.lookup_or_add(I);
  
  // Collapse PHI nodes
  if (PHINode* p = dyn_cast<PHINode>(I)) {
    Value* constVal = CollapsePhi(p);
    
    if (constVal) {
      for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
           PI != PE; ++PI)
        if (PI->second.count(p))
          PI->second.erase(p);
        
      p->replaceAllUsesWith(constVal);
      toErase.push_back(p);
    }
  // Perform value-number based elimination
  } else if (currAvail.test(num)) {
    Value* repl = find_leader(currAvail, num);
    
    if (CallInst* CI = dyn_cast<CallInst>(I)) {
      AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
      if (!AA.doesNotAccessMemory(CI)) {
        MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
        if (cast<Instruction>(repl)->getParent() != CI->getParent() ||
            MD.getDependency(CI) != MD.getDependency(cast<CallInst>(repl))) {
          // There must be an intervening may-alias store, so nothing from
          // this point on will be able to be replaced with the preceding call
          currAvail.erase(repl);
          currAvail.insert(I);
          
          return false;
        }
      }
    }
    
    // Remove it!
    MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
    MD.removeInstruction(I);
    
    VN.erase(I);
    I->replaceAllUsesWith(repl);
    toErase.push_back(I);
    return true;
  } else if (!I->isTerminator()) {
    currAvail.set(num);
    currAvail.insert(I);
  }
  
  return false;
}

// GVN::runOnFunction - This is the main transformation entry point for a
// function.
//
bool GVN::runOnFunction(Function& F) {
  VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
  
  bool changed = false;
  bool shouldContinue = true;
  
  while (shouldContinue) {
    shouldContinue = iterateOnFunction(F);
    changed |= shouldContinue;
  }
  
  return changed;
}


// GVN::iterateOnFunction - Executes one iteration of GVN
bool GVN::iterateOnFunction(Function &F) {
  // Clean out global sets from any previous functions
  VN.clear();
  availableOut.clear();
  phiMap.clear();
 
  bool changed_function = false;
  
  DominatorTree &DT = getAnalysis<DominatorTree>();   
  
  SmallVector<Instruction*, 4> toErase;
  
  // Top-down walk of the dominator tree
  for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
         E = df_end(DT.getRootNode()); DI != E; ++DI) {
    
    // Get the set to update for this block
    ValueNumberedSet& currAvail = availableOut[DI->getBlock()];     
    DenseMap<Value*, LoadInst*> lastSeenLoad;
    
    BasicBlock* BB = DI->getBlock();
  
    // A block inherits AVAIL_OUT from its dominator
    if (DI->getIDom() != 0)
      currAvail = availableOut[DI->getIDom()->getBlock()];

    for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
         BI != BE; ) {
      changed_function |= processInstruction(BI, currAvail,
                                             lastSeenLoad, toErase);
      
      NumGVNInstr += toErase.size();
      
      // Avoid iterator invalidation
      ++BI;

      for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
           E = toErase.end(); I != E; ++I) {
        (*I)->eraseFromParent();
      }

      toErase.clear();
    }
  }
  
  return changed_function;
}