path: root/lib/Transforms/Utils/PromoteMemoryToRegister.cpp

//===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
// 
//                     The LLVM Compiler Infrastructure
//
// This file was developed by the LLVM research group and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
// 
//===----------------------------------------------------------------------===//
//
// This file promote memory references to be register references.  It promotes
// alloca instructions which only have loads and stores as uses.  An alloca is
// transformed by using dominator frontiers to place PHI nodes, then traversing
// the function in depth-first order to rewrite loads and stores as appropriate.
// This is just the standard SSA construction algorithm to construct "pruned"
// SSA form.
//
//===----------------------------------------------------------------------===//

#include "llvm/Transforms/Utils/PromoteMemToReg.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/iMemory.h"
#include "llvm/iPHINode.h"
#include "llvm/Function.h"
#include "llvm/Constant.h"
#include "llvm/Support/CFG.h"
#include "Support/StringExtras.h"

/// isAllocaPromotable - Return true if this alloca is legal for promotion.
/// This is true if there are only loads and stores to the alloca...
///
bool isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) {
  // FIXME: If the memory unit is of pointer or integer type, we can permit
  // assignments to subsections of the memory unit.

  // Only allow direct loads and stores...
  for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
       UI != UE; ++UI)     // Loop over all of the uses of the alloca
    if (!isa<LoadInst>(*UI))
      if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
        if (SI->getOperand(0) == AI)
          return false;   // Don't allow a store of the AI, only INTO the AI.
      } else {
        return false;   // Not a load or store?
      }
  
  return true;
}

namespace {
  struct PromoteMem2Reg {
    // Allocas - The alloca instructions being promoted
    std::vector<AllocaInst*> Allocas;
    DominatorTree &DT;
    DominanceFrontier &DF;
    const TargetData &TD;

    // AllocaLookup - Reverse mapping of Allocas
    std::map<AllocaInst*, unsigned>  AllocaLookup;

    // NewPhiNodes - The PhiNodes we're adding.
    std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes;

    // Visited - The set of basic blocks the renamer has already visited.
    std::set<BasicBlock*> Visited;

  public:
    PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
                   DominanceFrontier &df, const TargetData &td)
      : Allocas(A), DT(dt), DF(df), TD(td) {}

    void run();

  private:
    void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
                               std::set<PHINode*> &DeadPHINodes);
    void PromoteLocallyUsedAlloca(AllocaInst *AI);

    void RenamePass(BasicBlock *BB, BasicBlock *Pred,
                    std::vector<Value*> &IncVals);
    bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
                      std::set<PHINode*> &InsertedPHINodes);
  };
}  // end of anonymous namespace

void PromoteMem2Reg::run() {
  Function &F = *DF.getRoot()->getParent();

  for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
    AllocaInst *AI = Allocas[AllocaNum];

    assert(isAllocaPromotable(AI, TD) &&
           "Cannot promote non-promotable alloca!");
    assert(AI->getParent()->getParent() == &F &&
           "All allocas should be in the same function, which is same as DF!");

    if (AI->use_empty()) {
      // If there are no uses of the alloca, just delete it now.
      AI->getParent()->getInstList().erase(AI);

      // Remove the alloca from the Allocas list, since it has been processed
      Allocas[AllocaNum] = Allocas.back();
      Allocas.pop_back();
      --AllocaNum;
      continue;
    }

    // Calculate the set of read and write-locations for each alloca.  This is
    // analogous to counting the number of 'uses' and 'definitions' of each
    // variable.
    std::vector<BasicBlock*> DefiningBlocks;
    std::vector<BasicBlock*> UsingBlocks;

    BasicBlock *OnlyBlock = 0;
    bool OnlyUsedInOneBlock = true;

    // As we scan the uses of the alloca instruction, keep track of stores, and
    // decide whether all of the loads and stores to the alloca are within the
    // same basic block.
    for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){
      Instruction *User = cast<Instruction>(*U);
      if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
        // Remember the basic blocks which define new values for the alloca
        DefiningBlocks.push_back(SI->getParent());
      } else {
        // Otherwise it must be a load instruction, keep track of variable reads
        UsingBlocks.push_back(cast<LoadInst>(User)->getParent());
      }

      if (OnlyUsedInOneBlock) {
        if (OnlyBlock == 0)
          OnlyBlock = User->getParent();
        else if (OnlyBlock != User->getParent())
          OnlyUsedInOneBlock = false;
      }
    }

    // If the alloca is only read and written in one basic block, just perform a
    // linear sweep over the block to eliminate it.
    if (OnlyUsedInOneBlock) {
      PromoteLocallyUsedAlloca(AI);

      // Remove the alloca from the Allocas list, since it has been processed
      Allocas[AllocaNum] = Allocas.back();
      Allocas.pop_back();
      --AllocaNum;
      continue;
    }

    // Compute the locations where PhiNodes need to be inserted.  Look at the
    // dominance frontier of EACH basic-block we have a write in.
    //
    unsigned CurrentVersion = 0;
    std::set<PHINode*> InsertedPHINodes;
    while (!DefiningBlocks.empty()) {
      BasicBlock *BB = DefiningBlocks.back();
      DefiningBlocks.pop_back();

      // Look up the DF for this write, add it to PhiNodes
      DominanceFrontier::const_iterator it = DF.find(BB);
      if (it != DF.end()) {
        const DominanceFrontier::DomSetType &S = it->second;
        for (DominanceFrontier::DomSetType::iterator P = S.begin(),PE = S.end();
             P != PE; ++P)
          if (QueuePhiNode(*P, AllocaNum, CurrentVersion, InsertedPHINodes))
            DefiningBlocks.push_back(*P);
      }
    }

    // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
    // of the writes to the variable, scan through the reads of the variable,
    // marking PHI nodes which are actually necessary as alive (by removing them
    // from the InsertedPHINodes set).  This is not perfect: there may PHI
    // marked alive because of loads which are dominated by stores, but there
    // will be no unmarked PHI nodes which are actually used.
    //
    for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
      MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes);
    UsingBlocks.clear();

    // If there are any PHI nodes which are now known to be dead, remove them!
    for (std::set<PHINode*>::iterator I = InsertedPHINodes.begin(),
           E = InsertedPHINodes.end(); I != E; ++I) {
      PHINode *PN = *I;
      std::vector<PHINode*> &BBPNs = NewPhiNodes[PN->getParent()];
      BBPNs[AllocaNum] = 0;

      // Check to see if we just removed the last inserted PHI node from this
      // basic block.  If so, remove the entry for the basic block.
      bool HasOtherPHIs = false;
      for (unsigned i = 0, e = BBPNs.size(); i != e; ++i)
        if (BBPNs[i]) {
          HasOtherPHIs = true;
          break;
        }
      if (!HasOtherPHIs)
        NewPhiNodes.erase(PN->getParent());

      PN->getParent()->getInstList().erase(PN);      
    }

    // Keep the reverse mapping of the 'Allocas' array. 
    AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
  }
  
  if (Allocas.empty())
    return; // All of the allocas must have been trivial!

  // Set the incoming values for the basic block to be null values for all of
  // the alloca's.  We do this in case there is a load of a value that has not
  // been stored yet.  In this case, it will get this null value.
  //
  std::vector<Value *> Values(Allocas.size());
  for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
    Values[i] = Constant::getNullValue(Allocas[