path: root/lib/Analysis/LoadValueNumbering.cpp

//===- LoadValueNumbering.cpp - Load Value #'ing Implementation -*- C++ -*-===//
// 
//                     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 implements a value numbering pass that value #'s load instructions.
// To do this, it finds lexically identical load instructions, and uses alias
// analysis to determine which loads are guaranteed to produce the same value.
//
// This pass builds off of another value numbering pass to implement value
// numbering for non-load instructions.  It uses Alias Analysis so that it can
// disambiguate the load instructions.  The more powerful these base analyses
// are, the more powerful the resultant analysis will be.
//
//===----------------------------------------------------------------------===//

#include "llvm/Analysis/LoadValueNumbering.h"
#include "llvm/Analysis/ValueNumbering.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Pass.h"
#include "llvm/Type.h"
#include "llvm/iMemory.h"
#include "llvm/BasicBlock.h"
#include "llvm/Support/CFG.h"
#include <set>
using namespace llvm;

namespace {
  // FIXME: This should not be a FunctionPass.
  struct LoadVN : public FunctionPass, public ValueNumbering {
    
    /// Pass Implementation stuff.  This doesn't do any analysis.
    ///
    bool runOnFunction(Function &) { return false; }
    
    /// getAnalysisUsage - Does not modify anything.  It uses Value Numbering
    /// and Alias Analysis.
    ///
    virtual void getAnalysisUsage(AnalysisUsage &AU) const;
    
    /// getEqualNumberNodes - Return nodes with the same value number as the
    /// specified Value.  This fills in the argument vector with any equal
    /// values.
    ///
    virtual void getEqualNumberNodes(Value *V1,
                                     std::vector<Value*> &RetVals) const;
  };

  // Register this pass...
  RegisterOpt<LoadVN> X("load-vn", "Load Value Numbering");

  // Declare that we implement the ValueNumbering interface
  RegisterAnalysisGroup<ValueNumbering, LoadVN> Y;
}

Pass *llvm::createLoadValueNumberingPass() { return new LoadVN(); }


/// getAnalysisUsage - Does not modify anything.  It uses Value Numbering and
/// Alias Analysis.
///
void LoadVN::getAnalysisUsage(AnalysisUsage &AU) const {
  AU.setPreservesAll();
  AU.addRequired<AliasAnalysis>();
  AU.addRequired<ValueNumbering>();
  AU.addRequired<DominatorSet>();
  AU.addRequired<TargetData>();
}

static bool isPathTransparentTo(BasicBlock *CurBlock, BasicBlock *Dom,
                                Value *Ptr, unsigned Size, AliasAnalysis &AA,
                                std::set<BasicBlock*> &Visited,
                                std::map<BasicBlock*, bool> &TransparentBlocks){
  // If we have already checked out this path, or if we reached our destination,
  // stop searching, returning success.
  if (CurBlock == Dom || !Visited.insert(CurBlock).second)
    return true;
  
  // Check whether this block is known transparent or not.
  std::map<BasicBlock*, bool>::iterator TBI =
    TransparentBlocks.lower_bound(CurBlock);

  if (TBI == TransparentBlocks.end() || TBI->first != CurBlock) {
    // If this basic block can modify the memory location, then the path is not
    // transparent!
    if (AA.canBasicBlockModify(*CurBlock, Ptr, Size)) {
      TransparentBlocks.insert(TBI, std::make_pair(CurBlock, false));
      return false;
    }
      TransparentBlocks.insert(TBI, std::make_pair(CurBlock, true));
  } else if (!TBI->second)
    // This block is known non-transparent, so that path can't be either.
    return false;
  
  // The current block is known to be transparent.  The entire path is
  // transparent if all of the predecessors paths to the parent is also
  // transparent to the memory location.
  for (pred_iterator PI = pred_begin(CurBlock), E = pred_end(CurBlock);
       PI != E; ++PI)
    if (!isPathTransparentTo(*PI, Dom, Ptr, Size, AA, Visited,
                             TransparentBlocks))
      return false;
  return true;
}


// getEqualNumberNodes - Return nodes with the same value number as the
// specified Value.  This fills in the argument vector with any equal values.
//
void LoadVN::getEqualNumberNodes(Value *V,
                                 std::vector<Value*> &RetVals) const {
  // If the alias analysis has any must alias information to share with us, we
  // can definitely use it.
  if (isa<PointerType>(V->getType()))
    getAnalysis<AliasAnalysis>().getMustAliases(V, RetVals);

  if (!isa<LoadInst>(V)) {
    // Not a load instruction?  Just chain to the base value numbering
    // implementation to satisfy the request...
    assert(&getAnalysis<ValueNumbering>() != (ValueNumbering*)this &&
           "getAnalysis() returned this!");

    return getAnalysis<ValueNumbering>().getEqualNumberNodes(V, RetVals);
  }

  // Volatile loads cannot be replaced with the value of other loads.
  LoadInst *LI = cast<LoadInst>(V);
  if (LI->isVolatile())
    return getAnalysis<ValueNumbering>().getEqualNumberNodes(V, RetVals);
  
  // If we have a load instruction, find all of the load and store instructions
  // that use the same source operand.  We implement this recursively, because
  // there could be a load of a load of a load that are all identical.  We are
  // guaranteed that this cannot be an infinite recursion because load
  // instructions would have to pass through a PHI node in order for there to be
  // a cycle.  The PHI node would be handled by the else case here, breaking the
  // infinite recursion.
  //
  std::vector<Value*> PointerSources;
  getEqualNumberNodes(LI->getOperand(0), PointerSources);
  PointerSources.push_back(LI->getOperand(0));
  
  BasicBlock *LoadBB = LI->getParent();
  Function *F = LoadBB->getParent();
  
  // Now that we know the set of equivalent source pointers for the load
  // instruction, look to see if there are any load or store candidates that are
  // identical.
  //
  std::map<BasicBlock*, std::vector<LoadInst*> >  CandidateLoads;
  std::map<BasicBlock*, std::vector<StoreInst*> > CandidateStores;
  
  while (!PointerSources.empty()) {
    Value *Source = PointerSources.back();
    PointerSources.pop_back();                // Get a source pointer...
    
    for (Value::use_iterator UI = Source->use_begin(), UE = Source->use_end();
         UI != UE; ++UI)
      if (LoadInst *Cand = dyn_cast<LoadInst>(*UI)) {// Is a load of source?
        if (Cand->getParent()->getParent() == F &&   // In the same function?
            Cand != LI && !Cand->isVolatile())       // Not LI itself?
          CandidateLoads[Cand->getParent()].push_back(Cand);     // Got one...
      } else if (StoreInst *Cand = dyn_cast<StoreInst>(*UI)) {
        if (Cand->getParent()->getParent() == F && !Cand->isVolatile() &&
            Cand->getOperand(1) == Source)  // It's a store THROUGH the ptr...
          CandidateStores[Cand->getParent()].push_back(Cand);
      }
  }
  
  // Get alias analysis & dominators.
  AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
  DominatorSet &DomSetInfo = getAnalysis<DominatorSet>();
  Value *LoadPtr = LI->getOperand(0);
  // Find out how many bytes of memory are loaded by the load instruction...
  unsigned LoadSize = getAnalysis<TargetData>().getTypeSize(LI->getType());

  // Find all of the candidate loads and stores that are in the same block as
  // the defining instruction.
  std::set<Instruction*> Instrs;
  Instrs.insert(CandidateLoads[LoadBB].begin(), CandidateLoads[LoadBB].end());
  CandidateLoads.erase(LoadBB);
  Instrs.insert(CandidateStores[LoadBB].begin(), CandidateStores[LoadBB].end());
  CandidateStores.erase(LoadBB);

  // Figure out if the load is invalidated from the entry of the block it is in
  // until the actual instruction.  This scans the block backwards from LI.  If
  // we see any candidate load or store instructions, then we know that the
  // candidates have the same value # as LI.
  bool LoadInvalidatedInBBBefore = false;
  for (BasicBlock::iterator I = LI; I != LoadBB->begin(); ) {
    --I;
    // If this instruction is a candidate load before LI, we know there are no
    // invalidating instructions between it and LI, so they have the same value
    // number.
    if (isa<LoadInst>(I) && Instrs.count(I)) {
      RetVals.push_back(I);
      Instrs.erase(I);
    }

    if (AA.getModRefInfo(I, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
      // If the invalidating instruction is a store, and its in our candidate
      // set, then we can do store-load forwarding: the load has the same value
      // # as the stored value.
      if (isa<StoreInst>(I) && Instrs.count(I)) {
        Instrs.erase(I);
        RetVals.push_back(I->getOperand(0));
      }

      LoadInvalidatedInBBBefore = true;
      break;
    }
  }

  // Figure out if the load is invalidated between the load and the exit of the
  // block it is defined in.  While we are scanning the current basic block, if
  // we see any candidate loads, then we know they have the same value # as LI.
  //
  bool LoadInvalidatedInBBAfter = false;
  for (BasicBlock::iterator I = LI->getNext(); I != LoadBB->end(); ++I) {
    // If this instruction is a load,