path: root/lib/CodeGen/TailDuplication.cpp

//===-- TailDuplication.cpp - Duplicate blocks into predecessors' tails ---===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass duplicates basic blocks ending in unconditional branches into
// the tails of their predecessors.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "tailduplication"
#include "llvm/Function.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/MachineSSAUpdater.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/Statistic.h"
using namespace llvm;

STATISTIC(NumTailDups  , "Number of tail duplicated blocks");
STATISTIC(NumInstrDups , "Additional instructions due to tail duplication");
STATISTIC(NumDeadBlocks, "Number of dead blocks removed");

// Heuristic for tail duplication.
static cl::opt<unsigned>
TailDuplicateSize("tail-dup-size",
                  cl::desc("Maximum instructions to consider tail duplicating"),
                  cl::init(2), cl::Hidden);

typedef std::vector<unsigned> AvailableValsTy;

namespace {
  /// TailDuplicatePass - Perform tail duplication.
  class TailDuplicatePass : public MachineFunctionPass {
    const TargetInstrInfo *TII;
    MachineModuleInfo *MMI;
    MachineRegisterInfo *MRI;

    // SSAUpdateVRs - A list of virtual registers for which to update SSA form.
    SmallVector<unsigned, 16> SSAUpdateVRs;

    // SSAUpdateVals - For each virtual register in SSAUpdateVals keep a list of
    // source virtual registers.
    DenseMap<unsigned, AvailableValsTy> SSAUpdateVals;

  public:
    static char ID;
    explicit TailDuplicatePass() : MachineFunctionPass(&ID) {}

    virtual bool runOnMachineFunction(MachineFunction &MF);
    virtual const char *getPassName() const { return "Tail Duplication"; }

  private:
    void AddSSAUpdateEntry(unsigned OrigReg, unsigned NewReg);
    bool TailDuplicateBlocks(MachineFunction &MF);
    bool TailDuplicate(MachineBasicBlock *TailBB, MachineFunction &MF);
    void RemoveDeadBlock(MachineBasicBlock *MBB);
  };

  char TailDuplicatePass::ID = 0;
}

FunctionPass *llvm::createTailDuplicatePass() {
  return new TailDuplicatePass();
}

bool TailDuplicatePass::runOnMachineFunction(MachineFunction &MF) {
  TII = MF.getTarget().getInstrInfo();
  MRI = &MF.getRegInfo();
  MMI = getAnalysisIfAvailable<MachineModuleInfo>();

  bool MadeChange = false;
  bool MadeChangeThisIteration = true;
  while (MadeChangeThisIteration) {
    MadeChangeThisIteration = false;
    MadeChangeThisIteration |= TailDuplicateBlocks(MF);
    MadeChange |= MadeChangeThisIteration;
  }

  return MadeChange;
}

/// TailDuplicateBlocks - Look for small blocks that are unconditionally
/// branched to and do not fall through. Tail-duplicate their instructions
/// into their predecessors to eliminate (dynamic) branches.
bool TailDuplicatePass::TailDuplicateBlocks(MachineFunction &MF) {
  bool MadeChange = false;

  SSAUpdateVRs.clear();
  SSAUpdateVals.clear();

  for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
    MachineBasicBlock *MBB = I++;

    // Only duplicate blocks that end with unconditional branches.
    if (MBB->canFallThrough())
      continue;

    MadeChange |= TailDuplicate(MBB, MF);

    // If it is dead, remove it.
    if (MBB->pred_empty()) {
      NumInstrDups -= MBB->size();
      RemoveDeadBlock(MBB);
      MadeChange = true;
      ++NumDeadBlocks;
    }
  }

  if (!SSAUpdateVRs.empty()) {
    // Update SSA form.
    MachineSSAUpdater SSAUpdate(MF);

    for (unsigned i = 0, e = SSAUpdateVRs.size(); i != e; ++i) {
      unsigned VReg = SSAUpdateVRs[i];
      SSAUpdate.Initialize(VReg);

      // If the original definition is still around, add it as an available
      // value.
      MachineInstr *DefMI = MRI->getVRegDef(VReg);
      MachineBasicBlock *DefBB = 0;
      if (DefMI) {
        DefBB = DefMI->getParent();
        SSAUpdate.AddAvailableValue(DefBB, VReg);
      }

      // Add the new vregs as available values.
      DenseMap<unsigned, AvailableValsTy>::iterator LI =
        SSAUpdateVals.find(VReg);  
      for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) {
        unsigned NewReg = LI->second[j];
        MachineInstr *DefMI = MRI->getVRegDef(NewReg);
        SSAUpdate.AddAvailableValue(DefMI->getParent(), NewReg);
      }

      // Rewrite uses that are outside of the original def's block.
      for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(VReg),
             UE = MRI->use_end(); UI != UE; ++UI) {
        MachineInstr *UseMI = &*UI;
        if (UseMI->getParent() != DefBB)
          SSAUpdate.RewriteUse(UI.getOperand());
      }
    }
  }

  return MadeChange;
}

static bool isDefLiveOut(unsigned Reg, MachineBasicBlock *BB,
                         const MachineRegisterInfo *MRI) {
  for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
         UE = MRI->use_end(); UI != UE; ++UI) {
    MachineInstr *UseMI = &*UI;
    if (UseMI->getParent() != BB)
      return true;
  }
  return false;
}

static unsigned getPHISrcRegOpIdx(MachineInstr *MI, MachineBasicBlock *SrcBB) {
  for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2)
    if (MI->getOperand(i+1).getMBB() == SrcBB)
      return i;
  return 0;
}

/// AddSSAUpdateEntry - Add a definition and source virtual registers pair for
/// SSA update.
void TailDuplicatePass::AddSSAUpdateEntry(unsigned OrigReg, unsigned NewReg) {
  DenseMap<unsigned, AvailableValsTy>::iterator LI =
    SSAUpdateVals.find(OrigReg);
  if (LI != SSAUpdateVals.end())
    LI->second.push_back(NewReg);
  else {
    AvailableValsTy Vals;
    Vals.push_back(NewReg);
    SSAUpdateVals.insert(std::make_pair(OrigReg, Vals));
    SSAUpdateVRs.push_back(OrigReg);
  }
}

/// TailDuplicate - If it is profitable, duplicate TailBB's contents in each
/// of its predecessors.
bool TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB,
                                      MachineFunction &MF) {
  // Don't try to tail-duplicate single-block loops.
  if (TailBB->isSuccessor(TailBB))
    return false;

  // Set the limit on the number of instructions to duplicate, with a default
  // of one less than the tail-merge threshold. When optimizing for size,
  // duplicate only one, because one branch instruction can be eliminated to
  // compensate for the duplication.
  unsigned MaxDuplicateCount;
  if (!TailBB->empty() && TailBB->back().getDesc().isIndirectBranch())
    // If the target has hardware branch prediction that can handle indirect
    // branches, duplicating them can often make them predictable when there
    // are common paths through the code.  The limit needs to be high enough
    // to allow undoing the effects of tail merging.
    MaxDuplicateCount = 20;
  else if (MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize))
    MaxDuplicateCount = 1;
  else
    MaxDuplicateCount = TailDuplicateSize;

  // Check the instructions in the block to determine whether tail-duplication
  // is invalid or unlikely to be profitable.
  unsigned InstrCount = 0;
  bool HasCall = false;
  for (MachineBasicBlock::iterator I = TailBB->begin();
       I != TailBB->end(); ++I) {
    // Non-duplicable things shouldn't be tail-duplicated.
    if (I->getDesc().isNotDuplicable()) return false;
    // Don't duplicate more than the threshold.
    if (InstrCount == MaxDuplicateCount) return false;
    // Remember if we saw a call.
    if (I->getDesc().isCall()) HasCall = true;
    if (I->getOpcode() != TargetInstrInfo::PHI)
      InstrCount += 1;
  }
  // Heuristically, don't tail-duplicate calls if it would expand code size,
  // as it's less likely to be worth the extra cost.
  if (InstrCount > 1 && HasCall)
    return false;

  // Iterate through all the unique predecessors and tail-duplicate this
  // block into them, if possible. Copying the list ahead of time also
  // avoids trouble with the predecessor list reallocating.
  bool Changed = false;
  SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
                                               TailBB->pred_end());
  for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),