BranchFolding.cpp

//===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass forwards branches to unconditional branches to make them branch
// directly to the target block.  This pass often results in dead MBB's, which
// it then removes.
//
// Note that this pass must be run after register allocation, it cannot handle
// SSA form.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "branchfolding"
#include "BranchFolding.h"
#include "llvm/Function.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
#include <algorithm>
using namespace llvm;

STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
STATISTIC(NumBranchOpts, "Number of branches optimized");
STATISTIC(NumTailMerge , "Number of block tails merged");

static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
                              cl::init(cl::BOU_UNSET), cl::Hidden);

// Throttle for huge numbers of predecessors (compile speed problems)
static cl::opt<unsigned>
TailMergeThreshold("tail-merge-threshold",
          cl::desc("Max number of predecessors to consider tail merging"),
          cl::init(150), cl::Hidden);

// Heuristic for tail merging (and, inversely, tail duplication).
// TODO: This should be replaced with a target query.
static cl::opt<unsigned>
TailMergeSize("tail-merge-size",
          cl::desc("Min number of instructions to consider tail merging"),
                              cl::init(3), cl::Hidden);

namespace {
  /// BranchFolderPass - Wrap branch folder in a machine function pass.
  class BranchFolderPass : public MachineFunctionPass,
                           public BranchFolder {
  public:
    static char ID;
    explicit BranchFolderPass(bool defaultEnableTailMerge)
      : MachineFunctionPass(ID), BranchFolder(defaultEnableTailMerge) {}

    virtual bool runOnMachineFunction(MachineFunction &MF);
    virtual const char *getPassName() const { return "Control Flow Optimizer"; }
  };
}

char BranchFolderPass::ID = 0;

FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
  return new BranchFolderPass(DefaultEnableTailMerge);
}

bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
  return OptimizeFunction(MF,
                          MF.getTarget().getInstrInfo(),
                          MF.getTarget().getRegisterInfo(),
                          getAnalysisIfAvailable<MachineModuleInfo>());
}


BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
  switch (FlagEnableTailMerge) {
  case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
  case cl::BOU_TRUE: EnableTailMerge = true; break;
  case cl::BOU_FALSE: EnableTailMerge = false; break;
  }
}

/// RemoveDeadBlock - Remove the specified dead machine basic block from the
/// function, updating the CFG.
void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
  assert(MBB->pred_empty() && "MBB must be dead!");
  DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);

  MachineFunction *MF = MBB->getParent();
  // drop all successors.
  while (!MBB->succ_empty())
    MBB->removeSuccessor(MBB->succ_end()-1);

  // Remove the block.
  MF->erase(MBB);
}

/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
/// followed by terminators, and if the implicitly defined registers are not
/// used by the terminators, remove those implicit_def's. e.g.
/// BB1:
///   r0 = implicit_def
///   r1 = implicit_def
///   br
/// This block can be optimized away later if the implicit instructions are
/// removed.
bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
  SmallSet<unsigned, 4> ImpDefRegs;
  MachineBasicBlock::iterator I = MBB->begin();
  while (I != MBB->end()) {
    if (!I->isImplicitDef())
      break;
    unsigned Reg = I->getOperand(0).getReg();
    ImpDefRegs.insert(Reg);
    for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
         unsigned SubReg = *SubRegs; ++SubRegs)
      ImpDefRegs.insert(SubReg);
    ++I;
  }
  if (ImpDefRegs.empty())
    return false;

  MachineBasicBlock::iterator FirstTerm = I;
  while (I != MBB->end()) {
    if (!TII->isUnpredicatedTerminator(I))
      return false;
    // See if it uses any of the implicitly defined registers.
    for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
      MachineOperand &MO = I->getOperand(i);
      if (!MO.isReg() || !MO.isUse())
        continue;
      unsigned Reg = MO.getReg();
      if (ImpDefRegs.count(Reg))
        return false;
    }
    ++I;
  }

  I = MBB->begin();
  while (I != FirstTerm) {
    MachineInstr *ImpDefMI = &*I;
    ++I;
    MBB->erase(ImpDefMI);
  }

  return true;
}

/// OptimizeFunction - Perhaps branch folding, tail merging and other
/// CFG optimizations on the given function.
bool BranchFolder::OptimizeFunction(MachineFunction &MF,
                                    const TargetInstrInfo *tii,
                                    const TargetRegisterInfo *tri,
                                    MachineModuleInfo *mmi) {
  if (!tii) return false;

  TII = tii;
  TRI = tri;
  MMI = mmi;

  RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;

  // Fix CFG.  The later algorithms expect it to be right.
  bool MadeChange = false;
  for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
    MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
    SmallVector<MachineOperand, 4> Cond;
    if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
      MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
    MadeChange |= OptimizeImpDefsBlock(MBB);
  }

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

  // See if any jump tables have become dead as the code generator
  // did its thing.
  MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
  if (JTI == 0) {
    delete RS;
    return MadeChange;
  }
  
  // Walk the function to find jump tables that are live.
  BitVector JTIsLive(JTI->getJumpTables().size());
  for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
       BB != E; ++BB) {
    for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
         I != E; ++I)
      for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
        MachineOperand &Op = I->getOperand(op);
        if (!Op.isJTI()) continue;

        // Remember that this JT is live.
        JTIsLive.set(Op.getIndex());
      }
  }

  // Finally, remove dead jump tables.  This happens when the
  // indirect jump was unreachable (and thus deleted).
  for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
    if (!JTIsLive.test(i)) {
      JTI->RemoveJumpTable(i);
      MadeChange = true;
    }

  delete RS;
  return MadeChange;
}

//===----------------------------------------------------------------------===//
//  Tail Merging of Blocks
//===----------------------------------------------------------------------===//

/// HashMachineInstr - Compute a hash value for MI and its operands.
static unsigned HashMachineInstr(const MachineInstr *MI) {
  unsigned Hash = MI->getOpcode();
  for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
    const MachineOperand &Op = MI->getOperand(i);

    // Merge in bits from the operand if easy.
    unsigned OperandHash = 0;
    switch (Op.getType()) {
    case MachineOperand::MO_Register:          OperandHash = Op.getReg(); break;
    case MachineOperand::MO_Immediate:         OperandHash = Op.getImm(); break;
    case MachineOperand::MO_MachineBasicBlock:
      OperandHash = Op.getMBB()->getNumber();
      break;
    case MachineOperand::MO_FrameIndex:
    case MachineOperand::MO_ConstantPoolIndex:
    case MachineOperand::MO_JumpTableIndex:
      OperandHash = Op.getIndex();
      break;
    case MachineOperand::MO_GlobalAddress:
    case MachineOperand::MO_ExternalSymbol:
      // Global address / external symbol are too hard, don't bother, but do
      // pull in the offset.
      OperandHash = Op.getOffset();
      break;
    default: break;
    }

    Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
  }
  return Hash;
}

/// HashEndOfMBB - Hash the last instruction in the MBB.
static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
  MachineBasicBlock::const_iterator I = MBB->end();
  if (I == MBB->begin())
    return 0;   // Empty MBB.

  --I;
  // Skip debug info so it will not affect codegen.
  while (I->isDebugValue()) {
    if (I==MBB->begin())
      return 0;      // MBB empty except for debug info.
    --I;
  }

  return HashMachineInstr(I);
}

/// ComputeCommonTailLength - Given two machine basic blocks, compute the number
/// of instructions they actually have in common together at their end.  Return
/// iterators for the first shared instruction in each block.
static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
                                        MachineBasicBlock *MBB2,
                                        MachineBasicBlock::iterator &I1,
                                        MachineBasicBlock::iterator &I2) {
  I1 = MBB1->end();
  I2 = MBB2->end();

  unsigned TailLen = 0;
  while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
    --I1; --I2;
    // Skip debugging pseudos; necessary to avoid changing the code.
    while (I1->isDebugValue()) {
      if (I1==MBB1->begin()) {
        while (I2->isDebugValue()) {
          if (I2==MBB2->begin())
            // I1==DBG at begin; I2==DBG at begin
            return TailLen;
          --I2;
        }
        ++I2;
        // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
        return TailLen;
      }
      --I1;
    }
    // I1==first (untested) non-DBG preceding known match
    while (I2->isDebugValue()) {
      if (I2==MBB2->begin()) {
        ++I1;
        // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
        return TailLen;
      }
      --I2;
    }
    // I1, I2==first (untested) non-DBGs preceding known match
    if (!I1->isIdenticalTo(I2) ||
        // FIXME: This check is dubious. It's used to get around a problem where
        // people incorrectly expect inline asm directives to remain in the same
        // relative order. This is untenable because normal compiler
        // optimizations (like this one) may reorder and/or merge these
        // directives.
        I1->isInlineAsm()) {
      ++I1; ++I2;
      break;
    }
    ++TailLen;
  }
  // Back past possible debugging pseudos at beginning of block.  This matters
  // when one block differs from the other only by whether debugging pseudos
  // are present at the beginning.  (This way, the various checks later for
  // I1==MBB1->begin() work as expected.)
  if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
    --I2;
    while (I2->isDebugValue()) {
      if (I2 == MBB2->begin()) {
        return TailLen;
        }
      --I2;
    }
    ++I2;
  }
  if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
    --I1;
    while (I1->isDebugValue()) {
      if (I1 == MBB1->begin())
        return TailLen;
      --I1;
    }
    ++I1;
  }
  return TailLen;
}

/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
/// after it, replacing it with an unconditional branch to NewDest.
void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
                                           MachineBasicBlock *NewDest) {
  TII->ReplaceTailWithBranchTo(OldInst, NewDest);
  ++NumTailMerge;
}

/// SplitMBBAt - Given a machine basic block and an iterator into it, split the
/// MBB so that the part before the iterator falls into the part starting at the
/// iterator.  This returns the new MBB.
MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
                                            MachineBasicBlock::iterator BBI1) {
  if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
    return 0;

  MachineFunction &MF = *CurMBB.getParent();

  // Create the fall-through block.
  MachineFunction::iterator MBBI = &CurMBB;
  MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
  CurMBB.getParent()->insert(++MBBI, NewMBB);

  // Move all the successors of this block to the specified block.
  NewMBB->transferSuccessors(&CurMBB);

  // Add an edge from CurMBB to NewMBB for the fall-through.
  CurMBB.addSuccessor(NewMBB);

  // Splice the code over.
  NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());

  // For targets that use the register scavenger, we must maintain LiveIns.
  if (RS) {
    RS->enterBasicBlock(&CurMBB);
    if (!CurMBB.empty())
      RS->forward(prior(CurMBB.end()));
    BitVector RegsLiveAtExit(TRI->getNumRegs());
    RS->getRegsUsed(RegsLiveAtExit, false);
    for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
      if (RegsLiveAtExit[i])
        NewMBB->addLiveIn(i);
  }

  return NewMBB;
}

/// EstimateRuntime - Make a rough estimate for how long it will take to run
/// the specified code.
static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
                                MachineBasicBlock::iterator E) {
  unsigned Time = 0;
  for (; I != E; ++I) {
    if (I->isDebugValue())
      continue;
    const TargetInstrDesc &TID = I->getDesc();
    if (TID.isCall())
      Time += 10;
    else if (TID.mayLoad() || TID.mayStore())
      Time += 2;
    else
      ++Time;
  }
  return Time;
}

// CurMBB needs to add an unconditional branch to SuccMBB (we removed these
// branches temporarily for tail merging).  In the case where CurMBB ends
// with a conditional branch to the next block, optimize by reversing the
// test and conditionally branching to SuccMBB instead.
static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
                    const TargetInstrInfo *TII) {
  MachineFunction *MF = CurMBB->getParent();
  MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB));
  MachineBasicBlock *TBB = 0, *FBB = 0;
  SmallVector<MachineOperand, 4> Cond;
  DebugLoc dl;  // FIXME: this is nowhere
  if (I != MF->end() &&
      !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
    MachineBasicBlock *NextBB = I;
    if (TBB == NextBB && !Cond.empty() && !FBB) {
      if (!TII->ReverseBranchCondition(Cond)) {
        TII->RemoveBranch(*CurMBB);
        TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl);
        return;
      }
    }
  }
  TII->InsertBranch(*CurMBB, SuccBB, NULL,
                    SmallVector<MachineOperand, 0>(), dl);
}

bool
BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
  if (getHash() < o.getHash())
    return true;
   else if (getHash() > o.getHash())
    return false;
  else if (getBlock()->getNumber() < o.getBlock()->getNumber())
    return true;
  else if (getBlock()->getNumber() > o.getBlock()->getNumber())
    return false;
  else {
    // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
    // an object with itself.
#ifndef _GLIBCXX_DEBUG
    llvm_unreachable("Predecessor appears twice");
#endif
    return false;
  }
}

/// CountTerminators - Count the number of terminators in the given
/// block and set I to the position of the first non-terminator, if there
/// is one, or MBB->end() otherwise.
static unsigned CountTerminators(MachineBasicBlock *MBB,
                                 MachineBasicBlock::iterator &I) {
  I = MBB->end();
  unsigned NumTerms = 0;
  for (;;) {
    if (I == MBB->begin()) {
      I = MBB->end();
      break;
    }
    --I;
    if (!I->getDesc().isTerminator()) break;
    ++NumTerms;
  }
  return NumTerms;
}

/// ProfitableToMerge - Check if two machine basic blocks have a common tail
/// and decide if it would be profitable to merge those tails.  Return the
/// length of the common tail and iterators to the first common instruction
/// in each block.
static bool ProfitableToMerge(MachineBasicBlock *MBB1,
                              MachineBasicBlock *MBB2,
                              unsigned minCommonTailLength,
                              unsigned &CommonTailLen,
                              MachineBasicBlock::iterator &I1,
                              MachineBasicBlock::iterator &I2,
                              MachineBasicBlock *SuccBB,
                              MachineBasicBlock *PredBB) {
  CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
  if (CommonTailLen == 0)
    return false;
  DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
               << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
               << '\n');

  // It's almost always profitable to merge any number of non-terminator
  // instructions with the block that falls through into the common successor.
  if (MBB1 == PredBB || MBB2 == PredBB) {
    MachineBasicBlock::iterator I;
    unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
    if (CommonTailLen > NumTerms)
      return true;
  }

  // If one of the blocks can be completely merged and happens to be in
  // a position where the other could fall through into it, merge any number
  // of instructions, because it can be done without a branch.
  // TODO: If the blocks are not adjacent, move one of them so that they are?
  if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
    return true;
  if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
    return true;

  // If both blocks have an unconditional branch temporarily stripped out,
  // count that as an additional common instruction for the following
  // heuristics.
  unsigned EffectiveTailLen = CommonTailLen;
  if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
      !MBB1->back().getDesc().isBarrier() &&
      !MBB2->back().getDesc().isBarrier())
    ++EffectiveTailLen;

  // Check if the common tail is long enough to be worthwhile.
  if (EffectiveTailLen >= minCommonTailLength)
    return true;

  // If we are optimizing for code size, 2 instructions in common is enough if
  // we don't have to split a block.  At worst we will be introducing 1 new
  // branch instruction, which is likely to be smaller than the 2
  // instructions that would be deleted in the merge.
  MachineFunction *MF = MBB1->getParent();
  if (EffectiveTailLen >= 2 &&
      MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
      (I1 == MBB1->begin() || I2 == MBB2->begin()))
    return true;

  return false;
}

/// ComputeSameTails - Look through all the blocks in MergePotentials that have
/// hash CurHash (guaranteed to match the last element).  Build the vector
/// SameTails of all those that have the (same) largest number of instructions
/// in common of any pair of these blocks.  SameTails entries contain an
/// iterator into MergePotentials (from which the MachineBasicBlock can be
/// found) and a MachineBasicBlock::iterator into that MBB indicating the
/// instruction where the matching code sequence begins.
/// Order of elements in SameTails is the reverse of the order in which
/// those blocks appear in MergePotentials (where they are not necessarily
/// consecutive).
unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
                                        unsigned minCommonTailLength,
                                        MachineBasicBlock *SuccBB,
                                        MachineBasicBlock *PredBB) {
  unsigned maxCommonTailLength = 0U;
  SameTails.clear();
  MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
  MPIterator HighestMPIter = prior(MergePotentials.end());
  for (MPIterator CurMPIter = prior(MergePotentials.end()),
                  B = MergePotentials.begin();
       CurMPIter != B && CurMPIter->getHash() == CurHash;
       --CurMPIter) {
    for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
      unsigned CommonTailLen;
      if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
                            minCommonTailLength,
                            CommonTailLen, TrialBBI1, TrialBBI2,
                            SuccBB, PredBB)) {
        if (CommonTailLen > maxCommonTailLength) {
          SameTails.clear();
          maxCommonTailLength = CommonTailLen;
          HighestMPIter = CurMPIter;
          SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
        }
        if (HighestMPIter == CurMPIter &&
            CommonTailLen == maxCommonTailLength)
          SameTails.push_back(SameTailElt(I, TrialBBI2));
      }
      if (I == B)
        break;
    }
  }
  return maxCommonTailLength;
}

/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
/// MergePotentials, restoring branches at ends of blocks as appropriate.
void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
                                        MachineBasicBlock *SuccBB,
                                        MachineBasicBlock *PredBB) {
  MPIterator CurMPIter, B;
  for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
       CurMPIter->getHash() == CurHash;
       --CurMPIter) {
    // Put the unconditional branch back, if we need one.
    MachineBasicBlock *CurMBB = CurMPIter->getBlock();
    if (SuccBB && CurMBB != PredBB)
      FixTail(CurMBB, SuccBB, TII);
    if (CurMPIter == B)
      break;
  }
  if (CurMPIter->getHash() != CurHash)
    CurMPIter++;
  MergePotentials.erase(CurMPIter, MergePotentials.end());
}

/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
/// only of the common tail.  Create a block that does by splitting one.
bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
                                             unsigned maxCommonTailLength,
                                             unsigned &commonTailIndex) {
  commonTailIndex = 0;
  unsigned TimeEstimate = ~0U;
  for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
    // Use PredBB if possible; that doesn't require a new branch.
    if (SameTails[i].getBlock() == PredBB) {
      commonTailIndex = i;
      break;
    }
    // Otherwise, make a (fairly bogus) choice based on estimate of
    // how long it will take the various blocks to execute.
    unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
                                 SameTails[i].getTailStartPos());
    if (t <= TimeEstimate) {
      TimeEstimate = t;
      commonTailIndex = i;
    }
  }

  MachineBasicBlock::iterator BBI =
    SameTails[commonTailIndex].getTailStartPos();
  MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();

  // If the common tail includes any debug info we will take it pretty
  // randomly from one of the inputs.  Might be better to remove it?
  DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
               << maxCommonTailLength);

  MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
  if (!newMBB) {
    DEBUG(dbgs() << "... failed!");
    return false;
  }

  SameTails[commonTailIndex].setBlock(newMBB);
  SameTails[commonTailIndex].setTailStartPos(newMBB->begin());

  // If we split PredBB, newMBB is the new predecessor.
  if (PredBB == MBB)
    PredBB = newMBB;

  return true;
}

// See if any of the blocks in MergePotentials (which all have a common single
// successor, or all have no successor) can be tail-merged.  If there is a
// successor, any blocks in MergePotentials that are not tail-merged and
// are not immediately before Succ must have an unconditional branch to
// Succ added (but the predecessor/successor lists need no adjustment).
// The lone predecessor of Succ that falls through into Succ,
// if any, is given in PredBB.

bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
                                      MachineBasicBlock *PredBB) {
  bool MadeChange = false;

  // Except for the special cases below, tail-merge if there are at least
  // this many instructions in common.
  unsigned minCommonTailLength = TailMergeSize;

  DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
        for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
          dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
                 << (i == e-1 ? "" : ", ");
        dbgs() << "\n";
        if (SuccBB) {
          dbgs() << "  with successor BB#" << SuccBB->getNumber() << '\n';
          if (PredBB)
            dbgs() << "  which has fall-through from BB#"
                   << PredBB->getNumber() << "\n";
        }
        dbgs() << "Looking for common tails of at least "
               << minCommonTailLength << " instruction"
               << (minCommonTailLength == 1 ? "" : "s") << '\n';
       );

  // Sort by hash value so that blocks with identical end sequences sort
  // together.
  std::stable_sort(MergePotentials.begin(), MergePotentials.end());

  // Walk through equivalence sets looking for actual exact matches.
  while (MergePotentials.size() > 1) {
    unsigned CurHash = MergePotentials.back().getHash();

    // Build SameTails, identifying the set of blocks with this hash code
    // and with the maximum number of instructions in common.
    unsigned maxCommonTailLength = ComputeSameTails(CurHash,
                                                    minCommonTailLength,
                                                    SuccBB, PredBB);

    // If we didn't find any pair that has at least minCommonTailLength
    // instructions in common, remove all blocks with this hash code and retry.
    if (SameTails.empty()) {
      RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
      continue;
    }

    // If one of the blocks is the entire common tail (and not the entry
    // block, which we can't jump to), we can treat all blocks with this same
    // tail at once.  Use PredBB if that is one of the possibilities, as that
    // will not introduce any extra branches.
    MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
                                 getParent()->begin();
    unsigned commonTailIndex = SameTails.size();
    // If there are two blocks, check to see if one can be made to fall through
    // into the other.
    if (SameTails.size() == 2 &&
        SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
        SameTails[1].tailIsWholeBlock())
      commonTailIndex = 1;
    else if (SameTails.size() == 2 &&
             SameTails[1].getBlock()->isLayoutSuccessor(
                                                     SameTails[0].getBlock()) &&
             SameTails[0].tailIsWholeBlock())
      commonTailIndex = 0;
    else {
      // Otherwise just pick one, favoring the fall-through predecessor if
      // there is one.
      for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
        MachineBasicBlock *MBB = SameTails[i].getBlock();
        if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
          continue;
        if (MBB == PredBB) {
          commonTailIndex = i;
          break;
        }
        if (SameTails[i].tailIsWholeBlock())
          commonTailIndex = i;
      }
    }

    if (commonTailIndex == SameTails.size() ||
        (SameTails[commonTailIndex].getBlock() == PredBB &&
         !SameTails[commonTailIndex].tailIsWholeBlock())) {
      // None of the blocks consist entirely of the common tail.
      // Split a block so that one does.
      if (!CreateCommonTailOnlyBlock(PredBB,
                                     maxCommonTailLength, commonTailIndex)) {
        RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
        continue;
      }
    }

    MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
    // MBB is common tail.  Adjust all other BB's to jump to this one.
    // Traversal must be forwards so erases work.
    DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
                 << " for ");
    for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
      if (commonTailIndex == i)
        continue;
      DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
                   << (i == e-1 ? "" : ", "));
      // Hack the end off BB i, making it jump to BB commonTailIndex instead.
      ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
      // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
      MergePotentials.erase(SameTails[i].getMPIter());
    }
    DEBUG(dbgs() << "\n");
    // We leave commonTailIndex in the worklist in case there are other blocks
    // that match it with a smaller number of instructions.
    MadeChange = true;
  }
  return MadeChange;
}

bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {

  if (!EnableTailMerge) return false;

  bool MadeChange = false;

  // First find blocks with no successors.
  MergePotentials.clear();
  for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
    if (I->succ_empty())
      MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
  }

  // See if we can do any tail merging on those.
  if (MergePotentials.size() < TailMergeThreshold &&
      MergePotentials.size() >= 2)
    MadeChange |= TryTailMergeBlocks(NULL, NULL);

  // Look at blocks (IBB) with multiple predecessors (PBB).
  // We change each predecessor to a canonical form, by
  // (1) temporarily removing any unconditional branch from the predecessor
  // to IBB, and
  // (2) alter conditional branches so they branch to the other block
  // not IBB; this may require adding back an unconditional branch to IBB
  // later, where there wasn't one coming in.  E.g.
  //   Bcc IBB
  //   fallthrough to QBB
  // here becomes
  //   Bncc QBB
  // with a conceptual B to IBB after that, which never actually exists.
  // With those changes, we see whether the predecessors' tails match,
  // and merge them if so.  We change things out of canonical form and
  // back to the way they were later in the process.  (OptimizeBranches
  // would undo some of this, but we can't use it, because we'd get into
  // a compile-time infinite loop repeatedly doing and undoing the same
  // transformations.)

  for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
       I != E; ++I) {
    if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
      SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
      MachineBasicBlock *IBB = I;
      MachineBasicBlock *PredBB = prior(I);
      MergePotentials.clear();
      for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
                                            E2 = I->pred_end();
           P != E2; ++P) {
        MachineBasicBlock *PBB = *P;
        // Skip blocks that loop to themselves, can't tail merge these.
        if (PBB == IBB)
          continue;
        // Visit each predecessor only once.
        if (!UniquePreds.insert(PBB))
          continue;
        MachineBasicBlock *TBB = 0, *FBB = 0;
        SmallVector<MachineOperand, 4> Cond;
        if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
          // Failing case:  IBB is the target of a cbr, and
          // we cannot reverse the branch.
          SmallVector<MachineOperand, 4> NewCond(Cond);
          if (!Cond.empty() && TBB == IBB) {
            if (TII->ReverseBranchCondition(NewCond))
              continue;
            // This is the QBB case described above
            if (!FBB)
              FBB = llvm::next(MachineFunction::iterator(PBB));
          }
          // Failing case:  the only way IBB can be reached from PBB is via
          // exception handling.  Happens for landing pads.  Would be nice
          // to have a bit in the edge so we didn't have to do all this.
          if (IBB->isLandingPad()) {
            MachineFunction::iterator IP = PBB;  IP++;
            MachineBasicBlock *PredNextBB = NULL;
            if (IP != MF.end())
              PredNextBB = IP;
            if (TBB == NULL) {
              if (IBB != PredNextBB)      // fallthrough
                continue;
            } else if (FBB) {
              if (TBB != IBB && FBB != IBB)   // cbr then ubr
                continue;
            } else if (Cond.empty()) {
              if (TBB != IBB)               // ubr
                continue;
            } else {
              if (TBB != IBB && IBB != PredNextBB)  // cbr
                continue;
            }
          }
          // Remove the unconditional branch at the end, if any.
          if (TBB && (Cond.empty() || FBB)) {
            DebugLoc dl;  // FIXME: this is nowhere
            TII->RemoveBranch(*PBB);
            if (!Cond.empty())
              // reinsert conditional branch only, for now
              TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl);
          }
          MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
        }
      }
      if (MergePotentials.size() >= 2)
        MadeChange |= TryTailMergeBlocks(IBB, PredBB);
      // Reinsert an unconditional branch if needed.
      // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
      PredBB = prior(I);      // this may have been changed in TryTailMergeBlocks
      if (MergePotentials.size() == 1 &&
          MergePotentials.begin()->getBlock() != PredBB)
        FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
    }
  }
  return MadeChange;
}

//===----------------------------------------------------------------------===//
//  Branch Optimization
//===----------------------------------------------------------------------===//

bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
  bool MadeChange = false;

  // Make sure blocks are numbered in order
  MF.RenumberBlocks();

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

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

// Blocks should be considered empty if they contain only debug info;
// else the debug info would affect codegen.
static bool IsEmptyBlock(MachineBasicBlock *MBB) {
  if (MBB->empty())
    return true;
  for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
       MBBI!=MBBE; ++MBBI) {
    if (!MBBI->isDebugValue())
      return false;
  }
  return true;
}

// Blocks with only debug info and branches should be considered the same
// as blocks with only branches.
static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
  MachineBasicBlock::iterator MBBI, MBBE;
  for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
    if (!MBBI->isDebugValue())
      break;
  }
  return (MBBI->getDesc().isBranch());
}

/// IsBetterFallthrough - Return true if it would be clearly better to
/// fall-through to MBB1 than to fall through into MBB2.  This has to return
/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
/// result in infinite loops.
static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
                                MachineBasicBlock *MBB2) {
  // Right now, we use a simple heuristic.  If MBB2 ends with a call, and
  // MBB1 doesn't, we prefer to fall through into MBB1.  This allows us to
  // optimize branches that branch to either a return block or an assert block
  // into a fallthrough to the return.
  if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;

  // If there is a clear successor ordering we make sure that one block
  // will fall through to the next
  if (MBB1->isSuccessor(MBB2)) return true;
  if (MBB2->isSuccessor(MBB1)) return false;

  // Neither block consists entirely of debug info (per IsEmptyBlock check),
  // so we needn't test for falling off the beginning here.
  MachineBasicBlock::iterator MBB1I = --MBB1->end();
  while (MBB1I->isDebugValue())
    --MBB1I;
  MachineBasicBlock::iterator MBB2I = --MBB2->end();
  while (MBB2I->isDebugValue())
    --MBB2I;
  return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
}

/// OptimizeBlock - Analyze and optimize control flow related to the specified
/// block.  This is never called on the entry block.
bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
  bool MadeChange = false;
  MachineFunction &MF = *MBB->getParent();
  DebugLoc dl;  // FIXME: this is nowhere
ReoptimizeBlock:

  MachineFunction::iterator FallThrough = MBB;
  ++FallThrough;

  // If this block is empty, make everyone use its fall-through, not the block
  // explicitly.  Landing pads should not do this since the landing-pad table
  // points to this block.  Blocks with their addresses taken shouldn't be
  // optimized away.
  if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
    // Dead block?  Leave for cleanup later.
    if (MBB->pred_empty()) return MadeChange;

    if (FallThrough == MF.end()) {
      // TODO: Simplify preds to not branch here if possible!
    } else {