SplitKit.cpp

//===----------------------------------------------------------------------===//
//                               Loop Splitting
//===----------------------------------------------------------------------===//

void SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
  SplitAnalysis::LoopBlocks Blocks;
  sa_.getLoopBlocks(Loop, Blocks);

  DEBUG({
    dbgs() << "  splitAround"; sa_.print(Blocks, dbgs()); dbgs() << '\n';
  });

  // Break critical edges as needed.
  SplitAnalysis::BlockPtrSet CriticalExits;
  sa_.getCriticalExits(Blocks, CriticalExits);
  assert(CriticalExits.empty() && "Cannot break critical exits yet");

  // Get critical predecessors so computeRemainder can deal with them.
  sa_.getCriticalPreds(Blocks, criticalPreds_);

  // Create new live interval for the loop.
  openIntv();

  // Insert copies in the predecessors.
  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
       E = Blocks.Preds.end(); I != E; ++I) {
    MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
    enterIntvAtEnd(MBB);
  }

  // Switch all loop blocks.
  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
       E = Blocks.Loop.end(); I != E; ++I)
     useIntv(**I);

  // Insert back copies in the exit blocks.
  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
       E = Blocks.Exits.end(); I != E; ++I) {
    MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
    leaveIntvAtTop(MBB);
  }

  // Done.
  closeIntv();
  finish();
}


//===----------------------------------------------------------------------===//
//                            Single Block Splitting
//===----------------------------------------------------------------------===//

/// getMultiUseBlocks - if curli has more than one use in a basic block, it
/// may be an advantage to split curli for the duration of the block.
bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
  // If curli is local to one block, there is no point to splitting it.
  if (usingBlocks_.size() <= 1)
    return false;
  // Add blocks with multiple uses.
  for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
       I != E; ++I)
    switch (I->second) {
    case 0:
    case 1:
      continue;
    case 2: {
      // When there are only two uses and curli is both live in and live out,
      // we don't really win anything by isolating the block since we would be
      // inserting two copies.
      // The remaing register would still have two uses in the block. (Unless it
      // separates into disconnected components).
      if (lis_.isLiveInToMBB(*curli_, I->first) &&
          lis_.isLiveOutOfMBB(*curli_, I->first))
        continue;
    } // Fall through.
    default:
      Blocks.insert(I->first);
    }
  return !Blocks.empty();
}

/// splitSingleBlocks - Split curli into a separate live interval inside each
/// basic block in Blocks.
void SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
  DEBUG(dbgs() << "  splitSingleBlocks for " << Blocks.size() << " blocks.\n");
  // Determine the first and last instruction using curli in each block.
  typedef std::pair<SlotIndex,SlotIndex> IndexPair;
  typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
  IndexPairMap MBBRange;
  for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
       E = sa_.usingInstrs_.end(); I != E; ++I) {
    const MachineBasicBlock *MBB = (*I)->getParent();
    if (!Blocks.count(MBB))
      continue;
    SlotIndex Idx = lis_.getInstructionIndex(*I);
    DEBUG(dbgs() << "  BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
    IndexPair &IP = MBBRange[MBB];
    if (!IP.first.isValid() || Idx < IP.first)
      IP.first = Idx;
    if (!IP.second.isValid() || Idx > IP.second)
      IP.second = Idx;
  }

  // Create a new interval for each block.
  for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
       E = Blocks.end(); I != E; ++I) {
    IndexPair &IP = MBBRange[*I];
    DEBUG(dbgs() << "  splitting for BB#" << (*I)->getNumber() << ": ["
                 << IP.first << ';' << IP.second << ")\n");
    assert(IP.first.isValid() && IP.second.isValid());

    openIntv();
    enterIntvBefore(IP.first);
    useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
    leaveIntvAfter(IP.second);
    closeIntv();
  }
  finish();
}


//===----------------------------------------------------------------------===//
//                            Sub Block Splitting
//===----------------------------------------------------------------------===//

/// getBlockForInsideSplit - If curli is contained inside a single basic block,
/// and it wou pay to subdivide the interval inside that block, return it.
/// Otherwise return NULL. The returned block can be passed to
/// SplitEditor::splitInsideBlock.
const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
  // The interval must be exclusive to one block.
  if (usingBlocks_.size() != 1)
    return 0;
  // Don't to this for less than 4 instructions. We want to be sure that
  // splitting actually reduces the instruction count per interval.
  if (usingInstrs_.size() < 4)
    return 0;
  return usingBlocks_.begin()->first;
}

/// splitInsideBlock - Split curli into multiple intervals inside MBB.
void SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
  SmallVector<SlotIndex, 32> Uses;
  Uses.reserve(sa_.usingInstrs_.size());
  for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
       E = sa_.usingInstrs_.end(); I != E; ++I)
    if ((*I)->getParent() == MBB)
      Uses.push_back(lis_.getInstructionIndex(*I));
  DEBUG(dbgs() << "  splitInsideBlock BB#" << MBB->getNumber() << " for "
               << Uses.size() << " instructions.\n");
  assert(Uses.size() >= 3 && "Need at least 3 instructions");
  array_pod_sort(Uses.begin(), Uses.end());

  // Simple algorithm: Find the largest gap between uses as determined by slot
  // indices. Create new intervals for instructions before the gap and after the
  // gap.
  unsigned bestPos = 0;
  int bestGap = 0;
  DEBUG(dbgs() << "    dist (" << Uses[0]);
  for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
    int g = Uses[i-1].distance(Uses[i]);
    DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
    if (g > bestGap)
      bestPos = i, bestGap = g;
  }
  DEBUG(dbgs() << "), best: -" << bestGap << "-\n");

  // bestPos points to the first use after the best gap.
  assert(bestPos > 0 && "Invalid gap");

  // FIXME: Don't create intervals for low densities.

  // First interval before the gap. Don't create single-instr intervals.
  if (bestPos > 1) {
    openIntv();
    enterIntvBefore(Uses.front());
    useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
    leaveIntvAfter(Uses[bestPos-1]);
    closeIntv();
  }

  // Second interval after the gap.
  if (bestPos < Uses.size()-1) {
    openIntv();
    enterIntvBefore(Uses[bestPos]);
    useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
    leaveIntvAfter(Uses.back());
    closeIntv();
  }

  finish();
}