PreAllocSplitting.cpp

bool PreAllocSplitting::SplitRegLiveInterval(LiveInterval *LI) {
  CurrLI = LI;

  // Find live range where current interval cross the barrier.
  LiveInterval::iterator LR =
    CurrLI->FindLiveRangeContaining(LIs->getUseIndex(BarrierIdx));
  VNInfo *ValNo = LR->valno;

  if (ValNo->def == ~1U) {
    // Defined by a dead def? How can this be?
    assert(0 && "Val# is defined by a dead def?");
    abort();
  }

  MachineInstr *DefMI = (ValNo->def != ~0U)
    ? LIs->getInstructionFromIndex(ValNo->def) : NULL;

  // If this would create a new join point, do not split.
  if (DefMI && createsNewJoin(LR, DefMI->getParent(), Barrier->getParent()))
    return false;

  // Find all references in the barrier mbb.
  SmallPtrSet<MachineInstr*, 4> RefsInMBB;
  for (MachineRegisterInfo::reg_iterator I = MRI->reg_begin(CurrLI->reg),
         E = MRI->reg_end(); I != E; ++I) {
    MachineInstr *RefMI = &*I;
    if (RefMI->getParent() == BarrierMBB)
      RefsInMBB.insert(RefMI);
  }

  // Find a point to restore the value after the barrier.
  unsigned RestoreIndex = 0;
  MachineBasicBlock::iterator RestorePt =
    findRestorePoint(BarrierMBB, Barrier, LR->end, RefsInMBB, RestoreIndex);
  if (RestorePt == BarrierMBB->end())
    return false;

  if (DefMI && LIs->isReMaterializable(*LI, ValNo, DefMI))
    if (Rematerialize(LI->reg, ValNo, DefMI, RestorePt,
                      RestoreIndex, RefsInMBB))
    return true;

  // Add a spill either before the barrier or after the definition.
  MachineBasicBlock *DefMBB = DefMI ? DefMI->getParent() : NULL;
  const TargetRegisterClass *RC = MRI->getRegClass(CurrLI->reg);
  unsigned SpillIndex = 0;
  MachineInstr *SpillMI = NULL;
  int SS = -1;
  if (ValNo->def == ~0U) {
    // If it's defined by a phi, we must split just before the barrier.
    if ((SpillMI = FoldSpill(LI->reg, RC, 0, Barrier,
                            BarrierMBB, SS, RefsInMBB))) {
      SpillIndex = LIs->getInstructionIndex(SpillMI);
    } else {
      MachineBasicBlock::iterator SpillPt = 
        findSpillPoint(BarrierMBB, Barrier, NULL, RefsInMBB, SpillIndex);
      if (SpillPt == BarrierMBB->begin())
        return false; // No gap to insert spill.
      // Add spill.
    
      SS = CreateSpillStackSlot(CurrLI->reg, RC);
      TII->storeRegToStackSlot(*BarrierMBB, SpillPt, CurrLI->reg, true, SS, RC);
      SpillMI = prior(SpillPt);
      LIs->InsertMachineInstrInMaps(SpillMI, SpillIndex);
    }
  } else if (!IsAvailableInStack(DefMBB, CurrLI->reg, ValNo->def,
                                 RestoreIndex, SpillIndex, SS)) {
    // If it's already split, just restore the value. There is no need to spill
    // the def again.
    if (!DefMI)
      return false; // Def is dead. Do nothing.
    
    if ((SpillMI = FoldSpill(LI->reg, RC, DefMI, Barrier,
                            BarrierMBB, SS, RefsInMBB))) {
      SpillIndex = LIs->getInstructionIndex(SpillMI);
    } else {
      // Check if it's possible to insert a spill after the def MI.
      MachineBasicBlock::iterator SpillPt;
      if (DefMBB == BarrierMBB) {
        // Add spill after the def and the last use before the barrier.
        SpillPt = findSpillPoint(BarrierMBB, Barrier, DefMI,
                                 RefsInMBB, SpillIndex);
        if (SpillPt == DefMBB->begin())
          return false; // No gap to insert spill.
      } else {
        SpillPt = findNextEmptySlot(DefMBB, DefMI, SpillIndex);
        if (SpillPt == DefMBB->end())
          return false; // No gap to insert spill.
      }
      // Add spill. The store instruction kills the register if def is before
      // the barrier in the barrier block.
      SS = CreateSpillStackSlot(CurrLI->reg, RC);
      TII->storeRegToStackSlot(*DefMBB, SpillPt, CurrLI->reg,
                               DefMBB == BarrierMBB, SS, RC);
      SpillMI = prior(SpillPt);
      LIs->InsertMachineInstrInMaps(SpillMI, SpillIndex);
    }
  }

  // Remember def instruction index to spill index mapping.
  if (DefMI && SpillMI)
    Def2SpillMap[ValNo->def] = SpillIndex;

  // Add restore.
  TII->loadRegFromStackSlot(*BarrierMBB, RestorePt, CurrLI->reg, SS, RC);
  MachineInstr *LoadMI = prior(RestorePt);
  LIs->InsertMachineInstrInMaps(LoadMI, RestoreIndex);

  // Update spill stack slot live interval.
  UpdateSpillSlotInterval(ValNo, LIs->getUseIndex(SpillIndex)+1,
                          LIs->getDefIndex(RestoreIndex));

  ReconstructLiveInterval(CurrLI);
  unsigned RestoreIdx = LIs->getInstructionIndex(prior(RestorePt));
  RestoreIdx = LiveIntervals::getDefIndex(RestoreIdx);
  RenumberValno(CurrLI->findDefinedVNInfo(RestoreIdx));
  
  ++NumSplits;
  return true;
}

/// SplitRegLiveIntervals - Split all register live intervals that cross the
/// barrier that's being processed.
bool
PreAllocSplitting::SplitRegLiveIntervals(const TargetRegisterClass **RCs,
                                         SmallPtrSet<LiveInterval*, 8>& Split) {
  // First find all the virtual registers whose live intervals are intercepted
  // by the current barrier.
  SmallVector<LiveInterval*, 8> Intervals;
  for (const TargetRegisterClass **RC = RCs; *RC; ++RC) {
    // FIXME: If it's not safe to move any instruction that defines the barrier
    // register class, then it means there are some special dependencies which
    // codegen is not modelling. Ignore these barriers for now.
    if (!TII->isSafeToMoveRegClassDefs(*RC))
      continue;
    std::vector<unsigned> &VRs = MRI->getRegClassVirtRegs(*RC);
    for (unsigned i = 0, e = VRs.size(); i != e; ++i) {
      unsigned Reg = VRs[i];
      if (!LIs->hasInterval(Reg))
        continue;
      LiveInterval *LI = &LIs->getInterval(Reg);
      if (LI->liveAt(BarrierIdx) && !Barrier->readsRegister(Reg))
        // Virtual register live interval is intercepted by the barrier. We
        // should split and shrink wrap its interval if possible.
        Intervals.push_back(LI);
    }
  }

  // Process the affected live intervals.
  bool Change = false;
  while (!Intervals.empty()) {
    if (PreSplitLimit != -1 && (int)NumSplits == PreSplitLimit)
      break;
    else if (NumSplits == 4)
      Change |= Change;
    LiveInterval *LI = Intervals.back();
    Intervals.pop_back();
    bool result = SplitRegLiveInterval(LI);
    if (result) Split.insert(LI);
    Change |= result;
  }

  return Change;
}

unsigned PreAllocSplitting::getNumberOfNonSpills(
                                  SmallPtrSet<MachineInstr*, 4>& MIs,
                                  unsigned Reg, int FrameIndex,
                                  bool& FeedsTwoAddr) {
  unsigned NonSpills = 0;
  for (SmallPtrSet<MachineInstr*, 4>::iterator UI = MIs.begin(), UE = MIs.end();
       UI != UE; ++UI) {
    int StoreFrameIndex;
    unsigned StoreVReg = TII->isStoreToStackSlot(*UI, StoreFrameIndex);
    if (StoreVReg != Reg || StoreFrameIndex != FrameIndex)
      NonSpills++;
    
    int DefIdx = (*UI)->findRegisterDefOperandIdx(Reg);
    if (DefIdx != -1 && (*UI)->isRegReDefinedByTwoAddr(DefIdx))
      FeedsTwoAddr = true;
  }
  
  return NonSpills;
}

/// removeDeadSpills - After doing splitting, filter through all intervals we've
/// split, and see if any of the spills are unnecessary.  If so, remove them.
bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
  bool changed = false;
  
  // Walk over all of the live intervals that were touched by the splitter,
  // and see if we can do any DCE and/or folding.
  for (SmallPtrSet<LiveInterval*, 8>::iterator LI = split.begin(),
       LE = split.end(); LI != LE; ++LI) {
    DenseMap<VNInfo*, SmallPtrSet<MachineInstr*, 4> > VNUseCount;
    
    // First, collect all the uses of the vreg, and sort them by their
    // reaching definition (VNInfo).
    for (MachineRegisterInfo::use_iterator UI = MRI->use_begin((*LI)->reg),
         UE = MRI->use_end(); UI != UE; ++UI) {
      unsigned index = LIs->getInstructionIndex(&*UI);
      index = LiveIntervals::getUseIndex(index);
      
      const LiveRange* LR = (*LI)->getLiveRangeContaining(index);
      VNUseCount[LR->valno].insert(&*UI);
    }
    
    // Now, take the definitions (VNInfo's) one at a time and try to DCE 
    // and/or fold them away.
    for (LiveInterval::vni_iterator VI = (*LI)->vni_begin(),
         VE = (*LI)->vni_end(); VI != VE; ++VI) {
      
      if (DeadSplitLimit != -1 && (int)NumDeadSpills == DeadSplitLimit) 
        return changed;
      
      VNInfo* CurrVN = *VI;
      
      // We don't currently try to handle definitions with PHI kills, because
      // it would involve processing more than one VNInfo at once.
      if (CurrVN->hasPHIKill) continue;
      
      // We also don't try to handle the results of PHI joins, since there's
      // no defining instruction to analyze.
      unsigned DefIdx = CurrVN->def;
      if (DefIdx == ~0U || DefIdx == ~1U) continue;
    
      // We're only interested in eliminating cruft introduced by the splitter,
      // is of the form load-use or load-use-store.  First, check that the
      // definition is a load, and remember what stack slot we loaded it from.
      MachineInstr* DefMI = LIs->getInstructionFromIndex(DefIdx);
      int FrameIndex;
      if (!TII->isLoadFromStackSlot(DefMI, FrameIndex)) continue;
      
      // If the definition has no uses at all, just DCE it.
      if (VNUseCount[CurrVN].size() == 0) {
        LIs->RemoveMachineInstrFromMaps(DefMI);
        (*LI)->removeValNo(CurrVN);
        DefMI->eraseFromParent();
        VNUseCount.erase(CurrVN);
        NumDeadSpills++;
        changed = true;
        continue;
      }
      
      // Second, get the number of non-store uses of the definition, as well as
      // a flag indicating whether it feeds into a later two-address definition.
      bool FeedsTwoAddr = false;
      unsigned NonSpillCount = getNumberOfNonSpills(VNUseCount[CurrVN],
                                                    (*LI)->reg, FrameIndex,
                                                    FeedsTwoAddr);
      
      // If there's one non-store use and it doesn't feed a two-addr, then
      // this is a load-use-store case that we can try to fold.
      if (NonSpillCount == 1 && !FeedsTwoAddr) {
        // Start by finding the non-store use MachineInstr.
        SmallPtrSet<MachineInstr*, 4>::iterator UI = VNUseCount[CurrVN].begin();
        int StoreFrameIndex;
        unsigned StoreVReg = TII->isStoreToStackSlot(*UI, StoreFrameIndex);
        while (UI != VNUseCount[CurrVN].end() &&
               (StoreVReg == (*LI)->reg && StoreFrameIndex == FrameIndex)) {
          ++UI;
          if (UI != VNUseCount[CurrVN].end())
            StoreVReg = TII->isStoreToStackSlot(*UI, StoreFrameIndex);
        }
        if (UI == VNUseCount[CurrVN].end()) continue;
        
        MachineInstr* use = *UI;
        
        // Attempt to fold it away!
        int OpIdx = use->findRegisterUseOperandIdx((*LI)->reg, false);
        if (OpIdx == -1) continue;
        SmallVector<unsigned, 1> Ops;
        Ops.push_back(OpIdx);
        if (!TII->canFoldMemoryOperand(use, Ops)) continue;

        MachineInstr* NewMI =
                          TII->foldMemoryOperand(*use->getParent()->getParent(),  
                                                 use, Ops, FrameIndex);

        if (!NewMI) continue;

        // Update relevant analyses.
        LIs->RemoveMachineInstrFromMaps(DefMI);
        LIs->ReplaceMachineInstrInMaps(use, NewMI);
        (*LI)->removeValNo(CurrVN);

        DefMI->eraseFromParent();
        MachineBasicBlock* MBB = use->getParent();
        NewMI = MBB->insert(MBB->erase(use), NewMI);
        VNUseCount[CurrVN].erase(use);
        
        // Remove deleted instructions.  Note that we need to remove them from 
        // the VNInfo->use map as well, just to be safe.
        for (SmallPtrSet<MachineInstr*, 4>::iterator II = 
             VNUseCount[CurrVN].begin(), IE = VNUseCount[CurrVN].end();
             II != IE; ++II) {
          for (DenseMap<VNInfo*, SmallPtrSet<MachineInstr*, 4> >::iterator
               VNI = VNUseCount.begin(), VNE = VNUseCount.end(); VNI != VNE; 
               ++VNI)
            if (VNI->first != CurrVN)
              VNI->second.erase(*II);
          LIs->RemoveMachineInstrFromMaps(*II);
          (*II)->eraseFromParent();
        }
        
        VNUseCount.erase(CurrVN);

        for (DenseMap<VNInfo*, SmallPtrSet<MachineInstr*, 4> >::iterator
             VI = VNUseCount.begin(), VE = VNUseCount.end(); VI != VE; ++VI)
          if (VI->second.erase(use))
            VI->second.insert(NewMI);

        NumDeadSpills++;
        changed = true;
        continue;
      }
      
      // If there's more than one non-store instruction, we can't profitably
      // fold it, so bail.
      if (NonSpillCount) continue;
        
      // Otherwise, this is a load-store case, so DCE them.
      for (SmallPtrSet<MachineInstr*, 4>::iterator UI = 
           VNUseCount[CurrVN].begin(), UE = VNUseCount[CurrVN].end();
           UI != UI; ++UI) {
        LIs->RemoveMachineInstrFromMaps(*UI);
        (*UI)->eraseFromParent();
      }
        
      VNUseCount.erase(CurrVN);
        
      LIs->RemoveMachineInstrFromMaps(DefMI);
      (*LI)->removeValNo(CurrVN);
      DefMI->eraseFromParent();
      NumDeadSpills++;
      changed = true;
    }
  }
  
  return changed;
}

bool PreAllocSplitting::createsNewJoin(LiveRange* LR,
                                       MachineBasicBlock* DefMBB,
                                       MachineBasicBlock* BarrierMBB) {
  if (DefMBB == BarrierMBB)
    return false;
  
  if (LR->valno->hasPHIKill)
    return false;
  
  unsigned MBBEnd = LIs->getMBBEndIdx(BarrierMBB);
  if (LR->end < MBBEnd)
    return false;
  
  MachineLoopInfo& MLI = getAnalysis<MachineLoopInfo>();
  if (MLI.getLoopFor(DefMBB) != MLI.getLoopFor(BarrierMBB))
    return true;
  
  MachineDominatorTree& MDT = getAnalysis<MachineDominatorTree>();
  SmallPtrSet<MachineBasicBlock*, 4> Visited;
  typedef std::pair<MachineBasicBlock*,
                    MachineBasicBlock::succ_iterator> ItPair;
  SmallVector<ItPair, 4> Stack;
  Stack.push_back(std::make_pair(BarrierMBB, BarrierMBB->succ_begin()));
  
  while (!Stack.empty()) {
    ItPair P = Stack.back();
    Stack.pop_back();
    
    MachineBasicBlock* PredMBB = P.first;
    MachineBasicBlock::succ_iterator S = P.second;
    
    if (S == PredMBB->succ_end())
      continue;
    else if (Visited.count(*S)) {
      Stack.push_back(std::make_pair(PredMBB, ++S));
      continue;
    } else
      Stack.push_back(std::make_pair(PredMBB, S+1));
    
    MachineBasicBlock* MBB = *S;
    Visited.insert(MBB);
    
    if (MBB == BarrierMBB)
      return true;
    
    MachineDomTreeNode* DefMDTN = MDT.getNode(DefMBB);
    MachineDomTreeNode* BarrierMDTN = MDT.getNode(BarrierMBB);
    MachineDomTreeNode* MDTN = MDT.getNode(MBB)->getIDom();
    while (MDTN) {
      if (MDTN == DefMDTN)
        return true;
      else if (MDTN == BarrierMDTN)
        break;
      MDTN = MDTN->getIDom();
    }
    
    MBBEnd = LIs->getMBBEndIdx(MBB);
    if (LR->end > MBBEnd)
      Stack.push_back(std::make_pair(MBB, MBB->succ_begin()));
  }
  
  return false;
} 
  

bool PreAllocSplitting::runOnMachineFunction(MachineFunction &MF) {
  CurrMF = &MF;
  TM     = &MF.getTarget();
  TRI    = TM->getRegisterInfo();
  TII    = TM->getInstrInfo();
  MFI    = MF.getFrameInfo();
  MRI    = &MF.getRegInfo();
  LIs    = &getAnalysis<LiveIntervals>();
  LSs    = &getAnalysis<LiveStacks>();

  bool MadeChange = false;

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

  MachineBasicBlock *Entry = MF.begin();
  SmallPtrSet<MachineBasicBlock*,16> Visited;

  SmallPtrSet<LiveInterval*, 8> Split;

  for (df_ext_iterator<MachineBasicBlock*, SmallPtrSet<MachineBasicBlock*,16> >
         DFI = df_ext_begin(Entry, Visited), E = df_ext_end(Entry, Visited);
       DFI != E; ++DFI) {
    BarrierMBB = *DFI;
    for (MachineBasicBlock::iterator I = BarrierMBB->begin(),
           E = BarrierMBB->end(); I != E; ++I) {
      Barrier = &*I;
      const TargetRegisterClass **BarrierRCs =
        Barrier->getDesc().getRegClassBarriers();
      if (!BarrierRCs)
        continue;
      BarrierIdx = LIs->getInstructionIndex(Barrier);
      MadeChange |= SplitRegLiveIntervals(BarrierRCs, Split);
    }
  }

  MadeChange |= removeDeadSpills(Split);

  return MadeChange;
}