SplitKit.cpp revision b5f327b30f048b989c6f08d3bf84568d541b7644
1//===---------- SplitKit.cpp - Toolkit for splitting live ranges ----------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file contains the SplitAnalysis class as well as mutator functions for
11// live range splitting.
12//
13//===----------------------------------------------------------------------===//
14
15#define DEBUG_TYPE "regalloc"
16#include "SplitKit.h"
17#include "LiveRangeEdit.h"
18#include "VirtRegMap.h"
19#include "llvm/CodeGen/CalcSpillWeights.h"
20#include "llvm/CodeGen/LiveIntervalAnalysis.h"
21#include "llvm/CodeGen/MachineDominators.h"
22#include "llvm/CodeGen/MachineInstrBuilder.h"
23#include "llvm/CodeGen/MachineLoopInfo.h"
24#include "llvm/CodeGen/MachineRegisterInfo.h"
25#include "llvm/Support/CommandLine.h"
26#include "llvm/Support/Debug.h"
27#include "llvm/Support/raw_ostream.h"
28#include "llvm/Target/TargetInstrInfo.h"
29#include "llvm/Target/TargetMachine.h"
30
31using namespace llvm;
32
33static cl::opt<bool>
34AllowSplit("spiller-splits-edges",
35           cl::desc("Allow critical edge splitting during spilling"));
36
37//===----------------------------------------------------------------------===//
38//                                 Split Analysis
39//===----------------------------------------------------------------------===//
40
41SplitAnalysis::SplitAnalysis(const MachineFunction &mf,
42                             const LiveIntervals &lis,
43                             const MachineLoopInfo &mli)
44  : mf_(mf),
45    lis_(lis),
46    loops_(mli),
47    tii_(*mf.getTarget().getInstrInfo()),
48    curli_(0) {}
49
50void SplitAnalysis::clear() {
51  usingInstrs_.clear();
52  usingBlocks_.clear();
53  usingLoops_.clear();
54  curli_ = 0;
55}
56
57bool SplitAnalysis::canAnalyzeBranch(const MachineBasicBlock *MBB) {
58  MachineBasicBlock *T, *F;
59  SmallVector<MachineOperand, 4> Cond;
60  return !tii_.AnalyzeBranch(const_cast<MachineBasicBlock&>(*MBB), T, F, Cond);
61}
62
63/// analyzeUses - Count instructions, basic blocks, and loops using curli.
64void SplitAnalysis::analyzeUses() {
65  const MachineRegisterInfo &MRI = mf_.getRegInfo();
66  for (MachineRegisterInfo::reg_iterator I = MRI.reg_begin(curli_->reg);
67       MachineInstr *MI = I.skipInstruction();) {
68    if (MI->isDebugValue() || !usingInstrs_.insert(MI))
69      continue;
70    MachineBasicBlock *MBB = MI->getParent();
71    if (usingBlocks_[MBB]++)
72      continue;
73    for (MachineLoop *Loop = loops_.getLoopFor(MBB); Loop;
74         Loop = Loop->getParentLoop())
75      usingLoops_[Loop]++;
76  }
77  DEBUG(dbgs() << "  counted "
78               << usingInstrs_.size() << " instrs, "
79               << usingBlocks_.size() << " blocks, "
80               << usingLoops_.size()  << " loops.\n");
81}
82
83void SplitAnalysis::print(const BlockPtrSet &B, raw_ostream &OS) const {
84  for (BlockPtrSet::const_iterator I = B.begin(), E = B.end(); I != E; ++I) {
85    unsigned count = usingBlocks_.lookup(*I);
86    OS << " BB#" << (*I)->getNumber();
87    if (count)
88      OS << '(' << count << ')';
89  }
90}
91
92// Get three sets of basic blocks surrounding a loop: Blocks inside the loop,
93// predecessor blocks, and exit blocks.
94void SplitAnalysis::getLoopBlocks(const MachineLoop *Loop, LoopBlocks &Blocks) {
95  Blocks.clear();
96
97  // Blocks in the loop.
98  Blocks.Loop.insert(Loop->block_begin(), Loop->block_end());
99
100  // Predecessor blocks.
101  const MachineBasicBlock *Header = Loop->getHeader();
102  for (MachineBasicBlock::const_pred_iterator I = Header->pred_begin(),
103       E = Header->pred_end(); I != E; ++I)
104    if (!Blocks.Loop.count(*I))
105      Blocks.Preds.insert(*I);
106
107  // Exit blocks.
108  for (MachineLoop::block_iterator I = Loop->block_begin(),
109       E = Loop->block_end(); I != E; ++I) {
110    const MachineBasicBlock *MBB = *I;
111    for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(),
112       SE = MBB->succ_end(); SI != SE; ++SI)
113      if (!Blocks.Loop.count(*SI))
114        Blocks.Exits.insert(*SI);
115  }
116}
117
118void SplitAnalysis::print(const LoopBlocks &B, raw_ostream &OS) const {
119  OS << "Loop:";
120  print(B.Loop, OS);
121  OS << ", preds:";
122  print(B.Preds, OS);
123  OS << ", exits:";
124  print(B.Exits, OS);
125}
126
127/// analyzeLoopPeripheralUse - Return an enum describing how curli_ is used in
128/// and around the Loop.
129SplitAnalysis::LoopPeripheralUse SplitAnalysis::
130analyzeLoopPeripheralUse(const SplitAnalysis::LoopBlocks &Blocks) {
131  LoopPeripheralUse use = ContainedInLoop;
132  for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
133       I != E; ++I) {
134    const MachineBasicBlock *MBB = I->first;
135    // Is this a peripheral block?
136    if (use < MultiPeripheral &&
137        (Blocks.Preds.count(MBB) || Blocks.Exits.count(MBB))) {
138      if (I->second > 1) use = MultiPeripheral;
139      else               use = SinglePeripheral;
140      continue;
141    }
142    // Is it a loop block?
143    if (Blocks.Loop.count(MBB))
144      continue;
145    // It must be an unrelated block.
146    DEBUG(dbgs() << ", outside: BB#" << MBB->getNumber());
147    return OutsideLoop;
148  }
149  return use;
150}
151
152/// getCriticalExits - It may be necessary to partially break critical edges
153/// leaving the loop if an exit block has predecessors from outside the loop
154/// periphery.
155void SplitAnalysis::getCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
156                                     BlockPtrSet &CriticalExits) {
157  CriticalExits.clear();
158
159  // A critical exit block has curli live-in, and has a predecessor that is not
160  // in the loop nor a loop predecessor. For such an exit block, the edges
161  // carrying the new variable must be moved to a new pre-exit block.
162  for (BlockPtrSet::iterator I = Blocks.Exits.begin(), E = Blocks.Exits.end();
163       I != E; ++I) {
164    const MachineBasicBlock *Exit = *I;
165    // A single-predecessor exit block is definitely not a critical edge.
166    if (Exit->pred_size() == 1)
167      continue;
168    // This exit may not have curli live in at all. No need to split.
169    if (!lis_.isLiveInToMBB(*curli_, Exit))
170      continue;
171    // Does this exit block have a predecessor that is not a loop block or loop
172    // predecessor?
173    for (MachineBasicBlock::const_pred_iterator PI = Exit->pred_begin(),
174         PE = Exit->pred_end(); PI != PE; ++PI) {
175      const MachineBasicBlock *Pred = *PI;
176      if (Blocks.Loop.count(Pred) || Blocks.Preds.count(Pred))
177        continue;
178      // This is a critical exit block, and we need to split the exit edge.
179      CriticalExits.insert(Exit);
180      break;
181    }
182  }
183}
184
185void SplitAnalysis::getCriticalPreds(const SplitAnalysis::LoopBlocks &Blocks,
186                                     BlockPtrSet &CriticalPreds) {
187  CriticalPreds.clear();
188
189  // A critical predecessor block has curli live-out, and has a successor that
190  // has curli live-in and is not in the loop nor a loop exit block. For such a
191  // predecessor block, we must carry the value in both the 'inside' and
192  // 'outside' registers.
193  for (BlockPtrSet::iterator I = Blocks.Preds.begin(), E = Blocks.Preds.end();
194       I != E; ++I) {
195    const MachineBasicBlock *Pred = *I;
196    // Definitely not a critical edge.
197    if (Pred->succ_size() == 1)
198      continue;
199    // This block may not have curli live out at all if there is a PHI.
200    if (!lis_.isLiveOutOfMBB(*curli_, Pred))
201      continue;
202    // Does this block have a successor outside the loop?
203    for (MachineBasicBlock::const_pred_iterator SI = Pred->succ_begin(),
204         SE = Pred->succ_end(); SI != SE; ++SI) {
205      const MachineBasicBlock *Succ = *SI;
206      if (Blocks.Loop.count(Succ) || Blocks.Exits.count(Succ))
207        continue;
208      if (!lis_.isLiveInToMBB(*curli_, Succ))
209        continue;
210      // This is a critical predecessor block.
211      CriticalPreds.insert(Pred);
212      break;
213    }
214  }
215}
216
217/// canSplitCriticalExits - Return true if it is possible to insert new exit
218/// blocks before the blocks in CriticalExits.
219bool
220SplitAnalysis::canSplitCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
221                                     BlockPtrSet &CriticalExits) {
222  // If we don't allow critical edge splitting, require no critical exits.
223  if (!AllowSplit)
224    return CriticalExits.empty();
225
226  for (BlockPtrSet::iterator I = CriticalExits.begin(), E = CriticalExits.end();
227       I != E; ++I) {
228    const MachineBasicBlock *Succ = *I;
229    // We want to insert a new pre-exit MBB before Succ, and change all the
230    // in-loop blocks to branch to the pre-exit instead of Succ.
231    // Check that all the in-loop predecessors can be changed.
232    for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
233         PE = Succ->pred_end(); PI != PE; ++PI) {
234      const MachineBasicBlock *Pred = *PI;
235      // The external predecessors won't be altered.
236      if (!Blocks.Loop.count(Pred) && !Blocks.Preds.count(Pred))
237        continue;
238      if (!canAnalyzeBranch(Pred))
239        return false;
240    }
241
242    // If Succ's layout predecessor falls through, that too must be analyzable.
243    // We need to insert the pre-exit block in the gap.
244    MachineFunction::const_iterator MFI = Succ;
245    if (MFI == mf_.begin())
246      continue;
247    if (!canAnalyzeBranch(--MFI))
248      return false;
249  }
250  // No problems found.
251  return true;
252}
253
254void SplitAnalysis::analyze(const LiveInterval *li) {
255  clear();
256  curli_ = li;
257  analyzeUses();
258}
259
260const MachineLoop *SplitAnalysis::getBestSplitLoop() {
261  assert(curli_ && "Call analyze() before getBestSplitLoop");
262  if (usingLoops_.empty())
263    return 0;
264
265  LoopPtrSet Loops;
266  LoopBlocks Blocks;
267  BlockPtrSet CriticalExits;
268
269  // We split around loops where curli is used outside the periphery.
270  for (LoopCountMap::const_iterator I = usingLoops_.begin(),
271       E = usingLoops_.end(); I != E; ++I) {
272    const MachineLoop *Loop = I->first;
273    getLoopBlocks(Loop, Blocks);
274    DEBUG({ dbgs() << "  "; print(Blocks, dbgs()); });
275
276    switch(analyzeLoopPeripheralUse(Blocks)) {
277    case OutsideLoop:
278      break;
279    case MultiPeripheral:
280      // FIXME: We could split a live range with multiple uses in a peripheral
281      // block and still make progress. However, it is possible that splitting
282      // another live range will insert copies into a peripheral block, and
283      // there is a small chance we can enter an infinity loop, inserting copies
284      // forever.
285      // For safety, stick to splitting live ranges with uses outside the
286      // periphery.
287      DEBUG(dbgs() << ": multiple peripheral uses\n");
288      break;
289    case ContainedInLoop:
290      DEBUG(dbgs() << ": fully contained\n");
291      continue;
292    case SinglePeripheral:
293      DEBUG(dbgs() << ": single peripheral use\n");
294      continue;
295    }
296    // Will it be possible to split around this loop?
297    getCriticalExits(Blocks, CriticalExits);
298    DEBUG(dbgs() << ": " << CriticalExits.size() << " critical exits\n");
299    if (!canSplitCriticalExits(Blocks, CriticalExits))
300      continue;
301    // This is a possible split.
302    Loops.insert(Loop);
303  }
304
305  DEBUG(dbgs() << "  getBestSplitLoop found " << Loops.size()
306               << " candidate loops.\n");
307
308  if (Loops.empty())
309    return 0;
310
311  // Pick the earliest loop.
312  // FIXME: Are there other heuristics to consider?
313  const MachineLoop *Best = 0;
314  SlotIndex BestIdx;
315  for (LoopPtrSet::const_iterator I = Loops.begin(), E = Loops.end(); I != E;
316       ++I) {
317    SlotIndex Idx = lis_.getMBBStartIdx((*I)->getHeader());
318    if (!Best || Idx < BestIdx)
319      Best = *I, BestIdx = Idx;
320  }
321  DEBUG(dbgs() << "  getBestSplitLoop found " << *Best);
322  return Best;
323}
324
325//===----------------------------------------------------------------------===//
326//                               LiveIntervalMap
327//===----------------------------------------------------------------------===//
328
329// Work around the fact that the std::pair constructors are broken for pointer
330// pairs in some implementations. makeVV(x, 0) works.
331static inline std::pair<const VNInfo*, VNInfo*>
332makeVV(const VNInfo *a, VNInfo *b) {
333  return std::make_pair(a, b);
334}
335
336void LiveIntervalMap::reset(LiveInterval *li) {
337  li_ = li;
338  valueMap_.clear();
339  liveOutCache_.clear();
340}
341
342bool LiveIntervalMap::isComplexMapped(const VNInfo *ParentVNI) const {
343  ValueMap::const_iterator i = valueMap_.find(ParentVNI);
344  return i != valueMap_.end() && i->second == 0;
345}
346
347// defValue - Introduce a li_ def for ParentVNI that could be later than
348// ParentVNI->def.
349VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) {
350  assert(li_ && "call reset first");
351  assert(ParentVNI && "Mapping  NULL value");
352  assert(Idx.isValid() && "Invalid SlotIndex");
353  assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
354
355  // Create a new value.
356  VNInfo *VNI = li_->getNextValue(Idx, 0, lis_.getVNInfoAllocator());
357
358  // Preserve the PHIDef bit.
359  if (ParentVNI->isPHIDef() && Idx == ParentVNI->def)
360    VNI->setIsPHIDef(true);
361
362  // Use insert for lookup, so we can add missing values with a second lookup.
363  std::pair<ValueMap::iterator,bool> InsP =
364    valueMap_.insert(makeVV(ParentVNI, Idx == ParentVNI->def ? VNI : 0));
365
366  // This is now a complex def. Mark with a NULL in valueMap.
367  if (!InsP.second)
368    InsP.first->second = 0;
369
370  return VNI;
371}
372
373
374// mapValue - Find the mapped value for ParentVNI at Idx.
375// Potentially create phi-def values.
376VNInfo *LiveIntervalMap::mapValue(const VNInfo *ParentVNI, SlotIndex Idx,
377                                  bool *simple) {
378  assert(li_ && "call reset first");
379  assert(ParentVNI && "Mapping  NULL value");
380  assert(Idx.isValid() && "Invalid SlotIndex");
381  assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
382
383  // Use insert for lookup, so we can add missing values with a second lookup.
384  std::pair<ValueMap::iterator,bool> InsP =
385    valueMap_.insert(makeVV(ParentVNI, 0));
386
387  // This was an unknown value. Create a simple mapping.
388  if (InsP.second) {
389    if (simple) *simple = true;
390    return InsP.first->second = li_->createValueCopy(ParentVNI,
391                                                     lis_.getVNInfoAllocator());
392  }
393
394  // This was a simple mapped value.
395  if (InsP.first->second) {
396    if (simple) *simple = true;
397    return InsP.first->second;
398  }
399
400  // This is a complex mapped value. There may be multiple defs, and we may need
401  // to create phi-defs.
402  if (simple) *simple = false;
403  MachineBasicBlock *IdxMBB = lis_.getMBBFromIndex(Idx);
404  assert(IdxMBB && "No MBB at Idx");
405
406  // Is there a def in the same MBB we can extend?
407  if (VNInfo *VNI = extendTo(IdxMBB, Idx))
408    return VNI;
409
410  // Now for the fun part. We know that ParentVNI potentially has multiple defs,
411  // and we may need to create even more phi-defs to preserve VNInfo SSA form.
412  // Perform a search for all predecessor blocks where we know the dominating
413  // VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
414  DEBUG(dbgs() << "\n  Reaching defs for BB#" << IdxMBB->getNumber()
415               << " at " << Idx << " in " << *li_ << '\n');
416
417  // Blocks where li_ should be live-in.
418  SmallVector<MachineDomTreeNode*, 16> LiveIn;
419  LiveIn.push_back(mdt_[IdxMBB]);
420
421  // Using liveOutCache_ as a visited set, perform a BFS for all reaching defs.
422  for (unsigned i = 0; i != LiveIn.size(); ++i) {
423    MachineBasicBlock *MBB = LiveIn[i]->getBlock();
424    for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
425           PE = MBB->pred_end(); PI != PE; ++PI) {
426       MachineBasicBlock *Pred = *PI;
427       // Is this a known live-out block?
428       std::pair<LiveOutMap::iterator,bool> LOIP =
429         liveOutCache_.insert(std::make_pair(Pred, LiveOutPair()));
430       // Yes, we have been here before.
431       if (!LOIP.second) {
432         DEBUG(if (VNInfo *VNI = LOIP.first->second.first)
433                 dbgs() << "    known valno #" << VNI->id
434                        << " at BB#" << Pred->getNumber() << '\n');
435         continue;
436       }
437
438       // Does Pred provide a live-out value?
439       SlotIndex Last = lis_.getMBBEndIdx(Pred).getPrevSlot();
440       if (VNInfo *VNI = extendTo(Pred, Last)) {
441         MachineBasicBlock *DefMBB = lis_.getMBBFromIndex(VNI->def);
442         DEBUG(dbgs() << "    found valno #" << VNI->id
443                      << " from BB#" << DefMBB->getNumber()
444                      << " at BB#" << Pred->getNumber() << '\n');
445         LiveOutPair &LOP = LOIP.first->second;
446         LOP.first = VNI;
447         LOP.second = mdt_[DefMBB];
448         continue;
449       }
450       // No, we need a live-in value for Pred as well
451       if (Pred != IdxMBB)
452         LiveIn.push_back(mdt_[Pred]);
453    }
454  }
455
456  // We may need to add phi-def values to preserve the SSA form.
457  // This is essentially the same iterative algorithm that SSAUpdater uses,
458  // except we already have a dominator tree, so we don't have to recompute it.
459  VNInfo *IdxVNI = 0;
460  unsigned Changes;
461  do {
462    Changes = 0;
463    DEBUG(dbgs() << "  Iterating over " << LiveIn.size() << " blocks.\n");
464    // Propagate live-out values down the dominator tree, inserting phi-defs when
465    // necessary. Since LiveIn was created by a BFS, going backwards makes it more
466    // likely for us to visit immediate dominators before their children.
467    for (unsigned i = LiveIn.size(); i; --i) {
468      MachineDomTreeNode *Node = LiveIn[i-1];
469      MachineBasicBlock *MBB = Node->getBlock();
470      MachineDomTreeNode *IDom = Node->getIDom();
471      LiveOutPair IDomValue;
472      // We need a live-in value to a block with no immediate dominator?
473      // This is probably an unreachable block that has survived somehow.
474      bool needPHI = !IDom;
475
476      // Get the IDom live-out value.
477      if (!needPHI) {
478        LiveOutMap::iterator I = liveOutCache_.find(IDom->getBlock());
479        if (I != liveOutCache_.end())
480          IDomValue = I->second;
481        else
482          // If IDom is outside our set of live-out blocks, there must be new
483          // defs, and we need a phi-def here.
484          needPHI = true;
485      }
486
487      // IDom dominates all of our predecessors, but it may not be the immediate
488      // dominator. Check if any of them have live-out values that are properly
489      // dominated by IDom. If so, we need a phi-def here.
490      if (!needPHI) {
491        for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
492               PE = MBB->pred_end(); PI != PE; ++PI) {
493          LiveOutPair Value = liveOutCache_[*PI];
494          if (!Value.first || Value.first == IDomValue.first)
495            continue;
496          // This predecessor is carrying something other than IDomValue.
497          // It could be because IDomValue hasn't propagated yet, or it could be
498          // because MBB is in the dominance frontier of that value.
499          if (mdt_.dominates(IDom, Value.second)) {
500            needPHI = true;
501            break;
502          }
503        }
504      }
505
506      // Create a phi-def if required.
507      if (needPHI) {
508        ++Changes;
509        SlotIndex Start = lis_.getMBBStartIdx(MBB);
510        VNInfo *VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator());
511        VNI->setIsPHIDef(true);
512        DEBUG(dbgs() << "    - BB#" << MBB->getNumber()
513                     << " phi-def #" << VNI->id << " at " << Start << '\n');
514        // We no longer need li_ to be live-in.
515        LiveIn.erase(LiveIn.begin()+(i-1));
516        // Blocks in LiveIn are either IdxMBB, or have a value live-through.
517        if (MBB == IdxMBB)
518          IdxVNI = VNI;
519        // Check if we need to update live-out info.
520        LiveOutMap::iterator I = liveOutCache_.find(MBB);
521        if (I == liveOutCache_.end() || I->second.second == Node) {
522          // We already have a live-out defined in MBB, so this must be IdxMBB.
523          assert(MBB == IdxMBB && "Adding phi-def to known live-out");
524          li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
525        } else {
526          // This phi-def is also live-out, so color the whole block.
527          li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
528          I->second = LiveOutPair(VNI, Node);
529        }
530      } else if (IDomValue.first) {
531        // No phi-def here. Remember incoming value for IdxMBB.
532        if (MBB == IdxMBB)
533          IdxVNI = IDomValue.first;
534        // Propagate IDomValue if needed:
535        // MBB is live-out and doesn't define its own value.
536        LiveOutMap::iterator I = liveOutCache_.find(MBB);
537        if (I != liveOutCache_.end() && I->second.second != Node &&
538            I->second.first != IDomValue.first) {
539          ++Changes;
540          I->second = IDomValue;
541          DEBUG(dbgs() << "    - BB#" << MBB->getNumber()
542                       << " idom valno #" << IDomValue.first->id
543                       << " from BB#" << IDom->getBlock()->getNumber() << '\n');
544        }
545      }
546    }
547    DEBUG(dbgs() << "  - made " << Changes << " changes.\n");
548  } while (Changes);
549
550  assert(IdxVNI && "Didn't find value for Idx");
551
552#ifndef NDEBUG
553  // Check the liveOutCache_ invariants.
554  for (LiveOutMap::iterator I = liveOutCache_.begin(), E = liveOutCache_.end();
555         I != E; ++I) {
556    assert(I->first && "Null MBB entry in cache");
557    assert(I->second.first && "Null VNInfo in cache");
558    assert(I->second.second && "Null DomTreeNode in cache");
559    if (I->second.second->getBlock() == I->first)
560      continue;
561    for (MachineBasicBlock::pred_iterator PI = I->first->pred_begin(),
562           PE = I->first->pred_end(); PI != PE; ++PI)
563      assert(liveOutCache_.lookup(*PI) == I->second && "Bad invariant");
564  }
565#endif
566
567  // Since we went through the trouble of a full BFS visiting all reaching defs,
568  // the values in LiveIn are now accurate. No more phi-defs are needed
569  // for these blocks, so we can color the live ranges.
570  // This makes the next mapValue call much faster.
571  for (unsigned i = 0, e = LiveIn.size(); i != e; ++i) {
572    MachineBasicBlock *MBB = LiveIn[i]->getBlock();
573    SlotIndex Start = lis_.getMBBStartIdx(MBB);
574    if (MBB == IdxMBB) {
575      li_->addRange(LiveRange(Start, Idx.getNextSlot(), IdxVNI));
576      continue;
577    }
578    // Anything in LiveIn other than IdxMBB is live-through.
579    VNInfo *VNI = liveOutCache_.lookup(MBB).first;
580    assert(VNI && "Missing block value");
581    li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
582  }
583
584  return IdxVNI;
585}
586
587// extendTo - Find the last li_ value defined in MBB at or before Idx. The
588// parentli_ is assumed to be live at Idx. Extend the live range to Idx.
589// Return the found VNInfo, or NULL.
590VNInfo *LiveIntervalMap::extendTo(const MachineBasicBlock *MBB, SlotIndex Idx) {
591  assert(li_ && "call reset first");
592  LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx);
593  if (I == li_->begin())
594    return 0;
595  --I;
596  if (I->end <= lis_.getMBBStartIdx(MBB))
597    return 0;
598  if (I->end <= Idx)
599    I->end = Idx.getNextSlot();
600  return I->valno;
601}
602
603// addSimpleRange - Add a simple range from parentli_ to li_.
604// ParentVNI must be live in the [Start;End) interval.
605void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End,
606                                     const VNInfo *ParentVNI) {
607  assert(li_ && "call reset first");
608  bool simple;
609  VNInfo *VNI = mapValue(ParentVNI, Start, &simple);
610  // A simple mapping is easy.
611  if (simple) {
612    li_->addRange(LiveRange(Start, End, VNI));
613    return;
614  }
615
616  // ParentVNI is a complex value. We must map per MBB.
617  MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
618  MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End.getPrevSlot());
619
620  if (MBB == MBBE) {
621    li_->addRange(LiveRange(Start, End, VNI));
622    return;
623  }
624
625  // First block.
626  li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
627
628  // Run sequence of full blocks.
629  for (++MBB; MBB != MBBE; ++MBB) {
630    Start = lis_.getMBBStartIdx(MBB);
631    li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB),
632                            mapValue(ParentVNI, Start)));
633  }
634
635  // Final block.
636  Start = lis_.getMBBStartIdx(MBB);
637  if (Start != End)
638    li_->addRange(LiveRange(Start, End, mapValue(ParentVNI, Start)));
639}
640
641/// addRange - Add live ranges to li_ where [Start;End) intersects parentli_.
642/// All needed values whose def is not inside [Start;End) must be defined
643/// beforehand so mapValue will work.
644void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) {
645  assert(li_ && "call reset first");
646  LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end();
647  LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
648
649  // Check if --I begins before Start and overlaps.
650  if (I != B) {
651    --I;
652    if (I->end > Start)
653      addSimpleRange(Start, std::min(End, I->end), I->valno);
654    ++I;
655  }
656
657  // The remaining ranges begin after Start.
658  for (;I != E && I->start < End; ++I)
659    addSimpleRange(I->start, std::min(End, I->end), I->valno);
660}
661
662
663//===----------------------------------------------------------------------===//
664//                               Split Editor
665//===----------------------------------------------------------------------===//
666
667/// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
668SplitEditor::SplitEditor(SplitAnalysis &sa,
669                         LiveIntervals &lis,
670                         VirtRegMap &vrm,
671                         MachineDominatorTree &mdt,
672                         LiveRangeEdit &edit)
673  : sa_(sa), lis_(lis), vrm_(vrm),
674    mri_(vrm.getMachineFunction().getRegInfo()),
675    tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
676    tri_(*vrm.getMachineFunction().getTarget().getRegisterInfo()),
677    edit_(edit),
678    dupli_(lis_, mdt, edit.getParent()),
679    openli_(lis_, mdt, edit.getParent())
680{
681  // We don't need an AliasAnalysis since we will only be performing
682  // cheap-as-a-copy remats anyway.
683  edit_.anyRematerializable(lis_, tii_, 0);
684}
685
686bool SplitEditor::intervalsLiveAt(SlotIndex Idx) const {
687  for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
688    if (*I != dupli_.getLI() && (*I)->liveAt(Idx))
689      return true;
690  return false;
691}
692
693VNInfo *SplitEditor::defFromParent(LiveIntervalMap &Reg,
694                                   VNInfo *ParentVNI,
695                                   SlotIndex UseIdx,
696                                   MachineBasicBlock &MBB,
697                                   MachineBasicBlock::iterator I) {
698  VNInfo *VNI = 0;
699  MachineInstr *CopyMI = 0;
700  SlotIndex Def;
701
702  // Attempt cheap-as-a-copy rematerialization.
703  LiveRangeEdit::Remat RM(ParentVNI);
704  if (edit_.canRematerializeAt(RM, UseIdx, true, lis_)) {
705    Def = edit_.rematerializeAt(MBB, I, Reg.getLI()->reg, RM,
706                                          lis_, tii_, tri_);
707  } else {
708    // Can't remat, just insert a copy from parent.
709    CopyMI = BuildMI(MBB, I, DebugLoc(), tii_.get(TargetOpcode::COPY),
710                     Reg.getLI()->reg).addReg(edit_.getReg());
711    Def = lis_.InsertMachineInstrInMaps(CopyMI).getDefIndex();
712  }
713
714  // Define the value in Reg.
715  VNI = Reg.defValue(ParentVNI, Def);
716  VNI->setCopy(CopyMI);
717
718  // Add minimal liveness for the new value.
719  if (UseIdx < Def)
720    UseIdx = Def;
721  Reg.getLI()->addRange(LiveRange(Def, UseIdx.getNextSlot(), VNI));
722  return VNI;
723}
724
725/// Create a new virtual register and live interval.
726void SplitEditor::openIntv() {
727  assert(!openli_.getLI() && "Previous LI not closed before openIntv");
728  if (!dupli_.getLI())
729    dupli_.reset(&edit_.create(mri_, lis_, vrm_));
730
731  openli_.reset(&edit_.create(mri_, lis_, vrm_));
732}
733
734/// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
735/// not live before Idx, a COPY is not inserted.
736void SplitEditor::enterIntvBefore(SlotIndex Idx) {
737  assert(openli_.getLI() && "openIntv not called before enterIntvBefore");
738  Idx = Idx.getUseIndex();
739  DEBUG(dbgs() << "    enterIntvBefore " << Idx);
740  VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx);
741  if (!ParentVNI) {
742    DEBUG(dbgs() << ": not live\n");
743    return;
744  }
745  DEBUG(dbgs() << ": valno " << ParentVNI->id);
746  truncatedValues.insert(ParentVNI);
747  MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
748  assert(MI && "enterIntvBefore called with invalid index");
749
750  defFromParent(openli_, ParentVNI, Idx, *MI->getParent(), MI);
751
752  DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
753}
754
755/// enterIntvAtEnd - Enter openli at the end of MBB.
756void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) {
757  assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd");
758  SlotIndex End = lis_.getMBBEndIdx(&MBB).getPrevSlot();
759  DEBUG(dbgs() << "    enterIntvAtEnd BB#" << MBB.getNumber() << ", " << End);
760  VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(End);
761  if (!ParentVNI) {
762    DEBUG(dbgs() << ": not live\n");
763    return;
764  }
765  DEBUG(dbgs() << ": valno " << ParentVNI->id);
766  truncatedValues.insert(ParentVNI);
767  defFromParent(openli_, ParentVNI, End, MBB, MBB.getFirstTerminator());
768  DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
769}
770
771/// useIntv - indicate that all instructions in MBB should use openli.
772void SplitEditor::useIntv(const MachineBasicBlock &MBB) {
773  useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB));
774}
775
776void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) {
777  assert(openli_.getLI() && "openIntv not called before useIntv");
778  openli_.addRange(Start, End);
779  DEBUG(dbgs() << "    use [" << Start << ';' << End << "): "
780               << *openli_.getLI() << '\n');
781}
782
783/// leaveIntvAfter - Leave openli after the instruction at Idx.
784void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
785  assert(openli_.getLI() && "openIntv not called before leaveIntvAfter");
786  DEBUG(dbgs() << "    leaveIntvAfter " << Idx);
787
788  // The interval must be live beyond the instruction at Idx.
789  Idx = Idx.getBoundaryIndex();
790  VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx);
791  if (!ParentVNI) {
792    DEBUG(dbgs() << ": not live\n");
793    return;
794  }
795  DEBUG(dbgs() << ": valno " << ParentVNI->id);
796
797  MachineBasicBlock::iterator MII = lis_.getInstructionFromIndex(Idx);
798  VNInfo *VNI = defFromParent(dupli_, ParentVNI, Idx,
799                              *MII->getParent(), llvm::next(MII));
800
801  // Make sure that openli is properly extended from Idx to the new copy.
802  // FIXME: This shouldn't be necessary for remats.
803  openli_.addSimpleRange(Idx, VNI->def, ParentVNI);
804
805  DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
806}
807
808/// leaveIntvAtTop - Leave the interval at the top of MBB.
809/// Currently, only one value can leave the interval.
810void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
811  assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop");
812  SlotIndex Start = lis_.getMBBStartIdx(&MBB);
813  DEBUG(dbgs() << "    leaveIntvAtTop BB#" << MBB.getNumber() << ", " << Start);
814
815  VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Start);
816  if (!ParentVNI) {
817    DEBUG(dbgs() << ": not live\n");
818    return;
819  }
820
821  VNInfo *VNI = defFromParent(dupli_, ParentVNI, Start, MBB,
822                              MBB.SkipPHIsAndLabels(MBB.begin()));
823
824  // Finally we must make sure that openli is properly extended from Start to
825  // the new copy.
826  openli_.addSimpleRange(Start, VNI->def, ParentVNI);
827  DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
828}
829
830/// closeIntv - Indicate that we are done editing the currently open
831/// LiveInterval, and ranges can be trimmed.
832void SplitEditor::closeIntv() {
833  assert(openli_.getLI() && "openIntv not called before closeIntv");
834
835  DEBUG(dbgs() << "    closeIntv cleaning up\n");
836  DEBUG(dbgs() << "    open " << *openli_.getLI() << '\n');
837  openli_.reset(0);
838}
839
840/// rewrite - Rewrite all uses of reg to use the new registers.
841void SplitEditor::rewrite(unsigned reg) {
842  for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(reg),
843       RE = mri_.reg_end(); RI != RE;) {
844    MachineOperand &MO = RI.getOperand();
845    unsigned OpNum = RI.getOperandNo();
846    MachineInstr *MI = MO.getParent();
847    ++RI;
848    if (MI->isDebugValue()) {
849      DEBUG(dbgs() << "Zapping " << *MI);
850      // FIXME: We can do much better with debug values.
851      MO.setReg(0);
852      continue;
853    }
854    SlotIndex Idx = lis_.getInstructionIndex(MI);
855    Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
856    LiveInterval *LI = 0;
857    for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E;
858         ++I) {
859      LiveInterval *testli = *I;
860      if (testli->liveAt(Idx)) {
861        LI = testli;
862        break;
863      }
864    }
865    DEBUG(dbgs() << "  rewr BB#" << MI->getParent()->getNumber() << '\t'<< Idx);
866    assert(LI && "No register was live at use");
867    MO.setReg(LI->reg);
868    if (MO.isUse() && !MI->isRegTiedToDefOperand(OpNum))
869      MO.setIsKill(LI->killedAt(Idx.getDefIndex()));
870    DEBUG(dbgs() << '\t' << *MI);
871  }
872}
873
874void
875SplitEditor::addTruncSimpleRange(SlotIndex Start, SlotIndex End, VNInfo *VNI) {
876  // Build vector of iterator pairs from the intervals.
877  typedef std::pair<LiveInterval::const_iterator,
878                    LiveInterval::const_iterator> IIPair;
879  SmallVector<IIPair, 8> Iters;
880  for (LiveRangeEdit::iterator LI = edit_.begin(), LE = edit_.end(); LI != LE;
881       ++LI) {
882    if (*LI == dupli_.getLI())
883      continue;
884    LiveInterval::const_iterator I = (*LI)->find(Start);
885    LiveInterval::const_iterator E = (*LI)->end();
886    if (I != E)
887      Iters.push_back(std::make_pair(I, E));
888  }
889
890  SlotIndex sidx = Start;
891  // Break [Start;End) into segments that don't overlap any intervals.
892  for (;;) {
893    SlotIndex next = sidx, eidx = End;
894    // Find overlapping intervals.
895    for (unsigned i = 0; i != Iters.size() && sidx < eidx; ++i) {
896      LiveInterval::const_iterator I = Iters[i].first;
897      // Interval I is overlapping [sidx;eidx). Trim sidx.
898      if (I->start <= sidx) {
899        sidx = I->end;
900        // Move to the next run, remove iters when all are consumed.
901        I = ++Iters[i].first;
902        if (I == Iters[i].second) {
903          Iters.erase(Iters.begin() + i);
904          --i;
905          continue;
906        }
907      }
908      // Trim eidx too if needed.
909      if (I->start >= eidx)
910        continue;
911      eidx = I->start;
912      next = I->end;
913    }
914    // Now, [sidx;eidx) doesn't overlap anything in intervals_.
915    if (sidx < eidx)
916      dupli_.addSimpleRange(sidx, eidx, VNI);
917    // If the interval end was truncated, we can try again from next.
918    if (next <= sidx)
919      break;
920    sidx = next;
921  }
922}
923
924void SplitEditor::computeRemainder() {
925  // First we need to fill in the live ranges in dupli.
926  // If values were redefined, we need a full recoloring with SSA update.
927  // If values were truncated, we only need to truncate the ranges.
928  // If values were partially rematted, we should shrink to uses.
929  // If values were fully rematted, they should be omitted.
930  // FIXME: If a single value is redefined, just move the def and truncate.
931  LiveInterval &parent = edit_.getParent();
932
933  // Values that are fully contained in the split intervals.
934  SmallPtrSet<const VNInfo*, 8> deadValues;
935  // Map all curli values that should have live defs in dupli.
936  for (LiveInterval::const_vni_iterator I = parent.vni_begin(),
937       E = parent.vni_end(); I != E; ++I) {
938    const VNInfo *VNI = *I;
939    // Don't transfer unused values to the new intervals.
940    if (VNI->isUnused())
941      continue;
942    // Original def is contained in the split intervals.
943    if (intervalsLiveAt(VNI->def)) {
944      // Did this value escape?
945      if (dupli_.isMapped(VNI))
946        truncatedValues.insert(VNI);
947      else
948        deadValues.insert(VNI);
949      continue;
950    }
951    // Add minimal live range at the definition.
952    VNInfo *DVNI = dupli_.defValue(VNI, VNI->def);
953    dupli_.getLI()->addRange(LiveRange(VNI->def, VNI->def.getNextSlot(), DVNI));
954  }
955
956  // Add all ranges to dupli.
957  for (LiveInterval::const_iterator I = parent.begin(), E = parent.end();
958       I != E; ++I) {
959    const LiveRange &LR = *I;
960    if (truncatedValues.count(LR.valno)) {
961      // recolor after removing intervals_.
962      addTruncSimpleRange(LR.start, LR.end, LR.valno);
963    } else if (!deadValues.count(LR.valno)) {
964      // recolor without truncation.
965      dupli_.addSimpleRange(LR.start, LR.end, LR.valno);
966    }
967  }
968
969  // Extend dupli_ to be live out of any critical loop predecessors.
970  // This means we have multiple registers live out of those blocks.
971  // The alternative would be to split the critical edges.
972  if (criticalPreds_.empty())
973    return;
974  for (SplitAnalysis::BlockPtrSet::iterator I = criticalPreds_.begin(),
975       E = criticalPreds_.end(); I != E; ++I)
976     dupli_.extendTo(*I, lis_.getMBBEndIdx(*I).getPrevSlot());
977   criticalPreds_.clear();
978}
979
980void SplitEditor::finish() {
981  assert(!openli_.getLI() && "Previous LI not closed before rewrite");
982  assert(dupli_.getLI() && "No dupli for rewrite. Noop spilt?");
983
984  // Complete dupli liveness.
985  computeRemainder();
986
987  // Get rid of unused values and set phi-kill flags.
988  for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
989    (*I)->RenumberValues(lis_);
990
991  // Rewrite instructions.
992  rewrite(edit_.getReg());
993
994  // Now check if any registers were separated into multiple components.
995  ConnectedVNInfoEqClasses ConEQ(lis_);
996  for (unsigned i = 0, e = edit_.size(); i != e; ++i) {
997    // Don't use iterators, they are invalidated by create() below.
998    LiveInterval *li = edit_.get(i);
999    unsigned NumComp = ConEQ.Classify(li);
1000    if (NumComp <= 1)
1001      continue;
1002    DEBUG(dbgs() << "  " << NumComp << " components: " << *li << '\n');
1003    SmallVector<LiveInterval*, 8> dups;
1004    dups.push_back(li);
1005    for (unsigned i = 1; i != NumComp; ++i)
1006      dups.push_back(&edit_.create(mri_, lis_, vrm_));
1007    ConEQ.Distribute(&dups[0]);
1008    // Rewrite uses to the new regs.
1009    rewrite(li->reg);
1010  }
1011
1012  // Calculate spill weight and allocation hints for new intervals.
1013  VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
1014  for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I){
1015    LiveInterval &li = **I;
1016    vrai.CalculateRegClass(li.reg);
1017    vrai.CalculateWeightAndHint(li);
1018    DEBUG(dbgs() << "  new interval " << mri_.getRegClass(li.reg)->getName()
1019                 << ":" << li << '\n');
1020  }
1021}
1022
1023
1024//===----------------------------------------------------------------------===//
1025//                               Loop Splitting
1026//===----------------------------------------------------------------------===//
1027
1028void SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
1029  SplitAnalysis::LoopBlocks Blocks;
1030  sa_.getLoopBlocks(Loop, Blocks);
1031
1032  DEBUG({
1033    dbgs() << "  splitAround"; sa_.print(Blocks, dbgs()); dbgs() << '\n';
1034  });
1035
1036  // Break critical edges as needed.
1037  SplitAnalysis::BlockPtrSet CriticalExits;
1038  sa_.getCriticalExits(Blocks, CriticalExits);
1039  assert(CriticalExits.empty() && "Cannot break critical exits yet");
1040
1041  // Get critical predecessors so computeRemainder can deal with them.
1042  sa_.getCriticalPreds(Blocks, criticalPreds_);
1043
1044  // Create new live interval for the loop.
1045  openIntv();
1046
1047  // Insert copies in the predecessors.
1048  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
1049       E = Blocks.Preds.end(); I != E; ++I) {
1050    MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
1051    enterIntvAtEnd(MBB);
1052  }
1053
1054  // Switch all loop blocks.
1055  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
1056       E = Blocks.Loop.end(); I != E; ++I)
1057     useIntv(**I);
1058
1059  // Insert back copies in the exit blocks.
1060  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
1061       E = Blocks.Exits.end(); I != E; ++I) {
1062    MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
1063    leaveIntvAtTop(MBB);
1064  }
1065
1066  // Done.
1067  closeIntv();
1068  finish();
1069}
1070
1071
1072//===----------------------------------------------------------------------===//
1073//                            Single Block Splitting
1074//===----------------------------------------------------------------------===//
1075
1076/// getMultiUseBlocks - if curli has more than one use in a basic block, it
1077/// may be an advantage to split curli for the duration of the block.
1078bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
1079  // If curli is local to one block, there is no point to splitting it.
1080  if (usingBlocks_.size() <= 1)
1081    return false;
1082  // Add blocks with multiple uses.
1083  for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
1084       I != E; ++I)
1085    switch (I->second) {
1086    case 0:
1087    case 1:
1088      continue;
1089    case 2: {
1090      // When there are only two uses and curli is both live in and live out,
1091      // we don't really win anything by isolating the block since we would be
1092      // inserting two copies.
1093      // The remaing register would still have two uses in the block. (Unless it
1094      // separates into disconnected components).
1095      if (lis_.isLiveInToMBB(*curli_, I->first) &&
1096          lis_.isLiveOutOfMBB(*curli_, I->first))
1097        continue;
1098    } // Fall through.
1099    default:
1100      Blocks.insert(I->first);
1101    }
1102  return !Blocks.empty();
1103}
1104
1105/// splitSingleBlocks - Split curli into a separate live interval inside each
1106/// basic block in Blocks.
1107void SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
1108  DEBUG(dbgs() << "  splitSingleBlocks for " << Blocks.size() << " blocks.\n");
1109  // Determine the first and last instruction using curli in each block.
1110  typedef std::pair<SlotIndex,SlotIndex> IndexPair;
1111  typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
1112  IndexPairMap MBBRange;
1113  for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1114       E = sa_.usingInstrs_.end(); I != E; ++I) {
1115    const MachineBasicBlock *MBB = (*I)->getParent();
1116    if (!Blocks.count(MBB))
1117      continue;
1118    SlotIndex Idx = lis_.getInstructionIndex(*I);
1119    DEBUG(dbgs() << "  BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
1120    IndexPair &IP = MBBRange[MBB];
1121    if (!IP.first.isValid() || Idx < IP.first)
1122      IP.first = Idx;
1123    if (!IP.second.isValid() || Idx > IP.second)
1124      IP.second = Idx;
1125  }
1126
1127  // Create a new interval for each block.
1128  for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
1129       E = Blocks.end(); I != E; ++I) {
1130    IndexPair &IP = MBBRange[*I];
1131    DEBUG(dbgs() << "  splitting for BB#" << (*I)->getNumber() << ": ["
1132                 << IP.first << ';' << IP.second << ")\n");
1133    assert(IP.first.isValid() && IP.second.isValid());
1134
1135    openIntv();
1136    enterIntvBefore(IP.first);
1137    useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
1138    leaveIntvAfter(IP.second);
1139    closeIntv();
1140  }
1141  finish();
1142}
1143
1144
1145//===----------------------------------------------------------------------===//
1146//                            Sub Block Splitting
1147//===----------------------------------------------------------------------===//
1148
1149/// getBlockForInsideSplit - If curli is contained inside a single basic block,
1150/// and it wou pay to subdivide the interval inside that block, return it.
1151/// Otherwise return NULL. The returned block can be passed to
1152/// SplitEditor::splitInsideBlock.
1153const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
1154  // The interval must be exclusive to one block.
1155  if (usingBlocks_.size() != 1)
1156    return 0;
1157  // Don't to this for less than 4 instructions. We want to be sure that
1158  // splitting actually reduces the instruction count per interval.
1159  if (usingInstrs_.size() < 4)
1160    return 0;
1161  return usingBlocks_.begin()->first;
1162}
1163
1164/// splitInsideBlock - Split curli into multiple intervals inside MBB.
1165void SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
1166  SmallVector<SlotIndex, 32> Uses;
1167  Uses.reserve(sa_.usingInstrs_.size());
1168  for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1169       E = sa_.usingInstrs_.end(); I != E; ++I)
1170    if ((*I)->getParent() == MBB)
1171      Uses.push_back(lis_.getInstructionIndex(*I));
1172  DEBUG(dbgs() << "  splitInsideBlock BB#" << MBB->getNumber() << " for "
1173               << Uses.size() << " instructions.\n");
1174  assert(Uses.size() >= 3 && "Need at least 3 instructions");
1175  array_pod_sort(Uses.begin(), Uses.end());
1176
1177  // Simple algorithm: Find the largest gap between uses as determined by slot
1178  // indices. Create new intervals for instructions before the gap and after the
1179  // gap.
1180  unsigned bestPos = 0;
1181  int bestGap = 0;
1182  DEBUG(dbgs() << "    dist (" << Uses[0]);
1183  for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
1184    int g = Uses[i-1].distance(Uses[i]);
1185    DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
1186    if (g > bestGap)
1187      bestPos = i, bestGap = g;
1188  }
1189  DEBUG(dbgs() << "), best: -" << bestGap << "-\n");
1190
1191  // bestPos points to the first use after the best gap.
1192  assert(bestPos > 0 && "Invalid gap");
1193
1194  // FIXME: Don't create intervals for low densities.
1195
1196  // First interval before the gap. Don't create single-instr intervals.
1197  if (bestPos > 1) {
1198    openIntv();
1199    enterIntvBefore(Uses.front());
1200    useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
1201    leaveIntvAfter(Uses[bestPos-1]);
1202    closeIntv();
1203  }
1204
1205  // Second interval after the gap.
1206  if (bestPos < Uses.size()-1) {
1207    openIntv();
1208    enterIntvBefore(Uses[bestPos]);
1209    useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
1210    leaveIntvAfter(Uses.back());
1211    closeIntv();
1212  }
1213
1214  finish();
1215}
1216