SplitKit.cpp revision a17768f5822ab62bc18608e5ba473187bf726b84
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 "splitter"
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/MachineInstrBuilder.h"
22#include "llvm/CodeGen/MachineLoopInfo.h"
23#include "llvm/CodeGen/MachineRegisterInfo.h"
24#include "llvm/Support/CommandLine.h"
25#include "llvm/Support/Debug.h"
26#include "llvm/Support/raw_ostream.h"
27#include "llvm/Target/TargetInstrInfo.h"
28#include "llvm/Target/TargetMachine.h"
29
30using namespace llvm;
31
32static cl::opt<bool>
33AllowSplit("spiller-splits-edges",
34           cl::desc("Allow critical edge splitting during spilling"));
35
36//===----------------------------------------------------------------------===//
37//                                 Split Analysis
38//===----------------------------------------------------------------------===//
39
40SplitAnalysis::SplitAnalysis(const MachineFunction &mf,
41                             const LiveIntervals &lis,
42                             const MachineLoopInfo &mli)
43  : mf_(mf),
44    lis_(lis),
45    loops_(mli),
46    tii_(*mf.getTarget().getInstrInfo()),
47    curli_(0) {}
48
49void SplitAnalysis::clear() {
50  usingInstrs_.clear();
51  usingBlocks_.clear();
52  usingLoops_.clear();
53  curli_ = 0;
54}
55
56bool SplitAnalysis::canAnalyzeBranch(const MachineBasicBlock *MBB) {
57  MachineBasicBlock *T, *F;
58  SmallVector<MachineOperand, 4> Cond;
59  return !tii_.AnalyzeBranch(const_cast<MachineBasicBlock&>(*MBB), T, F, Cond);
60}
61
62/// analyzeUses - Count instructions, basic blocks, and loops using curli.
63void SplitAnalysis::analyzeUses() {
64  const MachineRegisterInfo &MRI = mf_.getRegInfo();
65  for (MachineRegisterInfo::reg_iterator I = MRI.reg_begin(curli_->reg);
66       MachineInstr *MI = I.skipInstruction();) {
67    if (MI->isDebugValue() || !usingInstrs_.insert(MI))
68      continue;
69    MachineBasicBlock *MBB = MI->getParent();
70    if (usingBlocks_[MBB]++)
71      continue;
72    for (MachineLoop *Loop = loops_.getLoopFor(MBB); Loop;
73         Loop = Loop->getParentLoop())
74      usingLoops_[Loop]++;
75  }
76  DEBUG(dbgs() << "  counted "
77               << usingInstrs_.size() << " instrs, "
78               << usingBlocks_.size() << " blocks, "
79               << usingLoops_.size()  << " loops.\n");
80}
81
82// Get three sets of basic blocks surrounding a loop: Blocks inside the loop,
83// predecessor blocks, and exit blocks.
84void SplitAnalysis::getLoopBlocks(const MachineLoop *Loop, LoopBlocks &Blocks) {
85  Blocks.clear();
86
87  // Blocks in the loop.
88  Blocks.Loop.insert(Loop->block_begin(), Loop->block_end());
89
90  // Predecessor blocks.
91  const MachineBasicBlock *Header = Loop->getHeader();
92  for (MachineBasicBlock::const_pred_iterator I = Header->pred_begin(),
93       E = Header->pred_end(); I != E; ++I)
94    if (!Blocks.Loop.count(*I))
95      Blocks.Preds.insert(*I);
96
97  // Exit blocks.
98  for (MachineLoop::block_iterator I = Loop->block_begin(),
99       E = Loop->block_end(); I != E; ++I) {
100    const MachineBasicBlock *MBB = *I;
101    for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(),
102       SE = MBB->succ_end(); SI != SE; ++SI)
103      if (!Blocks.Loop.count(*SI))
104        Blocks.Exits.insert(*SI);
105  }
106}
107
108/// analyzeLoopPeripheralUse - Return an enum describing how curli_ is used in
109/// and around the Loop.
110SplitAnalysis::LoopPeripheralUse SplitAnalysis::
111analyzeLoopPeripheralUse(const SplitAnalysis::LoopBlocks &Blocks) {
112  LoopPeripheralUse use = ContainedInLoop;
113  for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
114       I != E; ++I) {
115    const MachineBasicBlock *MBB = I->first;
116    // Is this a peripheral block?
117    if (use < MultiPeripheral &&
118        (Blocks.Preds.count(MBB) || Blocks.Exits.count(MBB))) {
119      if (I->second > 1) use = MultiPeripheral;
120      else               use = SinglePeripheral;
121      continue;
122    }
123    // Is it a loop block?
124    if (Blocks.Loop.count(MBB))
125      continue;
126    // It must be an unrelated block.
127    return OutsideLoop;
128  }
129  return use;
130}
131
132/// getCriticalExits - It may be necessary to partially break critical edges
133/// leaving the loop if an exit block has phi uses of curli. Collect the exit
134/// blocks that need special treatment into CriticalExits.
135void SplitAnalysis::getCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
136                                     BlockPtrSet &CriticalExits) {
137  CriticalExits.clear();
138
139  // A critical exit block contains a phi def of curli, and has a predecessor
140  // that is not in the loop nor a loop predecessor.
141  // For such an exit block, the edges carrying the new variable must be moved
142  // to a new pre-exit block.
143  for (BlockPtrSet::iterator I = Blocks.Exits.begin(), E = Blocks.Exits.end();
144       I != E; ++I) {
145    const MachineBasicBlock *Succ = *I;
146    SlotIndex SuccIdx = lis_.getMBBStartIdx(Succ);
147    VNInfo *SuccVNI = curli_->getVNInfoAt(SuccIdx);
148    // This exit may not have curli live in at all. No need to split.
149    if (!SuccVNI)
150      continue;
151    // If this is not a PHI def, it is either using a value from before the
152    // loop, or a value defined inside the loop. Both are safe.
153    if (!SuccVNI->isPHIDef() || SuccVNI->def.getBaseIndex() != SuccIdx)
154      continue;
155    // This exit block does have a PHI. Does it also have a predecessor that is
156    // not a loop block or loop predecessor?
157    for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
158         PE = Succ->pred_end(); PI != PE; ++PI) {
159      const MachineBasicBlock *Pred = *PI;
160      if (Blocks.Loop.count(Pred) || Blocks.Preds.count(Pred))
161        continue;
162      // This is a critical exit block, and we need to split the exit edge.
163      CriticalExits.insert(Succ);
164      break;
165    }
166  }
167}
168
169/// canSplitCriticalExits - Return true if it is possible to insert new exit
170/// blocks before the blocks in CriticalExits.
171bool
172SplitAnalysis::canSplitCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
173                                     BlockPtrSet &CriticalExits) {
174  // If we don't allow critical edge splitting, require no critical exits.
175  if (!AllowSplit)
176    return CriticalExits.empty();
177
178  for (BlockPtrSet::iterator I = CriticalExits.begin(), E = CriticalExits.end();
179       I != E; ++I) {
180    const MachineBasicBlock *Succ = *I;
181    // We want to insert a new pre-exit MBB before Succ, and change all the
182    // in-loop blocks to branch to the pre-exit instead of Succ.
183    // Check that all the in-loop predecessors can be changed.
184    for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
185         PE = Succ->pred_end(); PI != PE; ++PI) {
186      const MachineBasicBlock *Pred = *PI;
187      // The external predecessors won't be altered.
188      if (!Blocks.Loop.count(Pred) && !Blocks.Preds.count(Pred))
189        continue;
190      if (!canAnalyzeBranch(Pred))
191        return false;
192    }
193
194    // If Succ's layout predecessor falls through, that too must be analyzable.
195    // We need to insert the pre-exit block in the gap.
196    MachineFunction::const_iterator MFI = Succ;
197    if (MFI == mf_.begin())
198      continue;
199    if (!canAnalyzeBranch(--MFI))
200      return false;
201  }
202  // No problems found.
203  return true;
204}
205
206void SplitAnalysis::analyze(const LiveInterval *li) {
207  clear();
208  curli_ = li;
209  analyzeUses();
210}
211
212const MachineLoop *SplitAnalysis::getBestSplitLoop() {
213  assert(curli_ && "Call analyze() before getBestSplitLoop");
214  if (usingLoops_.empty())
215    return 0;
216
217  LoopPtrSet Loops;
218  LoopBlocks Blocks;
219  BlockPtrSet CriticalExits;
220
221  // We split around loops where curli is used outside the periphery.
222  for (LoopCountMap::const_iterator I = usingLoops_.begin(),
223       E = usingLoops_.end(); I != E; ++I) {
224    const MachineLoop *Loop = I->first;
225    getLoopBlocks(Loop, Blocks);
226
227    switch(analyzeLoopPeripheralUse(Blocks)) {
228    case OutsideLoop:
229      break;
230    case MultiPeripheral:
231      // FIXME: We could split a live range with multiple uses in a peripheral
232      // block and still make progress. However, it is possible that splitting
233      // another live range will insert copies into a peripheral block, and
234      // there is a small chance we can enter an infinity loop, inserting copies
235      // forever.
236      // For safety, stick to splitting live ranges with uses outside the
237      // periphery.
238      DEBUG(dbgs() << "  multiple peripheral uses in " << *Loop);
239      break;
240    case ContainedInLoop:
241      DEBUG(dbgs() << "  contained in " << *Loop);
242      continue;
243    case SinglePeripheral:
244      DEBUG(dbgs() << "  single peripheral use in " << *Loop);
245      continue;
246    }
247    // Will it be possible to split around this loop?
248    getCriticalExits(Blocks, CriticalExits);
249    DEBUG(dbgs() << "  " << CriticalExits.size() << " critical exits from "
250                 << *Loop);
251    if (!canSplitCriticalExits(Blocks, CriticalExits))
252      continue;
253    // This is a possible split.
254    Loops.insert(Loop);
255  }
256
257  DEBUG(dbgs() << "  getBestSplitLoop found " << Loops.size()
258               << " candidate loops.\n");
259
260  if (Loops.empty())
261    return 0;
262
263  // Pick the earliest loop.
264  // FIXME: Are there other heuristics to consider?
265  const MachineLoop *Best = 0;
266  SlotIndex BestIdx;
267  for (LoopPtrSet::const_iterator I = Loops.begin(), E = Loops.end(); I != E;
268       ++I) {
269    SlotIndex Idx = lis_.getMBBStartIdx((*I)->getHeader());
270    if (!Best || Idx < BestIdx)
271      Best = *I, BestIdx = Idx;
272  }
273  DEBUG(dbgs() << "  getBestSplitLoop found " << *Best);
274  return Best;
275}
276
277/// getMultiUseBlocks - if curli has more than one use in a basic block, it
278/// may be an advantage to split curli for the duration of the block.
279bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
280  // If curli is local to one block, there is no point to splitting it.
281  if (usingBlocks_.size() <= 1)
282    return false;
283  // Add blocks with multiple uses.
284  for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
285       I != E; ++I)
286    switch (I->second) {
287    case 0:
288    case 1:
289      continue;
290    case 2: {
291      // It doesn't pay to split a 2-instr block if it redefines curli.
292      VNInfo *VN1 = curli_->getVNInfoAt(lis_.getMBBStartIdx(I->first));
293      VNInfo *VN2 =
294        curli_->getVNInfoAt(lis_.getMBBEndIdx(I->first).getPrevIndex());
295      // live-in and live-out with a different value.
296      if (VN1 && VN2 && VN1 != VN2)
297        continue;
298    } // Fall through.
299    default:
300      Blocks.insert(I->first);
301    }
302  return !Blocks.empty();
303}
304
305//===----------------------------------------------------------------------===//
306//                               LiveIntervalMap
307//===----------------------------------------------------------------------===//
308
309// Work around the fact that the std::pair constructors are broken for pointer
310// pairs in some implementations. makeVV(x, 0) works.
311static inline std::pair<const VNInfo*, VNInfo*>
312makeVV(const VNInfo *a, VNInfo *b) {
313  return std::make_pair(a, b);
314}
315
316void LiveIntervalMap::reset(LiveInterval *li) {
317  li_ = li;
318  valueMap_.clear();
319}
320
321bool LiveIntervalMap::isComplexMapped(const VNInfo *ParentVNI) const {
322  ValueMap::const_iterator i = valueMap_.find(ParentVNI);
323  return i != valueMap_.end() && i->second == 0;
324}
325
326// defValue - Introduce a li_ def for ParentVNI that could be later than
327// ParentVNI->def.
328VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) {
329  assert(li_ && "call reset first");
330  assert(ParentVNI && "Mapping  NULL value");
331  assert(Idx.isValid() && "Invalid SlotIndex");
332  assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
333
334  // Create a new value.
335  VNInfo *VNI = li_->getNextValue(Idx, 0, lis_.getVNInfoAllocator());
336
337  // Use insert for lookup, so we can add missing values with a second lookup.
338  std::pair<ValueMap::iterator,bool> InsP =
339    valueMap_.insert(makeVV(ParentVNI, Idx == ParentVNI->def ? VNI : 0));
340
341  // This is now a complex def. Mark with a NULL in valueMap.
342  if (!InsP.second)
343    InsP.first->second = 0;
344
345  return VNI;
346}
347
348
349// mapValue - Find the mapped value for ParentVNI at Idx.
350// Potentially create phi-def values.
351VNInfo *LiveIntervalMap::mapValue(const VNInfo *ParentVNI, SlotIndex Idx,
352                                  bool *simple) {
353  assert(li_ && "call reset first");
354  assert(ParentVNI && "Mapping  NULL value");
355  assert(Idx.isValid() && "Invalid SlotIndex");
356  assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
357
358  // Use insert for lookup, so we can add missing values with a second lookup.
359  std::pair<ValueMap::iterator,bool> InsP =
360    valueMap_.insert(makeVV(ParentVNI, 0));
361
362  // This was an unknown value. Create a simple mapping.
363  if (InsP.second) {
364    if (simple) *simple = true;
365    return InsP.first->second = li_->createValueCopy(ParentVNI,
366                                                     lis_.getVNInfoAllocator());
367  }
368
369  // This was a simple mapped value.
370  if (InsP.first->second) {
371    if (simple) *simple = true;
372    return InsP.first->second;
373  }
374
375  // This is a complex mapped value. There may be multiple defs, and we may need
376  // to create phi-defs.
377  if (simple) *simple = false;
378  MachineBasicBlock *IdxMBB = lis_.getMBBFromIndex(Idx);
379  assert(IdxMBB && "No MBB at Idx");
380
381  // Is there a def in the same MBB we can extend?
382  if (VNInfo *VNI = extendTo(IdxMBB, Idx))
383    return VNI;
384
385  // Now for the fun part. We know that ParentVNI potentially has multiple defs,
386  // and we may need to create even more phi-defs to preserve VNInfo SSA form.
387  // Perform a depth-first search for predecessor blocks where we know the
388  // dominating VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
389
390  // Track MBBs where we have created or learned the dominating value.
391  // This may change during the DFS as we create new phi-defs.
392  typedef DenseMap<MachineBasicBlock*, VNInfo*> MBBValueMap;
393  MBBValueMap DomValue;
394  typedef SplitAnalysis::BlockPtrSet BlockPtrSet;
395  BlockPtrSet Visited;
396
397  // Iterate over IdxMBB predecessors in a depth-first order.
398  // Skip begin() since that is always IdxMBB.
399  for (idf_ext_iterator<MachineBasicBlock*, BlockPtrSet>
400         IDFI = llvm::next(idf_ext_begin(IdxMBB, Visited)),
401         IDFE = idf_ext_end(IdxMBB, Visited); IDFI != IDFE;) {
402    MachineBasicBlock *MBB = *IDFI;
403    SlotIndex End = lis_.getMBBEndIdx(MBB).getPrevSlot();
404
405    // We are operating on the restricted CFG where ParentVNI is live.
406    if (parentli_.getVNInfoAt(End) != ParentVNI) {
407      IDFI.skipChildren();
408      continue;
409    }
410
411    // Do we have a dominating value in this block?
412    VNInfo *VNI = extendTo(MBB, End);
413    if (!VNI) {
414      ++IDFI;
415      continue;
416    }
417
418    // Yes, VNI dominates MBB. Make sure we visit MBB again from other paths.
419    Visited.erase(MBB);
420
421    // Track the path back to IdxMBB, creating phi-defs
422    // as needed along the way.
423    for (unsigned PI = IDFI.getPathLength()-1; PI != 0; --PI) {
424      // Start from MBB's immediate successor. End at IdxMBB.
425      MachineBasicBlock *Succ = IDFI.getPath(PI-1);
426      std::pair<MBBValueMap::iterator, bool> InsP =
427        DomValue.insert(MBBValueMap::value_type(Succ, VNI));
428
429      // This is the first time we backtrack to Succ.
430      if (InsP.second)
431        continue;
432
433      // We reached Succ again with the same VNI. Nothing is going to change.
434      VNInfo *OVNI = InsP.first->second;
435      if (OVNI == VNI)
436        break;
437
438      // Succ already has a phi-def. No need to continue.
439      SlotIndex Start = lis_.getMBBStartIdx(Succ);
440      if (OVNI->def == Start)
441        break;
442
443      // We have a collision between the old and new VNI at Succ. That means
444      // neither dominates and we need a new phi-def.
445      VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator());
446      VNI->setIsPHIDef(true);
447      InsP.first->second = VNI;
448
449      // Replace OVNI with VNI in the remaining path.
450      for (; PI > 1 ; --PI) {
451        MBBValueMap::iterator I = DomValue.find(IDFI.getPath(PI-2));
452        if (I == DomValue.end() || I->second != OVNI)
453          break;
454        I->second = VNI;
455      }
456    }
457
458    // No need to search the children, we found a dominating value.
459    IDFI.skipChildren();
460  }
461
462  // The search should at least find a dominating value for IdxMBB.
463  assert(!DomValue.empty() && "Couldn't find a reaching definition");
464
465  // Since we went through the trouble of a full DFS visiting all reaching defs,
466  // the values in DomValue are now accurate. No more phi-defs are needed for
467  // these blocks, so we can color the live ranges.
468  // This makes the next mapValue call much faster.
469  VNInfo *IdxVNI = 0;
470  for (MBBValueMap::iterator I = DomValue.begin(), E = DomValue.end(); I != E;
471       ++I) {
472     MachineBasicBlock *MBB = I->first;
473     VNInfo *VNI = I->second;
474     SlotIndex Start = lis_.getMBBStartIdx(MBB);
475     if (MBB == IdxMBB) {
476       // Don't add full liveness to IdxMBB, stop at Idx.
477       if (Start != Idx)
478         li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
479       // The caller had better add some liveness to IdxVNI, or it leaks.
480       IdxVNI = VNI;
481     } else
482      li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
483  }
484
485  assert(IdxVNI && "Didn't find value for Idx");
486  return IdxVNI;
487}
488
489// extendTo - Find the last li_ value defined in MBB at or before Idx. The
490// parentli_ is assumed to be live at Idx. Extend the live range to Idx.
491// Return the found VNInfo, or NULL.
492VNInfo *LiveIntervalMap::extendTo(MachineBasicBlock *MBB, SlotIndex Idx) {
493  assert(li_ && "call reset first");
494  LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx);
495  if (I == li_->begin())
496    return 0;
497  --I;
498  if (I->end <= lis_.getMBBStartIdx(MBB))
499    return 0;
500  if (I->end <= Idx)
501    I->end = Idx.getNextSlot();
502  return I->valno;
503}
504
505// addSimpleRange - Add a simple range from parentli_ to li_.
506// ParentVNI must be live in the [Start;End) interval.
507void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End,
508                                     const VNInfo *ParentVNI) {
509  assert(li_ && "call reset first");
510  bool simple;
511  VNInfo *VNI = mapValue(ParentVNI, Start, &simple);
512  // A simple mapping is easy.
513  if (simple) {
514    li_->addRange(LiveRange(Start, End, VNI));
515    return;
516  }
517
518  // ParentVNI is a complex value. We must map per MBB.
519  MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
520  MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End.getPrevSlot());
521
522  if (MBB == MBBE) {
523    li_->addRange(LiveRange(Start, End, VNI));
524    return;
525  }
526
527  // First block.
528  li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
529
530  // Run sequence of full blocks.
531  for (++MBB; MBB != MBBE; ++MBB) {
532    Start = lis_.getMBBStartIdx(MBB);
533    li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB),
534                            mapValue(ParentVNI, Start)));
535  }
536
537  // Final block.
538  Start = lis_.getMBBStartIdx(MBB);
539  if (Start != End)
540    li_->addRange(LiveRange(Start, End, mapValue(ParentVNI, Start)));
541}
542
543/// addRange - Add live ranges to li_ where [Start;End) intersects parentli_.
544/// All needed values whose def is not inside [Start;End) must be defined
545/// beforehand so mapValue will work.
546void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) {
547  assert(li_ && "call reset first");
548  LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end();
549  LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
550
551  // Check if --I begins before Start and overlaps.
552  if (I != B) {
553    --I;
554    if (I->end > Start)
555      addSimpleRange(Start, std::min(End, I->end), I->valno);
556    ++I;
557  }
558
559  // The remaining ranges begin after Start.
560  for (;I != E && I->start < End; ++I)
561    addSimpleRange(I->start, std::min(End, I->end), I->valno);
562}
563
564VNInfo *LiveIntervalMap::defByCopyFrom(unsigned Reg,
565                                       const VNInfo *ParentVNI,
566                                       MachineBasicBlock &MBB,
567                                       MachineBasicBlock::iterator I) {
568  const TargetInstrDesc &TID = MBB.getParent()->getTarget().getInstrInfo()->
569    get(TargetOpcode::COPY);
570  MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), TID, li_->reg).addReg(Reg);
571  SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
572  VNInfo *VNI = defValue(ParentVNI, DefIdx);
573  VNI->setCopy(MI);
574  li_->addRange(LiveRange(DefIdx, DefIdx.getNextSlot(), VNI));
575  return VNI;
576}
577
578//===----------------------------------------------------------------------===//
579//                               Split Editor
580//===----------------------------------------------------------------------===//
581
582/// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
583SplitEditor::SplitEditor(SplitAnalysis &sa, LiveIntervals &lis, VirtRegMap &vrm,
584                         LiveRangeEdit &edit)
585  : sa_(sa), lis_(lis), vrm_(vrm),
586    mri_(vrm.getMachineFunction().getRegInfo()),
587    tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
588    edit_(edit),
589    curli_(sa_.getCurLI()),
590    dupli_(lis_, *curli_),
591    openli_(lis_, *curli_)
592{
593  assert(curli_ && "SplitEditor created from empty SplitAnalysis");
594
595  // Make sure curli_ is assigned a stack slot, so all our intervals get the
596  // same slot as curli_.
597  if (vrm_.getStackSlot(curli_->reg) == VirtRegMap::NO_STACK_SLOT)
598    vrm_.assignVirt2StackSlot(curli_->reg);
599
600}
601
602bool SplitEditor::intervalsLiveAt(SlotIndex Idx) const {
603  for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
604    if (*I != dupli_.getLI() && (*I)->liveAt(Idx))
605      return true;
606  return false;
607}
608
609/// Create a new virtual register and live interval.
610void SplitEditor::openIntv() {
611  assert(!openli_.getLI() && "Previous LI not closed before openIntv");
612
613  if (!dupli_.getLI())
614    dupli_.reset(&edit_.create(mri_, lis_, vrm_));
615
616  openli_.reset(&edit_.create(mri_, lis_, vrm_));
617}
618
619/// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
620/// not live before Idx, a COPY is not inserted.
621void SplitEditor::enterIntvBefore(SlotIndex Idx) {
622  assert(openli_.getLI() && "openIntv not called before enterIntvBefore");
623  DEBUG(dbgs() << "    enterIntvBefore " << Idx);
624  VNInfo *ParentVNI = curli_->getVNInfoAt(Idx.getUseIndex());
625  if (!ParentVNI) {
626    DEBUG(dbgs() << ": not live\n");
627    return;
628  }
629  DEBUG(dbgs() << ": valno " << ParentVNI->id);
630  truncatedValues.insert(ParentVNI);
631  MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
632  assert(MI && "enterIntvBefore called with invalid index");
633  VNInfo *VNI = openli_.defByCopyFrom(curli_->reg, ParentVNI,
634                                      *MI->getParent(), MI);
635  openli_.getLI()->addRange(LiveRange(VNI->def, Idx.getDefIndex(), VNI));
636  DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
637}
638
639/// enterIntvAtEnd - Enter openli at the end of MBB.
640void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) {
641  assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd");
642  SlotIndex End = lis_.getMBBEndIdx(&MBB);
643  DEBUG(dbgs() << "    enterIntvAtEnd BB#" << MBB.getNumber() << ", " << End);
644  VNInfo *ParentVNI = curli_->getVNInfoAt(End.getPrevSlot());
645  if (!ParentVNI) {
646    DEBUG(dbgs() << ": not live\n");
647    return;
648  }
649  DEBUG(dbgs() << ": valno " << ParentVNI->id);
650  truncatedValues.insert(ParentVNI);
651  VNInfo *VNI = openli_.defByCopyFrom(curli_->reg, ParentVNI,
652                                      MBB, MBB.getFirstTerminator());
653  // Make sure openli is live out of MBB.
654  openli_.getLI()->addRange(LiveRange(VNI->def, End, VNI));
655  DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
656}
657
658/// useIntv - indicate that all instructions in MBB should use openli.
659void SplitEditor::useIntv(const MachineBasicBlock &MBB) {
660  useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB));
661}
662
663void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) {
664  assert(openli_.getLI() && "openIntv not called before useIntv");
665  openli_.addRange(Start, End);
666  DEBUG(dbgs() << "    use [" << Start << ';' << End << "): "
667               << *openli_.getLI() << '\n');
668}
669
670/// leaveIntvAfter - Leave openli after the instruction at Idx.
671void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
672  assert(openli_.getLI() && "openIntv not called before leaveIntvAfter");
673  DEBUG(dbgs() << "    leaveIntvAfter " << Idx);
674
675  // The interval must be live beyond the instruction at Idx.
676  VNInfo *ParentVNI = curli_->getVNInfoAt(Idx.getBoundaryIndex());
677  if (!ParentVNI) {
678    DEBUG(dbgs() << ": not live\n");
679    return;
680  }
681  DEBUG(dbgs() << ": valno " << ParentVNI->id);
682
683  MachineBasicBlock::iterator MII = lis_.getInstructionFromIndex(Idx);
684  MachineBasicBlock *MBB = MII->getParent();
685  VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI, *MBB,
686                                     llvm::next(MII));
687
688  // Finally we must make sure that openli is properly extended from Idx to the
689  // new copy.
690  openli_.addSimpleRange(Idx.getBoundaryIndex(), VNI->def, ParentVNI);
691  DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
692}
693
694/// leaveIntvAtTop - Leave the interval at the top of MBB.
695/// Currently, only one value can leave the interval.
696void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
697  assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop");
698  SlotIndex Start = lis_.getMBBStartIdx(&MBB);
699  DEBUG(dbgs() << "    leaveIntvAtTop BB#" << MBB.getNumber() << ", " << Start);
700
701  VNInfo *ParentVNI = curli_->getVNInfoAt(Start);
702  if (!ParentVNI) {
703    DEBUG(dbgs() << ": not live\n");
704    return;
705  }
706
707  // We are going to insert a back copy, so we must have a dupli_.
708  VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
709                                     MBB, MBB.begin());
710
711  // Finally we must make sure that openli is properly extended from Start to
712  // the new copy.
713  openli_.addSimpleRange(Start, VNI->def, ParentVNI);
714  DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
715}
716
717/// closeIntv - Indicate that we are done editing the currently open
718/// LiveInterval, and ranges can be trimmed.
719void SplitEditor::closeIntv() {
720  assert(openli_.getLI() && "openIntv not called before closeIntv");
721
722  DEBUG(dbgs() << "    closeIntv cleaning up\n");
723  DEBUG(dbgs() << "    open " << *openli_.getLI() << '\n');
724  openli_.reset(0);
725}
726
727/// rewrite - Rewrite all uses of reg to use the new registers.
728void SplitEditor::rewrite(unsigned reg) {
729  for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(reg),
730       RE = mri_.reg_end(); RI != RE;) {
731    MachineOperand &MO = RI.getOperand();
732    MachineInstr *MI = MO.getParent();
733    ++RI;
734    if (MI->isDebugValue()) {
735      DEBUG(dbgs() << "Zapping " << *MI);
736      // FIXME: We can do much better with debug values.
737      MO.setReg(0);
738      continue;
739    }
740    SlotIndex Idx = lis_.getInstructionIndex(MI);
741    Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
742    LiveInterval *LI = 0;
743    for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E;
744         ++I) {
745      LiveInterval *testli = *I;
746      if (testli->liveAt(Idx)) {
747        LI = testli;
748        break;
749      }
750    }
751    assert(LI && "No register was live at use");
752    MO.setReg(LI->reg);
753    DEBUG(dbgs() << "  rewrite BB#" << MI->getParent()->getNumber() << '\t'
754                 << Idx << '\t' << *MI);
755  }
756}
757
758void
759SplitEditor::addTruncSimpleRange(SlotIndex Start, SlotIndex End, VNInfo *VNI) {
760  // Build vector of iterator pairs from the intervals.
761  typedef std::pair<LiveInterval::const_iterator,
762                    LiveInterval::const_iterator> IIPair;
763  SmallVector<IIPair, 8> Iters;
764  for (LiveRangeEdit::iterator LI = edit_.begin(), LE = edit_.end(); LI != LE;
765       ++LI) {
766    LiveInterval::const_iterator I = (*LI)->find(Start);
767    LiveInterval::const_iterator E = (*LI)->end();
768    if (I != E)
769      Iters.push_back(std::make_pair(I, E));
770  }
771
772  SlotIndex sidx = Start;
773  // Break [Start;End) into segments that don't overlap any intervals.
774  for (;;) {
775    SlotIndex next = sidx, eidx = End;
776    // Find overlapping intervals.
777    for (unsigned i = 0; i != Iters.size() && sidx < eidx; ++i) {
778      LiveInterval::const_iterator I = Iters[i].first;
779      // Interval I is overlapping [sidx;eidx). Trim sidx.
780      if (I->start <= sidx) {
781        sidx = I->end;
782        // Move to the next run, remove iters when all are consumed.
783        I = ++Iters[i].first;
784        if (I == Iters[i].second) {
785          Iters.erase(Iters.begin() + i);
786          --i;
787          continue;
788        }
789      }
790      // Trim eidx too if needed.
791      if (I->start >= eidx)
792        continue;
793      eidx = I->start;
794      next = I->end;
795    }
796    // Now, [sidx;eidx) doesn't overlap anything in intervals_.
797    if (sidx < eidx)
798      dupli_.addSimpleRange(sidx, eidx, VNI);
799    // If the interval end was truncated, we can try again from next.
800    if (next <= sidx)
801      break;
802    sidx = next;
803  }
804}
805
806void SplitEditor::computeRemainder() {
807  // First we need to fill in the live ranges in dupli.
808  // If values were redefined, we need a full recoloring with SSA update.
809  // If values were truncated, we only need to truncate the ranges.
810  // If values were partially rematted, we should shrink to uses.
811  // If values were fully rematted, they should be omitted.
812  // FIXME: If a single value is redefined, just move the def and truncate.
813
814  // Values that are fully contained in the split intervals.
815  SmallPtrSet<const VNInfo*, 8> deadValues;
816
817  // Map all curli values that should have live defs in dupli.
818  for (LiveInterval::const_vni_iterator I = curli_->vni_begin(),
819       E = curli_->vni_end(); I != E; ++I) {
820    const VNInfo *VNI = *I;
821    // Original def is contained in the split intervals.
822    if (intervalsLiveAt(VNI->def)) {
823      // Did this value escape?
824      if (dupli_.isMapped(VNI))
825        truncatedValues.insert(VNI);
826      else
827        deadValues.insert(VNI);
828      continue;
829    }
830    // Add minimal live range at the definition.
831    VNInfo *DVNI = dupli_.defValue(VNI, VNI->def);
832    dupli_.getLI()->addRange(LiveRange(VNI->def, VNI->def.getNextSlot(), DVNI));
833  }
834
835  // Add all ranges to dupli.
836  for (LiveInterval::const_iterator I = curli_->begin(), E = curli_->end();
837       I != E; ++I) {
838    const LiveRange &LR = *I;
839    if (truncatedValues.count(LR.valno)) {
840      // recolor after removing intervals_.
841      addTruncSimpleRange(LR.start, LR.end, LR.valno);
842    } else if (!deadValues.count(LR.valno)) {
843      // recolor without truncation.
844      dupli_.addSimpleRange(LR.start, LR.end, LR.valno);
845    }
846  }
847}
848
849void SplitEditor::finish() {
850  assert(!openli_.getLI() && "Previous LI not closed before rewrite");
851  assert(dupli_.getLI() && "No dupli for rewrite. Noop spilt?");
852
853  // Complete dupli liveness.
854  computeRemainder();
855
856  // Get rid of unused values and set phi-kill flags.
857  dupli_.getLI()->RenumberValues(lis_);
858
859  // Now check if dupli was separated into multiple connected components.
860  ConnectedVNInfoEqClasses ConEQ(lis_);
861  if (unsigned NumComp = ConEQ.Classify(dupli_.getLI())) {
862    DEBUG(dbgs() << "  Remainder has " << NumComp << " connected components: "
863                 << *dupli_.getLI() << '\n');
864    // Did the remainder break up? Create intervals for all the components.
865    if (NumComp > 1) {
866      SmallVector<LiveInterval*, 8> dups;
867      dups.push_back(dupli_.getLI());
868      for (unsigned i = 1; i != NumComp; ++i)
869        dups.push_back(&edit_.create(mri_, lis_, vrm_));
870      ConEQ.Distribute(&dups[0]);
871      // Rewrite uses to the new regs.
872      rewrite(dupli_.getLI()->reg);
873    }
874  }
875
876  // Rewrite instructions.
877  rewrite(curli_->reg);
878
879  // Calculate spill weight and allocation hints for new intervals.
880  VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
881  for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I){
882    LiveInterval &li = **I;
883    vrai.CalculateRegClass(li.reg);
884    vrai.CalculateWeightAndHint(li);
885    DEBUG(dbgs() << "  new interval " << mri_.getRegClass(li.reg)->getName()
886                 << ":" << li << '\n');
887  }
888}
889
890
891//===----------------------------------------------------------------------===//
892//                               Loop Splitting
893//===----------------------------------------------------------------------===//
894
895void SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
896  SplitAnalysis::LoopBlocks Blocks;
897  sa_.getLoopBlocks(Loop, Blocks);
898
899  DEBUG({
900    dbgs() << "  splitAroundLoop";
901    for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
902         E = Blocks.Loop.end(); I != E; ++I)
903      dbgs() << " BB#" << (*I)->getNumber();
904    dbgs() << ", preds:";
905    for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
906         E = Blocks.Preds.end(); I != E; ++I)
907      dbgs() << " BB#" << (*I)->getNumber();
908    dbgs() << ", exits:";
909    for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
910         E = Blocks.Exits.end(); I != E; ++I)
911      dbgs() << " BB#" << (*I)->getNumber();
912    dbgs() << '\n';
913  });
914
915  // Break critical edges as needed.
916  SplitAnalysis::BlockPtrSet CriticalExits;
917  sa_.getCriticalExits(Blocks, CriticalExits);
918  assert(CriticalExits.empty() && "Cannot break critical exits yet");
919
920  // Create new live interval for the loop.
921  openIntv();
922
923  // Insert copies in the predecessors.
924  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
925       E = Blocks.Preds.end(); I != E; ++I) {
926    MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
927    enterIntvAtEnd(MBB);
928  }
929
930  // Switch all loop blocks.
931  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
932       E = Blocks.Loop.end(); I != E; ++I)
933     useIntv(**I);
934
935  // Insert back copies in the exit blocks.
936  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
937       E = Blocks.Exits.end(); I != E; ++I) {
938    MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
939    leaveIntvAtTop(MBB);
940  }
941
942  // Done.
943  closeIntv();
944  finish();
945}
946
947
948//===----------------------------------------------------------------------===//
949//                            Single Block Splitting
950//===----------------------------------------------------------------------===//
951
952/// splitSingleBlocks - Split curli into a separate live interval inside each
953/// basic block in Blocks.
954void SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
955  DEBUG(dbgs() << "  splitSingleBlocks for " << Blocks.size() << " blocks.\n");
956  // Determine the first and last instruction using curli in each block.
957  typedef std::pair<SlotIndex,SlotIndex> IndexPair;
958  typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
959  IndexPairMap MBBRange;
960  for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
961       E = sa_.usingInstrs_.end(); I != E; ++I) {
962    const MachineBasicBlock *MBB = (*I)->getParent();
963    if (!Blocks.count(MBB))
964      continue;
965    SlotIndex Idx = lis_.getInstructionIndex(*I);
966    DEBUG(dbgs() << "  BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
967    IndexPair &IP = MBBRange[MBB];
968    if (!IP.first.isValid() || Idx < IP.first)
969      IP.first = Idx;
970    if (!IP.second.isValid() || Idx > IP.second)
971      IP.second = Idx;
972  }
973
974  // Create a new interval for each block.
975  for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
976       E = Blocks.end(); I != E; ++I) {
977    IndexPair &IP = MBBRange[*I];
978    DEBUG(dbgs() << "  splitting for BB#" << (*I)->getNumber() << ": ["
979                 << IP.first << ';' << IP.second << ")\n");
980    assert(IP.first.isValid() && IP.second.isValid());
981
982    openIntv();
983    enterIntvBefore(IP.first);
984    useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
985    leaveIntvAfter(IP.second);
986    closeIntv();
987  }
988  finish();
989}
990
991
992//===----------------------------------------------------------------------===//
993//                            Sub Block Splitting
994//===----------------------------------------------------------------------===//
995
996/// getBlockForInsideSplit - If curli is contained inside a single basic block,
997/// and it wou pay to subdivide the interval inside that block, return it.
998/// Otherwise return NULL. The returned block can be passed to
999/// SplitEditor::splitInsideBlock.
1000const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
1001  // The interval must be exclusive to one block.
1002  if (usingBlocks_.size() != 1)
1003    return 0;
1004  // Don't to this for less than 4 instructions. We want to be sure that
1005  // splitting actually reduces the instruction count per interval.
1006  if (usingInstrs_.size() < 4)
1007    return 0;
1008  return usingBlocks_.begin()->first;
1009}
1010
1011/// splitInsideBlock - Split curli into multiple intervals inside MBB.
1012void SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
1013  SmallVector<SlotIndex, 32> Uses;
1014  Uses.reserve(sa_.usingInstrs_.size());
1015  for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1016       E = sa_.usingInstrs_.end(); I != E; ++I)
1017    if ((*I)->getParent() == MBB)
1018      Uses.push_back(lis_.getInstructionIndex(*I));
1019  DEBUG(dbgs() << "  splitInsideBlock BB#" << MBB->getNumber() << " for "
1020               << Uses.size() << " instructions.\n");
1021  assert(Uses.size() >= 3 && "Need at least 3 instructions");
1022  array_pod_sort(Uses.begin(), Uses.end());
1023
1024  // Simple algorithm: Find the largest gap between uses as determined by slot
1025  // indices. Create new intervals for instructions before the gap and after the
1026  // gap.
1027  unsigned bestPos = 0;
1028  int bestGap = 0;
1029  DEBUG(dbgs() << "    dist (" << Uses[0]);
1030  for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
1031    int g = Uses[i-1].distance(Uses[i]);
1032    DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
1033    if (g > bestGap)
1034      bestPos = i, bestGap = g;
1035  }
1036  DEBUG(dbgs() << "), best: -" << bestGap << "-\n");
1037
1038  // bestPos points to the first use after the best gap.
1039  assert(bestPos > 0 && "Invalid gap");
1040
1041  // FIXME: Don't create intervals for low densities.
1042
1043  // First interval before the gap. Don't create single-instr intervals.
1044  if (bestPos > 1) {
1045    openIntv();
1046    enterIntvBefore(Uses.front());
1047    useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
1048    leaveIntvAfter(Uses[bestPos-1]);
1049    closeIntv();
1050  }
1051
1052  // Second interval after the gap.
1053  if (bestPos < Uses.size()-1) {
1054    openIntv();
1055    enterIntvBefore(Uses[bestPos]);
1056    useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
1057    leaveIntvAfter(Uses.back());
1058    closeIntv();
1059  }
1060
1061  finish();
1062}
1063