SplitKit.cpp revision 6715be20e23ac19efd20e16d43fdd1f608e3e415
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 "VirtRegMap.h"
18#include "llvm/CodeGen/CalcSpillWeights.h"
19#include "llvm/CodeGen/LiveIntervalAnalysis.h"
20#include "llvm/CodeGen/MachineInstrBuilder.h"
21#include "llvm/CodeGen/MachineLoopInfo.h"
22#include "llvm/CodeGen/MachineRegisterInfo.h"
23#include "llvm/Support/CommandLine.h"
24#include "llvm/Support/Debug.h"
25#include "llvm/Support/raw_ostream.h"
26#include "llvm/Target/TargetInstrInfo.h"
27#include "llvm/Target/TargetMachine.h"
28
29using namespace llvm;
30
31static cl::opt<bool>
32AllowSplit("spiller-splits-edges",
33           cl::desc("Allow critical edge splitting during spilling"));
34
35//===----------------------------------------------------------------------===//
36//                                 Split Analysis
37//===----------------------------------------------------------------------===//
38
39SplitAnalysis::SplitAnalysis(const MachineFunction &mf,
40                             const LiveIntervals &lis,
41                             const MachineLoopInfo &mli)
42  : mf_(mf),
43    lis_(lis),
44    loops_(mli),
45    tii_(*mf.getTarget().getInstrInfo()),
46    curli_(0) {}
47
48void SplitAnalysis::clear() {
49  usingInstrs_.clear();
50  usingBlocks_.clear();
51  usingLoops_.clear();
52  curli_ = 0;
53}
54
55bool SplitAnalysis::canAnalyzeBranch(const MachineBasicBlock *MBB) {
56  MachineBasicBlock *T, *F;
57  SmallVector<MachineOperand, 4> Cond;
58  return !tii_.AnalyzeBranch(const_cast<MachineBasicBlock&>(*MBB), T, F, Cond);
59}
60
61/// analyzeUses - Count instructions, basic blocks, and loops using curli.
62void SplitAnalysis::analyzeUses() {
63  const MachineRegisterInfo &MRI = mf_.getRegInfo();
64  for (MachineRegisterInfo::reg_iterator I = MRI.reg_begin(curli_->reg);
65       MachineInstr *MI = I.skipInstruction();) {
66    if (MI->isDebugValue() || !usingInstrs_.insert(MI))
67      continue;
68    MachineBasicBlock *MBB = MI->getParent();
69    if (usingBlocks_[MBB]++)
70      continue;
71    if (MachineLoop *Loop = loops_.getLoopFor(MBB))
72      usingLoops_[Loop]++;
73  }
74  DEBUG(dbgs() << "  counted "
75               << usingInstrs_.size() << " instrs, "
76               << usingBlocks_.size() << " blocks, "
77               << usingLoops_.size()  << " loops.\n");
78}
79
80/// removeUse - Update statistics by noting that MI no longer uses curli.
81void SplitAnalysis::removeUse(const MachineInstr *MI) {
82  if (!usingInstrs_.erase(MI))
83    return;
84
85  // Decrement MBB count.
86  const MachineBasicBlock *MBB = MI->getParent();
87  BlockCountMap::iterator bi = usingBlocks_.find(MBB);
88  assert(bi != usingBlocks_.end() && "MBB missing");
89  assert(bi->second && "0 count in map");
90  if (--bi->second)
91    return;
92  // No more uses in MBB.
93  usingBlocks_.erase(bi);
94
95  // Decrement loop count.
96  MachineLoop *Loop = loops_.getLoopFor(MBB);
97  if (!Loop)
98    return;
99  LoopCountMap::iterator li = usingLoops_.find(Loop);
100  assert(li != usingLoops_.end() && "Loop missing");
101  assert(li->second && "0 count in map");
102  if (--li->second)
103    return;
104  // No more blocks in Loop.
105  usingLoops_.erase(li);
106}
107
108// Get three sets of basic blocks surrounding a loop: Blocks inside the loop,
109// predecessor blocks, and exit blocks.
110void SplitAnalysis::getLoopBlocks(const MachineLoop *Loop, LoopBlocks &Blocks) {
111  Blocks.clear();
112
113  // Blocks in the loop.
114  Blocks.Loop.insert(Loop->block_begin(), Loop->block_end());
115
116  // Predecessor blocks.
117  const MachineBasicBlock *Header = Loop->getHeader();
118  for (MachineBasicBlock::const_pred_iterator I = Header->pred_begin(),
119       E = Header->pred_end(); I != E; ++I)
120    if (!Blocks.Loop.count(*I))
121      Blocks.Preds.insert(*I);
122
123  // Exit blocks.
124  for (MachineLoop::block_iterator I = Loop->block_begin(),
125       E = Loop->block_end(); I != E; ++I) {
126    const MachineBasicBlock *MBB = *I;
127    for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(),
128       SE = MBB->succ_end(); SI != SE; ++SI)
129      if (!Blocks.Loop.count(*SI))
130        Blocks.Exits.insert(*SI);
131  }
132}
133
134/// analyzeLoopPeripheralUse - Return an enum describing how curli_ is used in
135/// and around the Loop.
136SplitAnalysis::LoopPeripheralUse SplitAnalysis::
137analyzeLoopPeripheralUse(const SplitAnalysis::LoopBlocks &Blocks) {
138  LoopPeripheralUse use = ContainedInLoop;
139  for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
140       I != E; ++I) {
141    const MachineBasicBlock *MBB = I->first;
142    // Is this a peripheral block?
143    if (use < MultiPeripheral &&
144        (Blocks.Preds.count(MBB) || Blocks.Exits.count(MBB))) {
145      if (I->second > 1) use = MultiPeripheral;
146      else               use = SinglePeripheral;
147      continue;
148    }
149    // Is it a loop block?
150    if (Blocks.Loop.count(MBB))
151      continue;
152    // It must be an unrelated block.
153    return OutsideLoop;
154  }
155  return use;
156}
157
158/// getCriticalExits - It may be necessary to partially break critical edges
159/// leaving the loop if an exit block has phi uses of curli. Collect the exit
160/// blocks that need special treatment into CriticalExits.
161void SplitAnalysis::getCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
162                                     BlockPtrSet &CriticalExits) {
163  CriticalExits.clear();
164
165  // A critical exit block contains a phi def of curli, and has a predecessor
166  // that is not in the loop nor a loop predecessor.
167  // For such an exit block, the edges carrying the new variable must be moved
168  // to a new pre-exit block.
169  for (BlockPtrSet::iterator I = Blocks.Exits.begin(), E = Blocks.Exits.end();
170       I != E; ++I) {
171    const MachineBasicBlock *Succ = *I;
172    SlotIndex SuccIdx = lis_.getMBBStartIdx(Succ);
173    VNInfo *SuccVNI = curli_->getVNInfoAt(SuccIdx);
174    // This exit may not have curli live in at all. No need to split.
175    if (!SuccVNI)
176      continue;
177    // If this is not a PHI def, it is either using a value from before the
178    // loop, or a value defined inside the loop. Both are safe.
179    if (!SuccVNI->isPHIDef() || SuccVNI->def.getBaseIndex() != SuccIdx)
180      continue;
181    // This exit block does have a PHI. Does it also have a predecessor that is
182    // not a loop block or loop predecessor?
183    for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
184         PE = Succ->pred_end(); PI != PE; ++PI) {
185      const MachineBasicBlock *Pred = *PI;
186      if (Blocks.Loop.count(Pred) || Blocks.Preds.count(Pred))
187        continue;
188      // This is a critical exit block, and we need to split the exit edge.
189      CriticalExits.insert(Succ);
190      break;
191    }
192  }
193}
194
195/// canSplitCriticalExits - Return true if it is possible to insert new exit
196/// blocks before the blocks in CriticalExits.
197bool
198SplitAnalysis::canSplitCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
199                                     BlockPtrSet &CriticalExits) {
200  // If we don't allow critical edge splitting, require no critical exits.
201  if (!AllowSplit)
202    return CriticalExits.empty();
203
204  for (BlockPtrSet::iterator I = CriticalExits.begin(), E = CriticalExits.end();
205       I != E; ++I) {
206    const MachineBasicBlock *Succ = *I;
207    // We want to insert a new pre-exit MBB before Succ, and change all the
208    // in-loop blocks to branch to the pre-exit instead of Succ.
209    // Check that all the in-loop predecessors can be changed.
210    for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
211         PE = Succ->pred_end(); PI != PE; ++PI) {
212      const MachineBasicBlock *Pred = *PI;
213      // The external predecessors won't be altered.
214      if (!Blocks.Loop.count(Pred) && !Blocks.Preds.count(Pred))
215        continue;
216      if (!canAnalyzeBranch(Pred))
217        return false;
218    }
219
220    // If Succ's layout predecessor falls through, that too must be analyzable.
221    // We need to insert the pre-exit block in the gap.
222    MachineFunction::const_iterator MFI = Succ;
223    if (MFI == mf_.begin())
224      continue;
225    if (!canAnalyzeBranch(--MFI))
226      return false;
227  }
228  // No problems found.
229  return true;
230}
231
232void SplitAnalysis::analyze(const LiveInterval *li) {
233  clear();
234  curli_ = li;
235  analyzeUses();
236}
237
238const MachineLoop *SplitAnalysis::getBestSplitLoop() {
239  assert(curli_ && "Call analyze() before getBestSplitLoop");
240  if (usingLoops_.empty())
241    return 0;
242
243  LoopPtrSet Loops, SecondLoops;
244  LoopBlocks Blocks;
245  BlockPtrSet CriticalExits;
246
247  // Find first-class and second class candidate loops.
248  // We prefer to split around loops where curli is used outside the periphery.
249  for (LoopCountMap::const_iterator I = usingLoops_.begin(),
250       E = usingLoops_.end(); I != E; ++I) {
251    const MachineLoop *Loop = I->first;
252    getLoopBlocks(Loop, Blocks);
253
254    // FIXME: We need an SSA updater to properly handle multiple exit blocks.
255    if (Blocks.Exits.size() > 1) {
256      DEBUG(dbgs() << "  multiple exits from " << *Loop);
257      continue;
258    }
259
260    LoopPtrSet *LPS = 0;
261    switch(analyzeLoopPeripheralUse(Blocks)) {
262    case OutsideLoop:
263      LPS = &Loops;
264      break;
265    case MultiPeripheral:
266      LPS = &SecondLoops;
267      break;
268    case ContainedInLoop:
269      DEBUG(dbgs() << "  contained in " << *Loop);
270      continue;
271    case SinglePeripheral:
272      DEBUG(dbgs() << "  single peripheral use in " << *Loop);
273      continue;
274    }
275    // Will it be possible to split around this loop?
276    getCriticalExits(Blocks, CriticalExits);
277    DEBUG(dbgs() << "  " << CriticalExits.size() << " critical exits from "
278                 << *Loop);
279    if (!canSplitCriticalExits(Blocks, CriticalExits))
280      continue;
281    // This is a possible split.
282    assert(LPS);
283    LPS->insert(Loop);
284  }
285
286  DEBUG(dbgs() << "  getBestSplitLoop found " << Loops.size() << " + "
287               << SecondLoops.size() << " candidate loops.\n");
288
289  // If there are no first class loops available, look at second class loops.
290  if (Loops.empty())
291    Loops = SecondLoops;
292
293  if (Loops.empty())
294    return 0;
295
296  // Pick the earliest loop.
297  // FIXME: Are there other heuristics to consider?
298  const MachineLoop *Best = 0;
299  SlotIndex BestIdx;
300  for (LoopPtrSet::const_iterator I = Loops.begin(), E = Loops.end(); I != E;
301       ++I) {
302    SlotIndex Idx = lis_.getMBBStartIdx((*I)->getHeader());
303    if (!Best || Idx < BestIdx)
304      Best = *I, BestIdx = Idx;
305  }
306  DEBUG(dbgs() << "  getBestSplitLoop found " << *Best);
307  return Best;
308}
309
310/// getMultiUseBlocks - if curli has more than one use in a basic block, it
311/// may be an advantage to split curli for the duration of the block.
312bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
313  // If curli is local to one block, there is no point to splitting it.
314  if (usingBlocks_.size() <= 1)
315    return false;
316  // Add blocks with multiple uses.
317  for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
318       I != E; ++I)
319    switch (I->second) {
320    case 0:
321    case 1:
322      continue;
323    case 2: {
324      // It doesn't pay to split a 2-instr block if it redefines curli.
325      VNInfo *VN1 = curli_->getVNInfoAt(lis_.getMBBStartIdx(I->first));
326      VNInfo *VN2 =
327        curli_->getVNInfoAt(lis_.getMBBEndIdx(I->first).getPrevIndex());
328      // live-in and live-out with a different value.
329      if (VN1 && VN2 && VN1 != VN2)
330        continue;
331    } // Fall through.
332    default:
333      Blocks.insert(I->first);
334    }
335  return !Blocks.empty();
336}
337
338//===----------------------------------------------------------------------===//
339//                               LiveIntervalMap
340//===----------------------------------------------------------------------===//
341
342// defValue - Introduce a li_ def for ParentVNI that could be later than
343// ParentVNI->def.
344VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) {
345  assert(ParentVNI && "Mapping  NULL value");
346  assert(Idx.isValid() && "Invalid SlotIndex");
347  assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
348
349  // Is this a simple 1-1 mapping? Not likely.
350  if (Idx == ParentVNI->def)
351    return mapValue(ParentVNI, Idx);
352
353  // This is a complex def. Mark with a NULL in valueMap.
354  VNInfo *OldVNI =
355    valueMap_.insert(
356      ValueMap::value_type(ParentVNI, static_cast<VNInfo *>(0))).first->second;
357      // The static_cast<VNInfo *> is only needed to work around a bug in an
358      // old version of the C++0x standard which the following compilers
359      // implemented and have yet to fix:
360      //
361      // Microsoft Visual Studio 2010 Version 10.0.30319.1 RTMRel
362      // Microsoft (R) 32-bit C/C++ Optimizing Compiler Version 16.00.30319.01
363      //
364      // If/When we move to C++0x, this can be replaced by nullptr.
365  (void)OldVNI;
366  assert(OldVNI == 0 && "Simple/Complex values mixed");
367
368  // Should we insert a minimal snippet of VNI LiveRange, or can we count on
369  // callers to do that? We need it for lookups of complex values.
370  VNInfo *VNI = li_.getNextValue(Idx, 0, true, lis_.getVNInfoAllocator());
371  return VNI;
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  assert(ParentVNI && "Mapping  NULL value");
378  assert(Idx.isValid() && "Invalid SlotIndex");
379  assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
380
381  // Use insert for lookup, so we can add missing values with a second lookup.
382  std::pair<ValueMap::iterator,bool> InsP =
383    valueMap_.insert(ValueMap::value_type(ParentVNI, static_cast<VNInfo *>(0)));
384    // The static_cast<VNInfo *> is only needed to work around a bug in an
385    // old version of the C++0x standard which the following compilers
386    // implemented and have yet to fix:
387    //
388    // Microsoft Visual Studio 2010 Version 10.0.30319.1 RTMRel
389    // Microsoft (R) 32-bit C/C++ Optimizing Compiler Version 16.00.30319.01
390    //
391    // If/When we move to C++0x, this can be replaced by nullptr.
392
393  // This was an unknown value. Create a simple mapping.
394  if (InsP.second)
395    return InsP.first->second = li_.createValueCopy(ParentVNI,
396                                                    lis_.getVNInfoAllocator());
397  // This was a simple mapped value.
398  if (InsP.first->second)
399    return InsP.first->second;
400
401  // This is a complex mapped value. There may be multiple defs, and we may need
402  // to create phi-defs.
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 depth-first search for predecessor blocks where we know the
413  // dominating VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
414
415  // Track MBBs where we have created or learned the dominating value.
416  // This may change during the DFS as we create new phi-defs.
417  typedef DenseMap<MachineBasicBlock*, VNInfo*> MBBValueMap;
418  MBBValueMap DomValue;
419
420  for (idf_iterator<MachineBasicBlock*>
421         IDFI = idf_begin(IdxMBB),
422         IDFE = idf_end(IdxMBB); IDFI != IDFE;) {
423    MachineBasicBlock *MBB = *IDFI;
424    SlotIndex End = lis_.getMBBEndIdx(MBB);
425
426    // We are operating on the restricted CFG where ParentVNI is live.
427    if (parentli_.getVNInfoAt(End.getPrevSlot()) != ParentVNI) {
428      IDFI.skipChildren();
429      continue;
430    }
431
432    // Do we have a dominating value in this block?
433    VNInfo *VNI = extendTo(MBB, End);
434    if (!VNI) {
435      ++IDFI;
436      continue;
437    }
438
439    // Yes, VNI dominates MBB. Track the path back to IdxMBB, creating phi-defs
440    // as needed along the way.
441    for (unsigned PI = IDFI.getPathLength()-1; PI != 0; --PI) {
442      // Start from MBB's immediate successor. End at IdxMBB.
443      MachineBasicBlock *Succ = IDFI.getPath(PI-1);
444      std::pair<MBBValueMap::iterator, bool> InsP =
445        DomValue.insert(MBBValueMap::value_type(Succ, VNI));
446
447      // This is the first time we backtrack to Succ.
448      if (InsP.second)
449        continue;
450
451      // We reached Succ again with the same VNI. Nothing is going to change.
452      VNInfo *OVNI = InsP.first->second;
453      if (OVNI == VNI)
454        break;
455
456      // Succ already has a phi-def. No need to continue.
457      SlotIndex Start = lis_.getMBBStartIdx(Succ);
458      if (OVNI->def == Start)
459        break;
460
461      // We have a collision between the old and new VNI at Succ. That means
462      // neither dominates and we need a new phi-def.
463      VNI = li_.getNextValue(Start, 0, true, lis_.getVNInfoAllocator());
464      VNI->setIsPHIDef(true);
465      InsP.first->second = VNI;
466
467      // Replace OVNI with VNI in the remaining path.
468      for (; PI > 1 ; --PI) {
469        MBBValueMap::iterator I = DomValue.find(IDFI.getPath(PI-2));
470        if (I == DomValue.end() || I->second != OVNI)
471          break;
472        I->second = VNI;
473      }
474    }
475
476    // No need to search the children, we found a dominating value.
477    IDFI.skipChildren();
478  }
479
480  // The search should at least find a dominating value for IdxMBB.
481  assert(!DomValue.empty() && "Couldn't find a reaching definition");
482
483  // Since we went through the trouble of a full DFS visiting all reaching defs,
484  // the values in DomValue are now accurate. No more phi-defs are needed for
485  // these blocks, so we can color the live ranges.
486  // This makes the next mapValue call much faster.
487  VNInfo *IdxVNI = 0;
488  for (MBBValueMap::iterator I = DomValue.begin(), E = DomValue.end(); I != E;
489       ++I) {
490     MachineBasicBlock *MBB = I->first;
491     VNInfo *VNI = I->second;
492     SlotIndex Start = lis_.getMBBStartIdx(MBB);
493     if (MBB == IdxMBB) {
494       // Don't add full liveness to IdxMBB, stop at Idx.
495       if (Start != Idx)
496         li_.addRange(LiveRange(Start, Idx, VNI));
497       // The caller had better add some liveness to IdxVNI, or it leaks.
498       IdxVNI = VNI;
499     } else
500      li_.addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
501  }
502
503  assert(IdxVNI && "Didn't find value for Idx");
504  return IdxVNI;
505}
506
507// extendTo - Find the last li_ value defined in MBB at or before Idx. The
508// parentli_ is assumed to be live at Idx. Extend the live range to Idx.
509// Return the found VNInfo, or NULL.
510VNInfo *LiveIntervalMap::extendTo(MachineBasicBlock *MBB, SlotIndex Idx) {
511  LiveInterval::iterator I = std::upper_bound(li_.begin(), li_.end(), Idx);
512  if (I == li_.begin())
513    return 0;
514  --I;
515  if (I->start < lis_.getMBBStartIdx(MBB))
516    return 0;
517  if (I->end < Idx)
518    I->end = Idx;
519  return I->valno;
520}
521
522// addSimpleRange - Add a simple range from parentli_ to li_.
523// ParentVNI must be live in the [Start;End) interval.
524void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End,
525                                     const VNInfo *ParentVNI) {
526  VNInfo *VNI = mapValue(ParentVNI, Start);
527  // A simple mappoing is easy.
528  if (VNI->def == ParentVNI->def) {
529    li_.addRange(LiveRange(Start, End, VNI));
530    return;
531  }
532
533  // ParentVNI is a complex value. We must map per MBB.
534  MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
535  MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End);
536
537  if (MBB == MBBE) {
538    li_.addRange(LiveRange(Start, End, VNI));
539    return;
540  }
541
542  // First block.
543  li_.addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
544
545  // Run sequence of full blocks.
546  for (++MBB; MBB != MBBE; ++MBB) {
547    Start = lis_.getMBBStartIdx(MBB);
548    li_.addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB),
549                           mapValue(ParentVNI, Start)));
550  }
551
552  // Final block.
553  Start = lis_.getMBBStartIdx(MBB);
554  if (Start != End)
555    li_.addRange(LiveRange(Start, End, mapValue(ParentVNI, Start)));
556}
557
558/// addRange - Add live ranges to li_ where [Start;End) intersects parentli_.
559/// All needed values whose def is not inside [Start;End) must be defined
560/// beforehand so mapValue will work.
561void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) {
562  LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end();
563  LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
564
565  // Check if --I begins before Start and overlaps.
566  if (I != B) {
567    --I;
568    if (I->end > Start)
569      addSimpleRange(Start, std::min(End, I->end), I->valno);
570    ++I;
571  }
572
573  // The remaining ranges begin after Start.
574  for (;I != E && I->start < End; ++I)
575    addSimpleRange(I->start, std::min(End, I->end), I->valno);
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                         SmallVectorImpl<LiveInterval*> &intervals)
585  : sa_(sa), lis_(lis), vrm_(vrm),
586    mri_(vrm.getMachineFunction().getRegInfo()),
587    tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
588    curli_(sa_.getCurLI()),
589    dupli_(0), openli_(0),
590    intervals_(intervals),
591    firstInterval(intervals_.size())
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
602LiveInterval *SplitEditor::createInterval() {
603  unsigned curli = sa_.getCurLI()->reg;
604  unsigned Reg = mri_.createVirtualRegister(mri_.getRegClass(curli));
605  LiveInterval &Intv = lis_.getOrCreateInterval(Reg);
606  vrm_.grow();
607  vrm_.assignVirt2StackSlot(Reg, vrm_.getStackSlot(curli));
608  return &Intv;
609}
610
611LiveInterval *SplitEditor::getDupLI() {
612  if (!dupli_) {
613    // Create an interval for dupli that is a copy of curli.
614    dupli_ = createInterval();
615    dupli_->Copy(*curli_, &mri_, lis_.getVNInfoAllocator());
616  }
617  return dupli_;
618}
619
620VNInfo *SplitEditor::mapValue(const VNInfo *curliVNI) {
621  VNInfo *&VNI = valueMap_[curliVNI];
622  if (!VNI)
623    VNI = openli_->createValueCopy(curliVNI, lis_.getVNInfoAllocator());
624  return VNI;
625}
626
627/// Insert a COPY instruction curli -> li. Allocate a new value from li
628/// defined by the COPY. Note that rewrite() will deal with the curli
629/// register, so this function can be used to copy from any interval - openli,
630/// curli, or dupli.
631VNInfo *SplitEditor::insertCopy(LiveInterval &LI,
632                                MachineBasicBlock &MBB,
633                                MachineBasicBlock::iterator I) {
634  MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), tii_.get(TargetOpcode::COPY),
635                             LI.reg).addReg(curli_->reg);
636  SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
637  return LI.getNextValue(DefIdx, MI, true, lis_.getVNInfoAllocator());
638}
639
640/// Create a new virtual register and live interval.
641void SplitEditor::openIntv() {
642  assert(!openli_ && "Previous LI not closed before openIntv");
643  openli_ = createInterval();
644  intervals_.push_back(openli_);
645  liveThrough_ = false;
646}
647
648/// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
649/// not live before Idx, a COPY is not inserted.
650void SplitEditor::enterIntvBefore(SlotIndex Idx) {
651  assert(openli_ && "openIntv not called before enterIntvBefore");
652
653  // Copy from curli_ if it is live.
654  if (VNInfo *CurVNI = curli_->getVNInfoAt(Idx.getUseIndex())) {
655    MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
656    assert(MI && "enterIntvBefore called with invalid index");
657    VNInfo *VNI = insertCopy(*openli_, *MI->getParent(), MI);
658    openli_->addRange(LiveRange(VNI->def, Idx.getDefIndex(), VNI));
659
660    // Make sure CurVNI is properly mapped.
661    VNInfo *&mapVNI = valueMap_[CurVNI];
662    // We dont have SSA update yet, so only one entry per value is allowed.
663    assert(!mapVNI && "enterIntvBefore called more than once for the same value");
664    mapVNI = VNI;
665  }
666  DEBUG(dbgs() << "    enterIntvBefore " << Idx << ": " << *openli_ << '\n');
667}
668
669/// enterIntvAtEnd - Enter openli at the end of MBB.
670/// PhiMBB is a successor inside openli where a PHI value is created.
671/// Currently, all entries must share the same PhiMBB.
672void SplitEditor::enterIntvAtEnd(MachineBasicBlock &A, MachineBasicBlock &B) {
673  assert(openli_ && "openIntv not called before enterIntvAtEnd");
674
675  SlotIndex EndA = lis_.getMBBEndIdx(&A);
676  VNInfo *CurVNIA = curli_->getVNInfoAt(EndA.getPrevIndex());
677  if (!CurVNIA) {
678    DEBUG(dbgs() << "    enterIntvAtEnd, curli not live out of BB#"
679                 << A.getNumber() << ".\n");
680    return;
681  }
682
683  // Add a phi kill value and live range out of A.
684  VNInfo *VNIA = insertCopy(*openli_, A, A.getFirstTerminator());
685  openli_->addRange(LiveRange(VNIA->def, EndA, VNIA));
686
687  // FIXME: If this is the only entry edge, we don't need the extra PHI value.
688  // FIXME: If there are multiple entry blocks (so not a loop), we need proper
689  // SSA update.
690
691  // Now look at the start of B.
692  SlotIndex StartB = lis_.getMBBStartIdx(&B);
693  SlotIndex EndB = lis_.getMBBEndIdx(&B);
694  const LiveRange *CurB = curli_->getLiveRangeContaining(StartB);
695  if (!CurB) {
696    DEBUG(dbgs() << "    enterIntvAtEnd: curli not live in to BB#"
697                 << B.getNumber() << ".\n");
698    return;
699  }
700
701  VNInfo *VNIB = openli_->getVNInfoAt(StartB);
702  if (!VNIB) {
703    // Create a phi value.
704    VNIB = openli_->getNextValue(SlotIndex(StartB, true), 0, false,
705                                 lis_.getVNInfoAllocator());
706    VNIB->setIsPHIDef(true);
707    VNInfo *&mapVNI = valueMap_[CurB->valno];
708    if (mapVNI) {
709      // Multiple copies - must create PHI value.
710      abort();
711    } else {
712      // This is the first copy of dupLR. Mark the mapping.
713      mapVNI = VNIB;
714    }
715
716  }
717
718  DEBUG(dbgs() << "    enterIntvAtEnd: " << *openli_ << '\n');
719}
720
721/// useIntv - indicate that all instructions in MBB should use openli.
722void SplitEditor::useIntv(const MachineBasicBlock &MBB) {
723  useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB));
724}
725
726void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) {
727  assert(openli_ && "openIntv not called before useIntv");
728
729  // Map the curli values from the interval into openli_
730  LiveInterval::const_iterator B = curli_->begin(), E = curli_->end();
731  LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
732
733  if (I != B) {
734    --I;
735    // I begins before Start, but overlaps.
736    if (I->end > Start)
737      openli_->addRange(LiveRange(Start, std::min(End, I->end),
738                        mapValue(I->valno)));
739    ++I;
740  }
741
742  // The remaining ranges begin after Start.
743  for (;I != E && I->start < End; ++I)
744    openli_->addRange(LiveRange(I->start, std::min(End, I->end),
745                                mapValue(I->valno)));
746  DEBUG(dbgs() << "    use [" << Start << ';' << End << "): " << *openli_
747               << '\n');
748}
749
750/// leaveIntvAfter - Leave openli after the instruction at Idx.
751void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
752  assert(openli_ && "openIntv not called before leaveIntvAfter");
753
754  const LiveRange *CurLR = curli_->getLiveRangeContaining(Idx.getDefIndex());
755  if (!CurLR || CurLR->end <= Idx.getBoundaryIndex()) {
756    DEBUG(dbgs() << "    leaveIntvAfter " << Idx << ": not live\n");
757    return;
758  }
759
760  // Was this value of curli live through openli?
761  if (!openli_->liveAt(CurLR->valno->def)) {
762    DEBUG(dbgs() << "    leaveIntvAfter " << Idx << ": using external value\n");
763    liveThrough_ = true;
764    return;
765  }
766
767  // We are going to insert a back copy, so we must have a dupli_.
768  LiveRange *DupLR = getDupLI()->getLiveRangeContaining(Idx.getDefIndex());
769  assert(DupLR && "dupli not live into black, but curli is?");
770
771  // Insert the COPY instruction.
772  MachineBasicBlock::iterator I = lis_.getInstructionFromIndex(Idx);
773  MachineInstr *MI = BuildMI(*I->getParent(), llvm::next(I), I->getDebugLoc(),
774                             tii_.get(TargetOpcode::COPY), dupli_->reg)
775                       .addReg(openli_->reg);
776  SlotIndex CopyIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
777  openli_->addRange(LiveRange(Idx.getDefIndex(), CopyIdx,
778                    mapValue(CurLR->valno)));
779  DupLR->valno->def = CopyIdx;
780  DEBUG(dbgs() << "    leaveIntvAfter " << Idx << ": " << *openli_ << '\n');
781}
782
783/// leaveIntvAtTop - Leave the interval at the top of MBB.
784/// Currently, only one value can leave the interval.
785void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
786  assert(openli_ && "openIntv not called before leaveIntvAtTop");
787
788  SlotIndex Start = lis_.getMBBStartIdx(&MBB);
789  const LiveRange *CurLR = curli_->getLiveRangeContaining(Start);
790
791  // Is curli even live-in to MBB?
792  if (!CurLR) {
793    DEBUG(dbgs() << "    leaveIntvAtTop at " << Start << ": not live\n");
794    return;
795  }
796
797  // Is curli defined by PHI at the beginning of MBB?
798  bool isPHIDef = CurLR->valno->isPHIDef() &&
799                  CurLR->valno->def.getBaseIndex() == Start;
800
801  // If MBB is using a value of curli that was defined outside the openli range,
802  // we don't want to copy it back here.
803  if (!isPHIDef && !openli_->liveAt(CurLR->valno->def)) {
804    DEBUG(dbgs() << "    leaveIntvAtTop at " << Start
805                 << ": using external value\n");
806    liveThrough_ = true;
807    return;
808  }
809
810  // We are going to insert a back copy, so we must have a dupli_.
811  LiveRange *DupLR = getDupLI()->getLiveRangeContaining(Start);
812  assert(DupLR && "dupli not live into black, but curli is?");
813
814  // Insert the COPY instruction.
815  MachineInstr *MI = BuildMI(MBB, MBB.begin(), DebugLoc(),
816                             tii_.get(TargetOpcode::COPY), dupli_->reg)
817                       .addReg(openli_->reg);
818  SlotIndex Idx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
819
820  // Adjust dupli and openli values.
821  if (isPHIDef) {
822    // dupli was already a PHI on entry to MBB. Simply insert an openli PHI,
823    // and shift the dupli def down to the COPY.
824    VNInfo *VNI = openli_->getNextValue(SlotIndex(Start, true), 0, false,
825                                        lis_.getVNInfoAllocator());
826    VNI->setIsPHIDef(true);
827    openli_->addRange(LiveRange(VNI->def, Idx, VNI));
828
829    dupli_->removeRange(Start, Idx);
830    DupLR->valno->def = Idx;
831    DupLR->valno->setIsPHIDef(false);
832  } else {
833    // The dupli value was defined somewhere inside the openli range.
834    DEBUG(dbgs() << "    leaveIntvAtTop source value defined at "
835                 << DupLR->valno->def << "\n");
836    // FIXME: We may not need a PHI here if all predecessors have the same
837    // value.
838    VNInfo *VNI = openli_->getNextValue(SlotIndex(Start, true), 0, false,
839                                        lis_.getVNInfoAllocator());
840    VNI->setIsPHIDef(true);
841    openli_->addRange(LiveRange(VNI->def, Idx, VNI));
842
843    // FIXME: What if DupLR->valno is used by multiple exits? SSA Update.
844
845    // closeIntv is going to remove the superfluous live ranges.
846    DupLR->valno->def = Idx;
847    DupLR->valno->setIsPHIDef(false);
848  }
849
850  DEBUG(dbgs() << "    leaveIntvAtTop at " << Idx << ": " << *openli_ << '\n');
851}
852
853/// closeIntv - Indicate that we are done editing the currently open
854/// LiveInterval, and ranges can be trimmed.
855void SplitEditor::closeIntv() {
856  assert(openli_ && "openIntv not called before closeIntv");
857
858  DEBUG(dbgs() << "    closeIntv cleaning up\n");
859  DEBUG(dbgs() << "    open " << *openli_ << '\n');
860
861  if (liveThrough_) {
862    DEBUG(dbgs() << "    value live through region, leaving dupli as is.\n");
863  } else {
864    // live out with copies inserted, or killed by region. Either way we need to
865    // remove the overlapping region from dupli.
866    getDupLI();
867    for (LiveInterval::iterator I = openli_->begin(), E = openli_->end();
868         I != E; ++I) {
869      dupli_->removeRange(I->start, I->end);
870    }
871    // FIXME: A block branching to the entry block may also branch elsewhere
872    // curli is live. We need both openli and curli to be live in that case.
873    DEBUG(dbgs() << "    dup2 " << *dupli_ << '\n');
874  }
875  openli_ = 0;
876  valueMap_.clear();
877}
878
879/// rewrite - after all the new live ranges have been created, rewrite
880/// instructions using curli to use the new intervals.
881void SplitEditor::rewrite() {
882  assert(!openli_ && "Previous LI not closed before rewrite");
883  const LiveInterval *curli = sa_.getCurLI();
884  for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(curli->reg),
885       RE = mri_.reg_end(); RI != RE;) {
886    MachineOperand &MO = RI.getOperand();
887    MachineInstr *MI = MO.getParent();
888    ++RI;
889    if (MI->isDebugValue()) {
890      DEBUG(dbgs() << "Zapping " << *MI);
891      // FIXME: We can do much better with debug values.
892      MO.setReg(0);
893      continue;
894    }
895    SlotIndex Idx = lis_.getInstructionIndex(MI);
896    Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
897    LiveInterval *LI = dupli_;
898    for (unsigned i = firstInterval, e = intervals_.size(); i != e; ++i) {
899      LiveInterval *testli = intervals_[i];
900      if (testli->liveAt(Idx)) {
901        LI = testli;
902        break;
903      }
904    }
905    if (LI) {
906      MO.setReg(LI->reg);
907      sa_.removeUse(MI);
908      DEBUG(dbgs() << "  rewrite " << Idx << '\t' << *MI);
909    }
910  }
911
912  // dupli_ goes in last, after rewriting.
913  if (dupli_) {
914    if (dupli_->empty()) {
915      DEBUG(dbgs() << "  dupli became empty?\n");
916      lis_.removeInterval(dupli_->reg);
917      dupli_ = 0;
918    } else {
919      dupli_->RenumberValues(lis_);
920      intervals_.push_back(dupli_);
921    }
922  }
923
924  // Calculate spill weight and allocation hints for new intervals.
925  VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
926  for (unsigned i = firstInterval, e = intervals_.size(); i != e; ++i) {
927    LiveInterval &li = *intervals_[i];
928    vrai.CalculateRegClass(li.reg);
929    vrai.CalculateWeightAndHint(li);
930    DEBUG(dbgs() << "  new interval " << mri_.getRegClass(li.reg)->getName()
931                 << ":" << li << '\n');
932  }
933}
934
935
936//===----------------------------------------------------------------------===//
937//                               Loop Splitting
938//===----------------------------------------------------------------------===//
939
940bool SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
941  SplitAnalysis::LoopBlocks Blocks;
942  sa_.getLoopBlocks(Loop, Blocks);
943
944  // Break critical edges as needed.
945  SplitAnalysis::BlockPtrSet CriticalExits;
946  sa_.getCriticalExits(Blocks, CriticalExits);
947  assert(CriticalExits.empty() && "Cannot break critical exits yet");
948
949  // Create new live interval for the loop.
950  openIntv();
951
952  // Insert copies in the predecessors.
953  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
954       E = Blocks.Preds.end(); I != E; ++I) {
955    MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
956    enterIntvAtEnd(MBB, *Loop->getHeader());
957  }
958
959  // Switch all loop blocks.
960  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
961       E = Blocks.Loop.end(); I != E; ++I)
962     useIntv(**I);
963
964  // Insert back copies in the exit blocks.
965  for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
966       E = Blocks.Exits.end(); I != E; ++I) {
967    MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
968    leaveIntvAtTop(MBB);
969  }
970
971  // Done.
972  closeIntv();
973  rewrite();
974  return dupli_;
975}
976
977
978//===----------------------------------------------------------------------===//
979//                            Single Block Splitting
980//===----------------------------------------------------------------------===//
981
982/// splitSingleBlocks - Split curli into a separate live interval inside each
983/// basic block in Blocks. Return true if curli has been completely replaced,
984/// false if curli is still intact, and needs to be spilled or split further.
985bool SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
986  DEBUG(dbgs() << "  splitSingleBlocks for " << Blocks.size() << " blocks.\n");
987  // Determine the first and last instruction using curli in each block.
988  typedef std::pair<SlotIndex,SlotIndex> IndexPair;
989  typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
990  IndexPairMap MBBRange;
991  for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
992       E = sa_.usingInstrs_.end(); I != E; ++I) {
993    const MachineBasicBlock *MBB = (*I)->getParent();
994    if (!Blocks.count(MBB))
995      continue;
996    SlotIndex Idx = lis_.getInstructionIndex(*I);
997    DEBUG(dbgs() << "  BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
998    IndexPair &IP = MBBRange[MBB];
999    if (!IP.first.isValid() || Idx < IP.first)
1000      IP.first = Idx;
1001    if (!IP.second.isValid() || Idx > IP.second)
1002      IP.second = Idx;
1003  }
1004
1005  // Create a new interval for each block.
1006  for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
1007       E = Blocks.end(); I != E; ++I) {
1008    IndexPair &IP = MBBRange[*I];
1009    DEBUG(dbgs() << "  splitting for BB#" << (*I)->getNumber() << ": ["
1010                 << IP.first << ';' << IP.second << ")\n");
1011    assert(IP.first.isValid() && IP.second.isValid());
1012
1013    openIntv();
1014    enterIntvBefore(IP.first);
1015    useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
1016    leaveIntvAfter(IP.second);
1017    closeIntv();
1018  }
1019  rewrite();
1020  return dupli_;
1021}
1022
1023
1024//===----------------------------------------------------------------------===//
1025//                            Sub Block Splitting
1026//===----------------------------------------------------------------------===//
1027
1028/// getBlockForInsideSplit - If curli is contained inside a single basic block,
1029/// and it wou pay to subdivide the interval inside that block, return it.
1030/// Otherwise return NULL. The returned block can be passed to
1031/// SplitEditor::splitInsideBlock.
1032const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
1033  // The interval must be exclusive to one block.
1034  if (usingBlocks_.size() != 1)
1035    return 0;
1036  // Don't to this for less than 4 instructions. We want to be sure that
1037  // splitting actually reduces the instruction count per interval.
1038  if (usingInstrs_.size() < 4)
1039    return 0;
1040  return usingBlocks_.begin()->first;
1041}
1042
1043/// splitInsideBlock - Split curli into multiple intervals inside MBB. Return
1044/// true if curli has been completely replaced, false if curli is still
1045/// intact, and needs to be spilled or split further.
1046bool SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
1047  SmallVector<SlotIndex, 32> Uses;
1048  Uses.reserve(sa_.usingInstrs_.size());
1049  for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1050       E = sa_.usingInstrs_.end(); I != E; ++I)
1051    if ((*I)->getParent() == MBB)
1052      Uses.push_back(lis_.getInstructionIndex(*I));
1053  DEBUG(dbgs() << "  splitInsideBlock BB#" << MBB->getNumber() << " for "
1054               << Uses.size() << " instructions.\n");
1055  assert(Uses.size() >= 3 && "Need at least 3 instructions");
1056  array_pod_sort(Uses.begin(), Uses.end());
1057
1058  // Simple algorithm: Find the largest gap between uses as determined by slot
1059  // indices. Create new intervals for instructions before the gap and after the
1060  // gap.
1061  unsigned bestPos = 0;
1062  int bestGap = 0;
1063  DEBUG(dbgs() << "    dist (" << Uses[0]);
1064  for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
1065    int g = Uses[i-1].distance(Uses[i]);
1066    DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
1067    if (g > bestGap)
1068      bestPos = i, bestGap = g;
1069  }
1070  DEBUG(dbgs() << "), best: -" << bestGap << "-\n");
1071
1072  // bestPos points to the first use after the best gap.
1073  assert(bestPos > 0 && "Invalid gap");
1074
1075  // FIXME: Don't create intervals for low densities.
1076
1077  // First interval before the gap. Don't create single-instr intervals.
1078  if (bestPos > 1) {
1079    openIntv();
1080    enterIntvBefore(Uses.front());
1081    useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
1082    leaveIntvAfter(Uses[bestPos-1]);
1083    closeIntv();
1084  }
1085
1086  // Second interval after the gap.
1087  if (bestPos < Uses.size()-1) {
1088    openIntv();
1089    enterIntvBefore(Uses[bestPos]);
1090    useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
1091    leaveIntvAfter(Uses.back());
1092    closeIntv();
1093  }
1094
1095  rewrite();
1096  return dupli_;
1097}
1098