TailDuplication.cpp revision 2c189061184925c6a8ecbb5a19e648b230a41c0e 
1//===-- TailDuplication.cpp - Duplicate blocks into predecessors' tails ---===//
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 pass duplicates basic blocks ending in unconditional branches into
11// the tails of their predecessors.
12//
13//===----------------------------------------------------------------------===//
14
15#define DEBUG_TYPE "tailduplication"
16#include "llvm/Function.h"
17#include "llvm/CodeGen/Passes.h"
18#include "llvm/CodeGen/MachineModuleInfo.h"
19#include "llvm/CodeGen/MachineFunctionPass.h"
20#include "llvm/CodeGen/MachineInstrBuilder.h"
21#include "llvm/CodeGen/MachineRegisterInfo.h"
22#include "llvm/CodeGen/MachineSSAUpdater.h"
23#include "llvm/CodeGen/RegisterScavenging.h"
24#include "llvm/Target/TargetInstrInfo.h"
25#include "llvm/Target/TargetRegisterInfo.h"
26#include "llvm/Support/CommandLine.h"
27#include "llvm/Support/Debug.h"
28#include "llvm/Support/ErrorHandling.h"
29#include "llvm/Support/raw_ostream.h"
30#include "llvm/ADT/DenseSet.h"
31#include "llvm/ADT/OwningPtr.h"
32#include "llvm/ADT/SmallSet.h"
33#include "llvm/ADT/SetVector.h"
34#include "llvm/ADT/Statistic.h"
35using namespace llvm;
36
37STATISTIC(NumTails     , "Number of tails duplicated");
38STATISTIC(NumTailDups  , "Number of tail duplicated blocks");
39STATISTIC(NumInstrDups , "Additional instructions due to tail duplication");
40STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
41STATISTIC(NumAddedPHIs , "Number of phis added");
42
43// Heuristic for tail duplication.
44static cl::opt<unsigned>
45TailDuplicateSize("tail-dup-size",
46                  cl::desc("Maximum instructions to consider tail duplicating"),
47                  cl::init(2), cl::Hidden);
48
49static cl::opt<bool>
50TailDupVerify("tail-dup-verify",
51              cl::desc("Verify sanity of PHI instructions during taildup"),
52              cl::init(false), cl::Hidden);
53
54static cl::opt<unsigned>
55TailDupLimit("tail-dup-limit", cl::init(~0U), cl::Hidden);
56
57typedef std::vector<std::pair<MachineBasicBlock*,unsigned> > AvailableValsTy;
58
59namespace {
60  /// TailDuplicatePass - Perform tail duplication.
61  class TailDuplicatePass : public MachineFunctionPass {
62    const TargetInstrInfo *TII;
63    const TargetRegisterInfo *TRI;
64    MachineModuleInfo *MMI;
65    MachineRegisterInfo *MRI;
66    OwningPtr<RegScavenger> RS;
67    bool PreRegAlloc;
68
69    // SSAUpdateVRs - A list of virtual registers for which to update SSA form.
70    SmallVector<unsigned, 16> SSAUpdateVRs;
71
72    // SSAUpdateVals - For each virtual register in SSAUpdateVals keep a list of
73    // source virtual registers.
74    DenseMap<unsigned, AvailableValsTy> SSAUpdateVals;
75
76  public:
77    static char ID;
78    explicit TailDuplicatePass() :
79      MachineFunctionPass(ID), PreRegAlloc(false) {}
80
81    virtual bool runOnMachineFunction(MachineFunction &MF);
82
83  private:
84    void AddSSAUpdateEntry(unsigned OrigReg, unsigned NewReg,
85                           MachineBasicBlock *BB);
86    void ProcessPHI(MachineInstr *MI, MachineBasicBlock *TailBB,
87                    MachineBasicBlock *PredBB,
88                    DenseMap<unsigned, unsigned> &LocalVRMap,
89                    SmallVector<std::pair<unsigned,unsigned>, 4> &Copies,
90                    const DenseSet<unsigned> &UsedByPhi,
91                    bool Remove);
92    void DuplicateInstruction(MachineInstr *MI,
93                              MachineBasicBlock *TailBB,
94                              MachineBasicBlock *PredBB,
95                              MachineFunction &MF,
96                              DenseMap<unsigned, unsigned> &LocalVRMap,
97                              const DenseSet<unsigned> &UsedByPhi);
98    void UpdateSuccessorsPHIs(MachineBasicBlock *FromBB, bool isDead,
99                              SmallVector<MachineBasicBlock*, 8> &TDBBs,
100                              SmallSetVector<MachineBasicBlock*, 8> &Succs);
101    bool TailDuplicateBlocks(MachineFunction &MF);
102    bool shouldTailDuplicate(const MachineFunction &MF,
103                             bool IsSimple, MachineBasicBlock &TailBB);
104    bool isSimpleBB(MachineBasicBlock *TailBB);
105    bool canCompletelyDuplicateBB(MachineBasicBlock &BB);
106    bool duplicateSimpleBB(MachineBasicBlock *TailBB,
107                           SmallVector<MachineBasicBlock*, 8> &TDBBs,
108                           const DenseSet<unsigned> &RegsUsedByPhi,
109                           SmallVector<MachineInstr*, 16> &Copies);
110    bool TailDuplicate(MachineBasicBlock *TailBB,
111                       bool IsSimple,
112                       MachineFunction &MF,
113                       SmallVector<MachineBasicBlock*, 8> &TDBBs,
114                       SmallVector<MachineInstr*, 16> &Copies);
115    bool TailDuplicateAndUpdate(MachineBasicBlock *MBB,
116                                bool IsSimple,
117                                MachineFunction &MF);
118
119    void RemoveDeadBlock(MachineBasicBlock *MBB);
120  };
121
122  char TailDuplicatePass::ID = 0;
123}
124
125char &llvm::TailDuplicateID = TailDuplicatePass::ID;
126
127INITIALIZE_PASS(TailDuplicatePass, "tailduplication", "Tail Duplication",
128                false, false)
129
130bool TailDuplicatePass::runOnMachineFunction(MachineFunction &MF) {
131  TII = MF.getTarget().getInstrInfo();
132  TRI = MF.getTarget().getRegisterInfo();
133  MRI = &MF.getRegInfo();
134  MMI = getAnalysisIfAvailable<MachineModuleInfo>();
135  PreRegAlloc = MRI->isSSA();
136  RS.reset();
137  if (MRI->tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
138    RS.reset(new RegScavenger());
139
140  bool MadeChange = false;
141  while (TailDuplicateBlocks(MF))
142    MadeChange = true;
143
144  return MadeChange;
145}
146
147static void VerifyPHIs(MachineFunction &MF, bool CheckExtra) {
148  for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ++I) {
149    MachineBasicBlock *MBB = I;
150    SmallSetVector<MachineBasicBlock*, 8> Preds(MBB->pred_begin(),
151                                                MBB->pred_end());
152    MachineBasicBlock::iterator MI = MBB->begin();
153    while (MI != MBB->end()) {
154      if (!MI->isPHI())
155        break;
156      for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
157             PE = Preds.end(); PI != PE; ++PI) {
158        MachineBasicBlock *PredBB = *PI;
159        bool Found = false;
160        for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2) {
161          MachineBasicBlock *PHIBB = MI->getOperand(i+1).getMBB();
162          if (PHIBB == PredBB) {
163            Found = true;
164            break;
165          }
166        }
167        if (!Found) {
168          dbgs() << "Malformed PHI in BB#" << MBB->getNumber() << ": " << *MI;
169          dbgs() << "  missing input from predecessor BB#"
170                 << PredBB->getNumber() << '\n';
171          llvm_unreachable(0);
172        }
173      }
174
175      for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2) {
176        MachineBasicBlock *PHIBB = MI->getOperand(i+1).getMBB();
177        if (CheckExtra && !Preds.count(PHIBB)) {
178          dbgs() << "Warning: malformed PHI in BB#" << MBB->getNumber()
179                 << ": " << *MI;
180          dbgs() << "  extra input from predecessor BB#"
181                 << PHIBB->getNumber() << '\n';
182          llvm_unreachable(0);
183        }
184        if (PHIBB->getNumber() < 0) {
185          dbgs() << "Malformed PHI in BB#" << MBB->getNumber() << ": " << *MI;
186          dbgs() << "  non-existing BB#" << PHIBB->getNumber() << '\n';
187          llvm_unreachable(0);
188        }
189      }
190      ++MI;
191    }
192  }
193}
194
195/// TailDuplicateAndUpdate - Tail duplicate the block and cleanup.
196bool
197TailDuplicatePass::TailDuplicateAndUpdate(MachineBasicBlock *MBB,
198                                          bool IsSimple,
199                                          MachineFunction &MF) {
200  // Save the successors list.
201  SmallSetVector<MachineBasicBlock*, 8> Succs(MBB->succ_begin(),
202                                              MBB->succ_end());
203
204  SmallVector<MachineBasicBlock*, 8> TDBBs;
205  SmallVector<MachineInstr*, 16> Copies;
206  if (!TailDuplicate(MBB, IsSimple, MF, TDBBs, Copies))
207    return false;
208
209  ++NumTails;
210
211  SmallVector<MachineInstr*, 8> NewPHIs;
212  MachineSSAUpdater SSAUpdate(MF, &NewPHIs);
213
214  // TailBB's immediate successors are now successors of those predecessors
215  // which duplicated TailBB. Add the predecessors as sources to the PHI
216  // instructions.
217  bool isDead = MBB->pred_empty() && !MBB->hasAddressTaken();
218  if (PreRegAlloc)
219    UpdateSuccessorsPHIs(MBB, isDead, TDBBs, Succs);
220
221  // If it is dead, remove it.
222  if (isDead) {
223    NumInstrDups -= MBB->size();
224    RemoveDeadBlock(MBB);
225    ++NumDeadBlocks;
226  }
227
228  // Update SSA form.
229  if (!SSAUpdateVRs.empty()) {
230    for (unsigned i = 0, e = SSAUpdateVRs.size(); i != e; ++i) {
231      unsigned VReg = SSAUpdateVRs[i];
232      SSAUpdate.Initialize(VReg);
233
234      // If the original definition is still around, add it as an available
235      // value.
236      MachineInstr *DefMI = MRI->getVRegDef(VReg);
237      MachineBasicBlock *DefBB = 0;
238      if (DefMI) {
239        DefBB = DefMI->getParent();
240        SSAUpdate.AddAvailableValue(DefBB, VReg);
241      }
242
243      // Add the new vregs as available values.
244      DenseMap<unsigned, AvailableValsTy>::iterator LI =
245        SSAUpdateVals.find(VReg);
246      for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) {
247        MachineBasicBlock *SrcBB = LI->second[j].first;
248        unsigned SrcReg = LI->second[j].second;
249        SSAUpdate.AddAvailableValue(SrcBB, SrcReg);
250      }
251
252      // Rewrite uses that are outside of the original def's block.
253      MachineRegisterInfo::use_iterator UI = MRI->use_begin(VReg);
254      while (UI != MRI->use_end()) {
255        MachineOperand &UseMO = UI.getOperand();
256        MachineInstr *UseMI = &*UI;
257        ++UI;
258        if (UseMI->isDebugValue()) {
259          // SSAUpdate can replace the use with an undef. That creates
260          // a debug instruction that is a kill.
261          // FIXME: Should it SSAUpdate job to delete debug instructions
262          // instead of replacing the use with undef?
263          UseMI->eraseFromParent();
264          continue;
265        }
266        if (UseMI->getParent() == DefBB && !UseMI->isPHI())
267          continue;
268        SSAUpdate.RewriteUse(UseMO);
269      }
270    }
271
272    SSAUpdateVRs.clear();
273    SSAUpdateVals.clear();
274  }
275
276  // Eliminate some of the copies inserted by tail duplication to maintain
277  // SSA form.
278  for (unsigned i = 0, e = Copies.size(); i != e; ++i) {
279    MachineInstr *Copy = Copies[i];
280    if (!Copy->isCopy())
281      continue;
282    unsigned Dst = Copy->getOperand(0).getReg();
283    unsigned Src = Copy->getOperand(1).getReg();
284    if (MRI->hasOneNonDBGUse(Src) &&
285        MRI->constrainRegClass(Src, MRI->getRegClass(Dst))) {
286      // Copy is the only use. Do trivial copy propagation here.
287      MRI->replaceRegWith(Dst, Src);
288      Copy->eraseFromParent();
289    }
290  }
291
292  if (NewPHIs.size())
293    NumAddedPHIs += NewPHIs.size();
294
295  return true;
296}
297
298/// TailDuplicateBlocks - Look for small blocks that are unconditionally
299/// branched to and do not fall through. Tail-duplicate their instructions
300/// into their predecessors to eliminate (dynamic) branches.
301bool TailDuplicatePass::TailDuplicateBlocks(MachineFunction &MF) {
302  bool MadeChange = false;
303
304  if (PreRegAlloc && TailDupVerify) {
305    DEBUG(dbgs() << "\n*** Before tail-duplicating\n");
306    VerifyPHIs(MF, true);
307  }
308
309  for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
310    MachineBasicBlock *MBB = I++;
311
312    if (NumTails == TailDupLimit)
313      break;
314
315    bool IsSimple = isSimpleBB(MBB);
316
317    if (!shouldTailDuplicate(MF, IsSimple, *MBB))
318      continue;
319
320    MadeChange |= TailDuplicateAndUpdate(MBB, IsSimple, MF);
321  }
322
323  if (PreRegAlloc && TailDupVerify)
324    VerifyPHIs(MF, false);
325
326  return MadeChange;
327}
328
329static bool isDefLiveOut(unsigned Reg, MachineBasicBlock *BB,
330                         const MachineRegisterInfo *MRI) {
331  for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
332         UE = MRI->use_end(); UI != UE; ++UI) {
333    MachineInstr *UseMI = &*UI;
334    if (UseMI->isDebugValue())
335      continue;
336    if (UseMI->getParent() != BB)
337      return true;
338  }
339  return false;
340}
341
342static unsigned getPHISrcRegOpIdx(MachineInstr *MI, MachineBasicBlock *SrcBB) {
343  for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2)
344    if (MI->getOperand(i+1).getMBB() == SrcBB)
345      return i;
346  return 0;
347}
348
349
350// Remember which registers are used by phis in this block. This is
351// used to determine which registers are liveout while modifying the
352// block (which is why we need to copy the information).
353static void getRegsUsedByPHIs(const MachineBasicBlock &BB,
354                              DenseSet<unsigned> *UsedByPhi) {
355  for(MachineBasicBlock::const_iterator I = BB.begin(), E = BB.end();
356      I != E; ++I) {
357    const MachineInstr &MI = *I;
358    if (!MI.isPHI())
359      break;
360    for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) {
361      unsigned SrcReg = MI.getOperand(i).getReg();
362      UsedByPhi->insert(SrcReg);
363    }
364  }
365}
366
367/// AddSSAUpdateEntry - Add a definition and source virtual registers pair for
368/// SSA update.
369void TailDuplicatePass::AddSSAUpdateEntry(unsigned OrigReg, unsigned NewReg,
370                                          MachineBasicBlock *BB) {
371  DenseMap<unsigned, AvailableValsTy>::iterator LI= SSAUpdateVals.find(OrigReg);
372  if (LI != SSAUpdateVals.end())
373    LI->second.push_back(std::make_pair(BB, NewReg));
374  else {
375    AvailableValsTy Vals;
376    Vals.push_back(std::make_pair(BB, NewReg));
377    SSAUpdateVals.insert(std::make_pair(OrigReg, Vals));
378    SSAUpdateVRs.push_back(OrigReg);
379  }
380}
381
382/// ProcessPHI - Process PHI node in TailBB by turning it into a copy in PredBB.
383/// Remember the source register that's contributed by PredBB and update SSA
384/// update map.
385void TailDuplicatePass::ProcessPHI(MachineInstr *MI,
386                                   MachineBasicBlock *TailBB,
387                                   MachineBasicBlock *PredBB,
388                                   DenseMap<unsigned, unsigned> &LocalVRMap,
389                           SmallVector<std::pair<unsigned,unsigned>, 4> &Copies,
390                                   const DenseSet<unsigned> &RegsUsedByPhi,
391                                   bool Remove) {
392  unsigned DefReg = MI->getOperand(0).getReg();
393  unsigned SrcOpIdx = getPHISrcRegOpIdx(MI, PredBB);
394  assert(SrcOpIdx && "Unable to find matching PHI source?");
395  unsigned SrcReg = MI->getOperand(SrcOpIdx).getReg();
396  const TargetRegisterClass *RC = MRI->getRegClass(DefReg);
397  LocalVRMap.insert(std::make_pair(DefReg, SrcReg));
398
399  // Insert a copy from source to the end of the block. The def register is the
400  // available value liveout of the block.
401  unsigned NewDef = MRI->createVirtualRegister(RC);
402  Copies.push_back(std::make_pair(NewDef, SrcReg));
403  if (isDefLiveOut(DefReg, TailBB, MRI) || RegsUsedByPhi.count(DefReg))
404    AddSSAUpdateEntry(DefReg, NewDef, PredBB);
405
406  if (!Remove)
407    return;
408
409  // Remove PredBB from the PHI node.
410  MI->RemoveOperand(SrcOpIdx+1);
411  MI->RemoveOperand(SrcOpIdx);
412  if (MI->getNumOperands() == 1)
413    MI->eraseFromParent();
414}
415
416/// DuplicateInstruction - Duplicate a TailBB instruction to PredBB and update
417/// the source operands due to earlier PHI translation.
418void TailDuplicatePass::DuplicateInstruction(MachineInstr *MI,
419                                     MachineBasicBlock *TailBB,
420                                     MachineBasicBlock *PredBB,
421                                     MachineFunction &MF,
422                                     DenseMap<unsigned, unsigned> &LocalVRMap,
423                                     const DenseSet<unsigned> &UsedByPhi) {
424  MachineInstr *NewMI = TII->duplicate(MI, MF);
425  for (unsigned i = 0, e = NewMI->getNumOperands(); i != e; ++i) {
426    MachineOperand &MO = NewMI->getOperand(i);
427    if (!MO.isReg())
428      continue;
429    unsigned Reg = MO.getReg();
430    if (!TargetRegisterInfo::isVirtualRegister(Reg))
431      continue;
432    if (MO.isDef()) {
433      const TargetRegisterClass *RC = MRI->getRegClass(Reg);
434      unsigned NewReg = MRI->createVirtualRegister(RC);
435      MO.setReg(NewReg);
436      LocalVRMap.insert(std::make_pair(Reg, NewReg));
437      if (isDefLiveOut(Reg, TailBB, MRI) || UsedByPhi.count(Reg))
438        AddSSAUpdateEntry(Reg, NewReg, PredBB);
439    } else {
440      DenseMap<unsigned, unsigned>::iterator VI = LocalVRMap.find(Reg);
441      if (VI != LocalVRMap.end()) {
442        MO.setReg(VI->second);
443        MRI->constrainRegClass(VI->second, MRI->getRegClass(Reg));
444      }
445    }
446  }
447  PredBB->insert(PredBB->instr_end(), NewMI);
448}
449
450/// UpdateSuccessorsPHIs - After FromBB is tail duplicated into its predecessor
451/// blocks, the successors have gained new predecessors. Update the PHI
452/// instructions in them accordingly.
453void
454TailDuplicatePass::UpdateSuccessorsPHIs(MachineBasicBlock *FromBB, bool isDead,
455                                  SmallVector<MachineBasicBlock*, 8> &TDBBs,
456                                  SmallSetVector<MachineBasicBlock*,8> &Succs) {
457  for (SmallSetVector<MachineBasicBlock*, 8>::iterator SI = Succs.begin(),
458         SE = Succs.end(); SI != SE; ++SI) {
459    MachineBasicBlock *SuccBB = *SI;
460    for (MachineBasicBlock::iterator II = SuccBB->begin(), EE = SuccBB->end();
461         II != EE; ++II) {
462      if (!II->isPHI())
463        break;
464      unsigned Idx = 0;
465      for (unsigned i = 1, e = II->getNumOperands(); i != e; i += 2) {
466        MachineOperand &MO = II->getOperand(i+1);
467        if (MO.getMBB() == FromBB) {
468          Idx = i;
469          break;
470        }
471      }
472
473      assert(Idx != 0);
474      MachineOperand &MO0 = II->getOperand(Idx);
475      unsigned Reg = MO0.getReg();
476      if (isDead) {
477        // Folded into the previous BB.
478        // There could be duplicate phi source entries. FIXME: Should sdisel
479        // or earlier pass fixed this?
480        for (unsigned i = II->getNumOperands()-2; i != Idx; i -= 2) {
481          MachineOperand &MO = II->getOperand(i+1);
482          if (MO.getMBB() == FromBB) {
483            II->RemoveOperand(i+1);
484            II->RemoveOperand(i);
485          }
486        }
487      } else
488        Idx = 0;
489
490      // If Idx is set, the operands at Idx and Idx+1 must be removed.
491      // We reuse the location to avoid expensive RemoveOperand calls.
492
493      DenseMap<unsigned,AvailableValsTy>::iterator LI=SSAUpdateVals.find(Reg);
494      if (LI != SSAUpdateVals.end()) {
495        // This register is defined in the tail block.
496        for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) {
497          MachineBasicBlock *SrcBB = LI->second[j].first;
498          // If we didn't duplicate a bb into a particular predecessor, we
499          // might still have added an entry to SSAUpdateVals to correcly
500          // recompute SSA. If that case, avoid adding a dummy extra argument
501          // this PHI.
502          if (!SrcBB->isSuccessor(SuccBB))
503            continue;
504
505          unsigned SrcReg = LI->second[j].second;
506          if (Idx != 0) {
507            II->getOperand(Idx).setReg(SrcReg);
508            II->getOperand(Idx+1).setMBB(SrcBB);
509            Idx = 0;
510          } else {
511            II->addOperand(MachineOperand::CreateReg(SrcReg, false));
512            II->addOperand(MachineOperand::CreateMBB(SrcBB));
513          }
514        }
515      } else {
516        // Live in tail block, must also be live in predecessors.
517        for (unsigned j = 0, ee = TDBBs.size(); j != ee; ++j) {
518          MachineBasicBlock *SrcBB = TDBBs[j];
519          if (Idx != 0) {
520            II->getOperand(Idx).setReg(Reg);
521            II->getOperand(Idx+1).setMBB(SrcBB);
522            Idx = 0;
523          } else {
524            II->addOperand(MachineOperand::CreateReg(Reg, false));
525            II->addOperand(MachineOperand::CreateMBB(SrcBB));
526          }
527        }
528      }
529      if (Idx != 0) {
530        II->RemoveOperand(Idx+1);
531        II->RemoveOperand(Idx);
532      }
533    }
534  }
535}
536
537/// shouldTailDuplicate - Determine if it is profitable to duplicate this block.
538bool
539TailDuplicatePass::shouldTailDuplicate(const MachineFunction &MF,
540                                       bool IsSimple,
541                                       MachineBasicBlock &TailBB) {
542  // Only duplicate blocks that end with unconditional branches.
543  if (TailBB.canFallThrough())
544    return false;
545
546  // Don't try to tail-duplicate single-block loops.
547  if (TailBB.isSuccessor(&TailBB))
548    return false;
549
550  // Set the limit on the cost to duplicate. When optimizing for size,
551  // duplicate only one, because one branch instruction can be eliminated to
552  // compensate for the duplication.
553  unsigned MaxDuplicateCount;
554  if (TailDuplicateSize.getNumOccurrences() == 0 &&
555      MF.getFunction()->getFnAttributes().hasOptimizeForSizeAttr())
556    MaxDuplicateCount = 1;
557  else
558    MaxDuplicateCount = TailDuplicateSize;
559
560  // If the target has hardware branch prediction that can handle indirect
561  // branches, duplicating them can often make them predictable when there
562  // are common paths through the code.  The limit needs to be high enough
563  // to allow undoing the effects of tail merging and other optimizations
564  // that rearrange the predecessors of the indirect branch.
565
566  bool HasIndirectbr = false;
567  if (!TailBB.empty())
568    HasIndirectbr = TailBB.back().isIndirectBranch();
569
570  if (HasIndirectbr && PreRegAlloc)
571    MaxDuplicateCount = 20;
572
573  // Check the instructions in the block to determine whether tail-duplication
574  // is invalid or unlikely to be profitable.
575  unsigned InstrCount = 0;
576  for (MachineBasicBlock::iterator I = TailBB.begin(); I != TailBB.end(); ++I) {
577    // Non-duplicable things shouldn't be tail-duplicated.
578    if (I->isNotDuplicable())
579      return false;
580
581    // Do not duplicate 'return' instructions if this is a pre-regalloc run.
582    // A return may expand into a lot more instructions (e.g. reload of callee
583    // saved registers) after PEI.
584    if (PreRegAlloc && I->isReturn())
585      return false;
586
587    // Avoid duplicating calls before register allocation. Calls presents a
588    // barrier to register allocation so duplicating them may end up increasing
589    // spills.
590    if (PreRegAlloc && I->isCall())
591      return false;
592
593    if (!I->isPHI() && !I->isDebugValue())
594      InstrCount += 1;
595
596    if (InstrCount > MaxDuplicateCount)
597      return false;
598  }
599
600  if (HasIndirectbr && PreRegAlloc)
601    return true;
602
603  if (IsSimple)
604    return true;
605
606  if (!PreRegAlloc)
607    return true;
608
609  return canCompletelyDuplicateBB(TailBB);
610}
611
612/// isSimpleBB - True if this BB has only one unconditional jump.
613bool
614TailDuplicatePass::isSimpleBB(MachineBasicBlock *TailBB) {
615  if (TailBB->succ_size() != 1)
616    return false;
617  if (TailBB->pred_empty())
618    return false;
619  MachineBasicBlock::iterator I = TailBB->begin();
620  MachineBasicBlock::iterator E = TailBB->end();
621  while (I != E && I->isDebugValue())
622    ++I;
623  if (I == E)
624    return true;
625  return I->isUnconditionalBranch();
626}
627
628static bool
629bothUsedInPHI(const MachineBasicBlock &A,
630              SmallPtrSet<MachineBasicBlock*, 8> SuccsB) {
631  for (MachineBasicBlock::const_succ_iterator SI = A.succ_begin(),
632         SE = A.succ_end(); SI != SE; ++SI) {
633    MachineBasicBlock *BB = *SI;
634    if (SuccsB.count(BB) && !BB->empty() && BB->begin()->isPHI())
635      return true;
636  }
637
638  return false;
639}
640
641bool
642TailDuplicatePass::canCompletelyDuplicateBB(MachineBasicBlock &BB) {
643  SmallPtrSet<MachineBasicBlock*, 8> Succs(BB.succ_begin(), BB.succ_end());
644
645  for (MachineBasicBlock::pred_iterator PI = BB.pred_begin(),
646       PE = BB.pred_end(); PI != PE; ++PI) {
647    MachineBasicBlock *PredBB = *PI;
648
649    if (PredBB->succ_size() > 1)
650      return false;
651
652    MachineBasicBlock *PredTBB = NULL, *PredFBB = NULL;
653    SmallVector<MachineOperand, 4> PredCond;
654    if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
655      return false;
656
657    if (!PredCond.empty())
658      return false;
659  }
660  return true;
661}
662
663bool
664TailDuplicatePass::duplicateSimpleBB(MachineBasicBlock *TailBB,
665                                     SmallVector<MachineBasicBlock*, 8> &TDBBs,
666                                     const DenseSet<unsigned> &UsedByPhi,
667                                     SmallVector<MachineInstr*, 16> &Copies) {
668  SmallPtrSet<MachineBasicBlock*, 8> Succs(TailBB->succ_begin(),
669                                           TailBB->succ_end());
670  SmallVector<MachineBasicBlock*, 8> Preds(TailBB->pred_begin(),
671                                           TailBB->pred_end());
672  bool Changed = false;
673  for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
674       PE = Preds.end(); PI != PE; ++PI) {
675    MachineBasicBlock *PredBB = *PI;
676
677    if (PredBB->getLandingPadSuccessor())
678      continue;
679
680    if (bothUsedInPHI(*PredBB, Succs))
681      continue;
682
683    MachineBasicBlock *PredTBB = NULL, *PredFBB = NULL;
684    SmallVector<MachineOperand, 4> PredCond;
685    if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
686      continue;
687
688    Changed = true;
689    DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB
690                 << "From simple Succ: " << *TailBB);
691
692    MachineBasicBlock *NewTarget = *TailBB->succ_begin();
693    MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(PredBB));
694
695    // Make PredFBB explicit.
696    if (PredCond.empty())
697      PredFBB = PredTBB;
698
699    // Make fall through explicit.
700    if (!PredTBB)
701      PredTBB = NextBB;
702    if (!PredFBB)
703      PredFBB = NextBB;
704
705    // Redirect
706    if (PredFBB == TailBB)
707      PredFBB = NewTarget;
708    if (PredTBB == TailBB)
709      PredTBB = NewTarget;
710
711    // Make the branch unconditional if possible
712    if (PredTBB == PredFBB) {
713      PredCond.clear();
714      PredFBB = NULL;
715    }
716
717    // Avoid adding fall through branches.
718    if (PredFBB == NextBB)
719      PredFBB = NULL;
720    if (PredTBB == NextBB && PredFBB == NULL)
721      PredTBB = NULL;
722
723    TII->RemoveBranch(*PredBB);
724
725    if (PredTBB)
726      TII->InsertBranch(*PredBB, PredTBB, PredFBB, PredCond, DebugLoc());
727
728    PredBB->removeSuccessor(TailBB);
729    unsigned NumSuccessors = PredBB->succ_size();
730    assert(NumSuccessors <= 1);
731    if (NumSuccessors == 0 || *PredBB->succ_begin() != NewTarget)
732      PredBB->addSuccessor(NewTarget);
733
734    TDBBs.push_back(PredBB);
735  }
736  return Changed;
737}
738
739/// TailDuplicate - If it is profitable, duplicate TailBB's contents in each
740/// of its predecessors.
741bool
742TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB,
743                                 bool IsSimple,
744                                 MachineFunction &MF,
745                                 SmallVector<MachineBasicBlock*, 8> &TDBBs,
746                                 SmallVector<MachineInstr*, 16> &Copies) {
747  DEBUG(dbgs() << "\n*** Tail-duplicating BB#" << TailBB->getNumber() << '\n');
748
749  DenseSet<unsigned> UsedByPhi;
750  getRegsUsedByPHIs(*TailBB, &UsedByPhi);
751
752  if (IsSimple)
753    return duplicateSimpleBB(TailBB, TDBBs, UsedByPhi, Copies);
754
755  // Iterate through all the unique predecessors and tail-duplicate this
756  // block into them, if possible. Copying the list ahead of time also
757  // avoids trouble with the predecessor list reallocating.
758  bool Changed = false;
759  SmallSetVector<MachineBasicBlock*, 8> Preds(TailBB->pred_begin(),
760                                              TailBB->pred_end());
761  for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
762       PE = Preds.end(); PI != PE; ++PI) {
763    MachineBasicBlock *PredBB = *PI;
764
765    assert(TailBB != PredBB &&
766           "Single-block loop should have been rejected earlier!");
767    // EH edges are ignored by AnalyzeBranch.
768    if (PredBB->succ_size() > 1)
769      continue;
770
771    MachineBasicBlock *PredTBB, *PredFBB;
772    SmallVector<MachineOperand, 4> PredCond;
773    if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
774      continue;
775    if (!PredCond.empty())
776      continue;
777    // Don't duplicate into a fall-through predecessor (at least for now).
778    if (PredBB->isLayoutSuccessor(TailBB) && PredBB->canFallThrough())
779      continue;
780
781    DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB
782                 << "From Succ: " << *TailBB);
783
784    TDBBs.push_back(PredBB);
785
786    // Remove PredBB's unconditional branch.
787    TII->RemoveBranch(*PredBB);
788
789    if (RS && !TailBB->livein_empty()) {
790      // Update PredBB livein.
791      RS->enterBasicBlock(PredBB);
792      if (!PredBB->empty())
793        RS->forward(prior(PredBB->end()));
794      BitVector RegsLiveAtExit(TRI->getNumRegs());
795      RS->getRegsUsed(RegsLiveAtExit, false);
796      for (MachineBasicBlock::livein_iterator I = TailBB->livein_begin(),
797             E = TailBB->livein_end(); I != E; ++I) {
798        if (!RegsLiveAtExit[*I])
799          // If a register is previously livein to the tail but it's not live
800          // at the end of predecessor BB, then it should be added to its
801          // livein list.
802          PredBB->addLiveIn(*I);
803      }
804    }
805
806    // Clone the contents of TailBB into PredBB.
807    DenseMap<unsigned, unsigned> LocalVRMap;
808    SmallVector<std::pair<unsigned,unsigned>, 4> CopyInfos;
809    // Use instr_iterator here to properly handle bundles, e.g.
810    // ARM Thumb2 IT block.
811    MachineBasicBlock::instr_iterator I = TailBB->instr_begin();
812    while (I != TailBB->instr_end()) {
813      MachineInstr *MI = &*I;
814      ++I;
815      if (MI->isPHI()) {
816        // Replace the uses of the def of the PHI with the register coming
817        // from PredBB.
818        ProcessPHI(MI, TailBB, PredBB, LocalVRMap, CopyInfos, UsedByPhi, true);
819      } else {
820        // Replace def of virtual registers with new registers, and update
821        // uses with PHI source register or the new registers.
822        DuplicateInstruction(MI, TailBB, PredBB, MF, LocalVRMap, UsedByPhi);
823      }
824    }
825    MachineBasicBlock::iterator Loc = PredBB->getFirstTerminator();
826    for (unsigned i = 0, e = CopyInfos.size(); i != e; ++i) {
827      Copies.push_back(BuildMI(*PredBB, Loc, DebugLoc(),
828                               TII->get(TargetOpcode::COPY),
829                               CopyInfos[i].first).addReg(CopyInfos[i].second));
830    }
831
832    // Simplify
833    TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true);
834
835    NumInstrDups += TailBB->size() - 1; // subtract one for removed branch
836
837    // Update the CFG.
838    PredBB->removeSuccessor(PredBB->succ_begin());
839    assert(PredBB->succ_empty() &&
840           "TailDuplicate called on block with multiple successors!");
841    for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
842           E = TailBB->succ_end(); I != E; ++I)
843      PredBB->addSuccessor(*I);
844
845    Changed = true;
846    ++NumTailDups;
847  }
848
849  // If TailBB was duplicated into all its predecessors except for the prior
850  // block, which falls through unconditionally, move the contents of this
851  // block into the prior block.
852  MachineBasicBlock *PrevBB = prior(MachineFunction::iterator(TailBB));
853  MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
854  SmallVector<MachineOperand, 4> PriorCond;
855  // This has to check PrevBB->succ_size() because EH edges are ignored by
856  // AnalyzeBranch.
857  if (PrevBB->succ_size() == 1 &&
858      !TII->AnalyzeBranch(*PrevBB, PriorTBB, PriorFBB, PriorCond, true) &&
859      PriorCond.empty() && !PriorTBB && TailBB->pred_size() == 1 &&
860      !TailBB->hasAddressTaken()) {
861    DEBUG(dbgs() << "\nMerging into block: " << *PrevBB
862          << "From MBB: " << *TailBB);
863    if (PreRegAlloc) {
864      DenseMap<unsigned, unsigned> LocalVRMap;
865      SmallVector<std::pair<unsigned,unsigned>, 4> CopyInfos;
866      MachineBasicBlock::iterator I = TailBB->begin();
867      // Process PHI instructions first.
868      while (I != TailBB->end() && I->isPHI()) {
869        // Replace the uses of the def of the PHI with the register coming
870        // from PredBB.
871        MachineInstr *MI = &*I++;
872        ProcessPHI(MI, TailBB, PrevBB, LocalVRMap, CopyInfos, UsedByPhi, true);
873        if (MI->getParent())
874          MI->eraseFromParent();
875      }
876
877      // Now copy the non-PHI instructions.
878      while (I != TailBB->end()) {
879        // Replace def of virtual registers with new registers, and update
880        // uses with PHI source register or the new registers.
881        MachineInstr *MI = &*I++;
882        assert(!MI->isBundle() && "Not expecting bundles before regalloc!");
883        DuplicateInstruction(MI, TailBB, PrevBB, MF, LocalVRMap, UsedByPhi);
884        MI->eraseFromParent();
885      }
886      MachineBasicBlock::iterator Loc = PrevBB->getFirstTerminator();
887      for (unsigned i = 0, e = CopyInfos.size(); i != e; ++i) {
888        Copies.push_back(BuildMI(*PrevBB, Loc, DebugLoc(),
889                                 TII->get(TargetOpcode::COPY),
890                                 CopyInfos[i].first)
891                           .addReg(CopyInfos[i].second));
892      }
893    } else {
894      // No PHIs to worry about, just splice the instructions over.
895      PrevBB->splice(PrevBB->end(), TailBB, TailBB->begin(), TailBB->end());
896    }
897    PrevBB->removeSuccessor(PrevBB->succ_begin());
898    assert(PrevBB->succ_empty());
899    PrevBB->transferSuccessors(TailBB);
900    TDBBs.push_back(PrevBB);
901    Changed = true;
902  }
903
904  // If this is after register allocation, there are no phis to fix.
905  if (!PreRegAlloc)
906    return Changed;
907
908  // If we made no changes so far, we are safe.
909  if (!Changed)
910    return Changed;
911
912
913  // Handle the nasty case in that we duplicated a block that is part of a loop
914  // into some but not all of its predecessors. For example:
915  //    1 -> 2 <-> 3                 |
916  //          \                      |
917  //           \---> rest            |
918  // if we duplicate 2 into 1 but not into 3, we end up with
919  // 12 -> 3 <-> 2 -> rest           |
920  //   \             /               |
921  //    \----->-----/                |
922  // If there was a "var = phi(1, 3)" in 2, it has to be ultimately replaced
923  // with a phi in 3 (which now dominates 2).
924  // What we do here is introduce a copy in 3 of the register defined by the
925  // phi, just like when we are duplicating 2 into 3, but we don't copy any
926  // real instructions or remove the 3 -> 2 edge from the phi in 2.
927  for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
928       PE = Preds.end(); PI != PE; ++PI) {
929    MachineBasicBlock *PredBB = *PI;
930    if (std::find(TDBBs.begin(), TDBBs.end(), PredBB) != TDBBs.end())
931      continue;
932
933    // EH edges
934    if (PredBB->succ_size() != 1)
935      continue;
936
937    DenseMap<unsigned, unsigned> LocalVRMap;
938    SmallVector<std::pair<unsigned,unsigned>, 4> CopyInfos;
939    MachineBasicBlock::iterator I = TailBB->begin();
940    // Process PHI instructions first.
941    while (I != TailBB->end() && I->isPHI()) {
942      // Replace the uses of the def of the PHI with the register coming
943      // from PredBB.
944      MachineInstr *MI = &*I++;
945      ProcessPHI(MI, TailBB, PredBB, LocalVRMap, CopyInfos, UsedByPhi, false);
946    }
947    MachineBasicBlock::iterator Loc = PredBB->getFirstTerminator();
948    for (unsigned i = 0, e = CopyInfos.size(); i != e; ++i) {
949      Copies.push_back(BuildMI(*PredBB, Loc, DebugLoc(),
950                               TII->get(TargetOpcode::COPY),
951                               CopyInfos[i].first).addReg(CopyInfos[i].second));
952    }
953  }
954
955  return Changed;
956}
957
958/// RemoveDeadBlock - Remove the specified dead machine basic block from the
959/// function, updating the CFG.
960void TailDuplicatePass::RemoveDeadBlock(MachineBasicBlock *MBB) {
961  assert(MBB->pred_empty() && "MBB must be dead!");
962  DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
963
964  // Remove all successors.
965  while (!MBB->succ_empty())
966    MBB->removeSuccessor(MBB->succ_end()-1);
967
968  // Remove the block.
969  MBB->eraseFromParent();
970}
971