1 //===-- TailDuplication.cpp - Duplicate blocks into predecessors' tails ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass duplicates basic blocks ending in unconditional branches into
11 // the tails of their predecessors.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/CodeGen/Passes.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
21 #include "llvm/CodeGen/MachineFunctionPass.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/MachineSSAUpdater.h"
26 #include "llvm/CodeGen/RegisterScavenging.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Target/TargetInstrInfo.h"
33 #include "llvm/Target/TargetRegisterInfo.h"
34 #include "llvm/Target/TargetSubtargetInfo.h"
37 #define DEBUG_TYPE "tailduplication"
39 STATISTIC(NumTails , "Number of tails duplicated");
40 STATISTIC(NumTailDups , "Number of tail duplicated blocks");
41 STATISTIC(NumInstrDups , "Additional instructions due to tail duplication");
42 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
43 STATISTIC(NumAddedPHIs , "Number of phis added");
45 // Heuristic for tail duplication.
46 static cl::opt<unsigned>
47 TailDuplicateSize("tail-dup-size",
48 cl::desc("Maximum instructions to consider tail duplicating"),
49 cl::init(2), cl::Hidden);
52 TailDupVerify("tail-dup-verify",
53 cl::desc("Verify sanity of PHI instructions during taildup"),
54 cl::init(false), cl::Hidden);
56 static cl::opt<unsigned>
57 TailDupLimit("tail-dup-limit", cl::init(~0U), cl::Hidden);
59 typedef std::vector<std::pair<MachineBasicBlock*,unsigned> > AvailableValsTy;
62 /// Perform tail duplication.
63 class TailDuplicatePass : public MachineFunctionPass {
64 const TargetInstrInfo *TII;
65 const TargetRegisterInfo *TRI;
66 const MachineBranchProbabilityInfo *MBPI;
67 MachineModuleInfo *MMI;
68 MachineRegisterInfo *MRI;
69 std::unique_ptr<RegScavenger> RS;
72 // A list of virtual registers for which to update SSA form.
73 SmallVector<unsigned, 16> SSAUpdateVRs;
75 // For each virtual register in SSAUpdateVals keep a list of source virtual
77 DenseMap<unsigned, AvailableValsTy> SSAUpdateVals;
81 explicit TailDuplicatePass() :
82 MachineFunctionPass(ID), PreRegAlloc(false) {}
84 bool runOnMachineFunction(MachineFunction &MF) override;
86 void getAnalysisUsage(AnalysisUsage &AU) const override;
89 void AddSSAUpdateEntry(unsigned OrigReg, unsigned NewReg,
90 MachineBasicBlock *BB);
91 void ProcessPHI(MachineInstr *MI, MachineBasicBlock *TailBB,
92 MachineBasicBlock *PredBB,
93 DenseMap<unsigned, unsigned> &LocalVRMap,
94 SmallVectorImpl<std::pair<unsigned,unsigned> > &Copies,
95 const DenseSet<unsigned> &UsedByPhi,
97 void DuplicateInstruction(MachineInstr *MI,
98 MachineBasicBlock *TailBB,
99 MachineBasicBlock *PredBB,
101 DenseMap<unsigned, unsigned> &LocalVRMap,
102 const DenseSet<unsigned> &UsedByPhi);
103 void UpdateSuccessorsPHIs(MachineBasicBlock *FromBB, bool isDead,
104 SmallVectorImpl<MachineBasicBlock *> &TDBBs,
105 SmallSetVector<MachineBasicBlock*, 8> &Succs);
106 bool TailDuplicateBlocks(MachineFunction &MF);
107 bool shouldTailDuplicate(const MachineFunction &MF,
108 bool IsSimple, MachineBasicBlock &TailBB);
109 bool isSimpleBB(MachineBasicBlock *TailBB);
110 bool canCompletelyDuplicateBB(MachineBasicBlock &BB);
111 bool duplicateSimpleBB(MachineBasicBlock *TailBB,
112 SmallVectorImpl<MachineBasicBlock *> &TDBBs,
113 const DenseSet<unsigned> &RegsUsedByPhi,
114 SmallVectorImpl<MachineInstr *> &Copies);
115 bool TailDuplicate(MachineBasicBlock *TailBB,
118 SmallVectorImpl<MachineBasicBlock *> &TDBBs,
119 SmallVectorImpl<MachineInstr *> &Copies);
120 bool TailDuplicateAndUpdate(MachineBasicBlock *MBB,
122 MachineFunction &MF);
124 void RemoveDeadBlock(MachineBasicBlock *MBB);
127 char TailDuplicatePass::ID = 0;
130 char &llvm::TailDuplicateID = TailDuplicatePass::ID;
132 INITIALIZE_PASS(TailDuplicatePass, "tailduplication", "Tail Duplication",
135 bool TailDuplicatePass::runOnMachineFunction(MachineFunction &MF) {
136 if (skipOptnoneFunction(*MF.getFunction()))
139 TII = MF.getSubtarget().getInstrInfo();
140 TRI = MF.getSubtarget().getRegisterInfo();
141 MRI = &MF.getRegInfo();
142 MMI = getAnalysisIfAvailable<MachineModuleInfo>();
143 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
145 PreRegAlloc = MRI->isSSA();
147 if (MRI->tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
148 RS.reset(new RegScavenger());
150 bool MadeChange = false;
151 while (TailDuplicateBlocks(MF))
157 void TailDuplicatePass::getAnalysisUsage(AnalysisUsage &AU) const {
158 AU.addRequired<MachineBranchProbabilityInfo>();
159 MachineFunctionPass::getAnalysisUsage(AU);
162 static void VerifyPHIs(MachineFunction &MF, bool CheckExtra) {
163 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ++I) {
164 MachineBasicBlock *MBB = I;
165 SmallSetVector<MachineBasicBlock*, 8> Preds(MBB->pred_begin(),
167 MachineBasicBlock::iterator MI = MBB->begin();
168 while (MI != MBB->end()) {
171 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
172 PE = Preds.end(); PI != PE; ++PI) {
173 MachineBasicBlock *PredBB = *PI;
175 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2) {
176 MachineBasicBlock *PHIBB = MI->getOperand(i+1).getMBB();
177 if (PHIBB == PredBB) {
183 dbgs() << "Malformed PHI in BB#" << MBB->getNumber() << ": " << *MI;
184 dbgs() << " missing input from predecessor BB#"
185 << PredBB->getNumber() << '\n';
186 llvm_unreachable(nullptr);
190 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2) {
191 MachineBasicBlock *PHIBB = MI->getOperand(i+1).getMBB();
192 if (CheckExtra && !Preds.count(PHIBB)) {
193 dbgs() << "Warning: malformed PHI in BB#" << MBB->getNumber()
195 dbgs() << " extra input from predecessor BB#"
196 << PHIBB->getNumber() << '\n';
197 llvm_unreachable(nullptr);
199 if (PHIBB->getNumber() < 0) {
200 dbgs() << "Malformed PHI in BB#" << MBB->getNumber() << ": " << *MI;
201 dbgs() << " non-existing BB#" << PHIBB->getNumber() << '\n';
202 llvm_unreachable(nullptr);
210 /// Tail duplicate the block and cleanup.
212 TailDuplicatePass::TailDuplicateAndUpdate(MachineBasicBlock *MBB,
214 MachineFunction &MF) {
215 // Save the successors list.
216 SmallSetVector<MachineBasicBlock*, 8> Succs(MBB->succ_begin(),
219 SmallVector<MachineBasicBlock*, 8> TDBBs;
220 SmallVector<MachineInstr*, 16> Copies;
221 if (!TailDuplicate(MBB, IsSimple, MF, TDBBs, Copies))
226 SmallVector<MachineInstr*, 8> NewPHIs;
227 MachineSSAUpdater SSAUpdate(MF, &NewPHIs);
229 // TailBB's immediate successors are now successors of those predecessors
230 // which duplicated TailBB. Add the predecessors as sources to the PHI
232 bool isDead = MBB->pred_empty() && !MBB->hasAddressTaken();
234 UpdateSuccessorsPHIs(MBB, isDead, TDBBs, Succs);
236 // If it is dead, remove it.
238 NumInstrDups -= MBB->size();
239 RemoveDeadBlock(MBB);
244 if (!SSAUpdateVRs.empty()) {
245 for (unsigned i = 0, e = SSAUpdateVRs.size(); i != e; ++i) {
246 unsigned VReg = SSAUpdateVRs[i];
247 SSAUpdate.Initialize(VReg);
249 // If the original definition is still around, add it as an available
251 MachineInstr *DefMI = MRI->getVRegDef(VReg);
252 MachineBasicBlock *DefBB = nullptr;
254 DefBB = DefMI->getParent();
255 SSAUpdate.AddAvailableValue(DefBB, VReg);
258 // Add the new vregs as available values.
259 DenseMap<unsigned, AvailableValsTy>::iterator LI =
260 SSAUpdateVals.find(VReg);
261 for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) {
262 MachineBasicBlock *SrcBB = LI->second[j].first;
263 unsigned SrcReg = LI->second[j].second;
264 SSAUpdate.AddAvailableValue(SrcBB, SrcReg);
267 // Rewrite uses that are outside of the original def's block.
268 MachineRegisterInfo::use_iterator UI = MRI->use_begin(VReg);
269 while (UI != MRI->use_end()) {
270 MachineOperand &UseMO = *UI;
271 MachineInstr *UseMI = UseMO.getParent();
273 if (UseMI->isDebugValue()) {
274 // SSAUpdate can replace the use with an undef. That creates
275 // a debug instruction that is a kill.
276 // FIXME: Should it SSAUpdate job to delete debug instructions
277 // instead of replacing the use with undef?
278 UseMI->eraseFromParent();
281 if (UseMI->getParent() == DefBB && !UseMI->isPHI())
283 SSAUpdate.RewriteUse(UseMO);
287 SSAUpdateVRs.clear();
288 SSAUpdateVals.clear();
291 // Eliminate some of the copies inserted by tail duplication to maintain
293 for (unsigned i = 0, e = Copies.size(); i != e; ++i) {
294 MachineInstr *Copy = Copies[i];
297 unsigned Dst = Copy->getOperand(0).getReg();
298 unsigned Src = Copy->getOperand(1).getReg();
299 if (MRI->hasOneNonDBGUse(Src) &&
300 MRI->constrainRegClass(Src, MRI->getRegClass(Dst))) {
301 // Copy is the only use. Do trivial copy propagation here.
302 MRI->replaceRegWith(Dst, Src);
303 Copy->eraseFromParent();
308 NumAddedPHIs += NewPHIs.size();
313 /// Look for small blocks that are unconditionally branched to and do not fall
314 /// through. Tail-duplicate their instructions into their predecessors to
315 /// eliminate (dynamic) branches.
316 bool TailDuplicatePass::TailDuplicateBlocks(MachineFunction &MF) {
317 bool MadeChange = false;
319 if (PreRegAlloc && TailDupVerify) {
320 DEBUG(dbgs() << "\n*** Before tail-duplicating\n");
321 VerifyPHIs(MF, true);
324 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
325 MachineBasicBlock *MBB = I++;
327 if (NumTails == TailDupLimit)
330 bool IsSimple = isSimpleBB(MBB);
332 if (!shouldTailDuplicate(MF, IsSimple, *MBB))
335 MadeChange |= TailDuplicateAndUpdate(MBB, IsSimple, MF);
338 if (PreRegAlloc && TailDupVerify)
339 VerifyPHIs(MF, false);
344 static bool isDefLiveOut(unsigned Reg, MachineBasicBlock *BB,
345 const MachineRegisterInfo *MRI) {
346 for (MachineInstr &UseMI : MRI->use_instructions(Reg)) {
347 if (UseMI.isDebugValue())
349 if (UseMI.getParent() != BB)
355 static unsigned getPHISrcRegOpIdx(MachineInstr *MI, MachineBasicBlock *SrcBB) {
356 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2)
357 if (MI->getOperand(i+1).getMBB() == SrcBB)
363 // Remember which registers are used by phis in this block. This is
364 // used to determine which registers are liveout while modifying the
365 // block (which is why we need to copy the information).
366 static void getRegsUsedByPHIs(const MachineBasicBlock &BB,
367 DenseSet<unsigned> *UsedByPhi) {
368 for (const auto &MI : BB) {
371 for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) {
372 unsigned SrcReg = MI.getOperand(i).getReg();
373 UsedByPhi->insert(SrcReg);
378 /// Add a definition and source virtual registers pair for SSA update.
379 void TailDuplicatePass::AddSSAUpdateEntry(unsigned OrigReg, unsigned NewReg,
380 MachineBasicBlock *BB) {
381 DenseMap<unsigned, AvailableValsTy>::iterator LI= SSAUpdateVals.find(OrigReg);
382 if (LI != SSAUpdateVals.end())
383 LI->second.push_back(std::make_pair(BB, NewReg));
385 AvailableValsTy Vals;
386 Vals.push_back(std::make_pair(BB, NewReg));
387 SSAUpdateVals.insert(std::make_pair(OrigReg, Vals));
388 SSAUpdateVRs.push_back(OrigReg);
392 /// Process PHI node in TailBB by turning it into a copy in PredBB. Remember the
393 /// source register that's contributed by PredBB and update SSA update map.
394 void TailDuplicatePass::ProcessPHI(
395 MachineInstr *MI, MachineBasicBlock *TailBB, MachineBasicBlock *PredBB,
396 DenseMap<unsigned, unsigned> &LocalVRMap,
397 SmallVectorImpl<std::pair<unsigned, unsigned> > &Copies,
398 const DenseSet<unsigned> &RegsUsedByPhi, bool Remove) {
399 unsigned DefReg = MI->getOperand(0).getReg();
400 unsigned SrcOpIdx = getPHISrcRegOpIdx(MI, PredBB);
401 assert(SrcOpIdx && "Unable to find matching PHI source?");
402 unsigned SrcReg = MI->getOperand(SrcOpIdx).getReg();
403 const TargetRegisterClass *RC = MRI->getRegClass(DefReg);
404 LocalVRMap.insert(std::make_pair(DefReg, SrcReg));
406 // Insert a copy from source to the end of the block. The def register is the
407 // available value liveout of the block.
408 unsigned NewDef = MRI->createVirtualRegister(RC);
409 Copies.push_back(std::make_pair(NewDef, SrcReg));
410 if (isDefLiveOut(DefReg, TailBB, MRI) || RegsUsedByPhi.count(DefReg))
411 AddSSAUpdateEntry(DefReg, NewDef, PredBB);
416 // Remove PredBB from the PHI node.
417 MI->RemoveOperand(SrcOpIdx+1);
418 MI->RemoveOperand(SrcOpIdx);
419 if (MI->getNumOperands() == 1)
420 MI->eraseFromParent();
423 /// Duplicate a TailBB instruction to PredBB and update
424 /// the source operands due to earlier PHI translation.
425 void TailDuplicatePass::DuplicateInstruction(MachineInstr *MI,
426 MachineBasicBlock *TailBB,
427 MachineBasicBlock *PredBB,
429 DenseMap<unsigned, unsigned> &LocalVRMap,
430 const DenseSet<unsigned> &UsedByPhi) {
431 MachineInstr *NewMI = TII->duplicate(MI, MF);
432 for (unsigned i = 0, e = NewMI->getNumOperands(); i != e; ++i) {
433 MachineOperand &MO = NewMI->getOperand(i);
436 unsigned Reg = MO.getReg();
437 if (!TargetRegisterInfo::isVirtualRegister(Reg))
440 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
441 unsigned NewReg = MRI->createVirtualRegister(RC);
443 LocalVRMap.insert(std::make_pair(Reg, NewReg));
444 if (isDefLiveOut(Reg, TailBB, MRI) || UsedByPhi.count(Reg))
445 AddSSAUpdateEntry(Reg, NewReg, PredBB);
447 DenseMap<unsigned, unsigned>::iterator VI = LocalVRMap.find(Reg);
448 if (VI != LocalVRMap.end()) {
449 MO.setReg(VI->second);
450 // Clear any kill flags from this operand. The new register could have
451 // uses after this one, so kills are not valid here.
453 MRI->constrainRegClass(VI->second, MRI->getRegClass(Reg));
457 PredBB->insert(PredBB->instr_end(), NewMI);
460 /// After FromBB is tail duplicated into its predecessor blocks, the successors
461 /// have gained new predecessors. Update the PHI instructions in them
464 TailDuplicatePass::UpdateSuccessorsPHIs(MachineBasicBlock *FromBB, bool isDead,
465 SmallVectorImpl<MachineBasicBlock *> &TDBBs,
466 SmallSetVector<MachineBasicBlock*,8> &Succs) {
467 for (SmallSetVector<MachineBasicBlock*, 8>::iterator SI = Succs.begin(),
468 SE = Succs.end(); SI != SE; ++SI) {
469 MachineBasicBlock *SuccBB = *SI;
470 for (MachineBasicBlock::iterator II = SuccBB->begin(), EE = SuccBB->end();
474 MachineInstrBuilder MIB(*FromBB->getParent(), II);
476 for (unsigned i = 1, e = II->getNumOperands(); i != e; i += 2) {
477 MachineOperand &MO = II->getOperand(i+1);
478 if (MO.getMBB() == FromBB) {
485 MachineOperand &MO0 = II->getOperand(Idx);
486 unsigned Reg = MO0.getReg();
488 // Folded into the previous BB.
489 // There could be duplicate phi source entries. FIXME: Should sdisel
490 // or earlier pass fixed this?
491 for (unsigned i = II->getNumOperands()-2; i != Idx; i -= 2) {
492 MachineOperand &MO = II->getOperand(i+1);
493 if (MO.getMBB() == FromBB) {
494 II->RemoveOperand(i+1);
495 II->RemoveOperand(i);
501 // If Idx is set, the operands at Idx and Idx+1 must be removed.
502 // We reuse the location to avoid expensive RemoveOperand calls.
504 DenseMap<unsigned,AvailableValsTy>::iterator LI=SSAUpdateVals.find(Reg);
505 if (LI != SSAUpdateVals.end()) {
506 // This register is defined in the tail block.
507 for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) {
508 MachineBasicBlock *SrcBB = LI->second[j].first;
509 // If we didn't duplicate a bb into a particular predecessor, we
510 // might still have added an entry to SSAUpdateVals to correcly
511 // recompute SSA. If that case, avoid adding a dummy extra argument
513 if (!SrcBB->isSuccessor(SuccBB))
516 unsigned SrcReg = LI->second[j].second;
518 II->getOperand(Idx).setReg(SrcReg);
519 II->getOperand(Idx+1).setMBB(SrcBB);
522 MIB.addReg(SrcReg).addMBB(SrcBB);
526 // Live in tail block, must also be live in predecessors.
527 for (unsigned j = 0, ee = TDBBs.size(); j != ee; ++j) {
528 MachineBasicBlock *SrcBB = TDBBs[j];
530 II->getOperand(Idx).setReg(Reg);
531 II->getOperand(Idx+1).setMBB(SrcBB);
534 MIB.addReg(Reg).addMBB(SrcBB);
539 II->RemoveOperand(Idx+1);
540 II->RemoveOperand(Idx);
546 /// Determine if it is profitable to duplicate this block.
548 TailDuplicatePass::shouldTailDuplicate(const MachineFunction &MF,
550 MachineBasicBlock &TailBB) {
551 // Only duplicate blocks that end with unconditional branches.
552 if (TailBB.canFallThrough())
555 // Don't try to tail-duplicate single-block loops.
556 if (TailBB.isSuccessor(&TailBB))
559 // Set the limit on the cost to duplicate. When optimizing for size,
560 // duplicate only one, because one branch instruction can be eliminated to
561 // compensate for the duplication.
562 unsigned MaxDuplicateCount;
563 if (TailDuplicateSize.getNumOccurrences() == 0 &&
564 // FIXME: Use Function::optForSize().
565 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize))
566 MaxDuplicateCount = 1;
568 MaxDuplicateCount = TailDuplicateSize;
570 // If the target has hardware branch prediction that can handle indirect
571 // branches, duplicating them can often make them predictable when there
572 // are common paths through the code. The limit needs to be high enough
573 // to allow undoing the effects of tail merging and other optimizations
574 // that rearrange the predecessors of the indirect branch.
576 bool HasIndirectbr = false;
578 HasIndirectbr = TailBB.back().isIndirectBranch();
580 if (HasIndirectbr && PreRegAlloc)
581 MaxDuplicateCount = 20;
583 // Check the instructions in the block to determine whether tail-duplication
584 // is invalid or unlikely to be profitable.
585 unsigned InstrCount = 0;
586 for (MachineInstr &MI : TailBB) {
587 // Non-duplicable things shouldn't be tail-duplicated.
588 if (MI.isNotDuplicable())
591 // Do not duplicate 'return' instructions if this is a pre-regalloc run.
592 // A return may expand into a lot more instructions (e.g. reload of callee
593 // saved registers) after PEI.
594 if (PreRegAlloc && MI.isReturn())
597 // Avoid duplicating calls before register allocation. Calls presents a
598 // barrier to register allocation so duplicating them may end up increasing
600 if (PreRegAlloc && MI.isCall())
603 if (!MI.isPHI() && !MI.isDebugValue())
606 if (InstrCount > MaxDuplicateCount)
610 if (HasIndirectbr && PreRegAlloc)
619 return canCompletelyDuplicateBB(TailBB);
622 /// True if this BB has only one unconditional jump.
624 TailDuplicatePass::isSimpleBB(MachineBasicBlock *TailBB) {
625 if (TailBB->succ_size() != 1)
627 if (TailBB->pred_empty())
629 MachineBasicBlock::iterator I = TailBB->getFirstNonDebugInstr();
630 if (I == TailBB->end())
632 return I->isUnconditionalBranch();
636 bothUsedInPHI(const MachineBasicBlock &A,
637 SmallPtrSet<MachineBasicBlock*, 8> SuccsB) {
638 for (MachineBasicBlock *BB : A.successors())
639 if (SuccsB.count(BB) && !BB->empty() && BB->begin()->isPHI())
646 TailDuplicatePass::canCompletelyDuplicateBB(MachineBasicBlock &BB) {
647 for (MachineBasicBlock *PredBB : BB.predecessors()) {
648 if (PredBB->succ_size() > 1)
651 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
652 SmallVector<MachineOperand, 4> PredCond;
653 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
656 if (!PredCond.empty())
663 TailDuplicatePass::duplicateSimpleBB(MachineBasicBlock *TailBB,
664 SmallVectorImpl<MachineBasicBlock *> &TDBBs,
665 const DenseSet<unsigned> &UsedByPhi,
666 SmallVectorImpl<MachineInstr *> &Copies) {
667 SmallPtrSet<MachineBasicBlock*, 8> Succs(TailBB->succ_begin(),
669 SmallVector<MachineBasicBlock*, 8> Preds(TailBB->pred_begin(),
671 bool Changed = false;
672 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
673 PE = Preds.end(); PI != PE; ++PI) {
674 MachineBasicBlock *PredBB = *PI;
676 if (PredBB->getLandingPadSuccessor())
679 if (bothUsedInPHI(*PredBB, Succs))
682 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
683 SmallVector<MachineOperand, 4> PredCond;
684 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
688 DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB
689 << "From simple Succ: " << *TailBB);
691 MachineBasicBlock *NewTarget = *TailBB->succ_begin();
692 MachineBasicBlock *NextBB = std::next(MachineFunction::iterator(PredBB));
694 // Make PredFBB explicit.
695 if (PredCond.empty())
698 // Make fall through explicit.
705 if (PredFBB == TailBB)
707 if (PredTBB == TailBB)
710 // Make the branch unconditional if possible
711 if (PredTBB == PredFBB) {
716 // Avoid adding fall through branches.
717 if (PredFBB == NextBB)
719 if (PredTBB == NextBB && PredFBB == nullptr)
722 TII->RemoveBranch(*PredBB);
725 TII->InsertBranch(*PredBB, PredTBB, PredFBB, PredCond, DebugLoc());
727 uint32_t Weight = MBPI->getEdgeWeight(PredBB, TailBB);
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, Weight);
734 TDBBs.push_back(PredBB);
739 /// If it is profitable, duplicate TailBB's contents in each
740 /// of its predecessors.
742 TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB,
745 SmallVectorImpl<MachineBasicBlock *> &TDBBs,
746 SmallVectorImpl<MachineInstr *> &Copies) {
747 DEBUG(dbgs() << "\n*** Tail-duplicating BB#" << TailBB->getNumber() << '\n');
749 DenseSet<unsigned> UsedByPhi;
750 getRegsUsedByPHIs(*TailBB, &UsedByPhi);
753 return duplicateSimpleBB(TailBB, TDBBs, UsedByPhi, Copies);
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(),
761 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
762 PE = Preds.end(); PI != PE; ++PI) {
763 MachineBasicBlock *PredBB = *PI;
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)
771 MachineBasicBlock *PredTBB, *PredFBB;
772 SmallVector<MachineOperand, 4> PredCond;
773 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
775 if (!PredCond.empty())
777 // Don't duplicate into a fall-through predecessor (at least for now).
778 if (PredBB->isLayoutSuccessor(TailBB) && PredBB->canFallThrough())
781 DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB
782 << "From Succ: " << *TailBB);
784 TDBBs.push_back(PredBB);
786 // Remove PredBB's unconditional branch.
787 TII->RemoveBranch(*PredBB);
789 if (RS && !TailBB->livein_empty()) {
790 // Update PredBB livein.
791 RS->enterBasicBlock(PredBB);
792 if (!PredBB->empty())
793 RS->forward(std::prev(PredBB->end()));
794 for (MachineBasicBlock::livein_iterator I = TailBB->livein_begin(),
795 E = TailBB->livein_end(); I != E; ++I) {
796 if (!RS->isRegUsed(*I, false))
797 // If a register is previously livein to the tail but it's not live
798 // at the end of predecessor BB, then it should be added to its
800 PredBB->addLiveIn(*I);
804 // Clone the contents of TailBB into PredBB.
805 DenseMap<unsigned, unsigned> LocalVRMap;
806 SmallVector<std::pair<unsigned,unsigned>, 4> CopyInfos;
807 // Use instr_iterator here to properly handle bundles, e.g.
808 // ARM Thumb2 IT block.
809 MachineBasicBlock::instr_iterator I = TailBB->instr_begin();
810 while (I != TailBB->instr_end()) {
811 MachineInstr *MI = &*I;
814 // Replace the uses of the def of the PHI with the register coming
816 ProcessPHI(MI, TailBB, PredBB, LocalVRMap, CopyInfos, UsedByPhi, true);
818 // Replace def of virtual registers with new registers, and update
819 // uses with PHI source register or the new registers.
820 DuplicateInstruction(MI, TailBB, PredBB, MF, LocalVRMap, UsedByPhi);
823 MachineBasicBlock::iterator Loc = PredBB->getFirstTerminator();
824 for (unsigned i = 0, e = CopyInfos.size(); i != e; ++i) {
825 Copies.push_back(BuildMI(*PredBB, Loc, DebugLoc(),
826 TII->get(TargetOpcode::COPY),
827 CopyInfos[i].first).addReg(CopyInfos[i].second));
831 TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true);
833 NumInstrDups += TailBB->size() - 1; // subtract one for removed branch
836 PredBB->removeSuccessor(PredBB->succ_begin());
837 assert(PredBB->succ_empty() &&
838 "TailDuplicate called on block with multiple successors!");
839 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
840 E = TailBB->succ_end(); I != E; ++I)
841 PredBB->addSuccessor(*I, MBPI->getEdgeWeight(TailBB, I));
847 // If TailBB was duplicated into all its predecessors except for the prior
848 // block, which falls through unconditionally, move the contents of this
849 // block into the prior block.
850 MachineBasicBlock *PrevBB = std::prev(MachineFunction::iterator(TailBB));
851 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
852 SmallVector<MachineOperand, 4> PriorCond;
853 // This has to check PrevBB->succ_size() because EH edges are ignored by
855 if (PrevBB->succ_size() == 1 &&
856 !TII->AnalyzeBranch(*PrevBB, PriorTBB, PriorFBB, PriorCond, true) &&
857 PriorCond.empty() && !PriorTBB && TailBB->pred_size() == 1 &&
858 !TailBB->hasAddressTaken()) {
859 DEBUG(dbgs() << "\nMerging into block: " << *PrevBB
860 << "From MBB: " << *TailBB);
862 DenseMap<unsigned, unsigned> LocalVRMap;
863 SmallVector<std::pair<unsigned,unsigned>, 4> CopyInfos;
864 MachineBasicBlock::iterator I = TailBB->begin();
865 // Process PHI instructions first.
866 while (I != TailBB->end() && I->isPHI()) {
867 // Replace the uses of the def of the PHI with the register coming
869 MachineInstr *MI = &*I++;
870 ProcessPHI(MI, TailBB, PrevBB, LocalVRMap, CopyInfos, UsedByPhi, true);
872 MI->eraseFromParent();
875 // Now copy the non-PHI instructions.
876 while (I != TailBB->end()) {
877 // Replace def of virtual registers with new registers, and update
878 // uses with PHI source register or the new registers.
879 MachineInstr *MI = &*I++;
880 assert(!MI->isBundle() && "Not expecting bundles before regalloc!");
881 DuplicateInstruction(MI, TailBB, PrevBB, MF, LocalVRMap, UsedByPhi);
882 MI->eraseFromParent();
884 MachineBasicBlock::iterator Loc = PrevBB->getFirstTerminator();
885 for (unsigned i = 0, e = CopyInfos.size(); i != e; ++i) {
886 Copies.push_back(BuildMI(*PrevBB, Loc, DebugLoc(),
887 TII->get(TargetOpcode::COPY),
889 .addReg(CopyInfos[i].second));
892 // No PHIs to worry about, just splice the instructions over.
893 PrevBB->splice(PrevBB->end(), TailBB, TailBB->begin(), TailBB->end());
895 PrevBB->removeSuccessor(PrevBB->succ_begin());
896 assert(PrevBB->succ_empty());
897 PrevBB->transferSuccessors(TailBB);
898 TDBBs.push_back(PrevBB);
902 // If this is after register allocation, there are no phis to fix.
906 // If we made no changes so far, we are safe.
911 // Handle the nasty case in that we duplicated a block that is part of a loop
912 // into some but not all of its predecessors. For example:
916 // if we duplicate 2 into 1 but not into 3, we end up with
917 // 12 -> 3 <-> 2 -> rest |
920 // If there was a "var = phi(1, 3)" in 2, it has to be ultimately replaced
921 // with a phi in 3 (which now dominates 2).
922 // What we do here is introduce a copy in 3 of the register defined by the
923 // phi, just like when we are duplicating 2 into 3, but we don't copy any
924 // real instructions or remove the 3 -> 2 edge from the phi in 2.
925 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
926 PE = Preds.end(); PI != PE; ++PI) {
927 MachineBasicBlock *PredBB = *PI;
928 if (std::find(TDBBs.begin(), TDBBs.end(), PredBB) != TDBBs.end())
932 if (PredBB->succ_size() != 1)
935 DenseMap<unsigned, unsigned> LocalVRMap;
936 SmallVector<std::pair<unsigned,unsigned>, 4> CopyInfos;
937 MachineBasicBlock::iterator I = TailBB->begin();
938 // Process PHI instructions first.
939 while (I != TailBB->end() && I->isPHI()) {
940 // Replace the uses of the def of the PHI with the register coming
942 MachineInstr *MI = &*I++;
943 ProcessPHI(MI, TailBB, PredBB, LocalVRMap, CopyInfos, UsedByPhi, false);
945 MachineBasicBlock::iterator Loc = PredBB->getFirstTerminator();
946 for (unsigned i = 0, e = CopyInfos.size(); i != e; ++i) {
947 Copies.push_back(BuildMI(*PredBB, Loc, DebugLoc(),
948 TII->get(TargetOpcode::COPY),
949 CopyInfos[i].first).addReg(CopyInfos[i].second));
956 /// Remove the specified dead machine basic block from the function, updating
958 void TailDuplicatePass::RemoveDeadBlock(MachineBasicBlock *MBB) {
959 assert(MBB->pred_empty() && "MBB must be dead!");
960 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
962 // Remove all successors.
963 while (!MBB->succ_empty())
964 MBB->removeSuccessor(MBB->succ_end()-1);
967 MBB->eraseFromParent();