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 /// TailDuplicatePass - 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 // SSAUpdateVRs - A list of virtual registers for which to update SSA form.
73 SmallVector<unsigned, 16> SSAUpdateVRs;
75 // SSAUpdateVals - For each virtual register in SSAUpdateVals keep a list of
76 // source virtual registers.
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 /// TailDuplicateAndUpdate - 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 /// TailDuplicateBlocks - Look for small blocks that are unconditionally
314 /// branched to and do not fall through. Tail-duplicate their instructions
315 /// into their predecessors to 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 /// AddSSAUpdateEntry - Add a definition and source virtual registers pair for
380 void TailDuplicatePass::AddSSAUpdateEntry(unsigned OrigReg, unsigned NewReg,
381 MachineBasicBlock *BB) {
382 DenseMap<unsigned, AvailableValsTy>::iterator LI= SSAUpdateVals.find(OrigReg);
383 if (LI != SSAUpdateVals.end())
384 LI->second.push_back(std::make_pair(BB, NewReg));
386 AvailableValsTy Vals;
387 Vals.push_back(std::make_pair(BB, NewReg));
388 SSAUpdateVals.insert(std::make_pair(OrigReg, Vals));
389 SSAUpdateVRs.push_back(OrigReg);
393 /// ProcessPHI - Process PHI node in TailBB by turning it into a copy in PredBB.
394 /// Remember the source register that's contributed by PredBB and update SSA
396 void TailDuplicatePass::ProcessPHI(
397 MachineInstr *MI, MachineBasicBlock *TailBB, MachineBasicBlock *PredBB,
398 DenseMap<unsigned, unsigned> &LocalVRMap,
399 SmallVectorImpl<std::pair<unsigned, unsigned> > &Copies,
400 const DenseSet<unsigned> &RegsUsedByPhi, bool Remove) {
401 unsigned DefReg = MI->getOperand(0).getReg();
402 unsigned SrcOpIdx = getPHISrcRegOpIdx(MI, PredBB);
403 assert(SrcOpIdx && "Unable to find matching PHI source?");
404 unsigned SrcReg = MI->getOperand(SrcOpIdx).getReg();
405 const TargetRegisterClass *RC = MRI->getRegClass(DefReg);
406 LocalVRMap.insert(std::make_pair(DefReg, SrcReg));
408 // Insert a copy from source to the end of the block. The def register is the
409 // available value liveout of the block.
410 unsigned NewDef = MRI->createVirtualRegister(RC);
411 Copies.push_back(std::make_pair(NewDef, SrcReg));
412 if (isDefLiveOut(DefReg, TailBB, MRI) || RegsUsedByPhi.count(DefReg))
413 AddSSAUpdateEntry(DefReg, NewDef, PredBB);
418 // Remove PredBB from the PHI node.
419 MI->RemoveOperand(SrcOpIdx+1);
420 MI->RemoveOperand(SrcOpIdx);
421 if (MI->getNumOperands() == 1)
422 MI->eraseFromParent();
425 /// DuplicateInstruction - Duplicate a TailBB instruction to PredBB and update
426 /// the source operands due to earlier PHI translation.
427 void TailDuplicatePass::DuplicateInstruction(MachineInstr *MI,
428 MachineBasicBlock *TailBB,
429 MachineBasicBlock *PredBB,
431 DenseMap<unsigned, unsigned> &LocalVRMap,
432 const DenseSet<unsigned> &UsedByPhi) {
433 MachineInstr *NewMI = TII->duplicate(MI, MF);
434 for (unsigned i = 0, e = NewMI->getNumOperands(); i != e; ++i) {
435 MachineOperand &MO = NewMI->getOperand(i);
438 unsigned Reg = MO.getReg();
439 if (!TargetRegisterInfo::isVirtualRegister(Reg))
442 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
443 unsigned NewReg = MRI->createVirtualRegister(RC);
445 LocalVRMap.insert(std::make_pair(Reg, NewReg));
446 if (isDefLiveOut(Reg, TailBB, MRI) || UsedByPhi.count(Reg))
447 AddSSAUpdateEntry(Reg, NewReg, PredBB);
449 DenseMap<unsigned, unsigned>::iterator VI = LocalVRMap.find(Reg);
450 if (VI != LocalVRMap.end()) {
451 MO.setReg(VI->second);
452 // Clear any kill flags from this operand. The new register could have
453 // uses after this one, so kills are not valid here.
455 MRI->constrainRegClass(VI->second, MRI->getRegClass(Reg));
459 PredBB->insert(PredBB->instr_end(), NewMI);
462 /// UpdateSuccessorsPHIs - After FromBB is tail duplicated into its predecessor
463 /// blocks, the successors have gained new predecessors. Update the PHI
464 /// instructions in them accordingly.
466 TailDuplicatePass::UpdateSuccessorsPHIs(MachineBasicBlock *FromBB, bool isDead,
467 SmallVectorImpl<MachineBasicBlock *> &TDBBs,
468 SmallSetVector<MachineBasicBlock*,8> &Succs) {
469 for (SmallSetVector<MachineBasicBlock*, 8>::iterator SI = Succs.begin(),
470 SE = Succs.end(); SI != SE; ++SI) {
471 MachineBasicBlock *SuccBB = *SI;
472 for (MachineBasicBlock::iterator II = SuccBB->begin(), EE = SuccBB->end();
476 MachineInstrBuilder MIB(*FromBB->getParent(), II);
478 for (unsigned i = 1, e = II->getNumOperands(); i != e; i += 2) {
479 MachineOperand &MO = II->getOperand(i+1);
480 if (MO.getMBB() == FromBB) {
487 MachineOperand &MO0 = II->getOperand(Idx);
488 unsigned Reg = MO0.getReg();
490 // Folded into the previous BB.
491 // There could be duplicate phi source entries. FIXME: Should sdisel
492 // or earlier pass fixed this?
493 for (unsigned i = II->getNumOperands()-2; i != Idx; i -= 2) {
494 MachineOperand &MO = II->getOperand(i+1);
495 if (MO.getMBB() == FromBB) {
496 II->RemoveOperand(i+1);
497 II->RemoveOperand(i);
503 // If Idx is set, the operands at Idx and Idx+1 must be removed.
504 // We reuse the location to avoid expensive RemoveOperand calls.
506 DenseMap<unsigned,AvailableValsTy>::iterator LI=SSAUpdateVals.find(Reg);
507 if (LI != SSAUpdateVals.end()) {
508 // This register is defined in the tail block.
509 for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) {
510 MachineBasicBlock *SrcBB = LI->second[j].first;
511 // If we didn't duplicate a bb into a particular predecessor, we
512 // might still have added an entry to SSAUpdateVals to correcly
513 // recompute SSA. If that case, avoid adding a dummy extra argument
515 if (!SrcBB->isSuccessor(SuccBB))
518 unsigned SrcReg = LI->second[j].second;
520 II->getOperand(Idx).setReg(SrcReg);
521 II->getOperand(Idx+1).setMBB(SrcBB);
524 MIB.addReg(SrcReg).addMBB(SrcBB);
528 // Live in tail block, must also be live in predecessors.
529 for (unsigned j = 0, ee = TDBBs.size(); j != ee; ++j) {
530 MachineBasicBlock *SrcBB = TDBBs[j];
532 II->getOperand(Idx).setReg(Reg);
533 II->getOperand(Idx+1).setMBB(SrcBB);
536 MIB.addReg(Reg).addMBB(SrcBB);
541 II->RemoveOperand(Idx+1);
542 II->RemoveOperand(Idx);
548 /// shouldTailDuplicate - Determine if it is profitable to duplicate this block.
550 TailDuplicatePass::shouldTailDuplicate(const MachineFunction &MF,
552 MachineBasicBlock &TailBB) {
553 // Only duplicate blocks that end with unconditional branches.
554 if (TailBB.canFallThrough())
557 // Don't try to tail-duplicate single-block loops.
558 if (TailBB.isSuccessor(&TailBB))
561 // Set the limit on the cost to duplicate. When optimizing for size,
562 // duplicate only one, because one branch instruction can be eliminated to
563 // compensate for the duplication.
564 unsigned MaxDuplicateCount;
565 if (TailDuplicateSize.getNumOccurrences() == 0 &&
566 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize))
567 MaxDuplicateCount = 1;
569 MaxDuplicateCount = TailDuplicateSize;
571 // If the target has hardware branch prediction that can handle indirect
572 // branches, duplicating them can often make them predictable when there
573 // are common paths through the code. The limit needs to be high enough
574 // to allow undoing the effects of tail merging and other optimizations
575 // that rearrange the predecessors of the indirect branch.
577 bool HasIndirectbr = false;
579 HasIndirectbr = TailBB.back().isIndirectBranch();
581 if (HasIndirectbr && PreRegAlloc)
582 MaxDuplicateCount = 20;
584 // Check the instructions in the block to determine whether tail-duplication
585 // is invalid or unlikely to be profitable.
586 unsigned InstrCount = 0;
587 for (MachineBasicBlock::iterator I = TailBB.begin(); I != TailBB.end(); ++I) {
588 // Non-duplicable things shouldn't be tail-duplicated.
589 if (I->isNotDuplicable())
592 // Do not duplicate 'return' instructions if this is a pre-regalloc run.
593 // A return may expand into a lot more instructions (e.g. reload of callee
594 // saved registers) after PEI.
595 if (PreRegAlloc && I->isReturn())
598 // Avoid duplicating calls before register allocation. Calls presents a
599 // barrier to register allocation so duplicating them may end up increasing
601 if (PreRegAlloc && I->isCall())
604 if (!I->isPHI() && !I->isDebugValue())
607 if (InstrCount > MaxDuplicateCount)
611 if (HasIndirectbr && PreRegAlloc)
620 return canCompletelyDuplicateBB(TailBB);
623 /// isSimpleBB - True if this BB has only one unconditional jump.
625 TailDuplicatePass::isSimpleBB(MachineBasicBlock *TailBB) {
626 if (TailBB->succ_size() != 1)
628 if (TailBB->pred_empty())
630 MachineBasicBlock::iterator I = TailBB->getFirstNonDebugInstr();
631 if (I == TailBB->end())
633 return I->isUnconditionalBranch();
637 bothUsedInPHI(const MachineBasicBlock &A,
638 SmallPtrSet<MachineBasicBlock*, 8> SuccsB) {
639 for (MachineBasicBlock::const_succ_iterator SI = A.succ_begin(),
640 SE = A.succ_end(); SI != SE; ++SI) {
641 MachineBasicBlock *BB = *SI;
642 if (SuccsB.count(BB) && !BB->empty() && BB->begin()->isPHI())
650 TailDuplicatePass::canCompletelyDuplicateBB(MachineBasicBlock &BB) {
651 for (MachineBasicBlock::pred_iterator PI = BB.pred_begin(),
652 PE = BB.pred_end(); PI != PE; ++PI) {
653 MachineBasicBlock *PredBB = *PI;
655 if (PredBB->succ_size() > 1)
658 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
659 SmallVector<MachineOperand, 4> PredCond;
660 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
663 if (!PredCond.empty())
670 TailDuplicatePass::duplicateSimpleBB(MachineBasicBlock *TailBB,
671 SmallVectorImpl<MachineBasicBlock *> &TDBBs,
672 const DenseSet<unsigned> &UsedByPhi,
673 SmallVectorImpl<MachineInstr *> &Copies) {
674 SmallPtrSet<MachineBasicBlock*, 8> Succs(TailBB->succ_begin(),
676 SmallVector<MachineBasicBlock*, 8> Preds(TailBB->pred_begin(),
678 bool Changed = false;
679 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
680 PE = Preds.end(); PI != PE; ++PI) {
681 MachineBasicBlock *PredBB = *PI;
683 if (PredBB->getLandingPadSuccessor())
686 if (bothUsedInPHI(*PredBB, Succs))
689 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
690 SmallVector<MachineOperand, 4> PredCond;
691 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
695 DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB
696 << "From simple Succ: " << *TailBB);
698 MachineBasicBlock *NewTarget = *TailBB->succ_begin();
699 MachineBasicBlock *NextBB = std::next(MachineFunction::iterator(PredBB));
701 // Make PredFBB explicit.
702 if (PredCond.empty())
705 // Make fall through explicit.
712 if (PredFBB == TailBB)
714 if (PredTBB == TailBB)
717 // Make the branch unconditional if possible
718 if (PredTBB == PredFBB) {
723 // Avoid adding fall through branches.
724 if (PredFBB == NextBB)
726 if (PredTBB == NextBB && PredFBB == nullptr)
729 TII->RemoveBranch(*PredBB);
732 TII->InsertBranch(*PredBB, PredTBB, PredFBB, PredCond, DebugLoc());
734 uint32_t Weight = MBPI->getEdgeWeight(PredBB, TailBB);
735 PredBB->removeSuccessor(TailBB);
736 unsigned NumSuccessors = PredBB->succ_size();
737 assert(NumSuccessors <= 1);
738 if (NumSuccessors == 0 || *PredBB->succ_begin() != NewTarget)
739 PredBB->addSuccessor(NewTarget, Weight);
741 TDBBs.push_back(PredBB);
746 /// TailDuplicate - If it is profitable, duplicate TailBB's contents in each
747 /// of its predecessors.
749 TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB,
752 SmallVectorImpl<MachineBasicBlock *> &TDBBs,
753 SmallVectorImpl<MachineInstr *> &Copies) {
754 DEBUG(dbgs() << "\n*** Tail-duplicating BB#" << TailBB->getNumber() << '\n');
756 DenseSet<unsigned> UsedByPhi;
757 getRegsUsedByPHIs(*TailBB, &UsedByPhi);
760 return duplicateSimpleBB(TailBB, TDBBs, UsedByPhi, Copies);
762 // Iterate through all the unique predecessors and tail-duplicate this
763 // block into them, if possible. Copying the list ahead of time also
764 // avoids trouble with the predecessor list reallocating.
765 bool Changed = false;
766 SmallSetVector<MachineBasicBlock*, 8> Preds(TailBB->pred_begin(),
768 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
769 PE = Preds.end(); PI != PE; ++PI) {
770 MachineBasicBlock *PredBB = *PI;
772 assert(TailBB != PredBB &&
773 "Single-block loop should have been rejected earlier!");
774 // EH edges are ignored by AnalyzeBranch.
775 if (PredBB->succ_size() > 1)
778 MachineBasicBlock *PredTBB, *PredFBB;
779 SmallVector<MachineOperand, 4> PredCond;
780 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
782 if (!PredCond.empty())
784 // Don't duplicate into a fall-through predecessor (at least for now).
785 if (PredBB->isLayoutSuccessor(TailBB) && PredBB->canFallThrough())
788 DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB
789 << "From Succ: " << *TailBB);
791 TDBBs.push_back(PredBB);
793 // Remove PredBB's unconditional branch.
794 TII->RemoveBranch(*PredBB);
796 if (RS && !TailBB->livein_empty()) {
797 // Update PredBB livein.
798 RS->enterBasicBlock(PredBB);
799 if (!PredBB->empty())
800 RS->forward(std::prev(PredBB->end()));
801 for (MachineBasicBlock::livein_iterator I = TailBB->livein_begin(),
802 E = TailBB->livein_end(); I != E; ++I) {
803 if (!RS->isRegUsed(*I, false))
804 // If a register is previously livein to the tail but it's not live
805 // at the end of predecessor BB, then it should be added to its
807 PredBB->addLiveIn(*I);
811 // Clone the contents of TailBB into PredBB.
812 DenseMap<unsigned, unsigned> LocalVRMap;
813 SmallVector<std::pair<unsigned,unsigned>, 4> CopyInfos;
814 // Use instr_iterator here to properly handle bundles, e.g.
815 // ARM Thumb2 IT block.
816 MachineBasicBlock::instr_iterator I = TailBB->instr_begin();
817 while (I != TailBB->instr_end()) {
818 MachineInstr *MI = &*I;
821 // Replace the uses of the def of the PHI with the register coming
823 ProcessPHI(MI, TailBB, PredBB, LocalVRMap, CopyInfos, UsedByPhi, true);
825 // Replace def of virtual registers with new registers, and update
826 // uses with PHI source register or the new registers.
827 DuplicateInstruction(MI, TailBB, PredBB, MF, LocalVRMap, UsedByPhi);
830 MachineBasicBlock::iterator Loc = PredBB->getFirstTerminator();
831 for (unsigned i = 0, e = CopyInfos.size(); i != e; ++i) {
832 Copies.push_back(BuildMI(*PredBB, Loc, DebugLoc(),
833 TII->get(TargetOpcode::COPY),
834 CopyInfos[i].first).addReg(CopyInfos[i].second));
838 TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true);
840 NumInstrDups += TailBB->size() - 1; // subtract one for removed branch
843 PredBB->removeSuccessor(PredBB->succ_begin());
844 assert(PredBB->succ_empty() &&
845 "TailDuplicate called on block with multiple successors!");
846 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
847 E = TailBB->succ_end(); I != E; ++I)
848 PredBB->addSuccessor(*I, MBPI->getEdgeWeight(TailBB, I));
854 // If TailBB was duplicated into all its predecessors except for the prior
855 // block, which falls through unconditionally, move the contents of this
856 // block into the prior block.
857 MachineBasicBlock *PrevBB = std::prev(MachineFunction::iterator(TailBB));
858 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
859 SmallVector<MachineOperand, 4> PriorCond;
860 // This has to check PrevBB->succ_size() because EH edges are ignored by
862 if (PrevBB->succ_size() == 1 &&
863 !TII->AnalyzeBranch(*PrevBB, PriorTBB, PriorFBB, PriorCond, true) &&
864 PriorCond.empty() && !PriorTBB && TailBB->pred_size() == 1 &&
865 !TailBB->hasAddressTaken()) {
866 DEBUG(dbgs() << "\nMerging into block: " << *PrevBB
867 << "From MBB: " << *TailBB);
869 DenseMap<unsigned, unsigned> LocalVRMap;
870 SmallVector<std::pair<unsigned,unsigned>, 4> CopyInfos;
871 MachineBasicBlock::iterator I = TailBB->begin();
872 // Process PHI instructions first.
873 while (I != TailBB->end() && I->isPHI()) {
874 // Replace the uses of the def of the PHI with the register coming
876 MachineInstr *MI = &*I++;
877 ProcessPHI(MI, TailBB, PrevBB, LocalVRMap, CopyInfos, UsedByPhi, true);
879 MI->eraseFromParent();
882 // Now copy the non-PHI instructions.
883 while (I != TailBB->end()) {
884 // Replace def of virtual registers with new registers, and update
885 // uses with PHI source register or the new registers.
886 MachineInstr *MI = &*I++;
887 assert(!MI->isBundle() && "Not expecting bundles before regalloc!");
888 DuplicateInstruction(MI, TailBB, PrevBB, MF, LocalVRMap, UsedByPhi);
889 MI->eraseFromParent();
891 MachineBasicBlock::iterator Loc = PrevBB->getFirstTerminator();
892 for (unsigned i = 0, e = CopyInfos.size(); i != e; ++i) {
893 Copies.push_back(BuildMI(*PrevBB, Loc, DebugLoc(),
894 TII->get(TargetOpcode::COPY),
896 .addReg(CopyInfos[i].second));
899 // No PHIs to worry about, just splice the instructions over.
900 PrevBB->splice(PrevBB->end(), TailBB, TailBB->begin(), TailBB->end());
902 PrevBB->removeSuccessor(PrevBB->succ_begin());
903 assert(PrevBB->succ_empty());
904 PrevBB->transferSuccessors(TailBB);
905 TDBBs.push_back(PrevBB);
909 // If this is after register allocation, there are no phis to fix.
913 // If we made no changes so far, we are safe.
918 // Handle the nasty case in that we duplicated a block that is part of a loop
919 // into some but not all of its predecessors. For example:
923 // if we duplicate 2 into 1 but not into 3, we end up with
924 // 12 -> 3 <-> 2 -> rest |
927 // If there was a "var = phi(1, 3)" in 2, it has to be ultimately replaced
928 // with a phi in 3 (which now dominates 2).
929 // What we do here is introduce a copy in 3 of the register defined by the
930 // phi, just like when we are duplicating 2 into 3, but we don't copy any
931 // real instructions or remove the 3 -> 2 edge from the phi in 2.
932 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
933 PE = Preds.end(); PI != PE; ++PI) {
934 MachineBasicBlock *PredBB = *PI;
935 if (std::find(TDBBs.begin(), TDBBs.end(), PredBB) != TDBBs.end())
939 if (PredBB->succ_size() != 1)
942 DenseMap<unsigned, unsigned> LocalVRMap;
943 SmallVector<std::pair<unsigned,unsigned>, 4> CopyInfos;
944 MachineBasicBlock::iterator I = TailBB->begin();
945 // Process PHI instructions first.
946 while (I != TailBB->end() && I->isPHI()) {
947 // Replace the uses of the def of the PHI with the register coming
949 MachineInstr *MI = &*I++;
950 ProcessPHI(MI, TailBB, PredBB, LocalVRMap, CopyInfos, UsedByPhi, false);
952 MachineBasicBlock::iterator Loc = PredBB->getFirstTerminator();
953 for (unsigned i = 0, e = CopyInfos.size(); i != e; ++i) {
954 Copies.push_back(BuildMI(*PredBB, Loc, DebugLoc(),
955 TII->get(TargetOpcode::COPY),
956 CopyInfos[i].first).addReg(CopyInfos[i].second));
963 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
964 /// function, updating the CFG.
965 void TailDuplicatePass::RemoveDeadBlock(MachineBasicBlock *MBB) {
966 assert(MBB->pred_empty() && "MBB must be dead!");
967 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
969 // Remove all successors.
970 while (!MBB->succ_empty())
971 MBB->removeSuccessor(MBB->succ_end()-1);
974 MBB->eraseFromParent();