1 //===- BranchFolding.cpp - Fold machine code branch instructions ----------===//
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 forwards branches to unconditional branches to make them branch
11 // directly to the target block. This pass often results in dead MBB's, which
14 // Note that this pass must be run after register allocation, it cannot handle
15 // SSA form. It also must handle virtual registers for targets that emit virtual
18 //===----------------------------------------------------------------------===//
20 #include "BranchFolding.h"
21 #include "llvm/ADT/BitVector.h"
22 #include "llvm/ADT/DenseMap.h"
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/ADT/SmallPtrSet.h"
25 #include "llvm/ADT/SmallSet.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/CodeGen/Analysis.h"
29 #include "llvm/CodeGen/LivePhysRegs.h"
30 #include "llvm/CodeGen/MachineBasicBlock.h"
31 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
32 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineFunctionPass.h"
35 #include "llvm/CodeGen/MachineInstr.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineJumpTableInfo.h"
38 #include "llvm/CodeGen/MachineLoopInfo.h"
39 #include "llvm/CodeGen/MachineModuleInfo.h"
40 #include "llvm/CodeGen/MachineOperand.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/TargetInstrInfo.h"
43 #include "llvm/CodeGen/TargetOpcodes.h"
44 #include "llvm/CodeGen/TargetPassConfig.h"
45 #include "llvm/CodeGen/TargetRegisterInfo.h"
46 #include "llvm/CodeGen/TargetSubtargetInfo.h"
47 #include "llvm/IR/DebugInfoMetadata.h"
48 #include "llvm/IR/DebugLoc.h"
49 #include "llvm/IR/Function.h"
50 #include "llvm/MC/LaneBitmask.h"
51 #include "llvm/MC/MCRegisterInfo.h"
52 #include "llvm/Pass.h"
53 #include "llvm/Support/BlockFrequency.h"
54 #include "llvm/Support/BranchProbability.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/ErrorHandling.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/Target/TargetMachine.h"
68 #define DEBUG_TYPE "branch-folder"
70 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
71 STATISTIC(NumBranchOpts, "Number of branches optimized");
72 STATISTIC(NumTailMerge , "Number of block tails merged");
73 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
74 STATISTIC(NumTailCalls, "Number of tail calls optimized");
76 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
77 cl::init(cl::BOU_UNSET), cl::Hidden);
79 // Throttle for huge numbers of predecessors (compile speed problems)
80 static cl::opt<unsigned>
81 TailMergeThreshold("tail-merge-threshold",
82 cl::desc("Max number of predecessors to consider tail merging"),
83 cl::init(150), cl::Hidden);
85 // Heuristic for tail merging (and, inversely, tail duplication).
86 // TODO: This should be replaced with a target query.
87 static cl::opt<unsigned>
88 TailMergeSize("tail-merge-size",
89 cl::desc("Min number of instructions to consider tail merging"),
90 cl::init(3), cl::Hidden);
94 /// BranchFolderPass - Wrap branch folder in a machine function pass.
95 class BranchFolderPass : public MachineFunctionPass {
99 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
101 bool runOnMachineFunction(MachineFunction &MF) override;
103 void getAnalysisUsage(AnalysisUsage &AU) const override {
104 AU.addRequired<MachineBlockFrequencyInfo>();
105 AU.addRequired<MachineBranchProbabilityInfo>();
106 AU.addRequired<TargetPassConfig>();
107 MachineFunctionPass::getAnalysisUsage(AU);
111 } // end anonymous namespace
113 char BranchFolderPass::ID = 0;
115 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
117 INITIALIZE_PASS(BranchFolderPass, DEBUG_TYPE,
118 "Control Flow Optimizer", false, false)
120 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
121 if (skipFunction(MF.getFunction()))
124 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
125 // TailMerge can create jump into if branches that make CFG irreducible for
126 // HW that requires structurized CFG.
127 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
128 PassConfig->getEnableTailMerge();
129 BranchFolder::MBFIWrapper MBBFreqInfo(
130 getAnalysis<MachineBlockFrequencyInfo>());
131 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo,
132 getAnalysis<MachineBranchProbabilityInfo>());
133 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
134 MF.getSubtarget().getRegisterInfo(),
135 getAnalysisIfAvailable<MachineModuleInfo>());
138 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
139 MBFIWrapper &FreqInfo,
140 const MachineBranchProbabilityInfo &ProbInfo,
141 unsigned MinTailLength)
142 : EnableHoistCommonCode(CommonHoist), MinCommonTailLength(MinTailLength),
143 MBBFreqInfo(FreqInfo), MBPI(ProbInfo) {
144 if (MinCommonTailLength == 0)
145 MinCommonTailLength = TailMergeSize;
146 switch (FlagEnableTailMerge) {
147 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
148 case cl::BOU_TRUE: EnableTailMerge = true; break;
149 case cl::BOU_FALSE: EnableTailMerge = false; break;
153 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
154 assert(MBB->pred_empty() && "MBB must be dead!");
155 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
157 MachineFunction *MF = MBB->getParent();
158 // drop all successors.
159 while (!MBB->succ_empty())
160 MBB->removeSuccessor(MBB->succ_end()-1);
162 // Avoid matching if this pointer gets reused.
163 TriedMerging.erase(MBB);
167 FuncletMembership.erase(MBB);
169 MLI->removeBlock(MBB);
172 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
173 const TargetInstrInfo *tii,
174 const TargetRegisterInfo *tri,
175 MachineModuleInfo *mmi,
176 MachineLoopInfo *mli, bool AfterPlacement) {
177 if (!tii) return false;
179 TriedMerging.clear();
181 MachineRegisterInfo &MRI = MF.getRegInfo();
182 AfterBlockPlacement = AfterPlacement;
189 UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF);
191 MRI.invalidateLiveness();
193 // Fix CFG. The later algorithms expect it to be right.
194 bool MadeChange = false;
195 for (MachineBasicBlock &MBB : MF) {
196 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
197 SmallVector<MachineOperand, 4> Cond;
198 if (!TII->analyzeBranch(MBB, TBB, FBB, Cond, true))
199 MadeChange |= MBB.CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
202 // Recalculate funclet membership.
203 FuncletMembership = getFuncletMembership(MF);
205 bool MadeChangeThisIteration = true;
206 while (MadeChangeThisIteration) {
207 MadeChangeThisIteration = TailMergeBlocks(MF);
208 // No need to clean up if tail merging does not change anything after the
210 if (!AfterBlockPlacement || MadeChangeThisIteration)
211 MadeChangeThisIteration |= OptimizeBranches(MF);
212 if (EnableHoistCommonCode)
213 MadeChangeThisIteration |= HoistCommonCode(MF);
214 MadeChange |= MadeChangeThisIteration;
217 // See if any jump tables have become dead as the code generator
219 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
223 // Walk the function to find jump tables that are live.
224 BitVector JTIsLive(JTI->getJumpTables().size());
225 for (const MachineBasicBlock &BB : MF) {
226 for (const MachineInstr &I : BB)
227 for (const MachineOperand &Op : I.operands()) {
228 if (!Op.isJTI()) continue;
230 // Remember that this JT is live.
231 JTIsLive.set(Op.getIndex());
235 // Finally, remove dead jump tables. This happens when the
236 // indirect jump was unreachable (and thus deleted).
237 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
238 if (!JTIsLive.test(i)) {
239 JTI->RemoveJumpTable(i);
246 //===----------------------------------------------------------------------===//
247 // Tail Merging of Blocks
248 //===----------------------------------------------------------------------===//
250 /// HashMachineInstr - Compute a hash value for MI and its operands.
251 static unsigned HashMachineInstr(const MachineInstr &MI) {
252 unsigned Hash = MI.getOpcode();
253 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
254 const MachineOperand &Op = MI.getOperand(i);
256 // Merge in bits from the operand if easy. We can't use MachineOperand's
257 // hash_code here because it's not deterministic and we sort by hash value
259 unsigned OperandHash = 0;
260 switch (Op.getType()) {
261 case MachineOperand::MO_Register:
262 OperandHash = Op.getReg();
264 case MachineOperand::MO_Immediate:
265 OperandHash = Op.getImm();
267 case MachineOperand::MO_MachineBasicBlock:
268 OperandHash = Op.getMBB()->getNumber();
270 case MachineOperand::MO_FrameIndex:
271 case MachineOperand::MO_ConstantPoolIndex:
272 case MachineOperand::MO_JumpTableIndex:
273 OperandHash = Op.getIndex();
275 case MachineOperand::MO_GlobalAddress:
276 case MachineOperand::MO_ExternalSymbol:
277 // Global address / external symbol are too hard, don't bother, but do
278 // pull in the offset.
279 OperandHash = Op.getOffset();
285 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
290 /// HashEndOfMBB - Hash the last instruction in the MBB.
291 static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) {
292 MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr();
296 return HashMachineInstr(*I);
299 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
300 /// of instructions they actually have in common together at their end. Return
301 /// iterators for the first shared instruction in each block.
302 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
303 MachineBasicBlock *MBB2,
304 MachineBasicBlock::iterator &I1,
305 MachineBasicBlock::iterator &I2) {
309 unsigned TailLen = 0;
310 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
312 // Skip debugging pseudos; necessary to avoid changing the code.
313 while (I1->isDebugValue()) {
314 if (I1==MBB1->begin()) {
315 while (I2->isDebugValue()) {
316 if (I2==MBB2->begin())
317 // I1==DBG at begin; I2==DBG at begin
322 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
327 // I1==first (untested) non-DBG preceding known match
328 while (I2->isDebugValue()) {
329 if (I2==MBB2->begin()) {
331 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
336 // I1, I2==first (untested) non-DBGs preceding known match
337 if (!I1->isIdenticalTo(*I2) ||
338 // FIXME: This check is dubious. It's used to get around a problem where
339 // people incorrectly expect inline asm directives to remain in the same
340 // relative order. This is untenable because normal compiler
341 // optimizations (like this one) may reorder and/or merge these
349 // Back past possible debugging pseudos at beginning of block. This matters
350 // when one block differs from the other only by whether debugging pseudos
351 // are present at the beginning. (This way, the various checks later for
352 // I1==MBB1->begin() work as expected.)
353 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
355 while (I2->isDebugValue()) {
356 if (I2 == MBB2->begin())
362 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
364 while (I1->isDebugValue()) {
365 if (I1 == MBB1->begin())
374 void BranchFolder::replaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
375 MachineBasicBlock &NewDest) {
377 // OldInst should always point to an instruction.
378 MachineBasicBlock &OldMBB = *OldInst->getParent();
380 LiveRegs.addLiveOuts(OldMBB);
381 // Move backward to the place where will insert the jump.
382 MachineBasicBlock::iterator I = OldMBB.end();
385 LiveRegs.stepBackward(*I);
386 } while (I != OldInst);
388 // Merging the tails may have switched some undef operand to non-undef ones.
389 // Add IMPLICIT_DEFS into OldMBB as necessary to have a definition of the
391 for (MachineBasicBlock::RegisterMaskPair P : NewDest.liveins()) {
392 // We computed the liveins with computeLiveIn earlier and should only see
394 assert(P.LaneMask == LaneBitmask::getAll() &&
395 "Can only handle full register.");
396 MCPhysReg Reg = P.PhysReg;
397 if (!LiveRegs.available(*MRI, Reg))
400 BuildMI(OldMBB, OldInst, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Reg);
404 TII->ReplaceTailWithBranchTo(OldInst, &NewDest);
408 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
409 MachineBasicBlock::iterator BBI1,
410 const BasicBlock *BB) {
411 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
414 MachineFunction &MF = *CurMBB.getParent();
416 // Create the fall-through block.
417 MachineFunction::iterator MBBI = CurMBB.getIterator();
418 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(BB);
419 CurMBB.getParent()->insert(++MBBI, NewMBB);
421 // Move all the successors of this block to the specified block.
422 NewMBB->transferSuccessors(&CurMBB);
424 // Add an edge from CurMBB to NewMBB for the fall-through.
425 CurMBB.addSuccessor(NewMBB);
427 // Splice the code over.
428 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
430 // NewMBB belongs to the same loop as CurMBB.
432 if (MachineLoop *ML = MLI->getLoopFor(&CurMBB))
433 ML->addBasicBlockToLoop(NewMBB, MLI->getBase());
435 // NewMBB inherits CurMBB's block frequency.
436 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
439 computeAndAddLiveIns(LiveRegs, *NewMBB);
441 // Add the new block to the funclet.
442 const auto &FuncletI = FuncletMembership.find(&CurMBB);
443 if (FuncletI != FuncletMembership.end()) {
444 auto n = FuncletI->second;
445 FuncletMembership[NewMBB] = n;
451 /// EstimateRuntime - Make a rough estimate for how long it will take to run
452 /// the specified code.
453 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
454 MachineBasicBlock::iterator E) {
456 for (; I != E; ++I) {
457 if (I->isDebugValue())
461 else if (I->mayLoad() || I->mayStore())
469 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
470 // branches temporarily for tail merging). In the case where CurMBB ends
471 // with a conditional branch to the next block, optimize by reversing the
472 // test and conditionally branching to SuccMBB instead.
473 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
474 const TargetInstrInfo *TII) {
475 MachineFunction *MF = CurMBB->getParent();
476 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
477 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
478 SmallVector<MachineOperand, 4> Cond;
479 DebugLoc dl = CurMBB->findBranchDebugLoc();
480 if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
481 MachineBasicBlock *NextBB = &*I;
482 if (TBB == NextBB && !Cond.empty() && !FBB) {
483 if (!TII->reverseBranchCondition(Cond)) {
484 TII->removeBranch(*CurMBB);
485 TII->insertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
490 TII->insertBranch(*CurMBB, SuccBB, nullptr,
491 SmallVector<MachineOperand, 0>(), dl);
495 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
496 if (getHash() < o.getHash())
498 if (getHash() > o.getHash())
500 if (getBlock()->getNumber() < o.getBlock()->getNumber())
502 if (getBlock()->getNumber() > o.getBlock()->getNumber())
504 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
505 // an object with itself.
506 #ifndef _GLIBCXX_DEBUG
507 llvm_unreachable("Predecessor appears twice");
514 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
515 auto I = MergedBBFreq.find(MBB);
517 if (I != MergedBBFreq.end())
520 return MBFI.getBlockFreq(MBB);
523 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
525 MergedBBFreq[MBB] = F;
529 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
530 const MachineBasicBlock *MBB) const {
531 return MBFI.printBlockFreq(OS, getBlockFreq(MBB));
535 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
536 const BlockFrequency Freq) const {
537 return MBFI.printBlockFreq(OS, Freq);
540 void BranchFolder::MBFIWrapper::view(const Twine &Name, bool isSimple) {
541 MBFI.view(Name, isSimple);
545 BranchFolder::MBFIWrapper::getEntryFreq() const {
546 return MBFI.getEntryFreq();
549 /// CountTerminators - Count the number of terminators in the given
550 /// block and set I to the position of the first non-terminator, if there
551 /// is one, or MBB->end() otherwise.
552 static unsigned CountTerminators(MachineBasicBlock *MBB,
553 MachineBasicBlock::iterator &I) {
555 unsigned NumTerms = 0;
557 if (I == MBB->begin()) {
562 if (!I->isTerminator()) break;
568 /// A no successor, non-return block probably ends in unreachable and is cold.
569 /// Also consider a block that ends in an indirect branch to be a return block,
570 /// since many targets use plain indirect branches to return.
571 static bool blockEndsInUnreachable(const MachineBasicBlock *MBB) {
572 if (!MBB->succ_empty())
576 return !(MBB->back().isReturn() || MBB->back().isIndirectBranch());
579 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
580 /// and decide if it would be profitable to merge those tails. Return the
581 /// length of the common tail and iterators to the first common instruction
583 /// MBB1, MBB2 The blocks to check
584 /// MinCommonTailLength Minimum size of tail block to be merged.
585 /// CommonTailLen Out parameter to record the size of the shared tail between
587 /// I1, I2 Iterator references that will be changed to point to the first
588 /// instruction in the common tail shared by MBB1,MBB2
589 /// SuccBB A common successor of MBB1, MBB2 which are in a canonical form
590 /// relative to SuccBB
591 /// PredBB The layout predecessor of SuccBB, if any.
592 /// FuncletMembership map from block to funclet #.
593 /// AfterPlacement True if we are merging blocks after layout. Stricter
594 /// thresholds apply to prevent undoing tail-duplication.
596 ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2,
597 unsigned MinCommonTailLength, unsigned &CommonTailLen,
598 MachineBasicBlock::iterator &I1,
599 MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB,
600 MachineBasicBlock *PredBB,
601 DenseMap<const MachineBasicBlock *, int> &FuncletMembership,
602 bool AfterPlacement) {
603 // It is never profitable to tail-merge blocks from two different funclets.
604 if (!FuncletMembership.empty()) {
605 auto Funclet1 = FuncletMembership.find(MBB1);
606 assert(Funclet1 != FuncletMembership.end());
607 auto Funclet2 = FuncletMembership.find(MBB2);
608 assert(Funclet2 != FuncletMembership.end());
609 if (Funclet1->second != Funclet2->second)
613 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
614 if (CommonTailLen == 0)
616 DEBUG(dbgs() << "Common tail length of " << printMBBReference(*MBB1)
617 << " and " << printMBBReference(*MBB2) << " is " << CommonTailLen
620 // It's almost always profitable to merge any number of non-terminator
621 // instructions with the block that falls through into the common successor.
622 // This is true only for a single successor. For multiple successors, we are
623 // trading a conditional branch for an unconditional one.
624 // TODO: Re-visit successor size for non-layout tail merging.
625 if ((MBB1 == PredBB || MBB2 == PredBB) &&
626 (!AfterPlacement || MBB1->succ_size() == 1)) {
627 MachineBasicBlock::iterator I;
628 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
629 if (CommonTailLen > NumTerms)
633 // If these are identical non-return blocks with no successors, merge them.
634 // Such blocks are typically cold calls to noreturn functions like abort, and
635 // are unlikely to become a fallthrough target after machine block placement.
636 // Tail merging these blocks is unlikely to create additional unconditional
637 // branches, and will reduce the size of this cold code.
638 if (I1 == MBB1->begin() && I2 == MBB2->begin() &&
639 blockEndsInUnreachable(MBB1) && blockEndsInUnreachable(MBB2))
642 // If one of the blocks can be completely merged and happens to be in
643 // a position where the other could fall through into it, merge any number
644 // of instructions, because it can be done without a branch.
645 // TODO: If the blocks are not adjacent, move one of them so that they are?
646 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
648 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
651 // If both blocks are identical and end in a branch, merge them unless they
652 // both have a fallthrough predecessor and successor.
653 // We can only do this after block placement because it depends on whether
654 // there are fallthroughs, and we don't know until after layout.
655 if (AfterPlacement && I1 == MBB1->begin() && I2 == MBB2->begin()) {
656 auto BothFallThrough = [](MachineBasicBlock *MBB) {
657 if (MBB->succ_size() != 0 && !MBB->canFallThrough())
659 MachineFunction::iterator I(MBB);
660 MachineFunction *MF = MBB->getParent();
661 return (MBB != &*MF->begin()) && std::prev(I)->canFallThrough();
663 if (!BothFallThrough(MBB1) || !BothFallThrough(MBB2))
667 // If both blocks have an unconditional branch temporarily stripped out,
668 // count that as an additional common instruction for the following
669 // heuristics. This heuristic is only accurate for single-succ blocks, so to
670 // make sure that during layout merging and duplicating don't crash, we check
671 // for that when merging during layout.
672 unsigned EffectiveTailLen = CommonTailLen;
673 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
674 (MBB1->succ_size() == 1 || !AfterPlacement) &&
675 !MBB1->back().isBarrier() &&
676 !MBB2->back().isBarrier())
679 // Check if the common tail is long enough to be worthwhile.
680 if (EffectiveTailLen >= MinCommonTailLength)
683 // If we are optimizing for code size, 2 instructions in common is enough if
684 // we don't have to split a block. At worst we will be introducing 1 new
685 // branch instruction, which is likely to be smaller than the 2
686 // instructions that would be deleted in the merge.
687 MachineFunction *MF = MBB1->getParent();
688 return EffectiveTailLen >= 2 && MF->getFunction().optForSize() &&
689 (I1 == MBB1->begin() || I2 == MBB2->begin());
692 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
693 unsigned MinCommonTailLength,
694 MachineBasicBlock *SuccBB,
695 MachineBasicBlock *PredBB) {
696 unsigned maxCommonTailLength = 0U;
698 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
699 MPIterator HighestMPIter = std::prev(MergePotentials.end());
700 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
701 B = MergePotentials.begin();
702 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
703 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
704 unsigned CommonTailLen;
705 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
707 CommonTailLen, TrialBBI1, TrialBBI2,
710 AfterBlockPlacement)) {
711 if (CommonTailLen > maxCommonTailLength) {
713 maxCommonTailLength = CommonTailLen;
714 HighestMPIter = CurMPIter;
715 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
717 if (HighestMPIter == CurMPIter &&
718 CommonTailLen == maxCommonTailLength)
719 SameTails.push_back(SameTailElt(I, TrialBBI2));
725 return maxCommonTailLength;
728 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
729 MachineBasicBlock *SuccBB,
730 MachineBasicBlock *PredBB) {
731 MPIterator CurMPIter, B;
732 for (CurMPIter = std::prev(MergePotentials.end()),
733 B = MergePotentials.begin();
734 CurMPIter->getHash() == CurHash; --CurMPIter) {
735 // Put the unconditional branch back, if we need one.
736 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
737 if (SuccBB && CurMBB != PredBB)
738 FixTail(CurMBB, SuccBB, TII);
742 if (CurMPIter->getHash() != CurHash)
744 MergePotentials.erase(CurMPIter, MergePotentials.end());
747 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
748 MachineBasicBlock *SuccBB,
749 unsigned maxCommonTailLength,
750 unsigned &commonTailIndex) {
752 unsigned TimeEstimate = ~0U;
753 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
754 // Use PredBB if possible; that doesn't require a new branch.
755 if (SameTails[i].getBlock() == PredBB) {
759 // Otherwise, make a (fairly bogus) choice based on estimate of
760 // how long it will take the various blocks to execute.
761 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
762 SameTails[i].getTailStartPos());
763 if (t <= TimeEstimate) {
769 MachineBasicBlock::iterator BBI =
770 SameTails[commonTailIndex].getTailStartPos();
771 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
773 DEBUG(dbgs() << "\nSplitting " << printMBBReference(*MBB) << ", size "
774 << maxCommonTailLength);
776 // If the split block unconditionally falls-thru to SuccBB, it will be
777 // merged. In control flow terms it should then take SuccBB's name. e.g. If
778 // SuccBB is an inner loop, the common tail is still part of the inner loop.
779 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
780 SuccBB->getBasicBlock() : MBB->getBasicBlock();
781 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
783 DEBUG(dbgs() << "... failed!");
787 SameTails[commonTailIndex].setBlock(newMBB);
788 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
790 // If we split PredBB, newMBB is the new predecessor.
798 mergeOperations(MachineBasicBlock::iterator MBBIStartPos,
799 MachineBasicBlock &MBBCommon) {
800 MachineBasicBlock *MBB = MBBIStartPos->getParent();
801 // Note CommonTailLen does not necessarily matches the size of
802 // the common BB nor all its instructions because of debug
803 // instructions differences.
804 unsigned CommonTailLen = 0;
805 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
808 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
809 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
810 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
811 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
813 while (CommonTailLen--) {
814 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
817 if (MBBI->isDebugValue()) {
822 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
825 assert(MBBICommon != MBBIECommon &&
826 "Reached BB end within common tail length!");
827 assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!");
829 // Merge MMOs from memory operations in the common block.
830 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
831 MBBICommon->setMemRefs(MBBICommon->mergeMemRefsWith(*MBBI));
832 // Drop undef flags if they aren't present in all merged instructions.
833 for (unsigned I = 0, E = MBBICommon->getNumOperands(); I != E; ++I) {
834 MachineOperand &MO = MBBICommon->getOperand(I);
835 if (MO.isReg() && MO.isUndef()) {
836 const MachineOperand &OtherMO = MBBI->getOperand(I);
837 if (!OtherMO.isUndef())
838 MO.setIsUndef(false);
847 void BranchFolder::mergeCommonTails(unsigned commonTailIndex) {
848 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
850 std::vector<MachineBasicBlock::iterator> NextCommonInsts(SameTails.size());
851 for (unsigned int i = 0 ; i != SameTails.size() ; ++i) {
852 if (i != commonTailIndex) {
853 NextCommonInsts[i] = SameTails[i].getTailStartPos();
854 mergeOperations(SameTails[i].getTailStartPos(), *MBB);
856 assert(SameTails[i].getTailStartPos() == MBB->begin() &&
857 "MBB is not a common tail only block");
861 for (auto &MI : *MBB) {
862 if (MI.isDebugValue())
864 DebugLoc DL = MI.getDebugLoc();
865 for (unsigned int i = 0 ; i < NextCommonInsts.size() ; i++) {
866 if (i == commonTailIndex)
869 auto &Pos = NextCommonInsts[i];
870 assert(Pos != SameTails[i].getBlock()->end() &&
871 "Reached BB end within common tail");
872 while (Pos->isDebugValue()) {
874 assert(Pos != SameTails[i].getBlock()->end() &&
875 "Reached BB end within common tail");
877 assert(MI.isIdenticalTo(*Pos) && "Expected matching MIIs!");
878 DL = DILocation::getMergedLocation(DL, Pos->getDebugLoc());
879 NextCommonInsts[i] = ++Pos;
885 LivePhysRegs NewLiveIns(*TRI);
886 computeLiveIns(NewLiveIns, *MBB);
889 // The flag merging may lead to some register uses no longer using the
890 // <undef> flag, add IMPLICIT_DEFs in the predecessors as necessary.
891 for (MachineBasicBlock *Pred : MBB->predecessors()) {
893 LiveRegs.addLiveOuts(*Pred);
894 MachineBasicBlock::iterator InsertBefore = Pred->getFirstTerminator();
895 for (unsigned Reg : NewLiveIns) {
896 if (!LiveRegs.available(*MRI, Reg))
899 BuildMI(*Pred, InsertBefore, DL, TII->get(TargetOpcode::IMPLICIT_DEF),
905 addLiveIns(*MBB, NewLiveIns);
909 // See if any of the blocks in MergePotentials (which all have SuccBB as a
910 // successor, or all have no successor if it is null) can be tail-merged.
911 // If there is a successor, any blocks in MergePotentials that are not
912 // tail-merged and are not immediately before Succ must have an unconditional
913 // branch to Succ added (but the predecessor/successor lists need no
914 // adjustment). The lone predecessor of Succ that falls through into Succ,
915 // if any, is given in PredBB.
916 // MinCommonTailLength - Except for the special cases below, tail-merge if
917 // there are at least this many instructions in common.
918 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
919 MachineBasicBlock *PredBB,
920 unsigned MinCommonTailLength) {
921 bool MadeChange = false;
923 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
924 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) dbgs()
925 << printMBBReference(*MergePotentials[i].getBlock())
926 << (i == e - 1 ? "" : ", ");
927 dbgs() << "\n"; if (SuccBB) {
928 dbgs() << " with successor " << printMBBReference(*SuccBB) << '\n';
930 dbgs() << " which has fall-through from "
931 << printMBBReference(*PredBB) << "\n";
932 } dbgs() << "Looking for common tails of at least "
933 << MinCommonTailLength << " instruction"
934 << (MinCommonTailLength == 1 ? "" : "s") << '\n';);
936 // Sort by hash value so that blocks with identical end sequences sort
938 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
940 // Walk through equivalence sets looking for actual exact matches.
941 while (MergePotentials.size() > 1) {
942 unsigned CurHash = MergePotentials.back().getHash();
944 // Build SameTails, identifying the set of blocks with this hash code
945 // and with the maximum number of instructions in common.
946 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
950 // If we didn't find any pair that has at least MinCommonTailLength
951 // instructions in common, remove all blocks with this hash code and retry.
952 if (SameTails.empty()) {
953 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
957 // If one of the blocks is the entire common tail (and not the entry
958 // block, which we can't jump to), we can treat all blocks with this same
959 // tail at once. Use PredBB if that is one of the possibilities, as that
960 // will not introduce any extra branches.
961 MachineBasicBlock *EntryBB =
962 &MergePotentials.front().getBlock()->getParent()->front();
963 unsigned commonTailIndex = SameTails.size();
964 // If there are two blocks, check to see if one can be made to fall through
966 if (SameTails.size() == 2 &&
967 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
968 SameTails[1].tailIsWholeBlock())
970 else if (SameTails.size() == 2 &&
971 SameTails[1].getBlock()->isLayoutSuccessor(
972 SameTails[0].getBlock()) &&
973 SameTails[0].tailIsWholeBlock())
976 // Otherwise just pick one, favoring the fall-through predecessor if
978 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
979 MachineBasicBlock *MBB = SameTails[i].getBlock();
980 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
986 if (SameTails[i].tailIsWholeBlock())
991 if (commonTailIndex == SameTails.size() ||
992 (SameTails[commonTailIndex].getBlock() == PredBB &&
993 !SameTails[commonTailIndex].tailIsWholeBlock())) {
994 // None of the blocks consist entirely of the common tail.
995 // Split a block so that one does.
996 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
997 maxCommonTailLength, commonTailIndex)) {
998 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
1003 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
1005 // Recompute common tail MBB's edge weights and block frequency.
1006 setCommonTailEdgeWeights(*MBB);
1008 // Merge debug locations, MMOs and undef flags across identical instructions
1010 mergeCommonTails(commonTailIndex);
1012 // MBB is common tail. Adjust all other BB's to jump to this one.
1013 // Traversal must be forwards so erases work.
1014 DEBUG(dbgs() << "\nUsing common tail in " << printMBBReference(*MBB)
1016 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
1017 if (commonTailIndex == i)
1019 DEBUG(dbgs() << printMBBReference(*SameTails[i].getBlock())
1020 << (i == e - 1 ? "" : ", "));
1021 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
1022 replaceTailWithBranchTo(SameTails[i].getTailStartPos(), *MBB);
1023 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
1024 MergePotentials.erase(SameTails[i].getMPIter());
1026 DEBUG(dbgs() << "\n");
1027 // We leave commonTailIndex in the worklist in case there are other blocks
1028 // that match it with a smaller number of instructions.
1034 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
1035 bool MadeChange = false;
1036 if (!EnableTailMerge) return MadeChange;
1038 // First find blocks with no successors.
1039 // Block placement does not create new tail merging opportunities for these
1041 if (!AfterBlockPlacement) {
1042 MergePotentials.clear();
1043 for (MachineBasicBlock &MBB : MF) {
1044 if (MergePotentials.size() == TailMergeThreshold)
1046 if (!TriedMerging.count(&MBB) && MBB.succ_empty())
1047 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB));
1050 // If this is a large problem, avoid visiting the same basic blocks
1052 if (MergePotentials.size() == TailMergeThreshold)
1053 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1054 TriedMerging.insert(MergePotentials[i].getBlock());
1056 // See if we can do any tail merging on those.
1057 if (MergePotentials.size() >= 2)
1058 MadeChange |= TryTailMergeBlocks(nullptr, nullptr, MinCommonTailLength);
1061 // Look at blocks (IBB) with multiple predecessors (PBB).
1062 // We change each predecessor to a canonical form, by
1063 // (1) temporarily removing any unconditional branch from the predecessor
1065 // (2) alter conditional branches so they branch to the other block
1066 // not IBB; this may require adding back an unconditional branch to IBB
1067 // later, where there wasn't one coming in. E.g.
1069 // fallthrough to QBB
1072 // with a conceptual B to IBB after that, which never actually exists.
1073 // With those changes, we see whether the predecessors' tails match,
1074 // and merge them if so. We change things out of canonical form and
1075 // back to the way they were later in the process. (OptimizeBranches
1076 // would undo some of this, but we can't use it, because we'd get into
1077 // a compile-time infinite loop repeatedly doing and undoing the same
1078 // transformations.)
1080 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1082 if (I->pred_size() < 2) continue;
1083 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
1084 MachineBasicBlock *IBB = &*I;
1085 MachineBasicBlock *PredBB = &*std::prev(I);
1086 MergePotentials.clear();
1089 // Bail if merging after placement and IBB is the loop header because
1090 // -- If merging predecessors that belong to the same loop as IBB, the
1091 // common tail of merged predecessors may become the loop top if block
1092 // placement is called again and the predecessors may branch to this common
1093 // tail and require more branches. This can be relaxed if
1094 // MachineBlockPlacement::findBestLoopTop is more flexible.
1095 // --If merging predecessors that do not belong to the same loop as IBB, the
1096 // loop info of IBB's loop and the other loops may be affected. Calling the
1097 // block placement again may make big change to the layout and eliminate the
1098 // reason to do tail merging here.
1099 if (AfterBlockPlacement && MLI) {
1100 ML = MLI->getLoopFor(IBB);
1101 if (ML && IBB == ML->getHeader())
1105 for (MachineBasicBlock *PBB : I->predecessors()) {
1106 if (MergePotentials.size() == TailMergeThreshold)
1109 if (TriedMerging.count(PBB))
1112 // Skip blocks that loop to themselves, can't tail merge these.
1116 // Visit each predecessor only once.
1117 if (!UniquePreds.insert(PBB).second)
1120 // Skip blocks which may jump to a landing pad. Can't tail merge these.
1121 if (PBB->hasEHPadSuccessor())
1124 // After block placement, only consider predecessors that belong to the
1125 // same loop as IBB. The reason is the same as above when skipping loop
1127 if (AfterBlockPlacement && MLI)
1128 if (ML != MLI->getLoopFor(PBB))
1131 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1132 SmallVector<MachineOperand, 4> Cond;
1133 if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) {
1134 // Failing case: IBB is the target of a cbr, and we cannot reverse the
1136 SmallVector<MachineOperand, 4> NewCond(Cond);
1137 if (!Cond.empty() && TBB == IBB) {
1138 if (TII->reverseBranchCondition(NewCond))
1140 // This is the QBB case described above
1142 auto Next = ++PBB->getIterator();
1143 if (Next != MF.end())
1148 // Failing case: the only way IBB can be reached from PBB is via
1149 // exception handling. Happens for landing pads. Would be nice to have
1150 // a bit in the edge so we didn't have to do all this.
1151 if (IBB->isEHPad()) {
1152 MachineFunction::iterator IP = ++PBB->getIterator();
1153 MachineBasicBlock *PredNextBB = nullptr;
1157 if (IBB != PredNextBB) // fallthrough
1160 if (TBB != IBB && FBB != IBB) // cbr then ubr
1162 } else if (Cond.empty()) {
1163 if (TBB != IBB) // ubr
1166 if (TBB != IBB && IBB != PredNextBB) // cbr
1171 // Remove the unconditional branch at the end, if any.
1172 if (TBB && (Cond.empty() || FBB)) {
1173 DebugLoc dl = PBB->findBranchDebugLoc();
1174 TII->removeBranch(*PBB);
1176 // reinsert conditional branch only, for now
1177 TII->insertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1181 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB));
1185 // If this is a large problem, avoid visiting the same basic blocks multiple
1187 if (MergePotentials.size() == TailMergeThreshold)
1188 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1189 TriedMerging.insert(MergePotentials[i].getBlock());
1191 if (MergePotentials.size() >= 2)
1192 MadeChange |= TryTailMergeBlocks(IBB, PredBB, MinCommonTailLength);
1194 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1195 // result of removing blocks in TryTailMergeBlocks.
1196 PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks
1197 if (MergePotentials.size() == 1 &&
1198 MergePotentials.begin()->getBlock() != PredBB)
1199 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1205 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1206 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1207 BlockFrequency AccumulatedMBBFreq;
1209 // Aggregate edge frequency of successor edge j:
1210 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1211 // where bb is a basic block that is in SameTails.
1212 for (const auto &Src : SameTails) {
1213 const MachineBasicBlock *SrcMBB = Src.getBlock();
1214 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1215 AccumulatedMBBFreq += BlockFreq;
1217 // It is not necessary to recompute edge weights if TailBB has less than two
1219 if (TailMBB.succ_size() <= 1)
1222 auto EdgeFreq = EdgeFreqLs.begin();
1224 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1225 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1226 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1229 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1231 if (TailMBB.succ_size() <= 1)
1235 std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0))
1237 auto EdgeFreq = EdgeFreqLs.begin();
1239 if (SumEdgeFreq > 0) {
1240 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1241 SuccI != SuccE; ++SuccI, ++EdgeFreq) {
1242 auto Prob = BranchProbability::getBranchProbability(
1243 EdgeFreq->getFrequency(), SumEdgeFreq);
1244 TailMBB.setSuccProbability(SuccI, Prob);
1249 //===----------------------------------------------------------------------===//
1250 // Branch Optimization
1251 //===----------------------------------------------------------------------===//
1253 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1254 bool MadeChange = false;
1256 // Make sure blocks are numbered in order
1257 MF.RenumberBlocks();
1258 // Renumbering blocks alters funclet membership, recalculate it.
1259 FuncletMembership = getFuncletMembership(MF);
1261 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1263 MachineBasicBlock *MBB = &*I++;
1264 MadeChange |= OptimizeBlock(MBB);
1266 // If it is dead, remove it.
1267 if (MBB->pred_empty()) {
1268 RemoveDeadBlock(MBB);
1277 // Blocks should be considered empty if they contain only debug info;
1278 // else the debug info would affect codegen.
1279 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1280 return MBB->getFirstNonDebugInstr() == MBB->end();
1283 // Blocks with only debug info and branches should be considered the same
1284 // as blocks with only branches.
1285 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1286 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1287 assert(I != MBB->end() && "empty block!");
1288 return I->isBranch();
1291 /// IsBetterFallthrough - Return true if it would be clearly better to
1292 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1293 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1294 /// result in infinite loops.
1295 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1296 MachineBasicBlock *MBB2) {
1297 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1298 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1299 // optimize branches that branch to either a return block or an assert block
1300 // into a fallthrough to the return.
1301 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1302 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1303 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1306 // If there is a clear successor ordering we make sure that one block
1307 // will fall through to the next
1308 if (MBB1->isSuccessor(MBB2)) return true;
1309 if (MBB2->isSuccessor(MBB1)) return false;
1311 return MBB2I->isCall() && !MBB1I->isCall();
1314 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1315 /// instructions on the block.
1316 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1317 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1318 if (I != MBB.end() && I->isBranch())
1319 return I->getDebugLoc();
1323 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1324 bool MadeChange = false;
1325 MachineFunction &MF = *MBB->getParent();
1328 MachineFunction::iterator FallThrough = MBB->getIterator();
1331 // Make sure MBB and FallThrough belong to the same funclet.
1332 bool SameFunclet = true;
1333 if (!FuncletMembership.empty() && FallThrough != MF.end()) {
1334 auto MBBFunclet = FuncletMembership.find(MBB);
1335 assert(MBBFunclet != FuncletMembership.end());
1336 auto FallThroughFunclet = FuncletMembership.find(&*FallThrough);
1337 assert(FallThroughFunclet != FuncletMembership.end());
1338 SameFunclet = MBBFunclet->second == FallThroughFunclet->second;
1341 // If this block is empty, make everyone use its fall-through, not the block
1342 // explicitly. Landing pads should not do this since the landing-pad table
1343 // points to this block. Blocks with their addresses taken shouldn't be
1345 if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() &&
1347 // Dead block? Leave for cleanup later.
1348 if (MBB->pred_empty()) return MadeChange;
1350 if (FallThrough == MF.end()) {
1351 // TODO: Simplify preds to not branch here if possible!
1352 } else if (FallThrough->isEHPad()) {
1353 // Don't rewrite to a landing pad fallthough. That could lead to the case
1354 // where a BB jumps to more than one landing pad.
1355 // TODO: Is it ever worth rewriting predecessors which don't already
1356 // jump to a landing pad, and so can safely jump to the fallthrough?
1357 } else if (MBB->isSuccessor(&*FallThrough)) {
1358 // Rewrite all predecessors of the old block to go to the fallthrough
1360 while (!MBB->pred_empty()) {
1361 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1362 Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough);
1364 // If MBB was the target of a jump table, update jump tables to go to the
1365 // fallthrough instead.
1366 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1367 MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough);
1373 // Check to see if we can simplify the terminator of the block before this
1375 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1377 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1378 SmallVector<MachineOperand, 4> PriorCond;
1379 bool PriorUnAnalyzable =
1380 TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1381 if (!PriorUnAnalyzable) {
1382 // If the CFG for the prior block has extra edges, remove them.
1383 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1384 !PriorCond.empty());
1386 // If the previous branch is conditional and both conditions go to the same
1387 // destination, remove the branch, replacing it with an unconditional one or
1389 if (PriorTBB && PriorTBB == PriorFBB) {
1390 DebugLoc dl = getBranchDebugLoc(PrevBB);
1391 TII->removeBranch(PrevBB);
1393 if (PriorTBB != MBB)
1394 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1397 goto ReoptimizeBlock;
1400 // If the previous block unconditionally falls through to this block and
1401 // this block has no other predecessors, move the contents of this block
1402 // into the prior block. This doesn't usually happen when SimplifyCFG
1403 // has been used, but it can happen if tail merging splits a fall-through
1404 // predecessor of a block.
1405 // This has to check PrevBB->succ_size() because EH edges are ignored by
1407 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1408 PrevBB.succ_size() == 1 &&
1409 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1410 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1411 << "From MBB: " << *MBB);
1412 // Remove redundant DBG_VALUEs first.
1413 if (PrevBB.begin() != PrevBB.end()) {
1414 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1416 MachineBasicBlock::iterator MBBIter = MBB->begin();
1417 // Check if DBG_VALUE at the end of PrevBB is identical to the
1418 // DBG_VALUE at the beginning of MBB.
1419 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1420 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1421 if (!MBBIter->isIdenticalTo(*PrevBBIter))
1423 MachineInstr &DuplicateDbg = *MBBIter;
1424 ++MBBIter; -- PrevBBIter;
1425 DuplicateDbg.eraseFromParent();
1428 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1429 PrevBB.removeSuccessor(PrevBB.succ_begin());
1430 assert(PrevBB.succ_empty());
1431 PrevBB.transferSuccessors(MBB);
1436 // If the previous branch *only* branches to *this* block (conditional or
1437 // not) remove the branch.
1438 if (PriorTBB == MBB && !PriorFBB) {
1439 TII->removeBranch(PrevBB);
1442 goto ReoptimizeBlock;
1445 // If the prior block branches somewhere else on the condition and here if
1446 // the condition is false, remove the uncond second branch.
1447 if (PriorFBB == MBB) {
1448 DebugLoc dl = getBranchDebugLoc(PrevBB);
1449 TII->removeBranch(PrevBB);
1450 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1453 goto ReoptimizeBlock;
1456 // If the prior block branches here on true and somewhere else on false, and
1457 // if the branch condition is reversible, reverse the branch to create a
1459 if (PriorTBB == MBB) {
1460 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1461 if (!TII->reverseBranchCondition(NewPriorCond)) {
1462 DebugLoc dl = getBranchDebugLoc(PrevBB);
1463 TII->removeBranch(PrevBB);
1464 TII->insertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1467 goto ReoptimizeBlock;
1471 // If this block has no successors (e.g. it is a return block or ends with
1472 // a call to a no-return function like abort or __cxa_throw) and if the pred
1473 // falls through into this block, and if it would otherwise fall through
1474 // into the block after this, move this block to the end of the function.
1476 // We consider it more likely that execution will stay in the function (e.g.
1477 // due to loops) than it is to exit it. This asserts in loops etc, moving
1478 // the assert condition out of the loop body.
1479 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1480 MachineFunction::iterator(PriorTBB) == FallThrough &&
1481 !MBB->canFallThrough()) {
1482 bool DoTransform = true;
1484 // We have to be careful that the succs of PredBB aren't both no-successor
1485 // blocks. If neither have successors and if PredBB is the second from
1486 // last block in the function, we'd just keep swapping the two blocks for
1487 // last. Only do the swap if one is clearly better to fall through than
1489 if (FallThrough == --MF.end() &&
1490 !IsBetterFallthrough(PriorTBB, MBB))
1491 DoTransform = false;
1494 // Reverse the branch so we will fall through on the previous true cond.
1495 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1496 if (!TII->reverseBranchCondition(NewPriorCond)) {
1497 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1498 << "To make fallthrough to: " << *PriorTBB << "\n");
1500 DebugLoc dl = getBranchDebugLoc(PrevBB);
1501 TII->removeBranch(PrevBB);
1502 TII->insertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1504 // Move this block to the end of the function.
1505 MBB->moveAfter(&MF.back());
1514 if (!IsEmptyBlock(MBB) && MBB->pred_size() == 1 &&
1515 MF.getFunction().optForSize()) {
1516 // Changing "Jcc foo; foo: jmp bar;" into "Jcc bar;" might change the branch
1517 // direction, thereby defeating careful block placement and regressing
1518 // performance. Therefore, only consider this for optsize functions.
1519 MachineInstr &TailCall = *MBB->getFirstNonDebugInstr();
1520 if (TII->isUnconditionalTailCall(TailCall)) {
1521 MachineBasicBlock *Pred = *MBB->pred_begin();
1522 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1523 SmallVector<MachineOperand, 4> PredCond;
1524 bool PredAnalyzable =
1525 !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true);
1527 if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB &&
1528 PredTBB != PredFBB) {
1529 // The predecessor has a conditional branch to this block which consists
1530 // of only a tail call. Try to fold the tail call into the conditional
1532 if (TII->canMakeTailCallConditional(PredCond, TailCall)) {
1533 // TODO: It would be nice if analyzeBranch() could provide a pointer
1534 // to the branch instruction so replaceBranchWithTailCall() doesn't
1535 // have to search for it.
1536 TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall);
1538 Pred->removeSuccessor(MBB);
1543 // If the predecessor is falling through to this block, we could reverse
1544 // the branch condition and fold the tail call into that. However, after
1545 // that we might have to re-arrange the CFG to fall through to the other
1546 // block and there is a high risk of regressing code size rather than
1551 // Analyze the branch in the current block.
1552 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1553 SmallVector<MachineOperand, 4> CurCond;
1554 bool CurUnAnalyzable =
1555 TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1556 if (!CurUnAnalyzable) {
1557 // If the CFG for the prior block has extra edges, remove them.
1558 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1560 // If this is a two-way branch, and the FBB branches to this block, reverse
1561 // the condition so the single-basic-block loop is faster. Instead of:
1562 // Loop: xxx; jcc Out; jmp Loop
1564 // Loop: xxx; jncc Loop; jmp Out
1565 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1566 SmallVector<MachineOperand, 4> NewCond(CurCond);
1567 if (!TII->reverseBranchCondition(NewCond)) {
1568 DebugLoc dl = getBranchDebugLoc(*MBB);
1569 TII->removeBranch(*MBB);
1570 TII->insertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1573 goto ReoptimizeBlock;
1577 // If this branch is the only thing in its block, see if we can forward
1578 // other blocks across it.
1579 if (CurTBB && CurCond.empty() && !CurFBB &&
1580 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1581 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1582 DebugLoc dl = getBranchDebugLoc(*MBB);
1583 // This block may contain just an unconditional branch. Because there can
1584 // be 'non-branch terminators' in the block, try removing the branch and
1585 // then seeing if the block is empty.
1586 TII->removeBranch(*MBB);
1587 // If the only things remaining in the block are debug info, remove these
1588 // as well, so this will behave the same as an empty block in non-debug
1590 if (IsEmptyBlock(MBB)) {
1591 // Make the block empty, losing the debug info (we could probably
1592 // improve this in some cases.)
1593 MBB->erase(MBB->begin(), MBB->end());
1595 // If this block is just an unconditional branch to CurTBB, we can
1596 // usually completely eliminate the block. The only case we cannot
1597 // completely eliminate the block is when the block before this one
1598 // falls through into MBB and we can't understand the prior block's branch
1601 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1602 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1603 !PrevBB.isSuccessor(MBB)) {
1604 // If the prior block falls through into us, turn it into an
1605 // explicit branch to us to make updates simpler.
1606 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1607 PriorTBB != MBB && PriorFBB != MBB) {
1609 assert(PriorCond.empty() && !PriorFBB &&
1610 "Bad branch analysis");
1613 assert(!PriorFBB && "Machine CFG out of date!");
1616 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1617 TII->removeBranch(PrevBB);
1618 TII->insertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1621 // Iterate through all the predecessors, revectoring each in-turn.
1623 bool DidChange = false;
1624 bool HasBranchToSelf = false;
1625 while(PI != MBB->pred_size()) {
1626 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1628 // If this block has an uncond branch to itself, leave it.
1630 HasBranchToSelf = true;
1633 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1634 // If this change resulted in PMBB ending in a conditional
1635 // branch where both conditions go to the same destination,
1636 // change this to an unconditional branch (and fix the CFG).
1637 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1638 SmallVector<MachineOperand, 4> NewCurCond;
1639 bool NewCurUnAnalyzable = TII->analyzeBranch(
1640 *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true);
1641 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1642 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1643 TII->removeBranch(*PMBB);
1645 TII->insertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1648 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1653 // Change any jumptables to go to the new MBB.
1654 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1655 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1659 if (!HasBranchToSelf) return MadeChange;
1664 // Add the branch back if the block is more than just an uncond branch.
1665 TII->insertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1669 // If the prior block doesn't fall through into this block, and if this
1670 // block doesn't fall through into some other block, see if we can find a
1671 // place to move this block where a fall-through will happen.
1672 if (!PrevBB.canFallThrough()) {
1673 // Now we know that there was no fall-through into this block, check to
1674 // see if it has a fall-through into its successor.
1675 bool CurFallsThru = MBB->canFallThrough();
1677 if (!MBB->isEHPad()) {
1678 // Check all the predecessors of this block. If one of them has no fall
1679 // throughs, move this block right after it.
1680 for (MachineBasicBlock *PredBB : MBB->predecessors()) {
1681 // Analyze the branch at the end of the pred.
1682 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1683 SmallVector<MachineOperand, 4> PredCond;
1684 if (PredBB != MBB && !PredBB->canFallThrough() &&
1685 !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) &&
1686 (!CurFallsThru || !CurTBB || !CurFBB) &&
1687 (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1688 // If the current block doesn't fall through, just move it.
1689 // If the current block can fall through and does not end with a
1690 // conditional branch, we need to append an unconditional jump to
1691 // the (current) next block. To avoid a possible compile-time
1692 // infinite loop, move blocks only backward in this case.
1693 // Also, if there are already 2 branches here, we cannot add a third;
1694 // this means we have the case
1699 MachineBasicBlock *NextBB = &*std::next(MBB->getIterator());
1701 TII->insertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1703 MBB->moveAfter(PredBB);
1705 goto ReoptimizeBlock;
1710 if (!CurFallsThru) {
1711 // Check all successors to see if we can move this block before it.
1712 for (MachineBasicBlock *SuccBB : MBB->successors()) {
1713 // Analyze the branch at the end of the block before the succ.
1714 MachineFunction::iterator SuccPrev = --SuccBB->getIterator();
1716 // If this block doesn't already fall-through to that successor, and if
1717 // the succ doesn't already have a block that can fall through into it,
1718 // and if the successor isn't an EH destination, we can arrange for the
1719 // fallthrough to happen.
1720 if (SuccBB != MBB && &*SuccPrev != MBB &&
1721 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1722 !SuccBB->isEHPad()) {
1723 MBB->moveBefore(SuccBB);
1725 goto ReoptimizeBlock;
1729 // Okay, there is no really great place to put this block. If, however,
1730 // the block before this one would be a fall-through if this block were
1731 // removed, move this block to the end of the function. There is no real
1732 // advantage in "falling through" to an EH block, so we don't want to
1733 // perform this transformation for that case.
1735 // Also, Windows EH introduced the possibility of an arbitrary number of
1736 // successors to a given block. The analyzeBranch call does not consider
1737 // exception handling and so we can get in a state where a block
1738 // containing a call is followed by multiple EH blocks that would be
1739 // rotated infinitely at the end of the function if the transformation
1740 // below were performed for EH "FallThrough" blocks. Therefore, even if
1741 // that appears not to be happening anymore, we should assume that it is
1742 // possible and not remove the "!FallThrough()->isEHPad" condition below.
1743 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1744 SmallVector<MachineOperand, 4> PrevCond;
1745 if (FallThrough != MF.end() &&
1746 !FallThrough->isEHPad() &&
1747 !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1748 PrevBB.isSuccessor(&*FallThrough)) {
1749 MBB->moveAfter(&MF.back());
1759 //===----------------------------------------------------------------------===//
1760 // Hoist Common Code
1761 //===----------------------------------------------------------------------===//
1763 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1764 bool MadeChange = false;
1765 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1766 MachineBasicBlock *MBB = &*I++;
1767 MadeChange |= HoistCommonCodeInSuccs(MBB);
1773 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1774 /// its 'true' successor.
1775 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1776 MachineBasicBlock *TrueBB) {
1777 for (MachineBasicBlock *SuccBB : BB->successors())
1778 if (SuccBB != TrueBB)
1783 template <class Container>
1784 static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
1786 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1787 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1794 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1795 /// in successors to. The location is usually just before the terminator,
1796 /// however if the terminator is a conditional branch and its previous
1797 /// instruction is the flag setting instruction, the previous instruction is
1798 /// the preferred location. This function also gathers uses and defs of the
1799 /// instructions from the insertion point to the end of the block. The data is
1800 /// used by HoistCommonCodeInSuccs to ensure safety.
1802 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1803 const TargetInstrInfo *TII,
1804 const TargetRegisterInfo *TRI,
1805 SmallSet<unsigned,4> &Uses,
1806 SmallSet<unsigned,4> &Defs) {
1807 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1808 if (!TII->isUnpredicatedTerminator(*Loc))
1811 for (const MachineOperand &MO : Loc->operands()) {
1814 unsigned Reg = MO.getReg();
1818 addRegAndItsAliases(Reg, TRI, Uses);
1821 // Don't try to hoist code in the rare case the terminator defines a
1822 // register that is later used.
1825 // If the terminator defines a register, make sure we don't hoist
1826 // the instruction whose def might be clobbered by the terminator.
1827 addRegAndItsAliases(Reg, TRI, Defs);
1833 if (Loc == MBB->begin())
1836 // The terminator is probably a conditional branch, try not to separate the
1837 // branch from condition setting instruction.
1838 MachineBasicBlock::iterator PI =
1839 skipDebugInstructionsBackward(std::prev(Loc), MBB->begin());
1842 for (const MachineOperand &MO : PI->operands()) {
1843 // If PI has a regmask operand, it is probably a call. Separate away.
1846 if (!MO.isReg() || MO.isUse())
1848 unsigned Reg = MO.getReg();
1851 if (Uses.count(Reg)) {
1857 // The condition setting instruction is not just before the conditional
1861 // Be conservative, don't insert instruction above something that may have
1862 // side-effects. And since it's potentially bad to separate flag setting
1863 // instruction from the conditional branch, just abort the optimization
1865 // Also avoid moving code above predicated instruction since it's hard to
1866 // reason about register liveness with predicated instruction.
1867 bool DontMoveAcrossStore = true;
1868 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI))
1871 // Find out what registers are live. Note this routine is ignoring other live
1872 // registers which are only used by instructions in successor blocks.
1873 for (const MachineOperand &MO : PI->operands()) {
1876 unsigned Reg = MO.getReg();
1880 addRegAndItsAliases(Reg, TRI, Uses);
1882 if (Uses.erase(Reg)) {
1883 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1884 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1885 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1888 addRegAndItsAliases(Reg, TRI, Defs);
1895 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1896 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1897 SmallVector<MachineOperand, 4> Cond;
1898 if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1901 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1903 // Malformed bcc? True and false blocks are the same?
1906 // Restrict the optimization to cases where MBB is the only predecessor,
1907 // it is an obvious win.
1908 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1911 // Find a suitable position to hoist the common instructions to. Also figure
1912 // out which registers are used or defined by instructions from the insertion
1913 // point to the end of the block.
1914 SmallSet<unsigned, 4> Uses, Defs;
1915 MachineBasicBlock::iterator Loc =
1916 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1917 if (Loc == MBB->end())
1920 bool HasDups = false;
1921 SmallVector<unsigned, 4> LocalDefs, LocalKills;
1922 SmallSet<unsigned, 4> ActiveDefsSet, AllDefsSet;
1923 MachineBasicBlock::iterator TIB = TBB->begin();
1924 MachineBasicBlock::iterator FIB = FBB->begin();
1925 MachineBasicBlock::iterator TIE = TBB->end();
1926 MachineBasicBlock::iterator FIE = FBB->end();
1927 while (TIB != TIE && FIB != FIE) {
1928 // Skip dbg_value instructions. These do not count.
1929 TIB = skipDebugInstructionsForward(TIB, TIE);
1930 FIB = skipDebugInstructionsForward(FIB, FIE);
1931 if (TIB == TIE || FIB == FIE)
1934 if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead))
1937 if (TII->isPredicated(*TIB))
1938 // Hard to reason about register liveness with predicated instruction.
1942 for (MachineOperand &MO : TIB->operands()) {
1943 // Don't attempt to hoist instructions with register masks.
1944 if (MO.isRegMask()) {
1950 unsigned Reg = MO.getReg();
1954 if (Uses.count(Reg)) {
1955 // Avoid clobbering a register that's used by the instruction at
1956 // the point of insertion.
1961 if (Defs.count(Reg) && !MO.isDead()) {
1962 // Don't hoist the instruction if the def would be clobber by the
1963 // instruction at the point insertion. FIXME: This is overly
1964 // conservative. It should be possible to hoist the instructions
1965 // in BB2 in the following example:
1967 // r1, eflag = op1 r2, r3
1976 } else if (!ActiveDefsSet.count(Reg)) {
1977 if (Defs.count(Reg)) {
1978 // Use is defined by the instruction at the point of insertion.
1983 if (MO.isKill() && Uses.count(Reg))
1984 // Kills a register that's read by the instruction at the point of
1985 // insertion. Remove the kill marker.
1986 MO.setIsKill(false);
1992 bool DontMoveAcrossStore = true;
1993 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1996 // Remove kills from ActiveDefsSet, these registers had short live ranges.
1997 for (const MachineOperand &MO : TIB->operands()) {
1998 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
2000 unsigned Reg = MO.getReg();
2003 if (!AllDefsSet.count(Reg)) {
2004 LocalKills.push_back(Reg);
2007 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
2008 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
2009 ActiveDefsSet.erase(*AI);
2011 ActiveDefsSet.erase(Reg);
2015 // Track local defs so we can update liveins.
2016 for (const MachineOperand &MO : TIB->operands()) {
2017 if (!MO.isReg() || !MO.isDef() || MO.isDead())
2019 unsigned Reg = MO.getReg();
2020 if (!Reg || TargetRegisterInfo::isVirtualRegister(Reg))
2022 LocalDefs.push_back(Reg);
2023 addRegAndItsAliases(Reg, TRI, ActiveDefsSet);
2024 addRegAndItsAliases(Reg, TRI, AllDefsSet);
2035 MBB->splice(Loc, TBB, TBB->begin(), TIB);
2036 FBB->erase(FBB->begin(), FIB);
2039 bool ChangedLiveIns = false;
2040 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
2041 unsigned Def = LocalDefs[i];
2042 if (ActiveDefsSet.count(Def)) {
2043 TBB->addLiveIn(Def);
2044 FBB->addLiveIn(Def);
2045 ChangedLiveIns = true;
2048 for (unsigned K : LocalKills) {
2049 TBB->removeLiveIn(K);
2050 FBB->removeLiveIn(K);
2051 ChangedLiveIns = true;
2054 if (ChangedLiveIns) {
2055 TBB->sortUniqueLiveIns();
2056 FBB->sortUniqueLiveIns();