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/STLExtras.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/CodeGen/Analysis.h"
25 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
26 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
27 #include "llvm/CodeGen/MachineFunctionPass.h"
28 #include "llvm/CodeGen/MachineJumpTableInfo.h"
29 #include "llvm/CodeGen/MachineMemOperand.h"
30 #include "llvm/CodeGen/MachineLoopInfo.h"
31 #include "llvm/CodeGen/MachineModuleInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/Passes.h"
34 #include "llvm/CodeGen/TargetPassConfig.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Target/TargetInstrInfo.h"
41 #include "llvm/Target/TargetRegisterInfo.h"
42 #include "llvm/Target/TargetSubtargetInfo.h"
46 #define DEBUG_TYPE "branchfolding"
48 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
49 STATISTIC(NumBranchOpts, "Number of branches optimized");
50 STATISTIC(NumTailMerge , "Number of block tails merged");
51 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
53 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
54 cl::init(cl::BOU_UNSET), cl::Hidden);
56 // Throttle for huge numbers of predecessors (compile speed problems)
57 static cl::opt<unsigned>
58 TailMergeThreshold("tail-merge-threshold",
59 cl::desc("Max number of predecessors to consider tail merging"),
60 cl::init(150), cl::Hidden);
62 // Heuristic for tail merging (and, inversely, tail duplication).
63 // TODO: This should be replaced with a target query.
64 static cl::opt<unsigned>
65 TailMergeSize("tail-merge-size",
66 cl::desc("Min number of instructions to consider tail merging"),
67 cl::init(3), cl::Hidden);
70 /// BranchFolderPass - Wrap branch folder in a machine function pass.
71 class BranchFolderPass : public MachineFunctionPass {
74 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
76 bool runOnMachineFunction(MachineFunction &MF) override;
78 void getAnalysisUsage(AnalysisUsage &AU) const override {
79 AU.addRequired<MachineBlockFrequencyInfo>();
80 AU.addRequired<MachineBranchProbabilityInfo>();
81 AU.addRequired<TargetPassConfig>();
82 MachineFunctionPass::getAnalysisUsage(AU);
87 char BranchFolderPass::ID = 0;
88 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
90 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
91 "Control Flow Optimizer", false, false)
93 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
94 if (skipFunction(*MF.getFunction()))
97 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
98 // TailMerge can create jump into if branches that make CFG irreducible for
99 // HW that requires structurized CFG.
100 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
101 PassConfig->getEnableTailMerge();
102 BranchFolder::MBFIWrapper MBBFreqInfo(
103 getAnalysis<MachineBlockFrequencyInfo>());
104 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo,
105 getAnalysis<MachineBranchProbabilityInfo>());
106 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
107 MF.getSubtarget().getRegisterInfo(),
108 getAnalysisIfAvailable<MachineModuleInfo>());
111 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
112 MBFIWrapper &FreqInfo,
113 const MachineBranchProbabilityInfo &ProbInfo)
114 : EnableHoistCommonCode(CommonHoist), MBBFreqInfo(FreqInfo),
116 switch (FlagEnableTailMerge) {
117 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
118 case cl::BOU_TRUE: EnableTailMerge = true; break;
119 case cl::BOU_FALSE: EnableTailMerge = false; break;
123 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
124 /// function, updating the CFG.
125 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
126 assert(MBB->pred_empty() && "MBB must be dead!");
127 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
129 MachineFunction *MF = MBB->getParent();
130 // drop all successors.
131 while (!MBB->succ_empty())
132 MBB->removeSuccessor(MBB->succ_end()-1);
134 // Avoid matching if this pointer gets reused.
135 TriedMerging.erase(MBB);
139 FuncletMembership.erase(MBB);
141 MLI->removeBlock(MBB);
144 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
145 /// followed by terminators, and if the implicitly defined registers are not
146 /// used by the terminators, remove those implicit_def's. e.g.
148 /// r0 = implicit_def
149 /// r1 = implicit_def
151 /// This block can be optimized away later if the implicit instructions are
153 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
154 SmallSet<unsigned, 4> ImpDefRegs;
155 MachineBasicBlock::iterator I = MBB->begin();
156 while (I != MBB->end()) {
157 if (!I->isImplicitDef())
159 unsigned Reg = I->getOperand(0).getReg();
160 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
161 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
162 SubRegs.isValid(); ++SubRegs)
163 ImpDefRegs.insert(*SubRegs);
165 ImpDefRegs.insert(Reg);
169 if (ImpDefRegs.empty())
172 MachineBasicBlock::iterator FirstTerm = I;
173 while (I != MBB->end()) {
174 if (!TII->isUnpredicatedTerminator(*I))
176 // See if it uses any of the implicitly defined registers.
177 for (const MachineOperand &MO : I->operands()) {
178 if (!MO.isReg() || !MO.isUse())
180 unsigned Reg = MO.getReg();
181 if (ImpDefRegs.count(Reg))
188 while (I != FirstTerm) {
189 MachineInstr *ImpDefMI = &*I;
191 MBB->erase(ImpDefMI);
197 /// OptimizeFunction - Perhaps branch folding, tail merging and other
198 /// CFG optimizations on the given function. Block placement changes the layout
199 /// and may create new tail merging opportunities.
200 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
201 const TargetInstrInfo *tii,
202 const TargetRegisterInfo *tri,
203 MachineModuleInfo *mmi,
204 MachineLoopInfo *mli, bool AfterPlacement) {
205 if (!tii) return false;
207 TriedMerging.clear();
209 AfterBlockPlacement = AfterPlacement;
215 MachineRegisterInfo &MRI = MF.getRegInfo();
216 UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF);
218 MRI.invalidateLiveness();
220 // Fix CFG. The later algorithms expect it to be right.
221 bool MadeChange = false;
222 for (MachineBasicBlock &MBB : MF) {
223 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
224 SmallVector<MachineOperand, 4> Cond;
225 if (!TII->analyzeBranch(MBB, TBB, FBB, Cond, true))
226 MadeChange |= MBB.CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
227 MadeChange |= OptimizeImpDefsBlock(&MBB);
230 // Recalculate funclet membership.
231 FuncletMembership = getFuncletMembership(MF);
233 bool MadeChangeThisIteration = true;
234 while (MadeChangeThisIteration) {
235 MadeChangeThisIteration = TailMergeBlocks(MF);
236 // No need to clean up if tail merging does not change anything after the
238 if (!AfterBlockPlacement || MadeChangeThisIteration)
239 MadeChangeThisIteration |= OptimizeBranches(MF);
240 if (EnableHoistCommonCode)
241 MadeChangeThisIteration |= HoistCommonCode(MF);
242 MadeChange |= MadeChangeThisIteration;
245 // See if any jump tables have become dead as the code generator
247 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
251 // Walk the function to find jump tables that are live.
252 BitVector JTIsLive(JTI->getJumpTables().size());
253 for (const MachineBasicBlock &BB : MF) {
254 for (const MachineInstr &I : BB)
255 for (const MachineOperand &Op : I.operands()) {
256 if (!Op.isJTI()) continue;
258 // Remember that this JT is live.
259 JTIsLive.set(Op.getIndex());
263 // Finally, remove dead jump tables. This happens when the
264 // indirect jump was unreachable (and thus deleted).
265 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
266 if (!JTIsLive.test(i)) {
267 JTI->RemoveJumpTable(i);
274 //===----------------------------------------------------------------------===//
275 // Tail Merging of Blocks
276 //===----------------------------------------------------------------------===//
278 /// HashMachineInstr - Compute a hash value for MI and its operands.
279 static unsigned HashMachineInstr(const MachineInstr &MI) {
280 unsigned Hash = MI.getOpcode();
281 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
282 const MachineOperand &Op = MI.getOperand(i);
284 // Merge in bits from the operand if easy. We can't use MachineOperand's
285 // hash_code here because it's not deterministic and we sort by hash value
287 unsigned OperandHash = 0;
288 switch (Op.getType()) {
289 case MachineOperand::MO_Register:
290 OperandHash = Op.getReg();
292 case MachineOperand::MO_Immediate:
293 OperandHash = Op.getImm();
295 case MachineOperand::MO_MachineBasicBlock:
296 OperandHash = Op.getMBB()->getNumber();
298 case MachineOperand::MO_FrameIndex:
299 case MachineOperand::MO_ConstantPoolIndex:
300 case MachineOperand::MO_JumpTableIndex:
301 OperandHash = Op.getIndex();
303 case MachineOperand::MO_GlobalAddress:
304 case MachineOperand::MO_ExternalSymbol:
305 // Global address / external symbol are too hard, don't bother, but do
306 // pull in the offset.
307 OperandHash = Op.getOffset();
313 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
318 /// HashEndOfMBB - Hash the last instruction in the MBB.
319 static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) {
320 MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr();
324 return HashMachineInstr(*I);
327 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
328 /// of instructions they actually have in common together at their end. Return
329 /// iterators for the first shared instruction in each block.
330 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
331 MachineBasicBlock *MBB2,
332 MachineBasicBlock::iterator &I1,
333 MachineBasicBlock::iterator &I2) {
337 unsigned TailLen = 0;
338 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
340 // Skip debugging pseudos; necessary to avoid changing the code.
341 while (I1->isDebugValue()) {
342 if (I1==MBB1->begin()) {
343 while (I2->isDebugValue()) {
344 if (I2==MBB2->begin())
345 // I1==DBG at begin; I2==DBG at begin
350 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
355 // I1==first (untested) non-DBG preceding known match
356 while (I2->isDebugValue()) {
357 if (I2==MBB2->begin()) {
359 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
364 // I1, I2==first (untested) non-DBGs preceding known match
365 if (!I1->isIdenticalTo(*I2) ||
366 // FIXME: This check is dubious. It's used to get around a problem where
367 // people incorrectly expect inline asm directives to remain in the same
368 // relative order. This is untenable because normal compiler
369 // optimizations (like this one) may reorder and/or merge these
377 // Back past possible debugging pseudos at beginning of block. This matters
378 // when one block differs from the other only by whether debugging pseudos
379 // are present at the beginning. (This way, the various checks later for
380 // I1==MBB1->begin() work as expected.)
381 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
383 while (I2->isDebugValue()) {
384 if (I2 == MBB2->begin())
390 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
392 while (I1->isDebugValue()) {
393 if (I1 == MBB1->begin())
402 void BranchFolder::computeLiveIns(MachineBasicBlock &MBB) {
407 LiveRegs.addLiveOutsNoPristines(MBB);
408 for (MachineInstr &MI : make_range(MBB.rbegin(), MBB.rend()))
409 LiveRegs.stepBackward(MI);
411 for (unsigned Reg : LiveRegs) {
412 // Skip the register if we are about to add one of its super registers.
413 bool ContainsSuperReg = false;
414 for (MCSuperRegIterator SReg(Reg, TRI); SReg.isValid(); ++SReg) {
415 if (LiveRegs.contains(*SReg)) {
416 ContainsSuperReg = true;
420 if (ContainsSuperReg)
426 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
427 /// after it, replacing it with an unconditional branch to NewDest.
428 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
429 MachineBasicBlock *NewDest) {
430 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
432 computeLiveIns(*NewDest);
437 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
438 /// MBB so that the part before the iterator falls into the part starting at the
439 /// iterator. This returns the new MBB.
440 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
441 MachineBasicBlock::iterator BBI1,
442 const BasicBlock *BB) {
443 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
446 MachineFunction &MF = *CurMBB.getParent();
448 // Create the fall-through block.
449 MachineFunction::iterator MBBI = CurMBB.getIterator();
450 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
451 CurMBB.getParent()->insert(++MBBI, NewMBB);
453 // Move all the successors of this block to the specified block.
454 NewMBB->transferSuccessors(&CurMBB);
456 // Add an edge from CurMBB to NewMBB for the fall-through.
457 CurMBB.addSuccessor(NewMBB);
459 // Splice the code over.
460 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
462 // NewMBB belongs to the same loop as CurMBB.
464 if (MachineLoop *ML = MLI->getLoopFor(&CurMBB))
465 ML->addBasicBlockToLoop(NewMBB, MLI->getBase());
467 // NewMBB inherits CurMBB's block frequency.
468 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
470 computeLiveIns(*NewMBB);
472 // Add the new block to the funclet.
473 const auto &FuncletI = FuncletMembership.find(&CurMBB);
474 if (FuncletI != FuncletMembership.end()) {
475 auto n = FuncletI->second;
476 FuncletMembership[NewMBB] = n;
482 /// EstimateRuntime - Make a rough estimate for how long it will take to run
483 /// the specified code.
484 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
485 MachineBasicBlock::iterator E) {
487 for (; I != E; ++I) {
488 if (I->isDebugValue())
492 else if (I->mayLoad() || I->mayStore())
500 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
501 // branches temporarily for tail merging). In the case where CurMBB ends
502 // with a conditional branch to the next block, optimize by reversing the
503 // test and conditionally branching to SuccMBB instead.
504 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
505 const TargetInstrInfo *TII) {
506 MachineFunction *MF = CurMBB->getParent();
507 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
508 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
509 SmallVector<MachineOperand, 4> Cond;
510 DebugLoc dl; // FIXME: this is nowhere
511 if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
512 MachineBasicBlock *NextBB = &*I;
513 if (TBB == NextBB && !Cond.empty() && !FBB) {
514 if (!TII->ReverseBranchCondition(Cond)) {
515 TII->RemoveBranch(*CurMBB);
516 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
521 TII->InsertBranch(*CurMBB, SuccBB, nullptr,
522 SmallVector<MachineOperand, 0>(), dl);
526 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
527 if (getHash() < o.getHash())
529 if (getHash() > o.getHash())
531 if (getBlock()->getNumber() < o.getBlock()->getNumber())
533 if (getBlock()->getNumber() > o.getBlock()->getNumber())
535 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
536 // an object with itself.
537 #ifndef _GLIBCXX_DEBUG
538 llvm_unreachable("Predecessor appears twice");
545 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
546 auto I = MergedBBFreq.find(MBB);
548 if (I != MergedBBFreq.end())
551 return MBFI.getBlockFreq(MBB);
554 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
556 MergedBBFreq[MBB] = F;
560 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
561 const MachineBasicBlock *MBB) const {
562 return MBFI.printBlockFreq(OS, getBlockFreq(MBB));
566 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
567 const BlockFrequency Freq) const {
568 return MBFI.printBlockFreq(OS, Freq);
571 /// CountTerminators - Count the number of terminators in the given
572 /// block and set I to the position of the first non-terminator, if there
573 /// is one, or MBB->end() otherwise.
574 static unsigned CountTerminators(MachineBasicBlock *MBB,
575 MachineBasicBlock::iterator &I) {
577 unsigned NumTerms = 0;
579 if (I == MBB->begin()) {
584 if (!I->isTerminator()) break;
590 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
591 /// and decide if it would be profitable to merge those tails. Return the
592 /// length of the common tail and iterators to the first common instruction
595 ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2,
596 unsigned minCommonTailLength, unsigned &CommonTailLen,
597 MachineBasicBlock::iterator &I1,
598 MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB,
599 MachineBasicBlock *PredBB,
600 DenseMap<const MachineBasicBlock *, int> &FuncletMembership) {
601 // It is never profitable to tail-merge blocks from two different funclets.
602 if (!FuncletMembership.empty()) {
603 auto Funclet1 = FuncletMembership.find(MBB1);
604 assert(Funclet1 != FuncletMembership.end());
605 auto Funclet2 = FuncletMembership.find(MBB2);
606 assert(Funclet2 != FuncletMembership.end());
607 if (Funclet1->second != Funclet2->second)
611 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
612 if (CommonTailLen == 0)
614 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
615 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
618 // It's almost always profitable to merge any number of non-terminator
619 // instructions with the block that falls through into the common successor.
620 if (MBB1 == PredBB || MBB2 == PredBB) {
621 MachineBasicBlock::iterator I;
622 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
623 if (CommonTailLen > NumTerms)
627 // If one of the blocks can be completely merged and happens to be in
628 // a position where the other could fall through into it, merge any number
629 // of instructions, because it can be done without a branch.
630 // TODO: If the blocks are not adjacent, move one of them so that they are?
631 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
633 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
636 // If both blocks have an unconditional branch temporarily stripped out,
637 // count that as an additional common instruction for the following
639 unsigned EffectiveTailLen = CommonTailLen;
640 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
641 !MBB1->back().isBarrier() &&
642 !MBB2->back().isBarrier())
645 // Check if the common tail is long enough to be worthwhile.
646 if (EffectiveTailLen >= minCommonTailLength)
649 // If we are optimizing for code size, 2 instructions in common is enough if
650 // we don't have to split a block. At worst we will be introducing 1 new
651 // branch instruction, which is likely to be smaller than the 2
652 // instructions that would be deleted in the merge.
653 MachineFunction *MF = MBB1->getParent();
654 return EffectiveTailLen >= 2 && MF->getFunction()->optForSize() &&
655 (I1 == MBB1->begin() || I2 == MBB2->begin());
658 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
659 /// hash CurHash (guaranteed to match the last element). Build the vector
660 /// SameTails of all those that have the (same) largest number of instructions
661 /// in common of any pair of these blocks. SameTails entries contain an
662 /// iterator into MergePotentials (from which the MachineBasicBlock can be
663 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
664 /// instruction where the matching code sequence begins.
665 /// Order of elements in SameTails is the reverse of the order in which
666 /// those blocks appear in MergePotentials (where they are not necessarily
668 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
669 unsigned minCommonTailLength,
670 MachineBasicBlock *SuccBB,
671 MachineBasicBlock *PredBB) {
672 unsigned maxCommonTailLength = 0U;
674 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
675 MPIterator HighestMPIter = std::prev(MergePotentials.end());
676 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
677 B = MergePotentials.begin();
678 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
679 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
680 unsigned CommonTailLen;
681 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
683 CommonTailLen, TrialBBI1, TrialBBI2,
685 FuncletMembership)) {
686 if (CommonTailLen > maxCommonTailLength) {
688 maxCommonTailLength = CommonTailLen;
689 HighestMPIter = CurMPIter;
690 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
692 if (HighestMPIter == CurMPIter &&
693 CommonTailLen == maxCommonTailLength)
694 SameTails.push_back(SameTailElt(I, TrialBBI2));
700 return maxCommonTailLength;
703 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
704 /// MergePotentials, restoring branches at ends of blocks as appropriate.
705 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
706 MachineBasicBlock *SuccBB,
707 MachineBasicBlock *PredBB) {
708 MPIterator CurMPIter, B;
709 for (CurMPIter = std::prev(MergePotentials.end()),
710 B = MergePotentials.begin();
711 CurMPIter->getHash() == CurHash; --CurMPIter) {
712 // Put the unconditional branch back, if we need one.
713 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
714 if (SuccBB && CurMBB != PredBB)
715 FixTail(CurMBB, SuccBB, TII);
719 if (CurMPIter->getHash() != CurHash)
721 MergePotentials.erase(CurMPIter, MergePotentials.end());
724 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
725 /// only of the common tail. Create a block that does by splitting one.
726 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
727 MachineBasicBlock *SuccBB,
728 unsigned maxCommonTailLength,
729 unsigned &commonTailIndex) {
731 unsigned TimeEstimate = ~0U;
732 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
733 // Use PredBB if possible; that doesn't require a new branch.
734 if (SameTails[i].getBlock() == PredBB) {
738 // Otherwise, make a (fairly bogus) choice based on estimate of
739 // how long it will take the various blocks to execute.
740 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
741 SameTails[i].getTailStartPos());
742 if (t <= TimeEstimate) {
748 MachineBasicBlock::iterator BBI =
749 SameTails[commonTailIndex].getTailStartPos();
750 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
752 // If the common tail includes any debug info we will take it pretty
753 // randomly from one of the inputs. Might be better to remove it?
754 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
755 << maxCommonTailLength);
757 // If the split block unconditionally falls-thru to SuccBB, it will be
758 // merged. In control flow terms it should then take SuccBB's name. e.g. If
759 // SuccBB is an inner loop, the common tail is still part of the inner loop.
760 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
761 SuccBB->getBasicBlock() : MBB->getBasicBlock();
762 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
764 DEBUG(dbgs() << "... failed!");
768 SameTails[commonTailIndex].setBlock(newMBB);
769 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
771 // If we split PredBB, newMBB is the new predecessor.
779 mergeOperations(MachineBasicBlock::iterator MBBIStartPos,
780 MachineBasicBlock &MBBCommon) {
781 MachineBasicBlock *MBB = MBBIStartPos->getParent();
782 // Note CommonTailLen does not necessarily matches the size of
783 // the common BB nor all its instructions because of debug
784 // instructions differences.
785 unsigned CommonTailLen = 0;
786 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
789 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
790 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
791 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
792 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
794 while (CommonTailLen--) {
795 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
798 if (MBBI->isDebugValue()) {
803 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
806 assert(MBBICommon != MBBIECommon &&
807 "Reached BB end within common tail length!");
808 assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!");
810 // Merge MMOs from memory operations in the common block.
811 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
812 MBBICommon->setMemRefs(MBBICommon->mergeMemRefsWith(*MBBI));
813 // Drop undef flags if they aren't present in all merged instructions.
814 for (unsigned I = 0, E = MBBICommon->getNumOperands(); I != E; ++I) {
815 MachineOperand &MO = MBBICommon->getOperand(I);
816 if (MO.isReg() && MO.isUndef()) {
817 const MachineOperand &OtherMO = MBBI->getOperand(I);
818 if (!OtherMO.isUndef())
819 MO.setIsUndef(false);
828 // See if any of the blocks in MergePotentials (which all have SuccBB as a
829 // successor, or all have no successor if it is null) can be tail-merged.
830 // If there is a successor, any blocks in MergePotentials that are not
831 // tail-merged and are not immediately before Succ must have an unconditional
832 // branch to Succ added (but the predecessor/successor lists need no
833 // adjustment). The lone predecessor of Succ that falls through into Succ,
834 // if any, is given in PredBB.
835 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
836 MachineBasicBlock *PredBB) {
837 bool MadeChange = false;
839 // Except for the special cases below, tail-merge if there are at least
840 // this many instructions in common.
841 unsigned minCommonTailLength = TailMergeSize;
843 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
844 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
845 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
846 << (i == e-1 ? "" : ", ");
849 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
851 dbgs() << " which has fall-through from BB#"
852 << PredBB->getNumber() << "\n";
854 dbgs() << "Looking for common tails of at least "
855 << minCommonTailLength << " instruction"
856 << (minCommonTailLength == 1 ? "" : "s") << '\n';
859 // Sort by hash value so that blocks with identical end sequences sort
861 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
863 // Walk through equivalence sets looking for actual exact matches.
864 while (MergePotentials.size() > 1) {
865 unsigned CurHash = MergePotentials.back().getHash();
867 // Build SameTails, identifying the set of blocks with this hash code
868 // and with the maximum number of instructions in common.
869 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
873 // If we didn't find any pair that has at least minCommonTailLength
874 // instructions in common, remove all blocks with this hash code and retry.
875 if (SameTails.empty()) {
876 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
880 // If one of the blocks is the entire common tail (and not the entry
881 // block, which we can't jump to), we can treat all blocks with this same
882 // tail at once. Use PredBB if that is one of the possibilities, as that
883 // will not introduce any extra branches.
884 MachineBasicBlock *EntryBB =
885 &MergePotentials.front().getBlock()->getParent()->front();
886 unsigned commonTailIndex = SameTails.size();
887 // If there are two blocks, check to see if one can be made to fall through
889 if (SameTails.size() == 2 &&
890 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
891 SameTails[1].tailIsWholeBlock())
893 else if (SameTails.size() == 2 &&
894 SameTails[1].getBlock()->isLayoutSuccessor(
895 SameTails[0].getBlock()) &&
896 SameTails[0].tailIsWholeBlock())
899 // Otherwise just pick one, favoring the fall-through predecessor if
901 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
902 MachineBasicBlock *MBB = SameTails[i].getBlock();
903 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
909 if (SameTails[i].tailIsWholeBlock())
914 if (commonTailIndex == SameTails.size() ||
915 (SameTails[commonTailIndex].getBlock() == PredBB &&
916 !SameTails[commonTailIndex].tailIsWholeBlock())) {
917 // None of the blocks consist entirely of the common tail.
918 // Split a block so that one does.
919 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
920 maxCommonTailLength, commonTailIndex)) {
921 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
926 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
928 // Recompute common tail MBB's edge weights and block frequency.
929 setCommonTailEdgeWeights(*MBB);
931 // MBB is common tail. Adjust all other BB's to jump to this one.
932 // Traversal must be forwards so erases work.
933 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
935 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
936 if (commonTailIndex == i)
938 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
939 << (i == e-1 ? "" : ", "));
940 // Merge operations (MMOs, undef flags)
941 mergeOperations(SameTails[i].getTailStartPos(), *MBB);
942 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
943 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
944 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
945 MergePotentials.erase(SameTails[i].getMPIter());
947 DEBUG(dbgs() << "\n");
948 // We leave commonTailIndex in the worklist in case there are other blocks
949 // that match it with a smaller number of instructions.
955 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
956 bool MadeChange = false;
957 if (!EnableTailMerge) return MadeChange;
959 // First find blocks with no successors.
960 // Block placement does not create new tail merging opportunities for these
962 if (!AfterBlockPlacement) {
963 MergePotentials.clear();
964 for (MachineBasicBlock &MBB : MF) {
965 if (MergePotentials.size() == TailMergeThreshold)
967 if (!TriedMerging.count(&MBB) && MBB.succ_empty())
968 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB));
971 // If this is a large problem, avoid visiting the same basic blocks
973 if (MergePotentials.size() == TailMergeThreshold)
974 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
975 TriedMerging.insert(MergePotentials[i].getBlock());
977 // See if we can do any tail merging on those.
978 if (MergePotentials.size() >= 2)
979 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
982 // Look at blocks (IBB) with multiple predecessors (PBB).
983 // We change each predecessor to a canonical form, by
984 // (1) temporarily removing any unconditional branch from the predecessor
986 // (2) alter conditional branches so they branch to the other block
987 // not IBB; this may require adding back an unconditional branch to IBB
988 // later, where there wasn't one coming in. E.g.
990 // fallthrough to QBB
993 // with a conceptual B to IBB after that, which never actually exists.
994 // With those changes, we see whether the predecessors' tails match,
995 // and merge them if so. We change things out of canonical form and
996 // back to the way they were later in the process. (OptimizeBranches
997 // would undo some of this, but we can't use it, because we'd get into
998 // a compile-time infinite loop repeatedly doing and undoing the same
1001 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1003 if (I->pred_size() < 2) continue;
1004 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
1005 MachineBasicBlock *IBB = &*I;
1006 MachineBasicBlock *PredBB = &*std::prev(I);
1007 MergePotentials.clear();
1010 // Bail if merging after placement and IBB is the loop header because
1011 // -- If merging predecessors that belong to the same loop as IBB, the
1012 // common tail of merged predecessors may become the loop top if block
1013 // placement is called again and the predecessors may branch to this common
1014 // tail and require more branches. This can be relaxed if
1015 // MachineBlockPlacement::findBestLoopTop is more flexible.
1016 // --If merging predecessors that do not belong to the same loop as IBB, the
1017 // loop info of IBB's loop and the other loops may be affected. Calling the
1018 // block placement again may make big change to the layout and eliminate the
1019 // reason to do tail merging here.
1020 if (AfterBlockPlacement && MLI) {
1021 ML = MLI->getLoopFor(IBB);
1022 if (ML && IBB == ML->getHeader())
1026 for (MachineBasicBlock *PBB : I->predecessors()) {
1027 if (MergePotentials.size() == TailMergeThreshold)
1030 if (TriedMerging.count(PBB))
1033 // Skip blocks that loop to themselves, can't tail merge these.
1037 // Visit each predecessor only once.
1038 if (!UniquePreds.insert(PBB).second)
1041 // Skip blocks which may jump to a landing pad. Can't tail merge these.
1042 if (PBB->hasEHPadSuccessor())
1045 // After block placement, only consider predecessors that belong to the
1046 // same loop as IBB. The reason is the same as above when skipping loop
1048 if (AfterBlockPlacement && MLI)
1049 if (ML != MLI->getLoopFor(PBB))
1052 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1053 SmallVector<MachineOperand, 4> Cond;
1054 if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) {
1055 // Failing case: IBB is the target of a cbr, and we cannot reverse the
1057 SmallVector<MachineOperand, 4> NewCond(Cond);
1058 if (!Cond.empty() && TBB == IBB) {
1059 if (TII->ReverseBranchCondition(NewCond))
1061 // This is the QBB case described above
1063 auto Next = ++PBB->getIterator();
1064 if (Next != MF.end())
1069 // Failing case: the only way IBB can be reached from PBB is via
1070 // exception handling. Happens for landing pads. Would be nice to have
1071 // a bit in the edge so we didn't have to do all this.
1072 if (IBB->isEHPad()) {
1073 MachineFunction::iterator IP = ++PBB->getIterator();
1074 MachineBasicBlock *PredNextBB = nullptr;
1078 if (IBB != PredNextBB) // fallthrough
1081 if (TBB != IBB && FBB != IBB) // cbr then ubr
1083 } else if (Cond.empty()) {
1084 if (TBB != IBB) // ubr
1087 if (TBB != IBB && IBB != PredNextBB) // cbr
1092 // Remove the unconditional branch at the end, if any.
1093 if (TBB && (Cond.empty() || FBB)) {
1094 DebugLoc dl; // FIXME: this is nowhere
1095 TII->RemoveBranch(*PBB);
1097 // reinsert conditional branch only, for now
1098 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1102 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB));
1106 // If this is a large problem, avoid visiting the same basic blocks multiple
1108 if (MergePotentials.size() == TailMergeThreshold)
1109 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1110 TriedMerging.insert(MergePotentials[i].getBlock());
1112 if (MergePotentials.size() >= 2)
1113 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
1115 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1116 // result of removing blocks in TryTailMergeBlocks.
1117 PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks
1118 if (MergePotentials.size() == 1 &&
1119 MergePotentials.begin()->getBlock() != PredBB)
1120 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1126 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1127 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1128 BlockFrequency AccumulatedMBBFreq;
1130 // Aggregate edge frequency of successor edge j:
1131 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1132 // where bb is a basic block that is in SameTails.
1133 for (const auto &Src : SameTails) {
1134 const MachineBasicBlock *SrcMBB = Src.getBlock();
1135 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1136 AccumulatedMBBFreq += BlockFreq;
1138 // It is not necessary to recompute edge weights if TailBB has less than two
1140 if (TailMBB.succ_size() <= 1)
1143 auto EdgeFreq = EdgeFreqLs.begin();
1145 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1146 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1147 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1150 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1152 if (TailMBB.succ_size() <= 1)
1156 std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0))
1158 auto EdgeFreq = EdgeFreqLs.begin();
1160 if (SumEdgeFreq > 0) {
1161 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1162 SuccI != SuccE; ++SuccI, ++EdgeFreq) {
1163 auto Prob = BranchProbability::getBranchProbability(
1164 EdgeFreq->getFrequency(), SumEdgeFreq);
1165 TailMBB.setSuccProbability(SuccI, Prob);
1170 //===----------------------------------------------------------------------===//
1171 // Branch Optimization
1172 //===----------------------------------------------------------------------===//
1174 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1175 bool MadeChange = false;
1177 // Make sure blocks are numbered in order
1178 MF.RenumberBlocks();
1179 // Renumbering blocks alters funclet membership, recalculate it.
1180 FuncletMembership = getFuncletMembership(MF);
1182 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1184 MachineBasicBlock *MBB = &*I++;
1185 MadeChange |= OptimizeBlock(MBB);
1187 // If it is dead, remove it.
1188 if (MBB->pred_empty()) {
1189 RemoveDeadBlock(MBB);
1198 // Blocks should be considered empty if they contain only debug info;
1199 // else the debug info would affect codegen.
1200 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1201 return MBB->getFirstNonDebugInstr() == MBB->end();
1204 // Blocks with only debug info and branches should be considered the same
1205 // as blocks with only branches.
1206 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1207 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1208 assert(I != MBB->end() && "empty block!");
1209 return I->isBranch();
1212 /// IsBetterFallthrough - Return true if it would be clearly better to
1213 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1214 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1215 /// result in infinite loops.
1216 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1217 MachineBasicBlock *MBB2) {
1218 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1219 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1220 // optimize branches that branch to either a return block or an assert block
1221 // into a fallthrough to the return.
1222 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1223 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1224 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1227 // If there is a clear successor ordering we make sure that one block
1228 // will fall through to the next
1229 if (MBB1->isSuccessor(MBB2)) return true;
1230 if (MBB2->isSuccessor(MBB1)) return false;
1232 return MBB2I->isCall() && !MBB1I->isCall();
1235 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1236 /// instructions on the block.
1237 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1238 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1239 if (I != MBB.end() && I->isBranch())
1240 return I->getDebugLoc();
1244 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1245 /// block. This is never called on the entry block.
1246 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1247 bool MadeChange = false;
1248 MachineFunction &MF = *MBB->getParent();
1251 MachineFunction::iterator FallThrough = MBB->getIterator();
1254 // Make sure MBB and FallThrough belong to the same funclet.
1255 bool SameFunclet = true;
1256 if (!FuncletMembership.empty() && FallThrough != MF.end()) {
1257 auto MBBFunclet = FuncletMembership.find(MBB);
1258 assert(MBBFunclet != FuncletMembership.end());
1259 auto FallThroughFunclet = FuncletMembership.find(&*FallThrough);
1260 assert(FallThroughFunclet != FuncletMembership.end());
1261 SameFunclet = MBBFunclet->second == FallThroughFunclet->second;
1264 // If this block is empty, make everyone use its fall-through, not the block
1265 // explicitly. Landing pads should not do this since the landing-pad table
1266 // points to this block. Blocks with their addresses taken shouldn't be
1268 if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() &&
1270 // Dead block? Leave for cleanup later.
1271 if (MBB->pred_empty()) return MadeChange;
1273 if (FallThrough == MF.end()) {
1274 // TODO: Simplify preds to not branch here if possible!
1275 } else if (FallThrough->isEHPad()) {
1276 // Don't rewrite to a landing pad fallthough. That could lead to the case
1277 // where a BB jumps to more than one landing pad.
1278 // TODO: Is it ever worth rewriting predecessors which don't already
1279 // jump to a landing pad, and so can safely jump to the fallthrough?
1280 } else if (MBB->isSuccessor(&*FallThrough)) {
1281 // Rewrite all predecessors of the old block to go to the fallthrough
1283 while (!MBB->pred_empty()) {
1284 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1285 Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough);
1287 // If MBB was the target of a jump table, update jump tables to go to the
1288 // fallthrough instead.
1289 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1290 MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough);
1296 // Check to see if we can simplify the terminator of the block before this
1298 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1300 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1301 SmallVector<MachineOperand, 4> PriorCond;
1302 bool PriorUnAnalyzable =
1303 TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1304 if (!PriorUnAnalyzable) {
1305 // If the CFG for the prior block has extra edges, remove them.
1306 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1307 !PriorCond.empty());
1309 // If the previous branch is conditional and both conditions go to the same
1310 // destination, remove the branch, replacing it with an unconditional one or
1312 if (PriorTBB && PriorTBB == PriorFBB) {
1313 DebugLoc dl = getBranchDebugLoc(PrevBB);
1314 TII->RemoveBranch(PrevBB);
1316 if (PriorTBB != MBB)
1317 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1320 goto ReoptimizeBlock;
1323 // If the previous block unconditionally falls through to this block and
1324 // this block has no other predecessors, move the contents of this block
1325 // into the prior block. This doesn't usually happen when SimplifyCFG
1326 // has been used, but it can happen if tail merging splits a fall-through
1327 // predecessor of a block.
1328 // This has to check PrevBB->succ_size() because EH edges are ignored by
1330 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1331 PrevBB.succ_size() == 1 &&
1332 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1333 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1334 << "From MBB: " << *MBB);
1335 // Remove redundant DBG_VALUEs first.
1336 if (PrevBB.begin() != PrevBB.end()) {
1337 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1339 MachineBasicBlock::iterator MBBIter = MBB->begin();
1340 // Check if DBG_VALUE at the end of PrevBB is identical to the
1341 // DBG_VALUE at the beginning of MBB.
1342 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1343 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1344 if (!MBBIter->isIdenticalTo(*PrevBBIter))
1346 MachineInstr &DuplicateDbg = *MBBIter;
1347 ++MBBIter; -- PrevBBIter;
1348 DuplicateDbg.eraseFromParent();
1351 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1352 PrevBB.removeSuccessor(PrevBB.succ_begin());
1353 assert(PrevBB.succ_empty());
1354 PrevBB.transferSuccessors(MBB);
1359 // If the previous branch *only* branches to *this* block (conditional or
1360 // not) remove the branch.
1361 if (PriorTBB == MBB && !PriorFBB) {
1362 TII->RemoveBranch(PrevBB);
1365 goto ReoptimizeBlock;
1368 // If the prior block branches somewhere else on the condition and here if
1369 // the condition is false, remove the uncond second branch.
1370 if (PriorFBB == MBB) {
1371 DebugLoc dl = getBranchDebugLoc(PrevBB);
1372 TII->RemoveBranch(PrevBB);
1373 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1376 goto ReoptimizeBlock;
1379 // If the prior block branches here on true and somewhere else on false, and
1380 // if the branch condition is reversible, reverse the branch to create a
1382 if (PriorTBB == MBB) {
1383 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1384 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1385 DebugLoc dl = getBranchDebugLoc(PrevBB);
1386 TII->RemoveBranch(PrevBB);
1387 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1390 goto ReoptimizeBlock;
1394 // If this block has no successors (e.g. it is a return block or ends with
1395 // a call to a no-return function like abort or __cxa_throw) and if the pred
1396 // falls through into this block, and if it would otherwise fall through
1397 // into the block after this, move this block to the end of the function.
1399 // We consider it more likely that execution will stay in the function (e.g.
1400 // due to loops) than it is to exit it. This asserts in loops etc, moving
1401 // the assert condition out of the loop body.
1402 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1403 MachineFunction::iterator(PriorTBB) == FallThrough &&
1404 !MBB->canFallThrough()) {
1405 bool DoTransform = true;
1407 // We have to be careful that the succs of PredBB aren't both no-successor
1408 // blocks. If neither have successors and if PredBB is the second from
1409 // last block in the function, we'd just keep swapping the two blocks for
1410 // last. Only do the swap if one is clearly better to fall through than
1412 if (FallThrough == --MF.end() &&
1413 !IsBetterFallthrough(PriorTBB, MBB))
1414 DoTransform = false;
1417 // Reverse the branch so we will fall through on the previous true cond.
1418 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1419 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1420 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1421 << "To make fallthrough to: " << *PriorTBB << "\n");
1423 DebugLoc dl = getBranchDebugLoc(PrevBB);
1424 TII->RemoveBranch(PrevBB);
1425 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1427 // Move this block to the end of the function.
1428 MBB->moveAfter(&MF.back());
1437 // Analyze the branch in the current block.
1438 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1439 SmallVector<MachineOperand, 4> CurCond;
1440 bool CurUnAnalyzable =
1441 TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1442 if (!CurUnAnalyzable) {
1443 // If the CFG for the prior block has extra edges, remove them.
1444 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1446 // If this is a two-way branch, and the FBB branches to this block, reverse
1447 // the condition so the single-basic-block loop is faster. Instead of:
1448 // Loop: xxx; jcc Out; jmp Loop
1450 // Loop: xxx; jncc Loop; jmp Out
1451 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1452 SmallVector<MachineOperand, 4> NewCond(CurCond);
1453 if (!TII->ReverseBranchCondition(NewCond)) {
1454 DebugLoc dl = getBranchDebugLoc(*MBB);
1455 TII->RemoveBranch(*MBB);
1456 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1459 goto ReoptimizeBlock;
1463 // If this branch is the only thing in its block, see if we can forward
1464 // other blocks across it.
1465 if (CurTBB && CurCond.empty() && !CurFBB &&
1466 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1467 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1468 DebugLoc dl = getBranchDebugLoc(*MBB);
1469 // This block may contain just an unconditional branch. Because there can
1470 // be 'non-branch terminators' in the block, try removing the branch and
1471 // then seeing if the block is empty.
1472 TII->RemoveBranch(*MBB);
1473 // If the only things remaining in the block are debug info, remove these
1474 // as well, so this will behave the same as an empty block in non-debug
1476 if (IsEmptyBlock(MBB)) {
1477 // Make the block empty, losing the debug info (we could probably
1478 // improve this in some cases.)
1479 MBB->erase(MBB->begin(), MBB->end());
1481 // If this block is just an unconditional branch to CurTBB, we can
1482 // usually completely eliminate the block. The only case we cannot
1483 // completely eliminate the block is when the block before this one
1484 // falls through into MBB and we can't understand the prior block's branch
1487 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1488 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1489 !PrevBB.isSuccessor(MBB)) {
1490 // If the prior block falls through into us, turn it into an
1491 // explicit branch to us to make updates simpler.
1492 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1493 PriorTBB != MBB && PriorFBB != MBB) {
1495 assert(PriorCond.empty() && !PriorFBB &&
1496 "Bad branch analysis");
1499 assert(!PriorFBB && "Machine CFG out of date!");
1502 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1503 TII->RemoveBranch(PrevBB);
1504 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1507 // Iterate through all the predecessors, revectoring each in-turn.
1509 bool DidChange = false;
1510 bool HasBranchToSelf = false;
1511 while(PI != MBB->pred_size()) {
1512 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1514 // If this block has an uncond branch to itself, leave it.
1516 HasBranchToSelf = true;
1519 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1520 // If this change resulted in PMBB ending in a conditional
1521 // branch where both conditions go to the same destination,
1522 // change this to an unconditional branch (and fix the CFG).
1523 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1524 SmallVector<MachineOperand, 4> NewCurCond;
1525 bool NewCurUnAnalyzable = TII->analyzeBranch(
1526 *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true);
1527 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1528 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1529 TII->RemoveBranch(*PMBB);
1531 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1534 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1539 // Change any jumptables to go to the new MBB.
1540 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1541 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1545 if (!HasBranchToSelf) return MadeChange;
1550 // Add the branch back if the block is more than just an uncond branch.
1551 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1555 // If the prior block doesn't fall through into this block, and if this
1556 // block doesn't fall through into some other block, see if we can find a
1557 // place to move this block where a fall-through will happen.
1558 if (!PrevBB.canFallThrough()) {
1560 // Now we know that there was no fall-through into this block, check to
1561 // see if it has a fall-through into its successor.
1562 bool CurFallsThru = MBB->canFallThrough();
1564 if (!MBB->isEHPad()) {
1565 // Check all the predecessors of this block. If one of them has no fall
1566 // throughs, move this block right after it.
1567 for (MachineBasicBlock *PredBB : MBB->predecessors()) {
1568 // Analyze the branch at the end of the pred.
1569 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1570 SmallVector<MachineOperand, 4> PredCond;
1571 if (PredBB != MBB && !PredBB->canFallThrough() &&
1572 !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) &&
1573 (!CurFallsThru || !CurTBB || !CurFBB) &&
1574 (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1575 // If the current block doesn't fall through, just move it.
1576 // If the current block can fall through and does not end with a
1577 // conditional branch, we need to append an unconditional jump to
1578 // the (current) next block. To avoid a possible compile-time
1579 // infinite loop, move blocks only backward in this case.
1580 // Also, if there are already 2 branches here, we cannot add a third;
1581 // this means we have the case
1586 MachineBasicBlock *NextBB = &*std::next(MBB->getIterator());
1588 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1590 MBB->moveAfter(PredBB);
1592 goto ReoptimizeBlock;
1597 if (!CurFallsThru) {
1598 // Check all successors to see if we can move this block before it.
1599 for (MachineBasicBlock *SuccBB : MBB->successors()) {
1600 // Analyze the branch at the end of the block before the succ.
1601 MachineFunction::iterator SuccPrev = --SuccBB->getIterator();
1603 // If this block doesn't already fall-through to that successor, and if
1604 // the succ doesn't already have a block that can fall through into it,
1605 // and if the successor isn't an EH destination, we can arrange for the
1606 // fallthrough to happen.
1607 if (SuccBB != MBB && &*SuccPrev != MBB &&
1608 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1609 !SuccBB->isEHPad()) {
1610 MBB->moveBefore(SuccBB);
1612 goto ReoptimizeBlock;
1616 // Okay, there is no really great place to put this block. If, however,
1617 // the block before this one would be a fall-through if this block were
1618 // removed, move this block to the end of the function.
1619 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1620 SmallVector<MachineOperand, 4> PrevCond;
1621 // We're looking for cases where PrevBB could possibly fall through to
1622 // FallThrough, but if FallThrough is an EH pad that wouldn't be useful
1623 // so here we skip over any EH pads so we might have a chance to find
1624 // a branch target from PrevBB.
1625 while (FallThrough != MF.end() && FallThrough->isEHPad())
1627 // Now check to see if the current block is sitting between PrevBB and
1628 // a block to which it could fall through.
1629 if (FallThrough != MF.end() &&
1630 !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1631 PrevBB.isSuccessor(&*FallThrough)) {
1632 MBB->moveAfter(&MF.back());
1642 //===----------------------------------------------------------------------===//
1643 // Hoist Common Code
1644 //===----------------------------------------------------------------------===//
1646 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1647 /// blocks to their common predecessor.
1648 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1649 bool MadeChange = false;
1650 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1651 MachineBasicBlock *MBB = &*I++;
1652 MadeChange |= HoistCommonCodeInSuccs(MBB);
1658 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1659 /// its 'true' successor.
1660 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1661 MachineBasicBlock *TrueBB) {
1662 for (MachineBasicBlock *SuccBB : BB->successors())
1663 if (SuccBB != TrueBB)
1668 template <class Container>
1669 static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
1671 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1672 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1679 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1680 /// in successors to. The location is usually just before the terminator,
1681 /// however if the terminator is a conditional branch and its previous
1682 /// instruction is the flag setting instruction, the previous instruction is
1683 /// the preferred location. This function also gathers uses and defs of the
1684 /// instructions from the insertion point to the end of the block. The data is
1685 /// used by HoistCommonCodeInSuccs to ensure safety.
1687 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1688 const TargetInstrInfo *TII,
1689 const TargetRegisterInfo *TRI,
1690 SmallSet<unsigned,4> &Uses,
1691 SmallSet<unsigned,4> &Defs) {
1692 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1693 if (!TII->isUnpredicatedTerminator(*Loc))
1696 for (const MachineOperand &MO : Loc->operands()) {
1699 unsigned Reg = MO.getReg();
1703 addRegAndItsAliases(Reg, TRI, Uses);
1706 // Don't try to hoist code in the rare case the terminator defines a
1707 // register that is later used.
1710 // If the terminator defines a register, make sure we don't hoist
1711 // the instruction whose def might be clobbered by the terminator.
1712 addRegAndItsAliases(Reg, TRI, Defs);
1718 if (Loc == MBB->begin())
1721 // The terminator is probably a conditional branch, try not to separate the
1722 // branch from condition setting instruction.
1723 MachineBasicBlock::iterator PI = Loc;
1725 while (PI != MBB->begin() && PI->isDebugValue())
1729 for (const MachineOperand &MO : PI->operands()) {
1730 // If PI has a regmask operand, it is probably a call. Separate away.
1733 if (!MO.isReg() || MO.isUse())
1735 unsigned Reg = MO.getReg();
1738 if (Uses.count(Reg)) {
1744 // The condition setting instruction is not just before the conditional
1748 // Be conservative, don't insert instruction above something that may have
1749 // side-effects. And since it's potentially bad to separate flag setting
1750 // instruction from the conditional branch, just abort the optimization
1752 // Also avoid moving code above predicated instruction since it's hard to
1753 // reason about register liveness with predicated instruction.
1754 bool DontMoveAcrossStore = true;
1755 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI))
1759 // Find out what registers are live. Note this routine is ignoring other live
1760 // registers which are only used by instructions in successor blocks.
1761 for (const MachineOperand &MO : PI->operands()) {
1764 unsigned Reg = MO.getReg();
1768 addRegAndItsAliases(Reg, TRI, Uses);
1770 if (Uses.erase(Reg)) {
1771 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1772 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1773 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1776 addRegAndItsAliases(Reg, TRI, Defs);
1783 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1784 /// sequence at the start of the function, move the instructions before MBB
1785 /// terminator if it's legal.
1786 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1787 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1788 SmallVector<MachineOperand, 4> Cond;
1789 if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1792 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1794 // Malformed bcc? True and false blocks are the same?
1797 // Restrict the optimization to cases where MBB is the only predecessor,
1798 // it is an obvious win.
1799 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1802 // Find a suitable position to hoist the common instructions to. Also figure
1803 // out which registers are used or defined by instructions from the insertion
1804 // point to the end of the block.
1805 SmallSet<unsigned, 4> Uses, Defs;
1806 MachineBasicBlock::iterator Loc =
1807 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1808 if (Loc == MBB->end())
1811 bool HasDups = false;
1812 SmallVector<unsigned, 4> LocalDefs;
1813 SmallSet<unsigned, 4> LocalDefsSet;
1814 MachineBasicBlock::iterator TIB = TBB->begin();
1815 MachineBasicBlock::iterator FIB = FBB->begin();
1816 MachineBasicBlock::iterator TIE = TBB->end();
1817 MachineBasicBlock::iterator FIE = FBB->end();
1818 while (TIB != TIE && FIB != FIE) {
1819 // Skip dbg_value instructions. These do not count.
1820 if (TIB->isDebugValue()) {
1821 while (TIB != TIE && TIB->isDebugValue())
1826 if (FIB->isDebugValue()) {
1827 while (FIB != FIE && FIB->isDebugValue())
1832 if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead))
1835 if (TII->isPredicated(*TIB))
1836 // Hard to reason about register liveness with predicated instruction.
1840 for (MachineOperand &MO : TIB->operands()) {
1841 // Don't attempt to hoist instructions with register masks.
1842 if (MO.isRegMask()) {
1848 unsigned Reg = MO.getReg();
1852 if (Uses.count(Reg)) {
1853 // Avoid clobbering a register that's used by the instruction at
1854 // the point of insertion.
1859 if (Defs.count(Reg) && !MO.isDead()) {
1860 // Don't hoist the instruction if the def would be clobber by the
1861 // instruction at the point insertion. FIXME: This is overly
1862 // conservative. It should be possible to hoist the instructions
1863 // in BB2 in the following example:
1865 // r1, eflag = op1 r2, r3
1874 } else if (!LocalDefsSet.count(Reg)) {
1875 if (Defs.count(Reg)) {
1876 // Use is defined by the instruction at the point of insertion.
1881 if (MO.isKill() && Uses.count(Reg))
1882 // Kills a register that's read by the instruction at the point of
1883 // insertion. Remove the kill marker.
1884 MO.setIsKill(false);
1890 bool DontMoveAcrossStore = true;
1891 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1894 // Remove kills from LocalDefsSet, these registers had short live ranges.
1895 for (const MachineOperand &MO : TIB->operands()) {
1896 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1898 unsigned Reg = MO.getReg();
1899 if (!Reg || !LocalDefsSet.count(Reg))
1901 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1902 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1903 LocalDefsSet.erase(*AI);
1905 LocalDefsSet.erase(Reg);
1909 // Track local defs so we can update liveins.
1910 for (const MachineOperand &MO : TIB->operands()) {
1911 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1913 unsigned Reg = MO.getReg();
1914 if (!Reg || TargetRegisterInfo::isVirtualRegister(Reg))
1916 LocalDefs.push_back(Reg);
1917 addRegAndItsAliases(Reg, TRI, LocalDefsSet);
1928 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1929 FBB->erase(FBB->begin(), FIB);
1932 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1933 unsigned Def = LocalDefs[i];
1934 if (LocalDefsSet.count(Def)) {
1935 TBB->addLiveIn(Def);
1936 FBB->addLiveIn(Def);