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/MachineModuleInfo.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/Passes.h"
33 #include "llvm/CodeGen/RegisterScavenging.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetInstrInfo.h"
40 #include "llvm/Target/TargetRegisterInfo.h"
41 #include "llvm/Target/TargetSubtargetInfo.h"
45 #define DEBUG_TYPE "branchfolding"
47 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
48 STATISTIC(NumBranchOpts, "Number of branches optimized");
49 STATISTIC(NumTailMerge , "Number of block tails merged");
50 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
52 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
53 cl::init(cl::BOU_UNSET), cl::Hidden);
55 // Throttle for huge numbers of predecessors (compile speed problems)
56 static cl::opt<unsigned>
57 TailMergeThreshold("tail-merge-threshold",
58 cl::desc("Max number of predecessors to consider tail merging"),
59 cl::init(150), cl::Hidden);
61 // Heuristic for tail merging (and, inversely, tail duplication).
62 // TODO: This should be replaced with a target query.
63 static cl::opt<unsigned>
64 TailMergeSize("tail-merge-size",
65 cl::desc("Min number of instructions to consider tail merging"),
66 cl::init(3), cl::Hidden);
69 /// BranchFolderPass - Wrap branch folder in a machine function pass.
70 class BranchFolderPass : public MachineFunctionPass {
73 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
75 bool runOnMachineFunction(MachineFunction &MF) override;
77 void getAnalysisUsage(AnalysisUsage &AU) const override {
78 AU.addRequired<MachineBlockFrequencyInfo>();
79 AU.addRequired<MachineBranchProbabilityInfo>();
80 AU.addRequired<TargetPassConfig>();
81 MachineFunctionPass::getAnalysisUsage(AU);
86 char BranchFolderPass::ID = 0;
87 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
89 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
90 "Control Flow Optimizer", false, false)
92 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
93 if (skipOptnoneFunction(*MF.getFunction()))
96 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
97 // TailMerge can create jump into if branches that make CFG irreducible for
98 // HW that requires structurized CFG.
99 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
100 PassConfig->getEnableTailMerge();
101 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true,
102 getAnalysis<MachineBlockFrequencyInfo>(),
103 getAnalysis<MachineBranchProbabilityInfo>());
104 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
105 MF.getSubtarget().getRegisterInfo(),
106 getAnalysisIfAvailable<MachineModuleInfo>());
109 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
110 const MachineBlockFrequencyInfo &FreqInfo,
111 const MachineBranchProbabilityInfo &ProbInfo)
112 : EnableHoistCommonCode(CommonHoist), MBBFreqInfo(FreqInfo),
114 switch (FlagEnableTailMerge) {
115 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
116 case cl::BOU_TRUE: EnableTailMerge = true; break;
117 case cl::BOU_FALSE: EnableTailMerge = false; break;
121 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
122 /// function, updating the CFG.
123 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
124 assert(MBB->pred_empty() && "MBB must be dead!");
125 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
127 MachineFunction *MF = MBB->getParent();
128 // drop all successors.
129 while (!MBB->succ_empty())
130 MBB->removeSuccessor(MBB->succ_end()-1);
132 // Avoid matching if this pointer gets reused.
133 TriedMerging.erase(MBB);
137 FuncletMembership.erase(MBB);
140 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
141 /// followed by terminators, and if the implicitly defined registers are not
142 /// used by the terminators, remove those implicit_def's. e.g.
144 /// r0 = implicit_def
145 /// r1 = implicit_def
147 /// This block can be optimized away later if the implicit instructions are
149 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
150 SmallSet<unsigned, 4> ImpDefRegs;
151 MachineBasicBlock::iterator I = MBB->begin();
152 while (I != MBB->end()) {
153 if (!I->isImplicitDef())
155 unsigned Reg = I->getOperand(0).getReg();
156 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
157 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
158 SubRegs.isValid(); ++SubRegs)
159 ImpDefRegs.insert(*SubRegs);
161 ImpDefRegs.insert(Reg);
165 if (ImpDefRegs.empty())
168 MachineBasicBlock::iterator FirstTerm = I;
169 while (I != MBB->end()) {
170 if (!TII->isUnpredicatedTerminator(I))
172 // See if it uses any of the implicitly defined registers.
173 for (const MachineOperand &MO : I->operands()) {
174 if (!MO.isReg() || !MO.isUse())
176 unsigned Reg = MO.getReg();
177 if (ImpDefRegs.count(Reg))
184 while (I != FirstTerm) {
185 MachineInstr *ImpDefMI = &*I;
187 MBB->erase(ImpDefMI);
193 /// OptimizeFunction - Perhaps branch folding, tail merging and other
194 /// CFG optimizations on the given function.
195 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
196 const TargetInstrInfo *tii,
197 const TargetRegisterInfo *tri,
198 MachineModuleInfo *mmi) {
199 if (!tii) return false;
201 TriedMerging.clear();
208 // Use a RegScavenger to help update liveness when required.
209 MachineRegisterInfo &MRI = MF.getRegInfo();
210 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
211 RS = new RegScavenger();
213 MRI.invalidateLiveness();
215 // Fix CFG. The later algorithms expect it to be right.
216 bool MadeChange = false;
217 for (MachineBasicBlock &MBB : MF) {
218 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
219 SmallVector<MachineOperand, 4> Cond;
220 if (!TII->AnalyzeBranch(MBB, TBB, FBB, Cond, true))
221 MadeChange |= MBB.CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
222 MadeChange |= OptimizeImpDefsBlock(&MBB);
225 // Recalculate funclet membership.
226 FuncletMembership = getFuncletMembership(MF);
228 bool MadeChangeThisIteration = true;
229 while (MadeChangeThisIteration) {
230 MadeChangeThisIteration = TailMergeBlocks(MF);
231 MadeChangeThisIteration |= OptimizeBranches(MF);
232 if (EnableHoistCommonCode)
233 MadeChangeThisIteration |= HoistCommonCode(MF);
234 MadeChange |= MadeChangeThisIteration;
237 // See if any jump tables have become dead as the code generator
239 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
245 // Walk the function to find jump tables that are live.
246 BitVector JTIsLive(JTI->getJumpTables().size());
247 for (const MachineBasicBlock &BB : MF) {
248 for (const MachineInstr &I : BB)
249 for (const MachineOperand &Op : I.operands()) {
250 if (!Op.isJTI()) continue;
252 // Remember that this JT is live.
253 JTIsLive.set(Op.getIndex());
257 // Finally, remove dead jump tables. This happens when the
258 // indirect jump was unreachable (and thus deleted).
259 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
260 if (!JTIsLive.test(i)) {
261 JTI->RemoveJumpTable(i);
269 //===----------------------------------------------------------------------===//
270 // Tail Merging of Blocks
271 //===----------------------------------------------------------------------===//
273 /// HashMachineInstr - Compute a hash value for MI and its operands.
274 static unsigned HashMachineInstr(const MachineInstr *MI) {
275 unsigned Hash = MI->getOpcode();
276 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
277 const MachineOperand &Op = MI->getOperand(i);
279 // Merge in bits from the operand if easy. We can't use MachineOperand's
280 // hash_code here because it's not deterministic and we sort by hash value
282 unsigned OperandHash = 0;
283 switch (Op.getType()) {
284 case MachineOperand::MO_Register:
285 OperandHash = Op.getReg();
287 case MachineOperand::MO_Immediate:
288 OperandHash = Op.getImm();
290 case MachineOperand::MO_MachineBasicBlock:
291 OperandHash = Op.getMBB()->getNumber();
293 case MachineOperand::MO_FrameIndex:
294 case MachineOperand::MO_ConstantPoolIndex:
295 case MachineOperand::MO_JumpTableIndex:
296 OperandHash = Op.getIndex();
298 case MachineOperand::MO_GlobalAddress:
299 case MachineOperand::MO_ExternalSymbol:
300 // Global address / external symbol are too hard, don't bother, but do
301 // pull in the offset.
302 OperandHash = Op.getOffset();
308 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
313 /// HashEndOfMBB - Hash the last instruction in the MBB.
314 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
315 MachineBasicBlock::const_iterator I = MBB->getLastNonDebugInstr();
319 return HashMachineInstr(I);
322 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
323 /// of instructions they actually have in common together at their end. Return
324 /// iterators for the first shared instruction in each block.
325 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
326 MachineBasicBlock *MBB2,
327 MachineBasicBlock::iterator &I1,
328 MachineBasicBlock::iterator &I2) {
332 unsigned TailLen = 0;
333 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
335 // Skip debugging pseudos; necessary to avoid changing the code.
336 while (I1->isDebugValue()) {
337 if (I1==MBB1->begin()) {
338 while (I2->isDebugValue()) {
339 if (I2==MBB2->begin())
340 // I1==DBG at begin; I2==DBG at begin
345 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
350 // I1==first (untested) non-DBG preceding known match
351 while (I2->isDebugValue()) {
352 if (I2==MBB2->begin()) {
354 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
359 // I1, I2==first (untested) non-DBGs preceding known match
360 if (!I1->isIdenticalTo(I2) ||
361 // FIXME: This check is dubious. It's used to get around a problem where
362 // people incorrectly expect inline asm directives to remain in the same
363 // relative order. This is untenable because normal compiler
364 // optimizations (like this one) may reorder and/or merge these
372 // Back past possible debugging pseudos at beginning of block. This matters
373 // when one block differs from the other only by whether debugging pseudos
374 // are present at the beginning. (This way, the various checks later for
375 // I1==MBB1->begin() work as expected.)
376 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
378 while (I2->isDebugValue()) {
379 if (I2 == MBB2->begin())
385 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
387 while (I1->isDebugValue()) {
388 if (I1 == MBB1->begin())
397 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
398 MachineBasicBlock *NewMBB) {
400 RS->enterBasicBlock(CurMBB);
401 if (!CurMBB->empty())
402 RS->forward(std::prev(CurMBB->end()));
403 for (unsigned int i = 1, e = TRI->getNumRegs(); i != e; i++)
404 if (RS->isRegUsed(i, false))
405 NewMBB->addLiveIn(i);
409 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
410 /// after it, replacing it with an unconditional branch to NewDest.
411 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
412 MachineBasicBlock *NewDest) {
413 MachineBasicBlock *CurMBB = OldInst->getParent();
415 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
417 // For targets that use the register scavenger, we must maintain LiveIns.
418 MaintainLiveIns(CurMBB, NewDest);
423 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
424 /// MBB so that the part before the iterator falls into the part starting at the
425 /// iterator. This returns the new MBB.
426 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
427 MachineBasicBlock::iterator BBI1,
428 const BasicBlock *BB) {
429 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
432 MachineFunction &MF = *CurMBB.getParent();
434 // Create the fall-through block.
435 MachineFunction::iterator MBBI = CurMBB.getIterator();
436 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
437 CurMBB.getParent()->insert(++MBBI, NewMBB);
439 // Move all the successors of this block to the specified block.
440 NewMBB->transferSuccessors(&CurMBB);
442 // Add an edge from CurMBB to NewMBB for the fall-through.
443 CurMBB.addSuccessor(NewMBB);
445 // Splice the code over.
446 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
448 // NewMBB inherits CurMBB's block frequency.
449 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
451 // For targets that use the register scavenger, we must maintain LiveIns.
452 MaintainLiveIns(&CurMBB, NewMBB);
454 // Add the new block to the funclet.
455 const auto &FuncletI = FuncletMembership.find(&CurMBB);
456 if (FuncletI != FuncletMembership.end())
457 FuncletMembership[NewMBB] = FuncletI->second;
462 /// EstimateRuntime - Make a rough estimate for how long it will take to run
463 /// the specified code.
464 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
465 MachineBasicBlock::iterator E) {
467 for (; I != E; ++I) {
468 if (I->isDebugValue())
472 else if (I->mayLoad() || I->mayStore())
480 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
481 // branches temporarily for tail merging). In the case where CurMBB ends
482 // with a conditional branch to the next block, optimize by reversing the
483 // test and conditionally branching to SuccMBB instead.
484 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
485 const TargetInstrInfo *TII) {
486 MachineFunction *MF = CurMBB->getParent();
487 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
488 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
489 SmallVector<MachineOperand, 4> Cond;
490 DebugLoc dl; // FIXME: this is nowhere
491 if (I != MF->end() &&
492 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
493 MachineBasicBlock *NextBB = &*I;
494 if (TBB == NextBB && !Cond.empty() && !FBB) {
495 if (!TII->ReverseBranchCondition(Cond)) {
496 TII->RemoveBranch(*CurMBB);
497 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
502 TII->InsertBranch(*CurMBB, SuccBB, nullptr,
503 SmallVector<MachineOperand, 0>(), dl);
507 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
508 if (getHash() < o.getHash())
510 if (getHash() > o.getHash())
512 if (getBlock()->getNumber() < o.getBlock()->getNumber())
514 if (getBlock()->getNumber() > o.getBlock()->getNumber())
516 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
517 // an object with itself.
518 #ifndef _GLIBCXX_DEBUG
519 llvm_unreachable("Predecessor appears twice");
526 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
527 auto I = MergedBBFreq.find(MBB);
529 if (I != MergedBBFreq.end())
532 return MBFI.getBlockFreq(MBB);
535 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
537 MergedBBFreq[MBB] = F;
540 /// CountTerminators - Count the number of terminators in the given
541 /// block and set I to the position of the first non-terminator, if there
542 /// is one, or MBB->end() otherwise.
543 static unsigned CountTerminators(MachineBasicBlock *MBB,
544 MachineBasicBlock::iterator &I) {
546 unsigned NumTerms = 0;
548 if (I == MBB->begin()) {
553 if (!I->isTerminator()) break;
559 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
560 /// and decide if it would be profitable to merge those tails. Return the
561 /// length of the common tail and iterators to the first common instruction
564 ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2,
565 unsigned minCommonTailLength, unsigned &CommonTailLen,
566 MachineBasicBlock::iterator &I1,
567 MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB,
568 MachineBasicBlock *PredBB,
569 DenseMap<const MachineBasicBlock *, int> &FuncletMembership) {
570 // It is never profitable to tail-merge blocks from two different funclets.
571 if (!FuncletMembership.empty()) {
572 auto Funclet1 = FuncletMembership.find(MBB1);
573 assert(Funclet1 != FuncletMembership.end());
574 auto Funclet2 = FuncletMembership.find(MBB2);
575 assert(Funclet2 != FuncletMembership.end());
576 if (Funclet1->second != Funclet2->second)
580 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
581 if (CommonTailLen == 0)
583 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
584 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
587 // It's almost always profitable to merge any number of non-terminator
588 // instructions with the block that falls through into the common successor.
589 if (MBB1 == PredBB || MBB2 == PredBB) {
590 MachineBasicBlock::iterator I;
591 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
592 if (CommonTailLen > NumTerms)
596 // If one of the blocks can be completely merged and happens to be in
597 // a position where the other could fall through into it, merge any number
598 // of instructions, because it can be done without a branch.
599 // TODO: If the blocks are not adjacent, move one of them so that they are?
600 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
602 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
605 // If both blocks have an unconditional branch temporarily stripped out,
606 // count that as an additional common instruction for the following
608 unsigned EffectiveTailLen = CommonTailLen;
609 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
610 !MBB1->back().isBarrier() &&
611 !MBB2->back().isBarrier())
614 // Check if the common tail is long enough to be worthwhile.
615 if (EffectiveTailLen >= minCommonTailLength)
618 // If we are optimizing for code size, 2 instructions in common is enough if
619 // we don't have to split a block. At worst we will be introducing 1 new
620 // branch instruction, which is likely to be smaller than the 2
621 // instructions that would be deleted in the merge.
622 MachineFunction *MF = MBB1->getParent();
623 return EffectiveTailLen >= 2 && MF->getFunction()->optForSize() &&
624 (I1 == MBB1->begin() || I2 == MBB2->begin());
627 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
628 /// hash CurHash (guaranteed to match the last element). Build the vector
629 /// SameTails of all those that have the (same) largest number of instructions
630 /// in common of any pair of these blocks. SameTails entries contain an
631 /// iterator into MergePotentials (from which the MachineBasicBlock can be
632 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
633 /// instruction where the matching code sequence begins.
634 /// Order of elements in SameTails is the reverse of the order in which
635 /// those blocks appear in MergePotentials (where they are not necessarily
637 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
638 unsigned minCommonTailLength,
639 MachineBasicBlock *SuccBB,
640 MachineBasicBlock *PredBB) {
641 unsigned maxCommonTailLength = 0U;
643 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
644 MPIterator HighestMPIter = std::prev(MergePotentials.end());
645 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
646 B = MergePotentials.begin();
647 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
648 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
649 unsigned CommonTailLen;
650 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
652 CommonTailLen, TrialBBI1, TrialBBI2,
654 FuncletMembership)) {
655 if (CommonTailLen > maxCommonTailLength) {
657 maxCommonTailLength = CommonTailLen;
658 HighestMPIter = CurMPIter;
659 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
661 if (HighestMPIter == CurMPIter &&
662 CommonTailLen == maxCommonTailLength)
663 SameTails.push_back(SameTailElt(I, TrialBBI2));
669 return maxCommonTailLength;
672 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
673 /// MergePotentials, restoring branches at ends of blocks as appropriate.
674 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
675 MachineBasicBlock *SuccBB,
676 MachineBasicBlock *PredBB) {
677 MPIterator CurMPIter, B;
678 for (CurMPIter = std::prev(MergePotentials.end()),
679 B = MergePotentials.begin();
680 CurMPIter->getHash() == CurHash; --CurMPIter) {
681 // Put the unconditional branch back, if we need one.
682 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
683 if (SuccBB && CurMBB != PredBB)
684 FixTail(CurMBB, SuccBB, TII);
688 if (CurMPIter->getHash() != CurHash)
690 MergePotentials.erase(CurMPIter, MergePotentials.end());
693 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
694 /// only of the common tail. Create a block that does by splitting one.
695 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
696 MachineBasicBlock *SuccBB,
697 unsigned maxCommonTailLength,
698 unsigned &commonTailIndex) {
700 unsigned TimeEstimate = ~0U;
701 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
702 // Use PredBB if possible; that doesn't require a new branch.
703 if (SameTails[i].getBlock() == PredBB) {
707 // Otherwise, make a (fairly bogus) choice based on estimate of
708 // how long it will take the various blocks to execute.
709 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
710 SameTails[i].getTailStartPos());
711 if (t <= TimeEstimate) {
717 MachineBasicBlock::iterator BBI =
718 SameTails[commonTailIndex].getTailStartPos();
719 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
721 // If the common tail includes any debug info we will take it pretty
722 // randomly from one of the inputs. Might be better to remove it?
723 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
724 << maxCommonTailLength);
726 // If the split block unconditionally falls-thru to SuccBB, it will be
727 // merged. In control flow terms it should then take SuccBB's name. e.g. If
728 // SuccBB is an inner loop, the common tail is still part of the inner loop.
729 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
730 SuccBB->getBasicBlock() : MBB->getBasicBlock();
731 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
733 DEBUG(dbgs() << "... failed!");
737 SameTails[commonTailIndex].setBlock(newMBB);
738 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
740 // If we split PredBB, newMBB is the new predecessor.
747 // Add MI1's MMOs to MI2's MMOs while excluding any duplicates. The MI scheduler
748 // currently doesn't handle multiple MMOs, so duplicates would likely pessimize
750 static void mergeMMOs(MachineInstr *MI1, MachineInstr *MI2) {
751 auto I1 = MI1->memoperands_begin(), E1 = MI1->memoperands_end();
752 auto I2 = MI2->memoperands_begin(), E2 = MI2->memoperands_end();
753 MachineFunction *MF = MI1->getParent()->getParent();
755 // Mostly, MI1's MMO count is 1 or zero. So we don't have to use
757 for (; I1 != E1; ++I1) {
758 bool IsDupMMO = false;
759 for (I2 = MI2->memoperands_begin(); I2 != E2; ++I2) {
765 if (IsDupMMO == false) {
766 MI2->addMemOperand(*MF, *I1);
767 E2 = MI2->memoperands_end();
773 mergeMMOsFromMemoryOperations(MachineBasicBlock::iterator MBBIStartPos,
774 MachineBasicBlock &MBBCommon) {
775 // Merge MMOs from memory operations in the common block
776 MachineBasicBlock *MBB = MBBIStartPos->getParent();
777 // Note CommonTailLen does not necessarily matches the size of
778 // the common BB nor all its instructions because of debug
779 // instructions differences.
780 unsigned CommonTailLen = 0;
781 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
784 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
785 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
786 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
787 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
789 while (CommonTailLen--) {
790 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
793 if (MBBI->isDebugValue()) {
798 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
801 assert(MBBICommon != MBBIECommon &&
802 "Reached BB end within common tail length!");
803 assert(MBBICommon->isIdenticalTo(&*MBBI) && "Expected matching MIIs!");
805 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
806 mergeMMOs(&*MBBI, &*MBBICommon);
813 // See if any of the blocks in MergePotentials (which all have a common single
814 // successor, or all have no successor) can be tail-merged. If there is a
815 // successor, any blocks in MergePotentials that are not tail-merged and
816 // are not immediately before Succ must have an unconditional branch to
817 // Succ added (but the predecessor/successor lists need no adjustment).
818 // The lone predecessor of Succ that falls through into Succ,
819 // if any, is given in PredBB.
821 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
822 MachineBasicBlock *PredBB) {
823 bool MadeChange = false;
825 // Except for the special cases below, tail-merge if there are at least
826 // this many instructions in common.
827 unsigned minCommonTailLength = TailMergeSize;
829 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
830 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
831 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
832 << (i == e-1 ? "" : ", ");
835 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
837 dbgs() << " which has fall-through from BB#"
838 << PredBB->getNumber() << "\n";
840 dbgs() << "Looking for common tails of at least "
841 << minCommonTailLength << " instruction"
842 << (minCommonTailLength == 1 ? "" : "s") << '\n';
845 // Sort by hash value so that blocks with identical end sequences sort
847 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
849 // Walk through equivalence sets looking for actual exact matches.
850 while (MergePotentials.size() > 1) {
851 unsigned CurHash = MergePotentials.back().getHash();
853 // Build SameTails, identifying the set of blocks with this hash code
854 // and with the maximum number of instructions in common.
855 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
859 // If we didn't find any pair that has at least minCommonTailLength
860 // instructions in common, remove all blocks with this hash code and retry.
861 if (SameTails.empty()) {
862 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
866 // If one of the blocks is the entire common tail (and not the entry
867 // block, which we can't jump to), we can treat all blocks with this same
868 // tail at once. Use PredBB if that is one of the possibilities, as that
869 // will not introduce any extra branches.
870 MachineBasicBlock *EntryBB =
871 &MergePotentials.front().getBlock()->getParent()->front();
872 unsigned commonTailIndex = SameTails.size();
873 // If there are two blocks, check to see if one can be made to fall through
875 if (SameTails.size() == 2 &&
876 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
877 SameTails[1].tailIsWholeBlock())
879 else if (SameTails.size() == 2 &&
880 SameTails[1].getBlock()->isLayoutSuccessor(
881 SameTails[0].getBlock()) &&
882 SameTails[0].tailIsWholeBlock())
885 // Otherwise just pick one, favoring the fall-through predecessor if
887 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
888 MachineBasicBlock *MBB = SameTails[i].getBlock();
889 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
895 if (SameTails[i].tailIsWholeBlock())
900 if (commonTailIndex == SameTails.size() ||
901 (SameTails[commonTailIndex].getBlock() == PredBB &&
902 !SameTails[commonTailIndex].tailIsWholeBlock())) {
903 // None of the blocks consist entirely of the common tail.
904 // Split a block so that one does.
905 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
906 maxCommonTailLength, commonTailIndex)) {
907 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
912 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
914 // Recompute commont tail MBB's edge weights and block frequency.
915 setCommonTailEdgeWeights(*MBB);
917 // MBB is common tail. Adjust all other BB's to jump to this one.
918 // Traversal must be forwards so erases work.
919 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
921 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
922 if (commonTailIndex == i)
924 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
925 << (i == e-1 ? "" : ", "));
926 // Merge MMOs from memory operations as needed.
927 mergeMMOsFromMemoryOperations(SameTails[i].getTailStartPos(), *MBB);
928 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
929 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
930 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
931 MergePotentials.erase(SameTails[i].getMPIter());
933 DEBUG(dbgs() << "\n");
934 // We leave commonTailIndex in the worklist in case there are other blocks
935 // that match it with a smaller number of instructions.
941 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
942 bool MadeChange = false;
943 if (!EnableTailMerge) return MadeChange;
945 // First find blocks with no successors.
946 MergePotentials.clear();
947 for (MachineBasicBlock &MBB : MF) {
948 if (MergePotentials.size() == TailMergeThreshold)
950 if (!TriedMerging.count(&MBB) && MBB.succ_empty())
951 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(&MBB), &MBB));
954 // If this is a large problem, avoid visiting the same basic blocks
956 if (MergePotentials.size() == TailMergeThreshold)
957 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
958 TriedMerging.insert(MergePotentials[i].getBlock());
960 // See if we can do any tail merging on those.
961 if (MergePotentials.size() >= 2)
962 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
964 // Look at blocks (IBB) with multiple predecessors (PBB).
965 // We change each predecessor to a canonical form, by
966 // (1) temporarily removing any unconditional branch from the predecessor
968 // (2) alter conditional branches so they branch to the other block
969 // not IBB; this may require adding back an unconditional branch to IBB
970 // later, where there wasn't one coming in. E.g.
972 // fallthrough to QBB
975 // with a conceptual B to IBB after that, which never actually exists.
976 // With those changes, we see whether the predecessors' tails match,
977 // and merge them if so. We change things out of canonical form and
978 // back to the way they were later in the process. (OptimizeBranches
979 // would undo some of this, but we can't use it, because we'd get into
980 // a compile-time infinite loop repeatedly doing and undoing the same
983 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
985 if (I->pred_size() < 2) continue;
986 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
987 MachineBasicBlock *IBB = &*I;
988 MachineBasicBlock *PredBB = &*std::prev(I);
989 MergePotentials.clear();
990 for (MachineBasicBlock *PBB : I->predecessors()) {
991 if (MergePotentials.size() == TailMergeThreshold)
994 if (TriedMerging.count(PBB))
997 // Skip blocks that loop to themselves, can't tail merge these.
1001 // Visit each predecessor only once.
1002 if (!UniquePreds.insert(PBB).second)
1005 // Skip blocks which may jump to a landing pad. Can't tail merge these.
1006 if (PBB->hasEHPadSuccessor())
1009 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1010 SmallVector<MachineOperand, 4> Cond;
1011 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
1012 // Failing case: IBB is the target of a cbr, and we cannot reverse the
1014 SmallVector<MachineOperand, 4> NewCond(Cond);
1015 if (!Cond.empty() && TBB == IBB) {
1016 if (TII->ReverseBranchCondition(NewCond))
1018 // This is the QBB case described above
1020 auto Next = ++PBB->getIterator();
1021 if (Next != MF.end())
1026 // Failing case: the only way IBB can be reached from PBB is via
1027 // exception handling. Happens for landing pads. Would be nice to have
1028 // a bit in the edge so we didn't have to do all this.
1029 if (IBB->isEHPad()) {
1030 MachineFunction::iterator IP = ++PBB->getIterator();
1031 MachineBasicBlock *PredNextBB = nullptr;
1035 if (IBB != PredNextBB) // fallthrough
1038 if (TBB != IBB && FBB != IBB) // cbr then ubr
1040 } else if (Cond.empty()) {
1041 if (TBB != IBB) // ubr
1044 if (TBB != IBB && IBB != PredNextBB) // cbr
1049 // Remove the unconditional branch at the end, if any.
1050 if (TBB && (Cond.empty() || FBB)) {
1051 DebugLoc dl; // FIXME: this is nowhere
1052 TII->RemoveBranch(*PBB);
1054 // reinsert conditional branch only, for now
1055 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1059 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), PBB));
1063 // If this is a large problem, avoid visiting the same basic blocks multiple
1065 if (MergePotentials.size() == TailMergeThreshold)
1066 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1067 TriedMerging.insert(MergePotentials[i].getBlock());
1069 if (MergePotentials.size() >= 2)
1070 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
1072 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1073 // result of removing blocks in TryTailMergeBlocks.
1074 PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks
1075 if (MergePotentials.size() == 1 &&
1076 MergePotentials.begin()->getBlock() != PredBB)
1077 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1083 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1084 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1085 BlockFrequency AccumulatedMBBFreq;
1087 // Aggregate edge frequency of successor edge j:
1088 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1089 // where bb is a basic block that is in SameTails.
1090 for (const auto &Src : SameTails) {
1091 const MachineBasicBlock *SrcMBB = Src.getBlock();
1092 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1093 AccumulatedMBBFreq += BlockFreq;
1095 // It is not necessary to recompute edge weights if TailBB has less than two
1097 if (TailMBB.succ_size() <= 1)
1100 auto EdgeFreq = EdgeFreqLs.begin();
1102 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1103 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1104 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1107 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1109 if (TailMBB.succ_size() <= 1)
1113 std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0))
1115 auto EdgeFreq = EdgeFreqLs.begin();
1117 if (SumEdgeFreq > 0) {
1118 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1119 SuccI != SuccE; ++SuccI, ++EdgeFreq) {
1120 auto Prob = BranchProbability::getBranchProbability(
1121 EdgeFreq->getFrequency(), SumEdgeFreq);
1122 TailMBB.setSuccProbability(SuccI, Prob);
1127 //===----------------------------------------------------------------------===//
1128 // Branch Optimization
1129 //===----------------------------------------------------------------------===//
1131 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1132 bool MadeChange = false;
1134 // Make sure blocks are numbered in order
1135 MF.RenumberBlocks();
1136 // Renumbering blocks alters funclet membership, recalculate it.
1137 FuncletMembership = getFuncletMembership(MF);
1139 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1141 MachineBasicBlock *MBB = &*I++;
1142 MadeChange |= OptimizeBlock(MBB);
1144 // If it is dead, remove it.
1145 if (MBB->pred_empty()) {
1146 RemoveDeadBlock(MBB);
1155 // Blocks should be considered empty if they contain only debug info;
1156 // else the debug info would affect codegen.
1157 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1158 return MBB->getFirstNonDebugInstr() == MBB->end();
1161 // Blocks with only debug info and branches should be considered the same
1162 // as blocks with only branches.
1163 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1164 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1165 assert(I != MBB->end() && "empty block!");
1166 return I->isBranch();
1169 /// IsBetterFallthrough - Return true if it would be clearly better to
1170 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1171 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1172 /// result in infinite loops.
1173 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1174 MachineBasicBlock *MBB2) {
1175 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1176 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1177 // optimize branches that branch to either a return block or an assert block
1178 // into a fallthrough to the return.
1179 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1180 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1181 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1184 // If there is a clear successor ordering we make sure that one block
1185 // will fall through to the next
1186 if (MBB1->isSuccessor(MBB2)) return true;
1187 if (MBB2->isSuccessor(MBB1)) return false;
1189 return MBB2I->isCall() && !MBB1I->isCall();
1192 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1193 /// instructions on the block.
1194 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1195 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1196 if (I != MBB.end() && I->isBranch())
1197 return I->getDebugLoc();
1201 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1202 /// block. This is never called on the entry block.
1203 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1204 bool MadeChange = false;
1205 MachineFunction &MF = *MBB->getParent();
1208 MachineFunction::iterator FallThrough = MBB->getIterator();
1211 // Make sure MBB and FallThrough belong to the same funclet.
1212 bool SameFunclet = true;
1213 if (!FuncletMembership.empty() && FallThrough != MF.end()) {
1214 auto MBBFunclet = FuncletMembership.find(MBB);
1215 assert(MBBFunclet != FuncletMembership.end());
1216 auto FallThroughFunclet = FuncletMembership.find(&*FallThrough);
1217 assert(FallThroughFunclet != FuncletMembership.end());
1218 SameFunclet = MBBFunclet->second == FallThroughFunclet->second;
1221 // If this block is empty, make everyone use its fall-through, not the block
1222 // explicitly. Landing pads should not do this since the landing-pad table
1223 // points to this block. Blocks with their addresses taken shouldn't be
1225 if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() &&
1227 // Dead block? Leave for cleanup later.
1228 if (MBB->pred_empty()) return MadeChange;
1230 if (FallThrough == MF.end()) {
1231 // TODO: Simplify preds to not branch here if possible!
1232 } else if (FallThrough->isEHPad()) {
1233 // Don't rewrite to a landing pad fallthough. That could lead to the case
1234 // where a BB jumps to more than one landing pad.
1235 // TODO: Is it ever worth rewriting predecessors which don't already
1236 // jump to a landing pad, and so can safely jump to the fallthrough?
1238 // Rewrite all predecessors of the old block to go to the fallthrough
1240 while (!MBB->pred_empty()) {
1241 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1242 Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough);
1244 // If MBB was the target of a jump table, update jump tables to go to the
1245 // fallthrough instead.
1246 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1247 MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough);
1253 // Check to see if we can simplify the terminator of the block before this
1255 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1257 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1258 SmallVector<MachineOperand, 4> PriorCond;
1259 bool PriorUnAnalyzable =
1260 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1261 if (!PriorUnAnalyzable) {
1262 // If the CFG for the prior block has extra edges, remove them.
1263 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1264 !PriorCond.empty());
1266 // If the previous branch is conditional and both conditions go to the same
1267 // destination, remove the branch, replacing it with an unconditional one or
1269 if (PriorTBB && PriorTBB == PriorFBB) {
1270 DebugLoc dl = getBranchDebugLoc(PrevBB);
1271 TII->RemoveBranch(PrevBB);
1273 if (PriorTBB != MBB)
1274 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1277 goto ReoptimizeBlock;
1280 // If the previous block unconditionally falls through to this block and
1281 // this block has no other predecessors, move the contents of this block
1282 // into the prior block. This doesn't usually happen when SimplifyCFG
1283 // has been used, but it can happen if tail merging splits a fall-through
1284 // predecessor of a block.
1285 // This has to check PrevBB->succ_size() because EH edges are ignored by
1287 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1288 PrevBB.succ_size() == 1 &&
1289 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1290 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1291 << "From MBB: " << *MBB);
1292 // Remove redundant DBG_VALUEs first.
1293 if (PrevBB.begin() != PrevBB.end()) {
1294 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1296 MachineBasicBlock::iterator MBBIter = MBB->begin();
1297 // Check if DBG_VALUE at the end of PrevBB is identical to the
1298 // DBG_VALUE at the beginning of MBB.
1299 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1300 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1301 if (!MBBIter->isIdenticalTo(PrevBBIter))
1303 MachineInstr *DuplicateDbg = MBBIter;
1304 ++MBBIter; -- PrevBBIter;
1305 DuplicateDbg->eraseFromParent();
1308 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1309 PrevBB.removeSuccessor(PrevBB.succ_begin());
1310 assert(PrevBB.succ_empty());
1311 PrevBB.transferSuccessors(MBB);
1316 // If the previous branch *only* branches to *this* block (conditional or
1317 // not) remove the branch.
1318 if (PriorTBB == MBB && !PriorFBB) {
1319 TII->RemoveBranch(PrevBB);
1322 goto ReoptimizeBlock;
1325 // If the prior block branches somewhere else on the condition and here if
1326 // the condition is false, remove the uncond second branch.
1327 if (PriorFBB == MBB) {
1328 DebugLoc dl = getBranchDebugLoc(PrevBB);
1329 TII->RemoveBranch(PrevBB);
1330 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1333 goto ReoptimizeBlock;
1336 // If the prior block branches here on true and somewhere else on false, and
1337 // if the branch condition is reversible, reverse the branch to create a
1339 if (PriorTBB == MBB) {
1340 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1341 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1342 DebugLoc dl = getBranchDebugLoc(PrevBB);
1343 TII->RemoveBranch(PrevBB);
1344 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1347 goto ReoptimizeBlock;
1351 // If this block has no successors (e.g. it is a return block or ends with
1352 // a call to a no-return function like abort or __cxa_throw) and if the pred
1353 // falls through into this block, and if it would otherwise fall through
1354 // into the block after this, move this block to the end of the function.
1356 // We consider it more likely that execution will stay in the function (e.g.
1357 // due to loops) than it is to exit it. This asserts in loops etc, moving
1358 // the assert condition out of the loop body.
1359 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1360 MachineFunction::iterator(PriorTBB) == FallThrough &&
1361 !MBB->canFallThrough()) {
1362 bool DoTransform = true;
1364 // We have to be careful that the succs of PredBB aren't both no-successor
1365 // blocks. If neither have successors and if PredBB is the second from
1366 // last block in the function, we'd just keep swapping the two blocks for
1367 // last. Only do the swap if one is clearly better to fall through than
1369 if (FallThrough == --MF.end() &&
1370 !IsBetterFallthrough(PriorTBB, MBB))
1371 DoTransform = false;
1374 // Reverse the branch so we will fall through on the previous true cond.
1375 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1376 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1377 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1378 << "To make fallthrough to: " << *PriorTBB << "\n");
1380 DebugLoc dl = getBranchDebugLoc(PrevBB);
1381 TII->RemoveBranch(PrevBB);
1382 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1384 // Move this block to the end of the function.
1385 MBB->moveAfter(&MF.back());
1394 // Analyze the branch in the current block.
1395 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1396 SmallVector<MachineOperand, 4> CurCond;
1397 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1398 if (!CurUnAnalyzable) {
1399 // If the CFG for the prior block has extra edges, remove them.
1400 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1402 // If this is a two-way branch, and the FBB branches to this block, reverse
1403 // the condition so the single-basic-block loop is faster. Instead of:
1404 // Loop: xxx; jcc Out; jmp Loop
1406 // Loop: xxx; jncc Loop; jmp Out
1407 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1408 SmallVector<MachineOperand, 4> NewCond(CurCond);
1409 if (!TII->ReverseBranchCondition(NewCond)) {
1410 DebugLoc dl = getBranchDebugLoc(*MBB);
1411 TII->RemoveBranch(*MBB);
1412 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1415 goto ReoptimizeBlock;
1419 // If this branch is the only thing in its block, see if we can forward
1420 // other blocks across it.
1421 if (CurTBB && CurCond.empty() && !CurFBB &&
1422 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1423 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1424 DebugLoc dl = getBranchDebugLoc(*MBB);
1425 // This block may contain just an unconditional branch. Because there can
1426 // be 'non-branch terminators' in the block, try removing the branch and
1427 // then seeing if the block is empty.
1428 TII->RemoveBranch(*MBB);
1429 // If the only things remaining in the block are debug info, remove these
1430 // as well, so this will behave the same as an empty block in non-debug
1432 if (IsEmptyBlock(MBB)) {
1433 // Make the block empty, losing the debug info (we could probably
1434 // improve this in some cases.)
1435 MBB->erase(MBB->begin(), MBB->end());
1437 // If this block is just an unconditional branch to CurTBB, we can
1438 // usually completely eliminate the block. The only case we cannot
1439 // completely eliminate the block is when the block before this one
1440 // falls through into MBB and we can't understand the prior block's branch
1443 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1444 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1445 !PrevBB.isSuccessor(MBB)) {
1446 // If the prior block falls through into us, turn it into an
1447 // explicit branch to us to make updates simpler.
1448 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1449 PriorTBB != MBB && PriorFBB != MBB) {
1451 assert(PriorCond.empty() && !PriorFBB &&
1452 "Bad branch analysis");
1455 assert(!PriorFBB && "Machine CFG out of date!");
1458 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1459 TII->RemoveBranch(PrevBB);
1460 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1463 // Iterate through all the predecessors, revectoring each in-turn.
1465 bool DidChange = false;
1466 bool HasBranchToSelf = false;
1467 while(PI != MBB->pred_size()) {
1468 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1470 // If this block has an uncond branch to itself, leave it.
1472 HasBranchToSelf = true;
1475 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1476 // If this change resulted in PMBB ending in a conditional
1477 // branch where both conditions go to the same destination,
1478 // change this to an unconditional branch (and fix the CFG).
1479 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1480 SmallVector<MachineOperand, 4> NewCurCond;
1481 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1482 NewCurFBB, NewCurCond, true);
1483 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1484 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1485 TII->RemoveBranch(*PMBB);
1487 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1490 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1495 // Change any jumptables to go to the new MBB.
1496 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1497 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1501 if (!HasBranchToSelf) return MadeChange;
1506 // Add the branch back if the block is more than just an uncond branch.
1507 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1511 // If the prior block doesn't fall through into this block, and if this
1512 // block doesn't fall through into some other block, see if we can find a
1513 // place to move this block where a fall-through will happen.
1514 if (!PrevBB.canFallThrough()) {
1516 // Now we know that there was no fall-through into this block, check to
1517 // see if it has a fall-through into its successor.
1518 bool CurFallsThru = MBB->canFallThrough();
1520 if (!MBB->isEHPad()) {
1521 // Check all the predecessors of this block. If one of them has no fall
1522 // throughs, move this block right after it.
1523 for (MachineBasicBlock *PredBB : MBB->predecessors()) {
1524 // Analyze the branch at the end of the pred.
1525 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1526 SmallVector<MachineOperand, 4> PredCond;
1527 if (PredBB != MBB && !PredBB->canFallThrough() &&
1528 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1529 && (!CurFallsThru || !CurTBB || !CurFBB)
1530 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1531 // If the current block doesn't fall through, just move it.
1532 // If the current block can fall through and does not end with a
1533 // conditional branch, we need to append an unconditional jump to
1534 // the (current) next block. To avoid a possible compile-time
1535 // infinite loop, move blocks only backward in this case.
1536 // Also, if there are already 2 branches here, we cannot add a third;
1537 // this means we have the case
1542 MachineBasicBlock *NextBB = &*std::next(MBB->getIterator());
1544 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1546 MBB->moveAfter(PredBB);
1548 goto ReoptimizeBlock;
1553 if (!CurFallsThru) {
1554 // Check all successors to see if we can move this block before it.
1555 for (MachineBasicBlock *SuccBB : MBB->successors()) {
1556 // Analyze the branch at the end of the block before the succ.
1557 MachineFunction::iterator SuccPrev = --SuccBB->getIterator();
1559 // If this block doesn't already fall-through to that successor, and if
1560 // the succ doesn't already have a block that can fall through into it,
1561 // and if the successor isn't an EH destination, we can arrange for the
1562 // fallthrough to happen.
1563 if (SuccBB != MBB && &*SuccPrev != MBB &&
1564 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1565 !SuccBB->isEHPad()) {
1566 MBB->moveBefore(SuccBB);
1568 goto ReoptimizeBlock;
1572 // Okay, there is no really great place to put this block. If, however,
1573 // the block before this one would be a fall-through if this block were
1574 // removed, move this block to the end of the function.
1575 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1576 SmallVector<MachineOperand, 4> PrevCond;
1577 // We're looking for cases where PrevBB could possibly fall through to
1578 // FallThrough, but if FallThrough is an EH pad that wouldn't be useful
1579 // so here we skip over any EH pads so we might have a chance to find
1580 // a branch target from PrevBB.
1581 while (FallThrough != MF.end() && FallThrough->isEHPad())
1583 // Now check to see if the current block is sitting between PrevBB and
1584 // a block to which it could fall through.
1585 if (FallThrough != MF.end() &&
1586 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1587 PrevBB.isSuccessor(&*FallThrough)) {
1588 MBB->moveAfter(&MF.back());
1598 //===----------------------------------------------------------------------===//
1599 // Hoist Common Code
1600 //===----------------------------------------------------------------------===//
1602 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1603 /// blocks to their common predecessor.
1604 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1605 bool MadeChange = false;
1606 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1607 MachineBasicBlock *MBB = &*I++;
1608 MadeChange |= HoistCommonCodeInSuccs(MBB);
1614 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1615 /// its 'true' successor.
1616 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1617 MachineBasicBlock *TrueBB) {
1618 for (MachineBasicBlock *SuccBB : BB->successors())
1619 if (SuccBB != TrueBB)
1624 template <class Container>
1625 static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
1627 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1628 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1635 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1636 /// in successors to. The location is usually just before the terminator,
1637 /// however if the terminator is a conditional branch and its previous
1638 /// instruction is the flag setting instruction, the previous instruction is
1639 /// the preferred location. This function also gathers uses and defs of the
1640 /// instructions from the insertion point to the end of the block. The data is
1641 /// used by HoistCommonCodeInSuccs to ensure safety.
1643 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1644 const TargetInstrInfo *TII,
1645 const TargetRegisterInfo *TRI,
1646 SmallSet<unsigned,4> &Uses,
1647 SmallSet<unsigned,4> &Defs) {
1648 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1649 if (!TII->isUnpredicatedTerminator(Loc))
1652 for (const MachineOperand &MO : Loc->operands()) {
1655 unsigned Reg = MO.getReg();
1659 addRegAndItsAliases(Reg, TRI, Uses);
1662 // Don't try to hoist code in the rare case the terminator defines a
1663 // register that is later used.
1666 // If the terminator defines a register, make sure we don't hoist
1667 // the instruction whose def might be clobbered by the terminator.
1668 addRegAndItsAliases(Reg, TRI, Defs);
1674 if (Loc == MBB->begin())
1677 // The terminator is probably a conditional branch, try not to separate the
1678 // branch from condition setting instruction.
1679 MachineBasicBlock::iterator PI = Loc;
1681 while (PI != MBB->begin() && PI->isDebugValue())
1685 for (const MachineOperand &MO : PI->operands()) {
1686 // If PI has a regmask operand, it is probably a call. Separate away.
1689 if (!MO.isReg() || MO.isUse())
1691 unsigned Reg = MO.getReg();
1694 if (Uses.count(Reg)) {
1700 // The condition setting instruction is not just before the conditional
1704 // Be conservative, don't insert instruction above something that may have
1705 // side-effects. And since it's potentially bad to separate flag setting
1706 // instruction from the conditional branch, just abort the optimization
1708 // Also avoid moving code above predicated instruction since it's hard to
1709 // reason about register liveness with predicated instruction.
1710 bool DontMoveAcrossStore = true;
1711 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(PI))
1715 // Find out what registers are live. Note this routine is ignoring other live
1716 // registers which are only used by instructions in successor blocks.
1717 for (const MachineOperand &MO : PI->operands()) {
1720 unsigned Reg = MO.getReg();
1724 addRegAndItsAliases(Reg, TRI, Uses);
1726 if (Uses.erase(Reg)) {
1727 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1728 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1729 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1732 addRegAndItsAliases(Reg, TRI, Defs);
1739 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1740 /// sequence at the start of the function, move the instructions before MBB
1741 /// terminator if it's legal.
1742 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1743 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1744 SmallVector<MachineOperand, 4> Cond;
1745 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1748 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1750 // Malformed bcc? True and false blocks are the same?
1753 // Restrict the optimization to cases where MBB is the only predecessor,
1754 // it is an obvious win.
1755 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1758 // Find a suitable position to hoist the common instructions to. Also figure
1759 // out which registers are used or defined by instructions from the insertion
1760 // point to the end of the block.
1761 SmallSet<unsigned, 4> Uses, Defs;
1762 MachineBasicBlock::iterator Loc =
1763 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1764 if (Loc == MBB->end())
1767 bool HasDups = false;
1768 SmallVector<unsigned, 4> LocalDefs;
1769 SmallSet<unsigned, 4> LocalDefsSet;
1770 MachineBasicBlock::iterator TIB = TBB->begin();
1771 MachineBasicBlock::iterator FIB = FBB->begin();
1772 MachineBasicBlock::iterator TIE = TBB->end();
1773 MachineBasicBlock::iterator FIE = FBB->end();
1774 while (TIB != TIE && FIB != FIE) {
1775 // Skip dbg_value instructions. These do not count.
1776 if (TIB->isDebugValue()) {
1777 while (TIB != TIE && TIB->isDebugValue())
1782 if (FIB->isDebugValue()) {
1783 while (FIB != FIE && FIB->isDebugValue())
1788 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1791 if (TII->isPredicated(TIB))
1792 // Hard to reason about register liveness with predicated instruction.
1796 for (MachineOperand &MO : TIB->operands()) {
1797 // Don't attempt to hoist instructions with register masks.
1798 if (MO.isRegMask()) {
1804 unsigned Reg = MO.getReg();
1808 if (Uses.count(Reg)) {
1809 // Avoid clobbering a register that's used by the instruction at
1810 // the point of insertion.
1815 if (Defs.count(Reg) && !MO.isDead()) {
1816 // Don't hoist the instruction if the def would be clobber by the
1817 // instruction at the point insertion. FIXME: This is overly
1818 // conservative. It should be possible to hoist the instructions
1819 // in BB2 in the following example:
1821 // r1, eflag = op1 r2, r3
1830 } else if (!LocalDefsSet.count(Reg)) {
1831 if (Defs.count(Reg)) {
1832 // Use is defined by the instruction at the point of insertion.
1837 if (MO.isKill() && Uses.count(Reg))
1838 // Kills a register that's read by the instruction at the point of
1839 // insertion. Remove the kill marker.
1840 MO.setIsKill(false);
1846 bool DontMoveAcrossStore = true;
1847 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1850 // Remove kills from LocalDefsSet, these registers had short live ranges.
1851 for (const MachineOperand &MO : TIB->operands()) {
1852 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1854 unsigned Reg = MO.getReg();
1855 if (!Reg || !LocalDefsSet.count(Reg))
1857 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1858 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1859 LocalDefsSet.erase(*AI);
1861 LocalDefsSet.erase(Reg);
1865 // Track local defs so we can update liveins.
1866 for (const MachineOperand &MO : TIB->operands()) {
1867 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1869 unsigned Reg = MO.getReg();
1872 LocalDefs.push_back(Reg);
1873 addRegAndItsAliases(Reg, TRI, LocalDefsSet);
1884 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1885 FBB->erase(FBB->begin(), FIB);
1888 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1889 unsigned Def = LocalDefs[i];
1890 if (LocalDefsSet.count(Def)) {
1891 TBB->addLiveIn(Def);
1892 FBB->addLiveIn(Def);