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
17 //===----------------------------------------------------------------------===//
19 #include "BranchFolding.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
24 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
25 #include "llvm/CodeGen/MachineFunctionPass.h"
26 #include "llvm/CodeGen/MachineJumpTableInfo.h"
27 #include "llvm/CodeGen/MachineMemOperand.h"
28 #include "llvm/CodeGen/MachineModuleInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/Passes.h"
31 #include "llvm/CodeGen/RegisterScavenging.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Target/TargetInstrInfo.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "branchfolding"
45 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
46 STATISTIC(NumBranchOpts, "Number of branches optimized");
47 STATISTIC(NumTailMerge , "Number of block tails merged");
48 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
50 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
51 cl::init(cl::BOU_UNSET), cl::Hidden);
53 // Throttle for huge numbers of predecessors (compile speed problems)
54 static cl::opt<unsigned>
55 TailMergeThreshold("tail-merge-threshold",
56 cl::desc("Max number of predecessors to consider tail merging"),
57 cl::init(150), cl::Hidden);
59 // Heuristic for tail merging (and, inversely, tail duplication).
60 // TODO: This should be replaced with a target query.
61 static cl::opt<unsigned>
62 TailMergeSize("tail-merge-size",
63 cl::desc("Min number of instructions to consider tail merging"),
64 cl::init(3), cl::Hidden);
67 /// BranchFolderPass - Wrap branch folder in a machine function pass.
68 class BranchFolderPass : public MachineFunctionPass {
71 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
73 bool runOnMachineFunction(MachineFunction &MF) override;
75 void getAnalysisUsage(AnalysisUsage &AU) const override {
76 AU.addRequired<MachineBlockFrequencyInfo>();
77 AU.addRequired<MachineBranchProbabilityInfo>();
78 AU.addRequired<TargetPassConfig>();
79 MachineFunctionPass::getAnalysisUsage(AU);
84 char BranchFolderPass::ID = 0;
85 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
87 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
88 "Control Flow Optimizer", false, false)
90 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
91 if (skipOptnoneFunction(*MF.getFunction()))
94 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
95 // TailMerge can create jump into if branches that make CFG irreducible for
96 // HW that requires structurized CFG.
97 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
98 PassConfig->getEnableTailMerge();
99 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true,
100 getAnalysis<MachineBlockFrequencyInfo>(),
101 getAnalysis<MachineBranchProbabilityInfo>());
102 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
103 MF.getSubtarget().getRegisterInfo(),
104 getAnalysisIfAvailable<MachineModuleInfo>());
107 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
108 const MachineBlockFrequencyInfo &FreqInfo,
109 const MachineBranchProbabilityInfo &ProbInfo)
110 : EnableHoistCommonCode(CommonHoist), MBBFreqInfo(FreqInfo),
112 switch (FlagEnableTailMerge) {
113 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
114 case cl::BOU_TRUE: EnableTailMerge = true; break;
115 case cl::BOU_FALSE: EnableTailMerge = false; break;
119 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
120 /// function, updating the CFG.
121 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
122 assert(MBB->pred_empty() && "MBB must be dead!");
123 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
125 MachineFunction *MF = MBB->getParent();
126 // drop all successors.
127 while (!MBB->succ_empty())
128 MBB->removeSuccessor(MBB->succ_end()-1);
130 // Avoid matching if this pointer gets reused.
131 TriedMerging.erase(MBB);
137 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
138 /// followed by terminators, and if the implicitly defined registers are not
139 /// used by the terminators, remove those implicit_def's. e.g.
141 /// r0 = implicit_def
142 /// r1 = implicit_def
144 /// This block can be optimized away later if the implicit instructions are
146 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
147 SmallSet<unsigned, 4> ImpDefRegs;
148 MachineBasicBlock::iterator I = MBB->begin();
149 while (I != MBB->end()) {
150 if (!I->isImplicitDef())
152 unsigned Reg = I->getOperand(0).getReg();
153 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
154 SubRegs.isValid(); ++SubRegs)
155 ImpDefRegs.insert(*SubRegs);
158 if (ImpDefRegs.empty())
161 MachineBasicBlock::iterator FirstTerm = I;
162 while (I != MBB->end()) {
163 if (!TII->isUnpredicatedTerminator(I))
165 // See if it uses any of the implicitly defined registers.
166 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
167 MachineOperand &MO = I->getOperand(i);
168 if (!MO.isReg() || !MO.isUse())
170 unsigned Reg = MO.getReg();
171 if (ImpDefRegs.count(Reg))
178 while (I != FirstTerm) {
179 MachineInstr *ImpDefMI = &*I;
181 MBB->erase(ImpDefMI);
187 /// OptimizeFunction - Perhaps branch folding, tail merging and other
188 /// CFG optimizations on the given function.
189 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
190 const TargetInstrInfo *tii,
191 const TargetRegisterInfo *tri,
192 MachineModuleInfo *mmi) {
193 if (!tii) return false;
195 TriedMerging.clear();
202 // Use a RegScavenger to help update liveness when required.
203 MachineRegisterInfo &MRI = MF.getRegInfo();
204 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
205 RS = new RegScavenger();
207 MRI.invalidateLiveness();
209 // Fix CFG. The later algorithms expect it to be right.
210 bool MadeChange = false;
211 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
212 MachineBasicBlock *MBB = I, *TBB = nullptr, *FBB = nullptr;
213 SmallVector<MachineOperand, 4> Cond;
214 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
215 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
216 MadeChange |= OptimizeImpDefsBlock(MBB);
219 bool MadeChangeThisIteration = true;
220 while (MadeChangeThisIteration) {
221 MadeChangeThisIteration = TailMergeBlocks(MF);
222 MadeChangeThisIteration |= OptimizeBranches(MF);
223 if (EnableHoistCommonCode)
224 MadeChangeThisIteration |= HoistCommonCode(MF);
225 MadeChange |= MadeChangeThisIteration;
228 // See if any jump tables have become dead as the code generator
230 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
236 // Walk the function to find jump tables that are live.
237 BitVector JTIsLive(JTI->getJumpTables().size());
238 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
240 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
242 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
243 MachineOperand &Op = I->getOperand(op);
244 if (!Op.isJTI()) continue;
246 // Remember that this JT is live.
247 JTIsLive.set(Op.getIndex());
251 // Finally, remove dead jump tables. This happens when the
252 // indirect jump was unreachable (and thus deleted).
253 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
254 if (!JTIsLive.test(i)) {
255 JTI->RemoveJumpTable(i);
263 //===----------------------------------------------------------------------===//
264 // Tail Merging of Blocks
265 //===----------------------------------------------------------------------===//
267 /// HashMachineInstr - Compute a hash value for MI and its operands.
268 static unsigned HashMachineInstr(const MachineInstr *MI) {
269 unsigned Hash = MI->getOpcode();
270 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
271 const MachineOperand &Op = MI->getOperand(i);
273 // Merge in bits from the operand if easy.
274 unsigned OperandHash = 0;
275 switch (Op.getType()) {
276 case MachineOperand::MO_Register:
277 OperandHash = Op.getReg();
279 case MachineOperand::MO_Immediate:
280 OperandHash = Op.getImm();
282 case MachineOperand::MO_MachineBasicBlock:
283 OperandHash = Op.getMBB()->getNumber();
285 case MachineOperand::MO_FrameIndex:
286 case MachineOperand::MO_ConstantPoolIndex:
287 case MachineOperand::MO_JumpTableIndex:
288 OperandHash = Op.getIndex();
290 case MachineOperand::MO_GlobalAddress:
291 case MachineOperand::MO_ExternalSymbol:
292 // Global address / external symbol are too hard, don't bother, but do
293 // pull in the offset.
294 OperandHash = Op.getOffset();
300 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
305 /// HashEndOfMBB - Hash the last instruction in the MBB.
306 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
307 MachineBasicBlock::const_iterator I = MBB->end();
308 if (I == MBB->begin())
309 return 0; // Empty MBB.
312 // Skip debug info so it will not affect codegen.
313 while (I->isDebugValue()) {
314 if (I == MBB->begin())
315 return 0; // MBB empty except for debug info.
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;
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);
457 /// EstimateRuntime - Make a rough estimate for how long it will take to run
458 /// the specified code.
459 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
460 MachineBasicBlock::iterator E) {
462 for (; I != E; ++I) {
463 if (I->isDebugValue())
467 else if (I->mayLoad() || I->mayStore())
475 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
476 // branches temporarily for tail merging). In the case where CurMBB ends
477 // with a conditional branch to the next block, optimize by reversing the
478 // test and conditionally branching to SuccMBB instead.
479 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
480 const TargetInstrInfo *TII) {
481 MachineFunction *MF = CurMBB->getParent();
482 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
483 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
484 SmallVector<MachineOperand, 4> Cond;
485 DebugLoc dl; // FIXME: this is nowhere
486 if (I != MF->end() &&
487 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
488 MachineBasicBlock *NextBB = I;
489 if (TBB == NextBB && !Cond.empty() && !FBB) {
490 if (!TII->ReverseBranchCondition(Cond)) {
491 TII->RemoveBranch(*CurMBB);
492 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
497 TII->InsertBranch(*CurMBB, SuccBB, nullptr,
498 SmallVector<MachineOperand, 0>(), dl);
502 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
503 if (getHash() < o.getHash())
505 if (getHash() > o.getHash())
507 if (getBlock()->getNumber() < o.getBlock()->getNumber())
509 if (getBlock()->getNumber() > o.getBlock()->getNumber())
511 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
512 // an object with itself.
513 #ifndef _GLIBCXX_DEBUG
514 llvm_unreachable("Predecessor appears twice");
521 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
522 auto I = MergedBBFreq.find(MBB);
524 if (I != MergedBBFreq.end())
527 return MBFI.getBlockFreq(MBB);
530 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
532 MergedBBFreq[MBB] = F;
535 /// CountTerminators - Count the number of terminators in the given
536 /// block and set I to the position of the first non-terminator, if there
537 /// is one, or MBB->end() otherwise.
538 static unsigned CountTerminators(MachineBasicBlock *MBB,
539 MachineBasicBlock::iterator &I) {
541 unsigned NumTerms = 0;
543 if (I == MBB->begin()) {
548 if (!I->isTerminator()) break;
554 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
555 /// and decide if it would be profitable to merge those tails. Return the
556 /// length of the common tail and iterators to the first common instruction
558 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
559 MachineBasicBlock *MBB2,
560 unsigned minCommonTailLength,
561 unsigned &CommonTailLen,
562 MachineBasicBlock::iterator &I1,
563 MachineBasicBlock::iterator &I2,
564 MachineBasicBlock *SuccBB,
565 MachineBasicBlock *PredBB) {
566 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
567 if (CommonTailLen == 0)
569 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
570 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
573 // It's almost always profitable to merge any number of non-terminator
574 // instructions with the block that falls through into the common successor.
575 if (MBB1 == PredBB || MBB2 == PredBB) {
576 MachineBasicBlock::iterator I;
577 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
578 if (CommonTailLen > NumTerms)
582 // If one of the blocks can be completely merged and happens to be in
583 // a position where the other could fall through into it, merge any number
584 // of instructions, because it can be done without a branch.
585 // TODO: If the blocks are not adjacent, move one of them so that they are?
586 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
588 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
591 // If both blocks have an unconditional branch temporarily stripped out,
592 // count that as an additional common instruction for the following
594 unsigned EffectiveTailLen = CommonTailLen;
595 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
596 !MBB1->back().isBarrier() &&
597 !MBB2->back().isBarrier())
600 // Check if the common tail is long enough to be worthwhile.
601 if (EffectiveTailLen >= minCommonTailLength)
604 // If we are optimizing for code size, 2 instructions in common is enough if
605 // we don't have to split a block. At worst we will be introducing 1 new
606 // branch instruction, which is likely to be smaller than the 2
607 // instructions that would be deleted in the merge.
608 MachineFunction *MF = MBB1->getParent();
609 if (EffectiveTailLen >= 2 &&
610 MF->getFunction()->hasFnAttribute(Attribute::OptimizeForSize) &&
611 (I1 == MBB1->begin() || I2 == MBB2->begin()))
617 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
618 /// hash CurHash (guaranteed to match the last element). Build the vector
619 /// SameTails of all those that have the (same) largest number of instructions
620 /// in common of any pair of these blocks. SameTails entries contain an
621 /// iterator into MergePotentials (from which the MachineBasicBlock can be
622 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
623 /// instruction where the matching code sequence begins.
624 /// Order of elements in SameTails is the reverse of the order in which
625 /// those blocks appear in MergePotentials (where they are not necessarily
627 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
628 unsigned minCommonTailLength,
629 MachineBasicBlock *SuccBB,
630 MachineBasicBlock *PredBB) {
631 unsigned maxCommonTailLength = 0U;
633 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
634 MPIterator HighestMPIter = std::prev(MergePotentials.end());
635 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
636 B = MergePotentials.begin();
637 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
638 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
639 unsigned CommonTailLen;
640 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
642 CommonTailLen, TrialBBI1, TrialBBI2,
644 if (CommonTailLen > maxCommonTailLength) {
646 maxCommonTailLength = CommonTailLen;
647 HighestMPIter = CurMPIter;
648 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
650 if (HighestMPIter == CurMPIter &&
651 CommonTailLen == maxCommonTailLength)
652 SameTails.push_back(SameTailElt(I, TrialBBI2));
658 return maxCommonTailLength;
661 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
662 /// MergePotentials, restoring branches at ends of blocks as appropriate.
663 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
664 MachineBasicBlock *SuccBB,
665 MachineBasicBlock *PredBB) {
666 MPIterator CurMPIter, B;
667 for (CurMPIter = std::prev(MergePotentials.end()),
668 B = MergePotentials.begin();
669 CurMPIter->getHash() == CurHash; --CurMPIter) {
670 // Put the unconditional branch back, if we need one.
671 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
672 if (SuccBB && CurMBB != PredBB)
673 FixTail(CurMBB, SuccBB, TII);
677 if (CurMPIter->getHash() != CurHash)
679 MergePotentials.erase(CurMPIter, MergePotentials.end());
682 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
683 /// only of the common tail. Create a block that does by splitting one.
684 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
685 MachineBasicBlock *SuccBB,
686 unsigned maxCommonTailLength,
687 unsigned &commonTailIndex) {
689 unsigned TimeEstimate = ~0U;
690 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
691 // Use PredBB if possible; that doesn't require a new branch.
692 if (SameTails[i].getBlock() == PredBB) {
696 // Otherwise, make a (fairly bogus) choice based on estimate of
697 // how long it will take the various blocks to execute.
698 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
699 SameTails[i].getTailStartPos());
700 if (t <= TimeEstimate) {
706 MachineBasicBlock::iterator BBI =
707 SameTails[commonTailIndex].getTailStartPos();
708 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
710 // If the common tail includes any debug info we will take it pretty
711 // randomly from one of the inputs. Might be better to remove it?
712 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
713 << maxCommonTailLength);
715 // If the split block unconditionally falls-thru to SuccBB, it will be
716 // merged. In control flow terms it should then take SuccBB's name. e.g. If
717 // SuccBB is an inner loop, the common tail is still part of the inner loop.
718 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
719 SuccBB->getBasicBlock() : MBB->getBasicBlock();
720 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
722 DEBUG(dbgs() << "... failed!");
726 SameTails[commonTailIndex].setBlock(newMBB);
727 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
729 // If we split PredBB, newMBB is the new predecessor.
736 static bool hasIdenticalMMOs(const MachineInstr *MI1, const MachineInstr *MI2) {
737 auto I1 = MI1->memoperands_begin(), E1 = MI1->memoperands_end();
738 auto I2 = MI2->memoperands_begin(), E2 = MI2->memoperands_end();
739 if ((E1 - I1) != (E2 - I2))
741 for (; I1 != E1; ++I1, ++I2) {
749 removeMMOsFromMemoryOperations(MachineBasicBlock::iterator MBBIStartPos,
750 MachineBasicBlock &MBBCommon) {
751 // Remove MMOs from memory operations in the common block
752 // when they do not match the ones from the block being tail-merged.
753 // This ensures later passes conservatively compute dependencies.
754 MachineBasicBlock *MBB = MBBIStartPos->getParent();
755 // Note CommonTailLen does not necessarily matches the size of
756 // the common BB nor all its instructions because of debug
757 // instructions differences.
758 unsigned CommonTailLen = 0;
759 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
762 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
763 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
764 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
765 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
767 while (CommonTailLen--) {
768 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
771 if (MBBI->isDebugValue()) {
776 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
779 assert(MBBICommon != MBBIECommon &&
780 "Reached BB end within common tail length!");
781 assert(MBBICommon->isIdenticalTo(&*MBBI) && "Expected matching MIIs!");
783 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
784 if (!hasIdenticalMMOs(&*MBBI, &*MBBICommon))
785 MBBICommon->clearMemRefs();
792 // See if any of the blocks in MergePotentials (which all have a common single
793 // successor, or all have no successor) can be tail-merged. If there is a
794 // successor, any blocks in MergePotentials that are not tail-merged and
795 // are not immediately before Succ must have an unconditional branch to
796 // Succ added (but the predecessor/successor lists need no adjustment).
797 // The lone predecessor of Succ that falls through into Succ,
798 // if any, is given in PredBB.
800 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
801 MachineBasicBlock *PredBB) {
802 bool MadeChange = false;
804 // Except for the special cases below, tail-merge if there are at least
805 // this many instructions in common.
806 unsigned minCommonTailLength = TailMergeSize;
808 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
809 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
810 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
811 << (i == e-1 ? "" : ", ");
814 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
816 dbgs() << " which has fall-through from BB#"
817 << PredBB->getNumber() << "\n";
819 dbgs() << "Looking for common tails of at least "
820 << minCommonTailLength << " instruction"
821 << (minCommonTailLength == 1 ? "" : "s") << '\n';
824 // Sort by hash value so that blocks with identical end sequences sort
826 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
828 // Walk through equivalence sets looking for actual exact matches.
829 while (MergePotentials.size() > 1) {
830 unsigned CurHash = MergePotentials.back().getHash();
832 // Build SameTails, identifying the set of blocks with this hash code
833 // and with the maximum number of instructions in common.
834 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
838 // If we didn't find any pair that has at least minCommonTailLength
839 // instructions in common, remove all blocks with this hash code and retry.
840 if (SameTails.empty()) {
841 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
845 // If one of the blocks is the entire common tail (and not the entry
846 // block, which we can't jump to), we can treat all blocks with this same
847 // tail at once. Use PredBB if that is one of the possibilities, as that
848 // will not introduce any extra branches.
849 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
850 getParent()->begin();
851 unsigned commonTailIndex = SameTails.size();
852 // If there are two blocks, check to see if one can be made to fall through
854 if (SameTails.size() == 2 &&
855 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
856 SameTails[1].tailIsWholeBlock())
858 else if (SameTails.size() == 2 &&
859 SameTails[1].getBlock()->isLayoutSuccessor(
860 SameTails[0].getBlock()) &&
861 SameTails[0].tailIsWholeBlock())
864 // Otherwise just pick one, favoring the fall-through predecessor if
866 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
867 MachineBasicBlock *MBB = SameTails[i].getBlock();
868 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
874 if (SameTails[i].tailIsWholeBlock())
879 if (commonTailIndex == SameTails.size() ||
880 (SameTails[commonTailIndex].getBlock() == PredBB &&
881 !SameTails[commonTailIndex].tailIsWholeBlock())) {
882 // None of the blocks consist entirely of the common tail.
883 // Split a block so that one does.
884 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
885 maxCommonTailLength, commonTailIndex)) {
886 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
891 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
893 // Recompute commont tail MBB's edge weights and block frequency.
894 setCommonTailEdgeWeights(*MBB);
896 // MBB is common tail. Adjust all other BB's to jump to this one.
897 // Traversal must be forwards so erases work.
898 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
900 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
901 if (commonTailIndex == i)
903 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
904 << (i == e-1 ? "" : ", "));
905 // Remove MMOs from memory operations as needed.
906 removeMMOsFromMemoryOperations(SameTails[i].getTailStartPos(), *MBB);
907 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
908 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
909 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
910 MergePotentials.erase(SameTails[i].getMPIter());
912 DEBUG(dbgs() << "\n");
913 // We leave commonTailIndex in the worklist in case there are other blocks
914 // that match it with a smaller number of instructions.
920 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
921 bool MadeChange = false;
922 if (!EnableTailMerge) return MadeChange;
924 // First find blocks with no successors.
925 MergePotentials.clear();
926 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
927 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
928 if (TriedMerging.count(I))
931 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
934 // If this is a large problem, avoid visiting the same basic blocks
936 if (MergePotentials.size() == TailMergeThreshold)
937 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
938 TriedMerging.insert(MergePotentials[i].getBlock());
940 // See if we can do any tail merging on those.
941 if (MergePotentials.size() >= 2)
942 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
944 // Look at blocks (IBB) with multiple predecessors (PBB).
945 // We change each predecessor to a canonical form, by
946 // (1) temporarily removing any unconditional branch from the predecessor
948 // (2) alter conditional branches so they branch to the other block
949 // not IBB; this may require adding back an unconditional branch to IBB
950 // later, where there wasn't one coming in. E.g.
952 // fallthrough to QBB
955 // with a conceptual B to IBB after that, which never actually exists.
956 // With those changes, we see whether the predecessors' tails match,
957 // and merge them if so. We change things out of canonical form and
958 // back to the way they were later in the process. (OptimizeBranches
959 // would undo some of this, but we can't use it, because we'd get into
960 // a compile-time infinite loop repeatedly doing and undoing the same
963 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
965 if (I->pred_size() < 2) continue;
966 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
967 MachineBasicBlock *IBB = I;
968 MachineBasicBlock *PredBB = std::prev(I);
969 MergePotentials.clear();
970 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
972 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
973 MachineBasicBlock *PBB = *P;
974 if (TriedMerging.count(PBB))
977 // Skip blocks that loop to themselves, can't tail merge these.
981 // Visit each predecessor only once.
982 if (!UniquePreds.insert(PBB).second)
985 // Skip blocks which may jump to a landing pad. Can't tail merge these.
986 if (PBB->getLandingPadSuccessor())
989 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
990 SmallVector<MachineOperand, 4> Cond;
991 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
992 // Failing case: IBB is the target of a cbr, and we cannot reverse the
994 SmallVector<MachineOperand, 4> NewCond(Cond);
995 if (!Cond.empty() && TBB == IBB) {
996 if (TII->ReverseBranchCondition(NewCond))
998 // This is the QBB case described above
1000 FBB = std::next(MachineFunction::iterator(PBB));
1003 // Failing case: the only way IBB can be reached from PBB is via
1004 // exception handling. Happens for landing pads. Would be nice to have
1005 // a bit in the edge so we didn't have to do all this.
1006 if (IBB->isLandingPad()) {
1007 MachineFunction::iterator IP = PBB; IP++;
1008 MachineBasicBlock *PredNextBB = nullptr;
1012 if (IBB != PredNextBB) // fallthrough
1015 if (TBB != IBB && FBB != IBB) // cbr then ubr
1017 } else if (Cond.empty()) {
1018 if (TBB != IBB) // ubr
1021 if (TBB != IBB && IBB != PredNextBB) // cbr
1026 // Remove the unconditional branch at the end, if any.
1027 if (TBB && (Cond.empty() || FBB)) {
1028 DebugLoc dl; // FIXME: this is nowhere
1029 TII->RemoveBranch(*PBB);
1031 // reinsert conditional branch only, for now
1032 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1036 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
1040 // If this is a large problem, avoid visiting the same basic blocks multiple
1042 if (MergePotentials.size() == TailMergeThreshold)
1043 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1044 TriedMerging.insert(MergePotentials[i].getBlock());
1046 if (MergePotentials.size() >= 2)
1047 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
1049 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1050 // result of removing blocks in TryTailMergeBlocks.
1051 PredBB = std::prev(I); // this may have been changed in TryTailMergeBlocks
1052 if (MergePotentials.size() == 1 &&
1053 MergePotentials.begin()->getBlock() != PredBB)
1054 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1060 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1061 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1062 BlockFrequency AccumulatedMBBFreq;
1064 // Aggregate edge frequency of successor edge j:
1065 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1066 // where bb is a basic block that is in SameTails.
1067 for (const auto &Src : SameTails) {
1068 const MachineBasicBlock *SrcMBB = Src.getBlock();
1069 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1070 AccumulatedMBBFreq += BlockFreq;
1072 // It is not necessary to recompute edge weights if TailBB has less than two
1074 if (TailMBB.succ_size() <= 1)
1077 auto EdgeFreq = EdgeFreqLs.begin();
1079 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1080 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1081 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1084 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1086 if (TailMBB.succ_size() <= 1)
1089 auto MaxEdgeFreq = *std::max_element(EdgeFreqLs.begin(), EdgeFreqLs.end());
1090 uint64_t Scale = MaxEdgeFreq.getFrequency() / UINT32_MAX + 1;
1091 auto EdgeFreq = EdgeFreqLs.begin();
1093 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1094 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1095 TailMBB.setSuccWeight(SuccI, EdgeFreq->getFrequency() / Scale);
1098 //===----------------------------------------------------------------------===//
1099 // Branch Optimization
1100 //===----------------------------------------------------------------------===//
1102 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1103 bool MadeChange = false;
1105 // Make sure blocks are numbered in order
1106 MF.RenumberBlocks();
1108 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1110 MachineBasicBlock *MBB = I++;
1111 MadeChange |= OptimizeBlock(MBB);
1113 // If it is dead, remove it.
1114 if (MBB->pred_empty()) {
1115 RemoveDeadBlock(MBB);
1123 // Blocks should be considered empty if they contain only debug info;
1124 // else the debug info would affect codegen.
1125 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1128 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
1129 MBBI!=MBBE; ++MBBI) {
1130 if (!MBBI->isDebugValue())
1136 // Blocks with only debug info and branches should be considered the same
1137 // as blocks with only branches.
1138 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1139 MachineBasicBlock::iterator MBBI, MBBE;
1140 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
1141 if (!MBBI->isDebugValue())
1144 return (MBBI->isBranch());
1147 /// IsBetterFallthrough - Return true if it would be clearly better to
1148 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1149 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1150 /// result in infinite loops.
1151 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1152 MachineBasicBlock *MBB2) {
1153 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1154 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1155 // optimize branches that branch to either a return block or an assert block
1156 // into a fallthrough to the return.
1157 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
1159 // If there is a clear successor ordering we make sure that one block
1160 // will fall through to the next
1161 if (MBB1->isSuccessor(MBB2)) return true;
1162 if (MBB2->isSuccessor(MBB1)) return false;
1164 // Neither block consists entirely of debug info (per IsEmptyBlock check),
1165 // so we needn't test for falling off the beginning here.
1166 MachineBasicBlock::iterator MBB1I = --MBB1->end();
1167 while (MBB1I->isDebugValue())
1169 MachineBasicBlock::iterator MBB2I = --MBB2->end();
1170 while (MBB2I->isDebugValue())
1172 return MBB2I->isCall() && !MBB1I->isCall();
1175 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1176 /// instructions on the block. Always use the DebugLoc of the first
1177 /// branching instruction found unless its absent, in which case use the
1178 /// DebugLoc of the second if present.
1179 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1180 MachineBasicBlock::iterator I = MBB.end();
1181 if (I == MBB.begin())
1184 while (I->isDebugValue() && I != MBB.begin())
1187 return I->getDebugLoc();
1191 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1192 /// block. This is never called on the entry block.
1193 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1194 bool MadeChange = false;
1195 MachineFunction &MF = *MBB->getParent();
1198 MachineFunction::iterator FallThrough = MBB;
1201 // If this block is empty, make everyone use its fall-through, not the block
1202 // explicitly. Landing pads should not do this since the landing-pad table
1203 // points to this block. Blocks with their addresses taken shouldn't be
1205 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1206 // Dead block? Leave for cleanup later.
1207 if (MBB->pred_empty()) return MadeChange;
1209 if (FallThrough == MF.end()) {
1210 // TODO: Simplify preds to not branch here if possible!
1211 } else if (FallThrough->isLandingPad()) {
1212 // Don't rewrite to a landing pad fallthough. That could lead to the case
1213 // where a BB jumps to more than one landing pad.
1214 // TODO: Is it ever worth rewriting predecessors which don't already
1215 // jump to a landing pad, and so can safely jump to the fallthrough?
1217 // Rewrite all predecessors of the old block to go to the fallthrough
1219 while (!MBB->pred_empty()) {
1220 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1221 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1223 // If MBB was the target of a jump table, update jump tables to go to the
1224 // fallthrough instead.
1225 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1226 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1232 // Check to see if we can simplify the terminator of the block before this
1234 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1236 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1237 SmallVector<MachineOperand, 4> PriorCond;
1238 bool PriorUnAnalyzable =
1239 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1240 if (!PriorUnAnalyzable) {
1241 // If the CFG for the prior block has extra edges, remove them.
1242 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1243 !PriorCond.empty());
1245 // If the previous branch is conditional and both conditions go to the same
1246 // destination, remove the branch, replacing it with an unconditional one or
1248 if (PriorTBB && PriorTBB == PriorFBB) {
1249 DebugLoc dl = getBranchDebugLoc(PrevBB);
1250 TII->RemoveBranch(PrevBB);
1252 if (PriorTBB != MBB)
1253 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1256 goto ReoptimizeBlock;
1259 // If the previous block unconditionally falls through to this block and
1260 // this block has no other predecessors, move the contents of this block
1261 // into the prior block. This doesn't usually happen when SimplifyCFG
1262 // has been used, but it can happen if tail merging splits a fall-through
1263 // predecessor of a block.
1264 // This has to check PrevBB->succ_size() because EH edges are ignored by
1266 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1267 PrevBB.succ_size() == 1 &&
1268 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1269 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1270 << "From MBB: " << *MBB);
1271 // Remove redundant DBG_VALUEs first.
1272 if (PrevBB.begin() != PrevBB.end()) {
1273 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1275 MachineBasicBlock::iterator MBBIter = MBB->begin();
1276 // Check if DBG_VALUE at the end of PrevBB is identical to the
1277 // DBG_VALUE at the beginning of MBB.
1278 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1279 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1280 if (!MBBIter->isIdenticalTo(PrevBBIter))
1282 MachineInstr *DuplicateDbg = MBBIter;
1283 ++MBBIter; -- PrevBBIter;
1284 DuplicateDbg->eraseFromParent();
1287 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1288 PrevBB.removeSuccessor(PrevBB.succ_begin());
1289 assert(PrevBB.succ_empty());
1290 PrevBB.transferSuccessors(MBB);
1295 // If the previous branch *only* branches to *this* block (conditional or
1296 // not) remove the branch.
1297 if (PriorTBB == MBB && !PriorFBB) {
1298 TII->RemoveBranch(PrevBB);
1301 goto ReoptimizeBlock;
1304 // If the prior block branches somewhere else on the condition and here if
1305 // the condition is false, remove the uncond second branch.
1306 if (PriorFBB == MBB) {
1307 DebugLoc dl = getBranchDebugLoc(PrevBB);
1308 TII->RemoveBranch(PrevBB);
1309 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1312 goto ReoptimizeBlock;
1315 // If the prior block branches here on true and somewhere else on false, and
1316 // if the branch condition is reversible, reverse the branch to create a
1318 if (PriorTBB == MBB) {
1319 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1320 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1321 DebugLoc dl = getBranchDebugLoc(PrevBB);
1322 TII->RemoveBranch(PrevBB);
1323 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1326 goto ReoptimizeBlock;
1330 // If this block has no successors (e.g. it is a return block or ends with
1331 // a call to a no-return function like abort or __cxa_throw) and if the pred
1332 // falls through into this block, and if it would otherwise fall through
1333 // into the block after this, move this block to the end of the function.
1335 // We consider it more likely that execution will stay in the function (e.g.
1336 // due to loops) than it is to exit it. This asserts in loops etc, moving
1337 // the assert condition out of the loop body.
1338 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1339 MachineFunction::iterator(PriorTBB) == FallThrough &&
1340 !MBB->canFallThrough()) {
1341 bool DoTransform = true;
1343 // We have to be careful that the succs of PredBB aren't both no-successor
1344 // blocks. If neither have successors and if PredBB is the second from
1345 // last block in the function, we'd just keep swapping the two blocks for
1346 // last. Only do the swap if one is clearly better to fall through than
1348 if (FallThrough == --MF.end() &&
1349 !IsBetterFallthrough(PriorTBB, MBB))
1350 DoTransform = false;
1353 // Reverse the branch so we will fall through on the previous true cond.
1354 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1355 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1356 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1357 << "To make fallthrough to: " << *PriorTBB << "\n");
1359 DebugLoc dl = getBranchDebugLoc(PrevBB);
1360 TII->RemoveBranch(PrevBB);
1361 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1363 // Move this block to the end of the function.
1364 MBB->moveAfter(--MF.end());
1373 // Analyze the branch in the current block.
1374 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1375 SmallVector<MachineOperand, 4> CurCond;
1376 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1377 if (!CurUnAnalyzable) {
1378 // If the CFG for the prior block has extra edges, remove them.
1379 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1381 // If this is a two-way branch, and the FBB branches to this block, reverse
1382 // the condition so the single-basic-block loop is faster. Instead of:
1383 // Loop: xxx; jcc Out; jmp Loop
1385 // Loop: xxx; jncc Loop; jmp Out
1386 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1387 SmallVector<MachineOperand, 4> NewCond(CurCond);
1388 if (!TII->ReverseBranchCondition(NewCond)) {
1389 DebugLoc dl = getBranchDebugLoc(*MBB);
1390 TII->RemoveBranch(*MBB);
1391 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1394 goto ReoptimizeBlock;
1398 // If this branch is the only thing in its block, see if we can forward
1399 // other blocks across it.
1400 if (CurTBB && CurCond.empty() && !CurFBB &&
1401 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1402 !MBB->hasAddressTaken()) {
1403 DebugLoc dl = getBranchDebugLoc(*MBB);
1404 // This block may contain just an unconditional branch. Because there can
1405 // be 'non-branch terminators' in the block, try removing the branch and
1406 // then seeing if the block is empty.
1407 TII->RemoveBranch(*MBB);
1408 // If the only things remaining in the block are debug info, remove these
1409 // as well, so this will behave the same as an empty block in non-debug
1411 if (!MBB->empty()) {
1412 bool NonDebugInfoFound = false;
1413 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1415 if (!I->isDebugValue()) {
1416 NonDebugInfoFound = true;
1420 if (!NonDebugInfoFound)
1421 // Make the block empty, losing the debug info (we could probably
1422 // improve this in some cases.)
1423 MBB->erase(MBB->begin(), MBB->end());
1425 // If this block is just an unconditional branch to CurTBB, we can
1426 // usually completely eliminate the block. The only case we cannot
1427 // completely eliminate the block is when the block before this one
1428 // falls through into MBB and we can't understand the prior block's branch
1431 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1432 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1433 !PrevBB.isSuccessor(MBB)) {
1434 // If the prior block falls through into us, turn it into an
1435 // explicit branch to us to make updates simpler.
1436 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1437 PriorTBB != MBB && PriorFBB != MBB) {
1439 assert(PriorCond.empty() && !PriorFBB &&
1440 "Bad branch analysis");
1443 assert(!PriorFBB && "Machine CFG out of date!");
1446 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1447 TII->RemoveBranch(PrevBB);
1448 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1451 // Iterate through all the predecessors, revectoring each in-turn.
1453 bool DidChange = false;
1454 bool HasBranchToSelf = false;
1455 while(PI != MBB->pred_size()) {
1456 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1458 // If this block has an uncond branch to itself, leave it.
1460 HasBranchToSelf = true;
1463 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1464 // If this change resulted in PMBB ending in a conditional
1465 // branch where both conditions go to the same destination,
1466 // change this to an unconditional branch (and fix the CFG).
1467 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1468 SmallVector<MachineOperand, 4> NewCurCond;
1469 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1470 NewCurFBB, NewCurCond, true);
1471 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1472 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1473 TII->RemoveBranch(*PMBB);
1475 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1478 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1483 // Change any jumptables to go to the new MBB.
1484 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1485 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1489 if (!HasBranchToSelf) return MadeChange;
1494 // Add the branch back if the block is more than just an uncond branch.
1495 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1499 // If the prior block doesn't fall through into this block, and if this
1500 // block doesn't fall through into some other block, see if we can find a
1501 // place to move this block where a fall-through will happen.
1502 if (!PrevBB.canFallThrough()) {
1504 // Now we know that there was no fall-through into this block, check to
1505 // see if it has a fall-through into its successor.
1506 bool CurFallsThru = MBB->canFallThrough();
1508 if (!MBB->isLandingPad()) {
1509 // Check all the predecessors of this block. If one of them has no fall
1510 // throughs, move this block right after it.
1511 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1512 E = MBB->pred_end(); PI != E; ++PI) {
1513 // Analyze the branch at the end of the pred.
1514 MachineBasicBlock *PredBB = *PI;
1515 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1516 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1517 SmallVector<MachineOperand, 4> PredCond;
1518 if (PredBB != MBB && !PredBB->canFallThrough() &&
1519 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1520 && (!CurFallsThru || !CurTBB || !CurFBB)
1521 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1522 // If the current block doesn't fall through, just move it.
1523 // If the current block can fall through and does not end with a
1524 // conditional branch, we need to append an unconditional jump to
1525 // the (current) next block. To avoid a possible compile-time
1526 // infinite loop, move blocks only backward in this case.
1527 // Also, if there are already 2 branches here, we cannot add a third;
1528 // this means we have the case
1533 MachineBasicBlock *NextBB =
1534 std::next(MachineFunction::iterator(MBB));
1536 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1538 MBB->moveAfter(PredBB);
1540 goto ReoptimizeBlock;
1545 if (!CurFallsThru) {
1546 // Check all successors to see if we can move this block before it.
1547 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1548 E = MBB->succ_end(); SI != E; ++SI) {
1549 // Analyze the branch at the end of the block before the succ.
1550 MachineBasicBlock *SuccBB = *SI;
1551 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1553 // If this block doesn't already fall-through to that successor, and if
1554 // the succ doesn't already have a block that can fall through into it,
1555 // and if the successor isn't an EH destination, we can arrange for the
1556 // fallthrough to happen.
1557 if (SuccBB != MBB && &*SuccPrev != MBB &&
1558 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1559 !SuccBB->isLandingPad()) {
1560 MBB->moveBefore(SuccBB);
1562 goto ReoptimizeBlock;
1566 // Okay, there is no really great place to put this block. If, however,
1567 // the block before this one would be a fall-through if this block were
1568 // removed, move this block to the end of the function.
1569 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1570 SmallVector<MachineOperand, 4> PrevCond;
1571 if (FallThrough != MF.end() &&
1572 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1573 PrevBB.isSuccessor(FallThrough)) {
1574 MBB->moveAfter(--MF.end());
1584 //===----------------------------------------------------------------------===//
1585 // Hoist Common Code
1586 //===----------------------------------------------------------------------===//
1588 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1589 /// blocks to their common predecessor.
1590 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1591 bool MadeChange = false;
1592 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1593 MachineBasicBlock *MBB = I++;
1594 MadeChange |= HoistCommonCodeInSuccs(MBB);
1600 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1601 /// its 'true' successor.
1602 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1603 MachineBasicBlock *TrueBB) {
1604 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1605 E = BB->succ_end(); SI != E; ++SI) {
1606 MachineBasicBlock *SuccBB = *SI;
1607 if (SuccBB != TrueBB)
1613 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1614 /// in successors to. The location is usually just before the terminator,
1615 /// however if the terminator is a conditional branch and its previous
1616 /// instruction is the flag setting instruction, the previous instruction is
1617 /// the preferred location. This function also gathers uses and defs of the
1618 /// instructions from the insertion point to the end of the block. The data is
1619 /// used by HoistCommonCodeInSuccs to ensure safety.
1621 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1622 const TargetInstrInfo *TII,
1623 const TargetRegisterInfo *TRI,
1624 SmallSet<unsigned,4> &Uses,
1625 SmallSet<unsigned,4> &Defs) {
1626 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1627 if (!TII->isUnpredicatedTerminator(Loc))
1630 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1631 const MachineOperand &MO = Loc->getOperand(i);
1634 unsigned Reg = MO.getReg();
1638 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1642 // Don't try to hoist code in the rare case the terminator defines a
1643 // register that is later used.
1646 // If the terminator defines a register, make sure we don't hoist
1647 // the instruction whose def might be clobbered by the terminator.
1648 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1655 if (Loc == MBB->begin())
1658 // The terminator is probably a conditional branch, try not to separate the
1659 // branch from condition setting instruction.
1660 MachineBasicBlock::iterator PI = Loc;
1662 while (PI != MBB->begin() && PI->isDebugValue())
1666 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1667 const MachineOperand &MO = PI->getOperand(i);
1668 // If PI has a regmask operand, it is probably a call. Separate away.
1671 if (!MO.isReg() || MO.isUse())
1673 unsigned Reg = MO.getReg();
1676 if (Uses.count(Reg))
1680 // The condition setting instruction is not just before the conditional
1684 // Be conservative, don't insert instruction above something that may have
1685 // side-effects. And since it's potentially bad to separate flag setting
1686 // instruction from the conditional branch, just abort the optimization
1688 // Also avoid moving code above predicated instruction since it's hard to
1689 // reason about register liveness with predicated instruction.
1690 bool DontMoveAcrossStore = true;
1691 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(PI))
1695 // Find out what registers are live. Note this routine is ignoring other live
1696 // registers which are only used by instructions in successor blocks.
1697 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1698 const MachineOperand &MO = PI->getOperand(i);
1701 unsigned Reg = MO.getReg();
1705 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1708 if (Uses.erase(Reg)) {
1709 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1710 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1712 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1720 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1721 /// sequence at the start of the function, move the instructions before MBB
1722 /// terminator if it's legal.
1723 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1724 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1725 SmallVector<MachineOperand, 4> Cond;
1726 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1729 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1731 // Malformed bcc? True and false blocks are the same?
1734 // Restrict the optimization to cases where MBB is the only predecessor,
1735 // it is an obvious win.
1736 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1739 // Find a suitable position to hoist the common instructions to. Also figure
1740 // out which registers are used or defined by instructions from the insertion
1741 // point to the end of the block.
1742 SmallSet<unsigned, 4> Uses, Defs;
1743 MachineBasicBlock::iterator Loc =
1744 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1745 if (Loc == MBB->end())
1748 bool HasDups = false;
1749 SmallVector<unsigned, 4> LocalDefs;
1750 SmallSet<unsigned, 4> LocalDefsSet;
1751 MachineBasicBlock::iterator TIB = TBB->begin();
1752 MachineBasicBlock::iterator FIB = FBB->begin();
1753 MachineBasicBlock::iterator TIE = TBB->end();
1754 MachineBasicBlock::iterator FIE = FBB->end();
1755 while (TIB != TIE && FIB != FIE) {
1756 // Skip dbg_value instructions. These do not count.
1757 if (TIB->isDebugValue()) {
1758 while (TIB != TIE && TIB->isDebugValue())
1763 if (FIB->isDebugValue()) {
1764 while (FIB != FIE && FIB->isDebugValue())
1769 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1772 if (TII->isPredicated(TIB))
1773 // Hard to reason about register liveness with predicated instruction.
1777 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1778 MachineOperand &MO = TIB->getOperand(i);
1779 // Don't attempt to hoist instructions with register masks.
1780 if (MO.isRegMask()) {
1786 unsigned Reg = MO.getReg();
1790 if (Uses.count(Reg)) {
1791 // Avoid clobbering a register that's used by the instruction at
1792 // the point of insertion.
1797 if (Defs.count(Reg) && !MO.isDead()) {
1798 // Don't hoist the instruction if the def would be clobber by the
1799 // instruction at the point insertion. FIXME: This is overly
1800 // conservative. It should be possible to hoist the instructions
1801 // in BB2 in the following example:
1803 // r1, eflag = op1 r2, r3
1812 } else if (!LocalDefsSet.count(Reg)) {
1813 if (Defs.count(Reg)) {
1814 // Use is defined by the instruction at the point of insertion.
1819 if (MO.isKill() && Uses.count(Reg))
1820 // Kills a register that's read by the instruction at the point of
1821 // insertion. Remove the kill marker.
1822 MO.setIsKill(false);
1828 bool DontMoveAcrossStore = true;
1829 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1832 // Remove kills from LocalDefsSet, these registers had short live ranges.
1833 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1834 MachineOperand &MO = TIB->getOperand(i);
1835 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1837 unsigned Reg = MO.getReg();
1838 if (!Reg || !LocalDefsSet.count(Reg))
1840 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1841 LocalDefsSet.erase(*AI);
1844 // Track local defs so we can update liveins.
1845 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1846 MachineOperand &MO = TIB->getOperand(i);
1847 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1849 unsigned Reg = MO.getReg();
1852 LocalDefs.push_back(Reg);
1853 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1854 LocalDefsSet.insert(*AI);
1865 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1866 FBB->erase(FBB->begin(), FIB);
1869 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1870 unsigned Def = LocalDefs[i];
1871 if (LocalDefsSet.count(Def)) {
1872 TBB->addLiveIn(Def);
1873 FBB->addLiveIn(Def);