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/MachineModuleInfo.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/Passes.h"
30 #include "llvm/CodeGen/RegisterScavenging.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetInstrInfo.h"
37 #include "llvm/Target/TargetMachine.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: OperandHash = Op.getReg(); break;
277 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
278 case MachineOperand::MO_MachineBasicBlock:
279 OperandHash = Op.getMBB()->getNumber();
281 case MachineOperand::MO_FrameIndex:
282 case MachineOperand::MO_ConstantPoolIndex:
283 case MachineOperand::MO_JumpTableIndex:
284 OperandHash = Op.getIndex();
286 case MachineOperand::MO_GlobalAddress:
287 case MachineOperand::MO_ExternalSymbol:
288 // Global address / external symbol are too hard, don't bother, but do
289 // pull in the offset.
290 OperandHash = Op.getOffset();
295 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
300 /// HashEndOfMBB - Hash the last instruction in the MBB.
301 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
302 MachineBasicBlock::const_iterator I = MBB->end();
303 if (I == MBB->begin())
304 return 0; // Empty MBB.
307 // Skip debug info so it will not affect codegen.
308 while (I->isDebugValue()) {
310 return 0; // MBB empty except for debug info.
314 return HashMachineInstr(I);
317 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
318 /// of instructions they actually have in common together at their end. Return
319 /// iterators for the first shared instruction in each block.
320 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
321 MachineBasicBlock *MBB2,
322 MachineBasicBlock::iterator &I1,
323 MachineBasicBlock::iterator &I2) {
327 unsigned TailLen = 0;
328 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
330 // Skip debugging pseudos; necessary to avoid changing the code.
331 while (I1->isDebugValue()) {
332 if (I1==MBB1->begin()) {
333 while (I2->isDebugValue()) {
334 if (I2==MBB2->begin())
335 // I1==DBG at begin; I2==DBG at begin
340 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
345 // I1==first (untested) non-DBG preceding known match
346 while (I2->isDebugValue()) {
347 if (I2==MBB2->begin()) {
349 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
354 // I1, I2==first (untested) non-DBGs preceding known match
355 if (!I1->isIdenticalTo(I2) ||
356 // FIXME: This check is dubious. It's used to get around a problem where
357 // people incorrectly expect inline asm directives to remain in the same
358 // relative order. This is untenable because normal compiler
359 // optimizations (like this one) may reorder and/or merge these
367 // Back past possible debugging pseudos at beginning of block. This matters
368 // when one block differs from the other only by whether debugging pseudos
369 // are present at the beginning. (This way, the various checks later for
370 // I1==MBB1->begin() work as expected.)
371 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
373 while (I2->isDebugValue()) {
374 if (I2 == MBB2->begin())
380 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
382 while (I1->isDebugValue()) {
383 if (I1 == MBB1->begin())
392 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
393 MachineBasicBlock *NewMBB) {
395 RS->enterBasicBlock(CurMBB);
396 if (!CurMBB->empty())
397 RS->forward(std::prev(CurMBB->end()));
398 for (unsigned int i = 1, e = TRI->getNumRegs(); i != e; i++)
399 if (RS->isRegUsed(i, false))
400 NewMBB->addLiveIn(i);
404 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
405 /// after it, replacing it with an unconditional branch to NewDest.
406 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
407 MachineBasicBlock *NewDest) {
408 MachineBasicBlock *CurMBB = OldInst->getParent();
410 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
412 // For targets that use the register scavenger, we must maintain LiveIns.
413 MaintainLiveIns(CurMBB, NewDest);
418 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
419 /// MBB so that the part before the iterator falls into the part starting at the
420 /// iterator. This returns the new MBB.
421 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
422 MachineBasicBlock::iterator BBI1,
423 const BasicBlock *BB) {
424 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
427 MachineFunction &MF = *CurMBB.getParent();
429 // Create the fall-through block.
430 MachineFunction::iterator MBBI = &CurMBB;
431 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
432 CurMBB.getParent()->insert(++MBBI, NewMBB);
434 // Move all the successors of this block to the specified block.
435 NewMBB->transferSuccessors(&CurMBB);
437 // Add an edge from CurMBB to NewMBB for the fall-through.
438 CurMBB.addSuccessor(NewMBB);
440 // Splice the code over.
441 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
443 // NewMBB inherits CurMBB's block frequency.
444 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
446 // For targets that use the register scavenger, we must maintain LiveIns.
447 MaintainLiveIns(&CurMBB, NewMBB);
452 /// EstimateRuntime - Make a rough estimate for how long it will take to run
453 /// the specified code.
454 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
455 MachineBasicBlock::iterator E) {
457 for (; I != E; ++I) {
458 if (I->isDebugValue())
462 else if (I->mayLoad() || I->mayStore())
470 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
471 // branches temporarily for tail merging). In the case where CurMBB ends
472 // with a conditional branch to the next block, optimize by reversing the
473 // test and conditionally branching to SuccMBB instead.
474 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
475 const TargetInstrInfo *TII) {
476 MachineFunction *MF = CurMBB->getParent();
477 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
478 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
479 SmallVector<MachineOperand, 4> Cond;
480 DebugLoc dl; // FIXME: this is nowhere
481 if (I != MF->end() &&
482 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
483 MachineBasicBlock *NextBB = I;
484 if (TBB == NextBB && !Cond.empty() && !FBB) {
485 if (!TII->ReverseBranchCondition(Cond)) {
486 TII->RemoveBranch(*CurMBB);
487 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
492 TII->InsertBranch(*CurMBB, SuccBB, nullptr,
493 SmallVector<MachineOperand, 0>(), dl);
497 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
498 if (getHash() < o.getHash())
500 if (getHash() > o.getHash())
502 if (getBlock()->getNumber() < o.getBlock()->getNumber())
504 if (getBlock()->getNumber() > o.getBlock()->getNumber())
506 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
507 // an object with itself.
508 #ifndef _GLIBCXX_DEBUG
509 llvm_unreachable("Predecessor appears twice");
516 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
517 auto I = MergedBBFreq.find(MBB);
519 if (I != MergedBBFreq.end())
522 return MBFI.getBlockFreq(MBB);
525 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
527 MergedBBFreq[MBB] = F;
530 /// CountTerminators - Count the number of terminators in the given
531 /// block and set I to the position of the first non-terminator, if there
532 /// is one, or MBB->end() otherwise.
533 static unsigned CountTerminators(MachineBasicBlock *MBB,
534 MachineBasicBlock::iterator &I) {
536 unsigned NumTerms = 0;
538 if (I == MBB->begin()) {
543 if (!I->isTerminator()) break;
549 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
550 /// and decide if it would be profitable to merge those tails. Return the
551 /// length of the common tail and iterators to the first common instruction
553 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
554 MachineBasicBlock *MBB2,
555 unsigned minCommonTailLength,
556 unsigned &CommonTailLen,
557 MachineBasicBlock::iterator &I1,
558 MachineBasicBlock::iterator &I2,
559 MachineBasicBlock *SuccBB,
560 MachineBasicBlock *PredBB) {
561 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
562 if (CommonTailLen == 0)
564 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
565 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
568 // It's almost always profitable to merge any number of non-terminator
569 // instructions with the block that falls through into the common successor.
570 if (MBB1 == PredBB || MBB2 == PredBB) {
571 MachineBasicBlock::iterator I;
572 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
573 if (CommonTailLen > NumTerms)
577 // If one of the blocks can be completely merged and happens to be in
578 // a position where the other could fall through into it, merge any number
579 // of instructions, because it can be done without a branch.
580 // TODO: If the blocks are not adjacent, move one of them so that they are?
581 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
583 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
586 // If both blocks have an unconditional branch temporarily stripped out,
587 // count that as an additional common instruction for the following
589 unsigned EffectiveTailLen = CommonTailLen;
590 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
591 !MBB1->back().isBarrier() &&
592 !MBB2->back().isBarrier())
595 // Check if the common tail is long enough to be worthwhile.
596 if (EffectiveTailLen >= minCommonTailLength)
599 // If we are optimizing for code size, 2 instructions in common is enough if
600 // we don't have to split a block. At worst we will be introducing 1 new
601 // branch instruction, which is likely to be smaller than the 2
602 // instructions that would be deleted in the merge.
603 MachineFunction *MF = MBB1->getParent();
604 if (EffectiveTailLen >= 2 &&
605 MF->getFunction()->getAttributes().
606 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
607 (I1 == MBB1->begin() || I2 == MBB2->begin()))
613 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
614 /// hash CurHash (guaranteed to match the last element). Build the vector
615 /// SameTails of all those that have the (same) largest number of instructions
616 /// in common of any pair of these blocks. SameTails entries contain an
617 /// iterator into MergePotentials (from which the MachineBasicBlock can be
618 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
619 /// instruction where the matching code sequence begins.
620 /// Order of elements in SameTails is the reverse of the order in which
621 /// those blocks appear in MergePotentials (where they are not necessarily
623 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
624 unsigned minCommonTailLength,
625 MachineBasicBlock *SuccBB,
626 MachineBasicBlock *PredBB) {
627 unsigned maxCommonTailLength = 0U;
629 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
630 MPIterator HighestMPIter = std::prev(MergePotentials.end());
631 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
632 B = MergePotentials.begin();
633 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
634 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
635 unsigned CommonTailLen;
636 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
638 CommonTailLen, TrialBBI1, TrialBBI2,
640 if (CommonTailLen > maxCommonTailLength) {
642 maxCommonTailLength = CommonTailLen;
643 HighestMPIter = CurMPIter;
644 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
646 if (HighestMPIter == CurMPIter &&
647 CommonTailLen == maxCommonTailLength)
648 SameTails.push_back(SameTailElt(I, TrialBBI2));
654 return maxCommonTailLength;
657 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
658 /// MergePotentials, restoring branches at ends of blocks as appropriate.
659 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
660 MachineBasicBlock *SuccBB,
661 MachineBasicBlock *PredBB) {
662 MPIterator CurMPIter, B;
663 for (CurMPIter = std::prev(MergePotentials.end()),
664 B = MergePotentials.begin();
665 CurMPIter->getHash() == CurHash; --CurMPIter) {
666 // Put the unconditional branch back, if we need one.
667 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
668 if (SuccBB && CurMBB != PredBB)
669 FixTail(CurMBB, SuccBB, TII);
673 if (CurMPIter->getHash() != CurHash)
675 MergePotentials.erase(CurMPIter, MergePotentials.end());
678 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
679 /// only of the common tail. Create a block that does by splitting one.
680 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
681 MachineBasicBlock *SuccBB,
682 unsigned maxCommonTailLength,
683 unsigned &commonTailIndex) {
685 unsigned TimeEstimate = ~0U;
686 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
687 // Use PredBB if possible; that doesn't require a new branch.
688 if (SameTails[i].getBlock() == PredBB) {
692 // Otherwise, make a (fairly bogus) choice based on estimate of
693 // how long it will take the various blocks to execute.
694 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
695 SameTails[i].getTailStartPos());
696 if (t <= TimeEstimate) {
702 MachineBasicBlock::iterator BBI =
703 SameTails[commonTailIndex].getTailStartPos();
704 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
706 // If the common tail includes any debug info we will take it pretty
707 // randomly from one of the inputs. Might be better to remove it?
708 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
709 << maxCommonTailLength);
711 // If the split block unconditionally falls-thru to SuccBB, it will be
712 // merged. In control flow terms it should then take SuccBB's name. e.g. If
713 // SuccBB is an inner loop, the common tail is still part of the inner loop.
714 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
715 SuccBB->getBasicBlock() : MBB->getBasicBlock();
716 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
718 DEBUG(dbgs() << "... failed!");
722 SameTails[commonTailIndex].setBlock(newMBB);
723 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
725 // If we split PredBB, newMBB is the new predecessor.
732 // See if any of the blocks in MergePotentials (which all have a common single
733 // successor, or all have no successor) can be tail-merged. If there is a
734 // successor, any blocks in MergePotentials that are not tail-merged and
735 // are not immediately before Succ must have an unconditional branch to
736 // Succ added (but the predecessor/successor lists need no adjustment).
737 // The lone predecessor of Succ that falls through into Succ,
738 // if any, is given in PredBB.
740 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
741 MachineBasicBlock *PredBB) {
742 bool MadeChange = false;
744 // Except for the special cases below, tail-merge if there are at least
745 // this many instructions in common.
746 unsigned minCommonTailLength = TailMergeSize;
748 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
749 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
750 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
751 << (i == e-1 ? "" : ", ");
754 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
756 dbgs() << " which has fall-through from BB#"
757 << PredBB->getNumber() << "\n";
759 dbgs() << "Looking for common tails of at least "
760 << minCommonTailLength << " instruction"
761 << (minCommonTailLength == 1 ? "" : "s") << '\n';
764 // Sort by hash value so that blocks with identical end sequences sort
766 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
768 // Walk through equivalence sets looking for actual exact matches.
769 while (MergePotentials.size() > 1) {
770 unsigned CurHash = MergePotentials.back().getHash();
772 // Build SameTails, identifying the set of blocks with this hash code
773 // and with the maximum number of instructions in common.
774 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
778 // If we didn't find any pair that has at least minCommonTailLength
779 // instructions in common, remove all blocks with this hash code and retry.
780 if (SameTails.empty()) {
781 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
785 // If one of the blocks is the entire common tail (and not the entry
786 // block, which we can't jump to), we can treat all blocks with this same
787 // tail at once. Use PredBB if that is one of the possibilities, as that
788 // will not introduce any extra branches.
789 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
790 getParent()->begin();
791 unsigned commonTailIndex = SameTails.size();
792 // If there are two blocks, check to see if one can be made to fall through
794 if (SameTails.size() == 2 &&
795 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
796 SameTails[1].tailIsWholeBlock())
798 else if (SameTails.size() == 2 &&
799 SameTails[1].getBlock()->isLayoutSuccessor(
800 SameTails[0].getBlock()) &&
801 SameTails[0].tailIsWholeBlock())
804 // Otherwise just pick one, favoring the fall-through predecessor if
806 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
807 MachineBasicBlock *MBB = SameTails[i].getBlock();
808 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
814 if (SameTails[i].tailIsWholeBlock())
819 if (commonTailIndex == SameTails.size() ||
820 (SameTails[commonTailIndex].getBlock() == PredBB &&
821 !SameTails[commonTailIndex].tailIsWholeBlock())) {
822 // None of the blocks consist entirely of the common tail.
823 // Split a block so that one does.
824 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
825 maxCommonTailLength, commonTailIndex)) {
826 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
831 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
833 // Recompute commont tail MBB's edge weights and block frequency.
834 setCommonTailEdgeWeights(*MBB);
836 // MBB is common tail. Adjust all other BB's to jump to this one.
837 // Traversal must be forwards so erases work.
838 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
840 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
841 if (commonTailIndex == i)
843 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
844 << (i == e-1 ? "" : ", "));
845 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
846 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
847 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
848 MergePotentials.erase(SameTails[i].getMPIter());
850 DEBUG(dbgs() << "\n");
851 // We leave commonTailIndex in the worklist in case there are other blocks
852 // that match it with a smaller number of instructions.
858 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
859 bool MadeChange = false;
860 if (!EnableTailMerge) return MadeChange;
862 // First find blocks with no successors.
863 MergePotentials.clear();
864 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
865 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
866 if (TriedMerging.count(I))
869 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
872 // If this is a large problem, avoid visiting the same basic blocks
874 if (MergePotentials.size() == TailMergeThreshold)
875 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
876 TriedMerging.insert(MergePotentials[i].getBlock());
878 // See if we can do any tail merging on those.
879 if (MergePotentials.size() >= 2)
880 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
882 // Look at blocks (IBB) with multiple predecessors (PBB).
883 // We change each predecessor to a canonical form, by
884 // (1) temporarily removing any unconditional branch from the predecessor
886 // (2) alter conditional branches so they branch to the other block
887 // not IBB; this may require adding back an unconditional branch to IBB
888 // later, where there wasn't one coming in. E.g.
890 // fallthrough to QBB
893 // with a conceptual B to IBB after that, which never actually exists.
894 // With those changes, we see whether the predecessors' tails match,
895 // and merge them if so. We change things out of canonical form and
896 // back to the way they were later in the process. (OptimizeBranches
897 // would undo some of this, but we can't use it, because we'd get into
898 // a compile-time infinite loop repeatedly doing and undoing the same
901 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
903 if (I->pred_size() < 2) continue;
904 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
905 MachineBasicBlock *IBB = I;
906 MachineBasicBlock *PredBB = std::prev(I);
907 MergePotentials.clear();
908 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
910 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
911 MachineBasicBlock *PBB = *P;
912 if (TriedMerging.count(PBB))
915 // Skip blocks that loop to themselves, can't tail merge these.
919 // Visit each predecessor only once.
920 if (!UniquePreds.insert(PBB))
923 // Skip blocks which may jump to a landing pad. Can't tail merge these.
924 if (PBB->getLandingPadSuccessor())
927 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
928 SmallVector<MachineOperand, 4> Cond;
929 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
930 // Failing case: IBB is the target of a cbr, and we cannot reverse the
932 SmallVector<MachineOperand, 4> NewCond(Cond);
933 if (!Cond.empty() && TBB == IBB) {
934 if (TII->ReverseBranchCondition(NewCond))
936 // This is the QBB case described above
938 FBB = std::next(MachineFunction::iterator(PBB));
941 // Failing case: the only way IBB can be reached from PBB is via
942 // exception handling. Happens for landing pads. Would be nice to have
943 // a bit in the edge so we didn't have to do all this.
944 if (IBB->isLandingPad()) {
945 MachineFunction::iterator IP = PBB; IP++;
946 MachineBasicBlock *PredNextBB = nullptr;
950 if (IBB != PredNextBB) // fallthrough
953 if (TBB != IBB && FBB != IBB) // cbr then ubr
955 } else if (Cond.empty()) {
956 if (TBB != IBB) // ubr
959 if (TBB != IBB && IBB != PredNextBB) // cbr
964 // Remove the unconditional branch at the end, if any.
965 if (TBB && (Cond.empty() || FBB)) {
966 DebugLoc dl; // FIXME: this is nowhere
967 TII->RemoveBranch(*PBB);
969 // reinsert conditional branch only, for now
970 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
974 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
978 // If this is a large problem, avoid visiting the same basic blocks multiple
980 if (MergePotentials.size() == TailMergeThreshold)
981 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
982 TriedMerging.insert(MergePotentials[i].getBlock());
984 if (MergePotentials.size() >= 2)
985 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
987 // Reinsert an unconditional branch if needed. The 1 below can occur as a
988 // result of removing blocks in TryTailMergeBlocks.
989 PredBB = std::prev(I); // this may have been changed in TryTailMergeBlocks
990 if (MergePotentials.size() == 1 &&
991 MergePotentials.begin()->getBlock() != PredBB)
992 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
998 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
999 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1000 BlockFrequency AccumulatedMBBFreq;
1002 // Aggregate edge frequency of successor edge j:
1003 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1004 // where bb is a basic block that is in SameTails.
1005 for (const auto &Src : SameTails) {
1006 const MachineBasicBlock *SrcMBB = Src.getBlock();
1007 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1008 AccumulatedMBBFreq += BlockFreq;
1010 // It is not necessary to recompute edge weights if TailBB has less than two
1012 if (TailMBB.succ_size() <= 1)
1015 auto EdgeFreq = EdgeFreqLs.begin();
1017 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1018 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1019 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1022 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1024 if (TailMBB.succ_size() <= 1)
1027 auto MaxEdgeFreq = *std::max_element(EdgeFreqLs.begin(), EdgeFreqLs.end());
1028 uint64_t Scale = MaxEdgeFreq.getFrequency() / UINT32_MAX + 1;
1029 auto EdgeFreq = EdgeFreqLs.begin();
1031 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1032 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1033 TailMBB.setSuccWeight(SuccI, EdgeFreq->getFrequency() / Scale);
1036 //===----------------------------------------------------------------------===//
1037 // Branch Optimization
1038 //===----------------------------------------------------------------------===//
1040 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1041 bool MadeChange = false;
1043 // Make sure blocks are numbered in order
1044 MF.RenumberBlocks();
1046 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1048 MachineBasicBlock *MBB = I++;
1049 MadeChange |= OptimizeBlock(MBB);
1051 // If it is dead, remove it.
1052 if (MBB->pred_empty()) {
1053 RemoveDeadBlock(MBB);
1061 // Blocks should be considered empty if they contain only debug info;
1062 // else the debug info would affect codegen.
1063 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1066 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
1067 MBBI!=MBBE; ++MBBI) {
1068 if (!MBBI->isDebugValue())
1074 // Blocks with only debug info and branches should be considered the same
1075 // as blocks with only branches.
1076 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1077 MachineBasicBlock::iterator MBBI, MBBE;
1078 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
1079 if (!MBBI->isDebugValue())
1082 return (MBBI->isBranch());
1085 /// IsBetterFallthrough - Return true if it would be clearly better to
1086 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1087 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1088 /// result in infinite loops.
1089 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1090 MachineBasicBlock *MBB2) {
1091 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1092 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1093 // optimize branches that branch to either a return block or an assert block
1094 // into a fallthrough to the return.
1095 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
1097 // If there is a clear successor ordering we make sure that one block
1098 // will fall through to the next
1099 if (MBB1->isSuccessor(MBB2)) return true;
1100 if (MBB2->isSuccessor(MBB1)) return false;
1102 // Neither block consists entirely of debug info (per IsEmptyBlock check),
1103 // so we needn't test for falling off the beginning here.
1104 MachineBasicBlock::iterator MBB1I = --MBB1->end();
1105 while (MBB1I->isDebugValue())
1107 MachineBasicBlock::iterator MBB2I = --MBB2->end();
1108 while (MBB2I->isDebugValue())
1110 return MBB2I->isCall() && !MBB1I->isCall();
1113 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1114 /// instructions on the block. Always use the DebugLoc of the first
1115 /// branching instruction found unless its absent, in which case use the
1116 /// DebugLoc of the second if present.
1117 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1118 MachineBasicBlock::iterator I = MBB.end();
1119 if (I == MBB.begin())
1122 while (I->isDebugValue() && I != MBB.begin())
1125 return I->getDebugLoc();
1129 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1130 /// block. This is never called on the entry block.
1131 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1132 bool MadeChange = false;
1133 MachineFunction &MF = *MBB->getParent();
1136 MachineFunction::iterator FallThrough = MBB;
1139 // If this block is empty, make everyone use its fall-through, not the block
1140 // explicitly. Landing pads should not do this since the landing-pad table
1141 // points to this block. Blocks with their addresses taken shouldn't be
1143 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1144 // Dead block? Leave for cleanup later.
1145 if (MBB->pred_empty()) return MadeChange;
1147 if (FallThrough == MF.end()) {
1148 // TODO: Simplify preds to not branch here if possible!
1150 // Rewrite all predecessors of the old block to go to the fallthrough
1152 while (!MBB->pred_empty()) {
1153 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1154 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1156 // If MBB was the target of a jump table, update jump tables to go to the
1157 // fallthrough instead.
1158 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1159 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1165 // Check to see if we can simplify the terminator of the block before this
1167 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1169 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1170 SmallVector<MachineOperand, 4> PriorCond;
1171 bool PriorUnAnalyzable =
1172 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1173 if (!PriorUnAnalyzable) {
1174 // If the CFG for the prior block has extra edges, remove them.
1175 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1176 !PriorCond.empty());
1178 // If the previous branch is conditional and both conditions go to the same
1179 // destination, remove the branch, replacing it with an unconditional one or
1181 if (PriorTBB && PriorTBB == PriorFBB) {
1182 DebugLoc dl = getBranchDebugLoc(PrevBB);
1183 TII->RemoveBranch(PrevBB);
1185 if (PriorTBB != MBB)
1186 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1189 goto ReoptimizeBlock;
1192 // If the previous block unconditionally falls through to this block and
1193 // this block has no other predecessors, move the contents of this block
1194 // into the prior block. This doesn't usually happen when SimplifyCFG
1195 // has been used, but it can happen if tail merging splits a fall-through
1196 // predecessor of a block.
1197 // This has to check PrevBB->succ_size() because EH edges are ignored by
1199 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1200 PrevBB.succ_size() == 1 &&
1201 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1202 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1203 << "From MBB: " << *MBB);
1204 // Remove redundant DBG_VALUEs first.
1205 if (PrevBB.begin() != PrevBB.end()) {
1206 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1208 MachineBasicBlock::iterator MBBIter = MBB->begin();
1209 // Check if DBG_VALUE at the end of PrevBB is identical to the
1210 // DBG_VALUE at the beginning of MBB.
1211 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1212 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1213 if (!MBBIter->isIdenticalTo(PrevBBIter))
1215 MachineInstr *DuplicateDbg = MBBIter;
1216 ++MBBIter; -- PrevBBIter;
1217 DuplicateDbg->eraseFromParent();
1220 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1221 PrevBB.removeSuccessor(PrevBB.succ_begin());
1222 assert(PrevBB.succ_empty());
1223 PrevBB.transferSuccessors(MBB);
1228 // If the previous branch *only* branches to *this* block (conditional or
1229 // not) remove the branch.
1230 if (PriorTBB == MBB && !PriorFBB) {
1231 TII->RemoveBranch(PrevBB);
1234 goto ReoptimizeBlock;
1237 // If the prior block branches somewhere else on the condition and here if
1238 // the condition is false, remove the uncond second branch.
1239 if (PriorFBB == MBB) {
1240 DebugLoc dl = getBranchDebugLoc(PrevBB);
1241 TII->RemoveBranch(PrevBB);
1242 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1245 goto ReoptimizeBlock;
1248 // If the prior block branches here on true and somewhere else on false, and
1249 // if the branch condition is reversible, reverse the branch to create a
1251 if (PriorTBB == MBB) {
1252 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1253 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1254 DebugLoc dl = getBranchDebugLoc(PrevBB);
1255 TII->RemoveBranch(PrevBB);
1256 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1259 goto ReoptimizeBlock;
1263 // If this block has no successors (e.g. it is a return block or ends with
1264 // a call to a no-return function like abort or __cxa_throw) and if the pred
1265 // falls through into this block, and if it would otherwise fall through
1266 // into the block after this, move this block to the end of the function.
1268 // We consider it more likely that execution will stay in the function (e.g.
1269 // due to loops) than it is to exit it. This asserts in loops etc, moving
1270 // the assert condition out of the loop body.
1271 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1272 MachineFunction::iterator(PriorTBB) == FallThrough &&
1273 !MBB->canFallThrough()) {
1274 bool DoTransform = true;
1276 // We have to be careful that the succs of PredBB aren't both no-successor
1277 // blocks. If neither have successors and if PredBB is the second from
1278 // last block in the function, we'd just keep swapping the two blocks for
1279 // last. Only do the swap if one is clearly better to fall through than
1281 if (FallThrough == --MF.end() &&
1282 !IsBetterFallthrough(PriorTBB, MBB))
1283 DoTransform = false;
1286 // Reverse the branch so we will fall through on the previous true cond.
1287 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1288 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1289 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1290 << "To make fallthrough to: " << *PriorTBB << "\n");
1292 DebugLoc dl = getBranchDebugLoc(PrevBB);
1293 TII->RemoveBranch(PrevBB);
1294 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1296 // Move this block to the end of the function.
1297 MBB->moveAfter(--MF.end());
1306 // Analyze the branch in the current block.
1307 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1308 SmallVector<MachineOperand, 4> CurCond;
1309 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1310 if (!CurUnAnalyzable) {
1311 // If the CFG for the prior block has extra edges, remove them.
1312 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1314 // If this is a two-way branch, and the FBB branches to this block, reverse
1315 // the condition so the single-basic-block loop is faster. Instead of:
1316 // Loop: xxx; jcc Out; jmp Loop
1318 // Loop: xxx; jncc Loop; jmp Out
1319 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1320 SmallVector<MachineOperand, 4> NewCond(CurCond);
1321 if (!TII->ReverseBranchCondition(NewCond)) {
1322 DebugLoc dl = getBranchDebugLoc(*MBB);
1323 TII->RemoveBranch(*MBB);
1324 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1327 goto ReoptimizeBlock;
1331 // If this branch is the only thing in its block, see if we can forward
1332 // other blocks across it.
1333 if (CurTBB && CurCond.empty() && !CurFBB &&
1334 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1335 !MBB->hasAddressTaken()) {
1336 DebugLoc dl = getBranchDebugLoc(*MBB);
1337 // This block may contain just an unconditional branch. Because there can
1338 // be 'non-branch terminators' in the block, try removing the branch and
1339 // then seeing if the block is empty.
1340 TII->RemoveBranch(*MBB);
1341 // If the only things remaining in the block are debug info, remove these
1342 // as well, so this will behave the same as an empty block in non-debug
1344 if (!MBB->empty()) {
1345 bool NonDebugInfoFound = false;
1346 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1348 if (!I->isDebugValue()) {
1349 NonDebugInfoFound = true;
1353 if (!NonDebugInfoFound)
1354 // Make the block empty, losing the debug info (we could probably
1355 // improve this in some cases.)
1356 MBB->erase(MBB->begin(), MBB->end());
1358 // If this block is just an unconditional branch to CurTBB, we can
1359 // usually completely eliminate the block. The only case we cannot
1360 // completely eliminate the block is when the block before this one
1361 // falls through into MBB and we can't understand the prior block's branch
1364 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1365 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1366 !PrevBB.isSuccessor(MBB)) {
1367 // If the prior block falls through into us, turn it into an
1368 // explicit branch to us to make updates simpler.
1369 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1370 PriorTBB != MBB && PriorFBB != MBB) {
1372 assert(PriorCond.empty() && !PriorFBB &&
1373 "Bad branch analysis");
1376 assert(!PriorFBB && "Machine CFG out of date!");
1379 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1380 TII->RemoveBranch(PrevBB);
1381 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1384 // Iterate through all the predecessors, revectoring each in-turn.
1386 bool DidChange = false;
1387 bool HasBranchToSelf = false;
1388 while(PI != MBB->pred_size()) {
1389 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1391 // If this block has an uncond branch to itself, leave it.
1393 HasBranchToSelf = true;
1396 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1397 // If this change resulted in PMBB ending in a conditional
1398 // branch where both conditions go to the same destination,
1399 // change this to an unconditional branch (and fix the CFG).
1400 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1401 SmallVector<MachineOperand, 4> NewCurCond;
1402 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1403 NewCurFBB, NewCurCond, true);
1404 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1405 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1406 TII->RemoveBranch(*PMBB);
1408 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1411 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1416 // Change any jumptables to go to the new MBB.
1417 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1418 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1422 if (!HasBranchToSelf) return MadeChange;
1427 // Add the branch back if the block is more than just an uncond branch.
1428 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1432 // If the prior block doesn't fall through into this block, and if this
1433 // block doesn't fall through into some other block, see if we can find a
1434 // place to move this block where a fall-through will happen.
1435 if (!PrevBB.canFallThrough()) {
1437 // Now we know that there was no fall-through into this block, check to
1438 // see if it has a fall-through into its successor.
1439 bool CurFallsThru = MBB->canFallThrough();
1441 if (!MBB->isLandingPad()) {
1442 // Check all the predecessors of this block. If one of them has no fall
1443 // throughs, move this block right after it.
1444 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1445 E = MBB->pred_end(); PI != E; ++PI) {
1446 // Analyze the branch at the end of the pred.
1447 MachineBasicBlock *PredBB = *PI;
1448 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1449 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1450 SmallVector<MachineOperand, 4> PredCond;
1451 if (PredBB != MBB && !PredBB->canFallThrough() &&
1452 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1453 && (!CurFallsThru || !CurTBB || !CurFBB)
1454 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1455 // If the current block doesn't fall through, just move it.
1456 // If the current block can fall through and does not end with a
1457 // conditional branch, we need to append an unconditional jump to
1458 // the (current) next block. To avoid a possible compile-time
1459 // infinite loop, move blocks only backward in this case.
1460 // Also, if there are already 2 branches here, we cannot add a third;
1461 // this means we have the case
1466 MachineBasicBlock *NextBB =
1467 std::next(MachineFunction::iterator(MBB));
1469 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1471 MBB->moveAfter(PredBB);
1473 goto ReoptimizeBlock;
1478 if (!CurFallsThru) {
1479 // Check all successors to see if we can move this block before it.
1480 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1481 E = MBB->succ_end(); SI != E; ++SI) {
1482 // Analyze the branch at the end of the block before the succ.
1483 MachineBasicBlock *SuccBB = *SI;
1484 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1486 // If this block doesn't already fall-through to that successor, and if
1487 // the succ doesn't already have a block that can fall through into it,
1488 // and if the successor isn't an EH destination, we can arrange for the
1489 // fallthrough to happen.
1490 if (SuccBB != MBB && &*SuccPrev != MBB &&
1491 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1492 !SuccBB->isLandingPad()) {
1493 MBB->moveBefore(SuccBB);
1495 goto ReoptimizeBlock;
1499 // Okay, there is no really great place to put this block. If, however,
1500 // the block before this one would be a fall-through if this block were
1501 // removed, move this block to the end of the function.
1502 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1503 SmallVector<MachineOperand, 4> PrevCond;
1504 if (FallThrough != MF.end() &&
1505 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1506 PrevBB.isSuccessor(FallThrough)) {
1507 MBB->moveAfter(--MF.end());
1517 //===----------------------------------------------------------------------===//
1518 // Hoist Common Code
1519 //===----------------------------------------------------------------------===//
1521 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1522 /// blocks to their common predecessor.
1523 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1524 bool MadeChange = false;
1525 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1526 MachineBasicBlock *MBB = I++;
1527 MadeChange |= HoistCommonCodeInSuccs(MBB);
1533 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1534 /// its 'true' successor.
1535 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1536 MachineBasicBlock *TrueBB) {
1537 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1538 E = BB->succ_end(); SI != E; ++SI) {
1539 MachineBasicBlock *SuccBB = *SI;
1540 if (SuccBB != TrueBB)
1546 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1547 /// in successors to. The location is usually just before the terminator,
1548 /// however if the terminator is a conditional branch and its previous
1549 /// instruction is the flag setting instruction, the previous instruction is
1550 /// the preferred location. This function also gathers uses and defs of the
1551 /// instructions from the insertion point to the end of the block. The data is
1552 /// used by HoistCommonCodeInSuccs to ensure safety.
1554 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1555 const TargetInstrInfo *TII,
1556 const TargetRegisterInfo *TRI,
1557 SmallSet<unsigned,4> &Uses,
1558 SmallSet<unsigned,4> &Defs) {
1559 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1560 if (!TII->isUnpredicatedTerminator(Loc))
1563 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1564 const MachineOperand &MO = Loc->getOperand(i);
1567 unsigned Reg = MO.getReg();
1571 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1575 // Don't try to hoist code in the rare case the terminator defines a
1576 // register that is later used.
1579 // If the terminator defines a register, make sure we don't hoist
1580 // the instruction whose def might be clobbered by the terminator.
1581 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1588 if (Loc == MBB->begin())
1591 // The terminator is probably a conditional branch, try not to separate the
1592 // branch from condition setting instruction.
1593 MachineBasicBlock::iterator PI = Loc;
1595 while (PI != MBB->begin() && PI->isDebugValue())
1599 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1600 const MachineOperand &MO = PI->getOperand(i);
1601 // If PI has a regmask operand, it is probably a call. Separate away.
1604 if (!MO.isReg() || MO.isUse())
1606 unsigned Reg = MO.getReg();
1609 if (Uses.count(Reg))
1613 // The condition setting instruction is not just before the conditional
1617 // Be conservative, don't insert instruction above something that may have
1618 // side-effects. And since it's potentially bad to separate flag setting
1619 // instruction from the conditional branch, just abort the optimization
1621 // Also avoid moving code above predicated instruction since it's hard to
1622 // reason about register liveness with predicated instruction.
1623 bool DontMoveAcrossStore = true;
1624 if (!PI->isSafeToMove(TII, nullptr, DontMoveAcrossStore) ||
1625 TII->isPredicated(PI))
1629 // Find out what registers are live. Note this routine is ignoring other live
1630 // registers which are only used by instructions in successor blocks.
1631 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1632 const MachineOperand &MO = PI->getOperand(i);
1635 unsigned Reg = MO.getReg();
1639 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1642 if (Uses.erase(Reg)) {
1643 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1644 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1646 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1654 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1655 /// sequence at the start of the function, move the instructions before MBB
1656 /// terminator if it's legal.
1657 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1658 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1659 SmallVector<MachineOperand, 4> Cond;
1660 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1663 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1665 // Malformed bcc? True and false blocks are the same?
1668 // Restrict the optimization to cases where MBB is the only predecessor,
1669 // it is an obvious win.
1670 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1673 // Find a suitable position to hoist the common instructions to. Also figure
1674 // out which registers are used or defined by instructions from the insertion
1675 // point to the end of the block.
1676 SmallSet<unsigned, 4> Uses, Defs;
1677 MachineBasicBlock::iterator Loc =
1678 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1679 if (Loc == MBB->end())
1682 bool HasDups = false;
1683 SmallVector<unsigned, 4> LocalDefs;
1684 SmallSet<unsigned, 4> LocalDefsSet;
1685 MachineBasicBlock::iterator TIB = TBB->begin();
1686 MachineBasicBlock::iterator FIB = FBB->begin();
1687 MachineBasicBlock::iterator TIE = TBB->end();
1688 MachineBasicBlock::iterator FIE = FBB->end();
1689 while (TIB != TIE && FIB != FIE) {
1690 // Skip dbg_value instructions. These do not count.
1691 if (TIB->isDebugValue()) {
1692 while (TIB != TIE && TIB->isDebugValue())
1697 if (FIB->isDebugValue()) {
1698 while (FIB != FIE && FIB->isDebugValue())
1703 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1706 if (TII->isPredicated(TIB))
1707 // Hard to reason about register liveness with predicated instruction.
1711 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1712 MachineOperand &MO = TIB->getOperand(i);
1713 // Don't attempt to hoist instructions with register masks.
1714 if (MO.isRegMask()) {
1720 unsigned Reg = MO.getReg();
1724 if (Uses.count(Reg)) {
1725 // Avoid clobbering a register that's used by the instruction at
1726 // the point of insertion.
1731 if (Defs.count(Reg) && !MO.isDead()) {
1732 // Don't hoist the instruction if the def would be clobber by the
1733 // instruction at the point insertion. FIXME: This is overly
1734 // conservative. It should be possible to hoist the instructions
1735 // in BB2 in the following example:
1737 // r1, eflag = op1 r2, r3
1746 } else if (!LocalDefsSet.count(Reg)) {
1747 if (Defs.count(Reg)) {
1748 // Use is defined by the instruction at the point of insertion.
1753 if (MO.isKill() && Uses.count(Reg))
1754 // Kills a register that's read by the instruction at the point of
1755 // insertion. Remove the kill marker.
1756 MO.setIsKill(false);
1762 bool DontMoveAcrossStore = true;
1763 if (!TIB->isSafeToMove(TII, nullptr, DontMoveAcrossStore))
1766 // Remove kills from LocalDefsSet, these registers had short live ranges.
1767 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1768 MachineOperand &MO = TIB->getOperand(i);
1769 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1771 unsigned Reg = MO.getReg();
1772 if (!Reg || !LocalDefsSet.count(Reg))
1774 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1775 LocalDefsSet.erase(*AI);
1778 // Track local defs so we can update liveins.
1779 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1780 MachineOperand &MO = TIB->getOperand(i);
1781 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1783 unsigned Reg = MO.getReg();
1786 LocalDefs.push_back(Reg);
1787 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1788 LocalDefsSet.insert(*AI);
1799 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1800 FBB->erase(FBB->begin(), FIB);
1803 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1804 unsigned Def = LocalDefs[i];
1805 if (LocalDefsSet.count(Def)) {
1806 TBB->addLiveIn(Def);
1807 FBB->addLiveIn(Def);