1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass forwards branches to unconditional branches to make them branch
11 // directly to the target block. This pass often results in dead MBB's, which
14 // Note that this pass must be run after register allocation, it cannot handle
15 // SSA form. It also must handle virtual registers for targets that emit virtual
18 //===----------------------------------------------------------------------===//
20 #include "BranchFolding.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
25 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
26 #include "llvm/CodeGen/MachineFunctionPass.h"
27 #include "llvm/CodeGen/MachineJumpTableInfo.h"
28 #include "llvm/CodeGen/MachineMemOperand.h"
29 #include "llvm/CodeGen/MachineModuleInfo.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/Passes.h"
32 #include "llvm/CodeGen/RegisterScavenging.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/Target/TargetInstrInfo.h"
39 #include "llvm/Target/TargetRegisterInfo.h"
40 #include "llvm/Target/TargetSubtargetInfo.h"
44 #define DEBUG_TYPE "branchfolding"
46 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
47 STATISTIC(NumBranchOpts, "Number of branches optimized");
48 STATISTIC(NumTailMerge , "Number of block tails merged");
49 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
51 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
52 cl::init(cl::BOU_UNSET), cl::Hidden);
54 // Throttle for huge numbers of predecessors (compile speed problems)
55 static cl::opt<unsigned>
56 TailMergeThreshold("tail-merge-threshold",
57 cl::desc("Max number of predecessors to consider tail merging"),
58 cl::init(150), cl::Hidden);
60 // Heuristic for tail merging (and, inversely, tail duplication).
61 // TODO: This should be replaced with a target query.
62 static cl::opt<unsigned>
63 TailMergeSize("tail-merge-size",
64 cl::desc("Min number of instructions to consider tail merging"),
65 cl::init(3), cl::Hidden);
68 /// BranchFolderPass - Wrap branch folder in a machine function pass.
69 class BranchFolderPass : public MachineFunctionPass {
72 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
74 bool runOnMachineFunction(MachineFunction &MF) override;
76 void getAnalysisUsage(AnalysisUsage &AU) const override {
77 AU.addRequired<MachineBlockFrequencyInfo>();
78 AU.addRequired<MachineBranchProbabilityInfo>();
79 AU.addRequired<TargetPassConfig>();
80 MachineFunctionPass::getAnalysisUsage(AU);
85 char BranchFolderPass::ID = 0;
86 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
88 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
89 "Control Flow Optimizer", false, false)
91 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
92 if (skipOptnoneFunction(*MF.getFunction()))
95 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
96 // TailMerge can create jump into if branches that make CFG irreducible for
97 // HW that requires structurized CFG.
98 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
99 PassConfig->getEnableTailMerge();
100 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true,
101 getAnalysis<MachineBlockFrequencyInfo>(),
102 getAnalysis<MachineBranchProbabilityInfo>());
103 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
104 MF.getSubtarget().getRegisterInfo(),
105 getAnalysisIfAvailable<MachineModuleInfo>());
108 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
109 const MachineBlockFrequencyInfo &FreqInfo,
110 const MachineBranchProbabilityInfo &ProbInfo)
111 : EnableHoistCommonCode(CommonHoist), MBBFreqInfo(FreqInfo),
113 switch (FlagEnableTailMerge) {
114 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
115 case cl::BOU_TRUE: EnableTailMerge = true; break;
116 case cl::BOU_FALSE: EnableTailMerge = false; break;
120 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
121 /// function, updating the CFG.
122 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
123 assert(MBB->pred_empty() && "MBB must be dead!");
124 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
126 MachineFunction *MF = MBB->getParent();
127 // drop all successors.
128 while (!MBB->succ_empty())
129 MBB->removeSuccessor(MBB->succ_end()-1);
131 // Avoid matching if this pointer gets reused.
132 TriedMerging.erase(MBB);
138 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
139 /// followed by terminators, and if the implicitly defined registers are not
140 /// used by the terminators, remove those implicit_def's. e.g.
142 /// r0 = implicit_def
143 /// r1 = implicit_def
145 /// This block can be optimized away later if the implicit instructions are
147 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
148 SmallSet<unsigned, 4> ImpDefRegs;
149 MachineBasicBlock::iterator I = MBB->begin();
150 while (I != MBB->end()) {
151 if (!I->isImplicitDef())
153 unsigned Reg = I->getOperand(0).getReg();
154 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
155 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
156 SubRegs.isValid(); ++SubRegs)
157 ImpDefRegs.insert(*SubRegs);
159 ImpDefRegs.insert(Reg);
163 if (ImpDefRegs.empty())
166 MachineBasicBlock::iterator FirstTerm = I;
167 while (I != MBB->end()) {
168 if (!TII->isUnpredicatedTerminator(I))
170 // See if it uses any of the implicitly defined registers.
171 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
172 MachineOperand &MO = I->getOperand(i);
173 if (!MO.isReg() || !MO.isUse())
175 unsigned Reg = MO.getReg();
176 if (ImpDefRegs.count(Reg))
183 while (I != FirstTerm) {
184 MachineInstr *ImpDefMI = &*I;
186 MBB->erase(ImpDefMI);
192 /// OptimizeFunction - Perhaps branch folding, tail merging and other
193 /// CFG optimizations on the given function.
194 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
195 const TargetInstrInfo *tii,
196 const TargetRegisterInfo *tri,
197 MachineModuleInfo *mmi) {
198 if (!tii) return false;
200 TriedMerging.clear();
207 // Use a RegScavenger to help update liveness when required.
208 MachineRegisterInfo &MRI = MF.getRegInfo();
209 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
210 RS = new RegScavenger();
212 MRI.invalidateLiveness();
214 // Fix CFG. The later algorithms expect it to be right.
215 bool MadeChange = false;
216 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
217 MachineBasicBlock *MBB = I, *TBB = nullptr, *FBB = nullptr;
218 SmallVector<MachineOperand, 4> Cond;
219 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
220 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
221 MadeChange |= OptimizeImpDefsBlock(MBB);
224 bool MadeChangeThisIteration = true;
225 while (MadeChangeThisIteration) {
226 MadeChangeThisIteration = TailMergeBlocks(MF);
227 MadeChangeThisIteration |= OptimizeBranches(MF);
228 if (EnableHoistCommonCode)
229 MadeChangeThisIteration |= HoistCommonCode(MF);
230 MadeChange |= MadeChangeThisIteration;
233 // See if any jump tables have become dead as the code generator
235 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
241 // Walk the function to find jump tables that are live.
242 BitVector JTIsLive(JTI->getJumpTables().size());
243 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
245 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
247 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
248 MachineOperand &Op = I->getOperand(op);
249 if (!Op.isJTI()) continue;
251 // Remember that this JT is live.
252 JTIsLive.set(Op.getIndex());
256 // Finally, remove dead jump tables. This happens when the
257 // indirect jump was unreachable (and thus deleted).
258 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
259 if (!JTIsLive.test(i)) {
260 JTI->RemoveJumpTable(i);
268 //===----------------------------------------------------------------------===//
269 // Tail Merging of Blocks
270 //===----------------------------------------------------------------------===//
272 /// HashMachineInstr - Compute a hash value for MI and its operands.
273 static unsigned HashMachineInstr(const MachineInstr *MI) {
274 unsigned Hash = MI->getOpcode();
275 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
276 const MachineOperand &Op = MI->getOperand(i);
278 // Merge in bits from the operand if easy. We can't use MachineOperand's
279 // hash_code here because it's not deterministic and we sort by hash value
281 unsigned OperandHash = 0;
282 switch (Op.getType()) {
283 case MachineOperand::MO_Register:
284 OperandHash = Op.getReg();
286 case MachineOperand::MO_Immediate:
287 OperandHash = Op.getImm();
289 case MachineOperand::MO_MachineBasicBlock:
290 OperandHash = Op.getMBB()->getNumber();
292 case MachineOperand::MO_FrameIndex:
293 case MachineOperand::MO_ConstantPoolIndex:
294 case MachineOperand::MO_JumpTableIndex:
295 OperandHash = Op.getIndex();
297 case MachineOperand::MO_GlobalAddress:
298 case MachineOperand::MO_ExternalSymbol:
299 // Global address / external symbol are too hard, don't bother, but do
300 // pull in the offset.
301 OperandHash = Op.getOffset();
307 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
312 /// HashEndOfMBB - Hash the last instruction in the MBB.
313 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
314 MachineBasicBlock::const_iterator I = MBB->getLastNonDebugInstr();
318 return HashMachineInstr(I);
321 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
322 /// of instructions they actually have in common together at their end. Return
323 /// iterators for the first shared instruction in each block.
324 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
325 MachineBasicBlock *MBB2,
326 MachineBasicBlock::iterator &I1,
327 MachineBasicBlock::iterator &I2) {
331 unsigned TailLen = 0;
332 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
334 // Skip debugging pseudos; necessary to avoid changing the code.
335 while (I1->isDebugValue()) {
336 if (I1==MBB1->begin()) {
337 while (I2->isDebugValue()) {
338 if (I2==MBB2->begin())
339 // I1==DBG at begin; I2==DBG at begin
344 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
349 // I1==first (untested) non-DBG preceding known match
350 while (I2->isDebugValue()) {
351 if (I2==MBB2->begin()) {
353 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
358 // I1, I2==first (untested) non-DBGs preceding known match
359 if (!I1->isIdenticalTo(I2) ||
360 // FIXME: This check is dubious. It's used to get around a problem where
361 // people incorrectly expect inline asm directives to remain in the same
362 // relative order. This is untenable because normal compiler
363 // optimizations (like this one) may reorder and/or merge these
371 // Back past possible debugging pseudos at beginning of block. This matters
372 // when one block differs from the other only by whether debugging pseudos
373 // are present at the beginning. (This way, the various checks later for
374 // I1==MBB1->begin() work as expected.)
375 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
377 while (I2->isDebugValue()) {
378 if (I2 == MBB2->begin())
384 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
386 while (I1->isDebugValue()) {
387 if (I1 == MBB1->begin())
396 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
397 MachineBasicBlock *NewMBB) {
399 RS->enterBasicBlock(CurMBB);
400 if (!CurMBB->empty())
401 RS->forward(std::prev(CurMBB->end()));
402 for (unsigned int i = 1, e = TRI->getNumRegs(); i != e; i++)
403 if (RS->isRegUsed(i, false))
404 NewMBB->addLiveIn(i);
408 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
409 /// after it, replacing it with an unconditional branch to NewDest.
410 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
411 MachineBasicBlock *NewDest) {
412 MachineBasicBlock *CurMBB = OldInst->getParent();
414 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
416 // For targets that use the register scavenger, we must maintain LiveIns.
417 MaintainLiveIns(CurMBB, NewDest);
422 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
423 /// MBB so that the part before the iterator falls into the part starting at the
424 /// iterator. This returns the new MBB.
425 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
426 MachineBasicBlock::iterator BBI1,
427 const BasicBlock *BB) {
428 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
431 MachineFunction &MF = *CurMBB.getParent();
433 // Create the fall-through block.
434 MachineFunction::iterator MBBI = &CurMBB;
435 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
436 CurMBB.getParent()->insert(++MBBI, NewMBB);
438 // Move all the successors of this block to the specified block.
439 NewMBB->transferSuccessors(&CurMBB);
441 // Add an edge from CurMBB to NewMBB for the fall-through.
442 CurMBB.addSuccessor(NewMBB);
444 // Splice the code over.
445 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
447 // NewMBB inherits CurMBB's block frequency.
448 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
450 // For targets that use the register scavenger, we must maintain LiveIns.
451 MaintainLiveIns(&CurMBB, NewMBB);
456 /// EstimateRuntime - Make a rough estimate for how long it will take to run
457 /// the specified code.
458 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
459 MachineBasicBlock::iterator E) {
461 for (; I != E; ++I) {
462 if (I->isDebugValue())
466 else if (I->mayLoad() || I->mayStore())
474 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
475 // branches temporarily for tail merging). In the case where CurMBB ends
476 // with a conditional branch to the next block, optimize by reversing the
477 // test and conditionally branching to SuccMBB instead.
478 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
479 const TargetInstrInfo *TII) {
480 MachineFunction *MF = CurMBB->getParent();
481 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
482 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
483 SmallVector<MachineOperand, 4> Cond;
484 DebugLoc dl; // FIXME: this is nowhere
485 if (I != MF->end() &&
486 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
487 MachineBasicBlock *NextBB = I;
488 if (TBB == NextBB && !Cond.empty() && !FBB) {
489 if (!TII->ReverseBranchCondition(Cond)) {
490 TII->RemoveBranch(*CurMBB);
491 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
496 TII->InsertBranch(*CurMBB, SuccBB, nullptr,
497 SmallVector<MachineOperand, 0>(), dl);
501 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
502 if (getHash() < o.getHash())
504 if (getHash() > o.getHash())
506 if (getBlock()->getNumber() < o.getBlock()->getNumber())
508 if (getBlock()->getNumber() > o.getBlock()->getNumber())
510 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
511 // an object with itself.
512 #ifndef _GLIBCXX_DEBUG
513 llvm_unreachable("Predecessor appears twice");
520 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
521 auto I = MergedBBFreq.find(MBB);
523 if (I != MergedBBFreq.end())
526 return MBFI.getBlockFreq(MBB);
529 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
531 MergedBBFreq[MBB] = F;
534 /// CountTerminators - Count the number of terminators in the given
535 /// block and set I to the position of the first non-terminator, if there
536 /// is one, or MBB->end() otherwise.
537 static unsigned CountTerminators(MachineBasicBlock *MBB,
538 MachineBasicBlock::iterator &I) {
540 unsigned NumTerms = 0;
542 if (I == MBB->begin()) {
547 if (!I->isTerminator()) break;
553 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
554 /// and decide if it would be profitable to merge those tails. Return the
555 /// length of the common tail and iterators to the first common instruction
557 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
558 MachineBasicBlock *MBB2,
559 unsigned minCommonTailLength,
560 unsigned &CommonTailLen,
561 MachineBasicBlock::iterator &I1,
562 MachineBasicBlock::iterator &I2,
563 MachineBasicBlock *SuccBB,
564 MachineBasicBlock *PredBB) {
565 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
566 if (CommonTailLen == 0)
568 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
569 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
572 // It's almost always profitable to merge any number of non-terminator
573 // instructions with the block that falls through into the common successor.
574 if (MBB1 == PredBB || MBB2 == PredBB) {
575 MachineBasicBlock::iterator I;
576 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
577 if (CommonTailLen > NumTerms)
581 // If one of the blocks can be completely merged and happens to be in
582 // a position where the other could fall through into it, merge any number
583 // of instructions, because it can be done without a branch.
584 // TODO: If the blocks are not adjacent, move one of them so that they are?
585 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
587 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
590 // If both blocks have an unconditional branch temporarily stripped out,
591 // count that as an additional common instruction for the following
593 unsigned EffectiveTailLen = CommonTailLen;
594 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
595 !MBB1->back().isBarrier() &&
596 !MBB2->back().isBarrier())
599 // Check if the common tail is long enough to be worthwhile.
600 if (EffectiveTailLen >= minCommonTailLength)
603 // If we are optimizing for code size, 2 instructions in common is enough if
604 // we don't have to split a block. At worst we will be introducing 1 new
605 // branch instruction, which is likely to be smaller than the 2
606 // instructions that would be deleted in the merge.
607 MachineFunction *MF = MBB1->getParent();
608 if (EffectiveTailLen >= 2 &&
609 MF->getFunction()->hasFnAttribute(Attribute::OptimizeForSize) &&
610 (I1 == MBB1->begin() || I2 == MBB2->begin()))
616 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
617 /// hash CurHash (guaranteed to match the last element). Build the vector
618 /// SameTails of all those that have the (same) largest number of instructions
619 /// in common of any pair of these blocks. SameTails entries contain an
620 /// iterator into MergePotentials (from which the MachineBasicBlock can be
621 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
622 /// instruction where the matching code sequence begins.
623 /// Order of elements in SameTails is the reverse of the order in which
624 /// those blocks appear in MergePotentials (where they are not necessarily
626 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
627 unsigned minCommonTailLength,
628 MachineBasicBlock *SuccBB,
629 MachineBasicBlock *PredBB) {
630 unsigned maxCommonTailLength = 0U;
632 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
633 MPIterator HighestMPIter = std::prev(MergePotentials.end());
634 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
635 B = MergePotentials.begin();
636 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
637 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
638 unsigned CommonTailLen;
639 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
641 CommonTailLen, TrialBBI1, TrialBBI2,
643 if (CommonTailLen > maxCommonTailLength) {
645 maxCommonTailLength = CommonTailLen;
646 HighestMPIter = CurMPIter;
647 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
649 if (HighestMPIter == CurMPIter &&
650 CommonTailLen == maxCommonTailLength)
651 SameTails.push_back(SameTailElt(I, TrialBBI2));
657 return maxCommonTailLength;
660 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
661 /// MergePotentials, restoring branches at ends of blocks as appropriate.
662 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
663 MachineBasicBlock *SuccBB,
664 MachineBasicBlock *PredBB) {
665 MPIterator CurMPIter, B;
666 for (CurMPIter = std::prev(MergePotentials.end()),
667 B = MergePotentials.begin();
668 CurMPIter->getHash() == CurHash; --CurMPIter) {
669 // Put the unconditional branch back, if we need one.
670 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
671 if (SuccBB && CurMBB != PredBB)
672 FixTail(CurMBB, SuccBB, TII);
676 if (CurMPIter->getHash() != CurHash)
678 MergePotentials.erase(CurMPIter, MergePotentials.end());
681 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
682 /// only of the common tail. Create a block that does by splitting one.
683 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
684 MachineBasicBlock *SuccBB,
685 unsigned maxCommonTailLength,
686 unsigned &commonTailIndex) {
688 unsigned TimeEstimate = ~0U;
689 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
690 // Use PredBB if possible; that doesn't require a new branch.
691 if (SameTails[i].getBlock() == PredBB) {
695 // Otherwise, make a (fairly bogus) choice based on estimate of
696 // how long it will take the various blocks to execute.
697 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
698 SameTails[i].getTailStartPos());
699 if (t <= TimeEstimate) {
705 MachineBasicBlock::iterator BBI =
706 SameTails[commonTailIndex].getTailStartPos();
707 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
709 // If the common tail includes any debug info we will take it pretty
710 // randomly from one of the inputs. Might be better to remove it?
711 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
712 << maxCommonTailLength);
714 // If the split block unconditionally falls-thru to SuccBB, it will be
715 // merged. In control flow terms it should then take SuccBB's name. e.g. If
716 // SuccBB is an inner loop, the common tail is still part of the inner loop.
717 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
718 SuccBB->getBasicBlock() : MBB->getBasicBlock();
719 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
721 DEBUG(dbgs() << "... failed!");
725 SameTails[commonTailIndex].setBlock(newMBB);
726 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
728 // If we split PredBB, newMBB is the new predecessor.
735 static bool hasIdenticalMMOs(const MachineInstr *MI1, const MachineInstr *MI2) {
736 auto I1 = MI1->memoperands_begin(), E1 = MI1->memoperands_end();
737 auto I2 = MI2->memoperands_begin(), E2 = MI2->memoperands_end();
738 if ((E1 - I1) != (E2 - I2))
740 for (; I1 != E1; ++I1, ++I2) {
748 removeMMOsFromMemoryOperations(MachineBasicBlock::iterator MBBIStartPos,
749 MachineBasicBlock &MBBCommon) {
750 // Remove MMOs from memory operations in the common block
751 // when they do not match the ones from the block being tail-merged.
752 // This ensures later passes conservatively compute dependencies.
753 MachineBasicBlock *MBB = MBBIStartPos->getParent();
754 // Note CommonTailLen does not necessarily matches the size of
755 // the common BB nor all its instructions because of debug
756 // instructions differences.
757 unsigned CommonTailLen = 0;
758 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
761 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
762 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
763 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
764 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
766 while (CommonTailLen--) {
767 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
770 if (MBBI->isDebugValue()) {
775 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
778 assert(MBBICommon != MBBIECommon &&
779 "Reached BB end within common tail length!");
780 assert(MBBICommon->isIdenticalTo(&*MBBI) && "Expected matching MIIs!");
782 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
783 if (!hasIdenticalMMOs(&*MBBI, &*MBBICommon))
784 MBBICommon->clearMemRefs();
791 // See if any of the blocks in MergePotentials (which all have a common single
792 // successor, or all have no successor) can be tail-merged. If there is a
793 // successor, any blocks in MergePotentials that are not tail-merged and
794 // are not immediately before Succ must have an unconditional branch to
795 // Succ added (but the predecessor/successor lists need no adjustment).
796 // The lone predecessor of Succ that falls through into Succ,
797 // if any, is given in PredBB.
799 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
800 MachineBasicBlock *PredBB) {
801 bool MadeChange = false;
803 // Except for the special cases below, tail-merge if there are at least
804 // this many instructions in common.
805 unsigned minCommonTailLength = TailMergeSize;
807 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
808 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
809 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
810 << (i == e-1 ? "" : ", ");
813 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
815 dbgs() << " which has fall-through from BB#"
816 << PredBB->getNumber() << "\n";
818 dbgs() << "Looking for common tails of at least "
819 << minCommonTailLength << " instruction"
820 << (minCommonTailLength == 1 ? "" : "s") << '\n';
823 // Sort by hash value so that blocks with identical end sequences sort
825 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
827 // Walk through equivalence sets looking for actual exact matches.
828 while (MergePotentials.size() > 1) {
829 unsigned CurHash = MergePotentials.back().getHash();
831 // Build SameTails, identifying the set of blocks with this hash code
832 // and with the maximum number of instructions in common.
833 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
837 // If we didn't find any pair that has at least minCommonTailLength
838 // instructions in common, remove all blocks with this hash code and retry.
839 if (SameTails.empty()) {
840 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
844 // If one of the blocks is the entire common tail (and not the entry
845 // block, which we can't jump to), we can treat all blocks with this same
846 // tail at once. Use PredBB if that is one of the possibilities, as that
847 // will not introduce any extra branches.
848 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
849 getParent()->begin();
850 unsigned commonTailIndex = SameTails.size();
851 // If there are two blocks, check to see if one can be made to fall through
853 if (SameTails.size() == 2 &&
854 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
855 SameTails[1].tailIsWholeBlock())
857 else if (SameTails.size() == 2 &&
858 SameTails[1].getBlock()->isLayoutSuccessor(
859 SameTails[0].getBlock()) &&
860 SameTails[0].tailIsWholeBlock())
863 // Otherwise just pick one, favoring the fall-through predecessor if
865 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
866 MachineBasicBlock *MBB = SameTails[i].getBlock();
867 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
873 if (SameTails[i].tailIsWholeBlock())
878 if (commonTailIndex == SameTails.size() ||
879 (SameTails[commonTailIndex].getBlock() == PredBB &&
880 !SameTails[commonTailIndex].tailIsWholeBlock())) {
881 // None of the blocks consist entirely of the common tail.
882 // Split a block so that one does.
883 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
884 maxCommonTailLength, commonTailIndex)) {
885 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
890 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
892 // Recompute commont tail MBB's edge weights and block frequency.
893 setCommonTailEdgeWeights(*MBB);
895 // MBB is common tail. Adjust all other BB's to jump to this one.
896 // Traversal must be forwards so erases work.
897 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
899 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
900 if (commonTailIndex == i)
902 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
903 << (i == e-1 ? "" : ", "));
904 // Remove MMOs from memory operations as needed.
905 removeMMOsFromMemoryOperations(SameTails[i].getTailStartPos(), *MBB);
906 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
907 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
908 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
909 MergePotentials.erase(SameTails[i].getMPIter());
911 DEBUG(dbgs() << "\n");
912 // We leave commonTailIndex in the worklist in case there are other blocks
913 // that match it with a smaller number of instructions.
919 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
920 bool MadeChange = false;
921 if (!EnableTailMerge) return MadeChange;
923 // First find blocks with no successors.
924 MergePotentials.clear();
925 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
926 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
927 if (TriedMerging.count(I))
930 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
933 // If this is a large problem, avoid visiting the same basic blocks
935 if (MergePotentials.size() == TailMergeThreshold)
936 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
937 TriedMerging.insert(MergePotentials[i].getBlock());
939 // See if we can do any tail merging on those.
940 if (MergePotentials.size() >= 2)
941 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
943 // Look at blocks (IBB) with multiple predecessors (PBB).
944 // We change each predecessor to a canonical form, by
945 // (1) temporarily removing any unconditional branch from the predecessor
947 // (2) alter conditional branches so they branch to the other block
948 // not IBB; this may require adding back an unconditional branch to IBB
949 // later, where there wasn't one coming in. E.g.
951 // fallthrough to QBB
954 // with a conceptual B to IBB after that, which never actually exists.
955 // With those changes, we see whether the predecessors' tails match,
956 // and merge them if so. We change things out of canonical form and
957 // back to the way they were later in the process. (OptimizeBranches
958 // would undo some of this, but we can't use it, because we'd get into
959 // a compile-time infinite loop repeatedly doing and undoing the same
962 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
964 if (I->pred_size() < 2) continue;
965 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
966 MachineBasicBlock *IBB = I;
967 MachineBasicBlock *PredBB = std::prev(I);
968 MergePotentials.clear();
969 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
971 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
972 MachineBasicBlock *PBB = *P;
973 if (TriedMerging.count(PBB))
976 // Skip blocks that loop to themselves, can't tail merge these.
980 // Visit each predecessor only once.
981 if (!UniquePreds.insert(PBB).second)
984 // Skip blocks which may jump to a landing pad. Can't tail merge these.
985 if (PBB->getLandingPadSuccessor())
988 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
989 SmallVector<MachineOperand, 4> Cond;
990 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
991 // Failing case: IBB is the target of a cbr, and we cannot reverse the
993 SmallVector<MachineOperand, 4> NewCond(Cond);
994 if (!Cond.empty() && TBB == IBB) {
995 if (TII->ReverseBranchCondition(NewCond))
997 // This is the QBB case described above
999 FBB = std::next(MachineFunction::iterator(PBB));
1002 // Failing case: the only way IBB can be reached from PBB is via
1003 // exception handling. Happens for landing pads. Would be nice to have
1004 // a bit in the edge so we didn't have to do all this.
1005 if (IBB->isLandingPad()) {
1006 MachineFunction::iterator IP = PBB; IP++;
1007 MachineBasicBlock *PredNextBB = nullptr;
1011 if (IBB != PredNextBB) // fallthrough
1014 if (TBB != IBB && FBB != IBB) // cbr then ubr
1016 } else if (Cond.empty()) {
1017 if (TBB != IBB) // ubr
1020 if (TBB != IBB && IBB != PredNextBB) // cbr
1025 // Remove the unconditional branch at the end, if any.
1026 if (TBB && (Cond.empty() || FBB)) {
1027 DebugLoc dl; // FIXME: this is nowhere
1028 TII->RemoveBranch(*PBB);
1030 // reinsert conditional branch only, for now
1031 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1035 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
1039 // If this is a large problem, avoid visiting the same basic blocks multiple
1041 if (MergePotentials.size() == TailMergeThreshold)
1042 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1043 TriedMerging.insert(MergePotentials[i].getBlock());
1045 if (MergePotentials.size() >= 2)
1046 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
1048 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1049 // result of removing blocks in TryTailMergeBlocks.
1050 PredBB = std::prev(I); // this may have been changed in TryTailMergeBlocks
1051 if (MergePotentials.size() == 1 &&
1052 MergePotentials.begin()->getBlock() != PredBB)
1053 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1059 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1060 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1061 BlockFrequency AccumulatedMBBFreq;
1063 // Aggregate edge frequency of successor edge j:
1064 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1065 // where bb is a basic block that is in SameTails.
1066 for (const auto &Src : SameTails) {
1067 const MachineBasicBlock *SrcMBB = Src.getBlock();
1068 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1069 AccumulatedMBBFreq += BlockFreq;
1071 // It is not necessary to recompute edge weights if TailBB has less than two
1073 if (TailMBB.succ_size() <= 1)
1076 auto EdgeFreq = EdgeFreqLs.begin();
1078 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1079 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1080 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1083 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1085 if (TailMBB.succ_size() <= 1)
1088 auto MaxEdgeFreq = *std::max_element(EdgeFreqLs.begin(), EdgeFreqLs.end());
1089 uint64_t Scale = MaxEdgeFreq.getFrequency() / UINT32_MAX + 1;
1090 auto EdgeFreq = EdgeFreqLs.begin();
1092 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1093 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1094 TailMBB.setSuccWeight(SuccI, EdgeFreq->getFrequency() / Scale);
1097 //===----------------------------------------------------------------------===//
1098 // Branch Optimization
1099 //===----------------------------------------------------------------------===//
1101 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1102 bool MadeChange = false;
1104 // Make sure blocks are numbered in order
1105 MF.RenumberBlocks();
1107 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1109 MachineBasicBlock *MBB = I++;
1110 MadeChange |= OptimizeBlock(MBB);
1112 // If it is dead, remove it.
1113 if (MBB->pred_empty()) {
1114 RemoveDeadBlock(MBB);
1122 // Blocks should be considered empty if they contain only debug info;
1123 // else the debug info would affect codegen.
1124 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1125 return MBB->getFirstNonDebugInstr() == MBB->end();
1128 // Blocks with only debug info and branches should be considered the same
1129 // as blocks with only branches.
1130 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1131 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1132 assert(I != MBB->end() && "empty block!");
1133 return I->isBranch();
1136 /// IsBetterFallthrough - Return true if it would be clearly better to
1137 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1138 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1139 /// result in infinite loops.
1140 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1141 MachineBasicBlock *MBB2) {
1142 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1143 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1144 // optimize branches that branch to either a return block or an assert block
1145 // into a fallthrough to the return.
1146 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1147 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1148 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1151 // If there is a clear successor ordering we make sure that one block
1152 // will fall through to the next
1153 if (MBB1->isSuccessor(MBB2)) return true;
1154 if (MBB2->isSuccessor(MBB1)) return false;
1156 return MBB2I->isCall() && !MBB1I->isCall();
1159 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1160 /// instructions on the block.
1161 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1162 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1163 if (I != MBB.end() && I->isBranch())
1164 return I->getDebugLoc();
1168 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1169 /// block. This is never called on the entry block.
1170 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1171 bool MadeChange = false;
1172 MachineFunction &MF = *MBB->getParent();
1175 MachineFunction::iterator FallThrough = MBB;
1178 // If this block is empty, make everyone use its fall-through, not the block
1179 // explicitly. Landing pads should not do this since the landing-pad table
1180 // points to this block. Blocks with their addresses taken shouldn't be
1182 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1183 // Dead block? Leave for cleanup later.
1184 if (MBB->pred_empty()) return MadeChange;
1186 if (FallThrough == MF.end()) {
1187 // TODO: Simplify preds to not branch here if possible!
1188 } else if (FallThrough->isLandingPad()) {
1189 // Don't rewrite to a landing pad fallthough. That could lead to the case
1190 // where a BB jumps to more than one landing pad.
1191 // TODO: Is it ever worth rewriting predecessors which don't already
1192 // jump to a landing pad, and so can safely jump to the fallthrough?
1194 // Rewrite all predecessors of the old block to go to the fallthrough
1196 while (!MBB->pred_empty()) {
1197 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1198 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1200 // If MBB was the target of a jump table, update jump tables to go to the
1201 // fallthrough instead.
1202 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1203 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1209 // Check to see if we can simplify the terminator of the block before this
1211 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1213 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1214 SmallVector<MachineOperand, 4> PriorCond;
1215 bool PriorUnAnalyzable =
1216 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1217 if (!PriorUnAnalyzable) {
1218 // If the CFG for the prior block has extra edges, remove them.
1219 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1220 !PriorCond.empty());
1222 // If the previous branch is conditional and both conditions go to the same
1223 // destination, remove the branch, replacing it with an unconditional one or
1225 if (PriorTBB && PriorTBB == PriorFBB) {
1226 DebugLoc dl = getBranchDebugLoc(PrevBB);
1227 TII->RemoveBranch(PrevBB);
1229 if (PriorTBB != MBB)
1230 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1233 goto ReoptimizeBlock;
1236 // If the previous block unconditionally falls through to this block and
1237 // this block has no other predecessors, move the contents of this block
1238 // into the prior block. This doesn't usually happen when SimplifyCFG
1239 // has been used, but it can happen if tail merging splits a fall-through
1240 // predecessor of a block.
1241 // This has to check PrevBB->succ_size() because EH edges are ignored by
1243 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1244 PrevBB.succ_size() == 1 &&
1245 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1246 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1247 << "From MBB: " << *MBB);
1248 // Remove redundant DBG_VALUEs first.
1249 if (PrevBB.begin() != PrevBB.end()) {
1250 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1252 MachineBasicBlock::iterator MBBIter = MBB->begin();
1253 // Check if DBG_VALUE at the end of PrevBB is identical to the
1254 // DBG_VALUE at the beginning of MBB.
1255 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1256 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1257 if (!MBBIter->isIdenticalTo(PrevBBIter))
1259 MachineInstr *DuplicateDbg = MBBIter;
1260 ++MBBIter; -- PrevBBIter;
1261 DuplicateDbg->eraseFromParent();
1264 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1265 PrevBB.removeSuccessor(PrevBB.succ_begin());
1266 assert(PrevBB.succ_empty());
1267 PrevBB.transferSuccessors(MBB);
1272 // If the previous branch *only* branches to *this* block (conditional or
1273 // not) remove the branch.
1274 if (PriorTBB == MBB && !PriorFBB) {
1275 TII->RemoveBranch(PrevBB);
1278 goto ReoptimizeBlock;
1281 // If the prior block branches somewhere else on the condition and here if
1282 // the condition is false, remove the uncond second branch.
1283 if (PriorFBB == MBB) {
1284 DebugLoc dl = getBranchDebugLoc(PrevBB);
1285 TII->RemoveBranch(PrevBB);
1286 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1289 goto ReoptimizeBlock;
1292 // If the prior block branches here on true and somewhere else on false, and
1293 // if the branch condition is reversible, reverse the branch to create a
1295 if (PriorTBB == MBB) {
1296 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1297 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1298 DebugLoc dl = getBranchDebugLoc(PrevBB);
1299 TII->RemoveBranch(PrevBB);
1300 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1303 goto ReoptimizeBlock;
1307 // If this block has no successors (e.g. it is a return block or ends with
1308 // a call to a no-return function like abort or __cxa_throw) and if the pred
1309 // falls through into this block, and if it would otherwise fall through
1310 // into the block after this, move this block to the end of the function.
1312 // We consider it more likely that execution will stay in the function (e.g.
1313 // due to loops) than it is to exit it. This asserts in loops etc, moving
1314 // the assert condition out of the loop body.
1315 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1316 MachineFunction::iterator(PriorTBB) == FallThrough &&
1317 !MBB->canFallThrough()) {
1318 bool DoTransform = true;
1320 // We have to be careful that the succs of PredBB aren't both no-successor
1321 // blocks. If neither have successors and if PredBB is the second from
1322 // last block in the function, we'd just keep swapping the two blocks for
1323 // last. Only do the swap if one is clearly better to fall through than
1325 if (FallThrough == --MF.end() &&
1326 !IsBetterFallthrough(PriorTBB, MBB))
1327 DoTransform = false;
1330 // Reverse the branch so we will fall through on the previous true cond.
1331 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1332 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1333 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1334 << "To make fallthrough to: " << *PriorTBB << "\n");
1336 DebugLoc dl = getBranchDebugLoc(PrevBB);
1337 TII->RemoveBranch(PrevBB);
1338 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1340 // Move this block to the end of the function.
1341 MBB->moveAfter(--MF.end());
1350 // Analyze the branch in the current block.
1351 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1352 SmallVector<MachineOperand, 4> CurCond;
1353 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1354 if (!CurUnAnalyzable) {
1355 // If the CFG for the prior block has extra edges, remove them.
1356 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1358 // If this is a two-way branch, and the FBB branches to this block, reverse
1359 // the condition so the single-basic-block loop is faster. Instead of:
1360 // Loop: xxx; jcc Out; jmp Loop
1362 // Loop: xxx; jncc Loop; jmp Out
1363 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1364 SmallVector<MachineOperand, 4> NewCond(CurCond);
1365 if (!TII->ReverseBranchCondition(NewCond)) {
1366 DebugLoc dl = getBranchDebugLoc(*MBB);
1367 TII->RemoveBranch(*MBB);
1368 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1371 goto ReoptimizeBlock;
1375 // If this branch is the only thing in its block, see if we can forward
1376 // other blocks across it.
1377 if (CurTBB && CurCond.empty() && !CurFBB &&
1378 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1379 !MBB->hasAddressTaken()) {
1380 DebugLoc dl = getBranchDebugLoc(*MBB);
1381 // This block may contain just an unconditional branch. Because there can
1382 // be 'non-branch terminators' in the block, try removing the branch and
1383 // then seeing if the block is empty.
1384 TII->RemoveBranch(*MBB);
1385 // If the only things remaining in the block are debug info, remove these
1386 // as well, so this will behave the same as an empty block in non-debug
1388 if (IsEmptyBlock(MBB)) {
1389 // Make the block empty, losing the debug info (we could probably
1390 // improve this in some cases.)
1391 MBB->erase(MBB->begin(), MBB->end());
1393 // If this block is just an unconditional branch to CurTBB, we can
1394 // usually completely eliminate the block. The only case we cannot
1395 // completely eliminate the block is when the block before this one
1396 // falls through into MBB and we can't understand the prior block's branch
1399 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1400 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1401 !PrevBB.isSuccessor(MBB)) {
1402 // If the prior block falls through into us, turn it into an
1403 // explicit branch to us to make updates simpler.
1404 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1405 PriorTBB != MBB && PriorFBB != MBB) {
1407 assert(PriorCond.empty() && !PriorFBB &&
1408 "Bad branch analysis");
1411 assert(!PriorFBB && "Machine CFG out of date!");
1414 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1415 TII->RemoveBranch(PrevBB);
1416 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1419 // Iterate through all the predecessors, revectoring each in-turn.
1421 bool DidChange = false;
1422 bool HasBranchToSelf = false;
1423 while(PI != MBB->pred_size()) {
1424 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1426 // If this block has an uncond branch to itself, leave it.
1428 HasBranchToSelf = true;
1431 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1432 // If this change resulted in PMBB ending in a conditional
1433 // branch where both conditions go to the same destination,
1434 // change this to an unconditional branch (and fix the CFG).
1435 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1436 SmallVector<MachineOperand, 4> NewCurCond;
1437 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1438 NewCurFBB, NewCurCond, true);
1439 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1440 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1441 TII->RemoveBranch(*PMBB);
1443 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1446 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1451 // Change any jumptables to go to the new MBB.
1452 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1453 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1457 if (!HasBranchToSelf) return MadeChange;
1462 // Add the branch back if the block is more than just an uncond branch.
1463 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1467 // If the prior block doesn't fall through into this block, and if this
1468 // block doesn't fall through into some other block, see if we can find a
1469 // place to move this block where a fall-through will happen.
1470 if (!PrevBB.canFallThrough()) {
1472 // Now we know that there was no fall-through into this block, check to
1473 // see if it has a fall-through into its successor.
1474 bool CurFallsThru = MBB->canFallThrough();
1476 if (!MBB->isLandingPad()) {
1477 // Check all the predecessors of this block. If one of them has no fall
1478 // throughs, move this block right after it.
1479 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1480 E = MBB->pred_end(); PI != E; ++PI) {
1481 // Analyze the branch at the end of the pred.
1482 MachineBasicBlock *PredBB = *PI;
1483 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1484 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1485 SmallVector<MachineOperand, 4> PredCond;
1486 if (PredBB != MBB && !PredBB->canFallThrough() &&
1487 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1488 && (!CurFallsThru || !CurTBB || !CurFBB)
1489 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1490 // If the current block doesn't fall through, just move it.
1491 // If the current block can fall through and does not end with a
1492 // conditional branch, we need to append an unconditional jump to
1493 // the (current) next block. To avoid a possible compile-time
1494 // infinite loop, move blocks only backward in this case.
1495 // Also, if there are already 2 branches here, we cannot add a third;
1496 // this means we have the case
1501 MachineBasicBlock *NextBB =
1502 std::next(MachineFunction::iterator(MBB));
1504 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1506 MBB->moveAfter(PredBB);
1508 goto ReoptimizeBlock;
1513 if (!CurFallsThru) {
1514 // Check all successors to see if we can move this block before it.
1515 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1516 E = MBB->succ_end(); SI != E; ++SI) {
1517 // Analyze the branch at the end of the block before the succ.
1518 MachineBasicBlock *SuccBB = *SI;
1519 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1521 // If this block doesn't already fall-through to that successor, and if
1522 // the succ doesn't already have a block that can fall through into it,
1523 // and if the successor isn't an EH destination, we can arrange for the
1524 // fallthrough to happen.
1525 if (SuccBB != MBB && &*SuccPrev != MBB &&
1526 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1527 !SuccBB->isLandingPad()) {
1528 MBB->moveBefore(SuccBB);
1530 goto ReoptimizeBlock;
1534 // Okay, there is no really great place to put this block. If, however,
1535 // the block before this one would be a fall-through if this block were
1536 // removed, move this block to the end of the function.
1537 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1538 SmallVector<MachineOperand, 4> PrevCond;
1539 if (FallThrough != MF.end() &&
1540 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1541 PrevBB.isSuccessor(FallThrough)) {
1542 MBB->moveAfter(--MF.end());
1552 //===----------------------------------------------------------------------===//
1553 // Hoist Common Code
1554 //===----------------------------------------------------------------------===//
1556 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1557 /// blocks to their common predecessor.
1558 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1559 bool MadeChange = false;
1560 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1561 MachineBasicBlock *MBB = I++;
1562 MadeChange |= HoistCommonCodeInSuccs(MBB);
1568 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1569 /// its 'true' successor.
1570 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1571 MachineBasicBlock *TrueBB) {
1572 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1573 E = BB->succ_end(); SI != E; ++SI) {
1574 MachineBasicBlock *SuccBB = *SI;
1575 if (SuccBB != TrueBB)
1581 template <class Container>
1582 static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
1584 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1585 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1592 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1593 /// in successors to. The location is usually just before the terminator,
1594 /// however if the terminator is a conditional branch and its previous
1595 /// instruction is the flag setting instruction, the previous instruction is
1596 /// the preferred location. This function also gathers uses and defs of the
1597 /// instructions from the insertion point to the end of the block. The data is
1598 /// used by HoistCommonCodeInSuccs to ensure safety.
1600 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1601 const TargetInstrInfo *TII,
1602 const TargetRegisterInfo *TRI,
1603 SmallSet<unsigned,4> &Uses,
1604 SmallSet<unsigned,4> &Defs) {
1605 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1606 if (!TII->isUnpredicatedTerminator(Loc))
1609 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1610 const MachineOperand &MO = Loc->getOperand(i);
1613 unsigned Reg = MO.getReg();
1617 addRegAndItsAliases(Reg, TRI, Uses);
1620 // Don't try to hoist code in the rare case the terminator defines a
1621 // register that is later used.
1624 // If the terminator defines a register, make sure we don't hoist
1625 // the instruction whose def might be clobbered by the terminator.
1626 addRegAndItsAliases(Reg, TRI, Defs);
1632 if (Loc == MBB->begin())
1635 // The terminator is probably a conditional branch, try not to separate the
1636 // branch from condition setting instruction.
1637 MachineBasicBlock::iterator PI = Loc;
1639 while (PI != MBB->begin() && PI->isDebugValue())
1643 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1644 const MachineOperand &MO = PI->getOperand(i);
1645 // If PI has a regmask operand, it is probably a call. Separate away.
1648 if (!MO.isReg() || MO.isUse())
1650 unsigned Reg = MO.getReg();
1653 if (Uses.count(Reg))
1657 // The condition setting instruction is not just before the conditional
1661 // Be conservative, don't insert instruction above something that may have
1662 // side-effects. And since it's potentially bad to separate flag setting
1663 // instruction from the conditional branch, just abort the optimization
1665 // Also avoid moving code above predicated instruction since it's hard to
1666 // reason about register liveness with predicated instruction.
1667 bool DontMoveAcrossStore = true;
1668 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(PI))
1672 // Find out what registers are live. Note this routine is ignoring other live
1673 // registers which are only used by instructions in successor blocks.
1674 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1675 const MachineOperand &MO = PI->getOperand(i);
1678 unsigned Reg = MO.getReg();
1682 addRegAndItsAliases(Reg, TRI, Uses);
1684 if (Uses.erase(Reg)) {
1685 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1686 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1687 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1690 addRegAndItsAliases(Reg, TRI, Defs);
1697 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1698 /// sequence at the start of the function, move the instructions before MBB
1699 /// terminator if it's legal.
1700 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1701 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1702 SmallVector<MachineOperand, 4> Cond;
1703 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1706 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1708 // Malformed bcc? True and false blocks are the same?
1711 // Restrict the optimization to cases where MBB is the only predecessor,
1712 // it is an obvious win.
1713 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1716 // Find a suitable position to hoist the common instructions to. Also figure
1717 // out which registers are used or defined by instructions from the insertion
1718 // point to the end of the block.
1719 SmallSet<unsigned, 4> Uses, Defs;
1720 MachineBasicBlock::iterator Loc =
1721 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1722 if (Loc == MBB->end())
1725 bool HasDups = false;
1726 SmallVector<unsigned, 4> LocalDefs;
1727 SmallSet<unsigned, 4> LocalDefsSet;
1728 MachineBasicBlock::iterator TIB = TBB->begin();
1729 MachineBasicBlock::iterator FIB = FBB->begin();
1730 MachineBasicBlock::iterator TIE = TBB->end();
1731 MachineBasicBlock::iterator FIE = FBB->end();
1732 while (TIB != TIE && FIB != FIE) {
1733 // Skip dbg_value instructions. These do not count.
1734 if (TIB->isDebugValue()) {
1735 while (TIB != TIE && TIB->isDebugValue())
1740 if (FIB->isDebugValue()) {
1741 while (FIB != FIE && FIB->isDebugValue())
1746 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1749 if (TII->isPredicated(TIB))
1750 // Hard to reason about register liveness with predicated instruction.
1754 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1755 MachineOperand &MO = TIB->getOperand(i);
1756 // Don't attempt to hoist instructions with register masks.
1757 if (MO.isRegMask()) {
1763 unsigned Reg = MO.getReg();
1767 if (Uses.count(Reg)) {
1768 // Avoid clobbering a register that's used by the instruction at
1769 // the point of insertion.
1774 if (Defs.count(Reg) && !MO.isDead()) {
1775 // Don't hoist the instruction if the def would be clobber by the
1776 // instruction at the point insertion. FIXME: This is overly
1777 // conservative. It should be possible to hoist the instructions
1778 // in BB2 in the following example:
1780 // r1, eflag = op1 r2, r3
1789 } else if (!LocalDefsSet.count(Reg)) {
1790 if (Defs.count(Reg)) {
1791 // Use is defined by the instruction at the point of insertion.
1796 if (MO.isKill() && Uses.count(Reg))
1797 // Kills a register that's read by the instruction at the point of
1798 // insertion. Remove the kill marker.
1799 MO.setIsKill(false);
1805 bool DontMoveAcrossStore = true;
1806 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1809 // Remove kills from LocalDefsSet, these registers had short live ranges.
1810 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1811 MachineOperand &MO = TIB->getOperand(i);
1812 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1814 unsigned Reg = MO.getReg();
1815 if (!Reg || !LocalDefsSet.count(Reg))
1817 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1818 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1819 LocalDefsSet.erase(*AI);
1821 LocalDefsSet.erase(Reg);
1825 // Track local defs so we can update liveins.
1826 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1827 MachineOperand &MO = TIB->getOperand(i);
1828 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1830 unsigned Reg = MO.getReg();
1833 LocalDefs.push_back(Reg);
1834 addRegAndItsAliases(Reg, TRI, LocalDefsSet);
1845 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1846 FBB->erase(FBB->begin(), FIB);
1849 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1850 unsigned Def = LocalDefs[i];
1851 if (LocalDefsSet.count(Def)) {
1852 TBB->addLiveIn(Def);
1853 FBB->addLiveIn(Def);