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 && MF->getFunction()->optForSize() &&
609 (I1 == MBB1->begin() || I2 == MBB2->begin()))
615 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
616 /// hash CurHash (guaranteed to match the last element). Build the vector
617 /// SameTails of all those that have the (same) largest number of instructions
618 /// in common of any pair of these blocks. SameTails entries contain an
619 /// iterator into MergePotentials (from which the MachineBasicBlock can be
620 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
621 /// instruction where the matching code sequence begins.
622 /// Order of elements in SameTails is the reverse of the order in which
623 /// those blocks appear in MergePotentials (where they are not necessarily
625 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
626 unsigned minCommonTailLength,
627 MachineBasicBlock *SuccBB,
628 MachineBasicBlock *PredBB) {
629 unsigned maxCommonTailLength = 0U;
631 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
632 MPIterator HighestMPIter = std::prev(MergePotentials.end());
633 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
634 B = MergePotentials.begin();
635 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
636 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
637 unsigned CommonTailLen;
638 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
640 CommonTailLen, TrialBBI1, TrialBBI2,
642 if (CommonTailLen > maxCommonTailLength) {
644 maxCommonTailLength = CommonTailLen;
645 HighestMPIter = CurMPIter;
646 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
648 if (HighestMPIter == CurMPIter &&
649 CommonTailLen == maxCommonTailLength)
650 SameTails.push_back(SameTailElt(I, TrialBBI2));
656 return maxCommonTailLength;
659 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
660 /// MergePotentials, restoring branches at ends of blocks as appropriate.
661 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
662 MachineBasicBlock *SuccBB,
663 MachineBasicBlock *PredBB) {
664 MPIterator CurMPIter, B;
665 for (CurMPIter = std::prev(MergePotentials.end()),
666 B = MergePotentials.begin();
667 CurMPIter->getHash() == CurHash; --CurMPIter) {
668 // Put the unconditional branch back, if we need one.
669 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
670 if (SuccBB && CurMBB != PredBB)
671 FixTail(CurMBB, SuccBB, TII);
675 if (CurMPIter->getHash() != CurHash)
677 MergePotentials.erase(CurMPIter, MergePotentials.end());
680 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
681 /// only of the common tail. Create a block that does by splitting one.
682 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
683 MachineBasicBlock *SuccBB,
684 unsigned maxCommonTailLength,
685 unsigned &commonTailIndex) {
687 unsigned TimeEstimate = ~0U;
688 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
689 // Use PredBB if possible; that doesn't require a new branch.
690 if (SameTails[i].getBlock() == PredBB) {
694 // Otherwise, make a (fairly bogus) choice based on estimate of
695 // how long it will take the various blocks to execute.
696 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
697 SameTails[i].getTailStartPos());
698 if (t <= TimeEstimate) {
704 MachineBasicBlock::iterator BBI =
705 SameTails[commonTailIndex].getTailStartPos();
706 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
708 // If the common tail includes any debug info we will take it pretty
709 // randomly from one of the inputs. Might be better to remove it?
710 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
711 << maxCommonTailLength);
713 // If the split block unconditionally falls-thru to SuccBB, it will be
714 // merged. In control flow terms it should then take SuccBB's name. e.g. If
715 // SuccBB is an inner loop, the common tail is still part of the inner loop.
716 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
717 SuccBB->getBasicBlock() : MBB->getBasicBlock();
718 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
720 DEBUG(dbgs() << "... failed!");
724 SameTails[commonTailIndex].setBlock(newMBB);
725 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
727 // If we split PredBB, newMBB is the new predecessor.
734 static bool hasIdenticalMMOs(const MachineInstr *MI1, const MachineInstr *MI2) {
735 auto I1 = MI1->memoperands_begin(), E1 = MI1->memoperands_end();
736 auto I2 = MI2->memoperands_begin(), E2 = MI2->memoperands_end();
737 if ((E1 - I1) != (E2 - I2))
739 for (; I1 != E1; ++I1, ++I2) {
747 removeMMOsFromMemoryOperations(MachineBasicBlock::iterator MBBIStartPos,
748 MachineBasicBlock &MBBCommon) {
749 // Remove MMOs from memory operations in the common block
750 // when they do not match the ones from the block being tail-merged.
751 // This ensures later passes conservatively compute dependencies.
752 MachineBasicBlock *MBB = MBBIStartPos->getParent();
753 // Note CommonTailLen does not necessarily matches the size of
754 // the common BB nor all its instructions because of debug
755 // instructions differences.
756 unsigned CommonTailLen = 0;
757 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
760 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
761 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
762 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
763 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
765 while (CommonTailLen--) {
766 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
769 if (MBBI->isDebugValue()) {
774 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
777 assert(MBBICommon != MBBIECommon &&
778 "Reached BB end within common tail length!");
779 assert(MBBICommon->isIdenticalTo(&*MBBI) && "Expected matching MIIs!");
781 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
782 if (!hasIdenticalMMOs(&*MBBI, &*MBBICommon))
783 MBBICommon->clearMemRefs();
790 // See if any of the blocks in MergePotentials (which all have a common single
791 // successor, or all have no successor) can be tail-merged. If there is a
792 // successor, any blocks in MergePotentials that are not tail-merged and
793 // are not immediately before Succ must have an unconditional branch to
794 // Succ added (but the predecessor/successor lists need no adjustment).
795 // The lone predecessor of Succ that falls through into Succ,
796 // if any, is given in PredBB.
798 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
799 MachineBasicBlock *PredBB) {
800 bool MadeChange = false;
802 // Except for the special cases below, tail-merge if there are at least
803 // this many instructions in common.
804 unsigned minCommonTailLength = TailMergeSize;
806 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
807 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
808 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
809 << (i == e-1 ? "" : ", ");
812 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
814 dbgs() << " which has fall-through from BB#"
815 << PredBB->getNumber() << "\n";
817 dbgs() << "Looking for common tails of at least "
818 << minCommonTailLength << " instruction"
819 << (minCommonTailLength == 1 ? "" : "s") << '\n';
822 // Sort by hash value so that blocks with identical end sequences sort
824 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
826 // Walk through equivalence sets looking for actual exact matches.
827 while (MergePotentials.size() > 1) {
828 unsigned CurHash = MergePotentials.back().getHash();
830 // Build SameTails, identifying the set of blocks with this hash code
831 // and with the maximum number of instructions in common.
832 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
836 // If we didn't find any pair that has at least minCommonTailLength
837 // instructions in common, remove all blocks with this hash code and retry.
838 if (SameTails.empty()) {
839 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
843 // If one of the blocks is the entire common tail (and not the entry
844 // block, which we can't jump to), we can treat all blocks with this same
845 // tail at once. Use PredBB if that is one of the possibilities, as that
846 // will not introduce any extra branches.
847 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
848 getParent()->begin();
849 unsigned commonTailIndex = SameTails.size();
850 // If there are two blocks, check to see if one can be made to fall through
852 if (SameTails.size() == 2 &&
853 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
854 SameTails[1].tailIsWholeBlock())
856 else if (SameTails.size() == 2 &&
857 SameTails[1].getBlock()->isLayoutSuccessor(
858 SameTails[0].getBlock()) &&
859 SameTails[0].tailIsWholeBlock())
862 // Otherwise just pick one, favoring the fall-through predecessor if
864 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
865 MachineBasicBlock *MBB = SameTails[i].getBlock();
866 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
872 if (SameTails[i].tailIsWholeBlock())
877 if (commonTailIndex == SameTails.size() ||
878 (SameTails[commonTailIndex].getBlock() == PredBB &&
879 !SameTails[commonTailIndex].tailIsWholeBlock())) {
880 // None of the blocks consist entirely of the common tail.
881 // Split a block so that one does.
882 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
883 maxCommonTailLength, commonTailIndex)) {
884 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
889 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
891 // Recompute commont tail MBB's edge weights and block frequency.
892 setCommonTailEdgeWeights(*MBB);
894 // MBB is common tail. Adjust all other BB's to jump to this one.
895 // Traversal must be forwards so erases work.
896 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
898 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
899 if (commonTailIndex == i)
901 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
902 << (i == e-1 ? "" : ", "));
903 // Remove MMOs from memory operations as needed.
904 removeMMOsFromMemoryOperations(SameTails[i].getTailStartPos(), *MBB);
905 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
906 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
907 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
908 MergePotentials.erase(SameTails[i].getMPIter());
910 DEBUG(dbgs() << "\n");
911 // We leave commonTailIndex in the worklist in case there are other blocks
912 // that match it with a smaller number of instructions.
918 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
919 bool MadeChange = false;
920 if (!EnableTailMerge) return MadeChange;
922 // First find blocks with no successors.
923 MergePotentials.clear();
924 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
925 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
926 if (TriedMerging.count(I))
929 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
932 // If this is a large problem, avoid visiting the same basic blocks
934 if (MergePotentials.size() == TailMergeThreshold)
935 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
936 TriedMerging.insert(MergePotentials[i].getBlock());
938 // See if we can do any tail merging on those.
939 if (MergePotentials.size() >= 2)
940 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
942 // Look at blocks (IBB) with multiple predecessors (PBB).
943 // We change each predecessor to a canonical form, by
944 // (1) temporarily removing any unconditional branch from the predecessor
946 // (2) alter conditional branches so they branch to the other block
947 // not IBB; this may require adding back an unconditional branch to IBB
948 // later, where there wasn't one coming in. E.g.
950 // fallthrough to QBB
953 // with a conceptual B to IBB after that, which never actually exists.
954 // With those changes, we see whether the predecessors' tails match,
955 // and merge them if so. We change things out of canonical form and
956 // back to the way they were later in the process. (OptimizeBranches
957 // would undo some of this, but we can't use it, because we'd get into
958 // a compile-time infinite loop repeatedly doing and undoing the same
961 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
963 if (I->pred_size() < 2) continue;
964 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
965 MachineBasicBlock *IBB = I;
966 MachineBasicBlock *PredBB = std::prev(I);
967 MergePotentials.clear();
968 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
970 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
971 MachineBasicBlock *PBB = *P;
972 if (TriedMerging.count(PBB))
975 // Skip blocks that loop to themselves, can't tail merge these.
979 // Visit each predecessor only once.
980 if (!UniquePreds.insert(PBB).second)
983 // Skip blocks which may jump to a landing pad. Can't tail merge these.
984 if (PBB->getLandingPadSuccessor())
987 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
988 SmallVector<MachineOperand, 4> Cond;
989 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
990 // Failing case: IBB is the target of a cbr, and we cannot reverse the
992 SmallVector<MachineOperand, 4> NewCond(Cond);
993 if (!Cond.empty() && TBB == IBB) {
994 if (TII->ReverseBranchCondition(NewCond))
996 // This is the QBB case described above
998 FBB = std::next(MachineFunction::iterator(PBB));
1001 // Failing case: the only way IBB can be reached from PBB is via
1002 // exception handling. Happens for landing pads. Would be nice to have
1003 // a bit in the edge so we didn't have to do all this.
1004 if (IBB->isLandingPad()) {
1005 MachineFunction::iterator IP = PBB; IP++;
1006 MachineBasicBlock *PredNextBB = nullptr;
1010 if (IBB != PredNextBB) // fallthrough
1013 if (TBB != IBB && FBB != IBB) // cbr then ubr
1015 } else if (Cond.empty()) {
1016 if (TBB != IBB) // ubr
1019 if (TBB != IBB && IBB != PredNextBB) // cbr
1024 // Remove the unconditional branch at the end, if any.
1025 if (TBB && (Cond.empty() || FBB)) {
1026 DebugLoc dl; // FIXME: this is nowhere
1027 TII->RemoveBranch(*PBB);
1029 // reinsert conditional branch only, for now
1030 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1034 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
1038 // If this is a large problem, avoid visiting the same basic blocks multiple
1040 if (MergePotentials.size() == TailMergeThreshold)
1041 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1042 TriedMerging.insert(MergePotentials[i].getBlock());
1044 if (MergePotentials.size() >= 2)
1045 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
1047 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1048 // result of removing blocks in TryTailMergeBlocks.
1049 PredBB = std::prev(I); // this may have been changed in TryTailMergeBlocks
1050 if (MergePotentials.size() == 1 &&
1051 MergePotentials.begin()->getBlock() != PredBB)
1052 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1058 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1059 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1060 BlockFrequency AccumulatedMBBFreq;
1062 // Aggregate edge frequency of successor edge j:
1063 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1064 // where bb is a basic block that is in SameTails.
1065 for (const auto &Src : SameTails) {
1066 const MachineBasicBlock *SrcMBB = Src.getBlock();
1067 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1068 AccumulatedMBBFreq += BlockFreq;
1070 // It is not necessary to recompute edge weights if TailBB has less than two
1072 if (TailMBB.succ_size() <= 1)
1075 auto EdgeFreq = EdgeFreqLs.begin();
1077 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1078 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1079 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1082 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1084 if (TailMBB.succ_size() <= 1)
1087 auto MaxEdgeFreq = *std::max_element(EdgeFreqLs.begin(), EdgeFreqLs.end());
1088 uint64_t Scale = MaxEdgeFreq.getFrequency() / UINT32_MAX + 1;
1089 auto EdgeFreq = EdgeFreqLs.begin();
1091 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1092 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1093 TailMBB.setSuccWeight(SuccI, EdgeFreq->getFrequency() / Scale);
1096 //===----------------------------------------------------------------------===//
1097 // Branch Optimization
1098 //===----------------------------------------------------------------------===//
1100 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1101 bool MadeChange = false;
1103 // Make sure blocks are numbered in order
1104 MF.RenumberBlocks();
1106 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1108 MachineBasicBlock *MBB = I++;
1109 MadeChange |= OptimizeBlock(MBB);
1111 // If it is dead, remove it.
1112 if (MBB->pred_empty()) {
1113 RemoveDeadBlock(MBB);
1121 // Blocks should be considered empty if they contain only debug info;
1122 // else the debug info would affect codegen.
1123 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1124 return MBB->getFirstNonDebugInstr() == MBB->end();
1127 // Blocks with only debug info and branches should be considered the same
1128 // as blocks with only branches.
1129 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1130 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1131 assert(I != MBB->end() && "empty block!");
1132 return I->isBranch();
1135 /// IsBetterFallthrough - Return true if it would be clearly better to
1136 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1137 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1138 /// result in infinite loops.
1139 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1140 MachineBasicBlock *MBB2) {
1141 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1142 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1143 // optimize branches that branch to either a return block or an assert block
1144 // into a fallthrough to the return.
1145 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1146 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1147 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1150 // If there is a clear successor ordering we make sure that one block
1151 // will fall through to the next
1152 if (MBB1->isSuccessor(MBB2)) return true;
1153 if (MBB2->isSuccessor(MBB1)) return false;
1155 return MBB2I->isCall() && !MBB1I->isCall();
1158 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1159 /// instructions on the block.
1160 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1161 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1162 if (I != MBB.end() && I->isBranch())
1163 return I->getDebugLoc();
1167 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1168 /// block. This is never called on the entry block.
1169 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1170 bool MadeChange = false;
1171 MachineFunction &MF = *MBB->getParent();
1174 MachineFunction::iterator FallThrough = MBB;
1177 // If this block is empty, make everyone use its fall-through, not the block
1178 // explicitly. Landing pads should not do this since the landing-pad table
1179 // points to this block. Blocks with their addresses taken shouldn't be
1181 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1182 // Dead block? Leave for cleanup later.
1183 if (MBB->pred_empty()) return MadeChange;
1185 if (FallThrough == MF.end()) {
1186 // TODO: Simplify preds to not branch here if possible!
1187 } else if (FallThrough->isLandingPad()) {
1188 // Don't rewrite to a landing pad fallthough. That could lead to the case
1189 // where a BB jumps to more than one landing pad.
1190 // TODO: Is it ever worth rewriting predecessors which don't already
1191 // jump to a landing pad, and so can safely jump to the fallthrough?
1193 // Rewrite all predecessors of the old block to go to the fallthrough
1195 while (!MBB->pred_empty()) {
1196 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1197 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1199 // If MBB was the target of a jump table, update jump tables to go to the
1200 // fallthrough instead.
1201 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1202 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1208 // Check to see if we can simplify the terminator of the block before this
1210 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1212 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1213 SmallVector<MachineOperand, 4> PriorCond;
1214 bool PriorUnAnalyzable =
1215 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1216 if (!PriorUnAnalyzable) {
1217 // If the CFG for the prior block has extra edges, remove them.
1218 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1219 !PriorCond.empty());
1221 // If the previous branch is conditional and both conditions go to the same
1222 // destination, remove the branch, replacing it with an unconditional one or
1224 if (PriorTBB && PriorTBB == PriorFBB) {
1225 DebugLoc dl = getBranchDebugLoc(PrevBB);
1226 TII->RemoveBranch(PrevBB);
1228 if (PriorTBB != MBB)
1229 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1232 goto ReoptimizeBlock;
1235 // If the previous block unconditionally falls through to this block and
1236 // this block has no other predecessors, move the contents of this block
1237 // into the prior block. This doesn't usually happen when SimplifyCFG
1238 // has been used, but it can happen if tail merging splits a fall-through
1239 // predecessor of a block.
1240 // This has to check PrevBB->succ_size() because EH edges are ignored by
1242 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1243 PrevBB.succ_size() == 1 &&
1244 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1245 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1246 << "From MBB: " << *MBB);
1247 // Remove redundant DBG_VALUEs first.
1248 if (PrevBB.begin() != PrevBB.end()) {
1249 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1251 MachineBasicBlock::iterator MBBIter = MBB->begin();
1252 // Check if DBG_VALUE at the end of PrevBB is identical to the
1253 // DBG_VALUE at the beginning of MBB.
1254 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1255 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1256 if (!MBBIter->isIdenticalTo(PrevBBIter))
1258 MachineInstr *DuplicateDbg = MBBIter;
1259 ++MBBIter; -- PrevBBIter;
1260 DuplicateDbg->eraseFromParent();
1263 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1264 PrevBB.removeSuccessor(PrevBB.succ_begin());
1265 assert(PrevBB.succ_empty());
1266 PrevBB.transferSuccessors(MBB);
1271 // If the previous branch *only* branches to *this* block (conditional or
1272 // not) remove the branch.
1273 if (PriorTBB == MBB && !PriorFBB) {
1274 TII->RemoveBranch(PrevBB);
1277 goto ReoptimizeBlock;
1280 // If the prior block branches somewhere else on the condition and here if
1281 // the condition is false, remove the uncond second branch.
1282 if (PriorFBB == MBB) {
1283 DebugLoc dl = getBranchDebugLoc(PrevBB);
1284 TII->RemoveBranch(PrevBB);
1285 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1288 goto ReoptimizeBlock;
1291 // If the prior block branches here on true and somewhere else on false, and
1292 // if the branch condition is reversible, reverse the branch to create a
1294 if (PriorTBB == MBB) {
1295 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1296 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1297 DebugLoc dl = getBranchDebugLoc(PrevBB);
1298 TII->RemoveBranch(PrevBB);
1299 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1302 goto ReoptimizeBlock;
1306 // If this block has no successors (e.g. it is a return block or ends with
1307 // a call to a no-return function like abort or __cxa_throw) and if the pred
1308 // falls through into this block, and if it would otherwise fall through
1309 // into the block after this, move this block to the end of the function.
1311 // We consider it more likely that execution will stay in the function (e.g.
1312 // due to loops) than it is to exit it. This asserts in loops etc, moving
1313 // the assert condition out of the loop body.
1314 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1315 MachineFunction::iterator(PriorTBB) == FallThrough &&
1316 !MBB->canFallThrough()) {
1317 bool DoTransform = true;
1319 // We have to be careful that the succs of PredBB aren't both no-successor
1320 // blocks. If neither have successors and if PredBB is the second from
1321 // last block in the function, we'd just keep swapping the two blocks for
1322 // last. Only do the swap if one is clearly better to fall through than
1324 if (FallThrough == --MF.end() &&
1325 !IsBetterFallthrough(PriorTBB, MBB))
1326 DoTransform = false;
1329 // Reverse the branch so we will fall through on the previous true cond.
1330 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1331 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1332 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1333 << "To make fallthrough to: " << *PriorTBB << "\n");
1335 DebugLoc dl = getBranchDebugLoc(PrevBB);
1336 TII->RemoveBranch(PrevBB);
1337 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1339 // Move this block to the end of the function.
1340 MBB->moveAfter(--MF.end());
1349 // Analyze the branch in the current block.
1350 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1351 SmallVector<MachineOperand, 4> CurCond;
1352 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1353 if (!CurUnAnalyzable) {
1354 // If the CFG for the prior block has extra edges, remove them.
1355 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1357 // If this is a two-way branch, and the FBB branches to this block, reverse
1358 // the condition so the single-basic-block loop is faster. Instead of:
1359 // Loop: xxx; jcc Out; jmp Loop
1361 // Loop: xxx; jncc Loop; jmp Out
1362 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1363 SmallVector<MachineOperand, 4> NewCond(CurCond);
1364 if (!TII->ReverseBranchCondition(NewCond)) {
1365 DebugLoc dl = getBranchDebugLoc(*MBB);
1366 TII->RemoveBranch(*MBB);
1367 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1370 goto ReoptimizeBlock;
1374 // If this branch is the only thing in its block, see if we can forward
1375 // other blocks across it.
1376 if (CurTBB && CurCond.empty() && !CurFBB &&
1377 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1378 !MBB->hasAddressTaken()) {
1379 DebugLoc dl = getBranchDebugLoc(*MBB);
1380 // This block may contain just an unconditional branch. Because there can
1381 // be 'non-branch terminators' in the block, try removing the branch and
1382 // then seeing if the block is empty.
1383 TII->RemoveBranch(*MBB);
1384 // If the only things remaining in the block are debug info, remove these
1385 // as well, so this will behave the same as an empty block in non-debug
1387 if (IsEmptyBlock(MBB)) {
1388 // Make the block empty, losing the debug info (we could probably
1389 // improve this in some cases.)
1390 MBB->erase(MBB->begin(), MBB->end());
1392 // If this block is just an unconditional branch to CurTBB, we can
1393 // usually completely eliminate the block. The only case we cannot
1394 // completely eliminate the block is when the block before this one
1395 // falls through into MBB and we can't understand the prior block's branch
1398 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1399 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1400 !PrevBB.isSuccessor(MBB)) {
1401 // If the prior block falls through into us, turn it into an
1402 // explicit branch to us to make updates simpler.
1403 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1404 PriorTBB != MBB && PriorFBB != MBB) {
1406 assert(PriorCond.empty() && !PriorFBB &&
1407 "Bad branch analysis");
1410 assert(!PriorFBB && "Machine CFG out of date!");
1413 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1414 TII->RemoveBranch(PrevBB);
1415 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1418 // Iterate through all the predecessors, revectoring each in-turn.
1420 bool DidChange = false;
1421 bool HasBranchToSelf = false;
1422 while(PI != MBB->pred_size()) {
1423 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1425 // If this block has an uncond branch to itself, leave it.
1427 HasBranchToSelf = true;
1430 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1431 // If this change resulted in PMBB ending in a conditional
1432 // branch where both conditions go to the same destination,
1433 // change this to an unconditional branch (and fix the CFG).
1434 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1435 SmallVector<MachineOperand, 4> NewCurCond;
1436 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1437 NewCurFBB, NewCurCond, true);
1438 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1439 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1440 TII->RemoveBranch(*PMBB);
1442 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1445 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1450 // Change any jumptables to go to the new MBB.
1451 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1452 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1456 if (!HasBranchToSelf) return MadeChange;
1461 // Add the branch back if the block is more than just an uncond branch.
1462 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1466 // If the prior block doesn't fall through into this block, and if this
1467 // block doesn't fall through into some other block, see if we can find a
1468 // place to move this block where a fall-through will happen.
1469 if (!PrevBB.canFallThrough()) {
1471 // Now we know that there was no fall-through into this block, check to
1472 // see if it has a fall-through into its successor.
1473 bool CurFallsThru = MBB->canFallThrough();
1475 if (!MBB->isLandingPad()) {
1476 // Check all the predecessors of this block. If one of them has no fall
1477 // throughs, move this block right after it.
1478 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1479 E = MBB->pred_end(); PI != E; ++PI) {
1480 // Analyze the branch at the end of the pred.
1481 MachineBasicBlock *PredBB = *PI;
1482 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1483 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1484 SmallVector<MachineOperand, 4> PredCond;
1485 if (PredBB != MBB && !PredBB->canFallThrough() &&
1486 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1487 && (!CurFallsThru || !CurTBB || !CurFBB)
1488 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1489 // If the current block doesn't fall through, just move it.
1490 // If the current block can fall through and does not end with a
1491 // conditional branch, we need to append an unconditional jump to
1492 // the (current) next block. To avoid a possible compile-time
1493 // infinite loop, move blocks only backward in this case.
1494 // Also, if there are already 2 branches here, we cannot add a third;
1495 // this means we have the case
1500 MachineBasicBlock *NextBB =
1501 std::next(MachineFunction::iterator(MBB));
1503 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1505 MBB->moveAfter(PredBB);
1507 goto ReoptimizeBlock;
1512 if (!CurFallsThru) {
1513 // Check all successors to see if we can move this block before it.
1514 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1515 E = MBB->succ_end(); SI != E; ++SI) {
1516 // Analyze the branch at the end of the block before the succ.
1517 MachineBasicBlock *SuccBB = *SI;
1518 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1520 // If this block doesn't already fall-through to that successor, and if
1521 // the succ doesn't already have a block that can fall through into it,
1522 // and if the successor isn't an EH destination, we can arrange for the
1523 // fallthrough to happen.
1524 if (SuccBB != MBB && &*SuccPrev != MBB &&
1525 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1526 !SuccBB->isLandingPad()) {
1527 MBB->moveBefore(SuccBB);
1529 goto ReoptimizeBlock;
1533 // Okay, there is no really great place to put this block. If, however,
1534 // the block before this one would be a fall-through if this block were
1535 // removed, move this block to the end of the function.
1536 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1537 SmallVector<MachineOperand, 4> PrevCond;
1538 if (FallThrough != MF.end() &&
1539 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1540 PrevBB.isSuccessor(FallThrough)) {
1541 MBB->moveAfter(--MF.end());
1551 //===----------------------------------------------------------------------===//
1552 // Hoist Common Code
1553 //===----------------------------------------------------------------------===//
1555 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1556 /// blocks to their common predecessor.
1557 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1558 bool MadeChange = false;
1559 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1560 MachineBasicBlock *MBB = I++;
1561 MadeChange |= HoistCommonCodeInSuccs(MBB);
1567 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1568 /// its 'true' successor.
1569 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1570 MachineBasicBlock *TrueBB) {
1571 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1572 E = BB->succ_end(); SI != E; ++SI) {
1573 MachineBasicBlock *SuccBB = *SI;
1574 if (SuccBB != TrueBB)
1580 template <class Container>
1581 static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
1583 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1584 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1591 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1592 /// in successors to. The location is usually just before the terminator,
1593 /// however if the terminator is a conditional branch and its previous
1594 /// instruction is the flag setting instruction, the previous instruction is
1595 /// the preferred location. This function also gathers uses and defs of the
1596 /// instructions from the insertion point to the end of the block. The data is
1597 /// used by HoistCommonCodeInSuccs to ensure safety.
1599 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1600 const TargetInstrInfo *TII,
1601 const TargetRegisterInfo *TRI,
1602 SmallSet<unsigned,4> &Uses,
1603 SmallSet<unsigned,4> &Defs) {
1604 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1605 if (!TII->isUnpredicatedTerminator(Loc))
1608 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1609 const MachineOperand &MO = Loc->getOperand(i);
1612 unsigned Reg = MO.getReg();
1616 addRegAndItsAliases(Reg, TRI, Uses);
1619 // Don't try to hoist code in the rare case the terminator defines a
1620 // register that is later used.
1623 // If the terminator defines a register, make sure we don't hoist
1624 // the instruction whose def might be clobbered by the terminator.
1625 addRegAndItsAliases(Reg, TRI, Defs);
1631 if (Loc == MBB->begin())
1634 // The terminator is probably a conditional branch, try not to separate the
1635 // branch from condition setting instruction.
1636 MachineBasicBlock::iterator PI = Loc;
1638 while (PI != MBB->begin() && PI->isDebugValue())
1642 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1643 const MachineOperand &MO = PI->getOperand(i);
1644 // If PI has a regmask operand, it is probably a call. Separate away.
1647 if (!MO.isReg() || MO.isUse())
1649 unsigned Reg = MO.getReg();
1652 if (Uses.count(Reg))
1656 // The condition setting instruction is not just before the conditional
1660 // Be conservative, don't insert instruction above something that may have
1661 // side-effects. And since it's potentially bad to separate flag setting
1662 // instruction from the conditional branch, just abort the optimization
1664 // Also avoid moving code above predicated instruction since it's hard to
1665 // reason about register liveness with predicated instruction.
1666 bool DontMoveAcrossStore = true;
1667 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(PI))
1671 // Find out what registers are live. Note this routine is ignoring other live
1672 // registers which are only used by instructions in successor blocks.
1673 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1674 const MachineOperand &MO = PI->getOperand(i);
1677 unsigned Reg = MO.getReg();
1681 addRegAndItsAliases(Reg, TRI, Uses);
1683 if (Uses.erase(Reg)) {
1684 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1685 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1686 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1689 addRegAndItsAliases(Reg, TRI, Defs);
1696 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1697 /// sequence at the start of the function, move the instructions before MBB
1698 /// terminator if it's legal.
1699 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1700 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1701 SmallVector<MachineOperand, 4> Cond;
1702 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1705 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1707 // Malformed bcc? True and false blocks are the same?
1710 // Restrict the optimization to cases where MBB is the only predecessor,
1711 // it is an obvious win.
1712 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1715 // Find a suitable position to hoist the common instructions to. Also figure
1716 // out which registers are used or defined by instructions from the insertion
1717 // point to the end of the block.
1718 SmallSet<unsigned, 4> Uses, Defs;
1719 MachineBasicBlock::iterator Loc =
1720 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1721 if (Loc == MBB->end())
1724 bool HasDups = false;
1725 SmallVector<unsigned, 4> LocalDefs;
1726 SmallSet<unsigned, 4> LocalDefsSet;
1727 MachineBasicBlock::iterator TIB = TBB->begin();
1728 MachineBasicBlock::iterator FIB = FBB->begin();
1729 MachineBasicBlock::iterator TIE = TBB->end();
1730 MachineBasicBlock::iterator FIE = FBB->end();
1731 while (TIB != TIE && FIB != FIE) {
1732 // Skip dbg_value instructions. These do not count.
1733 if (TIB->isDebugValue()) {
1734 while (TIB != TIE && TIB->isDebugValue())
1739 if (FIB->isDebugValue()) {
1740 while (FIB != FIE && FIB->isDebugValue())
1745 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1748 if (TII->isPredicated(TIB))
1749 // Hard to reason about register liveness with predicated instruction.
1753 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1754 MachineOperand &MO = TIB->getOperand(i);
1755 // Don't attempt to hoist instructions with register masks.
1756 if (MO.isRegMask()) {
1762 unsigned Reg = MO.getReg();
1766 if (Uses.count(Reg)) {
1767 // Avoid clobbering a register that's used by the instruction at
1768 // the point of insertion.
1773 if (Defs.count(Reg) && !MO.isDead()) {
1774 // Don't hoist the instruction if the def would be clobber by the
1775 // instruction at the point insertion. FIXME: This is overly
1776 // conservative. It should be possible to hoist the instructions
1777 // in BB2 in the following example:
1779 // r1, eflag = op1 r2, r3
1788 } else if (!LocalDefsSet.count(Reg)) {
1789 if (Defs.count(Reg)) {
1790 // Use is defined by the instruction at the point of insertion.
1795 if (MO.isKill() && Uses.count(Reg))
1796 // Kills a register that's read by the instruction at the point of
1797 // insertion. Remove the kill marker.
1798 MO.setIsKill(false);
1804 bool DontMoveAcrossStore = true;
1805 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1808 // Remove kills from LocalDefsSet, these registers had short live ranges.
1809 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1810 MachineOperand &MO = TIB->getOperand(i);
1811 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1813 unsigned Reg = MO.getReg();
1814 if (!Reg || !LocalDefsSet.count(Reg))
1816 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1817 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1818 LocalDefsSet.erase(*AI);
1820 LocalDefsSet.erase(Reg);
1824 // Track local defs so we can update liveins.
1825 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1826 MachineOperand &MO = TIB->getOperand(i);
1827 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1829 unsigned Reg = MO.getReg();
1832 LocalDefs.push_back(Reg);
1833 addRegAndItsAliases(Reg, TRI, LocalDefsSet);
1844 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1845 FBB->erase(FBB->begin(), FIB);
1848 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1849 unsigned Def = LocalDefs[i];
1850 if (LocalDefsSet.count(Def)) {
1851 TBB->addLiveIn(Def);
1852 FBB->addLiveIn(Def);