1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass forwards branches to unconditional branches to make them branch
11 // directly to the target block. This pass often results in dead MBB's, which
14 // Note that this pass must be run after register allocation, it cannot handle
17 //===----------------------------------------------------------------------===//
19 #include "BranchFolding.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
24 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
25 #include "llvm/CodeGen/MachineFunctionPass.h"
26 #include "llvm/CodeGen/MachineJumpTableInfo.h"
27 #include "llvm/CodeGen/MachineMemOperand.h"
28 #include "llvm/CodeGen/MachineModuleInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/Passes.h"
31 #include "llvm/CodeGen/RegisterScavenging.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Target/TargetInstrInfo.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "branchfolding"
45 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
46 STATISTIC(NumBranchOpts, "Number of branches optimized");
47 STATISTIC(NumTailMerge , "Number of block tails merged");
48 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
50 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
51 cl::init(cl::BOU_UNSET), cl::Hidden);
53 // Throttle for huge numbers of predecessors (compile speed problems)
54 static cl::opt<unsigned>
55 TailMergeThreshold("tail-merge-threshold",
56 cl::desc("Max number of predecessors to consider tail merging"),
57 cl::init(150), cl::Hidden);
59 // Heuristic for tail merging (and, inversely, tail duplication).
60 // TODO: This should be replaced with a target query.
61 static cl::opt<unsigned>
62 TailMergeSize("tail-merge-size",
63 cl::desc("Min number of instructions to consider tail merging"),
64 cl::init(3), cl::Hidden);
67 /// BranchFolderPass - Wrap branch folder in a machine function pass.
68 class BranchFolderPass : public MachineFunctionPass {
71 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
73 bool runOnMachineFunction(MachineFunction &MF) override;
75 void getAnalysisUsage(AnalysisUsage &AU) const override {
76 AU.addRequired<MachineBlockFrequencyInfo>();
77 AU.addRequired<MachineBranchProbabilityInfo>();
78 AU.addRequired<TargetPassConfig>();
79 MachineFunctionPass::getAnalysisUsage(AU);
84 char BranchFolderPass::ID = 0;
85 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
87 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
88 "Control Flow Optimizer", false, false)
90 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
91 if (skipOptnoneFunction(*MF.getFunction()))
94 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
95 // TailMerge can create jump into if branches that make CFG irreducible for
96 // HW that requires structurized CFG.
97 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
98 PassConfig->getEnableTailMerge();
99 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true,
100 getAnalysis<MachineBlockFrequencyInfo>(),
101 getAnalysis<MachineBranchProbabilityInfo>());
102 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
103 MF.getSubtarget().getRegisterInfo(),
104 getAnalysisIfAvailable<MachineModuleInfo>());
107 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
108 const MachineBlockFrequencyInfo &FreqInfo,
109 const MachineBranchProbabilityInfo &ProbInfo)
110 : EnableHoistCommonCode(CommonHoist), MBBFreqInfo(FreqInfo),
112 switch (FlagEnableTailMerge) {
113 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
114 case cl::BOU_TRUE: EnableTailMerge = true; break;
115 case cl::BOU_FALSE: EnableTailMerge = false; break;
119 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
120 /// function, updating the CFG.
121 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
122 assert(MBB->pred_empty() && "MBB must be dead!");
123 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
125 MachineFunction *MF = MBB->getParent();
126 // drop all successors.
127 while (!MBB->succ_empty())
128 MBB->removeSuccessor(MBB->succ_end()-1);
130 // Avoid matching if this pointer gets reused.
131 TriedMerging.erase(MBB);
137 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
138 /// followed by terminators, and if the implicitly defined registers are not
139 /// used by the terminators, remove those implicit_def's. e.g.
141 /// r0 = implicit_def
142 /// r1 = implicit_def
144 /// This block can be optimized away later if the implicit instructions are
146 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
147 SmallSet<unsigned, 4> ImpDefRegs;
148 MachineBasicBlock::iterator I = MBB->begin();
149 while (I != MBB->end()) {
150 if (!I->isImplicitDef())
152 unsigned Reg = I->getOperand(0).getReg();
153 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
154 SubRegs.isValid(); ++SubRegs)
155 ImpDefRegs.insert(*SubRegs);
158 if (ImpDefRegs.empty())
161 MachineBasicBlock::iterator FirstTerm = I;
162 while (I != MBB->end()) {
163 if (!TII->isUnpredicatedTerminator(I))
165 // See if it uses any of the implicitly defined registers.
166 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
167 MachineOperand &MO = I->getOperand(i);
168 if (!MO.isReg() || !MO.isUse())
170 unsigned Reg = MO.getReg();
171 if (ImpDefRegs.count(Reg))
178 while (I != FirstTerm) {
179 MachineInstr *ImpDefMI = &*I;
181 MBB->erase(ImpDefMI);
187 /// OptimizeFunction - Perhaps branch folding, tail merging and other
188 /// CFG optimizations on the given function.
189 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
190 const TargetInstrInfo *tii,
191 const TargetRegisterInfo *tri,
192 MachineModuleInfo *mmi) {
193 if (!tii) return false;
195 TriedMerging.clear();
202 // Use a RegScavenger to help update liveness when required.
203 MachineRegisterInfo &MRI = MF.getRegInfo();
204 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
205 RS = new RegScavenger();
207 MRI.invalidateLiveness();
209 // Fix CFG. The later algorithms expect it to be right.
210 bool MadeChange = false;
211 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
212 MachineBasicBlock *MBB = I, *TBB = nullptr, *FBB = nullptr;
213 SmallVector<MachineOperand, 4> Cond;
214 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
215 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
216 MadeChange |= OptimizeImpDefsBlock(MBB);
219 bool MadeChangeThisIteration = true;
220 while (MadeChangeThisIteration) {
221 MadeChangeThisIteration = TailMergeBlocks(MF);
222 MadeChangeThisIteration |= OptimizeBranches(MF);
223 if (EnableHoistCommonCode)
224 MadeChangeThisIteration |= HoistCommonCode(MF);
225 MadeChange |= MadeChangeThisIteration;
228 // See if any jump tables have become dead as the code generator
230 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
236 // Walk the function to find jump tables that are live.
237 BitVector JTIsLive(JTI->getJumpTables().size());
238 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
240 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
242 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
243 MachineOperand &Op = I->getOperand(op);
244 if (!Op.isJTI()) continue;
246 // Remember that this JT is live.
247 JTIsLive.set(Op.getIndex());
251 // Finally, remove dead jump tables. This happens when the
252 // indirect jump was unreachable (and thus deleted).
253 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
254 if (!JTIsLive.test(i)) {
255 JTI->RemoveJumpTable(i);
263 //===----------------------------------------------------------------------===//
264 // Tail Merging of Blocks
265 //===----------------------------------------------------------------------===//
267 /// HashMachineInstr - Compute a hash value for MI and its operands.
268 static unsigned HashMachineInstr(const MachineInstr *MI) {
269 unsigned Hash = MI->getOpcode();
270 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
271 const MachineOperand &Op = MI->getOperand(i);
273 // Merge in bits from the operand if easy. We can't use MachineOperand's
274 // hash_code here because it's not deterministic and we sort by hash value
276 unsigned OperandHash = 0;
277 switch (Op.getType()) {
278 case MachineOperand::MO_Register:
279 OperandHash = Op.getReg();
281 case MachineOperand::MO_Immediate:
282 OperandHash = Op.getImm();
284 case MachineOperand::MO_MachineBasicBlock:
285 OperandHash = Op.getMBB()->getNumber();
287 case MachineOperand::MO_FrameIndex:
288 case MachineOperand::MO_ConstantPoolIndex:
289 case MachineOperand::MO_JumpTableIndex:
290 OperandHash = Op.getIndex();
292 case MachineOperand::MO_GlobalAddress:
293 case MachineOperand::MO_ExternalSymbol:
294 // Global address / external symbol are too hard, don't bother, but do
295 // pull in the offset.
296 OperandHash = Op.getOffset();
302 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
307 /// HashEndOfMBB - Hash the last instruction in the MBB.
308 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
309 MachineBasicBlock::const_iterator I = MBB->getLastNonDebugInstr();
313 return HashMachineInstr(I);
316 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
317 /// of instructions they actually have in common together at their end. Return
318 /// iterators for the first shared instruction in each block.
319 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
320 MachineBasicBlock *MBB2,
321 MachineBasicBlock::iterator &I1,
322 MachineBasicBlock::iterator &I2) {
326 unsigned TailLen = 0;
327 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
329 // Skip debugging pseudos; necessary to avoid changing the code.
330 while (I1->isDebugValue()) {
331 if (I1==MBB1->begin()) {
332 while (I2->isDebugValue()) {
333 if (I2==MBB2->begin())
334 // I1==DBG at begin; I2==DBG at begin
339 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
344 // I1==first (untested) non-DBG preceding known match
345 while (I2->isDebugValue()) {
346 if (I2==MBB2->begin()) {
348 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
353 // I1, I2==first (untested) non-DBGs preceding known match
354 if (!I1->isIdenticalTo(I2) ||
355 // FIXME: This check is dubious. It's used to get around a problem where
356 // people incorrectly expect inline asm directives to remain in the same
357 // relative order. This is untenable because normal compiler
358 // optimizations (like this one) may reorder and/or merge these
366 // Back past possible debugging pseudos at beginning of block. This matters
367 // when one block differs from the other only by whether debugging pseudos
368 // are present at the beginning. (This way, the various checks later for
369 // I1==MBB1->begin() work as expected.)
370 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
372 while (I2->isDebugValue()) {
373 if (I2 == MBB2->begin())
379 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
381 while (I1->isDebugValue()) {
382 if (I1 == MBB1->begin())
391 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
392 MachineBasicBlock *NewMBB) {
394 RS->enterBasicBlock(CurMBB);
395 if (!CurMBB->empty())
396 RS->forward(std::prev(CurMBB->end()));
397 for (unsigned int i = 1, e = TRI->getNumRegs(); i != e; i++)
398 if (RS->isRegUsed(i, false))
399 NewMBB->addLiveIn(i);
403 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
404 /// after it, replacing it with an unconditional branch to NewDest.
405 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
406 MachineBasicBlock *NewDest) {
407 MachineBasicBlock *CurMBB = OldInst->getParent();
409 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
411 // For targets that use the register scavenger, we must maintain LiveIns.
412 MaintainLiveIns(CurMBB, NewDest);
417 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
418 /// MBB so that the part before the iterator falls into the part starting at the
419 /// iterator. This returns the new MBB.
420 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
421 MachineBasicBlock::iterator BBI1,
422 const BasicBlock *BB) {
423 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
426 MachineFunction &MF = *CurMBB.getParent();
428 // Create the fall-through block.
429 MachineFunction::iterator MBBI = &CurMBB;
430 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
431 CurMBB.getParent()->insert(++MBBI, NewMBB);
433 // Move all the successors of this block to the specified block.
434 NewMBB->transferSuccessors(&CurMBB);
436 // Add an edge from CurMBB to NewMBB for the fall-through.
437 CurMBB.addSuccessor(NewMBB);
439 // Splice the code over.
440 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
442 // NewMBB inherits CurMBB's block frequency.
443 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
445 // For targets that use the register scavenger, we must maintain LiveIns.
446 MaintainLiveIns(&CurMBB, NewMBB);
451 /// EstimateRuntime - Make a rough estimate for how long it will take to run
452 /// the specified code.
453 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
454 MachineBasicBlock::iterator E) {
456 for (; I != E; ++I) {
457 if (I->isDebugValue())
461 else if (I->mayLoad() || I->mayStore())
469 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
470 // branches temporarily for tail merging). In the case where CurMBB ends
471 // with a conditional branch to the next block, optimize by reversing the
472 // test and conditionally branching to SuccMBB instead.
473 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
474 const TargetInstrInfo *TII) {
475 MachineFunction *MF = CurMBB->getParent();
476 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
477 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
478 SmallVector<MachineOperand, 4> Cond;
479 DebugLoc dl; // FIXME: this is nowhere
480 if (I != MF->end() &&
481 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
482 MachineBasicBlock *NextBB = I;
483 if (TBB == NextBB && !Cond.empty() && !FBB) {
484 if (!TII->ReverseBranchCondition(Cond)) {
485 TII->RemoveBranch(*CurMBB);
486 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
491 TII->InsertBranch(*CurMBB, SuccBB, nullptr,
492 SmallVector<MachineOperand, 0>(), dl);
496 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
497 if (getHash() < o.getHash())
499 if (getHash() > o.getHash())
501 if (getBlock()->getNumber() < o.getBlock()->getNumber())
503 if (getBlock()->getNumber() > o.getBlock()->getNumber())
505 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
506 // an object with itself.
507 #ifndef _GLIBCXX_DEBUG
508 llvm_unreachable("Predecessor appears twice");
515 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
516 auto I = MergedBBFreq.find(MBB);
518 if (I != MergedBBFreq.end())
521 return MBFI.getBlockFreq(MBB);
524 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
526 MergedBBFreq[MBB] = F;
529 /// CountTerminators - Count the number of terminators in the given
530 /// block and set I to the position of the first non-terminator, if there
531 /// is one, or MBB->end() otherwise.
532 static unsigned CountTerminators(MachineBasicBlock *MBB,
533 MachineBasicBlock::iterator &I) {
535 unsigned NumTerms = 0;
537 if (I == MBB->begin()) {
542 if (!I->isTerminator()) break;
548 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
549 /// and decide if it would be profitable to merge those tails. Return the
550 /// length of the common tail and iterators to the first common instruction
552 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
553 MachineBasicBlock *MBB2,
554 unsigned minCommonTailLength,
555 unsigned &CommonTailLen,
556 MachineBasicBlock::iterator &I1,
557 MachineBasicBlock::iterator &I2,
558 MachineBasicBlock *SuccBB,
559 MachineBasicBlock *PredBB) {
560 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
561 if (CommonTailLen == 0)
563 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
564 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
567 // It's almost always profitable to merge any number of non-terminator
568 // instructions with the block that falls through into the common successor.
569 if (MBB1 == PredBB || MBB2 == PredBB) {
570 MachineBasicBlock::iterator I;
571 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
572 if (CommonTailLen > NumTerms)
576 // If one of the blocks can be completely merged and happens to be in
577 // a position where the other could fall through into it, merge any number
578 // of instructions, because it can be done without a branch.
579 // TODO: If the blocks are not adjacent, move one of them so that they are?
580 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
582 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
585 // If both blocks have an unconditional branch temporarily stripped out,
586 // count that as an additional common instruction for the following
588 unsigned EffectiveTailLen = CommonTailLen;
589 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
590 !MBB1->back().isBarrier() &&
591 !MBB2->back().isBarrier())
594 // Check if the common tail is long enough to be worthwhile.
595 if (EffectiveTailLen >= minCommonTailLength)
598 // If we are optimizing for code size, 2 instructions in common is enough if
599 // we don't have to split a block. At worst we will be introducing 1 new
600 // branch instruction, which is likely to be smaller than the 2
601 // instructions that would be deleted in the merge.
602 MachineFunction *MF = MBB1->getParent();
603 if (EffectiveTailLen >= 2 &&
604 MF->getFunction()->hasFnAttribute(Attribute::OptimizeForSize) &&
605 (I1 == MBB1->begin() || I2 == MBB2->begin()))
611 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
612 /// hash CurHash (guaranteed to match the last element). Build the vector
613 /// SameTails of all those that have the (same) largest number of instructions
614 /// in common of any pair of these blocks. SameTails entries contain an
615 /// iterator into MergePotentials (from which the MachineBasicBlock can be
616 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
617 /// instruction where the matching code sequence begins.
618 /// Order of elements in SameTails is the reverse of the order in which
619 /// those blocks appear in MergePotentials (where they are not necessarily
621 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
622 unsigned minCommonTailLength,
623 MachineBasicBlock *SuccBB,
624 MachineBasicBlock *PredBB) {
625 unsigned maxCommonTailLength = 0U;
627 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
628 MPIterator HighestMPIter = std::prev(MergePotentials.end());
629 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
630 B = MergePotentials.begin();
631 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
632 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
633 unsigned CommonTailLen;
634 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
636 CommonTailLen, TrialBBI1, TrialBBI2,
638 if (CommonTailLen > maxCommonTailLength) {
640 maxCommonTailLength = CommonTailLen;
641 HighestMPIter = CurMPIter;
642 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
644 if (HighestMPIter == CurMPIter &&
645 CommonTailLen == maxCommonTailLength)
646 SameTails.push_back(SameTailElt(I, TrialBBI2));
652 return maxCommonTailLength;
655 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
656 /// MergePotentials, restoring branches at ends of blocks as appropriate.
657 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
658 MachineBasicBlock *SuccBB,
659 MachineBasicBlock *PredBB) {
660 MPIterator CurMPIter, B;
661 for (CurMPIter = std::prev(MergePotentials.end()),
662 B = MergePotentials.begin();
663 CurMPIter->getHash() == CurHash; --CurMPIter) {
664 // Put the unconditional branch back, if we need one.
665 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
666 if (SuccBB && CurMBB != PredBB)
667 FixTail(CurMBB, SuccBB, TII);
671 if (CurMPIter->getHash() != CurHash)
673 MergePotentials.erase(CurMPIter, MergePotentials.end());
676 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
677 /// only of the common tail. Create a block that does by splitting one.
678 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
679 MachineBasicBlock *SuccBB,
680 unsigned maxCommonTailLength,
681 unsigned &commonTailIndex) {
683 unsigned TimeEstimate = ~0U;
684 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
685 // Use PredBB if possible; that doesn't require a new branch.
686 if (SameTails[i].getBlock() == PredBB) {
690 // Otherwise, make a (fairly bogus) choice based on estimate of
691 // how long it will take the various blocks to execute.
692 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
693 SameTails[i].getTailStartPos());
694 if (t <= TimeEstimate) {
700 MachineBasicBlock::iterator BBI =
701 SameTails[commonTailIndex].getTailStartPos();
702 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
704 // If the common tail includes any debug info we will take it pretty
705 // randomly from one of the inputs. Might be better to remove it?
706 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
707 << maxCommonTailLength);
709 // If the split block unconditionally falls-thru to SuccBB, it will be
710 // merged. In control flow terms it should then take SuccBB's name. e.g. If
711 // SuccBB is an inner loop, the common tail is still part of the inner loop.
712 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
713 SuccBB->getBasicBlock() : MBB->getBasicBlock();
714 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
716 DEBUG(dbgs() << "... failed!");
720 SameTails[commonTailIndex].setBlock(newMBB);
721 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
723 // If we split PredBB, newMBB is the new predecessor.
730 static bool hasIdenticalMMOs(const MachineInstr *MI1, const MachineInstr *MI2) {
731 auto I1 = MI1->memoperands_begin(), E1 = MI1->memoperands_end();
732 auto I2 = MI2->memoperands_begin(), E2 = MI2->memoperands_end();
733 if ((E1 - I1) != (E2 - I2))
735 for (; I1 != E1; ++I1, ++I2) {
743 removeMMOsFromMemoryOperations(MachineBasicBlock::iterator MBBIStartPos,
744 MachineBasicBlock &MBBCommon) {
745 // Remove MMOs from memory operations in the common block
746 // when they do not match the ones from the block being tail-merged.
747 // This ensures later passes conservatively compute dependencies.
748 MachineBasicBlock *MBB = MBBIStartPos->getParent();
749 // Note CommonTailLen does not necessarily matches the size of
750 // the common BB nor all its instructions because of debug
751 // instructions differences.
752 unsigned CommonTailLen = 0;
753 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
756 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
757 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
758 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
759 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
761 while (CommonTailLen--) {
762 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
765 if (MBBI->isDebugValue()) {
770 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
773 assert(MBBICommon != MBBIECommon &&
774 "Reached BB end within common tail length!");
775 assert(MBBICommon->isIdenticalTo(&*MBBI) && "Expected matching MIIs!");
777 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
778 if (!hasIdenticalMMOs(&*MBBI, &*MBBICommon))
779 MBBICommon->clearMemRefs();
786 // See if any of the blocks in MergePotentials (which all have a common single
787 // successor, or all have no successor) can be tail-merged. If there is a
788 // successor, any blocks in MergePotentials that are not tail-merged and
789 // are not immediately before Succ must have an unconditional branch to
790 // Succ added (but the predecessor/successor lists need no adjustment).
791 // The lone predecessor of Succ that falls through into Succ,
792 // if any, is given in PredBB.
794 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
795 MachineBasicBlock *PredBB) {
796 bool MadeChange = false;
798 // Except for the special cases below, tail-merge if there are at least
799 // this many instructions in common.
800 unsigned minCommonTailLength = TailMergeSize;
802 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
803 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
804 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
805 << (i == e-1 ? "" : ", ");
808 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
810 dbgs() << " which has fall-through from BB#"
811 << PredBB->getNumber() << "\n";
813 dbgs() << "Looking for common tails of at least "
814 << minCommonTailLength << " instruction"
815 << (minCommonTailLength == 1 ? "" : "s") << '\n';
818 // Sort by hash value so that blocks with identical end sequences sort
820 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
822 // Walk through equivalence sets looking for actual exact matches.
823 while (MergePotentials.size() > 1) {
824 unsigned CurHash = MergePotentials.back().getHash();
826 // Build SameTails, identifying the set of blocks with this hash code
827 // and with the maximum number of instructions in common.
828 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
832 // If we didn't find any pair that has at least minCommonTailLength
833 // instructions in common, remove all blocks with this hash code and retry.
834 if (SameTails.empty()) {
835 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
839 // If one of the blocks is the entire common tail (and not the entry
840 // block, which we can't jump to), we can treat all blocks with this same
841 // tail at once. Use PredBB if that is one of the possibilities, as that
842 // will not introduce any extra branches.
843 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
844 getParent()->begin();
845 unsigned commonTailIndex = SameTails.size();
846 // If there are two blocks, check to see if one can be made to fall through
848 if (SameTails.size() == 2 &&
849 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
850 SameTails[1].tailIsWholeBlock())
852 else if (SameTails.size() == 2 &&
853 SameTails[1].getBlock()->isLayoutSuccessor(
854 SameTails[0].getBlock()) &&
855 SameTails[0].tailIsWholeBlock())
858 // Otherwise just pick one, favoring the fall-through predecessor if
860 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
861 MachineBasicBlock *MBB = SameTails[i].getBlock();
862 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
868 if (SameTails[i].tailIsWholeBlock())
873 if (commonTailIndex == SameTails.size() ||
874 (SameTails[commonTailIndex].getBlock() == PredBB &&
875 !SameTails[commonTailIndex].tailIsWholeBlock())) {
876 // None of the blocks consist entirely of the common tail.
877 // Split a block so that one does.
878 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
879 maxCommonTailLength, commonTailIndex)) {
880 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
885 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
887 // Recompute commont tail MBB's edge weights and block frequency.
888 setCommonTailEdgeWeights(*MBB);
890 // MBB is common tail. Adjust all other BB's to jump to this one.
891 // Traversal must be forwards so erases work.
892 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
894 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
895 if (commonTailIndex == i)
897 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
898 << (i == e-1 ? "" : ", "));
899 // Remove MMOs from memory operations as needed.
900 removeMMOsFromMemoryOperations(SameTails[i].getTailStartPos(), *MBB);
901 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
902 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
903 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
904 MergePotentials.erase(SameTails[i].getMPIter());
906 DEBUG(dbgs() << "\n");
907 // We leave commonTailIndex in the worklist in case there are other blocks
908 // that match it with a smaller number of instructions.
914 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
915 bool MadeChange = false;
916 if (!EnableTailMerge) return MadeChange;
918 // First find blocks with no successors.
919 MergePotentials.clear();
920 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
921 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
922 if (TriedMerging.count(I))
925 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
928 // If this is a large problem, avoid visiting the same basic blocks
930 if (MergePotentials.size() == TailMergeThreshold)
931 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
932 TriedMerging.insert(MergePotentials[i].getBlock());
934 // See if we can do any tail merging on those.
935 if (MergePotentials.size() >= 2)
936 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
938 // Look at blocks (IBB) with multiple predecessors (PBB).
939 // We change each predecessor to a canonical form, by
940 // (1) temporarily removing any unconditional branch from the predecessor
942 // (2) alter conditional branches so they branch to the other block
943 // not IBB; this may require adding back an unconditional branch to IBB
944 // later, where there wasn't one coming in. E.g.
946 // fallthrough to QBB
949 // with a conceptual B to IBB after that, which never actually exists.
950 // With those changes, we see whether the predecessors' tails match,
951 // and merge them if so. We change things out of canonical form and
952 // back to the way they were later in the process. (OptimizeBranches
953 // would undo some of this, but we can't use it, because we'd get into
954 // a compile-time infinite loop repeatedly doing and undoing the same
957 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
959 if (I->pred_size() < 2) continue;
960 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
961 MachineBasicBlock *IBB = I;
962 MachineBasicBlock *PredBB = std::prev(I);
963 MergePotentials.clear();
964 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
966 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
967 MachineBasicBlock *PBB = *P;
968 if (TriedMerging.count(PBB))
971 // Skip blocks that loop to themselves, can't tail merge these.
975 // Visit each predecessor only once.
976 if (!UniquePreds.insert(PBB).second)
979 // Skip blocks which may jump to a landing pad. Can't tail merge these.
980 if (PBB->getLandingPadSuccessor())
983 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
984 SmallVector<MachineOperand, 4> Cond;
985 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
986 // Failing case: IBB is the target of a cbr, and we cannot reverse the
988 SmallVector<MachineOperand, 4> NewCond(Cond);
989 if (!Cond.empty() && TBB == IBB) {
990 if (TII->ReverseBranchCondition(NewCond))
992 // This is the QBB case described above
994 FBB = std::next(MachineFunction::iterator(PBB));
997 // Failing case: the only way IBB can be reached from PBB is via
998 // exception handling. Happens for landing pads. Would be nice to have
999 // a bit in the edge so we didn't have to do all this.
1000 if (IBB->isLandingPad()) {
1001 MachineFunction::iterator IP = PBB; IP++;
1002 MachineBasicBlock *PredNextBB = nullptr;
1006 if (IBB != PredNextBB) // fallthrough
1009 if (TBB != IBB && FBB != IBB) // cbr then ubr
1011 } else if (Cond.empty()) {
1012 if (TBB != IBB) // ubr
1015 if (TBB != IBB && IBB != PredNextBB) // cbr
1020 // Remove the unconditional branch at the end, if any.
1021 if (TBB && (Cond.empty() || FBB)) {
1022 DebugLoc dl; // FIXME: this is nowhere
1023 TII->RemoveBranch(*PBB);
1025 // reinsert conditional branch only, for now
1026 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1030 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
1034 // If this is a large problem, avoid visiting the same basic blocks multiple
1036 if (MergePotentials.size() == TailMergeThreshold)
1037 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1038 TriedMerging.insert(MergePotentials[i].getBlock());
1040 if (MergePotentials.size() >= 2)
1041 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
1043 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1044 // result of removing blocks in TryTailMergeBlocks.
1045 PredBB = std::prev(I); // this may have been changed in TryTailMergeBlocks
1046 if (MergePotentials.size() == 1 &&
1047 MergePotentials.begin()->getBlock() != PredBB)
1048 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1054 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1055 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1056 BlockFrequency AccumulatedMBBFreq;
1058 // Aggregate edge frequency of successor edge j:
1059 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1060 // where bb is a basic block that is in SameTails.
1061 for (const auto &Src : SameTails) {
1062 const MachineBasicBlock *SrcMBB = Src.getBlock();
1063 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1064 AccumulatedMBBFreq += BlockFreq;
1066 // It is not necessary to recompute edge weights if TailBB has less than two
1068 if (TailMBB.succ_size() <= 1)
1071 auto EdgeFreq = EdgeFreqLs.begin();
1073 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1074 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1075 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1078 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1080 if (TailMBB.succ_size() <= 1)
1083 auto MaxEdgeFreq = *std::max_element(EdgeFreqLs.begin(), EdgeFreqLs.end());
1084 uint64_t Scale = MaxEdgeFreq.getFrequency() / UINT32_MAX + 1;
1085 auto EdgeFreq = EdgeFreqLs.begin();
1087 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1088 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1089 TailMBB.setSuccWeight(SuccI, EdgeFreq->getFrequency() / Scale);
1092 //===----------------------------------------------------------------------===//
1093 // Branch Optimization
1094 //===----------------------------------------------------------------------===//
1096 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1097 bool MadeChange = false;
1099 // Make sure blocks are numbered in order
1100 MF.RenumberBlocks();
1102 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1104 MachineBasicBlock *MBB = I++;
1105 MadeChange |= OptimizeBlock(MBB);
1107 // If it is dead, remove it.
1108 if (MBB->pred_empty()) {
1109 RemoveDeadBlock(MBB);
1117 // Blocks should be considered empty if they contain only debug info;
1118 // else the debug info would affect codegen.
1119 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1120 return MBB->getFirstNonDebugInstr() == MBB->end();
1123 // Blocks with only debug info and branches should be considered the same
1124 // as blocks with only branches.
1125 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1126 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1127 assert(I != MBB->end() && "empty block!");
1128 return I->isBranch();
1131 /// IsBetterFallthrough - Return true if it would be clearly better to
1132 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1133 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1134 /// result in infinite loops.
1135 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1136 MachineBasicBlock *MBB2) {
1137 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1138 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1139 // optimize branches that branch to either a return block or an assert block
1140 // into a fallthrough to the return.
1141 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1142 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1143 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1146 // If there is a clear successor ordering we make sure that one block
1147 // will fall through to the next
1148 if (MBB1->isSuccessor(MBB2)) return true;
1149 if (MBB2->isSuccessor(MBB1)) return false;
1151 return MBB2I->isCall() && !MBB1I->isCall();
1154 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1155 /// instructions on the block.
1156 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1157 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1158 if (I != MBB.end() && I->isBranch())
1159 return I->getDebugLoc();
1163 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1164 /// block. This is never called on the entry block.
1165 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1166 bool MadeChange = false;
1167 MachineFunction &MF = *MBB->getParent();
1170 MachineFunction::iterator FallThrough = MBB;
1173 // If this block is empty, make everyone use its fall-through, not the block
1174 // explicitly. Landing pads should not do this since the landing-pad table
1175 // points to this block. Blocks with their addresses taken shouldn't be
1177 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1178 // Dead block? Leave for cleanup later.
1179 if (MBB->pred_empty()) return MadeChange;
1181 if (FallThrough == MF.end()) {
1182 // TODO: Simplify preds to not branch here if possible!
1183 } else if (FallThrough->isLandingPad()) {
1184 // Don't rewrite to a landing pad fallthough. That could lead to the case
1185 // where a BB jumps to more than one landing pad.
1186 // TODO: Is it ever worth rewriting predecessors which don't already
1187 // jump to a landing pad, and so can safely jump to the fallthrough?
1189 // Rewrite all predecessors of the old block to go to the fallthrough
1191 while (!MBB->pred_empty()) {
1192 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1193 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1195 // If MBB was the target of a jump table, update jump tables to go to the
1196 // fallthrough instead.
1197 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1198 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1204 // Check to see if we can simplify the terminator of the block before this
1206 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1208 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1209 SmallVector<MachineOperand, 4> PriorCond;
1210 bool PriorUnAnalyzable =
1211 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1212 if (!PriorUnAnalyzable) {
1213 // If the CFG for the prior block has extra edges, remove them.
1214 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1215 !PriorCond.empty());
1217 // If the previous branch is conditional and both conditions go to the same
1218 // destination, remove the branch, replacing it with an unconditional one or
1220 if (PriorTBB && PriorTBB == PriorFBB) {
1221 DebugLoc dl = getBranchDebugLoc(PrevBB);
1222 TII->RemoveBranch(PrevBB);
1224 if (PriorTBB != MBB)
1225 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1228 goto ReoptimizeBlock;
1231 // If the previous block unconditionally falls through to this block and
1232 // this block has no other predecessors, move the contents of this block
1233 // into the prior block. This doesn't usually happen when SimplifyCFG
1234 // has been used, but it can happen if tail merging splits a fall-through
1235 // predecessor of a block.
1236 // This has to check PrevBB->succ_size() because EH edges are ignored by
1238 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1239 PrevBB.succ_size() == 1 &&
1240 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1241 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1242 << "From MBB: " << *MBB);
1243 // Remove redundant DBG_VALUEs first.
1244 if (PrevBB.begin() != PrevBB.end()) {
1245 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1247 MachineBasicBlock::iterator MBBIter = MBB->begin();
1248 // Check if DBG_VALUE at the end of PrevBB is identical to the
1249 // DBG_VALUE at the beginning of MBB.
1250 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1251 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1252 if (!MBBIter->isIdenticalTo(PrevBBIter))
1254 MachineInstr *DuplicateDbg = MBBIter;
1255 ++MBBIter; -- PrevBBIter;
1256 DuplicateDbg->eraseFromParent();
1259 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1260 PrevBB.removeSuccessor(PrevBB.succ_begin());
1261 assert(PrevBB.succ_empty());
1262 PrevBB.transferSuccessors(MBB);
1267 // If the previous branch *only* branches to *this* block (conditional or
1268 // not) remove the branch.
1269 if (PriorTBB == MBB && !PriorFBB) {
1270 TII->RemoveBranch(PrevBB);
1273 goto ReoptimizeBlock;
1276 // If the prior block branches somewhere else on the condition and here if
1277 // the condition is false, remove the uncond second branch.
1278 if (PriorFBB == MBB) {
1279 DebugLoc dl = getBranchDebugLoc(PrevBB);
1280 TII->RemoveBranch(PrevBB);
1281 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1284 goto ReoptimizeBlock;
1287 // If the prior block branches here on true and somewhere else on false, and
1288 // if the branch condition is reversible, reverse the branch to create a
1290 if (PriorTBB == MBB) {
1291 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1292 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1293 DebugLoc dl = getBranchDebugLoc(PrevBB);
1294 TII->RemoveBranch(PrevBB);
1295 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1298 goto ReoptimizeBlock;
1302 // If this block has no successors (e.g. it is a return block or ends with
1303 // a call to a no-return function like abort or __cxa_throw) and if the pred
1304 // falls through into this block, and if it would otherwise fall through
1305 // into the block after this, move this block to the end of the function.
1307 // We consider it more likely that execution will stay in the function (e.g.
1308 // due to loops) than it is to exit it. This asserts in loops etc, moving
1309 // the assert condition out of the loop body.
1310 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1311 MachineFunction::iterator(PriorTBB) == FallThrough &&
1312 !MBB->canFallThrough()) {
1313 bool DoTransform = true;
1315 // We have to be careful that the succs of PredBB aren't both no-successor
1316 // blocks. If neither have successors and if PredBB is the second from
1317 // last block in the function, we'd just keep swapping the two blocks for
1318 // last. Only do the swap if one is clearly better to fall through than
1320 if (FallThrough == --MF.end() &&
1321 !IsBetterFallthrough(PriorTBB, MBB))
1322 DoTransform = false;
1325 // Reverse the branch so we will fall through on the previous true cond.
1326 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1327 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1328 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1329 << "To make fallthrough to: " << *PriorTBB << "\n");
1331 DebugLoc dl = getBranchDebugLoc(PrevBB);
1332 TII->RemoveBranch(PrevBB);
1333 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1335 // Move this block to the end of the function.
1336 MBB->moveAfter(--MF.end());
1345 // Analyze the branch in the current block.
1346 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1347 SmallVector<MachineOperand, 4> CurCond;
1348 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1349 if (!CurUnAnalyzable) {
1350 // If the CFG for the prior block has extra edges, remove them.
1351 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1353 // If this is a two-way branch, and the FBB branches to this block, reverse
1354 // the condition so the single-basic-block loop is faster. Instead of:
1355 // Loop: xxx; jcc Out; jmp Loop
1357 // Loop: xxx; jncc Loop; jmp Out
1358 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1359 SmallVector<MachineOperand, 4> NewCond(CurCond);
1360 if (!TII->ReverseBranchCondition(NewCond)) {
1361 DebugLoc dl = getBranchDebugLoc(*MBB);
1362 TII->RemoveBranch(*MBB);
1363 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1366 goto ReoptimizeBlock;
1370 // If this branch is the only thing in its block, see if we can forward
1371 // other blocks across it.
1372 if (CurTBB && CurCond.empty() && !CurFBB &&
1373 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1374 !MBB->hasAddressTaken()) {
1375 DebugLoc dl = getBranchDebugLoc(*MBB);
1376 // This block may contain just an unconditional branch. Because there can
1377 // be 'non-branch terminators' in the block, try removing the branch and
1378 // then seeing if the block is empty.
1379 TII->RemoveBranch(*MBB);
1380 // If the only things remaining in the block are debug info, remove these
1381 // as well, so this will behave the same as an empty block in non-debug
1383 if (IsEmptyBlock(MBB)) {
1384 // Make the block empty, losing the debug info (we could probably
1385 // improve this in some cases.)
1386 MBB->erase(MBB->begin(), MBB->end());
1388 // If this block is just an unconditional branch to CurTBB, we can
1389 // usually completely eliminate the block. The only case we cannot
1390 // completely eliminate the block is when the block before this one
1391 // falls through into MBB and we can't understand the prior block's branch
1394 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1395 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1396 !PrevBB.isSuccessor(MBB)) {
1397 // If the prior block falls through into us, turn it into an
1398 // explicit branch to us to make updates simpler.
1399 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1400 PriorTBB != MBB && PriorFBB != MBB) {
1402 assert(PriorCond.empty() && !PriorFBB &&
1403 "Bad branch analysis");
1406 assert(!PriorFBB && "Machine CFG out of date!");
1409 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1410 TII->RemoveBranch(PrevBB);
1411 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1414 // Iterate through all the predecessors, revectoring each in-turn.
1416 bool DidChange = false;
1417 bool HasBranchToSelf = false;
1418 while(PI != MBB->pred_size()) {
1419 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1421 // If this block has an uncond branch to itself, leave it.
1423 HasBranchToSelf = true;
1426 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1427 // If this change resulted in PMBB ending in a conditional
1428 // branch where both conditions go to the same destination,
1429 // change this to an unconditional branch (and fix the CFG).
1430 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1431 SmallVector<MachineOperand, 4> NewCurCond;
1432 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1433 NewCurFBB, NewCurCond, true);
1434 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1435 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1436 TII->RemoveBranch(*PMBB);
1438 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1441 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1446 // Change any jumptables to go to the new MBB.
1447 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1448 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1452 if (!HasBranchToSelf) return MadeChange;
1457 // Add the branch back if the block is more than just an uncond branch.
1458 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1462 // If the prior block doesn't fall through into this block, and if this
1463 // block doesn't fall through into some other block, see if we can find a
1464 // place to move this block where a fall-through will happen.
1465 if (!PrevBB.canFallThrough()) {
1467 // Now we know that there was no fall-through into this block, check to
1468 // see if it has a fall-through into its successor.
1469 bool CurFallsThru = MBB->canFallThrough();
1471 if (!MBB->isLandingPad()) {
1472 // Check all the predecessors of this block. If one of them has no fall
1473 // throughs, move this block right after it.
1474 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1475 E = MBB->pred_end(); PI != E; ++PI) {
1476 // Analyze the branch at the end of the pred.
1477 MachineBasicBlock *PredBB = *PI;
1478 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1479 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1480 SmallVector<MachineOperand, 4> PredCond;
1481 if (PredBB != MBB && !PredBB->canFallThrough() &&
1482 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1483 && (!CurFallsThru || !CurTBB || !CurFBB)
1484 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1485 // If the current block doesn't fall through, just move it.
1486 // If the current block can fall through and does not end with a
1487 // conditional branch, we need to append an unconditional jump to
1488 // the (current) next block. To avoid a possible compile-time
1489 // infinite loop, move blocks only backward in this case.
1490 // Also, if there are already 2 branches here, we cannot add a third;
1491 // this means we have the case
1496 MachineBasicBlock *NextBB =
1497 std::next(MachineFunction::iterator(MBB));
1499 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1501 MBB->moveAfter(PredBB);
1503 goto ReoptimizeBlock;
1508 if (!CurFallsThru) {
1509 // Check all successors to see if we can move this block before it.
1510 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1511 E = MBB->succ_end(); SI != E; ++SI) {
1512 // Analyze the branch at the end of the block before the succ.
1513 MachineBasicBlock *SuccBB = *SI;
1514 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1516 // If this block doesn't already fall-through to that successor, and if
1517 // the succ doesn't already have a block that can fall through into it,
1518 // and if the successor isn't an EH destination, we can arrange for the
1519 // fallthrough to happen.
1520 if (SuccBB != MBB && &*SuccPrev != MBB &&
1521 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1522 !SuccBB->isLandingPad()) {
1523 MBB->moveBefore(SuccBB);
1525 goto ReoptimizeBlock;
1529 // Okay, there is no really great place to put this block. If, however,
1530 // the block before this one would be a fall-through if this block were
1531 // removed, move this block to the end of the function.
1532 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1533 SmallVector<MachineOperand, 4> PrevCond;
1534 if (FallThrough != MF.end() &&
1535 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1536 PrevBB.isSuccessor(FallThrough)) {
1537 MBB->moveAfter(--MF.end());
1547 //===----------------------------------------------------------------------===//
1548 // Hoist Common Code
1549 //===----------------------------------------------------------------------===//
1551 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1552 /// blocks to their common predecessor.
1553 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1554 bool MadeChange = false;
1555 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1556 MachineBasicBlock *MBB = I++;
1557 MadeChange |= HoistCommonCodeInSuccs(MBB);
1563 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1564 /// its 'true' successor.
1565 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1566 MachineBasicBlock *TrueBB) {
1567 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1568 E = BB->succ_end(); SI != E; ++SI) {
1569 MachineBasicBlock *SuccBB = *SI;
1570 if (SuccBB != TrueBB)
1576 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1577 /// in successors to. The location is usually just before the terminator,
1578 /// however if the terminator is a conditional branch and its previous
1579 /// instruction is the flag setting instruction, the previous instruction is
1580 /// the preferred location. This function also gathers uses and defs of the
1581 /// instructions from the insertion point to the end of the block. The data is
1582 /// used by HoistCommonCodeInSuccs to ensure safety.
1584 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1585 const TargetInstrInfo *TII,
1586 const TargetRegisterInfo *TRI,
1587 SmallSet<unsigned,4> &Uses,
1588 SmallSet<unsigned,4> &Defs) {
1589 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1590 if (!TII->isUnpredicatedTerminator(Loc))
1593 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1594 const MachineOperand &MO = Loc->getOperand(i);
1597 unsigned Reg = MO.getReg();
1601 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1605 // Don't try to hoist code in the rare case the terminator defines a
1606 // register that is later used.
1609 // If the terminator defines a register, make sure we don't hoist
1610 // the instruction whose def might be clobbered by the terminator.
1611 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1618 if (Loc == MBB->begin())
1621 // The terminator is probably a conditional branch, try not to separate the
1622 // branch from condition setting instruction.
1623 MachineBasicBlock::iterator PI = Loc;
1625 while (PI != MBB->begin() && PI->isDebugValue())
1629 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1630 const MachineOperand &MO = PI->getOperand(i);
1631 // If PI has a regmask operand, it is probably a call. Separate away.
1634 if (!MO.isReg() || MO.isUse())
1636 unsigned Reg = MO.getReg();
1639 if (Uses.count(Reg))
1643 // The condition setting instruction is not just before the conditional
1647 // Be conservative, don't insert instruction above something that may have
1648 // side-effects. And since it's potentially bad to separate flag setting
1649 // instruction from the conditional branch, just abort the optimization
1651 // Also avoid moving code above predicated instruction since it's hard to
1652 // reason about register liveness with predicated instruction.
1653 bool DontMoveAcrossStore = true;
1654 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(PI))
1658 // Find out what registers are live. Note this routine is ignoring other live
1659 // registers which are only used by instructions in successor blocks.
1660 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1661 const MachineOperand &MO = PI->getOperand(i);
1664 unsigned Reg = MO.getReg();
1668 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1671 if (Uses.erase(Reg)) {
1672 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1673 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1675 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1683 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1684 /// sequence at the start of the function, move the instructions before MBB
1685 /// terminator if it's legal.
1686 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1687 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1688 SmallVector<MachineOperand, 4> Cond;
1689 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1692 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1694 // Malformed bcc? True and false blocks are the same?
1697 // Restrict the optimization to cases where MBB is the only predecessor,
1698 // it is an obvious win.
1699 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1702 // Find a suitable position to hoist the common instructions to. Also figure
1703 // out which registers are used or defined by instructions from the insertion
1704 // point to the end of the block.
1705 SmallSet<unsigned, 4> Uses, Defs;
1706 MachineBasicBlock::iterator Loc =
1707 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1708 if (Loc == MBB->end())
1711 bool HasDups = false;
1712 SmallVector<unsigned, 4> LocalDefs;
1713 SmallSet<unsigned, 4> LocalDefsSet;
1714 MachineBasicBlock::iterator TIB = TBB->begin();
1715 MachineBasicBlock::iterator FIB = FBB->begin();
1716 MachineBasicBlock::iterator TIE = TBB->end();
1717 MachineBasicBlock::iterator FIE = FBB->end();
1718 while (TIB != TIE && FIB != FIE) {
1719 // Skip dbg_value instructions. These do not count.
1720 if (TIB->isDebugValue()) {
1721 while (TIB != TIE && TIB->isDebugValue())
1726 if (FIB->isDebugValue()) {
1727 while (FIB != FIE && FIB->isDebugValue())
1732 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1735 if (TII->isPredicated(TIB))
1736 // Hard to reason about register liveness with predicated instruction.
1740 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1741 MachineOperand &MO = TIB->getOperand(i);
1742 // Don't attempt to hoist instructions with register masks.
1743 if (MO.isRegMask()) {
1749 unsigned Reg = MO.getReg();
1753 if (Uses.count(Reg)) {
1754 // Avoid clobbering a register that's used by the instruction at
1755 // the point of insertion.
1760 if (Defs.count(Reg) && !MO.isDead()) {
1761 // Don't hoist the instruction if the def would be clobber by the
1762 // instruction at the point insertion. FIXME: This is overly
1763 // conservative. It should be possible to hoist the instructions
1764 // in BB2 in the following example:
1766 // r1, eflag = op1 r2, r3
1775 } else if (!LocalDefsSet.count(Reg)) {
1776 if (Defs.count(Reg)) {
1777 // Use is defined by the instruction at the point of insertion.
1782 if (MO.isKill() && Uses.count(Reg))
1783 // Kills a register that's read by the instruction at the point of
1784 // insertion. Remove the kill marker.
1785 MO.setIsKill(false);
1791 bool DontMoveAcrossStore = true;
1792 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1795 // Remove kills from LocalDefsSet, these registers had short live ranges.
1796 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1797 MachineOperand &MO = TIB->getOperand(i);
1798 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1800 unsigned Reg = MO.getReg();
1801 if (!Reg || !LocalDefsSet.count(Reg))
1803 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1804 LocalDefsSet.erase(*AI);
1807 // Track local defs so we can update liveins.
1808 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1809 MachineOperand &MO = TIB->getOperand(i);
1810 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1812 unsigned Reg = MO.getReg();
1815 LocalDefs.push_back(Reg);
1816 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1817 LocalDefsSet.insert(*AI);
1828 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1829 FBB->erase(FBB->begin(), FIB);
1832 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1833 unsigned Def = LocalDefs[i];
1834 if (LocalDefsSet.count(Def)) {
1835 TBB->addLiveIn(Def);
1836 FBB->addLiveIn(Def);