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 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
155 SubRegs.isValid(); ++SubRegs)
156 ImpDefRegs.insert(*SubRegs);
159 if (ImpDefRegs.empty())
162 MachineBasicBlock::iterator FirstTerm = I;
163 while (I != MBB->end()) {
164 if (!TII->isUnpredicatedTerminator(I))
166 // See if it uses any of the implicitly defined registers.
167 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
168 MachineOperand &MO = I->getOperand(i);
169 if (!MO.isReg() || !MO.isUse())
171 unsigned Reg = MO.getReg();
172 if (ImpDefRegs.count(Reg))
179 while (I != FirstTerm) {
180 MachineInstr *ImpDefMI = &*I;
182 MBB->erase(ImpDefMI);
188 /// OptimizeFunction - Perhaps branch folding, tail merging and other
189 /// CFG optimizations on the given function.
190 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
191 const TargetInstrInfo *tii,
192 const TargetRegisterInfo *tri,
193 MachineModuleInfo *mmi) {
194 if (!tii) return false;
196 TriedMerging.clear();
203 // Use a RegScavenger to help update liveness when required.
204 MachineRegisterInfo &MRI = MF.getRegInfo();
205 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
206 RS = new RegScavenger();
208 MRI.invalidateLiveness();
210 // Fix CFG. The later algorithms expect it to be right.
211 bool MadeChange = false;
212 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
213 MachineBasicBlock *MBB = I, *TBB = nullptr, *FBB = nullptr;
214 SmallVector<MachineOperand, 4> Cond;
215 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
216 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
217 MadeChange |= OptimizeImpDefsBlock(MBB);
220 bool MadeChangeThisIteration = true;
221 while (MadeChangeThisIteration) {
222 MadeChangeThisIteration = TailMergeBlocks(MF);
223 MadeChangeThisIteration |= OptimizeBranches(MF);
224 if (EnableHoistCommonCode)
225 MadeChangeThisIteration |= HoistCommonCode(MF);
226 MadeChange |= MadeChangeThisIteration;
229 // See if any jump tables have become dead as the code generator
231 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
237 // Walk the function to find jump tables that are live.
238 BitVector JTIsLive(JTI->getJumpTables().size());
239 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
241 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
243 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
244 MachineOperand &Op = I->getOperand(op);
245 if (!Op.isJTI()) continue;
247 // Remember that this JT is live.
248 JTIsLive.set(Op.getIndex());
252 // Finally, remove dead jump tables. This happens when the
253 // indirect jump was unreachable (and thus deleted).
254 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
255 if (!JTIsLive.test(i)) {
256 JTI->RemoveJumpTable(i);
264 //===----------------------------------------------------------------------===//
265 // Tail Merging of Blocks
266 //===----------------------------------------------------------------------===//
268 /// HashMachineInstr - Compute a hash value for MI and its operands.
269 static unsigned HashMachineInstr(const MachineInstr *MI) {
270 unsigned Hash = MI->getOpcode();
271 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
272 const MachineOperand &Op = MI->getOperand(i);
274 // Merge in bits from the operand if easy. We can't use MachineOperand's
275 // hash_code here because it's not deterministic and we sort by hash value
277 unsigned OperandHash = 0;
278 switch (Op.getType()) {
279 case MachineOperand::MO_Register:
280 OperandHash = Op.getReg();
282 case MachineOperand::MO_Immediate:
283 OperandHash = Op.getImm();
285 case MachineOperand::MO_MachineBasicBlock:
286 OperandHash = Op.getMBB()->getNumber();
288 case MachineOperand::MO_FrameIndex:
289 case MachineOperand::MO_ConstantPoolIndex:
290 case MachineOperand::MO_JumpTableIndex:
291 OperandHash = Op.getIndex();
293 case MachineOperand::MO_GlobalAddress:
294 case MachineOperand::MO_ExternalSymbol:
295 // Global address / external symbol are too hard, don't bother, but do
296 // pull in the offset.
297 OperandHash = Op.getOffset();
303 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
308 /// HashEndOfMBB - Hash the last instruction in the MBB.
309 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
310 MachineBasicBlock::const_iterator I = MBB->getLastNonDebugInstr();
314 return HashMachineInstr(I);
317 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
318 /// of instructions they actually have in common together at their end. Return
319 /// iterators for the first shared instruction in each block.
320 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
321 MachineBasicBlock *MBB2,
322 MachineBasicBlock::iterator &I1,
323 MachineBasicBlock::iterator &I2) {
327 unsigned TailLen = 0;
328 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
330 // Skip debugging pseudos; necessary to avoid changing the code.
331 while (I1->isDebugValue()) {
332 if (I1==MBB1->begin()) {
333 while (I2->isDebugValue()) {
334 if (I2==MBB2->begin())
335 // I1==DBG at begin; I2==DBG at begin
340 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
345 // I1==first (untested) non-DBG preceding known match
346 while (I2->isDebugValue()) {
347 if (I2==MBB2->begin()) {
349 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
354 // I1, I2==first (untested) non-DBGs preceding known match
355 if (!I1->isIdenticalTo(I2) ||
356 // FIXME: This check is dubious. It's used to get around a problem where
357 // people incorrectly expect inline asm directives to remain in the same
358 // relative order. This is untenable because normal compiler
359 // optimizations (like this one) may reorder and/or merge these
367 // Back past possible debugging pseudos at beginning of block. This matters
368 // when one block differs from the other only by whether debugging pseudos
369 // are present at the beginning. (This way, the various checks later for
370 // I1==MBB1->begin() work as expected.)
371 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
373 while (I2->isDebugValue()) {
374 if (I2 == MBB2->begin())
380 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
382 while (I1->isDebugValue()) {
383 if (I1 == MBB1->begin())
392 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
393 MachineBasicBlock *NewMBB) {
395 RS->enterBasicBlock(CurMBB);
396 if (!CurMBB->empty())
397 RS->forward(std::prev(CurMBB->end()));
398 for (unsigned int i = 1, e = TRI->getNumRegs(); i != e; i++)
399 if (RS->isRegUsed(i, false))
400 NewMBB->addLiveIn(i);
404 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
405 /// after it, replacing it with an unconditional branch to NewDest.
406 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
407 MachineBasicBlock *NewDest) {
408 MachineBasicBlock *CurMBB = OldInst->getParent();
410 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
412 // For targets that use the register scavenger, we must maintain LiveIns.
413 MaintainLiveIns(CurMBB, NewDest);
418 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
419 /// MBB so that the part before the iterator falls into the part starting at the
420 /// iterator. This returns the new MBB.
421 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
422 MachineBasicBlock::iterator BBI1,
423 const BasicBlock *BB) {
424 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
427 MachineFunction &MF = *CurMBB.getParent();
429 // Create the fall-through block.
430 MachineFunction::iterator MBBI = &CurMBB;
431 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
432 CurMBB.getParent()->insert(++MBBI, NewMBB);
434 // Move all the successors of this block to the specified block.
435 NewMBB->transferSuccessors(&CurMBB);
437 // Add an edge from CurMBB to NewMBB for the fall-through.
438 CurMBB.addSuccessor(NewMBB);
440 // Splice the code over.
441 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
443 // NewMBB inherits CurMBB's block frequency.
444 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
446 // For targets that use the register scavenger, we must maintain LiveIns.
447 MaintainLiveIns(&CurMBB, NewMBB);
452 /// EstimateRuntime - Make a rough estimate for how long it will take to run
453 /// the specified code.
454 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
455 MachineBasicBlock::iterator E) {
457 for (; I != E; ++I) {
458 if (I->isDebugValue())
462 else if (I->mayLoad() || I->mayStore())
470 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
471 // branches temporarily for tail merging). In the case where CurMBB ends
472 // with a conditional branch to the next block, optimize by reversing the
473 // test and conditionally branching to SuccMBB instead.
474 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
475 const TargetInstrInfo *TII) {
476 MachineFunction *MF = CurMBB->getParent();
477 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
478 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
479 SmallVector<MachineOperand, 4> Cond;
480 DebugLoc dl; // FIXME: this is nowhere
481 if (I != MF->end() &&
482 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
483 MachineBasicBlock *NextBB = I;
484 if (TBB == NextBB && !Cond.empty() && !FBB) {
485 if (!TII->ReverseBranchCondition(Cond)) {
486 TII->RemoveBranch(*CurMBB);
487 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
492 TII->InsertBranch(*CurMBB, SuccBB, nullptr,
493 SmallVector<MachineOperand, 0>(), dl);
497 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
498 if (getHash() < o.getHash())
500 if (getHash() > o.getHash())
502 if (getBlock()->getNumber() < o.getBlock()->getNumber())
504 if (getBlock()->getNumber() > o.getBlock()->getNumber())
506 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
507 // an object with itself.
508 #ifndef _GLIBCXX_DEBUG
509 llvm_unreachable("Predecessor appears twice");
516 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
517 auto I = MergedBBFreq.find(MBB);
519 if (I != MergedBBFreq.end())
522 return MBFI.getBlockFreq(MBB);
525 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
527 MergedBBFreq[MBB] = F;
530 /// CountTerminators - Count the number of terminators in the given
531 /// block and set I to the position of the first non-terminator, if there
532 /// is one, or MBB->end() otherwise.
533 static unsigned CountTerminators(MachineBasicBlock *MBB,
534 MachineBasicBlock::iterator &I) {
536 unsigned NumTerms = 0;
538 if (I == MBB->begin()) {
543 if (!I->isTerminator()) break;
549 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
550 /// and decide if it would be profitable to merge those tails. Return the
551 /// length of the common tail and iterators to the first common instruction
553 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
554 MachineBasicBlock *MBB2,
555 unsigned minCommonTailLength,
556 unsigned &CommonTailLen,
557 MachineBasicBlock::iterator &I1,
558 MachineBasicBlock::iterator &I2,
559 MachineBasicBlock *SuccBB,
560 MachineBasicBlock *PredBB) {
561 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
562 if (CommonTailLen == 0)
564 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
565 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
568 // It's almost always profitable to merge any number of non-terminator
569 // instructions with the block that falls through into the common successor.
570 if (MBB1 == PredBB || MBB2 == PredBB) {
571 MachineBasicBlock::iterator I;
572 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
573 if (CommonTailLen > NumTerms)
577 // If one of the blocks can be completely merged and happens to be in
578 // a position where the other could fall through into it, merge any number
579 // of instructions, because it can be done without a branch.
580 // TODO: If the blocks are not adjacent, move one of them so that they are?
581 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
583 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
586 // If both blocks have an unconditional branch temporarily stripped out,
587 // count that as an additional common instruction for the following
589 unsigned EffectiveTailLen = CommonTailLen;
590 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
591 !MBB1->back().isBarrier() &&
592 !MBB2->back().isBarrier())
595 // Check if the common tail is long enough to be worthwhile.
596 if (EffectiveTailLen >= minCommonTailLength)
599 // If we are optimizing for code size, 2 instructions in common is enough if
600 // we don't have to split a block. At worst we will be introducing 1 new
601 // branch instruction, which is likely to be smaller than the 2
602 // instructions that would be deleted in the merge.
603 MachineFunction *MF = MBB1->getParent();
604 if (EffectiveTailLen >= 2 &&
605 MF->getFunction()->hasFnAttribute(Attribute::OptimizeForSize) &&
606 (I1 == MBB1->begin() || I2 == MBB2->begin()))
612 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
613 /// hash CurHash (guaranteed to match the last element). Build the vector
614 /// SameTails of all those that have the (same) largest number of instructions
615 /// in common of any pair of these blocks. SameTails entries contain an
616 /// iterator into MergePotentials (from which the MachineBasicBlock can be
617 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
618 /// instruction where the matching code sequence begins.
619 /// Order of elements in SameTails is the reverse of the order in which
620 /// those blocks appear in MergePotentials (where they are not necessarily
622 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
623 unsigned minCommonTailLength,
624 MachineBasicBlock *SuccBB,
625 MachineBasicBlock *PredBB) {
626 unsigned maxCommonTailLength = 0U;
628 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
629 MPIterator HighestMPIter = std::prev(MergePotentials.end());
630 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
631 B = MergePotentials.begin();
632 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
633 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
634 unsigned CommonTailLen;
635 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
637 CommonTailLen, TrialBBI1, TrialBBI2,
639 if (CommonTailLen > maxCommonTailLength) {
641 maxCommonTailLength = CommonTailLen;
642 HighestMPIter = CurMPIter;
643 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
645 if (HighestMPIter == CurMPIter &&
646 CommonTailLen == maxCommonTailLength)
647 SameTails.push_back(SameTailElt(I, TrialBBI2));
653 return maxCommonTailLength;
656 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
657 /// MergePotentials, restoring branches at ends of blocks as appropriate.
658 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
659 MachineBasicBlock *SuccBB,
660 MachineBasicBlock *PredBB) {
661 MPIterator CurMPIter, B;
662 for (CurMPIter = std::prev(MergePotentials.end()),
663 B = MergePotentials.begin();
664 CurMPIter->getHash() == CurHash; --CurMPIter) {
665 // Put the unconditional branch back, if we need one.
666 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
667 if (SuccBB && CurMBB != PredBB)
668 FixTail(CurMBB, SuccBB, TII);
672 if (CurMPIter->getHash() != CurHash)
674 MergePotentials.erase(CurMPIter, MergePotentials.end());
677 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
678 /// only of the common tail. Create a block that does by splitting one.
679 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
680 MachineBasicBlock *SuccBB,
681 unsigned maxCommonTailLength,
682 unsigned &commonTailIndex) {
684 unsigned TimeEstimate = ~0U;
685 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
686 // Use PredBB if possible; that doesn't require a new branch.
687 if (SameTails[i].getBlock() == PredBB) {
691 // Otherwise, make a (fairly bogus) choice based on estimate of
692 // how long it will take the various blocks to execute.
693 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
694 SameTails[i].getTailStartPos());
695 if (t <= TimeEstimate) {
701 MachineBasicBlock::iterator BBI =
702 SameTails[commonTailIndex].getTailStartPos();
703 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
705 // If the common tail includes any debug info we will take it pretty
706 // randomly from one of the inputs. Might be better to remove it?
707 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
708 << maxCommonTailLength);
710 // If the split block unconditionally falls-thru to SuccBB, it will be
711 // merged. In control flow terms it should then take SuccBB's name. e.g. If
712 // SuccBB is an inner loop, the common tail is still part of the inner loop.
713 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
714 SuccBB->getBasicBlock() : MBB->getBasicBlock();
715 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
717 DEBUG(dbgs() << "... failed!");
721 SameTails[commonTailIndex].setBlock(newMBB);
722 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
724 // If we split PredBB, newMBB is the new predecessor.
731 static bool hasIdenticalMMOs(const MachineInstr *MI1, const MachineInstr *MI2) {
732 auto I1 = MI1->memoperands_begin(), E1 = MI1->memoperands_end();
733 auto I2 = MI2->memoperands_begin(), E2 = MI2->memoperands_end();
734 if ((E1 - I1) != (E2 - I2))
736 for (; I1 != E1; ++I1, ++I2) {
744 removeMMOsFromMemoryOperations(MachineBasicBlock::iterator MBBIStartPos,
745 MachineBasicBlock &MBBCommon) {
746 // Remove MMOs from memory operations in the common block
747 // when they do not match the ones from the block being tail-merged.
748 // This ensures later passes conservatively compute dependencies.
749 MachineBasicBlock *MBB = MBBIStartPos->getParent();
750 // Note CommonTailLen does not necessarily matches the size of
751 // the common BB nor all its instructions because of debug
752 // instructions differences.
753 unsigned CommonTailLen = 0;
754 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
757 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
758 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
759 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
760 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
762 while (CommonTailLen--) {
763 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
766 if (MBBI->isDebugValue()) {
771 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
774 assert(MBBICommon != MBBIECommon &&
775 "Reached BB end within common tail length!");
776 assert(MBBICommon->isIdenticalTo(&*MBBI) && "Expected matching MIIs!");
778 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
779 if (!hasIdenticalMMOs(&*MBBI, &*MBBICommon))
780 MBBICommon->clearMemRefs();
787 // See if any of the blocks in MergePotentials (which all have a common single
788 // successor, or all have no successor) can be tail-merged. If there is a
789 // successor, any blocks in MergePotentials that are not tail-merged and
790 // are not immediately before Succ must have an unconditional branch to
791 // Succ added (but the predecessor/successor lists need no adjustment).
792 // The lone predecessor of Succ that falls through into Succ,
793 // if any, is given in PredBB.
795 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
796 MachineBasicBlock *PredBB) {
797 bool MadeChange = false;
799 // Except for the special cases below, tail-merge if there are at least
800 // this many instructions in common.
801 unsigned minCommonTailLength = TailMergeSize;
803 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
804 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
805 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
806 << (i == e-1 ? "" : ", ");
809 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
811 dbgs() << " which has fall-through from BB#"
812 << PredBB->getNumber() << "\n";
814 dbgs() << "Looking for common tails of at least "
815 << minCommonTailLength << " instruction"
816 << (minCommonTailLength == 1 ? "" : "s") << '\n';
819 // Sort by hash value so that blocks with identical end sequences sort
821 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
823 // Walk through equivalence sets looking for actual exact matches.
824 while (MergePotentials.size() > 1) {
825 unsigned CurHash = MergePotentials.back().getHash();
827 // Build SameTails, identifying the set of blocks with this hash code
828 // and with the maximum number of instructions in common.
829 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
833 // If we didn't find any pair that has at least minCommonTailLength
834 // instructions in common, remove all blocks with this hash code and retry.
835 if (SameTails.empty()) {
836 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
840 // If one of the blocks is the entire common tail (and not the entry
841 // block, which we can't jump to), we can treat all blocks with this same
842 // tail at once. Use PredBB if that is one of the possibilities, as that
843 // will not introduce any extra branches.
844 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
845 getParent()->begin();
846 unsigned commonTailIndex = SameTails.size();
847 // If there are two blocks, check to see if one can be made to fall through
849 if (SameTails.size() == 2 &&
850 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
851 SameTails[1].tailIsWholeBlock())
853 else if (SameTails.size() == 2 &&
854 SameTails[1].getBlock()->isLayoutSuccessor(
855 SameTails[0].getBlock()) &&
856 SameTails[0].tailIsWholeBlock())
859 // Otherwise just pick one, favoring the fall-through predecessor if
861 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
862 MachineBasicBlock *MBB = SameTails[i].getBlock();
863 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
869 if (SameTails[i].tailIsWholeBlock())
874 if (commonTailIndex == SameTails.size() ||
875 (SameTails[commonTailIndex].getBlock() == PredBB &&
876 !SameTails[commonTailIndex].tailIsWholeBlock())) {
877 // None of the blocks consist entirely of the common tail.
878 // Split a block so that one does.
879 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
880 maxCommonTailLength, commonTailIndex)) {
881 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
886 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
888 // Recompute commont tail MBB's edge weights and block frequency.
889 setCommonTailEdgeWeights(*MBB);
891 // MBB is common tail. Adjust all other BB's to jump to this one.
892 // Traversal must be forwards so erases work.
893 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
895 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
896 if (commonTailIndex == i)
898 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
899 << (i == e-1 ? "" : ", "));
900 // Remove MMOs from memory operations as needed.
901 removeMMOsFromMemoryOperations(SameTails[i].getTailStartPos(), *MBB);
902 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
903 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
904 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
905 MergePotentials.erase(SameTails[i].getMPIter());
907 DEBUG(dbgs() << "\n");
908 // We leave commonTailIndex in the worklist in case there are other blocks
909 // that match it with a smaller number of instructions.
915 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
916 bool MadeChange = false;
917 if (!EnableTailMerge) return MadeChange;
919 // First find blocks with no successors.
920 MergePotentials.clear();
921 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
922 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
923 if (TriedMerging.count(I))
926 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
929 // If this is a large problem, avoid visiting the same basic blocks
931 if (MergePotentials.size() == TailMergeThreshold)
932 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
933 TriedMerging.insert(MergePotentials[i].getBlock());
935 // See if we can do any tail merging on those.
936 if (MergePotentials.size() >= 2)
937 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
939 // Look at blocks (IBB) with multiple predecessors (PBB).
940 // We change each predecessor to a canonical form, by
941 // (1) temporarily removing any unconditional branch from the predecessor
943 // (2) alter conditional branches so they branch to the other block
944 // not IBB; this may require adding back an unconditional branch to IBB
945 // later, where there wasn't one coming in. E.g.
947 // fallthrough to QBB
950 // with a conceptual B to IBB after that, which never actually exists.
951 // With those changes, we see whether the predecessors' tails match,
952 // and merge them if so. We change things out of canonical form and
953 // back to the way they were later in the process. (OptimizeBranches
954 // would undo some of this, but we can't use it, because we'd get into
955 // a compile-time infinite loop repeatedly doing and undoing the same
958 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
960 if (I->pred_size() < 2) continue;
961 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
962 MachineBasicBlock *IBB = I;
963 MachineBasicBlock *PredBB = std::prev(I);
964 MergePotentials.clear();
965 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
967 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
968 MachineBasicBlock *PBB = *P;
969 if (TriedMerging.count(PBB))
972 // Skip blocks that loop to themselves, can't tail merge these.
976 // Visit each predecessor only once.
977 if (!UniquePreds.insert(PBB).second)
980 // Skip blocks which may jump to a landing pad. Can't tail merge these.
981 if (PBB->getLandingPadSuccessor())
984 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
985 SmallVector<MachineOperand, 4> Cond;
986 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
987 // Failing case: IBB is the target of a cbr, and we cannot reverse the
989 SmallVector<MachineOperand, 4> NewCond(Cond);
990 if (!Cond.empty() && TBB == IBB) {
991 if (TII->ReverseBranchCondition(NewCond))
993 // This is the QBB case described above
995 FBB = std::next(MachineFunction::iterator(PBB));
998 // Failing case: the only way IBB can be reached from PBB is via
999 // exception handling. Happens for landing pads. Would be nice to have
1000 // a bit in the edge so we didn't have to do all this.
1001 if (IBB->isLandingPad()) {
1002 MachineFunction::iterator IP = PBB; IP++;
1003 MachineBasicBlock *PredNextBB = nullptr;
1007 if (IBB != PredNextBB) // fallthrough
1010 if (TBB != IBB && FBB != IBB) // cbr then ubr
1012 } else if (Cond.empty()) {
1013 if (TBB != IBB) // ubr
1016 if (TBB != IBB && IBB != PredNextBB) // cbr
1021 // Remove the unconditional branch at the end, if any.
1022 if (TBB && (Cond.empty() || FBB)) {
1023 DebugLoc dl; // FIXME: this is nowhere
1024 TII->RemoveBranch(*PBB);
1026 // reinsert conditional branch only, for now
1027 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1031 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
1035 // If this is a large problem, avoid visiting the same basic blocks multiple
1037 if (MergePotentials.size() == TailMergeThreshold)
1038 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1039 TriedMerging.insert(MergePotentials[i].getBlock());
1041 if (MergePotentials.size() >= 2)
1042 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
1044 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1045 // result of removing blocks in TryTailMergeBlocks.
1046 PredBB = std::prev(I); // this may have been changed in TryTailMergeBlocks
1047 if (MergePotentials.size() == 1 &&
1048 MergePotentials.begin()->getBlock() != PredBB)
1049 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1055 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1056 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1057 BlockFrequency AccumulatedMBBFreq;
1059 // Aggregate edge frequency of successor edge j:
1060 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1061 // where bb is a basic block that is in SameTails.
1062 for (const auto &Src : SameTails) {
1063 const MachineBasicBlock *SrcMBB = Src.getBlock();
1064 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1065 AccumulatedMBBFreq += BlockFreq;
1067 // It is not necessary to recompute edge weights if TailBB has less than two
1069 if (TailMBB.succ_size() <= 1)
1072 auto EdgeFreq = EdgeFreqLs.begin();
1074 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1075 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1076 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1079 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1081 if (TailMBB.succ_size() <= 1)
1084 auto MaxEdgeFreq = *std::max_element(EdgeFreqLs.begin(), EdgeFreqLs.end());
1085 uint64_t Scale = MaxEdgeFreq.getFrequency() / UINT32_MAX + 1;
1086 auto EdgeFreq = EdgeFreqLs.begin();
1088 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1089 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1090 TailMBB.setSuccWeight(SuccI, EdgeFreq->getFrequency() / Scale);
1093 //===----------------------------------------------------------------------===//
1094 // Branch Optimization
1095 //===----------------------------------------------------------------------===//
1097 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1098 bool MadeChange = false;
1100 // Make sure blocks are numbered in order
1101 MF.RenumberBlocks();
1103 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1105 MachineBasicBlock *MBB = I++;
1106 MadeChange |= OptimizeBlock(MBB);
1108 // If it is dead, remove it.
1109 if (MBB->pred_empty()) {
1110 RemoveDeadBlock(MBB);
1118 // Blocks should be considered empty if they contain only debug info;
1119 // else the debug info would affect codegen.
1120 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1121 return MBB->getFirstNonDebugInstr() == MBB->end();
1124 // Blocks with only debug info and branches should be considered the same
1125 // as blocks with only branches.
1126 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1127 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1128 assert(I != MBB->end() && "empty block!");
1129 return I->isBranch();
1132 /// IsBetterFallthrough - Return true if it would be clearly better to
1133 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1134 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1135 /// result in infinite loops.
1136 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1137 MachineBasicBlock *MBB2) {
1138 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1139 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1140 // optimize branches that branch to either a return block or an assert block
1141 // into a fallthrough to the return.
1142 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1143 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1144 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1147 // If there is a clear successor ordering we make sure that one block
1148 // will fall through to the next
1149 if (MBB1->isSuccessor(MBB2)) return true;
1150 if (MBB2->isSuccessor(MBB1)) return false;
1152 return MBB2I->isCall() && !MBB1I->isCall();
1155 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1156 /// instructions on the block.
1157 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1158 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1159 if (I != MBB.end() && I->isBranch())
1160 return I->getDebugLoc();
1164 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1165 /// block. This is never called on the entry block.
1166 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1167 bool MadeChange = false;
1168 MachineFunction &MF = *MBB->getParent();
1171 MachineFunction::iterator FallThrough = MBB;
1174 // If this block is empty, make everyone use its fall-through, not the block
1175 // explicitly. Landing pads should not do this since the landing-pad table
1176 // points to this block. Blocks with their addresses taken shouldn't be
1178 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1179 // Dead block? Leave for cleanup later.
1180 if (MBB->pred_empty()) return MadeChange;
1182 if (FallThrough == MF.end()) {
1183 // TODO: Simplify preds to not branch here if possible!
1184 } else if (FallThrough->isLandingPad()) {
1185 // Don't rewrite to a landing pad fallthough. That could lead to the case
1186 // where a BB jumps to more than one landing pad.
1187 // TODO: Is it ever worth rewriting predecessors which don't already
1188 // jump to a landing pad, and so can safely jump to the fallthrough?
1190 // Rewrite all predecessors of the old block to go to the fallthrough
1192 while (!MBB->pred_empty()) {
1193 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1194 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1196 // If MBB was the target of a jump table, update jump tables to go to the
1197 // fallthrough instead.
1198 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1199 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1205 // Check to see if we can simplify the terminator of the block before this
1207 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1209 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1210 SmallVector<MachineOperand, 4> PriorCond;
1211 bool PriorUnAnalyzable =
1212 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1213 if (!PriorUnAnalyzable) {
1214 // If the CFG for the prior block has extra edges, remove them.
1215 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1216 !PriorCond.empty());
1218 // If the previous branch is conditional and both conditions go to the same
1219 // destination, remove the branch, replacing it with an unconditional one or
1221 if (PriorTBB && PriorTBB == PriorFBB) {
1222 DebugLoc dl = getBranchDebugLoc(PrevBB);
1223 TII->RemoveBranch(PrevBB);
1225 if (PriorTBB != MBB)
1226 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1229 goto ReoptimizeBlock;
1232 // If the previous block unconditionally falls through to this block and
1233 // this block has no other predecessors, move the contents of this block
1234 // into the prior block. This doesn't usually happen when SimplifyCFG
1235 // has been used, but it can happen if tail merging splits a fall-through
1236 // predecessor of a block.
1237 // This has to check PrevBB->succ_size() because EH edges are ignored by
1239 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1240 PrevBB.succ_size() == 1 &&
1241 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1242 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1243 << "From MBB: " << *MBB);
1244 // Remove redundant DBG_VALUEs first.
1245 if (PrevBB.begin() != PrevBB.end()) {
1246 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1248 MachineBasicBlock::iterator MBBIter = MBB->begin();
1249 // Check if DBG_VALUE at the end of PrevBB is identical to the
1250 // DBG_VALUE at the beginning of MBB.
1251 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1252 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1253 if (!MBBIter->isIdenticalTo(PrevBBIter))
1255 MachineInstr *DuplicateDbg = MBBIter;
1256 ++MBBIter; -- PrevBBIter;
1257 DuplicateDbg->eraseFromParent();
1260 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1261 PrevBB.removeSuccessor(PrevBB.succ_begin());
1262 assert(PrevBB.succ_empty());
1263 PrevBB.transferSuccessors(MBB);
1268 // If the previous branch *only* branches to *this* block (conditional or
1269 // not) remove the branch.
1270 if (PriorTBB == MBB && !PriorFBB) {
1271 TII->RemoveBranch(PrevBB);
1274 goto ReoptimizeBlock;
1277 // If the prior block branches somewhere else on the condition and here if
1278 // the condition is false, remove the uncond second branch.
1279 if (PriorFBB == MBB) {
1280 DebugLoc dl = getBranchDebugLoc(PrevBB);
1281 TII->RemoveBranch(PrevBB);
1282 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1285 goto ReoptimizeBlock;
1288 // If the prior block branches here on true and somewhere else on false, and
1289 // if the branch condition is reversible, reverse the branch to create a
1291 if (PriorTBB == MBB) {
1292 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1293 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1294 DebugLoc dl = getBranchDebugLoc(PrevBB);
1295 TII->RemoveBranch(PrevBB);
1296 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1299 goto ReoptimizeBlock;
1303 // If this block has no successors (e.g. it is a return block or ends with
1304 // a call to a no-return function like abort or __cxa_throw) and if the pred
1305 // falls through into this block, and if it would otherwise fall through
1306 // into the block after this, move this block to the end of the function.
1308 // We consider it more likely that execution will stay in the function (e.g.
1309 // due to loops) than it is to exit it. This asserts in loops etc, moving
1310 // the assert condition out of the loop body.
1311 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1312 MachineFunction::iterator(PriorTBB) == FallThrough &&
1313 !MBB->canFallThrough()) {
1314 bool DoTransform = true;
1316 // We have to be careful that the succs of PredBB aren't both no-successor
1317 // blocks. If neither have successors and if PredBB is the second from
1318 // last block in the function, we'd just keep swapping the two blocks for
1319 // last. Only do the swap if one is clearly better to fall through than
1321 if (FallThrough == --MF.end() &&
1322 !IsBetterFallthrough(PriorTBB, MBB))
1323 DoTransform = false;
1326 // Reverse the branch so we will fall through on the previous true cond.
1327 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1328 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1329 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1330 << "To make fallthrough to: " << *PriorTBB << "\n");
1332 DebugLoc dl = getBranchDebugLoc(PrevBB);
1333 TII->RemoveBranch(PrevBB);
1334 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1336 // Move this block to the end of the function.
1337 MBB->moveAfter(--MF.end());
1346 // Analyze the branch in the current block.
1347 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1348 SmallVector<MachineOperand, 4> CurCond;
1349 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1350 if (!CurUnAnalyzable) {
1351 // If the CFG for the prior block has extra edges, remove them.
1352 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1354 // If this is a two-way branch, and the FBB branches to this block, reverse
1355 // the condition so the single-basic-block loop is faster. Instead of:
1356 // Loop: xxx; jcc Out; jmp Loop
1358 // Loop: xxx; jncc Loop; jmp Out
1359 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1360 SmallVector<MachineOperand, 4> NewCond(CurCond);
1361 if (!TII->ReverseBranchCondition(NewCond)) {
1362 DebugLoc dl = getBranchDebugLoc(*MBB);
1363 TII->RemoveBranch(*MBB);
1364 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1367 goto ReoptimizeBlock;
1371 // If this branch is the only thing in its block, see if we can forward
1372 // other blocks across it.
1373 if (CurTBB && CurCond.empty() && !CurFBB &&
1374 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1375 !MBB->hasAddressTaken()) {
1376 DebugLoc dl = getBranchDebugLoc(*MBB);
1377 // This block may contain just an unconditional branch. Because there can
1378 // be 'non-branch terminators' in the block, try removing the branch and
1379 // then seeing if the block is empty.
1380 TII->RemoveBranch(*MBB);
1381 // If the only things remaining in the block are debug info, remove these
1382 // as well, so this will behave the same as an empty block in non-debug
1384 if (IsEmptyBlock(MBB)) {
1385 // Make the block empty, losing the debug info (we could probably
1386 // improve this in some cases.)
1387 MBB->erase(MBB->begin(), MBB->end());
1389 // If this block is just an unconditional branch to CurTBB, we can
1390 // usually completely eliminate the block. The only case we cannot
1391 // completely eliminate the block is when the block before this one
1392 // falls through into MBB and we can't understand the prior block's branch
1395 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1396 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1397 !PrevBB.isSuccessor(MBB)) {
1398 // If the prior block falls through into us, turn it into an
1399 // explicit branch to us to make updates simpler.
1400 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1401 PriorTBB != MBB && PriorFBB != MBB) {
1403 assert(PriorCond.empty() && !PriorFBB &&
1404 "Bad branch analysis");
1407 assert(!PriorFBB && "Machine CFG out of date!");
1410 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1411 TII->RemoveBranch(PrevBB);
1412 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1415 // Iterate through all the predecessors, revectoring each in-turn.
1417 bool DidChange = false;
1418 bool HasBranchToSelf = false;
1419 while(PI != MBB->pred_size()) {
1420 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1422 // If this block has an uncond branch to itself, leave it.
1424 HasBranchToSelf = true;
1427 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1428 // If this change resulted in PMBB ending in a conditional
1429 // branch where both conditions go to the same destination,
1430 // change this to an unconditional branch (and fix the CFG).
1431 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1432 SmallVector<MachineOperand, 4> NewCurCond;
1433 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1434 NewCurFBB, NewCurCond, true);
1435 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1436 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1437 TII->RemoveBranch(*PMBB);
1439 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1442 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1447 // Change any jumptables to go to the new MBB.
1448 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1449 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1453 if (!HasBranchToSelf) return MadeChange;
1458 // Add the branch back if the block is more than just an uncond branch.
1459 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1463 // If the prior block doesn't fall through into this block, and if this
1464 // block doesn't fall through into some other block, see if we can find a
1465 // place to move this block where a fall-through will happen.
1466 if (!PrevBB.canFallThrough()) {
1468 // Now we know that there was no fall-through into this block, check to
1469 // see if it has a fall-through into its successor.
1470 bool CurFallsThru = MBB->canFallThrough();
1472 if (!MBB->isLandingPad()) {
1473 // Check all the predecessors of this block. If one of them has no fall
1474 // throughs, move this block right after it.
1475 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1476 E = MBB->pred_end(); PI != E; ++PI) {
1477 // Analyze the branch at the end of the pred.
1478 MachineBasicBlock *PredBB = *PI;
1479 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1480 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1481 SmallVector<MachineOperand, 4> PredCond;
1482 if (PredBB != MBB && !PredBB->canFallThrough() &&
1483 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1484 && (!CurFallsThru || !CurTBB || !CurFBB)
1485 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1486 // If the current block doesn't fall through, just move it.
1487 // If the current block can fall through and does not end with a
1488 // conditional branch, we need to append an unconditional jump to
1489 // the (current) next block. To avoid a possible compile-time
1490 // infinite loop, move blocks only backward in this case.
1491 // Also, if there are already 2 branches here, we cannot add a third;
1492 // this means we have the case
1497 MachineBasicBlock *NextBB =
1498 std::next(MachineFunction::iterator(MBB));
1500 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1502 MBB->moveAfter(PredBB);
1504 goto ReoptimizeBlock;
1509 if (!CurFallsThru) {
1510 // Check all successors to see if we can move this block before it.
1511 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1512 E = MBB->succ_end(); SI != E; ++SI) {
1513 // Analyze the branch at the end of the block before the succ.
1514 MachineBasicBlock *SuccBB = *SI;
1515 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1517 // If this block doesn't already fall-through to that successor, and if
1518 // the succ doesn't already have a block that can fall through into it,
1519 // and if the successor isn't an EH destination, we can arrange for the
1520 // fallthrough to happen.
1521 if (SuccBB != MBB && &*SuccPrev != MBB &&
1522 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1523 !SuccBB->isLandingPad()) {
1524 MBB->moveBefore(SuccBB);
1526 goto ReoptimizeBlock;
1530 // Okay, there is no really great place to put this block. If, however,
1531 // the block before this one would be a fall-through if this block were
1532 // removed, move this block to the end of the function.
1533 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1534 SmallVector<MachineOperand, 4> PrevCond;
1535 if (FallThrough != MF.end() &&
1536 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1537 PrevBB.isSuccessor(FallThrough)) {
1538 MBB->moveAfter(--MF.end());
1548 //===----------------------------------------------------------------------===//
1549 // Hoist Common Code
1550 //===----------------------------------------------------------------------===//
1552 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1553 /// blocks to their common predecessor.
1554 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1555 bool MadeChange = false;
1556 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1557 MachineBasicBlock *MBB = I++;
1558 MadeChange |= HoistCommonCodeInSuccs(MBB);
1564 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1565 /// its 'true' successor.
1566 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1567 MachineBasicBlock *TrueBB) {
1568 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1569 E = BB->succ_end(); SI != E; ++SI) {
1570 MachineBasicBlock *SuccBB = *SI;
1571 if (SuccBB != TrueBB)
1577 template <class Container>
1578 static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
1580 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1581 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1588 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1589 /// in successors to. The location is usually just before the terminator,
1590 /// however if the terminator is a conditional branch and its previous
1591 /// instruction is the flag setting instruction, the previous instruction is
1592 /// the preferred location. This function also gathers uses and defs of the
1593 /// instructions from the insertion point to the end of the block. The data is
1594 /// used by HoistCommonCodeInSuccs to ensure safety.
1596 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1597 const TargetInstrInfo *TII,
1598 const TargetRegisterInfo *TRI,
1599 SmallSet<unsigned,4> &Uses,
1600 SmallSet<unsigned,4> &Defs) {
1601 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1602 if (!TII->isUnpredicatedTerminator(Loc))
1605 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1606 const MachineOperand &MO = Loc->getOperand(i);
1609 unsigned Reg = MO.getReg();
1613 addRegAndItsAliases(Reg, TRI, Uses);
1616 // Don't try to hoist code in the rare case the terminator defines a
1617 // register that is later used.
1620 // If the terminator defines a register, make sure we don't hoist
1621 // the instruction whose def might be clobbered by the terminator.
1622 addRegAndItsAliases(Reg, TRI, Defs);
1628 if (Loc == MBB->begin())
1631 // The terminator is probably a conditional branch, try not to separate the
1632 // branch from condition setting instruction.
1633 MachineBasicBlock::iterator PI = Loc;
1635 while (PI != MBB->begin() && PI->isDebugValue())
1639 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1640 const MachineOperand &MO = PI->getOperand(i);
1641 // If PI has a regmask operand, it is probably a call. Separate away.
1644 if (!MO.isReg() || MO.isUse())
1646 unsigned Reg = MO.getReg();
1649 if (Uses.count(Reg))
1653 // The condition setting instruction is not just before the conditional
1657 // Be conservative, don't insert instruction above something that may have
1658 // side-effects. And since it's potentially bad to separate flag setting
1659 // instruction from the conditional branch, just abort the optimization
1661 // Also avoid moving code above predicated instruction since it's hard to
1662 // reason about register liveness with predicated instruction.
1663 bool DontMoveAcrossStore = true;
1664 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(PI))
1668 // Find out what registers are live. Note this routine is ignoring other live
1669 // registers which are only used by instructions in successor blocks.
1670 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1671 const MachineOperand &MO = PI->getOperand(i);
1674 unsigned Reg = MO.getReg();
1678 addRegAndItsAliases(Reg, TRI, Uses);
1680 if (Uses.erase(Reg)) {
1681 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1682 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1683 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1686 addRegAndItsAliases(Reg, TRI, Defs);
1693 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1694 /// sequence at the start of the function, move the instructions before MBB
1695 /// terminator if it's legal.
1696 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1697 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1698 SmallVector<MachineOperand, 4> Cond;
1699 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1702 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1704 // Malformed bcc? True and false blocks are the same?
1707 // Restrict the optimization to cases where MBB is the only predecessor,
1708 // it is an obvious win.
1709 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1712 // Find a suitable position to hoist the common instructions to. Also figure
1713 // out which registers are used or defined by instructions from the insertion
1714 // point to the end of the block.
1715 SmallSet<unsigned, 4> Uses, Defs;
1716 MachineBasicBlock::iterator Loc =
1717 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1718 if (Loc == MBB->end())
1721 bool HasDups = false;
1722 SmallVector<unsigned, 4> LocalDefs;
1723 SmallSet<unsigned, 4> LocalDefsSet;
1724 MachineBasicBlock::iterator TIB = TBB->begin();
1725 MachineBasicBlock::iterator FIB = FBB->begin();
1726 MachineBasicBlock::iterator TIE = TBB->end();
1727 MachineBasicBlock::iterator FIE = FBB->end();
1728 while (TIB != TIE && FIB != FIE) {
1729 // Skip dbg_value instructions. These do not count.
1730 if (TIB->isDebugValue()) {
1731 while (TIB != TIE && TIB->isDebugValue())
1736 if (FIB->isDebugValue()) {
1737 while (FIB != FIE && FIB->isDebugValue())
1742 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1745 if (TII->isPredicated(TIB))
1746 // Hard to reason about register liveness with predicated instruction.
1750 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1751 MachineOperand &MO = TIB->getOperand(i);
1752 // Don't attempt to hoist instructions with register masks.
1753 if (MO.isRegMask()) {
1759 unsigned Reg = MO.getReg();
1763 if (Uses.count(Reg)) {
1764 // Avoid clobbering a register that's used by the instruction at
1765 // the point of insertion.
1770 if (Defs.count(Reg) && !MO.isDead()) {
1771 // Don't hoist the instruction if the def would be clobber by the
1772 // instruction at the point insertion. FIXME: This is overly
1773 // conservative. It should be possible to hoist the instructions
1774 // in BB2 in the following example:
1776 // r1, eflag = op1 r2, r3
1785 } else if (!LocalDefsSet.count(Reg)) {
1786 if (Defs.count(Reg)) {
1787 // Use is defined by the instruction at the point of insertion.
1792 if (MO.isKill() && Uses.count(Reg))
1793 // Kills a register that's read by the instruction at the point of
1794 // insertion. Remove the kill marker.
1795 MO.setIsKill(false);
1801 bool DontMoveAcrossStore = true;
1802 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1805 // Remove kills from LocalDefsSet, these registers had short live ranges.
1806 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1807 MachineOperand &MO = TIB->getOperand(i);
1808 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1810 unsigned Reg = MO.getReg();
1811 if (!Reg || !LocalDefsSet.count(Reg))
1813 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1814 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1815 LocalDefsSet.erase(*AI);
1817 LocalDefsSet.erase(Reg);
1821 // Track local defs so we can update liveins.
1822 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1823 MachineOperand &MO = TIB->getOperand(i);
1824 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1826 unsigned Reg = MO.getReg();
1829 LocalDefs.push_back(Reg);
1830 addRegAndItsAliases(Reg, TRI, LocalDefsSet);
1841 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1842 FBB->erase(FBB->begin(), FIB);
1845 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1846 unsigned Def = LocalDefs[i];
1847 if (LocalDefsSet.count(Def)) {
1848 TBB->addLiveIn(Def);
1849 FBB->addLiveIn(Def);