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/Analysis.h"
25 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
26 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
27 #include "llvm/CodeGen/MachineFunctionPass.h"
28 #include "llvm/CodeGen/MachineJumpTableInfo.h"
29 #include "llvm/CodeGen/MachineMemOperand.h"
30 #include "llvm/CodeGen/MachineLoopInfo.h"
31 #include "llvm/CodeGen/MachineModuleInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/Passes.h"
34 #include "llvm/CodeGen/TargetPassConfig.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Target/TargetInstrInfo.h"
41 #include "llvm/Target/TargetRegisterInfo.h"
42 #include "llvm/Target/TargetSubtargetInfo.h"
46 #define DEBUG_TYPE "branchfolding"
48 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
49 STATISTIC(NumBranchOpts, "Number of branches optimized");
50 STATISTIC(NumTailMerge , "Number of block tails merged");
51 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
52 STATISTIC(NumTailCalls, "Number of tail calls optimized");
54 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
55 cl::init(cl::BOU_UNSET), cl::Hidden);
57 // Throttle for huge numbers of predecessors (compile speed problems)
58 static cl::opt<unsigned>
59 TailMergeThreshold("tail-merge-threshold",
60 cl::desc("Max number of predecessors to consider tail merging"),
61 cl::init(150), cl::Hidden);
63 // Heuristic for tail merging (and, inversely, tail duplication).
64 // TODO: This should be replaced with a target query.
65 static cl::opt<unsigned>
66 TailMergeSize("tail-merge-size",
67 cl::desc("Min number of instructions to consider tail merging"),
68 cl::init(3), cl::Hidden);
71 /// BranchFolderPass - Wrap branch folder in a machine function pass.
72 class BranchFolderPass : public MachineFunctionPass {
75 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
77 bool runOnMachineFunction(MachineFunction &MF) override;
79 void getAnalysisUsage(AnalysisUsage &AU) const override {
80 AU.addRequired<MachineBlockFrequencyInfo>();
81 AU.addRequired<MachineBranchProbabilityInfo>();
82 AU.addRequired<TargetPassConfig>();
83 MachineFunctionPass::getAnalysisUsage(AU);
88 char BranchFolderPass::ID = 0;
89 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
91 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
92 "Control Flow Optimizer", false, false)
94 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
95 if (skipFunction(*MF.getFunction()))
98 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
99 // TailMerge can create jump into if branches that make CFG irreducible for
100 // HW that requires structurized CFG.
101 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
102 PassConfig->getEnableTailMerge();
103 BranchFolder::MBFIWrapper MBBFreqInfo(
104 getAnalysis<MachineBlockFrequencyInfo>());
105 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo,
106 getAnalysis<MachineBranchProbabilityInfo>());
107 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
108 MF.getSubtarget().getRegisterInfo(),
109 getAnalysisIfAvailable<MachineModuleInfo>());
112 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
113 MBFIWrapper &FreqInfo,
114 const MachineBranchProbabilityInfo &ProbInfo,
115 unsigned MinTailLength)
116 : EnableHoistCommonCode(CommonHoist), MinCommonTailLength(MinTailLength),
117 MBBFreqInfo(FreqInfo), MBPI(ProbInfo) {
118 if (MinCommonTailLength == 0)
119 MinCommonTailLength = TailMergeSize;
120 switch (FlagEnableTailMerge) {
121 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
122 case cl::BOU_TRUE: EnableTailMerge = true; break;
123 case cl::BOU_FALSE: EnableTailMerge = false; break;
127 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
128 /// function, updating the CFG.
129 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
130 assert(MBB->pred_empty() && "MBB must be dead!");
131 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
133 MachineFunction *MF = MBB->getParent();
134 // drop all successors.
135 while (!MBB->succ_empty())
136 MBB->removeSuccessor(MBB->succ_end()-1);
138 // Avoid matching if this pointer gets reused.
139 TriedMerging.erase(MBB);
143 FuncletMembership.erase(MBB);
145 MLI->removeBlock(MBB);
148 /// OptimizeFunction - Perhaps branch folding, tail merging and other
149 /// CFG optimizations on the given function. Block placement changes the layout
150 /// and may create new tail merging opportunities.
151 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
152 const TargetInstrInfo *tii,
153 const TargetRegisterInfo *tri,
154 MachineModuleInfo *mmi,
155 MachineLoopInfo *mli, bool AfterPlacement) {
156 if (!tii) return false;
158 TriedMerging.clear();
160 AfterBlockPlacement = AfterPlacement;
166 MachineRegisterInfo &MRI = MF.getRegInfo();
167 UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF);
169 MRI.invalidateLiveness();
171 // Fix CFG. The later algorithms expect it to be right.
172 bool MadeChange = false;
173 for (MachineBasicBlock &MBB : MF) {
174 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
175 SmallVector<MachineOperand, 4> Cond;
176 if (!TII->analyzeBranch(MBB, TBB, FBB, Cond, true))
177 MadeChange |= MBB.CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
180 // Recalculate funclet membership.
181 FuncletMembership = getFuncletMembership(MF);
183 bool MadeChangeThisIteration = true;
184 while (MadeChangeThisIteration) {
185 MadeChangeThisIteration = TailMergeBlocks(MF);
186 // No need to clean up if tail merging does not change anything after the
188 if (!AfterBlockPlacement || MadeChangeThisIteration)
189 MadeChangeThisIteration |= OptimizeBranches(MF);
190 if (EnableHoistCommonCode)
191 MadeChangeThisIteration |= HoistCommonCode(MF);
192 MadeChange |= MadeChangeThisIteration;
195 // See if any jump tables have become dead as the code generator
197 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
201 // Walk the function to find jump tables that are live.
202 BitVector JTIsLive(JTI->getJumpTables().size());
203 for (const MachineBasicBlock &BB : MF) {
204 for (const MachineInstr &I : BB)
205 for (const MachineOperand &Op : I.operands()) {
206 if (!Op.isJTI()) continue;
208 // Remember that this JT is live.
209 JTIsLive.set(Op.getIndex());
213 // Finally, remove dead jump tables. This happens when the
214 // indirect jump was unreachable (and thus deleted).
215 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
216 if (!JTIsLive.test(i)) {
217 JTI->RemoveJumpTable(i);
224 //===----------------------------------------------------------------------===//
225 // Tail Merging of Blocks
226 //===----------------------------------------------------------------------===//
228 /// HashMachineInstr - Compute a hash value for MI and its operands.
229 static unsigned HashMachineInstr(const MachineInstr &MI) {
230 unsigned Hash = MI.getOpcode();
231 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
232 const MachineOperand &Op = MI.getOperand(i);
234 // Merge in bits from the operand if easy. We can't use MachineOperand's
235 // hash_code here because it's not deterministic and we sort by hash value
237 unsigned OperandHash = 0;
238 switch (Op.getType()) {
239 case MachineOperand::MO_Register:
240 OperandHash = Op.getReg();
242 case MachineOperand::MO_Immediate:
243 OperandHash = Op.getImm();
245 case MachineOperand::MO_MachineBasicBlock:
246 OperandHash = Op.getMBB()->getNumber();
248 case MachineOperand::MO_FrameIndex:
249 case MachineOperand::MO_ConstantPoolIndex:
250 case MachineOperand::MO_JumpTableIndex:
251 OperandHash = Op.getIndex();
253 case MachineOperand::MO_GlobalAddress:
254 case MachineOperand::MO_ExternalSymbol:
255 // Global address / external symbol are too hard, don't bother, but do
256 // pull in the offset.
257 OperandHash = Op.getOffset();
263 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
268 /// HashEndOfMBB - Hash the last instruction in the MBB.
269 static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) {
270 MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr();
274 return HashMachineInstr(*I);
277 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
278 /// of instructions they actually have in common together at their end. Return
279 /// iterators for the first shared instruction in each block.
280 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
281 MachineBasicBlock *MBB2,
282 MachineBasicBlock::iterator &I1,
283 MachineBasicBlock::iterator &I2) {
287 unsigned TailLen = 0;
288 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
290 // Skip debugging pseudos; necessary to avoid changing the code.
291 while (I1->isDebugValue()) {
292 if (I1==MBB1->begin()) {
293 while (I2->isDebugValue()) {
294 if (I2==MBB2->begin())
295 // I1==DBG at begin; I2==DBG at begin
300 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
305 // I1==first (untested) non-DBG preceding known match
306 while (I2->isDebugValue()) {
307 if (I2==MBB2->begin()) {
309 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
314 // I1, I2==first (untested) non-DBGs preceding known match
315 if (!I1->isIdenticalTo(*I2) ||
316 // FIXME: This check is dubious. It's used to get around a problem where
317 // people incorrectly expect inline asm directives to remain in the same
318 // relative order. This is untenable because normal compiler
319 // optimizations (like this one) may reorder and/or merge these
327 // Back past possible debugging pseudos at beginning of block. This matters
328 // when one block differs from the other only by whether debugging pseudos
329 // are present at the beginning. (This way, the various checks later for
330 // I1==MBB1->begin() work as expected.)
331 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
333 while (I2->isDebugValue()) {
334 if (I2 == MBB2->begin())
340 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
342 while (I1->isDebugValue()) {
343 if (I1 == MBB1->begin())
352 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
353 /// after it, replacing it with an unconditional branch to NewDest.
354 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
355 MachineBasicBlock *NewDest) {
356 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
359 NewDest->clearLiveIns();
360 computeLiveIns(LiveRegs, *TRI, *NewDest);
366 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
367 /// MBB so that the part before the iterator falls into the part starting at the
368 /// iterator. This returns the new MBB.
369 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
370 MachineBasicBlock::iterator BBI1,
371 const BasicBlock *BB) {
372 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
375 MachineFunction &MF = *CurMBB.getParent();
377 // Create the fall-through block.
378 MachineFunction::iterator MBBI = CurMBB.getIterator();
379 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
380 CurMBB.getParent()->insert(++MBBI, NewMBB);
382 // Move all the successors of this block to the specified block.
383 NewMBB->transferSuccessors(&CurMBB);
385 // Add an edge from CurMBB to NewMBB for the fall-through.
386 CurMBB.addSuccessor(NewMBB);
388 // Splice the code over.
389 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
391 // NewMBB belongs to the same loop as CurMBB.
393 if (MachineLoop *ML = MLI->getLoopFor(&CurMBB))
394 ML->addBasicBlockToLoop(NewMBB, MLI->getBase());
396 // NewMBB inherits CurMBB's block frequency.
397 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
400 computeLiveIns(LiveRegs, *TRI, *NewMBB);
402 // Add the new block to the funclet.
403 const auto &FuncletI = FuncletMembership.find(&CurMBB);
404 if (FuncletI != FuncletMembership.end()) {
405 auto n = FuncletI->second;
406 FuncletMembership[NewMBB] = n;
412 /// EstimateRuntime - Make a rough estimate for how long it will take to run
413 /// the specified code.
414 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
415 MachineBasicBlock::iterator E) {
417 for (; I != E; ++I) {
418 if (I->isDebugValue())
422 else if (I->mayLoad() || I->mayStore())
430 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
431 // branches temporarily for tail merging). In the case where CurMBB ends
432 // with a conditional branch to the next block, optimize by reversing the
433 // test and conditionally branching to SuccMBB instead.
434 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
435 const TargetInstrInfo *TII) {
436 MachineFunction *MF = CurMBB->getParent();
437 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
438 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
439 SmallVector<MachineOperand, 4> Cond;
440 DebugLoc dl; // FIXME: this is nowhere
441 if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
442 MachineBasicBlock *NextBB = &*I;
443 if (TBB == NextBB && !Cond.empty() && !FBB) {
444 if (!TII->reverseBranchCondition(Cond)) {
445 TII->removeBranch(*CurMBB);
446 TII->insertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
451 TII->insertBranch(*CurMBB, SuccBB, nullptr,
452 SmallVector<MachineOperand, 0>(), dl);
456 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
457 if (getHash() < o.getHash())
459 if (getHash() > o.getHash())
461 if (getBlock()->getNumber() < o.getBlock()->getNumber())
463 if (getBlock()->getNumber() > o.getBlock()->getNumber())
465 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
466 // an object with itself.
467 #ifndef _GLIBCXX_DEBUG
468 llvm_unreachable("Predecessor appears twice");
475 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
476 auto I = MergedBBFreq.find(MBB);
478 if (I != MergedBBFreq.end())
481 return MBFI.getBlockFreq(MBB);
484 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
486 MergedBBFreq[MBB] = F;
490 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
491 const MachineBasicBlock *MBB) const {
492 return MBFI.printBlockFreq(OS, getBlockFreq(MBB));
496 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
497 const BlockFrequency Freq) const {
498 return MBFI.printBlockFreq(OS, Freq);
501 /// CountTerminators - Count the number of terminators in the given
502 /// block and set I to the position of the first non-terminator, if there
503 /// is one, or MBB->end() otherwise.
504 static unsigned CountTerminators(MachineBasicBlock *MBB,
505 MachineBasicBlock::iterator &I) {
507 unsigned NumTerms = 0;
509 if (I == MBB->begin()) {
514 if (!I->isTerminator()) break;
520 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
521 /// and decide if it would be profitable to merge those tails. Return the
522 /// length of the common tail and iterators to the first common instruction
524 /// MBB1, MBB2 The blocks to check
525 /// MinCommonTailLength Minimum size of tail block to be merged.
526 /// CommonTailLen Out parameter to record the size of the shared tail between
528 /// I1, I2 Iterator references that will be changed to point to the first
529 /// instruction in the common tail shared by MBB1,MBB2
530 /// SuccBB A common successor of MBB1, MBB2 which are in a canonical form
531 /// relative to SuccBB
532 /// PredBB The layout predecessor of SuccBB, if any.
533 /// FuncletMembership map from block to funclet #.
534 /// AfterPlacement True if we are merging blocks after layout. Stricter
535 /// thresholds apply to prevent undoing tail-duplication.
537 ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2,
538 unsigned MinCommonTailLength, unsigned &CommonTailLen,
539 MachineBasicBlock::iterator &I1,
540 MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB,
541 MachineBasicBlock *PredBB,
542 DenseMap<const MachineBasicBlock *, int> &FuncletMembership,
543 bool AfterPlacement) {
544 // It is never profitable to tail-merge blocks from two different funclets.
545 if (!FuncletMembership.empty()) {
546 auto Funclet1 = FuncletMembership.find(MBB1);
547 assert(Funclet1 != FuncletMembership.end());
548 auto Funclet2 = FuncletMembership.find(MBB2);
549 assert(Funclet2 != FuncletMembership.end());
550 if (Funclet1->second != Funclet2->second)
554 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
555 if (CommonTailLen == 0)
557 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
558 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
561 // It's almost always profitable to merge any number of non-terminator
562 // instructions with the block that falls through into the common successor.
563 // This is true only for a single successor. For multiple successors, we are
564 // trading a conditional branch for an unconditional one.
565 // TODO: Re-visit successor size for non-layout tail merging.
566 if ((MBB1 == PredBB || MBB2 == PredBB) &&
567 (!AfterPlacement || MBB1->succ_size() == 1)) {
568 MachineBasicBlock::iterator I;
569 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
570 if (CommonTailLen > NumTerms)
574 // If one of the blocks can be completely merged and happens to be in
575 // a position where the other could fall through into it, merge any number
576 // of instructions, because it can be done without a branch.
577 // TODO: If the blocks are not adjacent, move one of them so that they are?
578 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
580 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
583 // If both blocks have an unconditional branch temporarily stripped out,
584 // count that as an additional common instruction for the following
585 // heuristics. This heuristic is only accurate for single-succ blocks, so to
586 // make sure that during layout merging and duplicating don't crash, we check
587 // for that when merging during layout.
588 unsigned EffectiveTailLen = CommonTailLen;
589 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
590 (MBB1->succ_size() == 1 || !AfterPlacement) &&
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 return EffectiveTailLen >= 2 && MF->getFunction()->optForSize() &&
605 (I1 == MBB1->begin() || I2 == MBB2->begin());
608 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
609 /// hash CurHash (guaranteed to match the last element). Build the vector
610 /// SameTails of all those that have the (same) largest number of instructions
611 /// in common of any pair of these blocks. SameTails entries contain an
612 /// iterator into MergePotentials (from which the MachineBasicBlock can be
613 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
614 /// instruction where the matching code sequence begins.
615 /// Order of elements in SameTails is the reverse of the order in which
616 /// those blocks appear in MergePotentials (where they are not necessarily
618 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
619 unsigned MinCommonTailLength,
620 MachineBasicBlock *SuccBB,
621 MachineBasicBlock *PredBB) {
622 unsigned maxCommonTailLength = 0U;
624 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
625 MPIterator HighestMPIter = std::prev(MergePotentials.end());
626 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
627 B = MergePotentials.begin();
628 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
629 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
630 unsigned CommonTailLen;
631 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
633 CommonTailLen, TrialBBI1, TrialBBI2,
636 AfterBlockPlacement)) {
637 if (CommonTailLen > maxCommonTailLength) {
639 maxCommonTailLength = CommonTailLen;
640 HighestMPIter = CurMPIter;
641 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
643 if (HighestMPIter == CurMPIter &&
644 CommonTailLen == maxCommonTailLength)
645 SameTails.push_back(SameTailElt(I, TrialBBI2));
651 return maxCommonTailLength;
654 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
655 /// MergePotentials, restoring branches at ends of blocks as appropriate.
656 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
657 MachineBasicBlock *SuccBB,
658 MachineBasicBlock *PredBB) {
659 MPIterator CurMPIter, B;
660 for (CurMPIter = std::prev(MergePotentials.end()),
661 B = MergePotentials.begin();
662 CurMPIter->getHash() == CurHash; --CurMPIter) {
663 // Put the unconditional branch back, if we need one.
664 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
665 if (SuccBB && CurMBB != PredBB)
666 FixTail(CurMBB, SuccBB, TII);
670 if (CurMPIter->getHash() != CurHash)
672 MergePotentials.erase(CurMPIter, MergePotentials.end());
675 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
676 /// only of the common tail. Create a block that does by splitting one.
677 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
678 MachineBasicBlock *SuccBB,
679 unsigned maxCommonTailLength,
680 unsigned &commonTailIndex) {
682 unsigned TimeEstimate = ~0U;
683 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
684 // Use PredBB if possible; that doesn't require a new branch.
685 if (SameTails[i].getBlock() == PredBB) {
689 // Otherwise, make a (fairly bogus) choice based on estimate of
690 // how long it will take the various blocks to execute.
691 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
692 SameTails[i].getTailStartPos());
693 if (t <= TimeEstimate) {
699 MachineBasicBlock::iterator BBI =
700 SameTails[commonTailIndex].getTailStartPos();
701 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
703 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
704 << maxCommonTailLength);
706 // If the split block unconditionally falls-thru to SuccBB, it will be
707 // merged. In control flow terms it should then take SuccBB's name. e.g. If
708 // SuccBB is an inner loop, the common tail is still part of the inner loop.
709 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
710 SuccBB->getBasicBlock() : MBB->getBasicBlock();
711 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
713 DEBUG(dbgs() << "... failed!");
717 SameTails[commonTailIndex].setBlock(newMBB);
718 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
720 // If we split PredBB, newMBB is the new predecessor.
728 mergeOperations(MachineBasicBlock::iterator MBBIStartPos,
729 MachineBasicBlock &MBBCommon) {
730 MachineBasicBlock *MBB = MBBIStartPos->getParent();
731 // Note CommonTailLen does not necessarily matches the size of
732 // the common BB nor all its instructions because of debug
733 // instructions differences.
734 unsigned CommonTailLen = 0;
735 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
738 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
739 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
740 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
741 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
743 while (CommonTailLen--) {
744 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
747 if (MBBI->isDebugValue()) {
752 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
755 assert(MBBICommon != MBBIECommon &&
756 "Reached BB end within common tail length!");
757 assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!");
759 // Merge MMOs from memory operations in the common block.
760 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
761 MBBICommon->setMemRefs(MBBICommon->mergeMemRefsWith(*MBBI));
762 // Drop undef flags if they aren't present in all merged instructions.
763 for (unsigned I = 0, E = MBBICommon->getNumOperands(); I != E; ++I) {
764 MachineOperand &MO = MBBICommon->getOperand(I);
765 if (MO.isReg() && MO.isUndef()) {
766 const MachineOperand &OtherMO = MBBI->getOperand(I);
767 if (!OtherMO.isUndef())
768 MO.setIsUndef(false);
777 // See if any of the blocks in MergePotentials (which all have SuccBB as a
778 // successor, or all have no successor if it is null) can be tail-merged.
779 // If there is a successor, any blocks in MergePotentials that are not
780 // tail-merged and are not immediately before Succ must have an unconditional
781 // branch to Succ added (but the predecessor/successor lists need no
782 // adjustment). The lone predecessor of Succ that falls through into Succ,
783 // if any, is given in PredBB.
784 // MinCommonTailLength - Except for the special cases below, tail-merge if
785 // there are at least this many instructions in common.
786 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
787 MachineBasicBlock *PredBB,
788 unsigned MinCommonTailLength) {
789 bool MadeChange = false;
791 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
792 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
793 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
794 << (i == e-1 ? "" : ", ");
797 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
799 dbgs() << " which has fall-through from BB#"
800 << PredBB->getNumber() << "\n";
802 dbgs() << "Looking for common tails of at least "
803 << MinCommonTailLength << " instruction"
804 << (MinCommonTailLength == 1 ? "" : "s") << '\n';
807 // Sort by hash value so that blocks with identical end sequences sort
809 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
811 // Walk through equivalence sets looking for actual exact matches.
812 while (MergePotentials.size() > 1) {
813 unsigned CurHash = MergePotentials.back().getHash();
815 // Build SameTails, identifying the set of blocks with this hash code
816 // and with the maximum number of instructions in common.
817 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
821 // If we didn't find any pair that has at least MinCommonTailLength
822 // instructions in common, remove all blocks with this hash code and retry.
823 if (SameTails.empty()) {
824 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
828 // If one of the blocks is the entire common tail (and not the entry
829 // block, which we can't jump to), we can treat all blocks with this same
830 // tail at once. Use PredBB if that is one of the possibilities, as that
831 // will not introduce any extra branches.
832 MachineBasicBlock *EntryBB =
833 &MergePotentials.front().getBlock()->getParent()->front();
834 unsigned commonTailIndex = SameTails.size();
835 // If there are two blocks, check to see if one can be made to fall through
837 if (SameTails.size() == 2 &&
838 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
839 SameTails[1].tailIsWholeBlock())
841 else if (SameTails.size() == 2 &&
842 SameTails[1].getBlock()->isLayoutSuccessor(
843 SameTails[0].getBlock()) &&
844 SameTails[0].tailIsWholeBlock())
847 // Otherwise just pick one, favoring the fall-through predecessor if
849 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
850 MachineBasicBlock *MBB = SameTails[i].getBlock();
851 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
857 if (SameTails[i].tailIsWholeBlock())
862 if (commonTailIndex == SameTails.size() ||
863 (SameTails[commonTailIndex].getBlock() == PredBB &&
864 !SameTails[commonTailIndex].tailIsWholeBlock())) {
865 // None of the blocks consist entirely of the common tail.
866 // Split a block so that one does.
867 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
868 maxCommonTailLength, commonTailIndex)) {
869 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
874 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
876 // Recompute common tail MBB's edge weights and block frequency.
877 setCommonTailEdgeWeights(*MBB);
879 // Remove the original debug location from the common tail.
880 for (auto &MI : *MBB)
881 if (!MI.isDebugValue())
882 MI.setDebugLoc(DebugLoc());
884 // MBB is common tail. Adjust all other BB's to jump to this one.
885 // Traversal must be forwards so erases work.
886 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
888 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
889 if (commonTailIndex == i)
891 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
892 << (i == e-1 ? "" : ", "));
893 // Merge operations (MMOs, undef flags)
894 mergeOperations(SameTails[i].getTailStartPos(), *MBB);
895 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
896 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
897 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
898 MergePotentials.erase(SameTails[i].getMPIter());
900 DEBUG(dbgs() << "\n");
901 // We leave commonTailIndex in the worklist in case there are other blocks
902 // that match it with a smaller number of instructions.
908 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
909 bool MadeChange = false;
910 if (!EnableTailMerge) return MadeChange;
912 // First find blocks with no successors.
913 // Block placement does not create new tail merging opportunities for these
915 if (!AfterBlockPlacement) {
916 MergePotentials.clear();
917 for (MachineBasicBlock &MBB : MF) {
918 if (MergePotentials.size() == TailMergeThreshold)
920 if (!TriedMerging.count(&MBB) && MBB.succ_empty())
921 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB));
924 // If this is a large problem, avoid visiting the same basic blocks
926 if (MergePotentials.size() == TailMergeThreshold)
927 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
928 TriedMerging.insert(MergePotentials[i].getBlock());
930 // See if we can do any tail merging on those.
931 if (MergePotentials.size() >= 2)
932 MadeChange |= TryTailMergeBlocks(nullptr, nullptr, MinCommonTailLength);
935 // Look at blocks (IBB) with multiple predecessors (PBB).
936 // We change each predecessor to a canonical form, by
937 // (1) temporarily removing any unconditional branch from the predecessor
939 // (2) alter conditional branches so they branch to the other block
940 // not IBB; this may require adding back an unconditional branch to IBB
941 // later, where there wasn't one coming in. E.g.
943 // fallthrough to QBB
946 // with a conceptual B to IBB after that, which never actually exists.
947 // With those changes, we see whether the predecessors' tails match,
948 // and merge them if so. We change things out of canonical form and
949 // back to the way they were later in the process. (OptimizeBranches
950 // would undo some of this, but we can't use it, because we'd get into
951 // a compile-time infinite loop repeatedly doing and undoing the same
954 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
956 if (I->pred_size() < 2) continue;
957 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
958 MachineBasicBlock *IBB = &*I;
959 MachineBasicBlock *PredBB = &*std::prev(I);
960 MergePotentials.clear();
963 // Bail if merging after placement and IBB is the loop header because
964 // -- If merging predecessors that belong to the same loop as IBB, the
965 // common tail of merged predecessors may become the loop top if block
966 // placement is called again and the predecessors may branch to this common
967 // tail and require more branches. This can be relaxed if
968 // MachineBlockPlacement::findBestLoopTop is more flexible.
969 // --If merging predecessors that do not belong to the same loop as IBB, the
970 // loop info of IBB's loop and the other loops may be affected. Calling the
971 // block placement again may make big change to the layout and eliminate the
972 // reason to do tail merging here.
973 if (AfterBlockPlacement && MLI) {
974 ML = MLI->getLoopFor(IBB);
975 if (ML && IBB == ML->getHeader())
979 for (MachineBasicBlock *PBB : I->predecessors()) {
980 if (MergePotentials.size() == TailMergeThreshold)
983 if (TriedMerging.count(PBB))
986 // Skip blocks that loop to themselves, can't tail merge these.
990 // Visit each predecessor only once.
991 if (!UniquePreds.insert(PBB).second)
994 // Skip blocks which may jump to a landing pad. Can't tail merge these.
995 if (PBB->hasEHPadSuccessor())
998 // After block placement, only consider predecessors that belong to the
999 // same loop as IBB. The reason is the same as above when skipping loop
1001 if (AfterBlockPlacement && MLI)
1002 if (ML != MLI->getLoopFor(PBB))
1005 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1006 SmallVector<MachineOperand, 4> Cond;
1007 if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) {
1008 // Failing case: IBB is the target of a cbr, and we cannot reverse the
1010 SmallVector<MachineOperand, 4> NewCond(Cond);
1011 if (!Cond.empty() && TBB == IBB) {
1012 if (TII->reverseBranchCondition(NewCond))
1014 // This is the QBB case described above
1016 auto Next = ++PBB->getIterator();
1017 if (Next != MF.end())
1022 // Failing case: the only way IBB can be reached from PBB is via
1023 // exception handling. Happens for landing pads. Would be nice to have
1024 // a bit in the edge so we didn't have to do all this.
1025 if (IBB->isEHPad()) {
1026 MachineFunction::iterator IP = ++PBB->getIterator();
1027 MachineBasicBlock *PredNextBB = nullptr;
1031 if (IBB != PredNextBB) // fallthrough
1034 if (TBB != IBB && FBB != IBB) // cbr then ubr
1036 } else if (Cond.empty()) {
1037 if (TBB != IBB) // ubr
1040 if (TBB != IBB && IBB != PredNextBB) // cbr
1045 // Remove the unconditional branch at the end, if any.
1046 if (TBB && (Cond.empty() || FBB)) {
1047 DebugLoc dl; // FIXME: this is nowhere
1048 TII->removeBranch(*PBB);
1050 // reinsert conditional branch only, for now
1051 TII->insertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1055 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB));
1059 // If this is a large problem, avoid visiting the same basic blocks multiple
1061 if (MergePotentials.size() == TailMergeThreshold)
1062 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1063 TriedMerging.insert(MergePotentials[i].getBlock());
1065 if (MergePotentials.size() >= 2)
1066 MadeChange |= TryTailMergeBlocks(IBB, PredBB, MinCommonTailLength);
1068 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1069 // result of removing blocks in TryTailMergeBlocks.
1070 PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks
1071 if (MergePotentials.size() == 1 &&
1072 MergePotentials.begin()->getBlock() != PredBB)
1073 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1079 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1080 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1081 BlockFrequency AccumulatedMBBFreq;
1083 // Aggregate edge frequency of successor edge j:
1084 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1085 // where bb is a basic block that is in SameTails.
1086 for (const auto &Src : SameTails) {
1087 const MachineBasicBlock *SrcMBB = Src.getBlock();
1088 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1089 AccumulatedMBBFreq += BlockFreq;
1091 // It is not necessary to recompute edge weights if TailBB has less than two
1093 if (TailMBB.succ_size() <= 1)
1096 auto EdgeFreq = EdgeFreqLs.begin();
1098 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1099 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1100 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1103 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1105 if (TailMBB.succ_size() <= 1)
1109 std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0))
1111 auto EdgeFreq = EdgeFreqLs.begin();
1113 if (SumEdgeFreq > 0) {
1114 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1115 SuccI != SuccE; ++SuccI, ++EdgeFreq) {
1116 auto Prob = BranchProbability::getBranchProbability(
1117 EdgeFreq->getFrequency(), SumEdgeFreq);
1118 TailMBB.setSuccProbability(SuccI, Prob);
1123 //===----------------------------------------------------------------------===//
1124 // Branch Optimization
1125 //===----------------------------------------------------------------------===//
1127 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1128 bool MadeChange = false;
1130 // Make sure blocks are numbered in order
1131 MF.RenumberBlocks();
1132 // Renumbering blocks alters funclet membership, recalculate it.
1133 FuncletMembership = getFuncletMembership(MF);
1135 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1137 MachineBasicBlock *MBB = &*I++;
1138 MadeChange |= OptimizeBlock(MBB);
1140 // If it is dead, remove it.
1141 if (MBB->pred_empty()) {
1142 RemoveDeadBlock(MBB);
1151 // Blocks should be considered empty if they contain only debug info;
1152 // else the debug info would affect codegen.
1153 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1154 return MBB->getFirstNonDebugInstr() == MBB->end();
1157 // Blocks with only debug info and branches should be considered the same
1158 // as blocks with only branches.
1159 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1160 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1161 assert(I != MBB->end() && "empty block!");
1162 return I->isBranch();
1165 /// IsBetterFallthrough - Return true if it would be clearly better to
1166 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1167 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1168 /// result in infinite loops.
1169 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1170 MachineBasicBlock *MBB2) {
1171 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1172 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1173 // optimize branches that branch to either a return block or an assert block
1174 // into a fallthrough to the return.
1175 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1176 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1177 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1180 // If there is a clear successor ordering we make sure that one block
1181 // will fall through to the next
1182 if (MBB1->isSuccessor(MBB2)) return true;
1183 if (MBB2->isSuccessor(MBB1)) return false;
1185 return MBB2I->isCall() && !MBB1I->isCall();
1188 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1189 /// instructions on the block.
1190 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1191 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1192 if (I != MBB.end() && I->isBranch())
1193 return I->getDebugLoc();
1197 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1198 /// block. This is never called on the entry block.
1199 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1200 bool MadeChange = false;
1201 MachineFunction &MF = *MBB->getParent();
1204 MachineFunction::iterator FallThrough = MBB->getIterator();
1207 // Make sure MBB and FallThrough belong to the same funclet.
1208 bool SameFunclet = true;
1209 if (!FuncletMembership.empty() && FallThrough != MF.end()) {
1210 auto MBBFunclet = FuncletMembership.find(MBB);
1211 assert(MBBFunclet != FuncletMembership.end());
1212 auto FallThroughFunclet = FuncletMembership.find(&*FallThrough);
1213 assert(FallThroughFunclet != FuncletMembership.end());
1214 SameFunclet = MBBFunclet->second == FallThroughFunclet->second;
1217 // If this block is empty, make everyone use its fall-through, not the block
1218 // explicitly. Landing pads should not do this since the landing-pad table
1219 // points to this block. Blocks with their addresses taken shouldn't be
1221 if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() &&
1223 // Dead block? Leave for cleanup later.
1224 if (MBB->pred_empty()) return MadeChange;
1226 if (FallThrough == MF.end()) {
1227 // TODO: Simplify preds to not branch here if possible!
1228 } else if (FallThrough->isEHPad()) {
1229 // Don't rewrite to a landing pad fallthough. That could lead to the case
1230 // where a BB jumps to more than one landing pad.
1231 // TODO: Is it ever worth rewriting predecessors which don't already
1232 // jump to a landing pad, and so can safely jump to the fallthrough?
1233 } else if (MBB->isSuccessor(&*FallThrough)) {
1234 // Rewrite all predecessors of the old block to go to the fallthrough
1236 while (!MBB->pred_empty()) {
1237 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1238 Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough);
1240 // If MBB was the target of a jump table, update jump tables to go to the
1241 // fallthrough instead.
1242 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1243 MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough);
1249 // Check to see if we can simplify the terminator of the block before this
1251 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1253 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1254 SmallVector<MachineOperand, 4> PriorCond;
1255 bool PriorUnAnalyzable =
1256 TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1257 if (!PriorUnAnalyzable) {
1258 // If the CFG for the prior block has extra edges, remove them.
1259 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1260 !PriorCond.empty());
1262 // If the previous branch is conditional and both conditions go to the same
1263 // destination, remove the branch, replacing it with an unconditional one or
1265 if (PriorTBB && PriorTBB == PriorFBB) {
1266 DebugLoc dl = getBranchDebugLoc(PrevBB);
1267 TII->removeBranch(PrevBB);
1269 if (PriorTBB != MBB)
1270 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1273 goto ReoptimizeBlock;
1276 // If the previous block unconditionally falls through to this block and
1277 // this block has no other predecessors, move the contents of this block
1278 // into the prior block. This doesn't usually happen when SimplifyCFG
1279 // has been used, but it can happen if tail merging splits a fall-through
1280 // predecessor of a block.
1281 // This has to check PrevBB->succ_size() because EH edges are ignored by
1283 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1284 PrevBB.succ_size() == 1 &&
1285 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1286 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1287 << "From MBB: " << *MBB);
1288 // Remove redundant DBG_VALUEs first.
1289 if (PrevBB.begin() != PrevBB.end()) {
1290 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1292 MachineBasicBlock::iterator MBBIter = MBB->begin();
1293 // Check if DBG_VALUE at the end of PrevBB is identical to the
1294 // DBG_VALUE at the beginning of MBB.
1295 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1296 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1297 if (!MBBIter->isIdenticalTo(*PrevBBIter))
1299 MachineInstr &DuplicateDbg = *MBBIter;
1300 ++MBBIter; -- PrevBBIter;
1301 DuplicateDbg.eraseFromParent();
1304 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1305 PrevBB.removeSuccessor(PrevBB.succ_begin());
1306 assert(PrevBB.succ_empty());
1307 PrevBB.transferSuccessors(MBB);
1312 // If the previous branch *only* branches to *this* block (conditional or
1313 // not) remove the branch.
1314 if (PriorTBB == MBB && !PriorFBB) {
1315 TII->removeBranch(PrevBB);
1318 goto ReoptimizeBlock;
1321 // If the prior block branches somewhere else on the condition and here if
1322 // the condition is false, remove the uncond second branch.
1323 if (PriorFBB == MBB) {
1324 DebugLoc dl = getBranchDebugLoc(PrevBB);
1325 TII->removeBranch(PrevBB);
1326 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1329 goto ReoptimizeBlock;
1332 // If the prior block branches here on true and somewhere else on false, and
1333 // if the branch condition is reversible, reverse the branch to create a
1335 if (PriorTBB == MBB) {
1336 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1337 if (!TII->reverseBranchCondition(NewPriorCond)) {
1338 DebugLoc dl = getBranchDebugLoc(PrevBB);
1339 TII->removeBranch(PrevBB);
1340 TII->insertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1343 goto ReoptimizeBlock;
1347 // If this block has no successors (e.g. it is a return block or ends with
1348 // a call to a no-return function like abort or __cxa_throw) and if the pred
1349 // falls through into this block, and if it would otherwise fall through
1350 // into the block after this, move this block to the end of the function.
1352 // We consider it more likely that execution will stay in the function (e.g.
1353 // due to loops) than it is to exit it. This asserts in loops etc, moving
1354 // the assert condition out of the loop body.
1355 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1356 MachineFunction::iterator(PriorTBB) == FallThrough &&
1357 !MBB->canFallThrough()) {
1358 bool DoTransform = true;
1360 // We have to be careful that the succs of PredBB aren't both no-successor
1361 // blocks. If neither have successors and if PredBB is the second from
1362 // last block in the function, we'd just keep swapping the two blocks for
1363 // last. Only do the swap if one is clearly better to fall through than
1365 if (FallThrough == --MF.end() &&
1366 !IsBetterFallthrough(PriorTBB, MBB))
1367 DoTransform = false;
1370 // Reverse the branch so we will fall through on the previous true cond.
1371 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1372 if (!TII->reverseBranchCondition(NewPriorCond)) {
1373 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1374 << "To make fallthrough to: " << *PriorTBB << "\n");
1376 DebugLoc dl = getBranchDebugLoc(PrevBB);
1377 TII->removeBranch(PrevBB);
1378 TII->insertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1380 // Move this block to the end of the function.
1381 MBB->moveAfter(&MF.back());
1390 if (!IsEmptyBlock(MBB) && MBB->pred_size() == 1 &&
1391 MF.getFunction()->optForSize()) {
1392 // Changing "Jcc foo; foo: jmp bar;" into "Jcc bar;" might change the branch
1393 // direction, thereby defeating careful block placement and regressing
1394 // performance. Therefore, only consider this for optsize functions.
1395 MachineInstr &TailCall = *MBB->getFirstNonDebugInstr();
1396 if (TII->isUnconditionalTailCall(TailCall)) {
1397 MachineBasicBlock *Pred = *MBB->pred_begin();
1398 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1399 SmallVector<MachineOperand, 4> PredCond;
1400 bool PredAnalyzable =
1401 !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true);
1403 if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB) {
1404 // The predecessor has a conditional branch to this block which consists
1405 // of only a tail call. Try to fold the tail call into the conditional
1407 if (TII->canMakeTailCallConditional(PredCond, TailCall)) {
1408 // TODO: It would be nice if analyzeBranch() could provide a pointer
1409 // to the branch insturction so replaceBranchWithTailCall() doesn't
1410 // have to search for it.
1411 TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall);
1413 Pred->removeSuccessor(MBB);
1418 // If the predecessor is falling through to this block, we could reverse
1419 // the branch condition and fold the tail call into that. However, after
1420 // that we might have to re-arrange the CFG to fall through to the other
1421 // block and there is a high risk of regressing code size rather than
1426 // Analyze the branch in the current block.
1427 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1428 SmallVector<MachineOperand, 4> CurCond;
1429 bool CurUnAnalyzable =
1430 TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1431 if (!CurUnAnalyzable) {
1432 // If the CFG for the prior block has extra edges, remove them.
1433 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1435 // If this is a two-way branch, and the FBB branches to this block, reverse
1436 // the condition so the single-basic-block loop is faster. Instead of:
1437 // Loop: xxx; jcc Out; jmp Loop
1439 // Loop: xxx; jncc Loop; jmp Out
1440 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1441 SmallVector<MachineOperand, 4> NewCond(CurCond);
1442 if (!TII->reverseBranchCondition(NewCond)) {
1443 DebugLoc dl = getBranchDebugLoc(*MBB);
1444 TII->removeBranch(*MBB);
1445 TII->insertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1448 goto ReoptimizeBlock;
1452 // If this branch is the only thing in its block, see if we can forward
1453 // other blocks across it.
1454 if (CurTBB && CurCond.empty() && !CurFBB &&
1455 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1456 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1457 DebugLoc dl = getBranchDebugLoc(*MBB);
1458 // This block may contain just an unconditional branch. Because there can
1459 // be 'non-branch terminators' in the block, try removing the branch and
1460 // then seeing if the block is empty.
1461 TII->removeBranch(*MBB);
1462 // If the only things remaining in the block are debug info, remove these
1463 // as well, so this will behave the same as an empty block in non-debug
1465 if (IsEmptyBlock(MBB)) {
1466 // Make the block empty, losing the debug info (we could probably
1467 // improve this in some cases.)
1468 MBB->erase(MBB->begin(), MBB->end());
1470 // If this block is just an unconditional branch to CurTBB, we can
1471 // usually completely eliminate the block. The only case we cannot
1472 // completely eliminate the block is when the block before this one
1473 // falls through into MBB and we can't understand the prior block's branch
1476 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1477 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1478 !PrevBB.isSuccessor(MBB)) {
1479 // If the prior block falls through into us, turn it into an
1480 // explicit branch to us to make updates simpler.
1481 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1482 PriorTBB != MBB && PriorFBB != MBB) {
1484 assert(PriorCond.empty() && !PriorFBB &&
1485 "Bad branch analysis");
1488 assert(!PriorFBB && "Machine CFG out of date!");
1491 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1492 TII->removeBranch(PrevBB);
1493 TII->insertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1496 // Iterate through all the predecessors, revectoring each in-turn.
1498 bool DidChange = false;
1499 bool HasBranchToSelf = false;
1500 while(PI != MBB->pred_size()) {
1501 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1503 // If this block has an uncond branch to itself, leave it.
1505 HasBranchToSelf = true;
1508 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1509 // If this change resulted in PMBB ending in a conditional
1510 // branch where both conditions go to the same destination,
1511 // change this to an unconditional branch (and fix the CFG).
1512 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1513 SmallVector<MachineOperand, 4> NewCurCond;
1514 bool NewCurUnAnalyzable = TII->analyzeBranch(
1515 *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true);
1516 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1517 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1518 TII->removeBranch(*PMBB);
1520 TII->insertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1523 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1528 // Change any jumptables to go to the new MBB.
1529 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1530 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1534 if (!HasBranchToSelf) return MadeChange;
1539 // Add the branch back if the block is more than just an uncond branch.
1540 TII->insertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1544 // If the prior block doesn't fall through into this block, and if this
1545 // block doesn't fall through into some other block, see if we can find a
1546 // place to move this block where a fall-through will happen.
1547 if (!PrevBB.canFallThrough()) {
1549 // Now we know that there was no fall-through into this block, check to
1550 // see if it has a fall-through into its successor.
1551 bool CurFallsThru = MBB->canFallThrough();
1553 if (!MBB->isEHPad()) {
1554 // Check all the predecessors of this block. If one of them has no fall
1555 // throughs, move this block right after it.
1556 for (MachineBasicBlock *PredBB : MBB->predecessors()) {
1557 // Analyze the branch at the end of the pred.
1558 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1559 SmallVector<MachineOperand, 4> PredCond;
1560 if (PredBB != MBB && !PredBB->canFallThrough() &&
1561 !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) &&
1562 (!CurFallsThru || !CurTBB || !CurFBB) &&
1563 (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1564 // If the current block doesn't fall through, just move it.
1565 // If the current block can fall through and does not end with a
1566 // conditional branch, we need to append an unconditional jump to
1567 // the (current) next block. To avoid a possible compile-time
1568 // infinite loop, move blocks only backward in this case.
1569 // Also, if there are already 2 branches here, we cannot add a third;
1570 // this means we have the case
1575 MachineBasicBlock *NextBB = &*std::next(MBB->getIterator());
1577 TII->insertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1579 MBB->moveAfter(PredBB);
1581 goto ReoptimizeBlock;
1586 if (!CurFallsThru) {
1587 // Check all successors to see if we can move this block before it.
1588 for (MachineBasicBlock *SuccBB : MBB->successors()) {
1589 // Analyze the branch at the end of the block before the succ.
1590 MachineFunction::iterator SuccPrev = --SuccBB->getIterator();
1592 // If this block doesn't already fall-through to that successor, and if
1593 // the succ doesn't already have a block that can fall through into it,
1594 // and if the successor isn't an EH destination, we can arrange for the
1595 // fallthrough to happen.
1596 if (SuccBB != MBB && &*SuccPrev != MBB &&
1597 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1598 !SuccBB->isEHPad()) {
1599 MBB->moveBefore(SuccBB);
1601 goto ReoptimizeBlock;
1605 // Okay, there is no really great place to put this block. If, however,
1606 // the block before this one would be a fall-through if this block were
1607 // removed, move this block to the end of the function. There is no real
1608 // advantage in "falling through" to an EH block, so we don't want to
1609 // perform this transformation for that case.
1611 // Also, Windows EH introduced the possibility of an arbitrary number of
1612 // successors to a given block. The analyzeBranch call does not consider
1613 // exception handling and so we can get in a state where a block
1614 // containing a call is followed by multiple EH blocks that would be
1615 // rotated infinitely at the end of the function if the transformation
1616 // below were performed for EH "FallThrough" blocks. Therefore, even if
1617 // that appears not to be happening anymore, we should assume that it is
1618 // possible and not remove the "!FallThrough()->isEHPad" condition below.
1619 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1620 SmallVector<MachineOperand, 4> PrevCond;
1621 if (FallThrough != MF.end() &&
1622 !FallThrough->isEHPad() &&
1623 !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1624 PrevBB.isSuccessor(&*FallThrough)) {
1625 MBB->moveAfter(&MF.back());
1635 //===----------------------------------------------------------------------===//
1636 // Hoist Common Code
1637 //===----------------------------------------------------------------------===//
1639 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1640 /// blocks to their common predecessor.
1641 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1642 bool MadeChange = false;
1643 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1644 MachineBasicBlock *MBB = &*I++;
1645 MadeChange |= HoistCommonCodeInSuccs(MBB);
1651 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1652 /// its 'true' successor.
1653 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1654 MachineBasicBlock *TrueBB) {
1655 for (MachineBasicBlock *SuccBB : BB->successors())
1656 if (SuccBB != TrueBB)
1661 template <class Container>
1662 static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
1664 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1665 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1672 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1673 /// in successors to. The location is usually just before the terminator,
1674 /// however if the terminator is a conditional branch and its previous
1675 /// instruction is the flag setting instruction, the previous instruction is
1676 /// the preferred location. This function also gathers uses and defs of the
1677 /// instructions from the insertion point to the end of the block. The data is
1678 /// used by HoistCommonCodeInSuccs to ensure safety.
1680 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1681 const TargetInstrInfo *TII,
1682 const TargetRegisterInfo *TRI,
1683 SmallSet<unsigned,4> &Uses,
1684 SmallSet<unsigned,4> &Defs) {
1685 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1686 if (!TII->isUnpredicatedTerminator(*Loc))
1689 for (const MachineOperand &MO : Loc->operands()) {
1692 unsigned Reg = MO.getReg();
1696 addRegAndItsAliases(Reg, TRI, Uses);
1699 // Don't try to hoist code in the rare case the terminator defines a
1700 // register that is later used.
1703 // If the terminator defines a register, make sure we don't hoist
1704 // the instruction whose def might be clobbered by the terminator.
1705 addRegAndItsAliases(Reg, TRI, Defs);
1711 if (Loc == MBB->begin())
1714 // The terminator is probably a conditional branch, try not to separate the
1715 // branch from condition setting instruction.
1716 MachineBasicBlock::iterator PI =
1717 skipDebugInstructionsBackward(std::prev(Loc), MBB->begin());
1720 for (const MachineOperand &MO : PI->operands()) {
1721 // If PI has a regmask operand, it is probably a call. Separate away.
1724 if (!MO.isReg() || MO.isUse())
1726 unsigned Reg = MO.getReg();
1729 if (Uses.count(Reg)) {
1735 // The condition setting instruction is not just before the conditional
1739 // Be conservative, don't insert instruction above something that may have
1740 // side-effects. And since it's potentially bad to separate flag setting
1741 // instruction from the conditional branch, just abort the optimization
1743 // Also avoid moving code above predicated instruction since it's hard to
1744 // reason about register liveness with predicated instruction.
1745 bool DontMoveAcrossStore = true;
1746 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI))
1750 // Find out what registers are live. Note this routine is ignoring other live
1751 // registers which are only used by instructions in successor blocks.
1752 for (const MachineOperand &MO : PI->operands()) {
1755 unsigned Reg = MO.getReg();
1759 addRegAndItsAliases(Reg, TRI, Uses);
1761 if (Uses.erase(Reg)) {
1762 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1763 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1764 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1767 addRegAndItsAliases(Reg, TRI, Defs);
1774 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1775 /// sequence at the start of the function, move the instructions before MBB
1776 /// terminator if it's legal.
1777 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1778 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1779 SmallVector<MachineOperand, 4> Cond;
1780 if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1783 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1785 // Malformed bcc? True and false blocks are the same?
1788 // Restrict the optimization to cases where MBB is the only predecessor,
1789 // it is an obvious win.
1790 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1793 // Find a suitable position to hoist the common instructions to. Also figure
1794 // out which registers are used or defined by instructions from the insertion
1795 // point to the end of the block.
1796 SmallSet<unsigned, 4> Uses, Defs;
1797 MachineBasicBlock::iterator Loc =
1798 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1799 if (Loc == MBB->end())
1802 bool HasDups = false;
1803 SmallVector<unsigned, 4> LocalDefs;
1804 SmallSet<unsigned, 4> LocalDefsSet;
1805 MachineBasicBlock::iterator TIB = TBB->begin();
1806 MachineBasicBlock::iterator FIB = FBB->begin();
1807 MachineBasicBlock::iterator TIE = TBB->end();
1808 MachineBasicBlock::iterator FIE = FBB->end();
1809 while (TIB != TIE && FIB != FIE) {
1810 // Skip dbg_value instructions. These do not count.
1811 TIB = skipDebugInstructionsForward(TIB, TIE);
1812 FIB = skipDebugInstructionsForward(FIB, FIE);
1813 if (TIB == TIE || FIB == FIE)
1816 if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead))
1819 if (TII->isPredicated(*TIB))
1820 // Hard to reason about register liveness with predicated instruction.
1824 for (MachineOperand &MO : TIB->operands()) {
1825 // Don't attempt to hoist instructions with register masks.
1826 if (MO.isRegMask()) {
1832 unsigned Reg = MO.getReg();
1836 if (Uses.count(Reg)) {
1837 // Avoid clobbering a register that's used by the instruction at
1838 // the point of insertion.
1843 if (Defs.count(Reg) && !MO.isDead()) {
1844 // Don't hoist the instruction if the def would be clobber by the
1845 // instruction at the point insertion. FIXME: This is overly
1846 // conservative. It should be possible to hoist the instructions
1847 // in BB2 in the following example:
1849 // r1, eflag = op1 r2, r3
1858 } else if (!LocalDefsSet.count(Reg)) {
1859 if (Defs.count(Reg)) {
1860 // Use is defined by the instruction at the point of insertion.
1865 if (MO.isKill() && Uses.count(Reg))
1866 // Kills a register that's read by the instruction at the point of
1867 // insertion. Remove the kill marker.
1868 MO.setIsKill(false);
1874 bool DontMoveAcrossStore = true;
1875 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1878 // Remove kills from LocalDefsSet, these registers had short live ranges.
1879 for (const MachineOperand &MO : TIB->operands()) {
1880 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1882 unsigned Reg = MO.getReg();
1883 if (!Reg || !LocalDefsSet.count(Reg))
1885 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1886 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1887 LocalDefsSet.erase(*AI);
1889 LocalDefsSet.erase(Reg);
1893 // Track local defs so we can update liveins.
1894 for (const MachineOperand &MO : TIB->operands()) {
1895 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1897 unsigned Reg = MO.getReg();
1898 if (!Reg || TargetRegisterInfo::isVirtualRegister(Reg))
1900 LocalDefs.push_back(Reg);
1901 addRegAndItsAliases(Reg, TRI, LocalDefsSet);
1912 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1913 FBB->erase(FBB->begin(), FIB);
1916 bool AddedLiveIns = false;
1917 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1918 unsigned Def = LocalDefs[i];
1919 if (LocalDefsSet.count(Def)) {
1920 TBB->addLiveIn(Def);
1921 FBB->addLiveIn(Def);
1922 AddedLiveIns = true;
1927 TBB->sortUniqueLiveIns();
1928 FBB->sortUniqueLiveIns();