1 //===- BranchFolding.cpp - Fold machine code branch instructions ----------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This pass forwards branches to unconditional branches to make them branch
10 // directly to the target block. This pass often results in dead MBB's, which
13 // Note that this pass must be run after register allocation, it cannot handle
14 // SSA form. It also must handle virtual registers for targets that emit virtual
17 //===----------------------------------------------------------------------===//
19 #include "BranchFolding.h"
20 #include "llvm/ADT/BitVector.h"
21 #include "llvm/ADT/DenseMap.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/ProfileSummaryInfo.h"
28 #include "llvm/CodeGen/Analysis.h"
29 #include "llvm/CodeGen/LivePhysRegs.h"
30 #include "llvm/CodeGen/MachineBasicBlock.h"
31 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
32 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineFunctionPass.h"
35 #include "llvm/CodeGen/MachineInstr.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineJumpTableInfo.h"
38 #include "llvm/CodeGen/MachineLoopInfo.h"
39 #include "llvm/CodeGen/MachineModuleInfo.h"
40 #include "llvm/CodeGen/MachineOperand.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/MachineSizeOpts.h"
43 #include "llvm/CodeGen/TargetInstrInfo.h"
44 #include "llvm/CodeGen/TargetOpcodes.h"
45 #include "llvm/CodeGen/TargetPassConfig.h"
46 #include "llvm/CodeGen/TargetRegisterInfo.h"
47 #include "llvm/CodeGen/TargetSubtargetInfo.h"
48 #include "llvm/IR/DebugInfoMetadata.h"
49 #include "llvm/IR/DebugLoc.h"
50 #include "llvm/IR/Function.h"
51 #include "llvm/InitializePasses.h"
52 #include "llvm/MC/LaneBitmask.h"
53 #include "llvm/MC/MCRegisterInfo.h"
54 #include "llvm/Pass.h"
55 #include "llvm/Support/BlockFrequency.h"
56 #include "llvm/Support/BranchProbability.h"
57 #include "llvm/Support/CommandLine.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/ErrorHandling.h"
60 #include "llvm/Support/raw_ostream.h"
61 #include "llvm/Target/TargetMachine.h"
70 #define DEBUG_TYPE "branch-folder"
72 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
73 STATISTIC(NumBranchOpts, "Number of branches optimized");
74 STATISTIC(NumTailMerge , "Number of block tails merged");
75 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
76 STATISTIC(NumTailCalls, "Number of tail calls optimized");
78 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
79 cl::init(cl::BOU_UNSET), cl::Hidden);
81 // Throttle for huge numbers of predecessors (compile speed problems)
82 static cl::opt<unsigned>
83 TailMergeThreshold("tail-merge-threshold",
84 cl::desc("Max number of predecessors to consider tail merging"),
85 cl::init(150), cl::Hidden);
87 // Heuristic for tail merging (and, inversely, tail duplication).
88 // TODO: This should be replaced with a target query.
89 static cl::opt<unsigned>
90 TailMergeSize("tail-merge-size",
91 cl::desc("Min number of instructions to consider tail merging"),
92 cl::init(3), cl::Hidden);
96 /// BranchFolderPass - Wrap branch folder in a machine function pass.
97 class BranchFolderPass : public MachineFunctionPass {
101 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
103 bool runOnMachineFunction(MachineFunction &MF) override;
105 void getAnalysisUsage(AnalysisUsage &AU) const override {
106 AU.addRequired<MachineBlockFrequencyInfo>();
107 AU.addRequired<MachineBranchProbabilityInfo>();
108 AU.addRequired<ProfileSummaryInfoWrapperPass>();
109 AU.addRequired<TargetPassConfig>();
110 MachineFunctionPass::getAnalysisUsage(AU);
114 } // end anonymous namespace
116 char BranchFolderPass::ID = 0;
118 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
120 INITIALIZE_PASS(BranchFolderPass, DEBUG_TYPE,
121 "Control Flow Optimizer", false, false)
123 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
124 if (skipFunction(MF.getFunction()))
127 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
128 // TailMerge can create jump into if branches that make CFG irreducible for
129 // HW that requires structurized CFG.
130 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
131 PassConfig->getEnableTailMerge();
132 BranchFolder::MBFIWrapper MBBFreqInfo(
133 getAnalysis<MachineBlockFrequencyInfo>());
134 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo,
135 getAnalysis<MachineBranchProbabilityInfo>(),
136 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI());
137 auto *MMIWP = getAnalysisIfAvailable<MachineModuleInfoWrapperPass>();
138 return Folder.OptimizeFunction(
139 MF, MF.getSubtarget().getInstrInfo(), MF.getSubtarget().getRegisterInfo(),
140 MMIWP ? &MMIWP->getMMI() : nullptr);
143 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
144 MBFIWrapper &FreqInfo,
145 const MachineBranchProbabilityInfo &ProbInfo,
146 ProfileSummaryInfo *PSI,
147 unsigned MinTailLength)
148 : EnableHoistCommonCode(CommonHoist), MinCommonTailLength(MinTailLength),
149 MBBFreqInfo(FreqInfo), MBPI(ProbInfo), PSI(PSI) {
150 if (MinCommonTailLength == 0)
151 MinCommonTailLength = TailMergeSize;
152 switch (FlagEnableTailMerge) {
153 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
154 case cl::BOU_TRUE: EnableTailMerge = true; break;
155 case cl::BOU_FALSE: EnableTailMerge = false; break;
159 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
160 assert(MBB->pred_empty() && "MBB must be dead!");
161 LLVM_DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
163 MachineFunction *MF = MBB->getParent();
164 // drop all successors.
165 while (!MBB->succ_empty())
166 MBB->removeSuccessor(MBB->succ_end()-1);
168 // Avoid matching if this pointer gets reused.
169 TriedMerging.erase(MBB);
171 // Update call site info.
172 std::for_each(MBB->begin(), MBB->end(), [MF](const MachineInstr &MI) {
173 if (MI.isCall(MachineInstr::IgnoreBundle))
174 MF->eraseCallSiteInfo(&MI);
178 EHScopeMembership.erase(MBB);
180 MLI->removeBlock(MBB);
183 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
184 const TargetInstrInfo *tii,
185 const TargetRegisterInfo *tri,
186 MachineModuleInfo *mmi,
187 MachineLoopInfo *mli, bool AfterPlacement) {
188 if (!tii) return false;
190 TriedMerging.clear();
192 MachineRegisterInfo &MRI = MF.getRegInfo();
193 AfterBlockPlacement = AfterPlacement;
200 UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF);
202 MRI.invalidateLiveness();
204 // Fix CFG. The later algorithms expect it to be right.
205 bool MadeChange = false;
206 for (MachineBasicBlock &MBB : MF) {
207 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
208 SmallVector<MachineOperand, 4> Cond;
209 if (!TII->analyzeBranch(MBB, TBB, FBB, Cond, true))
210 MadeChange |= MBB.CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
213 // Recalculate EH scope membership.
214 EHScopeMembership = getEHScopeMembership(MF);
216 bool MadeChangeThisIteration = true;
217 while (MadeChangeThisIteration) {
218 MadeChangeThisIteration = TailMergeBlocks(MF);
219 // No need to clean up if tail merging does not change anything after the
221 if (!AfterBlockPlacement || MadeChangeThisIteration)
222 MadeChangeThisIteration |= OptimizeBranches(MF);
223 if (EnableHoistCommonCode)
224 MadeChangeThisIteration |= HoistCommonCode(MF);
225 MadeChange |= MadeChangeThisIteration;
228 // See if any jump tables have become dead as the code generator
230 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
234 // Walk the function to find jump tables that are live.
235 BitVector JTIsLive(JTI->getJumpTables().size());
236 for (const MachineBasicBlock &BB : MF) {
237 for (const MachineInstr &I : BB)
238 for (const MachineOperand &Op : I.operands()) {
239 if (!Op.isJTI()) continue;
241 // Remember that this JT is live.
242 JTIsLive.set(Op.getIndex());
246 // Finally, remove dead jump tables. This happens when the
247 // indirect jump was unreachable (and thus deleted).
248 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
249 if (!JTIsLive.test(i)) {
250 JTI->RemoveJumpTable(i);
257 //===----------------------------------------------------------------------===//
258 // Tail Merging of Blocks
259 //===----------------------------------------------------------------------===//
261 /// HashMachineInstr - Compute a hash value for MI and its operands.
262 static unsigned HashMachineInstr(const MachineInstr &MI) {
263 unsigned Hash = MI.getOpcode();
264 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
265 const MachineOperand &Op = MI.getOperand(i);
267 // Merge in bits from the operand if easy. We can't use MachineOperand's
268 // hash_code here because it's not deterministic and we sort by hash value
270 unsigned OperandHash = 0;
271 switch (Op.getType()) {
272 case MachineOperand::MO_Register:
273 OperandHash = Op.getReg();
275 case MachineOperand::MO_Immediate:
276 OperandHash = Op.getImm();
278 case MachineOperand::MO_MachineBasicBlock:
279 OperandHash = Op.getMBB()->getNumber();
281 case MachineOperand::MO_FrameIndex:
282 case MachineOperand::MO_ConstantPoolIndex:
283 case MachineOperand::MO_JumpTableIndex:
284 OperandHash = Op.getIndex();
286 case MachineOperand::MO_GlobalAddress:
287 case MachineOperand::MO_ExternalSymbol:
288 // Global address / external symbol are too hard, don't bother, but do
289 // pull in the offset.
290 OperandHash = Op.getOffset();
296 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
301 /// HashEndOfMBB - Hash the last instruction in the MBB.
302 static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) {
303 MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr();
307 return HashMachineInstr(*I);
310 /// Whether MI should be counted as an instruction when calculating common tail.
311 static bool countsAsInstruction(const MachineInstr &MI) {
312 return !(MI.isDebugInstr() || MI.isCFIInstruction());
315 /// Iterate backwards from the given iterator \p I, towards the beginning of the
316 /// block. If a MI satisfying 'countsAsInstruction' is found, return an iterator
317 /// pointing to that MI. If no such MI is found, return the end iterator.
318 static MachineBasicBlock::iterator
319 skipBackwardPastNonInstructions(MachineBasicBlock::iterator I,
320 MachineBasicBlock *MBB) {
321 while (I != MBB->begin()) {
323 if (countsAsInstruction(*I))
329 /// Given two machine basic blocks, return the number of instructions they
330 /// actually have in common together at their end. If a common tail is found (at
331 /// least by one instruction), then iterators for the first shared instruction
332 /// in each block are returned as well.
334 /// Non-instructions according to countsAsInstruction are ignored.
335 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
336 MachineBasicBlock *MBB2,
337 MachineBasicBlock::iterator &I1,
338 MachineBasicBlock::iterator &I2) {
339 MachineBasicBlock::iterator MBBI1 = MBB1->end();
340 MachineBasicBlock::iterator MBBI2 = MBB2->end();
342 unsigned TailLen = 0;
344 MBBI1 = skipBackwardPastNonInstructions(MBBI1, MBB1);
345 MBBI2 = skipBackwardPastNonInstructions(MBBI2, MBB2);
346 if (MBBI1 == MBB1->end() || MBBI2 == MBB2->end())
348 if (!MBBI1->isIdenticalTo(*MBBI2) ||
349 // FIXME: This check is dubious. It's used to get around a problem where
350 // people incorrectly expect inline asm directives to remain in the same
351 // relative order. This is untenable because normal compiler
352 // optimizations (like this one) may reorder and/or merge these
354 MBBI1->isInlineAsm()) {
365 void BranchFolder::replaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
366 MachineBasicBlock &NewDest) {
368 // OldInst should always point to an instruction.
369 MachineBasicBlock &OldMBB = *OldInst->getParent();
371 LiveRegs.addLiveOuts(OldMBB);
372 // Move backward to the place where will insert the jump.
373 MachineBasicBlock::iterator I = OldMBB.end();
376 LiveRegs.stepBackward(*I);
377 } while (I != OldInst);
379 // Merging the tails may have switched some undef operand to non-undef ones.
380 // Add IMPLICIT_DEFS into OldMBB as necessary to have a definition of the
382 for (MachineBasicBlock::RegisterMaskPair P : NewDest.liveins()) {
383 // We computed the liveins with computeLiveIn earlier and should only see
385 assert(P.LaneMask == LaneBitmask::getAll() &&
386 "Can only handle full register.");
387 MCPhysReg Reg = P.PhysReg;
388 if (!LiveRegs.available(*MRI, Reg))
391 BuildMI(OldMBB, OldInst, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Reg);
395 TII->ReplaceTailWithBranchTo(OldInst, &NewDest);
399 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
400 MachineBasicBlock::iterator BBI1,
401 const BasicBlock *BB) {
402 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
405 MachineFunction &MF = *CurMBB.getParent();
407 // Create the fall-through block.
408 MachineFunction::iterator MBBI = CurMBB.getIterator();
409 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(BB);
410 CurMBB.getParent()->insert(++MBBI, NewMBB);
412 // Move all the successors of this block to the specified block.
413 NewMBB->transferSuccessors(&CurMBB);
415 // Add an edge from CurMBB to NewMBB for the fall-through.
416 CurMBB.addSuccessor(NewMBB);
418 // Splice the code over.
419 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
421 // NewMBB belongs to the same loop as CurMBB.
423 if (MachineLoop *ML = MLI->getLoopFor(&CurMBB))
424 ML->addBasicBlockToLoop(NewMBB, MLI->getBase());
426 // NewMBB inherits CurMBB's block frequency.
427 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
430 computeAndAddLiveIns(LiveRegs, *NewMBB);
432 // Add the new block to the EH scope.
433 const auto &EHScopeI = EHScopeMembership.find(&CurMBB);
434 if (EHScopeI != EHScopeMembership.end()) {
435 auto n = EHScopeI->second;
436 EHScopeMembership[NewMBB] = n;
442 /// EstimateRuntime - Make a rough estimate for how long it will take to run
443 /// the specified code.
444 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
445 MachineBasicBlock::iterator E) {
447 for (; I != E; ++I) {
448 if (!countsAsInstruction(*I))
452 else if (I->mayLoadOrStore())
460 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
461 // branches temporarily for tail merging). In the case where CurMBB ends
462 // with a conditional branch to the next block, optimize by reversing the
463 // test and conditionally branching to SuccMBB instead.
464 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
465 const TargetInstrInfo *TII) {
466 MachineFunction *MF = CurMBB->getParent();
467 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
468 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
469 SmallVector<MachineOperand, 4> Cond;
470 DebugLoc dl = CurMBB->findBranchDebugLoc();
471 if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
472 MachineBasicBlock *NextBB = &*I;
473 if (TBB == NextBB && !Cond.empty() && !FBB) {
474 if (!TII->reverseBranchCondition(Cond)) {
475 TII->removeBranch(*CurMBB);
476 TII->insertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
481 TII->insertBranch(*CurMBB, SuccBB, nullptr,
482 SmallVector<MachineOperand, 0>(), dl);
486 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
487 if (getHash() < o.getHash())
489 if (getHash() > o.getHash())
491 if (getBlock()->getNumber() < o.getBlock()->getNumber())
493 if (getBlock()->getNumber() > o.getBlock()->getNumber())
495 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
496 // an object with itself.
497 #ifndef _GLIBCXX_DEBUG
498 llvm_unreachable("Predecessor appears twice");
505 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
506 auto I = MergedBBFreq.find(MBB);
508 if (I != MergedBBFreq.end())
511 return MBFI.getBlockFreq(MBB);
514 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
516 MergedBBFreq[MBB] = F;
520 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
521 const MachineBasicBlock *MBB) const {
522 return MBFI.printBlockFreq(OS, getBlockFreq(MBB));
526 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
527 const BlockFrequency Freq) const {
528 return MBFI.printBlockFreq(OS, Freq);
531 void BranchFolder::MBFIWrapper::view(const Twine &Name, bool isSimple) {
532 MBFI.view(Name, isSimple);
536 BranchFolder::MBFIWrapper::getEntryFreq() const {
537 return MBFI.getEntryFreq();
540 /// CountTerminators - Count the number of terminators in the given
541 /// block and set I to the position of the first non-terminator, if there
542 /// is one, or MBB->end() otherwise.
543 static unsigned CountTerminators(MachineBasicBlock *MBB,
544 MachineBasicBlock::iterator &I) {
546 unsigned NumTerms = 0;
548 if (I == MBB->begin()) {
553 if (!I->isTerminator()) break;
559 /// A no successor, non-return block probably ends in unreachable and is cold.
560 /// Also consider a block that ends in an indirect branch to be a return block,
561 /// since many targets use plain indirect branches to return.
562 static bool blockEndsInUnreachable(const MachineBasicBlock *MBB) {
563 if (!MBB->succ_empty())
567 return !(MBB->back().isReturn() || MBB->back().isIndirectBranch());
570 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
571 /// and decide if it would be profitable to merge those tails. Return the
572 /// length of the common tail and iterators to the first common instruction
574 /// MBB1, MBB2 The blocks to check
575 /// MinCommonTailLength Minimum size of tail block to be merged.
576 /// CommonTailLen Out parameter to record the size of the shared tail between
578 /// I1, I2 Iterator references that will be changed to point to the first
579 /// instruction in the common tail shared by MBB1,MBB2
580 /// SuccBB A common successor of MBB1, MBB2 which are in a canonical form
581 /// relative to SuccBB
582 /// PredBB The layout predecessor of SuccBB, if any.
583 /// EHScopeMembership map from block to EH scope #.
584 /// AfterPlacement True if we are merging blocks after layout. Stricter
585 /// thresholds apply to prevent undoing tail-duplication.
587 ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2,
588 unsigned MinCommonTailLength, unsigned &CommonTailLen,
589 MachineBasicBlock::iterator &I1,
590 MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB,
591 MachineBasicBlock *PredBB,
592 DenseMap<const MachineBasicBlock *, int> &EHScopeMembership,
594 BranchFolder::MBFIWrapper &MBBFreqInfo,
595 ProfileSummaryInfo *PSI) {
596 // It is never profitable to tail-merge blocks from two different EH scopes.
597 if (!EHScopeMembership.empty()) {
598 auto EHScope1 = EHScopeMembership.find(MBB1);
599 assert(EHScope1 != EHScopeMembership.end());
600 auto EHScope2 = EHScopeMembership.find(MBB2);
601 assert(EHScope2 != EHScopeMembership.end());
602 if (EHScope1->second != EHScope2->second)
606 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
607 if (CommonTailLen == 0)
609 LLVM_DEBUG(dbgs() << "Common tail length of " << printMBBReference(*MBB1)
610 << " and " << printMBBReference(*MBB2) << " is "
611 << CommonTailLen << '\n');
613 // Move the iterators to the beginning of the MBB if we only got debug
614 // instructions before the tail. This is to avoid splitting a block when we
615 // only got debug instructions before the tail (to be invariant on -g).
616 if (skipDebugInstructionsForward(MBB1->begin(), MBB1->end()) == I1)
618 if (skipDebugInstructionsForward(MBB2->begin(), MBB2->end()) == I2)
621 bool FullBlockTail1 = I1 == MBB1->begin();
622 bool FullBlockTail2 = I2 == MBB2->begin();
624 // It's almost always profitable to merge any number of non-terminator
625 // instructions with the block that falls through into the common successor.
626 // This is true only for a single successor. For multiple successors, we are
627 // trading a conditional branch for an unconditional one.
628 // TODO: Re-visit successor size for non-layout tail merging.
629 if ((MBB1 == PredBB || MBB2 == PredBB) &&
630 (!AfterPlacement || MBB1->succ_size() == 1)) {
631 MachineBasicBlock::iterator I;
632 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
633 if (CommonTailLen > NumTerms)
637 // If these are identical non-return blocks with no successors, merge them.
638 // Such blocks are typically cold calls to noreturn functions like abort, and
639 // are unlikely to become a fallthrough target after machine block placement.
640 // Tail merging these blocks is unlikely to create additional unconditional
641 // branches, and will reduce the size of this cold code.
642 if (FullBlockTail1 && FullBlockTail2 &&
643 blockEndsInUnreachable(MBB1) && blockEndsInUnreachable(MBB2))
646 // If one of the blocks can be completely merged and happens to be in
647 // a position where the other could fall through into it, merge any number
648 // of instructions, because it can be done without a branch.
649 // TODO: If the blocks are not adjacent, move one of them so that they are?
650 if (MBB1->isLayoutSuccessor(MBB2) && FullBlockTail2)
652 if (MBB2->isLayoutSuccessor(MBB1) && FullBlockTail1)
655 // If both blocks are identical and end in a branch, merge them unless they
656 // both have a fallthrough predecessor and successor.
657 // We can only do this after block placement because it depends on whether
658 // there are fallthroughs, and we don't know until after layout.
659 if (AfterPlacement && FullBlockTail1 && FullBlockTail2) {
660 auto BothFallThrough = [](MachineBasicBlock *MBB) {
661 if (MBB->succ_size() != 0 && !MBB->canFallThrough())
663 MachineFunction::iterator I(MBB);
664 MachineFunction *MF = MBB->getParent();
665 return (MBB != &*MF->begin()) && std::prev(I)->canFallThrough();
667 if (!BothFallThrough(MBB1) || !BothFallThrough(MBB2))
671 // If both blocks have an unconditional branch temporarily stripped out,
672 // count that as an additional common instruction for the following
673 // heuristics. This heuristic is only accurate for single-succ blocks, so to
674 // make sure that during layout merging and duplicating don't crash, we check
675 // for that when merging during layout.
676 unsigned EffectiveTailLen = CommonTailLen;
677 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
678 (MBB1->succ_size() == 1 || !AfterPlacement) &&
679 !MBB1->back().isBarrier() &&
680 !MBB2->back().isBarrier())
683 // Check if the common tail is long enough to be worthwhile.
684 if (EffectiveTailLen >= MinCommonTailLength)
687 // If we are optimizing for code size, 2 instructions in common is enough if
688 // we don't have to split a block. At worst we will be introducing 1 new
689 // branch instruction, which is likely to be smaller than the 2
690 // instructions that would be deleted in the merge.
691 MachineFunction *MF = MBB1->getParent();
693 MF->getFunction().hasOptSize() ||
694 (llvm::shouldOptimizeForSize(MBB1, PSI, &MBBFreqInfo.getMBFI()) &&
695 llvm::shouldOptimizeForSize(MBB2, PSI, &MBBFreqInfo.getMBFI()));
696 return EffectiveTailLen >= 2 && OptForSize &&
697 (FullBlockTail1 || FullBlockTail2);
700 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
701 unsigned MinCommonTailLength,
702 MachineBasicBlock *SuccBB,
703 MachineBasicBlock *PredBB) {
704 unsigned maxCommonTailLength = 0U;
706 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
707 MPIterator HighestMPIter = std::prev(MergePotentials.end());
708 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
709 B = MergePotentials.begin();
710 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
711 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
712 unsigned CommonTailLen;
713 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
715 CommonTailLen, TrialBBI1, TrialBBI2,
718 AfterBlockPlacement, MBBFreqInfo, PSI)) {
719 if (CommonTailLen > maxCommonTailLength) {
721 maxCommonTailLength = CommonTailLen;
722 HighestMPIter = CurMPIter;
723 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
725 if (HighestMPIter == CurMPIter &&
726 CommonTailLen == maxCommonTailLength)
727 SameTails.push_back(SameTailElt(I, TrialBBI2));
733 return maxCommonTailLength;
736 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
737 MachineBasicBlock *SuccBB,
738 MachineBasicBlock *PredBB) {
739 MPIterator CurMPIter, B;
740 for (CurMPIter = std::prev(MergePotentials.end()),
741 B = MergePotentials.begin();
742 CurMPIter->getHash() == CurHash; --CurMPIter) {
743 // Put the unconditional branch back, if we need one.
744 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
745 if (SuccBB && CurMBB != PredBB)
746 FixTail(CurMBB, SuccBB, TII);
750 if (CurMPIter->getHash() != CurHash)
752 MergePotentials.erase(CurMPIter, MergePotentials.end());
755 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
756 MachineBasicBlock *SuccBB,
757 unsigned maxCommonTailLength,
758 unsigned &commonTailIndex) {
760 unsigned TimeEstimate = ~0U;
761 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
762 // Use PredBB if possible; that doesn't require a new branch.
763 if (SameTails[i].getBlock() == PredBB) {
767 // Otherwise, make a (fairly bogus) choice based on estimate of
768 // how long it will take the various blocks to execute.
769 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
770 SameTails[i].getTailStartPos());
771 if (t <= TimeEstimate) {
777 MachineBasicBlock::iterator BBI =
778 SameTails[commonTailIndex].getTailStartPos();
779 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
781 LLVM_DEBUG(dbgs() << "\nSplitting " << printMBBReference(*MBB) << ", size "
782 << maxCommonTailLength);
784 // If the split block unconditionally falls-thru to SuccBB, it will be
785 // merged. In control flow terms it should then take SuccBB's name. e.g. If
786 // SuccBB is an inner loop, the common tail is still part of the inner loop.
787 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
788 SuccBB->getBasicBlock() : MBB->getBasicBlock();
789 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
791 LLVM_DEBUG(dbgs() << "... failed!");
795 SameTails[commonTailIndex].setBlock(newMBB);
796 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
798 // If we split PredBB, newMBB is the new predecessor.
806 mergeOperations(MachineBasicBlock::iterator MBBIStartPos,
807 MachineBasicBlock &MBBCommon) {
808 MachineBasicBlock *MBB = MBBIStartPos->getParent();
809 // Note CommonTailLen does not necessarily matches the size of
810 // the common BB nor all its instructions because of debug
811 // instructions differences.
812 unsigned CommonTailLen = 0;
813 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
816 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
817 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
818 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
819 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
821 while (CommonTailLen--) {
822 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
825 if (!countsAsInstruction(*MBBI)) {
830 while ((MBBICommon != MBBIECommon) && !countsAsInstruction(*MBBICommon))
833 assert(MBBICommon != MBBIECommon &&
834 "Reached BB end within common tail length!");
835 assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!");
837 // Merge MMOs from memory operations in the common block.
838 if (MBBICommon->mayLoadOrStore())
839 MBBICommon->cloneMergedMemRefs(*MBB->getParent(), {&*MBBICommon, &*MBBI});
840 // Drop undef flags if they aren't present in all merged instructions.
841 for (unsigned I = 0, E = MBBICommon->getNumOperands(); I != E; ++I) {
842 MachineOperand &MO = MBBICommon->getOperand(I);
843 if (MO.isReg() && MO.isUndef()) {
844 const MachineOperand &OtherMO = MBBI->getOperand(I);
845 if (!OtherMO.isUndef())
846 MO.setIsUndef(false);
855 void BranchFolder::mergeCommonTails(unsigned commonTailIndex) {
856 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
858 std::vector<MachineBasicBlock::iterator> NextCommonInsts(SameTails.size());
859 for (unsigned int i = 0 ; i != SameTails.size() ; ++i) {
860 if (i != commonTailIndex) {
861 NextCommonInsts[i] = SameTails[i].getTailStartPos();
862 mergeOperations(SameTails[i].getTailStartPos(), *MBB);
864 assert(SameTails[i].getTailStartPos() == MBB->begin() &&
865 "MBB is not a common tail only block");
869 for (auto &MI : *MBB) {
870 if (!countsAsInstruction(MI))
872 DebugLoc DL = MI.getDebugLoc();
873 for (unsigned int i = 0 ; i < NextCommonInsts.size() ; i++) {
874 if (i == commonTailIndex)
877 auto &Pos = NextCommonInsts[i];
878 assert(Pos != SameTails[i].getBlock()->end() &&
879 "Reached BB end within common tail");
880 while (!countsAsInstruction(*Pos)) {
882 assert(Pos != SameTails[i].getBlock()->end() &&
883 "Reached BB end within common tail");
885 assert(MI.isIdenticalTo(*Pos) && "Expected matching MIIs!");
886 DL = DILocation::getMergedLocation(DL, Pos->getDebugLoc());
887 NextCommonInsts[i] = ++Pos;
893 LivePhysRegs NewLiveIns(*TRI);
894 computeLiveIns(NewLiveIns, *MBB);
897 // The flag merging may lead to some register uses no longer using the
898 // <undef> flag, add IMPLICIT_DEFs in the predecessors as necessary.
899 for (MachineBasicBlock *Pred : MBB->predecessors()) {
901 LiveRegs.addLiveOuts(*Pred);
902 MachineBasicBlock::iterator InsertBefore = Pred->getFirstTerminator();
903 for (unsigned Reg : NewLiveIns) {
904 if (!LiveRegs.available(*MRI, Reg))
907 BuildMI(*Pred, InsertBefore, DL, TII->get(TargetOpcode::IMPLICIT_DEF),
913 addLiveIns(*MBB, NewLiveIns);
917 // See if any of the blocks in MergePotentials (which all have SuccBB as a
918 // successor, or all have no successor if it is null) can be tail-merged.
919 // If there is a successor, any blocks in MergePotentials that are not
920 // tail-merged and are not immediately before Succ must have an unconditional
921 // branch to Succ added (but the predecessor/successor lists need no
922 // adjustment). The lone predecessor of Succ that falls through into Succ,
923 // if any, is given in PredBB.
924 // MinCommonTailLength - Except for the special cases below, tail-merge if
925 // there are at least this many instructions in common.
926 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
927 MachineBasicBlock *PredBB,
928 unsigned MinCommonTailLength) {
929 bool MadeChange = false;
932 dbgs() << "\nTryTailMergeBlocks: ";
933 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) dbgs()
934 << printMBBReference(*MergePotentials[i].getBlock())
935 << (i == e - 1 ? "" : ", ");
936 dbgs() << "\n"; if (SuccBB) {
937 dbgs() << " with successor " << printMBBReference(*SuccBB) << '\n';
939 dbgs() << " which has fall-through from "
940 << printMBBReference(*PredBB) << "\n";
941 } dbgs() << "Looking for common tails of at least "
942 << MinCommonTailLength << " instruction"
943 << (MinCommonTailLength == 1 ? "" : "s") << '\n';);
945 // Sort by hash value so that blocks with identical end sequences sort
947 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
949 // Walk through equivalence sets looking for actual exact matches.
950 while (MergePotentials.size() > 1) {
951 unsigned CurHash = MergePotentials.back().getHash();
953 // Build SameTails, identifying the set of blocks with this hash code
954 // and with the maximum number of instructions in common.
955 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
959 // If we didn't find any pair that has at least MinCommonTailLength
960 // instructions in common, remove all blocks with this hash code and retry.
961 if (SameTails.empty()) {
962 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
966 // If one of the blocks is the entire common tail (and not the entry
967 // block, which we can't jump to), we can treat all blocks with this same
968 // tail at once. Use PredBB if that is one of the possibilities, as that
969 // will not introduce any extra branches.
970 MachineBasicBlock *EntryBB =
971 &MergePotentials.front().getBlock()->getParent()->front();
972 unsigned commonTailIndex = SameTails.size();
973 // If there are two blocks, check to see if one can be made to fall through
975 if (SameTails.size() == 2 &&
976 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
977 SameTails[1].tailIsWholeBlock())
979 else if (SameTails.size() == 2 &&
980 SameTails[1].getBlock()->isLayoutSuccessor(
981 SameTails[0].getBlock()) &&
982 SameTails[0].tailIsWholeBlock())
985 // Otherwise just pick one, favoring the fall-through predecessor if
987 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
988 MachineBasicBlock *MBB = SameTails[i].getBlock();
989 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
995 if (SameTails[i].tailIsWholeBlock())
1000 if (commonTailIndex == SameTails.size() ||
1001 (SameTails[commonTailIndex].getBlock() == PredBB &&
1002 !SameTails[commonTailIndex].tailIsWholeBlock())) {
1003 // None of the blocks consist entirely of the common tail.
1004 // Split a block so that one does.
1005 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
1006 maxCommonTailLength, commonTailIndex)) {
1007 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
1012 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
1014 // Recompute common tail MBB's edge weights and block frequency.
1015 setCommonTailEdgeWeights(*MBB);
1017 // Merge debug locations, MMOs and undef flags across identical instructions
1019 mergeCommonTails(commonTailIndex);
1021 // MBB is common tail. Adjust all other BB's to jump to this one.
1022 // Traversal must be forwards so erases work.
1023 LLVM_DEBUG(dbgs() << "\nUsing common tail in " << printMBBReference(*MBB)
1025 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
1026 if (commonTailIndex == i)
1028 LLVM_DEBUG(dbgs() << printMBBReference(*SameTails[i].getBlock())
1029 << (i == e - 1 ? "" : ", "));
1030 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
1031 replaceTailWithBranchTo(SameTails[i].getTailStartPos(), *MBB);
1032 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
1033 MergePotentials.erase(SameTails[i].getMPIter());
1035 LLVM_DEBUG(dbgs() << "\n");
1036 // We leave commonTailIndex in the worklist in case there are other blocks
1037 // that match it with a smaller number of instructions.
1043 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
1044 bool MadeChange = false;
1045 if (!EnableTailMerge)
1048 // First find blocks with no successors.
1049 // Block placement may create new tail merging opportunities for these blocks.
1050 MergePotentials.clear();
1051 for (MachineBasicBlock &MBB : MF) {
1052 if (MergePotentials.size() == TailMergeThreshold)
1054 if (!TriedMerging.count(&MBB) && MBB.succ_empty())
1055 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB));
1058 // If this is a large problem, avoid visiting the same basic blocks
1060 if (MergePotentials.size() == TailMergeThreshold)
1061 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1062 TriedMerging.insert(MergePotentials[i].getBlock());
1064 // See if we can do any tail merging on those.
1065 if (MergePotentials.size() >= 2)
1066 MadeChange |= TryTailMergeBlocks(nullptr, nullptr, MinCommonTailLength);
1068 // Look at blocks (IBB) with multiple predecessors (PBB).
1069 // We change each predecessor to a canonical form, by
1070 // (1) temporarily removing any unconditional branch from the predecessor
1072 // (2) alter conditional branches so they branch to the other block
1073 // not IBB; this may require adding back an unconditional branch to IBB
1074 // later, where there wasn't one coming in. E.g.
1076 // fallthrough to QBB
1079 // with a conceptual B to IBB after that, which never actually exists.
1080 // With those changes, we see whether the predecessors' tails match,
1081 // and merge them if so. We change things out of canonical form and
1082 // back to the way they were later in the process. (OptimizeBranches
1083 // would undo some of this, but we can't use it, because we'd get into
1084 // a compile-time infinite loop repeatedly doing and undoing the same
1085 // transformations.)
1087 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1089 if (I->pred_size() < 2) continue;
1090 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
1091 MachineBasicBlock *IBB = &*I;
1092 MachineBasicBlock *PredBB = &*std::prev(I);
1093 MergePotentials.clear();
1096 // Bail if merging after placement and IBB is the loop header because
1097 // -- If merging predecessors that belong to the same loop as IBB, the
1098 // common tail of merged predecessors may become the loop top if block
1099 // placement is called again and the predecessors may branch to this common
1100 // tail and require more branches. This can be relaxed if
1101 // MachineBlockPlacement::findBestLoopTop is more flexible.
1102 // --If merging predecessors that do not belong to the same loop as IBB, the
1103 // loop info of IBB's loop and the other loops may be affected. Calling the
1104 // block placement again may make big change to the layout and eliminate the
1105 // reason to do tail merging here.
1106 if (AfterBlockPlacement && MLI) {
1107 ML = MLI->getLoopFor(IBB);
1108 if (ML && IBB == ML->getHeader())
1112 for (MachineBasicBlock *PBB : I->predecessors()) {
1113 if (MergePotentials.size() == TailMergeThreshold)
1116 if (TriedMerging.count(PBB))
1119 // Skip blocks that loop to themselves, can't tail merge these.
1123 // Visit each predecessor only once.
1124 if (!UniquePreds.insert(PBB).second)
1127 // Skip blocks which may jump to a landing pad. Can't tail merge these.
1128 if (PBB->hasEHPadSuccessor())
1131 // After block placement, only consider predecessors that belong to the
1132 // same loop as IBB. The reason is the same as above when skipping loop
1134 if (AfterBlockPlacement && MLI)
1135 if (ML != MLI->getLoopFor(PBB))
1138 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1139 SmallVector<MachineOperand, 4> Cond;
1140 if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) {
1141 // Failing case: IBB is the target of a cbr, and we cannot reverse the
1143 SmallVector<MachineOperand, 4> NewCond(Cond);
1144 if (!Cond.empty() && TBB == IBB) {
1145 if (TII->reverseBranchCondition(NewCond))
1147 // This is the QBB case described above
1149 auto Next = ++PBB->getIterator();
1150 if (Next != MF.end())
1155 // Remove the unconditional branch at the end, if any.
1156 if (TBB && (Cond.empty() || FBB)) {
1157 DebugLoc dl = PBB->findBranchDebugLoc();
1158 TII->removeBranch(*PBB);
1160 // reinsert conditional branch only, for now
1161 TII->insertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1165 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB));
1169 // If this is a large problem, avoid visiting the same basic blocks multiple
1171 if (MergePotentials.size() == TailMergeThreshold)
1172 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1173 TriedMerging.insert(MergePotentials[i].getBlock());
1175 if (MergePotentials.size() >= 2)
1176 MadeChange |= TryTailMergeBlocks(IBB, PredBB, MinCommonTailLength);
1178 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1179 // result of removing blocks in TryTailMergeBlocks.
1180 PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks
1181 if (MergePotentials.size() == 1 &&
1182 MergePotentials.begin()->getBlock() != PredBB)
1183 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1189 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1190 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1191 BlockFrequency AccumulatedMBBFreq;
1193 // Aggregate edge frequency of successor edge j:
1194 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1195 // where bb is a basic block that is in SameTails.
1196 for (const auto &Src : SameTails) {
1197 const MachineBasicBlock *SrcMBB = Src.getBlock();
1198 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1199 AccumulatedMBBFreq += BlockFreq;
1201 // It is not necessary to recompute edge weights if TailBB has less than two
1203 if (TailMBB.succ_size() <= 1)
1206 auto EdgeFreq = EdgeFreqLs.begin();
1208 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1209 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1210 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1213 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1215 if (TailMBB.succ_size() <= 1)
1219 std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0))
1221 auto EdgeFreq = EdgeFreqLs.begin();
1223 if (SumEdgeFreq > 0) {
1224 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1225 SuccI != SuccE; ++SuccI, ++EdgeFreq) {
1226 auto Prob = BranchProbability::getBranchProbability(
1227 EdgeFreq->getFrequency(), SumEdgeFreq);
1228 TailMBB.setSuccProbability(SuccI, Prob);
1233 //===----------------------------------------------------------------------===//
1234 // Branch Optimization
1235 //===----------------------------------------------------------------------===//
1237 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1238 bool MadeChange = false;
1240 // Make sure blocks are numbered in order
1241 MF.RenumberBlocks();
1242 // Renumbering blocks alters EH scope membership, recalculate it.
1243 EHScopeMembership = getEHScopeMembership(MF);
1245 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1247 MachineBasicBlock *MBB = &*I++;
1248 MadeChange |= OptimizeBlock(MBB);
1250 // If it is dead, remove it.
1251 if (MBB->pred_empty()) {
1252 RemoveDeadBlock(MBB);
1261 // Blocks should be considered empty if they contain only debug info;
1262 // else the debug info would affect codegen.
1263 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1264 return MBB->getFirstNonDebugInstr() == MBB->end();
1267 // Blocks with only debug info and branches should be considered the same
1268 // as blocks with only branches.
1269 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1270 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1271 assert(I != MBB->end() && "empty block!");
1272 return I->isBranch();
1275 /// IsBetterFallthrough - Return true if it would be clearly better to
1276 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1277 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1278 /// result in infinite loops.
1279 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1280 MachineBasicBlock *MBB2) {
1281 assert(MBB1 && MBB2 && "Unknown MachineBasicBlock");
1283 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1284 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1285 // optimize branches that branch to either a return block or an assert block
1286 // into a fallthrough to the return.
1287 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1288 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1289 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1292 // If there is a clear successor ordering we make sure that one block
1293 // will fall through to the next
1294 if (MBB1->isSuccessor(MBB2)) return true;
1295 if (MBB2->isSuccessor(MBB1)) return false;
1297 return MBB2I->isCall() && !MBB1I->isCall();
1300 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1301 /// instructions on the block.
1302 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1303 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1304 if (I != MBB.end() && I->isBranch())
1305 return I->getDebugLoc();
1309 static void copyDebugInfoToPredecessor(const TargetInstrInfo *TII,
1310 MachineBasicBlock &MBB,
1311 MachineBasicBlock &PredMBB) {
1312 auto InsertBefore = PredMBB.getFirstTerminator();
1313 for (MachineInstr &MI : MBB.instrs())
1314 if (MI.isDebugInstr()) {
1315 TII->duplicate(PredMBB, InsertBefore, MI);
1316 LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to pred: "
1321 static void copyDebugInfoToSuccessor(const TargetInstrInfo *TII,
1322 MachineBasicBlock &MBB,
1323 MachineBasicBlock &SuccMBB) {
1324 auto InsertBefore = SuccMBB.SkipPHIsAndLabels(SuccMBB.begin());
1325 for (MachineInstr &MI : MBB.instrs())
1326 if (MI.isDebugInstr()) {
1327 TII->duplicate(SuccMBB, InsertBefore, MI);
1328 LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to succ: "
1333 // Try to salvage DBG_VALUE instructions from an otherwise empty block. If such
1334 // a basic block is removed we would lose the debug information unless we have
1335 // copied the information to a predecessor/successor.
1337 // TODO: This function only handles some simple cases. An alternative would be
1338 // to run a heavier analysis, such as the LiveDebugValues pass, before we do
1340 static void salvageDebugInfoFromEmptyBlock(const TargetInstrInfo *TII,
1341 MachineBasicBlock &MBB) {
1342 assert(IsEmptyBlock(&MBB) && "Expected an empty block (except debug info).");
1343 // If this MBB is the only predecessor of a successor it is legal to copy
1344 // DBG_VALUE instructions to the beginning of the successor.
1345 for (MachineBasicBlock *SuccBB : MBB.successors())
1346 if (SuccBB->pred_size() == 1)
1347 copyDebugInfoToSuccessor(TII, MBB, *SuccBB);
1348 // If this MBB is the only successor of a predecessor it is legal to copy the
1349 // DBG_VALUE instructions to the end of the predecessor (just before the
1350 // terminators, assuming that the terminator isn't affecting the DBG_VALUE).
1351 for (MachineBasicBlock *PredBB : MBB.predecessors())
1352 if (PredBB->succ_size() == 1)
1353 copyDebugInfoToPredecessor(TII, MBB, *PredBB);
1356 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1357 bool MadeChange = false;
1358 MachineFunction &MF = *MBB->getParent();
1361 MachineFunction::iterator FallThrough = MBB->getIterator();
1364 // Make sure MBB and FallThrough belong to the same EH scope.
1365 bool SameEHScope = true;
1366 if (!EHScopeMembership.empty() && FallThrough != MF.end()) {
1367 auto MBBEHScope = EHScopeMembership.find(MBB);
1368 assert(MBBEHScope != EHScopeMembership.end());
1369 auto FallThroughEHScope = EHScopeMembership.find(&*FallThrough);
1370 assert(FallThroughEHScope != EHScopeMembership.end());
1371 SameEHScope = MBBEHScope->second == FallThroughEHScope->second;
1374 // If this block is empty, make everyone use its fall-through, not the block
1375 // explicitly. Landing pads should not do this since the landing-pad table
1376 // points to this block. Blocks with their addresses taken shouldn't be
1378 if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() &&
1380 salvageDebugInfoFromEmptyBlock(TII, *MBB);
1381 // Dead block? Leave for cleanup later.
1382 if (MBB->pred_empty()) return MadeChange;
1384 if (FallThrough == MF.end()) {
1385 // TODO: Simplify preds to not branch here if possible!
1386 } else if (FallThrough->isEHPad()) {
1387 // Don't rewrite to a landing pad fallthough. That could lead to the case
1388 // where a BB jumps to more than one landing pad.
1389 // TODO: Is it ever worth rewriting predecessors which don't already
1390 // jump to a landing pad, and so can safely jump to the fallthrough?
1391 } else if (MBB->isSuccessor(&*FallThrough)) {
1392 // Rewrite all predecessors of the old block to go to the fallthrough
1394 while (!MBB->pred_empty()) {
1395 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1396 Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough);
1398 // If MBB was the target of a jump table, update jump tables to go to the
1399 // fallthrough instead.
1400 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1401 MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough);
1407 // Check to see if we can simplify the terminator of the block before this
1409 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1411 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1412 SmallVector<MachineOperand, 4> PriorCond;
1413 bool PriorUnAnalyzable =
1414 TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1415 if (!PriorUnAnalyzable) {
1416 // If the CFG for the prior block has extra edges, remove them.
1417 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1418 !PriorCond.empty());
1420 // If the previous branch is conditional and both conditions go to the same
1421 // destination, remove the branch, replacing it with an unconditional one or
1423 if (PriorTBB && PriorTBB == PriorFBB) {
1424 DebugLoc dl = getBranchDebugLoc(PrevBB);
1425 TII->removeBranch(PrevBB);
1427 if (PriorTBB != MBB)
1428 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1431 goto ReoptimizeBlock;
1434 // If the previous block unconditionally falls through to this block and
1435 // this block has no other predecessors, move the contents of this block
1436 // into the prior block. This doesn't usually happen when SimplifyCFG
1437 // has been used, but it can happen if tail merging splits a fall-through
1438 // predecessor of a block.
1439 // This has to check PrevBB->succ_size() because EH edges are ignored by
1441 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1442 PrevBB.succ_size() == 1 &&
1443 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1444 LLVM_DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1445 << "From MBB: " << *MBB);
1446 // Remove redundant DBG_VALUEs first.
1447 if (PrevBB.begin() != PrevBB.end()) {
1448 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1450 MachineBasicBlock::iterator MBBIter = MBB->begin();
1451 // Check if DBG_VALUE at the end of PrevBB is identical to the
1452 // DBG_VALUE at the beginning of MBB.
1453 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1454 && PrevBBIter->isDebugInstr() && MBBIter->isDebugInstr()) {
1455 if (!MBBIter->isIdenticalTo(*PrevBBIter))
1457 MachineInstr &DuplicateDbg = *MBBIter;
1458 ++MBBIter; -- PrevBBIter;
1459 DuplicateDbg.eraseFromParent();
1462 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1463 PrevBB.removeSuccessor(PrevBB.succ_begin());
1464 assert(PrevBB.succ_empty());
1465 PrevBB.transferSuccessors(MBB);
1470 // If the previous branch *only* branches to *this* block (conditional or
1471 // not) remove the branch.
1472 if (PriorTBB == MBB && !PriorFBB) {
1473 TII->removeBranch(PrevBB);
1476 goto ReoptimizeBlock;
1479 // If the prior block branches somewhere else on the condition and here if
1480 // the condition is false, remove the uncond second branch.
1481 if (PriorFBB == MBB) {
1482 DebugLoc dl = getBranchDebugLoc(PrevBB);
1483 TII->removeBranch(PrevBB);
1484 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1487 goto ReoptimizeBlock;
1490 // If the prior block branches here on true and somewhere else on false, and
1491 // if the branch condition is reversible, reverse the branch to create a
1493 if (PriorTBB == MBB) {
1494 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1495 if (!TII->reverseBranchCondition(NewPriorCond)) {
1496 DebugLoc dl = getBranchDebugLoc(PrevBB);
1497 TII->removeBranch(PrevBB);
1498 TII->insertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1501 goto ReoptimizeBlock;
1505 // If this block has no successors (e.g. it is a return block or ends with
1506 // a call to a no-return function like abort or __cxa_throw) and if the pred
1507 // falls through into this block, and if it would otherwise fall through
1508 // into the block after this, move this block to the end of the function.
1510 // We consider it more likely that execution will stay in the function (e.g.
1511 // due to loops) than it is to exit it. This asserts in loops etc, moving
1512 // the assert condition out of the loop body.
1513 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1514 MachineFunction::iterator(PriorTBB) == FallThrough &&
1515 !MBB->canFallThrough()) {
1516 bool DoTransform = true;
1518 // We have to be careful that the succs of PredBB aren't both no-successor
1519 // blocks. If neither have successors and if PredBB is the second from
1520 // last block in the function, we'd just keep swapping the two blocks for
1521 // last. Only do the swap if one is clearly better to fall through than
1523 if (FallThrough == --MF.end() &&
1524 !IsBetterFallthrough(PriorTBB, MBB))
1525 DoTransform = false;
1528 // Reverse the branch so we will fall through on the previous true cond.
1529 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1530 if (!TII->reverseBranchCondition(NewPriorCond)) {
1531 LLVM_DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1532 << "To make fallthrough to: " << *PriorTBB << "\n");
1534 DebugLoc dl = getBranchDebugLoc(PrevBB);
1535 TII->removeBranch(PrevBB);
1536 TII->insertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1538 // Move this block to the end of the function.
1539 MBB->moveAfter(&MF.back());
1549 MF.getFunction().hasOptSize() ||
1550 llvm::shouldOptimizeForSize(MBB, PSI, &MBBFreqInfo.getMBFI());
1551 if (!IsEmptyBlock(MBB) && MBB->pred_size() == 1 && OptForSize) {
1552 // Changing "Jcc foo; foo: jmp bar;" into "Jcc bar;" might change the branch
1553 // direction, thereby defeating careful block placement and regressing
1554 // performance. Therefore, only consider this for optsize functions.
1555 MachineInstr &TailCall = *MBB->getFirstNonDebugInstr();
1556 if (TII->isUnconditionalTailCall(TailCall)) {
1557 MachineBasicBlock *Pred = *MBB->pred_begin();
1558 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1559 SmallVector<MachineOperand, 4> PredCond;
1560 bool PredAnalyzable =
1561 !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true);
1563 if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB &&
1564 PredTBB != PredFBB) {
1565 // The predecessor has a conditional branch to this block which consists
1566 // of only a tail call. Try to fold the tail call into the conditional
1568 if (TII->canMakeTailCallConditional(PredCond, TailCall)) {
1569 // TODO: It would be nice if analyzeBranch() could provide a pointer
1570 // to the branch instruction so replaceBranchWithTailCall() doesn't
1571 // have to search for it.
1572 TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall);
1574 Pred->removeSuccessor(MBB);
1579 // If the predecessor is falling through to this block, we could reverse
1580 // the branch condition and fold the tail call into that. However, after
1581 // that we might have to re-arrange the CFG to fall through to the other
1582 // block and there is a high risk of regressing code size rather than
1587 // Analyze the branch in the current block.
1588 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1589 SmallVector<MachineOperand, 4> CurCond;
1590 bool CurUnAnalyzable =
1591 TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1592 if (!CurUnAnalyzable) {
1593 // If the CFG for the prior block has extra edges, remove them.
1594 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1596 // If this is a two-way branch, and the FBB branches to this block, reverse
1597 // the condition so the single-basic-block loop is faster. Instead of:
1598 // Loop: xxx; jcc Out; jmp Loop
1600 // Loop: xxx; jncc Loop; jmp Out
1601 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1602 SmallVector<MachineOperand, 4> NewCond(CurCond);
1603 if (!TII->reverseBranchCondition(NewCond)) {
1604 DebugLoc dl = getBranchDebugLoc(*MBB);
1605 TII->removeBranch(*MBB);
1606 TII->insertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1609 goto ReoptimizeBlock;
1613 // If this branch is the only thing in its block, see if we can forward
1614 // other blocks across it.
1615 if (CurTBB && CurCond.empty() && !CurFBB &&
1616 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1617 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1618 DebugLoc dl = getBranchDebugLoc(*MBB);
1619 // This block may contain just an unconditional branch. Because there can
1620 // be 'non-branch terminators' in the block, try removing the branch and
1621 // then seeing if the block is empty.
1622 TII->removeBranch(*MBB);
1623 // If the only things remaining in the block are debug info, remove these
1624 // as well, so this will behave the same as an empty block in non-debug
1626 if (IsEmptyBlock(MBB)) {
1627 // Make the block empty, losing the debug info (we could probably
1628 // improve this in some cases.)
1629 MBB->erase(MBB->begin(), MBB->end());
1631 // If this block is just an unconditional branch to CurTBB, we can
1632 // usually completely eliminate the block. The only case we cannot
1633 // completely eliminate the block is when the block before this one
1634 // falls through into MBB and we can't understand the prior block's branch
1637 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1638 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1639 !PrevBB.isSuccessor(MBB)) {
1640 // If the prior block falls through into us, turn it into an
1641 // explicit branch to us to make updates simpler.
1642 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1643 PriorTBB != MBB && PriorFBB != MBB) {
1645 assert(PriorCond.empty() && !PriorFBB &&
1646 "Bad branch analysis");
1649 assert(!PriorFBB && "Machine CFG out of date!");
1652 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1653 TII->removeBranch(PrevBB);
1654 TII->insertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1657 // Iterate through all the predecessors, revectoring each in-turn.
1659 bool DidChange = false;
1660 bool HasBranchToSelf = false;
1661 while(PI != MBB->pred_size()) {
1662 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1664 // If this block has an uncond branch to itself, leave it.
1666 HasBranchToSelf = true;
1669 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1670 // If this change resulted in PMBB ending in a conditional
1671 // branch where both conditions go to the same destination,
1672 // change this to an unconditional branch (and fix the CFG).
1673 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1674 SmallVector<MachineOperand, 4> NewCurCond;
1675 bool NewCurUnAnalyzable = TII->analyzeBranch(
1676 *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true);
1677 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1678 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1679 TII->removeBranch(*PMBB);
1681 TII->insertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1684 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1689 // Change any jumptables to go to the new MBB.
1690 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1691 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1695 if (!HasBranchToSelf) return MadeChange;
1700 // Add the branch back if the block is more than just an uncond branch.
1701 TII->insertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1705 // If the prior block doesn't fall through into this block, and if this
1706 // block doesn't fall through into some other block, see if we can find a
1707 // place to move this block where a fall-through will happen.
1708 if (!PrevBB.canFallThrough()) {
1709 // Now we know that there was no fall-through into this block, check to
1710 // see if it has a fall-through into its successor.
1711 bool CurFallsThru = MBB->canFallThrough();
1713 if (!MBB->isEHPad()) {
1714 // Check all the predecessors of this block. If one of them has no fall
1715 // throughs, move this block right after it.
1716 for (MachineBasicBlock *PredBB : MBB->predecessors()) {
1717 // Analyze the branch at the end of the pred.
1718 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1719 SmallVector<MachineOperand, 4> PredCond;
1720 if (PredBB != MBB && !PredBB->canFallThrough() &&
1721 !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) &&
1722 (!CurFallsThru || !CurTBB || !CurFBB) &&
1723 (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1724 // If the current block doesn't fall through, just move it.
1725 // If the current block can fall through and does not end with a
1726 // conditional branch, we need to append an unconditional jump to
1727 // the (current) next block. To avoid a possible compile-time
1728 // infinite loop, move blocks only backward in this case.
1729 // Also, if there are already 2 branches here, we cannot add a third;
1730 // this means we have the case
1735 MachineBasicBlock *NextBB = &*std::next(MBB->getIterator());
1737 TII->insertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1739 MBB->moveAfter(PredBB);
1741 goto ReoptimizeBlock;
1746 if (!CurFallsThru) {
1747 // Check all successors to see if we can move this block before it.
1748 for (MachineBasicBlock *SuccBB : MBB->successors()) {
1749 // Analyze the branch at the end of the block before the succ.
1750 MachineFunction::iterator SuccPrev = --SuccBB->getIterator();
1752 // If this block doesn't already fall-through to that successor, and if
1753 // the succ doesn't already have a block that can fall through into it,
1754 // and if the successor isn't an EH destination, we can arrange for the
1755 // fallthrough to happen.
1756 if (SuccBB != MBB && &*SuccPrev != MBB &&
1757 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1758 !SuccBB->isEHPad()) {
1759 MBB->moveBefore(SuccBB);
1761 goto ReoptimizeBlock;
1765 // Okay, there is no really great place to put this block. If, however,
1766 // the block before this one would be a fall-through if this block were
1767 // removed, move this block to the end of the function. There is no real
1768 // advantage in "falling through" to an EH block, so we don't want to
1769 // perform this transformation for that case.
1771 // Also, Windows EH introduced the possibility of an arbitrary number of
1772 // successors to a given block. The analyzeBranch call does not consider
1773 // exception handling and so we can get in a state where a block
1774 // containing a call is followed by multiple EH blocks that would be
1775 // rotated infinitely at the end of the function if the transformation
1776 // below were performed for EH "FallThrough" blocks. Therefore, even if
1777 // that appears not to be happening anymore, we should assume that it is
1778 // possible and not remove the "!FallThrough()->isEHPad" condition below.
1779 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1780 SmallVector<MachineOperand, 4> PrevCond;
1781 if (FallThrough != MF.end() &&
1782 !FallThrough->isEHPad() &&
1783 !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1784 PrevBB.isSuccessor(&*FallThrough)) {
1785 MBB->moveAfter(&MF.back());
1795 //===----------------------------------------------------------------------===//
1796 // Hoist Common Code
1797 //===----------------------------------------------------------------------===//
1799 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1800 bool MadeChange = false;
1801 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1802 MachineBasicBlock *MBB = &*I++;
1803 MadeChange |= HoistCommonCodeInSuccs(MBB);
1809 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1810 /// its 'true' successor.
1811 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1812 MachineBasicBlock *TrueBB) {
1813 for (MachineBasicBlock *SuccBB : BB->successors())
1814 if (SuccBB != TrueBB)
1819 template <class Container>
1820 static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
1822 if (Register::isPhysicalRegister(Reg)) {
1823 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1830 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1831 /// in successors to. The location is usually just before the terminator,
1832 /// however if the terminator is a conditional branch and its previous
1833 /// instruction is the flag setting instruction, the previous instruction is
1834 /// the preferred location. This function also gathers uses and defs of the
1835 /// instructions from the insertion point to the end of the block. The data is
1836 /// used by HoistCommonCodeInSuccs to ensure safety.
1838 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1839 const TargetInstrInfo *TII,
1840 const TargetRegisterInfo *TRI,
1841 SmallSet<unsigned,4> &Uses,
1842 SmallSet<unsigned,4> &Defs) {
1843 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1844 if (!TII->isUnpredicatedTerminator(*Loc))
1847 for (const MachineOperand &MO : Loc->operands()) {
1850 Register Reg = MO.getReg();
1854 addRegAndItsAliases(Reg, TRI, Uses);
1857 // Don't try to hoist code in the rare case the terminator defines a
1858 // register that is later used.
1861 // If the terminator defines a register, make sure we don't hoist
1862 // the instruction whose def might be clobbered by the terminator.
1863 addRegAndItsAliases(Reg, TRI, Defs);
1869 // If the terminator is the only instruction in the block and Uses is not
1870 // empty (or we would have returned above), we can still safely hoist
1871 // instructions just before the terminator as long as the Defs/Uses are not
1872 // violated (which is checked in HoistCommonCodeInSuccs).
1873 if (Loc == MBB->begin())
1876 // The terminator is probably a conditional branch, try not to separate the
1877 // branch from condition setting instruction.
1878 MachineBasicBlock::iterator PI =
1879 skipDebugInstructionsBackward(std::prev(Loc), MBB->begin());
1882 for (const MachineOperand &MO : PI->operands()) {
1883 // If PI has a regmask operand, it is probably a call. Separate away.
1886 if (!MO.isReg() || MO.isUse())
1888 Register Reg = MO.getReg();
1891 if (Uses.count(Reg)) {
1897 // The condition setting instruction is not just before the conditional
1901 // Be conservative, don't insert instruction above something that may have
1902 // side-effects. And since it's potentially bad to separate flag setting
1903 // instruction from the conditional branch, just abort the optimization
1905 // Also avoid moving code above predicated instruction since it's hard to
1906 // reason about register liveness with predicated instruction.
1907 bool DontMoveAcrossStore = true;
1908 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI))
1911 // Find out what registers are live. Note this routine is ignoring other live
1912 // registers which are only used by instructions in successor blocks.
1913 for (const MachineOperand &MO : PI->operands()) {
1916 Register Reg = MO.getReg();
1920 addRegAndItsAliases(Reg, TRI, Uses);
1922 if (Uses.erase(Reg)) {
1923 if (Register::isPhysicalRegister(Reg)) {
1924 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1925 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1928 addRegAndItsAliases(Reg, TRI, Defs);
1935 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1936 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1937 SmallVector<MachineOperand, 4> Cond;
1938 if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1941 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1943 // Malformed bcc? True and false blocks are the same?
1946 // Restrict the optimization to cases where MBB is the only predecessor,
1947 // it is an obvious win.
1948 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1951 // Find a suitable position to hoist the common instructions to. Also figure
1952 // out which registers are used or defined by instructions from the insertion
1953 // point to the end of the block.
1954 SmallSet<unsigned, 4> Uses, Defs;
1955 MachineBasicBlock::iterator Loc =
1956 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1957 if (Loc == MBB->end())
1960 bool HasDups = false;
1961 SmallSet<unsigned, 4> ActiveDefsSet, AllDefsSet;
1962 MachineBasicBlock::iterator TIB = TBB->begin();
1963 MachineBasicBlock::iterator FIB = FBB->begin();
1964 MachineBasicBlock::iterator TIE = TBB->end();
1965 MachineBasicBlock::iterator FIE = FBB->end();
1966 while (TIB != TIE && FIB != FIE) {
1967 // Skip dbg_value instructions. These do not count.
1968 TIB = skipDebugInstructionsForward(TIB, TIE);
1969 FIB = skipDebugInstructionsForward(FIB, FIE);
1970 if (TIB == TIE || FIB == FIE)
1973 if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead))
1976 if (TII->isPredicated(*TIB))
1977 // Hard to reason about register liveness with predicated instruction.
1981 for (MachineOperand &MO : TIB->operands()) {
1982 // Don't attempt to hoist instructions with register masks.
1983 if (MO.isRegMask()) {
1989 Register Reg = MO.getReg();
1993 if (Uses.count(Reg)) {
1994 // Avoid clobbering a register that's used by the instruction at
1995 // the point of insertion.
2000 if (Defs.count(Reg) && !MO.isDead()) {
2001 // Don't hoist the instruction if the def would be clobber by the
2002 // instruction at the point insertion. FIXME: This is overly
2003 // conservative. It should be possible to hoist the instructions
2004 // in BB2 in the following example:
2006 // r1, eflag = op1 r2, r3
2015 } else if (!ActiveDefsSet.count(Reg)) {
2016 if (Defs.count(Reg)) {
2017 // Use is defined by the instruction at the point of insertion.
2022 if (MO.isKill() && Uses.count(Reg))
2023 // Kills a register that's read by the instruction at the point of
2024 // insertion. Remove the kill marker.
2025 MO.setIsKill(false);
2031 bool DontMoveAcrossStore = true;
2032 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
2035 // Remove kills from ActiveDefsSet, these registers had short live ranges.
2036 for (const MachineOperand &MO : TIB->operands()) {
2037 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
2039 Register Reg = MO.getReg();
2042 if (!AllDefsSet.count(Reg)) {
2045 if (Register::isPhysicalRegister(Reg)) {
2046 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
2047 ActiveDefsSet.erase(*AI);
2049 ActiveDefsSet.erase(Reg);
2053 // Track local defs so we can update liveins.
2054 for (const MachineOperand &MO : TIB->operands()) {
2055 if (!MO.isReg() || !MO.isDef() || MO.isDead())
2057 Register Reg = MO.getReg();
2058 if (!Reg || Register::isVirtualRegister(Reg))
2060 addRegAndItsAliases(Reg, TRI, ActiveDefsSet);
2061 addRegAndItsAliases(Reg, TRI, AllDefsSet);
2072 MBB->splice(Loc, TBB, TBB->begin(), TIB);
2073 FBB->erase(FBB->begin(), FIB);
2075 if (UpdateLiveIns) {
2076 recomputeLiveIns(*TBB);
2077 recomputeLiveIns(*FBB);