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/MBFIWrapper.h"
44 #include "llvm/CodeGen/TargetInstrInfo.h"
45 #include "llvm/CodeGen/TargetOpcodes.h"
46 #include "llvm/CodeGen/TargetPassConfig.h"
47 #include "llvm/CodeGen/TargetRegisterInfo.h"
48 #include "llvm/CodeGen/TargetSubtargetInfo.h"
49 #include "llvm/IR/DebugInfoMetadata.h"
50 #include "llvm/IR/DebugLoc.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/InitializePasses.h"
53 #include "llvm/MC/LaneBitmask.h"
54 #include "llvm/MC/MCRegisterInfo.h"
55 #include "llvm/Pass.h"
56 #include "llvm/Support/BlockFrequency.h"
57 #include "llvm/Support/BranchProbability.h"
58 #include "llvm/Support/CommandLine.h"
59 #include "llvm/Support/Debug.h"
60 #include "llvm/Support/ErrorHandling.h"
61 #include "llvm/Support/raw_ostream.h"
62 #include "llvm/Target/TargetMachine.h"
71 #define DEBUG_TYPE "branch-folder"
73 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
74 STATISTIC(NumBranchOpts, "Number of branches optimized");
75 STATISTIC(NumTailMerge , "Number of block tails merged");
76 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
77 STATISTIC(NumTailCalls, "Number of tail calls optimized");
79 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
80 cl::init(cl::BOU_UNSET), cl::Hidden);
82 // Throttle for huge numbers of predecessors (compile speed problems)
83 static cl::opt<unsigned>
84 TailMergeThreshold("tail-merge-threshold",
85 cl::desc("Max number of predecessors to consider tail merging"),
86 cl::init(150), cl::Hidden);
88 // Heuristic for tail merging (and, inversely, tail duplication).
89 // TODO: This should be replaced with a target query.
90 static cl::opt<unsigned>
91 TailMergeSize("tail-merge-size",
92 cl::desc("Min number of instructions to consider tail merging"),
93 cl::init(3), cl::Hidden);
97 /// BranchFolderPass - Wrap branch folder in a machine function pass.
98 class BranchFolderPass : public MachineFunctionPass {
102 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
104 bool runOnMachineFunction(MachineFunction &MF) override;
106 void getAnalysisUsage(AnalysisUsage &AU) const override {
107 AU.addRequired<MachineBlockFrequencyInfo>();
108 AU.addRequired<MachineBranchProbabilityInfo>();
109 AU.addRequired<ProfileSummaryInfoWrapperPass>();
110 AU.addRequired<TargetPassConfig>();
111 MachineFunctionPass::getAnalysisUsage(AU);
115 } // end anonymous namespace
117 char BranchFolderPass::ID = 0;
119 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
121 INITIALIZE_PASS(BranchFolderPass, DEBUG_TYPE,
122 "Control Flow Optimizer", false, false)
124 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
125 if (skipFunction(MF.getFunction()))
128 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
129 // TailMerge can create jump into if branches that make CFG irreducible for
130 // HW that requires structurized CFG.
131 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
132 PassConfig->getEnableTailMerge();
133 MBFIWrapper MBBFreqInfo(
134 getAnalysis<MachineBlockFrequencyInfo>());
135 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo,
136 getAnalysis<MachineBranchProbabilityInfo>(),
137 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI());
138 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
139 MF.getSubtarget().getRegisterInfo());
142 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
143 MBFIWrapper &FreqInfo,
144 const MachineBranchProbabilityInfo &ProbInfo,
145 ProfileSummaryInfo *PSI,
146 unsigned MinTailLength)
147 : EnableHoistCommonCode(CommonHoist), MinCommonTailLength(MinTailLength),
148 MBBFreqInfo(FreqInfo), MBPI(ProbInfo), PSI(PSI) {
149 if (MinCommonTailLength == 0)
150 MinCommonTailLength = TailMergeSize;
151 switch (FlagEnableTailMerge) {
152 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
153 case cl::BOU_TRUE: EnableTailMerge = true; break;
154 case cl::BOU_FALSE: EnableTailMerge = false; break;
158 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
159 assert(MBB->pred_empty() && "MBB must be dead!");
160 LLVM_DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
162 MachineFunction *MF = MBB->getParent();
163 // drop all successors.
164 while (!MBB->succ_empty())
165 MBB->removeSuccessor(MBB->succ_end()-1);
167 // Avoid matching if this pointer gets reused.
168 TriedMerging.erase(MBB);
170 // Update call site info.
171 std::for_each(MBB->begin(), MBB->end(), [MF](const MachineInstr &MI) {
172 if (MI.shouldUpdateCallSiteInfo())
173 MF->eraseCallSiteInfo(&MI);
177 EHScopeMembership.erase(MBB);
179 MLI->removeBlock(MBB);
182 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
183 const TargetInstrInfo *tii,
184 const TargetRegisterInfo *tri,
185 MachineLoopInfo *mli, bool AfterPlacement) {
186 if (!tii) return false;
188 TriedMerging.clear();
190 MachineRegisterInfo &MRI = MF.getRegInfo();
191 AfterBlockPlacement = AfterPlacement;
197 UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF);
199 MRI.invalidateLiveness();
201 bool MadeChange = false;
203 // Recalculate EH scope membership.
204 EHScopeMembership = getEHScopeMembership(MF);
206 bool MadeChangeThisIteration = true;
207 while (MadeChangeThisIteration) {
208 MadeChangeThisIteration = TailMergeBlocks(MF);
209 // No need to clean up if tail merging does not change anything after the
211 if (!AfterBlockPlacement || MadeChangeThisIteration)
212 MadeChangeThisIteration |= OptimizeBranches(MF);
213 if (EnableHoistCommonCode)
214 MadeChangeThisIteration |= HoistCommonCode(MF);
215 MadeChange |= MadeChangeThisIteration;
218 // See if any jump tables have become dead as the code generator
220 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
224 // Walk the function to find jump tables that are live.
225 BitVector JTIsLive(JTI->getJumpTables().size());
226 for (const MachineBasicBlock &BB : MF) {
227 for (const MachineInstr &I : BB)
228 for (const MachineOperand &Op : I.operands()) {
229 if (!Op.isJTI()) continue;
231 // Remember that this JT is live.
232 JTIsLive.set(Op.getIndex());
236 // Finally, remove dead jump tables. This happens when the
237 // indirect jump was unreachable (and thus deleted).
238 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
239 if (!JTIsLive.test(i)) {
240 JTI->RemoveJumpTable(i);
247 //===----------------------------------------------------------------------===//
248 // Tail Merging of Blocks
249 //===----------------------------------------------------------------------===//
251 /// HashMachineInstr - Compute a hash value for MI and its operands.
252 static unsigned HashMachineInstr(const MachineInstr &MI) {
253 unsigned Hash = MI.getOpcode();
254 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
255 const MachineOperand &Op = MI.getOperand(i);
257 // Merge in bits from the operand if easy. We can't use MachineOperand's
258 // hash_code here because it's not deterministic and we sort by hash value
260 unsigned OperandHash = 0;
261 switch (Op.getType()) {
262 case MachineOperand::MO_Register:
263 OperandHash = Op.getReg();
265 case MachineOperand::MO_Immediate:
266 OperandHash = Op.getImm();
268 case MachineOperand::MO_MachineBasicBlock:
269 OperandHash = Op.getMBB()->getNumber();
271 case MachineOperand::MO_FrameIndex:
272 case MachineOperand::MO_ConstantPoolIndex:
273 case MachineOperand::MO_JumpTableIndex:
274 OperandHash = Op.getIndex();
276 case MachineOperand::MO_GlobalAddress:
277 case MachineOperand::MO_ExternalSymbol:
278 // Global address / external symbol are too hard, don't bother, but do
279 // pull in the offset.
280 OperandHash = Op.getOffset();
286 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
291 /// HashEndOfMBB - Hash the last instruction in the MBB.
292 static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) {
293 MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr();
297 return HashMachineInstr(*I);
300 /// Whether MI should be counted as an instruction when calculating common tail.
301 static bool countsAsInstruction(const MachineInstr &MI) {
302 return !(MI.isDebugInstr() || MI.isCFIInstruction());
305 /// Iterate backwards from the given iterator \p I, towards the beginning of the
306 /// block. If a MI satisfying 'countsAsInstruction' is found, return an iterator
307 /// pointing to that MI. If no such MI is found, return the end iterator.
308 static MachineBasicBlock::iterator
309 skipBackwardPastNonInstructions(MachineBasicBlock::iterator I,
310 MachineBasicBlock *MBB) {
311 while (I != MBB->begin()) {
313 if (countsAsInstruction(*I))
319 /// Given two machine basic blocks, return the number of instructions they
320 /// actually have in common together at their end. If a common tail is found (at
321 /// least by one instruction), then iterators for the first shared instruction
322 /// in each block are returned as well.
324 /// Non-instructions according to countsAsInstruction are ignored.
325 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
326 MachineBasicBlock *MBB2,
327 MachineBasicBlock::iterator &I1,
328 MachineBasicBlock::iterator &I2) {
329 MachineBasicBlock::iterator MBBI1 = MBB1->end();
330 MachineBasicBlock::iterator MBBI2 = MBB2->end();
332 unsigned TailLen = 0;
334 MBBI1 = skipBackwardPastNonInstructions(MBBI1, MBB1);
335 MBBI2 = skipBackwardPastNonInstructions(MBBI2, MBB2);
336 if (MBBI1 == MBB1->end() || MBBI2 == MBB2->end())
338 if (!MBBI1->isIdenticalTo(*MBBI2) ||
339 // FIXME: This check is dubious. It's used to get around a problem where
340 // people incorrectly expect inline asm directives to remain in the same
341 // relative order. This is untenable because normal compiler
342 // optimizations (like this one) may reorder and/or merge these
344 MBBI1->isInlineAsm()) {
347 if (MBBI1->getFlag(MachineInstr::NoMerge) ||
348 MBBI2->getFlag(MachineInstr::NoMerge))
358 void BranchFolder::replaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
359 MachineBasicBlock &NewDest) {
361 // OldInst should always point to an instruction.
362 MachineBasicBlock &OldMBB = *OldInst->getParent();
364 LiveRegs.addLiveOuts(OldMBB);
365 // Move backward to the place where will insert the jump.
366 MachineBasicBlock::iterator I = OldMBB.end();
369 LiveRegs.stepBackward(*I);
370 } while (I != OldInst);
372 // Merging the tails may have switched some undef operand to non-undef ones.
373 // Add IMPLICIT_DEFS into OldMBB as necessary to have a definition of the
375 for (MachineBasicBlock::RegisterMaskPair P : NewDest.liveins()) {
376 // We computed the liveins with computeLiveIn earlier and should only see
378 assert(P.LaneMask == LaneBitmask::getAll() &&
379 "Can only handle full register.");
380 MCPhysReg Reg = P.PhysReg;
381 if (!LiveRegs.available(*MRI, Reg))
384 BuildMI(OldMBB, OldInst, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Reg);
388 TII->ReplaceTailWithBranchTo(OldInst, &NewDest);
392 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
393 MachineBasicBlock::iterator BBI1,
394 const BasicBlock *BB) {
395 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
398 MachineFunction &MF = *CurMBB.getParent();
400 // Create the fall-through block.
401 MachineFunction::iterator MBBI = CurMBB.getIterator();
402 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(BB);
403 CurMBB.getParent()->insert(++MBBI, NewMBB);
405 // Move all the successors of this block to the specified block.
406 NewMBB->transferSuccessors(&CurMBB);
408 // Add an edge from CurMBB to NewMBB for the fall-through.
409 CurMBB.addSuccessor(NewMBB);
411 // Splice the code over.
412 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
414 // NewMBB belongs to the same loop as CurMBB.
416 if (MachineLoop *ML = MLI->getLoopFor(&CurMBB))
417 ML->addBasicBlockToLoop(NewMBB, MLI->getBase());
419 // NewMBB inherits CurMBB's block frequency.
420 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
423 computeAndAddLiveIns(LiveRegs, *NewMBB);
425 // Add the new block to the EH scope.
426 const auto &EHScopeI = EHScopeMembership.find(&CurMBB);
427 if (EHScopeI != EHScopeMembership.end()) {
428 auto n = EHScopeI->second;
429 EHScopeMembership[NewMBB] = n;
435 /// EstimateRuntime - Make a rough estimate for how long it will take to run
436 /// the specified code.
437 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
438 MachineBasicBlock::iterator E) {
440 for (; I != E; ++I) {
441 if (!countsAsInstruction(*I))
445 else if (I->mayLoadOrStore())
453 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
454 // branches temporarily for tail merging). In the case where CurMBB ends
455 // with a conditional branch to the next block, optimize by reversing the
456 // test and conditionally branching to SuccMBB instead.
457 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
458 const TargetInstrInfo *TII) {
459 MachineFunction *MF = CurMBB->getParent();
460 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
461 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
462 SmallVector<MachineOperand, 4> Cond;
463 DebugLoc dl = CurMBB->findBranchDebugLoc();
464 if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
465 MachineBasicBlock *NextBB = &*I;
466 if (TBB == NextBB && !Cond.empty() && !FBB) {
467 if (!TII->reverseBranchCondition(Cond)) {
468 TII->removeBranch(*CurMBB);
469 TII->insertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
474 TII->insertBranch(*CurMBB, SuccBB, nullptr,
475 SmallVector<MachineOperand, 0>(), dl);
479 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
480 if (getHash() < o.getHash())
482 if (getHash() > o.getHash())
484 if (getBlock()->getNumber() < o.getBlock()->getNumber())
486 if (getBlock()->getNumber() > o.getBlock()->getNumber())
488 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
489 // an object with itself.
490 #ifndef _GLIBCXX_DEBUG
491 llvm_unreachable("Predecessor appears twice");
497 /// CountTerminators - Count the number of terminators in the given
498 /// block and set I to the position of the first non-terminator, if there
499 /// is one, or MBB->end() otherwise.
500 static unsigned CountTerminators(MachineBasicBlock *MBB,
501 MachineBasicBlock::iterator &I) {
503 unsigned NumTerms = 0;
505 if (I == MBB->begin()) {
510 if (!I->isTerminator()) break;
516 /// A no successor, non-return block probably ends in unreachable and is cold.
517 /// Also consider a block that ends in an indirect branch to be a return block,
518 /// since many targets use plain indirect branches to return.
519 static bool blockEndsInUnreachable(const MachineBasicBlock *MBB) {
520 if (!MBB->succ_empty())
524 return !(MBB->back().isReturn() || MBB->back().isIndirectBranch());
527 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
528 /// and decide if it would be profitable to merge those tails. Return the
529 /// length of the common tail and iterators to the first common instruction
531 /// MBB1, MBB2 The blocks to check
532 /// MinCommonTailLength Minimum size of tail block to be merged.
533 /// CommonTailLen Out parameter to record the size of the shared tail between
535 /// I1, I2 Iterator references that will be changed to point to the first
536 /// instruction in the common tail shared by MBB1,MBB2
537 /// SuccBB A common successor of MBB1, MBB2 which are in a canonical form
538 /// relative to SuccBB
539 /// PredBB The layout predecessor of SuccBB, if any.
540 /// EHScopeMembership map from block to EH scope #.
541 /// AfterPlacement True if we are merging blocks after layout. Stricter
542 /// thresholds apply to prevent undoing tail-duplication.
544 ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2,
545 unsigned MinCommonTailLength, unsigned &CommonTailLen,
546 MachineBasicBlock::iterator &I1,
547 MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB,
548 MachineBasicBlock *PredBB,
549 DenseMap<const MachineBasicBlock *, int> &EHScopeMembership,
551 MBFIWrapper &MBBFreqInfo,
552 ProfileSummaryInfo *PSI) {
553 // It is never profitable to tail-merge blocks from two different EH scopes.
554 if (!EHScopeMembership.empty()) {
555 auto EHScope1 = EHScopeMembership.find(MBB1);
556 assert(EHScope1 != EHScopeMembership.end());
557 auto EHScope2 = EHScopeMembership.find(MBB2);
558 assert(EHScope2 != EHScopeMembership.end());
559 if (EHScope1->second != EHScope2->second)
563 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
564 if (CommonTailLen == 0)
566 LLVM_DEBUG(dbgs() << "Common tail length of " << printMBBReference(*MBB1)
567 << " and " << printMBBReference(*MBB2) << " is "
568 << CommonTailLen << '\n');
570 // Move the iterators to the beginning of the MBB if we only got debug
571 // instructions before the tail. This is to avoid splitting a block when we
572 // only got debug instructions before the tail (to be invariant on -g).
573 if (skipDebugInstructionsForward(MBB1->begin(), MBB1->end()) == I1)
575 if (skipDebugInstructionsForward(MBB2->begin(), MBB2->end()) == I2)
578 bool FullBlockTail1 = I1 == MBB1->begin();
579 bool FullBlockTail2 = I2 == MBB2->begin();
581 // It's almost always profitable to merge any number of non-terminator
582 // instructions with the block that falls through into the common successor.
583 // This is true only for a single successor. For multiple successors, we are
584 // trading a conditional branch for an unconditional one.
585 // TODO: Re-visit successor size for non-layout tail merging.
586 if ((MBB1 == PredBB || MBB2 == PredBB) &&
587 (!AfterPlacement || MBB1->succ_size() == 1)) {
588 MachineBasicBlock::iterator I;
589 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
590 if (CommonTailLen > NumTerms)
594 // If these are identical non-return blocks with no successors, merge them.
595 // Such blocks are typically cold calls to noreturn functions like abort, and
596 // are unlikely to become a fallthrough target after machine block placement.
597 // Tail merging these blocks is unlikely to create additional unconditional
598 // branches, and will reduce the size of this cold code.
599 if (FullBlockTail1 && FullBlockTail2 &&
600 blockEndsInUnreachable(MBB1) && blockEndsInUnreachable(MBB2))
603 // If one of the blocks can be completely merged and happens to be in
604 // a position where the other could fall through into it, merge any number
605 // of instructions, because it can be done without a branch.
606 // TODO: If the blocks are not adjacent, move one of them so that they are?
607 if (MBB1->isLayoutSuccessor(MBB2) && FullBlockTail2)
609 if (MBB2->isLayoutSuccessor(MBB1) && FullBlockTail1)
612 // If both blocks are identical and end in a branch, merge them unless they
613 // both have a fallthrough predecessor and successor.
614 // We can only do this after block placement because it depends on whether
615 // there are fallthroughs, and we don't know until after layout.
616 if (AfterPlacement && FullBlockTail1 && FullBlockTail2) {
617 auto BothFallThrough = [](MachineBasicBlock *MBB) {
618 if (MBB->succ_size() != 0 && !MBB->canFallThrough())
620 MachineFunction::iterator I(MBB);
621 MachineFunction *MF = MBB->getParent();
622 return (MBB != &*MF->begin()) && std::prev(I)->canFallThrough();
624 if (!BothFallThrough(MBB1) || !BothFallThrough(MBB2))
628 // If both blocks have an unconditional branch temporarily stripped out,
629 // count that as an additional common instruction for the following
630 // heuristics. This heuristic is only accurate for single-succ blocks, so to
631 // make sure that during layout merging and duplicating don't crash, we check
632 // for that when merging during layout.
633 unsigned EffectiveTailLen = CommonTailLen;
634 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
635 (MBB1->succ_size() == 1 || !AfterPlacement) &&
636 !MBB1->back().isBarrier() &&
637 !MBB2->back().isBarrier())
640 // Check if the common tail is long enough to be worthwhile.
641 if (EffectiveTailLen >= MinCommonTailLength)
644 // If we are optimizing for code size, 2 instructions in common is enough if
645 // we don't have to split a block. At worst we will be introducing 1 new
646 // branch instruction, which is likely to be smaller than the 2
647 // instructions that would be deleted in the merge.
648 MachineFunction *MF = MBB1->getParent();
650 MF->getFunction().hasOptSize() ||
651 (llvm::shouldOptimizeForSize(MBB1, PSI, &MBBFreqInfo) &&
652 llvm::shouldOptimizeForSize(MBB2, PSI, &MBBFreqInfo));
653 return EffectiveTailLen >= 2 && OptForSize &&
654 (FullBlockTail1 || FullBlockTail2);
657 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
658 unsigned MinCommonTailLength,
659 MachineBasicBlock *SuccBB,
660 MachineBasicBlock *PredBB) {
661 unsigned maxCommonTailLength = 0U;
663 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
664 MPIterator HighestMPIter = std::prev(MergePotentials.end());
665 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
666 B = MergePotentials.begin();
667 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
668 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
669 unsigned CommonTailLen;
670 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
672 CommonTailLen, TrialBBI1, TrialBBI2,
675 AfterBlockPlacement, MBBFreqInfo, PSI)) {
676 if (CommonTailLen > maxCommonTailLength) {
678 maxCommonTailLength = CommonTailLen;
679 HighestMPIter = CurMPIter;
680 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
682 if (HighestMPIter == CurMPIter &&
683 CommonTailLen == maxCommonTailLength)
684 SameTails.push_back(SameTailElt(I, TrialBBI2));
690 return maxCommonTailLength;
693 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
694 MachineBasicBlock *SuccBB,
695 MachineBasicBlock *PredBB) {
696 MPIterator CurMPIter, B;
697 for (CurMPIter = std::prev(MergePotentials.end()),
698 B = MergePotentials.begin();
699 CurMPIter->getHash() == CurHash; --CurMPIter) {
700 // Put the unconditional branch back, if we need one.
701 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
702 if (SuccBB && CurMBB != PredBB)
703 FixTail(CurMBB, SuccBB, TII);
707 if (CurMPIter->getHash() != CurHash)
709 MergePotentials.erase(CurMPIter, MergePotentials.end());
712 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
713 MachineBasicBlock *SuccBB,
714 unsigned maxCommonTailLength,
715 unsigned &commonTailIndex) {
717 unsigned TimeEstimate = ~0U;
718 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
719 // Use PredBB if possible; that doesn't require a new branch.
720 if (SameTails[i].getBlock() == PredBB) {
724 // Otherwise, make a (fairly bogus) choice based on estimate of
725 // how long it will take the various blocks to execute.
726 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
727 SameTails[i].getTailStartPos());
728 if (t <= TimeEstimate) {
734 MachineBasicBlock::iterator BBI =
735 SameTails[commonTailIndex].getTailStartPos();
736 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
738 LLVM_DEBUG(dbgs() << "\nSplitting " << printMBBReference(*MBB) << ", size "
739 << maxCommonTailLength);
741 // If the split block unconditionally falls-thru to SuccBB, it will be
742 // merged. In control flow terms it should then take SuccBB's name. e.g. If
743 // SuccBB is an inner loop, the common tail is still part of the inner loop.
744 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
745 SuccBB->getBasicBlock() : MBB->getBasicBlock();
746 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
748 LLVM_DEBUG(dbgs() << "... failed!");
752 SameTails[commonTailIndex].setBlock(newMBB);
753 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
755 // If we split PredBB, newMBB is the new predecessor.
763 mergeOperations(MachineBasicBlock::iterator MBBIStartPos,
764 MachineBasicBlock &MBBCommon) {
765 MachineBasicBlock *MBB = MBBIStartPos->getParent();
766 // Note CommonTailLen does not necessarily matches the size of
767 // the common BB nor all its instructions because of debug
768 // instructions differences.
769 unsigned CommonTailLen = 0;
770 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
773 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
774 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
775 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
776 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
778 while (CommonTailLen--) {
779 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
782 if (!countsAsInstruction(*MBBI)) {
787 while ((MBBICommon != MBBIECommon) && !countsAsInstruction(*MBBICommon))
790 assert(MBBICommon != MBBIECommon &&
791 "Reached BB end within common tail length!");
792 assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!");
794 // Merge MMOs from memory operations in the common block.
795 if (MBBICommon->mayLoadOrStore())
796 MBBICommon->cloneMergedMemRefs(*MBB->getParent(), {&*MBBICommon, &*MBBI});
797 // Drop undef flags if they aren't present in all merged instructions.
798 for (unsigned I = 0, E = MBBICommon->getNumOperands(); I != E; ++I) {
799 MachineOperand &MO = MBBICommon->getOperand(I);
800 if (MO.isReg() && MO.isUndef()) {
801 const MachineOperand &OtherMO = MBBI->getOperand(I);
802 if (!OtherMO.isUndef())
803 MO.setIsUndef(false);
812 void BranchFolder::mergeCommonTails(unsigned commonTailIndex) {
813 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
815 std::vector<MachineBasicBlock::iterator> NextCommonInsts(SameTails.size());
816 for (unsigned int i = 0 ; i != SameTails.size() ; ++i) {
817 if (i != commonTailIndex) {
818 NextCommonInsts[i] = SameTails[i].getTailStartPos();
819 mergeOperations(SameTails[i].getTailStartPos(), *MBB);
821 assert(SameTails[i].getTailStartPos() == MBB->begin() &&
822 "MBB is not a common tail only block");
826 for (auto &MI : *MBB) {
827 if (!countsAsInstruction(MI))
829 DebugLoc DL = MI.getDebugLoc();
830 for (unsigned int i = 0 ; i < NextCommonInsts.size() ; i++) {
831 if (i == commonTailIndex)
834 auto &Pos = NextCommonInsts[i];
835 assert(Pos != SameTails[i].getBlock()->end() &&
836 "Reached BB end within common tail");
837 while (!countsAsInstruction(*Pos)) {
839 assert(Pos != SameTails[i].getBlock()->end() &&
840 "Reached BB end within common tail");
842 assert(MI.isIdenticalTo(*Pos) && "Expected matching MIIs!");
843 DL = DILocation::getMergedLocation(DL, Pos->getDebugLoc());
844 NextCommonInsts[i] = ++Pos;
850 LivePhysRegs NewLiveIns(*TRI);
851 computeLiveIns(NewLiveIns, *MBB);
854 // The flag merging may lead to some register uses no longer using the
855 // <undef> flag, add IMPLICIT_DEFs in the predecessors as necessary.
856 for (MachineBasicBlock *Pred : MBB->predecessors()) {
858 LiveRegs.addLiveOuts(*Pred);
859 MachineBasicBlock::iterator InsertBefore = Pred->getFirstTerminator();
860 for (Register Reg : NewLiveIns) {
861 if (!LiveRegs.available(*MRI, Reg))
864 BuildMI(*Pred, InsertBefore, DL, TII->get(TargetOpcode::IMPLICIT_DEF),
870 addLiveIns(*MBB, NewLiveIns);
874 // See if any of the blocks in MergePotentials (which all have SuccBB as a
875 // successor, or all have no successor if it is null) can be tail-merged.
876 // If there is a successor, any blocks in MergePotentials that are not
877 // tail-merged and are not immediately before Succ must have an unconditional
878 // branch to Succ added (but the predecessor/successor lists need no
879 // adjustment). The lone predecessor of Succ that falls through into Succ,
880 // if any, is given in PredBB.
881 // MinCommonTailLength - Except for the special cases below, tail-merge if
882 // there are at least this many instructions in common.
883 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
884 MachineBasicBlock *PredBB,
885 unsigned MinCommonTailLength) {
886 bool MadeChange = false;
889 dbgs() << "\nTryTailMergeBlocks: ";
890 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) dbgs()
891 << printMBBReference(*MergePotentials[i].getBlock())
892 << (i == e - 1 ? "" : ", ");
893 dbgs() << "\n"; if (SuccBB) {
894 dbgs() << " with successor " << printMBBReference(*SuccBB) << '\n';
896 dbgs() << " which has fall-through from "
897 << printMBBReference(*PredBB) << "\n";
898 } dbgs() << "Looking for common tails of at least "
899 << MinCommonTailLength << " instruction"
900 << (MinCommonTailLength == 1 ? "" : "s") << '\n';);
902 // Sort by hash value so that blocks with identical end sequences sort
904 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
906 // Walk through equivalence sets looking for actual exact matches.
907 while (MergePotentials.size() > 1) {
908 unsigned CurHash = MergePotentials.back().getHash();
910 // Build SameTails, identifying the set of blocks with this hash code
911 // and with the maximum number of instructions in common.
912 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
916 // If we didn't find any pair that has at least MinCommonTailLength
917 // instructions in common, remove all blocks with this hash code and retry.
918 if (SameTails.empty()) {
919 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
923 // If one of the blocks is the entire common tail (and is not the entry
924 // block/an EH pad, which we can't jump to), we can treat all blocks with
925 // this same tail at once. Use PredBB if that is one of the possibilities,
926 // as that will not introduce any extra branches.
927 MachineBasicBlock *EntryBB =
928 &MergePotentials.front().getBlock()->getParent()->front();
929 unsigned commonTailIndex = SameTails.size();
930 // If there are two blocks, check to see if one can be made to fall through
932 if (SameTails.size() == 2 &&
933 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
934 SameTails[1].tailIsWholeBlock() && !SameTails[1].getBlock()->isEHPad())
936 else if (SameTails.size() == 2 &&
937 SameTails[1].getBlock()->isLayoutSuccessor(
938 SameTails[0].getBlock()) &&
939 SameTails[0].tailIsWholeBlock() &&
940 !SameTails[0].getBlock()->isEHPad())
943 // Otherwise just pick one, favoring the fall-through predecessor if
945 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
946 MachineBasicBlock *MBB = SameTails[i].getBlock();
947 if ((MBB == EntryBB || MBB->isEHPad()) &&
948 SameTails[i].tailIsWholeBlock())
954 if (SameTails[i].tailIsWholeBlock())
959 if (commonTailIndex == SameTails.size() ||
960 (SameTails[commonTailIndex].getBlock() == PredBB &&
961 !SameTails[commonTailIndex].tailIsWholeBlock())) {
962 // None of the blocks consist entirely of the common tail.
963 // Split a block so that one does.
964 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
965 maxCommonTailLength, commonTailIndex)) {
966 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
971 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
973 // Recompute common tail MBB's edge weights and block frequency.
974 setCommonTailEdgeWeights(*MBB);
976 // Merge debug locations, MMOs and undef flags across identical instructions
978 mergeCommonTails(commonTailIndex);
980 // MBB is common tail. Adjust all other BB's to jump to this one.
981 // Traversal must be forwards so erases work.
982 LLVM_DEBUG(dbgs() << "\nUsing common tail in " << printMBBReference(*MBB)
984 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
985 if (commonTailIndex == i)
987 LLVM_DEBUG(dbgs() << printMBBReference(*SameTails[i].getBlock())
988 << (i == e - 1 ? "" : ", "));
989 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
990 replaceTailWithBranchTo(SameTails[i].getTailStartPos(), *MBB);
991 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
992 MergePotentials.erase(SameTails[i].getMPIter());
994 LLVM_DEBUG(dbgs() << "\n");
995 // We leave commonTailIndex in the worklist in case there are other blocks
996 // that match it with a smaller number of instructions.
1002 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
1003 bool MadeChange = false;
1004 if (!EnableTailMerge)
1007 // First find blocks with no successors.
1008 // Block placement may create new tail merging opportunities for these blocks.
1009 MergePotentials.clear();
1010 for (MachineBasicBlock &MBB : MF) {
1011 if (MergePotentials.size() == TailMergeThreshold)
1013 if (!TriedMerging.count(&MBB) && MBB.succ_empty())
1014 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB));
1017 // If this is a large problem, avoid visiting the same basic blocks
1019 if (MergePotentials.size() == TailMergeThreshold)
1020 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1021 TriedMerging.insert(MergePotentials[i].getBlock());
1023 // See if we can do any tail merging on those.
1024 if (MergePotentials.size() >= 2)
1025 MadeChange |= TryTailMergeBlocks(nullptr, nullptr, MinCommonTailLength);
1027 // Look at blocks (IBB) with multiple predecessors (PBB).
1028 // We change each predecessor to a canonical form, by
1029 // (1) temporarily removing any unconditional branch from the predecessor
1031 // (2) alter conditional branches so they branch to the other block
1032 // not IBB; this may require adding back an unconditional branch to IBB
1033 // later, where there wasn't one coming in. E.g.
1035 // fallthrough to QBB
1038 // with a conceptual B to IBB after that, which never actually exists.
1039 // With those changes, we see whether the predecessors' tails match,
1040 // and merge them if so. We change things out of canonical form and
1041 // back to the way they were later in the process. (OptimizeBranches
1042 // would undo some of this, but we can't use it, because we'd get into
1043 // a compile-time infinite loop repeatedly doing and undoing the same
1044 // transformations.)
1046 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1048 if (I->pred_size() < 2) continue;
1049 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
1050 MachineBasicBlock *IBB = &*I;
1051 MachineBasicBlock *PredBB = &*std::prev(I);
1052 MergePotentials.clear();
1055 // Bail if merging after placement and IBB is the loop header because
1056 // -- If merging predecessors that belong to the same loop as IBB, the
1057 // common tail of merged predecessors may become the loop top if block
1058 // placement is called again and the predecessors may branch to this common
1059 // tail and require more branches. This can be relaxed if
1060 // MachineBlockPlacement::findBestLoopTop is more flexible.
1061 // --If merging predecessors that do not belong to the same loop as IBB, the
1062 // loop info of IBB's loop and the other loops may be affected. Calling the
1063 // block placement again may make big change to the layout and eliminate the
1064 // reason to do tail merging here.
1065 if (AfterBlockPlacement && MLI) {
1066 ML = MLI->getLoopFor(IBB);
1067 if (ML && IBB == ML->getHeader())
1071 for (MachineBasicBlock *PBB : I->predecessors()) {
1072 if (MergePotentials.size() == TailMergeThreshold)
1075 if (TriedMerging.count(PBB))
1078 // Skip blocks that loop to themselves, can't tail merge these.
1082 // Visit each predecessor only once.
1083 if (!UniquePreds.insert(PBB).second)
1086 // Skip blocks which may jump to a landing pad or jump from an asm blob.
1087 // Can't tail merge these.
1088 if (PBB->hasEHPadSuccessor() || PBB->mayHaveInlineAsmBr())
1091 // After block placement, only consider predecessors that belong to the
1092 // same loop as IBB. The reason is the same as above when skipping loop
1094 if (AfterBlockPlacement && MLI)
1095 if (ML != MLI->getLoopFor(PBB))
1098 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1099 SmallVector<MachineOperand, 4> Cond;
1100 if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) {
1101 // Failing case: IBB is the target of a cbr, and we cannot reverse the
1103 SmallVector<MachineOperand, 4> NewCond(Cond);
1104 if (!Cond.empty() && TBB == IBB) {
1105 if (TII->reverseBranchCondition(NewCond))
1107 // This is the QBB case described above
1109 auto Next = ++PBB->getIterator();
1110 if (Next != MF.end())
1115 // Remove the unconditional branch at the end, if any.
1116 if (TBB && (Cond.empty() || FBB)) {
1117 DebugLoc dl = PBB->findBranchDebugLoc();
1118 TII->removeBranch(*PBB);
1120 // reinsert conditional branch only, for now
1121 TII->insertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1125 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB));
1129 // If this is a large problem, avoid visiting the same basic blocks multiple
1131 if (MergePotentials.size() == TailMergeThreshold)
1132 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1133 TriedMerging.insert(MergePotentials[i].getBlock());
1135 if (MergePotentials.size() >= 2)
1136 MadeChange |= TryTailMergeBlocks(IBB, PredBB, MinCommonTailLength);
1138 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1139 // result of removing blocks in TryTailMergeBlocks.
1140 PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks
1141 if (MergePotentials.size() == 1 &&
1142 MergePotentials.begin()->getBlock() != PredBB)
1143 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1149 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1150 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1151 BlockFrequency AccumulatedMBBFreq;
1153 // Aggregate edge frequency of successor edge j:
1154 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1155 // where bb is a basic block that is in SameTails.
1156 for (const auto &Src : SameTails) {
1157 const MachineBasicBlock *SrcMBB = Src.getBlock();
1158 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1159 AccumulatedMBBFreq += BlockFreq;
1161 // It is not necessary to recompute edge weights if TailBB has less than two
1163 if (TailMBB.succ_size() <= 1)
1166 auto EdgeFreq = EdgeFreqLs.begin();
1168 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1169 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1170 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1173 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1175 if (TailMBB.succ_size() <= 1)
1179 std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0))
1181 auto EdgeFreq = EdgeFreqLs.begin();
1183 if (SumEdgeFreq > 0) {
1184 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1185 SuccI != SuccE; ++SuccI, ++EdgeFreq) {
1186 auto Prob = BranchProbability::getBranchProbability(
1187 EdgeFreq->getFrequency(), SumEdgeFreq);
1188 TailMBB.setSuccProbability(SuccI, Prob);
1193 //===----------------------------------------------------------------------===//
1194 // Branch Optimization
1195 //===----------------------------------------------------------------------===//
1197 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1198 bool MadeChange = false;
1200 // Make sure blocks are numbered in order
1201 MF.RenumberBlocks();
1202 // Renumbering blocks alters EH scope membership, recalculate it.
1203 EHScopeMembership = getEHScopeMembership(MF);
1205 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1207 MachineBasicBlock *MBB = &*I++;
1208 MadeChange |= OptimizeBlock(MBB);
1210 // If it is dead, remove it.
1211 if (MBB->pred_empty()) {
1212 RemoveDeadBlock(MBB);
1221 // Blocks should be considered empty if they contain only debug info;
1222 // else the debug info would affect codegen.
1223 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1224 return MBB->getFirstNonDebugInstr() == MBB->end();
1227 // Blocks with only debug info and branches should be considered the same
1228 // as blocks with only branches.
1229 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1230 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1231 assert(I != MBB->end() && "empty block!");
1232 return I->isBranch();
1235 /// IsBetterFallthrough - Return true if it would be clearly better to
1236 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1237 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1238 /// result in infinite loops.
1239 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1240 MachineBasicBlock *MBB2) {
1241 assert(MBB1 && MBB2 && "Unknown MachineBasicBlock");
1243 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1244 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1245 // optimize branches that branch to either a return block or an assert block
1246 // into a fallthrough to the return.
1247 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1248 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1249 if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1252 // If there is a clear successor ordering we make sure that one block
1253 // will fall through to the next
1254 if (MBB1->isSuccessor(MBB2)) return true;
1255 if (MBB2->isSuccessor(MBB1)) return false;
1257 return MBB2I->isCall() && !MBB1I->isCall();
1260 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1261 /// instructions on the block.
1262 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1263 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1264 if (I != MBB.end() && I->isBranch())
1265 return I->getDebugLoc();
1269 static void copyDebugInfoToPredecessor(const TargetInstrInfo *TII,
1270 MachineBasicBlock &MBB,
1271 MachineBasicBlock &PredMBB) {
1272 auto InsertBefore = PredMBB.getFirstTerminator();
1273 for (MachineInstr &MI : MBB.instrs())
1274 if (MI.isDebugInstr()) {
1275 TII->duplicate(PredMBB, InsertBefore, MI);
1276 LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to pred: "
1281 static void copyDebugInfoToSuccessor(const TargetInstrInfo *TII,
1282 MachineBasicBlock &MBB,
1283 MachineBasicBlock &SuccMBB) {
1284 auto InsertBefore = SuccMBB.SkipPHIsAndLabels(SuccMBB.begin());
1285 for (MachineInstr &MI : MBB.instrs())
1286 if (MI.isDebugInstr()) {
1287 TII->duplicate(SuccMBB, InsertBefore, MI);
1288 LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to succ: "
1293 // Try to salvage DBG_VALUE instructions from an otherwise empty block. If such
1294 // a basic block is removed we would lose the debug information unless we have
1295 // copied the information to a predecessor/successor.
1297 // TODO: This function only handles some simple cases. An alternative would be
1298 // to run a heavier analysis, such as the LiveDebugValues pass, before we do
1300 static void salvageDebugInfoFromEmptyBlock(const TargetInstrInfo *TII,
1301 MachineBasicBlock &MBB) {
1302 assert(IsEmptyBlock(&MBB) && "Expected an empty block (except debug info).");
1303 // If this MBB is the only predecessor of a successor it is legal to copy
1304 // DBG_VALUE instructions to the beginning of the successor.
1305 for (MachineBasicBlock *SuccBB : MBB.successors())
1306 if (SuccBB->pred_size() == 1)
1307 copyDebugInfoToSuccessor(TII, MBB, *SuccBB);
1308 // If this MBB is the only successor of a predecessor it is legal to copy the
1309 // DBG_VALUE instructions to the end of the predecessor (just before the
1310 // terminators, assuming that the terminator isn't affecting the DBG_VALUE).
1311 for (MachineBasicBlock *PredBB : MBB.predecessors())
1312 if (PredBB->succ_size() == 1)
1313 copyDebugInfoToPredecessor(TII, MBB, *PredBB);
1316 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1317 bool MadeChange = false;
1318 MachineFunction &MF = *MBB->getParent();
1321 MachineFunction::iterator FallThrough = MBB->getIterator();
1324 // Make sure MBB and FallThrough belong to the same EH scope.
1325 bool SameEHScope = true;
1326 if (!EHScopeMembership.empty() && FallThrough != MF.end()) {
1327 auto MBBEHScope = EHScopeMembership.find(MBB);
1328 assert(MBBEHScope != EHScopeMembership.end());
1329 auto FallThroughEHScope = EHScopeMembership.find(&*FallThrough);
1330 assert(FallThroughEHScope != EHScopeMembership.end());
1331 SameEHScope = MBBEHScope->second == FallThroughEHScope->second;
1334 // Analyze the branch in the current block. As a side-effect, this may cause
1335 // the block to become empty.
1336 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1337 SmallVector<MachineOperand, 4> CurCond;
1338 bool CurUnAnalyzable =
1339 TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1341 // If this block is empty, make everyone use its fall-through, not the block
1342 // explicitly. Landing pads should not do this since the landing-pad table
1343 // points to this block. Blocks with their addresses taken shouldn't be
1345 if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() &&
1347 salvageDebugInfoFromEmptyBlock(TII, *MBB);
1348 // Dead block? Leave for cleanup later.
1349 if (MBB->pred_empty()) return MadeChange;
1351 if (FallThrough == MF.end()) {
1352 // TODO: Simplify preds to not branch here if possible!
1353 } else if (FallThrough->isEHPad()) {
1354 // Don't rewrite to a landing pad fallthough. That could lead to the case
1355 // where a BB jumps to more than one landing pad.
1356 // TODO: Is it ever worth rewriting predecessors which don't already
1357 // jump to a landing pad, and so can safely jump to the fallthrough?
1358 } else if (MBB->isSuccessor(&*FallThrough)) {
1359 // Rewrite all predecessors of the old block to go to the fallthrough
1361 while (!MBB->pred_empty()) {
1362 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1363 Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough);
1365 // If MBB was the target of a jump table, update jump tables to go to the
1366 // fallthrough instead.
1367 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1368 MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough);
1374 // Check to see if we can simplify the terminator of the block before this
1376 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1378 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1379 SmallVector<MachineOperand, 4> PriorCond;
1380 bool PriorUnAnalyzable =
1381 TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1382 if (!PriorUnAnalyzable) {
1383 // If the previous branch is conditional and both conditions go to the same
1384 // destination, remove the branch, replacing it with an unconditional one or
1386 if (PriorTBB && PriorTBB == PriorFBB) {
1387 DebugLoc dl = getBranchDebugLoc(PrevBB);
1388 TII->removeBranch(PrevBB);
1390 if (PriorTBB != MBB)
1391 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1394 goto ReoptimizeBlock;
1397 // If the previous block unconditionally falls through to this block and
1398 // this block has no other predecessors, move the contents of this block
1399 // into the prior block. This doesn't usually happen when SimplifyCFG
1400 // has been used, but it can happen if tail merging splits a fall-through
1401 // predecessor of a block.
1402 // This has to check PrevBB->succ_size() because EH edges are ignored by
1404 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1405 PrevBB.succ_size() == 1 &&
1406 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1407 LLVM_DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1408 << "From MBB: " << *MBB);
1409 // Remove redundant DBG_VALUEs first.
1410 if (PrevBB.begin() != PrevBB.end()) {
1411 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1413 MachineBasicBlock::iterator MBBIter = MBB->begin();
1414 // Check if DBG_VALUE at the end of PrevBB is identical to the
1415 // DBG_VALUE at the beginning of MBB.
1416 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1417 && PrevBBIter->isDebugInstr() && MBBIter->isDebugInstr()) {
1418 if (!MBBIter->isIdenticalTo(*PrevBBIter))
1420 MachineInstr &DuplicateDbg = *MBBIter;
1421 ++MBBIter; -- PrevBBIter;
1422 DuplicateDbg.eraseFromParent();
1425 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1426 PrevBB.removeSuccessor(PrevBB.succ_begin());
1427 assert(PrevBB.succ_empty());
1428 PrevBB.transferSuccessors(MBB);
1433 // If the previous branch *only* branches to *this* block (conditional or
1434 // not) remove the branch.
1435 if (PriorTBB == MBB && !PriorFBB) {
1436 TII->removeBranch(PrevBB);
1439 goto ReoptimizeBlock;
1442 // If the prior block branches somewhere else on the condition and here if
1443 // the condition is false, remove the uncond second branch.
1444 if (PriorFBB == MBB) {
1445 DebugLoc dl = getBranchDebugLoc(PrevBB);
1446 TII->removeBranch(PrevBB);
1447 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1450 goto ReoptimizeBlock;
1453 // If the prior block branches here on true and somewhere else on false, and
1454 // if the branch condition is reversible, reverse the branch to create a
1456 if (PriorTBB == MBB) {
1457 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1458 if (!TII->reverseBranchCondition(NewPriorCond)) {
1459 DebugLoc dl = getBranchDebugLoc(PrevBB);
1460 TII->removeBranch(PrevBB);
1461 TII->insertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1464 goto ReoptimizeBlock;
1468 // If this block has no successors (e.g. it is a return block or ends with
1469 // a call to a no-return function like abort or __cxa_throw) and if the pred
1470 // falls through into this block, and if it would otherwise fall through
1471 // into the block after this, move this block to the end of the function.
1473 // We consider it more likely that execution will stay in the function (e.g.
1474 // due to loops) than it is to exit it. This asserts in loops etc, moving
1475 // the assert condition out of the loop body.
1476 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1477 MachineFunction::iterator(PriorTBB) == FallThrough &&
1478 !MBB->canFallThrough()) {
1479 bool DoTransform = true;
1481 // We have to be careful that the succs of PredBB aren't both no-successor
1482 // blocks. If neither have successors and if PredBB is the second from
1483 // last block in the function, we'd just keep swapping the two blocks for
1484 // last. Only do the swap if one is clearly better to fall through than
1486 if (FallThrough == --MF.end() &&
1487 !IsBetterFallthrough(PriorTBB, MBB))
1488 DoTransform = false;
1491 // Reverse the branch so we will fall through on the previous true cond.
1492 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1493 if (!TII->reverseBranchCondition(NewPriorCond)) {
1494 LLVM_DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1495 << "To make fallthrough to: " << *PriorTBB << "\n");
1497 DebugLoc dl = getBranchDebugLoc(PrevBB);
1498 TII->removeBranch(PrevBB);
1499 TII->insertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1501 // Move this block to the end of the function.
1502 MBB->moveAfter(&MF.back());
1512 MF.getFunction().hasOptSize() ||
1513 llvm::shouldOptimizeForSize(MBB, PSI, &MBBFreqInfo);
1514 if (!IsEmptyBlock(MBB) && MBB->pred_size() == 1 && OptForSize) {
1515 // Changing "Jcc foo; foo: jmp bar;" into "Jcc bar;" might change the branch
1516 // direction, thereby defeating careful block placement and regressing
1517 // performance. Therefore, only consider this for optsize functions.
1518 MachineInstr &TailCall = *MBB->getFirstNonDebugInstr();
1519 if (TII->isUnconditionalTailCall(TailCall)) {
1520 MachineBasicBlock *Pred = *MBB->pred_begin();
1521 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1522 SmallVector<MachineOperand, 4> PredCond;
1523 bool PredAnalyzable =
1524 !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true);
1526 if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB &&
1527 PredTBB != PredFBB) {
1528 // The predecessor has a conditional branch to this block which consists
1529 // of only a tail call. Try to fold the tail call into the conditional
1531 if (TII->canMakeTailCallConditional(PredCond, TailCall)) {
1532 // TODO: It would be nice if analyzeBranch() could provide a pointer
1533 // to the branch instruction so replaceBranchWithTailCall() doesn't
1534 // have to search for it.
1535 TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall);
1537 Pred->removeSuccessor(MBB);
1542 // If the predecessor is falling through to this block, we could reverse
1543 // the branch condition and fold the tail call into that. However, after
1544 // that we might have to re-arrange the CFG to fall through to the other
1545 // block and there is a high risk of regressing code size rather than
1550 if (!CurUnAnalyzable) {
1551 // If this is a two-way branch, and the FBB branches to this block, reverse
1552 // the condition so the single-basic-block loop is faster. Instead of:
1553 // Loop: xxx; jcc Out; jmp Loop
1555 // Loop: xxx; jncc Loop; jmp Out
1556 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1557 SmallVector<MachineOperand, 4> NewCond(CurCond);
1558 if (!TII->reverseBranchCondition(NewCond)) {
1559 DebugLoc dl = getBranchDebugLoc(*MBB);
1560 TII->removeBranch(*MBB);
1561 TII->insertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1564 goto ReoptimizeBlock;
1568 // If this branch is the only thing in its block, see if we can forward
1569 // other blocks across it.
1570 if (CurTBB && CurCond.empty() && !CurFBB &&
1571 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1572 !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1573 DebugLoc dl = getBranchDebugLoc(*MBB);
1574 // This block may contain just an unconditional branch. Because there can
1575 // be 'non-branch terminators' in the block, try removing the branch and
1576 // then seeing if the block is empty.
1577 TII->removeBranch(*MBB);
1578 // If the only things remaining in the block are debug info, remove these
1579 // as well, so this will behave the same as an empty block in non-debug
1581 if (IsEmptyBlock(MBB)) {
1582 // Make the block empty, losing the debug info (we could probably
1583 // improve this in some cases.)
1584 MBB->erase(MBB->begin(), MBB->end());
1586 // If this block is just an unconditional branch to CurTBB, we can
1587 // usually completely eliminate the block. The only case we cannot
1588 // completely eliminate the block is when the block before this one
1589 // falls through into MBB and we can't understand the prior block's branch
1592 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1593 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1594 !PrevBB.isSuccessor(MBB)) {
1595 // If the prior block falls through into us, turn it into an
1596 // explicit branch to us to make updates simpler.
1597 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1598 PriorTBB != MBB && PriorFBB != MBB) {
1600 assert(PriorCond.empty() && !PriorFBB &&
1601 "Bad branch analysis");
1604 assert(!PriorFBB && "Machine CFG out of date!");
1607 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1608 TII->removeBranch(PrevBB);
1609 TII->insertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1612 // Iterate through all the predecessors, revectoring each in-turn.
1614 bool DidChange = false;
1615 bool HasBranchToSelf = false;
1616 while(PI != MBB->pred_size()) {
1617 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1619 // If this block has an uncond branch to itself, leave it.
1621 HasBranchToSelf = true;
1624 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1625 // If this change resulted in PMBB ending in a conditional
1626 // branch where both conditions go to the same destination,
1627 // change this to an unconditional branch.
1628 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1629 SmallVector<MachineOperand, 4> NewCurCond;
1630 bool NewCurUnAnalyzable = TII->analyzeBranch(
1631 *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true);
1632 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1633 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1634 TII->removeBranch(*PMBB);
1636 TII->insertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1643 // Change any jumptables to go to the new MBB.
1644 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1645 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1649 if (!HasBranchToSelf) return MadeChange;
1654 // Add the branch back if the block is more than just an uncond branch.
1655 TII->insertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1659 // If the prior block doesn't fall through into this block, and if this
1660 // block doesn't fall through into some other block, see if we can find a
1661 // place to move this block where a fall-through will happen.
1662 if (!PrevBB.canFallThrough()) {
1663 // Now we know that there was no fall-through into this block, check to
1664 // see if it has a fall-through into its successor.
1665 bool CurFallsThru = MBB->canFallThrough();
1667 if (!MBB->isEHPad()) {
1668 // Check all the predecessors of this block. If one of them has no fall
1669 // throughs, and analyzeBranch thinks it _could_ fallthrough to this
1670 // block, move this block right after it.
1671 for (MachineBasicBlock *PredBB : MBB->predecessors()) {
1672 // Analyze the branch at the end of the pred.
1673 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1674 SmallVector<MachineOperand, 4> PredCond;
1675 if (PredBB != MBB && !PredBB->canFallThrough() &&
1676 !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) &&
1677 (PredTBB == MBB || PredFBB == MBB) &&
1678 (!CurFallsThru || !CurTBB || !CurFBB) &&
1679 (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1680 // If the current block doesn't fall through, just move it.
1681 // If the current block can fall through and does not end with a
1682 // conditional branch, we need to append an unconditional jump to
1683 // the (current) next block. To avoid a possible compile-time
1684 // infinite loop, move blocks only backward in this case.
1685 // Also, if there are already 2 branches here, we cannot add a third;
1686 // this means we have the case
1691 MachineBasicBlock *NextBB = &*std::next(MBB->getIterator());
1693 TII->insertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1695 MBB->moveAfter(PredBB);
1697 goto ReoptimizeBlock;
1702 if (!CurFallsThru) {
1703 // Check analyzable branch-successors to see if we can move this block
1705 if (!CurUnAnalyzable) {
1706 for (MachineBasicBlock *SuccBB : {CurFBB, CurTBB}) {
1709 // Analyze the branch at the end of the block before the succ.
1710 MachineFunction::iterator SuccPrev = --SuccBB->getIterator();
1712 // If this block doesn't already fall-through to that successor, and
1713 // if the succ doesn't already have a block that can fall through into
1714 // it, we can arrange for the fallthrough to happen.
1715 if (SuccBB != MBB && &*SuccPrev != MBB &&
1716 !SuccPrev->canFallThrough()) {
1717 MBB->moveBefore(SuccBB);
1719 goto ReoptimizeBlock;
1724 // Okay, there is no really great place to put this block. If, however,
1725 // the block before this one would be a fall-through if this block were
1726 // removed, move this block to the end of the function. There is no real
1727 // advantage in "falling through" to an EH block, so we don't want to
1728 // perform this transformation for that case.
1730 // Also, Windows EH introduced the possibility of an arbitrary number of
1731 // successors to a given block. The analyzeBranch call does not consider
1732 // exception handling and so we can get in a state where a block
1733 // containing a call is followed by multiple EH blocks that would be
1734 // rotated infinitely at the end of the function if the transformation
1735 // below were performed for EH "FallThrough" blocks. Therefore, even if
1736 // that appears not to be happening anymore, we should assume that it is
1737 // possible and not remove the "!FallThrough()->isEHPad" condition below.
1738 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1739 SmallVector<MachineOperand, 4> PrevCond;
1740 if (FallThrough != MF.end() &&
1741 !FallThrough->isEHPad() &&
1742 !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1743 PrevBB.isSuccessor(&*FallThrough)) {
1744 MBB->moveAfter(&MF.back());
1754 //===----------------------------------------------------------------------===//
1755 // Hoist Common Code
1756 //===----------------------------------------------------------------------===//
1758 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1759 bool MadeChange = false;
1760 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1761 MachineBasicBlock *MBB = &*I++;
1762 MadeChange |= HoistCommonCodeInSuccs(MBB);
1768 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1769 /// its 'true' successor.
1770 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1771 MachineBasicBlock *TrueBB) {
1772 for (MachineBasicBlock *SuccBB : BB->successors())
1773 if (SuccBB != TrueBB)
1778 template <class Container>
1779 static void addRegAndItsAliases(Register Reg, const TargetRegisterInfo *TRI,
1781 if (Reg.isPhysical()) {
1782 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1789 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1790 /// in successors to. The location is usually just before the terminator,
1791 /// however if the terminator is a conditional branch and its previous
1792 /// instruction is the flag setting instruction, the previous instruction is
1793 /// the preferred location. This function also gathers uses and defs of the
1794 /// instructions from the insertion point to the end of the block. The data is
1795 /// used by HoistCommonCodeInSuccs to ensure safety.
1797 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1798 const TargetInstrInfo *TII,
1799 const TargetRegisterInfo *TRI,
1800 SmallSet<Register, 4> &Uses,
1801 SmallSet<Register, 4> &Defs) {
1802 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1803 if (!TII->isUnpredicatedTerminator(*Loc))
1806 for (const MachineOperand &MO : Loc->operands()) {
1809 Register Reg = MO.getReg();
1813 addRegAndItsAliases(Reg, TRI, Uses);
1816 // Don't try to hoist code in the rare case the terminator defines a
1817 // register that is later used.
1820 // If the terminator defines a register, make sure we don't hoist
1821 // the instruction whose def might be clobbered by the terminator.
1822 addRegAndItsAliases(Reg, TRI, Defs);
1828 // If the terminator is the only instruction in the block and Uses is not
1829 // empty (or we would have returned above), we can still safely hoist
1830 // instructions just before the terminator as long as the Defs/Uses are not
1831 // violated (which is checked in HoistCommonCodeInSuccs).
1832 if (Loc == MBB->begin())
1835 // The terminator is probably a conditional branch, try not to separate the
1836 // branch from condition setting instruction.
1837 MachineBasicBlock::iterator PI = prev_nodbg(Loc, MBB->begin());
1840 for (const MachineOperand &MO : PI->operands()) {
1841 // If PI has a regmask operand, it is probably a call. Separate away.
1844 if (!MO.isReg() || MO.isUse())
1846 Register Reg = MO.getReg();
1849 if (Uses.count(Reg)) {
1855 // The condition setting instruction is not just before the conditional
1859 // Be conservative, don't insert instruction above something that may have
1860 // side-effects. And since it's potentially bad to separate flag setting
1861 // instruction from the conditional branch, just abort the optimization
1863 // Also avoid moving code above predicated instruction since it's hard to
1864 // reason about register liveness with predicated instruction.
1865 bool DontMoveAcrossStore = true;
1866 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI))
1869 // Find out what registers are live. Note this routine is ignoring other live
1870 // registers which are only used by instructions in successor blocks.
1871 for (const MachineOperand &MO : PI->operands()) {
1874 Register Reg = MO.getReg();
1878 addRegAndItsAliases(Reg, TRI, Uses);
1880 if (Uses.erase(Reg)) {
1881 if (Register::isPhysicalRegister(Reg)) {
1882 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1883 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1886 addRegAndItsAliases(Reg, TRI, Defs);
1893 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1894 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1895 SmallVector<MachineOperand, 4> Cond;
1896 if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1899 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1901 // Malformed bcc? True and false blocks are the same?
1904 // Restrict the optimization to cases where MBB is the only predecessor,
1905 // it is an obvious win.
1906 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1909 // Find a suitable position to hoist the common instructions to. Also figure
1910 // out which registers are used or defined by instructions from the insertion
1911 // point to the end of the block.
1912 SmallSet<Register, 4> Uses, Defs;
1913 MachineBasicBlock::iterator Loc =
1914 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1915 if (Loc == MBB->end())
1918 bool HasDups = false;
1919 SmallSet<Register, 4> ActiveDefsSet, AllDefsSet;
1920 MachineBasicBlock::iterator TIB = TBB->begin();
1921 MachineBasicBlock::iterator FIB = FBB->begin();
1922 MachineBasicBlock::iterator TIE = TBB->end();
1923 MachineBasicBlock::iterator FIE = FBB->end();
1924 while (TIB != TIE && FIB != FIE) {
1925 // Skip dbg_value instructions. These do not count.
1926 TIB = skipDebugInstructionsForward(TIB, TIE);
1927 FIB = skipDebugInstructionsForward(FIB, FIE);
1928 if (TIB == TIE || FIB == FIE)
1931 if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead))
1934 if (TII->isPredicated(*TIB))
1935 // Hard to reason about register liveness with predicated instruction.
1939 for (MachineOperand &MO : TIB->operands()) {
1940 // Don't attempt to hoist instructions with register masks.
1941 if (MO.isRegMask()) {
1947 Register Reg = MO.getReg();
1951 if (Uses.count(Reg)) {
1952 // Avoid clobbering a register that's used by the instruction at
1953 // the point of insertion.
1958 if (Defs.count(Reg) && !MO.isDead()) {
1959 // Don't hoist the instruction if the def would be clobber by the
1960 // instruction at the point insertion. FIXME: This is overly
1961 // conservative. It should be possible to hoist the instructions
1962 // in BB2 in the following example:
1964 // r1, eflag = op1 r2, r3
1973 } else if (!ActiveDefsSet.count(Reg)) {
1974 if (Defs.count(Reg)) {
1975 // Use is defined by the instruction at the point of insertion.
1980 if (MO.isKill() && Uses.count(Reg))
1981 // Kills a register that's read by the instruction at the point of
1982 // insertion. Remove the kill marker.
1983 MO.setIsKill(false);
1989 bool DontMoveAcrossStore = true;
1990 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1993 // Remove kills from ActiveDefsSet, these registers had short live ranges.
1994 for (const MachineOperand &MO : TIB->operands()) {
1995 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1997 Register Reg = MO.getReg();
2000 if (!AllDefsSet.count(Reg)) {
2003 if (Reg.isPhysical()) {
2004 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
2005 ActiveDefsSet.erase(*AI);
2007 ActiveDefsSet.erase(Reg);
2011 // Track local defs so we can update liveins.
2012 for (const MachineOperand &MO : TIB->operands()) {
2013 if (!MO.isReg() || !MO.isDef() || MO.isDead())
2015 Register Reg = MO.getReg();
2016 if (!Reg || Reg.isVirtual())
2018 addRegAndItsAliases(Reg, TRI, ActiveDefsSet);
2019 addRegAndItsAliases(Reg, TRI, AllDefsSet);
2030 MBB->splice(Loc, TBB, TBB->begin(), TIB);
2031 FBB->erase(FBB->begin(), FIB);
2033 if (UpdateLiveIns) {
2034 recomputeLiveIns(*TBB);
2035 recomputeLiveIns(*FBB);