xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/BranchFolding.cpp (revision ebacd8013fe5f7fdf9f6a5b286f6680dd2891036)
1 //===- BranchFolding.cpp - Fold machine code branch instructions ----------===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
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
11 // it then removes.
12 //
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
15 // ISA (e.g. NVPTX).
16 //
17 //===----------------------------------------------------------------------===//
18 
19 #include "BranchFolding.h"
20 #include "llvm/ADT/BitVector.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/ProfileSummaryInfo.h"
26 #include "llvm/CodeGen/Analysis.h"
27 #include "llvm/CodeGen/MBFIWrapper.h"
28 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
29 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineFunctionPass.h"
32 #include "llvm/CodeGen/MachineInstr.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineLoopInfo.h"
36 #include "llvm/CodeGen/MachineOperand.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/MachineSizeOpts.h"
39 #include "llvm/CodeGen/TargetInstrInfo.h"
40 #include "llvm/CodeGen/TargetOpcodes.h"
41 #include "llvm/CodeGen/TargetPassConfig.h"
42 #include "llvm/CodeGen/TargetRegisterInfo.h"
43 #include "llvm/CodeGen/TargetSubtargetInfo.h"
44 #include "llvm/IR/DebugInfoMetadata.h"
45 #include "llvm/IR/DebugLoc.h"
46 #include "llvm/IR/Function.h"
47 #include "llvm/InitializePasses.h"
48 #include "llvm/MC/LaneBitmask.h"
49 #include "llvm/MC/MCRegisterInfo.h"
50 #include "llvm/Pass.h"
51 #include "llvm/Support/BlockFrequency.h"
52 #include "llvm/Support/BranchProbability.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/ErrorHandling.h"
56 #include "llvm/Support/raw_ostream.h"
57 #include "llvm/Target/TargetMachine.h"
58 #include <cassert>
59 #include <cstddef>
60 #include <iterator>
61 #include <numeric>
62 
63 using namespace llvm;
64 
65 #define DEBUG_TYPE "branch-folder"
66 
67 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
68 STATISTIC(NumBranchOpts, "Number of branches optimized");
69 STATISTIC(NumTailMerge , "Number of block tails merged");
70 STATISTIC(NumHoist     , "Number of times common instructions are hoisted");
71 STATISTIC(NumTailCalls,  "Number of tail calls optimized");
72 
73 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
74                               cl::init(cl::BOU_UNSET), cl::Hidden);
75 
76 // Throttle for huge numbers of predecessors (compile speed problems)
77 static cl::opt<unsigned>
78 TailMergeThreshold("tail-merge-threshold",
79           cl::desc("Max number of predecessors to consider tail merging"),
80           cl::init(150), cl::Hidden);
81 
82 // Heuristic for tail merging (and, inversely, tail duplication).
83 // TODO: This should be replaced with a target query.
84 static cl::opt<unsigned>
85 TailMergeSize("tail-merge-size",
86               cl::desc("Min number of instructions to consider tail merging"),
87               cl::init(3), cl::Hidden);
88 
89 namespace {
90 
91   /// BranchFolderPass - Wrap branch folder in a machine function pass.
92   class BranchFolderPass : public MachineFunctionPass {
93   public:
94     static char ID;
95 
96     explicit BranchFolderPass(): MachineFunctionPass(ID) {}
97 
98     bool runOnMachineFunction(MachineFunction &MF) override;
99 
100     void getAnalysisUsage(AnalysisUsage &AU) const override {
101       AU.addRequired<MachineBlockFrequencyInfo>();
102       AU.addRequired<MachineBranchProbabilityInfo>();
103       AU.addRequired<ProfileSummaryInfoWrapperPass>();
104       AU.addRequired<TargetPassConfig>();
105       MachineFunctionPass::getAnalysisUsage(AU);
106     }
107 
108     MachineFunctionProperties getRequiredProperties() const override {
109       return MachineFunctionProperties().set(
110           MachineFunctionProperties::Property::NoPHIs);
111     }
112   };
113 
114 } // end anonymous namespace
115 
116 char BranchFolderPass::ID = 0;
117 
118 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
119 
120 INITIALIZE_PASS(BranchFolderPass, DEBUG_TYPE,
121                 "Control Flow Optimizer", false, false)
122 
123 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
124   if (skipFunction(MF.getFunction()))
125     return false;
126 
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   MBFIWrapper MBBFreqInfo(
133       getAnalysis<MachineBlockFrequencyInfo>());
134   BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo,
135                       getAnalysis<MachineBranchProbabilityInfo>(),
136                       &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI());
137   return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
138                                  MF.getSubtarget().getRegisterInfo());
139 }
140 
141 BranchFolder::BranchFolder(bool DefaultEnableTailMerge, bool CommonHoist,
142                            MBFIWrapper &FreqInfo,
143                            const MachineBranchProbabilityInfo &ProbInfo,
144                            ProfileSummaryInfo *PSI, unsigned MinTailLength)
145     : EnableHoistCommonCode(CommonHoist), MinCommonTailLength(MinTailLength),
146       MBBFreqInfo(FreqInfo), MBPI(ProbInfo), PSI(PSI) {
147   if (MinCommonTailLength == 0)
148     MinCommonTailLength = TailMergeSize;
149   switch (FlagEnableTailMerge) {
150   case cl::BOU_UNSET:
151     EnableTailMerge = DefaultEnableTailMerge;
152     break;
153   case cl::BOU_TRUE: EnableTailMerge = true; break;
154   case cl::BOU_FALSE: EnableTailMerge = false; break;
155   }
156 }
157 
158 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
159   assert(MBB->pred_empty() && "MBB must be dead!");
160   LLVM_DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
161 
162   MachineFunction *MF = MBB->getParent();
163   // drop all successors.
164   while (!MBB->succ_empty())
165     MBB->removeSuccessor(MBB->succ_end()-1);
166 
167   // Avoid matching if this pointer gets reused.
168   TriedMerging.erase(MBB);
169 
170   // Update call site info.
171   for (const MachineInstr &MI : *MBB)
172     if (MI.shouldUpdateCallSiteInfo())
173       MF->eraseCallSiteInfo(&MI);
174 
175   // Remove the block.
176   MF->erase(MBB);
177   EHScopeMembership.erase(MBB);
178   if (MLI)
179     MLI->removeBlock(MBB);
180 }
181 
182 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
183                                     const TargetInstrInfo *tii,
184                                     const TargetRegisterInfo *tri,
185                                     MachineLoopInfo *mli, bool AfterPlacement) {
186   if (!tii) return false;
187 
188   TriedMerging.clear();
189 
190   MachineRegisterInfo &MRI = MF.getRegInfo();
191   AfterBlockPlacement = AfterPlacement;
192   TII = tii;
193   TRI = tri;
194   MLI = mli;
195   this->MRI = &MRI;
196 
197   UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF);
198   if (!UpdateLiveIns)
199     MRI.invalidateLiveness();
200 
201   bool MadeChange = false;
202 
203   // Recalculate EH scope membership.
204   EHScopeMembership = getEHScopeMembership(MF);
205 
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
210     // block placement.
211     if (!AfterBlockPlacement || MadeChangeThisIteration)
212       MadeChangeThisIteration |= OptimizeBranches(MF);
213     if (EnableHoistCommonCode)
214       MadeChangeThisIteration |= HoistCommonCode(MF);
215     MadeChange |= MadeChangeThisIteration;
216   }
217 
218   // See if any jump tables have become dead as the code generator
219   // did its thing.
220   MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
221   if (!JTI)
222     return MadeChange;
223 
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;
230 
231         // Remember that this JT is live.
232         JTIsLive.set(Op.getIndex());
233       }
234   }
235 
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);
241       MadeChange = true;
242     }
243 
244   return MadeChange;
245 }
246 
247 //===----------------------------------------------------------------------===//
248 //  Tail Merging of Blocks
249 //===----------------------------------------------------------------------===//
250 
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);
256 
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
259     // later.
260     unsigned OperandHash = 0;
261     switch (Op.getType()) {
262     case MachineOperand::MO_Register:
263       OperandHash = Op.getReg();
264       break;
265     case MachineOperand::MO_Immediate:
266       OperandHash = Op.getImm();
267       break;
268     case MachineOperand::MO_MachineBasicBlock:
269       OperandHash = Op.getMBB()->getNumber();
270       break;
271     case MachineOperand::MO_FrameIndex:
272     case MachineOperand::MO_ConstantPoolIndex:
273     case MachineOperand::MO_JumpTableIndex:
274       OperandHash = Op.getIndex();
275       break;
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();
281       break;
282     default:
283       break;
284     }
285 
286     Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
287   }
288   return Hash;
289 }
290 
291 /// HashEndOfMBB - Hash the last instruction in the MBB.
292 static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) {
293   MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr(false);
294   if (I == MBB.end())
295     return 0;
296 
297   return HashMachineInstr(*I);
298 }
299 
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());
303 }
304 
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()) {
312     --I;
313     if (countsAsInstruction(*I))
314       return I;
315   }
316   return MBB->end();
317 }
318 
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.
323 ///
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();
331 
332   unsigned TailLen = 0;
333   while (true) {
334     MBBI1 = skipBackwardPastNonInstructions(MBBI1, MBB1);
335     MBBI2 = skipBackwardPastNonInstructions(MBBI2, MBB2);
336     if (MBBI1 == MBB1->end() || MBBI2 == MBB2->end())
337       break;
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
343         // directives.
344         MBBI1->isInlineAsm()) {
345       break;
346     }
347     if (MBBI1->getFlag(MachineInstr::NoMerge) ||
348         MBBI2->getFlag(MachineInstr::NoMerge))
349       break;
350     ++TailLen;
351     I1 = MBBI1;
352     I2 = MBBI2;
353   }
354 
355   return TailLen;
356 }
357 
358 void BranchFolder::replaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
359                                            MachineBasicBlock &NewDest) {
360   if (UpdateLiveIns) {
361     // OldInst should always point to an instruction.
362     MachineBasicBlock &OldMBB = *OldInst->getParent();
363     LiveRegs.clear();
364     LiveRegs.addLiveOuts(OldMBB);
365     // Move backward to the place where will insert the jump.
366     MachineBasicBlock::iterator I = OldMBB.end();
367     do {
368       --I;
369       LiveRegs.stepBackward(*I);
370     } while (I != OldInst);
371 
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
374     // register.
375     for (MachineBasicBlock::RegisterMaskPair P : NewDest.liveins()) {
376       // We computed the liveins with computeLiveIn earlier and should only see
377       // full registers:
378       assert(P.LaneMask == LaneBitmask::getAll() &&
379              "Can only handle full register.");
380       MCPhysReg Reg = P.PhysReg;
381       if (!LiveRegs.available(*MRI, Reg))
382         continue;
383       DebugLoc DL;
384       BuildMI(OldMBB, OldInst, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Reg);
385     }
386   }
387 
388   TII->ReplaceTailWithBranchTo(OldInst, &NewDest);
389   ++NumTailMerge;
390 }
391 
392 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
393                                             MachineBasicBlock::iterator BBI1,
394                                             const BasicBlock *BB) {
395   if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
396     return nullptr;
397 
398   MachineFunction &MF = *CurMBB.getParent();
399 
400   // Create the fall-through block.
401   MachineFunction::iterator MBBI = CurMBB.getIterator();
402   MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(BB);
403   CurMBB.getParent()->insert(++MBBI, NewMBB);
404 
405   // Move all the successors of this block to the specified block.
406   NewMBB->transferSuccessors(&CurMBB);
407 
408   // Add an edge from CurMBB to NewMBB for the fall-through.
409   CurMBB.addSuccessor(NewMBB);
410 
411   // Splice the code over.
412   NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
413 
414   // NewMBB belongs to the same loop as CurMBB.
415   if (MLI)
416     if (MachineLoop *ML = MLI->getLoopFor(&CurMBB))
417       ML->addBasicBlockToLoop(NewMBB, MLI->getBase());
418 
419   // NewMBB inherits CurMBB's block frequency.
420   MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
421 
422   if (UpdateLiveIns)
423     computeAndAddLiveIns(LiveRegs, *NewMBB);
424 
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;
430   }
431 
432   return NewMBB;
433 }
434 
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) {
439   unsigned Time = 0;
440   for (; I != E; ++I) {
441     if (!countsAsInstruction(*I))
442       continue;
443     if (I->isCall())
444       Time += 10;
445     else if (I->mayLoadOrStore())
446       Time += 2;
447     else
448       ++Time;
449   }
450   return Time;
451 }
452 
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);
470         return;
471       }
472     }
473   }
474   TII->insertBranch(*CurMBB, SuccBB, nullptr,
475                     SmallVector<MachineOperand, 0>(), dl);
476 }
477 
478 bool
479 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
480   if (getHash() < o.getHash())
481     return true;
482   if (getHash() > o.getHash())
483     return false;
484   if (getBlock()->getNumber() < o.getBlock()->getNumber())
485     return true;
486   if (getBlock()->getNumber() > o.getBlock()->getNumber())
487     return false;
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");
492 #else
493   return false;
494 #endif
495 }
496 
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) {
502   I = MBB->end();
503   unsigned NumTerms = 0;
504   while (true) {
505     if (I == MBB->begin()) {
506       I = MBB->end();
507       break;
508     }
509     --I;
510     if (!I->isTerminator()) break;
511     ++NumTerms;
512   }
513   return NumTerms;
514 }
515 
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())
521     return false;
522   if (MBB->empty())
523     return true;
524   return !(MBB->back().isReturn() || MBB->back().isIndirectBranch());
525 }
526 
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
530 /// in each block.
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
534 ///                 MBB1 and MBB2
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.
543 static bool
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,
550                   bool AfterPlacement,
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)
560       return false;
561   }
562 
563   CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
564   if (CommonTailLen == 0)
565     return false;
566   LLVM_DEBUG(dbgs() << "Common tail length of " << printMBBReference(*MBB1)
567                     << " and " << printMBBReference(*MBB2) << " is "
568                     << CommonTailLen << '\n');
569 
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(), false) == I1)
574     I1 = MBB1->begin();
575   if (skipDebugInstructionsForward(MBB2->begin(), MBB2->end(), false) == I2)
576     I2 = MBB2->begin();
577 
578   bool FullBlockTail1 = I1 == MBB1->begin();
579   bool FullBlockTail2 = I2 == MBB2->begin();
580 
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)
591       return true;
592   }
593 
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))
601     return true;
602 
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)
608     return true;
609   if (MBB2->isLayoutSuccessor(MBB1) && FullBlockTail1)
610     return true;
611 
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_empty() && !MBB->canFallThrough())
619         return false;
620       MachineFunction::iterator I(MBB);
621       MachineFunction *MF = MBB->getParent();
622       return (MBB != &*MF->begin()) && std::prev(I)->canFallThrough();
623     };
624     if (!BothFallThrough(MBB1) || !BothFallThrough(MBB2))
625       return true;
626   }
627 
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())
638     ++EffectiveTailLen;
639 
640   // Check if the common tail is long enough to be worthwhile.
641   if (EffectiveTailLen >= MinCommonTailLength)
642     return true;
643 
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();
649   bool OptForSize =
650       MF->getFunction().hasOptSize() ||
651       (llvm::shouldOptimizeForSize(MBB1, PSI, &MBBFreqInfo) &&
652        llvm::shouldOptimizeForSize(MBB2, PSI, &MBBFreqInfo));
653   return EffectiveTailLen >= 2 && OptForSize &&
654          (FullBlockTail1 || FullBlockTail2);
655 }
656 
657 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
658                                         unsigned MinCommonTailLength,
659                                         MachineBasicBlock *SuccBB,
660                                         MachineBasicBlock *PredBB) {
661   unsigned maxCommonTailLength = 0U;
662   SameTails.clear();
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(),
671                             MinCommonTailLength,
672                             CommonTailLen, TrialBBI1, TrialBBI2,
673                             SuccBB, PredBB,
674                             EHScopeMembership,
675                             AfterBlockPlacement, MBBFreqInfo, PSI)) {
676         if (CommonTailLen > maxCommonTailLength) {
677           SameTails.clear();
678           maxCommonTailLength = CommonTailLen;
679           HighestMPIter = CurMPIter;
680           SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
681         }
682         if (HighestMPIter == CurMPIter &&
683             CommonTailLen == maxCommonTailLength)
684           SameTails.push_back(SameTailElt(I, TrialBBI2));
685       }
686       if (I == B)
687         break;
688     }
689   }
690   return maxCommonTailLength;
691 }
692 
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);
704     if (CurMPIter == B)
705       break;
706   }
707   if (CurMPIter->getHash() != CurHash)
708     CurMPIter++;
709   MergePotentials.erase(CurMPIter, MergePotentials.end());
710 }
711 
712 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
713                                              MachineBasicBlock *SuccBB,
714                                              unsigned maxCommonTailLength,
715                                              unsigned &commonTailIndex) {
716   commonTailIndex = 0;
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) {
721       commonTailIndex = i;
722       break;
723     }
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) {
729       TimeEstimate = t;
730       commonTailIndex = i;
731     }
732   }
733 
734   MachineBasicBlock::iterator BBI =
735     SameTails[commonTailIndex].getTailStartPos();
736   MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
737 
738   LLVM_DEBUG(dbgs() << "\nSplitting " << printMBBReference(*MBB) << ", size "
739                     << maxCommonTailLength);
740 
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);
747   if (!newMBB) {
748     LLVM_DEBUG(dbgs() << "... failed!");
749     return false;
750   }
751 
752   SameTails[commonTailIndex].setBlock(newMBB);
753   SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
754 
755   // If we split PredBB, newMBB is the new predecessor.
756   if (PredBB == MBB)
757     PredBB = newMBB;
758 
759   return true;
760 }
761 
762 static void
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)
771     ++CommonTailLen;
772 
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();
777 
778   while (CommonTailLen--) {
779     assert(MBBI != MBBIE && "Reached BB end within common tail length!");
780     (void)MBBIE;
781 
782     if (!countsAsInstruction(*MBBI)) {
783       ++MBBI;
784       continue;
785     }
786 
787     while ((MBBICommon != MBBIECommon) && !countsAsInstruction(*MBBICommon))
788       ++MBBICommon;
789 
790     assert(MBBICommon != MBBIECommon &&
791            "Reached BB end within common tail length!");
792     assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!");
793 
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);
804       }
805     }
806 
807     ++MBBI;
808     ++MBBICommon;
809   }
810 }
811 
812 void BranchFolder::mergeCommonTails(unsigned commonTailIndex) {
813   MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
814 
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);
820     } else {
821       assert(SameTails[i].getTailStartPos() == MBB->begin() &&
822           "MBB is not a common tail only block");
823     }
824   }
825 
826   for (auto &MI : *MBB) {
827     if (!countsAsInstruction(MI))
828       continue;
829     DebugLoc DL = MI.getDebugLoc();
830     for (unsigned int i = 0 ; i < NextCommonInsts.size() ; i++) {
831       if (i == commonTailIndex)
832         continue;
833 
834       auto &Pos = NextCommonInsts[i];
835       assert(Pos != SameTails[i].getBlock()->end() &&
836           "Reached BB end within common tail");
837       while (!countsAsInstruction(*Pos)) {
838         ++Pos;
839         assert(Pos != SameTails[i].getBlock()->end() &&
840             "Reached BB end within common tail");
841       }
842       assert(MI.isIdenticalTo(*Pos) && "Expected matching MIIs!");
843       DL = DILocation::getMergedLocation(DL, Pos->getDebugLoc());
844       NextCommonInsts[i] = ++Pos;
845     }
846     MI.setDebugLoc(DL);
847   }
848 
849   if (UpdateLiveIns) {
850     LivePhysRegs NewLiveIns(*TRI);
851     computeLiveIns(NewLiveIns, *MBB);
852     LiveRegs.init(*TRI);
853 
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()) {
857       LiveRegs.clear();
858       LiveRegs.addLiveOuts(*Pred);
859       MachineBasicBlock::iterator InsertBefore = Pred->getFirstTerminator();
860       for (Register Reg : NewLiveIns) {
861         if (!LiveRegs.available(*MRI, Reg))
862           continue;
863         DebugLoc DL;
864         BuildMI(*Pred, InsertBefore, DL, TII->get(TargetOpcode::IMPLICIT_DEF),
865                 Reg);
866       }
867     }
868 
869     MBB->clearLiveIns();
870     addLiveIns(*MBB, NewLiveIns);
871   }
872 }
873 
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;
887 
888   LLVM_DEBUG(
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';
895         if (PredBB)
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';);
901 
902   // Sort by hash value so that blocks with identical end sequences sort
903   // together.
904   array_pod_sort(MergePotentials.begin(), MergePotentials.end());
905 
906   // Walk through equivalence sets looking for actual exact matches.
907   while (MergePotentials.size() > 1) {
908     unsigned CurHash = MergePotentials.back().getHash();
909 
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,
913                                                     MinCommonTailLength,
914                                                     SuccBB, PredBB);
915 
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);
920       continue;
921     }
922 
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
931     // into the other.
932     if (SameTails.size() == 2 &&
933         SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
934         SameTails[1].tailIsWholeBlock() && !SameTails[1].getBlock()->isEHPad())
935       commonTailIndex = 1;
936     else if (SameTails.size() == 2 &&
937              SameTails[1].getBlock()->isLayoutSuccessor(
938                  SameTails[0].getBlock()) &&
939              SameTails[0].tailIsWholeBlock() &&
940              !SameTails[0].getBlock()->isEHPad())
941       commonTailIndex = 0;
942     else {
943       // Otherwise just pick one, favoring the fall-through predecessor if
944       // there is one.
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())
949           continue;
950         if (MBB == PredBB) {
951           commonTailIndex = i;
952           break;
953         }
954         if (SameTails[i].tailIsWholeBlock())
955           commonTailIndex = i;
956       }
957     }
958 
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);
967         continue;
968       }
969     }
970 
971     MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
972 
973     // Recompute common tail MBB's edge weights and block frequency.
974     setCommonTailEdgeWeights(*MBB);
975 
976     // Merge debug locations, MMOs and undef flags across identical instructions
977     // for common tail.
978     mergeCommonTails(commonTailIndex);
979 
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)
983                       << " for ");
984     for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
985       if (commonTailIndex == i)
986         continue;
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());
993     }
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.
997     MadeChange = true;
998   }
999   return MadeChange;
1000 }
1001 
1002 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
1003   bool MadeChange = false;
1004   if (!EnableTailMerge)
1005     return MadeChange;
1006 
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)
1012       break;
1013     if (!TriedMerging.count(&MBB) && MBB.succ_empty())
1014       MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB));
1015   }
1016 
1017   // If this is a large problem, avoid visiting the same basic blocks
1018   // multiple times.
1019   if (MergePotentials.size() == TailMergeThreshold)
1020     for (const MergePotentialsElt &Elt : MergePotentials)
1021       TriedMerging.insert(Elt.getBlock());
1022 
1023   // See if we can do any tail merging on those.
1024   if (MergePotentials.size() >= 2)
1025     MadeChange |= TryTailMergeBlocks(nullptr, nullptr, MinCommonTailLength);
1026 
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
1030   // to IBB, and
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.
1034   //   Bcc IBB
1035   //   fallthrough to QBB
1036   // here becomes
1037   //   Bncc 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.)
1045 
1046   for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1047        I != E; ++I) {
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();
1053     MachineLoop *ML;
1054 
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())
1068         continue;
1069     }
1070 
1071     for (MachineBasicBlock *PBB : I->predecessors()) {
1072       if (MergePotentials.size() == TailMergeThreshold)
1073         break;
1074 
1075       if (TriedMerging.count(PBB))
1076         continue;
1077 
1078       // Skip blocks that loop to themselves, can't tail merge these.
1079       if (PBB == IBB)
1080         continue;
1081 
1082       // Visit each predecessor only once.
1083       if (!UniquePreds.insert(PBB).second)
1084         continue;
1085 
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())
1089         continue;
1090 
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
1093       // header.
1094       if (AfterBlockPlacement && MLI)
1095         if (ML != MLI->getLoopFor(PBB))
1096           continue;
1097 
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
1102         // branch.
1103         SmallVector<MachineOperand, 4> NewCond(Cond);
1104         if (!Cond.empty() && TBB == IBB) {
1105           if (TII->reverseBranchCondition(NewCond))
1106             continue;
1107           // This is the QBB case described above
1108           if (!FBB) {
1109             auto Next = ++PBB->getIterator();
1110             if (Next != MF.end())
1111               FBB = &*Next;
1112           }
1113         }
1114 
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);
1119           if (!Cond.empty())
1120             // reinsert conditional branch only, for now
1121             TII->insertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1122                               NewCond, dl);
1123         }
1124 
1125         MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB));
1126       }
1127     }
1128 
1129     // If this is a large problem, avoid visiting the same basic blocks multiple
1130     // times.
1131     if (MergePotentials.size() == TailMergeThreshold)
1132       for (MergePotentialsElt &Elt : MergePotentials)
1133         TriedMerging.insert(Elt.getBlock());
1134 
1135     if (MergePotentials.size() >= 2)
1136       MadeChange |= TryTailMergeBlocks(IBB, PredBB, MinCommonTailLength);
1137 
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);
1144   }
1145 
1146   return MadeChange;
1147 }
1148 
1149 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1150   SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1151   BlockFrequency AccumulatedMBBFreq;
1152 
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;
1160 
1161     // It is not necessary to recompute edge weights if TailBB has less than two
1162     // successors.
1163     if (TailMBB.succ_size() <= 1)
1164       continue;
1165 
1166     auto EdgeFreq = EdgeFreqLs.begin();
1167 
1168     for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1169          SuccI != SuccE; ++SuccI, ++EdgeFreq)
1170       *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1171   }
1172 
1173   MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1174 
1175   if (TailMBB.succ_size() <= 1)
1176     return;
1177 
1178   auto SumEdgeFreq =
1179       std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0))
1180           .getFrequency();
1181   auto EdgeFreq = EdgeFreqLs.begin();
1182 
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);
1189     }
1190   }
1191 }
1192 
1193 //===----------------------------------------------------------------------===//
1194 //  Branch Optimization
1195 //===----------------------------------------------------------------------===//
1196 
1197 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1198   bool MadeChange = false;
1199 
1200   // Make sure blocks are numbered in order
1201   MF.RenumberBlocks();
1202   // Renumbering blocks alters EH scope membership, recalculate it.
1203   EHScopeMembership = getEHScopeMembership(MF);
1204 
1205   for (MachineBasicBlock &MBB :
1206        llvm::make_early_inc_range(llvm::drop_begin(MF))) {
1207     MadeChange |= OptimizeBlock(&MBB);
1208 
1209     // If it is dead, remove it.
1210     if (MBB.pred_empty()) {
1211       RemoveDeadBlock(&MBB);
1212       MadeChange = true;
1213       ++NumDeadBlocks;
1214     }
1215   }
1216 
1217   return MadeChange;
1218 }
1219 
1220 // Blocks should be considered empty if they contain only debug info;
1221 // else the debug info would affect codegen.
1222 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1223   return MBB->getFirstNonDebugInstr(true) == MBB->end();
1224 }
1225 
1226 // Blocks with only debug info and branches should be considered the same
1227 // as blocks with only branches.
1228 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1229   MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1230   assert(I != MBB->end() && "empty block!");
1231   return I->isBranch();
1232 }
1233 
1234 /// IsBetterFallthrough - Return true if it would be clearly better to
1235 /// fall-through to MBB1 than to fall through into MBB2.  This has to return
1236 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1237 /// result in infinite loops.
1238 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1239                                 MachineBasicBlock *MBB2) {
1240   assert(MBB1 && MBB2 && "Unknown MachineBasicBlock");
1241 
1242   // Right now, we use a simple heuristic.  If MBB2 ends with a call, and
1243   // MBB1 doesn't, we prefer to fall through into MBB1.  This allows us to
1244   // optimize branches that branch to either a return block or an assert block
1245   // into a fallthrough to the return.
1246   MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1247   MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1248   if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1249     return false;
1250 
1251   // If there is a clear successor ordering we make sure that one block
1252   // will fall through to the next
1253   if (MBB1->isSuccessor(MBB2)) return true;
1254   if (MBB2->isSuccessor(MBB1)) return false;
1255 
1256   return MBB2I->isCall() && !MBB1I->isCall();
1257 }
1258 
1259 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1260 /// instructions on the block.
1261 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1262   MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1263   if (I != MBB.end() && I->isBranch())
1264     return I->getDebugLoc();
1265   return DebugLoc();
1266 }
1267 
1268 static void copyDebugInfoToPredecessor(const TargetInstrInfo *TII,
1269                                        MachineBasicBlock &MBB,
1270                                        MachineBasicBlock &PredMBB) {
1271   auto InsertBefore = PredMBB.getFirstTerminator();
1272   for (MachineInstr &MI : MBB.instrs())
1273     if (MI.isDebugInstr()) {
1274       TII->duplicate(PredMBB, InsertBefore, MI);
1275       LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to pred: "
1276                         << MI);
1277     }
1278 }
1279 
1280 static void copyDebugInfoToSuccessor(const TargetInstrInfo *TII,
1281                                      MachineBasicBlock &MBB,
1282                                      MachineBasicBlock &SuccMBB) {
1283   auto InsertBefore = SuccMBB.SkipPHIsAndLabels(SuccMBB.begin());
1284   for (MachineInstr &MI : MBB.instrs())
1285     if (MI.isDebugInstr()) {
1286       TII->duplicate(SuccMBB, InsertBefore, MI);
1287       LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to succ: "
1288                         << MI);
1289     }
1290 }
1291 
1292 // Try to salvage DBG_VALUE instructions from an otherwise empty block. If such
1293 // a basic block is removed we would lose the debug information unless we have
1294 // copied the information to a predecessor/successor.
1295 //
1296 // TODO: This function only handles some simple cases. An alternative would be
1297 // to run a heavier analysis, such as the LiveDebugValues pass, before we do
1298 // branch folding.
1299 static void salvageDebugInfoFromEmptyBlock(const TargetInstrInfo *TII,
1300                                            MachineBasicBlock &MBB) {
1301   assert(IsEmptyBlock(&MBB) && "Expected an empty block (except debug info).");
1302   // If this MBB is the only predecessor of a successor it is legal to copy
1303   // DBG_VALUE instructions to the beginning of the successor.
1304   for (MachineBasicBlock *SuccBB : MBB.successors())
1305     if (SuccBB->pred_size() == 1)
1306       copyDebugInfoToSuccessor(TII, MBB, *SuccBB);
1307   // If this MBB is the only successor of a predecessor it is legal to copy the
1308   // DBG_VALUE instructions to the end of the predecessor (just before the
1309   // terminators, assuming that the terminator isn't affecting the DBG_VALUE).
1310   for (MachineBasicBlock *PredBB : MBB.predecessors())
1311     if (PredBB->succ_size() == 1)
1312       copyDebugInfoToPredecessor(TII, MBB, *PredBB);
1313 }
1314 
1315 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1316   bool MadeChange = false;
1317   MachineFunction &MF = *MBB->getParent();
1318 ReoptimizeBlock:
1319 
1320   MachineFunction::iterator FallThrough = MBB->getIterator();
1321   ++FallThrough;
1322 
1323   // Make sure MBB and FallThrough belong to the same EH scope.
1324   bool SameEHScope = true;
1325   if (!EHScopeMembership.empty() && FallThrough != MF.end()) {
1326     auto MBBEHScope = EHScopeMembership.find(MBB);
1327     assert(MBBEHScope != EHScopeMembership.end());
1328     auto FallThroughEHScope = EHScopeMembership.find(&*FallThrough);
1329     assert(FallThroughEHScope != EHScopeMembership.end());
1330     SameEHScope = MBBEHScope->second == FallThroughEHScope->second;
1331   }
1332 
1333   // Analyze the branch in the current block. As a side-effect, this may cause
1334   // the block to become empty.
1335   MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1336   SmallVector<MachineOperand, 4> CurCond;
1337   bool CurUnAnalyzable =
1338       TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1339 
1340   // If this block is empty, make everyone use its fall-through, not the block
1341   // explicitly.  Landing pads should not do this since the landing-pad table
1342   // points to this block.  Blocks with their addresses taken shouldn't be
1343   // optimized away.
1344   if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() &&
1345       SameEHScope) {
1346     salvageDebugInfoFromEmptyBlock(TII, *MBB);
1347     // Dead block?  Leave for cleanup later.
1348     if (MBB->pred_empty()) return MadeChange;
1349 
1350     if (FallThrough == MF.end()) {
1351       // TODO: Simplify preds to not branch here if possible!
1352     } else if (FallThrough->isEHPad()) {
1353       // Don't rewrite to a landing pad fallthough.  That could lead to the case
1354       // where a BB jumps to more than one landing pad.
1355       // TODO: Is it ever worth rewriting predecessors which don't already
1356       // jump to a landing pad, and so can safely jump to the fallthrough?
1357     } else if (MBB->isSuccessor(&*FallThrough)) {
1358       // Rewrite all predecessors of the old block to go to the fallthrough
1359       // instead.
1360       while (!MBB->pred_empty()) {
1361         MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1362         Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough);
1363       }
1364       // If MBB was the target of a jump table, update jump tables to go to the
1365       // fallthrough instead.
1366       if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1367         MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough);
1368       MadeChange = true;
1369     }
1370     return MadeChange;
1371   }
1372 
1373   // Check to see if we can simplify the terminator of the block before this
1374   // one.
1375   MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1376 
1377   MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1378   SmallVector<MachineOperand, 4> PriorCond;
1379   bool PriorUnAnalyzable =
1380       TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1381   if (!PriorUnAnalyzable) {
1382     // If the previous branch is conditional and both conditions go to the same
1383     // destination, remove the branch, replacing it with an unconditional one or
1384     // a fall-through.
1385     if (PriorTBB && PriorTBB == PriorFBB) {
1386       DebugLoc dl = getBranchDebugLoc(PrevBB);
1387       TII->removeBranch(PrevBB);
1388       PriorCond.clear();
1389       if (PriorTBB != MBB)
1390         TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1391       MadeChange = true;
1392       ++NumBranchOpts;
1393       goto ReoptimizeBlock;
1394     }
1395 
1396     // If the previous block unconditionally falls through to this block and
1397     // this block has no other predecessors, move the contents of this block
1398     // into the prior block. This doesn't usually happen when SimplifyCFG
1399     // has been used, but it can happen if tail merging splits a fall-through
1400     // predecessor of a block.
1401     // This has to check PrevBB->succ_size() because EH edges are ignored by
1402     // analyzeBranch.
1403     if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1404         PrevBB.succ_size() == 1 &&
1405         !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1406       LLVM_DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1407                         << "From MBB: " << *MBB);
1408       // Remove redundant DBG_VALUEs first.
1409       if (!PrevBB.empty()) {
1410         MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1411         --PrevBBIter;
1412         MachineBasicBlock::iterator MBBIter = MBB->begin();
1413         // Check if DBG_VALUE at the end of PrevBB is identical to the
1414         // DBG_VALUE at the beginning of MBB.
1415         while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1416                && PrevBBIter->isDebugInstr() && MBBIter->isDebugInstr()) {
1417           if (!MBBIter->isIdenticalTo(*PrevBBIter))
1418             break;
1419           MachineInstr &DuplicateDbg = *MBBIter;
1420           ++MBBIter; -- PrevBBIter;
1421           DuplicateDbg.eraseFromParent();
1422         }
1423       }
1424       PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1425       PrevBB.removeSuccessor(PrevBB.succ_begin());
1426       assert(PrevBB.succ_empty());
1427       PrevBB.transferSuccessors(MBB);
1428       MadeChange = true;
1429       return MadeChange;
1430     }
1431 
1432     // If the previous branch *only* branches to *this* block (conditional or
1433     // not) remove the branch.
1434     if (PriorTBB == MBB && !PriorFBB) {
1435       TII->removeBranch(PrevBB);
1436       MadeChange = true;
1437       ++NumBranchOpts;
1438       goto ReoptimizeBlock;
1439     }
1440 
1441     // If the prior block branches somewhere else on the condition and here if
1442     // the condition is false, remove the uncond second branch.
1443     if (PriorFBB == MBB) {
1444       DebugLoc dl = getBranchDebugLoc(PrevBB);
1445       TII->removeBranch(PrevBB);
1446       TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1447       MadeChange = true;
1448       ++NumBranchOpts;
1449       goto ReoptimizeBlock;
1450     }
1451 
1452     // If the prior block branches here on true and somewhere else on false, and
1453     // if the branch condition is reversible, reverse the branch to create a
1454     // fall-through.
1455     if (PriorTBB == MBB) {
1456       SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1457       if (!TII->reverseBranchCondition(NewPriorCond)) {
1458         DebugLoc dl = getBranchDebugLoc(PrevBB);
1459         TII->removeBranch(PrevBB);
1460         TII->insertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1461         MadeChange = true;
1462         ++NumBranchOpts;
1463         goto ReoptimizeBlock;
1464       }
1465     }
1466 
1467     // If this block has no successors (e.g. it is a return block or ends with
1468     // a call to a no-return function like abort or __cxa_throw) and if the pred
1469     // falls through into this block, and if it would otherwise fall through
1470     // into the block after this, move this block to the end of the function.
1471     //
1472     // We consider it more likely that execution will stay in the function (e.g.
1473     // due to loops) than it is to exit it.  This asserts in loops etc, moving
1474     // the assert condition out of the loop body.
1475     if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1476         MachineFunction::iterator(PriorTBB) == FallThrough &&
1477         !MBB->canFallThrough()) {
1478       bool DoTransform = true;
1479 
1480       // We have to be careful that the succs of PredBB aren't both no-successor
1481       // blocks.  If neither have successors and if PredBB is the second from
1482       // last block in the function, we'd just keep swapping the two blocks for
1483       // last.  Only do the swap if one is clearly better to fall through than
1484       // the other.
1485       if (FallThrough == --MF.end() &&
1486           !IsBetterFallthrough(PriorTBB, MBB))
1487         DoTransform = false;
1488 
1489       if (DoTransform) {
1490         // Reverse the branch so we will fall through on the previous true cond.
1491         SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1492         if (!TII->reverseBranchCondition(NewPriorCond)) {
1493           LLVM_DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1494                             << "To make fallthrough to: " << *PriorTBB << "\n");
1495 
1496           DebugLoc dl = getBranchDebugLoc(PrevBB);
1497           TII->removeBranch(PrevBB);
1498           TII->insertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1499 
1500           // Move this block to the end of the function.
1501           MBB->moveAfter(&MF.back());
1502           MadeChange = true;
1503           ++NumBranchOpts;
1504           return MadeChange;
1505         }
1506       }
1507     }
1508   }
1509 
1510   bool OptForSize =
1511       MF.getFunction().hasOptSize() ||
1512       llvm::shouldOptimizeForSize(MBB, PSI, &MBBFreqInfo);
1513   if (!IsEmptyBlock(MBB) && MBB->pred_size() == 1 && OptForSize) {
1514     // Changing "Jcc foo; foo: jmp bar;" into "Jcc bar;" might change the branch
1515     // direction, thereby defeating careful block placement and regressing
1516     // performance. Therefore, only consider this for optsize functions.
1517     MachineInstr &TailCall = *MBB->getFirstNonDebugInstr();
1518     if (TII->isUnconditionalTailCall(TailCall)) {
1519       MachineBasicBlock *Pred = *MBB->pred_begin();
1520       MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1521       SmallVector<MachineOperand, 4> PredCond;
1522       bool PredAnalyzable =
1523           !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true);
1524 
1525       if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB &&
1526           PredTBB != PredFBB) {
1527         // The predecessor has a conditional branch to this block which consists
1528         // of only a tail call. Try to fold the tail call into the conditional
1529         // branch.
1530         if (TII->canMakeTailCallConditional(PredCond, TailCall)) {
1531           // TODO: It would be nice if analyzeBranch() could provide a pointer
1532           // to the branch instruction so replaceBranchWithTailCall() doesn't
1533           // have to search for it.
1534           TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall);
1535           ++NumTailCalls;
1536           Pred->removeSuccessor(MBB);
1537           MadeChange = true;
1538           return MadeChange;
1539         }
1540       }
1541       // If the predecessor is falling through to this block, we could reverse
1542       // the branch condition and fold the tail call into that. However, after
1543       // that we might have to re-arrange the CFG to fall through to the other
1544       // block and there is a high risk of regressing code size rather than
1545       // improving it.
1546     }
1547   }
1548 
1549   if (!CurUnAnalyzable) {
1550     // If this is a two-way branch, and the FBB branches to this block, reverse
1551     // the condition so the single-basic-block loop is faster.  Instead of:
1552     //    Loop: xxx; jcc Out; jmp Loop
1553     // we want:
1554     //    Loop: xxx; jncc Loop; jmp Out
1555     if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1556       SmallVector<MachineOperand, 4> NewCond(CurCond);
1557       if (!TII->reverseBranchCondition(NewCond)) {
1558         DebugLoc dl = getBranchDebugLoc(*MBB);
1559         TII->removeBranch(*MBB);
1560         TII->insertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1561         MadeChange = true;
1562         ++NumBranchOpts;
1563         goto ReoptimizeBlock;
1564       }
1565     }
1566 
1567     // If this branch is the only thing in its block, see if we can forward
1568     // other blocks across it.
1569     if (CurTBB && CurCond.empty() && !CurFBB &&
1570         IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1571         !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1572       DebugLoc dl = getBranchDebugLoc(*MBB);
1573       // This block may contain just an unconditional branch.  Because there can
1574       // be 'non-branch terminators' in the block, try removing the branch and
1575       // then seeing if the block is empty.
1576       TII->removeBranch(*MBB);
1577       // If the only things remaining in the block are debug info, remove these
1578       // as well, so this will behave the same as an empty block in non-debug
1579       // mode.
1580       if (IsEmptyBlock(MBB)) {
1581         // Make the block empty, losing the debug info (we could probably
1582         // improve this in some cases.)
1583         MBB->erase(MBB->begin(), MBB->end());
1584       }
1585       // If this block is just an unconditional branch to CurTBB, we can
1586       // usually completely eliminate the block.  The only case we cannot
1587       // completely eliminate the block is when the block before this one
1588       // falls through into MBB and we can't understand the prior block's branch
1589       // condition.
1590       if (MBB->empty()) {
1591         bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1592         if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1593             !PrevBB.isSuccessor(MBB)) {
1594           // If the prior block falls through into us, turn it into an
1595           // explicit branch to us to make updates simpler.
1596           if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1597               PriorTBB != MBB && PriorFBB != MBB) {
1598             if (!PriorTBB) {
1599               assert(PriorCond.empty() && !PriorFBB &&
1600                      "Bad branch analysis");
1601               PriorTBB = MBB;
1602             } else {
1603               assert(!PriorFBB && "Machine CFG out of date!");
1604               PriorFBB = MBB;
1605             }
1606             DebugLoc pdl = getBranchDebugLoc(PrevBB);
1607             TII->removeBranch(PrevBB);
1608             TII->insertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1609           }
1610 
1611           // Iterate through all the predecessors, revectoring each in-turn.
1612           size_t PI = 0;
1613           bool DidChange = false;
1614           bool HasBranchToSelf = false;
1615           while(PI != MBB->pred_size()) {
1616             MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1617             if (PMBB == MBB) {
1618               // If this block has an uncond branch to itself, leave it.
1619               ++PI;
1620               HasBranchToSelf = true;
1621             } else {
1622               DidChange = true;
1623               PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1624               // If this change resulted in PMBB ending in a conditional
1625               // branch where both conditions go to the same destination,
1626               // change this to an unconditional branch.
1627               MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1628               SmallVector<MachineOperand, 4> NewCurCond;
1629               bool NewCurUnAnalyzable = TII->analyzeBranch(
1630                   *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true);
1631               if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1632                 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1633                 TII->removeBranch(*PMBB);
1634                 NewCurCond.clear();
1635                 TII->insertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1636                 MadeChange = true;
1637                 ++NumBranchOpts;
1638               }
1639             }
1640           }
1641 
1642           // Change any jumptables to go to the new MBB.
1643           if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1644             MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1645           if (DidChange) {
1646             ++NumBranchOpts;
1647             MadeChange = true;
1648             if (!HasBranchToSelf) return MadeChange;
1649           }
1650         }
1651       }
1652 
1653       // Add the branch back if the block is more than just an uncond branch.
1654       TII->insertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1655     }
1656   }
1657 
1658   // If the prior block doesn't fall through into this block, and if this
1659   // block doesn't fall through into some other block, see if we can find a
1660   // place to move this block where a fall-through will happen.
1661   if (!PrevBB.canFallThrough()) {
1662     // Now we know that there was no fall-through into this block, check to
1663     // see if it has a fall-through into its successor.
1664     bool CurFallsThru = MBB->canFallThrough();
1665 
1666     if (!MBB->isEHPad()) {
1667       // Check all the predecessors of this block.  If one of them has no fall
1668       // throughs, and analyzeBranch thinks it _could_ fallthrough to this
1669       // block, move this block right after it.
1670       for (MachineBasicBlock *PredBB : MBB->predecessors()) {
1671         // Analyze the branch at the end of the pred.
1672         MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1673         SmallVector<MachineOperand, 4> PredCond;
1674         if (PredBB != MBB && !PredBB->canFallThrough() &&
1675             !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) &&
1676             (PredTBB == MBB || PredFBB == MBB) &&
1677             (!CurFallsThru || !CurTBB || !CurFBB) &&
1678             (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1679           // If the current block doesn't fall through, just move it.
1680           // If the current block can fall through and does not end with a
1681           // conditional branch, we need to append an unconditional jump to
1682           // the (current) next block.  To avoid a possible compile-time
1683           // infinite loop, move blocks only backward in this case.
1684           // Also, if there are already 2 branches here, we cannot add a third;
1685           // this means we have the case
1686           // Bcc next
1687           // B elsewhere
1688           // next:
1689           if (CurFallsThru) {
1690             MachineBasicBlock *NextBB = &*std::next(MBB->getIterator());
1691             CurCond.clear();
1692             TII->insertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1693           }
1694           MBB->moveAfter(PredBB);
1695           MadeChange = true;
1696           goto ReoptimizeBlock;
1697         }
1698       }
1699     }
1700 
1701     if (!CurFallsThru) {
1702       // Check analyzable branch-successors to see if we can move this block
1703       // before one.
1704       if (!CurUnAnalyzable) {
1705         for (MachineBasicBlock *SuccBB : {CurFBB, CurTBB}) {
1706           if (!SuccBB)
1707             continue;
1708           // Analyze the branch at the end of the block before the succ.
1709           MachineFunction::iterator SuccPrev = --SuccBB->getIterator();
1710 
1711           // If this block doesn't already fall-through to that successor, and
1712           // if the succ doesn't already have a block that can fall through into
1713           // it, we can arrange for the fallthrough to happen.
1714           if (SuccBB != MBB && &*SuccPrev != MBB &&
1715               !SuccPrev->canFallThrough()) {
1716             MBB->moveBefore(SuccBB);
1717             MadeChange = true;
1718             goto ReoptimizeBlock;
1719           }
1720         }
1721       }
1722 
1723       // Okay, there is no really great place to put this block.  If, however,
1724       // the block before this one would be a fall-through if this block were
1725       // removed, move this block to the end of the function. There is no real
1726       // advantage in "falling through" to an EH block, so we don't want to
1727       // perform this transformation for that case.
1728       //
1729       // Also, Windows EH introduced the possibility of an arbitrary number of
1730       // successors to a given block.  The analyzeBranch call does not consider
1731       // exception handling and so we can get in a state where a block
1732       // containing a call is followed by multiple EH blocks that would be
1733       // rotated infinitely at the end of the function if the transformation
1734       // below were performed for EH "FallThrough" blocks.  Therefore, even if
1735       // that appears not to be happening anymore, we should assume that it is
1736       // possible and not remove the "!FallThrough()->isEHPad" condition below.
1737       MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1738       SmallVector<MachineOperand, 4> PrevCond;
1739       if (FallThrough != MF.end() &&
1740           !FallThrough->isEHPad() &&
1741           !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1742           PrevBB.isSuccessor(&*FallThrough)) {
1743         MBB->moveAfter(&MF.back());
1744         MadeChange = true;
1745         return MadeChange;
1746       }
1747     }
1748   }
1749 
1750   return MadeChange;
1751 }
1752 
1753 //===----------------------------------------------------------------------===//
1754 //  Hoist Common Code
1755 //===----------------------------------------------------------------------===//
1756 
1757 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1758   bool MadeChange = false;
1759   for (MachineBasicBlock &MBB : llvm::make_early_inc_range(MF))
1760     MadeChange |= HoistCommonCodeInSuccs(&MBB);
1761 
1762   return MadeChange;
1763 }
1764 
1765 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1766 /// its 'true' successor.
1767 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1768                                          MachineBasicBlock *TrueBB) {
1769   for (MachineBasicBlock *SuccBB : BB->successors())
1770     if (SuccBB != TrueBB)
1771       return SuccBB;
1772   return nullptr;
1773 }
1774 
1775 template <class Container>
1776 static void addRegAndItsAliases(Register Reg, const TargetRegisterInfo *TRI,
1777                                 Container &Set) {
1778   if (Reg.isPhysical()) {
1779     for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1780       Set.insert(*AI);
1781   } else {
1782     Set.insert(Reg);
1783   }
1784 }
1785 
1786 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1787 /// in successors to. The location is usually just before the terminator,
1788 /// however if the terminator is a conditional branch and its previous
1789 /// instruction is the flag setting instruction, the previous instruction is
1790 /// the preferred location. This function also gathers uses and defs of the
1791 /// instructions from the insertion point to the end of the block. The data is
1792 /// used by HoistCommonCodeInSuccs to ensure safety.
1793 static
1794 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1795                                                   const TargetInstrInfo *TII,
1796                                                   const TargetRegisterInfo *TRI,
1797                                                   SmallSet<Register, 4> &Uses,
1798                                                   SmallSet<Register, 4> &Defs) {
1799   MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1800   if (!TII->isUnpredicatedTerminator(*Loc))
1801     return MBB->end();
1802 
1803   for (const MachineOperand &MO : Loc->operands()) {
1804     if (!MO.isReg())
1805       continue;
1806     Register Reg = MO.getReg();
1807     if (!Reg)
1808       continue;
1809     if (MO.isUse()) {
1810       addRegAndItsAliases(Reg, TRI, Uses);
1811     } else {
1812       if (!MO.isDead())
1813         // Don't try to hoist code in the rare case the terminator defines a
1814         // register that is later used.
1815         return MBB->end();
1816 
1817       // If the terminator defines a register, make sure we don't hoist
1818       // the instruction whose def might be clobbered by the terminator.
1819       addRegAndItsAliases(Reg, TRI, Defs);
1820     }
1821   }
1822 
1823   if (Uses.empty())
1824     return Loc;
1825   // If the terminator is the only instruction in the block and Uses is not
1826   // empty (or we would have returned above), we can still safely hoist
1827   // instructions just before the terminator as long as the Defs/Uses are not
1828   // violated (which is checked in HoistCommonCodeInSuccs).
1829   if (Loc == MBB->begin())
1830     return Loc;
1831 
1832   // The terminator is probably a conditional branch, try not to separate the
1833   // branch from condition setting instruction.
1834   MachineBasicBlock::iterator PI = prev_nodbg(Loc, MBB->begin());
1835 
1836   bool IsDef = false;
1837   for (const MachineOperand &MO : PI->operands()) {
1838     // If PI has a regmask operand, it is probably a call. Separate away.
1839     if (MO.isRegMask())
1840       return Loc;
1841     if (!MO.isReg() || MO.isUse())
1842       continue;
1843     Register Reg = MO.getReg();
1844     if (!Reg)
1845       continue;
1846     if (Uses.count(Reg)) {
1847       IsDef = true;
1848       break;
1849     }
1850   }
1851   if (!IsDef)
1852     // The condition setting instruction is not just before the conditional
1853     // branch.
1854     return Loc;
1855 
1856   // Be conservative, don't insert instruction above something that may have
1857   // side-effects. And since it's potentially bad to separate flag setting
1858   // instruction from the conditional branch, just abort the optimization
1859   // completely.
1860   // Also avoid moving code above predicated instruction since it's hard to
1861   // reason about register liveness with predicated instruction.
1862   bool DontMoveAcrossStore = true;
1863   if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI))
1864     return MBB->end();
1865 
1866   // Find out what registers are live. Note this routine is ignoring other live
1867   // registers which are only used by instructions in successor blocks.
1868   for (const MachineOperand &MO : PI->operands()) {
1869     if (!MO.isReg())
1870       continue;
1871     Register Reg = MO.getReg();
1872     if (!Reg)
1873       continue;
1874     if (MO.isUse()) {
1875       addRegAndItsAliases(Reg, TRI, Uses);
1876     } else {
1877       if (Uses.erase(Reg)) {
1878         if (Register::isPhysicalRegister(Reg)) {
1879           for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1880             Uses.erase(*SubRegs); // Use sub-registers to be conservative
1881         }
1882       }
1883       addRegAndItsAliases(Reg, TRI, Defs);
1884     }
1885   }
1886 
1887   return PI;
1888 }
1889 
1890 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1891   MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1892   SmallVector<MachineOperand, 4> Cond;
1893   if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1894     return false;
1895 
1896   if (!FBB) FBB = findFalseBlock(MBB, TBB);
1897   if (!FBB)
1898     // Malformed bcc? True and false blocks are the same?
1899     return false;
1900 
1901   // Restrict the optimization to cases where MBB is the only predecessor,
1902   // it is an obvious win.
1903   if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1904     return false;
1905 
1906   // Find a suitable position to hoist the common instructions to. Also figure
1907   // out which registers are used or defined by instructions from the insertion
1908   // point to the end of the block.
1909   SmallSet<Register, 4> Uses, Defs;
1910   MachineBasicBlock::iterator Loc =
1911     findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1912   if (Loc == MBB->end())
1913     return false;
1914 
1915   bool HasDups = false;
1916   SmallSet<Register, 4> ActiveDefsSet, AllDefsSet;
1917   MachineBasicBlock::iterator TIB = TBB->begin();
1918   MachineBasicBlock::iterator FIB = FBB->begin();
1919   MachineBasicBlock::iterator TIE = TBB->end();
1920   MachineBasicBlock::iterator FIE = FBB->end();
1921   while (TIB != TIE && FIB != FIE) {
1922     // Skip dbg_value instructions. These do not count.
1923     TIB = skipDebugInstructionsForward(TIB, TIE, false);
1924     FIB = skipDebugInstructionsForward(FIB, FIE, false);
1925     if (TIB == TIE || FIB == FIE)
1926       break;
1927 
1928     if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead))
1929       break;
1930 
1931     if (TII->isPredicated(*TIB))
1932       // Hard to reason about register liveness with predicated instruction.
1933       break;
1934 
1935     bool IsSafe = true;
1936     for (MachineOperand &MO : TIB->operands()) {
1937       // Don't attempt to hoist instructions with register masks.
1938       if (MO.isRegMask()) {
1939         IsSafe = false;
1940         break;
1941       }
1942       if (!MO.isReg())
1943         continue;
1944       Register Reg = MO.getReg();
1945       if (!Reg)
1946         continue;
1947       if (MO.isDef()) {
1948         if (Uses.count(Reg)) {
1949           // Avoid clobbering a register that's used by the instruction at
1950           // the point of insertion.
1951           IsSafe = false;
1952           break;
1953         }
1954 
1955         if (Defs.count(Reg) && !MO.isDead()) {
1956           // Don't hoist the instruction if the def would be clobber by the
1957           // instruction at the point insertion. FIXME: This is overly
1958           // conservative. It should be possible to hoist the instructions
1959           // in BB2 in the following example:
1960           // BB1:
1961           // r1, eflag = op1 r2, r3
1962           // brcc eflag
1963           //
1964           // BB2:
1965           // r1 = op2, ...
1966           //    = op3, killed r1
1967           IsSafe = false;
1968           break;
1969         }
1970       } else if (!ActiveDefsSet.count(Reg)) {
1971         if (Defs.count(Reg)) {
1972           // Use is defined by the instruction at the point of insertion.
1973           IsSafe = false;
1974           break;
1975         }
1976 
1977         if (MO.isKill() && Uses.count(Reg))
1978           // Kills a register that's read by the instruction at the point of
1979           // insertion. Remove the kill marker.
1980           MO.setIsKill(false);
1981       }
1982     }
1983     if (!IsSafe)
1984       break;
1985 
1986     bool DontMoveAcrossStore = true;
1987     if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
1988       break;
1989 
1990     // Remove kills from ActiveDefsSet, these registers had short live ranges.
1991     for (const MachineOperand &MO : TIB->operands()) {
1992       if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1993         continue;
1994       Register Reg = MO.getReg();
1995       if (!Reg)
1996         continue;
1997       if (!AllDefsSet.count(Reg)) {
1998         continue;
1999       }
2000       if (Reg.isPhysical()) {
2001         for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
2002           ActiveDefsSet.erase(*AI);
2003       } else {
2004         ActiveDefsSet.erase(Reg);
2005       }
2006     }
2007 
2008     // Track local defs so we can update liveins.
2009     for (const MachineOperand &MO : TIB->operands()) {
2010       if (!MO.isReg() || !MO.isDef() || MO.isDead())
2011         continue;
2012       Register Reg = MO.getReg();
2013       if (!Reg || Reg.isVirtual())
2014         continue;
2015       addRegAndItsAliases(Reg, TRI, ActiveDefsSet);
2016       addRegAndItsAliases(Reg, TRI, AllDefsSet);
2017     }
2018 
2019     HasDups = true;
2020     ++TIB;
2021     ++FIB;
2022   }
2023 
2024   if (!HasDups)
2025     return false;
2026 
2027   MBB->splice(Loc, TBB, TBB->begin(), TIB);
2028   FBB->erase(FBB->begin(), FIB);
2029 
2030   if (UpdateLiveIns) {
2031     recomputeLiveIns(*TBB);
2032     recomputeLiveIns(*FBB);
2033   }
2034 
2035   ++NumHoist;
2036   return true;
2037 }
2038