xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/BranchFolding.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
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 static cl::opt<unsigned>
84 TailMergeSize("tail-merge-size",
85               cl::desc("Min number of instructions to consider tail merging"),
86               cl::init(3), cl::Hidden);
87 
88 namespace {
89 
90   /// BranchFolderPass - Wrap branch folder in a machine function pass.
91   class BranchFolderPass : public MachineFunctionPass {
92   public:
93     static char ID;
94 
BranchFolderPass()95     explicit BranchFolderPass(): MachineFunctionPass(ID) {}
96 
97     bool runOnMachineFunction(MachineFunction &MF) override;
98 
getAnalysisUsage(AnalysisUsage & AU) const99     void getAnalysisUsage(AnalysisUsage &AU) const override {
100       AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
101       AU.addRequired<MachineBranchProbabilityInfoWrapperPass>();
102       AU.addRequired<ProfileSummaryInfoWrapperPass>();
103       AU.addRequired<TargetPassConfig>();
104       MachineFunctionPass::getAnalysisUsage(AU);
105     }
106 
getRequiredProperties() const107     MachineFunctionProperties getRequiredProperties() const override {
108       return MachineFunctionProperties().set(
109           MachineFunctionProperties::Property::NoPHIs);
110     }
111   };
112 
113 } // end anonymous namespace
114 
115 char BranchFolderPass::ID = 0;
116 
117 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
118 
119 INITIALIZE_PASS(BranchFolderPass, DEBUG_TYPE,
120                 "Control Flow Optimizer", false, false)
121 
runOnMachineFunction(MachineFunction & MF)122 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
123   if (skipFunction(MF.getFunction()))
124     return false;
125 
126   TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
127   // TailMerge can create jump into if branches that make CFG irreducible for
128   // HW that requires structurized CFG.
129   bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
130                          PassConfig->getEnableTailMerge();
131   MBFIWrapper MBBFreqInfo(
132       getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI());
133   BranchFolder Folder(
134       EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo,
135       getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI(),
136       &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI());
137   return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
138                                  MF.getSubtarget().getRegisterInfo());
139 }
140 
BranchFolder(bool DefaultEnableTailMerge,bool CommonHoist,MBFIWrapper & FreqInfo,const MachineBranchProbabilityInfo & ProbInfo,ProfileSummaryInfo * PSI,unsigned MinTailLength)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   switch (FlagEnableTailMerge) {
148   case cl::BOU_UNSET:
149     EnableTailMerge = DefaultEnableTailMerge;
150     break;
151   case cl::BOU_TRUE: EnableTailMerge = true; break;
152   case cl::BOU_FALSE: EnableTailMerge = false; break;
153   }
154 }
155 
RemoveDeadBlock(MachineBasicBlock * MBB)156 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
157   assert(MBB->pred_empty() && "MBB must be dead!");
158   LLVM_DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
159 
160   MachineFunction *MF = MBB->getParent();
161   // drop all successors.
162   while (!MBB->succ_empty())
163     MBB->removeSuccessor(MBB->succ_end()-1);
164 
165   // Avoid matching if this pointer gets reused.
166   TriedMerging.erase(MBB);
167 
168   // Update call site info.
169   for (const MachineInstr &MI : *MBB)
170     if (MI.shouldUpdateCallSiteInfo())
171       MF->eraseCallSiteInfo(&MI);
172 
173   // Remove the block.
174   MF->erase(MBB);
175   EHScopeMembership.erase(MBB);
176   if (MLI)
177     MLI->removeBlock(MBB);
178 }
179 
OptimizeFunction(MachineFunction & MF,const TargetInstrInfo * tii,const TargetRegisterInfo * tri,MachineLoopInfo * mli,bool AfterPlacement)180 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
181                                     const TargetInstrInfo *tii,
182                                     const TargetRegisterInfo *tri,
183                                     MachineLoopInfo *mli, bool AfterPlacement) {
184   if (!tii) return false;
185 
186   TriedMerging.clear();
187 
188   MachineRegisterInfo &MRI = MF.getRegInfo();
189   AfterBlockPlacement = AfterPlacement;
190   TII = tii;
191   TRI = tri;
192   MLI = mli;
193   this->MRI = &MRI;
194 
195   if (MinCommonTailLength == 0) {
196     MinCommonTailLength = TailMergeSize.getNumOccurrences() > 0
197                               ? TailMergeSize
198                               : TII->getTailMergeSize(MF);
199   }
200 
201   UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF);
202   if (!UpdateLiveIns)
203     MRI.invalidateLiveness();
204 
205   bool MadeChange = false;
206 
207   // Recalculate EH scope membership.
208   EHScopeMembership = getEHScopeMembership(MF);
209 
210   bool MadeChangeThisIteration = true;
211   while (MadeChangeThisIteration) {
212     MadeChangeThisIteration    = TailMergeBlocks(MF);
213     // No need to clean up if tail merging does not change anything after the
214     // block placement.
215     if (!AfterBlockPlacement || MadeChangeThisIteration)
216       MadeChangeThisIteration |= OptimizeBranches(MF);
217     if (EnableHoistCommonCode)
218       MadeChangeThisIteration |= HoistCommonCode(MF);
219     MadeChange |= MadeChangeThisIteration;
220   }
221 
222   // See if any jump tables have become dead as the code generator
223   // did its thing.
224   MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
225   if (!JTI)
226     return MadeChange;
227 
228   // Walk the function to find jump tables that are live.
229   BitVector JTIsLive(JTI->getJumpTables().size());
230   for (const MachineBasicBlock &BB : MF) {
231     for (const MachineInstr &I : BB)
232       for (const MachineOperand &Op : I.operands()) {
233         if (!Op.isJTI()) continue;
234 
235         // Remember that this JT is live.
236         JTIsLive.set(Op.getIndex());
237       }
238   }
239 
240   // Finally, remove dead jump tables.  This happens when the
241   // indirect jump was unreachable (and thus deleted).
242   for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
243     if (!JTIsLive.test(i)) {
244       JTI->RemoveJumpTable(i);
245       MadeChange = true;
246     }
247 
248   return MadeChange;
249 }
250 
251 //===----------------------------------------------------------------------===//
252 //  Tail Merging of Blocks
253 //===----------------------------------------------------------------------===//
254 
255 /// HashMachineInstr - Compute a hash value for MI and its operands.
HashMachineInstr(const MachineInstr & MI)256 static unsigned HashMachineInstr(const MachineInstr &MI) {
257   unsigned Hash = MI.getOpcode();
258   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
259     const MachineOperand &Op = MI.getOperand(i);
260 
261     // Merge in bits from the operand if easy. We can't use MachineOperand's
262     // hash_code here because it's not deterministic and we sort by hash value
263     // later.
264     unsigned OperandHash = 0;
265     switch (Op.getType()) {
266     case MachineOperand::MO_Register:
267       OperandHash = Op.getReg();
268       break;
269     case MachineOperand::MO_Immediate:
270       OperandHash = Op.getImm();
271       break;
272     case MachineOperand::MO_MachineBasicBlock:
273       OperandHash = Op.getMBB()->getNumber();
274       break;
275     case MachineOperand::MO_FrameIndex:
276     case MachineOperand::MO_ConstantPoolIndex:
277     case MachineOperand::MO_JumpTableIndex:
278       OperandHash = Op.getIndex();
279       break;
280     case MachineOperand::MO_GlobalAddress:
281     case MachineOperand::MO_ExternalSymbol:
282       // Global address / external symbol are too hard, don't bother, but do
283       // pull in the offset.
284       OperandHash = Op.getOffset();
285       break;
286     default:
287       break;
288     }
289 
290     Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
291   }
292   return Hash;
293 }
294 
295 /// HashEndOfMBB - Hash the last instruction in the MBB.
HashEndOfMBB(const MachineBasicBlock & MBB)296 static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) {
297   MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr(false);
298   if (I == MBB.end())
299     return 0;
300 
301   return HashMachineInstr(*I);
302 }
303 
304 /// Whether MI should be counted as an instruction when calculating common tail.
countsAsInstruction(const MachineInstr & MI)305 static bool countsAsInstruction(const MachineInstr &MI) {
306   return !(MI.isDebugInstr() || MI.isCFIInstruction());
307 }
308 
309 /// Iterate backwards from the given iterator \p I, towards the beginning of the
310 /// block. If a MI satisfying 'countsAsInstruction' is found, return an iterator
311 /// pointing to that MI. If no such MI is found, return the end iterator.
312 static MachineBasicBlock::iterator
skipBackwardPastNonInstructions(MachineBasicBlock::iterator I,MachineBasicBlock * MBB)313 skipBackwardPastNonInstructions(MachineBasicBlock::iterator I,
314                                 MachineBasicBlock *MBB) {
315   while (I != MBB->begin()) {
316     --I;
317     if (countsAsInstruction(*I))
318       return I;
319   }
320   return MBB->end();
321 }
322 
323 /// Given two machine basic blocks, return the number of instructions they
324 /// actually have in common together at their end. If a common tail is found (at
325 /// least by one instruction), then iterators for the first shared instruction
326 /// in each block are returned as well.
327 ///
328 /// Non-instructions according to countsAsInstruction are ignored.
ComputeCommonTailLength(MachineBasicBlock * MBB1,MachineBasicBlock * MBB2,MachineBasicBlock::iterator & I1,MachineBasicBlock::iterator & I2)329 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
330                                         MachineBasicBlock *MBB2,
331                                         MachineBasicBlock::iterator &I1,
332                                         MachineBasicBlock::iterator &I2) {
333   MachineBasicBlock::iterator MBBI1 = MBB1->end();
334   MachineBasicBlock::iterator MBBI2 = MBB2->end();
335 
336   unsigned TailLen = 0;
337   while (true) {
338     MBBI1 = skipBackwardPastNonInstructions(MBBI1, MBB1);
339     MBBI2 = skipBackwardPastNonInstructions(MBBI2, MBB2);
340     if (MBBI1 == MBB1->end() || MBBI2 == MBB2->end())
341       break;
342     if (!MBBI1->isIdenticalTo(*MBBI2) ||
343         // FIXME: This check is dubious. It's used to get around a problem where
344         // people incorrectly expect inline asm directives to remain in the same
345         // relative order. This is untenable because normal compiler
346         // optimizations (like this one) may reorder and/or merge these
347         // directives.
348         MBBI1->isInlineAsm()) {
349       break;
350     }
351     if (MBBI1->getFlag(MachineInstr::NoMerge) ||
352         MBBI2->getFlag(MachineInstr::NoMerge))
353       break;
354     ++TailLen;
355     I1 = MBBI1;
356     I2 = MBBI2;
357   }
358 
359   return TailLen;
360 }
361 
replaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,MachineBasicBlock & NewDest)362 void BranchFolder::replaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
363                                            MachineBasicBlock &NewDest) {
364   if (UpdateLiveIns) {
365     // OldInst should always point to an instruction.
366     MachineBasicBlock &OldMBB = *OldInst->getParent();
367     LiveRegs.clear();
368     LiveRegs.addLiveOuts(OldMBB);
369     // Move backward to the place where will insert the jump.
370     MachineBasicBlock::iterator I = OldMBB.end();
371     do {
372       --I;
373       LiveRegs.stepBackward(*I);
374     } while (I != OldInst);
375 
376     // Merging the tails may have switched some undef operand to non-undef ones.
377     // Add IMPLICIT_DEFS into OldMBB as necessary to have a definition of the
378     // register.
379     for (MachineBasicBlock::RegisterMaskPair P : NewDest.liveins()) {
380       // We computed the liveins with computeLiveIn earlier and should only see
381       // full registers:
382       assert(P.LaneMask == LaneBitmask::getAll() &&
383              "Can only handle full register.");
384       MCPhysReg Reg = P.PhysReg;
385       if (!LiveRegs.available(*MRI, Reg))
386         continue;
387       DebugLoc DL;
388       BuildMI(OldMBB, OldInst, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Reg);
389     }
390   }
391 
392   TII->ReplaceTailWithBranchTo(OldInst, &NewDest);
393   ++NumTailMerge;
394 }
395 
SplitMBBAt(MachineBasicBlock & CurMBB,MachineBasicBlock::iterator BBI1,const BasicBlock * BB)396 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
397                                             MachineBasicBlock::iterator BBI1,
398                                             const BasicBlock *BB) {
399   if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
400     return nullptr;
401 
402   MachineFunction &MF = *CurMBB.getParent();
403 
404   // Create the fall-through block.
405   MachineFunction::iterator MBBI = CurMBB.getIterator();
406   MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(BB);
407   CurMBB.getParent()->insert(++MBBI, NewMBB);
408 
409   // Move all the successors of this block to the specified block.
410   NewMBB->transferSuccessors(&CurMBB);
411 
412   // Add an edge from CurMBB to NewMBB for the fall-through.
413   CurMBB.addSuccessor(NewMBB);
414 
415   // Splice the code over.
416   NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
417 
418   // NewMBB belongs to the same loop as CurMBB.
419   if (MLI)
420     if (MachineLoop *ML = MLI->getLoopFor(&CurMBB))
421       ML->addBasicBlockToLoop(NewMBB, *MLI);
422 
423   // NewMBB inherits CurMBB's block frequency.
424   MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
425 
426   if (UpdateLiveIns)
427     computeAndAddLiveIns(LiveRegs, *NewMBB);
428 
429   // Add the new block to the EH scope.
430   const auto &EHScopeI = EHScopeMembership.find(&CurMBB);
431   if (EHScopeI != EHScopeMembership.end()) {
432     auto n = EHScopeI->second;
433     EHScopeMembership[NewMBB] = n;
434   }
435 
436   return NewMBB;
437 }
438 
439 /// EstimateRuntime - Make a rough estimate for how long it will take to run
440 /// the specified code.
EstimateRuntime(MachineBasicBlock::iterator I,MachineBasicBlock::iterator E)441 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
442                                 MachineBasicBlock::iterator E) {
443   unsigned Time = 0;
444   for (; I != E; ++I) {
445     if (!countsAsInstruction(*I))
446       continue;
447     if (I->isCall())
448       Time += 10;
449     else if (I->mayLoadOrStore())
450       Time += 2;
451     else
452       ++Time;
453   }
454   return Time;
455 }
456 
457 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
458 // branches temporarily for tail merging).  In the case where CurMBB ends
459 // with a conditional branch to the next block, optimize by reversing the
460 // test and conditionally branching to SuccMBB instead.
FixTail(MachineBasicBlock * CurMBB,MachineBasicBlock * SuccBB,const TargetInstrInfo * TII,const DebugLoc & BranchDL)461 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
462                     const TargetInstrInfo *TII, const DebugLoc &BranchDL) {
463   MachineFunction *MF = CurMBB->getParent();
464   MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
465   MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
466   SmallVector<MachineOperand, 4> Cond;
467   DebugLoc dl = CurMBB->findBranchDebugLoc();
468   if (!dl)
469     dl = BranchDL;
470   if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
471     MachineBasicBlock *NextBB = &*I;
472     if (TBB == NextBB && !Cond.empty() && !FBB) {
473       if (!TII->reverseBranchCondition(Cond)) {
474         TII->removeBranch(*CurMBB);
475         TII->insertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
476         return;
477       }
478     }
479   }
480   TII->insertBranch(*CurMBB, SuccBB, nullptr,
481                     SmallVector<MachineOperand, 0>(), dl);
482 }
483 
484 bool
operator <(const MergePotentialsElt & o) const485 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
486   if (getHash() < o.getHash())
487     return true;
488   if (getHash() > o.getHash())
489     return false;
490   if (getBlock()->getNumber() < o.getBlock()->getNumber())
491     return true;
492   if (getBlock()->getNumber() > o.getBlock()->getNumber())
493     return false;
494   return false;
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.
CountTerminators(MachineBasicBlock * MBB,MachineBasicBlock::iterator & I)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.
blockEndsInUnreachable(const MachineBasicBlock * MBB)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
ProfitableToMerge(MachineBasicBlock * MBB1,MachineBasicBlock * MBB2,unsigned MinCommonTailLength,unsigned & CommonTailLen,MachineBasicBlock::iterator & I1,MachineBasicBlock::iterator & I2,MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB,DenseMap<const MachineBasicBlock *,int> & EHScopeMembership,bool AfterPlacement,MBFIWrapper & MBBFreqInfo,ProfileSummaryInfo * PSI)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 
ComputeSameTails(unsigned CurHash,unsigned MinCommonTailLength,MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB)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 
RemoveBlocksWithHash(unsigned CurHash,MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB,const DebugLoc & BranchDL)693 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
694                                         MachineBasicBlock *SuccBB,
695                                         MachineBasicBlock *PredBB,
696                                         const DebugLoc &BranchDL) {
697   MPIterator CurMPIter, B;
698   for (CurMPIter = std::prev(MergePotentials.end()),
699       B = MergePotentials.begin();
700        CurMPIter->getHash() == CurHash; --CurMPIter) {
701     // Put the unconditional branch back, if we need one.
702     MachineBasicBlock *CurMBB = CurMPIter->getBlock();
703     if (SuccBB && CurMBB != PredBB)
704       FixTail(CurMBB, SuccBB, TII, BranchDL);
705     if (CurMPIter == B)
706       break;
707   }
708   if (CurMPIter->getHash() != CurHash)
709     CurMPIter++;
710   MergePotentials.erase(CurMPIter, MergePotentials.end());
711 }
712 
CreateCommonTailOnlyBlock(MachineBasicBlock * & PredBB,MachineBasicBlock * SuccBB,unsigned maxCommonTailLength,unsigned & commonTailIndex)713 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
714                                              MachineBasicBlock *SuccBB,
715                                              unsigned maxCommonTailLength,
716                                              unsigned &commonTailIndex) {
717   commonTailIndex = 0;
718   unsigned TimeEstimate = ~0U;
719   for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
720     // Use PredBB if possible; that doesn't require a new branch.
721     if (SameTails[i].getBlock() == PredBB) {
722       commonTailIndex = i;
723       break;
724     }
725     // Otherwise, make a (fairly bogus) choice based on estimate of
726     // how long it will take the various blocks to execute.
727     unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
728                                  SameTails[i].getTailStartPos());
729     if (t <= TimeEstimate) {
730       TimeEstimate = t;
731       commonTailIndex = i;
732     }
733   }
734 
735   MachineBasicBlock::iterator BBI =
736     SameTails[commonTailIndex].getTailStartPos();
737   MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
738 
739   LLVM_DEBUG(dbgs() << "\nSplitting " << printMBBReference(*MBB) << ", size "
740                     << maxCommonTailLength);
741 
742   // If the split block unconditionally falls-thru to SuccBB, it will be
743   // merged. In control flow terms it should then take SuccBB's name. e.g. If
744   // SuccBB is an inner loop, the common tail is still part of the inner loop.
745   const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
746     SuccBB->getBasicBlock() : MBB->getBasicBlock();
747   MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
748   if (!newMBB) {
749     LLVM_DEBUG(dbgs() << "... failed!");
750     return false;
751   }
752 
753   SameTails[commonTailIndex].setBlock(newMBB);
754   SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
755 
756   // If we split PredBB, newMBB is the new predecessor.
757   if (PredBB == MBB)
758     PredBB = newMBB;
759 
760   return true;
761 }
762 
763 static void
mergeOperations(MachineBasicBlock::iterator MBBIStartPos,MachineBasicBlock & MBBCommon)764 mergeOperations(MachineBasicBlock::iterator MBBIStartPos,
765                 MachineBasicBlock &MBBCommon) {
766   MachineBasicBlock *MBB = MBBIStartPos->getParent();
767   // Note CommonTailLen does not necessarily matches the size of
768   // the common BB nor all its instructions because of debug
769   // instructions differences.
770   unsigned CommonTailLen = 0;
771   for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
772     ++CommonTailLen;
773 
774   MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
775   MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
776   MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
777   MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
778 
779   while (CommonTailLen--) {
780     assert(MBBI != MBBIE && "Reached BB end within common tail length!");
781     (void)MBBIE;
782 
783     if (!countsAsInstruction(*MBBI)) {
784       ++MBBI;
785       continue;
786     }
787 
788     while ((MBBICommon != MBBIECommon) && !countsAsInstruction(*MBBICommon))
789       ++MBBICommon;
790 
791     assert(MBBICommon != MBBIECommon &&
792            "Reached BB end within common tail length!");
793     assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!");
794 
795     // Merge MMOs from memory operations in the common block.
796     if (MBBICommon->mayLoadOrStore())
797       MBBICommon->cloneMergedMemRefs(*MBB->getParent(), {&*MBBICommon, &*MBBI});
798     // Drop undef flags if they aren't present in all merged instructions.
799     for (unsigned I = 0, E = MBBICommon->getNumOperands(); I != E; ++I) {
800       MachineOperand &MO = MBBICommon->getOperand(I);
801       if (MO.isReg() && MO.isUndef()) {
802         const MachineOperand &OtherMO = MBBI->getOperand(I);
803         if (!OtherMO.isUndef())
804           MO.setIsUndef(false);
805       }
806     }
807 
808     ++MBBI;
809     ++MBBICommon;
810   }
811 }
812 
mergeCommonTails(unsigned commonTailIndex)813 void BranchFolder::mergeCommonTails(unsigned commonTailIndex) {
814   MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
815 
816   std::vector<MachineBasicBlock::iterator> NextCommonInsts(SameTails.size());
817   for (unsigned int i = 0 ; i != SameTails.size() ; ++i) {
818     if (i != commonTailIndex) {
819       NextCommonInsts[i] = SameTails[i].getTailStartPos();
820       mergeOperations(SameTails[i].getTailStartPos(), *MBB);
821     } else {
822       assert(SameTails[i].getTailStartPos() == MBB->begin() &&
823           "MBB is not a common tail only block");
824     }
825   }
826 
827   for (auto &MI : *MBB) {
828     if (!countsAsInstruction(MI))
829       continue;
830     DebugLoc DL = MI.getDebugLoc();
831     for (unsigned int i = 0 ; i < NextCommonInsts.size() ; i++) {
832       if (i == commonTailIndex)
833         continue;
834 
835       auto &Pos = NextCommonInsts[i];
836       assert(Pos != SameTails[i].getBlock()->end() &&
837           "Reached BB end within common tail");
838       while (!countsAsInstruction(*Pos)) {
839         ++Pos;
840         assert(Pos != SameTails[i].getBlock()->end() &&
841             "Reached BB end within common tail");
842       }
843       assert(MI.isIdenticalTo(*Pos) && "Expected matching MIIs!");
844       DL = DILocation::getMergedLocation(DL, Pos->getDebugLoc());
845       NextCommonInsts[i] = ++Pos;
846     }
847     MI.setDebugLoc(DL);
848   }
849 
850   if (UpdateLiveIns) {
851     LivePhysRegs NewLiveIns(*TRI);
852     computeLiveIns(NewLiveIns, *MBB);
853     LiveRegs.init(*TRI);
854 
855     // The flag merging may lead to some register uses no longer using the
856     // <undef> flag, add IMPLICIT_DEFs in the predecessors as necessary.
857     for (MachineBasicBlock *Pred : MBB->predecessors()) {
858       LiveRegs.clear();
859       LiveRegs.addLiveOuts(*Pred);
860       MachineBasicBlock::iterator InsertBefore = Pred->getFirstTerminator();
861       for (Register Reg : NewLiveIns) {
862         if (!LiveRegs.available(*MRI, Reg))
863           continue;
864 
865         // Skip the register if we are about to add one of its super registers.
866         // TODO: Common this up with the same logic in addLineIns().
867         if (any_of(TRI->superregs(Reg), [&](MCPhysReg SReg) {
868               return NewLiveIns.contains(SReg) && !MRI->isReserved(SReg);
869             }))
870           continue;
871 
872         DebugLoc DL;
873         BuildMI(*Pred, InsertBefore, DL, TII->get(TargetOpcode::IMPLICIT_DEF),
874                 Reg);
875       }
876     }
877 
878     MBB->clearLiveIns();
879     addLiveIns(*MBB, NewLiveIns);
880   }
881 }
882 
883 // See if any of the blocks in MergePotentials (which all have SuccBB as a
884 // successor, or all have no successor if it is null) can be tail-merged.
885 // If there is a successor, any blocks in MergePotentials that are not
886 // tail-merged and are not immediately before Succ must have an unconditional
887 // branch to Succ added (but the predecessor/successor lists need no
888 // adjustment). The lone predecessor of Succ that falls through into Succ,
889 // if any, is given in PredBB.
890 // MinCommonTailLength - Except for the special cases below, tail-merge if
891 // there are at least this many instructions in common.
TryTailMergeBlocks(MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB,unsigned MinCommonTailLength)892 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
893                                       MachineBasicBlock *PredBB,
894                                       unsigned MinCommonTailLength) {
895   bool MadeChange = false;
896 
897   LLVM_DEBUG(
898       dbgs() << "\nTryTailMergeBlocks: ";
899       for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) dbgs()
900       << printMBBReference(*MergePotentials[i].getBlock())
901       << (i == e - 1 ? "" : ", ");
902       dbgs() << "\n"; if (SuccBB) {
903         dbgs() << "  with successor " << printMBBReference(*SuccBB) << '\n';
904         if (PredBB)
905           dbgs() << "  which has fall-through from "
906                  << printMBBReference(*PredBB) << "\n";
907       } dbgs() << "Looking for common tails of at least "
908                << MinCommonTailLength << " instruction"
909                << (MinCommonTailLength == 1 ? "" : "s") << '\n';);
910 
911   // Sort by hash value so that blocks with identical end sequences sort
912   // together.
913   array_pod_sort(MergePotentials.begin(), MergePotentials.end());
914 
915   // Walk through equivalence sets looking for actual exact matches.
916   while (MergePotentials.size() > 1) {
917     unsigned CurHash = MergePotentials.back().getHash();
918     const DebugLoc &BranchDL = MergePotentials.back().getBranchDebugLoc();
919 
920     // Build SameTails, identifying the set of blocks with this hash code
921     // and with the maximum number of instructions in common.
922     unsigned maxCommonTailLength = ComputeSameTails(CurHash,
923                                                     MinCommonTailLength,
924                                                     SuccBB, PredBB);
925 
926     // If we didn't find any pair that has at least MinCommonTailLength
927     // instructions in common, remove all blocks with this hash code and retry.
928     if (SameTails.empty()) {
929       RemoveBlocksWithHash(CurHash, SuccBB, PredBB, BranchDL);
930       continue;
931     }
932 
933     // If one of the blocks is the entire common tail (and is not the entry
934     // block/an EH pad, which we can't jump to), we can treat all blocks with
935     // this same tail at once.  Use PredBB if that is one of the possibilities,
936     // as that will not introduce any extra branches.
937     MachineBasicBlock *EntryBB =
938         &MergePotentials.front().getBlock()->getParent()->front();
939     unsigned commonTailIndex = SameTails.size();
940     // If there are two blocks, check to see if one can be made to fall through
941     // into the other.
942     if (SameTails.size() == 2 &&
943         SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
944         SameTails[1].tailIsWholeBlock() && !SameTails[1].getBlock()->isEHPad())
945       commonTailIndex = 1;
946     else if (SameTails.size() == 2 &&
947              SameTails[1].getBlock()->isLayoutSuccessor(
948                  SameTails[0].getBlock()) &&
949              SameTails[0].tailIsWholeBlock() &&
950              !SameTails[0].getBlock()->isEHPad())
951       commonTailIndex = 0;
952     else {
953       // Otherwise just pick one, favoring the fall-through predecessor if
954       // there is one.
955       for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
956         MachineBasicBlock *MBB = SameTails[i].getBlock();
957         if ((MBB == EntryBB || MBB->isEHPad()) &&
958             SameTails[i].tailIsWholeBlock())
959           continue;
960         if (MBB == PredBB) {
961           commonTailIndex = i;
962           break;
963         }
964         if (SameTails[i].tailIsWholeBlock())
965           commonTailIndex = i;
966       }
967     }
968 
969     if (commonTailIndex == SameTails.size() ||
970         (SameTails[commonTailIndex].getBlock() == PredBB &&
971          !SameTails[commonTailIndex].tailIsWholeBlock())) {
972       // None of the blocks consist entirely of the common tail.
973       // Split a block so that one does.
974       if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
975                                      maxCommonTailLength, commonTailIndex)) {
976         RemoveBlocksWithHash(CurHash, SuccBB, PredBB, BranchDL);
977         continue;
978       }
979     }
980 
981     MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
982 
983     // Recompute common tail MBB's edge weights and block frequency.
984     setCommonTailEdgeWeights(*MBB);
985 
986     // Merge debug locations, MMOs and undef flags across identical instructions
987     // for common tail.
988     mergeCommonTails(commonTailIndex);
989 
990     // MBB is common tail.  Adjust all other BB's to jump to this one.
991     // Traversal must be forwards so erases work.
992     LLVM_DEBUG(dbgs() << "\nUsing common tail in " << printMBBReference(*MBB)
993                       << " for ");
994     for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
995       if (commonTailIndex == i)
996         continue;
997       LLVM_DEBUG(dbgs() << printMBBReference(*SameTails[i].getBlock())
998                         << (i == e - 1 ? "" : ", "));
999       // Hack the end off BB i, making it jump to BB commonTailIndex instead.
1000       replaceTailWithBranchTo(SameTails[i].getTailStartPos(), *MBB);
1001       // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
1002       MergePotentials.erase(SameTails[i].getMPIter());
1003     }
1004     LLVM_DEBUG(dbgs() << "\n");
1005     // We leave commonTailIndex in the worklist in case there are other blocks
1006     // that match it with a smaller number of instructions.
1007     MadeChange = true;
1008   }
1009   return MadeChange;
1010 }
1011 
TailMergeBlocks(MachineFunction & MF)1012 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
1013   bool MadeChange = false;
1014   if (!EnableTailMerge)
1015     return MadeChange;
1016 
1017   // First find blocks with no successors.
1018   // Block placement may create new tail merging opportunities for these blocks.
1019   MergePotentials.clear();
1020   for (MachineBasicBlock &MBB : MF) {
1021     if (MergePotentials.size() == TailMergeThreshold)
1022       break;
1023     if (!TriedMerging.count(&MBB) && MBB.succ_empty())
1024       MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB,
1025                                                    MBB.findBranchDebugLoc()));
1026   }
1027 
1028   // If this is a large problem, avoid visiting the same basic blocks
1029   // multiple times.
1030   if (MergePotentials.size() == TailMergeThreshold)
1031     for (const MergePotentialsElt &Elt : MergePotentials)
1032       TriedMerging.insert(Elt.getBlock());
1033 
1034   // See if we can do any tail merging on those.
1035   if (MergePotentials.size() >= 2)
1036     MadeChange |= TryTailMergeBlocks(nullptr, nullptr, MinCommonTailLength);
1037 
1038   // Look at blocks (IBB) with multiple predecessors (PBB).
1039   // We change each predecessor to a canonical form, by
1040   // (1) temporarily removing any unconditional branch from the predecessor
1041   // to IBB, and
1042   // (2) alter conditional branches so they branch to the other block
1043   // not IBB; this may require adding back an unconditional branch to IBB
1044   // later, where there wasn't one coming in.  E.g.
1045   //   Bcc IBB
1046   //   fallthrough to QBB
1047   // here becomes
1048   //   Bncc QBB
1049   // with a conceptual B to IBB after that, which never actually exists.
1050   // With those changes, we see whether the predecessors' tails match,
1051   // and merge them if so.  We change things out of canonical form and
1052   // back to the way they were later in the process.  (OptimizeBranches
1053   // would undo some of this, but we can't use it, because we'd get into
1054   // a compile-time infinite loop repeatedly doing and undoing the same
1055   // transformations.)
1056 
1057   for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1058        I != E; ++I) {
1059     if (I->pred_size() < 2) continue;
1060     SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
1061     MachineBasicBlock *IBB = &*I;
1062     MachineBasicBlock *PredBB = &*std::prev(I);
1063     MergePotentials.clear();
1064     MachineLoop *ML;
1065 
1066     // Bail if merging after placement and IBB is the loop header because
1067     // -- If merging predecessors that belong to the same loop as IBB, the
1068     // common tail of merged predecessors may become the loop top if block
1069     // placement is called again and the predecessors may branch to this common
1070     // tail and require more branches. This can be relaxed if
1071     // MachineBlockPlacement::findBestLoopTop is more flexible.
1072     // --If merging predecessors that do not belong to the same loop as IBB, the
1073     // loop info of IBB's loop and the other loops may be affected. Calling the
1074     // block placement again may make big change to the layout and eliminate the
1075     // reason to do tail merging here.
1076     if (AfterBlockPlacement && MLI) {
1077       ML = MLI->getLoopFor(IBB);
1078       if (ML && IBB == ML->getHeader())
1079         continue;
1080     }
1081 
1082     for (MachineBasicBlock *PBB : I->predecessors()) {
1083       if (MergePotentials.size() == TailMergeThreshold)
1084         break;
1085 
1086       if (TriedMerging.count(PBB))
1087         continue;
1088 
1089       // Skip blocks that loop to themselves, can't tail merge these.
1090       if (PBB == IBB)
1091         continue;
1092 
1093       // Visit each predecessor only once.
1094       if (!UniquePreds.insert(PBB).second)
1095         continue;
1096 
1097       // Skip blocks which may jump to a landing pad or jump from an asm blob.
1098       // Can't tail merge these.
1099       if (PBB->hasEHPadSuccessor() || PBB->mayHaveInlineAsmBr())
1100         continue;
1101 
1102       // After block placement, only consider predecessors that belong to the
1103       // same loop as IBB.  The reason is the same as above when skipping loop
1104       // header.
1105       if (AfterBlockPlacement && MLI)
1106         if (ML != MLI->getLoopFor(PBB))
1107           continue;
1108 
1109       MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1110       SmallVector<MachineOperand, 4> Cond;
1111       if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) {
1112         // Failing case: IBB is the target of a cbr, and we cannot reverse the
1113         // branch.
1114         SmallVector<MachineOperand, 4> NewCond(Cond);
1115         if (!Cond.empty() && TBB == IBB) {
1116           if (TII->reverseBranchCondition(NewCond))
1117             continue;
1118           // This is the QBB case described above
1119           if (!FBB) {
1120             auto Next = ++PBB->getIterator();
1121             if (Next != MF.end())
1122               FBB = &*Next;
1123           }
1124         }
1125 
1126         // Remove the unconditional branch at the end, if any.
1127         DebugLoc dl = PBB->findBranchDebugLoc();
1128         if (TBB && (Cond.empty() || FBB)) {
1129           TII->removeBranch(*PBB);
1130           if (!Cond.empty())
1131             // reinsert conditional branch only, for now
1132             TII->insertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1133                               NewCond, dl);
1134         }
1135 
1136         MergePotentials.push_back(
1137             MergePotentialsElt(HashEndOfMBB(*PBB), PBB, dl));
1138       }
1139     }
1140 
1141     // If this is a large problem, avoid visiting the same basic blocks multiple
1142     // times.
1143     if (MergePotentials.size() == TailMergeThreshold)
1144       for (MergePotentialsElt &Elt : MergePotentials)
1145         TriedMerging.insert(Elt.getBlock());
1146 
1147     if (MergePotentials.size() >= 2)
1148       MadeChange |= TryTailMergeBlocks(IBB, PredBB, MinCommonTailLength);
1149 
1150     // Reinsert an unconditional branch if needed. The 1 below can occur as a
1151     // result of removing blocks in TryTailMergeBlocks.
1152     PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks
1153     if (MergePotentials.size() == 1 &&
1154         MergePotentials.begin()->getBlock() != PredBB)
1155       FixTail(MergePotentials.begin()->getBlock(), IBB, TII,
1156               MergePotentials.begin()->getBranchDebugLoc());
1157   }
1158 
1159   return MadeChange;
1160 }
1161 
setCommonTailEdgeWeights(MachineBasicBlock & TailMBB)1162 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1163   SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1164   BlockFrequency AccumulatedMBBFreq;
1165 
1166   // Aggregate edge frequency of successor edge j:
1167   //  edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1168   //  where bb is a basic block that is in SameTails.
1169   for (const auto &Src : SameTails) {
1170     const MachineBasicBlock *SrcMBB = Src.getBlock();
1171     BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1172     AccumulatedMBBFreq += BlockFreq;
1173 
1174     // It is not necessary to recompute edge weights if TailBB has less than two
1175     // successors.
1176     if (TailMBB.succ_size() <= 1)
1177       continue;
1178 
1179     auto EdgeFreq = EdgeFreqLs.begin();
1180 
1181     for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1182          SuccI != SuccE; ++SuccI, ++EdgeFreq)
1183       *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1184   }
1185 
1186   MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1187 
1188   if (TailMBB.succ_size() <= 1)
1189     return;
1190 
1191   auto SumEdgeFreq =
1192       std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0))
1193           .getFrequency();
1194   auto EdgeFreq = EdgeFreqLs.begin();
1195 
1196   if (SumEdgeFreq > 0) {
1197     for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1198          SuccI != SuccE; ++SuccI, ++EdgeFreq) {
1199       auto Prob = BranchProbability::getBranchProbability(
1200           EdgeFreq->getFrequency(), SumEdgeFreq);
1201       TailMBB.setSuccProbability(SuccI, Prob);
1202     }
1203   }
1204 }
1205 
1206 //===----------------------------------------------------------------------===//
1207 //  Branch Optimization
1208 //===----------------------------------------------------------------------===//
1209 
OptimizeBranches(MachineFunction & MF)1210 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1211   bool MadeChange = false;
1212 
1213   // Make sure blocks are numbered in order
1214   MF.RenumberBlocks();
1215   // Renumbering blocks alters EH scope membership, recalculate it.
1216   EHScopeMembership = getEHScopeMembership(MF);
1217 
1218   for (MachineBasicBlock &MBB :
1219        llvm::make_early_inc_range(llvm::drop_begin(MF))) {
1220     MadeChange |= OptimizeBlock(&MBB);
1221 
1222     // If it is dead, remove it.
1223     if (MBB.pred_empty() && !MBB.isMachineBlockAddressTaken()) {
1224       RemoveDeadBlock(&MBB);
1225       MadeChange = true;
1226       ++NumDeadBlocks;
1227     }
1228   }
1229 
1230   return MadeChange;
1231 }
1232 
1233 // Blocks should be considered empty if they contain only debug info;
1234 // else the debug info would affect codegen.
IsEmptyBlock(MachineBasicBlock * MBB)1235 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1236   return MBB->getFirstNonDebugInstr(true) == MBB->end();
1237 }
1238 
1239 // Blocks with only debug info and branches should be considered the same
1240 // as blocks with only branches.
IsBranchOnlyBlock(MachineBasicBlock * MBB)1241 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1242   MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1243   assert(I != MBB->end() && "empty block!");
1244   return I->isBranch();
1245 }
1246 
1247 /// IsBetterFallthrough - Return true if it would be clearly better to
1248 /// fall-through to MBB1 than to fall through into MBB2.  This has to return
1249 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1250 /// result in infinite loops.
IsBetterFallthrough(MachineBasicBlock * MBB1,MachineBasicBlock * MBB2)1251 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1252                                 MachineBasicBlock *MBB2) {
1253   assert(MBB1 && MBB2 && "Unknown MachineBasicBlock");
1254 
1255   // Right now, we use a simple heuristic.  If MBB2 ends with a call, and
1256   // MBB1 doesn't, we prefer to fall through into MBB1.  This allows us to
1257   // optimize branches that branch to either a return block or an assert block
1258   // into a fallthrough to the return.
1259   MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1260   MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1261   if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1262     return false;
1263 
1264   // If there is a clear successor ordering we make sure that one block
1265   // will fall through to the next
1266   if (MBB1->isSuccessor(MBB2)) return true;
1267   if (MBB2->isSuccessor(MBB1)) return false;
1268 
1269   return MBB2I->isCall() && !MBB1I->isCall();
1270 }
1271 
1272 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1273 /// instructions on the block.
getBranchDebugLoc(MachineBasicBlock & MBB)1274 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1275   MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1276   if (I != MBB.end() && I->isBranch())
1277     return I->getDebugLoc();
1278   return DebugLoc();
1279 }
1280 
copyDebugInfoToPredecessor(const TargetInstrInfo * TII,MachineBasicBlock & MBB,MachineBasicBlock & PredMBB)1281 static void copyDebugInfoToPredecessor(const TargetInstrInfo *TII,
1282                                        MachineBasicBlock &MBB,
1283                                        MachineBasicBlock &PredMBB) {
1284   auto InsertBefore = PredMBB.getFirstTerminator();
1285   for (MachineInstr &MI : MBB.instrs())
1286     if (MI.isDebugInstr()) {
1287       TII->duplicate(PredMBB, InsertBefore, MI);
1288       LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to pred: "
1289                         << MI);
1290     }
1291 }
1292 
copyDebugInfoToSuccessor(const TargetInstrInfo * TII,MachineBasicBlock & MBB,MachineBasicBlock & SuccMBB)1293 static void copyDebugInfoToSuccessor(const TargetInstrInfo *TII,
1294                                      MachineBasicBlock &MBB,
1295                                      MachineBasicBlock &SuccMBB) {
1296   auto InsertBefore = SuccMBB.SkipPHIsAndLabels(SuccMBB.begin());
1297   for (MachineInstr &MI : MBB.instrs())
1298     if (MI.isDebugInstr()) {
1299       TII->duplicate(SuccMBB, InsertBefore, MI);
1300       LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to succ: "
1301                         << MI);
1302     }
1303 }
1304 
1305 // Try to salvage DBG_VALUE instructions from an otherwise empty block. If such
1306 // a basic block is removed we would lose the debug information unless we have
1307 // copied the information to a predecessor/successor.
1308 //
1309 // TODO: This function only handles some simple cases. An alternative would be
1310 // to run a heavier analysis, such as the LiveDebugValues pass, before we do
1311 // branch folding.
salvageDebugInfoFromEmptyBlock(const TargetInstrInfo * TII,MachineBasicBlock & MBB)1312 static void salvageDebugInfoFromEmptyBlock(const TargetInstrInfo *TII,
1313                                            MachineBasicBlock &MBB) {
1314   assert(IsEmptyBlock(&MBB) && "Expected an empty block (except debug info).");
1315   // If this MBB is the only predecessor of a successor it is legal to copy
1316   // DBG_VALUE instructions to the beginning of the successor.
1317   for (MachineBasicBlock *SuccBB : MBB.successors())
1318     if (SuccBB->pred_size() == 1)
1319       copyDebugInfoToSuccessor(TII, MBB, *SuccBB);
1320   // If this MBB is the only successor of a predecessor it is legal to copy the
1321   // DBG_VALUE instructions to the end of the predecessor (just before the
1322   // terminators, assuming that the terminator isn't affecting the DBG_VALUE).
1323   for (MachineBasicBlock *PredBB : MBB.predecessors())
1324     if (PredBB->succ_size() == 1)
1325       copyDebugInfoToPredecessor(TII, MBB, *PredBB);
1326 }
1327 
OptimizeBlock(MachineBasicBlock * MBB)1328 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1329   bool MadeChange = false;
1330   MachineFunction &MF = *MBB->getParent();
1331 ReoptimizeBlock:
1332 
1333   MachineFunction::iterator FallThrough = MBB->getIterator();
1334   ++FallThrough;
1335 
1336   // Make sure MBB and FallThrough belong to the same EH scope.
1337   bool SameEHScope = true;
1338   if (!EHScopeMembership.empty() && FallThrough != MF.end()) {
1339     auto MBBEHScope = EHScopeMembership.find(MBB);
1340     assert(MBBEHScope != EHScopeMembership.end());
1341     auto FallThroughEHScope = EHScopeMembership.find(&*FallThrough);
1342     assert(FallThroughEHScope != EHScopeMembership.end());
1343     SameEHScope = MBBEHScope->second == FallThroughEHScope->second;
1344   }
1345 
1346   // Analyze the branch in the current block. As a side-effect, this may cause
1347   // the block to become empty.
1348   MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1349   SmallVector<MachineOperand, 4> CurCond;
1350   bool CurUnAnalyzable =
1351       TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1352 
1353   // If this block is empty, make everyone use its fall-through, not the block
1354   // explicitly.  Landing pads should not do this since the landing-pad table
1355   // points to this block.  Blocks with their addresses taken shouldn't be
1356   // optimized away.
1357   if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() &&
1358       SameEHScope) {
1359     salvageDebugInfoFromEmptyBlock(TII, *MBB);
1360     // Dead block?  Leave for cleanup later.
1361     if (MBB->pred_empty()) return MadeChange;
1362 
1363     if (FallThrough == MF.end()) {
1364       // TODO: Simplify preds to not branch here if possible!
1365     } else if (FallThrough->isEHPad()) {
1366       // Don't rewrite to a landing pad fallthough.  That could lead to the case
1367       // where a BB jumps to more than one landing pad.
1368       // TODO: Is it ever worth rewriting predecessors which don't already
1369       // jump to a landing pad, and so can safely jump to the fallthrough?
1370     } else if (MBB->isSuccessor(&*FallThrough)) {
1371       // Rewrite all predecessors of the old block to go to the fallthrough
1372       // instead.
1373       while (!MBB->pred_empty()) {
1374         MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1375         Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough);
1376       }
1377       // Add rest successors of MBB to successors of FallThrough. Those
1378       // successors are not directly reachable via MBB, so it should be
1379       // landing-pad.
1380       for (auto SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI)
1381         if (*SI != &*FallThrough && !FallThrough->isSuccessor(*SI)) {
1382           assert((*SI)->isEHPad() && "Bad CFG");
1383           FallThrough->copySuccessor(MBB, SI);
1384         }
1385       // If MBB was the target of a jump table, update jump tables to go to the
1386       // fallthrough instead.
1387       if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1388         MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough);
1389       MadeChange = true;
1390     }
1391     return MadeChange;
1392   }
1393 
1394   // Check to see if we can simplify the terminator of the block before this
1395   // one.
1396   MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1397 
1398   MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1399   SmallVector<MachineOperand, 4> PriorCond;
1400   bool PriorUnAnalyzable =
1401       TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1402   if (!PriorUnAnalyzable) {
1403     // If the previous branch is conditional and both conditions go to the same
1404     // destination, remove the branch, replacing it with an unconditional one or
1405     // a fall-through.
1406     if (PriorTBB && PriorTBB == PriorFBB) {
1407       DebugLoc dl = getBranchDebugLoc(PrevBB);
1408       TII->removeBranch(PrevBB);
1409       PriorCond.clear();
1410       if (PriorTBB != MBB)
1411         TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1412       MadeChange = true;
1413       ++NumBranchOpts;
1414       goto ReoptimizeBlock;
1415     }
1416 
1417     // If the previous block unconditionally falls through to this block and
1418     // this block has no other predecessors, move the contents of this block
1419     // into the prior block. This doesn't usually happen when SimplifyCFG
1420     // has been used, but it can happen if tail merging splits a fall-through
1421     // predecessor of a block.
1422     // This has to check PrevBB->succ_size() because EH edges are ignored by
1423     // analyzeBranch.
1424     if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1425         PrevBB.succ_size() == 1 && PrevBB.isSuccessor(MBB) &&
1426         !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1427       LLVM_DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1428                         << "From MBB: " << *MBB);
1429       // Remove redundant DBG_VALUEs first.
1430       if (!PrevBB.empty()) {
1431         MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1432         --PrevBBIter;
1433         MachineBasicBlock::iterator MBBIter = MBB->begin();
1434         // Check if DBG_VALUE at the end of PrevBB is identical to the
1435         // DBG_VALUE at the beginning of MBB.
1436         while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1437                && PrevBBIter->isDebugInstr() && MBBIter->isDebugInstr()) {
1438           if (!MBBIter->isIdenticalTo(*PrevBBIter))
1439             break;
1440           MachineInstr &DuplicateDbg = *MBBIter;
1441           ++MBBIter; -- PrevBBIter;
1442           DuplicateDbg.eraseFromParent();
1443         }
1444       }
1445       PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1446       PrevBB.removeSuccessor(PrevBB.succ_begin());
1447       assert(PrevBB.succ_empty());
1448       PrevBB.transferSuccessors(MBB);
1449       MadeChange = true;
1450       return MadeChange;
1451     }
1452 
1453     // If the previous branch *only* branches to *this* block (conditional or
1454     // not) remove the branch.
1455     if (PriorTBB == MBB && !PriorFBB) {
1456       TII->removeBranch(PrevBB);
1457       MadeChange = true;
1458       ++NumBranchOpts;
1459       goto ReoptimizeBlock;
1460     }
1461 
1462     // If the prior block branches somewhere else on the condition and here if
1463     // the condition is false, remove the uncond second branch.
1464     if (PriorFBB == MBB) {
1465       DebugLoc dl = getBranchDebugLoc(PrevBB);
1466       TII->removeBranch(PrevBB);
1467       TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1468       MadeChange = true;
1469       ++NumBranchOpts;
1470       goto ReoptimizeBlock;
1471     }
1472 
1473     // If the prior block branches here on true and somewhere else on false, and
1474     // if the branch condition is reversible, reverse the branch to create a
1475     // fall-through.
1476     if (PriorTBB == MBB) {
1477       SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1478       if (!TII->reverseBranchCondition(NewPriorCond)) {
1479         DebugLoc dl = getBranchDebugLoc(PrevBB);
1480         TII->removeBranch(PrevBB);
1481         TII->insertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1482         MadeChange = true;
1483         ++NumBranchOpts;
1484         goto ReoptimizeBlock;
1485       }
1486     }
1487 
1488     // If this block has no successors (e.g. it is a return block or ends with
1489     // a call to a no-return function like abort or __cxa_throw) and if the pred
1490     // falls through into this block, and if it would otherwise fall through
1491     // into the block after this, move this block to the end of the function.
1492     //
1493     // We consider it more likely that execution will stay in the function (e.g.
1494     // due to loops) than it is to exit it.  This asserts in loops etc, moving
1495     // the assert condition out of the loop body.
1496     if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1497         MachineFunction::iterator(PriorTBB) == FallThrough &&
1498         !MBB->canFallThrough()) {
1499       bool DoTransform = true;
1500 
1501       // We have to be careful that the succs of PredBB aren't both no-successor
1502       // blocks.  If neither have successors and if PredBB is the second from
1503       // last block in the function, we'd just keep swapping the two blocks for
1504       // last.  Only do the swap if one is clearly better to fall through than
1505       // the other.
1506       if (FallThrough == --MF.end() &&
1507           !IsBetterFallthrough(PriorTBB, MBB))
1508         DoTransform = false;
1509 
1510       if (DoTransform) {
1511         // Reverse the branch so we will fall through on the previous true cond.
1512         SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1513         if (!TII->reverseBranchCondition(NewPriorCond)) {
1514           LLVM_DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1515                             << "To make fallthrough to: " << *PriorTBB << "\n");
1516 
1517           DebugLoc dl = getBranchDebugLoc(PrevBB);
1518           TII->removeBranch(PrevBB);
1519           TII->insertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1520 
1521           // Move this block to the end of the function.
1522           MBB->moveAfter(&MF.back());
1523           MadeChange = true;
1524           ++NumBranchOpts;
1525           return MadeChange;
1526         }
1527       }
1528     }
1529   }
1530 
1531   if (!IsEmptyBlock(MBB)) {
1532     MachineInstr &TailCall = *MBB->getFirstNonDebugInstr();
1533     if (TII->isUnconditionalTailCall(TailCall)) {
1534       SmallVector<MachineBasicBlock *> PredsChanged;
1535       for (auto &Pred : MBB->predecessors()) {
1536         MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1537         SmallVector<MachineOperand, 4> PredCond;
1538         bool PredAnalyzable =
1539             !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true);
1540 
1541         // Only eliminate if MBB == TBB (Taken Basic Block)
1542         if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB &&
1543             PredTBB != PredFBB) {
1544           // The predecessor has a conditional branch to this block which
1545           // consists of only a tail call. Try to fold the tail call into the
1546           // conditional branch.
1547           if (TII->canMakeTailCallConditional(PredCond, TailCall)) {
1548             // TODO: It would be nice if analyzeBranch() could provide a pointer
1549             // to the branch instruction so replaceBranchWithTailCall() doesn't
1550             // have to search for it.
1551             TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall);
1552             PredsChanged.push_back(Pred);
1553           }
1554         }
1555         // If the predecessor is falling through to this block, we could reverse
1556         // the branch condition and fold the tail call into that. However, after
1557         // that we might have to re-arrange the CFG to fall through to the other
1558         // block and there is a high risk of regressing code size rather than
1559         // improving it.
1560       }
1561       if (!PredsChanged.empty()) {
1562         NumTailCalls += PredsChanged.size();
1563         for (auto &Pred : PredsChanged)
1564           Pred->removeSuccessor(MBB);
1565 
1566         return true;
1567       }
1568     }
1569   }
1570 
1571   if (!CurUnAnalyzable) {
1572     // If this is a two-way branch, and the FBB branches to this block, reverse
1573     // the condition so the single-basic-block loop is faster.  Instead of:
1574     //    Loop: xxx; jcc Out; jmp Loop
1575     // we want:
1576     //    Loop: xxx; jncc Loop; jmp Out
1577     if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1578       SmallVector<MachineOperand, 4> NewCond(CurCond);
1579       if (!TII->reverseBranchCondition(NewCond)) {
1580         DebugLoc dl = getBranchDebugLoc(*MBB);
1581         TII->removeBranch(*MBB);
1582         TII->insertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1583         MadeChange = true;
1584         ++NumBranchOpts;
1585         goto ReoptimizeBlock;
1586       }
1587     }
1588 
1589     // If this branch is the only thing in its block, see if we can forward
1590     // other blocks across it.
1591     if (CurTBB && CurCond.empty() && !CurFBB &&
1592         IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1593         !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1594       DebugLoc dl = getBranchDebugLoc(*MBB);
1595       // This block may contain just an unconditional branch.  Because there can
1596       // be 'non-branch terminators' in the block, try removing the branch and
1597       // then seeing if the block is empty.
1598       TII->removeBranch(*MBB);
1599       // If the only things remaining in the block are debug info, remove these
1600       // as well, so this will behave the same as an empty block in non-debug
1601       // mode.
1602       if (IsEmptyBlock(MBB)) {
1603         // Make the block empty, losing the debug info (we could probably
1604         // improve this in some cases.)
1605         MBB->erase(MBB->begin(), MBB->end());
1606       }
1607       // If this block is just an unconditional branch to CurTBB, we can
1608       // usually completely eliminate the block.  The only case we cannot
1609       // completely eliminate the block is when the block before this one
1610       // falls through into MBB and we can't understand the prior block's branch
1611       // condition.
1612       if (MBB->empty()) {
1613         bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1614         if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1615             !PrevBB.isSuccessor(MBB)) {
1616           // If the prior block falls through into us, turn it into an
1617           // explicit branch to us to make updates simpler.
1618           if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1619               PriorTBB != MBB && PriorFBB != MBB) {
1620             if (!PriorTBB) {
1621               assert(PriorCond.empty() && !PriorFBB &&
1622                      "Bad branch analysis");
1623               PriorTBB = MBB;
1624             } else {
1625               assert(!PriorFBB && "Machine CFG out of date!");
1626               PriorFBB = MBB;
1627             }
1628             DebugLoc pdl = getBranchDebugLoc(PrevBB);
1629             TII->removeBranch(PrevBB);
1630             TII->insertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1631           }
1632 
1633           // Iterate through all the predecessors, revectoring each in-turn.
1634           size_t PI = 0;
1635           bool DidChange = false;
1636           bool HasBranchToSelf = false;
1637           while(PI != MBB->pred_size()) {
1638             MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1639             if (PMBB == MBB) {
1640               // If this block has an uncond branch to itself, leave it.
1641               ++PI;
1642               HasBranchToSelf = true;
1643             } else {
1644               DidChange = true;
1645               PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1646               // Add rest successors of MBB to successors of CurTBB. Those
1647               // successors are not directly reachable via MBB, so it should be
1648               // landing-pad.
1649               for (auto SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE;
1650                    ++SI)
1651                 if (*SI != CurTBB && !CurTBB->isSuccessor(*SI)) {
1652                   assert((*SI)->isEHPad() && "Bad CFG");
1653                   CurTBB->copySuccessor(MBB, SI);
1654                 }
1655               // If this change resulted in PMBB ending in a conditional
1656               // branch where both conditions go to the same destination,
1657               // change this to an unconditional branch.
1658               MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1659               SmallVector<MachineOperand, 4> NewCurCond;
1660               bool NewCurUnAnalyzable = TII->analyzeBranch(
1661                   *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true);
1662               if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1663                 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1664                 TII->removeBranch(*PMBB);
1665                 NewCurCond.clear();
1666                 TII->insertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1667                 MadeChange = true;
1668                 ++NumBranchOpts;
1669               }
1670             }
1671           }
1672 
1673           // Change any jumptables to go to the new MBB.
1674           if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1675             MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1676           if (DidChange) {
1677             ++NumBranchOpts;
1678             MadeChange = true;
1679             if (!HasBranchToSelf) return MadeChange;
1680           }
1681         }
1682       }
1683 
1684       // Add the branch back if the block is more than just an uncond branch.
1685       TII->insertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1686     }
1687   }
1688 
1689   // If the prior block doesn't fall through into this block, and if this
1690   // block doesn't fall through into some other block, see if we can find a
1691   // place to move this block where a fall-through will happen.
1692   if (!PrevBB.canFallThrough()) {
1693     // Now we know that there was no fall-through into this block, check to
1694     // see if it has a fall-through into its successor.
1695     bool CurFallsThru = MBB->canFallThrough();
1696 
1697     if (!MBB->isEHPad()) {
1698       // Check all the predecessors of this block.  If one of them has no fall
1699       // throughs, and analyzeBranch thinks it _could_ fallthrough to this
1700       // block, move this block right after it.
1701       for (MachineBasicBlock *PredBB : MBB->predecessors()) {
1702         // Analyze the branch at the end of the pred.
1703         MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1704         SmallVector<MachineOperand, 4> PredCond;
1705         if (PredBB != MBB && !PredBB->canFallThrough() &&
1706             !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) &&
1707             (PredTBB == MBB || PredFBB == MBB) &&
1708             (!CurFallsThru || !CurTBB || !CurFBB) &&
1709             (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1710           // If the current block doesn't fall through, just move it.
1711           // If the current block can fall through and does not end with a
1712           // conditional branch, we need to append an unconditional jump to
1713           // the (current) next block.  To avoid a possible compile-time
1714           // infinite loop, move blocks only backward in this case.
1715           // Also, if there are already 2 branches here, we cannot add a third;
1716           // this means we have the case
1717           // Bcc next
1718           // B elsewhere
1719           // next:
1720           if (CurFallsThru) {
1721             MachineBasicBlock *NextBB = &*std::next(MBB->getIterator());
1722             CurCond.clear();
1723             TII->insertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1724           }
1725           MBB->moveAfter(PredBB);
1726           MadeChange = true;
1727           goto ReoptimizeBlock;
1728         }
1729       }
1730     }
1731 
1732     if (!CurFallsThru) {
1733       // Check analyzable branch-successors to see if we can move this block
1734       // before one.
1735       if (!CurUnAnalyzable) {
1736         for (MachineBasicBlock *SuccBB : {CurFBB, CurTBB}) {
1737           if (!SuccBB)
1738             continue;
1739           // Analyze the branch at the end of the block before the succ.
1740           MachineFunction::iterator SuccPrev = --SuccBB->getIterator();
1741 
1742           // If this block doesn't already fall-through to that successor, and
1743           // if the succ doesn't already have a block that can fall through into
1744           // it, we can arrange for the fallthrough to happen.
1745           if (SuccBB != MBB && &*SuccPrev != MBB &&
1746               !SuccPrev->canFallThrough()) {
1747             MBB->moveBefore(SuccBB);
1748             MadeChange = true;
1749             goto ReoptimizeBlock;
1750           }
1751         }
1752       }
1753 
1754       // Okay, there is no really great place to put this block.  If, however,
1755       // the block before this one would be a fall-through if this block were
1756       // removed, move this block to the end of the function. There is no real
1757       // advantage in "falling through" to an EH block, so we don't want to
1758       // perform this transformation for that case.
1759       //
1760       // Also, Windows EH introduced the possibility of an arbitrary number of
1761       // successors to a given block.  The analyzeBranch call does not consider
1762       // exception handling and so we can get in a state where a block
1763       // containing a call is followed by multiple EH blocks that would be
1764       // rotated infinitely at the end of the function if the transformation
1765       // below were performed for EH "FallThrough" blocks.  Therefore, even if
1766       // that appears not to be happening anymore, we should assume that it is
1767       // possible and not remove the "!FallThrough()->isEHPad" condition below.
1768       MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1769       SmallVector<MachineOperand, 4> PrevCond;
1770       if (FallThrough != MF.end() &&
1771           !FallThrough->isEHPad() &&
1772           !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1773           PrevBB.isSuccessor(&*FallThrough)) {
1774         MBB->moveAfter(&MF.back());
1775         MadeChange = true;
1776         return MadeChange;
1777       }
1778     }
1779   }
1780 
1781   return MadeChange;
1782 }
1783 
1784 //===----------------------------------------------------------------------===//
1785 //  Hoist Common Code
1786 //===----------------------------------------------------------------------===//
1787 
HoistCommonCode(MachineFunction & MF)1788 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1789   bool MadeChange = false;
1790   for (MachineBasicBlock &MBB : llvm::make_early_inc_range(MF))
1791     MadeChange |= HoistCommonCodeInSuccs(&MBB);
1792 
1793   return MadeChange;
1794 }
1795 
1796 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1797 /// its 'true' successor.
findFalseBlock(MachineBasicBlock * BB,MachineBasicBlock * TrueBB)1798 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1799                                          MachineBasicBlock *TrueBB) {
1800   for (MachineBasicBlock *SuccBB : BB->successors())
1801     if (SuccBB != TrueBB)
1802       return SuccBB;
1803   return nullptr;
1804 }
1805 
1806 template <class Container>
addRegAndItsAliases(Register Reg,const TargetRegisterInfo * TRI,Container & Set)1807 static void addRegAndItsAliases(Register Reg, const TargetRegisterInfo *TRI,
1808                                 Container &Set) {
1809   if (Reg.isPhysical()) {
1810     for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1811       Set.insert(*AI);
1812   } else {
1813     Set.insert(Reg);
1814   }
1815 }
1816 
1817 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1818 /// in successors to. The location is usually just before the terminator,
1819 /// however if the terminator is a conditional branch and its previous
1820 /// instruction is the flag setting instruction, the previous instruction is
1821 /// the preferred location. This function also gathers uses and defs of the
1822 /// instructions from the insertion point to the end of the block. The data is
1823 /// used by HoistCommonCodeInSuccs to ensure safety.
1824 static
findHoistingInsertPosAndDeps(MachineBasicBlock * MBB,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI,SmallSet<Register,4> & Uses,SmallSet<Register,4> & Defs)1825 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1826                                                   const TargetInstrInfo *TII,
1827                                                   const TargetRegisterInfo *TRI,
1828                                                   SmallSet<Register, 4> &Uses,
1829                                                   SmallSet<Register, 4> &Defs) {
1830   MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1831   if (!TII->isUnpredicatedTerminator(*Loc))
1832     return MBB->end();
1833 
1834   for (const MachineOperand &MO : Loc->operands()) {
1835     if (!MO.isReg())
1836       continue;
1837     Register Reg = MO.getReg();
1838     if (!Reg)
1839       continue;
1840     if (MO.isUse()) {
1841       addRegAndItsAliases(Reg, TRI, Uses);
1842     } else {
1843       if (!MO.isDead())
1844         // Don't try to hoist code in the rare case the terminator defines a
1845         // register that is later used.
1846         return MBB->end();
1847 
1848       // If the terminator defines a register, make sure we don't hoist
1849       // the instruction whose def might be clobbered by the terminator.
1850       addRegAndItsAliases(Reg, TRI, Defs);
1851     }
1852   }
1853 
1854   if (Uses.empty())
1855     return Loc;
1856   // If the terminator is the only instruction in the block and Uses is not
1857   // empty (or we would have returned above), we can still safely hoist
1858   // instructions just before the terminator as long as the Defs/Uses are not
1859   // violated (which is checked in HoistCommonCodeInSuccs).
1860   if (Loc == MBB->begin())
1861     return Loc;
1862 
1863   // The terminator is probably a conditional branch, try not to separate the
1864   // branch from condition setting instruction.
1865   MachineBasicBlock::iterator PI = prev_nodbg(Loc, MBB->begin());
1866 
1867   bool IsDef = false;
1868   for (const MachineOperand &MO : PI->operands()) {
1869     // If PI has a regmask operand, it is probably a call. Separate away.
1870     if (MO.isRegMask())
1871       return Loc;
1872     if (!MO.isReg() || MO.isUse())
1873       continue;
1874     Register Reg = MO.getReg();
1875     if (!Reg)
1876       continue;
1877     if (Uses.count(Reg)) {
1878       IsDef = true;
1879       break;
1880     }
1881   }
1882   if (!IsDef)
1883     // The condition setting instruction is not just before the conditional
1884     // branch.
1885     return Loc;
1886 
1887   // Be conservative, don't insert instruction above something that may have
1888   // side-effects. And since it's potentially bad to separate flag setting
1889   // instruction from the conditional branch, just abort the optimization
1890   // completely.
1891   // Also avoid moving code above predicated instruction since it's hard to
1892   // reason about register liveness with predicated instruction.
1893   bool DontMoveAcrossStore = true;
1894   if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI))
1895     return MBB->end();
1896 
1897   // Find out what registers are live. Note this routine is ignoring other live
1898   // registers which are only used by instructions in successor blocks.
1899   for (const MachineOperand &MO : PI->operands()) {
1900     if (!MO.isReg())
1901       continue;
1902     Register Reg = MO.getReg();
1903     if (!Reg)
1904       continue;
1905     if (MO.isUse()) {
1906       addRegAndItsAliases(Reg, TRI, Uses);
1907     } else {
1908       if (Uses.erase(Reg)) {
1909         if (Reg.isPhysical()) {
1910           for (MCPhysReg SubReg : TRI->subregs(Reg))
1911             Uses.erase(SubReg); // Use sub-registers to be conservative
1912         }
1913       }
1914       addRegAndItsAliases(Reg, TRI, Defs);
1915     }
1916   }
1917 
1918   return PI;
1919 }
1920 
HoistCommonCodeInSuccs(MachineBasicBlock * MBB)1921 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1922   MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1923   SmallVector<MachineOperand, 4> Cond;
1924   if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1925     return false;
1926 
1927   if (!FBB) FBB = findFalseBlock(MBB, TBB);
1928   if (!FBB)
1929     // Malformed bcc? True and false blocks are the same?
1930     return false;
1931 
1932   // Restrict the optimization to cases where MBB is the only predecessor,
1933   // it is an obvious win.
1934   if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1935     return false;
1936 
1937   // Find a suitable position to hoist the common instructions to. Also figure
1938   // out which registers are used or defined by instructions from the insertion
1939   // point to the end of the block.
1940   SmallSet<Register, 4> Uses, Defs;
1941   MachineBasicBlock::iterator Loc =
1942     findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1943   if (Loc == MBB->end())
1944     return false;
1945 
1946   bool HasDups = false;
1947   SmallSet<Register, 4> ActiveDefsSet, AllDefsSet;
1948   MachineBasicBlock::iterator TIB = TBB->begin();
1949   MachineBasicBlock::iterator FIB = FBB->begin();
1950   MachineBasicBlock::iterator TIE = TBB->end();
1951   MachineBasicBlock::iterator FIE = FBB->end();
1952   while (TIB != TIE && FIB != FIE) {
1953     // Skip dbg_value instructions. These do not count.
1954     TIB = skipDebugInstructionsForward(TIB, TIE, false);
1955     FIB = skipDebugInstructionsForward(FIB, FIE, false);
1956     if (TIB == TIE || FIB == FIE)
1957       break;
1958 
1959     if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead))
1960       break;
1961 
1962     if (TII->isPredicated(*TIB))
1963       // Hard to reason about register liveness with predicated instruction.
1964       break;
1965 
1966     bool IsSafe = true;
1967     for (MachineOperand &MO : TIB->operands()) {
1968       // Don't attempt to hoist instructions with register masks.
1969       if (MO.isRegMask()) {
1970         IsSafe = false;
1971         break;
1972       }
1973       if (!MO.isReg())
1974         continue;
1975       Register Reg = MO.getReg();
1976       if (!Reg)
1977         continue;
1978       if (MO.isDef()) {
1979         if (Uses.count(Reg)) {
1980           // Avoid clobbering a register that's used by the instruction at
1981           // the point of insertion.
1982           IsSafe = false;
1983           break;
1984         }
1985 
1986         if (Defs.count(Reg) && !MO.isDead()) {
1987           // Don't hoist the instruction if the def would be clobber by the
1988           // instruction at the point insertion. FIXME: This is overly
1989           // conservative. It should be possible to hoist the instructions
1990           // in BB2 in the following example:
1991           // BB1:
1992           // r1, eflag = op1 r2, r3
1993           // brcc eflag
1994           //
1995           // BB2:
1996           // r1 = op2, ...
1997           //    = op3, killed r1
1998           IsSafe = false;
1999           break;
2000         }
2001       } else if (!ActiveDefsSet.count(Reg)) {
2002         if (Defs.count(Reg)) {
2003           // Use is defined by the instruction at the point of insertion.
2004           IsSafe = false;
2005           break;
2006         }
2007 
2008         if (MO.isKill() && Uses.count(Reg))
2009           // Kills a register that's read by the instruction at the point of
2010           // insertion. Remove the kill marker.
2011           MO.setIsKill(false);
2012       }
2013     }
2014     if (!IsSafe)
2015       break;
2016 
2017     bool DontMoveAcrossStore = true;
2018     if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
2019       break;
2020 
2021     // Remove kills from ActiveDefsSet, these registers had short live ranges.
2022     for (const MachineOperand &MO : TIB->all_uses()) {
2023       if (!MO.isKill())
2024         continue;
2025       Register Reg = MO.getReg();
2026       if (!Reg)
2027         continue;
2028       if (!AllDefsSet.count(Reg)) {
2029         continue;
2030       }
2031       if (Reg.isPhysical()) {
2032         for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
2033           ActiveDefsSet.erase(*AI);
2034       } else {
2035         ActiveDefsSet.erase(Reg);
2036       }
2037     }
2038 
2039     // Track local defs so we can update liveins.
2040     for (const MachineOperand &MO : TIB->all_defs()) {
2041       if (MO.isDead())
2042         continue;
2043       Register Reg = MO.getReg();
2044       if (!Reg || Reg.isVirtual())
2045         continue;
2046       addRegAndItsAliases(Reg, TRI, ActiveDefsSet);
2047       addRegAndItsAliases(Reg, TRI, AllDefsSet);
2048     }
2049 
2050     HasDups = true;
2051     ++TIB;
2052     ++FIB;
2053   }
2054 
2055   if (!HasDups)
2056     return false;
2057 
2058   MBB->splice(Loc, TBB, TBB->begin(), TIB);
2059   FBB->erase(FBB->begin(), FIB);
2060 
2061   if (UpdateLiveIns)
2062     fullyRecomputeLiveIns({TBB, FBB});
2063 
2064   ++NumHoist;
2065   return true;
2066 }
2067