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