1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass forwards branches to unconditional branches to make them branch
11 // directly to the target block. This pass often results in dead MBB's, which
12 // it then removes.
13 //
14 // Note that this pass must be run after register allocation, it cannot handle
15 // SSA form.
16 //
17 //===----------------------------------------------------------------------===//
18
19 #include "BranchFolding.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
24 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
25 #include "llvm/CodeGen/MachineFunctionPass.h"
26 #include "llvm/CodeGen/MachineJumpTableInfo.h"
27 #include "llvm/CodeGen/MachineModuleInfo.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/Passes.h"
30 #include "llvm/CodeGen/RegisterScavenging.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetInstrInfo.h"
37 #include "llvm/Target/TargetRegisterInfo.h"
38 #include "llvm/Target/TargetSubtargetInfo.h"
39 #include <algorithm>
40 using namespace llvm;
41
42 #define DEBUG_TYPE "branchfolding"
43
44 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
45 STATISTIC(NumBranchOpts, "Number of branches optimized");
46 STATISTIC(NumTailMerge , "Number of block tails merged");
47 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
48
49 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
50 cl::init(cl::BOU_UNSET), cl::Hidden);
51
52 // Throttle for huge numbers of predecessors (compile speed problems)
53 static cl::opt<unsigned>
54 TailMergeThreshold("tail-merge-threshold",
55 cl::desc("Max number of predecessors to consider tail merging"),
56 cl::init(150), cl::Hidden);
57
58 // Heuristic for tail merging (and, inversely, tail duplication).
59 // TODO: This should be replaced with a target query.
60 static cl::opt<unsigned>
61 TailMergeSize("tail-merge-size",
62 cl::desc("Min number of instructions to consider tail merging"),
63 cl::init(3), cl::Hidden);
64
65 namespace {
66 /// BranchFolderPass - Wrap branch folder in a machine function pass.
67 class BranchFolderPass : public MachineFunctionPass {
68 public:
69 static char ID;
BranchFolderPass()70 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
71
72 bool runOnMachineFunction(MachineFunction &MF) override;
73
getAnalysisUsage(AnalysisUsage & AU) const74 void getAnalysisUsage(AnalysisUsage &AU) const override {
75 AU.addRequired<MachineBlockFrequencyInfo>();
76 AU.addRequired<MachineBranchProbabilityInfo>();
77 AU.addRequired<TargetPassConfig>();
78 MachineFunctionPass::getAnalysisUsage(AU);
79 }
80 };
81 }
82
83 char BranchFolderPass::ID = 0;
84 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
85
86 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
87 "Control Flow Optimizer", false, false)
88
runOnMachineFunction(MachineFunction & MF)89 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
90 if (skipOptnoneFunction(*MF.getFunction()))
91 return false;
92
93 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
94 // TailMerge can create jump into if branches that make CFG irreducible for
95 // HW that requires structurized CFG.
96 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
97 PassConfig->getEnableTailMerge();
98 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true,
99 getAnalysis<MachineBlockFrequencyInfo>(),
100 getAnalysis<MachineBranchProbabilityInfo>());
101 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
102 MF.getSubtarget().getRegisterInfo(),
103 getAnalysisIfAvailable<MachineModuleInfo>());
104 }
105
BranchFolder(bool defaultEnableTailMerge,bool CommonHoist,const MachineBlockFrequencyInfo & FreqInfo,const MachineBranchProbabilityInfo & ProbInfo)106 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
107 const MachineBlockFrequencyInfo &FreqInfo,
108 const MachineBranchProbabilityInfo &ProbInfo)
109 : EnableHoistCommonCode(CommonHoist), MBBFreqInfo(FreqInfo),
110 MBPI(ProbInfo) {
111 switch (FlagEnableTailMerge) {
112 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
113 case cl::BOU_TRUE: EnableTailMerge = true; break;
114 case cl::BOU_FALSE: EnableTailMerge = false; break;
115 }
116 }
117
118 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
119 /// function, updating the CFG.
RemoveDeadBlock(MachineBasicBlock * MBB)120 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
121 assert(MBB->pred_empty() && "MBB must be dead!");
122 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
123
124 MachineFunction *MF = MBB->getParent();
125 // drop all successors.
126 while (!MBB->succ_empty())
127 MBB->removeSuccessor(MBB->succ_end()-1);
128
129 // Avoid matching if this pointer gets reused.
130 TriedMerging.erase(MBB);
131
132 // Remove the block.
133 MF->erase(MBB);
134 }
135
136 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
137 /// followed by terminators, and if the implicitly defined registers are not
138 /// used by the terminators, remove those implicit_def's. e.g.
139 /// BB1:
140 /// r0 = implicit_def
141 /// r1 = implicit_def
142 /// br
143 /// This block can be optimized away later if the implicit instructions are
144 /// removed.
OptimizeImpDefsBlock(MachineBasicBlock * MBB)145 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
146 SmallSet<unsigned, 4> ImpDefRegs;
147 MachineBasicBlock::iterator I = MBB->begin();
148 while (I != MBB->end()) {
149 if (!I->isImplicitDef())
150 break;
151 unsigned Reg = I->getOperand(0).getReg();
152 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
153 SubRegs.isValid(); ++SubRegs)
154 ImpDefRegs.insert(*SubRegs);
155 ++I;
156 }
157 if (ImpDefRegs.empty())
158 return false;
159
160 MachineBasicBlock::iterator FirstTerm = I;
161 while (I != MBB->end()) {
162 if (!TII->isUnpredicatedTerminator(I))
163 return false;
164 // See if it uses any of the implicitly defined registers.
165 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
166 MachineOperand &MO = I->getOperand(i);
167 if (!MO.isReg() || !MO.isUse())
168 continue;
169 unsigned Reg = MO.getReg();
170 if (ImpDefRegs.count(Reg))
171 return false;
172 }
173 ++I;
174 }
175
176 I = MBB->begin();
177 while (I != FirstTerm) {
178 MachineInstr *ImpDefMI = &*I;
179 ++I;
180 MBB->erase(ImpDefMI);
181 }
182
183 return true;
184 }
185
186 /// OptimizeFunction - Perhaps branch folding, tail merging and other
187 /// CFG optimizations on the given function.
OptimizeFunction(MachineFunction & MF,const TargetInstrInfo * tii,const TargetRegisterInfo * tri,MachineModuleInfo * mmi)188 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
189 const TargetInstrInfo *tii,
190 const TargetRegisterInfo *tri,
191 MachineModuleInfo *mmi) {
192 if (!tii) return false;
193
194 TriedMerging.clear();
195
196 TII = tii;
197 TRI = tri;
198 MMI = mmi;
199 RS = nullptr;
200
201 // Use a RegScavenger to help update liveness when required.
202 MachineRegisterInfo &MRI = MF.getRegInfo();
203 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
204 RS = new RegScavenger();
205 else
206 MRI.invalidateLiveness();
207
208 // Fix CFG. The later algorithms expect it to be right.
209 bool MadeChange = false;
210 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
211 MachineBasicBlock *MBB = I, *TBB = nullptr, *FBB = nullptr;
212 SmallVector<MachineOperand, 4> Cond;
213 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
214 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
215 MadeChange |= OptimizeImpDefsBlock(MBB);
216 }
217
218 bool MadeChangeThisIteration = true;
219 while (MadeChangeThisIteration) {
220 MadeChangeThisIteration = TailMergeBlocks(MF);
221 MadeChangeThisIteration |= OptimizeBranches(MF);
222 if (EnableHoistCommonCode)
223 MadeChangeThisIteration |= HoistCommonCode(MF);
224 MadeChange |= MadeChangeThisIteration;
225 }
226
227 // See if any jump tables have become dead as the code generator
228 // did its thing.
229 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
230 if (!JTI) {
231 delete RS;
232 return MadeChange;
233 }
234
235 // Walk the function to find jump tables that are live.
236 BitVector JTIsLive(JTI->getJumpTables().size());
237 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
238 BB != E; ++BB) {
239 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
240 I != E; ++I)
241 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
242 MachineOperand &Op = I->getOperand(op);
243 if (!Op.isJTI()) continue;
244
245 // Remember that this JT is live.
246 JTIsLive.set(Op.getIndex());
247 }
248 }
249
250 // Finally, remove dead jump tables. This happens when the
251 // indirect jump was unreachable (and thus deleted).
252 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
253 if (!JTIsLive.test(i)) {
254 JTI->RemoveJumpTable(i);
255 MadeChange = true;
256 }
257
258 delete RS;
259 return MadeChange;
260 }
261
262 //===----------------------------------------------------------------------===//
263 // Tail Merging of Blocks
264 //===----------------------------------------------------------------------===//
265
266 /// HashMachineInstr - Compute a hash value for MI and its operands.
HashMachineInstr(const MachineInstr * MI)267 static unsigned HashMachineInstr(const MachineInstr *MI) {
268 unsigned Hash = MI->getOpcode();
269 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
270 const MachineOperand &Op = MI->getOperand(i);
271
272 // Merge in bits from the operand if easy.
273 unsigned OperandHash = 0;
274 switch (Op.getType()) {
275 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
276 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
277 case MachineOperand::MO_MachineBasicBlock:
278 OperandHash = Op.getMBB()->getNumber();
279 break;
280 case MachineOperand::MO_FrameIndex:
281 case MachineOperand::MO_ConstantPoolIndex:
282 case MachineOperand::MO_JumpTableIndex:
283 OperandHash = Op.getIndex();
284 break;
285 case MachineOperand::MO_GlobalAddress:
286 case MachineOperand::MO_ExternalSymbol:
287 // Global address / external symbol are too hard, don't bother, but do
288 // pull in the offset.
289 OperandHash = Op.getOffset();
290 break;
291 default: break;
292 }
293
294 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
295 }
296 return Hash;
297 }
298
299 /// HashEndOfMBB - Hash the last instruction in the MBB.
HashEndOfMBB(const MachineBasicBlock * MBB)300 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
301 MachineBasicBlock::const_iterator I = MBB->end();
302 if (I == MBB->begin())
303 return 0; // Empty MBB.
304
305 --I;
306 // Skip debug info so it will not affect codegen.
307 while (I->isDebugValue()) {
308 if (I==MBB->begin())
309 return 0; // MBB empty except for debug info.
310 --I;
311 }
312
313 return HashMachineInstr(I);
314 }
315
316 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
317 /// of instructions they actually have in common together at their end. Return
318 /// iterators for the first shared instruction in each block.
ComputeCommonTailLength(MachineBasicBlock * MBB1,MachineBasicBlock * MBB2,MachineBasicBlock::iterator & I1,MachineBasicBlock::iterator & I2)319 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
320 MachineBasicBlock *MBB2,
321 MachineBasicBlock::iterator &I1,
322 MachineBasicBlock::iterator &I2) {
323 I1 = MBB1->end();
324 I2 = MBB2->end();
325
326 unsigned TailLen = 0;
327 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
328 --I1; --I2;
329 // Skip debugging pseudos; necessary to avoid changing the code.
330 while (I1->isDebugValue()) {
331 if (I1==MBB1->begin()) {
332 while (I2->isDebugValue()) {
333 if (I2==MBB2->begin())
334 // I1==DBG at begin; I2==DBG at begin
335 return TailLen;
336 --I2;
337 }
338 ++I2;
339 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
340 return TailLen;
341 }
342 --I1;
343 }
344 // I1==first (untested) non-DBG preceding known match
345 while (I2->isDebugValue()) {
346 if (I2==MBB2->begin()) {
347 ++I1;
348 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
349 return TailLen;
350 }
351 --I2;
352 }
353 // I1, I2==first (untested) non-DBGs preceding known match
354 if (!I1->isIdenticalTo(I2) ||
355 // FIXME: This check is dubious. It's used to get around a problem where
356 // people incorrectly expect inline asm directives to remain in the same
357 // relative order. This is untenable because normal compiler
358 // optimizations (like this one) may reorder and/or merge these
359 // directives.
360 I1->isInlineAsm()) {
361 ++I1; ++I2;
362 break;
363 }
364 ++TailLen;
365 }
366 // Back past possible debugging pseudos at beginning of block. This matters
367 // when one block differs from the other only by whether debugging pseudos
368 // are present at the beginning. (This way, the various checks later for
369 // I1==MBB1->begin() work as expected.)
370 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
371 --I2;
372 while (I2->isDebugValue()) {
373 if (I2 == MBB2->begin())
374 return TailLen;
375 --I2;
376 }
377 ++I2;
378 }
379 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
380 --I1;
381 while (I1->isDebugValue()) {
382 if (I1 == MBB1->begin())
383 return TailLen;
384 --I1;
385 }
386 ++I1;
387 }
388 return TailLen;
389 }
390
MaintainLiveIns(MachineBasicBlock * CurMBB,MachineBasicBlock * NewMBB)391 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
392 MachineBasicBlock *NewMBB) {
393 if (RS) {
394 RS->enterBasicBlock(CurMBB);
395 if (!CurMBB->empty())
396 RS->forward(std::prev(CurMBB->end()));
397 for (unsigned int i = 1, e = TRI->getNumRegs(); i != e; i++)
398 if (RS->isRegUsed(i, false))
399 NewMBB->addLiveIn(i);
400 }
401 }
402
403 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
404 /// after it, replacing it with an unconditional branch to NewDest.
ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,MachineBasicBlock * NewDest)405 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
406 MachineBasicBlock *NewDest) {
407 MachineBasicBlock *CurMBB = OldInst->getParent();
408
409 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
410
411 // For targets that use the register scavenger, we must maintain LiveIns.
412 MaintainLiveIns(CurMBB, NewDest);
413
414 ++NumTailMerge;
415 }
416
417 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
418 /// MBB so that the part before the iterator falls into the part starting at the
419 /// iterator. This returns the new MBB.
SplitMBBAt(MachineBasicBlock & CurMBB,MachineBasicBlock::iterator BBI1,const BasicBlock * BB)420 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
421 MachineBasicBlock::iterator BBI1,
422 const BasicBlock *BB) {
423 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
424 return nullptr;
425
426 MachineFunction &MF = *CurMBB.getParent();
427
428 // Create the fall-through block.
429 MachineFunction::iterator MBBI = &CurMBB;
430 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
431 CurMBB.getParent()->insert(++MBBI, NewMBB);
432
433 // Move all the successors of this block to the specified block.
434 NewMBB->transferSuccessors(&CurMBB);
435
436 // Add an edge from CurMBB to NewMBB for the fall-through.
437 CurMBB.addSuccessor(NewMBB);
438
439 // Splice the code over.
440 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
441
442 // NewMBB inherits CurMBB's block frequency.
443 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
444
445 // For targets that use the register scavenger, we must maintain LiveIns.
446 MaintainLiveIns(&CurMBB, NewMBB);
447
448 return NewMBB;
449 }
450
451 /// EstimateRuntime - Make a rough estimate for how long it will take to run
452 /// the specified code.
EstimateRuntime(MachineBasicBlock::iterator I,MachineBasicBlock::iterator E)453 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
454 MachineBasicBlock::iterator E) {
455 unsigned Time = 0;
456 for (; I != E; ++I) {
457 if (I->isDebugValue())
458 continue;
459 if (I->isCall())
460 Time += 10;
461 else if (I->mayLoad() || I->mayStore())
462 Time += 2;
463 else
464 ++Time;
465 }
466 return Time;
467 }
468
469 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
470 // branches temporarily for tail merging). In the case where CurMBB ends
471 // with a conditional branch to the next block, optimize by reversing the
472 // test and conditionally branching to SuccMBB instead.
FixTail(MachineBasicBlock * CurMBB,MachineBasicBlock * SuccBB,const TargetInstrInfo * TII)473 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
474 const TargetInstrInfo *TII) {
475 MachineFunction *MF = CurMBB->getParent();
476 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
477 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
478 SmallVector<MachineOperand, 4> Cond;
479 DebugLoc dl; // FIXME: this is nowhere
480 if (I != MF->end() &&
481 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
482 MachineBasicBlock *NextBB = I;
483 if (TBB == NextBB && !Cond.empty() && !FBB) {
484 if (!TII->ReverseBranchCondition(Cond)) {
485 TII->RemoveBranch(*CurMBB);
486 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
487 return;
488 }
489 }
490 }
491 TII->InsertBranch(*CurMBB, SuccBB, nullptr,
492 SmallVector<MachineOperand, 0>(), dl);
493 }
494
495 bool
operator <(const MergePotentialsElt & o) const496 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
497 if (getHash() < o.getHash())
498 return true;
499 if (getHash() > o.getHash())
500 return false;
501 if (getBlock()->getNumber() < o.getBlock()->getNumber())
502 return true;
503 if (getBlock()->getNumber() > o.getBlock()->getNumber())
504 return false;
505 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
506 // an object with itself.
507 #ifndef _GLIBCXX_DEBUG
508 llvm_unreachable("Predecessor appears twice");
509 #else
510 return false;
511 #endif
512 }
513
514 BlockFrequency
getBlockFreq(const MachineBasicBlock * MBB) const515 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
516 auto I = MergedBBFreq.find(MBB);
517
518 if (I != MergedBBFreq.end())
519 return I->second;
520
521 return MBFI.getBlockFreq(MBB);
522 }
523
setBlockFreq(const MachineBasicBlock * MBB,BlockFrequency F)524 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
525 BlockFrequency F) {
526 MergedBBFreq[MBB] = F;
527 }
528
529 /// CountTerminators - Count the number of terminators in the given
530 /// block and set I to the position of the first non-terminator, if there
531 /// is one, or MBB->end() otherwise.
CountTerminators(MachineBasicBlock * MBB,MachineBasicBlock::iterator & I)532 static unsigned CountTerminators(MachineBasicBlock *MBB,
533 MachineBasicBlock::iterator &I) {
534 I = MBB->end();
535 unsigned NumTerms = 0;
536 for (;;) {
537 if (I == MBB->begin()) {
538 I = MBB->end();
539 break;
540 }
541 --I;
542 if (!I->isTerminator()) break;
543 ++NumTerms;
544 }
545 return NumTerms;
546 }
547
548 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
549 /// and decide if it would be profitable to merge those tails. Return the
550 /// length of the common tail and iterators to the first common instruction
551 /// in each block.
ProfitableToMerge(MachineBasicBlock * MBB1,MachineBasicBlock * MBB2,unsigned minCommonTailLength,unsigned & CommonTailLen,MachineBasicBlock::iterator & I1,MachineBasicBlock::iterator & I2,MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB)552 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
553 MachineBasicBlock *MBB2,
554 unsigned minCommonTailLength,
555 unsigned &CommonTailLen,
556 MachineBasicBlock::iterator &I1,
557 MachineBasicBlock::iterator &I2,
558 MachineBasicBlock *SuccBB,
559 MachineBasicBlock *PredBB) {
560 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
561 if (CommonTailLen == 0)
562 return false;
563 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
564 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
565 << '\n');
566
567 // It's almost always profitable to merge any number of non-terminator
568 // instructions with the block that falls through into the common successor.
569 if (MBB1 == PredBB || MBB2 == PredBB) {
570 MachineBasicBlock::iterator I;
571 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
572 if (CommonTailLen > NumTerms)
573 return true;
574 }
575
576 // If one of the blocks can be completely merged and happens to be in
577 // a position where the other could fall through into it, merge any number
578 // of instructions, because it can be done without a branch.
579 // TODO: If the blocks are not adjacent, move one of them so that they are?
580 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
581 return true;
582 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
583 return true;
584
585 // If both blocks have an unconditional branch temporarily stripped out,
586 // count that as an additional common instruction for the following
587 // heuristics.
588 unsigned EffectiveTailLen = CommonTailLen;
589 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
590 !MBB1->back().isBarrier() &&
591 !MBB2->back().isBarrier())
592 ++EffectiveTailLen;
593
594 // Check if the common tail is long enough to be worthwhile.
595 if (EffectiveTailLen >= minCommonTailLength)
596 return true;
597
598 // If we are optimizing for code size, 2 instructions in common is enough if
599 // we don't have to split a block. At worst we will be introducing 1 new
600 // branch instruction, which is likely to be smaller than the 2
601 // instructions that would be deleted in the merge.
602 MachineFunction *MF = MBB1->getParent();
603 if (EffectiveTailLen >= 2 &&
604 MF->getFunction()->getAttributes().
605 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
606 (I1 == MBB1->begin() || I2 == MBB2->begin()))
607 return true;
608
609 return false;
610 }
611
612 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
613 /// hash CurHash (guaranteed to match the last element). Build the vector
614 /// SameTails of all those that have the (same) largest number of instructions
615 /// in common of any pair of these blocks. SameTails entries contain an
616 /// iterator into MergePotentials (from which the MachineBasicBlock can be
617 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
618 /// instruction where the matching code sequence begins.
619 /// Order of elements in SameTails is the reverse of the order in which
620 /// those blocks appear in MergePotentials (where they are not necessarily
621 /// consecutive).
ComputeSameTails(unsigned CurHash,unsigned minCommonTailLength,MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB)622 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
623 unsigned minCommonTailLength,
624 MachineBasicBlock *SuccBB,
625 MachineBasicBlock *PredBB) {
626 unsigned maxCommonTailLength = 0U;
627 SameTails.clear();
628 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
629 MPIterator HighestMPIter = std::prev(MergePotentials.end());
630 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
631 B = MergePotentials.begin();
632 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
633 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
634 unsigned CommonTailLen;
635 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
636 minCommonTailLength,
637 CommonTailLen, TrialBBI1, TrialBBI2,
638 SuccBB, PredBB)) {
639 if (CommonTailLen > maxCommonTailLength) {
640 SameTails.clear();
641 maxCommonTailLength = CommonTailLen;
642 HighestMPIter = CurMPIter;
643 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
644 }
645 if (HighestMPIter == CurMPIter &&
646 CommonTailLen == maxCommonTailLength)
647 SameTails.push_back(SameTailElt(I, TrialBBI2));
648 }
649 if (I == B)
650 break;
651 }
652 }
653 return maxCommonTailLength;
654 }
655
656 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
657 /// MergePotentials, restoring branches at ends of blocks as appropriate.
RemoveBlocksWithHash(unsigned CurHash,MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB)658 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
659 MachineBasicBlock *SuccBB,
660 MachineBasicBlock *PredBB) {
661 MPIterator CurMPIter, B;
662 for (CurMPIter = std::prev(MergePotentials.end()),
663 B = MergePotentials.begin();
664 CurMPIter->getHash() == CurHash; --CurMPIter) {
665 // Put the unconditional branch back, if we need one.
666 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
667 if (SuccBB && CurMBB != PredBB)
668 FixTail(CurMBB, SuccBB, TII);
669 if (CurMPIter == B)
670 break;
671 }
672 if (CurMPIter->getHash() != CurHash)
673 CurMPIter++;
674 MergePotentials.erase(CurMPIter, MergePotentials.end());
675 }
676
677 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
678 /// only of the common tail. Create a block that does by splitting one.
CreateCommonTailOnlyBlock(MachineBasicBlock * & PredBB,MachineBasicBlock * SuccBB,unsigned maxCommonTailLength,unsigned & commonTailIndex)679 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
680 MachineBasicBlock *SuccBB,
681 unsigned maxCommonTailLength,
682 unsigned &commonTailIndex) {
683 commonTailIndex = 0;
684 unsigned TimeEstimate = ~0U;
685 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
686 // Use PredBB if possible; that doesn't require a new branch.
687 if (SameTails[i].getBlock() == PredBB) {
688 commonTailIndex = i;
689 break;
690 }
691 // Otherwise, make a (fairly bogus) choice based on estimate of
692 // how long it will take the various blocks to execute.
693 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
694 SameTails[i].getTailStartPos());
695 if (t <= TimeEstimate) {
696 TimeEstimate = t;
697 commonTailIndex = i;
698 }
699 }
700
701 MachineBasicBlock::iterator BBI =
702 SameTails[commonTailIndex].getTailStartPos();
703 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
704
705 // If the common tail includes any debug info we will take it pretty
706 // randomly from one of the inputs. Might be better to remove it?
707 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
708 << maxCommonTailLength);
709
710 // If the split block unconditionally falls-thru to SuccBB, it will be
711 // merged. In control flow terms it should then take SuccBB's name. e.g. If
712 // SuccBB is an inner loop, the common tail is still part of the inner loop.
713 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
714 SuccBB->getBasicBlock() : MBB->getBasicBlock();
715 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
716 if (!newMBB) {
717 DEBUG(dbgs() << "... failed!");
718 return false;
719 }
720
721 SameTails[commonTailIndex].setBlock(newMBB);
722 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
723
724 // If we split PredBB, newMBB is the new predecessor.
725 if (PredBB == MBB)
726 PredBB = newMBB;
727
728 return true;
729 }
730
731 // See if any of the blocks in MergePotentials (which all have a common single
732 // successor, or all have no successor) can be tail-merged. If there is a
733 // successor, any blocks in MergePotentials that are not tail-merged and
734 // are not immediately before Succ must have an unconditional branch to
735 // Succ added (but the predecessor/successor lists need no adjustment).
736 // The lone predecessor of Succ that falls through into Succ,
737 // if any, is given in PredBB.
738
TryTailMergeBlocks(MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB)739 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
740 MachineBasicBlock *PredBB) {
741 bool MadeChange = false;
742
743 // Except for the special cases below, tail-merge if there are at least
744 // this many instructions in common.
745 unsigned minCommonTailLength = TailMergeSize;
746
747 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
748 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
749 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
750 << (i == e-1 ? "" : ", ");
751 dbgs() << "\n";
752 if (SuccBB) {
753 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
754 if (PredBB)
755 dbgs() << " which has fall-through from BB#"
756 << PredBB->getNumber() << "\n";
757 }
758 dbgs() << "Looking for common tails of at least "
759 << minCommonTailLength << " instruction"
760 << (minCommonTailLength == 1 ? "" : "s") << '\n';
761 );
762
763 // Sort by hash value so that blocks with identical end sequences sort
764 // together.
765 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
766
767 // Walk through equivalence sets looking for actual exact matches.
768 while (MergePotentials.size() > 1) {
769 unsigned CurHash = MergePotentials.back().getHash();
770
771 // Build SameTails, identifying the set of blocks with this hash code
772 // and with the maximum number of instructions in common.
773 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
774 minCommonTailLength,
775 SuccBB, PredBB);
776
777 // If we didn't find any pair that has at least minCommonTailLength
778 // instructions in common, remove all blocks with this hash code and retry.
779 if (SameTails.empty()) {
780 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
781 continue;
782 }
783
784 // If one of the blocks is the entire common tail (and not the entry
785 // block, which we can't jump to), we can treat all blocks with this same
786 // tail at once. Use PredBB if that is one of the possibilities, as that
787 // will not introduce any extra branches.
788 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
789 getParent()->begin();
790 unsigned commonTailIndex = SameTails.size();
791 // If there are two blocks, check to see if one can be made to fall through
792 // into the other.
793 if (SameTails.size() == 2 &&
794 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
795 SameTails[1].tailIsWholeBlock())
796 commonTailIndex = 1;
797 else if (SameTails.size() == 2 &&
798 SameTails[1].getBlock()->isLayoutSuccessor(
799 SameTails[0].getBlock()) &&
800 SameTails[0].tailIsWholeBlock())
801 commonTailIndex = 0;
802 else {
803 // Otherwise just pick one, favoring the fall-through predecessor if
804 // there is one.
805 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
806 MachineBasicBlock *MBB = SameTails[i].getBlock();
807 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
808 continue;
809 if (MBB == PredBB) {
810 commonTailIndex = i;
811 break;
812 }
813 if (SameTails[i].tailIsWholeBlock())
814 commonTailIndex = i;
815 }
816 }
817
818 if (commonTailIndex == SameTails.size() ||
819 (SameTails[commonTailIndex].getBlock() == PredBB &&
820 !SameTails[commonTailIndex].tailIsWholeBlock())) {
821 // None of the blocks consist entirely of the common tail.
822 // Split a block so that one does.
823 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
824 maxCommonTailLength, commonTailIndex)) {
825 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
826 continue;
827 }
828 }
829
830 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
831
832 // Recompute commont tail MBB's edge weights and block frequency.
833 setCommonTailEdgeWeights(*MBB);
834
835 // MBB is common tail. Adjust all other BB's to jump to this one.
836 // Traversal must be forwards so erases work.
837 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
838 << " for ");
839 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
840 if (commonTailIndex == i)
841 continue;
842 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
843 << (i == e-1 ? "" : ", "));
844 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
845 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
846 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
847 MergePotentials.erase(SameTails[i].getMPIter());
848 }
849 DEBUG(dbgs() << "\n");
850 // We leave commonTailIndex in the worklist in case there are other blocks
851 // that match it with a smaller number of instructions.
852 MadeChange = true;
853 }
854 return MadeChange;
855 }
856
TailMergeBlocks(MachineFunction & MF)857 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
858 bool MadeChange = false;
859 if (!EnableTailMerge) return MadeChange;
860
861 // First find blocks with no successors.
862 MergePotentials.clear();
863 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
864 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
865 if (TriedMerging.count(I))
866 continue;
867 if (I->succ_empty())
868 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
869 }
870
871 // If this is a large problem, avoid visiting the same basic blocks
872 // multiple times.
873 if (MergePotentials.size() == TailMergeThreshold)
874 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
875 TriedMerging.insert(MergePotentials[i].getBlock());
876
877 // See if we can do any tail merging on those.
878 if (MergePotentials.size() >= 2)
879 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
880
881 // Look at blocks (IBB) with multiple predecessors (PBB).
882 // We change each predecessor to a canonical form, by
883 // (1) temporarily removing any unconditional branch from the predecessor
884 // to IBB, and
885 // (2) alter conditional branches so they branch to the other block
886 // not IBB; this may require adding back an unconditional branch to IBB
887 // later, where there wasn't one coming in. E.g.
888 // Bcc IBB
889 // fallthrough to QBB
890 // here becomes
891 // Bncc QBB
892 // with a conceptual B to IBB after that, which never actually exists.
893 // With those changes, we see whether the predecessors' tails match,
894 // and merge them if so. We change things out of canonical form and
895 // back to the way they were later in the process. (OptimizeBranches
896 // would undo some of this, but we can't use it, because we'd get into
897 // a compile-time infinite loop repeatedly doing and undoing the same
898 // transformations.)
899
900 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
901 I != E; ++I) {
902 if (I->pred_size() < 2) continue;
903 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
904 MachineBasicBlock *IBB = I;
905 MachineBasicBlock *PredBB = std::prev(I);
906 MergePotentials.clear();
907 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
908 E2 = I->pred_end();
909 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
910 MachineBasicBlock *PBB = *P;
911 if (TriedMerging.count(PBB))
912 continue;
913
914 // Skip blocks that loop to themselves, can't tail merge these.
915 if (PBB == IBB)
916 continue;
917
918 // Visit each predecessor only once.
919 if (!UniquePreds.insert(PBB).second)
920 continue;
921
922 // Skip blocks which may jump to a landing pad. Can't tail merge these.
923 if (PBB->getLandingPadSuccessor())
924 continue;
925
926 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
927 SmallVector<MachineOperand, 4> Cond;
928 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
929 // Failing case: IBB is the target of a cbr, and we cannot reverse the
930 // branch.
931 SmallVector<MachineOperand, 4> NewCond(Cond);
932 if (!Cond.empty() && TBB == IBB) {
933 if (TII->ReverseBranchCondition(NewCond))
934 continue;
935 // This is the QBB case described above
936 if (!FBB)
937 FBB = std::next(MachineFunction::iterator(PBB));
938 }
939
940 // Failing case: the only way IBB can be reached from PBB is via
941 // exception handling. Happens for landing pads. Would be nice to have
942 // a bit in the edge so we didn't have to do all this.
943 if (IBB->isLandingPad()) {
944 MachineFunction::iterator IP = PBB; IP++;
945 MachineBasicBlock *PredNextBB = nullptr;
946 if (IP != MF.end())
947 PredNextBB = IP;
948 if (!TBB) {
949 if (IBB != PredNextBB) // fallthrough
950 continue;
951 } else if (FBB) {
952 if (TBB != IBB && FBB != IBB) // cbr then ubr
953 continue;
954 } else if (Cond.empty()) {
955 if (TBB != IBB) // ubr
956 continue;
957 } else {
958 if (TBB != IBB && IBB != PredNextBB) // cbr
959 continue;
960 }
961 }
962
963 // Remove the unconditional branch at the end, if any.
964 if (TBB && (Cond.empty() || FBB)) {
965 DebugLoc dl; // FIXME: this is nowhere
966 TII->RemoveBranch(*PBB);
967 if (!Cond.empty())
968 // reinsert conditional branch only, for now
969 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
970 NewCond, dl);
971 }
972
973 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
974 }
975 }
976
977 // If this is a large problem, avoid visiting the same basic blocks multiple
978 // times.
979 if (MergePotentials.size() == TailMergeThreshold)
980 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
981 TriedMerging.insert(MergePotentials[i].getBlock());
982
983 if (MergePotentials.size() >= 2)
984 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
985
986 // Reinsert an unconditional branch if needed. The 1 below can occur as a
987 // result of removing blocks in TryTailMergeBlocks.
988 PredBB = std::prev(I); // this may have been changed in TryTailMergeBlocks
989 if (MergePotentials.size() == 1 &&
990 MergePotentials.begin()->getBlock() != PredBB)
991 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
992 }
993
994 return MadeChange;
995 }
996
setCommonTailEdgeWeights(MachineBasicBlock & TailMBB)997 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
998 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
999 BlockFrequency AccumulatedMBBFreq;
1000
1001 // Aggregate edge frequency of successor edge j:
1002 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1003 // where bb is a basic block that is in SameTails.
1004 for (const auto &Src : SameTails) {
1005 const MachineBasicBlock *SrcMBB = Src.getBlock();
1006 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1007 AccumulatedMBBFreq += BlockFreq;
1008
1009 // It is not necessary to recompute edge weights if TailBB has less than two
1010 // successors.
1011 if (TailMBB.succ_size() <= 1)
1012 continue;
1013
1014 auto EdgeFreq = EdgeFreqLs.begin();
1015
1016 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1017 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1018 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1019 }
1020
1021 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1022
1023 if (TailMBB.succ_size() <= 1)
1024 return;
1025
1026 auto MaxEdgeFreq = *std::max_element(EdgeFreqLs.begin(), EdgeFreqLs.end());
1027 uint64_t Scale = MaxEdgeFreq.getFrequency() / UINT32_MAX + 1;
1028 auto EdgeFreq = EdgeFreqLs.begin();
1029
1030 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1031 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1032 TailMBB.setSuccWeight(SuccI, EdgeFreq->getFrequency() / Scale);
1033 }
1034
1035 //===----------------------------------------------------------------------===//
1036 // Branch Optimization
1037 //===----------------------------------------------------------------------===//
1038
OptimizeBranches(MachineFunction & MF)1039 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1040 bool MadeChange = false;
1041
1042 // Make sure blocks are numbered in order
1043 MF.RenumberBlocks();
1044
1045 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1046 I != E; ) {
1047 MachineBasicBlock *MBB = I++;
1048 MadeChange |= OptimizeBlock(MBB);
1049
1050 // If it is dead, remove it.
1051 if (MBB->pred_empty()) {
1052 RemoveDeadBlock(MBB);
1053 MadeChange = true;
1054 ++NumDeadBlocks;
1055 }
1056 }
1057 return MadeChange;
1058 }
1059
1060 // Blocks should be considered empty if they contain only debug info;
1061 // else the debug info would affect codegen.
IsEmptyBlock(MachineBasicBlock * MBB)1062 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1063 if (MBB->empty())
1064 return true;
1065 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
1066 MBBI!=MBBE; ++MBBI) {
1067 if (!MBBI->isDebugValue())
1068 return false;
1069 }
1070 return true;
1071 }
1072
1073 // Blocks with only debug info and branches should be considered the same
1074 // as blocks with only branches.
IsBranchOnlyBlock(MachineBasicBlock * MBB)1075 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1076 MachineBasicBlock::iterator MBBI, MBBE;
1077 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
1078 if (!MBBI->isDebugValue())
1079 break;
1080 }
1081 return (MBBI->isBranch());
1082 }
1083
1084 /// IsBetterFallthrough - Return true if it would be clearly better to
1085 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1086 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1087 /// result in infinite loops.
IsBetterFallthrough(MachineBasicBlock * MBB1,MachineBasicBlock * MBB2)1088 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1089 MachineBasicBlock *MBB2) {
1090 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1091 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1092 // optimize branches that branch to either a return block or an assert block
1093 // into a fallthrough to the return.
1094 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
1095
1096 // If there is a clear successor ordering we make sure that one block
1097 // will fall through to the next
1098 if (MBB1->isSuccessor(MBB2)) return true;
1099 if (MBB2->isSuccessor(MBB1)) return false;
1100
1101 // Neither block consists entirely of debug info (per IsEmptyBlock check),
1102 // so we needn't test for falling off the beginning here.
1103 MachineBasicBlock::iterator MBB1I = --MBB1->end();
1104 while (MBB1I->isDebugValue())
1105 --MBB1I;
1106 MachineBasicBlock::iterator MBB2I = --MBB2->end();
1107 while (MBB2I->isDebugValue())
1108 --MBB2I;
1109 return MBB2I->isCall() && !MBB1I->isCall();
1110 }
1111
1112 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1113 /// instructions on the block. Always use the DebugLoc of the first
1114 /// branching instruction found unless its absent, in which case use the
1115 /// DebugLoc of the second if present.
getBranchDebugLoc(MachineBasicBlock & MBB)1116 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1117 MachineBasicBlock::iterator I = MBB.end();
1118 if (I == MBB.begin())
1119 return DebugLoc();
1120 --I;
1121 while (I->isDebugValue() && I != MBB.begin())
1122 --I;
1123 if (I->isBranch())
1124 return I->getDebugLoc();
1125 return DebugLoc();
1126 }
1127
1128 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1129 /// block. This is never called on the entry block.
OptimizeBlock(MachineBasicBlock * MBB)1130 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1131 bool MadeChange = false;
1132 MachineFunction &MF = *MBB->getParent();
1133 ReoptimizeBlock:
1134
1135 MachineFunction::iterator FallThrough = MBB;
1136 ++FallThrough;
1137
1138 // If this block is empty, make everyone use its fall-through, not the block
1139 // explicitly. Landing pads should not do this since the landing-pad table
1140 // points to this block. Blocks with their addresses taken shouldn't be
1141 // optimized away.
1142 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1143 // Dead block? Leave for cleanup later.
1144 if (MBB->pred_empty()) return MadeChange;
1145
1146 if (FallThrough == MF.end()) {
1147 // TODO: Simplify preds to not branch here if possible!
1148 } else {
1149 // Rewrite all predecessors of the old block to go to the fallthrough
1150 // instead.
1151 while (!MBB->pred_empty()) {
1152 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1153 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1154 }
1155 // If MBB was the target of a jump table, update jump tables to go to the
1156 // fallthrough instead.
1157 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1158 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1159 MadeChange = true;
1160 }
1161 return MadeChange;
1162 }
1163
1164 // Check to see if we can simplify the terminator of the block before this
1165 // one.
1166 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1167
1168 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1169 SmallVector<MachineOperand, 4> PriorCond;
1170 bool PriorUnAnalyzable =
1171 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1172 if (!PriorUnAnalyzable) {
1173 // If the CFG for the prior block has extra edges, remove them.
1174 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1175 !PriorCond.empty());
1176
1177 // If the previous branch is conditional and both conditions go to the same
1178 // destination, remove the branch, replacing it with an unconditional one or
1179 // a fall-through.
1180 if (PriorTBB && PriorTBB == PriorFBB) {
1181 DebugLoc dl = getBranchDebugLoc(PrevBB);
1182 TII->RemoveBranch(PrevBB);
1183 PriorCond.clear();
1184 if (PriorTBB != MBB)
1185 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1186 MadeChange = true;
1187 ++NumBranchOpts;
1188 goto ReoptimizeBlock;
1189 }
1190
1191 // If the previous block unconditionally falls through to this block and
1192 // this block has no other predecessors, move the contents of this block
1193 // into the prior block. This doesn't usually happen when SimplifyCFG
1194 // has been used, but it can happen if tail merging splits a fall-through
1195 // predecessor of a block.
1196 // This has to check PrevBB->succ_size() because EH edges are ignored by
1197 // AnalyzeBranch.
1198 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1199 PrevBB.succ_size() == 1 &&
1200 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1201 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1202 << "From MBB: " << *MBB);
1203 // Remove redundant DBG_VALUEs first.
1204 if (PrevBB.begin() != PrevBB.end()) {
1205 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1206 --PrevBBIter;
1207 MachineBasicBlock::iterator MBBIter = MBB->begin();
1208 // Check if DBG_VALUE at the end of PrevBB is identical to the
1209 // DBG_VALUE at the beginning of MBB.
1210 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1211 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1212 if (!MBBIter->isIdenticalTo(PrevBBIter))
1213 break;
1214 MachineInstr *DuplicateDbg = MBBIter;
1215 ++MBBIter; -- PrevBBIter;
1216 DuplicateDbg->eraseFromParent();
1217 }
1218 }
1219 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1220 PrevBB.removeSuccessor(PrevBB.succ_begin());
1221 assert(PrevBB.succ_empty());
1222 PrevBB.transferSuccessors(MBB);
1223 MadeChange = true;
1224 return MadeChange;
1225 }
1226
1227 // If the previous branch *only* branches to *this* block (conditional or
1228 // not) remove the branch.
1229 if (PriorTBB == MBB && !PriorFBB) {
1230 TII->RemoveBranch(PrevBB);
1231 MadeChange = true;
1232 ++NumBranchOpts;
1233 goto ReoptimizeBlock;
1234 }
1235
1236 // If the prior block branches somewhere else on the condition and here if
1237 // the condition is false, remove the uncond second branch.
1238 if (PriorFBB == MBB) {
1239 DebugLoc dl = getBranchDebugLoc(PrevBB);
1240 TII->RemoveBranch(PrevBB);
1241 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1242 MadeChange = true;
1243 ++NumBranchOpts;
1244 goto ReoptimizeBlock;
1245 }
1246
1247 // If the prior block branches here on true and somewhere else on false, and
1248 // if the branch condition is reversible, reverse the branch to create a
1249 // fall-through.
1250 if (PriorTBB == MBB) {
1251 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1252 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1253 DebugLoc dl = getBranchDebugLoc(PrevBB);
1254 TII->RemoveBranch(PrevBB);
1255 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1256 MadeChange = true;
1257 ++NumBranchOpts;
1258 goto ReoptimizeBlock;
1259 }
1260 }
1261
1262 // If this block has no successors (e.g. it is a return block or ends with
1263 // a call to a no-return function like abort or __cxa_throw) and if the pred
1264 // falls through into this block, and if it would otherwise fall through
1265 // into the block after this, move this block to the end of the function.
1266 //
1267 // We consider it more likely that execution will stay in the function (e.g.
1268 // due to loops) than it is to exit it. This asserts in loops etc, moving
1269 // the assert condition out of the loop body.
1270 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1271 MachineFunction::iterator(PriorTBB) == FallThrough &&
1272 !MBB->canFallThrough()) {
1273 bool DoTransform = true;
1274
1275 // We have to be careful that the succs of PredBB aren't both no-successor
1276 // blocks. If neither have successors and if PredBB is the second from
1277 // last block in the function, we'd just keep swapping the two blocks for
1278 // last. Only do the swap if one is clearly better to fall through than
1279 // the other.
1280 if (FallThrough == --MF.end() &&
1281 !IsBetterFallthrough(PriorTBB, MBB))
1282 DoTransform = false;
1283
1284 if (DoTransform) {
1285 // Reverse the branch so we will fall through on the previous true cond.
1286 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1287 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1288 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1289 << "To make fallthrough to: " << *PriorTBB << "\n");
1290
1291 DebugLoc dl = getBranchDebugLoc(PrevBB);
1292 TII->RemoveBranch(PrevBB);
1293 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1294
1295 // Move this block to the end of the function.
1296 MBB->moveAfter(--MF.end());
1297 MadeChange = true;
1298 ++NumBranchOpts;
1299 return MadeChange;
1300 }
1301 }
1302 }
1303 }
1304
1305 // Analyze the branch in the current block.
1306 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1307 SmallVector<MachineOperand, 4> CurCond;
1308 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1309 if (!CurUnAnalyzable) {
1310 // If the CFG for the prior block has extra edges, remove them.
1311 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1312
1313 // If this is a two-way branch, and the FBB branches to this block, reverse
1314 // the condition so the single-basic-block loop is faster. Instead of:
1315 // Loop: xxx; jcc Out; jmp Loop
1316 // we want:
1317 // Loop: xxx; jncc Loop; jmp Out
1318 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1319 SmallVector<MachineOperand, 4> NewCond(CurCond);
1320 if (!TII->ReverseBranchCondition(NewCond)) {
1321 DebugLoc dl = getBranchDebugLoc(*MBB);
1322 TII->RemoveBranch(*MBB);
1323 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1324 MadeChange = true;
1325 ++NumBranchOpts;
1326 goto ReoptimizeBlock;
1327 }
1328 }
1329
1330 // If this branch is the only thing in its block, see if we can forward
1331 // other blocks across it.
1332 if (CurTBB && CurCond.empty() && !CurFBB &&
1333 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1334 !MBB->hasAddressTaken()) {
1335 DebugLoc dl = getBranchDebugLoc(*MBB);
1336 // This block may contain just an unconditional branch. Because there can
1337 // be 'non-branch terminators' in the block, try removing the branch and
1338 // then seeing if the block is empty.
1339 TII->RemoveBranch(*MBB);
1340 // If the only things remaining in the block are debug info, remove these
1341 // as well, so this will behave the same as an empty block in non-debug
1342 // mode.
1343 if (!MBB->empty()) {
1344 bool NonDebugInfoFound = false;
1345 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1346 I != E; ++I) {
1347 if (!I->isDebugValue()) {
1348 NonDebugInfoFound = true;
1349 break;
1350 }
1351 }
1352 if (!NonDebugInfoFound)
1353 // Make the block empty, losing the debug info (we could probably
1354 // improve this in some cases.)
1355 MBB->erase(MBB->begin(), MBB->end());
1356 }
1357 // If this block is just an unconditional branch to CurTBB, we can
1358 // usually completely eliminate the block. The only case we cannot
1359 // completely eliminate the block is when the block before this one
1360 // falls through into MBB and we can't understand the prior block's branch
1361 // condition.
1362 if (MBB->empty()) {
1363 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1364 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1365 !PrevBB.isSuccessor(MBB)) {
1366 // If the prior block falls through into us, turn it into an
1367 // explicit branch to us to make updates simpler.
1368 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1369 PriorTBB != MBB && PriorFBB != MBB) {
1370 if (!PriorTBB) {
1371 assert(PriorCond.empty() && !PriorFBB &&
1372 "Bad branch analysis");
1373 PriorTBB = MBB;
1374 } else {
1375 assert(!PriorFBB && "Machine CFG out of date!");
1376 PriorFBB = MBB;
1377 }
1378 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1379 TII->RemoveBranch(PrevBB);
1380 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1381 }
1382
1383 // Iterate through all the predecessors, revectoring each in-turn.
1384 size_t PI = 0;
1385 bool DidChange = false;
1386 bool HasBranchToSelf = false;
1387 while(PI != MBB->pred_size()) {
1388 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1389 if (PMBB == MBB) {
1390 // If this block has an uncond branch to itself, leave it.
1391 ++PI;
1392 HasBranchToSelf = true;
1393 } else {
1394 DidChange = true;
1395 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1396 // If this change resulted in PMBB ending in a conditional
1397 // branch where both conditions go to the same destination,
1398 // change this to an unconditional branch (and fix the CFG).
1399 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1400 SmallVector<MachineOperand, 4> NewCurCond;
1401 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1402 NewCurFBB, NewCurCond, true);
1403 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1404 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1405 TII->RemoveBranch(*PMBB);
1406 NewCurCond.clear();
1407 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1408 MadeChange = true;
1409 ++NumBranchOpts;
1410 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1411 }
1412 }
1413 }
1414
1415 // Change any jumptables to go to the new MBB.
1416 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1417 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1418 if (DidChange) {
1419 ++NumBranchOpts;
1420 MadeChange = true;
1421 if (!HasBranchToSelf) return MadeChange;
1422 }
1423 }
1424 }
1425
1426 // Add the branch back if the block is more than just an uncond branch.
1427 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1428 }
1429 }
1430
1431 // If the prior block doesn't fall through into this block, and if this
1432 // block doesn't fall through into some other block, see if we can find a
1433 // place to move this block where a fall-through will happen.
1434 if (!PrevBB.canFallThrough()) {
1435
1436 // Now we know that there was no fall-through into this block, check to
1437 // see if it has a fall-through into its successor.
1438 bool CurFallsThru = MBB->canFallThrough();
1439
1440 if (!MBB->isLandingPad()) {
1441 // Check all the predecessors of this block. If one of them has no fall
1442 // throughs, move this block right after it.
1443 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1444 E = MBB->pred_end(); PI != E; ++PI) {
1445 // Analyze the branch at the end of the pred.
1446 MachineBasicBlock *PredBB = *PI;
1447 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1448 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1449 SmallVector<MachineOperand, 4> PredCond;
1450 if (PredBB != MBB && !PredBB->canFallThrough() &&
1451 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1452 && (!CurFallsThru || !CurTBB || !CurFBB)
1453 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1454 // If the current block doesn't fall through, just move it.
1455 // If the current block can fall through and does not end with a
1456 // conditional branch, we need to append an unconditional jump to
1457 // the (current) next block. To avoid a possible compile-time
1458 // infinite loop, move blocks only backward in this case.
1459 // Also, if there are already 2 branches here, we cannot add a third;
1460 // this means we have the case
1461 // Bcc next
1462 // B elsewhere
1463 // next:
1464 if (CurFallsThru) {
1465 MachineBasicBlock *NextBB =
1466 std::next(MachineFunction::iterator(MBB));
1467 CurCond.clear();
1468 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1469 }
1470 MBB->moveAfter(PredBB);
1471 MadeChange = true;
1472 goto ReoptimizeBlock;
1473 }
1474 }
1475 }
1476
1477 if (!CurFallsThru) {
1478 // Check all successors to see if we can move this block before it.
1479 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1480 E = MBB->succ_end(); SI != E; ++SI) {
1481 // Analyze the branch at the end of the block before the succ.
1482 MachineBasicBlock *SuccBB = *SI;
1483 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1484
1485 // If this block doesn't already fall-through to that successor, and if
1486 // the succ doesn't already have a block that can fall through into it,
1487 // and if the successor isn't an EH destination, we can arrange for the
1488 // fallthrough to happen.
1489 if (SuccBB != MBB && &*SuccPrev != MBB &&
1490 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1491 !SuccBB->isLandingPad()) {
1492 MBB->moveBefore(SuccBB);
1493 MadeChange = true;
1494 goto ReoptimizeBlock;
1495 }
1496 }
1497
1498 // Okay, there is no really great place to put this block. If, however,
1499 // the block before this one would be a fall-through if this block were
1500 // removed, move this block to the end of the function.
1501 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1502 SmallVector<MachineOperand, 4> PrevCond;
1503 if (FallThrough != MF.end() &&
1504 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1505 PrevBB.isSuccessor(FallThrough)) {
1506 MBB->moveAfter(--MF.end());
1507 MadeChange = true;
1508 return MadeChange;
1509 }
1510 }
1511 }
1512
1513 return MadeChange;
1514 }
1515
1516 //===----------------------------------------------------------------------===//
1517 // Hoist Common Code
1518 //===----------------------------------------------------------------------===//
1519
1520 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1521 /// blocks to their common predecessor.
HoistCommonCode(MachineFunction & MF)1522 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1523 bool MadeChange = false;
1524 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1525 MachineBasicBlock *MBB = I++;
1526 MadeChange |= HoistCommonCodeInSuccs(MBB);
1527 }
1528
1529 return MadeChange;
1530 }
1531
1532 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1533 /// its 'true' successor.
findFalseBlock(MachineBasicBlock * BB,MachineBasicBlock * TrueBB)1534 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1535 MachineBasicBlock *TrueBB) {
1536 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1537 E = BB->succ_end(); SI != E; ++SI) {
1538 MachineBasicBlock *SuccBB = *SI;
1539 if (SuccBB != TrueBB)
1540 return SuccBB;
1541 }
1542 return nullptr;
1543 }
1544
1545 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1546 /// in successors to. The location is usually just before the terminator,
1547 /// however if the terminator is a conditional branch and its previous
1548 /// instruction is the flag setting instruction, the previous instruction is
1549 /// the preferred location. This function also gathers uses and defs of the
1550 /// instructions from the insertion point to the end of the block. The data is
1551 /// used by HoistCommonCodeInSuccs to ensure safety.
1552 static
findHoistingInsertPosAndDeps(MachineBasicBlock * MBB,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI,SmallSet<unsigned,4> & Uses,SmallSet<unsigned,4> & Defs)1553 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1554 const TargetInstrInfo *TII,
1555 const TargetRegisterInfo *TRI,
1556 SmallSet<unsigned,4> &Uses,
1557 SmallSet<unsigned,4> &Defs) {
1558 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1559 if (!TII->isUnpredicatedTerminator(Loc))
1560 return MBB->end();
1561
1562 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1563 const MachineOperand &MO = Loc->getOperand(i);
1564 if (!MO.isReg())
1565 continue;
1566 unsigned Reg = MO.getReg();
1567 if (!Reg)
1568 continue;
1569 if (MO.isUse()) {
1570 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1571 Uses.insert(*AI);
1572 } else {
1573 if (!MO.isDead())
1574 // Don't try to hoist code in the rare case the terminator defines a
1575 // register that is later used.
1576 return MBB->end();
1577
1578 // If the terminator defines a register, make sure we don't hoist
1579 // the instruction whose def might be clobbered by the terminator.
1580 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1581 Defs.insert(*AI);
1582 }
1583 }
1584
1585 if (Uses.empty())
1586 return Loc;
1587 if (Loc == MBB->begin())
1588 return MBB->end();
1589
1590 // The terminator is probably a conditional branch, try not to separate the
1591 // branch from condition setting instruction.
1592 MachineBasicBlock::iterator PI = Loc;
1593 --PI;
1594 while (PI != MBB->begin() && PI->isDebugValue())
1595 --PI;
1596
1597 bool IsDef = false;
1598 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1599 const MachineOperand &MO = PI->getOperand(i);
1600 // If PI has a regmask operand, it is probably a call. Separate away.
1601 if (MO.isRegMask())
1602 return Loc;
1603 if (!MO.isReg() || MO.isUse())
1604 continue;
1605 unsigned Reg = MO.getReg();
1606 if (!Reg)
1607 continue;
1608 if (Uses.count(Reg))
1609 IsDef = true;
1610 }
1611 if (!IsDef)
1612 // The condition setting instruction is not just before the conditional
1613 // branch.
1614 return Loc;
1615
1616 // Be conservative, don't insert instruction above something that may have
1617 // side-effects. And since it's potentially bad to separate flag setting
1618 // instruction from the conditional branch, just abort the optimization
1619 // completely.
1620 // Also avoid moving code above predicated instruction since it's hard to
1621 // reason about register liveness with predicated instruction.
1622 bool DontMoveAcrossStore = true;
1623 if (!PI->isSafeToMove(TII, nullptr, DontMoveAcrossStore) ||
1624 TII->isPredicated(PI))
1625 return MBB->end();
1626
1627
1628 // Find out what registers are live. Note this routine is ignoring other live
1629 // registers which are only used by instructions in successor blocks.
1630 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1631 const MachineOperand &MO = PI->getOperand(i);
1632 if (!MO.isReg())
1633 continue;
1634 unsigned Reg = MO.getReg();
1635 if (!Reg)
1636 continue;
1637 if (MO.isUse()) {
1638 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1639 Uses.insert(*AI);
1640 } else {
1641 if (Uses.erase(Reg)) {
1642 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1643 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1644 }
1645 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1646 Defs.insert(*AI);
1647 }
1648 }
1649
1650 return PI;
1651 }
1652
1653 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1654 /// sequence at the start of the function, move the instructions before MBB
1655 /// terminator if it's legal.
HoistCommonCodeInSuccs(MachineBasicBlock * MBB)1656 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1657 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1658 SmallVector<MachineOperand, 4> Cond;
1659 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1660 return false;
1661
1662 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1663 if (!FBB)
1664 // Malformed bcc? True and false blocks are the same?
1665 return false;
1666
1667 // Restrict the optimization to cases where MBB is the only predecessor,
1668 // it is an obvious win.
1669 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1670 return false;
1671
1672 // Find a suitable position to hoist the common instructions to. Also figure
1673 // out which registers are used or defined by instructions from the insertion
1674 // point to the end of the block.
1675 SmallSet<unsigned, 4> Uses, Defs;
1676 MachineBasicBlock::iterator Loc =
1677 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1678 if (Loc == MBB->end())
1679 return false;
1680
1681 bool HasDups = false;
1682 SmallVector<unsigned, 4> LocalDefs;
1683 SmallSet<unsigned, 4> LocalDefsSet;
1684 MachineBasicBlock::iterator TIB = TBB->begin();
1685 MachineBasicBlock::iterator FIB = FBB->begin();
1686 MachineBasicBlock::iterator TIE = TBB->end();
1687 MachineBasicBlock::iterator FIE = FBB->end();
1688 while (TIB != TIE && FIB != FIE) {
1689 // Skip dbg_value instructions. These do not count.
1690 if (TIB->isDebugValue()) {
1691 while (TIB != TIE && TIB->isDebugValue())
1692 ++TIB;
1693 if (TIB == TIE)
1694 break;
1695 }
1696 if (FIB->isDebugValue()) {
1697 while (FIB != FIE && FIB->isDebugValue())
1698 ++FIB;
1699 if (FIB == FIE)
1700 break;
1701 }
1702 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1703 break;
1704
1705 if (TII->isPredicated(TIB))
1706 // Hard to reason about register liveness with predicated instruction.
1707 break;
1708
1709 bool IsSafe = true;
1710 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1711 MachineOperand &MO = TIB->getOperand(i);
1712 // Don't attempt to hoist instructions with register masks.
1713 if (MO.isRegMask()) {
1714 IsSafe = false;
1715 break;
1716 }
1717 if (!MO.isReg())
1718 continue;
1719 unsigned Reg = MO.getReg();
1720 if (!Reg)
1721 continue;
1722 if (MO.isDef()) {
1723 if (Uses.count(Reg)) {
1724 // Avoid clobbering a register that's used by the instruction at
1725 // the point of insertion.
1726 IsSafe = false;
1727 break;
1728 }
1729
1730 if (Defs.count(Reg) && !MO.isDead()) {
1731 // Don't hoist the instruction if the def would be clobber by the
1732 // instruction at the point insertion. FIXME: This is overly
1733 // conservative. It should be possible to hoist the instructions
1734 // in BB2 in the following example:
1735 // BB1:
1736 // r1, eflag = op1 r2, r3
1737 // brcc eflag
1738 //
1739 // BB2:
1740 // r1 = op2, ...
1741 // = op3, r1<kill>
1742 IsSafe = false;
1743 break;
1744 }
1745 } else if (!LocalDefsSet.count(Reg)) {
1746 if (Defs.count(Reg)) {
1747 // Use is defined by the instruction at the point of insertion.
1748 IsSafe = false;
1749 break;
1750 }
1751
1752 if (MO.isKill() && Uses.count(Reg))
1753 // Kills a register that's read by the instruction at the point of
1754 // insertion. Remove the kill marker.
1755 MO.setIsKill(false);
1756 }
1757 }
1758 if (!IsSafe)
1759 break;
1760
1761 bool DontMoveAcrossStore = true;
1762 if (!TIB->isSafeToMove(TII, nullptr, DontMoveAcrossStore))
1763 break;
1764
1765 // Remove kills from LocalDefsSet, these registers had short live ranges.
1766 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1767 MachineOperand &MO = TIB->getOperand(i);
1768 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1769 continue;
1770 unsigned Reg = MO.getReg();
1771 if (!Reg || !LocalDefsSet.count(Reg))
1772 continue;
1773 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1774 LocalDefsSet.erase(*AI);
1775 }
1776
1777 // Track local defs so we can update liveins.
1778 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1779 MachineOperand &MO = TIB->getOperand(i);
1780 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1781 continue;
1782 unsigned Reg = MO.getReg();
1783 if (!Reg)
1784 continue;
1785 LocalDefs.push_back(Reg);
1786 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1787 LocalDefsSet.insert(*AI);
1788 }
1789
1790 HasDups = true;
1791 ++TIB;
1792 ++FIB;
1793 }
1794
1795 if (!HasDups)
1796 return false;
1797
1798 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1799 FBB->erase(FBB->begin(), FIB);
1800
1801 // Update livein's.
1802 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1803 unsigned Def = LocalDefs[i];
1804 if (LocalDefsSet.count(Def)) {
1805 TBB->addLiveIn(Def);
1806 FBB->addLiveIn(Def);
1807 }
1808 }
1809
1810 ++NumHoist;
1811 return true;
1812 }
1813