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/MachineMemOperand.h"
28 #include "llvm/CodeGen/MachineModuleInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/Passes.h"
31 #include "llvm/CodeGen/RegisterScavenging.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Target/TargetInstrInfo.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
40 #include <algorithm>
41 using namespace llvm;
42
43 #define DEBUG_TYPE "branchfolding"
44
45 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
46 STATISTIC(NumBranchOpts, "Number of branches optimized");
47 STATISTIC(NumTailMerge , "Number of block tails merged");
48 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
49
50 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
51 cl::init(cl::BOU_UNSET), cl::Hidden);
52
53 // Throttle for huge numbers of predecessors (compile speed problems)
54 static cl::opt<unsigned>
55 TailMergeThreshold("tail-merge-threshold",
56 cl::desc("Max number of predecessors to consider tail merging"),
57 cl::init(150), cl::Hidden);
58
59 // Heuristic for tail merging (and, inversely, tail duplication).
60 // TODO: This should be replaced with a target query.
61 static cl::opt<unsigned>
62 TailMergeSize("tail-merge-size",
63 cl::desc("Min number of instructions to consider tail merging"),
64 cl::init(3), cl::Hidden);
65
66 namespace {
67 /// BranchFolderPass - Wrap branch folder in a machine function pass.
68 class BranchFolderPass : public MachineFunctionPass {
69 public:
70 static char ID;
BranchFolderPass()71 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
72
73 bool runOnMachineFunction(MachineFunction &MF) override;
74
getAnalysisUsage(AnalysisUsage & AU) const75 void getAnalysisUsage(AnalysisUsage &AU) const override {
76 AU.addRequired<MachineBlockFrequencyInfo>();
77 AU.addRequired<MachineBranchProbabilityInfo>();
78 AU.addRequired<TargetPassConfig>();
79 MachineFunctionPass::getAnalysisUsage(AU);
80 }
81 };
82 }
83
84 char BranchFolderPass::ID = 0;
85 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
86
87 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
88 "Control Flow Optimizer", false, false)
89
runOnMachineFunction(MachineFunction & MF)90 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
91 if (skipOptnoneFunction(*MF.getFunction()))
92 return false;
93
94 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
95 // TailMerge can create jump into if branches that make CFG irreducible for
96 // HW that requires structurized CFG.
97 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
98 PassConfig->getEnableTailMerge();
99 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true,
100 getAnalysis<MachineBlockFrequencyInfo>(),
101 getAnalysis<MachineBranchProbabilityInfo>());
102 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
103 MF.getSubtarget().getRegisterInfo(),
104 getAnalysisIfAvailable<MachineModuleInfo>());
105 }
106
BranchFolder(bool defaultEnableTailMerge,bool CommonHoist,const MachineBlockFrequencyInfo & FreqInfo,const MachineBranchProbabilityInfo & ProbInfo)107 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
108 const MachineBlockFrequencyInfo &FreqInfo,
109 const MachineBranchProbabilityInfo &ProbInfo)
110 : EnableHoistCommonCode(CommonHoist), MBBFreqInfo(FreqInfo),
111 MBPI(ProbInfo) {
112 switch (FlagEnableTailMerge) {
113 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
114 case cl::BOU_TRUE: EnableTailMerge = true; break;
115 case cl::BOU_FALSE: EnableTailMerge = false; break;
116 }
117 }
118
119 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
120 /// function, updating the CFG.
RemoveDeadBlock(MachineBasicBlock * MBB)121 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
122 assert(MBB->pred_empty() && "MBB must be dead!");
123 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
124
125 MachineFunction *MF = MBB->getParent();
126 // drop all successors.
127 while (!MBB->succ_empty())
128 MBB->removeSuccessor(MBB->succ_end()-1);
129
130 // Avoid matching if this pointer gets reused.
131 TriedMerging.erase(MBB);
132
133 // Remove the block.
134 MF->erase(MBB);
135 }
136
137 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
138 /// followed by terminators, and if the implicitly defined registers are not
139 /// used by the terminators, remove those implicit_def's. e.g.
140 /// BB1:
141 /// r0 = implicit_def
142 /// r1 = implicit_def
143 /// br
144 /// This block can be optimized away later if the implicit instructions are
145 /// removed.
OptimizeImpDefsBlock(MachineBasicBlock * MBB)146 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
147 SmallSet<unsigned, 4> ImpDefRegs;
148 MachineBasicBlock::iterator I = MBB->begin();
149 while (I != MBB->end()) {
150 if (!I->isImplicitDef())
151 break;
152 unsigned Reg = I->getOperand(0).getReg();
153 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
154 SubRegs.isValid(); ++SubRegs)
155 ImpDefRegs.insert(*SubRegs);
156 ++I;
157 }
158 if (ImpDefRegs.empty())
159 return false;
160
161 MachineBasicBlock::iterator FirstTerm = I;
162 while (I != MBB->end()) {
163 if (!TII->isUnpredicatedTerminator(I))
164 return false;
165 // See if it uses any of the implicitly defined registers.
166 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
167 MachineOperand &MO = I->getOperand(i);
168 if (!MO.isReg() || !MO.isUse())
169 continue;
170 unsigned Reg = MO.getReg();
171 if (ImpDefRegs.count(Reg))
172 return false;
173 }
174 ++I;
175 }
176
177 I = MBB->begin();
178 while (I != FirstTerm) {
179 MachineInstr *ImpDefMI = &*I;
180 ++I;
181 MBB->erase(ImpDefMI);
182 }
183
184 return true;
185 }
186
187 /// OptimizeFunction - Perhaps branch folding, tail merging and other
188 /// CFG optimizations on the given function.
OptimizeFunction(MachineFunction & MF,const TargetInstrInfo * tii,const TargetRegisterInfo * tri,MachineModuleInfo * mmi)189 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
190 const TargetInstrInfo *tii,
191 const TargetRegisterInfo *tri,
192 MachineModuleInfo *mmi) {
193 if (!tii) return false;
194
195 TriedMerging.clear();
196
197 TII = tii;
198 TRI = tri;
199 MMI = mmi;
200 RS = nullptr;
201
202 // Use a RegScavenger to help update liveness when required.
203 MachineRegisterInfo &MRI = MF.getRegInfo();
204 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
205 RS = new RegScavenger();
206 else
207 MRI.invalidateLiveness();
208
209 // Fix CFG. The later algorithms expect it to be right.
210 bool MadeChange = false;
211 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
212 MachineBasicBlock *MBB = I, *TBB = nullptr, *FBB = nullptr;
213 SmallVector<MachineOperand, 4> Cond;
214 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
215 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
216 MadeChange |= OptimizeImpDefsBlock(MBB);
217 }
218
219 bool MadeChangeThisIteration = true;
220 while (MadeChangeThisIteration) {
221 MadeChangeThisIteration = TailMergeBlocks(MF);
222 MadeChangeThisIteration |= OptimizeBranches(MF);
223 if (EnableHoistCommonCode)
224 MadeChangeThisIteration |= HoistCommonCode(MF);
225 MadeChange |= MadeChangeThisIteration;
226 }
227
228 // See if any jump tables have become dead as the code generator
229 // did its thing.
230 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
231 if (!JTI) {
232 delete RS;
233 return MadeChange;
234 }
235
236 // Walk the function to find jump tables that are live.
237 BitVector JTIsLive(JTI->getJumpTables().size());
238 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
239 BB != E; ++BB) {
240 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
241 I != E; ++I)
242 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
243 MachineOperand &Op = I->getOperand(op);
244 if (!Op.isJTI()) continue;
245
246 // Remember that this JT is live.
247 JTIsLive.set(Op.getIndex());
248 }
249 }
250
251 // Finally, remove dead jump tables. This happens when the
252 // indirect jump was unreachable (and thus deleted).
253 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
254 if (!JTIsLive.test(i)) {
255 JTI->RemoveJumpTable(i);
256 MadeChange = true;
257 }
258
259 delete RS;
260 return MadeChange;
261 }
262
263 //===----------------------------------------------------------------------===//
264 // Tail Merging of Blocks
265 //===----------------------------------------------------------------------===//
266
267 /// HashMachineInstr - Compute a hash value for MI and its operands.
HashMachineInstr(const MachineInstr * MI)268 static unsigned HashMachineInstr(const MachineInstr *MI) {
269 unsigned Hash = MI->getOpcode();
270 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
271 const MachineOperand &Op = MI->getOperand(i);
272
273 // Merge in bits from the operand if easy.
274 unsigned OperandHash = 0;
275 switch (Op.getType()) {
276 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
277 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
278 case MachineOperand::MO_MachineBasicBlock:
279 OperandHash = Op.getMBB()->getNumber();
280 break;
281 case MachineOperand::MO_FrameIndex:
282 case MachineOperand::MO_ConstantPoolIndex:
283 case MachineOperand::MO_JumpTableIndex:
284 OperandHash = Op.getIndex();
285 break;
286 case MachineOperand::MO_GlobalAddress:
287 case MachineOperand::MO_ExternalSymbol:
288 // Global address / external symbol are too hard, don't bother, but do
289 // pull in the offset.
290 OperandHash = Op.getOffset();
291 break;
292 default: break;
293 }
294
295 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
296 }
297 return Hash;
298 }
299
300 /// HashEndOfMBB - Hash the last instruction in the MBB.
HashEndOfMBB(const MachineBasicBlock * MBB)301 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
302 MachineBasicBlock::const_iterator I = MBB->end();
303 if (I == MBB->begin())
304 return 0; // Empty MBB.
305
306 --I;
307 // Skip debug info so it will not affect codegen.
308 while (I->isDebugValue()) {
309 if (I==MBB->begin())
310 return 0; // MBB empty except for debug info.
311 --I;
312 }
313
314 return HashMachineInstr(I);
315 }
316
317 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
318 /// of instructions they actually have in common together at their end. Return
319 /// iterators for the first shared instruction in each block.
ComputeCommonTailLength(MachineBasicBlock * MBB1,MachineBasicBlock * MBB2,MachineBasicBlock::iterator & I1,MachineBasicBlock::iterator & I2)320 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
321 MachineBasicBlock *MBB2,
322 MachineBasicBlock::iterator &I1,
323 MachineBasicBlock::iterator &I2) {
324 I1 = MBB1->end();
325 I2 = MBB2->end();
326
327 unsigned TailLen = 0;
328 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
329 --I1; --I2;
330 // Skip debugging pseudos; necessary to avoid changing the code.
331 while (I1->isDebugValue()) {
332 if (I1==MBB1->begin()) {
333 while (I2->isDebugValue()) {
334 if (I2==MBB2->begin())
335 // I1==DBG at begin; I2==DBG at begin
336 return TailLen;
337 --I2;
338 }
339 ++I2;
340 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
341 return TailLen;
342 }
343 --I1;
344 }
345 // I1==first (untested) non-DBG preceding known match
346 while (I2->isDebugValue()) {
347 if (I2==MBB2->begin()) {
348 ++I1;
349 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
350 return TailLen;
351 }
352 --I2;
353 }
354 // I1, I2==first (untested) non-DBGs preceding known match
355 if (!I1->isIdenticalTo(I2) ||
356 // FIXME: This check is dubious. It's used to get around a problem where
357 // people incorrectly expect inline asm directives to remain in the same
358 // relative order. This is untenable because normal compiler
359 // optimizations (like this one) may reorder and/or merge these
360 // directives.
361 I1->isInlineAsm()) {
362 ++I1; ++I2;
363 break;
364 }
365 ++TailLen;
366 }
367 // Back past possible debugging pseudos at beginning of block. This matters
368 // when one block differs from the other only by whether debugging pseudos
369 // are present at the beginning. (This way, the various checks later for
370 // I1==MBB1->begin() work as expected.)
371 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
372 --I2;
373 while (I2->isDebugValue()) {
374 if (I2 == MBB2->begin())
375 return TailLen;
376 --I2;
377 }
378 ++I2;
379 }
380 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
381 --I1;
382 while (I1->isDebugValue()) {
383 if (I1 == MBB1->begin())
384 return TailLen;
385 --I1;
386 }
387 ++I1;
388 }
389 return TailLen;
390 }
391
MaintainLiveIns(MachineBasicBlock * CurMBB,MachineBasicBlock * NewMBB)392 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
393 MachineBasicBlock *NewMBB) {
394 if (RS) {
395 RS->enterBasicBlock(CurMBB);
396 if (!CurMBB->empty())
397 RS->forward(std::prev(CurMBB->end()));
398 for (unsigned int i = 1, e = TRI->getNumRegs(); i != e; i++)
399 if (RS->isRegUsed(i, false))
400 NewMBB->addLiveIn(i);
401 }
402 }
403
404 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
405 /// after it, replacing it with an unconditional branch to NewDest.
ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,MachineBasicBlock * NewDest)406 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
407 MachineBasicBlock *NewDest) {
408 MachineBasicBlock *CurMBB = OldInst->getParent();
409
410 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
411
412 // For targets that use the register scavenger, we must maintain LiveIns.
413 MaintainLiveIns(CurMBB, NewDest);
414
415 ++NumTailMerge;
416 }
417
418 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
419 /// MBB so that the part before the iterator falls into the part starting at the
420 /// iterator. This returns the new MBB.
SplitMBBAt(MachineBasicBlock & CurMBB,MachineBasicBlock::iterator BBI1,const BasicBlock * BB)421 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
422 MachineBasicBlock::iterator BBI1,
423 const BasicBlock *BB) {
424 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
425 return nullptr;
426
427 MachineFunction &MF = *CurMBB.getParent();
428
429 // Create the fall-through block.
430 MachineFunction::iterator MBBI = &CurMBB;
431 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
432 CurMBB.getParent()->insert(++MBBI, NewMBB);
433
434 // Move all the successors of this block to the specified block.
435 NewMBB->transferSuccessors(&CurMBB);
436
437 // Add an edge from CurMBB to NewMBB for the fall-through.
438 CurMBB.addSuccessor(NewMBB);
439
440 // Splice the code over.
441 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
442
443 // NewMBB inherits CurMBB's block frequency.
444 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
445
446 // For targets that use the register scavenger, we must maintain LiveIns.
447 MaintainLiveIns(&CurMBB, NewMBB);
448
449 return NewMBB;
450 }
451
452 /// EstimateRuntime - Make a rough estimate for how long it will take to run
453 /// the specified code.
EstimateRuntime(MachineBasicBlock::iterator I,MachineBasicBlock::iterator E)454 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
455 MachineBasicBlock::iterator E) {
456 unsigned Time = 0;
457 for (; I != E; ++I) {
458 if (I->isDebugValue())
459 continue;
460 if (I->isCall())
461 Time += 10;
462 else if (I->mayLoad() || I->mayStore())
463 Time += 2;
464 else
465 ++Time;
466 }
467 return Time;
468 }
469
470 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
471 // branches temporarily for tail merging). In the case where CurMBB ends
472 // with a conditional branch to the next block, optimize by reversing the
473 // test and conditionally branching to SuccMBB instead.
FixTail(MachineBasicBlock * CurMBB,MachineBasicBlock * SuccBB,const TargetInstrInfo * TII)474 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
475 const TargetInstrInfo *TII) {
476 MachineFunction *MF = CurMBB->getParent();
477 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
478 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
479 SmallVector<MachineOperand, 4> Cond;
480 DebugLoc dl; // FIXME: this is nowhere
481 if (I != MF->end() &&
482 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
483 MachineBasicBlock *NextBB = I;
484 if (TBB == NextBB && !Cond.empty() && !FBB) {
485 if (!TII->ReverseBranchCondition(Cond)) {
486 TII->RemoveBranch(*CurMBB);
487 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
488 return;
489 }
490 }
491 }
492 TII->InsertBranch(*CurMBB, SuccBB, nullptr,
493 SmallVector<MachineOperand, 0>(), dl);
494 }
495
496 bool
operator <(const MergePotentialsElt & o) const497 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
498 if (getHash() < o.getHash())
499 return true;
500 if (getHash() > o.getHash())
501 return false;
502 if (getBlock()->getNumber() < o.getBlock()->getNumber())
503 return true;
504 if (getBlock()->getNumber() > o.getBlock()->getNumber())
505 return false;
506 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
507 // an object with itself.
508 #ifndef _GLIBCXX_DEBUG
509 llvm_unreachable("Predecessor appears twice");
510 #else
511 return false;
512 #endif
513 }
514
515 BlockFrequency
getBlockFreq(const MachineBasicBlock * MBB) const516 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
517 auto I = MergedBBFreq.find(MBB);
518
519 if (I != MergedBBFreq.end())
520 return I->second;
521
522 return MBFI.getBlockFreq(MBB);
523 }
524
setBlockFreq(const MachineBasicBlock * MBB,BlockFrequency F)525 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
526 BlockFrequency F) {
527 MergedBBFreq[MBB] = F;
528 }
529
530 /// CountTerminators - Count the number of terminators in the given
531 /// block and set I to the position of the first non-terminator, if there
532 /// is one, or MBB->end() otherwise.
CountTerminators(MachineBasicBlock * MBB,MachineBasicBlock::iterator & I)533 static unsigned CountTerminators(MachineBasicBlock *MBB,
534 MachineBasicBlock::iterator &I) {
535 I = MBB->end();
536 unsigned NumTerms = 0;
537 for (;;) {
538 if (I == MBB->begin()) {
539 I = MBB->end();
540 break;
541 }
542 --I;
543 if (!I->isTerminator()) break;
544 ++NumTerms;
545 }
546 return NumTerms;
547 }
548
549 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
550 /// and decide if it would be profitable to merge those tails. Return the
551 /// length of the common tail and iterators to the first common instruction
552 /// in each block.
ProfitableToMerge(MachineBasicBlock * MBB1,MachineBasicBlock * MBB2,unsigned minCommonTailLength,unsigned & CommonTailLen,MachineBasicBlock::iterator & I1,MachineBasicBlock::iterator & I2,MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB)553 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
554 MachineBasicBlock *MBB2,
555 unsigned minCommonTailLength,
556 unsigned &CommonTailLen,
557 MachineBasicBlock::iterator &I1,
558 MachineBasicBlock::iterator &I2,
559 MachineBasicBlock *SuccBB,
560 MachineBasicBlock *PredBB) {
561 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
562 if (CommonTailLen == 0)
563 return false;
564 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
565 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
566 << '\n');
567
568 // It's almost always profitable to merge any number of non-terminator
569 // instructions with the block that falls through into the common successor.
570 if (MBB1 == PredBB || MBB2 == PredBB) {
571 MachineBasicBlock::iterator I;
572 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
573 if (CommonTailLen > NumTerms)
574 return true;
575 }
576
577 // If one of the blocks can be completely merged and happens to be in
578 // a position where the other could fall through into it, merge any number
579 // of instructions, because it can be done without a branch.
580 // TODO: If the blocks are not adjacent, move one of them so that they are?
581 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
582 return true;
583 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
584 return true;
585
586 // If both blocks have an unconditional branch temporarily stripped out,
587 // count that as an additional common instruction for the following
588 // heuristics.
589 unsigned EffectiveTailLen = CommonTailLen;
590 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
591 !MBB1->back().isBarrier() &&
592 !MBB2->back().isBarrier())
593 ++EffectiveTailLen;
594
595 // Check if the common tail is long enough to be worthwhile.
596 if (EffectiveTailLen >= minCommonTailLength)
597 return true;
598
599 // If we are optimizing for code size, 2 instructions in common is enough if
600 // we don't have to split a block. At worst we will be introducing 1 new
601 // branch instruction, which is likely to be smaller than the 2
602 // instructions that would be deleted in the merge.
603 MachineFunction *MF = MBB1->getParent();
604 if (EffectiveTailLen >= 2 &&
605 MF->getFunction()->hasFnAttribute(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
hasIdenticalMMOs(const MachineInstr * MI1,const MachineInstr * MI2)731 static bool hasIdenticalMMOs(const MachineInstr *MI1, const MachineInstr *MI2) {
732 auto I1 = MI1->memoperands_begin(), E1 = MI1->memoperands_end();
733 auto I2 = MI2->memoperands_begin(), E2 = MI2->memoperands_end();
734 if ((E1 - I1) != (E2 - I2))
735 return false;
736 for (; I1 != E1; ++I1, ++I2) {
737 if (**I1 != **I2)
738 return false;
739 }
740 return true;
741 }
742
743 static void
removeMMOsFromMemoryOperations(MachineBasicBlock::iterator MBBIStartPos,MachineBasicBlock & MBBCommon)744 removeMMOsFromMemoryOperations(MachineBasicBlock::iterator MBBIStartPos,
745 MachineBasicBlock &MBBCommon) {
746 // Remove MMOs from memory operations in the common block
747 // when they do not match the ones from the block being tail-merged.
748 // This ensures later passes conservatively compute dependencies.
749 MachineBasicBlock *MBB = MBBIStartPos->getParent();
750 // Note CommonTailLen does not necessarily matches the size of
751 // the common BB nor all its instructions because of debug
752 // instructions differences.
753 unsigned CommonTailLen = 0;
754 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
755 ++CommonTailLen;
756
757 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
758 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
759 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
760 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
761
762 while (CommonTailLen--) {
763 assert(MBBI != MBBIE && "Reached BB end within common tail length!");
764 (void)MBBIE;
765
766 if (MBBI->isDebugValue()) {
767 ++MBBI;
768 continue;
769 }
770
771 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue())
772 ++MBBICommon;
773
774 assert(MBBICommon != MBBIECommon &&
775 "Reached BB end within common tail length!");
776 assert(MBBICommon->isIdenticalTo(&*MBBI) && "Expected matching MIIs!");
777
778 if (MBBICommon->mayLoad() || MBBICommon->mayStore())
779 if (!hasIdenticalMMOs(&*MBBI, &*MBBICommon))
780 MBBICommon->clearMemRefs();
781
782 ++MBBI;
783 ++MBBICommon;
784 }
785 }
786
787 // See if any of the blocks in MergePotentials (which all have a common single
788 // successor, or all have no successor) can be tail-merged. If there is a
789 // successor, any blocks in MergePotentials that are not tail-merged and
790 // are not immediately before Succ must have an unconditional branch to
791 // Succ added (but the predecessor/successor lists need no adjustment).
792 // The lone predecessor of Succ that falls through into Succ,
793 // if any, is given in PredBB.
794
TryTailMergeBlocks(MachineBasicBlock * SuccBB,MachineBasicBlock * PredBB)795 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
796 MachineBasicBlock *PredBB) {
797 bool MadeChange = false;
798
799 // Except for the special cases below, tail-merge if there are at least
800 // this many instructions in common.
801 unsigned minCommonTailLength = TailMergeSize;
802
803 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
804 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
805 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
806 << (i == e-1 ? "" : ", ");
807 dbgs() << "\n";
808 if (SuccBB) {
809 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
810 if (PredBB)
811 dbgs() << " which has fall-through from BB#"
812 << PredBB->getNumber() << "\n";
813 }
814 dbgs() << "Looking for common tails of at least "
815 << minCommonTailLength << " instruction"
816 << (minCommonTailLength == 1 ? "" : "s") << '\n';
817 );
818
819 // Sort by hash value so that blocks with identical end sequences sort
820 // together.
821 array_pod_sort(MergePotentials.begin(), MergePotentials.end());
822
823 // Walk through equivalence sets looking for actual exact matches.
824 while (MergePotentials.size() > 1) {
825 unsigned CurHash = MergePotentials.back().getHash();
826
827 // Build SameTails, identifying the set of blocks with this hash code
828 // and with the maximum number of instructions in common.
829 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
830 minCommonTailLength,
831 SuccBB, PredBB);
832
833 // If we didn't find any pair that has at least minCommonTailLength
834 // instructions in common, remove all blocks with this hash code and retry.
835 if (SameTails.empty()) {
836 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
837 continue;
838 }
839
840 // If one of the blocks is the entire common tail (and not the entry
841 // block, which we can't jump to), we can treat all blocks with this same
842 // tail at once. Use PredBB if that is one of the possibilities, as that
843 // will not introduce any extra branches.
844 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
845 getParent()->begin();
846 unsigned commonTailIndex = SameTails.size();
847 // If there are two blocks, check to see if one can be made to fall through
848 // into the other.
849 if (SameTails.size() == 2 &&
850 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
851 SameTails[1].tailIsWholeBlock())
852 commonTailIndex = 1;
853 else if (SameTails.size() == 2 &&
854 SameTails[1].getBlock()->isLayoutSuccessor(
855 SameTails[0].getBlock()) &&
856 SameTails[0].tailIsWholeBlock())
857 commonTailIndex = 0;
858 else {
859 // Otherwise just pick one, favoring the fall-through predecessor if
860 // there is one.
861 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
862 MachineBasicBlock *MBB = SameTails[i].getBlock();
863 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
864 continue;
865 if (MBB == PredBB) {
866 commonTailIndex = i;
867 break;
868 }
869 if (SameTails[i].tailIsWholeBlock())
870 commonTailIndex = i;
871 }
872 }
873
874 if (commonTailIndex == SameTails.size() ||
875 (SameTails[commonTailIndex].getBlock() == PredBB &&
876 !SameTails[commonTailIndex].tailIsWholeBlock())) {
877 // None of the blocks consist entirely of the common tail.
878 // Split a block so that one does.
879 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
880 maxCommonTailLength, commonTailIndex)) {
881 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
882 continue;
883 }
884 }
885
886 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
887
888 // Recompute commont tail MBB's edge weights and block frequency.
889 setCommonTailEdgeWeights(*MBB);
890
891 // MBB is common tail. Adjust all other BB's to jump to this one.
892 // Traversal must be forwards so erases work.
893 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
894 << " for ");
895 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
896 if (commonTailIndex == i)
897 continue;
898 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
899 << (i == e-1 ? "" : ", "));
900 // Remove MMOs from memory operations as needed.
901 removeMMOsFromMemoryOperations(SameTails[i].getTailStartPos(), *MBB);
902 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
903 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
904 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
905 MergePotentials.erase(SameTails[i].getMPIter());
906 }
907 DEBUG(dbgs() << "\n");
908 // We leave commonTailIndex in the worklist in case there are other blocks
909 // that match it with a smaller number of instructions.
910 MadeChange = true;
911 }
912 return MadeChange;
913 }
914
TailMergeBlocks(MachineFunction & MF)915 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
916 bool MadeChange = false;
917 if (!EnableTailMerge) return MadeChange;
918
919 // First find blocks with no successors.
920 MergePotentials.clear();
921 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
922 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
923 if (TriedMerging.count(I))
924 continue;
925 if (I->succ_empty())
926 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
927 }
928
929 // If this is a large problem, avoid visiting the same basic blocks
930 // multiple times.
931 if (MergePotentials.size() == TailMergeThreshold)
932 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
933 TriedMerging.insert(MergePotentials[i].getBlock());
934
935 // See if we can do any tail merging on those.
936 if (MergePotentials.size() >= 2)
937 MadeChange |= TryTailMergeBlocks(nullptr, nullptr);
938
939 // Look at blocks (IBB) with multiple predecessors (PBB).
940 // We change each predecessor to a canonical form, by
941 // (1) temporarily removing any unconditional branch from the predecessor
942 // to IBB, and
943 // (2) alter conditional branches so they branch to the other block
944 // not IBB; this may require adding back an unconditional branch to IBB
945 // later, where there wasn't one coming in. E.g.
946 // Bcc IBB
947 // fallthrough to QBB
948 // here becomes
949 // Bncc QBB
950 // with a conceptual B to IBB after that, which never actually exists.
951 // With those changes, we see whether the predecessors' tails match,
952 // and merge them if so. We change things out of canonical form and
953 // back to the way they were later in the process. (OptimizeBranches
954 // would undo some of this, but we can't use it, because we'd get into
955 // a compile-time infinite loop repeatedly doing and undoing the same
956 // transformations.)
957
958 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
959 I != E; ++I) {
960 if (I->pred_size() < 2) continue;
961 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
962 MachineBasicBlock *IBB = I;
963 MachineBasicBlock *PredBB = std::prev(I);
964 MergePotentials.clear();
965 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
966 E2 = I->pred_end();
967 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
968 MachineBasicBlock *PBB = *P;
969 if (TriedMerging.count(PBB))
970 continue;
971
972 // Skip blocks that loop to themselves, can't tail merge these.
973 if (PBB == IBB)
974 continue;
975
976 // Visit each predecessor only once.
977 if (!UniquePreds.insert(PBB).second)
978 continue;
979
980 // Skip blocks which may jump to a landing pad. Can't tail merge these.
981 if (PBB->getLandingPadSuccessor())
982 continue;
983
984 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
985 SmallVector<MachineOperand, 4> Cond;
986 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
987 // Failing case: IBB is the target of a cbr, and we cannot reverse the
988 // branch.
989 SmallVector<MachineOperand, 4> NewCond(Cond);
990 if (!Cond.empty() && TBB == IBB) {
991 if (TII->ReverseBranchCondition(NewCond))
992 continue;
993 // This is the QBB case described above
994 if (!FBB)
995 FBB = std::next(MachineFunction::iterator(PBB));
996 }
997
998 // Failing case: the only way IBB can be reached from PBB is via
999 // exception handling. Happens for landing pads. Would be nice to have
1000 // a bit in the edge so we didn't have to do all this.
1001 if (IBB->isLandingPad()) {
1002 MachineFunction::iterator IP = PBB; IP++;
1003 MachineBasicBlock *PredNextBB = nullptr;
1004 if (IP != MF.end())
1005 PredNextBB = IP;
1006 if (!TBB) {
1007 if (IBB != PredNextBB) // fallthrough
1008 continue;
1009 } else if (FBB) {
1010 if (TBB != IBB && FBB != IBB) // cbr then ubr
1011 continue;
1012 } else if (Cond.empty()) {
1013 if (TBB != IBB) // ubr
1014 continue;
1015 } else {
1016 if (TBB != IBB && IBB != PredNextBB) // cbr
1017 continue;
1018 }
1019 }
1020
1021 // Remove the unconditional branch at the end, if any.
1022 if (TBB && (Cond.empty() || FBB)) {
1023 DebugLoc dl; // FIXME: this is nowhere
1024 TII->RemoveBranch(*PBB);
1025 if (!Cond.empty())
1026 // reinsert conditional branch only, for now
1027 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1028 NewCond, dl);
1029 }
1030
1031 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
1032 }
1033 }
1034
1035 // If this is a large problem, avoid visiting the same basic blocks multiple
1036 // times.
1037 if (MergePotentials.size() == TailMergeThreshold)
1038 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1039 TriedMerging.insert(MergePotentials[i].getBlock());
1040
1041 if (MergePotentials.size() >= 2)
1042 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
1043
1044 // Reinsert an unconditional branch if needed. The 1 below can occur as a
1045 // result of removing blocks in TryTailMergeBlocks.
1046 PredBB = std::prev(I); // this may have been changed in TryTailMergeBlocks
1047 if (MergePotentials.size() == 1 &&
1048 MergePotentials.begin()->getBlock() != PredBB)
1049 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1050 }
1051
1052 return MadeChange;
1053 }
1054
setCommonTailEdgeWeights(MachineBasicBlock & TailMBB)1055 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1056 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1057 BlockFrequency AccumulatedMBBFreq;
1058
1059 // Aggregate edge frequency of successor edge j:
1060 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1061 // where bb is a basic block that is in SameTails.
1062 for (const auto &Src : SameTails) {
1063 const MachineBasicBlock *SrcMBB = Src.getBlock();
1064 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1065 AccumulatedMBBFreq += BlockFreq;
1066
1067 // It is not necessary to recompute edge weights if TailBB has less than two
1068 // successors.
1069 if (TailMBB.succ_size() <= 1)
1070 continue;
1071
1072 auto EdgeFreq = EdgeFreqLs.begin();
1073
1074 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1075 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1076 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1077 }
1078
1079 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1080
1081 if (TailMBB.succ_size() <= 1)
1082 return;
1083
1084 auto MaxEdgeFreq = *std::max_element(EdgeFreqLs.begin(), EdgeFreqLs.end());
1085 uint64_t Scale = MaxEdgeFreq.getFrequency() / UINT32_MAX + 1;
1086 auto EdgeFreq = EdgeFreqLs.begin();
1087
1088 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1089 SuccI != SuccE; ++SuccI, ++EdgeFreq)
1090 TailMBB.setSuccWeight(SuccI, EdgeFreq->getFrequency() / Scale);
1091 }
1092
1093 //===----------------------------------------------------------------------===//
1094 // Branch Optimization
1095 //===----------------------------------------------------------------------===//
1096
OptimizeBranches(MachineFunction & MF)1097 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1098 bool MadeChange = false;
1099
1100 // Make sure blocks are numbered in order
1101 MF.RenumberBlocks();
1102
1103 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1104 I != E; ) {
1105 MachineBasicBlock *MBB = I++;
1106 MadeChange |= OptimizeBlock(MBB);
1107
1108 // If it is dead, remove it.
1109 if (MBB->pred_empty()) {
1110 RemoveDeadBlock(MBB);
1111 MadeChange = true;
1112 ++NumDeadBlocks;
1113 }
1114 }
1115 return MadeChange;
1116 }
1117
1118 // Blocks should be considered empty if they contain only debug info;
1119 // else the debug info would affect codegen.
IsEmptyBlock(MachineBasicBlock * MBB)1120 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1121 if (MBB->empty())
1122 return true;
1123 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
1124 MBBI!=MBBE; ++MBBI) {
1125 if (!MBBI->isDebugValue())
1126 return false;
1127 }
1128 return true;
1129 }
1130
1131 // Blocks with only debug info and branches should be considered the same
1132 // as blocks with only branches.
IsBranchOnlyBlock(MachineBasicBlock * MBB)1133 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1134 MachineBasicBlock::iterator MBBI, MBBE;
1135 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
1136 if (!MBBI->isDebugValue())
1137 break;
1138 }
1139 return (MBBI->isBranch());
1140 }
1141
1142 /// IsBetterFallthrough - Return true if it would be clearly better to
1143 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1144 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1145 /// result in infinite loops.
IsBetterFallthrough(MachineBasicBlock * MBB1,MachineBasicBlock * MBB2)1146 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1147 MachineBasicBlock *MBB2) {
1148 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1149 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1150 // optimize branches that branch to either a return block or an assert block
1151 // into a fallthrough to the return.
1152 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
1153
1154 // If there is a clear successor ordering we make sure that one block
1155 // will fall through to the next
1156 if (MBB1->isSuccessor(MBB2)) return true;
1157 if (MBB2->isSuccessor(MBB1)) return false;
1158
1159 // Neither block consists entirely of debug info (per IsEmptyBlock check),
1160 // so we needn't test for falling off the beginning here.
1161 MachineBasicBlock::iterator MBB1I = --MBB1->end();
1162 while (MBB1I->isDebugValue())
1163 --MBB1I;
1164 MachineBasicBlock::iterator MBB2I = --MBB2->end();
1165 while (MBB2I->isDebugValue())
1166 --MBB2I;
1167 return MBB2I->isCall() && !MBB1I->isCall();
1168 }
1169
1170 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1171 /// instructions on the block. Always use the DebugLoc of the first
1172 /// branching instruction found unless its absent, in which case use the
1173 /// DebugLoc of the second if present.
getBranchDebugLoc(MachineBasicBlock & MBB)1174 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1175 MachineBasicBlock::iterator I = MBB.end();
1176 if (I == MBB.begin())
1177 return DebugLoc();
1178 --I;
1179 while (I->isDebugValue() && I != MBB.begin())
1180 --I;
1181 if (I->isBranch())
1182 return I->getDebugLoc();
1183 return DebugLoc();
1184 }
1185
1186 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1187 /// block. This is never called on the entry block.
OptimizeBlock(MachineBasicBlock * MBB)1188 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1189 bool MadeChange = false;
1190 MachineFunction &MF = *MBB->getParent();
1191 ReoptimizeBlock:
1192
1193 MachineFunction::iterator FallThrough = MBB;
1194 ++FallThrough;
1195
1196 // If this block is empty, make everyone use its fall-through, not the block
1197 // explicitly. Landing pads should not do this since the landing-pad table
1198 // points to this block. Blocks with their addresses taken shouldn't be
1199 // optimized away.
1200 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1201 // Dead block? Leave for cleanup later.
1202 if (MBB->pred_empty()) return MadeChange;
1203
1204 if (FallThrough == MF.end()) {
1205 // TODO: Simplify preds to not branch here if possible!
1206 } else {
1207 // Rewrite all predecessors of the old block to go to the fallthrough
1208 // instead.
1209 while (!MBB->pred_empty()) {
1210 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1211 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1212 }
1213 // If MBB was the target of a jump table, update jump tables to go to the
1214 // fallthrough instead.
1215 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1216 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1217 MadeChange = true;
1218 }
1219 return MadeChange;
1220 }
1221
1222 // Check to see if we can simplify the terminator of the block before this
1223 // one.
1224 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1225
1226 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1227 SmallVector<MachineOperand, 4> PriorCond;
1228 bool PriorUnAnalyzable =
1229 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1230 if (!PriorUnAnalyzable) {
1231 // If the CFG for the prior block has extra edges, remove them.
1232 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1233 !PriorCond.empty());
1234
1235 // If the previous branch is conditional and both conditions go to the same
1236 // destination, remove the branch, replacing it with an unconditional one or
1237 // a fall-through.
1238 if (PriorTBB && PriorTBB == PriorFBB) {
1239 DebugLoc dl = getBranchDebugLoc(PrevBB);
1240 TII->RemoveBranch(PrevBB);
1241 PriorCond.clear();
1242 if (PriorTBB != MBB)
1243 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1244 MadeChange = true;
1245 ++NumBranchOpts;
1246 goto ReoptimizeBlock;
1247 }
1248
1249 // If the previous block unconditionally falls through to this block and
1250 // this block has no other predecessors, move the contents of this block
1251 // into the prior block. This doesn't usually happen when SimplifyCFG
1252 // has been used, but it can happen if tail merging splits a fall-through
1253 // predecessor of a block.
1254 // This has to check PrevBB->succ_size() because EH edges are ignored by
1255 // AnalyzeBranch.
1256 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1257 PrevBB.succ_size() == 1 &&
1258 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1259 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1260 << "From MBB: " << *MBB);
1261 // Remove redundant DBG_VALUEs first.
1262 if (PrevBB.begin() != PrevBB.end()) {
1263 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1264 --PrevBBIter;
1265 MachineBasicBlock::iterator MBBIter = MBB->begin();
1266 // Check if DBG_VALUE at the end of PrevBB is identical to the
1267 // DBG_VALUE at the beginning of MBB.
1268 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1269 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1270 if (!MBBIter->isIdenticalTo(PrevBBIter))
1271 break;
1272 MachineInstr *DuplicateDbg = MBBIter;
1273 ++MBBIter; -- PrevBBIter;
1274 DuplicateDbg->eraseFromParent();
1275 }
1276 }
1277 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1278 PrevBB.removeSuccessor(PrevBB.succ_begin());
1279 assert(PrevBB.succ_empty());
1280 PrevBB.transferSuccessors(MBB);
1281 MadeChange = true;
1282 return MadeChange;
1283 }
1284
1285 // If the previous branch *only* branches to *this* block (conditional or
1286 // not) remove the branch.
1287 if (PriorTBB == MBB && !PriorFBB) {
1288 TII->RemoveBranch(PrevBB);
1289 MadeChange = true;
1290 ++NumBranchOpts;
1291 goto ReoptimizeBlock;
1292 }
1293
1294 // If the prior block branches somewhere else on the condition and here if
1295 // the condition is false, remove the uncond second branch.
1296 if (PriorFBB == MBB) {
1297 DebugLoc dl = getBranchDebugLoc(PrevBB);
1298 TII->RemoveBranch(PrevBB);
1299 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1300 MadeChange = true;
1301 ++NumBranchOpts;
1302 goto ReoptimizeBlock;
1303 }
1304
1305 // If the prior block branches here on true and somewhere else on false, and
1306 // if the branch condition is reversible, reverse the branch to create a
1307 // fall-through.
1308 if (PriorTBB == MBB) {
1309 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1310 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1311 DebugLoc dl = getBranchDebugLoc(PrevBB);
1312 TII->RemoveBranch(PrevBB);
1313 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1314 MadeChange = true;
1315 ++NumBranchOpts;
1316 goto ReoptimizeBlock;
1317 }
1318 }
1319
1320 // If this block has no successors (e.g. it is a return block or ends with
1321 // a call to a no-return function like abort or __cxa_throw) and if the pred
1322 // falls through into this block, and if it would otherwise fall through
1323 // into the block after this, move this block to the end of the function.
1324 //
1325 // We consider it more likely that execution will stay in the function (e.g.
1326 // due to loops) than it is to exit it. This asserts in loops etc, moving
1327 // the assert condition out of the loop body.
1328 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1329 MachineFunction::iterator(PriorTBB) == FallThrough &&
1330 !MBB->canFallThrough()) {
1331 bool DoTransform = true;
1332
1333 // We have to be careful that the succs of PredBB aren't both no-successor
1334 // blocks. If neither have successors and if PredBB is the second from
1335 // last block in the function, we'd just keep swapping the two blocks for
1336 // last. Only do the swap if one is clearly better to fall through than
1337 // the other.
1338 if (FallThrough == --MF.end() &&
1339 !IsBetterFallthrough(PriorTBB, MBB))
1340 DoTransform = false;
1341
1342 if (DoTransform) {
1343 // Reverse the branch so we will fall through on the previous true cond.
1344 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1345 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1346 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1347 << "To make fallthrough to: " << *PriorTBB << "\n");
1348
1349 DebugLoc dl = getBranchDebugLoc(PrevBB);
1350 TII->RemoveBranch(PrevBB);
1351 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1352
1353 // Move this block to the end of the function.
1354 MBB->moveAfter(--MF.end());
1355 MadeChange = true;
1356 ++NumBranchOpts;
1357 return MadeChange;
1358 }
1359 }
1360 }
1361 }
1362
1363 // Analyze the branch in the current block.
1364 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1365 SmallVector<MachineOperand, 4> CurCond;
1366 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1367 if (!CurUnAnalyzable) {
1368 // If the CFG for the prior block has extra edges, remove them.
1369 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1370
1371 // If this is a two-way branch, and the FBB branches to this block, reverse
1372 // the condition so the single-basic-block loop is faster. Instead of:
1373 // Loop: xxx; jcc Out; jmp Loop
1374 // we want:
1375 // Loop: xxx; jncc Loop; jmp Out
1376 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1377 SmallVector<MachineOperand, 4> NewCond(CurCond);
1378 if (!TII->ReverseBranchCondition(NewCond)) {
1379 DebugLoc dl = getBranchDebugLoc(*MBB);
1380 TII->RemoveBranch(*MBB);
1381 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1382 MadeChange = true;
1383 ++NumBranchOpts;
1384 goto ReoptimizeBlock;
1385 }
1386 }
1387
1388 // If this branch is the only thing in its block, see if we can forward
1389 // other blocks across it.
1390 if (CurTBB && CurCond.empty() && !CurFBB &&
1391 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1392 !MBB->hasAddressTaken()) {
1393 DebugLoc dl = getBranchDebugLoc(*MBB);
1394 // This block may contain just an unconditional branch. Because there can
1395 // be 'non-branch terminators' in the block, try removing the branch and
1396 // then seeing if the block is empty.
1397 TII->RemoveBranch(*MBB);
1398 // If the only things remaining in the block are debug info, remove these
1399 // as well, so this will behave the same as an empty block in non-debug
1400 // mode.
1401 if (!MBB->empty()) {
1402 bool NonDebugInfoFound = false;
1403 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1404 I != E; ++I) {
1405 if (!I->isDebugValue()) {
1406 NonDebugInfoFound = true;
1407 break;
1408 }
1409 }
1410 if (!NonDebugInfoFound)
1411 // Make the block empty, losing the debug info (we could probably
1412 // improve this in some cases.)
1413 MBB->erase(MBB->begin(), MBB->end());
1414 }
1415 // If this block is just an unconditional branch to CurTBB, we can
1416 // usually completely eliminate the block. The only case we cannot
1417 // completely eliminate the block is when the block before this one
1418 // falls through into MBB and we can't understand the prior block's branch
1419 // condition.
1420 if (MBB->empty()) {
1421 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1422 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1423 !PrevBB.isSuccessor(MBB)) {
1424 // If the prior block falls through into us, turn it into an
1425 // explicit branch to us to make updates simpler.
1426 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1427 PriorTBB != MBB && PriorFBB != MBB) {
1428 if (!PriorTBB) {
1429 assert(PriorCond.empty() && !PriorFBB &&
1430 "Bad branch analysis");
1431 PriorTBB = MBB;
1432 } else {
1433 assert(!PriorFBB && "Machine CFG out of date!");
1434 PriorFBB = MBB;
1435 }
1436 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1437 TII->RemoveBranch(PrevBB);
1438 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1439 }
1440
1441 // Iterate through all the predecessors, revectoring each in-turn.
1442 size_t PI = 0;
1443 bool DidChange = false;
1444 bool HasBranchToSelf = false;
1445 while(PI != MBB->pred_size()) {
1446 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1447 if (PMBB == MBB) {
1448 // If this block has an uncond branch to itself, leave it.
1449 ++PI;
1450 HasBranchToSelf = true;
1451 } else {
1452 DidChange = true;
1453 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1454 // If this change resulted in PMBB ending in a conditional
1455 // branch where both conditions go to the same destination,
1456 // change this to an unconditional branch (and fix the CFG).
1457 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1458 SmallVector<MachineOperand, 4> NewCurCond;
1459 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1460 NewCurFBB, NewCurCond, true);
1461 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1462 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1463 TII->RemoveBranch(*PMBB);
1464 NewCurCond.clear();
1465 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1466 MadeChange = true;
1467 ++NumBranchOpts;
1468 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1469 }
1470 }
1471 }
1472
1473 // Change any jumptables to go to the new MBB.
1474 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1475 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1476 if (DidChange) {
1477 ++NumBranchOpts;
1478 MadeChange = true;
1479 if (!HasBranchToSelf) return MadeChange;
1480 }
1481 }
1482 }
1483
1484 // Add the branch back if the block is more than just an uncond branch.
1485 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1486 }
1487 }
1488
1489 // If the prior block doesn't fall through into this block, and if this
1490 // block doesn't fall through into some other block, see if we can find a
1491 // place to move this block where a fall-through will happen.
1492 if (!PrevBB.canFallThrough()) {
1493
1494 // Now we know that there was no fall-through into this block, check to
1495 // see if it has a fall-through into its successor.
1496 bool CurFallsThru = MBB->canFallThrough();
1497
1498 if (!MBB->isLandingPad()) {
1499 // Check all the predecessors of this block. If one of them has no fall
1500 // throughs, move this block right after it.
1501 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1502 E = MBB->pred_end(); PI != E; ++PI) {
1503 // Analyze the branch at the end of the pred.
1504 MachineBasicBlock *PredBB = *PI;
1505 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1506 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1507 SmallVector<MachineOperand, 4> PredCond;
1508 if (PredBB != MBB && !PredBB->canFallThrough() &&
1509 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1510 && (!CurFallsThru || !CurTBB || !CurFBB)
1511 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1512 // If the current block doesn't fall through, just move it.
1513 // If the current block can fall through and does not end with a
1514 // conditional branch, we need to append an unconditional jump to
1515 // the (current) next block. To avoid a possible compile-time
1516 // infinite loop, move blocks only backward in this case.
1517 // Also, if there are already 2 branches here, we cannot add a third;
1518 // this means we have the case
1519 // Bcc next
1520 // B elsewhere
1521 // next:
1522 if (CurFallsThru) {
1523 MachineBasicBlock *NextBB =
1524 std::next(MachineFunction::iterator(MBB));
1525 CurCond.clear();
1526 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1527 }
1528 MBB->moveAfter(PredBB);
1529 MadeChange = true;
1530 goto ReoptimizeBlock;
1531 }
1532 }
1533 }
1534
1535 if (!CurFallsThru) {
1536 // Check all successors to see if we can move this block before it.
1537 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1538 E = MBB->succ_end(); SI != E; ++SI) {
1539 // Analyze the branch at the end of the block before the succ.
1540 MachineBasicBlock *SuccBB = *SI;
1541 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1542
1543 // If this block doesn't already fall-through to that successor, and if
1544 // the succ doesn't already have a block that can fall through into it,
1545 // and if the successor isn't an EH destination, we can arrange for the
1546 // fallthrough to happen.
1547 if (SuccBB != MBB && &*SuccPrev != MBB &&
1548 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1549 !SuccBB->isLandingPad()) {
1550 MBB->moveBefore(SuccBB);
1551 MadeChange = true;
1552 goto ReoptimizeBlock;
1553 }
1554 }
1555
1556 // Okay, there is no really great place to put this block. If, however,
1557 // the block before this one would be a fall-through if this block were
1558 // removed, move this block to the end of the function.
1559 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1560 SmallVector<MachineOperand, 4> PrevCond;
1561 if (FallThrough != MF.end() &&
1562 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1563 PrevBB.isSuccessor(FallThrough)) {
1564 MBB->moveAfter(--MF.end());
1565 MadeChange = true;
1566 return MadeChange;
1567 }
1568 }
1569 }
1570
1571 return MadeChange;
1572 }
1573
1574 //===----------------------------------------------------------------------===//
1575 // Hoist Common Code
1576 //===----------------------------------------------------------------------===//
1577
1578 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1579 /// blocks to their common predecessor.
HoistCommonCode(MachineFunction & MF)1580 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1581 bool MadeChange = false;
1582 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1583 MachineBasicBlock *MBB = I++;
1584 MadeChange |= HoistCommonCodeInSuccs(MBB);
1585 }
1586
1587 return MadeChange;
1588 }
1589
1590 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1591 /// its 'true' successor.
findFalseBlock(MachineBasicBlock * BB,MachineBasicBlock * TrueBB)1592 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1593 MachineBasicBlock *TrueBB) {
1594 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1595 E = BB->succ_end(); SI != E; ++SI) {
1596 MachineBasicBlock *SuccBB = *SI;
1597 if (SuccBB != TrueBB)
1598 return SuccBB;
1599 }
1600 return nullptr;
1601 }
1602
1603 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1604 /// in successors to. The location is usually just before the terminator,
1605 /// however if the terminator is a conditional branch and its previous
1606 /// instruction is the flag setting instruction, the previous instruction is
1607 /// the preferred location. This function also gathers uses and defs of the
1608 /// instructions from the insertion point to the end of the block. The data is
1609 /// used by HoistCommonCodeInSuccs to ensure safety.
1610 static
findHoistingInsertPosAndDeps(MachineBasicBlock * MBB,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI,SmallSet<unsigned,4> & Uses,SmallSet<unsigned,4> & Defs)1611 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1612 const TargetInstrInfo *TII,
1613 const TargetRegisterInfo *TRI,
1614 SmallSet<unsigned,4> &Uses,
1615 SmallSet<unsigned,4> &Defs) {
1616 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1617 if (!TII->isUnpredicatedTerminator(Loc))
1618 return MBB->end();
1619
1620 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1621 const MachineOperand &MO = Loc->getOperand(i);
1622 if (!MO.isReg())
1623 continue;
1624 unsigned Reg = MO.getReg();
1625 if (!Reg)
1626 continue;
1627 if (MO.isUse()) {
1628 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1629 Uses.insert(*AI);
1630 } else {
1631 if (!MO.isDead())
1632 // Don't try to hoist code in the rare case the terminator defines a
1633 // register that is later used.
1634 return MBB->end();
1635
1636 // If the terminator defines a register, make sure we don't hoist
1637 // the instruction whose def might be clobbered by the terminator.
1638 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1639 Defs.insert(*AI);
1640 }
1641 }
1642
1643 if (Uses.empty())
1644 return Loc;
1645 if (Loc == MBB->begin())
1646 return MBB->end();
1647
1648 // The terminator is probably a conditional branch, try not to separate the
1649 // branch from condition setting instruction.
1650 MachineBasicBlock::iterator PI = Loc;
1651 --PI;
1652 while (PI != MBB->begin() && PI->isDebugValue())
1653 --PI;
1654
1655 bool IsDef = false;
1656 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1657 const MachineOperand &MO = PI->getOperand(i);
1658 // If PI has a regmask operand, it is probably a call. Separate away.
1659 if (MO.isRegMask())
1660 return Loc;
1661 if (!MO.isReg() || MO.isUse())
1662 continue;
1663 unsigned Reg = MO.getReg();
1664 if (!Reg)
1665 continue;
1666 if (Uses.count(Reg))
1667 IsDef = true;
1668 }
1669 if (!IsDef)
1670 // The condition setting instruction is not just before the conditional
1671 // branch.
1672 return Loc;
1673
1674 // Be conservative, don't insert instruction above something that may have
1675 // side-effects. And since it's potentially bad to separate flag setting
1676 // instruction from the conditional branch, just abort the optimization
1677 // completely.
1678 // Also avoid moving code above predicated instruction since it's hard to
1679 // reason about register liveness with predicated instruction.
1680 bool DontMoveAcrossStore = true;
1681 if (!PI->isSafeToMove(TII, nullptr, DontMoveAcrossStore) ||
1682 TII->isPredicated(PI))
1683 return MBB->end();
1684
1685
1686 // Find out what registers are live. Note this routine is ignoring other live
1687 // registers which are only used by instructions in successor blocks.
1688 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1689 const MachineOperand &MO = PI->getOperand(i);
1690 if (!MO.isReg())
1691 continue;
1692 unsigned Reg = MO.getReg();
1693 if (!Reg)
1694 continue;
1695 if (MO.isUse()) {
1696 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1697 Uses.insert(*AI);
1698 } else {
1699 if (Uses.erase(Reg)) {
1700 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1701 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1702 }
1703 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1704 Defs.insert(*AI);
1705 }
1706 }
1707
1708 return PI;
1709 }
1710
1711 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1712 /// sequence at the start of the function, move the instructions before MBB
1713 /// terminator if it's legal.
HoistCommonCodeInSuccs(MachineBasicBlock * MBB)1714 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1715 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1716 SmallVector<MachineOperand, 4> Cond;
1717 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1718 return false;
1719
1720 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1721 if (!FBB)
1722 // Malformed bcc? True and false blocks are the same?
1723 return false;
1724
1725 // Restrict the optimization to cases where MBB is the only predecessor,
1726 // it is an obvious win.
1727 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1728 return false;
1729
1730 // Find a suitable position to hoist the common instructions to. Also figure
1731 // out which registers are used or defined by instructions from the insertion
1732 // point to the end of the block.
1733 SmallSet<unsigned, 4> Uses, Defs;
1734 MachineBasicBlock::iterator Loc =
1735 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1736 if (Loc == MBB->end())
1737 return false;
1738
1739 bool HasDups = false;
1740 SmallVector<unsigned, 4> LocalDefs;
1741 SmallSet<unsigned, 4> LocalDefsSet;
1742 MachineBasicBlock::iterator TIB = TBB->begin();
1743 MachineBasicBlock::iterator FIB = FBB->begin();
1744 MachineBasicBlock::iterator TIE = TBB->end();
1745 MachineBasicBlock::iterator FIE = FBB->end();
1746 while (TIB != TIE && FIB != FIE) {
1747 // Skip dbg_value instructions. These do not count.
1748 if (TIB->isDebugValue()) {
1749 while (TIB != TIE && TIB->isDebugValue())
1750 ++TIB;
1751 if (TIB == TIE)
1752 break;
1753 }
1754 if (FIB->isDebugValue()) {
1755 while (FIB != FIE && FIB->isDebugValue())
1756 ++FIB;
1757 if (FIB == FIE)
1758 break;
1759 }
1760 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1761 break;
1762
1763 if (TII->isPredicated(TIB))
1764 // Hard to reason about register liveness with predicated instruction.
1765 break;
1766
1767 bool IsSafe = true;
1768 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1769 MachineOperand &MO = TIB->getOperand(i);
1770 // Don't attempt to hoist instructions with register masks.
1771 if (MO.isRegMask()) {
1772 IsSafe = false;
1773 break;
1774 }
1775 if (!MO.isReg())
1776 continue;
1777 unsigned Reg = MO.getReg();
1778 if (!Reg)
1779 continue;
1780 if (MO.isDef()) {
1781 if (Uses.count(Reg)) {
1782 // Avoid clobbering a register that's used by the instruction at
1783 // the point of insertion.
1784 IsSafe = false;
1785 break;
1786 }
1787
1788 if (Defs.count(Reg) && !MO.isDead()) {
1789 // Don't hoist the instruction if the def would be clobber by the
1790 // instruction at the point insertion. FIXME: This is overly
1791 // conservative. It should be possible to hoist the instructions
1792 // in BB2 in the following example:
1793 // BB1:
1794 // r1, eflag = op1 r2, r3
1795 // brcc eflag
1796 //
1797 // BB2:
1798 // r1 = op2, ...
1799 // = op3, r1<kill>
1800 IsSafe = false;
1801 break;
1802 }
1803 } else if (!LocalDefsSet.count(Reg)) {
1804 if (Defs.count(Reg)) {
1805 // Use is defined by the instruction at the point of insertion.
1806 IsSafe = false;
1807 break;
1808 }
1809
1810 if (MO.isKill() && Uses.count(Reg))
1811 // Kills a register that's read by the instruction at the point of
1812 // insertion. Remove the kill marker.
1813 MO.setIsKill(false);
1814 }
1815 }
1816 if (!IsSafe)
1817 break;
1818
1819 bool DontMoveAcrossStore = true;
1820 if (!TIB->isSafeToMove(TII, nullptr, DontMoveAcrossStore))
1821 break;
1822
1823 // Remove kills from LocalDefsSet, these registers had short live ranges.
1824 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1825 MachineOperand &MO = TIB->getOperand(i);
1826 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1827 continue;
1828 unsigned Reg = MO.getReg();
1829 if (!Reg || !LocalDefsSet.count(Reg))
1830 continue;
1831 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1832 LocalDefsSet.erase(*AI);
1833 }
1834
1835 // Track local defs so we can update liveins.
1836 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1837 MachineOperand &MO = TIB->getOperand(i);
1838 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1839 continue;
1840 unsigned Reg = MO.getReg();
1841 if (!Reg)
1842 continue;
1843 LocalDefs.push_back(Reg);
1844 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1845 LocalDefsSet.insert(*AI);
1846 }
1847
1848 HasDups = true;
1849 ++TIB;
1850 ++FIB;
1851 }
1852
1853 if (!HasDups)
1854 return false;
1855
1856 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1857 FBB->erase(FBB->begin(), FIB);
1858
1859 // Update livein's.
1860 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1861 unsigned Def = LocalDefs[i];
1862 if (LocalDefsSet.count(Def)) {
1863 TBB->addLiveIn(Def);
1864 FBB->addLiveIn(Def);
1865 }
1866 }
1867
1868 ++NumHoist;
1869 return true;
1870 }
1871