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