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 #define DEBUG_TYPE "branchfolding"
20 #include "BranchFolding.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/MachineModuleInfo.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/Passes.h"
29 #include "llvm/CodeGen/RegisterScavenging.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetInstrInfo.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetRegisterInfo.h"
38 #include <algorithm>
39 using namespace llvm;
41 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
42 STATISTIC(NumBranchOpts, "Number of branches optimized");
43 STATISTIC(NumTailMerge , "Number of block tails merged");
44 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
46 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
47 cl::init(cl::BOU_UNSET), cl::Hidden);
49 // Throttle for huge numbers of predecessors (compile speed problems)
50 static cl::opt<unsigned>
51 TailMergeThreshold("tail-merge-threshold",
52 cl::desc("Max number of predecessors to consider tail merging"),
53 cl::init(150), cl::Hidden);
55 // Heuristic for tail merging (and, inversely, tail duplication).
56 // TODO: This should be replaced with a target query.
57 static cl::opt<unsigned>
58 TailMergeSize("tail-merge-size",
59 cl::desc("Min number of instructions to consider tail merging"),
60 cl::init(3), cl::Hidden);
62 namespace {
63 /// BranchFolderPass - Wrap branch folder in a machine function pass.
64 class BranchFolderPass : public MachineFunctionPass {
65 public:
66 static char ID;
67 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
69 virtual bool runOnMachineFunction(MachineFunction &MF);
71 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
72 AU.addRequired<TargetPassConfig>();
73 MachineFunctionPass::getAnalysisUsage(AU);
74 }
75 };
76 }
78 char BranchFolderPass::ID = 0;
79 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
81 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
82 "Control Flow Optimizer", false, false)
84 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
85 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
86 BranchFolder Folder(PassConfig->getEnableTailMerge(), /*CommonHoist=*/true);
87 return Folder.OptimizeFunction(MF,
88 MF.getTarget().getInstrInfo(),
89 MF.getTarget().getRegisterInfo(),
90 getAnalysisIfAvailable<MachineModuleInfo>());
91 }
94 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist) {
95 switch (FlagEnableTailMerge) {
96 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
97 case cl::BOU_TRUE: EnableTailMerge = true; break;
98 case cl::BOU_FALSE: EnableTailMerge = false; break;
99 }
101 EnableHoistCommonCode = CommonHoist;
102 }
104 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
105 /// function, updating the CFG.
106 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
107 assert(MBB->pred_empty() && "MBB must be dead!");
108 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
110 MachineFunction *MF = MBB->getParent();
111 // drop all successors.
112 while (!MBB->succ_empty())
113 MBB->removeSuccessor(MBB->succ_end()-1);
115 // Avoid matching if this pointer gets reused.
116 TriedMerging.erase(MBB);
118 // Remove the block.
119 MF->erase(MBB);
120 }
122 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
123 /// followed by terminators, and if the implicitly defined registers are not
124 /// used by the terminators, remove those implicit_def's. e.g.
125 /// BB1:
126 /// r0 = implicit_def
127 /// r1 = implicit_def
128 /// br
129 /// This block can be optimized away later if the implicit instructions are
130 /// removed.
131 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
132 SmallSet<unsigned, 4> ImpDefRegs;
133 MachineBasicBlock::iterator I = MBB->begin();
134 while (I != MBB->end()) {
135 if (!I->isImplicitDef())
136 break;
137 unsigned Reg = I->getOperand(0).getReg();
138 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
139 SubRegs.isValid(); ++SubRegs)
140 ImpDefRegs.insert(*SubRegs);
141 ++I;
142 }
143 if (ImpDefRegs.empty())
144 return false;
146 MachineBasicBlock::iterator FirstTerm = I;
147 while (I != MBB->end()) {
148 if (!TII->isUnpredicatedTerminator(I))
149 return false;
150 // See if it uses any of the implicitly defined registers.
151 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
152 MachineOperand &MO = I->getOperand(i);
153 if (!MO.isReg() || !MO.isUse())
154 continue;
155 unsigned Reg = MO.getReg();
156 if (ImpDefRegs.count(Reg))
157 return false;
158 }
159 ++I;
160 }
162 I = MBB->begin();
163 while (I != FirstTerm) {
164 MachineInstr *ImpDefMI = &*I;
165 ++I;
166 MBB->erase(ImpDefMI);
167 }
169 return true;
170 }
172 /// OptimizeFunction - Perhaps branch folding, tail merging and other
173 /// CFG optimizations on the given function.
174 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
175 const TargetInstrInfo *tii,
176 const TargetRegisterInfo *tri,
177 MachineModuleInfo *mmi) {
178 if (!tii) return false;
180 TriedMerging.clear();
182 TII = tii;
183 TRI = tri;
184 MMI = mmi;
185 RS = NULL;
187 // Use a RegScavenger to help update liveness when required.
188 MachineRegisterInfo &MRI = MF.getRegInfo();
189 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
190 RS = new RegScavenger();
191 else
192 MRI.invalidateLiveness();
194 // Fix CFG. The later algorithms expect it to be right.
195 bool MadeChange = false;
196 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
197 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
198 SmallVector<MachineOperand, 4> Cond;
199 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
200 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
201 MadeChange |= OptimizeImpDefsBlock(MBB);
202 }
204 bool MadeChangeThisIteration = true;
205 while (MadeChangeThisIteration) {
206 MadeChangeThisIteration = TailMergeBlocks(MF);
207 MadeChangeThisIteration |= OptimizeBranches(MF);
208 if (EnableHoistCommonCode)
209 MadeChangeThisIteration |= HoistCommonCode(MF);
210 MadeChange |= MadeChangeThisIteration;
211 }
213 // See if any jump tables have become dead as the code generator
214 // did its thing.
215 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
216 if (JTI == 0) {
217 delete RS;
218 return MadeChange;
219 }
221 // Walk the function to find jump tables that are live.
222 BitVector JTIsLive(JTI->getJumpTables().size());
223 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
224 BB != E; ++BB) {
225 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
226 I != E; ++I)
227 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
228 MachineOperand &Op = I->getOperand(op);
229 if (!Op.isJTI()) continue;
231 // Remember that this JT is live.
232 JTIsLive.set(Op.getIndex());
233 }
234 }
236 // Finally, remove dead jump tables. This happens when the
237 // indirect jump was unreachable (and thus deleted).
238 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
239 if (!JTIsLive.test(i)) {
240 JTI->RemoveJumpTable(i);
241 MadeChange = true;
242 }
244 delete RS;
245 return MadeChange;
246 }
248 //===----------------------------------------------------------------------===//
249 // Tail Merging of Blocks
250 //===----------------------------------------------------------------------===//
252 /// HashMachineInstr - Compute a hash value for MI and its operands.
253 static unsigned HashMachineInstr(const MachineInstr *MI) {
254 unsigned Hash = MI->getOpcode();
255 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
256 const MachineOperand &Op = MI->getOperand(i);
258 // Merge in bits from the operand if easy.
259 unsigned OperandHash = 0;
260 switch (Op.getType()) {
261 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
262 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
263 case MachineOperand::MO_MachineBasicBlock:
264 OperandHash = Op.getMBB()->getNumber();
265 break;
266 case MachineOperand::MO_FrameIndex:
267 case MachineOperand::MO_ConstantPoolIndex:
268 case MachineOperand::MO_JumpTableIndex:
269 OperandHash = Op.getIndex();
270 break;
271 case MachineOperand::MO_GlobalAddress:
272 case MachineOperand::MO_ExternalSymbol:
273 // Global address / external symbol are too hard, don't bother, but do
274 // pull in the offset.
275 OperandHash = Op.getOffset();
276 break;
277 default: break;
278 }
280 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
281 }
282 return Hash;
283 }
285 /// HashEndOfMBB - Hash the last instruction in the MBB.
286 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
287 MachineBasicBlock::const_iterator I = MBB->end();
288 if (I == MBB->begin())
289 return 0; // Empty MBB.
291 --I;
292 // Skip debug info so it will not affect codegen.
293 while (I->isDebugValue()) {
294 if (I==MBB->begin())
295 return 0; // MBB empty except for debug info.
296 --I;
297 }
299 return HashMachineInstr(I);
300 }
302 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
303 /// of instructions they actually have in common together at their end. Return
304 /// iterators for the first shared instruction in each block.
305 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
306 MachineBasicBlock *MBB2,
307 MachineBasicBlock::iterator &I1,
308 MachineBasicBlock::iterator &I2) {
309 I1 = MBB1->end();
310 I2 = MBB2->end();
312 unsigned TailLen = 0;
313 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
314 --I1; --I2;
315 // Skip debugging pseudos; necessary to avoid changing the code.
316 while (I1->isDebugValue()) {
317 if (I1==MBB1->begin()) {
318 while (I2->isDebugValue()) {
319 if (I2==MBB2->begin())
320 // I1==DBG at begin; I2==DBG at begin
321 return TailLen;
322 --I2;
323 }
324 ++I2;
325 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
326 return TailLen;
327 }
328 --I1;
329 }
330 // I1==first (untested) non-DBG preceding known match
331 while (I2->isDebugValue()) {
332 if (I2==MBB2->begin()) {
333 ++I1;
334 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
335 return TailLen;
336 }
337 --I2;
338 }
339 // I1, I2==first (untested) non-DBGs preceding known match
340 if (!I1->isIdenticalTo(I2) ||
341 // FIXME: This check is dubious. It's used to get around a problem where
342 // people incorrectly expect inline asm directives to remain in the same
343 // relative order. This is untenable because normal compiler
344 // optimizations (like this one) may reorder and/or merge these
345 // directives.
346 I1->isInlineAsm()) {
347 ++I1; ++I2;
348 break;
349 }
350 ++TailLen;
351 }
352 // Back past possible debugging pseudos at beginning of block. This matters
353 // when one block differs from the other only by whether debugging pseudos
354 // are present at the beginning. (This way, the various checks later for
355 // I1==MBB1->begin() work as expected.)
356 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
357 --I2;
358 while (I2->isDebugValue()) {
359 if (I2 == MBB2->begin())
360 return TailLen;
361 --I2;
362 }
363 ++I2;
364 }
365 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
366 --I1;
367 while (I1->isDebugValue()) {
368 if (I1 == MBB1->begin())
369 return TailLen;
370 --I1;
371 }
372 ++I1;
373 }
374 return TailLen;
375 }
377 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
378 MachineBasicBlock *NewMBB) {
379 if (RS) {
380 RS->enterBasicBlock(CurMBB);
381 if (!CurMBB->empty())
382 RS->forward(prior(CurMBB->end()));
383 BitVector RegsLiveAtExit(TRI->getNumRegs());
384 RS->getRegsUsed(RegsLiveAtExit, false);
385 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
386 if (RegsLiveAtExit[i])
387 NewMBB->addLiveIn(i);
388 }
389 }
391 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
392 /// after it, replacing it with an unconditional branch to NewDest.
393 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
394 MachineBasicBlock *NewDest) {
395 MachineBasicBlock *CurMBB = OldInst->getParent();
397 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
399 // For targets that use the register scavenger, we must maintain LiveIns.
400 MaintainLiveIns(CurMBB, NewDest);
402 ++NumTailMerge;
403 }
405 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
406 /// MBB so that the part before the iterator falls into the part starting at the
407 /// iterator. This returns the new MBB.
408 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
409 MachineBasicBlock::iterator BBI1,
410 const BasicBlock *BB) {
411 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
412 return 0;
414 MachineFunction &MF = *CurMBB.getParent();
416 // Create the fall-through block.
417 MachineFunction::iterator MBBI = &CurMBB;
418 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
419 CurMBB.getParent()->insert(++MBBI, NewMBB);
421 // Move all the successors of this block to the specified block.
422 NewMBB->transferSuccessors(&CurMBB);
424 // Add an edge from CurMBB to NewMBB for the fall-through.
425 CurMBB.addSuccessor(NewMBB);
427 // Splice the code over.
428 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
430 // For targets that use the register scavenger, we must maintain LiveIns.
431 MaintainLiveIns(&CurMBB, NewMBB);
433 return NewMBB;
434 }
436 /// EstimateRuntime - Make a rough estimate for how long it will take to run
437 /// the specified code.
438 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
439 MachineBasicBlock::iterator E) {
440 unsigned Time = 0;
441 for (; I != E; ++I) {
442 if (I->isDebugValue())
443 continue;
444 if (I->isCall())
445 Time += 10;
446 else if (I->mayLoad() || I->mayStore())
447 Time += 2;
448 else
449 ++Time;
450 }
451 return Time;
452 }
454 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
455 // branches temporarily for tail merging). In the case where CurMBB ends
456 // with a conditional branch to the next block, optimize by reversing the
457 // test and conditionally branching to SuccMBB instead.
458 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
459 const TargetInstrInfo *TII) {
460 MachineFunction *MF = CurMBB->getParent();
461 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB));
462 MachineBasicBlock *TBB = 0, *FBB = 0;
463 SmallVector<MachineOperand, 4> Cond;
464 DebugLoc dl; // FIXME: this is nowhere
465 if (I != MF->end() &&
466 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
467 MachineBasicBlock *NextBB = I;
468 if (TBB == NextBB && !Cond.empty() && !FBB) {
469 if (!TII->ReverseBranchCondition(Cond)) {
470 TII->RemoveBranch(*CurMBB);
471 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl);
472 return;
473 }
474 }
475 }
476 TII->InsertBranch(*CurMBB, SuccBB, NULL,
477 SmallVector<MachineOperand, 0>(), dl);
478 }
480 bool
481 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
482 if (getHash() < o.getHash())
483 return true;
484 if (getHash() > o.getHash())
485 return false;
486 if (getBlock()->getNumber() < o.getBlock()->getNumber())
487 return true;
488 if (getBlock()->getNumber() > o.getBlock()->getNumber())
489 return false;
490 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
491 // an object with itself.
492 #ifndef _GLIBCXX_DEBUG
493 llvm_unreachable("Predecessor appears twice");
494 #else
495 return false;
496 #endif
497 }
499 /// CountTerminators - Count the number of terminators in the given
500 /// block and set I to the position of the first non-terminator, if there
501 /// is one, or MBB->end() otherwise.
502 static unsigned CountTerminators(MachineBasicBlock *MBB,
503 MachineBasicBlock::iterator &I) {
504 I = MBB->end();
505 unsigned NumTerms = 0;
506 for (;;) {
507 if (I == MBB->begin()) {
508 I = MBB->end();
509 break;
510 }
511 --I;
512 if (!I->isTerminator()) break;
513 ++NumTerms;
514 }
515 return NumTerms;
516 }
518 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
519 /// and decide if it would be profitable to merge those tails. Return the
520 /// length of the common tail and iterators to the first common instruction
521 /// in each block.
522 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
523 MachineBasicBlock *MBB2,
524 unsigned minCommonTailLength,
525 unsigned &CommonTailLen,
526 MachineBasicBlock::iterator &I1,
527 MachineBasicBlock::iterator &I2,
528 MachineBasicBlock *SuccBB,
529 MachineBasicBlock *PredBB) {
530 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
531 if (CommonTailLen == 0)
532 return false;
533 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
534 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
535 << '\n');
537 // It's almost always profitable to merge any number of non-terminator
538 // instructions with the block that falls through into the common successor.
539 if (MBB1 == PredBB || MBB2 == PredBB) {
540 MachineBasicBlock::iterator I;
541 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
542 if (CommonTailLen > NumTerms)
543 return true;
544 }
546 // If one of the blocks can be completely merged and happens to be in
547 // a position where the other could fall through into it, merge any number
548 // of instructions, because it can be done without a branch.
549 // TODO: If the blocks are not adjacent, move one of them so that they are?
550 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
551 return true;
552 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
553 return true;
555 // If both blocks have an unconditional branch temporarily stripped out,
556 // count that as an additional common instruction for the following
557 // heuristics.
558 unsigned EffectiveTailLen = CommonTailLen;
559 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
560 !MBB1->back().isBarrier() &&
561 !MBB2->back().isBarrier())
562 ++EffectiveTailLen;
564 // Check if the common tail is long enough to be worthwhile.
565 if (EffectiveTailLen >= minCommonTailLength)
566 return true;
568 // If we are optimizing for code size, 2 instructions in common is enough if
569 // we don't have to split a block. At worst we will be introducing 1 new
570 // branch instruction, which is likely to be smaller than the 2
571 // instructions that would be deleted in the merge.
572 MachineFunction *MF = MBB1->getParent();
573 if (EffectiveTailLen >= 2 &&
574 MF->getFunction()->getAttributes().
575 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
576 (I1 == MBB1->begin() || I2 == MBB2->begin()))
577 return true;
579 return false;
580 }
582 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
583 /// hash CurHash (guaranteed to match the last element). Build the vector
584 /// SameTails of all those that have the (same) largest number of instructions
585 /// in common of any pair of these blocks. SameTails entries contain an
586 /// iterator into MergePotentials (from which the MachineBasicBlock can be
587 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
588 /// instruction where the matching code sequence begins.
589 /// Order of elements in SameTails is the reverse of the order in which
590 /// those blocks appear in MergePotentials (where they are not necessarily
591 /// consecutive).
592 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
593 unsigned minCommonTailLength,
594 MachineBasicBlock *SuccBB,
595 MachineBasicBlock *PredBB) {
596 unsigned maxCommonTailLength = 0U;
597 SameTails.clear();
598 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
599 MPIterator HighestMPIter = prior(MergePotentials.end());
600 for (MPIterator CurMPIter = prior(MergePotentials.end()),
601 B = MergePotentials.begin();
602 CurMPIter != B && CurMPIter->getHash() == CurHash;
603 --CurMPIter) {
604 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
605 unsigned CommonTailLen;
606 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
607 minCommonTailLength,
608 CommonTailLen, TrialBBI1, TrialBBI2,
609 SuccBB, PredBB)) {
610 if (CommonTailLen > maxCommonTailLength) {
611 SameTails.clear();
612 maxCommonTailLength = CommonTailLen;
613 HighestMPIter = CurMPIter;
614 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
615 }
616 if (HighestMPIter == CurMPIter &&
617 CommonTailLen == maxCommonTailLength)
618 SameTails.push_back(SameTailElt(I, TrialBBI2));
619 }
620 if (I == B)
621 break;
622 }
623 }
624 return maxCommonTailLength;
625 }
627 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
628 /// MergePotentials, restoring branches at ends of blocks as appropriate.
629 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
630 MachineBasicBlock *SuccBB,
631 MachineBasicBlock *PredBB) {
632 MPIterator CurMPIter, B;
633 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
634 CurMPIter->getHash() == CurHash;
635 --CurMPIter) {
636 // Put the unconditional branch back, if we need one.
637 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
638 if (SuccBB && CurMBB != PredBB)
639 FixTail(CurMBB, SuccBB, TII);
640 if (CurMPIter == B)
641 break;
642 }
643 if (CurMPIter->getHash() != CurHash)
644 CurMPIter++;
645 MergePotentials.erase(CurMPIter, MergePotentials.end());
646 }
648 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
649 /// only of the common tail. Create a block that does by splitting one.
650 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
651 MachineBasicBlock *SuccBB,
652 unsigned maxCommonTailLength,
653 unsigned &commonTailIndex) {
654 commonTailIndex = 0;
655 unsigned TimeEstimate = ~0U;
656 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
657 // Use PredBB if possible; that doesn't require a new branch.
658 if (SameTails[i].getBlock() == PredBB) {
659 commonTailIndex = i;
660 break;
661 }
662 // Otherwise, make a (fairly bogus) choice based on estimate of
663 // how long it will take the various blocks to execute.
664 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
665 SameTails[i].getTailStartPos());
666 if (t <= TimeEstimate) {
667 TimeEstimate = t;
668 commonTailIndex = i;
669 }
670 }
672 MachineBasicBlock::iterator BBI =
673 SameTails[commonTailIndex].getTailStartPos();
674 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
676 // If the common tail includes any debug info we will take it pretty
677 // randomly from one of the inputs. Might be better to remove it?
678 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
679 << maxCommonTailLength);
681 // If the split block unconditionally falls-thru to SuccBB, it will be
682 // merged. In control flow terms it should then take SuccBB's name. e.g. If
683 // SuccBB is an inner loop, the common tail is still part of the inner loop.
684 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
685 SuccBB->getBasicBlock() : MBB->getBasicBlock();
686 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
687 if (!newMBB) {
688 DEBUG(dbgs() << "... failed!");
689 return false;
690 }
692 SameTails[commonTailIndex].setBlock(newMBB);
693 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
695 // If we split PredBB, newMBB is the new predecessor.
696 if (PredBB == MBB)
697 PredBB = newMBB;
699 return true;
700 }
702 // See if any of the blocks in MergePotentials (which all have a common single
703 // successor, or all have no successor) can be tail-merged. If there is a
704 // successor, any blocks in MergePotentials that are not tail-merged and
705 // are not immediately before Succ must have an unconditional branch to
706 // Succ added (but the predecessor/successor lists need no adjustment).
707 // The lone predecessor of Succ that falls through into Succ,
708 // if any, is given in PredBB.
710 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
711 MachineBasicBlock *PredBB) {
712 bool MadeChange = false;
714 // Except for the special cases below, tail-merge if there are at least
715 // this many instructions in common.
716 unsigned minCommonTailLength = TailMergeSize;
718 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
719 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
720 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
721 << (i == e-1 ? "" : ", ");
722 dbgs() << "\n";
723 if (SuccBB) {
724 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
725 if (PredBB)
726 dbgs() << " which has fall-through from BB#"
727 << PredBB->getNumber() << "\n";
728 }
729 dbgs() << "Looking for common tails of at least "
730 << minCommonTailLength << " instruction"
731 << (minCommonTailLength == 1 ? "" : "s") << '\n';
732 );
734 // Sort by hash value so that blocks with identical end sequences sort
735 // together.
736 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
738 // Walk through equivalence sets looking for actual exact matches.
739 while (MergePotentials.size() > 1) {
740 unsigned CurHash = MergePotentials.back().getHash();
742 // Build SameTails, identifying the set of blocks with this hash code
743 // and with the maximum number of instructions in common.
744 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
745 minCommonTailLength,
746 SuccBB, PredBB);
748 // If we didn't find any pair that has at least minCommonTailLength
749 // instructions in common, remove all blocks with this hash code and retry.
750 if (SameTails.empty()) {
751 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
752 continue;
753 }
755 // If one of the blocks is the entire common tail (and not the entry
756 // block, which we can't jump to), we can treat all blocks with this same
757 // tail at once. Use PredBB if that is one of the possibilities, as that
758 // will not introduce any extra branches.
759 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
760 getParent()->begin();
761 unsigned commonTailIndex = SameTails.size();
762 // If there are two blocks, check to see if one can be made to fall through
763 // into the other.
764 if (SameTails.size() == 2 &&
765 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
766 SameTails[1].tailIsWholeBlock())
767 commonTailIndex = 1;
768 else if (SameTails.size() == 2 &&
769 SameTails[1].getBlock()->isLayoutSuccessor(
770 SameTails[0].getBlock()) &&
771 SameTails[0].tailIsWholeBlock())
772 commonTailIndex = 0;
773 else {
774 // Otherwise just pick one, favoring the fall-through predecessor if
775 // there is one.
776 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
777 MachineBasicBlock *MBB = SameTails[i].getBlock();
778 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
779 continue;
780 if (MBB == PredBB) {
781 commonTailIndex = i;
782 break;
783 }
784 if (SameTails[i].tailIsWholeBlock())
785 commonTailIndex = i;
786 }
787 }
789 if (commonTailIndex == SameTails.size() ||
790 (SameTails[commonTailIndex].getBlock() == PredBB &&
791 !SameTails[commonTailIndex].tailIsWholeBlock())) {
792 // None of the blocks consist entirely of the common tail.
793 // Split a block so that one does.
794 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
795 maxCommonTailLength, commonTailIndex)) {
796 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
797 continue;
798 }
799 }
801 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
802 // MBB is common tail. Adjust all other BB's to jump to this one.
803 // Traversal must be forwards so erases work.
804 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
805 << " for ");
806 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
807 if (commonTailIndex == i)
808 continue;
809 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
810 << (i == e-1 ? "" : ", "));
811 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
812 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
813 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
814 MergePotentials.erase(SameTails[i].getMPIter());
815 }
816 DEBUG(dbgs() << "\n");
817 // We leave commonTailIndex in the worklist in case there are other blocks
818 // that match it with a smaller number of instructions.
819 MadeChange = true;
820 }
821 return MadeChange;
822 }
824 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
825 bool MadeChange = false;
826 if (!EnableTailMerge) return MadeChange;
828 // First find blocks with no successors.
829 MergePotentials.clear();
830 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
831 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
832 if (TriedMerging.count(I))
833 continue;
834 if (I->succ_empty())
835 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
836 }
838 // If this is a large problem, avoid visiting the same basic blocks
839 // multiple times.
840 if (MergePotentials.size() == TailMergeThreshold)
841 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
842 TriedMerging.insert(MergePotentials[i].getBlock());
844 // See if we can do any tail merging on those.
845 if (MergePotentials.size() >= 2)
846 MadeChange |= TryTailMergeBlocks(NULL, NULL);
848 // Look at blocks (IBB) with multiple predecessors (PBB).
849 // We change each predecessor to a canonical form, by
850 // (1) temporarily removing any unconditional branch from the predecessor
851 // to IBB, and
852 // (2) alter conditional branches so they branch to the other block
853 // not IBB; this may require adding back an unconditional branch to IBB
854 // later, where there wasn't one coming in. E.g.
855 // Bcc IBB
856 // fallthrough to QBB
857 // here becomes
858 // Bncc QBB
859 // with a conceptual B to IBB after that, which never actually exists.
860 // With those changes, we see whether the predecessors' tails match,
861 // and merge them if so. We change things out of canonical form and
862 // back to the way they were later in the process. (OptimizeBranches
863 // would undo some of this, but we can't use it, because we'd get into
864 // a compile-time infinite loop repeatedly doing and undoing the same
865 // transformations.)
867 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
868 I != E; ++I) {
869 if (I->pred_size() < 2) continue;
870 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
871 MachineBasicBlock *IBB = I;
872 MachineBasicBlock *PredBB = prior(I);
873 MergePotentials.clear();
874 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
875 E2 = I->pred_end();
876 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
877 MachineBasicBlock *PBB = *P;
878 if (TriedMerging.count(PBB))
879 continue;
881 // Skip blocks that loop to themselves, can't tail merge these.
882 if (PBB == IBB)
883 continue;
885 // Visit each predecessor only once.
886 if (!UniquePreds.insert(PBB))
887 continue;
889 // Skip blocks which may jump to a landing pad. Can't tail merge these.
890 if (PBB->getLandingPadSuccessor())
891 continue;
893 MachineBasicBlock *TBB = 0, *FBB = 0;
894 SmallVector<MachineOperand, 4> Cond;
895 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
896 // Failing case: IBB is the target of a cbr, and we cannot reverse the
897 // branch.
898 SmallVector<MachineOperand, 4> NewCond(Cond);
899 if (!Cond.empty() && TBB == IBB) {
900 if (TII->ReverseBranchCondition(NewCond))
901 continue;
902 // This is the QBB case described above
903 if (!FBB)
904 FBB = llvm::next(MachineFunction::iterator(PBB));
905 }
907 // Failing case: the only way IBB can be reached from PBB is via
908 // exception handling. Happens for landing pads. Would be nice to have
909 // a bit in the edge so we didn't have to do all this.
910 if (IBB->isLandingPad()) {
911 MachineFunction::iterator IP = PBB; IP++;
912 MachineBasicBlock *PredNextBB = NULL;
913 if (IP != MF.end())
914 PredNextBB = IP;
915 if (TBB == NULL) {
916 if (IBB != PredNextBB) // fallthrough
917 continue;
918 } else if (FBB) {
919 if (TBB != IBB && FBB != IBB) // cbr then ubr
920 continue;
921 } else if (Cond.empty()) {
922 if (TBB != IBB) // ubr
923 continue;
924 } else {
925 if (TBB != IBB && IBB != PredNextBB) // cbr
926 continue;
927 }
928 }
930 // Remove the unconditional branch at the end, if any.
931 if (TBB && (Cond.empty() || FBB)) {
932 DebugLoc dl; // FIXME: this is nowhere
933 TII->RemoveBranch(*PBB);
934 if (!Cond.empty())
935 // reinsert conditional branch only, for now
936 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl);
937 }
939 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
940 }
941 }
943 // If this is a large problem, avoid visiting the same basic blocks multiple
944 // times.
945 if (MergePotentials.size() == TailMergeThreshold)
946 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
947 TriedMerging.insert(MergePotentials[i].getBlock());
949 if (MergePotentials.size() >= 2)
950 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
952 // Reinsert an unconditional branch if needed. The 1 below can occur as a
953 // result of removing blocks in TryTailMergeBlocks.
954 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
955 if (MergePotentials.size() == 1 &&
956 MergePotentials.begin()->getBlock() != PredBB)
957 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
958 }
960 return MadeChange;
961 }
963 //===----------------------------------------------------------------------===//
964 // Branch Optimization
965 //===----------------------------------------------------------------------===//
967 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
968 bool MadeChange = false;
970 // Make sure blocks are numbered in order
971 MF.RenumberBlocks();
973 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
974 I != E; ) {
975 MachineBasicBlock *MBB = I++;
976 MadeChange |= OptimizeBlock(MBB);
978 // If it is dead, remove it.
979 if (MBB->pred_empty()) {
980 RemoveDeadBlock(MBB);
981 MadeChange = true;
982 ++NumDeadBlocks;
983 }
984 }
985 return MadeChange;
986 }
988 // Blocks should be considered empty if they contain only debug info;
989 // else the debug info would affect codegen.
990 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
991 if (MBB->empty())
992 return true;
993 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
994 MBBI!=MBBE; ++MBBI) {
995 if (!MBBI->isDebugValue())
996 return false;
997 }
998 return true;
999 }
1001 // Blocks with only debug info and branches should be considered the same
1002 // as blocks with only branches.
1003 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1004 MachineBasicBlock::iterator MBBI, MBBE;
1005 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
1006 if (!MBBI->isDebugValue())
1007 break;
1008 }
1009 return (MBBI->isBranch());
1010 }
1012 /// IsBetterFallthrough - Return true if it would be clearly better to
1013 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1014 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1015 /// result in infinite loops.
1016 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1017 MachineBasicBlock *MBB2) {
1018 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1019 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1020 // optimize branches that branch to either a return block or an assert block
1021 // into a fallthrough to the return.
1022 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
1024 // If there is a clear successor ordering we make sure that one block
1025 // will fall through to the next
1026 if (MBB1->isSuccessor(MBB2)) return true;
1027 if (MBB2->isSuccessor(MBB1)) return false;
1029 // Neither block consists entirely of debug info (per IsEmptyBlock check),
1030 // so we needn't test for falling off the beginning here.
1031 MachineBasicBlock::iterator MBB1I = --MBB1->end();
1032 while (MBB1I->isDebugValue())
1033 --MBB1I;
1034 MachineBasicBlock::iterator MBB2I = --MBB2->end();
1035 while (MBB2I->isDebugValue())
1036 --MBB2I;
1037 return MBB2I->isCall() && !MBB1I->isCall();
1038 }
1040 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1041 /// instructions on the block. Always use the DebugLoc of the first
1042 /// branching instruction found unless its absent, in which case use the
1043 /// DebugLoc of the second if present.
1044 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1045 MachineBasicBlock::iterator I = MBB.end();
1046 if (I == MBB.begin())
1047 return DebugLoc();
1048 --I;
1049 while (I->isDebugValue() && I != MBB.begin())
1050 --I;
1051 if (I->isBranch())
1052 return I->getDebugLoc();
1053 return DebugLoc();
1054 }
1056 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1057 /// block. This is never called on the entry block.
1058 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1059 bool MadeChange = false;
1060 MachineFunction &MF = *MBB->getParent();
1061 ReoptimizeBlock:
1063 MachineFunction::iterator FallThrough = MBB;
1064 ++FallThrough;
1066 // If this block is empty, make everyone use its fall-through, not the block
1067 // explicitly. Landing pads should not do this since the landing-pad table
1068 // points to this block. Blocks with their addresses taken shouldn't be
1069 // optimized away.
1070 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1071 // Dead block? Leave for cleanup later.
1072 if (MBB->pred_empty()) return MadeChange;
1074 if (FallThrough == MF.end()) {
1075 // TODO: Simplify preds to not branch here if possible!
1076 } else {
1077 // Rewrite all predecessors of the old block to go to the fallthrough
1078 // instead.
1079 while (!MBB->pred_empty()) {
1080 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1081 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1082 }
1083 // If MBB was the target of a jump table, update jump tables to go to the
1084 // fallthrough instead.
1085 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1086 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1087 MadeChange = true;
1088 }
1089 return MadeChange;
1090 }
1092 // Check to see if we can simplify the terminator of the block before this
1093 // one.
1094 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1096 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1097 SmallVector<MachineOperand, 4> PriorCond;
1098 bool PriorUnAnalyzable =
1099 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1100 if (!PriorUnAnalyzable) {
1101 // If the CFG for the prior block has extra edges, remove them.
1102 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1103 !PriorCond.empty());
1105 // If the previous branch is conditional and both conditions go to the same
1106 // destination, remove the branch, replacing it with an unconditional one or
1107 // a fall-through.
1108 if (PriorTBB && PriorTBB == PriorFBB) {
1109 DebugLoc dl = getBranchDebugLoc(PrevBB);
1110 TII->RemoveBranch(PrevBB);
1111 PriorCond.clear();
1112 if (PriorTBB != MBB)
1113 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1114 MadeChange = true;
1115 ++NumBranchOpts;
1116 goto ReoptimizeBlock;
1117 }
1119 // If the previous block unconditionally falls through to this block and
1120 // this block has no other predecessors, move the contents of this block
1121 // into the prior block. This doesn't usually happen when SimplifyCFG
1122 // has been used, but it can happen if tail merging splits a fall-through
1123 // predecessor of a block.
1124 // This has to check PrevBB->succ_size() because EH edges are ignored by
1125 // AnalyzeBranch.
1126 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1127 PrevBB.succ_size() == 1 &&
1128 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1129 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1130 << "From MBB: " << *MBB);
1131 // Remove redundant DBG_VALUEs first.
1132 if (PrevBB.begin() != PrevBB.end()) {
1133 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1134 --PrevBBIter;
1135 MachineBasicBlock::iterator MBBIter = MBB->begin();
1136 // Check if DBG_VALUE at the end of PrevBB is identical to the
1137 // DBG_VALUE at the beginning of MBB.
1138 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1139 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1140 if (!MBBIter->isIdenticalTo(PrevBBIter))
1141 break;
1142 MachineInstr *DuplicateDbg = MBBIter;
1143 ++MBBIter; -- PrevBBIter;
1144 DuplicateDbg->eraseFromParent();
1145 }
1146 }
1147 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1148 PrevBB.removeSuccessor(PrevBB.succ_begin());
1149 assert(PrevBB.succ_empty());
1150 PrevBB.transferSuccessors(MBB);
1151 MadeChange = true;
1152 return MadeChange;
1153 }
1155 // If the previous branch *only* branches to *this* block (conditional or
1156 // not) remove the branch.
1157 if (PriorTBB == MBB && PriorFBB == 0) {
1158 TII->RemoveBranch(PrevBB);
1159 MadeChange = true;
1160 ++NumBranchOpts;
1161 goto ReoptimizeBlock;
1162 }
1164 // If the prior block branches somewhere else on the condition and here if
1165 // the condition is false, remove the uncond second branch.
1166 if (PriorFBB == MBB) {
1167 DebugLoc dl = getBranchDebugLoc(PrevBB);
1168 TII->RemoveBranch(PrevBB);
1169 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1170 MadeChange = true;
1171 ++NumBranchOpts;
1172 goto ReoptimizeBlock;
1173 }
1175 // If the prior block branches here on true and somewhere else on false, and
1176 // if the branch condition is reversible, reverse the branch to create a
1177 // fall-through.
1178 if (PriorTBB == MBB) {
1179 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1180 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1181 DebugLoc dl = getBranchDebugLoc(PrevBB);
1182 TII->RemoveBranch(PrevBB);
1183 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl);
1184 MadeChange = true;
1185 ++NumBranchOpts;
1186 goto ReoptimizeBlock;
1187 }
1188 }
1190 // If this block has no successors (e.g. it is a return block or ends with
1191 // a call to a no-return function like abort or __cxa_throw) and if the pred
1192 // falls through into this block, and if it would otherwise fall through
1193 // into the block after this, move this block to the end of the function.
1194 //
1195 // We consider it more likely that execution will stay in the function (e.g.
1196 // due to loops) than it is to exit it. This asserts in loops etc, moving
1197 // the assert condition out of the loop body.
1198 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1199 MachineFunction::iterator(PriorTBB) == FallThrough &&
1200 !MBB->canFallThrough()) {
1201 bool DoTransform = true;
1203 // We have to be careful that the succs of PredBB aren't both no-successor
1204 // blocks. If neither have successors and if PredBB is the second from
1205 // last block in the function, we'd just keep swapping the two blocks for
1206 // last. Only do the swap if one is clearly better to fall through than
1207 // the other.
1208 if (FallThrough == --MF.end() &&
1209 !IsBetterFallthrough(PriorTBB, MBB))
1210 DoTransform = false;
1212 if (DoTransform) {
1213 // Reverse the branch so we will fall through on the previous true cond.
1214 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1215 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1216 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1217 << "To make fallthrough to: " << *PriorTBB << "\n");
1219 DebugLoc dl = getBranchDebugLoc(PrevBB);
1220 TII->RemoveBranch(PrevBB);
1221 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl);
1223 // Move this block to the end of the function.
1224 MBB->moveAfter(--MF.end());
1225 MadeChange = true;
1226 ++NumBranchOpts;
1227 return MadeChange;
1228 }
1229 }
1230 }
1231 }
1233 // Analyze the branch in the current block.
1234 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1235 SmallVector<MachineOperand, 4> CurCond;
1236 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1237 if (!CurUnAnalyzable) {
1238 // If the CFG for the prior block has extra edges, remove them.
1239 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1241 // If this is a two-way branch, and the FBB branches to this block, reverse
1242 // the condition so the single-basic-block loop is faster. Instead of:
1243 // Loop: xxx; jcc Out; jmp Loop
1244 // we want:
1245 // Loop: xxx; jncc Loop; jmp Out
1246 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1247 SmallVector<MachineOperand, 4> NewCond(CurCond);
1248 if (!TII->ReverseBranchCondition(NewCond)) {
1249 DebugLoc dl = getBranchDebugLoc(*MBB);
1250 TII->RemoveBranch(*MBB);
1251 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1252 MadeChange = true;
1253 ++NumBranchOpts;
1254 goto ReoptimizeBlock;
1255 }
1256 }
1258 // If this branch is the only thing in its block, see if we can forward
1259 // other blocks across it.
1260 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1261 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1262 !MBB->hasAddressTaken()) {
1263 DebugLoc dl = getBranchDebugLoc(*MBB);
1264 // This block may contain just an unconditional branch. Because there can
1265 // be 'non-branch terminators' in the block, try removing the branch and
1266 // then seeing if the block is empty.
1267 TII->RemoveBranch(*MBB);
1268 // If the only things remaining in the block are debug info, remove these
1269 // as well, so this will behave the same as an empty block in non-debug
1270 // mode.
1271 if (!MBB->empty()) {
1272 bool NonDebugInfoFound = false;
1273 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1274 I != E; ++I) {
1275 if (!I->isDebugValue()) {
1276 NonDebugInfoFound = true;
1277 break;
1278 }
1279 }
1280 if (!NonDebugInfoFound)
1281 // Make the block empty, losing the debug info (we could probably
1282 // improve this in some cases.)
1283 MBB->erase(MBB->begin(), MBB->end());
1284 }
1285 // If this block is just an unconditional branch to CurTBB, we can
1286 // usually completely eliminate the block. The only case we cannot
1287 // completely eliminate the block is when the block before this one
1288 // falls through into MBB and we can't understand the prior block's branch
1289 // condition.
1290 if (MBB->empty()) {
1291 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1292 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1293 !PrevBB.isSuccessor(MBB)) {
1294 // If the prior block falls through into us, turn it into an
1295 // explicit branch to us to make updates simpler.
1296 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1297 PriorTBB != MBB && PriorFBB != MBB) {
1298 if (PriorTBB == 0) {
1299 assert(PriorCond.empty() && PriorFBB == 0 &&
1300 "Bad branch analysis");
1301 PriorTBB = MBB;
1302 } else {
1303 assert(PriorFBB == 0 && "Machine CFG out of date!");
1304 PriorFBB = MBB;
1305 }
1306 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1307 TII->RemoveBranch(PrevBB);
1308 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1309 }
1311 // Iterate through all the predecessors, revectoring each in-turn.
1312 size_t PI = 0;
1313 bool DidChange = false;
1314 bool HasBranchToSelf = false;
1315 while(PI != MBB->pred_size()) {
1316 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1317 if (PMBB == MBB) {
1318 // If this block has an uncond branch to itself, leave it.
1319 ++PI;
1320 HasBranchToSelf = true;
1321 } else {
1322 DidChange = true;
1323 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1324 // If this change resulted in PMBB ending in a conditional
1325 // branch where both conditions go to the same destination,
1326 // change this to an unconditional branch (and fix the CFG).
1327 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1328 SmallVector<MachineOperand, 4> NewCurCond;
1329 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1330 NewCurFBB, NewCurCond, true);
1331 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1332 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1333 TII->RemoveBranch(*PMBB);
1334 NewCurCond.clear();
1335 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, pdl);
1336 MadeChange = true;
1337 ++NumBranchOpts;
1338 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1339 }
1340 }
1341 }
1343 // Change any jumptables to go to the new MBB.
1344 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1345 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1346 if (DidChange) {
1347 ++NumBranchOpts;
1348 MadeChange = true;
1349 if (!HasBranchToSelf) return MadeChange;
1350 }
1351 }
1352 }
1354 // Add the branch back if the block is more than just an uncond branch.
1355 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl);
1356 }
1357 }
1359 // If the prior block doesn't fall through into this block, and if this
1360 // block doesn't fall through into some other block, see if we can find a
1361 // place to move this block where a fall-through will happen.
1362 if (!PrevBB.canFallThrough()) {
1364 // Now we know that there was no fall-through into this block, check to
1365 // see if it has a fall-through into its successor.
1366 bool CurFallsThru = MBB->canFallThrough();
1368 if (!MBB->isLandingPad()) {
1369 // Check all the predecessors of this block. If one of them has no fall
1370 // throughs, move this block right after it.
1371 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1372 E = MBB->pred_end(); PI != E; ++PI) {
1373 // Analyze the branch at the end of the pred.
1374 MachineBasicBlock *PredBB = *PI;
1375 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1376 MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
1377 SmallVector<MachineOperand, 4> PredCond;
1378 if (PredBB != MBB && !PredBB->canFallThrough() &&
1379 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1380 && (!CurFallsThru || !CurTBB || !CurFBB)
1381 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1382 // If the current block doesn't fall through, just move it.
1383 // If the current block can fall through and does not end with a
1384 // conditional branch, we need to append an unconditional jump to
1385 // the (current) next block. To avoid a possible compile-time
1386 // infinite loop, move blocks only backward in this case.
1387 // Also, if there are already 2 branches here, we cannot add a third;
1388 // this means we have the case
1389 // Bcc next
1390 // B elsewhere
1391 // next:
1392 if (CurFallsThru) {
1393 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
1394 CurCond.clear();
1395 TII->InsertBranch(*MBB, NextBB, 0, CurCond, DebugLoc());
1396 }
1397 MBB->moveAfter(PredBB);
1398 MadeChange = true;
1399 goto ReoptimizeBlock;
1400 }
1401 }
1402 }
1404 if (!CurFallsThru) {
1405 // Check all successors to see if we can move this block before it.
1406 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1407 E = MBB->succ_end(); SI != E; ++SI) {
1408 // Analyze the branch at the end of the block before the succ.
1409 MachineBasicBlock *SuccBB = *SI;
1410 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1412 // If this block doesn't already fall-through to that successor, and if
1413 // the succ doesn't already have a block that can fall through into it,
1414 // and if the successor isn't an EH destination, we can arrange for the
1415 // fallthrough to happen.
1416 if (SuccBB != MBB && &*SuccPrev != MBB &&
1417 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1418 !SuccBB->isLandingPad()) {
1419 MBB->moveBefore(SuccBB);
1420 MadeChange = true;
1421 goto ReoptimizeBlock;
1422 }
1423 }
1425 // Okay, there is no really great place to put this block. If, however,
1426 // the block before this one would be a fall-through if this block were
1427 // removed, move this block to the end of the function.
1428 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
1429 SmallVector<MachineOperand, 4> PrevCond;
1430 if (FallThrough != MF.end() &&
1431 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1432 PrevBB.isSuccessor(FallThrough)) {
1433 MBB->moveAfter(--MF.end());
1434 MadeChange = true;
1435 return MadeChange;
1436 }
1437 }
1438 }
1440 return MadeChange;
1441 }
1443 //===----------------------------------------------------------------------===//
1444 // Hoist Common Code
1445 //===----------------------------------------------------------------------===//
1447 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1448 /// blocks to their common predecessor.
1449 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1450 bool MadeChange = false;
1451 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1452 MachineBasicBlock *MBB = I++;
1453 MadeChange |= HoistCommonCodeInSuccs(MBB);
1454 }
1456 return MadeChange;
1457 }
1459 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1460 /// its 'true' successor.
1461 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1462 MachineBasicBlock *TrueBB) {
1463 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1464 E = BB->succ_end(); SI != E; ++SI) {
1465 MachineBasicBlock *SuccBB = *SI;
1466 if (SuccBB != TrueBB)
1467 return SuccBB;
1468 }
1469 return NULL;
1470 }
1472 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1473 /// in successors to. The location is usually just before the terminator,
1474 /// however if the terminator is a conditional branch and its previous
1475 /// instruction is the flag setting instruction, the previous instruction is
1476 /// the preferred location. This function also gathers uses and defs of the
1477 /// instructions from the insertion point to the end of the block. The data is
1478 /// used by HoistCommonCodeInSuccs to ensure safety.
1479 static
1480 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1481 const TargetInstrInfo *TII,
1482 const TargetRegisterInfo *TRI,
1483 SmallSet<unsigned,4> &Uses,
1484 SmallSet<unsigned,4> &Defs) {
1485 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1486 if (!TII->isUnpredicatedTerminator(Loc))
1487 return MBB->end();
1489 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1490 const MachineOperand &MO = Loc->getOperand(i);
1491 if (!MO.isReg())
1492 continue;
1493 unsigned Reg = MO.getReg();
1494 if (!Reg)
1495 continue;
1496 if (MO.isUse()) {
1497 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1498 Uses.insert(*AI);
1499 } else if (!MO.isDead())
1500 // Don't try to hoist code in the rare case the terminator defines a
1501 // register that is later used.
1502 return MBB->end();
1503 }
1505 if (Uses.empty())
1506 return Loc;
1507 if (Loc == MBB->begin())
1508 return MBB->end();
1510 // The terminator is probably a conditional branch, try not to separate the
1511 // branch from condition setting instruction.
1512 MachineBasicBlock::iterator PI = Loc;
1513 --PI;
1514 while (PI != MBB->begin() && Loc->isDebugValue())
1515 --PI;
1517 bool IsDef = false;
1518 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1519 const MachineOperand &MO = PI->getOperand(i);
1520 // If PI has a regmask operand, it is probably a call. Separate away.
1521 if (MO.isRegMask())
1522 return Loc;
1523 if (!MO.isReg() || MO.isUse())
1524 continue;
1525 unsigned Reg = MO.getReg();
1526 if (!Reg)
1527 continue;
1528 if (Uses.count(Reg))
1529 IsDef = true;
1530 }
1531 if (!IsDef)
1532 // The condition setting instruction is not just before the conditional
1533 // branch.
1534 return Loc;
1536 // Be conservative, don't insert instruction above something that may have
1537 // side-effects. And since it's potentially bad to separate flag setting
1538 // instruction from the conditional branch, just abort the optimization
1539 // completely.
1540 // Also avoid moving code above predicated instruction since it's hard to
1541 // reason about register liveness with predicated instruction.
1542 bool DontMoveAcrossStore = true;
1543 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) ||
1544 TII->isPredicated(PI))
1545 return MBB->end();
1548 // Find out what registers are live. Note this routine is ignoring other live
1549 // registers which are only used by instructions in successor blocks.
1550 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1551 const MachineOperand &MO = PI->getOperand(i);
1552 if (!MO.isReg())
1553 continue;
1554 unsigned Reg = MO.getReg();
1555 if (!Reg)
1556 continue;
1557 if (MO.isUse()) {
1558 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1559 Uses.insert(*AI);
1560 } else {
1561 if (Uses.erase(Reg)) {
1562 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1563 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1564 }
1565 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1566 Defs.insert(*AI);
1567 }
1568 }
1570 return PI;
1571 }
1573 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1574 /// sequence at the start of the function, move the instructions before MBB
1575 /// terminator if it's legal.
1576 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1577 MachineBasicBlock *TBB = 0, *FBB = 0;
1578 SmallVector<MachineOperand, 4> Cond;
1579 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1580 return false;
1582 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1583 if (!FBB)
1584 // Malformed bcc? True and false blocks are the same?
1585 return false;
1587 // Restrict the optimization to cases where MBB is the only predecessor,
1588 // it is an obvious win.
1589 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1590 return false;
1592 // Find a suitable position to hoist the common instructions to. Also figure
1593 // out which registers are used or defined by instructions from the insertion
1594 // point to the end of the block.
1595 SmallSet<unsigned, 4> Uses, Defs;
1596 MachineBasicBlock::iterator Loc =
1597 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1598 if (Loc == MBB->end())
1599 return false;
1601 bool HasDups = false;
1602 SmallVector<unsigned, 4> LocalDefs;
1603 SmallSet<unsigned, 4> LocalDefsSet;
1604 MachineBasicBlock::iterator TIB = TBB->begin();
1605 MachineBasicBlock::iterator FIB = FBB->begin();
1606 MachineBasicBlock::iterator TIE = TBB->end();
1607 MachineBasicBlock::iterator FIE = FBB->end();
1608 while (TIB != TIE && FIB != FIE) {
1609 // Skip dbg_value instructions. These do not count.
1610 if (TIB->isDebugValue()) {
1611 while (TIB != TIE && TIB->isDebugValue())
1612 ++TIB;
1613 if (TIB == TIE)
1614 break;
1615 }
1616 if (FIB->isDebugValue()) {
1617 while (FIB != FIE && FIB->isDebugValue())
1618 ++FIB;
1619 if (FIB == FIE)
1620 break;
1621 }
1622 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1623 break;
1625 if (TII->isPredicated(TIB))
1626 // Hard to reason about register liveness with predicated instruction.
1627 break;
1629 bool IsSafe = true;
1630 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1631 MachineOperand &MO = TIB->getOperand(i);
1632 // Don't attempt to hoist instructions with register masks.
1633 if (MO.isRegMask()) {
1634 IsSafe = false;
1635 break;
1636 }
1637 if (!MO.isReg())
1638 continue;
1639 unsigned Reg = MO.getReg();
1640 if (!Reg)
1641 continue;
1642 if (MO.isDef()) {
1643 if (Uses.count(Reg)) {
1644 // Avoid clobbering a register that's used by the instruction at
1645 // the point of insertion.
1646 IsSafe = false;
1647 break;
1648 }
1650 if (Defs.count(Reg) && !MO.isDead()) {
1651 // Don't hoist the instruction if the def would be clobber by the
1652 // instruction at the point insertion. FIXME: This is overly
1653 // conservative. It should be possible to hoist the instructions
1654 // in BB2 in the following example:
1655 // BB1:
1656 // r1, eflag = op1 r2, r3
1657 // brcc eflag
1658 //
1659 // BB2:
1660 // r1 = op2, ...
1661 // = op3, r1<kill>
1662 IsSafe = false;
1663 break;
1664 }
1665 } else if (!LocalDefsSet.count(Reg)) {
1666 if (Defs.count(Reg)) {
1667 // Use is defined by the instruction at the point of insertion.
1668 IsSafe = false;
1669 break;
1670 }
1672 if (MO.isKill() && Uses.count(Reg))
1673 // Kills a register that's read by the instruction at the point of
1674 // insertion. Remove the kill marker.
1675 MO.setIsKill(false);
1676 }
1677 }
1678 if (!IsSafe)
1679 break;
1681 bool DontMoveAcrossStore = true;
1682 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore))
1683 break;
1685 // Remove kills from LocalDefsSet, these registers had short live ranges.
1686 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1687 MachineOperand &MO = TIB->getOperand(i);
1688 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1689 continue;
1690 unsigned Reg = MO.getReg();
1691 if (!Reg || !LocalDefsSet.count(Reg))
1692 continue;
1693 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1694 LocalDefsSet.erase(*AI);
1695 }
1697 // Track local defs so we can update liveins.
1698 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1699 MachineOperand &MO = TIB->getOperand(i);
1700 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1701 continue;
1702 unsigned Reg = MO.getReg();
1703 if (!Reg)
1704 continue;
1705 LocalDefs.push_back(Reg);
1706 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1707 LocalDefsSet.insert(*AI);
1708 }
1710 HasDups = true;
1711 ++TIB;
1712 ++FIB;
1713 }
1715 if (!HasDups)
1716 return false;
1718 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1719 FBB->erase(FBB->begin(), FIB);
1721 // Update livein's.
1722 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1723 unsigned Def = LocalDefs[i];
1724 if (LocalDefsSet.count(Def)) {
1725 TBB->addLiveIn(Def);
1726 FBB->addLiveIn(Def);
1727 }
1728 }
1730 ++NumHoist;
1731 return true;
1732 }