1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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
14 // Note that this pass must be run after register allocation, it cannot handle
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "branchfolding"
20 #include "llvm/CodeGen/Passes.h"
21 #include "llvm/CodeGen/MachineModuleInfo.h"
22 #include "llvm/CodeGen/MachineFunctionPass.h"
23 #include "llvm/CodeGen/MachineJumpTableInfo.h"
24 #include "llvm/CodeGen/RegisterScavenging.h"
25 #include "llvm/Target/TargetInstrInfo.h"
26 #include "llvm/Target/TargetMachine.h"
27 #include "llvm/Target/MRegisterInfo.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/ADT/STLExtras.h"
35 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
36 STATISTIC(NumBranchOpts, "Number of branches optimized");
37 STATISTIC(NumTailMerge , "Number of block tails merged");
38 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
39 cl::init(cl::BOU_UNSET), cl::Hidden);
41 // Throttle for huge numbers of predecessors (compile speed problems)
43 TailMergeThreshold("tail-merge-threshold",
44 cl::desc("Max number of predecessors to consider tail merging"),
45 cl::init(100), cl::Hidden);
47 struct BranchFolder : public MachineFunctionPass {
49 explicit BranchFolder(bool defaultEnableTailMerge) :
50 MachineFunctionPass((intptr_t)&ID) {
51 switch (FlagEnableTailMerge) {
52 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
53 case cl::BOU_TRUE: EnableTailMerge = true; break;
54 case cl::BOU_FALSE: EnableTailMerge = false; break;
58 virtual bool runOnMachineFunction(MachineFunction &MF);
59 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
60 const TargetInstrInfo *TII;
61 MachineModuleInfo *MMI;
66 bool TailMergeBlocks(MachineFunction &MF);
67 bool TryMergeBlocks(MachineBasicBlock* SuccBB,
68 MachineBasicBlock* PredBB);
69 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
70 MachineBasicBlock *NewDest);
71 MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB,
72 MachineBasicBlock::iterator BBI1);
74 std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials;
75 const MRegisterInfo *RegInfo;
78 bool OptimizeBranches(MachineFunction &MF);
79 void OptimizeBlock(MachineBasicBlock *MBB);
80 void RemoveDeadBlock(MachineBasicBlock *MBB);
82 bool CanFallThrough(MachineBasicBlock *CurBB);
83 bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable,
84 MachineBasicBlock *TBB, MachineBasicBlock *FBB,
85 const std::vector<MachineOperand> &Cond);
87 char BranchFolder::ID = 0;
90 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
91 return new BranchFolder(DefaultEnableTailMerge); }
93 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
94 /// function, updating the CFG.
95 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
96 assert(MBB->pred_empty() && "MBB must be dead!");
97 DOUT << "\nRemoving MBB: " << *MBB;
99 MachineFunction *MF = MBB->getParent();
100 // drop all successors.
101 while (!MBB->succ_empty())
102 MBB->removeSuccessor(MBB->succ_end()-1);
104 // If there is DWARF info to active, check to see if there are any LABEL
105 // records in the basic block. If so, unregister them from MachineModuleInfo.
106 if (MMI && !MBB->empty()) {
107 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
109 if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) {
110 // The label ID # is always operand #0, an immediate.
111 MMI->InvalidateLabel(I->getOperand(0).getImm());
117 MF->getBasicBlockList().erase(MBB);
120 bool BranchFolder::runOnMachineFunction(MachineFunction &MF) {
121 TII = MF.getTarget().getInstrInfo();
122 if (!TII) return false;
124 // Fix CFG. The later algorithms expect it to be right.
125 bool EverMadeChange = false;
126 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
127 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
128 std::vector<MachineOperand> Cond;
129 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond))
130 EverMadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
133 RegInfo = MF.getTarget().getRegisterInfo();
134 RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
136 MMI = getAnalysisToUpdate<MachineModuleInfo>();
138 bool MadeChangeThisIteration = true;
139 while (MadeChangeThisIteration) {
140 MadeChangeThisIteration = false;
141 MadeChangeThisIteration |= TailMergeBlocks(MF);
142 MadeChangeThisIteration |= OptimizeBranches(MF);
143 EverMadeChange |= MadeChangeThisIteration;
146 // See if any jump tables have become mergable or dead as the code generator
148 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
149 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
151 // Figure out how these jump tables should be merged.
152 std::vector<unsigned> JTMapping;
153 JTMapping.reserve(JTs.size());
155 // We always keep the 0th jump table.
156 JTMapping.push_back(0);
158 // Scan the jump tables, seeing if there are any duplicates. Note that this
159 // is N^2, which should be fixed someday.
160 for (unsigned i = 1, e = JTs.size(); i != e; ++i)
161 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
163 // If a jump table was merge with another one, walk the function rewriting
164 // references to jump tables to reference the new JT ID's. Keep track of
165 // whether we see a jump table idx, if not, we can delete the JT.
166 std::vector<bool> JTIsLive;
167 JTIsLive.resize(JTs.size());
168 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
170 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
172 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
173 MachineOperand &Op = I->getOperand(op);
174 if (!Op.isJumpTableIndex()) continue;
175 unsigned NewIdx = JTMapping[Op.getJumpTableIndex()];
176 Op.setJumpTableIndex(NewIdx);
178 // Remember that this JT is live.
179 JTIsLive[NewIdx] = true;
183 // Finally, remove dead jump tables. This happens either because the
184 // indirect jump was unreachable (and thus deleted) or because the jump
185 // table was merged with some other one.
186 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
188 JTI->RemoveJumpTable(i);
189 EverMadeChange = true;
194 return EverMadeChange;
197 //===----------------------------------------------------------------------===//
198 // Tail Merging of Blocks
199 //===----------------------------------------------------------------------===//
201 /// HashMachineInstr - Compute a hash value for MI and its operands.
202 static unsigned HashMachineInstr(const MachineInstr *MI) {
203 unsigned Hash = MI->getOpcode();
204 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
205 const MachineOperand &Op = MI->getOperand(i);
207 // Merge in bits from the operand if easy.
208 unsigned OperandHash = 0;
209 switch (Op.getType()) {
210 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
211 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
212 case MachineOperand::MO_MachineBasicBlock:
213 OperandHash = Op.getMachineBasicBlock()->getNumber();
215 case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break;
216 case MachineOperand::MO_ConstantPoolIndex:
217 OperandHash = Op.getConstantPoolIndex();
219 case MachineOperand::MO_JumpTableIndex:
220 OperandHash = Op.getJumpTableIndex();
222 case MachineOperand::MO_GlobalAddress:
223 case MachineOperand::MO_ExternalSymbol:
224 // Global address / external symbol are too hard, don't bother, but do
225 // pull in the offset.
226 OperandHash = Op.getOffset();
231 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
236 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
237 /// with no successors, we hash two instructions, because cross-jumping
238 /// only saves code when at least two instructions are removed (since a
239 /// branch must be inserted). For blocks with a successor, one of the
240 /// two blocks to be tail-merged will end with a branch already, so
241 /// it gains to cross-jump even for one instruction.
243 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
244 unsigned minCommonTailLength) {
245 MachineBasicBlock::const_iterator I = MBB->end();
246 if (I == MBB->begin())
247 return 0; // Empty MBB.
250 unsigned Hash = HashMachineInstr(I);
252 if (I == MBB->begin() || minCommonTailLength == 1)
253 return Hash; // Single instr MBB.
256 // Hash in the second-to-last instruction.
257 Hash ^= HashMachineInstr(I) << 2;
261 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
262 /// of instructions they actually have in common together at their end. Return
263 /// iterators for the first shared instruction in each block.
264 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
265 MachineBasicBlock *MBB2,
266 MachineBasicBlock::iterator &I1,
267 MachineBasicBlock::iterator &I2) {
271 unsigned TailLen = 0;
272 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
274 if (!I1->isIdenticalTo(I2) ||
275 // FIXME: This check is dubious. It's used to get around a problem where
276 // people incorrectly expect inline asm directives to remain in the same
277 // relative order. This is untenable because normal compiler
278 // optimizations (like this one) may reorder and/or merge these
280 I1->getOpcode() == TargetInstrInfo::INLINEASM) {
289 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
290 /// after it, replacing it with an unconditional branch to NewDest. This
291 /// returns true if OldInst's block is modified, false if NewDest is modified.
292 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
293 MachineBasicBlock *NewDest) {
294 MachineBasicBlock *OldBB = OldInst->getParent();
296 // Remove all the old successors of OldBB from the CFG.
297 while (!OldBB->succ_empty())
298 OldBB->removeSuccessor(OldBB->succ_begin());
300 // Remove all the dead instructions from the end of OldBB.
301 OldBB->erase(OldInst, OldBB->end());
303 // If OldBB isn't immediately before OldBB, insert a branch to it.
304 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
305 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>());
306 OldBB->addSuccessor(NewDest);
310 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
311 /// MBB so that the part before the iterator falls into the part starting at the
312 /// iterator. This returns the new MBB.
313 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
314 MachineBasicBlock::iterator BBI1) {
315 // Create the fall-through block.
316 MachineFunction::iterator MBBI = &CurMBB;
317 MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock());
318 CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB);
320 // Move all the successors of this block to the specified block.
321 while (!CurMBB.succ_empty()) {
322 MachineBasicBlock *S = *(CurMBB.succ_end()-1);
323 NewMBB->addSuccessor(S);
324 CurMBB.removeSuccessor(S);
327 // Add an edge from CurMBB to NewMBB for the fall-through.
328 CurMBB.addSuccessor(NewMBB);
330 // Splice the code over.
331 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
333 // For targets that use the register scavenger, we must maintain LiveIns.
335 RS->enterBasicBlock(&CurMBB);
337 RS->forward(prior(CurMBB.end()));
338 BitVector RegsLiveAtExit(RegInfo->getNumRegs());
339 RS->getRegsUsed(RegsLiveAtExit, false);
340 for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++)
341 if (RegsLiveAtExit[i])
342 NewMBB->addLiveIn(i);
348 /// EstimateRuntime - Make a rough estimate for how long it will take to run
349 /// the specified code.
350 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
351 MachineBasicBlock::iterator E,
352 const TargetInstrInfo *TII) {
354 for (; I != E; ++I) {
355 const TargetInstrDescriptor &TID = TII->get(I->getOpcode());
356 if (TID.Flags & M_CALL_FLAG)
358 else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG))
366 /// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at
367 /// MBB2I and then insert an unconditional branch in the other block. Determine
368 /// which is the best to split
369 static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1,
370 MachineBasicBlock::iterator MBB1I,
371 MachineBasicBlock *MBB2,
372 MachineBasicBlock::iterator MBB2I,
373 const TargetInstrInfo *TII,
374 MachineBasicBlock *PredBB) {
375 // If one block is the entry block, split the other one; we can't generate
376 // a branch to the entry block, as its label is not emitted.
377 MachineBasicBlock *Entry = MBB1->getParent()->begin();
383 // If one block falls through into the common successor, choose that
384 // one to split; it is one instruction less to do that.
388 else if (MBB2 == PredBB)
391 // TODO: if we had some notion of which block was hotter, we could split
392 // the hot block, so it is the fall-through. Since we don't have profile info
393 // make a decision based on which will hurt most to split.
394 unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII);
395 unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII);
397 // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the
398 // MBB1 block so it falls through. This will penalize the MBB2 path, but will
399 // have a lower overall impact on the program execution.
400 return MBB1Time < MBB2Time;
403 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
404 // branches temporarily for tail merging). In the case where CurMBB ends
405 // with a conditional branch to the next block, optimize by reversing the
406 // test and conditionally branching to SuccMBB instead.
408 static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB,
409 const TargetInstrInfo *TII) {
410 MachineFunction *MF = CurMBB->getParent();
411 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
412 MachineBasicBlock *TBB = 0, *FBB = 0;
413 std::vector<MachineOperand> Cond;
414 if (I != MF->end() &&
415 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) {
416 MachineBasicBlock *NextBB = I;
417 if (TBB == NextBB && Cond.size() && !FBB) {
418 if (!TII->ReverseBranchCondition(Cond)) {
419 TII->RemoveBranch(*CurMBB);
420 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
425 TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>());
428 static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p,
429 const std::pair<unsigned,MachineBasicBlock*> &q) {
430 if (p.first < q.first)
432 else if (p.first > q.first)
434 else if (p.second->getNumber() < q.second->getNumber())
436 else if (p.second->getNumber() > q.second->getNumber())
439 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
440 // an object with itself.
441 #ifndef _GLIBCXX_DEBUG
442 assert(0 && "Predecessor appears twice");
448 // See if any of the blocks in MergePotentials (which all have a common single
449 // successor, or all have no successor) can be tail-merged. If there is a
450 // successor, any blocks in MergePotentials that are not tail-merged and
451 // are not immediately before Succ must have an unconditional branch to
452 // Succ added (but the predecessor/successor lists need no adjustment).
453 // The lone predecessor of Succ that falls through into Succ,
454 // if any, is given in PredBB.
456 bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB,
457 MachineBasicBlock* PredBB) {
458 unsigned minCommonTailLength = (SuccBB ? 1 : 2);
461 // Sort by hash value so that blocks with identical end sequences sort
463 std::stable_sort(MergePotentials.begin(), MergePotentials.end(), MergeCompare);
465 // Walk through equivalence sets looking for actual exact matches.
466 while (MergePotentials.size() > 1) {
467 unsigned CurHash = (MergePotentials.end()-1)->first;
468 unsigned PrevHash = (MergePotentials.end()-2)->first;
469 MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second;
471 // If there is nothing that matches the hash of the current basic block,
473 if (CurHash != PrevHash) {
474 if (SuccBB && CurMBB != PredBB)
475 FixTail(CurMBB, SuccBB, TII);
476 MergePotentials.pop_back();
480 // Look through all the pairs of blocks that have the same hash as this
481 // one, and find the pair that has the largest number of instructions in
483 // Since instructions may get combined later (e.g. single stores into
484 // store multiple) this measure is not particularly accurate.
485 MachineBasicBlock::iterator BBI1, BBI2;
487 unsigned FoundI = ~0U, FoundJ = ~0U;
488 unsigned maxCommonTailLength = 0U;
489 for (int i = MergePotentials.size()-1;
490 i != -1 && MergePotentials[i].first == CurHash; --i) {
492 j != -1 && MergePotentials[j].first == CurHash; --j) {
493 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
494 unsigned CommonTailLen = ComputeCommonTailLength(
495 MergePotentials[i].second,
496 MergePotentials[j].second,
497 TrialBBI1, TrialBBI2);
498 if (CommonTailLen >= minCommonTailLength &&
499 CommonTailLen > maxCommonTailLength) {
502 maxCommonTailLength = CommonTailLen;
509 // If we didn't find any pair that has at least minCommonTailLength
510 // instructions in common, bail out. All entries with this
511 // hash code can go away now.
513 for (int i = MergePotentials.size()-1;
514 i != -1 && MergePotentials[i].first == CurHash; --i) {
515 // Put the unconditional branch back, if we need one.
516 CurMBB = MergePotentials[i].second;
517 if (SuccBB && CurMBB != PredBB)
518 FixTail(CurMBB, SuccBB, TII);
519 MergePotentials.pop_back();
524 // Otherwise, move the block(s) to the right position(s). So that
525 // BBI1/2 will be valid, the last must be I and the next-to-last J.
526 if (FoundI != MergePotentials.size()-1)
527 std::swap(MergePotentials[FoundI], *(MergePotentials.end()-1));
528 if (FoundJ != MergePotentials.size()-2)
529 std::swap(MergePotentials[FoundJ], *(MergePotentials.end()-2));
531 CurMBB = (MergePotentials.end()-1)->second;
532 MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second;
534 // If neither block is the entire common tail, split the tail of one block
535 // to make it redundant with the other tail. Also, we cannot jump to the
536 // entry block, so if one block is the entry block, split the other one.
537 MachineBasicBlock *Entry = CurMBB->getParent()->begin();
538 if (CurMBB->begin() == BBI1 && CurMBB != Entry)
539 ; // CurMBB is common tail
540 else if (MBB2->begin() == BBI2 && MBB2 != Entry)
541 ; // MBB2 is common tail
543 if (0) { // Enable this to disable partial tail merges.
544 MergePotentials.pop_back();
548 // Decide whether we want to split CurMBB or MBB2.
549 if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII, PredBB)) {
550 CurMBB = SplitMBBAt(*CurMBB, BBI1);
551 BBI1 = CurMBB->begin();
552 MergePotentials.back().second = CurMBB;
554 MBB2 = SplitMBBAt(*MBB2, BBI2);
555 BBI2 = MBB2->begin();
556 (MergePotentials.end()-2)->second = MBB2;
560 if (MBB2->begin() == BBI2 && MBB2 != Entry) {
561 // Hack the end off CurMBB, making it jump to MBBI@ instead.
562 ReplaceTailWithBranchTo(BBI1, MBB2);
563 // This modifies CurMBB, so remove it from the worklist.
564 MergePotentials.pop_back();
566 assert(CurMBB->begin() == BBI1 && CurMBB != Entry &&
567 "Didn't split block correctly?");
568 // Hack the end off MBB2, making it jump to CurMBB instead.
569 ReplaceTailWithBranchTo(BBI2, CurMBB);
570 // This modifies MBB2, so remove it from the worklist.
571 MergePotentials.erase(MergePotentials.end()-2);
578 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
580 if (!EnableTailMerge) return false;
584 // First find blocks with no successors.
585 MergePotentials.clear();
586 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
588 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I));
590 // See if we can do any tail merging on those.
591 if (MergePotentials.size() < TailMergeThreshold)
592 MadeChange |= TryMergeBlocks(NULL, NULL);
594 // Look at blocks (IBB) with multiple predecessors (PBB).
595 // We change each predecessor to a canonical form, by
596 // (1) temporarily removing any unconditional branch from the predecessor
598 // (2) alter conditional branches so they branch to the other block
599 // not IBB; this may require adding back an unconditional branch to IBB
600 // later, where there wasn't one coming in. E.g.
602 // fallthrough to QBB
605 // with a conceptual B to IBB after that, which never actually exists.
606 // With those changes, we see whether the predecessors' tails match,
607 // and merge them if so. We change things out of canonical form and
608 // back to the way they were later in the process. (OptimizeBranches
609 // would undo some of this, but we can't use it, because we'd get into
610 // a compile-time infinite loop repeatedly doing and undoing the same
613 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
614 if (!I->succ_empty() && I->pred_size() >= 2 &&
615 I->pred_size() < TailMergeThreshold) {
616 MachineBasicBlock *IBB = I;
617 MachineBasicBlock *PredBB = prior(I);
618 MergePotentials.clear();
619 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
622 MachineBasicBlock* PBB = *P;
623 // Skip blocks that loop to themselves, can't tail merge these.
626 MachineBasicBlock *TBB = 0, *FBB = 0;
627 std::vector<MachineOperand> Cond;
628 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) {
629 // Failing case: IBB is the target of a cbr, and
630 // we cannot reverse the branch.
631 std::vector<MachineOperand> NewCond(Cond);
632 if (Cond.size() && TBB==IBB) {
633 if (TII->ReverseBranchCondition(NewCond))
635 // This is the QBB case described above
637 FBB = next(MachineFunction::iterator(PBB));
639 // Failing case: the only way IBB can be reached from PBB is via
640 // exception handling. Happens for landing pads. Would be nice
641 // to have a bit in the edge so we didn't have to do all this.
642 if (IBB->isLandingPad()) {
643 MachineFunction::iterator IP = PBB; IP++;
644 MachineBasicBlock* PredNextBB = NULL;
648 if (IBB!=PredNextBB) // fallthrough
651 if (TBB!=IBB && FBB!=IBB) // cbr then ubr
653 } else if (Cond.size() == 0) {
657 if (TBB!=IBB && IBB!=PredNextBB) // cbr
661 // Remove the unconditional branch at the end, if any.
662 if (TBB && (Cond.size()==0 || FBB)) {
663 TII->RemoveBranch(*PBB);
665 // reinsert conditional branch only, for now
666 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond);
668 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P));
671 if (MergePotentials.size() >= 2)
672 MadeChange |= TryMergeBlocks(I, PredBB);
673 // Reinsert an unconditional branch if needed.
674 // The 1 below can be either an original single predecessor, or a result
675 // of removing blocks in TryMergeBlocks.
676 PredBB = prior(I); // this may have been changed in TryMergeBlocks
677 if (MergePotentials.size()==1 &&
678 (MergePotentials.begin())->second != PredBB)
679 FixTail((MergePotentials.begin())->second, I, TII);
685 //===----------------------------------------------------------------------===//
686 // Branch Optimization
687 //===----------------------------------------------------------------------===//
689 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
692 // Make sure blocks are numbered in order
695 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
696 MachineBasicBlock *MBB = I++;
699 // If it is dead, remove it.
700 if (MBB->pred_empty()) {
701 RemoveDeadBlock(MBB);
710 /// CanFallThrough - Return true if the specified block (with the specified
711 /// branch condition) can implicitly transfer control to the block after it by
712 /// falling off the end of it. This should return false if it can reach the
713 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
715 /// True is a conservative answer.
717 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
718 bool BranchUnAnalyzable,
719 MachineBasicBlock *TBB, MachineBasicBlock *FBB,
720 const std::vector<MachineOperand> &Cond) {
721 MachineFunction::iterator Fallthrough = CurBB;
723 // If FallthroughBlock is off the end of the function, it can't fall through.
724 if (Fallthrough == CurBB->getParent()->end())
727 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
728 if (!CurBB->isSuccessor(Fallthrough))
731 // If we couldn't analyze the branch, assume it could fall through.
732 if (BranchUnAnalyzable) return true;
734 // If there is no branch, control always falls through.
735 if (TBB == 0) return true;
737 // If there is some explicit branch to the fallthrough block, it can obviously
738 // reach, even though the branch should get folded to fall through implicitly.
739 if (MachineFunction::iterator(TBB) == Fallthrough ||
740 MachineFunction::iterator(FBB) == Fallthrough)
743 // If it's an unconditional branch to some block not the fall through, it
744 // doesn't fall through.
745 if (Cond.empty()) return false;
747 // Otherwise, if it is conditional and has no explicit false block, it falls
752 /// CanFallThrough - Return true if the specified can implicitly transfer
753 /// control to the block after it by falling off the end of it. This should
754 /// return false if it can reach the block after it, but it uses an explicit
755 /// branch to do so (e.g. a table jump).
757 /// True is a conservative answer.
759 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
760 MachineBasicBlock *TBB = 0, *FBB = 0;
761 std::vector<MachineOperand> Cond;
762 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond);
763 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
766 /// IsBetterFallthrough - Return true if it would be clearly better to
767 /// fall-through to MBB1 than to fall through into MBB2. This has to return
768 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
769 /// result in infinite loops.
770 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
771 MachineBasicBlock *MBB2,
772 const TargetInstrInfo &TII) {
773 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
774 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
775 // optimize branches that branch to either a return block or an assert block
776 // into a fallthrough to the return.
777 if (MBB1->empty() || MBB2->empty()) return false;
779 MachineInstr *MBB1I = --MBB1->end();
780 MachineInstr *MBB2I = --MBB2->end();
781 return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode());
784 /// OptimizeBlock - Analyze and optimize control flow related to the specified
785 /// block. This is never called on the entry block.
786 void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
787 MachineFunction::iterator FallThrough = MBB;
790 // If this block is empty, make everyone use its fall-through, not the block
791 // explicitly. Landing pads should not do this since the landing-pad table
792 // points to this block.
793 if (MBB->empty() && !MBB->isLandingPad()) {
794 // Dead block? Leave for cleanup later.
795 if (MBB->pred_empty()) return;
797 if (FallThrough == MBB->getParent()->end()) {
798 // TODO: Simplify preds to not branch here if possible!
800 // Rewrite all predecessors of the old block to go to the fallthrough
802 while (!MBB->pred_empty()) {
803 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
804 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
807 // If MBB was the target of a jump table, update jump tables to go to the
808 // fallthrough instead.
809 MBB->getParent()->getJumpTableInfo()->
810 ReplaceMBBInJumpTables(MBB, FallThrough);
816 // Check to see if we can simplify the terminator of the block before this
818 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
820 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
821 std::vector<MachineOperand> PriorCond;
822 bool PriorUnAnalyzable =
823 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
824 if (!PriorUnAnalyzable) {
825 // If the CFG for the prior block has extra edges, remove them.
826 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
829 // If the previous branch is conditional and both conditions go to the same
830 // destination, remove the branch, replacing it with an unconditional one or
832 if (PriorTBB && PriorTBB == PriorFBB) {
833 TII->RemoveBranch(PrevBB);
836 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
839 return OptimizeBlock(MBB);
842 // If the previous branch *only* branches to *this* block (conditional or
843 // not) remove the branch.
844 if (PriorTBB == MBB && PriorFBB == 0) {
845 TII->RemoveBranch(PrevBB);
848 return OptimizeBlock(MBB);
851 // If the prior block branches somewhere else on the condition and here if
852 // the condition is false, remove the uncond second branch.
853 if (PriorFBB == MBB) {
854 TII->RemoveBranch(PrevBB);
855 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
858 return OptimizeBlock(MBB);
861 // If the prior block branches here on true and somewhere else on false, and
862 // if the branch condition is reversible, reverse the branch to create a
864 if (PriorTBB == MBB) {
865 std::vector<MachineOperand> NewPriorCond(PriorCond);
866 if (!TII->ReverseBranchCondition(NewPriorCond)) {
867 TII->RemoveBranch(PrevBB);
868 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
871 return OptimizeBlock(MBB);
875 // If this block doesn't fall through (e.g. it ends with an uncond branch or
876 // has no successors) and if the pred falls through into this block, and if
877 // it would otherwise fall through into the block after this, move this
878 // block to the end of the function.
880 // We consider it more likely that execution will stay in the function (e.g.
881 // due to loops) than it is to exit it. This asserts in loops etc, moving
882 // the assert condition out of the loop body.
883 if (!PriorCond.empty() && PriorFBB == 0 &&
884 MachineFunction::iterator(PriorTBB) == FallThrough &&
885 !CanFallThrough(MBB)) {
886 bool DoTransform = true;
888 // We have to be careful that the succs of PredBB aren't both no-successor
889 // blocks. If neither have successors and if PredBB is the second from
890 // last block in the function, we'd just keep swapping the two blocks for
891 // last. Only do the swap if one is clearly better to fall through than
893 if (FallThrough == --MBB->getParent()->end() &&
894 !IsBetterFallthrough(PriorTBB, MBB, *TII))
897 // We don't want to do this transformation if we have control flow like:
906 // In this case, we could actually be moving the return block *into* a
908 if (DoTransform && !MBB->succ_empty() &&
909 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
914 // Reverse the branch so we will fall through on the previous true cond.
915 std::vector<MachineOperand> NewPriorCond(PriorCond);
916 if (!TII->ReverseBranchCondition(NewPriorCond)) {
917 DOUT << "\nMoving MBB: " << *MBB;
918 DOUT << "To make fallthrough to: " << *PriorTBB << "\n";
920 TII->RemoveBranch(PrevBB);
921 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
923 // Move this block to the end of the function.
924 MBB->moveAfter(--MBB->getParent()->end());
933 // Analyze the branch in the current block.
934 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
935 std::vector<MachineOperand> CurCond;
936 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond);
937 if (!CurUnAnalyzable) {
938 // If the CFG for the prior block has extra edges, remove them.
939 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
941 // If this is a two-way branch, and the FBB branches to this block, reverse
942 // the condition so the single-basic-block loop is faster. Instead of:
943 // Loop: xxx; jcc Out; jmp Loop
945 // Loop: xxx; jncc Loop; jmp Out
946 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
947 std::vector<MachineOperand> NewCond(CurCond);
948 if (!TII->ReverseBranchCondition(NewCond)) {
949 TII->RemoveBranch(*MBB);
950 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
953 return OptimizeBlock(MBB);
958 // If this branch is the only thing in its block, see if we can forward
959 // other blocks across it.
960 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
961 TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) {
962 // This block may contain just an unconditional branch. Because there can
963 // be 'non-branch terminators' in the block, try removing the branch and
964 // then seeing if the block is empty.
965 TII->RemoveBranch(*MBB);
967 // If this block is just an unconditional branch to CurTBB, we can
968 // usually completely eliminate the block. The only case we cannot
969 // completely eliminate the block is when the block before this one
970 // falls through into MBB and we can't understand the prior block's branch
973 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
974 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
975 !PrevBB.isSuccessor(MBB)) {
976 // If the prior block falls through into us, turn it into an
977 // explicit branch to us to make updates simpler.
978 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
979 PriorTBB != MBB && PriorFBB != MBB) {
981 assert(PriorCond.empty() && PriorFBB == 0 &&
982 "Bad branch analysis");
985 assert(PriorFBB == 0 && "Machine CFG out of date!");
988 TII->RemoveBranch(PrevBB);
989 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
992 // Iterate through all the predecessors, revectoring each in-turn.
994 bool DidChange = false;
995 bool HasBranchToSelf = false;
996 while(PI != MBB->pred_size()) {
997 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
999 // If this block has an uncond branch to itself, leave it.
1001 HasBranchToSelf = true;
1004 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1008 // Change any jumptables to go to the new MBB.
1009 MBB->getParent()->getJumpTableInfo()->
1010 ReplaceMBBInJumpTables(MBB, CurTBB);
1014 if (!HasBranchToSelf) return;
1019 // Add the branch back if the block is more than just an uncond branch.
1020 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1024 // If the prior block doesn't fall through into this block, and if this
1025 // block doesn't fall through into some other block, see if we can find a
1026 // place to move this block where a fall-through will happen.
1027 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
1028 PriorTBB, PriorFBB, PriorCond)) {
1029 // Now we know that there was no fall-through into this block, check to
1030 // see if it has a fall-through into its successor.
1031 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1034 if (!MBB->isLandingPad()) {
1035 // Check all the predecessors of this block. If one of them has no fall
1036 // throughs, move this block right after it.
1037 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1038 E = MBB->pred_end(); PI != E; ++PI) {
1039 // Analyze the branch at the end of the pred.
1040 MachineBasicBlock *PredBB = *PI;
1041 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1042 if (PredBB != MBB && !CanFallThrough(PredBB)
1043 && (!CurFallsThru || !CurTBB || !CurFBB)
1044 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1045 // If the current block doesn't fall through, just move it.
1046 // If the current block can fall through and does not end with a
1047 // conditional branch, we need to append an unconditional jump to
1048 // the (current) next block. To avoid a possible compile-time
1049 // infinite loop, move blocks only backward in this case.
1050 // Also, if there are already 2 branches here, we cannot add a third;
1051 // this means we have the case
1056 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1058 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1060 MBB->moveAfter(PredBB);
1062 return OptimizeBlock(MBB);
1067 if (!CurFallsThru) {
1068 // Check all successors to see if we can move this block before it.
1069 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1070 E = MBB->succ_end(); SI != E; ++SI) {
1071 // Analyze the branch at the end of the block before the succ.
1072 MachineBasicBlock *SuccBB = *SI;
1073 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1074 std::vector<MachineOperand> SuccPrevCond;
1076 // If this block doesn't already fall-through to that successor, and if
1077 // the succ doesn't already have a block that can fall through into it,
1078 // and if the successor isn't an EH destination, we can arrange for the
1079 // fallthrough to happen.
1080 if (SuccBB != MBB && !CanFallThrough(SuccPrev) &&
1081 !SuccBB->isLandingPad()) {
1082 MBB->moveBefore(SuccBB);
1084 return OptimizeBlock(MBB);
1088 // Okay, there is no really great place to put this block. If, however,
1089 // the block before this one would be a fall-through if this block were
1090 // removed, move this block to the end of the function.
1091 if (FallThrough != MBB->getParent()->end() &&
1092 PrevBB.isSuccessor(FallThrough)) {
1093 MBB->moveAfter(--MBB->getParent()->end());