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<bool> EnableTailMerge("enable-tail-merge", cl::Hidden);
41 struct BranchFolder : public MachineFunctionPass {
43 BranchFolder() : MachineFunctionPass((intptr_t)&ID) {}
45 virtual bool runOnMachineFunction(MachineFunction &MF);
46 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
47 const TargetInstrInfo *TII;
48 MachineModuleInfo *MMI;
52 bool TailMergeBlocks(MachineFunction &MF);
53 bool TryMergeBlocks(MachineBasicBlock* SuccBB,
54 MachineBasicBlock* PredBB);
55 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
56 MachineBasicBlock *NewDest);
57 MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB,
58 MachineBasicBlock::iterator BBI1);
60 std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials;
61 const MRegisterInfo *RegInfo;
64 bool OptimizeBranches(MachineFunction &MF);
65 void OptimizeBlock(MachineBasicBlock *MBB);
66 void RemoveDeadBlock(MachineBasicBlock *MBB);
68 bool CanFallThrough(MachineBasicBlock *CurBB);
69 bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable,
70 MachineBasicBlock *TBB, MachineBasicBlock *FBB,
71 const std::vector<MachineOperand> &Cond);
73 char BranchFolder::ID = 0;
76 FunctionPass *llvm::createBranchFoldingPass() { return new BranchFolder(); }
78 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
79 /// function, updating the CFG.
80 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
81 assert(MBB->pred_empty() && "MBB must be dead!");
82 DOUT << "\nRemoving MBB: " << *MBB;
84 MachineFunction *MF = MBB->getParent();
85 // drop all successors.
86 while (!MBB->succ_empty())
87 MBB->removeSuccessor(MBB->succ_end()-1);
89 // If there is DWARF info to active, check to see if there are any LABEL
90 // records in the basic block. If so, unregister them from MachineModuleInfo.
91 if (MMI && !MBB->empty()) {
92 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
94 if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) {
95 // The label ID # is always operand #0, an immediate.
96 MMI->InvalidateLabel(I->getOperand(0).getImm());
102 MF->getBasicBlockList().erase(MBB);
105 bool BranchFolder::runOnMachineFunction(MachineFunction &MF) {
106 TII = MF.getTarget().getInstrInfo();
107 if (!TII) return false;
109 RegInfo = MF.getTarget().getRegisterInfo();
110 RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
112 MMI = getAnalysisToUpdate<MachineModuleInfo>();
114 bool EverMadeChange = false;
115 bool MadeChangeThisIteration = true;
116 while (MadeChangeThisIteration) {
117 MadeChangeThisIteration = false;
118 MadeChangeThisIteration |= TailMergeBlocks(MF);
119 MadeChangeThisIteration |= OptimizeBranches(MF);
120 EverMadeChange |= MadeChangeThisIteration;
123 // See if any jump tables have become mergable or dead as the code generator
125 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
126 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
128 // Figure out how these jump tables should be merged.
129 std::vector<unsigned> JTMapping;
130 JTMapping.reserve(JTs.size());
132 // We always keep the 0th jump table.
133 JTMapping.push_back(0);
135 // Scan the jump tables, seeing if there are any duplicates. Note that this
136 // is N^2, which should be fixed someday.
137 for (unsigned i = 1, e = JTs.size(); i != e; ++i)
138 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
140 // If a jump table was merge with another one, walk the function rewriting
141 // references to jump tables to reference the new JT ID's. Keep track of
142 // whether we see a jump table idx, if not, we can delete the JT.
143 std::vector<bool> JTIsLive;
144 JTIsLive.resize(JTs.size());
145 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
147 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
149 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
150 MachineOperand &Op = I->getOperand(op);
151 if (!Op.isJumpTableIndex()) continue;
152 unsigned NewIdx = JTMapping[Op.getJumpTableIndex()];
153 Op.setJumpTableIndex(NewIdx);
155 // Remember that this JT is live.
156 JTIsLive[NewIdx] = true;
160 // Finally, remove dead jump tables. This happens either because the
161 // indirect jump was unreachable (and thus deleted) or because the jump
162 // table was merged with some other one.
163 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
165 JTI->RemoveJumpTable(i);
166 EverMadeChange = true;
171 return EverMadeChange;
174 //===----------------------------------------------------------------------===//
175 // Tail Merging of Blocks
176 //===----------------------------------------------------------------------===//
178 /// HashMachineInstr - Compute a hash value for MI and its operands.
179 static unsigned HashMachineInstr(const MachineInstr *MI) {
180 unsigned Hash = MI->getOpcode();
181 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
182 const MachineOperand &Op = MI->getOperand(i);
184 // Merge in bits from the operand if easy.
185 unsigned OperandHash = 0;
186 switch (Op.getType()) {
187 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
188 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
189 case MachineOperand::MO_MachineBasicBlock:
190 OperandHash = Op.getMachineBasicBlock()->getNumber();
192 case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break;
193 case MachineOperand::MO_ConstantPoolIndex:
194 OperandHash = Op.getConstantPoolIndex();
196 case MachineOperand::MO_JumpTableIndex:
197 OperandHash = Op.getJumpTableIndex();
199 case MachineOperand::MO_GlobalAddress:
200 case MachineOperand::MO_ExternalSymbol:
201 // Global address / external symbol are too hard, don't bother, but do
202 // pull in the offset.
203 OperandHash = Op.getOffset();
208 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
213 /// HashEndOfMBB - Hash the last two instructions in the MBB. We hash two
214 /// instructions, because cross-jumping only saves code when at least two
215 /// instructions are removed (since a branch must be inserted).
216 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
217 MachineBasicBlock::const_iterator I = MBB->end();
218 if (I == MBB->begin())
219 return 0; // Empty MBB.
222 unsigned Hash = HashMachineInstr(I);
224 if (I == MBB->begin())
225 return Hash; // Single instr MBB.
228 // Hash in the second-to-last instruction.
229 Hash ^= HashMachineInstr(I) << 2;
233 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
234 /// of instructions they actually have in common together at their end. Return
235 /// iterators for the first shared instruction in each block.
236 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
237 MachineBasicBlock *MBB2,
238 MachineBasicBlock::iterator &I1,
239 MachineBasicBlock::iterator &I2) {
243 unsigned TailLen = 0;
244 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
246 if (!I1->isIdenticalTo(I2)) {
255 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
256 /// after it, replacing it with an unconditional branch to NewDest. This
257 /// returns true if OldInst's block is modified, false if NewDest is modified.
258 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
259 MachineBasicBlock *NewDest) {
260 MachineBasicBlock *OldBB = OldInst->getParent();
262 // Remove all the old successors of OldBB from the CFG.
263 while (!OldBB->succ_empty())
264 OldBB->removeSuccessor(OldBB->succ_begin());
266 // Remove all the dead instructions from the end of OldBB.
267 OldBB->erase(OldInst, OldBB->end());
269 // If OldBB isn't immediately before OldBB, insert a branch to it.
270 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
271 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>());
272 OldBB->addSuccessor(NewDest);
276 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
277 /// MBB so that the part before the iterator falls into the part starting at the
278 /// iterator. This returns the new MBB.
279 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
280 MachineBasicBlock::iterator BBI1) {
281 // Create the fall-through block.
282 MachineFunction::iterator MBBI = &CurMBB;
283 MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock());
284 CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB);
286 // Move all the successors of this block to the specified block.
287 while (!CurMBB.succ_empty()) {
288 MachineBasicBlock *S = *(CurMBB.succ_end()-1);
289 NewMBB->addSuccessor(S);
290 CurMBB.removeSuccessor(S);
293 // Add an edge from CurMBB to NewMBB for the fall-through.
294 CurMBB.addSuccessor(NewMBB);
296 // Splice the code over.
297 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
299 // For targets that use the register scavenger, we must maintain LiveIns.
301 RS->enterBasicBlock(&CurMBB);
303 RS->forward(prior(CurMBB.end()));
304 BitVector RegsLiveAtExit(RegInfo->getNumRegs());
305 RS->getRegsUsed(RegsLiveAtExit, false);
306 for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++)
307 if (RegsLiveAtExit[i])
308 NewMBB->addLiveIn(i);
314 /// EstimateRuntime - Make a rough estimate for how long it will take to run
315 /// the specified code.
316 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
317 MachineBasicBlock::iterator E,
318 const TargetInstrInfo *TII) {
320 for (; I != E; ++I) {
321 const TargetInstrDescriptor &TID = TII->get(I->getOpcode());
322 if (TID.Flags & M_CALL_FLAG)
324 else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG))
332 /// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at
333 /// MBB2I and then insert an unconditional branch in the other block. Determine
334 /// which is the best to split
335 static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1,
336 MachineBasicBlock::iterator MBB1I,
337 MachineBasicBlock *MBB2,
338 MachineBasicBlock::iterator MBB2I,
339 const TargetInstrInfo *TII,
340 MachineBasicBlock *PredBB) {
341 // If one block falls through into the common successor, choose that
342 // one to split; it is one instruction less to do that.
346 else if (MBB2 == PredBB)
349 // TODO: if we had some notion of which block was hotter, we could split
350 // the hot block, so it is the fall-through. Since we don't have profile info
351 // make a decision based on which will hurt most to split.
352 unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII);
353 unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII);
355 // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the
356 // MBB1 block so it falls through. This will penalize the MBB2 path, but will
357 // have a lower overall impact on the program execution.
358 return MBB1Time < MBB2Time;
361 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
362 // branches temporarily for tail merging). In the case where CurMBB ends
363 // with a conditional branch to the next block, optimize by reversing the
364 // test and conditionally branching to SuccMBB instead.
366 static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB,
367 const TargetInstrInfo *TII) {
368 MachineFunction *MF = CurMBB->getParent();
369 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
370 MachineBasicBlock *TBB = 0, *FBB = 0;
371 std::vector<MachineOperand> Cond;
372 if (I != MF->end() &&
373 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) {
374 MachineBasicBlock *NextBB = I;
375 if (TBB == NextBB && Cond.size() && !FBB) {
376 if (!TII->ReverseBranchCondition(Cond)) {
377 TII->RemoveBranch(*CurMBB);
378 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
383 TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>());
386 // See if any of the blocks in MergePotentials (which all have a common single
387 // successor, or all have no successor) can be tail-merged. If there is a
388 // successor, any blocks in MergePotentials that are not tail-merged and
389 // are not immediately before Succ must have an unconditional branch to
390 // Succ added (but the predecessor/successor lists need no adjustment).
391 // The lone predecessor of Succ that falls through into Succ,
392 // if any, is given in PredBB.
394 bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB,
395 MachineBasicBlock* PredBB) {
398 // Sort by hash value so that blocks with identical end sequences sort
400 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
402 // Walk through equivalence sets looking for actual exact matches.
403 while (MergePotentials.size() > 1) {
404 unsigned CurHash = (MergePotentials.end()-1)->first;
405 unsigned PrevHash = (MergePotentials.end()-2)->first;
406 MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second;
408 // If there is nothing that matches the hash of the current basic block,
410 if (CurHash != PrevHash) {
411 if (SuccBB && CurMBB != PredBB)
412 FixTail(CurMBB, SuccBB, TII);
413 MergePotentials.pop_back();
417 // Determine the actual length of the shared tail between these two basic
418 // blocks. Because the hash can have collisions, it's possible that this is
420 MachineBasicBlock::iterator BBI1, BBI2;
421 unsigned CommonTailLen =
422 ComputeCommonTailLength(CurMBB, (MergePotentials.end()-2)->second,
425 // If the tails don't have at least two instructions in common, see if there
426 // is anything else in the equivalence class that does match.
427 // Since instructions may get combined later (e.g. single stores into
428 // store multiple) this measure is not particularly accurate.
429 if (CommonTailLen < 2) {
430 unsigned FoundMatch = ~0U;
431 for (int i = MergePotentials.size()-2;
432 i != -1 && MergePotentials[i].first == CurHash; --i) {
433 CommonTailLen = ComputeCommonTailLength(CurMBB,
434 MergePotentials[i].second,
436 if (CommonTailLen >= 2) {
442 // If we didn't find anything that has at least two instructions matching
443 // this one, bail out.
444 if (FoundMatch == ~0U) {
445 // Put the unconditional branch back, if we need one.
446 if (SuccBB && CurMBB != PredBB)
447 FixTail(CurMBB, SuccBB, TII);
448 MergePotentials.pop_back();
452 // Otherwise, move the matching block to the right position.
453 std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2));
456 MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second;
458 // If neither block is the entire common tail, split the tail of one block
459 // to make it redundant with the other tail.
460 if (CurMBB->begin() != BBI1 && MBB2->begin() != BBI2) {
461 if (0) { // Enable this to disable partial tail merges.
462 MergePotentials.pop_back();
466 // Decide whether we want to split CurMBB or MBB2.
467 if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII, PredBB)) {
468 CurMBB = SplitMBBAt(*CurMBB, BBI1);
469 BBI1 = CurMBB->begin();
470 MergePotentials.back().second = CurMBB;
472 MBB2 = SplitMBBAt(*MBB2, BBI2);
473 BBI2 = MBB2->begin();
474 (MergePotentials.end()-2)->second = MBB2;
478 if (MBB2->begin() == BBI2) {
479 // Hack the end off CurMBB, making it jump to MBBI@ instead.
480 ReplaceTailWithBranchTo(BBI1, MBB2);
481 // This modifies CurMBB, so remove it from the worklist.
482 MergePotentials.pop_back();
484 assert(CurMBB->begin() == BBI1 && "Didn't split block correctly?");
485 // Hack the end off MBB2, making it jump to CurMBB instead.
486 ReplaceTailWithBranchTo(BBI2, CurMBB);
487 // This modifies MBB2, so remove it from the worklist.
488 MergePotentials.erase(MergePotentials.end()-2);
495 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
497 if (!EnableTailMerge) return false;
501 // First find blocks with no successors.
502 MergePotentials.clear();
503 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
505 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I), I));
507 // See if we can do any tail merging on those.
508 MadeChange |= TryMergeBlocks(NULL, NULL);
510 // Look at blocks (IBB) with multiple predecessors (PBB).
511 // We change each predecessor to a canonical form, by
512 // (1) temporarily removing any unconditional branch from the predecessor
514 // (2) alter conditional branches so they branch to the other block
515 // not IBB; this may require adding back an unconditional branch to IBB
516 // later, where there wasn't one coming in. E.g.
518 // fallthrough to QBB
521 // with a conceptual B to IBB after that, which never actually exists.
522 // With those changes, we see whether the predecessors' tails match,
523 // and merge them if so. We change things out of canonical form and
524 // back to the way they were later in the process. (OptimizeBranches
525 // would undo some of this, but we can't use it, because we'd get into
526 // a compile-time infinite loop repeatedly doing and undoing the same
529 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
530 if (!I->succ_empty() && I->pred_size() >= 2) {
531 MachineBasicBlock *IBB = I;
532 MachineBasicBlock *PredBB = prior(I);
533 MergePotentials.clear();
534 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), E2 = I->pred_end();
536 MachineBasicBlock* PBB = *P;
537 // Skip blocks that loop to themselves, can't tail merge these.
540 MachineBasicBlock *TBB = 0, *FBB = 0;
541 std::vector<MachineOperand> Cond;
542 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) {
543 // Failing case: IBB is the target of a cbr, and
544 // we cannot reverse the branch.
545 std::vector<MachineOperand> NewCond(Cond);
546 if (Cond.size() && TBB==IBB) {
547 if (TII->ReverseBranchCondition(NewCond))
549 // This is the QBB case described above
551 FBB = next(MachineFunction::iterator(PBB));
553 // Remove the unconditional branch at the end, if any.
554 if (TBB && (Cond.size()==0 || FBB)) {
555 TII->RemoveBranch(*PBB);
557 // reinsert conditional branch only, for now
558 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond);
560 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB), *P));
563 if (MergePotentials.size() >= 2)
564 MadeChange |= TryMergeBlocks(I, PredBB);
565 // Reinsert an unconditional branch if needed.
566 // The 1 below can be either an original single predecessor, or a result
567 // of removing blocks in TryMergeBlocks.
568 if (MergePotentials.size()==1 &&
569 (MergePotentials.begin())->second != PredBB)
570 FixTail((MergePotentials.begin())->second, I, TII);
576 //===----------------------------------------------------------------------===//
577 // Branch Optimization
578 //===----------------------------------------------------------------------===//
580 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
583 // Make sure blocks are numbered in order
586 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
587 MachineBasicBlock *MBB = I++;
590 // If it is dead, remove it.
591 if (MBB->pred_empty()) {
592 RemoveDeadBlock(MBB);
601 /// CorrectExtraCFGEdges - Various pieces of code can cause excess edges in the
602 /// CFG to be inserted. If we have proven that MBB can only branch to DestA and
603 /// DestB, remove any other MBB successors from the CFG. DestA and DestB can
605 static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB,
606 MachineBasicBlock *DestA,
607 MachineBasicBlock *DestB,
609 MachineFunction::iterator FallThru) {
610 bool MadeChange = false;
611 bool AddedFallThrough = false;
613 // If this block ends with a conditional branch that falls through to its
614 // successor, set DestB as the successor.
616 if (DestB == 0 && FallThru != MBB.getParent()->end()) {
618 AddedFallThrough = true;
621 // If this is an unconditional branch with no explicit dest, it must just be
622 // a fallthrough into DestB.
623 if (DestA == 0 && FallThru != MBB.getParent()->end()) {
625 AddedFallThrough = true;
629 MachineBasicBlock::pred_iterator SI = MBB.succ_begin();
630 while (SI != MBB.succ_end()) {
634 } else if (*SI == DestB) {
637 } else if ((*SI)->isLandingPad()) {
640 // Otherwise, this is a superfluous edge, remove it.
641 MBB.removeSuccessor(SI);
645 if (!AddedFallThrough) {
646 assert(DestA == 0 && DestB == 0 &&
647 "MachineCFG is missing edges!");
649 assert(DestA == 0 && "MachineCFG is missing edges!");
655 /// ReplaceUsesOfBlockWith - Given a machine basic block 'BB' that branched to
656 /// 'Old', change the code and CFG so that it branches to 'New' instead.
657 static void ReplaceUsesOfBlockWith(MachineBasicBlock *BB,
658 MachineBasicBlock *Old,
659 MachineBasicBlock *New,
660 const TargetInstrInfo *TII) {
661 assert(Old != New && "Cannot replace self with self!");
663 MachineBasicBlock::iterator I = BB->end();
664 while (I != BB->begin()) {
666 if (!TII->isTerminatorInstr(I->getOpcode())) break;
668 // Scan the operands of this machine instruction, replacing any uses of Old
670 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
671 if (I->getOperand(i).isMachineBasicBlock() &&
672 I->getOperand(i).getMachineBasicBlock() == Old)
673 I->getOperand(i).setMachineBasicBlock(New);
676 // Update the successor information.
677 std::vector<MachineBasicBlock*> Succs(BB->succ_begin(), BB->succ_end());
678 for (int i = Succs.size()-1; i >= 0; --i)
679 if (Succs[i] == Old) {
680 BB->removeSuccessor(Old);
681 BB->addSuccessor(New);
685 /// CanFallThrough - Return true if the specified block (with the specified
686 /// branch condition) can implicitly transfer control to the block after it by
687 /// falling off the end of it. This should return false if it can reach the
688 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
690 /// True is a conservative answer.
692 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
693 bool BranchUnAnalyzable,
694 MachineBasicBlock *TBB, MachineBasicBlock *FBB,
695 const std::vector<MachineOperand> &Cond) {
696 MachineFunction::iterator Fallthrough = CurBB;
698 // If FallthroughBlock is off the end of the function, it can't fall through.
699 if (Fallthrough == CurBB->getParent()->end())
702 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
703 if (!CurBB->isSuccessor(Fallthrough))
706 // If we couldn't analyze the branch, assume it could fall through.
707 if (BranchUnAnalyzable) return true;
709 // If there is no branch, control always falls through.
710 if (TBB == 0) return true;
712 // If there is some explicit branch to the fallthrough block, it can obviously
713 // reach, even though the branch should get folded to fall through implicitly.
714 if (MachineFunction::iterator(TBB) == Fallthrough ||
715 MachineFunction::iterator(FBB) == Fallthrough)
718 // If it's an unconditional branch to some block not the fall through, it
719 // doesn't fall through.
720 if (Cond.empty()) return false;
722 // Otherwise, if it is conditional and has no explicit false block, it falls
727 /// CanFallThrough - Return true if the specified can implicitly transfer
728 /// control to the block after it by falling off the end of it. This should
729 /// return false if it can reach the block after it, but it uses an explicit
730 /// branch to do so (e.g. a table jump).
732 /// True is a conservative answer.
734 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
735 MachineBasicBlock *TBB = 0, *FBB = 0;
736 std::vector<MachineOperand> Cond;
737 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond);
738 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
741 /// IsBetterFallthrough - Return true if it would be clearly better to
742 /// fall-through to MBB1 than to fall through into MBB2. This has to return
743 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
744 /// result in infinite loops.
745 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
746 MachineBasicBlock *MBB2,
747 const TargetInstrInfo &TII) {
748 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
749 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
750 // optimize branches that branch to either a return block or an assert block
751 // into a fallthrough to the return.
752 if (MBB1->empty() || MBB2->empty()) return false;
754 MachineInstr *MBB1I = --MBB1->end();
755 MachineInstr *MBB2I = --MBB2->end();
756 return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode());
759 /// OptimizeBlock - Analyze and optimize control flow related to the specified
760 /// block. This is never called on the entry block.
761 void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
762 MachineFunction::iterator FallThrough = MBB;
765 // If this block is empty, make everyone use its fall-through, not the block
768 // Dead block? Leave for cleanup later.
769 if (MBB->pred_empty()) return;
771 if (FallThrough == MBB->getParent()->end()) {
772 // TODO: Simplify preds to not branch here if possible!
774 // Rewrite all predecessors of the old block to go to the fallthrough
776 while (!MBB->pred_empty()) {
777 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
778 ReplaceUsesOfBlockWith(Pred, MBB, FallThrough, TII);
781 // If MBB was the target of a jump table, update jump tables to go to the
782 // fallthrough instead.
783 MBB->getParent()->getJumpTableInfo()->
784 ReplaceMBBInJumpTables(MBB, FallThrough);
790 // Check to see if we can simplify the terminator of the block before this
792 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
794 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
795 std::vector<MachineOperand> PriorCond;
796 bool PriorUnAnalyzable =
797 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
798 if (!PriorUnAnalyzable) {
799 // If the CFG for the prior block has extra edges, remove them.
800 MadeChange |= CorrectExtraCFGEdges(PrevBB, PriorTBB, PriorFBB,
801 !PriorCond.empty(), MBB);
803 // If the previous branch is conditional and both conditions go to the same
804 // destination, remove the branch, replacing it with an unconditional one or
806 if (PriorTBB && PriorTBB == PriorFBB) {
807 TII->RemoveBranch(PrevBB);
810 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
813 return OptimizeBlock(MBB);
816 // If the previous branch *only* branches to *this* block (conditional or
817 // not) remove the branch.
818 if (PriorTBB == MBB && PriorFBB == 0) {
819 TII->RemoveBranch(PrevBB);
822 return OptimizeBlock(MBB);
825 // If the prior block branches somewhere else on the condition and here if
826 // the condition is false, remove the uncond second branch.
827 if (PriorFBB == MBB) {
828 TII->RemoveBranch(PrevBB);
829 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
832 return OptimizeBlock(MBB);
835 // If the prior block branches here on true and somewhere else on false, and
836 // if the branch condition is reversible, reverse the branch to create a
838 if (PriorTBB == MBB) {
839 std::vector<MachineOperand> NewPriorCond(PriorCond);
840 if (!TII->ReverseBranchCondition(NewPriorCond)) {
841 TII->RemoveBranch(PrevBB);
842 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
845 return OptimizeBlock(MBB);
849 // If this block doesn't fall through (e.g. it ends with an uncond branch or
850 // has no successors) and if the pred falls through into this block, and if
851 // it would otherwise fall through into the block after this, move this
852 // block to the end of the function.
854 // We consider it more likely that execution will stay in the function (e.g.
855 // due to loops) than it is to exit it. This asserts in loops etc, moving
856 // the assert condition out of the loop body.
857 if (!PriorCond.empty() && PriorFBB == 0 &&
858 MachineFunction::iterator(PriorTBB) == FallThrough &&
859 !CanFallThrough(MBB)) {
860 bool DoTransform = true;
862 // We have to be careful that the succs of PredBB aren't both no-successor
863 // blocks. If neither have successors and if PredBB is the second from
864 // last block in the function, we'd just keep swapping the two blocks for
865 // last. Only do the swap if one is clearly better to fall through than
867 if (FallThrough == --MBB->getParent()->end() &&
868 !IsBetterFallthrough(PriorTBB, MBB, *TII))
871 // We don't want to do this transformation if we have control flow like:
880 // In this case, we could actually be moving the return block *into* a
882 if (DoTransform && !MBB->succ_empty() &&
883 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
888 // Reverse the branch so we will fall through on the previous true cond.
889 std::vector<MachineOperand> NewPriorCond(PriorCond);
890 if (!TII->ReverseBranchCondition(NewPriorCond)) {
891 DOUT << "\nMoving MBB: " << *MBB;
892 DOUT << "To make fallthrough to: " << *PriorTBB << "\n";
894 TII->RemoveBranch(PrevBB);
895 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
897 // Move this block to the end of the function.
898 MBB->moveAfter(--MBB->getParent()->end());
907 // Analyze the branch in the current block.
908 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
909 std::vector<MachineOperand> CurCond;
910 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond);
911 if (!CurUnAnalyzable) {
912 // If the CFG for the prior block has extra edges, remove them.
913 MadeChange |= CorrectExtraCFGEdges(*MBB, CurTBB, CurFBB,
915 ++MachineFunction::iterator(MBB));
917 // If this is a two-way branch, and the FBB branches to this block, reverse
918 // the condition so the single-basic-block loop is faster. Instead of:
919 // Loop: xxx; jcc Out; jmp Loop
921 // Loop: xxx; jncc Loop; jmp Out
922 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
923 std::vector<MachineOperand> NewCond(CurCond);
924 if (!TII->ReverseBranchCondition(NewCond)) {
925 TII->RemoveBranch(*MBB);
926 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
929 return OptimizeBlock(MBB);
934 // If this branch is the only thing in its block, see if we can forward
935 // other blocks across it.
936 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
937 TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) {
938 // This block may contain just an unconditional branch. Because there can
939 // be 'non-branch terminators' in the block, try removing the branch and
940 // then seeing if the block is empty.
941 TII->RemoveBranch(*MBB);
943 // If this block is just an unconditional branch to CurTBB, we can
944 // usually completely eliminate the block. The only case we cannot
945 // completely eliminate the block is when the block before this one
946 // falls through into MBB and we can't understand the prior block's branch
949 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
950 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
951 !PrevBB.isSuccessor(MBB)) {
952 // If the prior block falls through into us, turn it into an
953 // explicit branch to us to make updates simpler.
954 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
955 PriorTBB != MBB && PriorFBB != MBB) {
957 assert(PriorCond.empty() && PriorFBB == 0 &&
958 "Bad branch analysis");
961 assert(PriorFBB == 0 && "Machine CFG out of date!");
964 TII->RemoveBranch(PrevBB);
965 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
968 // Iterate through all the predecessors, revectoring each in-turn.
969 MachineBasicBlock::pred_iterator PI = MBB->pred_begin();
970 bool DidChange = false;
971 bool HasBranchToSelf = false;
972 while (PI != MBB->pred_end()) {
974 // If this block has an uncond branch to itself, leave it.
976 HasBranchToSelf = true;
979 ReplaceUsesOfBlockWith(*PI, MBB, CurTBB, TII);
983 // Change any jumptables to go to the new MBB.
984 MBB->getParent()->getJumpTableInfo()->
985 ReplaceMBBInJumpTables(MBB, CurTBB);
989 if (!HasBranchToSelf) return;
994 // Add the branch back if the block is more than just an uncond branch.
995 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
999 // If the prior block doesn't fall through into this block, and if this
1000 // block doesn't fall through into some other block, see if we can find a
1001 // place to move this block where a fall-through will happen.
1002 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
1003 PriorTBB, PriorFBB, PriorCond)) {
1004 // Now we know that there was no fall-through into this block, check to
1005 // see if it has a fall-through into its successor.
1006 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1009 if (!MBB->isLandingPad()) {
1010 // Check all the predecessors of this block. If one of them has no fall
1011 // throughs, move this block right after it.
1012 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1013 E = MBB->pred_end(); PI != E; ++PI) {
1014 // Analyze the branch at the end of the pred.
1015 MachineBasicBlock *PredBB = *PI;
1016 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1017 if (PredBB != MBB && !CanFallThrough(PredBB)
1018 && (!CurFallsThru || !CurTBB || !CurFBB)
1019 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1020 // If the current block doesn't fall through, just move it.
1021 // If the current block can fall through and does not end with a
1022 // conditional branch, we need to append an unconditional jump to
1023 // the (current) next block. To avoid a possible compile-time
1024 // infinite loop, move blocks only backward in this case.
1025 // Also, if there are already 2 branches here, we cannot add a third;
1026 // this means we have the case
1031 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1033 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1035 MBB->moveAfter(PredBB);
1037 return OptimizeBlock(MBB);
1042 if (!CurFallsThru) {
1043 // Check all successors to see if we can move this block before it.
1044 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1045 E = MBB->succ_end(); SI != E; ++SI) {
1046 // Analyze the branch at the end of the block before the succ.
1047 MachineBasicBlock *SuccBB = *SI;
1048 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1049 std::vector<MachineOperand> SuccPrevCond;
1051 // If this block doesn't already fall-through to that successor, and if
1052 // the succ doesn't already have a block that can fall through into it,
1053 // and if the successor isn't an EH destination, we can arrange for the
1054 // fallthrough to happen.
1055 if (SuccBB != MBB && !CanFallThrough(SuccPrev) &&
1056 !SuccBB->isLandingPad()) {
1057 MBB->moveBefore(SuccBB);
1059 return OptimizeBlock(MBB);
1063 // Okay, there is no really great place to put this block. If, however,
1064 // the block before this one would be a fall-through if this block were
1065 // removed, move this block to the end of the function.
1066 if (FallThrough != MBB->getParent()->end() &&
1067 PrevBB.isSuccessor(FallThrough)) {
1068 MBB->moveAfter(--MBB->getParent()->end());