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 struct BranchFolder : public MachineFunctionPass {
43 BranchFolder(bool defaultEnableTailMerge) :
44 MachineFunctionPass((intptr_t)&ID) {
45 switch (FlagEnableTailMerge) {
46 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
47 case cl::BOU_TRUE: EnableTailMerge = true; break;
48 case cl::BOU_FALSE: EnableTailMerge = false; break;
52 virtual bool runOnMachineFunction(MachineFunction &MF);
53 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
54 const TargetInstrInfo *TII;
55 MachineModuleInfo *MMI;
60 bool TailMergeBlocks(MachineFunction &MF);
61 bool TryMergeBlocks(MachineBasicBlock* SuccBB,
62 MachineBasicBlock* PredBB);
63 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
64 MachineBasicBlock *NewDest);
65 MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB,
66 MachineBasicBlock::iterator BBI1);
68 std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials;
69 const MRegisterInfo *RegInfo;
72 bool OptimizeBranches(MachineFunction &MF);
73 void OptimizeBlock(MachineBasicBlock *MBB);
74 void RemoveDeadBlock(MachineBasicBlock *MBB);
76 bool CanFallThrough(MachineBasicBlock *CurBB);
77 bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable,
78 MachineBasicBlock *TBB, MachineBasicBlock *FBB,
79 const std::vector<MachineOperand> &Cond);
81 char BranchFolder::ID = 0;
84 static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB,
85 MachineBasicBlock *DestA,
86 MachineBasicBlock *DestB,
88 MachineFunction::iterator FallThru);
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 |= CorrectExtraCFGEdges(*MBB, TBB, FBB,
131 !Cond.empty(), next(I));
134 RegInfo = MF.getTarget().getRegisterInfo();
135 RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
137 MMI = getAnalysisToUpdate<MachineModuleInfo>();
139 bool MadeChangeThisIteration = true;
140 while (MadeChangeThisIteration) {
141 MadeChangeThisIteration = false;
142 MadeChangeThisIteration |= TailMergeBlocks(MF);
143 MadeChangeThisIteration |= OptimizeBranches(MF);
144 EverMadeChange |= MadeChangeThisIteration;
147 // See if any jump tables have become mergable or dead as the code generator
149 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
150 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
152 // Figure out how these jump tables should be merged.
153 std::vector<unsigned> JTMapping;
154 JTMapping.reserve(JTs.size());
156 // We always keep the 0th jump table.
157 JTMapping.push_back(0);
159 // Scan the jump tables, seeing if there are any duplicates. Note that this
160 // is N^2, which should be fixed someday.
161 for (unsigned i = 1, e = JTs.size(); i != e; ++i)
162 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
164 // If a jump table was merge with another one, walk the function rewriting
165 // references to jump tables to reference the new JT ID's. Keep track of
166 // whether we see a jump table idx, if not, we can delete the JT.
167 std::vector<bool> JTIsLive;
168 JTIsLive.resize(JTs.size());
169 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
171 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
173 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
174 MachineOperand &Op = I->getOperand(op);
175 if (!Op.isJumpTableIndex()) continue;
176 unsigned NewIdx = JTMapping[Op.getJumpTableIndex()];
177 Op.setJumpTableIndex(NewIdx);
179 // Remember that this JT is live.
180 JTIsLive[NewIdx] = true;
184 // Finally, remove dead jump tables. This happens either because the
185 // indirect jump was unreachable (and thus deleted) or because the jump
186 // table was merged with some other one.
187 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
189 JTI->RemoveJumpTable(i);
190 EverMadeChange = true;
195 return EverMadeChange;
198 //===----------------------------------------------------------------------===//
199 // Tail Merging of Blocks
200 //===----------------------------------------------------------------------===//
202 /// HashMachineInstr - Compute a hash value for MI and its operands.
203 static unsigned HashMachineInstr(const MachineInstr *MI) {
204 unsigned Hash = MI->getOpcode();
205 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
206 const MachineOperand &Op = MI->getOperand(i);
208 // Merge in bits from the operand if easy.
209 unsigned OperandHash = 0;
210 switch (Op.getType()) {
211 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
212 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
213 case MachineOperand::MO_MachineBasicBlock:
214 OperandHash = Op.getMachineBasicBlock()->getNumber();
216 case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break;
217 case MachineOperand::MO_ConstantPoolIndex:
218 OperandHash = Op.getConstantPoolIndex();
220 case MachineOperand::MO_JumpTableIndex:
221 OperandHash = Op.getJumpTableIndex();
223 case MachineOperand::MO_GlobalAddress:
224 case MachineOperand::MO_ExternalSymbol:
225 // Global address / external symbol are too hard, don't bother, but do
226 // pull in the offset.
227 OperandHash = Op.getOffset();
232 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
237 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
238 /// with no successors, we hash two instructions, because cross-jumping
239 /// only saves code when at least two instructions are removed (since a
240 /// branch must be inserted). For blocks with a successor, one of the
241 /// two blocks to be tail-merged will end with a branch already, so
242 /// it gains to cross-jump even for one instruction.
244 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
245 unsigned minCommonTailLength) {
246 MachineBasicBlock::const_iterator I = MBB->end();
247 if (I == MBB->begin())
248 return 0; // Empty MBB.
251 unsigned Hash = HashMachineInstr(I);
253 if (I == MBB->begin() || minCommonTailLength == 1)
254 return Hash; // Single instr MBB.
257 // Hash in the second-to-last instruction.
258 Hash ^= HashMachineInstr(I) << 2;
262 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
263 /// of instructions they actually have in common together at their end. Return
264 /// iterators for the first shared instruction in each block.
265 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
266 MachineBasicBlock *MBB2,
267 MachineBasicBlock::iterator &I1,
268 MachineBasicBlock::iterator &I2) {
272 unsigned TailLen = 0;
273 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
275 if (!I1->isIdenticalTo(I2)) {
284 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
285 /// after it, replacing it with an unconditional branch to NewDest. This
286 /// returns true if OldInst's block is modified, false if NewDest is modified.
287 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
288 MachineBasicBlock *NewDest) {
289 MachineBasicBlock *OldBB = OldInst->getParent();
291 // Remove all the old successors of OldBB from the CFG.
292 while (!OldBB->succ_empty())
293 OldBB->removeSuccessor(OldBB->succ_begin());
295 // Remove all the dead instructions from the end of OldBB.
296 OldBB->erase(OldInst, OldBB->end());
298 // If OldBB isn't immediately before OldBB, insert a branch to it.
299 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
300 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>());
301 OldBB->addSuccessor(NewDest);
305 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
306 /// MBB so that the part before the iterator falls into the part starting at the
307 /// iterator. This returns the new MBB.
308 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
309 MachineBasicBlock::iterator BBI1) {
310 // Create the fall-through block.
311 MachineFunction::iterator MBBI = &CurMBB;
312 MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock());
313 CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB);
315 // Move all the successors of this block to the specified block.
316 while (!CurMBB.succ_empty()) {
317 MachineBasicBlock *S = *(CurMBB.succ_end()-1);
318 NewMBB->addSuccessor(S);
319 CurMBB.removeSuccessor(S);
322 // Add an edge from CurMBB to NewMBB for the fall-through.
323 CurMBB.addSuccessor(NewMBB);
325 // Splice the code over.
326 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
328 // For targets that use the register scavenger, we must maintain LiveIns.
330 RS->enterBasicBlock(&CurMBB);
332 RS->forward(prior(CurMBB.end()));
333 BitVector RegsLiveAtExit(RegInfo->getNumRegs());
334 RS->getRegsUsed(RegsLiveAtExit, false);
335 for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++)
336 if (RegsLiveAtExit[i])
337 NewMBB->addLiveIn(i);
343 /// EstimateRuntime - Make a rough estimate for how long it will take to run
344 /// the specified code.
345 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
346 MachineBasicBlock::iterator E,
347 const TargetInstrInfo *TII) {
349 for (; I != E; ++I) {
350 const TargetInstrDescriptor &TID = TII->get(I->getOpcode());
351 if (TID.Flags & M_CALL_FLAG)
353 else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG))
361 /// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at
362 /// MBB2I and then insert an unconditional branch in the other block. Determine
363 /// which is the best to split
364 static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1,
365 MachineBasicBlock::iterator MBB1I,
366 MachineBasicBlock *MBB2,
367 MachineBasicBlock::iterator MBB2I,
368 const TargetInstrInfo *TII,
369 MachineBasicBlock *PredBB) {
370 // If one block is the entry block, split the other one; we can't generate
371 // a branch to the entry block, as its label is not emitted.
372 MachineBasicBlock *Entry = MBB1->getParent()->begin();
378 // If one block falls through into the common successor, choose that
379 // one to split; it is one instruction less to do that.
383 else if (MBB2 == PredBB)
386 // TODO: if we had some notion of which block was hotter, we could split
387 // the hot block, so it is the fall-through. Since we don't have profile info
388 // make a decision based on which will hurt most to split.
389 unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII);
390 unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII);
392 // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the
393 // MBB1 block so it falls through. This will penalize the MBB2 path, but will
394 // have a lower overall impact on the program execution.
395 return MBB1Time < MBB2Time;
398 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
399 // branches temporarily for tail merging). In the case where CurMBB ends
400 // with a conditional branch to the next block, optimize by reversing the
401 // test and conditionally branching to SuccMBB instead.
403 static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB,
404 const TargetInstrInfo *TII) {
405 MachineFunction *MF = CurMBB->getParent();
406 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
407 MachineBasicBlock *TBB = 0, *FBB = 0;
408 std::vector<MachineOperand> Cond;
409 if (I != MF->end() &&
410 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) {
411 MachineBasicBlock *NextBB = I;
412 if (TBB == NextBB && Cond.size() && !FBB) {
413 if (!TII->ReverseBranchCondition(Cond)) {
414 TII->RemoveBranch(*CurMBB);
415 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
420 TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>());
423 static bool MergeCompare(std::pair<unsigned,MachineBasicBlock*> p,
424 std::pair<unsigned,MachineBasicBlock*> q) {
426 if (p.first < q.first)
428 else if (p.first > q.first)
430 else if (p.second->getNumber() < q.second->getNumber())
432 else if (p.second->getNumber() > q.second->getNumber())
435 assert(0 && "Predecessor appears twice");
438 // See if any of the blocks in MergePotentials (which all have a common single
439 // successor, or all have no successor) can be tail-merged. If there is a
440 // successor, any blocks in MergePotentials that are not tail-merged and
441 // are not immediately before Succ must have an unconditional branch to
442 // Succ added (but the predecessor/successor lists need no adjustment).
443 // The lone predecessor of Succ that falls through into Succ,
444 // if any, is given in PredBB.
446 bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB,
447 MachineBasicBlock* PredBB) {
448 unsigned minCommonTailLength = (SuccBB ? 1 : 2);
451 // Sort by hash value so that blocks with identical end sequences sort
453 std::stable_sort(MergePotentials.begin(), MergePotentials.end(), MergeCompare);
455 // Walk through equivalence sets looking for actual exact matches.
456 while (MergePotentials.size() > 1) {
457 unsigned CurHash = (MergePotentials.end()-1)->first;
458 unsigned PrevHash = (MergePotentials.end()-2)->first;
459 MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second;
461 // If there is nothing that matches the hash of the current basic block,
463 if (CurHash != PrevHash) {
464 if (SuccBB && CurMBB != PredBB)
465 FixTail(CurMBB, SuccBB, TII);
466 MergePotentials.pop_back();
470 // Look through all the blocks that have the same hash as this one, and
471 // find the one that has the largest number of instructions in common.
472 // Since instructions may get combined later (e.g. single stores into
473 // store multiple) this measure is not particularly accurate.
474 MachineBasicBlock::iterator BBI1, BBI2;
476 unsigned FoundMatch = ~0U;
477 unsigned maxCommonTailLength = 0U;
478 for (int i = MergePotentials.size()-2;
479 i != -1 && MergePotentials[i].first == CurHash; --i) {
480 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
481 unsigned CommonTailLen = ComputeCommonTailLength(CurMBB,
482 MergePotentials[i].second,
483 TrialBBI1, TrialBBI2);
484 if (CommonTailLen >= minCommonTailLength &&
485 CommonTailLen >= maxCommonTailLength) {
487 maxCommonTailLength = CommonTailLen;
493 // If we didn't find anything that has at least minCommonTailLength
494 // instructions matching this one, bail out.
495 if (FoundMatch == ~0U) {
496 // Put the unconditional branch back, if we need one.
497 if (SuccBB && CurMBB != PredBB)
498 FixTail(CurMBB, SuccBB, TII);
499 MergePotentials.pop_back();
503 // Otherwise, move the matching block to the right position.
504 if (FoundMatch != MergePotentials.size()-2)
505 std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2));
507 MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second;
509 // If neither block is the entire common tail, split the tail of one block
510 // to make it redundant with the other tail. Also, we cannot jump to the
511 // entry block, so if one block is the entry block, split the other one.
512 MachineBasicBlock *Entry = CurMBB->getParent()->begin();
513 if (CurMBB->begin() == BBI1 && CurMBB != Entry)
514 ; // CurMBB is common tail
515 else if (MBB2->begin() == BBI2 && MBB2 != Entry)
516 ; // MBB2 is common tail
518 if (0) { // Enable this to disable partial tail merges.
519 MergePotentials.pop_back();
523 // Decide whether we want to split CurMBB or MBB2.
524 if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII, PredBB)) {
525 CurMBB = SplitMBBAt(*CurMBB, BBI1);
526 BBI1 = CurMBB->begin();
527 MergePotentials.back().second = CurMBB;
529 MBB2 = SplitMBBAt(*MBB2, BBI2);
530 BBI2 = MBB2->begin();
531 (MergePotentials.end()-2)->second = MBB2;
535 if (MBB2->begin() == BBI2 && MBB2 != Entry) {
536 // Hack the end off CurMBB, making it jump to MBBI@ instead.
537 ReplaceTailWithBranchTo(BBI1, MBB2);
538 // This modifies CurMBB, so remove it from the worklist.
539 MergePotentials.pop_back();
541 assert(CurMBB->begin() == BBI1 && CurMBB != Entry &&
542 "Didn't split block correctly?");
543 // Hack the end off MBB2, making it jump to CurMBB instead.
544 ReplaceTailWithBranchTo(BBI2, CurMBB);
545 // This modifies MBB2, so remove it from the worklist.
546 MergePotentials.erase(MergePotentials.end()-2);
553 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
555 if (!EnableTailMerge) return false;
559 // First find blocks with no successors.
560 MergePotentials.clear();
561 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
563 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I));
565 // See if we can do any tail merging on those.
566 MadeChange |= TryMergeBlocks(NULL, NULL);
568 // Look at blocks (IBB) with multiple predecessors (PBB).
569 // We change each predecessor to a canonical form, by
570 // (1) temporarily removing any unconditional branch from the predecessor
572 // (2) alter conditional branches so they branch to the other block
573 // not IBB; this may require adding back an unconditional branch to IBB
574 // later, where there wasn't one coming in. E.g.
576 // fallthrough to QBB
579 // with a conceptual B to IBB after that, which never actually exists.
580 // With those changes, we see whether the predecessors' tails match,
581 // and merge them if so. We change things out of canonical form and
582 // back to the way they were later in the process. (OptimizeBranches
583 // would undo some of this, but we can't use it, because we'd get into
584 // a compile-time infinite loop repeatedly doing and undoing the same
587 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
588 if (!I->succ_empty() && I->pred_size() >= 2) {
589 MachineBasicBlock *IBB = I;
590 MachineBasicBlock *PredBB = prior(I);
591 MergePotentials.clear();
592 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), E2 = I->pred_end();
594 MachineBasicBlock* PBB = *P;
595 // Skip blocks that loop to themselves, can't tail merge these.
598 MachineBasicBlock *TBB = 0, *FBB = 0;
599 std::vector<MachineOperand> Cond;
600 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) {
601 // Failing case: IBB is the target of a cbr, and
602 // we cannot reverse the branch.
603 std::vector<MachineOperand> NewCond(Cond);
604 if (Cond.size() && TBB==IBB) {
605 if (TII->ReverseBranchCondition(NewCond))
607 // This is the QBB case described above
609 FBB = next(MachineFunction::iterator(PBB));
611 // Remove the unconditional branch at the end, if any.
612 if (TBB && (Cond.size()==0 || FBB)) {
613 TII->RemoveBranch(*PBB);
615 // reinsert conditional branch only, for now
616 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond);
618 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P));
621 if (MergePotentials.size() >= 2)
622 MadeChange |= TryMergeBlocks(I, PredBB);
623 // Reinsert an unconditional branch if needed.
624 // The 1 below can be either an original single predecessor, or a result
625 // of removing blocks in TryMergeBlocks.
626 PredBB = prior(I); // this may have been changed in TryMergeBlocks
627 if (MergePotentials.size()==1 &&
628 (MergePotentials.begin())->second != PredBB)
629 FixTail((MergePotentials.begin())->second, I, TII);
635 //===----------------------------------------------------------------------===//
636 // Branch Optimization
637 //===----------------------------------------------------------------------===//
639 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
642 // Make sure blocks are numbered in order
645 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
646 MachineBasicBlock *MBB = I++;
649 // If it is dead, remove it.
650 if (MBB->pred_empty()) {
651 RemoveDeadBlock(MBB);
660 /// CorrectExtraCFGEdges - Various pieces of code can cause excess edges in the
661 /// CFG to be inserted. If we have proven that MBB can only branch to DestA and
662 /// DestB, remove any other MBB successors from the CFG. DestA and DestB can
664 static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB,
665 MachineBasicBlock *DestA,
666 MachineBasicBlock *DestB,
668 MachineFunction::iterator FallThru) {
669 bool MadeChange = false;
670 bool AddedFallThrough = false;
672 // If this block ends with a conditional branch that falls through to its
673 // successor, set DestB as the successor.
675 if (DestB == 0 && FallThru != MBB.getParent()->end()) {
677 AddedFallThrough = true;
680 // If this is an unconditional branch with no explicit dest, it must just be
681 // a fallthrough into DestB.
682 if (DestA == 0 && FallThru != MBB.getParent()->end()) {
684 AddedFallThrough = true;
688 MachineBasicBlock::pred_iterator SI = MBB.succ_begin();
689 while (SI != MBB.succ_end()) {
690 if (*SI == DestA && DestA == DestB) {
693 } else if (*SI == DestA) {
696 } else if (*SI == DestB) {
699 } else if ((*SI)->isLandingPad()) {
702 // Otherwise, this is a superfluous edge, remove it.
703 MBB.removeSuccessor(SI);
707 if (!AddedFallThrough) {
708 assert(DestA == 0 && DestB == 0 &&
709 "MachineCFG is missing edges!");
711 assert(DestA == 0 && "MachineCFG is missing edges!");
717 /// ReplaceUsesOfBlockWith - Given a machine basic block 'BB' that branched to
718 /// 'Old', change the code and CFG so that it branches to 'New' instead.
719 static void ReplaceUsesOfBlockWith(MachineBasicBlock *BB,
720 MachineBasicBlock *Old,
721 MachineBasicBlock *New,
722 const TargetInstrInfo *TII) {
723 assert(Old != New && "Cannot replace self with self!");
725 MachineBasicBlock::iterator I = BB->end();
726 while (I != BB->begin()) {
728 if (!TII->isTerminatorInstr(I->getOpcode())) break;
730 // Scan the operands of this machine instruction, replacing any uses of Old
732 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
733 if (I->getOperand(i).isMachineBasicBlock() &&
734 I->getOperand(i).getMachineBasicBlock() == Old)
735 I->getOperand(i).setMachineBasicBlock(New);
738 // Update the successor information. If New was already a successor, just
739 // remove the link to Old instead of creating another one. PR 1444.
740 bool HadSuccessorNew = false;
741 std::vector<MachineBasicBlock*> Succs(BB->succ_begin(), BB->succ_end());
742 for (int i = Succs.size()-1; i >= 0; --i)
743 if (Succs[i] == New) {
744 HadSuccessorNew = true;
747 for (int i = Succs.size()-1; i >= 0; --i)
748 if (Succs[i] == Old) {
749 BB->removeSuccessor(Old);
750 if (!HadSuccessorNew)
751 BB->addSuccessor(New);
755 /// CanFallThrough - Return true if the specified block (with the specified
756 /// branch condition) can implicitly transfer control to the block after it by
757 /// falling off the end of it. This should return false if it can reach the
758 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
760 /// True is a conservative answer.
762 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
763 bool BranchUnAnalyzable,
764 MachineBasicBlock *TBB, MachineBasicBlock *FBB,
765 const std::vector<MachineOperand> &Cond) {
766 MachineFunction::iterator Fallthrough = CurBB;
768 // If FallthroughBlock is off the end of the function, it can't fall through.
769 if (Fallthrough == CurBB->getParent()->end())
772 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
773 if (!CurBB->isSuccessor(Fallthrough))
776 // If we couldn't analyze the branch, assume it could fall through.
777 if (BranchUnAnalyzable) return true;
779 // If there is no branch, control always falls through.
780 if (TBB == 0) return true;
782 // If there is some explicit branch to the fallthrough block, it can obviously
783 // reach, even though the branch should get folded to fall through implicitly.
784 if (MachineFunction::iterator(TBB) == Fallthrough ||
785 MachineFunction::iterator(FBB) == Fallthrough)
788 // If it's an unconditional branch to some block not the fall through, it
789 // doesn't fall through.
790 if (Cond.empty()) return false;
792 // Otherwise, if it is conditional and has no explicit false block, it falls
797 /// CanFallThrough - Return true if the specified can implicitly transfer
798 /// control to the block after it by falling off the end of it. This should
799 /// return false if it can reach the block after it, but it uses an explicit
800 /// branch to do so (e.g. a table jump).
802 /// True is a conservative answer.
804 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
805 MachineBasicBlock *TBB = 0, *FBB = 0;
806 std::vector<MachineOperand> Cond;
807 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond);
808 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
811 /// IsBetterFallthrough - Return true if it would be clearly better to
812 /// fall-through to MBB1 than to fall through into MBB2. This has to return
813 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
814 /// result in infinite loops.
815 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
816 MachineBasicBlock *MBB2,
817 const TargetInstrInfo &TII) {
818 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
819 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
820 // optimize branches that branch to either a return block or an assert block
821 // into a fallthrough to the return.
822 if (MBB1->empty() || MBB2->empty()) return false;
824 MachineInstr *MBB1I = --MBB1->end();
825 MachineInstr *MBB2I = --MBB2->end();
826 return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode());
829 /// OptimizeBlock - Analyze and optimize control flow related to the specified
830 /// block. This is never called on the entry block.
831 void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
832 MachineFunction::iterator FallThrough = MBB;
835 // If this block is empty, make everyone use its fall-through, not the block
838 // Dead block? Leave for cleanup later.
839 if (MBB->pred_empty()) return;
841 if (FallThrough == MBB->getParent()->end()) {
842 // TODO: Simplify preds to not branch here if possible!
844 // Rewrite all predecessors of the old block to go to the fallthrough
846 while (!MBB->pred_empty()) {
847 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
848 ReplaceUsesOfBlockWith(Pred, MBB, FallThrough, TII);
851 // If MBB was the target of a jump table, update jump tables to go to the
852 // fallthrough instead.
853 MBB->getParent()->getJumpTableInfo()->
854 ReplaceMBBInJumpTables(MBB, FallThrough);
860 // Check to see if we can simplify the terminator of the block before this
862 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
864 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
865 std::vector<MachineOperand> PriorCond;
866 bool PriorUnAnalyzable =
867 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
868 if (!PriorUnAnalyzable) {
869 // If the CFG for the prior block has extra edges, remove them.
870 MadeChange |= CorrectExtraCFGEdges(PrevBB, PriorTBB, PriorFBB,
871 !PriorCond.empty(), MBB);
873 // If the previous branch is conditional and both conditions go to the same
874 // destination, remove the branch, replacing it with an unconditional one or
876 if (PriorTBB && PriorTBB == PriorFBB) {
877 TII->RemoveBranch(PrevBB);
880 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
883 return OptimizeBlock(MBB);
886 // If the previous branch *only* branches to *this* block (conditional or
887 // not) remove the branch.
888 if (PriorTBB == MBB && PriorFBB == 0) {
889 TII->RemoveBranch(PrevBB);
892 return OptimizeBlock(MBB);
895 // If the prior block branches somewhere else on the condition and here if
896 // the condition is false, remove the uncond second branch.
897 if (PriorFBB == MBB) {
898 TII->RemoveBranch(PrevBB);
899 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
902 return OptimizeBlock(MBB);
905 // If the prior block branches here on true and somewhere else on false, and
906 // if the branch condition is reversible, reverse the branch to create a
908 if (PriorTBB == MBB) {
909 std::vector<MachineOperand> NewPriorCond(PriorCond);
910 if (!TII->ReverseBranchCondition(NewPriorCond)) {
911 TII->RemoveBranch(PrevBB);
912 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
915 return OptimizeBlock(MBB);
919 // If this block doesn't fall through (e.g. it ends with an uncond branch or
920 // has no successors) and if the pred falls through into this block, and if
921 // it would otherwise fall through into the block after this, move this
922 // block to the end of the function.
924 // We consider it more likely that execution will stay in the function (e.g.
925 // due to loops) than it is to exit it. This asserts in loops etc, moving
926 // the assert condition out of the loop body.
927 if (!PriorCond.empty() && PriorFBB == 0 &&
928 MachineFunction::iterator(PriorTBB) == FallThrough &&
929 !CanFallThrough(MBB)) {
930 bool DoTransform = true;
932 // We have to be careful that the succs of PredBB aren't both no-successor
933 // blocks. If neither have successors and if PredBB is the second from
934 // last block in the function, we'd just keep swapping the two blocks for
935 // last. Only do the swap if one is clearly better to fall through than
937 if (FallThrough == --MBB->getParent()->end() &&
938 !IsBetterFallthrough(PriorTBB, MBB, *TII))
941 // We don't want to do this transformation if we have control flow like:
950 // In this case, we could actually be moving the return block *into* a
952 if (DoTransform && !MBB->succ_empty() &&
953 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
958 // Reverse the branch so we will fall through on the previous true cond.
959 std::vector<MachineOperand> NewPriorCond(PriorCond);
960 if (!TII->ReverseBranchCondition(NewPriorCond)) {
961 DOUT << "\nMoving MBB: " << *MBB;
962 DOUT << "To make fallthrough to: " << *PriorTBB << "\n";
964 TII->RemoveBranch(PrevBB);
965 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
967 // Move this block to the end of the function.
968 MBB->moveAfter(--MBB->getParent()->end());
977 // Analyze the branch in the current block.
978 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
979 std::vector<MachineOperand> CurCond;
980 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond);
981 if (!CurUnAnalyzable) {
982 // If the CFG for the prior block has extra edges, remove them.
983 MadeChange |= CorrectExtraCFGEdges(*MBB, CurTBB, CurFBB,
985 ++MachineFunction::iterator(MBB));
987 // If this is a two-way branch, and the FBB branches to this block, reverse
988 // the condition so the single-basic-block loop is faster. Instead of:
989 // Loop: xxx; jcc Out; jmp Loop
991 // Loop: xxx; jncc Loop; jmp Out
992 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
993 std::vector<MachineOperand> NewCond(CurCond);
994 if (!TII->ReverseBranchCondition(NewCond)) {
995 TII->RemoveBranch(*MBB);
996 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
999 return OptimizeBlock(MBB);
1004 // If this branch is the only thing in its block, see if we can forward
1005 // other blocks across it.
1006 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1007 TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) {
1008 // This block may contain just an unconditional branch. Because there can
1009 // be 'non-branch terminators' in the block, try removing the branch and
1010 // then seeing if the block is empty.
1011 TII->RemoveBranch(*MBB);
1013 // If this block is just an unconditional branch to CurTBB, we can
1014 // usually completely eliminate the block. The only case we cannot
1015 // completely eliminate the block is when the block before this one
1016 // falls through into MBB and we can't understand the prior block's branch
1019 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
1020 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1021 !PrevBB.isSuccessor(MBB)) {
1022 // If the prior block falls through into us, turn it into an
1023 // explicit branch to us to make updates simpler.
1024 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1025 PriorTBB != MBB && PriorFBB != MBB) {
1026 if (PriorTBB == 0) {
1027 assert(PriorCond.empty() && PriorFBB == 0 &&
1028 "Bad branch analysis");
1031 assert(PriorFBB == 0 && "Machine CFG out of date!");
1034 TII->RemoveBranch(PrevBB);
1035 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1038 // Iterate through all the predecessors, revectoring each in-turn.
1039 MachineBasicBlock::pred_iterator PI = MBB->pred_begin();
1040 bool DidChange = false;
1041 bool HasBranchToSelf = false;
1042 while (PI != MBB->pred_end()) {
1044 // If this block has an uncond branch to itself, leave it.
1046 HasBranchToSelf = true;
1049 ReplaceUsesOfBlockWith(*PI, MBB, CurTBB, TII);
1053 // Change any jumptables to go to the new MBB.
1054 MBB->getParent()->getJumpTableInfo()->
1055 ReplaceMBBInJumpTables(MBB, CurTBB);
1059 if (!HasBranchToSelf) return;
1064 // Add the branch back if the block is more than just an uncond branch.
1065 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1069 // If the prior block doesn't fall through into this block, and if this
1070 // block doesn't fall through into some other block, see if we can find a
1071 // place to move this block where a fall-through will happen.
1072 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
1073 PriorTBB, PriorFBB, PriorCond)) {
1074 // Now we know that there was no fall-through into this block, check to
1075 // see if it has a fall-through into its successor.
1076 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1079 if (!MBB->isLandingPad()) {
1080 // Check all the predecessors of this block. If one of them has no fall
1081 // throughs, move this block right after it.
1082 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1083 E = MBB->pred_end(); PI != E; ++PI) {
1084 // Analyze the branch at the end of the pred.
1085 MachineBasicBlock *PredBB = *PI;
1086 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1087 if (PredBB != MBB && !CanFallThrough(PredBB)
1088 && (!CurFallsThru || !CurTBB || !CurFBB)
1089 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1090 // If the current block doesn't fall through, just move it.
1091 // If the current block can fall through and does not end with a
1092 // conditional branch, we need to append an unconditional jump to
1093 // the (current) next block. To avoid a possible compile-time
1094 // infinite loop, move blocks only backward in this case.
1095 // Also, if there are already 2 branches here, we cannot add a third;
1096 // this means we have the case
1101 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1103 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1105 MBB->moveAfter(PredBB);
1107 return OptimizeBlock(MBB);
1112 if (!CurFallsThru) {
1113 // Check all successors to see if we can move this block before it.
1114 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1115 E = MBB->succ_end(); SI != E; ++SI) {
1116 // Analyze the branch at the end of the block before the succ.
1117 MachineBasicBlock *SuccBB = *SI;
1118 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1119 std::vector<MachineOperand> SuccPrevCond;
1121 // If this block doesn't already fall-through to that successor, and if
1122 // the succ doesn't already have a block that can fall through into it,
1123 // and if the successor isn't an EH destination, we can arrange for the
1124 // fallthrough to happen.
1125 if (SuccBB != MBB && !CanFallThrough(SuccPrev) &&
1126 !SuccBB->isLandingPad()) {
1127 MBB->moveBefore(SuccBB);
1129 return OptimizeBlock(MBB);
1133 // Okay, there is no really great place to put this block. If, however,
1134 // the block before this one would be a fall-through if this block were
1135 // removed, move this block to the end of the function.
1136 if (FallThrough != MBB->getParent()->end() &&
1137 PrevBB.isSuccessor(FallThrough)) {
1138 MBB->moveAfter(--MBB->getParent()->end());