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 #include "llvm/CodeGen/Passes.h"
20 #include "llvm/CodeGen/MachineDebugInfo.h"
21 #include "llvm/CodeGen/MachineFunctionPass.h"
22 #include "llvm/CodeGen/MachineJumpTableInfo.h"
23 #include "llvm/Target/TargetInstrInfo.h"
24 #include "llvm/Target/TargetMachine.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/STLExtras.h"
30 static Statistic<> NumDeadBlocks("branchfold", "Number of dead blocks removed");
31 static Statistic<> NumBranchOpts("branchfold", "Number of branches optimized");
32 static Statistic<> NumTailMerge ("branchfold", "Number of block tails merged");
35 struct BranchFolder : public MachineFunctionPass {
36 virtual bool runOnMachineFunction(MachineFunction &MF);
37 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
38 const TargetInstrInfo *TII;
39 MachineDebugInfo *MDI;
43 bool TailMergeBlocks(MachineFunction &MF);
44 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
45 MachineBasicBlock *NewDest);
48 bool OptimizeBranches(MachineFunction &MF);
49 void OptimizeBlock(MachineBasicBlock *MBB);
50 void RemoveDeadBlock(MachineBasicBlock *MBB);
54 FunctionPass *llvm::createBranchFoldingPass() { return new BranchFolder(); }
56 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
57 /// function, updating the CFG.
58 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
59 assert(MBB->pred_empty() && "MBB must be dead!");
61 MachineFunction *MF = MBB->getParent();
62 // drop all successors.
63 while (!MBB->succ_empty())
64 MBB->removeSuccessor(MBB->succ_end()-1);
66 // If there is DWARF info to active, check to see if there are any DWARF_LABEL
67 // records in the basic block. If so, unregister them from MachineDebugInfo.
68 if (MDI && !MBB->empty()) {
69 unsigned DWARF_LABELOpc = TII->getDWARF_LABELOpcode();
70 assert(DWARF_LABELOpc &&
71 "Target supports dwarf but didn't implement getDWARF_LABELOpcode!");
73 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
75 if ((unsigned)I->getOpcode() == DWARF_LABELOpc) {
76 // The label ID # is always operand #0, an immediate.
77 MDI->InvalidateLabel(I->getOperand(0).getImm());
83 MF->getBasicBlockList().erase(MBB);
86 bool BranchFolder::runOnMachineFunction(MachineFunction &MF) {
87 TII = MF.getTarget().getInstrInfo();
88 if (!TII) return false;
90 MDI = getAnalysisToUpdate<MachineDebugInfo>();
92 bool EverMadeChange = false;
93 bool MadeChangeThisIteration = true;
94 while (MadeChangeThisIteration) {
95 MadeChangeThisIteration = false;
96 MadeChangeThisIteration |= TailMergeBlocks(MF);
97 MadeChangeThisIteration |= OptimizeBranches(MF);
98 EverMadeChange |= MadeChangeThisIteration;
101 return EverMadeChange;
104 //===----------------------------------------------------------------------===//
105 // Tail Merging of Blocks
106 //===----------------------------------------------------------------------===//
108 /// HashMachineInstr - Compute a hash value for MI and its operands.
109 static unsigned HashMachineInstr(const MachineInstr *MI) {
110 unsigned Hash = MI->getOpcode();
111 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
112 const MachineOperand &Op = MI->getOperand(i);
114 // Merge in bits from the operand if easy.
115 unsigned OperandHash = 0;
116 switch (Op.getType()) {
117 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
118 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
119 case MachineOperand::MO_MachineBasicBlock:
120 OperandHash = Op.getMachineBasicBlock()->getNumber();
122 case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break;
123 case MachineOperand::MO_ConstantPoolIndex:
124 OperandHash = Op.getConstantPoolIndex();
126 case MachineOperand::MO_JumpTableIndex:
127 OperandHash = Op.getJumpTableIndex();
129 case MachineOperand::MO_GlobalAddress:
130 case MachineOperand::MO_ExternalSymbol:
131 // Global address / external symbol are too hard, don't bother, but do
132 // pull in the offset.
133 OperandHash = Op.getOffset();
138 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
143 /// HashEndOfMBB - Hash the last two instructions in the MBB. We hash two
144 /// instructions, because cross-jumping only saves code when at least two
145 /// instructions are removed (since a branch must be inserted).
146 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
147 MachineBasicBlock::const_iterator I = MBB->end();
148 if (I == MBB->begin())
149 return 0; // Empty MBB.
152 unsigned Hash = HashMachineInstr(I);
154 if (I == MBB->begin())
155 return Hash; // Single instr MBB.
158 // Hash in the second-to-last instruction.
159 Hash ^= HashMachineInstr(I) << 2;
163 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
164 /// of instructions they actually have in common together at their end. Return
165 /// iterators for the first shared instruction in each block.
166 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
167 MachineBasicBlock *MBB2,
168 MachineBasicBlock::iterator &I1,
169 MachineBasicBlock::iterator &I2) {
173 unsigned TailLen = 0;
174 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
176 if (!I1->isIdenticalTo(I2)) {
185 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
186 /// after it, replacing it with an unconditional branch to NewDest. This
187 /// returns true if OldInst's block is modified, false if NewDest is modified.
188 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
189 MachineBasicBlock *NewDest) {
190 MachineBasicBlock *OldBB = OldInst->getParent();
192 // Remove all the old successors of OldBB from the CFG.
193 while (!OldBB->succ_empty())
194 OldBB->removeSuccessor(OldBB->succ_begin());
196 // Remove all the dead instructions from the end of OldBB.
197 OldBB->erase(OldInst, OldBB->end());
199 // If OldBB isn't immediately before OldBB, insert a branch to it.
200 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
201 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>());
202 OldBB->addSuccessor(NewDest);
206 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
211 // Find blocks with no successors.
212 std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials;
213 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
215 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I), I));
218 // Sort by hash value so that blocks with identical end sequences sort
220 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
222 // Walk through equivalence sets looking for actual exact matches.
223 while (MergePotentials.size() > 1) {
224 unsigned CurHash = (MergePotentials.end()-1)->first;
225 unsigned PrevHash = (MergePotentials.end()-2)->first;
226 MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second;
228 // If there is nothing that matches the hash of the current basic block,
230 if (CurHash != PrevHash) {
231 MergePotentials.pop_back();
235 // Determine the actual length of the shared tail between these two basic
236 // blocks. Because the hash can have collisions, it's possible that this is
238 MachineBasicBlock::iterator BBI1, BBI2;
239 unsigned CommonTailLen =
240 ComputeCommonTailLength(CurMBB, (MergePotentials.end()-2)->second,
243 // If the tails don't have at least two instructions in common, see if there
244 // is anything else in the equivalence class that does match.
245 if (CommonTailLen < 2) {
246 unsigned FoundMatch = ~0U;
247 for (int i = MergePotentials.size()-2;
248 i != -1 && MergePotentials[i].first == CurHash; --i) {
249 CommonTailLen = ComputeCommonTailLength(CurMBB,
250 MergePotentials[i].second,
252 if (CommonTailLen >= 2) {
258 // If we didn't find anything that has at least two instructions matching
259 // this one, bail out.
260 if (FoundMatch == ~0U) {
261 MergePotentials.pop_back();
265 // Otherwise, move the matching block to the right position.
266 std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2));
269 // If either block is the entire common tail, make the longer one branch to
271 MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second;
272 if (CurMBB->begin() == BBI1) {
273 // Hack the end off MBB2, making it jump to CurMBB instead.
274 ReplaceTailWithBranchTo(BBI2, CurMBB);
275 // This modifies MBB2, so remove it from the worklist.
276 MergePotentials.erase(MergePotentials.end()-2);
279 } else if (MBB2->begin() == BBI2) {
280 // Hack the end off CurMBB, making it jump to MBBI@ instead.
281 ReplaceTailWithBranchTo(BBI1, MBB2);
282 // This modifies CurMBB, so remove it from the worklist.
283 MergePotentials.pop_back();
288 MergePotentials.pop_back();
295 //===----------------------------------------------------------------------===//
296 // Branch Optimization
297 //===----------------------------------------------------------------------===//
299 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
302 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
303 MachineBasicBlock *MBB = I++;
306 // If it is dead, remove it.
307 if (MBB->pred_empty()) {
308 RemoveDeadBlock(MBB);
317 /// CorrectExtraCFGEdges - Various pieces of code can cause excess edges in the
318 /// CFG to be inserted. If we have proven that MBB can only branch to DestA and
319 /// DestB, remove any other MBB successors from the CFG. DestA and DestB can
321 static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB,
322 MachineBasicBlock *DestA,
323 MachineBasicBlock *DestB,
325 MachineFunction::iterator FallThru) {
326 bool MadeChange = false;
327 bool AddedFallThrough = false;
329 // If this block ends with a conditional branch that falls through to its
330 // successor, set DestB as the successor.
332 if (DestB == 0 && FallThru != MBB.getParent()->end()) {
334 AddedFallThrough = true;
337 // If this is an unconditional branch with no explicit dest, it must just be
338 // a fallthrough into DestB.
339 if (DestA == 0 && FallThru != MBB.getParent()->end()) {
341 AddedFallThrough = true;
345 MachineBasicBlock::pred_iterator SI = MBB.succ_begin();
346 while (SI != MBB.succ_end()) {
350 } else if (*SI == DestB) {
354 // Otherwise, this is a superfluous edge, remove it.
355 MBB.removeSuccessor(SI);
359 if (!AddedFallThrough) {
360 assert(DestA == 0 && DestB == 0 &&
361 "MachineCFG is missing edges!");
363 assert(DestA == 0 && "MachineCFG is missing edges!");
369 /// ReplaceUsesOfBlockWith - Given a machine basic block 'BB' that branched to
370 /// 'Old', change the code and CFG so that it branches to 'New' instead.
371 static void ReplaceUsesOfBlockWith(MachineBasicBlock *BB,
372 MachineBasicBlock *Old,
373 MachineBasicBlock *New,
374 const TargetInstrInfo *TII) {
375 assert(Old != New && "Cannot replace self with self!");
377 MachineBasicBlock::iterator I = BB->end();
378 while (I != BB->begin()) {
380 if (!TII->isTerminatorInstr(I->getOpcode())) break;
382 // Scan the operands of this machine instruction, replacing any uses of Old
384 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
385 if (I->getOperand(i).isMachineBasicBlock() &&
386 I->getOperand(i).getMachineBasicBlock() == Old)
387 I->getOperand(i).setMachineBasicBlock(New);
390 // Update the successor information.
391 std::vector<MachineBasicBlock*> Succs(BB->succ_begin(), BB->succ_end());
392 for (int i = Succs.size()-1; i >= 0; --i)
393 if (Succs[i] == Old) {
394 BB->removeSuccessor(Old);
395 BB->addSuccessor(New);
399 /// CanFallThrough - Return true of the specified branch condition can transfer
400 /// control to FallthroughBlock, the block immediately after the branch.
401 static bool CanFallThrough(MachineBasicBlock *TBB,
402 MachineBasicBlock *FBB,
403 const std::vector<MachineOperand> &Cond,
404 MachineFunction::iterator FallthroughBlock) {
405 // If there is no branch, control always falls through.
406 if (TBB == 0) return true;
408 // If there is some explicit branch to the fallthrough block, it can obviously
409 // reach, even though the branch should get folded to fall through implicitly.
410 if (MachineFunction::iterator(TBB) == FallthroughBlock ||
411 MachineFunction::iterator(FBB) == FallthroughBlock)
414 // If it's an unconditional branch to some block not the fall through, it
415 // doesn't fall through.
416 if (Cond.empty()) return false;
418 // Otherwise, if it is conditional and has no explicit false block, it falls
423 /// OptimizeBlock - Analyze and optimize control flow related to the specified
424 /// block. This is never called on the entry block.
425 void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
426 MachineFunction::iterator FallThrough = MBB;
429 // If this block is empty, make everyone use its fall-through, not the block
432 // Dead block? Leave for cleanup later.
433 if (MBB->pred_empty()) return;
435 if (FallThrough == MBB->getParent()->end()) {
436 // TODO: Simplify preds to not branch here if possible!
438 // Rewrite all predecessors of the old block to go to the fallthrough
440 while (!MBB->pred_empty()) {
441 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
442 ReplaceUsesOfBlockWith(Pred, MBB, FallThrough, TII);
445 // If MBB was the target of a jump table, update jump tables to go to the
446 // fallthrough instead.
447 MBB->getParent()->getJumpTableInfo()->ReplaceMBBInJumpTables(MBB,
454 // Check to see if we can simplify the terminator of the block before this
456 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
458 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
459 std::vector<MachineOperand> PriorCond;
460 bool PriorUnAnalyzable = false;
461 PriorUnAnalyzable = TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
462 if (!PriorUnAnalyzable) {
463 // If the CFG for the prior block has extra edges, remove them.
464 MadeChange |= CorrectExtraCFGEdges(PrevBB, PriorTBB, PriorFBB,
465 !PriorCond.empty(), MBB);
467 // If the previous branch is conditional and both conditions go to the same
468 // destination, remove the branch, replacing it with an unconditional one or
470 if (PriorTBB && PriorTBB == PriorFBB) {
471 TII->RemoveBranch(PrevBB);
474 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
477 return OptimizeBlock(MBB);
480 // If the previous branch *only* branches to *this* block (conditional or
481 // not) remove the branch.
482 if (PriorTBB == MBB && PriorFBB == 0) {
483 TII->RemoveBranch(PrevBB);
486 return OptimizeBlock(MBB);
489 // If the prior block branches somewhere else on the condition and here if
490 // the condition is false, remove the uncond second branch.
491 if (PriorFBB == MBB) {
492 TII->RemoveBranch(PrevBB);
493 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
496 return OptimizeBlock(MBB);
499 // If the prior block branches here on true and somewhere else on false, and
500 // if the branch condition is reversible, reverse the branch to create a
502 if (PriorTBB == MBB) {
503 std::vector<MachineOperand> NewPriorCond(PriorCond);
504 if (!TII->ReverseBranchCondition(NewPriorCond)) {
505 TII->RemoveBranch(PrevBB);
506 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
509 return OptimizeBlock(MBB);
514 // Analyze the branch in the current block.
515 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
516 std::vector<MachineOperand> CurCond;
517 if (!TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond)) {
518 // If the CFG for the prior block has extra edges, remove them.
519 MadeChange |= CorrectExtraCFGEdges(*MBB, CurTBB, CurFBB,
521 ++MachineFunction::iterator(MBB));
523 // If this branch is the only thing in its block, see if we can forward
524 // other blocks across it.
525 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
526 TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) {
527 // This block may contain just an unconditional branch. Because there can
528 // be 'non-branch terminators' in the block, try removing the branch and
529 // then seeing if the block is empty.
530 TII->RemoveBranch(*MBB);
532 // If this block is just an unconditional branch to CurTBB, we can
533 // usually completely eliminate the block. The only case we cannot
534 // completely eliminate the block is when the block before this one
535 // falls through into MBB and we can't understand the prior block's branch
538 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
539 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
540 !PrevBB.isSuccessor(MBB)) {
541 // If the prior block falls through into us, turn it into an
542 // explicit branch to us to make updates simpler.
543 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
544 PriorTBB != MBB && PriorFBB != MBB) {
546 assert(PriorCond.empty() && PriorFBB == 0 && "Bad branch analysis");
549 assert(PriorFBB == 0 && "Machine CFG out of date!");
552 TII->RemoveBranch(PrevBB);
553 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
556 // Iterate through all the predecessors, revectoring each in-turn.
557 MachineBasicBlock::pred_iterator PI = MBB->pred_begin();
558 bool DidChange = false;
559 bool HasBranchToSelf = false;
560 while (PI != MBB->pred_end()) {
562 // If this block has an uncond branch to itself, leave it.
564 HasBranchToSelf = true;
567 ReplaceUsesOfBlockWith(*PI, MBB, CurTBB, TII);
571 // Change any jumptables to go to the new MBB.
572 MBB->getParent()->getJumpTableInfo()->ReplaceMBBInJumpTables(MBB,
577 if (!HasBranchToSelf) return;
582 // Add the branch back if the block is more than just an uncond branch.
583 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
586 // If the prior block doesn't fall through into this block, and if this
587 // block doesn't fall through into some other block, see if we can find a
588 // place to move this block where a fall-through will happen.
589 if (!PriorUnAnalyzable && !CanFallThrough(PriorTBB, PriorFBB,
591 // Now we know that there was no fall-through into this block, check to
592 // see if it has fall-throughs.
593 if (!CanFallThrough(CurTBB, CurFBB, CurCond, FallThrough)) {
595 // Check all the predecessors of this block. If one of them has no fall
596 // throughs, move this block right after it.
597 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
598 E = MBB->pred_end(); PI != E; ++PI) {
599 // Analyze the branch at the end of the pred.
600 MachineBasicBlock *PredBB = *PI;
601 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
602 MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
603 std::vector<MachineOperand> PredCond;
605 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond) &&
606 !CanFallThrough(PredTBB, PredFBB, PredCond, PredFallthrough)) {
607 MBB->moveAfter(PredBB);
609 return OptimizeBlock(MBB);
613 // Check all successors to see if we can move this block before it.
614 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
615 E = MBB->succ_end(); SI != E; ++SI) {
616 // Analyze the branch at the end of the block before the succ.
617 MachineBasicBlock *SuccBB = *SI;
618 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
619 MachineBasicBlock *SuccPrevTBB = 0, *SuccPrevFBB = 0;
620 std::vector<MachineOperand> SuccPrevCond;
622 !TII->AnalyzeBranch(*SuccPrev, SuccPrevTBB, SuccPrevFBB,
624 !CanFallThrough(SuccPrevTBB, SuccPrevFBB, SuccPrevCond, SuccBB)) {
625 MBB->moveBefore(SuccBB);
627 return OptimizeBlock(MBB);
631 // Okay, there is no really great place to put this block. If, however,
632 // the block before this one would be a fall-through if this block were
633 // removed, move this block to the end of the function.
634 if (FallThrough != MBB->getParent()->end() &&
635 CanFallThrough(PriorTBB, PriorFBB, PriorCond, FallThrough)) {
636 MBB->moveAfter(--MBB->getParent()->end());