1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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 "BranchFolding.h"
21 #include "llvm/Function.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/RegisterScavenging.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Target/TargetRegisterInfo.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/ADT/SmallSet.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/ADT/STLExtras.h"
40 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
41 STATISTIC(NumBranchOpts, "Number of branches optimized");
42 STATISTIC(NumTailMerge , "Number of block tails merged");
43 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
44 cl::init(cl::BOU_UNSET), cl::Hidden);
45 // Throttle for huge numbers of predecessors (compile speed problems)
46 static cl::opt<unsigned>
47 TailMergeThreshold("tail-merge-threshold",
48 cl::desc("Max number of predecessors to consider tail merging"),
49 cl::init(150), cl::Hidden);
52 char BranchFolderPass::ID = 0;
54 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
55 return new BranchFolderPass(DefaultEnableTailMerge);
58 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
59 return OptimizeFunction(MF,
60 MF.getTarget().getInstrInfo(),
61 MF.getTarget().getRegisterInfo(),
62 getAnalysisIfAvailable<MachineModuleInfo>());
67 BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
68 switch (FlagEnableTailMerge) {
69 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
70 case cl::BOU_TRUE: EnableTailMerge = true; break;
71 case cl::BOU_FALSE: EnableTailMerge = false; break;
75 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
76 /// function, updating the CFG.
77 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
78 assert(MBB->pred_empty() && "MBB must be dead!");
79 DEBUG(errs() << "\nRemoving MBB: " << *MBB);
81 MachineFunction *MF = MBB->getParent();
82 // drop all successors.
83 while (!MBB->succ_empty())
84 MBB->removeSuccessor(MBB->succ_end()-1);
86 // If there are any labels in the basic block, unregister them from
88 if (MMI && !MBB->empty()) {
89 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
92 // The label ID # is always operand #0, an immediate.
93 MMI->InvalidateLabel(I->getOperand(0).getImm());
101 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
102 /// followed by terminators, and if the implicitly defined registers are not
103 /// used by the terminators, remove those implicit_def's. e.g.
105 /// r0 = implicit_def
106 /// r1 = implicit_def
108 /// This block can be optimized away later if the implicit instructions are
110 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
111 SmallSet<unsigned, 4> ImpDefRegs;
112 MachineBasicBlock::iterator I = MBB->begin();
113 while (I != MBB->end()) {
114 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF)
116 unsigned Reg = I->getOperand(0).getReg();
117 ImpDefRegs.insert(Reg);
118 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
119 unsigned SubReg = *SubRegs; ++SubRegs)
120 ImpDefRegs.insert(SubReg);
123 if (ImpDefRegs.empty())
126 MachineBasicBlock::iterator FirstTerm = I;
127 while (I != MBB->end()) {
128 if (!TII->isUnpredicatedTerminator(I))
130 // See if it uses any of the implicitly defined registers.
131 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
132 MachineOperand &MO = I->getOperand(i);
133 if (!MO.isReg() || !MO.isUse())
135 unsigned Reg = MO.getReg();
136 if (ImpDefRegs.count(Reg))
143 while (I != FirstTerm) {
144 MachineInstr *ImpDefMI = &*I;
146 MBB->erase(ImpDefMI);
152 /// OptimizeFunction - Perhaps branch folding, tail merging and other
153 /// CFG optimizations on the given function.
154 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
155 const TargetInstrInfo *tii,
156 const TargetRegisterInfo *tri,
157 MachineModuleInfo *mmi) {
158 if (!tii) return false;
164 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
166 // Fix CFG. The later algorithms expect it to be right.
167 bool MadeChange = false;
168 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
169 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
170 SmallVector<MachineOperand, 4> Cond;
171 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
172 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
173 MadeChange |= OptimizeImpDefsBlock(MBB);
177 bool MadeChangeThisIteration = true;
178 while (MadeChangeThisIteration) {
179 MadeChangeThisIteration = false;
180 MadeChangeThisIteration |= TailMergeBlocks(MF);
181 MadeChangeThisIteration |= OptimizeBranches(MF);
182 MadeChange |= MadeChangeThisIteration;
185 // See if any jump tables have become mergable or dead as the code generator
187 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
188 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
190 // Figure out how these jump tables should be merged.
191 std::vector<unsigned> JTMapping;
192 JTMapping.reserve(JTs.size());
194 // We always keep the 0th jump table.
195 JTMapping.push_back(0);
197 // Scan the jump tables, seeing if there are any duplicates. Note that this
198 // is N^2, which should be fixed someday.
199 for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
200 if (JTs[i].MBBs.empty())
201 JTMapping.push_back(i);
203 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
206 // If a jump table was merge with another one, walk the function rewriting
207 // references to jump tables to reference the new JT ID's. Keep track of
208 // whether we see a jump table idx, if not, we can delete the JT.
209 BitVector JTIsLive(JTs.size());
210 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
212 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
214 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
215 MachineOperand &Op = I->getOperand(op);
216 if (!Op.isJTI()) continue;
217 unsigned NewIdx = JTMapping[Op.getIndex()];
220 // Remember that this JT is live.
221 JTIsLive.set(NewIdx);
225 // Finally, remove dead jump tables. This happens either because the
226 // indirect jump was unreachable (and thus deleted) or because the jump
227 // table was merged with some other one.
228 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
229 if (!JTIsLive.test(i)) {
230 JTI->RemoveJumpTable(i);
239 //===----------------------------------------------------------------------===//
240 // Tail Merging of Blocks
241 //===----------------------------------------------------------------------===//
243 /// HashMachineInstr - Compute a hash value for MI and its operands.
244 static unsigned HashMachineInstr(const MachineInstr *MI) {
245 unsigned Hash = MI->getOpcode();
246 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
247 const MachineOperand &Op = MI->getOperand(i);
249 // Merge in bits from the operand if easy.
250 unsigned OperandHash = 0;
251 switch (Op.getType()) {
252 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
253 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
254 case MachineOperand::MO_MachineBasicBlock:
255 OperandHash = Op.getMBB()->getNumber();
257 case MachineOperand::MO_FrameIndex:
258 case MachineOperand::MO_ConstantPoolIndex:
259 case MachineOperand::MO_JumpTableIndex:
260 OperandHash = Op.getIndex();
262 case MachineOperand::MO_GlobalAddress:
263 case MachineOperand::MO_ExternalSymbol:
264 // Global address / external symbol are too hard, don't bother, but do
265 // pull in the offset.
266 OperandHash = Op.getOffset();
271 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
276 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
277 /// with no successors, we hash two instructions, because cross-jumping
278 /// only saves code when at least two instructions are removed (since a
279 /// branch must be inserted). For blocks with a successor, one of the
280 /// two blocks to be tail-merged will end with a branch already, so
281 /// it gains to cross-jump even for one instruction.
283 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
284 unsigned minCommonTailLength) {
285 MachineBasicBlock::const_iterator I = MBB->end();
286 if (I == MBB->begin())
287 return 0; // Empty MBB.
290 unsigned Hash = HashMachineInstr(I);
292 if (I == MBB->begin() || minCommonTailLength == 1)
293 return Hash; // Single instr MBB.
296 // Hash in the second-to-last instruction.
297 Hash ^= HashMachineInstr(I) << 2;
301 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
302 /// of instructions they actually have in common together at their end. Return
303 /// iterators for the first shared instruction in each block.
304 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
305 MachineBasicBlock *MBB2,
306 MachineBasicBlock::iterator &I1,
307 MachineBasicBlock::iterator &I2) {
311 unsigned TailLen = 0;
312 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
314 if (!I1->isIdenticalTo(I2) ||
315 // FIXME: This check is dubious. It's used to get around a problem where
316 // people incorrectly expect inline asm directives to remain in the same
317 // relative order. This is untenable because normal compiler
318 // optimizations (like this one) may reorder and/or merge these
320 I1->getOpcode() == TargetInstrInfo::INLINEASM) {
329 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
330 /// after it, replacing it with an unconditional branch to NewDest. This
331 /// returns true if OldInst's block is modified, false if NewDest is modified.
332 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
333 MachineBasicBlock *NewDest) {
334 MachineBasicBlock *OldBB = OldInst->getParent();
336 // Remove all the old successors of OldBB from the CFG.
337 while (!OldBB->succ_empty())
338 OldBB->removeSuccessor(OldBB->succ_begin());
340 // Remove all the dead instructions from the end of OldBB.
341 OldBB->erase(OldInst, OldBB->end());
343 // If OldBB isn't immediately before OldBB, insert a branch to it.
344 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
345 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
346 OldBB->addSuccessor(NewDest);
350 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
351 /// MBB so that the part before the iterator falls into the part starting at the
352 /// iterator. This returns the new MBB.
353 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
354 MachineBasicBlock::iterator BBI1) {
355 MachineFunction &MF = *CurMBB.getParent();
357 // Create the fall-through block.
358 MachineFunction::iterator MBBI = &CurMBB;
359 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
360 CurMBB.getParent()->insert(++MBBI, NewMBB);
362 // Move all the successors of this block to the specified block.
363 NewMBB->transferSuccessors(&CurMBB);
365 // Add an edge from CurMBB to NewMBB for the fall-through.
366 CurMBB.addSuccessor(NewMBB);
368 // Splice the code over.
369 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
371 // For targets that use the register scavenger, we must maintain LiveIns.
373 RS->enterBasicBlock(&CurMBB);
375 RS->forward(prior(CurMBB.end()));
376 BitVector RegsLiveAtExit(TRI->getNumRegs());
377 RS->getRegsUsed(RegsLiveAtExit, false);
378 for (unsigned int i=0, e=TRI->getNumRegs(); i!=e; i++)
379 if (RegsLiveAtExit[i])
380 NewMBB->addLiveIn(i);
386 /// EstimateRuntime - Make a rough estimate for how long it will take to run
387 /// the specified code.
388 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
389 MachineBasicBlock::iterator E) {
391 for (; I != E; ++I) {
392 const TargetInstrDesc &TID = I->getDesc();
395 else if (TID.mayLoad() || TID.mayStore())
403 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
404 // branches temporarily for tail merging). In the case where CurMBB ends
405 // with a conditional branch to the next block, optimize by reversing the
406 // test and conditionally branching to SuccMBB instead.
408 static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB,
409 const TargetInstrInfo *TII) {
410 MachineFunction *MF = CurMBB->getParent();
411 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
412 MachineBasicBlock *TBB = 0, *FBB = 0;
413 SmallVector<MachineOperand, 4> Cond;
414 if (I != MF->end() &&
415 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
416 MachineBasicBlock *NextBB = I;
417 if (TBB == NextBB && !Cond.empty() && !FBB) {
418 if (!TII->ReverseBranchCondition(Cond)) {
419 TII->RemoveBranch(*CurMBB);
420 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
425 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
428 static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p,
429 const std::pair<unsigned,MachineBasicBlock*> &q) {
430 if (p.first < q.first)
432 else if (p.first > q.first)
434 else if (p.second->getNumber() < q.second->getNumber())
436 else if (p.second->getNumber() > q.second->getNumber())
439 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
440 // an object with itself.
441 #ifndef _GLIBCXX_DEBUG
442 llvm_unreachable("Predecessor appears twice");
448 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
449 /// and decide if it would be profitable to merge those tails. Return the
450 /// length of the common tail and iterators to the first common instruction
452 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
453 MachineBasicBlock *MBB2,
454 unsigned minCommonTailLength,
455 unsigned &CommonTailLen,
456 MachineBasicBlock::iterator &I1,
457 MachineBasicBlock::iterator &I2) {
458 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
459 MachineFunction *MF = MBB1->getParent();
461 if (CommonTailLen >= minCommonTailLength)
464 if (CommonTailLen == 0)
467 // If we are optimizing for code size, 1 instruction in common is enough if
468 // we don't have to split a block. At worst we will be replacing a
469 // fallthrough into the common tail with a branch, which at worst breaks
470 // even with falling through into the duplicated common tail.
471 if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
472 (I1 == MBB1->begin() || I2 == MBB2->begin()))
478 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
479 /// hash CurHash (guaranteed to match the last element). Build the vector
480 /// SameTails of all those that have the (same) largest number of instructions
481 /// in common of any pair of these blocks. SameTails entries contain an
482 /// iterator into MergePotentials (from which the MachineBasicBlock can be
483 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
484 /// instruction where the matching code sequence begins.
485 /// Order of elements in SameTails is the reverse of the order in which
486 /// those blocks appear in MergePotentials (where they are not necessarily
488 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
489 unsigned minCommonTailLength) {
490 unsigned maxCommonTailLength = 0U;
492 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
493 MPIterator HighestMPIter = prior(MergePotentials.end());
494 for (MPIterator CurMPIter = prior(MergePotentials.end()),
495 B = MergePotentials.begin();
496 CurMPIter!=B && CurMPIter->first==CurHash;
498 for (MPIterator I = prior(CurMPIter); I->first==CurHash ; --I) {
499 unsigned CommonTailLen;
500 if (ProfitableToMerge(CurMPIter->second, I->second, minCommonTailLength,
501 CommonTailLen, TrialBBI1, TrialBBI2)) {
502 if (CommonTailLen > maxCommonTailLength) {
504 maxCommonTailLength = CommonTailLen;
505 HighestMPIter = CurMPIter;
506 SameTails.push_back(std::make_pair(CurMPIter, TrialBBI1));
508 if (HighestMPIter == CurMPIter &&
509 CommonTailLen == maxCommonTailLength)
510 SameTails.push_back(std::make_pair(I, TrialBBI2));
516 return maxCommonTailLength;
519 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
520 /// MergePotentials, restoring branches at ends of blocks as appropriate.
521 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
522 MachineBasicBlock* SuccBB,
523 MachineBasicBlock* PredBB) {
524 MPIterator CurMPIter, B;
525 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
526 CurMPIter->first==CurHash;
528 // Put the unconditional branch back, if we need one.
529 MachineBasicBlock *CurMBB = CurMPIter->second;
530 if (SuccBB && CurMBB != PredBB)
531 FixTail(CurMBB, SuccBB, TII);
535 if (CurMPIter->first!=CurHash)
537 MergePotentials.erase(CurMPIter, MergePotentials.end());
540 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
541 /// only of the common tail. Create a block that does by splitting one.
542 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
543 unsigned maxCommonTailLength) {
544 unsigned i, commonTailIndex;
545 unsigned TimeEstimate = ~0U;
546 for (i=0, commonTailIndex=0; i<SameTails.size(); i++) {
547 // Use PredBB if possible; that doesn't require a new branch.
548 if (SameTails[i].first->second==PredBB) {
552 // Otherwise, make a (fairly bogus) choice based on estimate of
553 // how long it will take the various blocks to execute.
554 unsigned t = EstimateRuntime(SameTails[i].first->second->begin(),
555 SameTails[i].second);
556 if (t<=TimeEstimate) {
562 MachineBasicBlock::iterator BBI = SameTails[commonTailIndex].second;
563 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second;
565 DEBUG(errs() << "\nSplitting " << MBB->getNumber() << ", size "
566 << maxCommonTailLength);
568 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
569 SameTails[commonTailIndex].first->second = newMBB;
570 SameTails[commonTailIndex].second = newMBB->begin();
571 // If we split PredBB, newMBB is the new predecessor.
575 return commonTailIndex;
578 // See if any of the blocks in MergePotentials (which all have a common single
579 // successor, or all have no successor) can be tail-merged. If there is a
580 // successor, any blocks in MergePotentials that are not tail-merged and
581 // are not immediately before Succ must have an unconditional branch to
582 // Succ added (but the predecessor/successor lists need no adjustment).
583 // The lone predecessor of Succ that falls through into Succ,
584 // if any, is given in PredBB.
586 bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB,
587 MachineBasicBlock* PredBB) {
588 bool MadeChange = false;
590 // It doesn't make sense to save a single instruction since tail merging
592 // FIXME: Ask the target to provide the threshold?
593 unsigned minCommonTailLength = (SuccBB ? 1 : 2) + 1;
595 DEBUG(errs() << "\nTryMergeBlocks " << MergePotentials.size() << '\n');
597 // Sort by hash value so that blocks with identical end sequences sort
599 std::stable_sort(MergePotentials.begin(), MergePotentials.end(),MergeCompare);
601 // Walk through equivalence sets looking for actual exact matches.
602 while (MergePotentials.size() > 1) {
603 unsigned CurHash = prior(MergePotentials.end())->first;
605 // Build SameTails, identifying the set of blocks with this hash code
606 // and with the maximum number of instructions in common.
607 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
608 minCommonTailLength);
610 // If we didn't find any pair that has at least minCommonTailLength
611 // instructions in common, remove all blocks with this hash code and retry.
612 if (SameTails.empty()) {
613 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
617 // If one of the blocks is the entire common tail (and not the entry
618 // block, which we can't jump to), we can treat all blocks with this same
619 // tail at once. Use PredBB if that is one of the possibilities, as that
620 // will not introduce any extra branches.
621 MachineBasicBlock *EntryBB = MergePotentials.begin()->second->
622 getParent()->begin();
623 unsigned int commonTailIndex, i;
624 for (commonTailIndex=SameTails.size(), i=0; i<SameTails.size(); i++) {
625 MachineBasicBlock *MBB = SameTails[i].first->second;
626 if (MBB->begin() == SameTails[i].second && MBB != EntryBB) {
633 if (commonTailIndex==SameTails.size()) {
634 // None of the blocks consist entirely of the common tail.
635 // Split a block so that one does.
636 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
639 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second;
640 // MBB is common tail. Adjust all other BB's to jump to this one.
641 // Traversal must be forwards so erases work.
642 DEBUG(errs() << "\nUsing common tail " << MBB->getNumber() << " for ");
643 for (unsigned int i=0; i<SameTails.size(); ++i) {
644 if (commonTailIndex==i)
646 DEBUG(errs() << SameTails[i].first->second->getNumber() << ",");
647 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
648 ReplaceTailWithBranchTo(SameTails[i].second, MBB);
649 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
650 MergePotentials.erase(SameTails[i].first);
652 DEBUG(errs() << "\n");
653 // We leave commonTailIndex in the worklist in case there are other blocks
654 // that match it with a smaller number of instructions.
660 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
662 if (!EnableTailMerge) return false;
664 bool MadeChange = false;
666 // First find blocks with no successors.
667 MergePotentials.clear();
668 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
670 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I));
672 // See if we can do any tail merging on those.
673 if (MergePotentials.size() < TailMergeThreshold &&
674 MergePotentials.size() >= 2)
675 MadeChange |= TryMergeBlocks(NULL, NULL);
677 // Look at blocks (IBB) with multiple predecessors (PBB).
678 // We change each predecessor to a canonical form, by
679 // (1) temporarily removing any unconditional branch from the predecessor
681 // (2) alter conditional branches so they branch to the other block
682 // not IBB; this may require adding back an unconditional branch to IBB
683 // later, where there wasn't one coming in. E.g.
685 // fallthrough to QBB
688 // with a conceptual B to IBB after that, which never actually exists.
689 // With those changes, we see whether the predecessors' tails match,
690 // and merge them if so. We change things out of canonical form and
691 // back to the way they were later in the process. (OptimizeBranches
692 // would undo some of this, but we can't use it, because we'd get into
693 // a compile-time infinite loop repeatedly doing and undoing the same
696 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
697 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
698 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
699 MachineBasicBlock *IBB = I;
700 MachineBasicBlock *PredBB = prior(I);
701 MergePotentials.clear();
702 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
705 MachineBasicBlock* PBB = *P;
706 // Skip blocks that loop to themselves, can't tail merge these.
709 // Visit each predecessor only once.
710 if (!UniquePreds.insert(PBB))
712 MachineBasicBlock *TBB = 0, *FBB = 0;
713 SmallVector<MachineOperand, 4> Cond;
714 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
715 // Failing case: IBB is the target of a cbr, and
716 // we cannot reverse the branch.
717 SmallVector<MachineOperand, 4> NewCond(Cond);
718 if (!Cond.empty() && TBB==IBB) {
719 if (TII->ReverseBranchCondition(NewCond))
721 // This is the QBB case described above
723 FBB = next(MachineFunction::iterator(PBB));
725 // Failing case: the only way IBB can be reached from PBB is via
726 // exception handling. Happens for landing pads. Would be nice
727 // to have a bit in the edge so we didn't have to do all this.
728 if (IBB->isLandingPad()) {
729 MachineFunction::iterator IP = PBB; IP++;
730 MachineBasicBlock* PredNextBB = NULL;
734 if (IBB!=PredNextBB) // fallthrough
737 if (TBB!=IBB && FBB!=IBB) // cbr then ubr
739 } else if (Cond.empty()) {
743 if (TBB!=IBB && IBB!=PredNextBB) // cbr
747 // Remove the unconditional branch at the end, if any.
748 if (TBB && (Cond.empty() || FBB)) {
749 TII->RemoveBranch(*PBB);
751 // reinsert conditional branch only, for now
752 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond);
754 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P));
757 if (MergePotentials.size() >= 2)
758 MadeChange |= TryMergeBlocks(I, PredBB);
759 // Reinsert an unconditional branch if needed.
760 // The 1 below can occur as a result of removing blocks in TryMergeBlocks.
761 PredBB = prior(I); // this may have been changed in TryMergeBlocks
762 if (MergePotentials.size()==1 &&
763 MergePotentials.begin()->second != PredBB)
764 FixTail(MergePotentials.begin()->second, I, TII);
770 //===----------------------------------------------------------------------===//
771 // Branch Optimization
772 //===----------------------------------------------------------------------===//
774 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
775 bool MadeChange = false;
777 // Make sure blocks are numbered in order
780 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
781 MachineBasicBlock *MBB = I++;
782 MadeChange |= OptimizeBlock(MBB);
784 // If it is dead, remove it.
785 if (MBB->pred_empty()) {
786 RemoveDeadBlock(MBB);
795 /// CanFallThrough - Return true if the specified block (with the specified
796 /// branch condition) can implicitly transfer control to the block after it by
797 /// falling off the end of it. This should return false if it can reach the
798 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
800 /// True is a conservative answer.
802 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
803 bool BranchUnAnalyzable,
804 MachineBasicBlock *TBB,
805 MachineBasicBlock *FBB,
806 const SmallVectorImpl<MachineOperand> &Cond) {
807 MachineFunction::iterator Fallthrough = CurBB;
809 // If FallthroughBlock is off the end of the function, it can't fall through.
810 if (Fallthrough == CurBB->getParent()->end())
813 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
814 if (!CurBB->isSuccessor(Fallthrough))
817 // If we couldn't analyze the branch, assume it could fall through.
818 if (BranchUnAnalyzable) return true;
820 // If there is no branch, control always falls through.
821 if (TBB == 0) return true;
823 // If there is some explicit branch to the fallthrough block, it can obviously
824 // reach, even though the branch should get folded to fall through implicitly.
825 if (MachineFunction::iterator(TBB) == Fallthrough ||
826 MachineFunction::iterator(FBB) == Fallthrough)
829 // If it's an unconditional branch to some block not the fall through, it
830 // doesn't fall through.
831 if (Cond.empty()) return false;
833 // Otherwise, if it is conditional and has no explicit false block, it falls
838 /// CanFallThrough - Return true if the specified can implicitly transfer
839 /// control to the block after it by falling off the end of it. This should
840 /// return false if it can reach the block after it, but it uses an explicit
841 /// branch to do so (e.g. a table jump).
843 /// True is a conservative answer.
845 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
846 MachineBasicBlock *TBB = 0, *FBB = 0;
847 SmallVector<MachineOperand, 4> Cond;
848 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true);
849 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
852 /// IsBetterFallthrough - Return true if it would be clearly better to
853 /// fall-through to MBB1 than to fall through into MBB2. This has to return
854 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
855 /// result in infinite loops.
856 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
857 MachineBasicBlock *MBB2) {
858 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
859 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
860 // optimize branches that branch to either a return block or an assert block
861 // into a fallthrough to the return.
862 if (MBB1->empty() || MBB2->empty()) return false;
864 // If there is a clear successor ordering we make sure that one block
865 // will fall through to the next
866 if (MBB1->isSuccessor(MBB2)) return true;
867 if (MBB2->isSuccessor(MBB1)) return false;
869 MachineInstr *MBB1I = --MBB1->end();
870 MachineInstr *MBB2I = --MBB2->end();
871 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
874 /// OptimizeBlock - Analyze and optimize control flow related to the specified
875 /// block. This is never called on the entry block.
876 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
877 bool MadeChange = false;
879 MachineFunction::iterator FallThrough = MBB;
882 // If this block is empty, make everyone use its fall-through, not the block
883 // explicitly. Landing pads should not do this since the landing-pad table
884 // points to this block. Blocks with their addresses taken shouldn't be
886 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
887 // Dead block? Leave for cleanup later.
888 if (MBB->pred_empty()) return MadeChange;
890 if (FallThrough == MBB->getParent()->end()) {
891 // TODO: Simplify preds to not branch here if possible!
893 // Rewrite all predecessors of the old block to go to the fallthrough
895 while (!MBB->pred_empty()) {
896 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
897 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
899 // If MBB was the target of a jump table, update jump tables to go to the
900 // fallthrough instead.
901 MBB->getParent()->getJumpTableInfo()->
902 ReplaceMBBInJumpTables(MBB, FallThrough);
908 // Check to see if we can simplify the terminator of the block before this
910 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
912 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
913 SmallVector<MachineOperand, 4> PriorCond;
914 bool PriorUnAnalyzable =
915 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
916 if (!PriorUnAnalyzable) {
917 // If the CFG for the prior block has extra edges, remove them.
918 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
921 // If the previous branch is conditional and both conditions go to the same
922 // destination, remove the branch, replacing it with an unconditional one or
924 if (PriorTBB && PriorTBB == PriorFBB) {
925 TII->RemoveBranch(PrevBB);
928 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
931 return OptimizeBlock(MBB);
934 // If the previous branch *only* branches to *this* block (conditional or
935 // not) remove the branch.
936 if (PriorTBB == MBB && PriorFBB == 0) {
937 TII->RemoveBranch(PrevBB);
940 return OptimizeBlock(MBB);
943 // If the prior block branches somewhere else on the condition and here if
944 // the condition is false, remove the uncond second branch.
945 if (PriorFBB == MBB) {
946 TII->RemoveBranch(PrevBB);
947 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
950 return OptimizeBlock(MBB);
953 // If the prior block branches here on true and somewhere else on false, and
954 // if the branch condition is reversible, reverse the branch to create a
956 if (PriorTBB == MBB) {
957 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
958 if (!TII->ReverseBranchCondition(NewPriorCond)) {
959 TII->RemoveBranch(PrevBB);
960 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
963 return OptimizeBlock(MBB);
967 // If this block has no successors (e.g. it is a return block or ends with
968 // a call to a no-return function like abort or __cxa_throw) and if the pred
969 // falls through into this block, and if it would otherwise fall through
970 // into the block after this, move this block to the end of the function.
972 // We consider it more likely that execution will stay in the function (e.g.
973 // due to loops) than it is to exit it. This asserts in loops etc, moving
974 // the assert condition out of the loop body.
975 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
976 MachineFunction::iterator(PriorTBB) == FallThrough &&
977 !CanFallThrough(MBB)) {
978 bool DoTransform = true;
980 // We have to be careful that the succs of PredBB aren't both no-successor
981 // blocks. If neither have successors and if PredBB is the second from
982 // last block in the function, we'd just keep swapping the two blocks for
983 // last. Only do the swap if one is clearly better to fall through than
985 if (FallThrough == --MBB->getParent()->end() &&
986 !IsBetterFallthrough(PriorTBB, MBB))
989 // We don't want to do this transformation if we have control flow like:
998 // In this case, we could actually be moving the return block *into* a
1000 if (DoTransform && !MBB->succ_empty() &&
1001 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
1002 DoTransform = false;
1006 // Reverse the branch so we will fall through on the previous true cond.
1007 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1008 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1009 DEBUG(errs() << "\nMoving MBB: " << *MBB
1010 << "To make fallthrough to: " << *PriorTBB << "\n");
1012 TII->RemoveBranch(PrevBB);
1013 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1015 // Move this block to the end of the function.
1016 MBB->moveAfter(--MBB->getParent()->end());
1025 // Analyze the branch in the current block.
1026 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1027 SmallVector<MachineOperand, 4> CurCond;
1028 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1029 if (!CurUnAnalyzable) {
1030 // If the CFG for the prior block has extra edges, remove them.
1031 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1033 // If this is a two-way branch, and the FBB branches to this block, reverse
1034 // the condition so the single-basic-block loop is faster. Instead of:
1035 // Loop: xxx; jcc Out; jmp Loop
1037 // Loop: xxx; jncc Loop; jmp Out
1038 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1039 SmallVector<MachineOperand, 4> NewCond(CurCond);
1040 if (!TII->ReverseBranchCondition(NewCond)) {
1041 TII->RemoveBranch(*MBB);
1042 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1045 return OptimizeBlock(MBB);
1050 // If this branch is the only thing in its block, see if we can forward
1051 // other blocks across it.
1052 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1053 MBB->begin()->getDesc().isBranch() && CurTBB != MBB &&
1054 !MBB->hasAddressTaken()) {
1055 // This block may contain just an unconditional branch. Because there can
1056 // be 'non-branch terminators' in the block, try removing the branch and
1057 // then seeing if the block is empty.
1058 TII->RemoveBranch(*MBB);
1060 // If this block is just an unconditional branch to CurTBB, we can
1061 // usually completely eliminate the block. The only case we cannot
1062 // completely eliminate the block is when the block before this one
1063 // falls through into MBB and we can't understand the prior block's branch
1066 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
1067 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1068 !PrevBB.isSuccessor(MBB)) {
1069 // If the prior block falls through into us, turn it into an
1070 // explicit branch to us to make updates simpler.
1071 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1072 PriorTBB != MBB && PriorFBB != MBB) {
1073 if (PriorTBB == 0) {
1074 assert(PriorCond.empty() && PriorFBB == 0 &&
1075 "Bad branch analysis");
1078 assert(PriorFBB == 0 && "Machine CFG out of date!");
1081 TII->RemoveBranch(PrevBB);
1082 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1085 // Iterate through all the predecessors, revectoring each in-turn.
1087 bool DidChange = false;
1088 bool HasBranchToSelf = false;
1089 while(PI != MBB->pred_size()) {
1090 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1092 // If this block has an uncond branch to itself, leave it.
1094 HasBranchToSelf = true;
1097 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1098 // If this change resulted in PMBB ending in a conditional
1099 // branch where both conditions go to the same destination,
1100 // change this to an unconditional branch (and fix the CFG).
1101 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1102 SmallVector<MachineOperand, 4> NewCurCond;
1103 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1104 NewCurFBB, NewCurCond, true);
1105 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1106 TII->RemoveBranch(*PMBB);
1108 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1111 PMBB->CorrectExtraCFGEdges(NewCurTBB, NewCurFBB, false);
1116 // Change any jumptables to go to the new MBB.
1117 MBB->getParent()->getJumpTableInfo()->
1118 ReplaceMBBInJumpTables(MBB, CurTBB);
1122 if (!HasBranchToSelf) return MadeChange;
1127 // Add the branch back if the block is more than just an uncond branch.
1128 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1132 // If the prior block doesn't fall through into this block, and if this
1133 // block doesn't fall through into some other block, see if we can find a
1134 // place to move this block where a fall-through will happen.
1135 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
1136 PriorTBB, PriorFBB, PriorCond)) {
1137 // Now we know that there was no fall-through into this block, check to
1138 // see if it has a fall-through into its successor.
1139 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1142 if (!MBB->isLandingPad()) {
1143 // Check all the predecessors of this block. If one of them has no fall
1144 // throughs, move this block right after it.
1145 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1146 E = MBB->pred_end(); PI != E; ++PI) {
1147 // Analyze the branch at the end of the pred.
1148 MachineBasicBlock *PredBB = *PI;
1149 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1150 if (PredBB != MBB && !CanFallThrough(PredBB)
1151 && (!CurFallsThru || !CurTBB || !CurFBB)
1152 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1153 // If the current block doesn't fall through, just move it.
1154 // If the current block can fall through and does not end with a
1155 // conditional branch, we need to append an unconditional jump to
1156 // the (current) next block. To avoid a possible compile-time
1157 // infinite loop, move blocks only backward in this case.
1158 // Also, if there are already 2 branches here, we cannot add a third;
1159 // this means we have the case
1164 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1166 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1168 MBB->moveAfter(PredBB);
1170 return OptimizeBlock(MBB);
1175 if (!CurFallsThru) {
1176 // Check all successors to see if we can move this block before it.
1177 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1178 E = MBB->succ_end(); SI != E; ++SI) {
1179 // Analyze the branch at the end of the block before the succ.
1180 MachineBasicBlock *SuccBB = *SI;
1181 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1182 std::vector<MachineOperand> SuccPrevCond;
1184 // If this block doesn't already fall-through to that successor, and if
1185 // the succ doesn't already have a block that can fall through into it,
1186 // and if the successor isn't an EH destination, we can arrange for the
1187 // fallthrough to happen.
1188 if (SuccBB != MBB && !CanFallThrough(SuccPrev) &&
1189 !SuccBB->isLandingPad()) {
1190 MBB->moveBefore(SuccBB);
1192 return OptimizeBlock(MBB);
1196 // Okay, there is no really great place to put this block. If, however,
1197 // the block before this one would be a fall-through if this block were
1198 // removed, move this block to the end of the function.
1199 if (FallThrough != MBB->getParent()->end() &&
1200 PrevBB.isSuccessor(FallThrough)) {
1201 MBB->moveAfter(--MBB->getParent()->end());