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/SetVector.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/ADT/STLExtras.h"
41 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
42 STATISTIC(NumBranchOpts, "Number of branches optimized");
43 STATISTIC(NumTailMerge , "Number of block tails merged");
44 STATISTIC(NumTailDups , "Number of tail duplicated blocks");
46 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
47 cl::init(cl::BOU_UNSET), cl::Hidden);
49 // Throttle for huge numbers of predecessors (compile speed problems)
50 static cl::opt<unsigned>
51 TailMergeThreshold("tail-merge-threshold",
52 cl::desc("Max number of predecessors to consider tail merging"),
53 cl::init(150), cl::Hidden);
55 // Heuristic for tail merging (and, inversely, tail duplication).
56 // TODO: This should be replaced with a target query.
57 static cl::opt<unsigned>
58 TailMergeSize("tail-merge-size",
59 cl::desc("Min number of instructions to consider tail merging"),
60 cl::init(3), cl::Hidden);
63 /// BranchFolderPass - Wrap branch folder in a machine function pass.
64 class BranchFolderPass : public MachineFunctionPass,
68 explicit BranchFolderPass(bool defaultEnableTailMerge)
69 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {}
71 virtual bool runOnMachineFunction(MachineFunction &MF);
72 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
76 char BranchFolderPass::ID = 0;
78 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
79 return new BranchFolderPass(DefaultEnableTailMerge);
82 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
83 return OptimizeFunction(MF,
84 MF.getTarget().getInstrInfo(),
85 MF.getTarget().getRegisterInfo(),
86 getAnalysisIfAvailable<MachineModuleInfo>());
90 BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
91 switch (FlagEnableTailMerge) {
92 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
93 case cl::BOU_TRUE: EnableTailMerge = true; break;
94 case cl::BOU_FALSE: EnableTailMerge = false; break;
98 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
99 /// function, updating the CFG.
100 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
101 assert(MBB->pred_empty() && "MBB must be dead!");
102 DEBUG(errs() << "\nRemoving MBB: " << *MBB);
104 MachineFunction *MF = MBB->getParent();
105 // drop all successors.
106 while (!MBB->succ_empty())
107 MBB->removeSuccessor(MBB->succ_end()-1);
109 // If there are any labels in the basic block, unregister them from
110 // MachineModuleInfo.
111 if (MMI && !MBB->empty()) {
112 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
115 // The label ID # is always operand #0, an immediate.
116 MMI->InvalidateLabel(I->getOperand(0).getImm());
124 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
125 /// followed by terminators, and if the implicitly defined registers are not
126 /// used by the terminators, remove those implicit_def's. e.g.
128 /// r0 = implicit_def
129 /// r1 = implicit_def
131 /// This block can be optimized away later if the implicit instructions are
133 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
134 SmallSet<unsigned, 4> ImpDefRegs;
135 MachineBasicBlock::iterator I = MBB->begin();
136 while (I != MBB->end()) {
137 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF)
139 unsigned Reg = I->getOperand(0).getReg();
140 ImpDefRegs.insert(Reg);
141 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
142 unsigned SubReg = *SubRegs; ++SubRegs)
143 ImpDefRegs.insert(SubReg);
146 if (ImpDefRegs.empty())
149 MachineBasicBlock::iterator FirstTerm = I;
150 while (I != MBB->end()) {
151 if (!TII->isUnpredicatedTerminator(I))
153 // See if it uses any of the implicitly defined registers.
154 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
155 MachineOperand &MO = I->getOperand(i);
156 if (!MO.isReg() || !MO.isUse())
158 unsigned Reg = MO.getReg();
159 if (ImpDefRegs.count(Reg))
166 while (I != FirstTerm) {
167 MachineInstr *ImpDefMI = &*I;
169 MBB->erase(ImpDefMI);
175 /// OptimizeFunction - Perhaps branch folding, tail merging and other
176 /// CFG optimizations on the given function.
177 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
178 const TargetInstrInfo *tii,
179 const TargetRegisterInfo *tri,
180 MachineModuleInfo *mmi) {
181 if (!tii) return false;
187 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
189 // Fix CFG. The later algorithms expect it to be right.
190 bool MadeChange = false;
191 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
192 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
193 SmallVector<MachineOperand, 4> Cond;
194 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
195 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
196 MadeChange |= OptimizeImpDefsBlock(MBB);
200 bool MadeChangeThisIteration = true;
201 while (MadeChangeThisIteration) {
202 MadeChangeThisIteration = false;
203 MadeChangeThisIteration |= TailMergeBlocks(MF);
204 MadeChangeThisIteration |= OptimizeBranches(MF);
205 MadeChange |= MadeChangeThisIteration;
208 // Do tail duplication once after tail merging is done. Otherwise it is
209 // tough to avoid situations where tail duplication and tail merging undo
210 // each other's transformations ad infinitum.
211 MadeChange |= TailDuplicateBlocks(MF);
213 // See if any jump tables have become mergable or dead as the code generator
215 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
216 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
218 // Figure out how these jump tables should be merged.
219 std::vector<unsigned> JTMapping;
220 JTMapping.reserve(JTs.size());
222 // We always keep the 0th jump table.
223 JTMapping.push_back(0);
225 // Scan the jump tables, seeing if there are any duplicates. Note that this
226 // is N^2, which should be fixed someday.
227 for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
228 if (JTs[i].MBBs.empty())
229 JTMapping.push_back(i);
231 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
234 // If a jump table was merge with another one, walk the function rewriting
235 // references to jump tables to reference the new JT ID's. Keep track of
236 // whether we see a jump table idx, if not, we can delete the JT.
237 BitVector JTIsLive(JTs.size());
238 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
240 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
242 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
243 MachineOperand &Op = I->getOperand(op);
244 if (!Op.isJTI()) continue;
245 unsigned NewIdx = JTMapping[Op.getIndex()];
248 // Remember that this JT is live.
249 JTIsLive.set(NewIdx);
253 // Finally, remove dead jump tables. This happens either because the
254 // indirect jump was unreachable (and thus deleted) or because the jump
255 // table was merged with some other one.
256 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
257 if (!JTIsLive.test(i)) {
258 JTI->RemoveJumpTable(i);
267 //===----------------------------------------------------------------------===//
268 // Tail Merging of Blocks
269 //===----------------------------------------------------------------------===//
271 /// HashMachineInstr - Compute a hash value for MI and its operands.
272 static unsigned HashMachineInstr(const MachineInstr *MI) {
273 unsigned Hash = MI->getOpcode();
274 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
275 const MachineOperand &Op = MI->getOperand(i);
277 // Merge in bits from the operand if easy.
278 unsigned OperandHash = 0;
279 switch (Op.getType()) {
280 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
281 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
282 case MachineOperand::MO_MachineBasicBlock:
283 OperandHash = Op.getMBB()->getNumber();
285 case MachineOperand::MO_FrameIndex:
286 case MachineOperand::MO_ConstantPoolIndex:
287 case MachineOperand::MO_JumpTableIndex:
288 OperandHash = Op.getIndex();
290 case MachineOperand::MO_GlobalAddress:
291 case MachineOperand::MO_ExternalSymbol:
292 // Global address / external symbol are too hard, don't bother, but do
293 // pull in the offset.
294 OperandHash = Op.getOffset();
299 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
304 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
305 /// with no successors, we hash two instructions, because cross-jumping
306 /// only saves code when at least two instructions are removed (since a
307 /// branch must be inserted). For blocks with a successor, one of the
308 /// two blocks to be tail-merged will end with a branch already, so
309 /// it gains to cross-jump even for one instruction.
310 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
311 unsigned minCommonTailLength) {
312 MachineBasicBlock::const_iterator I = MBB->end();
313 if (I == MBB->begin())
314 return 0; // Empty MBB.
317 unsigned Hash = HashMachineInstr(I);
319 if (I == MBB->begin() || minCommonTailLength == 1)
320 return Hash; // Single instr MBB.
323 // Hash in the second-to-last instruction.
324 Hash ^= HashMachineInstr(I) << 2;
328 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
329 /// of instructions they actually have in common together at their end. Return
330 /// iterators for the first shared instruction in each block.
331 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
332 MachineBasicBlock *MBB2,
333 MachineBasicBlock::iterator &I1,
334 MachineBasicBlock::iterator &I2) {
338 unsigned TailLen = 0;
339 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
341 if (!I1->isIdenticalTo(I2) ||
342 // FIXME: This check is dubious. It's used to get around a problem where
343 // people incorrectly expect inline asm directives to remain in the same
344 // relative order. This is untenable because normal compiler
345 // optimizations (like this one) may reorder and/or merge these
347 I1->getOpcode() == TargetInstrInfo::INLINEASM) {
356 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
357 /// after it, replacing it with an unconditional branch to NewDest. This
358 /// returns true if OldInst's block is modified, false if NewDest is modified.
359 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
360 MachineBasicBlock *NewDest) {
361 MachineBasicBlock *OldBB = OldInst->getParent();
363 // Remove all the old successors of OldBB from the CFG.
364 while (!OldBB->succ_empty())
365 OldBB->removeSuccessor(OldBB->succ_begin());
367 // Remove all the dead instructions from the end of OldBB.
368 OldBB->erase(OldInst, OldBB->end());
370 // If OldBB isn't immediately before OldBB, insert a branch to it.
371 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
372 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
373 OldBB->addSuccessor(NewDest);
377 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
378 /// MBB so that the part before the iterator falls into the part starting at the
379 /// iterator. This returns the new MBB.
380 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
381 MachineBasicBlock::iterator BBI1) {
382 MachineFunction &MF = *CurMBB.getParent();
384 // Create the fall-through block.
385 MachineFunction::iterator MBBI = &CurMBB;
386 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
387 CurMBB.getParent()->insert(++MBBI, NewMBB);
389 // Move all the successors of this block to the specified block.
390 NewMBB->transferSuccessors(&CurMBB);
392 // Add an edge from CurMBB to NewMBB for the fall-through.
393 CurMBB.addSuccessor(NewMBB);
395 // Splice the code over.
396 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
398 // For targets that use the register scavenger, we must maintain LiveIns.
400 RS->enterBasicBlock(&CurMBB);
402 RS->forward(prior(CurMBB.end()));
403 BitVector RegsLiveAtExit(TRI->getNumRegs());
404 RS->getRegsUsed(RegsLiveAtExit, false);
405 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
406 if (RegsLiveAtExit[i])
407 NewMBB->addLiveIn(i);
413 /// EstimateRuntime - Make a rough estimate for how long it will take to run
414 /// the specified code.
415 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
416 MachineBasicBlock::iterator E) {
418 for (; I != E; ++I) {
419 const TargetInstrDesc &TID = I->getDesc();
422 else if (TID.mayLoad() || TID.mayStore())
430 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
431 // branches temporarily for tail merging). In the case where CurMBB ends
432 // with a conditional branch to the next block, optimize by reversing the
433 // test and conditionally branching to SuccMBB instead.
434 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
435 const TargetInstrInfo *TII) {
436 MachineFunction *MF = CurMBB->getParent();
437 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
438 MachineBasicBlock *TBB = 0, *FBB = 0;
439 SmallVector<MachineOperand, 4> Cond;
440 if (I != MF->end() &&
441 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
442 MachineBasicBlock *NextBB = I;
443 if (TBB == NextBB && !Cond.empty() && !FBB) {
444 if (!TII->ReverseBranchCondition(Cond)) {
445 TII->RemoveBranch(*CurMBB);
446 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
451 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
455 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
456 if (getHash() < o.getHash())
458 else if (getHash() > o.getHash())
460 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
462 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
465 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
466 // an object with itself.
467 #ifndef _GLIBCXX_DEBUG
468 llvm_unreachable("Predecessor appears twice");
474 /// CountTerminators - Count the number of terminators in the given
475 /// block and set I to the position of the first non-terminator, if there
476 /// is one, or MBB->end() otherwise.
477 static unsigned CountTerminators(MachineBasicBlock *MBB,
478 MachineBasicBlock::iterator &I) {
480 unsigned NumTerms = 0;
482 if (I == MBB->begin()) {
487 if (!I->getDesc().isTerminator()) break;
493 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
494 /// and decide if it would be profitable to merge those tails. Return the
495 /// length of the common tail and iterators to the first common instruction
497 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
498 MachineBasicBlock *MBB2,
499 unsigned minCommonTailLength,
500 unsigned &CommonTailLen,
501 MachineBasicBlock::iterator &I1,
502 MachineBasicBlock::iterator &I2,
503 MachineBasicBlock *SuccBB,
504 MachineBasicBlock *PredBB) {
505 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
506 MachineFunction *MF = MBB1->getParent();
508 if (CommonTailLen == 0)
511 // It's almost always profitable to merge any number of non-terminator
512 // instructions with the block that falls through into the common successor.
513 if (MBB1 == PredBB || MBB2 == PredBB) {
514 MachineBasicBlock::iterator I;
515 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
516 if (CommonTailLen > NumTerms)
520 // If one of the blocks can be completely merged and happens to be in
521 // a position where the other could fall through into it, merge any number
522 // of instructions, because it can be done without a branch.
523 // TODO: If the blocks are not adjacent, move one of them so that they are?
524 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
526 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
529 // If both blocks have an unconditional branch temporarily stripped out,
530 // count that as an additional common instruction for the following
532 unsigned EffectiveTailLen = CommonTailLen;
533 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
534 !MBB1->back().getDesc().isBarrier() &&
535 !MBB2->back().getDesc().isBarrier())
538 // Check if the common tail is long enough to be worthwhile.
539 if (EffectiveTailLen >= minCommonTailLength)
542 // If we are optimizing for code size, 2 instructions in common is enough if
543 // we don't have to split a block. At worst we will be introducing 1 new
544 // branch instruction, which is likely to be smaller than the 2
545 // instructions that would be deleted in the merge.
546 if (EffectiveTailLen >= 2 &&
547 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
548 (I1 == MBB1->begin() || I2 == MBB2->begin()))
554 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
555 /// hash CurHash (guaranteed to match the last element). Build the vector
556 /// SameTails of all those that have the (same) largest number of instructions
557 /// in common of any pair of these blocks. SameTails entries contain an
558 /// iterator into MergePotentials (from which the MachineBasicBlock can be
559 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
560 /// instruction where the matching code sequence begins.
561 /// Order of elements in SameTails is the reverse of the order in which
562 /// those blocks appear in MergePotentials (where they are not necessarily
564 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
565 unsigned minCommonTailLength,
566 MachineBasicBlock *SuccBB,
567 MachineBasicBlock *PredBB) {
568 unsigned maxCommonTailLength = 0U;
570 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
571 MPIterator HighestMPIter = prior(MergePotentials.end());
572 for (MPIterator CurMPIter = prior(MergePotentials.end()),
573 B = MergePotentials.begin();
574 CurMPIter != B && CurMPIter->getHash() == CurHash;
576 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
577 unsigned CommonTailLen;
578 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
580 CommonTailLen, TrialBBI1, TrialBBI2,
582 if (CommonTailLen > maxCommonTailLength) {
584 maxCommonTailLength = CommonTailLen;
585 HighestMPIter = CurMPIter;
586 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
588 if (HighestMPIter == CurMPIter &&
589 CommonTailLen == maxCommonTailLength)
590 SameTails.push_back(SameTailElt(I, TrialBBI2));
596 return maxCommonTailLength;
599 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
600 /// MergePotentials, restoring branches at ends of blocks as appropriate.
601 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
602 MachineBasicBlock *SuccBB,
603 MachineBasicBlock *PredBB) {
604 MPIterator CurMPIter, B;
605 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
606 CurMPIter->getHash() == CurHash;
608 // Put the unconditional branch back, if we need one.
609 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
610 if (SuccBB && CurMBB != PredBB)
611 FixTail(CurMBB, SuccBB, TII);
615 if (CurMPIter->getHash() != CurHash)
617 MergePotentials.erase(CurMPIter, MergePotentials.end());
620 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
621 /// only of the common tail. Create a block that does by splitting one.
622 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
623 unsigned maxCommonTailLength) {
624 unsigned commonTailIndex = 0;
625 unsigned TimeEstimate = ~0U;
626 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
627 // Use PredBB if possible; that doesn't require a new branch.
628 if (SameTails[i].getBlock() == PredBB) {
632 // Otherwise, make a (fairly bogus) choice based on estimate of
633 // how long it will take the various blocks to execute.
634 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
635 SameTails[i].getTailStartPos());
636 if (t <= TimeEstimate) {
642 MachineBasicBlock::iterator BBI =
643 SameTails[commonTailIndex].getTailStartPos();
644 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
646 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
647 << maxCommonTailLength);
649 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
650 SameTails[commonTailIndex].setBlock(newMBB);
651 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
653 // If we split PredBB, newMBB is the new predecessor.
657 return commonTailIndex;
660 // See if any of the blocks in MergePotentials (which all have a common single
661 // successor, or all have no successor) can be tail-merged. If there is a
662 // successor, any blocks in MergePotentials that are not tail-merged and
663 // are not immediately before Succ must have an unconditional branch to
664 // Succ added (but the predecessor/successor lists need no adjustment).
665 // The lone predecessor of Succ that falls through into Succ,
666 // if any, is given in PredBB.
668 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
669 MachineBasicBlock *PredBB) {
670 bool MadeChange = false;
672 // Except for the special cases below, tail-merge if there are at least
673 // this many instructions in common.
674 unsigned minCommonTailLength = TailMergeSize;
676 DEBUG(errs() << "\nTryTailMergeBlocks: ";
677 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
678 errs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
679 << (i == e-1 ? "" : ", ");
682 errs() << " with successor BB#" << SuccBB->getNumber() << '\n';
684 errs() << " which has fall-through from BB#"
685 << PredBB->getNumber() << "\n";
687 errs() << "Looking for common tails of at least "
688 << minCommonTailLength << " instruction"
689 << (minCommonTailLength == 1 ? "" : "s") << '\n';
692 // Sort by hash value so that blocks with identical end sequences sort
694 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
696 // Walk through equivalence sets looking for actual exact matches.
697 while (MergePotentials.size() > 1) {
698 unsigned CurHash = MergePotentials.back().getHash();
700 // Build SameTails, identifying the set of blocks with this hash code
701 // and with the maximum number of instructions in common.
702 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
706 // If we didn't find any pair that has at least minCommonTailLength
707 // instructions in common, remove all blocks with this hash code and retry.
708 if (SameTails.empty()) {
709 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
713 // If one of the blocks is the entire common tail (and not the entry
714 // block, which we can't jump to), we can treat all blocks with this same
715 // tail at once. Use PredBB if that is one of the possibilities, as that
716 // will not introduce any extra branches.
717 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
718 getParent()->begin();
719 unsigned commonTailIndex = SameTails.size();
720 // If there are two blocks, check to see if one can be made to fall through
722 if (SameTails.size() == 2 &&
723 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
724 SameTails[1].tailIsWholeBlock())
726 else if (SameTails.size() == 2 &&
727 SameTails[1].getBlock()->isLayoutSuccessor(
728 SameTails[0].getBlock()) &&
729 SameTails[0].tailIsWholeBlock())
732 // Otherwise just pick one, favoring the fall-through predecessor if
734 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
735 MachineBasicBlock *MBB = SameTails[i].getBlock();
736 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
742 if (SameTails[i].tailIsWholeBlock())
747 if (commonTailIndex == SameTails.size() ||
748 (SameTails[commonTailIndex].getBlock() == PredBB &&
749 !SameTails[commonTailIndex].tailIsWholeBlock())) {
750 // None of the blocks consist entirely of the common tail.
751 // Split a block so that one does.
752 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
755 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
756 // MBB is common tail. Adjust all other BB's to jump to this one.
757 // Traversal must be forwards so erases work.
758 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber()
760 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
761 if (commonTailIndex == i)
763 DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber()
764 << (i == e-1 ? "" : ", "));
765 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
766 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
767 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
768 MergePotentials.erase(SameTails[i].getMPIter());
770 DEBUG(errs() << "\n");
771 // We leave commonTailIndex in the worklist in case there are other blocks
772 // that match it with a smaller number of instructions.
778 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
780 if (!EnableTailMerge) return false;
782 bool MadeChange = false;
784 // First find blocks with no successors.
785 MergePotentials.clear();
786 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
788 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
791 // See if we can do any tail merging on those.
792 if (MergePotentials.size() < TailMergeThreshold &&
793 MergePotentials.size() >= 2)
794 MadeChange |= TryTailMergeBlocks(NULL, NULL);
796 // Look at blocks (IBB) with multiple predecessors (PBB).
797 // We change each predecessor to a canonical form, by
798 // (1) temporarily removing any unconditional branch from the predecessor
800 // (2) alter conditional branches so they branch to the other block
801 // not IBB; this may require adding back an unconditional branch to IBB
802 // later, where there wasn't one coming in. E.g.
804 // fallthrough to QBB
807 // with a conceptual B to IBB after that, which never actually exists.
808 // With those changes, we see whether the predecessors' tails match,
809 // and merge them if so. We change things out of canonical form and
810 // back to the way they were later in the process. (OptimizeBranches
811 // would undo some of this, but we can't use it, because we'd get into
812 // a compile-time infinite loop repeatedly doing and undoing the same
815 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end();
817 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
818 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
819 MachineBasicBlock *IBB = I;
820 MachineBasicBlock *PredBB = prior(I);
821 MergePotentials.clear();
822 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
825 MachineBasicBlock *PBB = *P;
826 // Skip blocks that loop to themselves, can't tail merge these.
829 // Visit each predecessor only once.
830 if (!UniquePreds.insert(PBB))
832 MachineBasicBlock *TBB = 0, *FBB = 0;
833 SmallVector<MachineOperand, 4> Cond;
834 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
835 // Failing case: IBB is the target of a cbr, and
836 // we cannot reverse the branch.
837 SmallVector<MachineOperand, 4> NewCond(Cond);
838 if (!Cond.empty() && TBB == IBB) {
839 if (TII->ReverseBranchCondition(NewCond))
841 // This is the QBB case described above
843 FBB = next(MachineFunction::iterator(PBB));
845 // Failing case: the only way IBB can be reached from PBB is via
846 // exception handling. Happens for landing pads. Would be nice
847 // to have a bit in the edge so we didn't have to do all this.
848 if (IBB->isLandingPad()) {
849 MachineFunction::iterator IP = PBB; IP++;
850 MachineBasicBlock *PredNextBB = NULL;
854 if (IBB != PredNextBB) // fallthrough
857 if (TBB != IBB && FBB != IBB) // cbr then ubr
859 } else if (Cond.empty()) {
860 if (TBB != IBB) // ubr
863 if (TBB != IBB && IBB != PredNextBB) // cbr
867 // Remove the unconditional branch at the end, if any.
868 if (TBB && (Cond.empty() || FBB)) {
869 TII->RemoveBranch(*PBB);
871 // reinsert conditional branch only, for now
872 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
874 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
878 if (MergePotentials.size() >= 2)
879 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
880 // Reinsert an unconditional branch if needed.
881 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
882 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
883 if (MergePotentials.size() == 1 &&
884 MergePotentials.begin()->getBlock() != PredBB)
885 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
891 //===----------------------------------------------------------------------===//
892 // Branch Optimization
893 //===----------------------------------------------------------------------===//
895 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
896 bool MadeChange = false;
898 // Make sure blocks are numbered in order
901 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
902 MachineBasicBlock *MBB = I++;
903 MadeChange |= OptimizeBlock(MBB);
905 // If it is dead, remove it.
906 if (MBB->pred_empty()) {
907 RemoveDeadBlock(MBB);
916 /// CanFallThrough - Return true if the specified block (with the specified
917 /// branch condition) can implicitly transfer control to the block after it by
918 /// falling off the end of it. This should return false if it can reach the
919 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
921 /// True is a conservative answer.
923 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
924 bool BranchUnAnalyzable,
925 MachineBasicBlock *TBB,
926 MachineBasicBlock *FBB,
927 const SmallVectorImpl<MachineOperand> &Cond) {
928 MachineFunction::iterator Fallthrough = CurBB;
930 // If FallthroughBlock is off the end of the function, it can't fall through.
931 if (Fallthrough == CurBB->getParent()->end())
934 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
935 if (!CurBB->isSuccessor(Fallthrough))
938 // If we couldn't analyze the branch, examine the last instruction.
939 // If the block doesn't end in a known control barrier, assume fallthrough
940 // is possible. The isPredicable check is needed because this code can be
941 // called during IfConversion, where an instruction which is normally a
942 // Barrier is predicated and thus no longer an actual control barrier. This
943 // is over-conservative though, because if an instruction isn't actually
944 // predicated we could still treat it like a barrier.
945 if (BranchUnAnalyzable)
946 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() ||
947 CurBB->back().getDesc().isPredicable();
949 // If there is no branch, control always falls through.
950 if (TBB == 0) return true;
952 // If there is some explicit branch to the fallthrough block, it can obviously
953 // reach, even though the branch should get folded to fall through implicitly.
954 if (MachineFunction::iterator(TBB) == Fallthrough ||
955 MachineFunction::iterator(FBB) == Fallthrough)
958 // If it's an unconditional branch to some block not the fall through, it
959 // doesn't fall through.
960 if (Cond.empty()) return false;
962 // Otherwise, if it is conditional and has no explicit false block, it falls
967 /// CanFallThrough - Return true if the specified can implicitly transfer
968 /// control to the block after it by falling off the end of it. This should
969 /// return false if it can reach the block after it, but it uses an explicit
970 /// branch to do so (e.g. a table jump).
972 /// True is a conservative answer.
974 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
975 MachineBasicBlock *TBB = 0, *FBB = 0;
976 SmallVector<MachineOperand, 4> Cond;
977 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true);
978 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
981 /// IsBetterFallthrough - Return true if it would be clearly better to
982 /// fall-through to MBB1 than to fall through into MBB2. This has to return
983 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
984 /// result in infinite loops.
985 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
986 MachineBasicBlock *MBB2) {
987 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
988 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
989 // optimize branches that branch to either a return block or an assert block
990 // into a fallthrough to the return.
991 if (MBB1->empty() || MBB2->empty()) return false;
993 // If there is a clear successor ordering we make sure that one block
994 // will fall through to the next
995 if (MBB1->isSuccessor(MBB2)) return true;
996 if (MBB2->isSuccessor(MBB1)) return false;
998 MachineInstr *MBB1I = --MBB1->end();
999 MachineInstr *MBB2I = --MBB2->end();
1000 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
1003 /// TailDuplicateBlocks - Look for small blocks that are unconditionally
1004 /// branched to and do not fall through. Tail-duplicate their instructions
1005 /// into their predecessors to eliminate (dynamic) branches.
1006 bool BranchFolder::TailDuplicateBlocks(MachineFunction &MF) {
1007 bool MadeChange = false;
1009 // Make sure blocks are numbered in order
1010 MF.RenumberBlocks();
1012 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
1013 MachineBasicBlock *MBB = I++;
1015 // Only duplicate blocks that end with unconditional branches.
1016 if (CanFallThrough(MBB))
1019 MadeChange |= TailDuplicate(MBB, MF);
1021 // If it is dead, remove it.
1022 if (MBB->pred_empty()) {
1023 RemoveDeadBlock(MBB);
1031 /// TailDuplicate - If it is profitable, duplicate TailBB's contents in each
1032 /// of its predecessors.
1033 bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB,
1034 MachineFunction &MF) {
1035 // Don't try to tail-duplicate single-block loops.
1036 if (TailBB->isSuccessor(TailBB))
1039 // Set the limit on the number of instructions to duplicate, with a default
1040 // of one less than the tail-merge threshold. When optimizing for size,
1041 // duplicate only one, because one branch instruction can be eliminated to
1042 // compensate for the duplication.
1043 unsigned MaxDuplicateCount =
1044 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ?
1045 1 : TII->TailDuplicationLimit(*TailBB, TailMergeSize - 1);
1047 // Check the instructions in the block to determine whether tail-duplication
1048 // is invalid or unlikely to be profitable.
1050 bool HasCall = false;
1051 for (MachineBasicBlock::iterator I = TailBB->begin();
1052 I != TailBB->end(); ++I, ++i) {
1053 // Non-duplicable things shouldn't be tail-duplicated.
1054 if (I->getDesc().isNotDuplicable()) return false;
1055 // Don't duplicate more than the threshold.
1056 if (i == MaxDuplicateCount) return false;
1057 // Remember if we saw a call.
1058 if (I->getDesc().isCall()) HasCall = true;
1060 // Heuristically, don't tail-duplicate calls if it would expand code size,
1061 // as it's less likely to be worth the extra cost.
1062 if (i > 1 && HasCall)
1065 // Iterate through all the unique predecessors and tail-duplicate this
1066 // block into them, if possible. Copying the list ahead of time also
1067 // avoids trouble with the predecessor list reallocating.
1068 bool Changed = false;
1069 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
1070 TailBB->pred_end());
1071 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
1072 PE = Preds.end(); PI != PE; ++PI) {
1073 MachineBasicBlock *PredBB = *PI;
1075 assert(TailBB != PredBB &&
1076 "Single-block loop should have been rejected earlier!");
1077 if (PredBB->succ_size() > 1) continue;
1079 MachineBasicBlock *PredTBB, *PredFBB;
1080 SmallVector<MachineOperand, 4> PredCond;
1081 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
1083 if (!PredCond.empty())
1085 // EH edges are ignored by AnalyzeBranch.
1086 if (PredBB->succ_size() != 1)
1088 // Don't duplicate into a fall-through predecessor (at least for now).
1089 if (PredBB->isLayoutSuccessor(TailBB) && CanFallThrough(PredBB))
1092 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB
1093 << "From Succ: " << *TailBB);
1095 // Remove PredBB's unconditional branch.
1096 TII->RemoveBranch(*PredBB);
1097 // Clone the contents of TailBB into PredBB.
1098 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end();
1100 MachineInstr *NewMI = MF.CloneMachineInstr(I);
1101 PredBB->insert(PredBB->end(), NewMI);
1105 PredBB->removeSuccessor(PredBB->succ_begin());
1106 assert(PredBB->succ_empty() &&
1107 "TailDuplicate called on block with multiple successors!");
1108 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
1109 E = TailBB->succ_end(); I != E; ++I)
1110 PredBB->addSuccessor(*I);
1116 // If TailBB was duplicated into all its predecessors except for the prior
1117 // block, which falls through unconditionally, move the contents of this
1118 // block into the prior block.
1119 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(TailBB));
1120 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1121 SmallVector<MachineOperand, 4> PriorCond;
1122 bool PriorUnAnalyzable =
1123 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1124 // This has to check PrevBB->succ_size() because EH edges are ignored by
1126 if (!PriorUnAnalyzable && PriorCond.empty() && !PriorTBB &&
1127 TailBB->pred_size() == 1 && PrevBB.succ_size() == 1 &&
1128 !TailBB->hasAddressTaken()) {
1129 DEBUG(errs() << "\nMerging into block: " << PrevBB
1130 << "From MBB: " << *TailBB);
1131 PrevBB.splice(PrevBB.end(), TailBB, TailBB->begin(), TailBB->end());
1132 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1133 assert(PrevBB.succ_empty());
1134 PrevBB.transferSuccessors(TailBB);
1141 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1142 /// block. This is never called on the entry block.
1143 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1144 bool MadeChange = false;
1145 MachineFunction &MF = *MBB->getParent();
1148 MachineFunction::iterator FallThrough = MBB;
1151 // If this block is empty, make everyone use its fall-through, not the block
1152 // explicitly. Landing pads should not do this since the landing-pad table
1153 // points to this block. Blocks with their addresses taken shouldn't be
1155 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1156 // Dead block? Leave for cleanup later.
1157 if (MBB->pred_empty()) return MadeChange;
1159 if (FallThrough == MF.end()) {
1160 // TODO: Simplify preds to not branch here if possible!
1162 // Rewrite all predecessors of the old block to go to the fallthrough
1164 while (!MBB->pred_empty()) {
1165 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1166 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1168 // If MBB was the target of a jump table, update jump tables to go to the
1169 // fallthrough instead.
1170 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough);
1176 // Check to see if we can simplify the terminator of the block before this
1178 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1180 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1181 SmallVector<MachineOperand, 4> PriorCond;
1182 bool PriorUnAnalyzable =
1183 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1184 if (!PriorUnAnalyzable) {
1185 // If the CFG for the prior block has extra edges, remove them.
1186 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1187 !PriorCond.empty());
1189 // If the previous branch is conditional and both conditions go to the same
1190 // destination, remove the branch, replacing it with an unconditional one or
1192 if (PriorTBB && PriorTBB == PriorFBB) {
1193 TII->RemoveBranch(PrevBB);
1195 if (PriorTBB != MBB)
1196 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1199 goto ReoptimizeBlock;
1202 // If the previous block unconditionally falls through to this block and
1203 // this block has no other predecessors, move the contents of this block
1204 // into the prior block. This doesn't usually happen when SimplifyCFG
1205 // has been used, but it can happen if tail merging splits a fall-through
1206 // predecessor of a block.
1207 // This has to check PrevBB->succ_size() because EH edges are ignored by
1209 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1210 PrevBB.succ_size() == 1 &&
1211 !MBB->hasAddressTaken()) {
1212 DEBUG(errs() << "\nMerging into block: " << PrevBB
1213 << "From MBB: " << *MBB);
1214 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1215 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1216 assert(PrevBB.succ_empty());
1217 PrevBB.transferSuccessors(MBB);
1222 // If the previous branch *only* branches to *this* block (conditional or
1223 // not) remove the branch.
1224 if (PriorTBB == MBB && PriorFBB == 0) {
1225 TII->RemoveBranch(PrevBB);
1228 goto ReoptimizeBlock;
1231 // If the prior block branches somewhere else on the condition and here if
1232 // the condition is false, remove the uncond second branch.
1233 if (PriorFBB == MBB) {
1234 TII->RemoveBranch(PrevBB);
1235 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1238 goto ReoptimizeBlock;
1241 // If the prior block branches here on true and somewhere else on false, and
1242 // if the branch condition is reversible, reverse the branch to create a
1244 if (PriorTBB == MBB) {
1245 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1246 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1247 TII->RemoveBranch(PrevBB);
1248 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
1251 goto ReoptimizeBlock;
1255 // If this block has no successors (e.g. it is a return block or ends with
1256 // a call to a no-return function like abort or __cxa_throw) and if the pred
1257 // falls through into this block, and if it would otherwise fall through
1258 // into the block after this, move this block to the end of the function.
1260 // We consider it more likely that execution will stay in the function (e.g.
1261 // due to loops) than it is to exit it. This asserts in loops etc, moving
1262 // the assert condition out of the loop body.
1263 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1264 MachineFunction::iterator(PriorTBB) == FallThrough &&
1265 !CanFallThrough(MBB)) {
1266 bool DoTransform = true;
1268 // We have to be careful that the succs of PredBB aren't both no-successor
1269 // blocks. If neither have successors and if PredBB is the second from
1270 // last block in the function, we'd just keep swapping the two blocks for
1271 // last. Only do the swap if one is clearly better to fall through than
1273 if (FallThrough == --MF.end() &&
1274 !IsBetterFallthrough(PriorTBB, MBB))
1275 DoTransform = false;
1277 // We don't want to do this transformation if we have control flow like:
1286 // In this case, we could actually be moving the return block *into* a
1288 if (DoTransform && !MBB->succ_empty() &&
1289 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
1290 DoTransform = false;
1294 // Reverse the branch so we will fall through on the previous true cond.
1295 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1296 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1297 DEBUG(errs() << "\nMoving MBB: " << *MBB
1298 << "To make fallthrough to: " << *PriorTBB << "\n");
1300 TII->RemoveBranch(PrevBB);
1301 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1303 // Move this block to the end of the function.
1304 MBB->moveAfter(--MF.end());
1313 // Analyze the branch in the current block.
1314 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1315 SmallVector<MachineOperand, 4> CurCond;
1316 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1317 if (!CurUnAnalyzable) {
1318 // If the CFG for the prior block has extra edges, remove them.
1319 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1321 // If this is a two-way branch, and the FBB branches to this block, reverse
1322 // the condition so the single-basic-block loop is faster. Instead of:
1323 // Loop: xxx; jcc Out; jmp Loop
1325 // Loop: xxx; jncc Loop; jmp Out
1326 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1327 SmallVector<MachineOperand, 4> NewCond(CurCond);
1328 if (!TII->ReverseBranchCondition(NewCond)) {
1329 TII->RemoveBranch(*MBB);
1330 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1333 goto ReoptimizeBlock;
1337 // If this branch is the only thing in its block, see if we can forward
1338 // other blocks across it.
1339 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1340 MBB->begin()->getDesc().isBranch() && CurTBB != MBB &&
1341 !MBB->hasAddressTaken()) {
1342 // This block may contain just an unconditional branch. Because there can
1343 // be 'non-branch terminators' in the block, try removing the branch and
1344 // then seeing if the block is empty.
1345 TII->RemoveBranch(*MBB);
1347 // If this block is just an unconditional branch to CurTBB, we can
1348 // usually completely eliminate the block. The only case we cannot
1349 // completely eliminate the block is when the block before this one
1350 // falls through into MBB and we can't understand the prior block's branch
1353 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
1354 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1355 !PrevBB.isSuccessor(MBB)) {
1356 // If the prior block falls through into us, turn it into an
1357 // explicit branch to us to make updates simpler.
1358 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1359 PriorTBB != MBB && PriorFBB != MBB) {
1360 if (PriorTBB == 0) {
1361 assert(PriorCond.empty() && PriorFBB == 0 &&
1362 "Bad branch analysis");
1365 assert(PriorFBB == 0 && "Machine CFG out of date!");
1368 TII->RemoveBranch(PrevBB);
1369 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1372 // Iterate through all the predecessors, revectoring each in-turn.
1374 bool DidChange = false;
1375 bool HasBranchToSelf = false;
1376 while(PI != MBB->pred_size()) {
1377 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1379 // If this block has an uncond branch to itself, leave it.
1381 HasBranchToSelf = true;
1384 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1385 // If this change resulted in PMBB ending in a conditional
1386 // branch where both conditions go to the same destination,
1387 // change this to an unconditional branch (and fix the CFG).
1388 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1389 SmallVector<MachineOperand, 4> NewCurCond;
1390 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1391 NewCurFBB, NewCurCond, true);
1392 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1393 TII->RemoveBranch(*PMBB);
1395 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1398 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1403 // Change any jumptables to go to the new MBB.
1404 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB);
1408 if (!HasBranchToSelf) return MadeChange;
1413 // Add the branch back if the block is more than just an uncond branch.
1414 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1418 // If the prior block doesn't fall through into this block, and if this
1419 // block doesn't fall through into some other block, see if we can find a
1420 // place to move this block where a fall-through will happen.
1421 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
1422 PriorTBB, PriorFBB, PriorCond)) {
1424 // Now we know that there was no fall-through into this block, check to
1425 // see if it has a fall-through into its successor.
1426 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1429 if (!MBB->isLandingPad()) {
1430 // Check all the predecessors of this block. If one of them has no fall
1431 // throughs, move this block right after it.
1432 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1433 E = MBB->pred_end(); PI != E; ++PI) {
1434 // Analyze the branch at the end of the pred.
1435 MachineBasicBlock *PredBB = *PI;
1436 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1437 MachineBasicBlock *PredTBB, *PredFBB;
1438 SmallVector<MachineOperand, 4> PredCond;
1439 if (PredBB != MBB && !CanFallThrough(PredBB) &&
1440 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1441 && (!CurFallsThru || !CurTBB || !CurFBB)
1442 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1443 // If the current block doesn't fall through, just move it.
1444 // If the current block can fall through and does not end with a
1445 // conditional branch, we need to append an unconditional jump to
1446 // the (current) next block. To avoid a possible compile-time
1447 // infinite loop, move blocks only backward in this case.
1448 // Also, if there are already 2 branches here, we cannot add a third;
1449 // this means we have the case
1454 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1456 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1458 MBB->moveAfter(PredBB);
1460 goto ReoptimizeBlock;
1465 if (!CurFallsThru) {
1466 // Check all successors to see if we can move this block before it.
1467 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1468 E = MBB->succ_end(); SI != E; ++SI) {
1469 // Analyze the branch at the end of the block before the succ.
1470 MachineBasicBlock *SuccBB = *SI;
1471 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1473 // If this block doesn't already fall-through to that successor, and if
1474 // the succ doesn't already have a block that can fall through into it,
1475 // and if the successor isn't an EH destination, we can arrange for the
1476 // fallthrough to happen.
1477 if (SuccBB != MBB && &*SuccPrev != MBB &&
1478 !CanFallThrough(SuccPrev) && !CurUnAnalyzable &&
1479 !SuccBB->isLandingPad()) {
1480 MBB->moveBefore(SuccBB);
1482 goto ReoptimizeBlock;
1486 // Okay, there is no really great place to put this block. If, however,
1487 // the block before this one would be a fall-through if this block were
1488 // removed, move this block to the end of the function.
1489 MachineBasicBlock *PrevTBB, *PrevFBB;
1490 SmallVector<MachineOperand, 4> PrevCond;
1491 if (FallThrough != MF.end() &&
1492 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1493 PrevBB.isSuccessor(FallThrough)) {
1494 MBB->moveAfter(--MF.end());