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 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
45 cl::init(cl::BOU_UNSET), cl::Hidden);
46 // Throttle for huge numbers of predecessors (compile speed problems)
47 static cl::opt<unsigned>
48 TailMergeThreshold("tail-merge-threshold",
49 cl::desc("Max number of predecessors to consider tail merging"),
50 cl::init(150), cl::Hidden);
52 // Heuristic for tail merging (and, inversely, tail duplication).
53 // TODO: This should be replaced with a target query.
54 static cl::opt<unsigned>
55 TailMergeSize("tail-merge-size",
56 cl::desc("Min number of instructions to consider tail merging"),
57 cl::init(3), cl::Hidden);
60 /// BranchFolderPass - Wrap branch folder in a machine function pass.
61 class BranchFolderPass : public MachineFunctionPass,
65 explicit BranchFolderPass(bool defaultEnableTailMerge)
66 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {}
68 virtual bool runOnMachineFunction(MachineFunction &MF);
69 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
73 char BranchFolderPass::ID = 0;
75 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
76 return new BranchFolderPass(DefaultEnableTailMerge);
79 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
80 return OptimizeFunction(MF,
81 MF.getTarget().getInstrInfo(),
82 MF.getTarget().getRegisterInfo(),
83 getAnalysisIfAvailable<MachineModuleInfo>());
87 BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
88 switch (FlagEnableTailMerge) {
89 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
90 case cl::BOU_TRUE: EnableTailMerge = true; break;
91 case cl::BOU_FALSE: EnableTailMerge = false; break;
95 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
96 /// function, updating the CFG.
97 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
98 assert(MBB->pred_empty() && "MBB must be dead!");
99 DEBUG(errs() << "\nRemoving MBB: " << *MBB);
101 MachineFunction *MF = MBB->getParent();
102 // drop all successors.
103 while (!MBB->succ_empty())
104 MBB->removeSuccessor(MBB->succ_end()-1);
106 // If there are any labels in the basic block, unregister them from
107 // MachineModuleInfo.
108 if (MMI && !MBB->empty()) {
109 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
112 // The label ID # is always operand #0, an immediate.
113 MMI->InvalidateLabel(I->getOperand(0).getImm());
121 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
122 /// followed by terminators, and if the implicitly defined registers are not
123 /// used by the terminators, remove those implicit_def's. e.g.
125 /// r0 = implicit_def
126 /// r1 = implicit_def
128 /// This block can be optimized away later if the implicit instructions are
130 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
131 SmallSet<unsigned, 4> ImpDefRegs;
132 MachineBasicBlock::iterator I = MBB->begin();
133 while (I != MBB->end()) {
134 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF)
136 unsigned Reg = I->getOperand(0).getReg();
137 ImpDefRegs.insert(Reg);
138 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
139 unsigned SubReg = *SubRegs; ++SubRegs)
140 ImpDefRegs.insert(SubReg);
143 if (ImpDefRegs.empty())
146 MachineBasicBlock::iterator FirstTerm = I;
147 while (I != MBB->end()) {
148 if (!TII->isUnpredicatedTerminator(I))
150 // See if it uses any of the implicitly defined registers.
151 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
152 MachineOperand &MO = I->getOperand(i);
153 if (!MO.isReg() || !MO.isUse())
155 unsigned Reg = MO.getReg();
156 if (ImpDefRegs.count(Reg))
163 while (I != FirstTerm) {
164 MachineInstr *ImpDefMI = &*I;
166 MBB->erase(ImpDefMI);
172 /// OptimizeFunction - Perhaps branch folding, tail merging and other
173 /// CFG optimizations on the given function.
174 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
175 const TargetInstrInfo *tii,
176 const TargetRegisterInfo *tri,
177 MachineModuleInfo *mmi) {
178 if (!tii) return false;
184 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
186 // Fix CFG. The later algorithms expect it to be right.
187 bool MadeChange = false;
188 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
189 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
190 SmallVector<MachineOperand, 4> Cond;
191 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
192 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
193 MadeChange |= OptimizeImpDefsBlock(MBB);
197 bool MadeChangeThisIteration = true;
198 while (MadeChangeThisIteration) {
199 MadeChangeThisIteration = false;
200 MadeChangeThisIteration |= TailMergeBlocks(MF);
201 MadeChangeThisIteration |= OptimizeBranches(MF);
202 MadeChange |= MadeChangeThisIteration;
205 // See if any jump tables have become mergable or dead as the code generator
207 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
208 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
210 // Figure out how these jump tables should be merged.
211 std::vector<unsigned> JTMapping;
212 JTMapping.reserve(JTs.size());
214 // We always keep the 0th jump table.
215 JTMapping.push_back(0);
217 // Scan the jump tables, seeing if there are any duplicates. Note that this
218 // is N^2, which should be fixed someday.
219 for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
220 if (JTs[i].MBBs.empty())
221 JTMapping.push_back(i);
223 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
226 // If a jump table was merge with another one, walk the function rewriting
227 // references to jump tables to reference the new JT ID's. Keep track of
228 // whether we see a jump table idx, if not, we can delete the JT.
229 BitVector JTIsLive(JTs.size());
230 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
232 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
234 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
235 MachineOperand &Op = I->getOperand(op);
236 if (!Op.isJTI()) continue;
237 unsigned NewIdx = JTMapping[Op.getIndex()];
240 // Remember that this JT is live.
241 JTIsLive.set(NewIdx);
245 // Finally, remove dead jump tables. This happens either because the
246 // indirect jump was unreachable (and thus deleted) or because the jump
247 // table was merged with some other one.
248 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
249 if (!JTIsLive.test(i)) {
250 JTI->RemoveJumpTable(i);
259 //===----------------------------------------------------------------------===//
260 // Tail Merging of Blocks
261 //===----------------------------------------------------------------------===//
263 /// HashMachineInstr - Compute a hash value for MI and its operands.
264 static unsigned HashMachineInstr(const MachineInstr *MI) {
265 unsigned Hash = MI->getOpcode();
266 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
267 const MachineOperand &Op = MI->getOperand(i);
269 // Merge in bits from the operand if easy.
270 unsigned OperandHash = 0;
271 switch (Op.getType()) {
272 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
273 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
274 case MachineOperand::MO_MachineBasicBlock:
275 OperandHash = Op.getMBB()->getNumber();
277 case MachineOperand::MO_FrameIndex:
278 case MachineOperand::MO_ConstantPoolIndex:
279 case MachineOperand::MO_JumpTableIndex:
280 OperandHash = Op.getIndex();
282 case MachineOperand::MO_GlobalAddress:
283 case MachineOperand::MO_ExternalSymbol:
284 // Global address / external symbol are too hard, don't bother, but do
285 // pull in the offset.
286 OperandHash = Op.getOffset();
291 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
296 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
297 /// with no successors, we hash two instructions, because cross-jumping
298 /// only saves code when at least two instructions are removed (since a
299 /// branch must be inserted). For blocks with a successor, one of the
300 /// two blocks to be tail-merged will end with a branch already, so
301 /// it gains to cross-jump even for one instruction.
302 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
303 unsigned minCommonTailLength) {
304 MachineBasicBlock::const_iterator I = MBB->end();
305 if (I == MBB->begin())
306 return 0; // Empty MBB.
309 unsigned Hash = HashMachineInstr(I);
311 if (I == MBB->begin() || minCommonTailLength == 1)
312 return Hash; // Single instr MBB.
315 // Hash in the second-to-last instruction.
316 Hash ^= HashMachineInstr(I) << 2;
320 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
321 /// of instructions they actually have in common together at their end. Return
322 /// iterators for the first shared instruction in each block.
323 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
324 MachineBasicBlock *MBB2,
325 MachineBasicBlock::iterator &I1,
326 MachineBasicBlock::iterator &I2) {
330 unsigned TailLen = 0;
331 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
333 if (!I1->isIdenticalTo(I2) ||
334 // FIXME: This check is dubious. It's used to get around a problem where
335 // people incorrectly expect inline asm directives to remain in the same
336 // relative order. This is untenable because normal compiler
337 // optimizations (like this one) may reorder and/or merge these
339 I1->getOpcode() == TargetInstrInfo::INLINEASM) {
348 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
349 /// after it, replacing it with an unconditional branch to NewDest. This
350 /// returns true if OldInst's block is modified, false if NewDest is modified.
351 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
352 MachineBasicBlock *NewDest) {
353 MachineBasicBlock *OldBB = OldInst->getParent();
355 // Remove all the old successors of OldBB from the CFG.
356 while (!OldBB->succ_empty())
357 OldBB->removeSuccessor(OldBB->succ_begin());
359 // Remove all the dead instructions from the end of OldBB.
360 OldBB->erase(OldInst, OldBB->end());
362 // If OldBB isn't immediately before OldBB, insert a branch to it.
363 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
364 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
365 OldBB->addSuccessor(NewDest);
369 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
370 /// MBB so that the part before the iterator falls into the part starting at the
371 /// iterator. This returns the new MBB.
372 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
373 MachineBasicBlock::iterator BBI1) {
374 MachineFunction &MF = *CurMBB.getParent();
376 // Create the fall-through block.
377 MachineFunction::iterator MBBI = &CurMBB;
378 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
379 CurMBB.getParent()->insert(++MBBI, NewMBB);
381 // Move all the successors of this block to the specified block.
382 NewMBB->transferSuccessors(&CurMBB);
384 // Add an edge from CurMBB to NewMBB for the fall-through.
385 CurMBB.addSuccessor(NewMBB);
387 // Splice the code over.
388 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
390 // For targets that use the register scavenger, we must maintain LiveIns.
392 RS->enterBasicBlock(&CurMBB);
394 RS->forward(prior(CurMBB.end()));
395 BitVector RegsLiveAtExit(TRI->getNumRegs());
396 RS->getRegsUsed(RegsLiveAtExit, false);
397 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
398 if (RegsLiveAtExit[i])
399 NewMBB->addLiveIn(i);
405 /// EstimateRuntime - Make a rough estimate for how long it will take to run
406 /// the specified code.
407 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
408 MachineBasicBlock::iterator E) {
410 for (; I != E; ++I) {
411 const TargetInstrDesc &TID = I->getDesc();
414 else if (TID.mayLoad() || TID.mayStore())
422 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
423 // branches temporarily for tail merging). In the case where CurMBB ends
424 // with a conditional branch to the next block, optimize by reversing the
425 // test and conditionally branching to SuccMBB instead.
426 static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB,
427 const TargetInstrInfo *TII) {
428 MachineFunction *MF = CurMBB->getParent();
429 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
430 MachineBasicBlock *TBB = 0, *FBB = 0;
431 SmallVector<MachineOperand, 4> Cond;
432 if (I != MF->end() &&
433 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
434 MachineBasicBlock *NextBB = I;
435 if (TBB == NextBB && !Cond.empty() && !FBB) {
436 if (!TII->ReverseBranchCondition(Cond)) {
437 TII->RemoveBranch(*CurMBB);
438 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
443 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
447 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
448 if (getHash() < o.getHash())
450 else if (getHash() > o.getHash())
452 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
454 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
457 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
458 // an object with itself.
459 #ifndef _GLIBCXX_DEBUG
460 llvm_unreachable("Predecessor appears twice");
466 /// CountTerminators - Count the number of terminators in the given
467 /// block and set I to the position of the first non-terminator, if there
468 /// is one, or MBB->end() otherwise.
469 static unsigned CountTerminators(MachineBasicBlock *MBB,
470 MachineBasicBlock::iterator &I) {
472 unsigned NumTerms = 0;
474 if (I == MBB->begin()) {
479 if (!I->getDesc().isTerminator()) break;
485 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
486 /// and decide if it would be profitable to merge those tails. Return the
487 /// length of the common tail and iterators to the first common instruction
489 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
490 MachineBasicBlock *MBB2,
491 unsigned minCommonTailLength,
492 unsigned &CommonTailLen,
493 MachineBasicBlock::iterator &I1,
494 MachineBasicBlock::iterator &I2,
495 MachineBasicBlock *SuccBB,
496 MachineBasicBlock *PredBB) {
497 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
498 MachineFunction *MF = MBB1->getParent();
500 if (CommonTailLen == 0)
503 // It's almost always profitable to merge any number of non-terminator
504 // instructions with the block that falls through into the common successor.
505 if (MBB1 == PredBB || MBB2 == PredBB) {
506 MachineBasicBlock::iterator I;
507 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
508 if (CommonTailLen > NumTerms)
512 // If one of the blocks can be completely merged and happens to be in
513 // a position where the other could fall through into it, merge any number
514 // of instructions, because it can be done without a branch.
515 // TODO: If the blocks are not adjacent, move one of them so that they are?
516 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
518 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
521 // If both blocks have an unconditional branch temporarily stripped out,
522 // treat that as an additional common instruction.
523 if (MBB1 != PredBB && MBB2 != PredBB &&
524 !MBB1->back().getDesc().isBarrier() &&
525 !MBB2->back().getDesc().isBarrier())
526 --minCommonTailLength;
528 // Check if the common tail is long enough to be worthwhile.
529 if (CommonTailLen >= minCommonTailLength)
532 // If we are optimizing for code size, 1 instruction in common is enough if
533 // we don't have to split a block. At worst we will be replacing a
534 // fallthrough into the common tail with a branch, which at worst breaks
535 // even with falling through into the duplicated common tail.
536 if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
537 (I1 == MBB1->begin() || I2 == MBB2->begin()))
543 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
544 /// hash CurHash (guaranteed to match the last element). Build the vector
545 /// SameTails of all those that have the (same) largest number of instructions
546 /// in common of any pair of these blocks. SameTails entries contain an
547 /// iterator into MergePotentials (from which the MachineBasicBlock can be
548 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
549 /// instruction where the matching code sequence begins.
550 /// Order of elements in SameTails is the reverse of the order in which
551 /// those blocks appear in MergePotentials (where they are not necessarily
553 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
554 unsigned minCommonTailLength,
555 MachineBasicBlock *SuccBB,
556 MachineBasicBlock *PredBB) {
557 unsigned maxCommonTailLength = 0U;
559 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
560 MPIterator HighestMPIter = prior(MergePotentials.end());
561 for (MPIterator CurMPIter = prior(MergePotentials.end()),
562 B = MergePotentials.begin();
563 CurMPIter != B && CurMPIter->getHash() == CurHash;
565 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
566 unsigned CommonTailLen;
567 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
569 CommonTailLen, TrialBBI1, TrialBBI2,
571 if (CommonTailLen > maxCommonTailLength) {
573 maxCommonTailLength = CommonTailLen;
574 HighestMPIter = CurMPIter;
575 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
577 if (HighestMPIter == CurMPIter &&
578 CommonTailLen == maxCommonTailLength)
579 SameTails.push_back(SameTailElt(I, TrialBBI2));
585 return maxCommonTailLength;
588 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
589 /// MergePotentials, restoring branches at ends of blocks as appropriate.
590 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
591 MachineBasicBlock* SuccBB,
592 MachineBasicBlock* PredBB) {
593 MPIterator CurMPIter, B;
594 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
595 CurMPIter->getHash() == CurHash;
597 // Put the unconditional branch back, if we need one.
598 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
599 if (SuccBB && CurMBB != PredBB)
600 FixTail(CurMBB, SuccBB, TII);
604 if (CurMPIter->getHash() != CurHash)
606 MergePotentials.erase(CurMPIter, MergePotentials.end());
609 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
610 /// only of the common tail. Create a block that does by splitting one.
611 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
612 unsigned maxCommonTailLength) {
613 unsigned commonTailIndex = 0;
614 unsigned TimeEstimate = ~0U;
615 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
616 // Use PredBB if possible; that doesn't require a new branch.
617 if (SameTails[i].getBlock() == PredBB) {
621 // Otherwise, make a (fairly bogus) choice based on estimate of
622 // how long it will take the various blocks to execute.
623 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
624 SameTails[i].getTailStartPos());
625 if (t <= TimeEstimate) {
631 MachineBasicBlock::iterator BBI =
632 SameTails[commonTailIndex].getTailStartPos();
633 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
635 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
636 << maxCommonTailLength);
638 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
639 SameTails[commonTailIndex].setBlock(newMBB);
640 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
642 // If we split PredBB, newMBB is the new predecessor.
646 return commonTailIndex;
649 // See if any of the blocks in MergePotentials (which all have a common single
650 // successor, or all have no successor) can be tail-merged. If there is a
651 // successor, any blocks in MergePotentials that are not tail-merged and
652 // are not immediately before Succ must have an unconditional branch to
653 // Succ added (but the predecessor/successor lists need no adjustment).
654 // The lone predecessor of Succ that falls through into Succ,
655 // if any, is given in PredBB.
657 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
658 MachineBasicBlock* PredBB) {
659 bool MadeChange = false;
661 // Except for the special cases below, tail-merge if there are at least
662 // this many instructions in common.
663 unsigned minCommonTailLength = TailMergeSize;
665 // If there's a successor block, there are some cases which don't require
666 // new branching and as such are very likely to be profitable.
668 if (SuccBB->pred_size() == MergePotentials.size() &&
669 !MergePotentials[0].getBlock()->empty()) {
670 // If all the predecessors have at least one tail instruction in common,
671 // merging is very likely to be a win since it won't require an increase
672 // in static branches, and it will decrease the static instruction count.
673 bool AllPredsMatch = true;
674 MachineBasicBlock::iterator FirstNonTerm;
675 unsigned MinNumTerms = CountTerminators(MergePotentials[0].getBlock(),
677 if (FirstNonTerm != MergePotentials[0].getBlock()->end()) {
678 for (unsigned i = 1, e = MergePotentials.size(); i != e; ++i) {
679 MachineBasicBlock::iterator OtherFirstNonTerm;
680 unsigned NumTerms = CountTerminators(MergePotentials[0].getBlock(),
682 if (NumTerms < MinNumTerms)
683 MinNumTerms = NumTerms;
684 if (OtherFirstNonTerm == MergePotentials[i].getBlock()->end() ||
685 OtherFirstNonTerm->isIdenticalTo(FirstNonTerm)) {
686 AllPredsMatch = false;
691 // If they all have an instruction in common, do any amount of merging.
693 minCommonTailLength = MinNumTerms + 1;
698 DEBUG(errs() << "\nTryTailMergeBlocks: ";
699 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
700 errs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
701 << (i == e-1 ? "" : ", ");
704 errs() << " with successor BB#" << SuccBB->getNumber() << '\n';
706 errs() << " which has fall-through from BB#"
707 << PredBB->getNumber() << "\n";
709 errs() << "Looking for common tails of at least "
710 << minCommonTailLength << " instruction"
711 << (minCommonTailLength == 1 ? "" : "s") << '\n';
714 // Sort by hash value so that blocks with identical end sequences sort
716 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
718 // Walk through equivalence sets looking for actual exact matches.
719 while (MergePotentials.size() > 1) {
720 unsigned CurHash = MergePotentials.back().getHash();
722 // Build SameTails, identifying the set of blocks with this hash code
723 // and with the maximum number of instructions in common.
724 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
728 // If we didn't find any pair that has at least minCommonTailLength
729 // instructions in common, remove all blocks with this hash code and retry.
730 if (SameTails.empty()) {
731 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
735 // If one of the blocks is the entire common tail (and not the entry
736 // block, which we can't jump to), we can treat all blocks with this same
737 // tail at once. Use PredBB if that is one of the possibilities, as that
738 // will not introduce any extra branches.
739 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
740 getParent()->begin();
741 unsigned commonTailIndex = SameTails.size();
742 // If there are two blocks, check to see if one can be made to fall through
744 if (SameTails.size() == 2 &&
745 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
746 SameTails[1].tailIsWholeBlock())
748 else if (SameTails.size() == 2 &&
749 SameTails[1].getBlock()->isLayoutSuccessor(
750 SameTails[0].getBlock()) &&
751 SameTails[0].tailIsWholeBlock())
754 // Otherwise just pick one, favoring the fall-through predecessor if
756 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
757 MachineBasicBlock *MBB = SameTails[i].getBlock();
758 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
764 if (SameTails[i].tailIsWholeBlock())
769 if (commonTailIndex == SameTails.size() ||
770 (SameTails[commonTailIndex].getBlock() == PredBB &&
771 !SameTails[commonTailIndex].tailIsWholeBlock())) {
772 // None of the blocks consist entirely of the common tail.
773 // Split a block so that one does.
774 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
777 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
778 // MBB is common tail. Adjust all other BB's to jump to this one.
779 // Traversal must be forwards so erases work.
780 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber()
782 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
783 if (commonTailIndex == i)
785 DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber()
786 << (i == e-1 ? "" : ", "));
787 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
788 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
789 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
790 MergePotentials.erase(SameTails[i].getMPIter());
792 DEBUG(errs() << "\n");
793 // We leave commonTailIndex in the worklist in case there are other blocks
794 // that match it with a smaller number of instructions.
800 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
802 if (!EnableTailMerge) return false;
804 bool MadeChange = false;
806 // First find blocks with no successors.
807 MergePotentials.clear();
808 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
810 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
813 // See if we can do any tail merging on those.
814 if (MergePotentials.size() < TailMergeThreshold &&
815 MergePotentials.size() >= 2)
816 MadeChange |= TryTailMergeBlocks(NULL, NULL);
818 // Look at blocks (IBB) with multiple predecessors (PBB).
819 // We change each predecessor to a canonical form, by
820 // (1) temporarily removing any unconditional branch from the predecessor
822 // (2) alter conditional branches so they branch to the other block
823 // not IBB; this may require adding back an unconditional branch to IBB
824 // later, where there wasn't one coming in. E.g.
826 // fallthrough to QBB
829 // with a conceptual B to IBB after that, which never actually exists.
830 // With those changes, we see whether the predecessors' tails match,
831 // and merge them if so. We change things out of canonical form and
832 // back to the way they were later in the process. (OptimizeBranches
833 // would undo some of this, but we can't use it, because we'd get into
834 // a compile-time infinite loop repeatedly doing and undoing the same
837 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end();
839 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
840 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
841 MachineBasicBlock *IBB = I;
842 MachineBasicBlock *PredBB = prior(I);
843 MergePotentials.clear();
844 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
847 MachineBasicBlock* PBB = *P;
848 // Skip blocks that loop to themselves, can't tail merge these.
851 // Visit each predecessor only once.
852 if (!UniquePreds.insert(PBB))
854 MachineBasicBlock *TBB = 0, *FBB = 0;
855 SmallVector<MachineOperand, 4> Cond;
856 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
857 // Failing case: IBB is the target of a cbr, and
858 // we cannot reverse the branch.
859 SmallVector<MachineOperand, 4> NewCond(Cond);
860 if (!Cond.empty() && TBB == IBB) {
861 if (TII->ReverseBranchCondition(NewCond))
863 // This is the QBB case described above
865 FBB = next(MachineFunction::iterator(PBB));
867 // Failing case: the only way IBB can be reached from PBB is via
868 // exception handling. Happens for landing pads. Would be nice
869 // to have a bit in the edge so we didn't have to do all this.
870 if (IBB->isLandingPad()) {
871 MachineFunction::iterator IP = PBB; IP++;
872 MachineBasicBlock* PredNextBB = NULL;
876 if (IBB != PredNextBB) // fallthrough
879 if (TBB != IBB && FBB != IBB) // cbr then ubr
881 } else if (Cond.empty()) {
882 if (TBB != IBB) // ubr
885 if (TBB != IBB && IBB != PredNextBB) // cbr
889 // Remove the unconditional branch at the end, if any.
890 if (TBB && (Cond.empty() || FBB)) {
891 TII->RemoveBranch(*PBB);
893 // reinsert conditional branch only, for now
894 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
896 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
900 if (MergePotentials.size() >= 2)
901 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
902 // Reinsert an unconditional branch if needed.
903 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
904 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
905 if (MergePotentials.size() == 1 &&
906 MergePotentials.begin()->getBlock() != PredBB)
907 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
913 //===----------------------------------------------------------------------===//
914 // Branch Optimization
915 //===----------------------------------------------------------------------===//
917 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
918 bool MadeChange = false;
920 // Make sure blocks are numbered in order
923 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
924 MachineBasicBlock *MBB = I++;
925 MadeChange |= OptimizeBlock(MBB);
927 // If it is dead, remove it.
928 if (MBB->pred_empty()) {
929 RemoveDeadBlock(MBB);
938 /// CanFallThrough - Return true if the specified block (with the specified
939 /// branch condition) can implicitly transfer control to the block after it by
940 /// falling off the end of it. This should return false if it can reach the
941 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
943 /// True is a conservative answer.
945 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
946 bool BranchUnAnalyzable,
947 MachineBasicBlock *TBB,
948 MachineBasicBlock *FBB,
949 const SmallVectorImpl<MachineOperand> &Cond) {
950 MachineFunction::iterator Fallthrough = CurBB;
952 // If FallthroughBlock is off the end of the function, it can't fall through.
953 if (Fallthrough == CurBB->getParent()->end())
956 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
957 if (!CurBB->isSuccessor(Fallthrough))
960 // If we couldn't analyze the branch, examine the last instruction.
961 // If the block doesn't end in a known control barrier, assume fallthrough
962 // is possible. The isPredicable check is needed because this code can be
963 // called during IfConversion, where an instruction which is normally a
964 // Barrier is predicated and thus no longer an actual control barrier. This
965 // is over-conservative though, because if an instruction isn't actually
966 // predicated we could still treat it like a barrier.
967 if (BranchUnAnalyzable)
968 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() ||
969 CurBB->back().getDesc().isPredicable();
971 // If there is no branch, control always falls through.
972 if (TBB == 0) return true;
974 // If there is some explicit branch to the fallthrough block, it can obviously
975 // reach, even though the branch should get folded to fall through implicitly.
976 if (MachineFunction::iterator(TBB) == Fallthrough ||
977 MachineFunction::iterator(FBB) == Fallthrough)
980 // If it's an unconditional branch to some block not the fall through, it
981 // doesn't fall through.
982 if (Cond.empty()) return false;
984 // Otherwise, if it is conditional and has no explicit false block, it falls
989 /// CanFallThrough - Return true if the specified can implicitly transfer
990 /// control to the block after it by falling off the end of it. This should
991 /// return false if it can reach the block after it, but it uses an explicit
992 /// branch to do so (e.g. a table jump).
994 /// True is a conservative answer.
996 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
997 MachineBasicBlock *TBB = 0, *FBB = 0;
998 SmallVector<MachineOperand, 4> Cond;
999 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true);
1000 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
1003 /// IsBetterFallthrough - Return true if it would be clearly better to
1004 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1005 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1006 /// result in infinite loops.
1007 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1008 MachineBasicBlock *MBB2) {
1009 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1010 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1011 // optimize branches that branch to either a return block or an assert block
1012 // into a fallthrough to the return.
1013 if (MBB1->empty() || MBB2->empty()) return false;
1015 // If there is a clear successor ordering we make sure that one block
1016 // will fall through to the next
1017 if (MBB1->isSuccessor(MBB2)) return true;
1018 if (MBB2->isSuccessor(MBB1)) return false;
1020 MachineInstr *MBB1I = --MBB1->end();
1021 MachineInstr *MBB2I = --MBB2->end();
1022 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
1025 /// TailDuplicate - MBB unconditionally branches to SuccBB. If it is profitable,
1026 /// duplicate SuccBB's contents in MBB to eliminate the branch.
1027 bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB,
1028 bool PrevFallsThrough,
1029 MachineFunction &MF) {
1030 // Don't try to tail-duplicate single-block loops.
1031 if (TailBB->isSuccessor(TailBB))
1034 // Don't tail-duplicate a block which will soon be folded into its successor.
1035 if (TailBB->succ_size() == 1 &&
1036 TailBB->succ_begin()[0]->pred_size() == 1)
1039 // Duplicate up to one less that the tail-merge threshold, so that we don't
1040 // get into an infinite loop between duplicating and merging. When optimizing
1041 // for size, duplicate only one, because one branch instruction can be
1042 // eliminated to compensate for the duplication.
1043 unsigned MaxDuplicateCount =
1044 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ?
1045 1 : (TailMergeSize - 1);
1047 // Check the instructions in the block to determine whether tail-duplication
1048 // is invalid or unlikely to be unprofitable.
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 unless its the
1089 // only predecessor.
1090 if (PredBB->isLayoutSuccessor(TailBB) &&
1092 TailBB->pred_size() != 1)
1095 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB
1096 << "From Succ: " << *TailBB);
1098 // Remove PredBB's unconditional branch.
1099 TII->RemoveBranch(*PredBB);
1100 // Clone the contents of TailBB into PredBB.
1101 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end();
1103 MachineInstr *NewMI = MF.CloneMachineInstr(I);
1104 PredBB->insert(PredBB->end(), NewMI);
1108 PredBB->removeSuccessor(PredBB->succ_begin());
1109 assert(PredBB->succ_empty() &&
1110 "TailDuplicate called on block with multiple successors!");
1111 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
1112 E = TailBB->succ_end(); I != E; ++I)
1113 PredBB->addSuccessor(*I);
1121 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1122 /// block. This is never called on the entry block.
1123 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1124 bool MadeChange = false;
1125 MachineFunction &MF = *MBB->getParent();
1128 MachineFunction::iterator FallThrough = MBB;
1131 // If this block is empty, make everyone use its fall-through, not the block
1132 // explicitly. Landing pads should not do this since the landing-pad table
1133 // points to this block. Blocks with their addresses taken shouldn't be
1135 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1136 // Dead block? Leave for cleanup later.
1137 if (MBB->pred_empty()) return MadeChange;
1139 if (FallThrough == MF.end()) {
1140 // TODO: Simplify preds to not branch here if possible!
1142 // Rewrite all predecessors of the old block to go to the fallthrough
1144 while (!MBB->pred_empty()) {
1145 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1146 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1148 // If MBB was the target of a jump table, update jump tables to go to the
1149 // fallthrough instead.
1150 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough);
1156 // Check to see if we can simplify the terminator of the block before this
1158 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1160 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1161 SmallVector<MachineOperand, 4> PriorCond;
1162 bool PriorUnAnalyzable =
1163 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1164 if (!PriorUnAnalyzable) {
1165 // If the CFG for the prior block has extra edges, remove them.
1166 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1167 !PriorCond.empty());
1169 // If the previous branch is conditional and both conditions go to the same
1170 // destination, remove the branch, replacing it with an unconditional one or
1172 if (PriorTBB && PriorTBB == PriorFBB) {
1173 TII->RemoveBranch(PrevBB);
1175 if (PriorTBB != MBB)
1176 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1179 goto ReoptimizeBlock;
1182 // If the previous block unconditionally falls through to this block and
1183 // this block has no other predecessors, move the contents of this block
1184 // into the prior block. This doesn't usually happen when SimplifyCFG
1185 // has been used, but it can happen tail duplication eliminates all the
1186 // non-branch predecessors of a block leaving only the fall-through edge.
1187 // This has to check PrevBB->succ_size() because EH edges are ignored by
1189 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1190 PrevBB.succ_size() == 1 &&
1191 !MBB->hasAddressTaken()) {
1192 DEBUG(errs() << "\nMerging into block: " << PrevBB
1193 << "From MBB: " << *MBB);
1194 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1195 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1196 assert(PrevBB.succ_empty());
1197 PrevBB.transferSuccessors(MBB);
1202 // If the previous branch *only* branches to *this* block (conditional or
1203 // not) remove the branch.
1204 if (PriorTBB == MBB && PriorFBB == 0) {
1205 TII->RemoveBranch(PrevBB);
1208 goto ReoptimizeBlock;
1211 // If the prior block branches somewhere else on the condition and here if
1212 // the condition is false, remove the uncond second branch.
1213 if (PriorFBB == MBB) {
1214 TII->RemoveBranch(PrevBB);
1215 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1218 goto ReoptimizeBlock;
1221 // If the prior block branches here on true and somewhere else on false, and
1222 // if the branch condition is reversible, reverse the branch to create a
1224 if (PriorTBB == MBB) {
1225 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1226 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1227 TII->RemoveBranch(PrevBB);
1228 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
1231 goto ReoptimizeBlock;
1235 // If this block has no successors (e.g. it is a return block or ends with
1236 // a call to a no-return function like abort or __cxa_throw) and if the pred
1237 // falls through into this block, and if it would otherwise fall through
1238 // into the block after this, move this block to the end of the function.
1240 // We consider it more likely that execution will stay in the function (e.g.
1241 // due to loops) than it is to exit it. This asserts in loops etc, moving
1242 // the assert condition out of the loop body.
1243 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1244 MachineFunction::iterator(PriorTBB) == FallThrough &&
1245 !CanFallThrough(MBB)) {
1246 bool DoTransform = true;
1248 // We have to be careful that the succs of PredBB aren't both no-successor
1249 // blocks. If neither have successors and if PredBB is the second from
1250 // last block in the function, we'd just keep swapping the two blocks for
1251 // last. Only do the swap if one is clearly better to fall through than
1253 if (FallThrough == --MF.end() &&
1254 !IsBetterFallthrough(PriorTBB, MBB))
1255 DoTransform = false;
1257 // We don't want to do this transformation if we have control flow like:
1266 // In this case, we could actually be moving the return block *into* a
1268 if (DoTransform && !MBB->succ_empty() &&
1269 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
1270 DoTransform = false;
1274 // Reverse the branch so we will fall through on the previous true cond.
1275 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1276 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1277 DEBUG(errs() << "\nMoving MBB: " << *MBB
1278 << "To make fallthrough to: " << *PriorTBB << "\n");
1280 TII->RemoveBranch(PrevBB);
1281 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1283 // Move this block to the end of the function.
1284 MBB->moveAfter(--MF.end());
1293 // Analyze the branch in the current block.
1294 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1295 SmallVector<MachineOperand, 4> CurCond;
1296 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1297 if (!CurUnAnalyzable) {
1298 // If the CFG for the prior block has extra edges, remove them.
1299 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1301 // If this is a two-way branch, and the FBB branches to this block, reverse
1302 // the condition so the single-basic-block loop is faster. Instead of:
1303 // Loop: xxx; jcc Out; jmp Loop
1305 // Loop: xxx; jncc Loop; jmp Out
1306 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1307 SmallVector<MachineOperand, 4> NewCond(CurCond);
1308 if (!TII->ReverseBranchCondition(NewCond)) {
1309 TII->RemoveBranch(*MBB);
1310 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1313 goto ReoptimizeBlock;
1318 // If this branch is the only thing in its block, see if we can forward
1319 // other blocks across it.
1320 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1321 MBB->begin()->getDesc().isBranch() && CurTBB != MBB &&
1322 !MBB->hasAddressTaken()) {
1323 // This block may contain just an unconditional branch. Because there can
1324 // be 'non-branch terminators' in the block, try removing the branch and
1325 // then seeing if the block is empty.
1326 TII->RemoveBranch(*MBB);
1328 // If this block is just an unconditional branch to CurTBB, we can
1329 // usually completely eliminate the block. The only case we cannot
1330 // completely eliminate the block is when the block before this one
1331 // falls through into MBB and we can't understand the prior block's branch
1334 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
1335 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1336 !PrevBB.isSuccessor(MBB)) {
1337 // If the prior block falls through into us, turn it into an
1338 // explicit branch to us to make updates simpler.
1339 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1340 PriorTBB != MBB && PriorFBB != MBB) {
1341 if (PriorTBB == 0) {
1342 assert(PriorCond.empty() && PriorFBB == 0 &&
1343 "Bad branch analysis");
1346 assert(PriorFBB == 0 && "Machine CFG out of date!");
1349 TII->RemoveBranch(PrevBB);
1350 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1353 // Iterate through all the predecessors, revectoring each in-turn.
1355 bool DidChange = false;
1356 bool HasBranchToSelf = false;
1357 while(PI != MBB->pred_size()) {
1358 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1360 // If this block has an uncond branch to itself, leave it.
1362 HasBranchToSelf = true;
1365 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1366 // If this change resulted in PMBB ending in a conditional
1367 // branch where both conditions go to the same destination,
1368 // change this to an unconditional branch (and fix the CFG).
1369 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1370 SmallVector<MachineOperand, 4> NewCurCond;
1371 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1372 NewCurFBB, NewCurCond, true);
1373 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1374 TII->RemoveBranch(*PMBB);
1376 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1379 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1384 // Change any jumptables to go to the new MBB.
1385 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB);
1389 if (!HasBranchToSelf) return MadeChange;
1394 // Add the branch back if the block is more than just an uncond branch.
1395 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1399 // Now we know that there was no fall-through into this block, check to
1400 // see if it has a fall-through into its successor.
1401 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1403 bool PrevFallsThru = CanFallThrough(&PrevBB, PriorUnAnalyzable,
1404 PriorTBB, PriorFBB, PriorCond);
1406 // If this block is small, unconditionally branched to, and does not
1407 // fall through, tail-duplicate its instructions into its predecessors
1408 // to eliminate a (dynamic) branch.
1410 if (TailDuplicate(MBB, PrevFallsThru, MF)) {
1415 // If the prior block doesn't fall through into this block, and if this
1416 // block doesn't fall through into some other block, see if we can find a
1417 // place to move this block where a fall-through will happen.
1418 if (!PrevFallsThru) {
1419 if (!MBB->isLandingPad()) {
1420 // Check all the predecessors of this block. If one of them has no fall
1421 // throughs, move this block right after it.
1422 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1423 E = MBB->pred_end(); PI != E; ++PI) {
1424 // Analyze the branch at the end of the pred.
1425 MachineBasicBlock *PredBB = *PI;
1426 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1427 MachineBasicBlock *PredTBB, *PredFBB;
1428 SmallVector<MachineOperand, 4> PredCond;
1429 if (PredBB != MBB && !CanFallThrough(PredBB) &&
1430 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1431 && (!CurFallsThru || !CurTBB || !CurFBB)
1432 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1433 // If the current block doesn't fall through, just move it.
1434 // If the current block can fall through and does not end with a
1435 // conditional branch, we need to append an unconditional jump to
1436 // the (current) next block. To avoid a possible compile-time
1437 // infinite loop, move blocks only backward in this case.
1438 // Also, if there are already 2 branches here, we cannot add a third;
1439 // this means we have the case
1444 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1446 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1448 MBB->moveAfter(PredBB);
1450 goto ReoptimizeBlock;
1455 if (!CurFallsThru) {
1456 // Check all successors to see if we can move this block before it.
1457 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1458 E = MBB->succ_end(); SI != E; ++SI) {
1459 // Analyze the branch at the end of the block before the succ.
1460 MachineBasicBlock *SuccBB = *SI;
1461 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1463 // If this block doesn't already fall-through to that successor, and if
1464 // the succ doesn't already have a block that can fall through into it,
1465 // and if the successor isn't an EH destination, we can arrange for the
1466 // fallthrough to happen.
1467 if (SuccBB != MBB && &*SuccPrev != MBB &&
1468 !CanFallThrough(SuccPrev) && !CurUnAnalyzable &&
1469 !SuccBB->isLandingPad()) {
1470 MBB->moveBefore(SuccBB);
1472 goto ReoptimizeBlock;
1476 // Okay, there is no really great place to put this block. If, however,
1477 // the block before this one would be a fall-through if this block were
1478 // removed, move this block to the end of the function.
1479 MachineBasicBlock *PrevTBB, *PrevFBB;
1480 SmallVector<MachineOperand, 4> PrevCond;
1481 if (FallThrough != MF.end() &&
1482 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1483 PrevBB.isSuccessor(FallThrough)) {
1484 MBB->moveAfter(--MF.end());