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");
45 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
46 cl::init(cl::BOU_UNSET), cl::Hidden);
48 // Throttle for huge numbers of predecessors (compile speed problems)
49 static cl::opt<unsigned>
50 TailMergeThreshold("tail-merge-threshold",
51 cl::desc("Max number of predecessors to consider tail merging"),
52 cl::init(150), cl::Hidden);
54 // Heuristic for tail merging (and, inversely, tail duplication).
55 // TODO: This should be replaced with a target query.
56 static cl::opt<unsigned>
57 TailMergeSize("tail-merge-size",
58 cl::desc("Min number of instructions to consider tail merging"),
59 cl::init(3), cl::Hidden);
62 /// BranchFolderPass - Wrap branch folder in a machine function pass.
63 class BranchFolderPass : public MachineFunctionPass,
67 explicit BranchFolderPass(bool defaultEnableTailMerge)
68 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {}
70 virtual bool runOnMachineFunction(MachineFunction &MF);
71 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
75 char BranchFolderPass::ID = 0;
77 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
78 return new BranchFolderPass(DefaultEnableTailMerge);
81 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
82 return OptimizeFunction(MF,
83 MF.getTarget().getInstrInfo(),
84 MF.getTarget().getRegisterInfo(),
85 getAnalysisIfAvailable<MachineModuleInfo>());
89 BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
90 switch (FlagEnableTailMerge) {
91 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
92 case cl::BOU_TRUE: EnableTailMerge = true; break;
93 case cl::BOU_FALSE: EnableTailMerge = false; break;
97 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
98 /// function, updating the CFG.
99 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
100 assert(MBB->pred_empty() && "MBB must be dead!");
101 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
103 MachineFunction *MF = MBB->getParent();
104 // drop all successors.
105 while (!MBB->succ_empty())
106 MBB->removeSuccessor(MBB->succ_end()-1);
108 // If there are any labels in the basic block, unregister them from
109 // MachineModuleInfo.
110 if (MMI && !MBB->empty()) {
111 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
114 // The label ID # is always operand #0, an immediate.
115 MMI->InvalidateLabel(I->getOperand(0).getImm());
123 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
124 /// followed by terminators, and if the implicitly defined registers are not
125 /// used by the terminators, remove those implicit_def's. e.g.
127 /// r0 = implicit_def
128 /// r1 = implicit_def
130 /// This block can be optimized away later if the implicit instructions are
132 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
133 SmallSet<unsigned, 4> ImpDefRegs;
134 MachineBasicBlock::iterator I = MBB->begin();
135 while (I != MBB->end()) {
136 if (!I->isImplicitDef())
138 unsigned Reg = I->getOperand(0).getReg();
139 ImpDefRegs.insert(Reg);
140 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
141 unsigned SubReg = *SubRegs; ++SubRegs)
142 ImpDefRegs.insert(SubReg);
145 if (ImpDefRegs.empty())
148 MachineBasicBlock::iterator FirstTerm = I;
149 while (I != MBB->end()) {
150 if (!TII->isUnpredicatedTerminator(I))
152 // See if it uses any of the implicitly defined registers.
153 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
154 MachineOperand &MO = I->getOperand(i);
155 if (!MO.isReg() || !MO.isUse())
157 unsigned Reg = MO.getReg();
158 if (ImpDefRegs.count(Reg))
165 while (I != FirstTerm) {
166 MachineInstr *ImpDefMI = &*I;
168 MBB->erase(ImpDefMI);
174 /// OptimizeFunction - Perhaps branch folding, tail merging and other
175 /// CFG optimizations on the given function.
176 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
177 const TargetInstrInfo *tii,
178 const TargetRegisterInfo *tri,
179 MachineModuleInfo *mmi) {
180 if (!tii) return false;
186 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
188 // Fix CFG. The later algorithms expect it to be right.
189 bool MadeChange = false;
190 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
191 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
192 SmallVector<MachineOperand, 4> Cond;
193 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
194 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
195 MadeChange |= OptimizeImpDefsBlock(MBB);
198 bool MadeChangeThisIteration = true;
199 while (MadeChangeThisIteration) {
200 MadeChangeThisIteration = false;
201 MadeChangeThisIteration |= TailMergeBlocks(MF);
202 MadeChangeThisIteration |= OptimizeBranches(MF);
203 MadeChange |= MadeChangeThisIteration;
206 // See if any jump tables have become mergable or dead as the code generator
208 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
214 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
215 // Figure out how these jump tables should be merged.
216 std::vector<unsigned> JTMapping;
217 JTMapping.reserve(JTs.size());
219 // We always keep the 0th jump table.
220 JTMapping.push_back(0);
222 // Scan the jump tables, seeing if there are any duplicates. Note that this
223 // is N^2, which should be fixed someday.
224 for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
225 if (JTs[i].MBBs.empty())
226 JTMapping.push_back(i);
228 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
231 // If a jump table was merge with another one, walk the function rewriting
232 // references to jump tables to reference the new JT ID's. Keep track of
233 // whether we see a jump table idx, if not, we can delete the JT.
234 BitVector JTIsLive(JTs.size());
235 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
237 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
239 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
240 MachineOperand &Op = I->getOperand(op);
241 if (!Op.isJTI()) continue;
242 unsigned NewIdx = JTMapping[Op.getIndex()];
245 // Remember that this JT is live.
246 JTIsLive.set(NewIdx);
250 // Finally, remove dead jump tables. This happens either because the
251 // indirect jump was unreachable (and thus deleted) or because the jump
252 // table was merged with some other one.
253 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
254 if (!JTIsLive.test(i)) {
255 JTI->RemoveJumpTable(i);
263 //===----------------------------------------------------------------------===//
264 // Tail Merging of Blocks
265 //===----------------------------------------------------------------------===//
267 /// HashMachineInstr - Compute a hash value for MI and its operands.
268 static unsigned HashMachineInstr(const MachineInstr *MI) {
269 unsigned Hash = MI->getOpcode();
270 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
271 const MachineOperand &Op = MI->getOperand(i);
273 // Merge in bits from the operand if easy.
274 unsigned OperandHash = 0;
275 switch (Op.getType()) {
276 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
277 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
278 case MachineOperand::MO_MachineBasicBlock:
279 OperandHash = Op.getMBB()->getNumber();
281 case MachineOperand::MO_FrameIndex:
282 case MachineOperand::MO_ConstantPoolIndex:
283 case MachineOperand::MO_JumpTableIndex:
284 OperandHash = Op.getIndex();
286 case MachineOperand::MO_GlobalAddress:
287 case MachineOperand::MO_ExternalSymbol:
288 // Global address / external symbol are too hard, don't bother, but do
289 // pull in the offset.
290 OperandHash = Op.getOffset();
295 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
300 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
301 /// with no successors, we hash two instructions, because cross-jumping
302 /// only saves code when at least two instructions are removed (since a
303 /// branch must be inserted). For blocks with a successor, one of the
304 /// two blocks to be tail-merged will end with a branch already, so
305 /// it gains to cross-jump even for one instruction.
306 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
307 unsigned minCommonTailLength) {
308 MachineBasicBlock::const_iterator I = MBB->end();
309 if (I == MBB->begin())
310 return 0; // Empty MBB.
313 // Skip debug info so it will not affect codegen.
314 while (I->isDebugValue()) {
316 return 0; // MBB empty except for debug info.
319 unsigned Hash = HashMachineInstr(I);
321 if (I == MBB->begin() || minCommonTailLength == 1)
322 return Hash; // Single instr MBB.
325 while (I->isDebugValue()) {
327 return Hash; // MBB with single non-debug instr.
330 // Hash in the second-to-last instruction.
331 Hash ^= HashMachineInstr(I) << 2;
335 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
336 /// of instructions they actually have in common together at their end. Return
337 /// iterators for the first shared instruction in each block.
338 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
339 MachineBasicBlock *MBB2,
340 MachineBasicBlock::iterator &I1,
341 MachineBasicBlock::iterator &I2) {
345 unsigned TailLen = 0;
346 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
348 // Skip debugging pseudos; necessary to avoid changing the code.
349 while (I1->isDebugValue()) {
350 if (I1==MBB1->begin()) {
351 while (I2->isDebugValue()) {
352 if (I2==MBB2->begin())
353 // I1==DBG at begin; I2==DBG at begin
358 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
363 // I1==first (untested) non-DBG preceding known match
364 while (I2->isDebugValue()) {
365 if (I2==MBB2->begin()) {
367 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
372 // I1, I2==first (untested) non-DBGs preceding known match
373 if (!I1->isIdenticalTo(I2) ||
374 // FIXME: This check is dubious. It's used to get around a problem where
375 // people incorrectly expect inline asm directives to remain in the same
376 // relative order. This is untenable because normal compiler
377 // optimizations (like this one) may reorder and/or merge these
385 // Back past possible debugging pseudos at beginning of block. This matters
386 // when one block differs from the other only by whether debugging pseudos
387 // are present at the beginning. (This way, the various checks later for
388 // I1==MBB1->begin() work as expected.)
389 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
391 while (I2->isDebugValue()) {
392 if (I2 == MBB2->begin()) {
399 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
401 while (I1->isDebugValue()) {
402 if (I1 == MBB1->begin())
411 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
412 /// after it, replacing it with an unconditional branch to NewDest. This
413 /// returns true if OldInst's block is modified, false if NewDest is modified.
414 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
415 MachineBasicBlock *NewDest) {
416 MachineBasicBlock *OldBB = OldInst->getParent();
418 // Remove all the old successors of OldBB from the CFG.
419 while (!OldBB->succ_empty())
420 OldBB->removeSuccessor(OldBB->succ_begin());
422 // Remove all the dead instructions from the end of OldBB.
423 OldBB->erase(OldInst, OldBB->end());
425 // If OldBB isn't immediately before OldBB, insert a branch to it.
426 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
427 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
428 OldBB->addSuccessor(NewDest);
432 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
433 /// MBB so that the part before the iterator falls into the part starting at the
434 /// iterator. This returns the new MBB.
435 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
436 MachineBasicBlock::iterator BBI1) {
437 MachineFunction &MF = *CurMBB.getParent();
439 // Create the fall-through block.
440 MachineFunction::iterator MBBI = &CurMBB;
441 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
442 CurMBB.getParent()->insert(++MBBI, NewMBB);
444 // Move all the successors of this block to the specified block.
445 NewMBB->transferSuccessors(&CurMBB);
447 // Add an edge from CurMBB to NewMBB for the fall-through.
448 CurMBB.addSuccessor(NewMBB);
450 // Splice the code over.
451 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
453 // For targets that use the register scavenger, we must maintain LiveIns.
455 RS->enterBasicBlock(&CurMBB);
457 RS->forward(prior(CurMBB.end()));
458 BitVector RegsLiveAtExit(TRI->getNumRegs());
459 RS->getRegsUsed(RegsLiveAtExit, false);
460 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
461 if (RegsLiveAtExit[i])
462 NewMBB->addLiveIn(i);
468 /// EstimateRuntime - Make a rough estimate for how long it will take to run
469 /// the specified code.
470 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
471 MachineBasicBlock::iterator E) {
473 for (; I != E; ++I) {
474 if (I->isDebugValue())
476 const TargetInstrDesc &TID = I->getDesc();
479 else if (TID.mayLoad() || TID.mayStore())
487 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
488 // branches temporarily for tail merging). In the case where CurMBB ends
489 // with a conditional branch to the next block, optimize by reversing the
490 // test and conditionally branching to SuccMBB instead.
491 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
492 const TargetInstrInfo *TII) {
493 MachineFunction *MF = CurMBB->getParent();
494 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB));
495 MachineBasicBlock *TBB = 0, *FBB = 0;
496 SmallVector<MachineOperand, 4> Cond;
497 if (I != MF->end() &&
498 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
499 MachineBasicBlock *NextBB = I;
500 if (TBB == NextBB && !Cond.empty() && !FBB) {
501 if (!TII->ReverseBranchCondition(Cond)) {
502 TII->RemoveBranch(*CurMBB);
503 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
508 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
512 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
513 if (getHash() < o.getHash())
515 else if (getHash() > o.getHash())
517 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
519 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
522 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
523 // an object with itself.
524 #ifndef _GLIBCXX_DEBUG
525 llvm_unreachable("Predecessor appears twice");
531 /// CountTerminators - Count the number of terminators in the given
532 /// block and set I to the position of the first non-terminator, if there
533 /// is one, or MBB->end() otherwise.
534 static unsigned CountTerminators(MachineBasicBlock *MBB,
535 MachineBasicBlock::iterator &I) {
537 unsigned NumTerms = 0;
539 if (I == MBB->begin()) {
544 if (!I->getDesc().isTerminator()) break;
550 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
551 /// and decide if it would be profitable to merge those tails. Return the
552 /// length of the common tail and iterators to the first common instruction
554 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
555 MachineBasicBlock *MBB2,
556 unsigned minCommonTailLength,
557 unsigned &CommonTailLen,
558 MachineBasicBlock::iterator &I1,
559 MachineBasicBlock::iterator &I2,
560 MachineBasicBlock *SuccBB,
561 MachineBasicBlock *PredBB) {
562 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
563 MachineFunction *MF = MBB1->getParent();
565 if (CommonTailLen == 0)
568 // It's almost always profitable to merge any number of non-terminator
569 // instructions with the block that falls through into the common successor.
570 if (MBB1 == PredBB || MBB2 == PredBB) {
571 MachineBasicBlock::iterator I;
572 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
573 if (CommonTailLen > NumTerms)
577 // If one of the blocks can be completely merged and happens to be in
578 // a position where the other could fall through into it, merge any number
579 // of instructions, because it can be done without a branch.
580 // TODO: If the blocks are not adjacent, move one of them so that they are?
581 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
583 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
586 // If both blocks have an unconditional branch temporarily stripped out,
587 // count that as an additional common instruction for the following
589 unsigned EffectiveTailLen = CommonTailLen;
590 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
591 !MBB1->back().getDesc().isBarrier() &&
592 !MBB2->back().getDesc().isBarrier())
595 // Check if the common tail is long enough to be worthwhile.
596 if (EffectiveTailLen >= minCommonTailLength)
599 // If we are optimizing for code size, 2 instructions in common is enough if
600 // we don't have to split a block. At worst we will be introducing 1 new
601 // branch instruction, which is likely to be smaller than the 2
602 // instructions that would be deleted in the merge.
603 if (EffectiveTailLen >= 2 &&
604 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
605 (I1 == MBB1->begin() || I2 == MBB2->begin()))
611 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
612 /// hash CurHash (guaranteed to match the last element). Build the vector
613 /// SameTails of all those that have the (same) largest number of instructions
614 /// in common of any pair of these blocks. SameTails entries contain an
615 /// iterator into MergePotentials (from which the MachineBasicBlock can be
616 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
617 /// instruction where the matching code sequence begins.
618 /// Order of elements in SameTails is the reverse of the order in which
619 /// those blocks appear in MergePotentials (where they are not necessarily
621 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
622 unsigned minCommonTailLength,
623 MachineBasicBlock *SuccBB,
624 MachineBasicBlock *PredBB) {
625 unsigned maxCommonTailLength = 0U;
627 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
628 MPIterator HighestMPIter = prior(MergePotentials.end());
629 for (MPIterator CurMPIter = prior(MergePotentials.end()),
630 B = MergePotentials.begin();
631 CurMPIter != B && CurMPIter->getHash() == CurHash;
633 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
634 unsigned CommonTailLen;
635 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
637 CommonTailLen, TrialBBI1, TrialBBI2,
639 if (CommonTailLen > maxCommonTailLength) {
641 maxCommonTailLength = CommonTailLen;
642 HighestMPIter = CurMPIter;
643 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
645 if (HighestMPIter == CurMPIter &&
646 CommonTailLen == maxCommonTailLength)
647 SameTails.push_back(SameTailElt(I, TrialBBI2));
653 return maxCommonTailLength;
656 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
657 /// MergePotentials, restoring branches at ends of blocks as appropriate.
658 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
659 MachineBasicBlock *SuccBB,
660 MachineBasicBlock *PredBB) {
661 MPIterator CurMPIter, B;
662 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
663 CurMPIter->getHash() == CurHash;
665 // Put the unconditional branch back, if we need one.
666 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
667 if (SuccBB && CurMBB != PredBB)
668 FixTail(CurMBB, SuccBB, TII);
672 if (CurMPIter->getHash() != CurHash)
674 MergePotentials.erase(CurMPIter, MergePotentials.end());
677 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
678 /// only of the common tail. Create a block that does by splitting one.
679 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
680 unsigned maxCommonTailLength) {
681 unsigned commonTailIndex = 0;
682 unsigned TimeEstimate = ~0U;
683 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
684 // Use PredBB if possible; that doesn't require a new branch.
685 if (SameTails[i].getBlock() == PredBB) {
689 // Otherwise, make a (fairly bogus) choice based on estimate of
690 // how long it will take the various blocks to execute.
691 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
692 SameTails[i].getTailStartPos());
693 if (t <= TimeEstimate) {
699 MachineBasicBlock::iterator BBI =
700 SameTails[commonTailIndex].getTailStartPos();
701 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
703 // If the common tail includes any debug info we will take it pretty
704 // randomly from one of the inputs. Might be better to remove it?
705 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
706 << maxCommonTailLength);
708 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
709 SameTails[commonTailIndex].setBlock(newMBB);
710 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
712 // If we split PredBB, newMBB is the new predecessor.
716 return commonTailIndex;
719 // See if any of the blocks in MergePotentials (which all have a common single
720 // successor, or all have no successor) can be tail-merged. If there is a
721 // successor, any blocks in MergePotentials that are not tail-merged and
722 // are not immediately before Succ must have an unconditional branch to
723 // Succ added (but the predecessor/successor lists need no adjustment).
724 // The lone predecessor of Succ that falls through into Succ,
725 // if any, is given in PredBB.
727 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
728 MachineBasicBlock *PredBB) {
729 bool MadeChange = false;
731 // Except for the special cases below, tail-merge if there are at least
732 // this many instructions in common.
733 unsigned minCommonTailLength = TailMergeSize;
735 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
736 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
737 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
738 << (i == e-1 ? "" : ", ");
741 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
743 dbgs() << " which has fall-through from BB#"
744 << PredBB->getNumber() << "\n";
746 dbgs() << "Looking for common tails of at least "
747 << minCommonTailLength << " instruction"
748 << (minCommonTailLength == 1 ? "" : "s") << '\n';
751 // Sort by hash value so that blocks with identical end sequences sort
753 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
755 // Walk through equivalence sets looking for actual exact matches.
756 while (MergePotentials.size() > 1) {
757 unsigned CurHash = MergePotentials.back().getHash();
759 // Build SameTails, identifying the set of blocks with this hash code
760 // and with the maximum number of instructions in common.
761 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
765 // If we didn't find any pair that has at least minCommonTailLength
766 // instructions in common, remove all blocks with this hash code and retry.
767 if (SameTails.empty()) {
768 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
772 // If one of the blocks is the entire common tail (and not the entry
773 // block, which we can't jump to), we can treat all blocks with this same
774 // tail at once. Use PredBB if that is one of the possibilities, as that
775 // will not introduce any extra branches.
776 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
777 getParent()->begin();
778 unsigned commonTailIndex = SameTails.size();
779 // If there are two blocks, check to see if one can be made to fall through
781 if (SameTails.size() == 2 &&
782 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
783 SameTails[1].tailIsWholeBlock())
785 else if (SameTails.size() == 2 &&
786 SameTails[1].getBlock()->isLayoutSuccessor(
787 SameTails[0].getBlock()) &&
788 SameTails[0].tailIsWholeBlock())
791 // Otherwise just pick one, favoring the fall-through predecessor if
793 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
794 MachineBasicBlock *MBB = SameTails[i].getBlock();
795 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
801 if (SameTails[i].tailIsWholeBlock())
806 if (commonTailIndex == SameTails.size() ||
807 (SameTails[commonTailIndex].getBlock() == PredBB &&
808 !SameTails[commonTailIndex].tailIsWholeBlock())) {
809 // None of the blocks consist entirely of the common tail.
810 // Split a block so that one does.
811 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
814 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
815 // MBB is common tail. Adjust all other BB's to jump to this one.
816 // Traversal must be forwards so erases work.
817 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
819 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
820 if (commonTailIndex == i)
822 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
823 << (i == e-1 ? "" : ", "));
824 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
825 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
826 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
827 MergePotentials.erase(SameTails[i].getMPIter());
829 DEBUG(dbgs() << "\n");
830 // We leave commonTailIndex in the worklist in case there are other blocks
831 // that match it with a smaller number of instructions.
837 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
839 if (!EnableTailMerge) return false;
841 bool MadeChange = false;
843 // First find blocks with no successors.
844 MergePotentials.clear();
845 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
847 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
850 // See if we can do any tail merging on those.
851 if (MergePotentials.size() < TailMergeThreshold &&
852 MergePotentials.size() >= 2)
853 MadeChange |= TryTailMergeBlocks(NULL, NULL);
855 // Look at blocks (IBB) with multiple predecessors (PBB).
856 // We change each predecessor to a canonical form, by
857 // (1) temporarily removing any unconditional branch from the predecessor
859 // (2) alter conditional branches so they branch to the other block
860 // not IBB; this may require adding back an unconditional branch to IBB
861 // later, where there wasn't one coming in. E.g.
863 // fallthrough to QBB
866 // with a conceptual B to IBB after that, which never actually exists.
867 // With those changes, we see whether the predecessors' tails match,
868 // and merge them if so. We change things out of canonical form and
869 // back to the way they were later in the process. (OptimizeBranches
870 // would undo some of this, but we can't use it, because we'd get into
871 // a compile-time infinite loop repeatedly doing and undoing the same
874 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
876 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
877 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
878 MachineBasicBlock *IBB = I;
879 MachineBasicBlock *PredBB = prior(I);
880 MergePotentials.clear();
881 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
884 MachineBasicBlock *PBB = *P;
885 // Skip blocks that loop to themselves, can't tail merge these.
888 // Visit each predecessor only once.
889 if (!UniquePreds.insert(PBB))
891 MachineBasicBlock *TBB = 0, *FBB = 0;
892 SmallVector<MachineOperand, 4> Cond;
893 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
894 // Failing case: IBB is the target of a cbr, and
895 // we cannot reverse the branch.
896 SmallVector<MachineOperand, 4> NewCond(Cond);
897 if (!Cond.empty() && TBB == IBB) {
898 if (TII->ReverseBranchCondition(NewCond))
900 // This is the QBB case described above
902 FBB = llvm::next(MachineFunction::iterator(PBB));
904 // Failing case: the only way IBB can be reached from PBB is via
905 // exception handling. Happens for landing pads. Would be nice
906 // to have a bit in the edge so we didn't have to do all this.
907 if (IBB->isLandingPad()) {
908 MachineFunction::iterator IP = PBB; IP++;
909 MachineBasicBlock *PredNextBB = NULL;
913 if (IBB != PredNextBB) // fallthrough
916 if (TBB != IBB && FBB != IBB) // cbr then ubr
918 } else if (Cond.empty()) {
919 if (TBB != IBB) // ubr
922 if (TBB != IBB && IBB != PredNextBB) // cbr
926 // Remove the unconditional branch at the end, if any.
927 if (TBB && (Cond.empty() || FBB)) {
928 TII->RemoveBranch(*PBB);
930 // reinsert conditional branch only, for now
931 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
933 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
937 if (MergePotentials.size() >= 2)
938 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
939 // Reinsert an unconditional branch if needed.
940 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
941 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
942 if (MergePotentials.size() == 1 &&
943 MergePotentials.begin()->getBlock() != PredBB)
944 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
950 //===----------------------------------------------------------------------===//
951 // Branch Optimization
952 //===----------------------------------------------------------------------===//
954 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
955 bool MadeChange = false;
957 // Make sure blocks are numbered in order
960 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
961 MachineBasicBlock *MBB = I++;
962 MadeChange |= OptimizeBlock(MBB);
964 // If it is dead, remove it.
965 if (MBB->pred_empty()) {
966 RemoveDeadBlock(MBB);
974 // Blocks should be considered empty if they contain only debug info;
975 // else the debug info would affect codegen.
976 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
979 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
980 MBBI!=MBBE; ++MBBI) {
981 if (!MBBI->isDebugValue())
987 // Blocks with only debug info and branches should be considered the same
988 // as blocks with only branches.
989 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
990 MachineBasicBlock::iterator MBBI, MBBE;
991 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
992 if (!MBBI->isDebugValue())
995 return (MBBI->getDesc().isBranch());
998 /// IsBetterFallthrough - Return true if it would be clearly better to
999 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1000 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1001 /// result in infinite loops.
1002 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1003 MachineBasicBlock *MBB2) {
1004 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1005 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1006 // optimize branches that branch to either a return block or an assert block
1007 // into a fallthrough to the return.
1008 if (MBB1->empty() || MBB2->empty()) return false;
1010 // If there is a clear successor ordering we make sure that one block
1011 // will fall through to the next
1012 if (MBB1->isSuccessor(MBB2)) return true;
1013 if (MBB2->isSuccessor(MBB1)) return false;
1015 MachineInstr *MBB1I = --MBB1->end();
1016 MachineInstr *MBB2I = --MBB2->end();
1017 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
1020 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1021 /// block. This is never called on the entry block.
1022 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1023 bool MadeChange = false;
1024 MachineFunction &MF = *MBB->getParent();
1027 MachineFunction::iterator FallThrough = MBB;
1030 // If this block is empty, make everyone use its fall-through, not the block
1031 // explicitly. Landing pads should not do this since the landing-pad table
1032 // points to this block. Blocks with their addresses taken shouldn't be
1034 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1035 // Dead block? Leave for cleanup later.
1036 if (MBB->pred_empty()) return MadeChange;
1038 if (FallThrough == MF.end()) {
1039 // TODO: Simplify preds to not branch here if possible!
1041 // Rewrite all predecessors of the old block to go to the fallthrough
1043 while (!MBB->pred_empty()) {
1044 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1045 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1047 // If MBB was the target of a jump table, update jump tables to go to the
1048 // fallthrough instead.
1049 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1050 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1056 // Check to see if we can simplify the terminator of the block before this
1058 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1060 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1061 SmallVector<MachineOperand, 4> PriorCond;
1062 bool PriorUnAnalyzable =
1063 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1064 if (!PriorUnAnalyzable) {
1065 // If the CFG for the prior block has extra edges, remove them.
1066 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1067 !PriorCond.empty());
1069 // If the previous branch is conditional and both conditions go to the same
1070 // destination, remove the branch, replacing it with an unconditional one or
1072 if (PriorTBB && PriorTBB == PriorFBB) {
1073 TII->RemoveBranch(PrevBB);
1075 if (PriorTBB != MBB)
1076 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1079 goto ReoptimizeBlock;
1082 // If the previous block unconditionally falls through to this block and
1083 // this block has no other predecessors, move the contents of this block
1084 // into the prior block. This doesn't usually happen when SimplifyCFG
1085 // has been used, but it can happen if tail merging splits a fall-through
1086 // predecessor of a block.
1087 // This has to check PrevBB->succ_size() because EH edges are ignored by
1089 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1090 PrevBB.succ_size() == 1 &&
1091 !MBB->hasAddressTaken()) {
1092 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1093 << "From MBB: " << *MBB);
1094 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1095 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1096 assert(PrevBB.succ_empty());
1097 PrevBB.transferSuccessors(MBB);
1102 // If the previous branch *only* branches to *this* block (conditional or
1103 // not) remove the branch.
1104 if (PriorTBB == MBB && PriorFBB == 0) {
1105 TII->RemoveBranch(PrevBB);
1108 goto ReoptimizeBlock;
1111 // If the prior block branches somewhere else on the condition and here if
1112 // the condition is false, remove the uncond second branch.
1113 if (PriorFBB == MBB) {
1114 TII->RemoveBranch(PrevBB);
1115 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1118 goto ReoptimizeBlock;
1121 // If the prior block branches here on true and somewhere else on false, and
1122 // if the branch condition is reversible, reverse the branch to create a
1124 if (PriorTBB == MBB) {
1125 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1126 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1127 TII->RemoveBranch(PrevBB);
1128 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
1131 goto ReoptimizeBlock;
1135 // If this block has no successors (e.g. it is a return block or ends with
1136 // a call to a no-return function like abort or __cxa_throw) and if the pred
1137 // falls through into this block, and if it would otherwise fall through
1138 // into the block after this, move this block to the end of the function.
1140 // We consider it more likely that execution will stay in the function (e.g.
1141 // due to loops) than it is to exit it. This asserts in loops etc, moving
1142 // the assert condition out of the loop body.
1143 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1144 MachineFunction::iterator(PriorTBB) == FallThrough &&
1145 !MBB->canFallThrough()) {
1146 bool DoTransform = true;
1148 // We have to be careful that the succs of PredBB aren't both no-successor
1149 // blocks. If neither have successors and if PredBB is the second from
1150 // last block in the function, we'd just keep swapping the two blocks for
1151 // last. Only do the swap if one is clearly better to fall through than
1153 if (FallThrough == --MF.end() &&
1154 !IsBetterFallthrough(PriorTBB, MBB))
1155 DoTransform = false;
1157 // We don't want to do this transformation if we have control flow like:
1166 // In this case, we could actually be moving the return block *into* a
1168 if (DoTransform && !MBB->succ_empty() &&
1169 (!PriorTBB->canFallThrough() || PriorTBB->empty()))
1170 DoTransform = false;
1174 // Reverse the branch so we will fall through on the previous true cond.
1175 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1176 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1177 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1178 << "To make fallthrough to: " << *PriorTBB << "\n");
1180 TII->RemoveBranch(PrevBB);
1181 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1183 // Move this block to the end of the function.
1184 MBB->moveAfter(--MF.end());
1193 // Analyze the branch in the current block.
1194 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1195 SmallVector<MachineOperand, 4> CurCond;
1196 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1197 if (!CurUnAnalyzable) {
1198 // If the CFG for the prior block has extra edges, remove them.
1199 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1201 // If this is a two-way branch, and the FBB branches to this block, reverse
1202 // the condition so the single-basic-block loop is faster. Instead of:
1203 // Loop: xxx; jcc Out; jmp Loop
1205 // Loop: xxx; jncc Loop; jmp Out
1206 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1207 SmallVector<MachineOperand, 4> NewCond(CurCond);
1208 if (!TII->ReverseBranchCondition(NewCond)) {
1209 TII->RemoveBranch(*MBB);
1210 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1213 goto ReoptimizeBlock;
1217 // If this branch is the only thing in its block, see if we can forward
1218 // other blocks across it.
1219 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1220 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1221 !MBB->hasAddressTaken()) {
1222 // This block may contain just an unconditional branch. Because there can
1223 // be 'non-branch terminators' in the block, try removing the branch and
1224 // then seeing if the block is empty.
1225 TII->RemoveBranch(*MBB);
1226 // If the only things remaining in the block are debug info, remove these
1227 // as well, so this will behave the same as an empty block in non-debug
1229 if (!MBB->empty()) {
1230 bool NonDebugInfoFound = false;
1231 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1233 if (!I->isDebugValue()) {
1234 NonDebugInfoFound = true;
1238 if (!NonDebugInfoFound)
1239 // Make the block empty, losing the debug info (we could probably
1240 // improve this in some cases.)
1241 MBB->erase(MBB->begin(), MBB->end());
1243 // If this block is just an unconditional branch to CurTBB, we can
1244 // usually completely eliminate the block. The only case we cannot
1245 // completely eliminate the block is when the block before this one
1246 // falls through into MBB and we can't understand the prior block's branch
1249 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1250 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1251 !PrevBB.isSuccessor(MBB)) {
1252 // If the prior block falls through into us, turn it into an
1253 // explicit branch to us to make updates simpler.
1254 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1255 PriorTBB != MBB && PriorFBB != MBB) {
1256 if (PriorTBB == 0) {
1257 assert(PriorCond.empty() && PriorFBB == 0 &&
1258 "Bad branch analysis");
1261 assert(PriorFBB == 0 && "Machine CFG out of date!");
1264 TII->RemoveBranch(PrevBB);
1265 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1268 // Iterate through all the predecessors, revectoring each in-turn.
1270 bool DidChange = false;
1271 bool HasBranchToSelf = false;
1272 while(PI != MBB->pred_size()) {
1273 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1275 // If this block has an uncond branch to itself, leave it.
1277 HasBranchToSelf = true;
1280 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1281 // If this change resulted in PMBB ending in a conditional
1282 // branch where both conditions go to the same destination,
1283 // change this to an unconditional branch (and fix the CFG).
1284 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1285 SmallVector<MachineOperand, 4> NewCurCond;
1286 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1287 NewCurFBB, NewCurCond, true);
1288 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1289 TII->RemoveBranch(*PMBB);
1291 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1294 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1299 // Change any jumptables to go to the new MBB.
1300 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1301 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1305 if (!HasBranchToSelf) return MadeChange;
1310 // Add the branch back if the block is more than just an uncond branch.
1311 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1315 // If the prior block doesn't fall through into this block, and if this
1316 // block doesn't fall through into some other block, see if we can find a
1317 // place to move this block where a fall-through will happen.
1318 if (!PrevBB.canFallThrough()) {
1320 // Now we know that there was no fall-through into this block, check to
1321 // see if it has a fall-through into its successor.
1322 bool CurFallsThru = MBB->canFallThrough();
1324 if (!MBB->isLandingPad()) {
1325 // Check all the predecessors of this block. If one of them has no fall
1326 // throughs, move this block right after it.
1327 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1328 E = MBB->pred_end(); PI != E; ++PI) {
1329 // Analyze the branch at the end of the pred.
1330 MachineBasicBlock *PredBB = *PI;
1331 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1332 MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
1333 SmallVector<MachineOperand, 4> PredCond;
1334 if (PredBB != MBB && !PredBB->canFallThrough() &&
1335 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1336 && (!CurFallsThru || !CurTBB || !CurFBB)
1337 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1338 // If the current block doesn't fall through, just move it.
1339 // If the current block can fall through and does not end with a
1340 // conditional branch, we need to append an unconditional jump to
1341 // the (current) next block. To avoid a possible compile-time
1342 // infinite loop, move blocks only backward in this case.
1343 // Also, if there are already 2 branches here, we cannot add a third;
1344 // this means we have the case
1349 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
1351 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1353 MBB->moveAfter(PredBB);
1355 goto ReoptimizeBlock;
1360 if (!CurFallsThru) {
1361 // Check all successors to see if we can move this block before it.
1362 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1363 E = MBB->succ_end(); SI != E; ++SI) {
1364 // Analyze the branch at the end of the block before the succ.
1365 MachineBasicBlock *SuccBB = *SI;
1366 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1368 // If this block doesn't already fall-through to that successor, and if
1369 // the succ doesn't already have a block that can fall through into it,
1370 // and if the successor isn't an EH destination, we can arrange for the
1371 // fallthrough to happen.
1372 if (SuccBB != MBB && &*SuccPrev != MBB &&
1373 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1374 !SuccBB->isLandingPad()) {
1375 MBB->moveBefore(SuccBB);
1377 goto ReoptimizeBlock;
1381 // Okay, there is no really great place to put this block. If, however,
1382 // the block before this one would be a fall-through if this block were
1383 // removed, move this block to the end of the function.
1384 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
1385 SmallVector<MachineOperand, 4> PrevCond;
1386 if (FallThrough != MF.end() &&
1387 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1388 PrevBB.isSuccessor(FallThrough)) {
1389 MBB->moveAfter(--MF.end());