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);
112 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
113 /// followed by terminators, and if the implicitly defined registers are not
114 /// used by the terminators, remove those implicit_def's. e.g.
116 /// r0 = implicit_def
117 /// r1 = implicit_def
119 /// This block can be optimized away later if the implicit instructions are
121 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
122 SmallSet<unsigned, 4> ImpDefRegs;
123 MachineBasicBlock::iterator I = MBB->begin();
124 while (I != MBB->end()) {
125 if (!I->isImplicitDef())
127 unsigned Reg = I->getOperand(0).getReg();
128 ImpDefRegs.insert(Reg);
129 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
130 unsigned SubReg = *SubRegs; ++SubRegs)
131 ImpDefRegs.insert(SubReg);
134 if (ImpDefRegs.empty())
137 MachineBasicBlock::iterator FirstTerm = I;
138 while (I != MBB->end()) {
139 if (!TII->isUnpredicatedTerminator(I))
141 // See if it uses any of the implicitly defined registers.
142 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
143 MachineOperand &MO = I->getOperand(i);
144 if (!MO.isReg() || !MO.isUse())
146 unsigned Reg = MO.getReg();
147 if (ImpDefRegs.count(Reg))
154 while (I != FirstTerm) {
155 MachineInstr *ImpDefMI = &*I;
157 MBB->erase(ImpDefMI);
163 /// OptimizeFunction - Perhaps branch folding, tail merging and other
164 /// CFG optimizations on the given function.
165 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
166 const TargetInstrInfo *tii,
167 const TargetRegisterInfo *tri,
168 MachineModuleInfo *mmi) {
169 if (!tii) return false;
175 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
177 // Fix CFG. The later algorithms expect it to be right.
178 bool MadeChange = false;
179 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
180 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
181 SmallVector<MachineOperand, 4> Cond;
182 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
183 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
184 MadeChange |= OptimizeImpDefsBlock(MBB);
187 bool MadeChangeThisIteration = true;
188 while (MadeChangeThisIteration) {
189 MadeChangeThisIteration = false;
190 MadeChangeThisIteration |= TailMergeBlocks(MF);
191 MadeChangeThisIteration |= OptimizeBranches(MF);
192 MadeChange |= MadeChangeThisIteration;
195 // See if any jump tables have become mergable or dead as the code generator
197 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
203 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
204 // Figure out how these jump tables should be merged.
205 std::vector<unsigned> JTMapping;
206 JTMapping.reserve(JTs.size());
208 // We always keep the 0th jump table.
209 JTMapping.push_back(0);
211 // Scan the jump tables, seeing if there are any duplicates. Note that this
212 // is N^2, which should be fixed someday.
213 for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
214 if (JTs[i].MBBs.empty())
215 JTMapping.push_back(i);
217 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
220 // If a jump table was merge with another one, walk the function rewriting
221 // references to jump tables to reference the new JT ID's. Keep track of
222 // whether we see a jump table idx, if not, we can delete the JT.
223 BitVector JTIsLive(JTs.size());
224 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
226 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
228 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
229 MachineOperand &Op = I->getOperand(op);
230 if (!Op.isJTI()) continue;
231 unsigned NewIdx = JTMapping[Op.getIndex()];
234 // Remember that this JT is live.
235 JTIsLive.set(NewIdx);
239 // Finally, remove dead jump tables. This happens either because the
240 // indirect jump was unreachable (and thus deleted) or because the jump
241 // table was merged with some other one.
242 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
243 if (!JTIsLive.test(i)) {
244 JTI->RemoveJumpTable(i);
252 //===----------------------------------------------------------------------===//
253 // Tail Merging of Blocks
254 //===----------------------------------------------------------------------===//
256 /// HashMachineInstr - Compute a hash value for MI and its operands.
257 static unsigned HashMachineInstr(const MachineInstr *MI) {
258 unsigned Hash = MI->getOpcode();
259 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
260 const MachineOperand &Op = MI->getOperand(i);
262 // Merge in bits from the operand if easy.
263 unsigned OperandHash = 0;
264 switch (Op.getType()) {
265 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
266 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
267 case MachineOperand::MO_MachineBasicBlock:
268 OperandHash = Op.getMBB()->getNumber();
270 case MachineOperand::MO_FrameIndex:
271 case MachineOperand::MO_ConstantPoolIndex:
272 case MachineOperand::MO_JumpTableIndex:
273 OperandHash = Op.getIndex();
275 case MachineOperand::MO_GlobalAddress:
276 case MachineOperand::MO_ExternalSymbol:
277 // Global address / external symbol are too hard, don't bother, but do
278 // pull in the offset.
279 OperandHash = Op.getOffset();
284 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
289 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
290 /// with no successors, we hash two instructions, because cross-jumping
291 /// only saves code when at least two instructions are removed (since a
292 /// branch must be inserted). For blocks with a successor, one of the
293 /// two blocks to be tail-merged will end with a branch already, so
294 /// it gains to cross-jump even for one instruction.
295 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
296 unsigned minCommonTailLength) {
297 MachineBasicBlock::const_iterator I = MBB->end();
298 if (I == MBB->begin())
299 return 0; // Empty MBB.
302 // Skip debug info so it will not affect codegen.
303 while (I->isDebugValue()) {
305 return 0; // MBB empty except for debug info.
308 unsigned Hash = HashMachineInstr(I);
310 if (I == MBB->begin() || minCommonTailLength == 1)
311 return Hash; // Single instr MBB.
314 while (I->isDebugValue()) {
316 return Hash; // MBB with single non-debug instr.
319 // Hash in the second-to-last instruction.
320 Hash ^= HashMachineInstr(I) << 2;
324 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
325 /// of instructions they actually have in common together at their end. Return
326 /// iterators for the first shared instruction in each block.
327 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
328 MachineBasicBlock *MBB2,
329 MachineBasicBlock::iterator &I1,
330 MachineBasicBlock::iterator &I2) {
334 unsigned TailLen = 0;
335 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
337 // Skip debugging pseudos; necessary to avoid changing the code.
338 while (I1->isDebugValue()) {
339 if (I1==MBB1->begin()) {
340 while (I2->isDebugValue()) {
341 if (I2==MBB2->begin())
342 // I1==DBG at begin; I2==DBG at begin
347 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
352 // I1==first (untested) non-DBG preceding known match
353 while (I2->isDebugValue()) {
354 if (I2==MBB2->begin()) {
356 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
361 // I1, I2==first (untested) non-DBGs preceding known match
362 if (!I1->isIdenticalTo(I2) ||
363 // FIXME: This check is dubious. It's used to get around a problem where
364 // people incorrectly expect inline asm directives to remain in the same
365 // relative order. This is untenable because normal compiler
366 // optimizations (like this one) may reorder and/or merge these
374 // Back past possible debugging pseudos at beginning of block. This matters
375 // when one block differs from the other only by whether debugging pseudos
376 // are present at the beginning. (This way, the various checks later for
377 // I1==MBB1->begin() work as expected.)
378 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
380 while (I2->isDebugValue()) {
381 if (I2 == MBB2->begin()) {
388 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
390 while (I1->isDebugValue()) {
391 if (I1 == MBB1->begin())
400 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
401 /// after it, replacing it with an unconditional branch to NewDest. This
402 /// returns true if OldInst's block is modified, false if NewDest is modified.
403 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
404 MachineBasicBlock *NewDest) {
405 MachineBasicBlock *OldBB = OldInst->getParent();
407 // Remove all the old successors of OldBB from the CFG.
408 while (!OldBB->succ_empty())
409 OldBB->removeSuccessor(OldBB->succ_begin());
411 // Remove all the dead instructions from the end of OldBB.
412 OldBB->erase(OldInst, OldBB->end());
414 // If OldBB isn't immediately before OldBB, insert a branch to it.
415 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
416 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
417 OldBB->addSuccessor(NewDest);
421 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
422 /// MBB so that the part before the iterator falls into the part starting at the
423 /// iterator. This returns the new MBB.
424 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
425 MachineBasicBlock::iterator BBI1) {
426 MachineFunction &MF = *CurMBB.getParent();
428 // Create the fall-through block.
429 MachineFunction::iterator MBBI = &CurMBB;
430 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
431 CurMBB.getParent()->insert(++MBBI, NewMBB);
433 // Move all the successors of this block to the specified block.
434 NewMBB->transferSuccessors(&CurMBB);
436 // Add an edge from CurMBB to NewMBB for the fall-through.
437 CurMBB.addSuccessor(NewMBB);
439 // Splice the code over.
440 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
442 // For targets that use the register scavenger, we must maintain LiveIns.
444 RS->enterBasicBlock(&CurMBB);
446 RS->forward(prior(CurMBB.end()));
447 BitVector RegsLiveAtExit(TRI->getNumRegs());
448 RS->getRegsUsed(RegsLiveAtExit, false);
449 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
450 if (RegsLiveAtExit[i])
451 NewMBB->addLiveIn(i);
457 /// EstimateRuntime - Make a rough estimate for how long it will take to run
458 /// the specified code.
459 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
460 MachineBasicBlock::iterator E) {
462 for (; I != E; ++I) {
463 if (I->isDebugValue())
465 const TargetInstrDesc &TID = I->getDesc();
468 else if (TID.mayLoad() || TID.mayStore())
476 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
477 // branches temporarily for tail merging). In the case where CurMBB ends
478 // with a conditional branch to the next block, optimize by reversing the
479 // test and conditionally branching to SuccMBB instead.
480 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
481 const TargetInstrInfo *TII) {
482 MachineFunction *MF = CurMBB->getParent();
483 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB));
484 MachineBasicBlock *TBB = 0, *FBB = 0;
485 SmallVector<MachineOperand, 4> Cond;
486 if (I != MF->end() &&
487 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
488 MachineBasicBlock *NextBB = I;
489 if (TBB == NextBB && !Cond.empty() && !FBB) {
490 if (!TII->ReverseBranchCondition(Cond)) {
491 TII->RemoveBranch(*CurMBB);
492 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
497 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
501 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
502 if (getHash() < o.getHash())
504 else if (getHash() > o.getHash())
506 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
508 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
511 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
512 // an object with itself.
513 #ifndef _GLIBCXX_DEBUG
514 llvm_unreachable("Predecessor appears twice");
520 /// CountTerminators - Count the number of terminators in the given
521 /// block and set I to the position of the first non-terminator, if there
522 /// is one, or MBB->end() otherwise.
523 static unsigned CountTerminators(MachineBasicBlock *MBB,
524 MachineBasicBlock::iterator &I) {
526 unsigned NumTerms = 0;
528 if (I == MBB->begin()) {
533 if (!I->getDesc().isTerminator()) break;
539 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
540 /// and decide if it would be profitable to merge those tails. Return the
541 /// length of the common tail and iterators to the first common instruction
543 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
544 MachineBasicBlock *MBB2,
545 unsigned minCommonTailLength,
546 unsigned &CommonTailLen,
547 MachineBasicBlock::iterator &I1,
548 MachineBasicBlock::iterator &I2,
549 MachineBasicBlock *SuccBB,
550 MachineBasicBlock *PredBB) {
551 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
552 MachineFunction *MF = MBB1->getParent();
554 if (CommonTailLen == 0)
557 // It's almost always profitable to merge any number of non-terminator
558 // instructions with the block that falls through into the common successor.
559 if (MBB1 == PredBB || MBB2 == PredBB) {
560 MachineBasicBlock::iterator I;
561 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
562 if (CommonTailLen > NumTerms)
566 // If one of the blocks can be completely merged and happens to be in
567 // a position where the other could fall through into it, merge any number
568 // of instructions, because it can be done without a branch.
569 // TODO: If the blocks are not adjacent, move one of them so that they are?
570 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
572 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
575 // If both blocks have an unconditional branch temporarily stripped out,
576 // count that as an additional common instruction for the following
578 unsigned EffectiveTailLen = CommonTailLen;
579 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
580 !MBB1->back().getDesc().isBarrier() &&
581 !MBB2->back().getDesc().isBarrier())
584 // Check if the common tail is long enough to be worthwhile.
585 if (EffectiveTailLen >= minCommonTailLength)
588 // If we are optimizing for code size, 2 instructions in common is enough if
589 // we don't have to split a block. At worst we will be introducing 1 new
590 // branch instruction, which is likely to be smaller than the 2
591 // instructions that would be deleted in the merge.
592 if (EffectiveTailLen >= 2 &&
593 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
594 (I1 == MBB1->begin() || I2 == MBB2->begin()))
600 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
601 /// hash CurHash (guaranteed to match the last element). Build the vector
602 /// SameTails of all those that have the (same) largest number of instructions
603 /// in common of any pair of these blocks. SameTails entries contain an
604 /// iterator into MergePotentials (from which the MachineBasicBlock can be
605 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
606 /// instruction where the matching code sequence begins.
607 /// Order of elements in SameTails is the reverse of the order in which
608 /// those blocks appear in MergePotentials (where they are not necessarily
610 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
611 unsigned minCommonTailLength,
612 MachineBasicBlock *SuccBB,
613 MachineBasicBlock *PredBB) {
614 unsigned maxCommonTailLength = 0U;
616 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
617 MPIterator HighestMPIter = prior(MergePotentials.end());
618 for (MPIterator CurMPIter = prior(MergePotentials.end()),
619 B = MergePotentials.begin();
620 CurMPIter != B && CurMPIter->getHash() == CurHash;
622 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
623 unsigned CommonTailLen;
624 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
626 CommonTailLen, TrialBBI1, TrialBBI2,
628 if (CommonTailLen > maxCommonTailLength) {
630 maxCommonTailLength = CommonTailLen;
631 HighestMPIter = CurMPIter;
632 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
634 if (HighestMPIter == CurMPIter &&
635 CommonTailLen == maxCommonTailLength)
636 SameTails.push_back(SameTailElt(I, TrialBBI2));
642 return maxCommonTailLength;
645 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
646 /// MergePotentials, restoring branches at ends of blocks as appropriate.
647 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
648 MachineBasicBlock *SuccBB,
649 MachineBasicBlock *PredBB) {
650 MPIterator CurMPIter, B;
651 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
652 CurMPIter->getHash() == CurHash;
654 // Put the unconditional branch back, if we need one.
655 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
656 if (SuccBB && CurMBB != PredBB)
657 FixTail(CurMBB, SuccBB, TII);
661 if (CurMPIter->getHash() != CurHash)
663 MergePotentials.erase(CurMPIter, MergePotentials.end());
666 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
667 /// only of the common tail. Create a block that does by splitting one.
668 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
669 unsigned maxCommonTailLength) {
670 unsigned commonTailIndex = 0;
671 unsigned TimeEstimate = ~0U;
672 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
673 // Use PredBB if possible; that doesn't require a new branch.
674 if (SameTails[i].getBlock() == PredBB) {
678 // Otherwise, make a (fairly bogus) choice based on estimate of
679 // how long it will take the various blocks to execute.
680 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
681 SameTails[i].getTailStartPos());
682 if (t <= TimeEstimate) {
688 MachineBasicBlock::iterator BBI =
689 SameTails[commonTailIndex].getTailStartPos();
690 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
692 // If the common tail includes any debug info we will take it pretty
693 // randomly from one of the inputs. Might be better to remove it?
694 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
695 << maxCommonTailLength);
697 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
698 SameTails[commonTailIndex].setBlock(newMBB);
699 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
701 // If we split PredBB, newMBB is the new predecessor.
705 return commonTailIndex;
708 // See if any of the blocks in MergePotentials (which all have a common single
709 // successor, or all have no successor) can be tail-merged. If there is a
710 // successor, any blocks in MergePotentials that are not tail-merged and
711 // are not immediately before Succ must have an unconditional branch to
712 // Succ added (but the predecessor/successor lists need no adjustment).
713 // The lone predecessor of Succ that falls through into Succ,
714 // if any, is given in PredBB.
716 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
717 MachineBasicBlock *PredBB) {
718 bool MadeChange = false;
720 // Except for the special cases below, tail-merge if there are at least
721 // this many instructions in common.
722 unsigned minCommonTailLength = TailMergeSize;
724 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
725 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
726 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
727 << (i == e-1 ? "" : ", ");
730 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
732 dbgs() << " which has fall-through from BB#"
733 << PredBB->getNumber() << "\n";
735 dbgs() << "Looking for common tails of at least "
736 << minCommonTailLength << " instruction"
737 << (minCommonTailLength == 1 ? "" : "s") << '\n';
740 // Sort by hash value so that blocks with identical end sequences sort
742 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
744 // Walk through equivalence sets looking for actual exact matches.
745 while (MergePotentials.size() > 1) {
746 unsigned CurHash = MergePotentials.back().getHash();
748 // Build SameTails, identifying the set of blocks with this hash code
749 // and with the maximum number of instructions in common.
750 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
754 // If we didn't find any pair that has at least minCommonTailLength
755 // instructions in common, remove all blocks with this hash code and retry.
756 if (SameTails.empty()) {
757 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
761 // If one of the blocks is the entire common tail (and not the entry
762 // block, which we can't jump to), we can treat all blocks with this same
763 // tail at once. Use PredBB if that is one of the possibilities, as that
764 // will not introduce any extra branches.
765 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
766 getParent()->begin();
767 unsigned commonTailIndex = SameTails.size();
768 // If there are two blocks, check to see if one can be made to fall through
770 if (SameTails.size() == 2 &&
771 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
772 SameTails[1].tailIsWholeBlock())
774 else if (SameTails.size() == 2 &&
775 SameTails[1].getBlock()->isLayoutSuccessor(
776 SameTails[0].getBlock()) &&
777 SameTails[0].tailIsWholeBlock())
780 // Otherwise just pick one, favoring the fall-through predecessor if
782 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
783 MachineBasicBlock *MBB = SameTails[i].getBlock();
784 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
790 if (SameTails[i].tailIsWholeBlock())
795 if (commonTailIndex == SameTails.size() ||
796 (SameTails[commonTailIndex].getBlock() == PredBB &&
797 !SameTails[commonTailIndex].tailIsWholeBlock())) {
798 // None of the blocks consist entirely of the common tail.
799 // Split a block so that one does.
800 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
803 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
804 // MBB is common tail. Adjust all other BB's to jump to this one.
805 // Traversal must be forwards so erases work.
806 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
808 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
809 if (commonTailIndex == i)
811 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
812 << (i == e-1 ? "" : ", "));
813 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
814 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
815 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
816 MergePotentials.erase(SameTails[i].getMPIter());
818 DEBUG(dbgs() << "\n");
819 // We leave commonTailIndex in the worklist in case there are other blocks
820 // that match it with a smaller number of instructions.
826 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
828 if (!EnableTailMerge) return false;
830 bool MadeChange = false;
832 // First find blocks with no successors.
833 MergePotentials.clear();
834 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
836 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
839 // See if we can do any tail merging on those.
840 if (MergePotentials.size() < TailMergeThreshold &&
841 MergePotentials.size() >= 2)
842 MadeChange |= TryTailMergeBlocks(NULL, NULL);
844 // Look at blocks (IBB) with multiple predecessors (PBB).
845 // We change each predecessor to a canonical form, by
846 // (1) temporarily removing any unconditional branch from the predecessor
848 // (2) alter conditional branches so they branch to the other block
849 // not IBB; this may require adding back an unconditional branch to IBB
850 // later, where there wasn't one coming in. E.g.
852 // fallthrough to QBB
855 // with a conceptual B to IBB after that, which never actually exists.
856 // With those changes, we see whether the predecessors' tails match,
857 // and merge them if so. We change things out of canonical form and
858 // back to the way they were later in the process. (OptimizeBranches
859 // would undo some of this, but we can't use it, because we'd get into
860 // a compile-time infinite loop repeatedly doing and undoing the same
863 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
865 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
866 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
867 MachineBasicBlock *IBB = I;
868 MachineBasicBlock *PredBB = prior(I);
869 MergePotentials.clear();
870 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
873 MachineBasicBlock *PBB = *P;
874 // Skip blocks that loop to themselves, can't tail merge these.
877 // Visit each predecessor only once.
878 if (!UniquePreds.insert(PBB))
880 MachineBasicBlock *TBB = 0, *FBB = 0;
881 SmallVector<MachineOperand, 4> Cond;
882 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
883 // Failing case: IBB is the target of a cbr, and
884 // we cannot reverse the branch.
885 SmallVector<MachineOperand, 4> NewCond(Cond);
886 if (!Cond.empty() && TBB == IBB) {
887 if (TII->ReverseBranchCondition(NewCond))
889 // This is the QBB case described above
891 FBB = llvm::next(MachineFunction::iterator(PBB));
893 // Failing case: the only way IBB can be reached from PBB is via
894 // exception handling. Happens for landing pads. Would be nice
895 // to have a bit in the edge so we didn't have to do all this.
896 if (IBB->isLandingPad()) {
897 MachineFunction::iterator IP = PBB; IP++;
898 MachineBasicBlock *PredNextBB = NULL;
902 if (IBB != PredNextBB) // fallthrough
905 if (TBB != IBB && FBB != IBB) // cbr then ubr
907 } else if (Cond.empty()) {
908 if (TBB != IBB) // ubr
911 if (TBB != IBB && IBB != PredNextBB) // cbr
915 // Remove the unconditional branch at the end, if any.
916 if (TBB && (Cond.empty() || FBB)) {
917 TII->RemoveBranch(*PBB);
919 // reinsert conditional branch only, for now
920 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
922 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
926 if (MergePotentials.size() >= 2)
927 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
928 // Reinsert an unconditional branch if needed.
929 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
930 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
931 if (MergePotentials.size() == 1 &&
932 MergePotentials.begin()->getBlock() != PredBB)
933 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
939 //===----------------------------------------------------------------------===//
940 // Branch Optimization
941 //===----------------------------------------------------------------------===//
943 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
944 bool MadeChange = false;
946 // Make sure blocks are numbered in order
949 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
950 MachineBasicBlock *MBB = I++;
951 MadeChange |= OptimizeBlock(MBB);
953 // If it is dead, remove it.
954 if (MBB->pred_empty()) {
955 RemoveDeadBlock(MBB);
963 // Blocks should be considered empty if they contain only debug info;
964 // else the debug info would affect codegen.
965 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
968 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
969 MBBI!=MBBE; ++MBBI) {
970 if (!MBBI->isDebugValue())
976 // Blocks with only debug info and branches should be considered the same
977 // as blocks with only branches.
978 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
979 MachineBasicBlock::iterator MBBI, MBBE;
980 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
981 if (!MBBI->isDebugValue())
984 return (MBBI->getDesc().isBranch());
987 /// IsBetterFallthrough - Return true if it would be clearly better to
988 /// fall-through to MBB1 than to fall through into MBB2. This has to return
989 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
990 /// result in infinite loops.
991 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
992 MachineBasicBlock *MBB2) {
993 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
994 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
995 // optimize branches that branch to either a return block or an assert block
996 // into a fallthrough to the return.
997 if (MBB1->empty() || MBB2->empty()) return false;
999 // If there is a clear successor ordering we make sure that one block
1000 // will fall through to the next
1001 if (MBB1->isSuccessor(MBB2)) return true;
1002 if (MBB2->isSuccessor(MBB1)) return false;
1004 MachineInstr *MBB1I = --MBB1->end();
1005 MachineInstr *MBB2I = --MBB2->end();
1006 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
1009 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1010 /// block. This is never called on the entry block.
1011 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1012 bool MadeChange = false;
1013 MachineFunction &MF = *MBB->getParent();
1016 MachineFunction::iterator FallThrough = MBB;
1019 // If this block is empty, make everyone use its fall-through, not the block
1020 // explicitly. Landing pads should not do this since the landing-pad table
1021 // points to this block. Blocks with their addresses taken shouldn't be
1023 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1024 // Dead block? Leave for cleanup later.
1025 if (MBB->pred_empty()) return MadeChange;
1027 if (FallThrough == MF.end()) {
1028 // TODO: Simplify preds to not branch here if possible!
1030 // Rewrite all predecessors of the old block to go to the fallthrough
1032 while (!MBB->pred_empty()) {
1033 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1034 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1036 // If MBB was the target of a jump table, update jump tables to go to the
1037 // fallthrough instead.
1038 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1039 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1045 // Check to see if we can simplify the terminator of the block before this
1047 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1049 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1050 SmallVector<MachineOperand, 4> PriorCond;
1051 bool PriorUnAnalyzable =
1052 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1053 if (!PriorUnAnalyzable) {
1054 // If the CFG for the prior block has extra edges, remove them.
1055 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1056 !PriorCond.empty());
1058 // If the previous branch is conditional and both conditions go to the same
1059 // destination, remove the branch, replacing it with an unconditional one or
1061 if (PriorTBB && PriorTBB == PriorFBB) {
1062 TII->RemoveBranch(PrevBB);
1064 if (PriorTBB != MBB)
1065 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1068 goto ReoptimizeBlock;
1071 // If the previous block unconditionally falls through to this block and
1072 // this block has no other predecessors, move the contents of this block
1073 // into the prior block. This doesn't usually happen when SimplifyCFG
1074 // has been used, but it can happen if tail merging splits a fall-through
1075 // predecessor of a block.
1076 // This has to check PrevBB->succ_size() because EH edges are ignored by
1078 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1079 PrevBB.succ_size() == 1 &&
1080 !MBB->hasAddressTaken()) {
1081 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1082 << "From MBB: " << *MBB);
1083 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1084 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1085 assert(PrevBB.succ_empty());
1086 PrevBB.transferSuccessors(MBB);
1091 // If the previous branch *only* branches to *this* block (conditional or
1092 // not) remove the branch.
1093 if (PriorTBB == MBB && PriorFBB == 0) {
1094 TII->RemoveBranch(PrevBB);
1097 goto ReoptimizeBlock;
1100 // If the prior block branches somewhere else on the condition and here if
1101 // the condition is false, remove the uncond second branch.
1102 if (PriorFBB == MBB) {
1103 TII->RemoveBranch(PrevBB);
1104 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1107 goto ReoptimizeBlock;
1110 // If the prior block branches here on true and somewhere else on false, and
1111 // if the branch condition is reversible, reverse the branch to create a
1113 if (PriorTBB == MBB) {
1114 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1115 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1116 TII->RemoveBranch(PrevBB);
1117 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
1120 goto ReoptimizeBlock;
1124 // If this block has no successors (e.g. it is a return block or ends with
1125 // a call to a no-return function like abort or __cxa_throw) and if the pred
1126 // falls through into this block, and if it would otherwise fall through
1127 // into the block after this, move this block to the end of the function.
1129 // We consider it more likely that execution will stay in the function (e.g.
1130 // due to loops) than it is to exit it. This asserts in loops etc, moving
1131 // the assert condition out of the loop body.
1132 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1133 MachineFunction::iterator(PriorTBB) == FallThrough &&
1134 !MBB->canFallThrough()) {
1135 bool DoTransform = true;
1137 // We have to be careful that the succs of PredBB aren't both no-successor
1138 // blocks. If neither have successors and if PredBB is the second from
1139 // last block in the function, we'd just keep swapping the two blocks for
1140 // last. Only do the swap if one is clearly better to fall through than
1142 if (FallThrough == --MF.end() &&
1143 !IsBetterFallthrough(PriorTBB, MBB))
1144 DoTransform = false;
1146 // We don't want to do this transformation if we have control flow like:
1155 // In this case, we could actually be moving the return block *into* a
1157 if (DoTransform && !MBB->succ_empty() &&
1158 (!PriorTBB->canFallThrough() || PriorTBB->empty()))
1159 DoTransform = false;
1163 // Reverse the branch so we will fall through on the previous true cond.
1164 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1165 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1166 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1167 << "To make fallthrough to: " << *PriorTBB << "\n");
1169 TII->RemoveBranch(PrevBB);
1170 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1172 // Move this block to the end of the function.
1173 MBB->moveAfter(--MF.end());
1182 // Analyze the branch in the current block.
1183 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1184 SmallVector<MachineOperand, 4> CurCond;
1185 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1186 if (!CurUnAnalyzable) {
1187 // If the CFG for the prior block has extra edges, remove them.
1188 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1190 // If this is a two-way branch, and the FBB branches to this block, reverse
1191 // the condition so the single-basic-block loop is faster. Instead of:
1192 // Loop: xxx; jcc Out; jmp Loop
1194 // Loop: xxx; jncc Loop; jmp Out
1195 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1196 SmallVector<MachineOperand, 4> NewCond(CurCond);
1197 if (!TII->ReverseBranchCondition(NewCond)) {
1198 TII->RemoveBranch(*MBB);
1199 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1202 goto ReoptimizeBlock;
1206 // If this branch is the only thing in its block, see if we can forward
1207 // other blocks across it.
1208 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1209 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1210 !MBB->hasAddressTaken()) {
1211 // This block may contain just an unconditional branch. Because there can
1212 // be 'non-branch terminators' in the block, try removing the branch and
1213 // then seeing if the block is empty.
1214 TII->RemoveBranch(*MBB);
1215 // If the only things remaining in the block are debug info, remove these
1216 // as well, so this will behave the same as an empty block in non-debug
1218 if (!MBB->empty()) {
1219 bool NonDebugInfoFound = false;
1220 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1222 if (!I->isDebugValue()) {
1223 NonDebugInfoFound = true;
1227 if (!NonDebugInfoFound)
1228 // Make the block empty, losing the debug info (we could probably
1229 // improve this in some cases.)
1230 MBB->erase(MBB->begin(), MBB->end());
1232 // If this block is just an unconditional branch to CurTBB, we can
1233 // usually completely eliminate the block. The only case we cannot
1234 // completely eliminate the block is when the block before this one
1235 // falls through into MBB and we can't understand the prior block's branch
1238 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1239 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1240 !PrevBB.isSuccessor(MBB)) {
1241 // If the prior block falls through into us, turn it into an
1242 // explicit branch to us to make updates simpler.
1243 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1244 PriorTBB != MBB && PriorFBB != MBB) {
1245 if (PriorTBB == 0) {
1246 assert(PriorCond.empty() && PriorFBB == 0 &&
1247 "Bad branch analysis");
1250 assert(PriorFBB == 0 && "Machine CFG out of date!");
1253 TII->RemoveBranch(PrevBB);
1254 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1257 // Iterate through all the predecessors, revectoring each in-turn.
1259 bool DidChange = false;
1260 bool HasBranchToSelf = false;
1261 while(PI != MBB->pred_size()) {
1262 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1264 // If this block has an uncond branch to itself, leave it.
1266 HasBranchToSelf = true;
1269 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1270 // If this change resulted in PMBB ending in a conditional
1271 // branch where both conditions go to the same destination,
1272 // change this to an unconditional branch (and fix the CFG).
1273 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1274 SmallVector<MachineOperand, 4> NewCurCond;
1275 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1276 NewCurFBB, NewCurCond, true);
1277 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1278 TII->RemoveBranch(*PMBB);
1280 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1283 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1288 // Change any jumptables to go to the new MBB.
1289 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1290 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1294 if (!HasBranchToSelf) return MadeChange;
1299 // Add the branch back if the block is more than just an uncond branch.
1300 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1304 // If the prior block doesn't fall through into this block, and if this
1305 // block doesn't fall through into some other block, see if we can find a
1306 // place to move this block where a fall-through will happen.
1307 if (!PrevBB.canFallThrough()) {
1309 // Now we know that there was no fall-through into this block, check to
1310 // see if it has a fall-through into its successor.
1311 bool CurFallsThru = MBB->canFallThrough();
1313 if (!MBB->isLandingPad()) {
1314 // Check all the predecessors of this block. If one of them has no fall
1315 // throughs, move this block right after it.
1316 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1317 E = MBB->pred_end(); PI != E; ++PI) {
1318 // Analyze the branch at the end of the pred.
1319 MachineBasicBlock *PredBB = *PI;
1320 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1321 MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
1322 SmallVector<MachineOperand, 4> PredCond;
1323 if (PredBB != MBB && !PredBB->canFallThrough() &&
1324 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1325 && (!CurFallsThru || !CurTBB || !CurFBB)
1326 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1327 // If the current block doesn't fall through, just move it.
1328 // If the current block can fall through and does not end with a
1329 // conditional branch, we need to append an unconditional jump to
1330 // the (current) next block. To avoid a possible compile-time
1331 // infinite loop, move blocks only backward in this case.
1332 // Also, if there are already 2 branches here, we cannot add a third;
1333 // this means we have the case
1338 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
1340 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1342 MBB->moveAfter(PredBB);
1344 goto ReoptimizeBlock;
1349 if (!CurFallsThru) {
1350 // Check all successors to see if we can move this block before it.
1351 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1352 E = MBB->succ_end(); SI != E; ++SI) {
1353 // Analyze the branch at the end of the block before the succ.
1354 MachineBasicBlock *SuccBB = *SI;
1355 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1357 // If this block doesn't already fall-through to that successor, and if
1358 // the succ doesn't already have a block that can fall through into it,
1359 // and if the successor isn't an EH destination, we can arrange for the
1360 // fallthrough to happen.
1361 if (SuccBB != MBB && &*SuccPrev != MBB &&
1362 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1363 !SuccBB->isLandingPad()) {
1364 MBB->moveBefore(SuccBB);
1366 goto ReoptimizeBlock;
1370 // Okay, there is no really great place to put this block. If, however,
1371 // the block before this one would be a fall-through if this block were
1372 // removed, move this block to the end of the function.
1373 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
1374 SmallVector<MachineOperand, 4> PrevCond;
1375 if (FallThrough != MF.end() &&
1376 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1377 PrevBB.isSuccessor(FallThrough)) {
1378 MBB->moveAfter(--MF.end());