1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass forwards branches to unconditional branches to make them branch
11 // directly to the target block. This pass often results in dead MBB's, which
14 // Note that this pass must be run after register allocation, it cannot handle
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "branchfolding"
20 #include "BranchFolding.h"
21 #include "llvm/Function.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/RegisterScavenging.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Target/TargetRegisterInfo.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/ADT/SmallSet.h"
35 #include "llvm/ADT/SetVector.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/ADT/STLExtras.h"
41 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
42 STATISTIC(NumBranchOpts, "Number of branches optimized");
43 STATISTIC(NumTailMerge , "Number of block tails merged");
44 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
45 cl::init(cl::BOU_UNSET), cl::Hidden);
46 // Throttle for huge numbers of predecessors (compile speed problems)
47 static cl::opt<unsigned>
48 TailMergeThreshold("tail-merge-threshold",
49 cl::desc("Max number of predecessors to consider tail merging"),
50 cl::init(150), cl::Hidden);
52 // Heuristic for tail merging (and, inversely, tail duplication).
53 // TODO: This should be replaced with a target query.
54 static cl::opt<unsigned>
55 TailMergeSize("tail-merge-size",
56 cl::desc("Min number of instructions to consider tail merging"),
57 cl::init(3), cl::Hidden);
59 char BranchFolderPass::ID = 0;
61 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
62 return new BranchFolderPass(DefaultEnableTailMerge);
65 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
66 return OptimizeFunction(MF,
67 MF.getTarget().getInstrInfo(),
68 MF.getTarget().getRegisterInfo(),
69 getAnalysisIfAvailable<MachineModuleInfo>());
73 BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
74 switch (FlagEnableTailMerge) {
75 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
76 case cl::BOU_TRUE: EnableTailMerge = true; break;
77 case cl::BOU_FALSE: EnableTailMerge = false; break;
81 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
82 /// function, updating the CFG.
83 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
84 assert(MBB->pred_empty() && "MBB must be dead!");
85 DEBUG(errs() << "\nRemoving MBB: " << *MBB);
87 MachineFunction *MF = MBB->getParent();
88 // drop all successors.
89 while (!MBB->succ_empty())
90 MBB->removeSuccessor(MBB->succ_end()-1);
92 // If there are any labels in the basic block, unregister them from
94 if (MMI && !MBB->empty()) {
95 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
98 // The label ID # is always operand #0, an immediate.
99 MMI->InvalidateLabel(I->getOperand(0).getImm());
107 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
108 /// followed by terminators, and if the implicitly defined registers are not
109 /// used by the terminators, remove those implicit_def's. e.g.
111 /// r0 = implicit_def
112 /// r1 = implicit_def
114 /// This block can be optimized away later if the implicit instructions are
116 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
117 SmallSet<unsigned, 4> ImpDefRegs;
118 MachineBasicBlock::iterator I = MBB->begin();
119 while (I != MBB->end()) {
120 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF)
122 unsigned Reg = I->getOperand(0).getReg();
123 ImpDefRegs.insert(Reg);
124 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
125 unsigned SubReg = *SubRegs; ++SubRegs)
126 ImpDefRegs.insert(SubReg);
129 if (ImpDefRegs.empty())
132 MachineBasicBlock::iterator FirstTerm = I;
133 while (I != MBB->end()) {
134 if (!TII->isUnpredicatedTerminator(I))
136 // See if it uses any of the implicitly defined registers.
137 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
138 MachineOperand &MO = I->getOperand(i);
139 if (!MO.isReg() || !MO.isUse())
141 unsigned Reg = MO.getReg();
142 if (ImpDefRegs.count(Reg))
149 while (I != FirstTerm) {
150 MachineInstr *ImpDefMI = &*I;
152 MBB->erase(ImpDefMI);
158 /// OptimizeFunction - Perhaps branch folding, tail merging and other
159 /// CFG optimizations on the given function.
160 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
161 const TargetInstrInfo *tii,
162 const TargetRegisterInfo *tri,
163 MachineModuleInfo *mmi) {
164 if (!tii) return false;
170 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
172 // Fix CFG. The later algorithms expect it to be right.
173 bool MadeChange = false;
174 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
175 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
176 SmallVector<MachineOperand, 4> Cond;
177 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
178 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
179 MadeChange |= OptimizeImpDefsBlock(MBB);
183 bool MadeChangeThisIteration = true;
184 while (MadeChangeThisIteration) {
185 MadeChangeThisIteration = false;
186 MadeChangeThisIteration |= TailMergeBlocks(MF);
187 MadeChangeThisIteration |= OptimizeBranches(MF);
188 MadeChange |= MadeChangeThisIteration;
191 // See if any jump tables have become mergable or dead as the code generator
193 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
194 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
196 // Figure out how these jump tables should be merged.
197 std::vector<unsigned> JTMapping;
198 JTMapping.reserve(JTs.size());
200 // We always keep the 0th jump table.
201 JTMapping.push_back(0);
203 // Scan the jump tables, seeing if there are any duplicates. Note that this
204 // is N^2, which should be fixed someday.
205 for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
206 if (JTs[i].MBBs.empty())
207 JTMapping.push_back(i);
209 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
212 // If a jump table was merge with another one, walk the function rewriting
213 // references to jump tables to reference the new JT ID's. Keep track of
214 // whether we see a jump table idx, if not, we can delete the JT.
215 BitVector JTIsLive(JTs.size());
216 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
218 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
220 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
221 MachineOperand &Op = I->getOperand(op);
222 if (!Op.isJTI()) continue;
223 unsigned NewIdx = JTMapping[Op.getIndex()];
226 // Remember that this JT is live.
227 JTIsLive.set(NewIdx);
231 // Finally, remove dead jump tables. This happens either because the
232 // indirect jump was unreachable (and thus deleted) or because the jump
233 // table was merged with some other one.
234 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
235 if (!JTIsLive.test(i)) {
236 JTI->RemoveJumpTable(i);
245 //===----------------------------------------------------------------------===//
246 // Tail Merging of Blocks
247 //===----------------------------------------------------------------------===//
249 /// HashMachineInstr - Compute a hash value for MI and its operands.
250 static unsigned HashMachineInstr(const MachineInstr *MI) {
251 unsigned Hash = MI->getOpcode();
252 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
253 const MachineOperand &Op = MI->getOperand(i);
255 // Merge in bits from the operand if easy.
256 unsigned OperandHash = 0;
257 switch (Op.getType()) {
258 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
259 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
260 case MachineOperand::MO_MachineBasicBlock:
261 OperandHash = Op.getMBB()->getNumber();
263 case MachineOperand::MO_FrameIndex:
264 case MachineOperand::MO_ConstantPoolIndex:
265 case MachineOperand::MO_JumpTableIndex:
266 OperandHash = Op.getIndex();
268 case MachineOperand::MO_GlobalAddress:
269 case MachineOperand::MO_ExternalSymbol:
270 // Global address / external symbol are too hard, don't bother, but do
271 // pull in the offset.
272 OperandHash = Op.getOffset();
277 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
282 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
283 /// with no successors, we hash two instructions, because cross-jumping
284 /// only saves code when at least two instructions are removed (since a
285 /// branch must be inserted). For blocks with a successor, one of the
286 /// two blocks to be tail-merged will end with a branch already, so
287 /// it gains to cross-jump even for one instruction.
288 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
289 unsigned minCommonTailLength) {
290 MachineBasicBlock::const_iterator I = MBB->end();
291 if (I == MBB->begin())
292 return 0; // Empty MBB.
295 unsigned Hash = HashMachineInstr(I);
297 if (I == MBB->begin() || minCommonTailLength == 1)
298 return Hash; // Single instr MBB.
301 // Hash in the second-to-last instruction.
302 Hash ^= HashMachineInstr(I) << 2;
306 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
307 /// of instructions they actually have in common together at their end. Return
308 /// iterators for the first shared instruction in each block.
309 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
310 MachineBasicBlock *MBB2,
311 MachineBasicBlock::iterator &I1,
312 MachineBasicBlock::iterator &I2) {
316 unsigned TailLen = 0;
317 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
319 if (!I1->isIdenticalTo(I2) ||
320 // FIXME: This check is dubious. It's used to get around a problem where
321 // people incorrectly expect inline asm directives to remain in the same
322 // relative order. This is untenable because normal compiler
323 // optimizations (like this one) may reorder and/or merge these
325 I1->getOpcode() == TargetInstrInfo::INLINEASM) {
334 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
335 /// after it, replacing it with an unconditional branch to NewDest. This
336 /// returns true if OldInst's block is modified, false if NewDest is modified.
337 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
338 MachineBasicBlock *NewDest) {
339 MachineBasicBlock *OldBB = OldInst->getParent();
341 // Remove all the old successors of OldBB from the CFG.
342 while (!OldBB->succ_empty())
343 OldBB->removeSuccessor(OldBB->succ_begin());
345 // Remove all the dead instructions from the end of OldBB.
346 OldBB->erase(OldInst, OldBB->end());
348 // If OldBB isn't immediately before OldBB, insert a branch to it.
349 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
350 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
351 OldBB->addSuccessor(NewDest);
355 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
356 /// MBB so that the part before the iterator falls into the part starting at the
357 /// iterator. This returns the new MBB.
358 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
359 MachineBasicBlock::iterator BBI1) {
360 MachineFunction &MF = *CurMBB.getParent();
362 // Create the fall-through block.
363 MachineFunction::iterator MBBI = &CurMBB;
364 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
365 CurMBB.getParent()->insert(++MBBI, NewMBB);
367 // Move all the successors of this block to the specified block.
368 NewMBB->transferSuccessors(&CurMBB);
370 // Add an edge from CurMBB to NewMBB for the fall-through.
371 CurMBB.addSuccessor(NewMBB);
373 // Splice the code over.
374 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
376 // For targets that use the register scavenger, we must maintain LiveIns.
378 RS->enterBasicBlock(&CurMBB);
380 RS->forward(prior(CurMBB.end()));
381 BitVector RegsLiveAtExit(TRI->getNumRegs());
382 RS->getRegsUsed(RegsLiveAtExit, false);
383 for (unsigned int i=0, e=TRI->getNumRegs(); i!=e; i++)
384 if (RegsLiveAtExit[i])
385 NewMBB->addLiveIn(i);
391 /// EstimateRuntime - Make a rough estimate for how long it will take to run
392 /// the specified code.
393 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
394 MachineBasicBlock::iterator E) {
396 for (; I != E; ++I) {
397 const TargetInstrDesc &TID = I->getDesc();
400 else if (TID.mayLoad() || TID.mayStore())
408 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
409 // branches temporarily for tail merging). In the case where CurMBB ends
410 // with a conditional branch to the next block, optimize by reversing the
411 // test and conditionally branching to SuccMBB instead.
412 static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB,
413 const TargetInstrInfo *TII) {
414 MachineFunction *MF = CurMBB->getParent();
415 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
416 MachineBasicBlock *TBB = 0, *FBB = 0;
417 SmallVector<MachineOperand, 4> Cond;
418 if (I != MF->end() &&
419 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
420 MachineBasicBlock *NextBB = I;
421 if (TBB == NextBB && !Cond.empty() && !FBB) {
422 if (!TII->ReverseBranchCondition(Cond)) {
423 TII->RemoveBranch(*CurMBB);
424 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
429 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
433 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
434 if (getHash() < o.getHash())
436 else if (getHash() > o.getHash())
438 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
440 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
443 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
444 // an object with itself.
445 #ifndef _GLIBCXX_DEBUG
446 llvm_unreachable("Predecessor appears twice");
452 /// CountTerminators - Count the number of terminators in the given
453 /// block and set I to the position of the first non-terminator, if there
454 /// is one, or MBB->end() otherwise.
455 static unsigned CountTerminators(MachineBasicBlock *MBB,
456 MachineBasicBlock::iterator &I) {
458 unsigned NumTerms = 0;
460 if (I == MBB->begin()) {
465 if (!I->getDesc().isTerminator()) break;
471 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
472 /// and decide if it would be profitable to merge those tails. Return the
473 /// length of the common tail and iterators to the first common instruction
475 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
476 MachineBasicBlock *MBB2,
477 unsigned minCommonTailLength,
478 unsigned &CommonTailLen,
479 MachineBasicBlock::iterator &I1,
480 MachineBasicBlock::iterator &I2,
481 MachineBasicBlock *SuccBB,
482 MachineBasicBlock *PredBB) {
483 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
484 MachineFunction *MF = MBB1->getParent();
486 if (CommonTailLen == 0)
489 // It's almost always profitable to merge any number of non-terminator
490 // instructions with the block that falls through into the common successor.
491 if (MBB1 == PredBB || MBB2 == PredBB) {
492 MachineBasicBlock::iterator I;
493 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
494 if (CommonTailLen > NumTerms)
498 // If one of the blocks can be completely merged and happens to be in
499 // a position where the other could fall through into it, merge any number
500 // of instructions, because it can be done without a branch.
501 // TODO: If the blocks are not adjacent, move one of them so that they are?
502 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
504 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
507 // If both blocks have an unconditional branch temporarily stripped out,
508 // treat that as an additional common instruction.
509 if (MBB1 != PredBB && MBB2 != PredBB &&
510 !MBB1->back().getDesc().isBarrier() &&
511 !MBB2->back().getDesc().isBarrier())
512 --minCommonTailLength;
514 // Check if the common tail is long enough to be worthwhile.
515 if (CommonTailLen >= minCommonTailLength)
518 // If we are optimizing for code size, 1 instruction in common is enough if
519 // we don't have to split a block. At worst we will be replacing a
520 // fallthrough into the common tail with a branch, which at worst breaks
521 // even with falling through into the duplicated common tail.
522 if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
523 (I1 == MBB1->begin() || I2 == MBB2->begin()))
529 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
530 /// hash CurHash (guaranteed to match the last element). Build the vector
531 /// SameTails of all those that have the (same) largest number of instructions
532 /// in common of any pair of these blocks. SameTails entries contain an
533 /// iterator into MergePotentials (from which the MachineBasicBlock can be
534 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
535 /// instruction where the matching code sequence begins.
536 /// Order of elements in SameTails is the reverse of the order in which
537 /// those blocks appear in MergePotentials (where they are not necessarily
539 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
540 unsigned minCommonTailLength,
541 MachineBasicBlock *SuccBB,
542 MachineBasicBlock *PredBB) {
543 unsigned maxCommonTailLength = 0U;
545 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
546 MPIterator HighestMPIter = prior(MergePotentials.end());
547 for (MPIterator CurMPIter = prior(MergePotentials.end()),
548 B = MergePotentials.begin();
549 CurMPIter!=B && CurMPIter->getHash() == CurHash;
551 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
552 unsigned CommonTailLen;
553 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
555 CommonTailLen, TrialBBI1, TrialBBI2,
557 if (CommonTailLen > maxCommonTailLength) {
559 maxCommonTailLength = CommonTailLen;
560 HighestMPIter = CurMPIter;
561 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
563 if (HighestMPIter == CurMPIter &&
564 CommonTailLen == maxCommonTailLength)
565 SameTails.push_back(SameTailElt(I, TrialBBI2));
571 return maxCommonTailLength;
574 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
575 /// MergePotentials, restoring branches at ends of blocks as appropriate.
576 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
577 MachineBasicBlock* SuccBB,
578 MachineBasicBlock* PredBB) {
579 MPIterator CurMPIter, B;
580 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
581 CurMPIter->getHash() == CurHash;
583 // Put the unconditional branch back, if we need one.
584 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
585 if (SuccBB && CurMBB != PredBB)
586 FixTail(CurMBB, SuccBB, TII);
590 if (CurMPIter->getHash() != CurHash)
592 MergePotentials.erase(CurMPIter, MergePotentials.end());
595 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
596 /// only of the common tail. Create a block that does by splitting one.
597 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
598 unsigned maxCommonTailLength) {
599 unsigned i, commonTailIndex;
600 unsigned TimeEstimate = ~0U;
601 for (i=0, commonTailIndex=0; i<SameTails.size(); i++) {
602 // Use PredBB if possible; that doesn't require a new branch.
603 if (SameTails[i].getBlock() == PredBB) {
607 // Otherwise, make a (fairly bogus) choice based on estimate of
608 // how long it will take the various blocks to execute.
609 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
610 SameTails[i].getTailStartPos());
611 if (t <= TimeEstimate) {
617 MachineBasicBlock::iterator BBI =
618 SameTails[commonTailIndex].getTailStartPos();
619 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
621 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
622 << maxCommonTailLength);
624 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
625 SameTails[commonTailIndex].setBlock(newMBB);
626 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
628 // If we split PredBB, newMBB is the new predecessor.
632 return commonTailIndex;
635 // See if any of the blocks in MergePotentials (which all have a common single
636 // successor, or all have no successor) can be tail-merged. If there is a
637 // successor, any blocks in MergePotentials that are not tail-merged and
638 // are not immediately before Succ must have an unconditional branch to
639 // Succ added (but the predecessor/successor lists need no adjustment).
640 // The lone predecessor of Succ that falls through into Succ,
641 // if any, is given in PredBB.
643 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
644 MachineBasicBlock* PredBB) {
645 bool MadeChange = false;
647 // Except for the special cases below, tail-merge if there are at least
648 // this many instructions in common.
649 unsigned minCommonTailLength = TailMergeSize;
651 // If there's a successor block, there are some cases which don't require
652 // new branching and as such are very likely to be profitable.
654 if (SuccBB->pred_size() == MergePotentials.size() &&
655 !MergePotentials[0].getBlock()->empty()) {
656 // If all the predecessors have at least one tail instruction in common,
657 // merging is very likely to be a win since it won't require an increase
658 // in static branches, and it will decrease the static instruction count.
659 bool AllPredsMatch = true;
660 MachineBasicBlock::iterator FirstNonTerm;
661 unsigned MinNumTerms = CountTerminators(MergePotentials[0].getBlock(),
663 if (FirstNonTerm != MergePotentials[0].getBlock()->end()) {
664 for (unsigned i = 1, e = MergePotentials.size(); i != e; ++i) {
665 MachineBasicBlock::iterator OtherFirstNonTerm;
666 unsigned NumTerms = CountTerminators(MergePotentials[0].getBlock(),
668 if (NumTerms < MinNumTerms)
669 MinNumTerms = NumTerms;
670 if (OtherFirstNonTerm == MergePotentials[i].getBlock()->end() ||
671 OtherFirstNonTerm->isIdenticalTo(FirstNonTerm)) {
672 AllPredsMatch = false;
677 // If they all have an instruction in common, do any amount of merging.
679 minCommonTailLength = MinNumTerms + 1;
684 DEBUG(errs() << "\nTryTailMergeBlocks: ";
685 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
686 errs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
687 << (i == e-1 ? "" : ", ");
690 errs() << " with successor BB#" << SuccBB->getNumber() << '\n';
692 errs() << " which has fall-through from BB#"
693 << PredBB->getNumber() << "\n";
695 errs() << "Looking for common tails of at least "
696 << minCommonTailLength << " instruction"
697 << (minCommonTailLength == 1 ? "" : "s") << '\n';
700 // Sort by hash value so that blocks with identical end sequences sort
702 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
704 // Walk through equivalence sets looking for actual exact matches.
705 while (MergePotentials.size() > 1) {
706 unsigned CurHash = MergePotentials.back().getHash();
708 // Build SameTails, identifying the set of blocks with this hash code
709 // and with the maximum number of instructions in common.
710 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
714 // If we didn't find any pair that has at least minCommonTailLength
715 // instructions in common, remove all blocks with this hash code and retry.
716 if (SameTails.empty()) {
717 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
721 // If one of the blocks is the entire common tail (and not the entry
722 // block, which we can't jump to), we can treat all blocks with this same
723 // tail at once. Use PredBB if that is one of the possibilities, as that
724 // will not introduce any extra branches.
725 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
726 getParent()->begin();
727 unsigned commonTailIndex = SameTails.size();
728 // If there are two blocks, check to see if one can be made to fall through
730 if (SameTails.size() == 2 &&
731 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
732 SameTails[1].tailIsWholeBlock())
734 else if (SameTails.size() == 2 &&
735 SameTails[1].getBlock()->isLayoutSuccessor(
736 SameTails[0].getBlock()) &&
737 SameTails[0].tailIsWholeBlock())
740 // Otherwise just pick one, favoring the fall-through predecessor if
742 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
743 MachineBasicBlock *MBB = SameTails[i].getBlock();
744 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
750 if (SameTails[i].tailIsWholeBlock())
755 if (commonTailIndex == SameTails.size() ||
756 (SameTails[commonTailIndex].getBlock() == PredBB &&
757 !SameTails[commonTailIndex].tailIsWholeBlock())) {
758 // None of the blocks consist entirely of the common tail.
759 // Split a block so that one does.
760 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
763 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
764 // MBB is common tail. Adjust all other BB's to jump to this one.
765 // Traversal must be forwards so erases work.
766 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber()
768 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
769 if (commonTailIndex == i)
771 DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber()
772 << (i == e-1 ? "" : ", "));
773 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
774 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
775 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
776 MergePotentials.erase(SameTails[i].getMPIter());
778 DEBUG(errs() << "\n");
779 // We leave commonTailIndex in the worklist in case there are other blocks
780 // that match it with a smaller number of instructions.
786 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
788 if (!EnableTailMerge) return false;
790 bool MadeChange = false;
792 // First find blocks with no successors.
793 MergePotentials.clear();
794 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
796 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
799 // See if we can do any tail merging on those.
800 if (MergePotentials.size() < TailMergeThreshold &&
801 MergePotentials.size() >= 2)
802 MadeChange |= TryTailMergeBlocks(NULL, NULL);
804 // Look at blocks (IBB) with multiple predecessors (PBB).
805 // We change each predecessor to a canonical form, by
806 // (1) temporarily removing any unconditional branch from the predecessor
808 // (2) alter conditional branches so they branch to the other block
809 // not IBB; this may require adding back an unconditional branch to IBB
810 // later, where there wasn't one coming in. E.g.
812 // fallthrough to QBB
815 // with a conceptual B to IBB after that, which never actually exists.
816 // With those changes, we see whether the predecessors' tails match,
817 // and merge them if so. We change things out of canonical form and
818 // back to the way they were later in the process. (OptimizeBranches
819 // would undo some of this, but we can't use it, because we'd get into
820 // a compile-time infinite loop repeatedly doing and undoing the same
823 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end();
825 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
826 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
827 MachineBasicBlock *IBB = I;
828 MachineBasicBlock *PredBB = prior(I);
829 MergePotentials.clear();
830 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
833 MachineBasicBlock* PBB = *P;
834 // Skip blocks that loop to themselves, can't tail merge these.
837 // Visit each predecessor only once.
838 if (!UniquePreds.insert(PBB))
840 MachineBasicBlock *TBB = 0, *FBB = 0;
841 SmallVector<MachineOperand, 4> Cond;
842 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
843 // Failing case: IBB is the target of a cbr, and
844 // we cannot reverse the branch.
845 SmallVector<MachineOperand, 4> NewCond(Cond);
846 if (!Cond.empty() && TBB == IBB) {
847 if (TII->ReverseBranchCondition(NewCond))
849 // This is the QBB case described above
851 FBB = next(MachineFunction::iterator(PBB));
853 // Failing case: the only way IBB can be reached from PBB is via
854 // exception handling. Happens for landing pads. Would be nice
855 // to have a bit in the edge so we didn't have to do all this.
856 if (IBB->isLandingPad()) {
857 MachineFunction::iterator IP = PBB; IP++;
858 MachineBasicBlock* PredNextBB = NULL;
862 if (IBB!=PredNextBB) // fallthrough
865 if (TBB!=IBB && FBB!=IBB) // cbr then ubr
867 } else if (Cond.empty()) {
871 if (TBB!=IBB && IBB!=PredNextBB) // cbr
875 // Remove the unconditional branch at the end, if any.
876 if (TBB && (Cond.empty() || FBB)) {
877 TII->RemoveBranch(*PBB);
879 // reinsert conditional branch only, for now
880 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
882 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
886 if (MergePotentials.size() >= 2)
887 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
888 // Reinsert an unconditional branch if needed.
889 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
890 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
891 if (MergePotentials.size() == 1 &&
892 MergePotentials.begin()->getBlock() != PredBB)
893 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
899 //===----------------------------------------------------------------------===//
900 // Branch Optimization
901 //===----------------------------------------------------------------------===//
903 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
904 bool MadeChange = false;
906 // Make sure blocks are numbered in order
909 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
910 MachineBasicBlock *MBB = I++;
911 MadeChange |= OptimizeBlock(MBB);
913 // If it is dead, remove it.
914 if (MBB->pred_empty()) {
915 RemoveDeadBlock(MBB);
924 /// CanFallThrough - Return true if the specified block (with the specified
925 /// branch condition) can implicitly transfer control to the block after it by
926 /// falling off the end of it. This should return false if it can reach the
927 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
929 /// True is a conservative answer.
931 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
932 bool BranchUnAnalyzable,
933 MachineBasicBlock *TBB,
934 MachineBasicBlock *FBB,
935 const SmallVectorImpl<MachineOperand> &Cond) {
936 MachineFunction::iterator Fallthrough = CurBB;
938 // If FallthroughBlock is off the end of the function, it can't fall through.
939 if (Fallthrough == CurBB->getParent()->end())
942 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
943 if (!CurBB->isSuccessor(Fallthrough))
946 // If we couldn't analyze the branch, examine the last instruction.
947 // If the block doesn't end in a known control barrier, assume fallthrough
948 // is possible. The isPredicable check is needed because this code can be
949 // called during IfConversion, where an instruction which is normally a
950 // Barrier is predicated and thus no longer an actual control barrier. This
951 // is over-conservative though, because if an instruction isn't actually
952 // predicated we could still treat it like a barrier.
953 if (BranchUnAnalyzable)
954 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() ||
955 CurBB->back().getDesc().isPredicable();
957 // If there is no branch, control always falls through.
958 if (TBB == 0) return true;
960 // If there is some explicit branch to the fallthrough block, it can obviously
961 // reach, even though the branch should get folded to fall through implicitly.
962 if (MachineFunction::iterator(TBB) == Fallthrough ||
963 MachineFunction::iterator(FBB) == Fallthrough)
966 // If it's an unconditional branch to some block not the fall through, it
967 // doesn't fall through.
968 if (Cond.empty()) return false;
970 // Otherwise, if it is conditional and has no explicit false block, it falls
975 /// CanFallThrough - Return true if the specified can implicitly transfer
976 /// control to the block after it by falling off the end of it. This should
977 /// return false if it can reach the block after it, but it uses an explicit
978 /// branch to do so (e.g. a table jump).
980 /// True is a conservative answer.
982 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
983 MachineBasicBlock *TBB = 0, *FBB = 0;
984 SmallVector<MachineOperand, 4> Cond;
985 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true);
986 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
989 /// IsBetterFallthrough - Return true if it would be clearly better to
990 /// fall-through to MBB1 than to fall through into MBB2. This has to return
991 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
992 /// result in infinite loops.
993 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
994 MachineBasicBlock *MBB2) {
995 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
996 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
997 // optimize branches that branch to either a return block or an assert block
998 // into a fallthrough to the return.
999 if (MBB1->empty() || MBB2->empty()) return false;
1001 // If there is a clear successor ordering we make sure that one block
1002 // will fall through to the next
1003 if (MBB1->isSuccessor(MBB2)) return true;
1004 if (MBB2->isSuccessor(MBB1)) return false;
1006 MachineInstr *MBB1I = --MBB1->end();
1007 MachineInstr *MBB2I = --MBB2->end();
1008 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
1011 /// TailDuplicate - MBB unconditionally branches to SuccBB. If it is profitable,
1012 /// duplicate SuccBB's contents in MBB to eliminate the branch.
1013 bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB,
1014 bool PrevFallsThrough,
1015 MachineFunction &MF) {
1016 // Don't try to tail-duplicate single-block loops.
1017 if (TailBB->isSuccessor(TailBB))
1020 // Don't tail-duplicate a block which will soon be folded into its successor.
1021 if (TailBB->succ_size() == 1 &&
1022 TailBB->succ_begin()[0]->pred_size() == 1)
1025 // Duplicate up to one less that the tail-merge threshold, so that we don't
1026 // get into an infinite loop between duplicating and merging. When optimizing
1027 // for size, duplicate only one, because one branch instruction can be
1028 // eliminated to compensate for the duplication.
1029 unsigned MaxDuplicateCount =
1030 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ?
1031 1 : (TailMergeSize - 1);
1033 // Check the instructions in the block to determine whether tail-duplication
1034 // is invalid or unlikely to be unprofitable.
1036 bool HasCall = false;
1037 for (MachineBasicBlock::iterator I = TailBB->begin();
1038 I != TailBB->end(); ++I, ++i) {
1039 // Non-duplicable things shouldn't be tail-duplicated.
1040 if (I->getDesc().isNotDuplicable()) return false;
1041 // Don't duplicate more than the threshold.
1042 if (i == MaxDuplicateCount) return false;
1043 // Remember if we saw a call.
1044 if (I->getDesc().isCall()) HasCall = true;
1046 // Heuristically, don't tail-duplicate calls if it would expand code size,
1047 // as it's less likely to be worth the extra cost.
1048 if (i > 1 && HasCall)
1051 // Iterate through all the unique predecessors and tail-duplicate this
1052 // block into them, if possible. Copying the list ahead of time also
1053 // avoids trouble with the predecessor list reallocating.
1054 bool Changed = false;
1055 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
1056 TailBB->pred_end());
1057 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
1058 PE = Preds.end(); PI != PE; ++PI) {
1059 MachineBasicBlock *PredBB = *PI;
1061 assert(TailBB != PredBB &&
1062 "Single-block loop should have been rejected earlier!");
1063 if (PredBB->succ_size() > 1) continue;
1065 MachineBasicBlock *PredTBB, *PredFBB;
1066 SmallVector<MachineOperand, 4> PredCond;
1067 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
1069 if (!PredCond.empty())
1071 // EH edges are ignored by AnalyzeBranch.
1072 if (PredBB->succ_size() != 1)
1074 // Don't duplicate into a fall-through predecessor unless its the
1075 // only predecessor.
1076 if (PredBB->isLayoutSuccessor(TailBB) &&
1078 TailBB->pred_size() != 1)
1081 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB
1082 << "From Succ: " << *TailBB);
1084 // Remove PredBB's unconditional branch.
1085 TII->RemoveBranch(*PredBB);
1086 // Clone the contents of TailBB into PredBB.
1087 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end();
1089 MachineInstr *NewMI = MF.CloneMachineInstr(I);
1090 PredBB->insert(PredBB->end(), NewMI);
1094 PredBB->removeSuccessor(PredBB->succ_begin());
1095 assert(PredBB->succ_empty() &&
1096 "TailDuplicate called on block with multiple successors!");
1097 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
1098 E = TailBB->succ_end(); I != E; ++I)
1099 PredBB->addSuccessor(*I);
1107 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1108 /// block. This is never called on the entry block.
1109 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1110 bool MadeChange = false;
1111 MachineFunction &MF = *MBB->getParent();
1114 MachineFunction::iterator FallThrough = MBB;
1117 // If this block is empty, make everyone use its fall-through, not the block
1118 // explicitly. Landing pads should not do this since the landing-pad table
1119 // points to this block. Blocks with their addresses taken shouldn't be
1121 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1122 // Dead block? Leave for cleanup later.
1123 if (MBB->pred_empty()) return MadeChange;
1125 if (FallThrough == MF.end()) {
1126 // TODO: Simplify preds to not branch here if possible!
1128 // Rewrite all predecessors of the old block to go to the fallthrough
1130 while (!MBB->pred_empty()) {
1131 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1132 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1134 // If MBB was the target of a jump table, update jump tables to go to the
1135 // fallthrough instead.
1136 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough);
1142 // Check to see if we can simplify the terminator of the block before this
1144 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1146 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1147 SmallVector<MachineOperand, 4> PriorCond;
1148 bool PriorUnAnalyzable =
1149 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1150 if (!PriorUnAnalyzable) {
1151 // If the CFG for the prior block has extra edges, remove them.
1152 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1153 !PriorCond.empty());
1155 // If the previous branch is conditional and both conditions go to the same
1156 // destination, remove the branch, replacing it with an unconditional one or
1158 if (PriorTBB && PriorTBB == PriorFBB) {
1159 TII->RemoveBranch(PrevBB);
1161 if (PriorTBB != MBB)
1162 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1165 goto ReoptimizeBlock;
1168 // If the previous block unconditionally falls through to this block and
1169 // this block has no other predecessors, move the contents of this block
1170 // into the prior block. This doesn't usually happen when SimplifyCFG
1171 // has been used, but it can happen tail duplication eliminates all the
1172 // non-branch predecessors of a block leaving only the fall-through edge.
1173 // This has to check PrevBB->succ_size() because EH edges are ignored by
1175 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1176 PrevBB.succ_size() == 1 &&
1177 !MBB->hasAddressTaken()) {
1178 DEBUG(errs() << "\nMerging into block: " << PrevBB
1179 << "From MBB: " << *MBB);
1180 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1181 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1182 assert(PrevBB.succ_empty());
1183 PrevBB.transferSuccessors(MBB);
1188 // If the previous branch *only* branches to *this* block (conditional or
1189 // not) remove the branch.
1190 if (PriorTBB == MBB && PriorFBB == 0) {
1191 TII->RemoveBranch(PrevBB);
1194 goto ReoptimizeBlock;
1197 // If the prior block branches somewhere else on the condition and here if
1198 // the condition is false, remove the uncond second branch.
1199 if (PriorFBB == MBB) {
1200 TII->RemoveBranch(PrevBB);
1201 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1204 goto ReoptimizeBlock;
1207 // If the prior block branches here on true and somewhere else on false, and
1208 // if the branch condition is reversible, reverse the branch to create a
1210 if (PriorTBB == MBB) {
1211 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1212 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1213 TII->RemoveBranch(PrevBB);
1214 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
1217 goto ReoptimizeBlock;
1221 // If this block has no successors (e.g. it is a return block or ends with
1222 // a call to a no-return function like abort or __cxa_throw) and if the pred
1223 // falls through into this block, and if it would otherwise fall through
1224 // into the block after this, move this block to the end of the function.
1226 // We consider it more likely that execution will stay in the function (e.g.
1227 // due to loops) than it is to exit it. This asserts in loops etc, moving
1228 // the assert condition out of the loop body.
1229 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1230 MachineFunction::iterator(PriorTBB) == FallThrough &&
1231 !CanFallThrough(MBB)) {
1232 bool DoTransform = true;
1234 // We have to be careful that the succs of PredBB aren't both no-successor
1235 // blocks. If neither have successors and if PredBB is the second from
1236 // last block in the function, we'd just keep swapping the two blocks for
1237 // last. Only do the swap if one is clearly better to fall through than
1239 if (FallThrough == --MF.end() &&
1240 !IsBetterFallthrough(PriorTBB, MBB))
1241 DoTransform = false;
1243 // We don't want to do this transformation if we have control flow like:
1252 // In this case, we could actually be moving the return block *into* a
1254 if (DoTransform && !MBB->succ_empty() &&
1255 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
1256 DoTransform = false;
1260 // Reverse the branch so we will fall through on the previous true cond.
1261 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1262 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1263 DEBUG(errs() << "\nMoving MBB: " << *MBB
1264 << "To make fallthrough to: " << *PriorTBB << "\n");
1266 TII->RemoveBranch(PrevBB);
1267 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1269 // Move this block to the end of the function.
1270 MBB->moveAfter(--MF.end());
1279 // Analyze the branch in the current block.
1280 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1281 SmallVector<MachineOperand, 4> CurCond;
1282 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1283 if (!CurUnAnalyzable) {
1284 // If the CFG for the prior block has extra edges, remove them.
1285 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1287 // If this is a two-way branch, and the FBB branches to this block, reverse
1288 // the condition so the single-basic-block loop is faster. Instead of:
1289 // Loop: xxx; jcc Out; jmp Loop
1291 // Loop: xxx; jncc Loop; jmp Out
1292 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1293 SmallVector<MachineOperand, 4> NewCond(CurCond);
1294 if (!TII->ReverseBranchCondition(NewCond)) {
1295 TII->RemoveBranch(*MBB);
1296 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1299 goto ReoptimizeBlock;
1304 // If this branch is the only thing in its block, see if we can forward
1305 // other blocks across it.
1306 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1307 MBB->begin()->getDesc().isBranch() && CurTBB != MBB &&
1308 !MBB->hasAddressTaken()) {
1309 // This block may contain just an unconditional branch. Because there can
1310 // be 'non-branch terminators' in the block, try removing the branch and
1311 // then seeing if the block is empty.
1312 TII->RemoveBranch(*MBB);
1314 // If this block is just an unconditional branch to CurTBB, we can
1315 // usually completely eliminate the block. The only case we cannot
1316 // completely eliminate the block is when the block before this one
1317 // falls through into MBB and we can't understand the prior block's branch
1320 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
1321 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1322 !PrevBB.isSuccessor(MBB)) {
1323 // If the prior block falls through into us, turn it into an
1324 // explicit branch to us to make updates simpler.
1325 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1326 PriorTBB != MBB && PriorFBB != MBB) {
1327 if (PriorTBB == 0) {
1328 assert(PriorCond.empty() && PriorFBB == 0 &&
1329 "Bad branch analysis");
1332 assert(PriorFBB == 0 && "Machine CFG out of date!");
1335 TII->RemoveBranch(PrevBB);
1336 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1339 // Iterate through all the predecessors, revectoring each in-turn.
1341 bool DidChange = false;
1342 bool HasBranchToSelf = false;
1343 while(PI != MBB->pred_size()) {
1344 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1346 // If this block has an uncond branch to itself, leave it.
1348 HasBranchToSelf = true;
1351 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1352 // If this change resulted in PMBB ending in a conditional
1353 // branch where both conditions go to the same destination,
1354 // change this to an unconditional branch (and fix the CFG).
1355 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1356 SmallVector<MachineOperand, 4> NewCurCond;
1357 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1358 NewCurFBB, NewCurCond, true);
1359 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1360 TII->RemoveBranch(*PMBB);
1362 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1365 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1370 // Change any jumptables to go to the new MBB.
1371 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB);
1375 if (!HasBranchToSelf) return MadeChange;
1380 // Add the branch back if the block is more than just an uncond branch.
1381 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1385 // Now we know that there was no fall-through into this block, check to
1386 // see if it has a fall-through into its successor.
1387 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1389 bool PrevFallsThru = CanFallThrough(&PrevBB, PriorUnAnalyzable,
1390 PriorTBB, PriorFBB, PriorCond);
1392 // If this block is small, unconditionally branched to, and does not
1393 // fall through, tail-duplicate its instructions into its predecessors
1394 // to eliminate a (dynamic) branch.
1396 if (TailDuplicate(MBB, PrevFallsThru, MF)) {
1401 // If the prior block doesn't fall through into this block, and if this
1402 // block doesn't fall through into some other block, see if we can find a
1403 // place to move this block where a fall-through will happen.
1404 if (!PrevFallsThru) {
1405 if (!MBB->isLandingPad()) {
1406 // Check all the predecessors of this block. If one of them has no fall
1407 // throughs, move this block right after it.
1408 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1409 E = MBB->pred_end(); PI != E; ++PI) {
1410 // Analyze the branch at the end of the pred.
1411 MachineBasicBlock *PredBB = *PI;
1412 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1413 MachineBasicBlock *PredTBB, *PredFBB;
1414 SmallVector<MachineOperand, 4> PredCond;
1415 if (PredBB != MBB && !CanFallThrough(PredBB) &&
1416 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1417 && (!CurFallsThru || !CurTBB || !CurFBB)
1418 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1419 // If the current block doesn't fall through, just move it.
1420 // If the current block can fall through and does not end with a
1421 // conditional branch, we need to append an unconditional jump to
1422 // the (current) next block. To avoid a possible compile-time
1423 // infinite loop, move blocks only backward in this case.
1424 // Also, if there are already 2 branches here, we cannot add a third;
1425 // this means we have the case
1430 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1432 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1434 MBB->moveAfter(PredBB);
1436 goto ReoptimizeBlock;
1441 if (!CurFallsThru) {
1442 // Check all successors to see if we can move this block before it.
1443 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1444 E = MBB->succ_end(); SI != E; ++SI) {
1445 // Analyze the branch at the end of the block before the succ.
1446 MachineBasicBlock *SuccBB = *SI;
1447 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1449 // If this block doesn't already fall-through to that successor, and if
1450 // the succ doesn't already have a block that can fall through into it,
1451 // and if the successor isn't an EH destination, we can arrange for the
1452 // fallthrough to happen.
1453 if (SuccBB != MBB && &*SuccPrev != MBB &&
1454 !CanFallThrough(SuccPrev) && !CurUnAnalyzable &&
1455 !SuccBB->isLandingPad()) {
1456 MBB->moveBefore(SuccBB);
1458 goto ReoptimizeBlock;
1462 // Okay, there is no really great place to put this block. If, however,
1463 // the block before this one would be a fall-through if this block were
1464 // removed, move this block to the end of the function.
1465 MachineBasicBlock *PrevTBB, *PrevFBB;
1466 SmallVector<MachineOperand, 4> PrevCond;
1467 if (FallThrough != MF.end() &&
1468 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1469 PrevBB.isSuccessor(FallThrough)) {
1470 MBB->moveAfter(--MF.end());