1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass forwards branches to unconditional branches to make them branch
11 // directly to the target block. This pass often results in dead MBB's, which
14 // Note that this pass must be run after register allocation, it cannot handle
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "branchfolding"
20 #include "BranchFolding.h"
21 #include "llvm/Function.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/RegisterScavenging.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Target/TargetRegisterInfo.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/ADT/SmallSet.h"
35 #include "llvm/ADT/SetVector.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/ADT/STLExtras.h"
41 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
42 STATISTIC(NumBranchOpts, "Number of branches optimized");
43 STATISTIC(NumTailMerge , "Number of block tails merged");
44 STATISTIC(NumTailDups , "Number of tail duplicated blocks");
45 STATISTIC(NumInstrDups , "Additional instructions due to tail duplication");
47 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
48 cl::init(cl::BOU_UNSET), cl::Hidden);
50 // Throttle for huge numbers of predecessors (compile speed problems)
51 static cl::opt<unsigned>
52 TailMergeThreshold("tail-merge-threshold",
53 cl::desc("Max number of predecessors to consider tail merging"),
54 cl::init(150), cl::Hidden);
56 // Heuristic for tail merging (and, inversely, tail duplication).
57 // TODO: This should be replaced with a target query.
58 static cl::opt<unsigned>
59 TailMergeSize("tail-merge-size",
60 cl::desc("Min number of instructions to consider tail merging"),
61 cl::init(3), cl::Hidden);
64 /// BranchFolderPass - Wrap branch folder in a machine function pass.
65 class BranchFolderPass : public MachineFunctionPass,
69 explicit BranchFolderPass(bool defaultEnableTailMerge)
70 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {}
72 virtual bool runOnMachineFunction(MachineFunction &MF);
73 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
77 char BranchFolderPass::ID = 0;
79 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
80 return new BranchFolderPass(DefaultEnableTailMerge);
83 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
84 return OptimizeFunction(MF,
85 MF.getTarget().getInstrInfo(),
86 MF.getTarget().getRegisterInfo(),
87 getAnalysisIfAvailable<MachineModuleInfo>());
91 BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
92 switch (FlagEnableTailMerge) {
93 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
94 case cl::BOU_TRUE: EnableTailMerge = true; break;
95 case cl::BOU_FALSE: EnableTailMerge = false; break;
99 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
100 /// function, updating the CFG.
101 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
102 assert(MBB->pred_empty() && "MBB must be dead!");
103 DEBUG(errs() << "\nRemoving MBB: " << *MBB);
105 MachineFunction *MF = MBB->getParent();
106 // drop all successors.
107 while (!MBB->succ_empty())
108 MBB->removeSuccessor(MBB->succ_end()-1);
110 // If there are any labels in the basic block, unregister them from
111 // MachineModuleInfo.
112 if (MMI && !MBB->empty()) {
113 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
116 // The label ID # is always operand #0, an immediate.
117 MMI->InvalidateLabel(I->getOperand(0).getImm());
125 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
126 /// followed by terminators, and if the implicitly defined registers are not
127 /// used by the terminators, remove those implicit_def's. e.g.
129 /// r0 = implicit_def
130 /// r1 = implicit_def
132 /// This block can be optimized away later if the implicit instructions are
134 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
135 SmallSet<unsigned, 4> ImpDefRegs;
136 MachineBasicBlock::iterator I = MBB->begin();
137 while (I != MBB->end()) {
138 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF)
140 unsigned Reg = I->getOperand(0).getReg();
141 ImpDefRegs.insert(Reg);
142 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
143 unsigned SubReg = *SubRegs; ++SubRegs)
144 ImpDefRegs.insert(SubReg);
147 if (ImpDefRegs.empty())
150 MachineBasicBlock::iterator FirstTerm = I;
151 while (I != MBB->end()) {
152 if (!TII->isUnpredicatedTerminator(I))
154 // See if it uses any of the implicitly defined registers.
155 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
156 MachineOperand &MO = I->getOperand(i);
157 if (!MO.isReg() || !MO.isUse())
159 unsigned Reg = MO.getReg();
160 if (ImpDefRegs.count(Reg))
167 while (I != FirstTerm) {
168 MachineInstr *ImpDefMI = &*I;
170 MBB->erase(ImpDefMI);
176 /// OptimizeFunction - Perhaps branch folding, tail merging and other
177 /// CFG optimizations on the given function.
178 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
179 const TargetInstrInfo *tii,
180 const TargetRegisterInfo *tri,
181 MachineModuleInfo *mmi) {
182 if (!tii) return false;
188 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
190 // Fix CFG. The later algorithms expect it to be right.
191 bool MadeChange = false;
192 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
193 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
194 SmallVector<MachineOperand, 4> Cond;
195 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
196 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
197 MadeChange |= OptimizeImpDefsBlock(MBB);
200 bool MadeChangeThisIteration = true;
201 while (MadeChangeThisIteration) {
202 MadeChangeThisIteration = false;
203 MadeChangeThisIteration |= TailMergeBlocks(MF);
204 MadeChangeThisIteration |= OptimizeBranches(MF);
205 MadeChange |= MadeChangeThisIteration;
208 // Do tail duplication after tail merging is done. Otherwise it is
209 // tough to avoid situations where tail duplication and tail merging undo
210 // each other's transformations ad infinitum.
211 MadeChangeThisIteration = true;
212 while (MadeChangeThisIteration) {
213 MadeChangeThisIteration = false;
214 MadeChangeThisIteration |= TailDuplicateBlocks(MF);
215 MadeChange |= MadeChangeThisIteration;
218 // See if any jump tables have become mergable or dead as the code generator
220 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
221 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
223 // Figure out how these jump tables should be merged.
224 std::vector<unsigned> JTMapping;
225 JTMapping.reserve(JTs.size());
227 // We always keep the 0th jump table.
228 JTMapping.push_back(0);
230 // Scan the jump tables, seeing if there are any duplicates. Note that this
231 // is N^2, which should be fixed someday.
232 for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
233 if (JTs[i].MBBs.empty())
234 JTMapping.push_back(i);
236 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
239 // If a jump table was merge with another one, walk the function rewriting
240 // references to jump tables to reference the new JT ID's. Keep track of
241 // whether we see a jump table idx, if not, we can delete the JT.
242 BitVector JTIsLive(JTs.size());
243 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
245 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
247 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
248 MachineOperand &Op = I->getOperand(op);
249 if (!Op.isJTI()) continue;
250 unsigned NewIdx = JTMapping[Op.getIndex()];
253 // Remember that this JT is live.
254 JTIsLive.set(NewIdx);
258 // Finally, remove dead jump tables. This happens either because the
259 // indirect jump was unreachable (and thus deleted) or because the jump
260 // table was merged with some other one.
261 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
262 if (!JTIsLive.test(i)) {
263 JTI->RemoveJumpTable(i);
272 //===----------------------------------------------------------------------===//
273 // Tail Merging of Blocks
274 //===----------------------------------------------------------------------===//
276 /// HashMachineInstr - Compute a hash value for MI and its operands.
277 static unsigned HashMachineInstr(const MachineInstr *MI) {
278 unsigned Hash = MI->getOpcode();
279 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
280 const MachineOperand &Op = MI->getOperand(i);
282 // Merge in bits from the operand if easy.
283 unsigned OperandHash = 0;
284 switch (Op.getType()) {
285 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
286 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
287 case MachineOperand::MO_MachineBasicBlock:
288 OperandHash = Op.getMBB()->getNumber();
290 case MachineOperand::MO_FrameIndex:
291 case MachineOperand::MO_ConstantPoolIndex:
292 case MachineOperand::MO_JumpTableIndex:
293 OperandHash = Op.getIndex();
295 case MachineOperand::MO_GlobalAddress:
296 case MachineOperand::MO_ExternalSymbol:
297 // Global address / external symbol are too hard, don't bother, but do
298 // pull in the offset.
299 OperandHash = Op.getOffset();
304 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
309 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
310 /// with no successors, we hash two instructions, because cross-jumping
311 /// only saves code when at least two instructions are removed (since a
312 /// branch must be inserted). For blocks with a successor, one of the
313 /// two blocks to be tail-merged will end with a branch already, so
314 /// it gains to cross-jump even for one instruction.
315 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
316 unsigned minCommonTailLength) {
317 MachineBasicBlock::const_iterator I = MBB->end();
318 if (I == MBB->begin())
319 return 0; // Empty MBB.
322 unsigned Hash = HashMachineInstr(I);
324 if (I == MBB->begin() || minCommonTailLength == 1)
325 return Hash; // Single instr MBB.
328 // Hash in the second-to-last instruction.
329 Hash ^= HashMachineInstr(I) << 2;
333 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
334 /// of instructions they actually have in common together at their end. Return
335 /// iterators for the first shared instruction in each block.
336 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
337 MachineBasicBlock *MBB2,
338 MachineBasicBlock::iterator &I1,
339 MachineBasicBlock::iterator &I2) {
343 unsigned TailLen = 0;
344 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
346 if (!I1->isIdenticalTo(I2) ||
347 // FIXME: This check is dubious. It's used to get around a problem where
348 // people incorrectly expect inline asm directives to remain in the same
349 // relative order. This is untenable because normal compiler
350 // optimizations (like this one) may reorder and/or merge these
352 I1->getOpcode() == TargetInstrInfo::INLINEASM) {
361 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
362 /// after it, replacing it with an unconditional branch to NewDest. This
363 /// returns true if OldInst's block is modified, false if NewDest is modified.
364 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
365 MachineBasicBlock *NewDest) {
366 MachineBasicBlock *OldBB = OldInst->getParent();
368 // Remove all the old successors of OldBB from the CFG.
369 while (!OldBB->succ_empty())
370 OldBB->removeSuccessor(OldBB->succ_begin());
372 // Remove all the dead instructions from the end of OldBB.
373 OldBB->erase(OldInst, OldBB->end());
375 // If OldBB isn't immediately before OldBB, insert a branch to it.
376 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
377 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
378 OldBB->addSuccessor(NewDest);
382 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
383 /// MBB so that the part before the iterator falls into the part starting at the
384 /// iterator. This returns the new MBB.
385 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
386 MachineBasicBlock::iterator BBI1) {
387 MachineFunction &MF = *CurMBB.getParent();
389 // Create the fall-through block.
390 MachineFunction::iterator MBBI = &CurMBB;
391 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
392 CurMBB.getParent()->insert(++MBBI, NewMBB);
394 // Move all the successors of this block to the specified block.
395 NewMBB->transferSuccessors(&CurMBB);
397 // Add an edge from CurMBB to NewMBB for the fall-through.
398 CurMBB.addSuccessor(NewMBB);
400 // Splice the code over.
401 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
403 // For targets that use the register scavenger, we must maintain LiveIns.
405 RS->enterBasicBlock(&CurMBB);
407 RS->forward(prior(CurMBB.end()));
408 BitVector RegsLiveAtExit(TRI->getNumRegs());
409 RS->getRegsUsed(RegsLiveAtExit, false);
410 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
411 if (RegsLiveAtExit[i])
412 NewMBB->addLiveIn(i);
418 /// EstimateRuntime - Make a rough estimate for how long it will take to run
419 /// the specified code.
420 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
421 MachineBasicBlock::iterator E) {
423 for (; I != E; ++I) {
424 const TargetInstrDesc &TID = I->getDesc();
427 else if (TID.mayLoad() || TID.mayStore())
435 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
436 // branches temporarily for tail merging). In the case where CurMBB ends
437 // with a conditional branch to the next block, optimize by reversing the
438 // test and conditionally branching to SuccMBB instead.
439 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
440 const TargetInstrInfo *TII) {
441 MachineFunction *MF = CurMBB->getParent();
442 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
443 MachineBasicBlock *TBB = 0, *FBB = 0;
444 SmallVector<MachineOperand, 4> Cond;
445 if (I != MF->end() &&
446 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
447 MachineBasicBlock *NextBB = I;
448 if (TBB == NextBB && !Cond.empty() && !FBB) {
449 if (!TII->ReverseBranchCondition(Cond)) {
450 TII->RemoveBranch(*CurMBB);
451 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
456 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
460 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
461 if (getHash() < o.getHash())
463 else if (getHash() > o.getHash())
465 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
467 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
470 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
471 // an object with itself.
472 #ifndef _GLIBCXX_DEBUG
473 llvm_unreachable("Predecessor appears twice");
479 /// CountTerminators - Count the number of terminators in the given
480 /// block and set I to the position of the first non-terminator, if there
481 /// is one, or MBB->end() otherwise.
482 static unsigned CountTerminators(MachineBasicBlock *MBB,
483 MachineBasicBlock::iterator &I) {
485 unsigned NumTerms = 0;
487 if (I == MBB->begin()) {
492 if (!I->getDesc().isTerminator()) break;
498 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
499 /// and decide if it would be profitable to merge those tails. Return the
500 /// length of the common tail and iterators to the first common instruction
502 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
503 MachineBasicBlock *MBB2,
504 unsigned minCommonTailLength,
505 unsigned &CommonTailLen,
506 MachineBasicBlock::iterator &I1,
507 MachineBasicBlock::iterator &I2,
508 MachineBasicBlock *SuccBB,
509 MachineBasicBlock *PredBB) {
510 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
511 MachineFunction *MF = MBB1->getParent();
513 if (CommonTailLen == 0)
516 // It's almost always profitable to merge any number of non-terminator
517 // instructions with the block that falls through into the common successor.
518 if (MBB1 == PredBB || MBB2 == PredBB) {
519 MachineBasicBlock::iterator I;
520 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
521 if (CommonTailLen > NumTerms)
525 // If one of the blocks can be completely merged and happens to be in
526 // a position where the other could fall through into it, merge any number
527 // of instructions, because it can be done without a branch.
528 // TODO: If the blocks are not adjacent, move one of them so that they are?
529 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
531 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
534 // If both blocks have an unconditional branch temporarily stripped out,
535 // count that as an additional common instruction for the following
537 unsigned EffectiveTailLen = CommonTailLen;
538 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
539 !MBB1->back().getDesc().isBarrier() &&
540 !MBB2->back().getDesc().isBarrier())
543 // Check if the common tail is long enough to be worthwhile.
544 if (EffectiveTailLen >= minCommonTailLength)
547 // If we are optimizing for code size, 2 instructions in common is enough if
548 // we don't have to split a block. At worst we will be introducing 1 new
549 // branch instruction, which is likely to be smaller than the 2
550 // instructions that would be deleted in the merge.
551 if (EffectiveTailLen >= 2 &&
552 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
553 (I1 == MBB1->begin() || I2 == MBB2->begin()))
559 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
560 /// hash CurHash (guaranteed to match the last element). Build the vector
561 /// SameTails of all those that have the (same) largest number of instructions
562 /// in common of any pair of these blocks. SameTails entries contain an
563 /// iterator into MergePotentials (from which the MachineBasicBlock can be
564 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
565 /// instruction where the matching code sequence begins.
566 /// Order of elements in SameTails is the reverse of the order in which
567 /// those blocks appear in MergePotentials (where they are not necessarily
569 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
570 unsigned minCommonTailLength,
571 MachineBasicBlock *SuccBB,
572 MachineBasicBlock *PredBB) {
573 unsigned maxCommonTailLength = 0U;
575 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
576 MPIterator HighestMPIter = prior(MergePotentials.end());
577 for (MPIterator CurMPIter = prior(MergePotentials.end()),
578 B = MergePotentials.begin();
579 CurMPIter != B && CurMPIter->getHash() == CurHash;
581 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
582 unsigned CommonTailLen;
583 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
585 CommonTailLen, TrialBBI1, TrialBBI2,
587 if (CommonTailLen > maxCommonTailLength) {
589 maxCommonTailLength = CommonTailLen;
590 HighestMPIter = CurMPIter;
591 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
593 if (HighestMPIter == CurMPIter &&
594 CommonTailLen == maxCommonTailLength)
595 SameTails.push_back(SameTailElt(I, TrialBBI2));
601 return maxCommonTailLength;
604 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
605 /// MergePotentials, restoring branches at ends of blocks as appropriate.
606 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
607 MachineBasicBlock *SuccBB,
608 MachineBasicBlock *PredBB) {
609 MPIterator CurMPIter, B;
610 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
611 CurMPIter->getHash() == CurHash;
613 // Put the unconditional branch back, if we need one.
614 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
615 if (SuccBB && CurMBB != PredBB)
616 FixTail(CurMBB, SuccBB, TII);
620 if (CurMPIter->getHash() != CurHash)
622 MergePotentials.erase(CurMPIter, MergePotentials.end());
625 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
626 /// only of the common tail. Create a block that does by splitting one.
627 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
628 unsigned maxCommonTailLength) {
629 unsigned commonTailIndex = 0;
630 unsigned TimeEstimate = ~0U;
631 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
632 // Use PredBB if possible; that doesn't require a new branch.
633 if (SameTails[i].getBlock() == PredBB) {
637 // Otherwise, make a (fairly bogus) choice based on estimate of
638 // how long it will take the various blocks to execute.
639 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
640 SameTails[i].getTailStartPos());
641 if (t <= TimeEstimate) {
647 MachineBasicBlock::iterator BBI =
648 SameTails[commonTailIndex].getTailStartPos();
649 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
651 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
652 << maxCommonTailLength);
654 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
655 SameTails[commonTailIndex].setBlock(newMBB);
656 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
658 // If we split PredBB, newMBB is the new predecessor.
662 return commonTailIndex;
665 // See if any of the blocks in MergePotentials (which all have a common single
666 // successor, or all have no successor) can be tail-merged. If there is a
667 // successor, any blocks in MergePotentials that are not tail-merged and
668 // are not immediately before Succ must have an unconditional branch to
669 // Succ added (but the predecessor/successor lists need no adjustment).
670 // The lone predecessor of Succ that falls through into Succ,
671 // if any, is given in PredBB.
673 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
674 MachineBasicBlock *PredBB) {
675 bool MadeChange = false;
677 // Except for the special cases below, tail-merge if there are at least
678 // this many instructions in common.
679 unsigned minCommonTailLength = TailMergeSize;
681 DEBUG(errs() << "\nTryTailMergeBlocks: ";
682 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
683 errs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
684 << (i == e-1 ? "" : ", ");
687 errs() << " with successor BB#" << SuccBB->getNumber() << '\n';
689 errs() << " which has fall-through from BB#"
690 << PredBB->getNumber() << "\n";
692 errs() << "Looking for common tails of at least "
693 << minCommonTailLength << " instruction"
694 << (minCommonTailLength == 1 ? "" : "s") << '\n';
697 // Sort by hash value so that blocks with identical end sequences sort
699 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
701 // Walk through equivalence sets looking for actual exact matches.
702 while (MergePotentials.size() > 1) {
703 unsigned CurHash = MergePotentials.back().getHash();
705 // Build SameTails, identifying the set of blocks with this hash code
706 // and with the maximum number of instructions in common.
707 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
711 // If we didn't find any pair that has at least minCommonTailLength
712 // instructions in common, remove all blocks with this hash code and retry.
713 if (SameTails.empty()) {
714 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
718 // If one of the blocks is the entire common tail (and not the entry
719 // block, which we can't jump to), we can treat all blocks with this same
720 // tail at once. Use PredBB if that is one of the possibilities, as that
721 // will not introduce any extra branches.
722 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
723 getParent()->begin();
724 unsigned commonTailIndex = SameTails.size();
725 // If there are two blocks, check to see if one can be made to fall through
727 if (SameTails.size() == 2 &&
728 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
729 SameTails[1].tailIsWholeBlock())
731 else if (SameTails.size() == 2 &&
732 SameTails[1].getBlock()->isLayoutSuccessor(
733 SameTails[0].getBlock()) &&
734 SameTails[0].tailIsWholeBlock())
737 // Otherwise just pick one, favoring the fall-through predecessor if
739 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
740 MachineBasicBlock *MBB = SameTails[i].getBlock();
741 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
747 if (SameTails[i].tailIsWholeBlock())
752 if (commonTailIndex == SameTails.size() ||
753 (SameTails[commonTailIndex].getBlock() == PredBB &&
754 !SameTails[commonTailIndex].tailIsWholeBlock())) {
755 // None of the blocks consist entirely of the common tail.
756 // Split a block so that one does.
757 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
760 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
761 // MBB is common tail. Adjust all other BB's to jump to this one.
762 // Traversal must be forwards so erases work.
763 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber()
765 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
766 if (commonTailIndex == i)
768 DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber()
769 << (i == e-1 ? "" : ", "));
770 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
771 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
772 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
773 MergePotentials.erase(SameTails[i].getMPIter());
775 DEBUG(errs() << "\n");
776 // We leave commonTailIndex in the worklist in case there are other blocks
777 // that match it with a smaller number of instructions.
783 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
785 if (!EnableTailMerge) return false;
787 bool MadeChange = false;
789 // First find blocks with no successors.
790 MergePotentials.clear();
791 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
793 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
796 // See if we can do any tail merging on those.
797 if (MergePotentials.size() < TailMergeThreshold &&
798 MergePotentials.size() >= 2)
799 MadeChange |= TryTailMergeBlocks(NULL, NULL);
801 // Look at blocks (IBB) with multiple predecessors (PBB).
802 // We change each predecessor to a canonical form, by
803 // (1) temporarily removing any unconditional branch from the predecessor
805 // (2) alter conditional branches so they branch to the other block
806 // not IBB; this may require adding back an unconditional branch to IBB
807 // later, where there wasn't one coming in. E.g.
809 // fallthrough to QBB
812 // with a conceptual B to IBB after that, which never actually exists.
813 // With those changes, we see whether the predecessors' tails match,
814 // and merge them if so. We change things out of canonical form and
815 // back to the way they were later in the process. (OptimizeBranches
816 // would undo some of this, but we can't use it, because we'd get into
817 // a compile-time infinite loop repeatedly doing and undoing the same
820 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end();
822 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
823 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
824 MachineBasicBlock *IBB = I;
825 MachineBasicBlock *PredBB = prior(I);
826 MergePotentials.clear();
827 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
830 MachineBasicBlock *PBB = *P;
831 // Skip blocks that loop to themselves, can't tail merge these.
834 // Visit each predecessor only once.
835 if (!UniquePreds.insert(PBB))
837 MachineBasicBlock *TBB = 0, *FBB = 0;
838 SmallVector<MachineOperand, 4> Cond;
839 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
840 // Failing case: IBB is the target of a cbr, and
841 // we cannot reverse the branch.
842 SmallVector<MachineOperand, 4> NewCond(Cond);
843 if (!Cond.empty() && TBB == IBB) {
844 if (TII->ReverseBranchCondition(NewCond))
846 // This is the QBB case described above
848 FBB = next(MachineFunction::iterator(PBB));
850 // Failing case: the only way IBB can be reached from PBB is via
851 // exception handling. Happens for landing pads. Would be nice
852 // to have a bit in the edge so we didn't have to do all this.
853 if (IBB->isLandingPad()) {
854 MachineFunction::iterator IP = PBB; IP++;
855 MachineBasicBlock *PredNextBB = NULL;
859 if (IBB != PredNextBB) // fallthrough
862 if (TBB != IBB && FBB != IBB) // cbr then ubr
864 } else if (Cond.empty()) {
865 if (TBB != IBB) // ubr
868 if (TBB != IBB && IBB != PredNextBB) // cbr
872 // Remove the unconditional branch at the end, if any.
873 if (TBB && (Cond.empty() || FBB)) {
874 TII->RemoveBranch(*PBB);
876 // reinsert conditional branch only, for now
877 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
879 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
883 if (MergePotentials.size() >= 2)
884 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
885 // Reinsert an unconditional branch if needed.
886 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
887 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
888 if (MergePotentials.size() == 1 &&
889 MergePotentials.begin()->getBlock() != PredBB)
890 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
896 //===----------------------------------------------------------------------===//
897 // Branch Optimization
898 //===----------------------------------------------------------------------===//
900 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
901 bool MadeChange = false;
903 // Make sure blocks are numbered in order
906 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
907 MachineBasicBlock *MBB = I++;
908 MadeChange |= OptimizeBlock(MBB);
910 // If it is dead, remove it.
911 if (MBB->pred_empty()) {
912 RemoveDeadBlock(MBB);
921 /// CanFallThrough - Return true if the specified block (with the specified
922 /// branch condition) can implicitly transfer control to the block after it by
923 /// falling off the end of it. This should return false if it can reach the
924 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
926 /// True is a conservative answer.
928 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
929 bool BranchUnAnalyzable,
930 MachineBasicBlock *TBB,
931 MachineBasicBlock *FBB,
932 const SmallVectorImpl<MachineOperand> &Cond) {
933 MachineFunction::iterator Fallthrough = CurBB;
935 // If FallthroughBlock is off the end of the function, it can't fall through.
936 if (Fallthrough == CurBB->getParent()->end())
939 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
940 if (!CurBB->isSuccessor(Fallthrough))
943 // If we couldn't analyze the branch, examine the last instruction.
944 // If the block doesn't end in a known control barrier, assume fallthrough
945 // is possible. The isPredicable check is needed because this code can be
946 // called during IfConversion, where an instruction which is normally a
947 // Barrier is predicated and thus no longer an actual control barrier. This
948 // is over-conservative though, because if an instruction isn't actually
949 // predicated we could still treat it like a barrier.
950 if (BranchUnAnalyzable)
951 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() ||
952 CurBB->back().getDesc().isPredicable();
954 // If there is no branch, control always falls through.
955 if (TBB == 0) return true;
957 // If there is some explicit branch to the fallthrough block, it can obviously
958 // reach, even though the branch should get folded to fall through implicitly.
959 if (MachineFunction::iterator(TBB) == Fallthrough ||
960 MachineFunction::iterator(FBB) == Fallthrough)
963 // If it's an unconditional branch to some block not the fall through, it
964 // doesn't fall through.
965 if (Cond.empty()) return false;
967 // Otherwise, if it is conditional and has no explicit false block, it falls
972 /// CanFallThrough - Return true if the specified can implicitly transfer
973 /// control to the block after it by falling off the end of it. This should
974 /// return false if it can reach the block after it, but it uses an explicit
975 /// branch to do so (e.g. a table jump).
977 /// True is a conservative answer.
979 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
980 MachineBasicBlock *TBB = 0, *FBB = 0;
981 SmallVector<MachineOperand, 4> Cond;
982 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true);
983 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
986 /// IsBetterFallthrough - Return true if it would be clearly better to
987 /// fall-through to MBB1 than to fall through into MBB2. This has to return
988 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
989 /// result in infinite loops.
990 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
991 MachineBasicBlock *MBB2) {
992 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
993 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
994 // optimize branches that branch to either a return block or an assert block
995 // into a fallthrough to the return.
996 if (MBB1->empty() || MBB2->empty()) return false;
998 // If there is a clear successor ordering we make sure that one block
999 // will fall through to the next
1000 if (MBB1->isSuccessor(MBB2)) return true;
1001 if (MBB2->isSuccessor(MBB1)) return false;
1003 MachineInstr *MBB1I = --MBB1->end();
1004 MachineInstr *MBB2I = --MBB2->end();
1005 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
1008 /// TailDuplicateBlocks - Look for small blocks that are unconditionally
1009 /// branched to and do not fall through. Tail-duplicate their instructions
1010 /// into their predecessors to eliminate (dynamic) branches.
1011 bool BranchFolder::TailDuplicateBlocks(MachineFunction &MF) {
1012 bool MadeChange = false;
1014 // Make sure blocks are numbered in order
1015 MF.RenumberBlocks();
1017 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
1018 MachineBasicBlock *MBB = I++;
1020 // Only duplicate blocks that end with unconditional branches.
1021 if (CanFallThrough(MBB))
1024 MadeChange |= TailDuplicate(MBB, MF);
1026 // If it is dead, remove it.
1027 if (MBB->pred_empty()) {
1028 NumInstrDups -= MBB->size();
1029 RemoveDeadBlock(MBB);
1037 /// TailDuplicate - If it is profitable, duplicate TailBB's contents in each
1038 /// of its predecessors.
1039 bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB,
1040 MachineFunction &MF) {
1041 // Don't try to tail-duplicate single-block loops.
1042 if (TailBB->isSuccessor(TailBB))
1045 // Set the limit on the number of instructions to duplicate, with a default
1046 // of one less than the tail-merge threshold. When optimizing for size,
1047 // duplicate only one, because one branch instruction can be eliminated to
1048 // compensate for the duplication.
1049 unsigned MaxDuplicateCount =
1050 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ?
1051 1 : TII->TailDuplicationLimit(*TailBB, TailMergeSize - 1);
1053 // Check the instructions in the block to determine whether tail-duplication
1054 // is invalid or unlikely to be profitable.
1056 bool HasCall = false;
1057 for (MachineBasicBlock::iterator I = TailBB->begin();
1058 I != TailBB->end(); ++I, ++i) {
1059 // Non-duplicable things shouldn't be tail-duplicated.
1060 if (I->getDesc().isNotDuplicable()) return false;
1061 // Don't duplicate more than the threshold.
1062 if (i == MaxDuplicateCount) return false;
1063 // Remember if we saw a call.
1064 if (I->getDesc().isCall()) HasCall = true;
1066 // Heuristically, don't tail-duplicate calls if it would expand code size,
1067 // as it's less likely to be worth the extra cost.
1068 if (i > 1 && HasCall)
1071 // Iterate through all the unique predecessors and tail-duplicate this
1072 // block into them, if possible. Copying the list ahead of time also
1073 // avoids trouble with the predecessor list reallocating.
1074 bool Changed = false;
1075 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
1076 TailBB->pred_end());
1077 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
1078 PE = Preds.end(); PI != PE; ++PI) {
1079 MachineBasicBlock *PredBB = *PI;
1081 assert(TailBB != PredBB &&
1082 "Single-block loop should have been rejected earlier!");
1083 if (PredBB->succ_size() > 1) continue;
1085 MachineBasicBlock *PredTBB, *PredFBB;
1086 SmallVector<MachineOperand, 4> PredCond;
1087 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
1089 if (!PredCond.empty())
1091 // EH edges are ignored by AnalyzeBranch.
1092 if (PredBB->succ_size() != 1)
1094 // Don't duplicate into a fall-through predecessor (at least for now).
1095 if (PredBB->isLayoutSuccessor(TailBB) && CanFallThrough(PredBB))
1098 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB
1099 << "From Succ: " << *TailBB);
1101 // Remove PredBB's unconditional branch.
1102 TII->RemoveBranch(*PredBB);
1103 // Clone the contents of TailBB into PredBB.
1104 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end();
1106 MachineInstr *NewMI = MF.CloneMachineInstr(I);
1107 PredBB->insert(PredBB->end(), NewMI);
1109 NumInstrDups += TailBB->size() - 1; // subtract one for removed branch
1112 PredBB->removeSuccessor(PredBB->succ_begin());
1113 assert(PredBB->succ_empty() &&
1114 "TailDuplicate called on block with multiple successors!");
1115 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
1116 E = TailBB->succ_end(); I != E; ++I)
1117 PredBB->addSuccessor(*I);
1123 // If TailBB was duplicated into all its predecessors except for the prior
1124 // block, which falls through unconditionally, move the contents of this
1125 // block into the prior block.
1126 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(TailBB));
1127 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1128 SmallVector<MachineOperand, 4> PriorCond;
1129 bool PriorUnAnalyzable =
1130 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1131 // This has to check PrevBB->succ_size() because EH edges are ignored by
1133 if (!PriorUnAnalyzable && PriorCond.empty() && !PriorTBB &&
1134 TailBB->pred_size() == 1 && PrevBB.succ_size() == 1 &&
1135 !TailBB->hasAddressTaken()) {
1136 DEBUG(errs() << "\nMerging into block: " << PrevBB
1137 << "From MBB: " << *TailBB);
1138 PrevBB.splice(PrevBB.end(), TailBB, TailBB->begin(), TailBB->end());
1139 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1140 assert(PrevBB.succ_empty());
1141 PrevBB.transferSuccessors(TailBB);
1148 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1149 /// block. This is never called on the entry block.
1150 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1151 bool MadeChange = false;
1152 MachineFunction &MF = *MBB->getParent();
1155 MachineFunction::iterator FallThrough = MBB;
1158 // If this block is empty, make everyone use its fall-through, not the block
1159 // explicitly. Landing pads should not do this since the landing-pad table
1160 // points to this block. Blocks with their addresses taken shouldn't be
1162 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1163 // Dead block? Leave for cleanup later.
1164 if (MBB->pred_empty()) return MadeChange;
1166 if (FallThrough == MF.end()) {
1167 // TODO: Simplify preds to not branch here if possible!
1169 // Rewrite all predecessors of the old block to go to the fallthrough
1171 while (!MBB->pred_empty()) {
1172 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1173 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1175 // If MBB was the target of a jump table, update jump tables to go to the
1176 // fallthrough instead.
1177 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough);
1183 // Check to see if we can simplify the terminator of the block before this
1185 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1187 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1188 SmallVector<MachineOperand, 4> PriorCond;
1189 bool PriorUnAnalyzable =
1190 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1191 if (!PriorUnAnalyzable) {
1192 // If the CFG for the prior block has extra edges, remove them.
1193 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1194 !PriorCond.empty());
1196 // If the previous branch is conditional and both conditions go to the same
1197 // destination, remove the branch, replacing it with an unconditional one or
1199 if (PriorTBB && PriorTBB == PriorFBB) {
1200 TII->RemoveBranch(PrevBB);
1202 if (PriorTBB != MBB)
1203 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1206 goto ReoptimizeBlock;
1209 // If the previous block unconditionally falls through to this block and
1210 // this block has no other predecessors, move the contents of this block
1211 // into the prior block. This doesn't usually happen when SimplifyCFG
1212 // has been used, but it can happen if tail merging splits a fall-through
1213 // predecessor of a block.
1214 // This has to check PrevBB->succ_size() because EH edges are ignored by
1216 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1217 PrevBB.succ_size() == 1 &&
1218 !MBB->hasAddressTaken()) {
1219 DEBUG(errs() << "\nMerging into block: " << PrevBB
1220 << "From MBB: " << *MBB);
1221 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1222 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1223 assert(PrevBB.succ_empty());
1224 PrevBB.transferSuccessors(MBB);
1229 // If the previous branch *only* branches to *this* block (conditional or
1230 // not) remove the branch.
1231 if (PriorTBB == MBB && PriorFBB == 0) {
1232 TII->RemoveBranch(PrevBB);
1235 goto ReoptimizeBlock;
1238 // If the prior block branches somewhere else on the condition and here if
1239 // the condition is false, remove the uncond second branch.
1240 if (PriorFBB == MBB) {
1241 TII->RemoveBranch(PrevBB);
1242 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1245 goto ReoptimizeBlock;
1248 // If the prior block branches here on true and somewhere else on false, and
1249 // if the branch condition is reversible, reverse the branch to create a
1251 if (PriorTBB == MBB) {
1252 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1253 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1254 TII->RemoveBranch(PrevBB);
1255 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
1258 goto ReoptimizeBlock;
1262 // If this block has no successors (e.g. it is a return block or ends with
1263 // a call to a no-return function like abort or __cxa_throw) and if the pred
1264 // falls through into this block, and if it would otherwise fall through
1265 // into the block after this, move this block to the end of the function.
1267 // We consider it more likely that execution will stay in the function (e.g.
1268 // due to loops) than it is to exit it. This asserts in loops etc, moving
1269 // the assert condition out of the loop body.
1270 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1271 MachineFunction::iterator(PriorTBB) == FallThrough &&
1272 !CanFallThrough(MBB)) {
1273 bool DoTransform = true;
1275 // We have to be careful that the succs of PredBB aren't both no-successor
1276 // blocks. If neither have successors and if PredBB is the second from
1277 // last block in the function, we'd just keep swapping the two blocks for
1278 // last. Only do the swap if one is clearly better to fall through than
1280 if (FallThrough == --MF.end() &&
1281 !IsBetterFallthrough(PriorTBB, MBB))
1282 DoTransform = false;
1284 // We don't want to do this transformation if we have control flow like:
1293 // In this case, we could actually be moving the return block *into* a
1295 if (DoTransform && !MBB->succ_empty() &&
1296 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
1297 DoTransform = false;
1301 // Reverse the branch so we will fall through on the previous true cond.
1302 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1303 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1304 DEBUG(errs() << "\nMoving MBB: " << *MBB
1305 << "To make fallthrough to: " << *PriorTBB << "\n");
1307 TII->RemoveBranch(PrevBB);
1308 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1310 // Move this block to the end of the function.
1311 MBB->moveAfter(--MF.end());
1320 // Analyze the branch in the current block.
1321 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1322 SmallVector<MachineOperand, 4> CurCond;
1323 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1324 if (!CurUnAnalyzable) {
1325 // If the CFG for the prior block has extra edges, remove them.
1326 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1328 // If this is a two-way branch, and the FBB branches to this block, reverse
1329 // the condition so the single-basic-block loop is faster. Instead of:
1330 // Loop: xxx; jcc Out; jmp Loop
1332 // Loop: xxx; jncc Loop; jmp Out
1333 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1334 SmallVector<MachineOperand, 4> NewCond(CurCond);
1335 if (!TII->ReverseBranchCondition(NewCond)) {
1336 TII->RemoveBranch(*MBB);
1337 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1340 goto ReoptimizeBlock;
1344 // If this branch is the only thing in its block, see if we can forward
1345 // other blocks across it.
1346 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1347 MBB->begin()->getDesc().isBranch() && CurTBB != MBB &&
1348 !MBB->hasAddressTaken()) {
1349 // This block may contain just an unconditional branch. Because there can
1350 // be 'non-branch terminators' in the block, try removing the branch and
1351 // then seeing if the block is empty.
1352 TII->RemoveBranch(*MBB);
1354 // If this block is just an unconditional branch to CurTBB, we can
1355 // usually completely eliminate the block. The only case we cannot
1356 // completely eliminate the block is when the block before this one
1357 // falls through into MBB and we can't understand the prior block's branch
1360 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
1361 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1362 !PrevBB.isSuccessor(MBB)) {
1363 // If the prior block falls through into us, turn it into an
1364 // explicit branch to us to make updates simpler.
1365 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1366 PriorTBB != MBB && PriorFBB != MBB) {
1367 if (PriorTBB == 0) {
1368 assert(PriorCond.empty() && PriorFBB == 0 &&
1369 "Bad branch analysis");
1372 assert(PriorFBB == 0 && "Machine CFG out of date!");
1375 TII->RemoveBranch(PrevBB);
1376 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1379 // Iterate through all the predecessors, revectoring each in-turn.
1381 bool DidChange = false;
1382 bool HasBranchToSelf = false;
1383 while(PI != MBB->pred_size()) {
1384 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1386 // If this block has an uncond branch to itself, leave it.
1388 HasBranchToSelf = true;
1391 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1392 // If this change resulted in PMBB ending in a conditional
1393 // branch where both conditions go to the same destination,
1394 // change this to an unconditional branch (and fix the CFG).
1395 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1396 SmallVector<MachineOperand, 4> NewCurCond;
1397 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1398 NewCurFBB, NewCurCond, true);
1399 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1400 TII->RemoveBranch(*PMBB);
1402 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1405 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1410 // Change any jumptables to go to the new MBB.
1411 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB);
1415 if (!HasBranchToSelf) return MadeChange;
1420 // Add the branch back if the block is more than just an uncond branch.
1421 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1425 // If the prior block doesn't fall through into this block, and if this
1426 // block doesn't fall through into some other block, see if we can find a
1427 // place to move this block where a fall-through will happen.
1428 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
1429 PriorTBB, PriorFBB, PriorCond)) {
1431 // Now we know that there was no fall-through into this block, check to
1432 // see if it has a fall-through into its successor.
1433 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1436 if (!MBB->isLandingPad()) {
1437 // Check all the predecessors of this block. If one of them has no fall
1438 // throughs, move this block right after it.
1439 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1440 E = MBB->pred_end(); PI != E; ++PI) {
1441 // Analyze the branch at the end of the pred.
1442 MachineBasicBlock *PredBB = *PI;
1443 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1444 MachineBasicBlock *PredTBB, *PredFBB;
1445 SmallVector<MachineOperand, 4> PredCond;
1446 if (PredBB != MBB && !CanFallThrough(PredBB) &&
1447 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1448 && (!CurFallsThru || !CurTBB || !CurFBB)
1449 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1450 // If the current block doesn't fall through, just move it.
1451 // If the current block can fall through and does not end with a
1452 // conditional branch, we need to append an unconditional jump to
1453 // the (current) next block. To avoid a possible compile-time
1454 // infinite loop, move blocks only backward in this case.
1455 // Also, if there are already 2 branches here, we cannot add a third;
1456 // this means we have the case
1461 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1463 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1465 MBB->moveAfter(PredBB);
1467 goto ReoptimizeBlock;
1472 if (!CurFallsThru) {
1473 // Check all successors to see if we can move this block before it.
1474 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1475 E = MBB->succ_end(); SI != E; ++SI) {
1476 // Analyze the branch at the end of the block before the succ.
1477 MachineBasicBlock *SuccBB = *SI;
1478 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1480 // If this block doesn't already fall-through to that successor, and if
1481 // the succ doesn't already have a block that can fall through into it,
1482 // and if the successor isn't an EH destination, we can arrange for the
1483 // fallthrough to happen.
1484 if (SuccBB != MBB && &*SuccPrev != MBB &&
1485 !CanFallThrough(SuccPrev) && !CurUnAnalyzable &&
1486 !SuccBB->isLandingPad()) {
1487 MBB->moveBefore(SuccBB);
1489 goto ReoptimizeBlock;
1493 // Okay, there is no really great place to put this block. If, however,
1494 // the block before this one would be a fall-through if this block were
1495 // removed, move this block to the end of the function.
1496 MachineBasicBlock *PrevTBB, *PrevFBB;
1497 SmallVector<MachineOperand, 4> PrevCond;
1498 if (FallThrough != MF.end() &&
1499 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1500 PrevBB.isSuccessor(FallThrough)) {
1501 MBB->moveAfter(--MF.end());