1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass forwards branches to unconditional branches to make them branch
11 // directly to the target block. This pass often results in dead MBB's, which
14 // Note that this pass must be run after register allocation, it cannot handle
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "branchfolding"
20 #include "BranchFolding.h"
21 #include "llvm/Function.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/RegisterScavenging.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Target/TargetRegisterInfo.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/ADT/SmallSet.h"
35 #include "llvm/ADT/SetVector.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/ADT/STLExtras.h"
41 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
42 STATISTIC(NumBranchOpts, "Number of branches optimized");
43 STATISTIC(NumTailMerge , "Number of block tails merged");
44 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
45 cl::init(cl::BOU_UNSET), cl::Hidden);
46 // Throttle for huge numbers of predecessors (compile speed problems)
47 static cl::opt<unsigned>
48 TailMergeThreshold("tail-merge-threshold",
49 cl::desc("Max number of predecessors to consider tail merging"),
50 cl::init(150), cl::Hidden);
52 // Heuristic for tail merging (and, inversely, tail duplication).
53 // TODO: This should be replaced with a target query.
54 static cl::opt<unsigned>
55 TailMergeSize("tail-merge-size",
56 cl::desc("Min number of instructions to consider tail merging"),
57 cl::init(3), cl::Hidden);
60 /// BranchFolderPass - Wrap branch folder in a machine function pass.
61 class BranchFolderPass : public MachineFunctionPass,
65 explicit BranchFolderPass(bool defaultEnableTailMerge)
66 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {}
68 virtual bool runOnMachineFunction(MachineFunction &MF);
69 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
73 char BranchFolderPass::ID = 0;
75 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
76 return new BranchFolderPass(DefaultEnableTailMerge);
79 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
80 return OptimizeFunction(MF,
81 MF.getTarget().getInstrInfo(),
82 MF.getTarget().getRegisterInfo(),
83 getAnalysisIfAvailable<MachineModuleInfo>());
87 BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
88 switch (FlagEnableTailMerge) {
89 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
90 case cl::BOU_TRUE: EnableTailMerge = true; break;
91 case cl::BOU_FALSE: EnableTailMerge = false; break;
95 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
96 /// function, updating the CFG.
97 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
98 assert(MBB->pred_empty() && "MBB must be dead!");
99 DEBUG(errs() << "\nRemoving MBB: " << *MBB);
101 MachineFunction *MF = MBB->getParent();
102 // drop all successors.
103 while (!MBB->succ_empty())
104 MBB->removeSuccessor(MBB->succ_end()-1);
106 // If there are any labels in the basic block, unregister them from
107 // MachineModuleInfo.
108 if (MMI && !MBB->empty()) {
109 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
112 // The label ID # is always operand #0, an immediate.
113 MMI->InvalidateLabel(I->getOperand(0).getImm());
121 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
122 /// followed by terminators, and if the implicitly defined registers are not
123 /// used by the terminators, remove those implicit_def's. e.g.
125 /// r0 = implicit_def
126 /// r1 = implicit_def
128 /// This block can be optimized away later if the implicit instructions are
130 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
131 SmallSet<unsigned, 4> ImpDefRegs;
132 MachineBasicBlock::iterator I = MBB->begin();
133 while (I != MBB->end()) {
134 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF)
136 unsigned Reg = I->getOperand(0).getReg();
137 ImpDefRegs.insert(Reg);
138 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
139 unsigned SubReg = *SubRegs; ++SubRegs)
140 ImpDefRegs.insert(SubReg);
143 if (ImpDefRegs.empty())
146 MachineBasicBlock::iterator FirstTerm = I;
147 while (I != MBB->end()) {
148 if (!TII->isUnpredicatedTerminator(I))
150 // See if it uses any of the implicitly defined registers.
151 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
152 MachineOperand &MO = I->getOperand(i);
153 if (!MO.isReg() || !MO.isUse())
155 unsigned Reg = MO.getReg();
156 if (ImpDefRegs.count(Reg))
163 while (I != FirstTerm) {
164 MachineInstr *ImpDefMI = &*I;
166 MBB->erase(ImpDefMI);
172 /// OptimizeFunction - Perhaps branch folding, tail merging and other
173 /// CFG optimizations on the given function.
174 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
175 const TargetInstrInfo *tii,
176 const TargetRegisterInfo *tri,
177 MachineModuleInfo *mmi) {
178 if (!tii) return false;
184 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
186 // Fix CFG. The later algorithms expect it to be right.
187 bool MadeChange = false;
188 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
189 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
190 SmallVector<MachineOperand, 4> Cond;
191 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
192 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
193 MadeChange |= OptimizeImpDefsBlock(MBB);
197 bool MadeChangeThisIteration = true;
198 while (MadeChangeThisIteration) {
199 MadeChangeThisIteration = false;
200 MadeChangeThisIteration |= TailMergeBlocks(MF);
201 MadeChangeThisIteration |= OptimizeBranches(MF);
202 MadeChange |= MadeChangeThisIteration;
205 // Do tail duplication once after tail merging is done. Otherwise it is
206 // tough to avoid situations where tail duplication and tail merging undo
207 // each other's transformations ad infinitum.
208 MadeChangeThisIteration |= TailDuplicateBlocks(MF);
210 // See if any jump tables have become mergable or dead as the code generator
212 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
213 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
215 // Figure out how these jump tables should be merged.
216 std::vector<unsigned> JTMapping;
217 JTMapping.reserve(JTs.size());
219 // We always keep the 0th jump table.
220 JTMapping.push_back(0);
222 // Scan the jump tables, seeing if there are any duplicates. Note that this
223 // is N^2, which should be fixed someday.
224 for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
225 if (JTs[i].MBBs.empty())
226 JTMapping.push_back(i);
228 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
231 // If a jump table was merge with another one, walk the function rewriting
232 // references to jump tables to reference the new JT ID's. Keep track of
233 // whether we see a jump table idx, if not, we can delete the JT.
234 BitVector JTIsLive(JTs.size());
235 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
237 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
239 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
240 MachineOperand &Op = I->getOperand(op);
241 if (!Op.isJTI()) continue;
242 unsigned NewIdx = JTMapping[Op.getIndex()];
245 // Remember that this JT is live.
246 JTIsLive.set(NewIdx);
250 // Finally, remove dead jump tables. This happens either because the
251 // indirect jump was unreachable (and thus deleted) or because the jump
252 // table was merged with some other one.
253 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
254 if (!JTIsLive.test(i)) {
255 JTI->RemoveJumpTable(i);
264 //===----------------------------------------------------------------------===//
265 // Tail Merging of Blocks
266 //===----------------------------------------------------------------------===//
268 /// HashMachineInstr - Compute a hash value for MI and its operands.
269 static unsigned HashMachineInstr(const MachineInstr *MI) {
270 unsigned Hash = MI->getOpcode();
271 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
272 const MachineOperand &Op = MI->getOperand(i);
274 // Merge in bits from the operand if easy.
275 unsigned OperandHash = 0;
276 switch (Op.getType()) {
277 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
278 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
279 case MachineOperand::MO_MachineBasicBlock:
280 OperandHash = Op.getMBB()->getNumber();
282 case MachineOperand::MO_FrameIndex:
283 case MachineOperand::MO_ConstantPoolIndex:
284 case MachineOperand::MO_JumpTableIndex:
285 OperandHash = Op.getIndex();
287 case MachineOperand::MO_GlobalAddress:
288 case MachineOperand::MO_ExternalSymbol:
289 // Global address / external symbol are too hard, don't bother, but do
290 // pull in the offset.
291 OperandHash = Op.getOffset();
296 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
301 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
302 /// with no successors, we hash two instructions, because cross-jumping
303 /// only saves code when at least two instructions are removed (since a
304 /// branch must be inserted). For blocks with a successor, one of the
305 /// two blocks to be tail-merged will end with a branch already, so
306 /// it gains to cross-jump even for one instruction.
307 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
308 unsigned minCommonTailLength) {
309 MachineBasicBlock::const_iterator I = MBB->end();
310 if (I == MBB->begin())
311 return 0; // Empty MBB.
314 unsigned Hash = HashMachineInstr(I);
316 if (I == MBB->begin() || minCommonTailLength == 1)
317 return Hash; // Single instr MBB.
320 // Hash in the second-to-last instruction.
321 Hash ^= HashMachineInstr(I) << 2;
325 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
326 /// of instructions they actually have in common together at their end. Return
327 /// iterators for the first shared instruction in each block.
328 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
329 MachineBasicBlock *MBB2,
330 MachineBasicBlock::iterator &I1,
331 MachineBasicBlock::iterator &I2) {
335 unsigned TailLen = 0;
336 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
338 if (!I1->isIdenticalTo(I2) ||
339 // FIXME: This check is dubious. It's used to get around a problem where
340 // people incorrectly expect inline asm directives to remain in the same
341 // relative order. This is untenable because normal compiler
342 // optimizations (like this one) may reorder and/or merge these
344 I1->getOpcode() == TargetInstrInfo::INLINEASM) {
353 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
354 /// after it, replacing it with an unconditional branch to NewDest. This
355 /// returns true if OldInst's block is modified, false if NewDest is modified.
356 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
357 MachineBasicBlock *NewDest) {
358 MachineBasicBlock *OldBB = OldInst->getParent();
360 // Remove all the old successors of OldBB from the CFG.
361 while (!OldBB->succ_empty())
362 OldBB->removeSuccessor(OldBB->succ_begin());
364 // Remove all the dead instructions from the end of OldBB.
365 OldBB->erase(OldInst, OldBB->end());
367 // If OldBB isn't immediately before OldBB, insert a branch to it.
368 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
369 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
370 OldBB->addSuccessor(NewDest);
374 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
375 /// MBB so that the part before the iterator falls into the part starting at the
376 /// iterator. This returns the new MBB.
377 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
378 MachineBasicBlock::iterator BBI1) {
379 MachineFunction &MF = *CurMBB.getParent();
381 // Create the fall-through block.
382 MachineFunction::iterator MBBI = &CurMBB;
383 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
384 CurMBB.getParent()->insert(++MBBI, NewMBB);
386 // Move all the successors of this block to the specified block.
387 NewMBB->transferSuccessors(&CurMBB);
389 // Add an edge from CurMBB to NewMBB for the fall-through.
390 CurMBB.addSuccessor(NewMBB);
392 // Splice the code over.
393 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
395 // For targets that use the register scavenger, we must maintain LiveIns.
397 RS->enterBasicBlock(&CurMBB);
399 RS->forward(prior(CurMBB.end()));
400 BitVector RegsLiveAtExit(TRI->getNumRegs());
401 RS->getRegsUsed(RegsLiveAtExit, false);
402 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
403 if (RegsLiveAtExit[i])
404 NewMBB->addLiveIn(i);
410 /// EstimateRuntime - Make a rough estimate for how long it will take to run
411 /// the specified code.
412 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
413 MachineBasicBlock::iterator E) {
415 for (; I != E; ++I) {
416 const TargetInstrDesc &TID = I->getDesc();
419 else if (TID.mayLoad() || TID.mayStore())
427 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
428 // branches temporarily for tail merging). In the case where CurMBB ends
429 // with a conditional branch to the next block, optimize by reversing the
430 // test and conditionally branching to SuccMBB instead.
431 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
432 const TargetInstrInfo *TII) {
433 MachineFunction *MF = CurMBB->getParent();
434 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
435 MachineBasicBlock *TBB = 0, *FBB = 0;
436 SmallVector<MachineOperand, 4> Cond;
437 if (I != MF->end() &&
438 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
439 MachineBasicBlock *NextBB = I;
440 if (TBB == NextBB && !Cond.empty() && !FBB) {
441 if (!TII->ReverseBranchCondition(Cond)) {
442 TII->RemoveBranch(*CurMBB);
443 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
448 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
452 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
453 if (getHash() < o.getHash())
455 else if (getHash() > o.getHash())
457 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
459 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
462 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
463 // an object with itself.
464 #ifndef _GLIBCXX_DEBUG
465 llvm_unreachable("Predecessor appears twice");
471 /// CountTerminators - Count the number of terminators in the given
472 /// block and set I to the position of the first non-terminator, if there
473 /// is one, or MBB->end() otherwise.
474 static unsigned CountTerminators(MachineBasicBlock *MBB,
475 MachineBasicBlock::iterator &I) {
477 unsigned NumTerms = 0;
479 if (I == MBB->begin()) {
484 if (!I->getDesc().isTerminator()) break;
490 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
491 /// and decide if it would be profitable to merge those tails. Return the
492 /// length of the common tail and iterators to the first common instruction
494 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
495 MachineBasicBlock *MBB2,
496 unsigned minCommonTailLength,
497 unsigned &CommonTailLen,
498 MachineBasicBlock::iterator &I1,
499 MachineBasicBlock::iterator &I2,
500 MachineBasicBlock *SuccBB,
501 MachineBasicBlock *PredBB) {
502 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
503 MachineFunction *MF = MBB1->getParent();
505 if (CommonTailLen == 0)
508 // It's almost always profitable to merge any number of non-terminator
509 // instructions with the block that falls through into the common successor.
510 if (MBB1 == PredBB || MBB2 == PredBB) {
511 MachineBasicBlock::iterator I;
512 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
513 if (CommonTailLen > NumTerms)
517 // If one of the blocks can be completely merged and happens to be in
518 // a position where the other could fall through into it, merge any number
519 // of instructions, because it can be done without a branch.
520 // TODO: If the blocks are not adjacent, move one of them so that they are?
521 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
523 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
526 // If both blocks have an unconditional branch temporarily stripped out,
527 // count that as an additional common instruction for the following
529 unsigned EffectiveTailLen = CommonTailLen;
530 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
531 !MBB1->back().getDesc().isBarrier() &&
532 !MBB2->back().getDesc().isBarrier())
535 // Check if the common tail is long enough to be worthwhile.
536 if (EffectiveTailLen >= minCommonTailLength)
539 // If we are optimizing for code size, 2 instructions in common is enough if
540 // we don't have to split a block. At worst we will be introducing 1 new
541 // branch instruction, which is likely to be smaller than the 2
542 // instructions that would be deleted in the merge.
543 if (EffectiveTailLen >= 2 &&
544 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
545 (I1 == MBB1->begin() || I2 == MBB2->begin()))
551 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
552 /// hash CurHash (guaranteed to match the last element). Build the vector
553 /// SameTails of all those that have the (same) largest number of instructions
554 /// in common of any pair of these blocks. SameTails entries contain an
555 /// iterator into MergePotentials (from which the MachineBasicBlock can be
556 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
557 /// instruction where the matching code sequence begins.
558 /// Order of elements in SameTails is the reverse of the order in which
559 /// those blocks appear in MergePotentials (where they are not necessarily
561 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
562 unsigned minCommonTailLength,
563 MachineBasicBlock *SuccBB,
564 MachineBasicBlock *PredBB) {
565 unsigned maxCommonTailLength = 0U;
567 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
568 MPIterator HighestMPIter = prior(MergePotentials.end());
569 for (MPIterator CurMPIter = prior(MergePotentials.end()),
570 B = MergePotentials.begin();
571 CurMPIter != B && CurMPIter->getHash() == CurHash;
573 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
574 unsigned CommonTailLen;
575 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
577 CommonTailLen, TrialBBI1, TrialBBI2,
579 if (CommonTailLen > maxCommonTailLength) {
581 maxCommonTailLength = CommonTailLen;
582 HighestMPIter = CurMPIter;
583 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
585 if (HighestMPIter == CurMPIter &&
586 CommonTailLen == maxCommonTailLength)
587 SameTails.push_back(SameTailElt(I, TrialBBI2));
593 return maxCommonTailLength;
596 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
597 /// MergePotentials, restoring branches at ends of blocks as appropriate.
598 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
599 MachineBasicBlock *SuccBB,
600 MachineBasicBlock *PredBB) {
601 MPIterator CurMPIter, B;
602 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
603 CurMPIter->getHash() == CurHash;
605 // Put the unconditional branch back, if we need one.
606 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
607 if (SuccBB && CurMBB != PredBB)
608 FixTail(CurMBB, SuccBB, TII);
612 if (CurMPIter->getHash() != CurHash)
614 MergePotentials.erase(CurMPIter, MergePotentials.end());
617 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
618 /// only of the common tail. Create a block that does by splitting one.
619 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
620 unsigned maxCommonTailLength) {
621 unsigned commonTailIndex = 0;
622 unsigned TimeEstimate = ~0U;
623 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
624 // Use PredBB if possible; that doesn't require a new branch.
625 if (SameTails[i].getBlock() == PredBB) {
629 // Otherwise, make a (fairly bogus) choice based on estimate of
630 // how long it will take the various blocks to execute.
631 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
632 SameTails[i].getTailStartPos());
633 if (t <= TimeEstimate) {
639 MachineBasicBlock::iterator BBI =
640 SameTails[commonTailIndex].getTailStartPos();
641 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
643 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
644 << maxCommonTailLength);
646 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
647 SameTails[commonTailIndex].setBlock(newMBB);
648 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
650 // If we split PredBB, newMBB is the new predecessor.
654 return commonTailIndex;
657 // See if any of the blocks in MergePotentials (which all have a common single
658 // successor, or all have no successor) can be tail-merged. If there is a
659 // successor, any blocks in MergePotentials that are not tail-merged and
660 // are not immediately before Succ must have an unconditional branch to
661 // Succ added (but the predecessor/successor lists need no adjustment).
662 // The lone predecessor of Succ that falls through into Succ,
663 // if any, is given in PredBB.
665 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
666 MachineBasicBlock *PredBB) {
667 bool MadeChange = false;
669 // Except for the special cases below, tail-merge if there are at least
670 // this many instructions in common.
671 unsigned minCommonTailLength = TailMergeSize;
673 // If there's a successor block, there are some cases which don't require
674 // new branching and as such are very likely to be profitable.
676 if (SuccBB->pred_size() == MergePotentials.size() &&
677 !MergePotentials[0].getBlock()->empty()) {
678 // If all the predecessors have at least one tail instruction in common,
679 // merging is very likely to be a win since it won't require an increase
680 // in static branches, and it will decrease the static instruction count.
681 bool AllPredsMatch = true;
682 MachineBasicBlock::iterator FirstNonTerm;
683 unsigned MinNumTerms = CountTerminators(MergePotentials[0].getBlock(),
685 if (FirstNonTerm != MergePotentials[0].getBlock()->end()) {
686 for (unsigned i = 1, e = MergePotentials.size(); i != e; ++i) {
687 MachineBasicBlock::iterator OtherFirstNonTerm;
688 unsigned NumTerms = CountTerminators(MergePotentials[0].getBlock(),
690 if (NumTerms < MinNumTerms)
691 MinNumTerms = NumTerms;
692 if (OtherFirstNonTerm == MergePotentials[i].getBlock()->end() ||
693 OtherFirstNonTerm->isIdenticalTo(FirstNonTerm)) {
694 AllPredsMatch = false;
699 // If they all have an instruction in common, do any amount of merging.
701 minCommonTailLength = MinNumTerms + 1;
706 DEBUG(errs() << "\nTryTailMergeBlocks: ";
707 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
708 errs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
709 << (i == e-1 ? "" : ", ");
712 errs() << " with successor BB#" << SuccBB->getNumber() << '\n';
714 errs() << " which has fall-through from BB#"
715 << PredBB->getNumber() << "\n";
717 errs() << "Looking for common tails of at least "
718 << minCommonTailLength << " instruction"
719 << (minCommonTailLength == 1 ? "" : "s") << '\n';
722 // Sort by hash value so that blocks with identical end sequences sort
724 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
726 // Walk through equivalence sets looking for actual exact matches.
727 while (MergePotentials.size() > 1) {
728 unsigned CurHash = MergePotentials.back().getHash();
730 // Build SameTails, identifying the set of blocks with this hash code
731 // and with the maximum number of instructions in common.
732 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
736 // If we didn't find any pair that has at least minCommonTailLength
737 // instructions in common, remove all blocks with this hash code and retry.
738 if (SameTails.empty()) {
739 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
743 // If one of the blocks is the entire common tail (and not the entry
744 // block, which we can't jump to), we can treat all blocks with this same
745 // tail at once. Use PredBB if that is one of the possibilities, as that
746 // will not introduce any extra branches.
747 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
748 getParent()->begin();
749 unsigned commonTailIndex = SameTails.size();
750 // If there are two blocks, check to see if one can be made to fall through
752 if (SameTails.size() == 2 &&
753 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
754 SameTails[1].tailIsWholeBlock())
756 else if (SameTails.size() == 2 &&
757 SameTails[1].getBlock()->isLayoutSuccessor(
758 SameTails[0].getBlock()) &&
759 SameTails[0].tailIsWholeBlock())
762 // Otherwise just pick one, favoring the fall-through predecessor if
764 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
765 MachineBasicBlock *MBB = SameTails[i].getBlock();
766 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
772 if (SameTails[i].tailIsWholeBlock())
777 if (commonTailIndex == SameTails.size() ||
778 (SameTails[commonTailIndex].getBlock() == PredBB &&
779 !SameTails[commonTailIndex].tailIsWholeBlock())) {
780 // None of the blocks consist entirely of the common tail.
781 // Split a block so that one does.
782 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
785 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
786 // MBB is common tail. Adjust all other BB's to jump to this one.
787 // Traversal must be forwards so erases work.
788 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber()
790 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
791 if (commonTailIndex == i)
793 DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber()
794 << (i == e-1 ? "" : ", "));
795 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
796 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
797 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
798 MergePotentials.erase(SameTails[i].getMPIter());
800 DEBUG(errs() << "\n");
801 // We leave commonTailIndex in the worklist in case there are other blocks
802 // that match it with a smaller number of instructions.
808 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
810 if (!EnableTailMerge) return false;
812 bool MadeChange = false;
814 // First find blocks with no successors.
815 MergePotentials.clear();
816 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
818 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
821 // See if we can do any tail merging on those.
822 if (MergePotentials.size() < TailMergeThreshold &&
823 MergePotentials.size() >= 2)
824 MadeChange |= TryTailMergeBlocks(NULL, NULL);
826 // Look at blocks (IBB) with multiple predecessors (PBB).
827 // We change each predecessor to a canonical form, by
828 // (1) temporarily removing any unconditional branch from the predecessor
830 // (2) alter conditional branches so they branch to the other block
831 // not IBB; this may require adding back an unconditional branch to IBB
832 // later, where there wasn't one coming in. E.g.
834 // fallthrough to QBB
837 // with a conceptual B to IBB after that, which never actually exists.
838 // With those changes, we see whether the predecessors' tails match,
839 // and merge them if so. We change things out of canonical form and
840 // back to the way they were later in the process. (OptimizeBranches
841 // would undo some of this, but we can't use it, because we'd get into
842 // a compile-time infinite loop repeatedly doing and undoing the same
845 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end();
847 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
848 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
849 MachineBasicBlock *IBB = I;
850 MachineBasicBlock *PredBB = prior(I);
851 MergePotentials.clear();
852 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
855 MachineBasicBlock *PBB = *P;
856 // Skip blocks that loop to themselves, can't tail merge these.
859 // Visit each predecessor only once.
860 if (!UniquePreds.insert(PBB))
862 MachineBasicBlock *TBB = 0, *FBB = 0;
863 SmallVector<MachineOperand, 4> Cond;
864 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
865 // Failing case: IBB is the target of a cbr, and
866 // we cannot reverse the branch.
867 SmallVector<MachineOperand, 4> NewCond(Cond);
868 if (!Cond.empty() && TBB == IBB) {
869 if (TII->ReverseBranchCondition(NewCond))
871 // This is the QBB case described above
873 FBB = next(MachineFunction::iterator(PBB));
875 // Failing case: the only way IBB can be reached from PBB is via
876 // exception handling. Happens for landing pads. Would be nice
877 // to have a bit in the edge so we didn't have to do all this.
878 if (IBB->isLandingPad()) {
879 MachineFunction::iterator IP = PBB; IP++;
880 MachineBasicBlock *PredNextBB = NULL;
884 if (IBB != PredNextBB) // fallthrough
887 if (TBB != IBB && FBB != IBB) // cbr then ubr
889 } else if (Cond.empty()) {
890 if (TBB != IBB) // ubr
893 if (TBB != IBB && IBB != PredNextBB) // cbr
897 // Remove the unconditional branch at the end, if any.
898 if (TBB && (Cond.empty() || FBB)) {
899 TII->RemoveBranch(*PBB);
901 // reinsert conditional branch only, for now
902 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
904 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
908 if (MergePotentials.size() >= 2)
909 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
910 // Reinsert an unconditional branch if needed.
911 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
912 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
913 if (MergePotentials.size() == 1 &&
914 MergePotentials.begin()->getBlock() != PredBB)
915 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
921 //===----------------------------------------------------------------------===//
922 // Branch Optimization
923 //===----------------------------------------------------------------------===//
925 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
926 bool MadeChange = false;
928 // Make sure blocks are numbered in order
931 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
932 MachineBasicBlock *MBB = I++;
933 MadeChange |= OptimizeBlock(MBB);
935 // If it is dead, remove it.
936 if (MBB->pred_empty()) {
937 RemoveDeadBlock(MBB);
946 /// CanFallThrough - Return true if the specified block (with the specified
947 /// branch condition) can implicitly transfer control to the block after it by
948 /// falling off the end of it. This should return false if it can reach the
949 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
951 /// True is a conservative answer.
953 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
954 bool BranchUnAnalyzable,
955 MachineBasicBlock *TBB,
956 MachineBasicBlock *FBB,
957 const SmallVectorImpl<MachineOperand> &Cond) {
958 MachineFunction::iterator Fallthrough = CurBB;
960 // If FallthroughBlock is off the end of the function, it can't fall through.
961 if (Fallthrough == CurBB->getParent()->end())
964 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
965 if (!CurBB->isSuccessor(Fallthrough))
968 // If we couldn't analyze the branch, examine the last instruction.
969 // If the block doesn't end in a known control barrier, assume fallthrough
970 // is possible. The isPredicable check is needed because this code can be
971 // called during IfConversion, where an instruction which is normally a
972 // Barrier is predicated and thus no longer an actual control barrier. This
973 // is over-conservative though, because if an instruction isn't actually
974 // predicated we could still treat it like a barrier.
975 if (BranchUnAnalyzable)
976 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() ||
977 CurBB->back().getDesc().isPredicable();
979 // If there is no branch, control always falls through.
980 if (TBB == 0) return true;
982 // If there is some explicit branch to the fallthrough block, it can obviously
983 // reach, even though the branch should get folded to fall through implicitly.
984 if (MachineFunction::iterator(TBB) == Fallthrough ||
985 MachineFunction::iterator(FBB) == Fallthrough)
988 // If it's an unconditional branch to some block not the fall through, it
989 // doesn't fall through.
990 if (Cond.empty()) return false;
992 // Otherwise, if it is conditional and has no explicit false block, it falls
997 /// CanFallThrough - Return true if the specified can implicitly transfer
998 /// control to the block after it by falling off the end of it. This should
999 /// return false if it can reach the block after it, but it uses an explicit
1000 /// branch to do so (e.g. a table jump).
1002 /// True is a conservative answer.
1004 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
1005 MachineBasicBlock *TBB = 0, *FBB = 0;
1006 SmallVector<MachineOperand, 4> Cond;
1007 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true);
1008 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
1011 /// IsBetterFallthrough - Return true if it would be clearly better to
1012 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1013 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1014 /// result in infinite loops.
1015 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1016 MachineBasicBlock *MBB2) {
1017 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1018 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1019 // optimize branches that branch to either a return block or an assert block
1020 // into a fallthrough to the return.
1021 if (MBB1->empty() || MBB2->empty()) return false;
1023 // If there is a clear successor ordering we make sure that one block
1024 // will fall through to the next
1025 if (MBB1->isSuccessor(MBB2)) return true;
1026 if (MBB2->isSuccessor(MBB1)) return false;
1028 MachineInstr *MBB1I = --MBB1->end();
1029 MachineInstr *MBB2I = --MBB2->end();
1030 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
1033 /// TailDuplicateBlocks - Look for small blocks that are unconditionally
1034 /// branched to and do not fall through. Tail-duplicate their instructions
1035 /// into their predecessors to eliminate (dynamic) branches.
1036 bool BranchFolder::TailDuplicateBlocks(MachineFunction &MF) {
1037 bool MadeChange = false;
1039 // Make sure blocks are numbered in order
1040 MF.RenumberBlocks();
1042 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
1043 MachineBasicBlock *MBB = I++;
1045 // Only duplicate blocks that end with unconditional branches.
1046 if (CanFallThrough(MBB))
1049 MadeChange |= TailDuplicate(MBB, MF);
1051 // If it is dead, remove it.
1052 if (MBB->pred_empty()) {
1053 RemoveDeadBlock(MBB);
1061 /// TailDuplicate - If it is profitable, duplicate TailBB's contents in each
1062 /// of its predecessors.
1063 bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB,
1064 MachineFunction &MF) {
1065 // Don't try to tail-duplicate single-block loops.
1066 if (TailBB->isSuccessor(TailBB))
1069 // Duplicate up to one less than the tail-merge threshold. When optimizing
1070 // for size, duplicate only one, because one branch instruction can be
1071 // eliminated to compensate for the duplication.
1072 unsigned MaxDuplicateCount =
1073 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ?
1074 1 : (TailMergeSize - 1);
1076 // Check the instructions in the block to determine whether tail-duplication
1077 // is invalid or unlikely to be profitable.
1079 bool HasCall = false;
1080 for (MachineBasicBlock::iterator I = TailBB->begin();
1081 I != TailBB->end(); ++I, ++i) {
1082 // Non-duplicable things shouldn't be tail-duplicated.
1083 if (I->getDesc().isNotDuplicable()) return false;
1084 // Don't duplicate more than the threshold.
1085 if (i == MaxDuplicateCount) return false;
1086 // Remember if we saw a call.
1087 if (I->getDesc().isCall()) HasCall = true;
1089 // Heuristically, don't tail-duplicate calls if it would expand code size,
1090 // as it's less likely to be worth the extra cost.
1091 if (i > 1 && HasCall)
1094 // Iterate through all the unique predecessors and tail-duplicate this
1095 // block into them, if possible. Copying the list ahead of time also
1096 // avoids trouble with the predecessor list reallocating.
1097 bool Changed = false;
1098 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
1099 TailBB->pred_end());
1100 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
1101 PE = Preds.end(); PI != PE; ++PI) {
1102 MachineBasicBlock *PredBB = *PI;
1104 assert(TailBB != PredBB &&
1105 "Single-block loop should have been rejected earlier!");
1106 if (PredBB->succ_size() > 1) continue;
1108 MachineBasicBlock *PredTBB, *PredFBB;
1109 SmallVector<MachineOperand, 4> PredCond;
1110 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
1112 if (!PredCond.empty())
1114 // EH edges are ignored by AnalyzeBranch.
1115 if (PredBB->succ_size() != 1)
1117 // Don't duplicate into a fall-through predecessor (at least for now).
1118 if (PredBB->isLayoutSuccessor(TailBB) && CanFallThrough(PredBB))
1121 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB
1122 << "From Succ: " << *TailBB);
1124 // Remove PredBB's unconditional branch.
1125 TII->RemoveBranch(*PredBB);
1126 // Clone the contents of TailBB into PredBB.
1127 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end();
1129 MachineInstr *NewMI = MF.CloneMachineInstr(I);
1130 PredBB->insert(PredBB->end(), NewMI);
1134 PredBB->removeSuccessor(PredBB->succ_begin());
1135 assert(PredBB->succ_empty() &&
1136 "TailDuplicate called on block with multiple successors!");
1137 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
1138 E = TailBB->succ_end(); I != E; ++I)
1139 PredBB->addSuccessor(*I);
1144 // If TailBB was duplicated into all its predecessors except for the prior
1145 // block, which falls through unconditionally, move the contents of this
1146 // block into the prior block.
1147 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(TailBB));
1148 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1149 SmallVector<MachineOperand, 4> PriorCond;
1150 bool PriorUnAnalyzable =
1151 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1152 // This has to check PrevBB->succ_size() because EH edges are ignored by
1154 if (!PriorUnAnalyzable && PriorCond.empty() && !PriorTBB &&
1155 TailBB->pred_size() == 1 && PrevBB.succ_size() == 1 &&
1156 !TailBB->hasAddressTaken()) {
1157 DEBUG(errs() << "\nMerging into block: " << PrevBB
1158 << "From MBB: " << *TailBB);
1159 PrevBB.splice(PrevBB.end(), TailBB, TailBB->begin(), TailBB->end());
1160 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1161 assert(PrevBB.succ_empty());
1162 PrevBB.transferSuccessors(TailBB);
1169 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1170 /// block. This is never called on the entry block.
1171 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1172 bool MadeChange = false;
1173 MachineFunction &MF = *MBB->getParent();
1176 MachineFunction::iterator FallThrough = MBB;
1179 // If this block is empty, make everyone use its fall-through, not the block
1180 // explicitly. Landing pads should not do this since the landing-pad table
1181 // points to this block. Blocks with their addresses taken shouldn't be
1183 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1184 // Dead block? Leave for cleanup later.
1185 if (MBB->pred_empty()) return MadeChange;
1187 if (FallThrough == MF.end()) {
1188 // TODO: Simplify preds to not branch here if possible!
1190 // Rewrite all predecessors of the old block to go to the fallthrough
1192 while (!MBB->pred_empty()) {
1193 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1194 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1196 // If MBB was the target of a jump table, update jump tables to go to the
1197 // fallthrough instead.
1198 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough);
1204 // Check to see if we can simplify the terminator of the block before this
1206 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1208 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1209 SmallVector<MachineOperand, 4> PriorCond;
1210 bool PriorUnAnalyzable =
1211 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1212 if (!PriorUnAnalyzable) {
1213 // If the CFG for the prior block has extra edges, remove them.
1214 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1215 !PriorCond.empty());
1217 // If the previous branch is conditional and both conditions go to the same
1218 // destination, remove the branch, replacing it with an unconditional one or
1220 if (PriorTBB && PriorTBB == PriorFBB) {
1221 TII->RemoveBranch(PrevBB);
1223 if (PriorTBB != MBB)
1224 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1227 goto ReoptimizeBlock;
1230 // If the previous block unconditionally falls through to this block and
1231 // this block has no other predecessors, move the contents of this block
1232 // into the prior block. This doesn't usually happen when SimplifyCFG
1233 // has been used, but it can happen if tail duplication eliminates all the
1234 // non-branch predecessors of a block leaving only the fall-through edge.
1235 // This has to check PrevBB->succ_size() because EH edges are ignored by
1237 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1238 PrevBB.succ_size() == 1 &&
1239 !MBB->hasAddressTaken()) {
1240 DEBUG(errs() << "\nMerging into block: " << PrevBB
1241 << "From MBB: " << *MBB);
1242 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1243 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1244 assert(PrevBB.succ_empty());
1245 PrevBB.transferSuccessors(MBB);
1250 // If the previous branch *only* branches to *this* block (conditional or
1251 // not) remove the branch.
1252 if (PriorTBB == MBB && PriorFBB == 0) {
1253 TII->RemoveBranch(PrevBB);
1256 goto ReoptimizeBlock;
1259 // If the prior block branches somewhere else on the condition and here if
1260 // the condition is false, remove the uncond second branch.
1261 if (PriorFBB == MBB) {
1262 TII->RemoveBranch(PrevBB);
1263 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1266 goto ReoptimizeBlock;
1269 // If the prior block branches here on true and somewhere else on false, and
1270 // if the branch condition is reversible, reverse the branch to create a
1272 if (PriorTBB == MBB) {
1273 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1274 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1275 TII->RemoveBranch(PrevBB);
1276 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
1279 goto ReoptimizeBlock;
1283 // If this block has no successors (e.g. it is a return block or ends with
1284 // a call to a no-return function like abort or __cxa_throw) and if the pred
1285 // falls through into this block, and if it would otherwise fall through
1286 // into the block after this, move this block to the end of the function.
1288 // We consider it more likely that execution will stay in the function (e.g.
1289 // due to loops) than it is to exit it. This asserts in loops etc, moving
1290 // the assert condition out of the loop body.
1291 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1292 MachineFunction::iterator(PriorTBB) == FallThrough &&
1293 !CanFallThrough(MBB)) {
1294 bool DoTransform = true;
1296 // We have to be careful that the succs of PredBB aren't both no-successor
1297 // blocks. If neither have successors and if PredBB is the second from
1298 // last block in the function, we'd just keep swapping the two blocks for
1299 // last. Only do the swap if one is clearly better to fall through than
1301 if (FallThrough == --MF.end() &&
1302 !IsBetterFallthrough(PriorTBB, MBB))
1303 DoTransform = false;
1305 // We don't want to do this transformation if we have control flow like:
1314 // In this case, we could actually be moving the return block *into* a
1316 if (DoTransform && !MBB->succ_empty() &&
1317 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
1318 DoTransform = false;
1322 // Reverse the branch so we will fall through on the previous true cond.
1323 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1324 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1325 DEBUG(errs() << "\nMoving MBB: " << *MBB
1326 << "To make fallthrough to: " << *PriorTBB << "\n");
1328 TII->RemoveBranch(PrevBB);
1329 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1331 // Move this block to the end of the function.
1332 MBB->moveAfter(--MF.end());
1341 // Analyze the branch in the current block.
1342 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1343 SmallVector<MachineOperand, 4> CurCond;
1344 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1345 if (!CurUnAnalyzable) {
1346 // If the CFG for the prior block has extra edges, remove them.
1347 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1349 // If this is a two-way branch, and the FBB branches to this block, reverse
1350 // the condition so the single-basic-block loop is faster. Instead of:
1351 // Loop: xxx; jcc Out; jmp Loop
1353 // Loop: xxx; jncc Loop; jmp Out
1354 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1355 SmallVector<MachineOperand, 4> NewCond(CurCond);
1356 if (!TII->ReverseBranchCondition(NewCond)) {
1357 TII->RemoveBranch(*MBB);
1358 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1361 goto ReoptimizeBlock;
1365 // If this branch is the only thing in its block, see if we can forward
1366 // other blocks across it.
1367 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1368 MBB->begin()->getDesc().isBranch() && CurTBB != MBB &&
1369 !MBB->hasAddressTaken()) {
1370 // This block may contain just an unconditional branch. Because there can
1371 // be 'non-branch terminators' in the block, try removing the branch and
1372 // then seeing if the block is empty.
1373 TII->RemoveBranch(*MBB);
1375 // If this block is just an unconditional branch to CurTBB, we can
1376 // usually completely eliminate the block. The only case we cannot
1377 // completely eliminate the block is when the block before this one
1378 // falls through into MBB and we can't understand the prior block's branch
1381 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
1382 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1383 !PrevBB.isSuccessor(MBB)) {
1384 // If the prior block falls through into us, turn it into an
1385 // explicit branch to us to make updates simpler.
1386 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1387 PriorTBB != MBB && PriorFBB != MBB) {
1388 if (PriorTBB == 0) {
1389 assert(PriorCond.empty() && PriorFBB == 0 &&
1390 "Bad branch analysis");
1393 assert(PriorFBB == 0 && "Machine CFG out of date!");
1396 TII->RemoveBranch(PrevBB);
1397 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1400 // Iterate through all the predecessors, revectoring each in-turn.
1402 bool DidChange = false;
1403 bool HasBranchToSelf = false;
1404 while(PI != MBB->pred_size()) {
1405 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1407 // If this block has an uncond branch to itself, leave it.
1409 HasBranchToSelf = true;
1412 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1413 // If this change resulted in PMBB ending in a conditional
1414 // branch where both conditions go to the same destination,
1415 // change this to an unconditional branch (and fix the CFG).
1416 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1417 SmallVector<MachineOperand, 4> NewCurCond;
1418 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1419 NewCurFBB, NewCurCond, true);
1420 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1421 TII->RemoveBranch(*PMBB);
1423 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1426 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1431 // Change any jumptables to go to the new MBB.
1432 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB);
1436 if (!HasBranchToSelf) return MadeChange;
1441 // Add the branch back if the block is more than just an uncond branch.
1442 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1446 // If the prior block doesn't fall through into this block, and if this
1447 // block doesn't fall through into some other block, see if we can find a
1448 // place to move this block where a fall-through will happen.
1449 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
1450 PriorTBB, PriorFBB, PriorCond)) {
1452 // Now we know that there was no fall-through into this block, check to
1453 // see if it has a fall-through into its successor.
1454 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1457 if (!MBB->isLandingPad()) {
1458 // Check all the predecessors of this block. If one of them has no fall
1459 // throughs, move this block right after it.
1460 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1461 E = MBB->pred_end(); PI != E; ++PI) {
1462 // Analyze the branch at the end of the pred.
1463 MachineBasicBlock *PredBB = *PI;
1464 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1465 MachineBasicBlock *PredTBB, *PredFBB;
1466 SmallVector<MachineOperand, 4> PredCond;
1467 if (PredBB != MBB && !CanFallThrough(PredBB) &&
1468 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1469 && (!CurFallsThru || !CurTBB || !CurFBB)
1470 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1471 // If the current block doesn't fall through, just move it.
1472 // If the current block can fall through and does not end with a
1473 // conditional branch, we need to append an unconditional jump to
1474 // the (current) next block. To avoid a possible compile-time
1475 // infinite loop, move blocks only backward in this case.
1476 // Also, if there are already 2 branches here, we cannot add a third;
1477 // this means we have the case
1482 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1484 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1486 MBB->moveAfter(PredBB);
1488 goto ReoptimizeBlock;
1493 if (!CurFallsThru) {
1494 // Check all successors to see if we can move this block before it.
1495 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1496 E = MBB->succ_end(); SI != E; ++SI) {
1497 // Analyze the branch at the end of the block before the succ.
1498 MachineBasicBlock *SuccBB = *SI;
1499 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1501 // If this block doesn't already fall-through to that successor, and if
1502 // the succ doesn't already have a block that can fall through into it,
1503 // and if the successor isn't an EH destination, we can arrange for the
1504 // fallthrough to happen.
1505 if (SuccBB != MBB && &*SuccPrev != MBB &&
1506 !CanFallThrough(SuccPrev) && !CurUnAnalyzable &&
1507 !SuccBB->isLandingPad()) {
1508 MBB->moveBefore(SuccBB);
1510 goto ReoptimizeBlock;
1514 // Okay, there is no really great place to put this block. If, however,
1515 // the block before this one would be a fall-through if this block were
1516 // removed, move this block to the end of the function.
1517 MachineBasicBlock *PrevTBB, *PrevFBB;
1518 SmallVector<MachineOperand, 4> PrevCond;
1519 if (FallThrough != MF.end() &&
1520 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1521 PrevBB.isSuccessor(FallThrough)) {
1522 MBB->moveAfter(--MF.end());