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);
201 bool MadeChangeThisIteration = true;
202 while (MadeChangeThisIteration) {
203 MadeChangeThisIteration = false;
204 MadeChangeThisIteration |= TailMergeBlocks(MF);
205 MadeChangeThisIteration |= OptimizeBranches(MF);
206 MadeChange |= MadeChangeThisIteration;
209 // Do tail duplication once after tail merging is done. Otherwise it is
210 // tough to avoid situations where tail duplication and tail merging undo
211 // each other's transformations ad infinitum.
212 MadeChange |= TailDuplicateBlocks(MF);
214 // See if any jump tables have become mergable or dead as the code generator
216 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
217 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
219 // Figure out how these jump tables should be merged.
220 std::vector<unsigned> JTMapping;
221 JTMapping.reserve(JTs.size());
223 // We always keep the 0th jump table.
224 JTMapping.push_back(0);
226 // Scan the jump tables, seeing if there are any duplicates. Note that this
227 // is N^2, which should be fixed someday.
228 for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
229 if (JTs[i].MBBs.empty())
230 JTMapping.push_back(i);
232 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
235 // If a jump table was merge with another one, walk the function rewriting
236 // references to jump tables to reference the new JT ID's. Keep track of
237 // whether we see a jump table idx, if not, we can delete the JT.
238 BitVector JTIsLive(JTs.size());
239 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
241 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
243 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
244 MachineOperand &Op = I->getOperand(op);
245 if (!Op.isJTI()) continue;
246 unsigned NewIdx = JTMapping[Op.getIndex()];
249 // Remember that this JT is live.
250 JTIsLive.set(NewIdx);
254 // Finally, remove dead jump tables. This happens either because the
255 // indirect jump was unreachable (and thus deleted) or because the jump
256 // table was merged with some other one.
257 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
258 if (!JTIsLive.test(i)) {
259 JTI->RemoveJumpTable(i);
268 //===----------------------------------------------------------------------===//
269 // Tail Merging of Blocks
270 //===----------------------------------------------------------------------===//
272 /// HashMachineInstr - Compute a hash value for MI and its operands.
273 static unsigned HashMachineInstr(const MachineInstr *MI) {
274 unsigned Hash = MI->getOpcode();
275 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
276 const MachineOperand &Op = MI->getOperand(i);
278 // Merge in bits from the operand if easy.
279 unsigned OperandHash = 0;
280 switch (Op.getType()) {
281 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
282 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
283 case MachineOperand::MO_MachineBasicBlock:
284 OperandHash = Op.getMBB()->getNumber();
286 case MachineOperand::MO_FrameIndex:
287 case MachineOperand::MO_ConstantPoolIndex:
288 case MachineOperand::MO_JumpTableIndex:
289 OperandHash = Op.getIndex();
291 case MachineOperand::MO_GlobalAddress:
292 case MachineOperand::MO_ExternalSymbol:
293 // Global address / external symbol are too hard, don't bother, but do
294 // pull in the offset.
295 OperandHash = Op.getOffset();
300 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
305 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
306 /// with no successors, we hash two instructions, because cross-jumping
307 /// only saves code when at least two instructions are removed (since a
308 /// branch must be inserted). For blocks with a successor, one of the
309 /// two blocks to be tail-merged will end with a branch already, so
310 /// it gains to cross-jump even for one instruction.
311 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
312 unsigned minCommonTailLength) {
313 MachineBasicBlock::const_iterator I = MBB->end();
314 if (I == MBB->begin())
315 return 0; // Empty MBB.
318 unsigned Hash = HashMachineInstr(I);
320 if (I == MBB->begin() || minCommonTailLength == 1)
321 return Hash; // Single instr MBB.
324 // Hash in the second-to-last instruction.
325 Hash ^= HashMachineInstr(I) << 2;
329 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
330 /// of instructions they actually have in common together at their end. Return
331 /// iterators for the first shared instruction in each block.
332 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
333 MachineBasicBlock *MBB2,
334 MachineBasicBlock::iterator &I1,
335 MachineBasicBlock::iterator &I2) {
339 unsigned TailLen = 0;
340 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
342 if (!I1->isIdenticalTo(I2) ||
343 // FIXME: This check is dubious. It's used to get around a problem where
344 // people incorrectly expect inline asm directives to remain in the same
345 // relative order. This is untenable because normal compiler
346 // optimizations (like this one) may reorder and/or merge these
348 I1->getOpcode() == TargetInstrInfo::INLINEASM) {
357 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
358 /// after it, replacing it with an unconditional branch to NewDest. This
359 /// returns true if OldInst's block is modified, false if NewDest is modified.
360 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
361 MachineBasicBlock *NewDest) {
362 MachineBasicBlock *OldBB = OldInst->getParent();
364 // Remove all the old successors of OldBB from the CFG.
365 while (!OldBB->succ_empty())
366 OldBB->removeSuccessor(OldBB->succ_begin());
368 // Remove all the dead instructions from the end of OldBB.
369 OldBB->erase(OldInst, OldBB->end());
371 // If OldBB isn't immediately before OldBB, insert a branch to it.
372 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
373 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
374 OldBB->addSuccessor(NewDest);
378 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
379 /// MBB so that the part before the iterator falls into the part starting at the
380 /// iterator. This returns the new MBB.
381 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
382 MachineBasicBlock::iterator BBI1) {
383 MachineFunction &MF = *CurMBB.getParent();
385 // Create the fall-through block.
386 MachineFunction::iterator MBBI = &CurMBB;
387 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
388 CurMBB.getParent()->insert(++MBBI, NewMBB);
390 // Move all the successors of this block to the specified block.
391 NewMBB->transferSuccessors(&CurMBB);
393 // Add an edge from CurMBB to NewMBB for the fall-through.
394 CurMBB.addSuccessor(NewMBB);
396 // Splice the code over.
397 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
399 // For targets that use the register scavenger, we must maintain LiveIns.
401 RS->enterBasicBlock(&CurMBB);
403 RS->forward(prior(CurMBB.end()));
404 BitVector RegsLiveAtExit(TRI->getNumRegs());
405 RS->getRegsUsed(RegsLiveAtExit, false);
406 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
407 if (RegsLiveAtExit[i])
408 NewMBB->addLiveIn(i);
414 /// EstimateRuntime - Make a rough estimate for how long it will take to run
415 /// the specified code.
416 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
417 MachineBasicBlock::iterator E) {
419 for (; I != E; ++I) {
420 const TargetInstrDesc &TID = I->getDesc();
423 else if (TID.mayLoad() || TID.mayStore())
431 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
432 // branches temporarily for tail merging). In the case where CurMBB ends
433 // with a conditional branch to the next block, optimize by reversing the
434 // test and conditionally branching to SuccMBB instead.
435 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
436 const TargetInstrInfo *TII) {
437 MachineFunction *MF = CurMBB->getParent();
438 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
439 MachineBasicBlock *TBB = 0, *FBB = 0;
440 SmallVector<MachineOperand, 4> Cond;
441 if (I != MF->end() &&
442 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
443 MachineBasicBlock *NextBB = I;
444 if (TBB == NextBB && !Cond.empty() && !FBB) {
445 if (!TII->ReverseBranchCondition(Cond)) {
446 TII->RemoveBranch(*CurMBB);
447 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
452 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
456 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
457 if (getHash() < o.getHash())
459 else if (getHash() > o.getHash())
461 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
463 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
466 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
467 // an object with itself.
468 #ifndef _GLIBCXX_DEBUG
469 llvm_unreachable("Predecessor appears twice");
475 /// CountTerminators - Count the number of terminators in the given
476 /// block and set I to the position of the first non-terminator, if there
477 /// is one, or MBB->end() otherwise.
478 static unsigned CountTerminators(MachineBasicBlock *MBB,
479 MachineBasicBlock::iterator &I) {
481 unsigned NumTerms = 0;
483 if (I == MBB->begin()) {
488 if (!I->getDesc().isTerminator()) break;
494 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
495 /// and decide if it would be profitable to merge those tails. Return the
496 /// length of the common tail and iterators to the first common instruction
498 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
499 MachineBasicBlock *MBB2,
500 unsigned minCommonTailLength,
501 unsigned &CommonTailLen,
502 MachineBasicBlock::iterator &I1,
503 MachineBasicBlock::iterator &I2,
504 MachineBasicBlock *SuccBB,
505 MachineBasicBlock *PredBB) {
506 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
507 MachineFunction *MF = MBB1->getParent();
509 if (CommonTailLen == 0)
512 // It's almost always profitable to merge any number of non-terminator
513 // instructions with the block that falls through into the common successor.
514 if (MBB1 == PredBB || MBB2 == PredBB) {
515 MachineBasicBlock::iterator I;
516 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
517 if (CommonTailLen > NumTerms)
521 // If one of the blocks can be completely merged and happens to be in
522 // a position where the other could fall through into it, merge any number
523 // of instructions, because it can be done without a branch.
524 // TODO: If the blocks are not adjacent, move one of them so that they are?
525 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
527 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
530 // If both blocks have an unconditional branch temporarily stripped out,
531 // count that as an additional common instruction for the following
533 unsigned EffectiveTailLen = CommonTailLen;
534 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
535 !MBB1->back().getDesc().isBarrier() &&
536 !MBB2->back().getDesc().isBarrier())
539 // Check if the common tail is long enough to be worthwhile.
540 if (EffectiveTailLen >= minCommonTailLength)
543 // If we are optimizing for code size, 2 instructions in common is enough if
544 // we don't have to split a block. At worst we will be introducing 1 new
545 // branch instruction, which is likely to be smaller than the 2
546 // instructions that would be deleted in the merge.
547 if (EffectiveTailLen >= 2 &&
548 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
549 (I1 == MBB1->begin() || I2 == MBB2->begin()))
555 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
556 /// hash CurHash (guaranteed to match the last element). Build the vector
557 /// SameTails of all those that have the (same) largest number of instructions
558 /// in common of any pair of these blocks. SameTails entries contain an
559 /// iterator into MergePotentials (from which the MachineBasicBlock can be
560 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
561 /// instruction where the matching code sequence begins.
562 /// Order of elements in SameTails is the reverse of the order in which
563 /// those blocks appear in MergePotentials (where they are not necessarily
565 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
566 unsigned minCommonTailLength,
567 MachineBasicBlock *SuccBB,
568 MachineBasicBlock *PredBB) {
569 unsigned maxCommonTailLength = 0U;
571 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
572 MPIterator HighestMPIter = prior(MergePotentials.end());
573 for (MPIterator CurMPIter = prior(MergePotentials.end()),
574 B = MergePotentials.begin();
575 CurMPIter != B && CurMPIter->getHash() == CurHash;
577 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
578 unsigned CommonTailLen;
579 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
581 CommonTailLen, TrialBBI1, TrialBBI2,
583 if (CommonTailLen > maxCommonTailLength) {
585 maxCommonTailLength = CommonTailLen;
586 HighestMPIter = CurMPIter;
587 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
589 if (HighestMPIter == CurMPIter &&
590 CommonTailLen == maxCommonTailLength)
591 SameTails.push_back(SameTailElt(I, TrialBBI2));
597 return maxCommonTailLength;
600 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
601 /// MergePotentials, restoring branches at ends of blocks as appropriate.
602 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
603 MachineBasicBlock *SuccBB,
604 MachineBasicBlock *PredBB) {
605 MPIterator CurMPIter, B;
606 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
607 CurMPIter->getHash() == CurHash;
609 // Put the unconditional branch back, if we need one.
610 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
611 if (SuccBB && CurMBB != PredBB)
612 FixTail(CurMBB, SuccBB, TII);
616 if (CurMPIter->getHash() != CurHash)
618 MergePotentials.erase(CurMPIter, MergePotentials.end());
621 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
622 /// only of the common tail. Create a block that does by splitting one.
623 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
624 unsigned maxCommonTailLength) {
625 unsigned commonTailIndex = 0;
626 unsigned TimeEstimate = ~0U;
627 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
628 // Use PredBB if possible; that doesn't require a new branch.
629 if (SameTails[i].getBlock() == PredBB) {
633 // Otherwise, make a (fairly bogus) choice based on estimate of
634 // how long it will take the various blocks to execute.
635 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
636 SameTails[i].getTailStartPos());
637 if (t <= TimeEstimate) {
643 MachineBasicBlock::iterator BBI =
644 SameTails[commonTailIndex].getTailStartPos();
645 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
647 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
648 << maxCommonTailLength);
650 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
651 SameTails[commonTailIndex].setBlock(newMBB);
652 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
654 // If we split PredBB, newMBB is the new predecessor.
658 return commonTailIndex;
661 // See if any of the blocks in MergePotentials (which all have a common single
662 // successor, or all have no successor) can be tail-merged. If there is a
663 // successor, any blocks in MergePotentials that are not tail-merged and
664 // are not immediately before Succ must have an unconditional branch to
665 // Succ added (but the predecessor/successor lists need no adjustment).
666 // The lone predecessor of Succ that falls through into Succ,
667 // if any, is given in PredBB.
669 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
670 MachineBasicBlock *PredBB) {
671 bool MadeChange = false;
673 // Except for the special cases below, tail-merge if there are at least
674 // this many instructions in common.
675 unsigned minCommonTailLength = TailMergeSize;
677 DEBUG(errs() << "\nTryTailMergeBlocks: ";
678 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
679 errs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
680 << (i == e-1 ? "" : ", ");
683 errs() << " with successor BB#" << SuccBB->getNumber() << '\n';
685 errs() << " which has fall-through from BB#"
686 << PredBB->getNumber() << "\n";
688 errs() << "Looking for common tails of at least "
689 << minCommonTailLength << " instruction"
690 << (minCommonTailLength == 1 ? "" : "s") << '\n';
693 // Sort by hash value so that blocks with identical end sequences sort
695 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
697 // Walk through equivalence sets looking for actual exact matches.
698 while (MergePotentials.size() > 1) {
699 unsigned CurHash = MergePotentials.back().getHash();
701 // Build SameTails, identifying the set of blocks with this hash code
702 // and with the maximum number of instructions in common.
703 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
707 // If we didn't find any pair that has at least minCommonTailLength
708 // instructions in common, remove all blocks with this hash code and retry.
709 if (SameTails.empty()) {
710 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
714 // If one of the blocks is the entire common tail (and not the entry
715 // block, which we can't jump to), we can treat all blocks with this same
716 // tail at once. Use PredBB if that is one of the possibilities, as that
717 // will not introduce any extra branches.
718 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
719 getParent()->begin();
720 unsigned commonTailIndex = SameTails.size();
721 // If there are two blocks, check to see if one can be made to fall through
723 if (SameTails.size() == 2 &&
724 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
725 SameTails[1].tailIsWholeBlock())
727 else if (SameTails.size() == 2 &&
728 SameTails[1].getBlock()->isLayoutSuccessor(
729 SameTails[0].getBlock()) &&
730 SameTails[0].tailIsWholeBlock())
733 // Otherwise just pick one, favoring the fall-through predecessor if
735 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
736 MachineBasicBlock *MBB = SameTails[i].getBlock();
737 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
743 if (SameTails[i].tailIsWholeBlock())
748 if (commonTailIndex == SameTails.size() ||
749 (SameTails[commonTailIndex].getBlock() == PredBB &&
750 !SameTails[commonTailIndex].tailIsWholeBlock())) {
751 // None of the blocks consist entirely of the common tail.
752 // Split a block so that one does.
753 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
756 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
757 // MBB is common tail. Adjust all other BB's to jump to this one.
758 // Traversal must be forwards so erases work.
759 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber()
761 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
762 if (commonTailIndex == i)
764 DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber()
765 << (i == e-1 ? "" : ", "));
766 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
767 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
768 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
769 MergePotentials.erase(SameTails[i].getMPIter());
771 DEBUG(errs() << "\n");
772 // We leave commonTailIndex in the worklist in case there are other blocks
773 // that match it with a smaller number of instructions.
779 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
781 if (!EnableTailMerge) return false;
783 bool MadeChange = false;
785 // First find blocks with no successors.
786 MergePotentials.clear();
787 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
789 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
792 // See if we can do any tail merging on those.
793 if (MergePotentials.size() < TailMergeThreshold &&
794 MergePotentials.size() >= 2)
795 MadeChange |= TryTailMergeBlocks(NULL, NULL);
797 // Look at blocks (IBB) with multiple predecessors (PBB).
798 // We change each predecessor to a canonical form, by
799 // (1) temporarily removing any unconditional branch from the predecessor
801 // (2) alter conditional branches so they branch to the other block
802 // not IBB; this may require adding back an unconditional branch to IBB
803 // later, where there wasn't one coming in. E.g.
805 // fallthrough to QBB
808 // with a conceptual B to IBB after that, which never actually exists.
809 // With those changes, we see whether the predecessors' tails match,
810 // and merge them if so. We change things out of canonical form and
811 // back to the way they were later in the process. (OptimizeBranches
812 // would undo some of this, but we can't use it, because we'd get into
813 // a compile-time infinite loop repeatedly doing and undoing the same
816 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end();
818 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
819 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
820 MachineBasicBlock *IBB = I;
821 MachineBasicBlock *PredBB = prior(I);
822 MergePotentials.clear();
823 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
826 MachineBasicBlock *PBB = *P;
827 // Skip blocks that loop to themselves, can't tail merge these.
830 // Visit each predecessor only once.
831 if (!UniquePreds.insert(PBB))
833 MachineBasicBlock *TBB = 0, *FBB = 0;
834 SmallVector<MachineOperand, 4> Cond;
835 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
836 // Failing case: IBB is the target of a cbr, and
837 // we cannot reverse the branch.
838 SmallVector<MachineOperand, 4> NewCond(Cond);
839 if (!Cond.empty() && TBB == IBB) {
840 if (TII->ReverseBranchCondition(NewCond))
842 // This is the QBB case described above
844 FBB = next(MachineFunction::iterator(PBB));
846 // Failing case: the only way IBB can be reached from PBB is via
847 // exception handling. Happens for landing pads. Would be nice
848 // to have a bit in the edge so we didn't have to do all this.
849 if (IBB->isLandingPad()) {
850 MachineFunction::iterator IP = PBB; IP++;
851 MachineBasicBlock *PredNextBB = NULL;
855 if (IBB != PredNextBB) // fallthrough
858 if (TBB != IBB && FBB != IBB) // cbr then ubr
860 } else if (Cond.empty()) {
861 if (TBB != IBB) // ubr
864 if (TBB != IBB && IBB != PredNextBB) // cbr
868 // Remove the unconditional branch at the end, if any.
869 if (TBB && (Cond.empty() || FBB)) {
870 TII->RemoveBranch(*PBB);
872 // reinsert conditional branch only, for now
873 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
875 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
879 if (MergePotentials.size() >= 2)
880 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
881 // Reinsert an unconditional branch if needed.
882 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
883 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
884 if (MergePotentials.size() == 1 &&
885 MergePotentials.begin()->getBlock() != PredBB)
886 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
892 //===----------------------------------------------------------------------===//
893 // Branch Optimization
894 //===----------------------------------------------------------------------===//
896 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
897 bool MadeChange = false;
899 // Make sure blocks are numbered in order
902 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
903 MachineBasicBlock *MBB = I++;
904 MadeChange |= OptimizeBlock(MBB);
906 // If it is dead, remove it.
907 if (MBB->pred_empty()) {
908 RemoveDeadBlock(MBB);
917 /// CanFallThrough - Return true if the specified block (with the specified
918 /// branch condition) can implicitly transfer control to the block after it by
919 /// falling off the end of it. This should return false if it can reach the
920 /// block after it, but it uses an explicit branch to do so (e.g. a table jump).
922 /// True is a conservative answer.
924 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
925 bool BranchUnAnalyzable,
926 MachineBasicBlock *TBB,
927 MachineBasicBlock *FBB,
928 const SmallVectorImpl<MachineOperand> &Cond) {
929 MachineFunction::iterator Fallthrough = CurBB;
931 // If FallthroughBlock is off the end of the function, it can't fall through.
932 if (Fallthrough == CurBB->getParent()->end())
935 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
936 if (!CurBB->isSuccessor(Fallthrough))
939 // If we couldn't analyze the branch, examine the last instruction.
940 // If the block doesn't end in a known control barrier, assume fallthrough
941 // is possible. The isPredicable check is needed because this code can be
942 // called during IfConversion, where an instruction which is normally a
943 // Barrier is predicated and thus no longer an actual control barrier. This
944 // is over-conservative though, because if an instruction isn't actually
945 // predicated we could still treat it like a barrier.
946 if (BranchUnAnalyzable)
947 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() ||
948 CurBB->back().getDesc().isPredicable();
950 // If there is no branch, control always falls through.
951 if (TBB == 0) return true;
953 // If there is some explicit branch to the fallthrough block, it can obviously
954 // reach, even though the branch should get folded to fall through implicitly.
955 if (MachineFunction::iterator(TBB) == Fallthrough ||
956 MachineFunction::iterator(FBB) == Fallthrough)
959 // If it's an unconditional branch to some block not the fall through, it
960 // doesn't fall through.
961 if (Cond.empty()) return false;
963 // Otherwise, if it is conditional and has no explicit false block, it falls
968 /// CanFallThrough - Return true if the specified can implicitly transfer
969 /// control to the block after it by falling off the end of it. This should
970 /// return false if it can reach the block after it, but it uses an explicit
971 /// branch to do so (e.g. a table jump).
973 /// True is a conservative answer.
975 bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
976 MachineBasicBlock *TBB = 0, *FBB = 0;
977 SmallVector<MachineOperand, 4> Cond;
978 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true);
979 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
982 /// IsBetterFallthrough - Return true if it would be clearly better to
983 /// fall-through to MBB1 than to fall through into MBB2. This has to return
984 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
985 /// result in infinite loops.
986 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
987 MachineBasicBlock *MBB2) {
988 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
989 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
990 // optimize branches that branch to either a return block or an assert block
991 // into a fallthrough to the return.
992 if (MBB1->empty() || MBB2->empty()) return false;
994 // If there is a clear successor ordering we make sure that one block
995 // will fall through to the next
996 if (MBB1->isSuccessor(MBB2)) return true;
997 if (MBB2->isSuccessor(MBB1)) return false;
999 MachineInstr *MBB1I = --MBB1->end();
1000 MachineInstr *MBB2I = --MBB2->end();
1001 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
1004 /// TailDuplicateBlocks - Look for small blocks that are unconditionally
1005 /// branched to and do not fall through. Tail-duplicate their instructions
1006 /// into their predecessors to eliminate (dynamic) branches.
1007 bool BranchFolder::TailDuplicateBlocks(MachineFunction &MF) {
1008 bool MadeChange = false;
1010 // Make sure blocks are numbered in order
1011 MF.RenumberBlocks();
1013 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
1014 MachineBasicBlock *MBB = I++;
1016 // Only duplicate blocks that end with unconditional branches.
1017 if (CanFallThrough(MBB))
1020 MadeChange |= TailDuplicate(MBB, MF);
1022 // If it is dead, remove it.
1023 if (MBB->pred_empty()) {
1024 NumInstrDups -= MBB->size();
1025 RemoveDeadBlock(MBB);
1033 /// TailDuplicate - If it is profitable, duplicate TailBB's contents in each
1034 /// of its predecessors.
1035 bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB,
1036 MachineFunction &MF) {
1037 // Don't try to tail-duplicate single-block loops.
1038 if (TailBB->isSuccessor(TailBB))
1041 // Set the limit on the number of instructions to duplicate, with a default
1042 // of one less than the tail-merge threshold. When optimizing for size,
1043 // duplicate only one, because one branch instruction can be eliminated to
1044 // compensate for the duplication.
1045 unsigned MaxDuplicateCount =
1046 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ?
1047 1 : TII->TailDuplicationLimit(*TailBB, TailMergeSize - 1);
1049 // Check the instructions in the block to determine whether tail-duplication
1050 // is invalid or unlikely to be profitable.
1052 bool HasCall = false;
1053 for (MachineBasicBlock::iterator I = TailBB->begin();
1054 I != TailBB->end(); ++I, ++i) {
1055 // Non-duplicable things shouldn't be tail-duplicated.
1056 if (I->getDesc().isNotDuplicable()) return false;
1057 // Don't duplicate more than the threshold.
1058 if (i == MaxDuplicateCount) return false;
1059 // Remember if we saw a call.
1060 if (I->getDesc().isCall()) HasCall = true;
1062 // Heuristically, don't tail-duplicate calls if it would expand code size,
1063 // as it's less likely to be worth the extra cost.
1064 if (i > 1 && HasCall)
1067 // Iterate through all the unique predecessors and tail-duplicate this
1068 // block into them, if possible. Copying the list ahead of time also
1069 // avoids trouble with the predecessor list reallocating.
1070 bool Changed = false;
1071 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
1072 TailBB->pred_end());
1073 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
1074 PE = Preds.end(); PI != PE; ++PI) {
1075 MachineBasicBlock *PredBB = *PI;
1077 assert(TailBB != PredBB &&
1078 "Single-block loop should have been rejected earlier!");
1079 if (PredBB->succ_size() > 1) continue;
1081 MachineBasicBlock *PredTBB, *PredFBB;
1082 SmallVector<MachineOperand, 4> PredCond;
1083 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
1085 if (!PredCond.empty())
1087 // EH edges are ignored by AnalyzeBranch.
1088 if (PredBB->succ_size() != 1)
1090 // Don't duplicate into a fall-through predecessor (at least for now).
1091 if (PredBB->isLayoutSuccessor(TailBB) && CanFallThrough(PredBB))
1094 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB
1095 << "From Succ: " << *TailBB);
1097 // Remove PredBB's unconditional branch.
1098 TII->RemoveBranch(*PredBB);
1099 // Clone the contents of TailBB into PredBB.
1100 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end();
1102 MachineInstr *NewMI = MF.CloneMachineInstr(I);
1103 PredBB->insert(PredBB->end(), NewMI);
1105 NumInstrDups += TailBB->size() - 1; // subtract one for removed branch
1108 PredBB->removeSuccessor(PredBB->succ_begin());
1109 assert(PredBB->succ_empty() &&
1110 "TailDuplicate called on block with multiple successors!");
1111 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
1112 E = TailBB->succ_end(); I != E; ++I)
1113 PredBB->addSuccessor(*I);
1119 // If TailBB was duplicated into all its predecessors except for the prior
1120 // block, which falls through unconditionally, move the contents of this
1121 // block into the prior block.
1122 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(TailBB));
1123 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1124 SmallVector<MachineOperand, 4> PriorCond;
1125 bool PriorUnAnalyzable =
1126 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1127 // This has to check PrevBB->succ_size() because EH edges are ignored by
1129 if (!PriorUnAnalyzable && PriorCond.empty() && !PriorTBB &&
1130 TailBB->pred_size() == 1 && PrevBB.succ_size() == 1 &&
1131 !TailBB->hasAddressTaken()) {
1132 DEBUG(errs() << "\nMerging into block: " << PrevBB
1133 << "From MBB: " << *TailBB);
1134 PrevBB.splice(PrevBB.end(), TailBB, TailBB->begin(), TailBB->end());
1135 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1136 assert(PrevBB.succ_empty());
1137 PrevBB.transferSuccessors(TailBB);
1144 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1145 /// block. This is never called on the entry block.
1146 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1147 bool MadeChange = false;
1148 MachineFunction &MF = *MBB->getParent();
1151 MachineFunction::iterator FallThrough = MBB;
1154 // If this block is empty, make everyone use its fall-through, not the block
1155 // explicitly. Landing pads should not do this since the landing-pad table
1156 // points to this block. Blocks with their addresses taken shouldn't be
1158 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1159 // Dead block? Leave for cleanup later.
1160 if (MBB->pred_empty()) return MadeChange;
1162 if (FallThrough == MF.end()) {
1163 // TODO: Simplify preds to not branch here if possible!
1165 // Rewrite all predecessors of the old block to go to the fallthrough
1167 while (!MBB->pred_empty()) {
1168 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1169 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1171 // If MBB was the target of a jump table, update jump tables to go to the
1172 // fallthrough instead.
1173 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough);
1179 // Check to see if we can simplify the terminator of the block before this
1181 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1183 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1184 SmallVector<MachineOperand, 4> PriorCond;
1185 bool PriorUnAnalyzable =
1186 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1187 if (!PriorUnAnalyzable) {
1188 // If the CFG for the prior block has extra edges, remove them.
1189 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1190 !PriorCond.empty());
1192 // If the previous branch is conditional and both conditions go to the same
1193 // destination, remove the branch, replacing it with an unconditional one or
1195 if (PriorTBB && PriorTBB == PriorFBB) {
1196 TII->RemoveBranch(PrevBB);
1198 if (PriorTBB != MBB)
1199 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1202 goto ReoptimizeBlock;
1205 // If the previous block unconditionally falls through to this block and
1206 // this block has no other predecessors, move the contents of this block
1207 // into the prior block. This doesn't usually happen when SimplifyCFG
1208 // has been used, but it can happen if tail merging splits a fall-through
1209 // predecessor of a block.
1210 // This has to check PrevBB->succ_size() because EH edges are ignored by
1212 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1213 PrevBB.succ_size() == 1 &&
1214 !MBB->hasAddressTaken()) {
1215 DEBUG(errs() << "\nMerging into block: " << PrevBB
1216 << "From MBB: " << *MBB);
1217 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1218 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1219 assert(PrevBB.succ_empty());
1220 PrevBB.transferSuccessors(MBB);
1225 // If the previous branch *only* branches to *this* block (conditional or
1226 // not) remove the branch.
1227 if (PriorTBB == MBB && PriorFBB == 0) {
1228 TII->RemoveBranch(PrevBB);
1231 goto ReoptimizeBlock;
1234 // If the prior block branches somewhere else on the condition and here if
1235 // the condition is false, remove the uncond second branch.
1236 if (PriorFBB == MBB) {
1237 TII->RemoveBranch(PrevBB);
1238 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1241 goto ReoptimizeBlock;
1244 // If the prior block branches here on true and somewhere else on false, and
1245 // if the branch condition is reversible, reverse the branch to create a
1247 if (PriorTBB == MBB) {
1248 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1249 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1250 TII->RemoveBranch(PrevBB);
1251 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
1254 goto ReoptimizeBlock;
1258 // If this block has no successors (e.g. it is a return block or ends with
1259 // a call to a no-return function like abort or __cxa_throw) and if the pred
1260 // falls through into this block, and if it would otherwise fall through
1261 // into the block after this, move this block to the end of the function.
1263 // We consider it more likely that execution will stay in the function (e.g.
1264 // due to loops) than it is to exit it. This asserts in loops etc, moving
1265 // the assert condition out of the loop body.
1266 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1267 MachineFunction::iterator(PriorTBB) == FallThrough &&
1268 !CanFallThrough(MBB)) {
1269 bool DoTransform = true;
1271 // We have to be careful that the succs of PredBB aren't both no-successor
1272 // blocks. If neither have successors and if PredBB is the second from
1273 // last block in the function, we'd just keep swapping the two blocks for
1274 // last. Only do the swap if one is clearly better to fall through than
1276 if (FallThrough == --MF.end() &&
1277 !IsBetterFallthrough(PriorTBB, MBB))
1278 DoTransform = false;
1280 // We don't want to do this transformation if we have control flow like:
1289 // In this case, we could actually be moving the return block *into* a
1291 if (DoTransform && !MBB->succ_empty() &&
1292 (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
1293 DoTransform = false;
1297 // Reverse the branch so we will fall through on the previous true cond.
1298 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1299 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1300 DEBUG(errs() << "\nMoving MBB: " << *MBB
1301 << "To make fallthrough to: " << *PriorTBB << "\n");
1303 TII->RemoveBranch(PrevBB);
1304 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1306 // Move this block to the end of the function.
1307 MBB->moveAfter(--MF.end());
1316 // Analyze the branch in the current block.
1317 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1318 SmallVector<MachineOperand, 4> CurCond;
1319 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1320 if (!CurUnAnalyzable) {
1321 // If the CFG for the prior block has extra edges, remove them.
1322 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1324 // If this is a two-way branch, and the FBB branches to this block, reverse
1325 // the condition so the single-basic-block loop is faster. Instead of:
1326 // Loop: xxx; jcc Out; jmp Loop
1328 // Loop: xxx; jncc Loop; jmp Out
1329 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1330 SmallVector<MachineOperand, 4> NewCond(CurCond);
1331 if (!TII->ReverseBranchCondition(NewCond)) {
1332 TII->RemoveBranch(*MBB);
1333 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1336 goto ReoptimizeBlock;
1340 // If this branch is the only thing in its block, see if we can forward
1341 // other blocks across it.
1342 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1343 MBB->begin()->getDesc().isBranch() && CurTBB != MBB &&
1344 !MBB->hasAddressTaken()) {
1345 // This block may contain just an unconditional branch. Because there can
1346 // be 'non-branch terminators' in the block, try removing the branch and
1347 // then seeing if the block is empty.
1348 TII->RemoveBranch(*MBB);
1350 // If this block is just an unconditional branch to CurTBB, we can
1351 // usually completely eliminate the block. The only case we cannot
1352 // completely eliminate the block is when the block before this one
1353 // falls through into MBB and we can't understand the prior block's branch
1356 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
1357 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1358 !PrevBB.isSuccessor(MBB)) {
1359 // If the prior block falls through into us, turn it into an
1360 // explicit branch to us to make updates simpler.
1361 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1362 PriorTBB != MBB && PriorFBB != MBB) {
1363 if (PriorTBB == 0) {
1364 assert(PriorCond.empty() && PriorFBB == 0 &&
1365 "Bad branch analysis");
1368 assert(PriorFBB == 0 && "Machine CFG out of date!");
1371 TII->RemoveBranch(PrevBB);
1372 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1375 // Iterate through all the predecessors, revectoring each in-turn.
1377 bool DidChange = false;
1378 bool HasBranchToSelf = false;
1379 while(PI != MBB->pred_size()) {
1380 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1382 // If this block has an uncond branch to itself, leave it.
1384 HasBranchToSelf = true;
1387 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1388 // If this change resulted in PMBB ending in a conditional
1389 // branch where both conditions go to the same destination,
1390 // change this to an unconditional branch (and fix the CFG).
1391 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1392 SmallVector<MachineOperand, 4> NewCurCond;
1393 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1394 NewCurFBB, NewCurCond, true);
1395 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1396 TII->RemoveBranch(*PMBB);
1398 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1401 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1406 // Change any jumptables to go to the new MBB.
1407 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB);
1411 if (!HasBranchToSelf) return MadeChange;
1416 // Add the branch back if the block is more than just an uncond branch.
1417 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1421 // If the prior block doesn't fall through into this block, and if this
1422 // block doesn't fall through into some other block, see if we can find a
1423 // place to move this block where a fall-through will happen.
1424 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
1425 PriorTBB, PriorFBB, PriorCond)) {
1427 // Now we know that there was no fall-through into this block, check to
1428 // see if it has a fall-through into its successor.
1429 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
1432 if (!MBB->isLandingPad()) {
1433 // Check all the predecessors of this block. If one of them has no fall
1434 // throughs, move this block right after it.
1435 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1436 E = MBB->pred_end(); PI != E; ++PI) {
1437 // Analyze the branch at the end of the pred.
1438 MachineBasicBlock *PredBB = *PI;
1439 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1440 MachineBasicBlock *PredTBB, *PredFBB;
1441 SmallVector<MachineOperand, 4> PredCond;
1442 if (PredBB != MBB && !CanFallThrough(PredBB) &&
1443 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1444 && (!CurFallsThru || !CurTBB || !CurFBB)
1445 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1446 // If the current block doesn't fall through, just move it.
1447 // If the current block can fall through and does not end with a
1448 // conditional branch, we need to append an unconditional jump to
1449 // the (current) next block. To avoid a possible compile-time
1450 // infinite loop, move blocks only backward in this case.
1451 // Also, if there are already 2 branches here, we cannot add a third;
1452 // this means we have the case
1457 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
1459 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1461 MBB->moveAfter(PredBB);
1463 goto ReoptimizeBlock;
1468 if (!CurFallsThru) {
1469 // Check all successors to see if we can move this block before it.
1470 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1471 E = MBB->succ_end(); SI != E; ++SI) {
1472 // Analyze the branch at the end of the block before the succ.
1473 MachineBasicBlock *SuccBB = *SI;
1474 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1476 // If this block doesn't already fall-through to that successor, and if
1477 // the succ doesn't already have a block that can fall through into it,
1478 // and if the successor isn't an EH destination, we can arrange for the
1479 // fallthrough to happen.
1480 if (SuccBB != MBB && &*SuccPrev != MBB &&
1481 !CanFallThrough(SuccPrev) && !CurUnAnalyzable &&
1482 !SuccBB->isLandingPad()) {
1483 MBB->moveBefore(SuccBB);
1485 goto ReoptimizeBlock;
1489 // Okay, there is no really great place to put this block. If, however,
1490 // the block before this one would be a fall-through if this block were
1491 // removed, move this block to the end of the function.
1492 MachineBasicBlock *PrevTBB, *PrevFBB;
1493 SmallVector<MachineOperand, 4> PrevCond;
1494 if (FallThrough != MF.end() &&
1495 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1496 PrevBB.isSuccessor(FallThrough)) {
1497 MBB->moveAfter(--MF.end());