1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass forwards branches to unconditional branches to make them branch
11 // directly to the target block. This pass often results in dead MBB's, which
14 // Note that this pass must be run after register allocation, it cannot handle
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "branchfolding"
20 #include "BranchFolding.h"
21 #include "llvm/Function.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/RegisterScavenging.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Target/TargetRegisterInfo.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/ADT/SmallSet.h"
35 #include "llvm/ADT/SetVector.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/ADT/STLExtras.h"
41 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
42 STATISTIC(NumBranchOpts, "Number of branches optimized");
43 STATISTIC(NumTailMerge , "Number of block tails merged");
45 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
46 cl::init(cl::BOU_UNSET), cl::Hidden);
48 // Throttle for huge numbers of predecessors (compile speed problems)
49 static cl::opt<unsigned>
50 TailMergeThreshold("tail-merge-threshold",
51 cl::desc("Max number of predecessors to consider tail merging"),
52 cl::init(150), cl::Hidden);
54 // Heuristic for tail merging (and, inversely, tail duplication).
55 // TODO: This should be replaced with a target query.
56 static cl::opt<unsigned>
57 TailMergeSize("tail-merge-size",
58 cl::desc("Min number of instructions to consider tail merging"),
59 cl::init(3), cl::Hidden);
62 /// BranchFolderPass - Wrap branch folder in a machine function pass.
63 class BranchFolderPass : public MachineFunctionPass,
67 explicit BranchFolderPass(bool defaultEnableTailMerge)
68 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {}
70 virtual bool runOnMachineFunction(MachineFunction &MF);
71 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
75 char BranchFolderPass::ID = 0;
77 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
78 return new BranchFolderPass(DefaultEnableTailMerge);
81 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
82 return OptimizeFunction(MF,
83 MF.getTarget().getInstrInfo(),
84 MF.getTarget().getRegisterInfo(),
85 getAnalysisIfAvailable<MachineModuleInfo>());
89 BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
90 switch (FlagEnableTailMerge) {
91 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
92 case cl::BOU_TRUE: EnableTailMerge = true; break;
93 case cl::BOU_FALSE: EnableTailMerge = false; break;
97 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
98 /// function, updating the CFG.
99 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
100 assert(MBB->pred_empty() && "MBB must be dead!");
101 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
103 MachineFunction *MF = MBB->getParent();
104 // drop all successors.
105 while (!MBB->succ_empty())
106 MBB->removeSuccessor(MBB->succ_end()-1);
108 // If there are any labels in the basic block, unregister them from
109 // MachineModuleInfo.
110 if (MMI && !MBB->empty()) {
111 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
114 // The label ID # is always operand #0, an immediate.
115 MMI->InvalidateLabel(I->getOperand(0).getImm());
123 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
124 /// followed by terminators, and if the implicitly defined registers are not
125 /// used by the terminators, remove those implicit_def's. e.g.
127 /// r0 = implicit_def
128 /// r1 = implicit_def
130 /// This block can be optimized away later if the implicit instructions are
132 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
133 SmallSet<unsigned, 4> ImpDefRegs;
134 MachineBasicBlock::iterator I = MBB->begin();
135 while (I != MBB->end()) {
136 if (!I->isImplicitDef())
138 unsigned Reg = I->getOperand(0).getReg();
139 ImpDefRegs.insert(Reg);
140 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
141 unsigned SubReg = *SubRegs; ++SubRegs)
142 ImpDefRegs.insert(SubReg);
145 if (ImpDefRegs.empty())
148 MachineBasicBlock::iterator FirstTerm = I;
149 while (I != MBB->end()) {
150 if (!TII->isUnpredicatedTerminator(I))
152 // See if it uses any of the implicitly defined registers.
153 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
154 MachineOperand &MO = I->getOperand(i);
155 if (!MO.isReg() || !MO.isUse())
157 unsigned Reg = MO.getReg();
158 if (ImpDefRegs.count(Reg))
165 while (I != FirstTerm) {
166 MachineInstr *ImpDefMI = &*I;
168 MBB->erase(ImpDefMI);
174 /// OptimizeFunction - Perhaps branch folding, tail merging and other
175 /// CFG optimizations on the given function.
176 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
177 const TargetInstrInfo *tii,
178 const TargetRegisterInfo *tri,
179 MachineModuleInfo *mmi) {
180 if (!tii) return false;
186 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
188 // Fix CFG. The later algorithms expect it to be right.
189 bool MadeChange = false;
190 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
191 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
192 SmallVector<MachineOperand, 4> Cond;
193 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
194 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
195 MadeChange |= OptimizeImpDefsBlock(MBB);
198 bool MadeChangeThisIteration = true;
199 while (MadeChangeThisIteration) {
200 MadeChangeThisIteration = false;
201 MadeChangeThisIteration |= TailMergeBlocks(MF);
202 MadeChangeThisIteration |= OptimizeBranches(MF);
203 MadeChange |= MadeChangeThisIteration;
206 // See if any jump tables have become mergable or dead as the code generator
208 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
214 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);
263 //===----------------------------------------------------------------------===//
264 // Tail Merging of Blocks
265 //===----------------------------------------------------------------------===//
267 /// HashMachineInstr - Compute a hash value for MI and its operands.
268 static unsigned HashMachineInstr(const MachineInstr *MI) {
269 unsigned Hash = MI->getOpcode();
270 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
271 const MachineOperand &Op = MI->getOperand(i);
273 // Merge in bits from the operand if easy.
274 unsigned OperandHash = 0;
275 switch (Op.getType()) {
276 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
277 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
278 case MachineOperand::MO_MachineBasicBlock:
279 OperandHash = Op.getMBB()->getNumber();
281 case MachineOperand::MO_FrameIndex:
282 case MachineOperand::MO_ConstantPoolIndex:
283 case MachineOperand::MO_JumpTableIndex:
284 OperandHash = Op.getIndex();
286 case MachineOperand::MO_GlobalAddress:
287 case MachineOperand::MO_ExternalSymbol:
288 // Global address / external symbol are too hard, don't bother, but do
289 // pull in the offset.
290 OperandHash = Op.getOffset();
295 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
300 /// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
301 /// with no successors, we hash two instructions, because cross-jumping
302 /// only saves code when at least two instructions are removed (since a
303 /// branch must be inserted). For blocks with a successor, one of the
304 /// two blocks to be tail-merged will end with a branch already, so
305 /// it gains to cross-jump even for one instruction.
306 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
307 unsigned minCommonTailLength) {
308 MachineBasicBlock::const_iterator I = MBB->end();
309 if (I == MBB->begin())
310 return 0; // Empty MBB.
313 unsigned Hash = HashMachineInstr(I);
315 if (I == MBB->begin() || minCommonTailLength == 1)
316 return Hash; // Single instr MBB.
319 // Hash in the second-to-last instruction.
320 Hash ^= HashMachineInstr(I) << 2;
324 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
325 /// of instructions they actually have in common together at their end. Return
326 /// iterators for the first shared instruction in each block.
327 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
328 MachineBasicBlock *MBB2,
329 MachineBasicBlock::iterator &I1,
330 MachineBasicBlock::iterator &I2) {
334 unsigned TailLen = 0;
335 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
337 if (!I1->isIdenticalTo(I2) ||
338 // FIXME: This check is dubious. It's used to get around a problem where
339 // people incorrectly expect inline asm directives to remain in the same
340 // relative order. This is untenable because normal compiler
341 // optimizations (like this one) may reorder and/or merge these
352 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
353 /// after it, replacing it with an unconditional branch to NewDest. This
354 /// returns true if OldInst's block is modified, false if NewDest is modified.
355 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
356 MachineBasicBlock *NewDest) {
357 MachineBasicBlock *OldBB = OldInst->getParent();
359 // Remove all the old successors of OldBB from the CFG.
360 while (!OldBB->succ_empty())
361 OldBB->removeSuccessor(OldBB->succ_begin());
363 // Remove all the dead instructions from the end of OldBB.
364 OldBB->erase(OldInst, OldBB->end());
366 // If OldBB isn't immediately before OldBB, insert a branch to it.
367 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
368 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
369 OldBB->addSuccessor(NewDest);
373 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
374 /// MBB so that the part before the iterator falls into the part starting at the
375 /// iterator. This returns the new MBB.
376 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
377 MachineBasicBlock::iterator BBI1) {
378 MachineFunction &MF = *CurMBB.getParent();
380 // Create the fall-through block.
381 MachineFunction::iterator MBBI = &CurMBB;
382 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
383 CurMBB.getParent()->insert(++MBBI, NewMBB);
385 // Move all the successors of this block to the specified block.
386 NewMBB->transferSuccessors(&CurMBB);
388 // Add an edge from CurMBB to NewMBB for the fall-through.
389 CurMBB.addSuccessor(NewMBB);
391 // Splice the code over.
392 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
394 // For targets that use the register scavenger, we must maintain LiveIns.
396 RS->enterBasicBlock(&CurMBB);
398 RS->forward(prior(CurMBB.end()));
399 BitVector RegsLiveAtExit(TRI->getNumRegs());
400 RS->getRegsUsed(RegsLiveAtExit, false);
401 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
402 if (RegsLiveAtExit[i])
403 NewMBB->addLiveIn(i);
409 /// EstimateRuntime - Make a rough estimate for how long it will take to run
410 /// the specified code.
411 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
412 MachineBasicBlock::iterator E) {
414 for (; I != E; ++I) {
415 const TargetInstrDesc &TID = I->getDesc();
418 else if (TID.mayLoad() || TID.mayStore())
426 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
427 // branches temporarily for tail merging). In the case where CurMBB ends
428 // with a conditional branch to the next block, optimize by reversing the
429 // test and conditionally branching to SuccMBB instead.
430 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
431 const TargetInstrInfo *TII) {
432 MachineFunction *MF = CurMBB->getParent();
433 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB));
434 MachineBasicBlock *TBB = 0, *FBB = 0;
435 SmallVector<MachineOperand, 4> Cond;
436 if (I != MF->end() &&
437 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
438 MachineBasicBlock *NextBB = I;
439 if (TBB == NextBB && !Cond.empty() && !FBB) {
440 if (!TII->ReverseBranchCondition(Cond)) {
441 TII->RemoveBranch(*CurMBB);
442 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
447 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
451 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
452 if (getHash() < o.getHash())
454 else if (getHash() > o.getHash())
456 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
458 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
461 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
462 // an object with itself.
463 #ifndef _GLIBCXX_DEBUG
464 llvm_unreachable("Predecessor appears twice");
470 /// CountTerminators - Count the number of terminators in the given
471 /// block and set I to the position of the first non-terminator, if there
472 /// is one, or MBB->end() otherwise.
473 static unsigned CountTerminators(MachineBasicBlock *MBB,
474 MachineBasicBlock::iterator &I) {
476 unsigned NumTerms = 0;
478 if (I == MBB->begin()) {
483 if (!I->getDesc().isTerminator()) break;
489 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
490 /// and decide if it would be profitable to merge those tails. Return the
491 /// length of the common tail and iterators to the first common instruction
493 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
494 MachineBasicBlock *MBB2,
495 unsigned minCommonTailLength,
496 unsigned &CommonTailLen,
497 MachineBasicBlock::iterator &I1,
498 MachineBasicBlock::iterator &I2,
499 MachineBasicBlock *SuccBB,
500 MachineBasicBlock *PredBB) {
501 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
502 MachineFunction *MF = MBB1->getParent();
504 if (CommonTailLen == 0)
507 // It's almost always profitable to merge any number of non-terminator
508 // instructions with the block that falls through into the common successor.
509 if (MBB1 == PredBB || MBB2 == PredBB) {
510 MachineBasicBlock::iterator I;
511 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
512 if (CommonTailLen > NumTerms)
516 // If one of the blocks can be completely merged and happens to be in
517 // a position where the other could fall through into it, merge any number
518 // of instructions, because it can be done without a branch.
519 // TODO: If the blocks are not adjacent, move one of them so that they are?
520 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
522 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
525 // If both blocks have an unconditional branch temporarily stripped out,
526 // count that as an additional common instruction for the following
528 unsigned EffectiveTailLen = CommonTailLen;
529 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
530 !MBB1->back().getDesc().isBarrier() &&
531 !MBB2->back().getDesc().isBarrier())
534 // Check if the common tail is long enough to be worthwhile.
535 if (EffectiveTailLen >= minCommonTailLength)
538 // If we are optimizing for code size, 2 instructions in common is enough if
539 // we don't have to split a block. At worst we will be introducing 1 new
540 // branch instruction, which is likely to be smaller than the 2
541 // instructions that would be deleted in the merge.
542 if (EffectiveTailLen >= 2 &&
543 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
544 (I1 == MBB1->begin() || I2 == MBB2->begin()))
550 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
551 /// hash CurHash (guaranteed to match the last element). Build the vector
552 /// SameTails of all those that have the (same) largest number of instructions
553 /// in common of any pair of these blocks. SameTails entries contain an
554 /// iterator into MergePotentials (from which the MachineBasicBlock can be
555 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
556 /// instruction where the matching code sequence begins.
557 /// Order of elements in SameTails is the reverse of the order in which
558 /// those blocks appear in MergePotentials (where they are not necessarily
560 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
561 unsigned minCommonTailLength,
562 MachineBasicBlock *SuccBB,
563 MachineBasicBlock *PredBB) {
564 unsigned maxCommonTailLength = 0U;
566 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
567 MPIterator HighestMPIter = prior(MergePotentials.end());
568 for (MPIterator CurMPIter = prior(MergePotentials.end()),
569 B = MergePotentials.begin();
570 CurMPIter != B && CurMPIter->getHash() == CurHash;
572 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
573 unsigned CommonTailLen;
574 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
576 CommonTailLen, TrialBBI1, TrialBBI2,
578 if (CommonTailLen > maxCommonTailLength) {
580 maxCommonTailLength = CommonTailLen;
581 HighestMPIter = CurMPIter;
582 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
584 if (HighestMPIter == CurMPIter &&
585 CommonTailLen == maxCommonTailLength)
586 SameTails.push_back(SameTailElt(I, TrialBBI2));
592 return maxCommonTailLength;
595 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
596 /// MergePotentials, restoring branches at ends of blocks as appropriate.
597 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
598 MachineBasicBlock *SuccBB,
599 MachineBasicBlock *PredBB) {
600 MPIterator CurMPIter, B;
601 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
602 CurMPIter->getHash() == CurHash;
604 // Put the unconditional branch back, if we need one.
605 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
606 if (SuccBB && CurMBB != PredBB)
607 FixTail(CurMBB, SuccBB, TII);
611 if (CurMPIter->getHash() != CurHash)
613 MergePotentials.erase(CurMPIter, MergePotentials.end());
616 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
617 /// only of the common tail. Create a block that does by splitting one.
618 unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
619 unsigned maxCommonTailLength) {
620 unsigned commonTailIndex = 0;
621 unsigned TimeEstimate = ~0U;
622 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
623 // Use PredBB if possible; that doesn't require a new branch.
624 if (SameTails[i].getBlock() == PredBB) {
628 // Otherwise, make a (fairly bogus) choice based on estimate of
629 // how long it will take the various blocks to execute.
630 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
631 SameTails[i].getTailStartPos());
632 if (t <= TimeEstimate) {
638 MachineBasicBlock::iterator BBI =
639 SameTails[commonTailIndex].getTailStartPos();
640 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
642 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
643 << maxCommonTailLength);
645 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
646 SameTails[commonTailIndex].setBlock(newMBB);
647 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
649 // If we split PredBB, newMBB is the new predecessor.
653 return commonTailIndex;
656 // See if any of the blocks in MergePotentials (which all have a common single
657 // successor, or all have no successor) can be tail-merged. If there is a
658 // successor, any blocks in MergePotentials that are not tail-merged and
659 // are not immediately before Succ must have an unconditional branch to
660 // Succ added (but the predecessor/successor lists need no adjustment).
661 // The lone predecessor of Succ that falls through into Succ,
662 // if any, is given in PredBB.
664 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
665 MachineBasicBlock *PredBB) {
666 bool MadeChange = false;
668 // Except for the special cases below, tail-merge if there are at least
669 // this many instructions in common.
670 unsigned minCommonTailLength = TailMergeSize;
672 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
673 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
674 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
675 << (i == e-1 ? "" : ", ");
678 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
680 dbgs() << " which has fall-through from BB#"
681 << PredBB->getNumber() << "\n";
683 dbgs() << "Looking for common tails of at least "
684 << minCommonTailLength << " instruction"
685 << (minCommonTailLength == 1 ? "" : "s") << '\n';
688 // Sort by hash value so that blocks with identical end sequences sort
690 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
692 // Walk through equivalence sets looking for actual exact matches.
693 while (MergePotentials.size() > 1) {
694 unsigned CurHash = MergePotentials.back().getHash();
696 // Build SameTails, identifying the set of blocks with this hash code
697 // and with the maximum number of instructions in common.
698 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
702 // If we didn't find any pair that has at least minCommonTailLength
703 // instructions in common, remove all blocks with this hash code and retry.
704 if (SameTails.empty()) {
705 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
709 // If one of the blocks is the entire common tail (and not the entry
710 // block, which we can't jump to), we can treat all blocks with this same
711 // tail at once. Use PredBB if that is one of the possibilities, as that
712 // will not introduce any extra branches.
713 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
714 getParent()->begin();
715 unsigned commonTailIndex = SameTails.size();
716 // If there are two blocks, check to see if one can be made to fall through
718 if (SameTails.size() == 2 &&
719 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
720 SameTails[1].tailIsWholeBlock())
722 else if (SameTails.size() == 2 &&
723 SameTails[1].getBlock()->isLayoutSuccessor(
724 SameTails[0].getBlock()) &&
725 SameTails[0].tailIsWholeBlock())
728 // Otherwise just pick one, favoring the fall-through predecessor if
730 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
731 MachineBasicBlock *MBB = SameTails[i].getBlock();
732 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
738 if (SameTails[i].tailIsWholeBlock())
743 if (commonTailIndex == SameTails.size() ||
744 (SameTails[commonTailIndex].getBlock() == PredBB &&
745 !SameTails[commonTailIndex].tailIsWholeBlock())) {
746 // None of the blocks consist entirely of the common tail.
747 // Split a block so that one does.
748 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
751 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
752 // MBB is common tail. Adjust all other BB's to jump to this one.
753 // Traversal must be forwards so erases work.
754 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
756 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
757 if (commonTailIndex == i)
759 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
760 << (i == e-1 ? "" : ", "));
761 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
762 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
763 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
764 MergePotentials.erase(SameTails[i].getMPIter());
766 DEBUG(dbgs() << "\n");
767 // We leave commonTailIndex in the worklist in case there are other blocks
768 // that match it with a smaller number of instructions.
774 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
776 if (!EnableTailMerge) return false;
778 bool MadeChange = false;
780 // First find blocks with no successors.
781 MergePotentials.clear();
782 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
784 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
787 // See if we can do any tail merging on those.
788 if (MergePotentials.size() < TailMergeThreshold &&
789 MergePotentials.size() >= 2)
790 MadeChange |= TryTailMergeBlocks(NULL, NULL);
792 // Look at blocks (IBB) with multiple predecessors (PBB).
793 // We change each predecessor to a canonical form, by
794 // (1) temporarily removing any unconditional branch from the predecessor
796 // (2) alter conditional branches so they branch to the other block
797 // not IBB; this may require adding back an unconditional branch to IBB
798 // later, where there wasn't one coming in. E.g.
800 // fallthrough to QBB
803 // with a conceptual B to IBB after that, which never actually exists.
804 // With those changes, we see whether the predecessors' tails match,
805 // and merge them if so. We change things out of canonical form and
806 // back to the way they were later in the process. (OptimizeBranches
807 // would undo some of this, but we can't use it, because we'd get into
808 // a compile-time infinite loop repeatedly doing and undoing the same
811 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
813 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
814 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
815 MachineBasicBlock *IBB = I;
816 MachineBasicBlock *PredBB = prior(I);
817 MergePotentials.clear();
818 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
821 MachineBasicBlock *PBB = *P;
822 // Skip blocks that loop to themselves, can't tail merge these.
825 // Visit each predecessor only once.
826 if (!UniquePreds.insert(PBB))
828 MachineBasicBlock *TBB = 0, *FBB = 0;
829 SmallVector<MachineOperand, 4> Cond;
830 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
831 // Failing case: IBB is the target of a cbr, and
832 // we cannot reverse the branch.
833 SmallVector<MachineOperand, 4> NewCond(Cond);
834 if (!Cond.empty() && TBB == IBB) {
835 if (TII->ReverseBranchCondition(NewCond))
837 // This is the QBB case described above
839 FBB = llvm::next(MachineFunction::iterator(PBB));
841 // Failing case: the only way IBB can be reached from PBB is via
842 // exception handling. Happens for landing pads. Would be nice
843 // to have a bit in the edge so we didn't have to do all this.
844 if (IBB->isLandingPad()) {
845 MachineFunction::iterator IP = PBB; IP++;
846 MachineBasicBlock *PredNextBB = NULL;
850 if (IBB != PredNextBB) // fallthrough
853 if (TBB != IBB && FBB != IBB) // cbr then ubr
855 } else if (Cond.empty()) {
856 if (TBB != IBB) // ubr
859 if (TBB != IBB && IBB != PredNextBB) // cbr
863 // Remove the unconditional branch at the end, if any.
864 if (TBB && (Cond.empty() || FBB)) {
865 TII->RemoveBranch(*PBB);
867 // reinsert conditional branch only, for now
868 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
870 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
874 if (MergePotentials.size() >= 2)
875 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
876 // Reinsert an unconditional branch if needed.
877 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
878 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
879 if (MergePotentials.size() == 1 &&
880 MergePotentials.begin()->getBlock() != PredBB)
881 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
887 //===----------------------------------------------------------------------===//
888 // Branch Optimization
889 //===----------------------------------------------------------------------===//
891 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
892 bool MadeChange = false;
894 // Make sure blocks are numbered in order
897 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
898 MachineBasicBlock *MBB = I++;
899 MadeChange |= OptimizeBlock(MBB);
901 // If it is dead, remove it.
902 if (MBB->pred_empty()) {
903 RemoveDeadBlock(MBB);
912 /// IsBetterFallthrough - Return true if it would be clearly better to
913 /// fall-through to MBB1 than to fall through into MBB2. This has to return
914 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
915 /// result in infinite loops.
916 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
917 MachineBasicBlock *MBB2) {
918 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
919 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
920 // optimize branches that branch to either a return block or an assert block
921 // into a fallthrough to the return.
922 if (MBB1->empty() || MBB2->empty()) return false;
924 // If there is a clear successor ordering we make sure that one block
925 // will fall through to the next
926 if (MBB1->isSuccessor(MBB2)) return true;
927 if (MBB2->isSuccessor(MBB1)) return false;
929 MachineInstr *MBB1I = --MBB1->end();
930 MachineInstr *MBB2I = --MBB2->end();
931 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
934 /// OptimizeBlock - Analyze and optimize control flow related to the specified
935 /// block. This is never called on the entry block.
936 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
937 bool MadeChange = false;
938 MachineFunction &MF = *MBB->getParent();
941 MachineFunction::iterator FallThrough = MBB;
944 // If this block is empty, make everyone use its fall-through, not the block
945 // explicitly. Landing pads should not do this since the landing-pad table
946 // points to this block. Blocks with their addresses taken shouldn't be
948 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
949 // Dead block? Leave for cleanup later.
950 if (MBB->pred_empty()) return MadeChange;
952 if (FallThrough == MF.end()) {
953 // TODO: Simplify preds to not branch here if possible!
955 // Rewrite all predecessors of the old block to go to the fallthrough
957 while (!MBB->pred_empty()) {
958 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
959 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
961 // If MBB was the target of a jump table, update jump tables to go to the
962 // fallthrough instead.
963 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
964 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
970 // Check to see if we can simplify the terminator of the block before this
972 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
974 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
975 SmallVector<MachineOperand, 4> PriorCond;
976 bool PriorUnAnalyzable =
977 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
978 if (!PriorUnAnalyzable) {
979 // If the CFG for the prior block has extra edges, remove them.
980 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
983 // If the previous branch is conditional and both conditions go to the same
984 // destination, remove the branch, replacing it with an unconditional one or
986 if (PriorTBB && PriorTBB == PriorFBB) {
987 TII->RemoveBranch(PrevBB);
990 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
993 goto ReoptimizeBlock;
996 // If the previous block unconditionally falls through to this block and
997 // this block has no other predecessors, move the contents of this block
998 // into the prior block. This doesn't usually happen when SimplifyCFG
999 // has been used, but it can happen if tail merging splits a fall-through
1000 // predecessor of a block.
1001 // This has to check PrevBB->succ_size() because EH edges are ignored by
1003 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1004 PrevBB.succ_size() == 1 &&
1005 !MBB->hasAddressTaken()) {
1006 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1007 << "From MBB: " << *MBB);
1008 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1009 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1010 assert(PrevBB.succ_empty());
1011 PrevBB.transferSuccessors(MBB);
1016 // If the previous branch *only* branches to *this* block (conditional or
1017 // not) remove the branch.
1018 if (PriorTBB == MBB && PriorFBB == 0) {
1019 TII->RemoveBranch(PrevBB);
1022 goto ReoptimizeBlock;
1025 // If the prior block branches somewhere else on the condition and here if
1026 // the condition is false, remove the uncond second branch.
1027 if (PriorFBB == MBB) {
1028 TII->RemoveBranch(PrevBB);
1029 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
1032 goto ReoptimizeBlock;
1035 // If the prior block branches here on true and somewhere else on false, and
1036 // if the branch condition is reversible, reverse the branch to create a
1038 if (PriorTBB == MBB) {
1039 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1040 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1041 TII->RemoveBranch(PrevBB);
1042 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
1045 goto ReoptimizeBlock;
1049 // If this block has no successors (e.g. it is a return block or ends with
1050 // a call to a no-return function like abort or __cxa_throw) and if the pred
1051 // falls through into this block, and if it would otherwise fall through
1052 // into the block after this, move this block to the end of the function.
1054 // We consider it more likely that execution will stay in the function (e.g.
1055 // due to loops) than it is to exit it. This asserts in loops etc, moving
1056 // the assert condition out of the loop body.
1057 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1058 MachineFunction::iterator(PriorTBB) == FallThrough &&
1059 !MBB->canFallThrough()) {
1060 bool DoTransform = true;
1062 // We have to be careful that the succs of PredBB aren't both no-successor
1063 // blocks. If neither have successors and if PredBB is the second from
1064 // last block in the function, we'd just keep swapping the two blocks for
1065 // last. Only do the swap if one is clearly better to fall through than
1067 if (FallThrough == --MF.end() &&
1068 !IsBetterFallthrough(PriorTBB, MBB))
1069 DoTransform = false;
1071 // We don't want to do this transformation if we have control flow like:
1080 // In this case, we could actually be moving the return block *into* a
1082 if (DoTransform && !MBB->succ_empty() &&
1083 (!PriorTBB->canFallThrough() || PriorTBB->empty()))
1084 DoTransform = false;
1088 // Reverse the branch so we will fall through on the previous true cond.
1089 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1090 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1091 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1092 << "To make fallthrough to: " << *PriorTBB << "\n");
1094 TII->RemoveBranch(PrevBB);
1095 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
1097 // Move this block to the end of the function.
1098 MBB->moveAfter(--MF.end());
1107 // Analyze the branch in the current block.
1108 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1109 SmallVector<MachineOperand, 4> CurCond;
1110 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1111 if (!CurUnAnalyzable) {
1112 // If the CFG for the prior block has extra edges, remove them.
1113 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1115 // If this is a two-way branch, and the FBB branches to this block, reverse
1116 // the condition so the single-basic-block loop is faster. Instead of:
1117 // Loop: xxx; jcc Out; jmp Loop
1119 // Loop: xxx; jncc Loop; jmp Out
1120 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1121 SmallVector<MachineOperand, 4> NewCond(CurCond);
1122 if (!TII->ReverseBranchCondition(NewCond)) {
1123 TII->RemoveBranch(*MBB);
1124 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
1127 goto ReoptimizeBlock;
1131 // If this branch is the only thing in its block, see if we can forward
1132 // other blocks across it.
1133 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1134 MBB->begin()->getDesc().isBranch() && CurTBB != MBB &&
1135 !MBB->hasAddressTaken()) {
1136 // This block may contain just an unconditional branch. Because there can
1137 // be 'non-branch terminators' in the block, try removing the branch and
1138 // then seeing if the block is empty.
1139 TII->RemoveBranch(*MBB);
1141 // If this block is just an unconditional branch to CurTBB, we can
1142 // usually completely eliminate the block. The only case we cannot
1143 // completely eliminate the block is when the block before this one
1144 // falls through into MBB and we can't understand the prior block's branch
1147 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1148 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1149 !PrevBB.isSuccessor(MBB)) {
1150 // If the prior block falls through into us, turn it into an
1151 // explicit branch to us to make updates simpler.
1152 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1153 PriorTBB != MBB && PriorFBB != MBB) {
1154 if (PriorTBB == 0) {
1155 assert(PriorCond.empty() && PriorFBB == 0 &&
1156 "Bad branch analysis");
1159 assert(PriorFBB == 0 && "Machine CFG out of date!");
1162 TII->RemoveBranch(PrevBB);
1163 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
1166 // Iterate through all the predecessors, revectoring each in-turn.
1168 bool DidChange = false;
1169 bool HasBranchToSelf = false;
1170 while(PI != MBB->pred_size()) {
1171 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1173 // If this block has an uncond branch to itself, leave it.
1175 HasBranchToSelf = true;
1178 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1179 // If this change resulted in PMBB ending in a conditional
1180 // branch where both conditions go to the same destination,
1181 // change this to an unconditional branch (and fix the CFG).
1182 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1183 SmallVector<MachineOperand, 4> NewCurCond;
1184 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1185 NewCurFBB, NewCurCond, true);
1186 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1187 TII->RemoveBranch(*PMBB);
1189 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
1192 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1197 // Change any jumptables to go to the new MBB.
1198 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1199 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1203 if (!HasBranchToSelf) return MadeChange;
1208 // Add the branch back if the block is more than just an uncond branch.
1209 TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
1213 // If the prior block doesn't fall through into this block, and if this
1214 // block doesn't fall through into some other block, see if we can find a
1215 // place to move this block where a fall-through will happen.
1216 if (!PrevBB.canFallThrough()) {
1218 // Now we know that there was no fall-through into this block, check to
1219 // see if it has a fall-through into its successor.
1220 bool CurFallsThru = MBB->canFallThrough();
1222 if (!MBB->isLandingPad()) {
1223 // Check all the predecessors of this block. If one of them has no fall
1224 // throughs, move this block right after it.
1225 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1226 E = MBB->pred_end(); PI != E; ++PI) {
1227 // Analyze the branch at the end of the pred.
1228 MachineBasicBlock *PredBB = *PI;
1229 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1230 MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
1231 SmallVector<MachineOperand, 4> PredCond;
1232 if (PredBB != MBB && !PredBB->canFallThrough() &&
1233 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1234 && (!CurFallsThru || !CurTBB || !CurFBB)
1235 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1236 // If the current block doesn't fall through, just move it.
1237 // If the current block can fall through and does not end with a
1238 // conditional branch, we need to append an unconditional jump to
1239 // the (current) next block. To avoid a possible compile-time
1240 // infinite loop, move blocks only backward in this case.
1241 // Also, if there are already 2 branches here, we cannot add a third;
1242 // this means we have the case
1247 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
1249 TII->InsertBranch(*MBB, NextBB, 0, CurCond);
1251 MBB->moveAfter(PredBB);
1253 goto ReoptimizeBlock;
1258 if (!CurFallsThru) {
1259 // Check all successors to see if we can move this block before it.
1260 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1261 E = MBB->succ_end(); SI != E; ++SI) {
1262 // Analyze the branch at the end of the block before the succ.
1263 MachineBasicBlock *SuccBB = *SI;
1264 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1266 // If this block doesn't already fall-through to that successor, and if
1267 // the succ doesn't already have a block that can fall through into it,
1268 // and if the successor isn't an EH destination, we can arrange for the
1269 // fallthrough to happen.
1270 if (SuccBB != MBB && &*SuccPrev != MBB &&
1271 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1272 !SuccBB->isLandingPad()) {
1273 MBB->moveBefore(SuccBB);
1275 goto ReoptimizeBlock;
1279 // Okay, there is no really great place to put this block. If, however,
1280 // the block before this one would be a fall-through if this block were
1281 // removed, move this block to the end of the function.
1282 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
1283 SmallVector<MachineOperand, 4> PrevCond;
1284 if (FallThrough != MF.end() &&
1285 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1286 PrevBB.isSuccessor(FallThrough)) {
1287 MBB->moveAfter(--MF.end());