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(NumHoist , "Number of times common instructions are hoisted");
46 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
47 cl::init(cl::BOU_UNSET), cl::Hidden);
49 // Throttle for huge numbers of predecessors (compile speed problems)
50 static cl::opt<unsigned>
51 TailMergeThreshold("tail-merge-threshold",
52 cl::desc("Max number of predecessors to consider tail merging"),
53 cl::init(150), cl::Hidden);
55 // Heuristic for tail merging (and, inversely, tail duplication).
56 // TODO: This should be replaced with a target query.
57 static cl::opt<unsigned>
58 TailMergeSize("tail-merge-size",
59 cl::desc("Min number of instructions to consider tail merging"),
60 cl::init(3), cl::Hidden);
63 /// BranchFolderPass - Wrap branch folder in a machine function pass.
64 class BranchFolderPass : public MachineFunctionPass,
68 explicit BranchFolderPass(bool defaultEnableTailMerge)
69 : MachineFunctionPass(ID), BranchFolder(defaultEnableTailMerge, true) {}
71 virtual bool runOnMachineFunction(MachineFunction &MF);
72 virtual const char *getPassName() const { return "Control Flow Optimizer"; }
76 char BranchFolderPass::ID = 0;
78 FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
79 return new BranchFolderPass(DefaultEnableTailMerge);
82 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
83 return OptimizeFunction(MF,
84 MF.getTarget().getInstrInfo(),
85 MF.getTarget().getRegisterInfo(),
86 getAnalysisIfAvailable<MachineModuleInfo>());
90 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist) {
91 switch (FlagEnableTailMerge) {
92 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
93 case cl::BOU_TRUE: EnableTailMerge = true; break;
94 case cl::BOU_FALSE: EnableTailMerge = false; break;
97 EnableHoistCommonCode = CommonHoist;
100 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
101 /// function, updating the CFG.
102 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
103 assert(MBB->pred_empty() && "MBB must be dead!");
104 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
106 MachineFunction *MF = MBB->getParent();
107 // drop all successors.
108 while (!MBB->succ_empty())
109 MBB->removeSuccessor(MBB->succ_end()-1);
111 // Avoid matching if this pointer gets reused.
112 TriedMerging.erase(MBB);
118 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
119 /// followed by terminators, and if the implicitly defined registers are not
120 /// used by the terminators, remove those implicit_def's. e.g.
122 /// r0 = implicit_def
123 /// r1 = implicit_def
125 /// This block can be optimized away later if the implicit instructions are
127 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
128 SmallSet<unsigned, 4> ImpDefRegs;
129 MachineBasicBlock::iterator I = MBB->begin();
130 while (I != MBB->end()) {
131 if (!I->isImplicitDef())
133 unsigned Reg = I->getOperand(0).getReg();
134 ImpDefRegs.insert(Reg);
135 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
136 unsigned SubReg = *SubRegs; ++SubRegs)
137 ImpDefRegs.insert(SubReg);
140 if (ImpDefRegs.empty())
143 MachineBasicBlock::iterator FirstTerm = I;
144 while (I != MBB->end()) {
145 if (!TII->isUnpredicatedTerminator(I))
147 // See if it uses any of the implicitly defined registers.
148 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
149 MachineOperand &MO = I->getOperand(i);
150 if (!MO.isReg() || !MO.isUse())
152 unsigned Reg = MO.getReg();
153 if (ImpDefRegs.count(Reg))
160 while (I != FirstTerm) {
161 MachineInstr *ImpDefMI = &*I;
163 MBB->erase(ImpDefMI);
169 /// OptimizeFunction - Perhaps branch folding, tail merging and other
170 /// CFG optimizations on the given function.
171 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
172 const TargetInstrInfo *tii,
173 const TargetRegisterInfo *tri,
174 MachineModuleInfo *mmi) {
175 if (!tii) return false;
177 TriedMerging.clear();
183 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
185 // Fix CFG. The later algorithms expect it to be right.
186 bool MadeChange = false;
187 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
188 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
189 SmallVector<MachineOperand, 4> Cond;
190 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
191 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
192 MadeChange |= OptimizeImpDefsBlock(MBB);
195 bool MadeChangeThisIteration = true;
196 while (MadeChangeThisIteration) {
197 MadeChangeThisIteration = TailMergeBlocks(MF);
198 MadeChangeThisIteration |= OptimizeBranches(MF);
199 if (EnableHoistCommonCode)
200 MadeChangeThisIteration |= HoistCommonCode(MF);
201 MadeChange |= MadeChangeThisIteration;
204 // See if any jump tables have become dead as the code generator
206 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
212 // Walk the function to find jump tables that are live.
213 BitVector JTIsLive(JTI->getJumpTables().size());
214 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
216 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
218 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
219 MachineOperand &Op = I->getOperand(op);
220 if (!Op.isJTI()) continue;
222 // Remember that this JT is live.
223 JTIsLive.set(Op.getIndex());
227 // Finally, remove dead jump tables. This happens when the
228 // indirect jump was unreachable (and thus deleted).
229 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
230 if (!JTIsLive.test(i)) {
231 JTI->RemoveJumpTable(i);
239 //===----------------------------------------------------------------------===//
240 // Tail Merging of Blocks
241 //===----------------------------------------------------------------------===//
243 /// HashMachineInstr - Compute a hash value for MI and its operands.
244 static unsigned HashMachineInstr(const MachineInstr *MI) {
245 unsigned Hash = MI->getOpcode();
246 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
247 const MachineOperand &Op = MI->getOperand(i);
249 // Merge in bits from the operand if easy.
250 unsigned OperandHash = 0;
251 switch (Op.getType()) {
252 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
253 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
254 case MachineOperand::MO_MachineBasicBlock:
255 OperandHash = Op.getMBB()->getNumber();
257 case MachineOperand::MO_FrameIndex:
258 case MachineOperand::MO_ConstantPoolIndex:
259 case MachineOperand::MO_JumpTableIndex:
260 OperandHash = Op.getIndex();
262 case MachineOperand::MO_GlobalAddress:
263 case MachineOperand::MO_ExternalSymbol:
264 // Global address / external symbol are too hard, don't bother, but do
265 // pull in the offset.
266 OperandHash = Op.getOffset();
271 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
276 /// HashEndOfMBB - Hash the last instruction in the MBB.
277 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
278 MachineBasicBlock::const_iterator I = MBB->end();
279 if (I == MBB->begin())
280 return 0; // Empty MBB.
283 // Skip debug info so it will not affect codegen.
284 while (I->isDebugValue()) {
286 return 0; // MBB empty except for debug info.
290 return HashMachineInstr(I);
293 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
294 /// of instructions they actually have in common together at their end. Return
295 /// iterators for the first shared instruction in each block.
296 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
297 MachineBasicBlock *MBB2,
298 MachineBasicBlock::iterator &I1,
299 MachineBasicBlock::iterator &I2) {
303 unsigned TailLen = 0;
304 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
306 // Skip debugging pseudos; necessary to avoid changing the code.
307 while (I1->isDebugValue()) {
308 if (I1==MBB1->begin()) {
309 while (I2->isDebugValue()) {
310 if (I2==MBB2->begin())
311 // I1==DBG at begin; I2==DBG at begin
316 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
321 // I1==first (untested) non-DBG preceding known match
322 while (I2->isDebugValue()) {
323 if (I2==MBB2->begin()) {
325 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
330 // I1, I2==first (untested) non-DBGs preceding known match
331 if (!I1->isIdenticalTo(I2) ||
332 // FIXME: This check is dubious. It's used to get around a problem where
333 // people incorrectly expect inline asm directives to remain in the same
334 // relative order. This is untenable because normal compiler
335 // optimizations (like this one) may reorder and/or merge these
343 // Back past possible debugging pseudos at beginning of block. This matters
344 // when one block differs from the other only by whether debugging pseudos
345 // are present at the beginning. (This way, the various checks later for
346 // I1==MBB1->begin() work as expected.)
347 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
349 while (I2->isDebugValue()) {
350 if (I2 == MBB2->begin()) {
357 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
359 while (I1->isDebugValue()) {
360 if (I1 == MBB1->begin())
369 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
370 MachineBasicBlock *NewMBB) {
372 RS->enterBasicBlock(CurMBB);
373 if (!CurMBB->empty())
374 RS->forward(prior(CurMBB->end()));
375 BitVector RegsLiveAtExit(TRI->getNumRegs());
376 RS->getRegsUsed(RegsLiveAtExit, false);
377 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
378 if (RegsLiveAtExit[i])
379 NewMBB->addLiveIn(i);
383 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
384 /// after it, replacing it with an unconditional branch to NewDest.
385 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
386 MachineBasicBlock *NewDest) {
387 MachineBasicBlock *CurMBB = OldInst->getParent();
389 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
391 // For targets that use the register scavenger, we must maintain LiveIns.
392 MaintainLiveIns(CurMBB, NewDest);
397 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
398 /// MBB so that the part before the iterator falls into the part starting at the
399 /// iterator. This returns the new MBB.
400 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
401 MachineBasicBlock::iterator BBI1) {
402 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
405 MachineFunction &MF = *CurMBB.getParent();
407 // Create the fall-through block.
408 MachineFunction::iterator MBBI = &CurMBB;
409 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
410 CurMBB.getParent()->insert(++MBBI, NewMBB);
412 // Move all the successors of this block to the specified block.
413 NewMBB->transferSuccessors(&CurMBB);
415 // Add an edge from CurMBB to NewMBB for the fall-through.
416 CurMBB.addSuccessor(NewMBB);
418 // Splice the code over.
419 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
421 // For targets that use the register scavenger, we must maintain LiveIns.
422 MaintainLiveIns(&CurMBB, NewMBB);
427 /// EstimateRuntime - Make a rough estimate for how long it will take to run
428 /// the specified code.
429 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
430 MachineBasicBlock::iterator E) {
432 for (; I != E; ++I) {
433 if (I->isDebugValue())
435 const MCInstrDesc &MCID = I->getDesc();
438 else if (MCID.mayLoad() || MCID.mayStore())
446 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
447 // branches temporarily for tail merging). In the case where CurMBB ends
448 // with a conditional branch to the next block, optimize by reversing the
449 // test and conditionally branching to SuccMBB instead.
450 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
451 const TargetInstrInfo *TII) {
452 MachineFunction *MF = CurMBB->getParent();
453 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB));
454 MachineBasicBlock *TBB = 0, *FBB = 0;
455 SmallVector<MachineOperand, 4> Cond;
456 DebugLoc dl; // FIXME: this is nowhere
457 if (I != MF->end() &&
458 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
459 MachineBasicBlock *NextBB = I;
460 if (TBB == NextBB && !Cond.empty() && !FBB) {
461 if (!TII->ReverseBranchCondition(Cond)) {
462 TII->RemoveBranch(*CurMBB);
463 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl);
468 TII->InsertBranch(*CurMBB, SuccBB, NULL,
469 SmallVector<MachineOperand, 0>(), dl);
473 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
474 if (getHash() < o.getHash())
476 else if (getHash() > o.getHash())
478 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
480 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
483 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
484 // an object with itself.
485 #ifndef _GLIBCXX_DEBUG
486 llvm_unreachable("Predecessor appears twice");
492 /// CountTerminators - Count the number of terminators in the given
493 /// block and set I to the position of the first non-terminator, if there
494 /// is one, or MBB->end() otherwise.
495 static unsigned CountTerminators(MachineBasicBlock *MBB,
496 MachineBasicBlock::iterator &I) {
498 unsigned NumTerms = 0;
500 if (I == MBB->begin()) {
505 if (!I->getDesc().isTerminator()) break;
511 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
512 /// and decide if it would be profitable to merge those tails. Return the
513 /// length of the common tail and iterators to the first common instruction
515 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
516 MachineBasicBlock *MBB2,
517 unsigned minCommonTailLength,
518 unsigned &CommonTailLen,
519 MachineBasicBlock::iterator &I1,
520 MachineBasicBlock::iterator &I2,
521 MachineBasicBlock *SuccBB,
522 MachineBasicBlock *PredBB) {
523 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
524 if (CommonTailLen == 0)
526 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
527 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
530 // It's almost always profitable to merge any number of non-terminator
531 // instructions with the block that falls through into the common successor.
532 if (MBB1 == PredBB || MBB2 == PredBB) {
533 MachineBasicBlock::iterator I;
534 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
535 if (CommonTailLen > NumTerms)
539 // If one of the blocks can be completely merged and happens to be in
540 // a position where the other could fall through into it, merge any number
541 // of instructions, because it can be done without a branch.
542 // TODO: If the blocks are not adjacent, move one of them so that they are?
543 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
545 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
548 // If both blocks have an unconditional branch temporarily stripped out,
549 // count that as an additional common instruction for the following
551 unsigned EffectiveTailLen = CommonTailLen;
552 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
553 !MBB1->back().getDesc().isBarrier() &&
554 !MBB2->back().getDesc().isBarrier())
557 // Check if the common tail is long enough to be worthwhile.
558 if (EffectiveTailLen >= minCommonTailLength)
561 // If we are optimizing for code size, 2 instructions in common is enough if
562 // we don't have to split a block. At worst we will be introducing 1 new
563 // branch instruction, which is likely to be smaller than the 2
564 // instructions that would be deleted in the merge.
565 MachineFunction *MF = MBB1->getParent();
566 if (EffectiveTailLen >= 2 &&
567 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
568 (I1 == MBB1->begin() || I2 == MBB2->begin()))
574 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
575 /// hash CurHash (guaranteed to match the last element). Build the vector
576 /// SameTails of all those that have the (same) largest number of instructions
577 /// in common of any pair of these blocks. SameTails entries contain an
578 /// iterator into MergePotentials (from which the MachineBasicBlock can be
579 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
580 /// instruction where the matching code sequence begins.
581 /// Order of elements in SameTails is the reverse of the order in which
582 /// those blocks appear in MergePotentials (where they are not necessarily
584 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
585 unsigned minCommonTailLength,
586 MachineBasicBlock *SuccBB,
587 MachineBasicBlock *PredBB) {
588 unsigned maxCommonTailLength = 0U;
590 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
591 MPIterator HighestMPIter = prior(MergePotentials.end());
592 for (MPIterator CurMPIter = prior(MergePotentials.end()),
593 B = MergePotentials.begin();
594 CurMPIter != B && CurMPIter->getHash() == CurHash;
596 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
597 unsigned CommonTailLen;
598 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
600 CommonTailLen, TrialBBI1, TrialBBI2,
602 if (CommonTailLen > maxCommonTailLength) {
604 maxCommonTailLength = CommonTailLen;
605 HighestMPIter = CurMPIter;
606 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
608 if (HighestMPIter == CurMPIter &&
609 CommonTailLen == maxCommonTailLength)
610 SameTails.push_back(SameTailElt(I, TrialBBI2));
616 return maxCommonTailLength;
619 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
620 /// MergePotentials, restoring branches at ends of blocks as appropriate.
621 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
622 MachineBasicBlock *SuccBB,
623 MachineBasicBlock *PredBB) {
624 MPIterator CurMPIter, B;
625 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
626 CurMPIter->getHash() == CurHash;
628 // Put the unconditional branch back, if we need one.
629 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
630 if (SuccBB && CurMBB != PredBB)
631 FixTail(CurMBB, SuccBB, TII);
635 if (CurMPIter->getHash() != CurHash)
637 MergePotentials.erase(CurMPIter, MergePotentials.end());
640 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
641 /// only of the common tail. Create a block that does by splitting one.
642 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
643 unsigned maxCommonTailLength,
644 unsigned &commonTailIndex) {
646 unsigned TimeEstimate = ~0U;
647 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
648 // Use PredBB if possible; that doesn't require a new branch.
649 if (SameTails[i].getBlock() == PredBB) {
653 // Otherwise, make a (fairly bogus) choice based on estimate of
654 // how long it will take the various blocks to execute.
655 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
656 SameTails[i].getTailStartPos());
657 if (t <= TimeEstimate) {
663 MachineBasicBlock::iterator BBI =
664 SameTails[commonTailIndex].getTailStartPos();
665 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
667 // If the common tail includes any debug info we will take it pretty
668 // randomly from one of the inputs. Might be better to remove it?
669 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
670 << maxCommonTailLength);
672 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
674 DEBUG(dbgs() << "... failed!");
678 SameTails[commonTailIndex].setBlock(newMBB);
679 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
681 // If we split PredBB, newMBB is the new predecessor.
688 // See if any of the blocks in MergePotentials (which all have a common single
689 // successor, or all have no successor) can be tail-merged. If there is a
690 // successor, any blocks in MergePotentials that are not tail-merged and
691 // are not immediately before Succ must have an unconditional branch to
692 // Succ added (but the predecessor/successor lists need no adjustment).
693 // The lone predecessor of Succ that falls through into Succ,
694 // if any, is given in PredBB.
696 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
697 MachineBasicBlock *PredBB) {
698 bool MadeChange = false;
700 // Except for the special cases below, tail-merge if there are at least
701 // this many instructions in common.
702 unsigned minCommonTailLength = TailMergeSize;
704 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
705 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
706 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
707 << (i == e-1 ? "" : ", ");
710 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
712 dbgs() << " which has fall-through from BB#"
713 << PredBB->getNumber() << "\n";
715 dbgs() << "Looking for common tails of at least "
716 << minCommonTailLength << " instruction"
717 << (minCommonTailLength == 1 ? "" : "s") << '\n';
720 // Sort by hash value so that blocks with identical end sequences sort
722 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
724 // Walk through equivalence sets looking for actual exact matches.
725 while (MergePotentials.size() > 1) {
726 unsigned CurHash = MergePotentials.back().getHash();
728 // Build SameTails, identifying the set of blocks with this hash code
729 // and with the maximum number of instructions in common.
730 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
734 // If we didn't find any pair that has at least minCommonTailLength
735 // instructions in common, remove all blocks with this hash code and retry.
736 if (SameTails.empty()) {
737 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
741 // If one of the blocks is the entire common tail (and not the entry
742 // block, which we can't jump to), we can treat all blocks with this same
743 // tail at once. Use PredBB if that is one of the possibilities, as that
744 // will not introduce any extra branches.
745 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
746 getParent()->begin();
747 unsigned commonTailIndex = SameTails.size();
748 // If there are two blocks, check to see if one can be made to fall through
750 if (SameTails.size() == 2 &&
751 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
752 SameTails[1].tailIsWholeBlock())
754 else if (SameTails.size() == 2 &&
755 SameTails[1].getBlock()->isLayoutSuccessor(
756 SameTails[0].getBlock()) &&
757 SameTails[0].tailIsWholeBlock())
760 // Otherwise just pick one, favoring the fall-through predecessor if
762 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
763 MachineBasicBlock *MBB = SameTails[i].getBlock();
764 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
770 if (SameTails[i].tailIsWholeBlock())
775 if (commonTailIndex == SameTails.size() ||
776 (SameTails[commonTailIndex].getBlock() == PredBB &&
777 !SameTails[commonTailIndex].tailIsWholeBlock())) {
778 // None of the blocks consist entirely of the common tail.
779 // Split a block so that one does.
780 if (!CreateCommonTailOnlyBlock(PredBB,
781 maxCommonTailLength, commonTailIndex)) {
782 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
787 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
788 // MBB is common tail. Adjust all other BB's to jump to this one.
789 // Traversal must be forwards so erases work.
790 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
792 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
793 if (commonTailIndex == i)
795 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
796 << (i == e-1 ? "" : ", "));
797 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
798 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
799 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
800 MergePotentials.erase(SameTails[i].getMPIter());
802 DEBUG(dbgs() << "\n");
803 // We leave commonTailIndex in the worklist in case there are other blocks
804 // that match it with a smaller number of instructions.
810 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
812 if (!EnableTailMerge) return false;
814 bool MadeChange = false;
816 // First find blocks with no successors.
817 MergePotentials.clear();
818 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
819 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
820 if (TriedMerging.count(I))
823 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
826 // If this is a large problem, avoid visiting the same basic blocks
828 if (MergePotentials.size() == TailMergeThreshold)
829 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
830 TriedMerging.insert(MergePotentials[i].getBlock());
831 // See if we can do any tail merging on those.
832 if (MergePotentials.size() >= 2)
833 MadeChange |= TryTailMergeBlocks(NULL, NULL);
835 // Look at blocks (IBB) with multiple predecessors (PBB).
836 // We change each predecessor to a canonical form, by
837 // (1) temporarily removing any unconditional branch from the predecessor
839 // (2) alter conditional branches so they branch to the other block
840 // not IBB; this may require adding back an unconditional branch to IBB
841 // later, where there wasn't one coming in. E.g.
843 // fallthrough to QBB
846 // with a conceptual B to IBB after that, which never actually exists.
847 // With those changes, we see whether the predecessors' tails match,
848 // and merge them if so. We change things out of canonical form and
849 // back to the way they were later in the process. (OptimizeBranches
850 // would undo some of this, but we can't use it, because we'd get into
851 // a compile-time infinite loop repeatedly doing and undoing the same
854 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
856 if (I->pred_size() >= 2) {
857 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
858 MachineBasicBlock *IBB = I;
859 MachineBasicBlock *PredBB = prior(I);
860 MergePotentials.clear();
861 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
863 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
864 MachineBasicBlock *PBB = *P;
865 if (TriedMerging.count(PBB))
867 // Skip blocks that loop to themselves, can't tail merge these.
870 // Visit each predecessor only once.
871 if (!UniquePreds.insert(PBB))
873 MachineBasicBlock *TBB = 0, *FBB = 0;
874 SmallVector<MachineOperand, 4> Cond;
875 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
876 // Failing case: IBB is the target of a cbr, and
877 // we cannot reverse the branch.
878 SmallVector<MachineOperand, 4> NewCond(Cond);
879 if (!Cond.empty() && TBB == IBB) {
880 if (TII->ReverseBranchCondition(NewCond))
882 // This is the QBB case described above
884 FBB = llvm::next(MachineFunction::iterator(PBB));
886 // Failing case: the only way IBB can be reached from PBB is via
887 // exception handling. Happens for landing pads. Would be nice
888 // to have a bit in the edge so we didn't have to do all this.
889 if (IBB->isLandingPad()) {
890 MachineFunction::iterator IP = PBB; IP++;
891 MachineBasicBlock *PredNextBB = NULL;
895 if (IBB != PredNextBB) // fallthrough
898 if (TBB != IBB && FBB != IBB) // cbr then ubr
900 } else if (Cond.empty()) {
901 if (TBB != IBB) // ubr
904 if (TBB != IBB && IBB != PredNextBB) // cbr
908 // Remove the unconditional branch at the end, if any.
909 if (TBB && (Cond.empty() || FBB)) {
910 DebugLoc dl; // FIXME: this is nowhere
911 TII->RemoveBranch(*PBB);
913 // reinsert conditional branch only, for now
914 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl);
916 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
919 // If this is a large problem, avoid visiting the same basic blocks
921 if (MergePotentials.size() == TailMergeThreshold)
922 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
923 TriedMerging.insert(MergePotentials[i].getBlock());
924 if (MergePotentials.size() >= 2)
925 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
926 // Reinsert an unconditional branch if needed.
927 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
928 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
929 if (MergePotentials.size() == 1 &&
930 MergePotentials.begin()->getBlock() != PredBB)
931 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
937 //===----------------------------------------------------------------------===//
938 // Branch Optimization
939 //===----------------------------------------------------------------------===//
941 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
942 bool MadeChange = false;
944 // Make sure blocks are numbered in order
947 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
949 MachineBasicBlock *MBB = I++;
950 MadeChange |= OptimizeBlock(MBB);
952 // If it is dead, remove it.
953 if (MBB->pred_empty()) {
954 RemoveDeadBlock(MBB);
962 // Blocks should be considered empty if they contain only debug info;
963 // else the debug info would affect codegen.
964 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
967 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
968 MBBI!=MBBE; ++MBBI) {
969 if (!MBBI->isDebugValue())
975 // Blocks with only debug info and branches should be considered the same
976 // as blocks with only branches.
977 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
978 MachineBasicBlock::iterator MBBI, MBBE;
979 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
980 if (!MBBI->isDebugValue())
983 return (MBBI->getDesc().isBranch());
986 /// IsBetterFallthrough - Return true if it would be clearly better to
987 /// fall-through to MBB1 than to fall through into MBB2. This has to return
988 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
989 /// result in infinite loops.
990 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
991 MachineBasicBlock *MBB2) {
992 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
993 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
994 // optimize branches that branch to either a return block or an assert block
995 // into a fallthrough to the return.
996 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
998 // If there is a clear successor ordering we make sure that one block
999 // will fall through to the next
1000 if (MBB1->isSuccessor(MBB2)) return true;
1001 if (MBB2->isSuccessor(MBB1)) return false;
1003 // Neither block consists entirely of debug info (per IsEmptyBlock check),
1004 // so we needn't test for falling off the beginning here.
1005 MachineBasicBlock::iterator MBB1I = --MBB1->end();
1006 while (MBB1I->isDebugValue())
1008 MachineBasicBlock::iterator MBB2I = --MBB2->end();
1009 while (MBB2I->isDebugValue())
1011 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
1014 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1015 /// block. This is never called on the entry block.
1016 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1017 bool MadeChange = false;
1018 MachineFunction &MF = *MBB->getParent();
1019 DebugLoc dl; // FIXME: this is nowhere
1022 MachineFunction::iterator FallThrough = MBB;
1025 // If this block is empty, make everyone use its fall-through, not the block
1026 // explicitly. Landing pads should not do this since the landing-pad table
1027 // points to this block. Blocks with their addresses taken shouldn't be
1029 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1030 // Dead block? Leave for cleanup later.
1031 if (MBB->pred_empty()) return MadeChange;
1033 if (FallThrough == MF.end()) {
1034 // TODO: Simplify preds to not branch here if possible!
1036 // Rewrite all predecessors of the old block to go to the fallthrough
1038 while (!MBB->pred_empty()) {
1039 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1040 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1042 // If MBB was the target of a jump table, update jump tables to go to the
1043 // fallthrough instead.
1044 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1045 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1051 // Check to see if we can simplify the terminator of the block before this
1053 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1055 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1056 SmallVector<MachineOperand, 4> PriorCond;
1057 bool PriorUnAnalyzable =
1058 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1059 if (!PriorUnAnalyzable) {
1060 // If the CFG for the prior block has extra edges, remove them.
1061 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1062 !PriorCond.empty());
1064 // If the previous branch is conditional and both conditions go to the same
1065 // destination, remove the branch, replacing it with an unconditional one or
1067 if (PriorTBB && PriorTBB == PriorFBB) {
1068 TII->RemoveBranch(PrevBB);
1070 if (PriorTBB != MBB)
1071 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1074 goto ReoptimizeBlock;
1077 // If the previous block unconditionally falls through to this block and
1078 // this block has no other predecessors, move the contents of this block
1079 // into the prior block. This doesn't usually happen when SimplifyCFG
1080 // has been used, but it can happen if tail merging splits a fall-through
1081 // predecessor of a block.
1082 // This has to check PrevBB->succ_size() because EH edges are ignored by
1084 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1085 PrevBB.succ_size() == 1 &&
1086 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1087 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1088 << "From MBB: " << *MBB);
1089 // Remove redundant DBG_VALUEs first.
1090 if (PrevBB.begin() != PrevBB.end()) {
1091 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1093 MachineBasicBlock::iterator MBBIter = MBB->begin();
1094 // Check if DBG_VALUE at the end of PrevBB is identical to the
1095 // DBG_VALUE at the beginning of MBB.
1096 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1097 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1098 if (!MBBIter->isIdenticalTo(PrevBBIter))
1100 MachineInstr *DuplicateDbg = MBBIter;
1101 ++MBBIter; -- PrevBBIter;
1102 DuplicateDbg->eraseFromParent();
1105 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1106 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1107 assert(PrevBB.succ_empty());
1108 PrevBB.transferSuccessors(MBB);
1113 // If the previous branch *only* branches to *this* block (conditional or
1114 // not) remove the branch.
1115 if (PriorTBB == MBB && PriorFBB == 0) {
1116 TII->RemoveBranch(PrevBB);
1119 goto ReoptimizeBlock;
1122 // If the prior block branches somewhere else on the condition and here if
1123 // the condition is false, remove the uncond second branch.
1124 if (PriorFBB == MBB) {
1125 TII->RemoveBranch(PrevBB);
1126 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1129 goto ReoptimizeBlock;
1132 // If the prior block branches here on true and somewhere else on false, and
1133 // if the branch condition is reversible, reverse the branch to create a
1135 if (PriorTBB == MBB) {
1136 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1137 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1138 TII->RemoveBranch(PrevBB);
1139 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl);
1142 goto ReoptimizeBlock;
1146 // If this block has no successors (e.g. it is a return block or ends with
1147 // a call to a no-return function like abort or __cxa_throw) and if the pred
1148 // falls through into this block, and if it would otherwise fall through
1149 // into the block after this, move this block to the end of the function.
1151 // We consider it more likely that execution will stay in the function (e.g.
1152 // due to loops) than it is to exit it. This asserts in loops etc, moving
1153 // the assert condition out of the loop body.
1154 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1155 MachineFunction::iterator(PriorTBB) == FallThrough &&
1156 !MBB->canFallThrough()) {
1157 bool DoTransform = true;
1159 // We have to be careful that the succs of PredBB aren't both no-successor
1160 // blocks. If neither have successors and if PredBB is the second from
1161 // last block in the function, we'd just keep swapping the two blocks for
1162 // last. Only do the swap if one is clearly better to fall through than
1164 if (FallThrough == --MF.end() &&
1165 !IsBetterFallthrough(PriorTBB, MBB))
1166 DoTransform = false;
1169 // Reverse the branch so we will fall through on the previous true cond.
1170 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1171 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1172 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1173 << "To make fallthrough to: " << *PriorTBB << "\n");
1175 TII->RemoveBranch(PrevBB);
1176 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl);
1178 // Move this block to the end of the function.
1179 MBB->moveAfter(--MF.end());
1188 // Analyze the branch in the current block.
1189 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1190 SmallVector<MachineOperand, 4> CurCond;
1191 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1192 if (!CurUnAnalyzable) {
1193 // If the CFG for the prior block has extra edges, remove them.
1194 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1196 // If this is a two-way branch, and the FBB branches to this block, reverse
1197 // the condition so the single-basic-block loop is faster. Instead of:
1198 // Loop: xxx; jcc Out; jmp Loop
1200 // Loop: xxx; jncc Loop; jmp Out
1201 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1202 SmallVector<MachineOperand, 4> NewCond(CurCond);
1203 if (!TII->ReverseBranchCondition(NewCond)) {
1204 TII->RemoveBranch(*MBB);
1205 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1208 goto ReoptimizeBlock;
1212 // If this branch is the only thing in its block, see if we can forward
1213 // other blocks across it.
1214 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1215 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1216 !MBB->hasAddressTaken()) {
1217 // This block may contain just an unconditional branch. Because there can
1218 // be 'non-branch terminators' in the block, try removing the branch and
1219 // then seeing if the block is empty.
1220 TII->RemoveBranch(*MBB);
1221 // If the only things remaining in the block are debug info, remove these
1222 // as well, so this will behave the same as an empty block in non-debug
1224 if (!MBB->empty()) {
1225 bool NonDebugInfoFound = false;
1226 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1228 if (!I->isDebugValue()) {
1229 NonDebugInfoFound = true;
1233 if (!NonDebugInfoFound)
1234 // Make the block empty, losing the debug info (we could probably
1235 // improve this in some cases.)
1236 MBB->erase(MBB->begin(), MBB->end());
1238 // If this block is just an unconditional branch to CurTBB, we can
1239 // usually completely eliminate the block. The only case we cannot
1240 // completely eliminate the block is when the block before this one
1241 // falls through into MBB and we can't understand the prior block's branch
1244 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1245 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1246 !PrevBB.isSuccessor(MBB)) {
1247 // If the prior block falls through into us, turn it into an
1248 // explicit branch to us to make updates simpler.
1249 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1250 PriorTBB != MBB && PriorFBB != MBB) {
1251 if (PriorTBB == 0) {
1252 assert(PriorCond.empty() && PriorFBB == 0 &&
1253 "Bad branch analysis");
1256 assert(PriorFBB == 0 && "Machine CFG out of date!");
1259 TII->RemoveBranch(PrevBB);
1260 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, dl);
1263 // Iterate through all the predecessors, revectoring each in-turn.
1265 bool DidChange = false;
1266 bool HasBranchToSelf = false;
1267 while(PI != MBB->pred_size()) {
1268 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1270 // If this block has an uncond branch to itself, leave it.
1272 HasBranchToSelf = true;
1275 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1276 // If this change resulted in PMBB ending in a conditional
1277 // branch where both conditions go to the same destination,
1278 // change this to an unconditional branch (and fix the CFG).
1279 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1280 SmallVector<MachineOperand, 4> NewCurCond;
1281 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1282 NewCurFBB, NewCurCond, true);
1283 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1284 TII->RemoveBranch(*PMBB);
1286 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, dl);
1289 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1294 // Change any jumptables to go to the new MBB.
1295 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1296 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1300 if (!HasBranchToSelf) return MadeChange;
1305 // Add the branch back if the block is more than just an uncond branch.
1306 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl);
1310 // If the prior block doesn't fall through into this block, and if this
1311 // block doesn't fall through into some other block, see if we can find a
1312 // place to move this block where a fall-through will happen.
1313 if (!PrevBB.canFallThrough()) {
1315 // Now we know that there was no fall-through into this block, check to
1316 // see if it has a fall-through into its successor.
1317 bool CurFallsThru = MBB->canFallThrough();
1319 if (!MBB->isLandingPad()) {
1320 // Check all the predecessors of this block. If one of them has no fall
1321 // throughs, move this block right after it.
1322 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1323 E = MBB->pred_end(); PI != E; ++PI) {
1324 // Analyze the branch at the end of the pred.
1325 MachineBasicBlock *PredBB = *PI;
1326 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1327 MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
1328 SmallVector<MachineOperand, 4> PredCond;
1329 if (PredBB != MBB && !PredBB->canFallThrough() &&
1330 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1331 && (!CurFallsThru || !CurTBB || !CurFBB)
1332 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1333 // If the current block doesn't fall through, just move it.
1334 // If the current block can fall through and does not end with a
1335 // conditional branch, we need to append an unconditional jump to
1336 // the (current) next block. To avoid a possible compile-time
1337 // infinite loop, move blocks only backward in this case.
1338 // Also, if there are already 2 branches here, we cannot add a third;
1339 // this means we have the case
1344 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
1346 TII->InsertBranch(*MBB, NextBB, 0, CurCond, dl);
1348 MBB->moveAfter(PredBB);
1350 goto ReoptimizeBlock;
1355 if (!CurFallsThru) {
1356 // Check all successors to see if we can move this block before it.
1357 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1358 E = MBB->succ_end(); SI != E; ++SI) {
1359 // Analyze the branch at the end of the block before the succ.
1360 MachineBasicBlock *SuccBB = *SI;
1361 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1363 // If this block doesn't already fall-through to that successor, and if
1364 // the succ doesn't already have a block that can fall through into it,
1365 // and if the successor isn't an EH destination, we can arrange for the
1366 // fallthrough to happen.
1367 if (SuccBB != MBB && &*SuccPrev != MBB &&
1368 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1369 !SuccBB->isLandingPad()) {
1370 MBB->moveBefore(SuccBB);
1372 goto ReoptimizeBlock;
1376 // Okay, there is no really great place to put this block. If, however,
1377 // the block before this one would be a fall-through if this block were
1378 // removed, move this block to the end of the function.
1379 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
1380 SmallVector<MachineOperand, 4> PrevCond;
1381 if (FallThrough != MF.end() &&
1382 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1383 PrevBB.isSuccessor(FallThrough)) {
1384 MBB->moveAfter(--MF.end());
1394 //===----------------------------------------------------------------------===//
1395 // Hoist Common Code
1396 //===----------------------------------------------------------------------===//
1398 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1399 /// blocks to their common predecessor.
1400 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1401 bool MadeChange = false;
1402 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1403 MachineBasicBlock *MBB = I++;
1404 MadeChange |= HoistCommonCodeInSuccs(MBB);
1410 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1411 /// its 'true' successor.
1412 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1413 MachineBasicBlock *TrueBB) {
1414 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1415 E = BB->succ_end(); SI != E; ++SI) {
1416 MachineBasicBlock *SuccBB = *SI;
1417 if (SuccBB != TrueBB)
1423 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1424 /// in successors to. The location is ususally just before the terminator,
1425 /// however if the terminator is a conditional branch and its previous
1426 /// instruction is the flag setting instruction, the previous instruction is
1427 /// the preferred location. This function also gathers uses and defs of the
1428 /// instructions from the insertion point to the end of the block. The data is
1429 /// used by HoistCommonCodeInSuccs to ensure safety.
1431 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1432 const TargetInstrInfo *TII,
1433 const TargetRegisterInfo *TRI,
1434 SmallSet<unsigned,4> &Uses,
1435 SmallSet<unsigned,4> &Defs) {
1436 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1437 if (!TII->isUnpredicatedTerminator(Loc))
1440 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1441 const MachineOperand &MO = Loc->getOperand(i);
1444 unsigned Reg = MO.getReg();
1449 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
1451 } else if (!MO.isDead())
1452 // Don't try to hoist code in the rare case the terminator defines a
1453 // register that is later used.
1459 if (Loc == MBB->begin())
1462 // The terminator is probably a conditional branch, try not to separate the
1463 // branch from condition setting instruction.
1464 MachineBasicBlock::iterator PI = Loc;
1466 while (PI != MBB->begin() && Loc->isDebugValue())
1470 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1471 const MachineOperand &MO = PI->getOperand(i);
1472 if (!MO.isReg() || MO.isUse())
1474 unsigned Reg = MO.getReg();
1477 if (Uses.count(Reg))
1481 // The condition setting instruction is not just before the conditional
1485 // Be conservative, don't insert instruction above something that may have
1486 // side-effects. And since it's potentially bad to separate flag setting
1487 // instruction from the conditional branch, just abort the optimization
1489 // Also avoid moving code above predicated instruction since it's hard to
1490 // reason about register liveness with predicated instruction.
1491 bool DontMoveAcrossStore = true;
1492 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) ||
1493 TII->isPredicated(PI))
1497 // Find out what registers are live. Note this routine is ignoring other live
1498 // registers which are only used by instructions in successor blocks.
1499 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1500 const MachineOperand &MO = PI->getOperand(i);
1503 unsigned Reg = MO.getReg();
1508 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
1511 if (Uses.count(Reg)) {
1513 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1514 Uses.erase(*SR); // Use getSubRegisters to be conservative
1517 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
1525 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1526 /// sequence at the start of the function, move the instructions before MBB
1527 /// terminator if it's legal.
1528 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1529 MachineBasicBlock *TBB = 0, *FBB = 0;
1530 SmallVector<MachineOperand, 4> Cond;
1531 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1534 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1536 // Malformed bcc? True and false blocks are the same?
1539 // Restrict the optimization to cases where MBB is the only predecessor,
1540 // it is an obvious win.
1541 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1544 // Find a suitable position to hoist the common instructions to. Also figure
1545 // out which registers are used or defined by instructions from the insertion
1546 // point to the end of the block.
1547 SmallSet<unsigned, 4> Uses, Defs;
1548 MachineBasicBlock::iterator Loc =
1549 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1550 if (Loc == MBB->end())
1553 bool HasDups = false;
1554 SmallVector<unsigned, 4> LocalDefs;
1555 SmallSet<unsigned, 4> LocalDefsSet;
1556 MachineBasicBlock::iterator TIB = TBB->begin();
1557 MachineBasicBlock::iterator FIB = FBB->begin();
1558 MachineBasicBlock::iterator TIE = TBB->end();
1559 MachineBasicBlock::iterator FIE = FBB->end();
1560 while (TIB != TIE && FIB != FIE) {
1561 // Skip dbg_value instructions. These do not count.
1562 if (TIB->isDebugValue()) {
1563 while (TIB != TIE && TIB->isDebugValue())
1568 if (FIB->isDebugValue()) {
1569 while (FIB != FIE && FIB->isDebugValue())
1574 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1577 if (TII->isPredicated(TIB))
1578 // Hard to reason about register liveness with predicated instruction.
1582 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1583 MachineOperand &MO = TIB->getOperand(i);
1586 unsigned Reg = MO.getReg();
1590 if (Uses.count(Reg)) {
1591 // Avoid clobbering a register that's used by the instruction at
1592 // the point of insertion.
1597 if (Defs.count(Reg) && !MO.isDead()) {
1598 // Don't hoist the instruction if the def would be clobber by the
1599 // instruction at the point insertion. FIXME: This is overly
1600 // conservative. It should be possible to hoist the instructions
1601 // in BB2 in the following example:
1603 // r1, eflag = op1 r2, r3
1612 } else if (!LocalDefsSet.count(Reg)) {
1613 if (Defs.count(Reg)) {
1614 // Use is defined by the instruction at the point of insertion.
1623 bool DontMoveAcrossStore = true;
1624 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore))
1627 // Track local defs so we can update liveins.
1628 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1629 MachineOperand &MO = TIB->getOperand(i);
1632 unsigned Reg = MO.getReg();
1637 LocalDefs.push_back(Reg);
1638 LocalDefsSet.insert(Reg);
1639 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1640 LocalDefsSet.insert(*SR);
1642 } else if (MO.isKill() && LocalDefsSet.count(Reg)) {
1643 LocalDefsSet.erase(Reg);
1644 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1645 LocalDefsSet.erase(*SR);
1657 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1658 FBB->erase(FBB->begin(), FIB);
1661 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1662 unsigned Def = LocalDefs[i];
1663 if (LocalDefsSet.count(Def)) {
1664 TBB->addLiveIn(Def);
1665 FBB->addLiveIn(Def);