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 = new RegScavenger();
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) {
371 RS->enterBasicBlock(CurMBB);
372 if (!CurMBB->empty())
373 RS->forward(prior(CurMBB->end()));
374 BitVector RegsLiveAtExit(TRI->getNumRegs());
375 RS->getRegsUsed(RegsLiveAtExit, false);
376 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
377 if (RegsLiveAtExit[i])
378 NewMBB->addLiveIn(i);
381 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
382 /// after it, replacing it with an unconditional branch to NewDest.
383 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
384 MachineBasicBlock *NewDest) {
385 MachineBasicBlock *CurMBB = OldInst->getParent();
387 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
389 // For targets that use the register scavenger, we must maintain LiveIns.
390 MaintainLiveIns(CurMBB, NewDest);
395 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
396 /// MBB so that the part before the iterator falls into the part starting at the
397 /// iterator. This returns the new MBB.
398 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
399 MachineBasicBlock::iterator BBI1) {
400 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
403 MachineFunction &MF = *CurMBB.getParent();
405 // Create the fall-through block.
406 MachineFunction::iterator MBBI = &CurMBB;
407 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
408 CurMBB.getParent()->insert(++MBBI, NewMBB);
410 // Move all the successors of this block to the specified block.
411 NewMBB->transferSuccessors(&CurMBB);
413 // Add an edge from CurMBB to NewMBB for the fall-through.
414 CurMBB.addSuccessor(NewMBB);
416 // Splice the code over.
417 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
419 // For targets that use the register scavenger, we must maintain LiveIns.
420 MaintainLiveIns(&CurMBB, NewMBB);
425 /// EstimateRuntime - Make a rough estimate for how long it will take to run
426 /// the specified code.
427 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
428 MachineBasicBlock::iterator E) {
430 for (; I != E; ++I) {
431 if (I->isDebugValue())
435 else if (I->mayLoad() || I->mayStore())
443 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
444 // branches temporarily for tail merging). In the case where CurMBB ends
445 // with a conditional branch to the next block, optimize by reversing the
446 // test and conditionally branching to SuccMBB instead.
447 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
448 const TargetInstrInfo *TII) {
449 MachineFunction *MF = CurMBB->getParent();
450 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB));
451 MachineBasicBlock *TBB = 0, *FBB = 0;
452 SmallVector<MachineOperand, 4> Cond;
453 DebugLoc dl; // FIXME: this is nowhere
454 if (I != MF->end() &&
455 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
456 MachineBasicBlock *NextBB = I;
457 if (TBB == NextBB && !Cond.empty() && !FBB) {
458 if (!TII->ReverseBranchCondition(Cond)) {
459 TII->RemoveBranch(*CurMBB);
460 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl);
465 TII->InsertBranch(*CurMBB, SuccBB, NULL,
466 SmallVector<MachineOperand, 0>(), dl);
470 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
471 if (getHash() < o.getHash())
473 else if (getHash() > o.getHash())
475 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
477 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
480 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
481 // an object with itself.
482 #ifndef _GLIBCXX_DEBUG
483 llvm_unreachable("Predecessor appears twice");
489 /// CountTerminators - Count the number of terminators in the given
490 /// block and set I to the position of the first non-terminator, if there
491 /// is one, or MBB->end() otherwise.
492 static unsigned CountTerminators(MachineBasicBlock *MBB,
493 MachineBasicBlock::iterator &I) {
495 unsigned NumTerms = 0;
497 if (I == MBB->begin()) {
502 if (!I->isTerminator()) break;
508 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
509 /// and decide if it would be profitable to merge those tails. Return the
510 /// length of the common tail and iterators to the first common instruction
512 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
513 MachineBasicBlock *MBB2,
514 unsigned minCommonTailLength,
515 unsigned &CommonTailLen,
516 MachineBasicBlock::iterator &I1,
517 MachineBasicBlock::iterator &I2,
518 MachineBasicBlock *SuccBB,
519 MachineBasicBlock *PredBB) {
520 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
521 if (CommonTailLen == 0)
523 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
524 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
527 // It's almost always profitable to merge any number of non-terminator
528 // instructions with the block that falls through into the common successor.
529 if (MBB1 == PredBB || MBB2 == PredBB) {
530 MachineBasicBlock::iterator I;
531 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
532 if (CommonTailLen > NumTerms)
536 // If one of the blocks can be completely merged and happens to be in
537 // a position where the other could fall through into it, merge any number
538 // of instructions, because it can be done without a branch.
539 // TODO: If the blocks are not adjacent, move one of them so that they are?
540 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
542 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
545 // If both blocks have an unconditional branch temporarily stripped out,
546 // count that as an additional common instruction for the following
548 unsigned EffectiveTailLen = CommonTailLen;
549 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
550 !MBB1->back().isBarrier() &&
551 !MBB2->back().isBarrier())
554 // Check if the common tail is long enough to be worthwhile.
555 if (EffectiveTailLen >= minCommonTailLength)
558 // If we are optimizing for code size, 2 instructions in common is enough if
559 // we don't have to split a block. At worst we will be introducing 1 new
560 // branch instruction, which is likely to be smaller than the 2
561 // instructions that would be deleted in the merge.
562 MachineFunction *MF = MBB1->getParent();
563 if (EffectiveTailLen >= 2 &&
564 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
565 (I1 == MBB1->begin() || I2 == MBB2->begin()))
571 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
572 /// hash CurHash (guaranteed to match the last element). Build the vector
573 /// SameTails of all those that have the (same) largest number of instructions
574 /// in common of any pair of these blocks. SameTails entries contain an
575 /// iterator into MergePotentials (from which the MachineBasicBlock can be
576 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
577 /// instruction where the matching code sequence begins.
578 /// Order of elements in SameTails is the reverse of the order in which
579 /// those blocks appear in MergePotentials (where they are not necessarily
581 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
582 unsigned minCommonTailLength,
583 MachineBasicBlock *SuccBB,
584 MachineBasicBlock *PredBB) {
585 unsigned maxCommonTailLength = 0U;
587 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
588 MPIterator HighestMPIter = prior(MergePotentials.end());
589 for (MPIterator CurMPIter = prior(MergePotentials.end()),
590 B = MergePotentials.begin();
591 CurMPIter != B && CurMPIter->getHash() == CurHash;
593 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
594 unsigned CommonTailLen;
595 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
597 CommonTailLen, TrialBBI1, TrialBBI2,
599 if (CommonTailLen > maxCommonTailLength) {
601 maxCommonTailLength = CommonTailLen;
602 HighestMPIter = CurMPIter;
603 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
605 if (HighestMPIter == CurMPIter &&
606 CommonTailLen == maxCommonTailLength)
607 SameTails.push_back(SameTailElt(I, TrialBBI2));
613 return maxCommonTailLength;
616 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
617 /// MergePotentials, restoring branches at ends of blocks as appropriate.
618 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
619 MachineBasicBlock *SuccBB,
620 MachineBasicBlock *PredBB) {
621 MPIterator CurMPIter, B;
622 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
623 CurMPIter->getHash() == CurHash;
625 // Put the unconditional branch back, if we need one.
626 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
627 if (SuccBB && CurMBB != PredBB)
628 FixTail(CurMBB, SuccBB, TII);
632 if (CurMPIter->getHash() != CurHash)
634 MergePotentials.erase(CurMPIter, MergePotentials.end());
637 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
638 /// only of the common tail. Create a block that does by splitting one.
639 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
640 unsigned maxCommonTailLength,
641 unsigned &commonTailIndex) {
643 unsigned TimeEstimate = ~0U;
644 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
645 // Use PredBB if possible; that doesn't require a new branch.
646 if (SameTails[i].getBlock() == PredBB) {
650 // Otherwise, make a (fairly bogus) choice based on estimate of
651 // how long it will take the various blocks to execute.
652 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
653 SameTails[i].getTailStartPos());
654 if (t <= TimeEstimate) {
660 MachineBasicBlock::iterator BBI =
661 SameTails[commonTailIndex].getTailStartPos();
662 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
664 // If the common tail includes any debug info we will take it pretty
665 // randomly from one of the inputs. Might be better to remove it?
666 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
667 << maxCommonTailLength);
669 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
671 DEBUG(dbgs() << "... failed!");
675 SameTails[commonTailIndex].setBlock(newMBB);
676 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
678 // If we split PredBB, newMBB is the new predecessor.
685 // See if any of the blocks in MergePotentials (which all have a common single
686 // successor, or all have no successor) can be tail-merged. If there is a
687 // successor, any blocks in MergePotentials that are not tail-merged and
688 // are not immediately before Succ must have an unconditional branch to
689 // Succ added (but the predecessor/successor lists need no adjustment).
690 // The lone predecessor of Succ that falls through into Succ,
691 // if any, is given in PredBB.
693 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
694 MachineBasicBlock *PredBB) {
695 bool MadeChange = false;
697 // Except for the special cases below, tail-merge if there are at least
698 // this many instructions in common.
699 unsigned minCommonTailLength = TailMergeSize;
701 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
702 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
703 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
704 << (i == e-1 ? "" : ", ");
707 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
709 dbgs() << " which has fall-through from BB#"
710 << PredBB->getNumber() << "\n";
712 dbgs() << "Looking for common tails of at least "
713 << minCommonTailLength << " instruction"
714 << (minCommonTailLength == 1 ? "" : "s") << '\n';
717 // Sort by hash value so that blocks with identical end sequences sort
719 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
721 // Walk through equivalence sets looking for actual exact matches.
722 while (MergePotentials.size() > 1) {
723 unsigned CurHash = MergePotentials.back().getHash();
725 // Build SameTails, identifying the set of blocks with this hash code
726 // and with the maximum number of instructions in common.
727 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
731 // If we didn't find any pair that has at least minCommonTailLength
732 // instructions in common, remove all blocks with this hash code and retry.
733 if (SameTails.empty()) {
734 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
738 // If one of the blocks is the entire common tail (and not the entry
739 // block, which we can't jump to), we can treat all blocks with this same
740 // tail at once. Use PredBB if that is one of the possibilities, as that
741 // will not introduce any extra branches.
742 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
743 getParent()->begin();
744 unsigned commonTailIndex = SameTails.size();
745 // If there are two blocks, check to see if one can be made to fall through
747 if (SameTails.size() == 2 &&
748 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
749 SameTails[1].tailIsWholeBlock())
751 else if (SameTails.size() == 2 &&
752 SameTails[1].getBlock()->isLayoutSuccessor(
753 SameTails[0].getBlock()) &&
754 SameTails[0].tailIsWholeBlock())
757 // Otherwise just pick one, favoring the fall-through predecessor if
759 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
760 MachineBasicBlock *MBB = SameTails[i].getBlock();
761 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
767 if (SameTails[i].tailIsWholeBlock())
772 if (commonTailIndex == SameTails.size() ||
773 (SameTails[commonTailIndex].getBlock() == PredBB &&
774 !SameTails[commonTailIndex].tailIsWholeBlock())) {
775 // None of the blocks consist entirely of the common tail.
776 // Split a block so that one does.
777 if (!CreateCommonTailOnlyBlock(PredBB,
778 maxCommonTailLength, commonTailIndex)) {
779 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
784 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
785 // MBB is common tail. Adjust all other BB's to jump to this one.
786 // Traversal must be forwards so erases work.
787 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
789 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
790 if (commonTailIndex == i)
792 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
793 << (i == e-1 ? "" : ", "));
794 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
795 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
796 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
797 MergePotentials.erase(SameTails[i].getMPIter());
799 DEBUG(dbgs() << "\n");
800 // We leave commonTailIndex in the worklist in case there are other blocks
801 // that match it with a smaller number of instructions.
807 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
809 if (!EnableTailMerge) return false;
811 bool MadeChange = false;
813 // First find blocks with no successors.
814 MergePotentials.clear();
815 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
816 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
817 if (TriedMerging.count(I))
820 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
823 // If this is a large problem, avoid visiting the same basic blocks
825 if (MergePotentials.size() == TailMergeThreshold)
826 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
827 TriedMerging.insert(MergePotentials[i].getBlock());
828 // See if we can do any tail merging on those.
829 if (MergePotentials.size() >= 2)
830 MadeChange |= TryTailMergeBlocks(NULL, NULL);
832 // Look at blocks (IBB) with multiple predecessors (PBB).
833 // We change each predecessor to a canonical form, by
834 // (1) temporarily removing any unconditional branch from the predecessor
836 // (2) alter conditional branches so they branch to the other block
837 // not IBB; this may require adding back an unconditional branch to IBB
838 // later, where there wasn't one coming in. E.g.
840 // fallthrough to QBB
843 // with a conceptual B to IBB after that, which never actually exists.
844 // With those changes, we see whether the predecessors' tails match,
845 // and merge them if so. We change things out of canonical form and
846 // back to the way they were later in the process. (OptimizeBranches
847 // would undo some of this, but we can't use it, because we'd get into
848 // a compile-time infinite loop repeatedly doing and undoing the same
851 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
853 if (I->pred_size() >= 2) {
854 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
855 MachineBasicBlock *IBB = I;
856 MachineBasicBlock *PredBB = prior(I);
857 MergePotentials.clear();
858 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
860 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
861 MachineBasicBlock *PBB = *P;
862 if (TriedMerging.count(PBB))
864 // Skip blocks that loop to themselves, can't tail merge these.
867 // Visit each predecessor only once.
868 if (!UniquePreds.insert(PBB))
870 // Skip blocks which may jump to a landing pad. Can't tail merge these.
871 if (PBB->getLandingPadSuccessor())
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
928 // TryTailMergeBlocks.
929 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
930 if (MergePotentials.size() == 1 &&
931 MergePotentials.begin()->getBlock() != PredBB)
932 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
938 //===----------------------------------------------------------------------===//
939 // Branch Optimization
940 //===----------------------------------------------------------------------===//
942 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
943 bool MadeChange = false;
945 // Make sure blocks are numbered in order
948 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
950 MachineBasicBlock *MBB = I++;
951 MadeChange |= OptimizeBlock(MBB);
953 // If it is dead, remove it.
954 if (MBB->pred_empty()) {
955 RemoveDeadBlock(MBB);
963 // Blocks should be considered empty if they contain only debug info;
964 // else the debug info would affect codegen.
965 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
968 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
969 MBBI!=MBBE; ++MBBI) {
970 if (!MBBI->isDebugValue())
976 // Blocks with only debug info and branches should be considered the same
977 // as blocks with only branches.
978 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
979 MachineBasicBlock::iterator MBBI, MBBE;
980 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
981 if (!MBBI->isDebugValue())
984 return (MBBI->isBranch());
987 /// IsBetterFallthrough - Return true if it would be clearly better to
988 /// fall-through to MBB1 than to fall through into MBB2. This has to return
989 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
990 /// result in infinite loops.
991 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
992 MachineBasicBlock *MBB2) {
993 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
994 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
995 // optimize branches that branch to either a return block or an assert block
996 // into a fallthrough to the return.
997 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
999 // If there is a clear successor ordering we make sure that one block
1000 // will fall through to the next
1001 if (MBB1->isSuccessor(MBB2)) return true;
1002 if (MBB2->isSuccessor(MBB1)) return false;
1004 // Neither block consists entirely of debug info (per IsEmptyBlock check),
1005 // so we needn't test for falling off the beginning here.
1006 MachineBasicBlock::iterator MBB1I = --MBB1->end();
1007 while (MBB1I->isDebugValue())
1009 MachineBasicBlock::iterator MBB2I = --MBB2->end();
1010 while (MBB2I->isDebugValue())
1012 return MBB2I->isCall() && !MBB1I->isCall();
1015 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1016 /// block. This is never called on the entry block.
1017 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1018 bool MadeChange = false;
1019 MachineFunction &MF = *MBB->getParent();
1020 DebugLoc dl; // FIXME: this is nowhere
1023 MachineFunction::iterator FallThrough = MBB;
1026 // If this block is empty, make everyone use its fall-through, not the block
1027 // explicitly. Landing pads should not do this since the landing-pad table
1028 // points to this block. Blocks with their addresses taken shouldn't be
1030 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1031 // Dead block? Leave for cleanup later.
1032 if (MBB->pred_empty()) return MadeChange;
1034 if (FallThrough == MF.end()) {
1035 // TODO: Simplify preds to not branch here if possible!
1037 // Rewrite all predecessors of the old block to go to the fallthrough
1039 while (!MBB->pred_empty()) {
1040 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1041 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1043 // If MBB was the target of a jump table, update jump tables to go to the
1044 // fallthrough instead.
1045 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1046 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1052 // Check to see if we can simplify the terminator of the block before this
1054 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1056 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1057 SmallVector<MachineOperand, 4> PriorCond;
1058 bool PriorUnAnalyzable =
1059 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1060 if (!PriorUnAnalyzable) {
1061 // If the CFG for the prior block has extra edges, remove them.
1062 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1063 !PriorCond.empty());
1065 // If the previous branch is conditional and both conditions go to the same
1066 // destination, remove the branch, replacing it with an unconditional one or
1068 if (PriorTBB && PriorTBB == PriorFBB) {
1069 TII->RemoveBranch(PrevBB);
1071 if (PriorTBB != MBB)
1072 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1075 goto ReoptimizeBlock;
1078 // If the previous block unconditionally falls through to this block and
1079 // this block has no other predecessors, move the contents of this block
1080 // into the prior block. This doesn't usually happen when SimplifyCFG
1081 // has been used, but it can happen if tail merging splits a fall-through
1082 // predecessor of a block.
1083 // This has to check PrevBB->succ_size() because EH edges are ignored by
1085 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1086 PrevBB.succ_size() == 1 &&
1087 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1088 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1089 << "From MBB: " << *MBB);
1090 // Remove redundant DBG_VALUEs first.
1091 if (PrevBB.begin() != PrevBB.end()) {
1092 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1094 MachineBasicBlock::iterator MBBIter = MBB->begin();
1095 // Check if DBG_VALUE at the end of PrevBB is identical to the
1096 // DBG_VALUE at the beginning of MBB.
1097 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1098 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1099 if (!MBBIter->isIdenticalTo(PrevBBIter))
1101 MachineInstr *DuplicateDbg = MBBIter;
1102 ++MBBIter; -- PrevBBIter;
1103 DuplicateDbg->eraseFromParent();
1106 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1107 PrevBB.removeSuccessor(PrevBB.succ_begin());;
1108 assert(PrevBB.succ_empty());
1109 PrevBB.transferSuccessors(MBB);
1114 // If the previous branch *only* branches to *this* block (conditional or
1115 // not) remove the branch.
1116 if (PriorTBB == MBB && PriorFBB == 0) {
1117 TII->RemoveBranch(PrevBB);
1120 goto ReoptimizeBlock;
1123 // If the prior block branches somewhere else on the condition and here if
1124 // the condition is false, remove the uncond second branch.
1125 if (PriorFBB == MBB) {
1126 TII->RemoveBranch(PrevBB);
1127 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1130 goto ReoptimizeBlock;
1133 // If the prior block branches here on true and somewhere else on false, and
1134 // if the branch condition is reversible, reverse the branch to create a
1136 if (PriorTBB == MBB) {
1137 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1138 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1139 TII->RemoveBranch(PrevBB);
1140 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl);
1143 goto ReoptimizeBlock;
1147 // If this block has no successors (e.g. it is a return block or ends with
1148 // a call to a no-return function like abort or __cxa_throw) and if the pred
1149 // falls through into this block, and if it would otherwise fall through
1150 // into the block after this, move this block to the end of the function.
1152 // We consider it more likely that execution will stay in the function (e.g.
1153 // due to loops) than it is to exit it. This asserts in loops etc, moving
1154 // the assert condition out of the loop body.
1155 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1156 MachineFunction::iterator(PriorTBB) == FallThrough &&
1157 !MBB->canFallThrough()) {
1158 bool DoTransform = true;
1160 // We have to be careful that the succs of PredBB aren't both no-successor
1161 // blocks. If neither have successors and if PredBB is the second from
1162 // last block in the function, we'd just keep swapping the two blocks for
1163 // last. Only do the swap if one is clearly better to fall through than
1165 if (FallThrough == --MF.end() &&
1166 !IsBetterFallthrough(PriorTBB, MBB))
1167 DoTransform = false;
1170 // Reverse the branch so we will fall through on the previous true cond.
1171 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1172 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1173 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1174 << "To make fallthrough to: " << *PriorTBB << "\n");
1176 TII->RemoveBranch(PrevBB);
1177 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl);
1179 // Move this block to the end of the function.
1180 MBB->moveAfter(--MF.end());
1189 // Analyze the branch in the current block.
1190 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1191 SmallVector<MachineOperand, 4> CurCond;
1192 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1193 if (!CurUnAnalyzable) {
1194 // If the CFG for the prior block has extra edges, remove them.
1195 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1197 // If this is a two-way branch, and the FBB branches to this block, reverse
1198 // the condition so the single-basic-block loop is faster. Instead of:
1199 // Loop: xxx; jcc Out; jmp Loop
1201 // Loop: xxx; jncc Loop; jmp Out
1202 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1203 SmallVector<MachineOperand, 4> NewCond(CurCond);
1204 if (!TII->ReverseBranchCondition(NewCond)) {
1205 TII->RemoveBranch(*MBB);
1206 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1209 goto ReoptimizeBlock;
1213 // If this branch is the only thing in its block, see if we can forward
1214 // other blocks across it.
1215 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1216 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1217 !MBB->hasAddressTaken()) {
1218 // This block may contain just an unconditional branch. Because there can
1219 // be 'non-branch terminators' in the block, try removing the branch and
1220 // then seeing if the block is empty.
1221 TII->RemoveBranch(*MBB);
1222 // If the only things remaining in the block are debug info, remove these
1223 // as well, so this will behave the same as an empty block in non-debug
1225 if (!MBB->empty()) {
1226 bool NonDebugInfoFound = false;
1227 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1229 if (!I->isDebugValue()) {
1230 NonDebugInfoFound = true;
1234 if (!NonDebugInfoFound)
1235 // Make the block empty, losing the debug info (we could probably
1236 // improve this in some cases.)
1237 MBB->erase(MBB->begin(), MBB->end());
1239 // If this block is just an unconditional branch to CurTBB, we can
1240 // usually completely eliminate the block. The only case we cannot
1241 // completely eliminate the block is when the block before this one
1242 // falls through into MBB and we can't understand the prior block's branch
1245 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1246 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1247 !PrevBB.isSuccessor(MBB)) {
1248 // If the prior block falls through into us, turn it into an
1249 // explicit branch to us to make updates simpler.
1250 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1251 PriorTBB != MBB && PriorFBB != MBB) {
1252 if (PriorTBB == 0) {
1253 assert(PriorCond.empty() && PriorFBB == 0 &&
1254 "Bad branch analysis");
1257 assert(PriorFBB == 0 && "Machine CFG out of date!");
1260 TII->RemoveBranch(PrevBB);
1261 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, dl);
1264 // Iterate through all the predecessors, revectoring each in-turn.
1266 bool DidChange = false;
1267 bool HasBranchToSelf = false;
1268 while(PI != MBB->pred_size()) {
1269 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1271 // If this block has an uncond branch to itself, leave it.
1273 HasBranchToSelf = true;
1276 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1277 // If this change resulted in PMBB ending in a conditional
1278 // branch where both conditions go to the same destination,
1279 // change this to an unconditional branch (and fix the CFG).
1280 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1281 SmallVector<MachineOperand, 4> NewCurCond;
1282 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1283 NewCurFBB, NewCurCond, true);
1284 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1285 TII->RemoveBranch(*PMBB);
1287 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, dl);
1290 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1295 // Change any jumptables to go to the new MBB.
1296 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1297 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1301 if (!HasBranchToSelf) return MadeChange;
1306 // Add the branch back if the block is more than just an uncond branch.
1307 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl);
1311 // If the prior block doesn't fall through into this block, and if this
1312 // block doesn't fall through into some other block, see if we can find a
1313 // place to move this block where a fall-through will happen.
1314 if (!PrevBB.canFallThrough()) {
1316 // Now we know that there was no fall-through into this block, check to
1317 // see if it has a fall-through into its successor.
1318 bool CurFallsThru = MBB->canFallThrough();
1320 if (!MBB->isLandingPad()) {
1321 // Check all the predecessors of this block. If one of them has no fall
1322 // throughs, move this block right after it.
1323 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1324 E = MBB->pred_end(); PI != E; ++PI) {
1325 // Analyze the branch at the end of the pred.
1326 MachineBasicBlock *PredBB = *PI;
1327 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1328 MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
1329 SmallVector<MachineOperand, 4> PredCond;
1330 if (PredBB != MBB && !PredBB->canFallThrough() &&
1331 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1332 && (!CurFallsThru || !CurTBB || !CurFBB)
1333 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1334 // If the current block doesn't fall through, just move it.
1335 // If the current block can fall through and does not end with a
1336 // conditional branch, we need to append an unconditional jump to
1337 // the (current) next block. To avoid a possible compile-time
1338 // infinite loop, move blocks only backward in this case.
1339 // Also, if there are already 2 branches here, we cannot add a third;
1340 // this means we have the case
1345 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
1347 TII->InsertBranch(*MBB, NextBB, 0, CurCond, dl);
1349 MBB->moveAfter(PredBB);
1351 goto ReoptimizeBlock;
1356 if (!CurFallsThru) {
1357 // Check all successors to see if we can move this block before it.
1358 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1359 E = MBB->succ_end(); SI != E; ++SI) {
1360 // Analyze the branch at the end of the block before the succ.
1361 MachineBasicBlock *SuccBB = *SI;
1362 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1364 // If this block doesn't already fall-through to that successor, and if
1365 // the succ doesn't already have a block that can fall through into it,
1366 // and if the successor isn't an EH destination, we can arrange for the
1367 // fallthrough to happen.
1368 if (SuccBB != MBB && &*SuccPrev != MBB &&
1369 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1370 !SuccBB->isLandingPad()) {
1371 MBB->moveBefore(SuccBB);
1373 goto ReoptimizeBlock;
1377 // Okay, there is no really great place to put this block. If, however,
1378 // the block before this one would be a fall-through if this block were
1379 // removed, move this block to the end of the function.
1380 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
1381 SmallVector<MachineOperand, 4> PrevCond;
1382 if (FallThrough != MF.end() &&
1383 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1384 PrevBB.isSuccessor(FallThrough)) {
1385 MBB->moveAfter(--MF.end());
1395 //===----------------------------------------------------------------------===//
1396 // Hoist Common Code
1397 //===----------------------------------------------------------------------===//
1399 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1400 /// blocks to their common predecessor.
1401 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1402 bool MadeChange = false;
1403 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1404 MachineBasicBlock *MBB = I++;
1405 MadeChange |= HoistCommonCodeInSuccs(MBB);
1411 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1412 /// its 'true' successor.
1413 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1414 MachineBasicBlock *TrueBB) {
1415 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1416 E = BB->succ_end(); SI != E; ++SI) {
1417 MachineBasicBlock *SuccBB = *SI;
1418 if (SuccBB != TrueBB)
1424 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1425 /// in successors to. The location is ususally just before the terminator,
1426 /// however if the terminator is a conditional branch and its previous
1427 /// instruction is the flag setting instruction, the previous instruction is
1428 /// the preferred location. This function also gathers uses and defs of the
1429 /// instructions from the insertion point to the end of the block. The data is
1430 /// used by HoistCommonCodeInSuccs to ensure safety.
1432 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1433 const TargetInstrInfo *TII,
1434 const TargetRegisterInfo *TRI,
1435 SmallSet<unsigned,4> &Uses,
1436 SmallSet<unsigned,4> &Defs) {
1437 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1438 if (!TII->isUnpredicatedTerminator(Loc))
1441 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1442 const MachineOperand &MO = Loc->getOperand(i);
1445 unsigned Reg = MO.getReg();
1450 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
1452 } else if (!MO.isDead())
1453 // Don't try to hoist code in the rare case the terminator defines a
1454 // register that is later used.
1460 if (Loc == MBB->begin())
1463 // The terminator is probably a conditional branch, try not to separate the
1464 // branch from condition setting instruction.
1465 MachineBasicBlock::iterator PI = Loc;
1467 while (PI != MBB->begin() && Loc->isDebugValue())
1471 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1472 const MachineOperand &MO = PI->getOperand(i);
1473 if (!MO.isReg() || MO.isUse())
1475 unsigned Reg = MO.getReg();
1478 if (Uses.count(Reg))
1482 // The condition setting instruction is not just before the conditional
1486 // Be conservative, don't insert instruction above something that may have
1487 // side-effects. And since it's potentially bad to separate flag setting
1488 // instruction from the conditional branch, just abort the optimization
1490 // Also avoid moving code above predicated instruction since it's hard to
1491 // reason about register liveness with predicated instruction.
1492 bool DontMoveAcrossStore = true;
1493 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) ||
1494 TII->isPredicated(PI))
1498 // Find out what registers are live. Note this routine is ignoring other live
1499 // registers which are only used by instructions in successor blocks.
1500 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1501 const MachineOperand &MO = PI->getOperand(i);
1504 unsigned Reg = MO.getReg();
1509 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
1512 if (Uses.count(Reg)) {
1514 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1515 Uses.erase(*SR); // Use getSubRegisters to be conservative
1518 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
1526 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1527 /// sequence at the start of the function, move the instructions before MBB
1528 /// terminator if it's legal.
1529 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1530 MachineBasicBlock *TBB = 0, *FBB = 0;
1531 SmallVector<MachineOperand, 4> Cond;
1532 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1535 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1537 // Malformed bcc? True and false blocks are the same?
1540 // Restrict the optimization to cases where MBB is the only predecessor,
1541 // it is an obvious win.
1542 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1545 // Find a suitable position to hoist the common instructions to. Also figure
1546 // out which registers are used or defined by instructions from the insertion
1547 // point to the end of the block.
1548 SmallSet<unsigned, 4> Uses, Defs;
1549 MachineBasicBlock::iterator Loc =
1550 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1551 if (Loc == MBB->end())
1554 bool HasDups = false;
1555 SmallVector<unsigned, 4> LocalDefs;
1556 SmallSet<unsigned, 4> LocalDefsSet;
1557 MachineBasicBlock::iterator TIB = TBB->begin();
1558 MachineBasicBlock::iterator FIB = FBB->begin();
1559 MachineBasicBlock::iterator TIE = TBB->end();
1560 MachineBasicBlock::iterator FIE = FBB->end();
1561 while (TIB != TIE && FIB != FIE) {
1562 // Skip dbg_value instructions. These do not count.
1563 if (TIB->isDebugValue()) {
1564 while (TIB != TIE && TIB->isDebugValue())
1569 if (FIB->isDebugValue()) {
1570 while (FIB != FIE && FIB->isDebugValue())
1575 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1578 if (TII->isPredicated(TIB))
1579 // Hard to reason about register liveness with predicated instruction.
1583 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1584 MachineOperand &MO = TIB->getOperand(i);
1587 unsigned Reg = MO.getReg();
1591 if (Uses.count(Reg)) {
1592 // Avoid clobbering a register that's used by the instruction at
1593 // the point of insertion.
1598 if (Defs.count(Reg) && !MO.isDead()) {
1599 // Don't hoist the instruction if the def would be clobber by the
1600 // instruction at the point insertion. FIXME: This is overly
1601 // conservative. It should be possible to hoist the instructions
1602 // in BB2 in the following example:
1604 // r1, eflag = op1 r2, r3
1613 } else if (!LocalDefsSet.count(Reg)) {
1614 if (Defs.count(Reg)) {
1615 // Use is defined by the instruction at the point of insertion.
1624 bool DontMoveAcrossStore = true;
1625 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore))
1628 // Remove kills from LocalDefsSet, these registers had short live ranges.
1629 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1630 MachineOperand &MO = TIB->getOperand(i);
1631 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1633 unsigned Reg = MO.getReg();
1634 if (!Reg || !LocalDefsSet.count(Reg))
1636 for (const unsigned *OR = TRI->getOverlaps(Reg); *OR; ++OR)
1637 LocalDefsSet.erase(*OR);
1640 // Track local defs so we can update liveins.
1641 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1642 MachineOperand &MO = TIB->getOperand(i);
1643 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1645 unsigned Reg = MO.getReg();
1648 LocalDefs.push_back(Reg);
1649 for (const unsigned *OR = TRI->getOverlaps(Reg); *OR; ++OR)
1650 LocalDefsSet.insert(*OR);
1661 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1662 FBB->erase(FBB->begin(), FIB);
1665 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1666 unsigned Def = LocalDefs[i];
1667 if (LocalDefsSet.count(Def)) {
1668 TBB->addLiveIn(Def);
1669 FBB->addLiveIn(Def);