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/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/MachineModuleInfo.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/Passes.h"
29 #include "llvm/CodeGen/RegisterScavenging.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetInstrInfo.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetRegisterInfo.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 {
67 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
69 bool runOnMachineFunction(MachineFunction &MF) override;
71 void getAnalysisUsage(AnalysisUsage &AU) const override {
72 AU.addRequired<TargetPassConfig>();
73 MachineFunctionPass::getAnalysisUsage(AU);
78 char BranchFolderPass::ID = 0;
79 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
81 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
82 "Control Flow Optimizer", false, false)
84 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
85 if (skipOptnoneFunction(*MF.getFunction()))
88 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
89 // TailMerge can create jump into if branches that make CFG irreducible for
90 // HW that requires structurized CFG.
91 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
92 PassConfig->getEnableTailMerge();
93 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true);
94 return Folder.OptimizeFunction(MF,
95 MF.getTarget().getInstrInfo(),
96 MF.getTarget().getRegisterInfo(),
97 getAnalysisIfAvailable<MachineModuleInfo>());
101 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist) {
102 switch (FlagEnableTailMerge) {
103 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
104 case cl::BOU_TRUE: EnableTailMerge = true; break;
105 case cl::BOU_FALSE: EnableTailMerge = false; break;
108 EnableHoistCommonCode = CommonHoist;
111 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
112 /// function, updating the CFG.
113 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
114 assert(MBB->pred_empty() && "MBB must be dead!");
115 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
117 MachineFunction *MF = MBB->getParent();
118 // drop all successors.
119 while (!MBB->succ_empty())
120 MBB->removeSuccessor(MBB->succ_end()-1);
122 // Avoid matching if this pointer gets reused.
123 TriedMerging.erase(MBB);
129 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
130 /// followed by terminators, and if the implicitly defined registers are not
131 /// used by the terminators, remove those implicit_def's. e.g.
133 /// r0 = implicit_def
134 /// r1 = implicit_def
136 /// This block can be optimized away later if the implicit instructions are
138 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
139 SmallSet<unsigned, 4> ImpDefRegs;
140 MachineBasicBlock::iterator I = MBB->begin();
141 while (I != MBB->end()) {
142 if (!I->isImplicitDef())
144 unsigned Reg = I->getOperand(0).getReg();
145 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
146 SubRegs.isValid(); ++SubRegs)
147 ImpDefRegs.insert(*SubRegs);
150 if (ImpDefRegs.empty())
153 MachineBasicBlock::iterator FirstTerm = I;
154 while (I != MBB->end()) {
155 if (!TII->isUnpredicatedTerminator(I))
157 // See if it uses any of the implicitly defined registers.
158 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
159 MachineOperand &MO = I->getOperand(i);
160 if (!MO.isReg() || !MO.isUse())
162 unsigned Reg = MO.getReg();
163 if (ImpDefRegs.count(Reg))
170 while (I != FirstTerm) {
171 MachineInstr *ImpDefMI = &*I;
173 MBB->erase(ImpDefMI);
179 /// OptimizeFunction - Perhaps branch folding, tail merging and other
180 /// CFG optimizations on the given function.
181 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
182 const TargetInstrInfo *tii,
183 const TargetRegisterInfo *tri,
184 MachineModuleInfo *mmi) {
185 if (!tii) return false;
187 TriedMerging.clear();
194 // Use a RegScavenger to help update liveness when required.
195 MachineRegisterInfo &MRI = MF.getRegInfo();
196 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
197 RS = new RegScavenger();
199 MRI.invalidateLiveness();
201 // Fix CFG. The later algorithms expect it to be right.
202 bool MadeChange = false;
203 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
204 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
205 SmallVector<MachineOperand, 4> Cond;
206 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
207 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
208 MadeChange |= OptimizeImpDefsBlock(MBB);
211 bool MadeChangeThisIteration = true;
212 while (MadeChangeThisIteration) {
213 MadeChangeThisIteration = TailMergeBlocks(MF);
214 MadeChangeThisIteration |= OptimizeBranches(MF);
215 if (EnableHoistCommonCode)
216 MadeChangeThisIteration |= HoistCommonCode(MF);
217 MadeChange |= MadeChangeThisIteration;
220 // See if any jump tables have become dead as the code generator
222 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
228 // Walk the function to find jump tables that are live.
229 BitVector JTIsLive(JTI->getJumpTables().size());
230 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
232 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
234 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
235 MachineOperand &Op = I->getOperand(op);
236 if (!Op.isJTI()) continue;
238 // Remember that this JT is live.
239 JTIsLive.set(Op.getIndex());
243 // Finally, remove dead jump tables. This happens when the
244 // indirect jump was unreachable (and thus deleted).
245 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
246 if (!JTIsLive.test(i)) {
247 JTI->RemoveJumpTable(i);
255 //===----------------------------------------------------------------------===//
256 // Tail Merging of Blocks
257 //===----------------------------------------------------------------------===//
259 /// HashMachineInstr - Compute a hash value for MI and its operands.
260 static unsigned HashMachineInstr(const MachineInstr *MI) {
261 unsigned Hash = MI->getOpcode();
262 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
263 const MachineOperand &Op = MI->getOperand(i);
265 // Merge in bits from the operand if easy.
266 unsigned OperandHash = 0;
267 switch (Op.getType()) {
268 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
269 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
270 case MachineOperand::MO_MachineBasicBlock:
271 OperandHash = Op.getMBB()->getNumber();
273 case MachineOperand::MO_FrameIndex:
274 case MachineOperand::MO_ConstantPoolIndex:
275 case MachineOperand::MO_JumpTableIndex:
276 OperandHash = Op.getIndex();
278 case MachineOperand::MO_GlobalAddress:
279 case MachineOperand::MO_ExternalSymbol:
280 // Global address / external symbol are too hard, don't bother, but do
281 // pull in the offset.
282 OperandHash = Op.getOffset();
287 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
292 /// HashEndOfMBB - Hash the last instruction in the MBB.
293 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
294 MachineBasicBlock::const_iterator I = MBB->end();
295 if (I == MBB->begin())
296 return 0; // Empty MBB.
299 // Skip debug info so it will not affect codegen.
300 while (I->isDebugValue()) {
302 return 0; // MBB empty except for debug info.
306 return HashMachineInstr(I);
309 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
310 /// of instructions they actually have in common together at their end. Return
311 /// iterators for the first shared instruction in each block.
312 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
313 MachineBasicBlock *MBB2,
314 MachineBasicBlock::iterator &I1,
315 MachineBasicBlock::iterator &I2) {
319 unsigned TailLen = 0;
320 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
322 // Skip debugging pseudos; necessary to avoid changing the code.
323 while (I1->isDebugValue()) {
324 if (I1==MBB1->begin()) {
325 while (I2->isDebugValue()) {
326 if (I2==MBB2->begin())
327 // I1==DBG at begin; I2==DBG at begin
332 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
337 // I1==first (untested) non-DBG preceding known match
338 while (I2->isDebugValue()) {
339 if (I2==MBB2->begin()) {
341 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
346 // I1, I2==first (untested) non-DBGs preceding known match
347 if (!I1->isIdenticalTo(I2) ||
348 // FIXME: This check is dubious. It's used to get around a problem where
349 // people incorrectly expect inline asm directives to remain in the same
350 // relative order. This is untenable because normal compiler
351 // optimizations (like this one) may reorder and/or merge these
359 // Back past possible debugging pseudos at beginning of block. This matters
360 // when one block differs from the other only by whether debugging pseudos
361 // are present at the beginning. (This way, the various checks later for
362 // I1==MBB1->begin() work as expected.)
363 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
365 while (I2->isDebugValue()) {
366 if (I2 == MBB2->begin())
372 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
374 while (I1->isDebugValue()) {
375 if (I1 == MBB1->begin())
384 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
385 MachineBasicBlock *NewMBB) {
387 RS->enterBasicBlock(CurMBB);
388 if (!CurMBB->empty())
389 RS->forward(std::prev(CurMBB->end()));
390 BitVector RegsLiveAtExit(TRI->getNumRegs());
391 RS->getRegsUsed(RegsLiveAtExit, false);
392 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
393 if (RegsLiveAtExit[i])
394 NewMBB->addLiveIn(i);
398 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
399 /// after it, replacing it with an unconditional branch to NewDest.
400 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
401 MachineBasicBlock *NewDest) {
402 MachineBasicBlock *CurMBB = OldInst->getParent();
404 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
406 // For targets that use the register scavenger, we must maintain LiveIns.
407 MaintainLiveIns(CurMBB, NewDest);
412 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
413 /// MBB so that the part before the iterator falls into the part starting at the
414 /// iterator. This returns the new MBB.
415 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
416 MachineBasicBlock::iterator BBI1,
417 const BasicBlock *BB) {
418 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
421 MachineFunction &MF = *CurMBB.getParent();
423 // Create the fall-through block.
424 MachineFunction::iterator MBBI = &CurMBB;
425 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB);
426 CurMBB.getParent()->insert(++MBBI, NewMBB);
428 // Move all the successors of this block to the specified block.
429 NewMBB->transferSuccessors(&CurMBB);
431 // Add an edge from CurMBB to NewMBB for the fall-through.
432 CurMBB.addSuccessor(NewMBB);
434 // Splice the code over.
435 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
437 // For targets that use the register scavenger, we must maintain LiveIns.
438 MaintainLiveIns(&CurMBB, NewMBB);
443 /// EstimateRuntime - Make a rough estimate for how long it will take to run
444 /// the specified code.
445 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
446 MachineBasicBlock::iterator E) {
448 for (; I != E; ++I) {
449 if (I->isDebugValue())
453 else if (I->mayLoad() || I->mayStore())
461 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
462 // branches temporarily for tail merging). In the case where CurMBB ends
463 // with a conditional branch to the next block, optimize by reversing the
464 // test and conditionally branching to SuccMBB instead.
465 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
466 const TargetInstrInfo *TII) {
467 MachineFunction *MF = CurMBB->getParent();
468 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
469 MachineBasicBlock *TBB = 0, *FBB = 0;
470 SmallVector<MachineOperand, 4> Cond;
471 DebugLoc dl; // FIXME: this is nowhere
472 if (I != MF->end() &&
473 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
474 MachineBasicBlock *NextBB = I;
475 if (TBB == NextBB && !Cond.empty() && !FBB) {
476 if (!TII->ReverseBranchCondition(Cond)) {
477 TII->RemoveBranch(*CurMBB);
478 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl);
483 TII->InsertBranch(*CurMBB, SuccBB, NULL,
484 SmallVector<MachineOperand, 0>(), dl);
488 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
489 if (getHash() < o.getHash())
491 if (getHash() > o.getHash())
493 if (getBlock()->getNumber() < o.getBlock()->getNumber())
495 if (getBlock()->getNumber() > o.getBlock()->getNumber())
497 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
498 // an object with itself.
499 #ifndef _GLIBCXX_DEBUG
500 llvm_unreachable("Predecessor appears twice");
506 /// CountTerminators - Count the number of terminators in the given
507 /// block and set I to the position of the first non-terminator, if there
508 /// is one, or MBB->end() otherwise.
509 static unsigned CountTerminators(MachineBasicBlock *MBB,
510 MachineBasicBlock::iterator &I) {
512 unsigned NumTerms = 0;
514 if (I == MBB->begin()) {
519 if (!I->isTerminator()) break;
525 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
526 /// and decide if it would be profitable to merge those tails. Return the
527 /// length of the common tail and iterators to the first common instruction
529 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
530 MachineBasicBlock *MBB2,
531 unsigned minCommonTailLength,
532 unsigned &CommonTailLen,
533 MachineBasicBlock::iterator &I1,
534 MachineBasicBlock::iterator &I2,
535 MachineBasicBlock *SuccBB,
536 MachineBasicBlock *PredBB) {
537 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
538 if (CommonTailLen == 0)
540 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
541 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
544 // It's almost always profitable to merge any number of non-terminator
545 // instructions with the block that falls through into the common successor.
546 if (MBB1 == PredBB || MBB2 == PredBB) {
547 MachineBasicBlock::iterator I;
548 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
549 if (CommonTailLen > NumTerms)
553 // If one of the blocks can be completely merged and happens to be in
554 // a position where the other could fall through into it, merge any number
555 // of instructions, because it can be done without a branch.
556 // TODO: If the blocks are not adjacent, move one of them so that they are?
557 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
559 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
562 // If both blocks have an unconditional branch temporarily stripped out,
563 // count that as an additional common instruction for the following
565 unsigned EffectiveTailLen = CommonTailLen;
566 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
567 !MBB1->back().isBarrier() &&
568 !MBB2->back().isBarrier())
571 // Check if the common tail is long enough to be worthwhile.
572 if (EffectiveTailLen >= minCommonTailLength)
575 // If we are optimizing for code size, 2 instructions in common is enough if
576 // we don't have to split a block. At worst we will be introducing 1 new
577 // branch instruction, which is likely to be smaller than the 2
578 // instructions that would be deleted in the merge.
579 MachineFunction *MF = MBB1->getParent();
580 if (EffectiveTailLen >= 2 &&
581 MF->getFunction()->getAttributes().
582 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
583 (I1 == MBB1->begin() || I2 == MBB2->begin()))
589 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
590 /// hash CurHash (guaranteed to match the last element). Build the vector
591 /// SameTails of all those that have the (same) largest number of instructions
592 /// in common of any pair of these blocks. SameTails entries contain an
593 /// iterator into MergePotentials (from which the MachineBasicBlock can be
594 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
595 /// instruction where the matching code sequence begins.
596 /// Order of elements in SameTails is the reverse of the order in which
597 /// those blocks appear in MergePotentials (where they are not necessarily
599 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
600 unsigned minCommonTailLength,
601 MachineBasicBlock *SuccBB,
602 MachineBasicBlock *PredBB) {
603 unsigned maxCommonTailLength = 0U;
605 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
606 MPIterator HighestMPIter = std::prev(MergePotentials.end());
607 for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
608 B = MergePotentials.begin();
609 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
610 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
611 unsigned CommonTailLen;
612 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
614 CommonTailLen, TrialBBI1, TrialBBI2,
616 if (CommonTailLen > maxCommonTailLength) {
618 maxCommonTailLength = CommonTailLen;
619 HighestMPIter = CurMPIter;
620 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
622 if (HighestMPIter == CurMPIter &&
623 CommonTailLen == maxCommonTailLength)
624 SameTails.push_back(SameTailElt(I, TrialBBI2));
630 return maxCommonTailLength;
633 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
634 /// MergePotentials, restoring branches at ends of blocks as appropriate.
635 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
636 MachineBasicBlock *SuccBB,
637 MachineBasicBlock *PredBB) {
638 MPIterator CurMPIter, B;
639 for (CurMPIter = std::prev(MergePotentials.end()),
640 B = MergePotentials.begin();
641 CurMPIter->getHash() == CurHash; --CurMPIter) {
642 // Put the unconditional branch back, if we need one.
643 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
644 if (SuccBB && CurMBB != PredBB)
645 FixTail(CurMBB, SuccBB, TII);
649 if (CurMPIter->getHash() != CurHash)
651 MergePotentials.erase(CurMPIter, MergePotentials.end());
654 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
655 /// only of the common tail. Create a block that does by splitting one.
656 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
657 MachineBasicBlock *SuccBB,
658 unsigned maxCommonTailLength,
659 unsigned &commonTailIndex) {
661 unsigned TimeEstimate = ~0U;
662 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
663 // Use PredBB if possible; that doesn't require a new branch.
664 if (SameTails[i].getBlock() == PredBB) {
668 // Otherwise, make a (fairly bogus) choice based on estimate of
669 // how long it will take the various blocks to execute.
670 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
671 SameTails[i].getTailStartPos());
672 if (t <= TimeEstimate) {
678 MachineBasicBlock::iterator BBI =
679 SameTails[commonTailIndex].getTailStartPos();
680 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
682 // If the common tail includes any debug info we will take it pretty
683 // randomly from one of the inputs. Might be better to remove it?
684 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
685 << maxCommonTailLength);
687 // If the split block unconditionally falls-thru to SuccBB, it will be
688 // merged. In control flow terms it should then take SuccBB's name. e.g. If
689 // SuccBB is an inner loop, the common tail is still part of the inner loop.
690 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
691 SuccBB->getBasicBlock() : MBB->getBasicBlock();
692 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
694 DEBUG(dbgs() << "... failed!");
698 SameTails[commonTailIndex].setBlock(newMBB);
699 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
701 // If we split PredBB, newMBB is the new predecessor.
708 // See if any of the blocks in MergePotentials (which all have a common single
709 // successor, or all have no successor) can be tail-merged. If there is a
710 // successor, any blocks in MergePotentials that are not tail-merged and
711 // are not immediately before Succ must have an unconditional branch to
712 // Succ added (but the predecessor/successor lists need no adjustment).
713 // The lone predecessor of Succ that falls through into Succ,
714 // if any, is given in PredBB.
716 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
717 MachineBasicBlock *PredBB) {
718 bool MadeChange = false;
720 // Except for the special cases below, tail-merge if there are at least
721 // this many instructions in common.
722 unsigned minCommonTailLength = TailMergeSize;
724 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
725 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
726 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
727 << (i == e-1 ? "" : ", ");
730 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
732 dbgs() << " which has fall-through from BB#"
733 << PredBB->getNumber() << "\n";
735 dbgs() << "Looking for common tails of at least "
736 << minCommonTailLength << " instruction"
737 << (minCommonTailLength == 1 ? "" : "s") << '\n';
740 // Sort by hash value so that blocks with identical end sequences sort
742 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
744 // Walk through equivalence sets looking for actual exact matches.
745 while (MergePotentials.size() > 1) {
746 unsigned CurHash = MergePotentials.back().getHash();
748 // Build SameTails, identifying the set of blocks with this hash code
749 // and with the maximum number of instructions in common.
750 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
754 // If we didn't find any pair that has at least minCommonTailLength
755 // instructions in common, remove all blocks with this hash code and retry.
756 if (SameTails.empty()) {
757 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
761 // If one of the blocks is the entire common tail (and not the entry
762 // block, which we can't jump to), we can treat all blocks with this same
763 // tail at once. Use PredBB if that is one of the possibilities, as that
764 // will not introduce any extra branches.
765 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
766 getParent()->begin();
767 unsigned commonTailIndex = SameTails.size();
768 // If there are two blocks, check to see if one can be made to fall through
770 if (SameTails.size() == 2 &&
771 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
772 SameTails[1].tailIsWholeBlock())
774 else if (SameTails.size() == 2 &&
775 SameTails[1].getBlock()->isLayoutSuccessor(
776 SameTails[0].getBlock()) &&
777 SameTails[0].tailIsWholeBlock())
780 // Otherwise just pick one, favoring the fall-through predecessor if
782 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
783 MachineBasicBlock *MBB = SameTails[i].getBlock();
784 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
790 if (SameTails[i].tailIsWholeBlock())
795 if (commonTailIndex == SameTails.size() ||
796 (SameTails[commonTailIndex].getBlock() == PredBB &&
797 !SameTails[commonTailIndex].tailIsWholeBlock())) {
798 // None of the blocks consist entirely of the common tail.
799 // Split a block so that one does.
800 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
801 maxCommonTailLength, commonTailIndex)) {
802 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
807 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
808 // MBB is common tail. Adjust all other BB's to jump to this one.
809 // Traversal must be forwards so erases work.
810 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
812 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
813 if (commonTailIndex == i)
815 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
816 << (i == e-1 ? "" : ", "));
817 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
818 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
819 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
820 MergePotentials.erase(SameTails[i].getMPIter());
822 DEBUG(dbgs() << "\n");
823 // We leave commonTailIndex in the worklist in case there are other blocks
824 // that match it with a smaller number of instructions.
830 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
831 bool MadeChange = false;
832 if (!EnableTailMerge) return MadeChange;
834 // First find blocks with no successors.
835 MergePotentials.clear();
836 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
837 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
838 if (TriedMerging.count(I))
841 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
844 // If this is a large problem, avoid visiting the same basic blocks
846 if (MergePotentials.size() == TailMergeThreshold)
847 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
848 TriedMerging.insert(MergePotentials[i].getBlock());
850 // See if we can do any tail merging on those.
851 if (MergePotentials.size() >= 2)
852 MadeChange |= TryTailMergeBlocks(NULL, NULL);
854 // Look at blocks (IBB) with multiple predecessors (PBB).
855 // We change each predecessor to a canonical form, by
856 // (1) temporarily removing any unconditional branch from the predecessor
858 // (2) alter conditional branches so they branch to the other block
859 // not IBB; this may require adding back an unconditional branch to IBB
860 // later, where there wasn't one coming in. E.g.
862 // fallthrough to QBB
865 // with a conceptual B to IBB after that, which never actually exists.
866 // With those changes, we see whether the predecessors' tails match,
867 // and merge them if so. We change things out of canonical form and
868 // back to the way they were later in the process. (OptimizeBranches
869 // would undo some of this, but we can't use it, because we'd get into
870 // a compile-time infinite loop repeatedly doing and undoing the same
873 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
875 if (I->pred_size() < 2) continue;
876 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
877 MachineBasicBlock *IBB = I;
878 MachineBasicBlock *PredBB = std::prev(I);
879 MergePotentials.clear();
880 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
882 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
883 MachineBasicBlock *PBB = *P;
884 if (TriedMerging.count(PBB))
887 // Skip blocks that loop to themselves, can't tail merge these.
891 // Visit each predecessor only once.
892 if (!UniquePreds.insert(PBB))
895 // Skip blocks which may jump to a landing pad. Can't tail merge these.
896 if (PBB->getLandingPadSuccessor())
899 MachineBasicBlock *TBB = 0, *FBB = 0;
900 SmallVector<MachineOperand, 4> Cond;
901 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
902 // Failing case: IBB is the target of a cbr, and we cannot reverse the
904 SmallVector<MachineOperand, 4> NewCond(Cond);
905 if (!Cond.empty() && TBB == IBB) {
906 if (TII->ReverseBranchCondition(NewCond))
908 // This is the QBB case described above
910 FBB = std::next(MachineFunction::iterator(PBB));
913 // Failing case: the only way IBB can be reached from PBB is via
914 // exception handling. Happens for landing pads. Would be nice to have
915 // a bit in the edge so we didn't have to do all this.
916 if (IBB->isLandingPad()) {
917 MachineFunction::iterator IP = PBB; IP++;
918 MachineBasicBlock *PredNextBB = NULL;
922 if (IBB != PredNextBB) // fallthrough
925 if (TBB != IBB && FBB != IBB) // cbr then ubr
927 } else if (Cond.empty()) {
928 if (TBB != IBB) // ubr
931 if (TBB != IBB && IBB != PredNextBB) // cbr
936 // Remove the unconditional branch at the end, if any.
937 if (TBB && (Cond.empty() || FBB)) {
938 DebugLoc dl; // FIXME: this is nowhere
939 TII->RemoveBranch(*PBB);
941 // reinsert conditional branch only, for now
942 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl);
945 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
949 // If this is a large problem, avoid visiting the same basic blocks multiple
951 if (MergePotentials.size() == TailMergeThreshold)
952 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
953 TriedMerging.insert(MergePotentials[i].getBlock());
955 if (MergePotentials.size() >= 2)
956 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
958 // Reinsert an unconditional branch if needed. The 1 below can occur as a
959 // result of removing blocks in TryTailMergeBlocks.
960 PredBB = std::prev(I); // this may have been changed in TryTailMergeBlocks
961 if (MergePotentials.size() == 1 &&
962 MergePotentials.begin()->getBlock() != PredBB)
963 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
969 //===----------------------------------------------------------------------===//
970 // Branch Optimization
971 //===----------------------------------------------------------------------===//
973 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
974 bool MadeChange = false;
976 // Make sure blocks are numbered in order
979 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
981 MachineBasicBlock *MBB = I++;
982 MadeChange |= OptimizeBlock(MBB);
984 // If it is dead, remove it.
985 if (MBB->pred_empty()) {
986 RemoveDeadBlock(MBB);
994 // Blocks should be considered empty if they contain only debug info;
995 // else the debug info would affect codegen.
996 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
999 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
1000 MBBI!=MBBE; ++MBBI) {
1001 if (!MBBI->isDebugValue())
1007 // Blocks with only debug info and branches should be considered the same
1008 // as blocks with only branches.
1009 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1010 MachineBasicBlock::iterator MBBI, MBBE;
1011 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
1012 if (!MBBI->isDebugValue())
1015 return (MBBI->isBranch());
1018 /// IsBetterFallthrough - Return true if it would be clearly better to
1019 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1020 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1021 /// result in infinite loops.
1022 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1023 MachineBasicBlock *MBB2) {
1024 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1025 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1026 // optimize branches that branch to either a return block or an assert block
1027 // into a fallthrough to the return.
1028 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
1030 // If there is a clear successor ordering we make sure that one block
1031 // will fall through to the next
1032 if (MBB1->isSuccessor(MBB2)) return true;
1033 if (MBB2->isSuccessor(MBB1)) return false;
1035 // Neither block consists entirely of debug info (per IsEmptyBlock check),
1036 // so we needn't test for falling off the beginning here.
1037 MachineBasicBlock::iterator MBB1I = --MBB1->end();
1038 while (MBB1I->isDebugValue())
1040 MachineBasicBlock::iterator MBB2I = --MBB2->end();
1041 while (MBB2I->isDebugValue())
1043 return MBB2I->isCall() && !MBB1I->isCall();
1046 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1047 /// instructions on the block. Always use the DebugLoc of the first
1048 /// branching instruction found unless its absent, in which case use the
1049 /// DebugLoc of the second if present.
1050 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1051 MachineBasicBlock::iterator I = MBB.end();
1052 if (I == MBB.begin())
1055 while (I->isDebugValue() && I != MBB.begin())
1058 return I->getDebugLoc();
1062 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1063 /// block. This is never called on the entry block.
1064 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1065 bool MadeChange = false;
1066 MachineFunction &MF = *MBB->getParent();
1069 MachineFunction::iterator FallThrough = MBB;
1072 // If this block is empty, make everyone use its fall-through, not the block
1073 // explicitly. Landing pads should not do this since the landing-pad table
1074 // points to this block. Blocks with their addresses taken shouldn't be
1076 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1077 // Dead block? Leave for cleanup later.
1078 if (MBB->pred_empty()) return MadeChange;
1080 if (FallThrough == MF.end()) {
1081 // TODO: Simplify preds to not branch here if possible!
1083 // Rewrite all predecessors of the old block to go to the fallthrough
1085 while (!MBB->pred_empty()) {
1086 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1087 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1089 // If MBB was the target of a jump table, update jump tables to go to the
1090 // fallthrough instead.
1091 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1092 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1098 // Check to see if we can simplify the terminator of the block before this
1100 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1102 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1103 SmallVector<MachineOperand, 4> PriorCond;
1104 bool PriorUnAnalyzable =
1105 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1106 if (!PriorUnAnalyzable) {
1107 // If the CFG for the prior block has extra edges, remove them.
1108 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1109 !PriorCond.empty());
1111 // If the previous branch is conditional and both conditions go to the same
1112 // destination, remove the branch, replacing it with an unconditional one or
1114 if (PriorTBB && PriorTBB == PriorFBB) {
1115 DebugLoc dl = getBranchDebugLoc(PrevBB);
1116 TII->RemoveBranch(PrevBB);
1118 if (PriorTBB != MBB)
1119 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1122 goto ReoptimizeBlock;
1125 // If the previous block unconditionally falls through to this block and
1126 // this block has no other predecessors, move the contents of this block
1127 // into the prior block. This doesn't usually happen when SimplifyCFG
1128 // has been used, but it can happen if tail merging splits a fall-through
1129 // predecessor of a block.
1130 // This has to check PrevBB->succ_size() because EH edges are ignored by
1132 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1133 PrevBB.succ_size() == 1 &&
1134 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1135 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1136 << "From MBB: " << *MBB);
1137 // Remove redundant DBG_VALUEs first.
1138 if (PrevBB.begin() != PrevBB.end()) {
1139 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1141 MachineBasicBlock::iterator MBBIter = MBB->begin();
1142 // Check if DBG_VALUE at the end of PrevBB is identical to the
1143 // DBG_VALUE at the beginning of MBB.
1144 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1145 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1146 if (!MBBIter->isIdenticalTo(PrevBBIter))
1148 MachineInstr *DuplicateDbg = MBBIter;
1149 ++MBBIter; -- PrevBBIter;
1150 DuplicateDbg->eraseFromParent();
1153 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1154 PrevBB.removeSuccessor(PrevBB.succ_begin());
1155 assert(PrevBB.succ_empty());
1156 PrevBB.transferSuccessors(MBB);
1161 // If the previous branch *only* branches to *this* block (conditional or
1162 // not) remove the branch.
1163 if (PriorTBB == MBB && PriorFBB == 0) {
1164 TII->RemoveBranch(PrevBB);
1167 goto ReoptimizeBlock;
1170 // If the prior block branches somewhere else on the condition and here if
1171 // the condition is false, remove the uncond second branch.
1172 if (PriorFBB == MBB) {
1173 DebugLoc dl = getBranchDebugLoc(PrevBB);
1174 TII->RemoveBranch(PrevBB);
1175 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1178 goto ReoptimizeBlock;
1181 // If the prior block branches here on true and somewhere else on false, and
1182 // if the branch condition is reversible, reverse the branch to create a
1184 if (PriorTBB == MBB) {
1185 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1186 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1187 DebugLoc dl = getBranchDebugLoc(PrevBB);
1188 TII->RemoveBranch(PrevBB);
1189 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl);
1192 goto ReoptimizeBlock;
1196 // If this block has no successors (e.g. it is a return block or ends with
1197 // a call to a no-return function like abort or __cxa_throw) and if the pred
1198 // falls through into this block, and if it would otherwise fall through
1199 // into the block after this, move this block to the end of the function.
1201 // We consider it more likely that execution will stay in the function (e.g.
1202 // due to loops) than it is to exit it. This asserts in loops etc, moving
1203 // the assert condition out of the loop body.
1204 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1205 MachineFunction::iterator(PriorTBB) == FallThrough &&
1206 !MBB->canFallThrough()) {
1207 bool DoTransform = true;
1209 // We have to be careful that the succs of PredBB aren't both no-successor
1210 // blocks. If neither have successors and if PredBB is the second from
1211 // last block in the function, we'd just keep swapping the two blocks for
1212 // last. Only do the swap if one is clearly better to fall through than
1214 if (FallThrough == --MF.end() &&
1215 !IsBetterFallthrough(PriorTBB, MBB))
1216 DoTransform = false;
1219 // Reverse the branch so we will fall through on the previous true cond.
1220 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1221 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1222 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1223 << "To make fallthrough to: " << *PriorTBB << "\n");
1225 DebugLoc dl = getBranchDebugLoc(PrevBB);
1226 TII->RemoveBranch(PrevBB);
1227 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl);
1229 // Move this block to the end of the function.
1230 MBB->moveAfter(--MF.end());
1239 // Analyze the branch in the current block.
1240 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1241 SmallVector<MachineOperand, 4> CurCond;
1242 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1243 if (!CurUnAnalyzable) {
1244 // If the CFG for the prior block has extra edges, remove them.
1245 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1247 // If this is a two-way branch, and the FBB branches to this block, reverse
1248 // the condition so the single-basic-block loop is faster. Instead of:
1249 // Loop: xxx; jcc Out; jmp Loop
1251 // Loop: xxx; jncc Loop; jmp Out
1252 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1253 SmallVector<MachineOperand, 4> NewCond(CurCond);
1254 if (!TII->ReverseBranchCondition(NewCond)) {
1255 DebugLoc dl = getBranchDebugLoc(*MBB);
1256 TII->RemoveBranch(*MBB);
1257 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1260 goto ReoptimizeBlock;
1264 // If this branch is the only thing in its block, see if we can forward
1265 // other blocks across it.
1266 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1267 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1268 !MBB->hasAddressTaken()) {
1269 DebugLoc dl = getBranchDebugLoc(*MBB);
1270 // This block may contain just an unconditional branch. Because there can
1271 // be 'non-branch terminators' in the block, try removing the branch and
1272 // then seeing if the block is empty.
1273 TII->RemoveBranch(*MBB);
1274 // If the only things remaining in the block are debug info, remove these
1275 // as well, so this will behave the same as an empty block in non-debug
1277 if (!MBB->empty()) {
1278 bool NonDebugInfoFound = false;
1279 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1281 if (!I->isDebugValue()) {
1282 NonDebugInfoFound = true;
1286 if (!NonDebugInfoFound)
1287 // Make the block empty, losing the debug info (we could probably
1288 // improve this in some cases.)
1289 MBB->erase(MBB->begin(), MBB->end());
1291 // If this block is just an unconditional branch to CurTBB, we can
1292 // usually completely eliminate the block. The only case we cannot
1293 // completely eliminate the block is when the block before this one
1294 // falls through into MBB and we can't understand the prior block's branch
1297 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1298 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1299 !PrevBB.isSuccessor(MBB)) {
1300 // If the prior block falls through into us, turn it into an
1301 // explicit branch to us to make updates simpler.
1302 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1303 PriorTBB != MBB && PriorFBB != MBB) {
1304 if (PriorTBB == 0) {
1305 assert(PriorCond.empty() && PriorFBB == 0 &&
1306 "Bad branch analysis");
1309 assert(PriorFBB == 0 && "Machine CFG out of date!");
1312 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1313 TII->RemoveBranch(PrevBB);
1314 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1317 // Iterate through all the predecessors, revectoring each in-turn.
1319 bool DidChange = false;
1320 bool HasBranchToSelf = false;
1321 while(PI != MBB->pred_size()) {
1322 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1324 // If this block has an uncond branch to itself, leave it.
1326 HasBranchToSelf = true;
1329 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1330 // If this change resulted in PMBB ending in a conditional
1331 // branch where both conditions go to the same destination,
1332 // change this to an unconditional branch (and fix the CFG).
1333 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1334 SmallVector<MachineOperand, 4> NewCurCond;
1335 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1336 NewCurFBB, NewCurCond, true);
1337 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1338 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1339 TII->RemoveBranch(*PMBB);
1341 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, pdl);
1344 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1349 // Change any jumptables to go to the new MBB.
1350 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1351 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1355 if (!HasBranchToSelf) return MadeChange;
1360 // Add the branch back if the block is more than just an uncond branch.
1361 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl);
1365 // If the prior block doesn't fall through into this block, and if this
1366 // block doesn't fall through into some other block, see if we can find a
1367 // place to move this block where a fall-through will happen.
1368 if (!PrevBB.canFallThrough()) {
1370 // Now we know that there was no fall-through into this block, check to
1371 // see if it has a fall-through into its successor.
1372 bool CurFallsThru = MBB->canFallThrough();
1374 if (!MBB->isLandingPad()) {
1375 // Check all the predecessors of this block. If one of them has no fall
1376 // throughs, move this block right after it.
1377 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1378 E = MBB->pred_end(); PI != E; ++PI) {
1379 // Analyze the branch at the end of the pred.
1380 MachineBasicBlock *PredBB = *PI;
1381 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1382 MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
1383 SmallVector<MachineOperand, 4> PredCond;
1384 if (PredBB != MBB && !PredBB->canFallThrough() &&
1385 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1386 && (!CurFallsThru || !CurTBB || !CurFBB)
1387 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1388 // If the current block doesn't fall through, just move it.
1389 // If the current block can fall through and does not end with a
1390 // conditional branch, we need to append an unconditional jump to
1391 // the (current) next block. To avoid a possible compile-time
1392 // infinite loop, move blocks only backward in this case.
1393 // Also, if there are already 2 branches here, we cannot add a third;
1394 // this means we have the case
1399 MachineBasicBlock *NextBB =
1400 std::next(MachineFunction::iterator(MBB));
1402 TII->InsertBranch(*MBB, NextBB, 0, CurCond, DebugLoc());
1404 MBB->moveAfter(PredBB);
1406 goto ReoptimizeBlock;
1411 if (!CurFallsThru) {
1412 // Check all successors to see if we can move this block before it.
1413 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1414 E = MBB->succ_end(); SI != E; ++SI) {
1415 // Analyze the branch at the end of the block before the succ.
1416 MachineBasicBlock *SuccBB = *SI;
1417 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1419 // If this block doesn't already fall-through to that successor, and if
1420 // the succ doesn't already have a block that can fall through into it,
1421 // and if the successor isn't an EH destination, we can arrange for the
1422 // fallthrough to happen.
1423 if (SuccBB != MBB && &*SuccPrev != MBB &&
1424 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1425 !SuccBB->isLandingPad()) {
1426 MBB->moveBefore(SuccBB);
1428 goto ReoptimizeBlock;
1432 // Okay, there is no really great place to put this block. If, however,
1433 // the block before this one would be a fall-through if this block were
1434 // removed, move this block to the end of the function.
1435 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
1436 SmallVector<MachineOperand, 4> PrevCond;
1437 if (FallThrough != MF.end() &&
1438 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1439 PrevBB.isSuccessor(FallThrough)) {
1440 MBB->moveAfter(--MF.end());
1450 //===----------------------------------------------------------------------===//
1451 // Hoist Common Code
1452 //===----------------------------------------------------------------------===//
1454 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1455 /// blocks to their common predecessor.
1456 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1457 bool MadeChange = false;
1458 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1459 MachineBasicBlock *MBB = I++;
1460 MadeChange |= HoistCommonCodeInSuccs(MBB);
1466 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1467 /// its 'true' successor.
1468 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1469 MachineBasicBlock *TrueBB) {
1470 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1471 E = BB->succ_end(); SI != E; ++SI) {
1472 MachineBasicBlock *SuccBB = *SI;
1473 if (SuccBB != TrueBB)
1479 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1480 /// in successors to. The location is usually just before the terminator,
1481 /// however if the terminator is a conditional branch and its previous
1482 /// instruction is the flag setting instruction, the previous instruction is
1483 /// the preferred location. This function also gathers uses and defs of the
1484 /// instructions from the insertion point to the end of the block. The data is
1485 /// used by HoistCommonCodeInSuccs to ensure safety.
1487 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1488 const TargetInstrInfo *TII,
1489 const TargetRegisterInfo *TRI,
1490 SmallSet<unsigned,4> &Uses,
1491 SmallSet<unsigned,4> &Defs) {
1492 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1493 if (!TII->isUnpredicatedTerminator(Loc))
1496 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1497 const MachineOperand &MO = Loc->getOperand(i);
1500 unsigned Reg = MO.getReg();
1504 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1506 } else if (!MO.isDead())
1507 // Don't try to hoist code in the rare case the terminator defines a
1508 // register that is later used.
1514 if (Loc == MBB->begin())
1517 // The terminator is probably a conditional branch, try not to separate the
1518 // branch from condition setting instruction.
1519 MachineBasicBlock::iterator PI = Loc;
1521 while (PI != MBB->begin() && PI->isDebugValue())
1525 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1526 const MachineOperand &MO = PI->getOperand(i);
1527 // If PI has a regmask operand, it is probably a call. Separate away.
1530 if (!MO.isReg() || MO.isUse())
1532 unsigned Reg = MO.getReg();
1535 if (Uses.count(Reg))
1539 // The condition setting instruction is not just before the conditional
1543 // Be conservative, don't insert instruction above something that may have
1544 // side-effects. And since it's potentially bad to separate flag setting
1545 // instruction from the conditional branch, just abort the optimization
1547 // Also avoid moving code above predicated instruction since it's hard to
1548 // reason about register liveness with predicated instruction.
1549 bool DontMoveAcrossStore = true;
1550 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) ||
1551 TII->isPredicated(PI))
1555 // Find out what registers are live. Note this routine is ignoring other live
1556 // registers which are only used by instructions in successor blocks.
1557 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1558 const MachineOperand &MO = PI->getOperand(i);
1561 unsigned Reg = MO.getReg();
1565 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1568 if (Uses.erase(Reg)) {
1569 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1570 Uses.erase(*SubRegs); // Use sub-registers to be conservative
1572 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1580 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1581 /// sequence at the start of the function, move the instructions before MBB
1582 /// terminator if it's legal.
1583 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1584 MachineBasicBlock *TBB = 0, *FBB = 0;
1585 SmallVector<MachineOperand, 4> Cond;
1586 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1589 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1591 // Malformed bcc? True and false blocks are the same?
1594 // Restrict the optimization to cases where MBB is the only predecessor,
1595 // it is an obvious win.
1596 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1599 // Find a suitable position to hoist the common instructions to. Also figure
1600 // out which registers are used or defined by instructions from the insertion
1601 // point to the end of the block.
1602 SmallSet<unsigned, 4> Uses, Defs;
1603 MachineBasicBlock::iterator Loc =
1604 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1605 if (Loc == MBB->end())
1608 bool HasDups = false;
1609 SmallVector<unsigned, 4> LocalDefs;
1610 SmallSet<unsigned, 4> LocalDefsSet;
1611 MachineBasicBlock::iterator TIB = TBB->begin();
1612 MachineBasicBlock::iterator FIB = FBB->begin();
1613 MachineBasicBlock::iterator TIE = TBB->end();
1614 MachineBasicBlock::iterator FIE = FBB->end();
1615 while (TIB != TIE && FIB != FIE) {
1616 // Skip dbg_value instructions. These do not count.
1617 if (TIB->isDebugValue()) {
1618 while (TIB != TIE && TIB->isDebugValue())
1623 if (FIB->isDebugValue()) {
1624 while (FIB != FIE && FIB->isDebugValue())
1629 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1632 if (TII->isPredicated(TIB))
1633 // Hard to reason about register liveness with predicated instruction.
1637 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1638 MachineOperand &MO = TIB->getOperand(i);
1639 // Don't attempt to hoist instructions with register masks.
1640 if (MO.isRegMask()) {
1646 unsigned Reg = MO.getReg();
1650 if (Uses.count(Reg)) {
1651 // Avoid clobbering a register that's used by the instruction at
1652 // the point of insertion.
1657 if (Defs.count(Reg) && !MO.isDead()) {
1658 // Don't hoist the instruction if the def would be clobber by the
1659 // instruction at the point insertion. FIXME: This is overly
1660 // conservative. It should be possible to hoist the instructions
1661 // in BB2 in the following example:
1663 // r1, eflag = op1 r2, r3
1672 } else if (!LocalDefsSet.count(Reg)) {
1673 if (Defs.count(Reg)) {
1674 // Use is defined by the instruction at the point of insertion.
1679 if (MO.isKill() && Uses.count(Reg))
1680 // Kills a register that's read by the instruction at the point of
1681 // insertion. Remove the kill marker.
1682 MO.setIsKill(false);
1688 bool DontMoveAcrossStore = true;
1689 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore))
1692 // Remove kills from LocalDefsSet, these registers had short live ranges.
1693 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1694 MachineOperand &MO = TIB->getOperand(i);
1695 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1697 unsigned Reg = MO.getReg();
1698 if (!Reg || !LocalDefsSet.count(Reg))
1700 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1701 LocalDefsSet.erase(*AI);
1704 // Track local defs so we can update liveins.
1705 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1706 MachineOperand &MO = TIB->getOperand(i);
1707 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1709 unsigned Reg = MO.getReg();
1712 LocalDefs.push_back(Reg);
1713 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1714 LocalDefsSet.insert(*AI);
1725 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1726 FBB->erase(FBB->begin(), FIB);
1729 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1730 unsigned Def = LocalDefs[i];
1731 if (LocalDefsSet.count(Def)) {
1732 TBB->addLiveIn(Def);
1733 FBB->addLiveIn(Def);