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/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/RegisterScavenging.h"
28 #include "llvm/Target/TargetInstrInfo.h"
29 #include "llvm/Target/TargetMachine.h"
30 #include "llvm/Target/TargetRegisterInfo.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/ADT/SmallSet.h"
36 #include "llvm/ADT/SetVector.h"
37 #include "llvm/ADT/Statistic.h"
38 #include "llvm/ADT/STLExtras.h"
42 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
43 STATISTIC(NumBranchOpts, "Number of branches optimized");
44 STATISTIC(NumTailMerge , "Number of block tails merged");
45 STATISTIC(NumHoist , "Number of times common instructions are hoisted");
47 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
48 cl::init(cl::BOU_UNSET), cl::Hidden);
50 // Throttle for huge numbers of predecessors (compile speed problems)
51 static cl::opt<unsigned>
52 TailMergeThreshold("tail-merge-threshold",
53 cl::desc("Max number of predecessors to consider tail merging"),
54 cl::init(150), cl::Hidden);
56 // Heuristic for tail merging (and, inversely, tail duplication).
57 // TODO: This should be replaced with a target query.
58 static cl::opt<unsigned>
59 TailMergeSize("tail-merge-size",
60 cl::desc("Min number of instructions to consider tail merging"),
61 cl::init(3), cl::Hidden);
64 /// BranchFolderPass - Wrap branch folder in a machine function pass.
65 class BranchFolderPass : public MachineFunctionPass {
68 explicit BranchFolderPass(): MachineFunctionPass(ID) {}
70 virtual bool runOnMachineFunction(MachineFunction &MF);
72 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
73 AU.addRequired<TargetPassConfig>();
74 MachineFunctionPass::getAnalysisUsage(AU);
79 char BranchFolderPass::ID = 0;
80 char &llvm::BranchFolderPassID = BranchFolderPass::ID;
82 INITIALIZE_PASS(BranchFolderPass, "branch-folder",
83 "Control Flow Optimizer", false, false)
85 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
86 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
87 BranchFolder Folder(PassConfig->getEnableTailMerge(), /*CommonHoist=*/true);
88 return Folder.OptimizeFunction(MF,
89 MF.getTarget().getInstrInfo(),
90 MF.getTarget().getRegisterInfo(),
91 getAnalysisIfAvailable<MachineModuleInfo>());
95 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist) {
96 switch (FlagEnableTailMerge) {
97 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
98 case cl::BOU_TRUE: EnableTailMerge = true; break;
99 case cl::BOU_FALSE: EnableTailMerge = false; break;
102 EnableHoistCommonCode = CommonHoist;
105 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
106 /// function, updating the CFG.
107 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
108 assert(MBB->pred_empty() && "MBB must be dead!");
109 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
111 MachineFunction *MF = MBB->getParent();
112 // drop all successors.
113 while (!MBB->succ_empty())
114 MBB->removeSuccessor(MBB->succ_end()-1);
116 // Avoid matching if this pointer gets reused.
117 TriedMerging.erase(MBB);
123 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
124 /// followed by terminators, and if the implicitly defined registers are not
125 /// used by the terminators, remove those implicit_def's. e.g.
127 /// r0 = implicit_def
128 /// r1 = implicit_def
130 /// This block can be optimized away later if the implicit instructions are
132 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
133 SmallSet<unsigned, 4> ImpDefRegs;
134 MachineBasicBlock::iterator I = MBB->begin();
135 while (I != MBB->end()) {
136 if (!I->isImplicitDef())
138 unsigned Reg = I->getOperand(0).getReg();
139 ImpDefRegs.insert(Reg);
140 for (const uint16_t *SubRegs = TRI->getSubRegisters(Reg);
141 unsigned SubReg = *SubRegs; ++SubRegs)
142 ImpDefRegs.insert(SubReg);
145 if (ImpDefRegs.empty())
148 MachineBasicBlock::iterator FirstTerm = I;
149 while (I != MBB->end()) {
150 if (!TII->isUnpredicatedTerminator(I))
152 // See if it uses any of the implicitly defined registers.
153 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
154 MachineOperand &MO = I->getOperand(i);
155 if (!MO.isReg() || !MO.isUse())
157 unsigned Reg = MO.getReg();
158 if (ImpDefRegs.count(Reg))
165 while (I != FirstTerm) {
166 MachineInstr *ImpDefMI = &*I;
168 MBB->erase(ImpDefMI);
174 /// OptimizeFunction - Perhaps branch folding, tail merging and other
175 /// CFG optimizations on the given function.
176 bool BranchFolder::OptimizeFunction(MachineFunction &MF,
177 const TargetInstrInfo *tii,
178 const TargetRegisterInfo *tri,
179 MachineModuleInfo *mmi) {
180 if (!tii) return false;
182 TriedMerging.clear();
189 // Use a RegScavenger to help update liveness when required.
190 MachineRegisterInfo &MRI = MF.getRegInfo();
191 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
192 RS = new RegScavenger();
194 MRI.invalidateLiveness();
196 // Fix CFG. The later algorithms expect it to be right.
197 bool MadeChange = false;
198 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
199 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
200 SmallVector<MachineOperand, 4> Cond;
201 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
202 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
203 MadeChange |= OptimizeImpDefsBlock(MBB);
206 bool MadeChangeThisIteration = true;
207 while (MadeChangeThisIteration) {
208 MadeChangeThisIteration = TailMergeBlocks(MF);
209 MadeChangeThisIteration |= OptimizeBranches(MF);
210 if (EnableHoistCommonCode)
211 MadeChangeThisIteration |= HoistCommonCode(MF);
212 MadeChange |= MadeChangeThisIteration;
215 // See if any jump tables have become dead as the code generator
217 MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
223 // Walk the function to find jump tables that are live.
224 BitVector JTIsLive(JTI->getJumpTables().size());
225 for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
227 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
229 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
230 MachineOperand &Op = I->getOperand(op);
231 if (!Op.isJTI()) continue;
233 // Remember that this JT is live.
234 JTIsLive.set(Op.getIndex());
238 // Finally, remove dead jump tables. This happens when the
239 // indirect jump was unreachable (and thus deleted).
240 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
241 if (!JTIsLive.test(i)) {
242 JTI->RemoveJumpTable(i);
250 //===----------------------------------------------------------------------===//
251 // Tail Merging of Blocks
252 //===----------------------------------------------------------------------===//
254 /// HashMachineInstr - Compute a hash value for MI and its operands.
255 static unsigned HashMachineInstr(const MachineInstr *MI) {
256 unsigned Hash = MI->getOpcode();
257 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
258 const MachineOperand &Op = MI->getOperand(i);
260 // Merge in bits from the operand if easy.
261 unsigned OperandHash = 0;
262 switch (Op.getType()) {
263 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
264 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
265 case MachineOperand::MO_MachineBasicBlock:
266 OperandHash = Op.getMBB()->getNumber();
268 case MachineOperand::MO_FrameIndex:
269 case MachineOperand::MO_ConstantPoolIndex:
270 case MachineOperand::MO_JumpTableIndex:
271 OperandHash = Op.getIndex();
273 case MachineOperand::MO_GlobalAddress:
274 case MachineOperand::MO_ExternalSymbol:
275 // Global address / external symbol are too hard, don't bother, but do
276 // pull in the offset.
277 OperandHash = Op.getOffset();
282 Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
287 /// HashEndOfMBB - Hash the last instruction in the MBB.
288 static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
289 MachineBasicBlock::const_iterator I = MBB->end();
290 if (I == MBB->begin())
291 return 0; // Empty MBB.
294 // Skip debug info so it will not affect codegen.
295 while (I->isDebugValue()) {
297 return 0; // MBB empty except for debug info.
301 return HashMachineInstr(I);
304 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
305 /// of instructions they actually have in common together at their end. Return
306 /// iterators for the first shared instruction in each block.
307 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
308 MachineBasicBlock *MBB2,
309 MachineBasicBlock::iterator &I1,
310 MachineBasicBlock::iterator &I2) {
314 unsigned TailLen = 0;
315 while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
317 // Skip debugging pseudos; necessary to avoid changing the code.
318 while (I1->isDebugValue()) {
319 if (I1==MBB1->begin()) {
320 while (I2->isDebugValue()) {
321 if (I2==MBB2->begin())
322 // I1==DBG at begin; I2==DBG at begin
327 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
332 // I1==first (untested) non-DBG preceding known match
333 while (I2->isDebugValue()) {
334 if (I2==MBB2->begin()) {
336 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
341 // I1, I2==first (untested) non-DBGs preceding known match
342 if (!I1->isIdenticalTo(I2) ||
343 // FIXME: This check is dubious. It's used to get around a problem where
344 // people incorrectly expect inline asm directives to remain in the same
345 // relative order. This is untenable because normal compiler
346 // optimizations (like this one) may reorder and/or merge these
354 // Back past possible debugging pseudos at beginning of block. This matters
355 // when one block differs from the other only by whether debugging pseudos
356 // are present at the beginning. (This way, the various checks later for
357 // I1==MBB1->begin() work as expected.)
358 if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
360 while (I2->isDebugValue()) {
361 if (I2 == MBB2->begin()) {
368 if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
370 while (I1->isDebugValue()) {
371 if (I1 == MBB1->begin())
380 void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
381 MachineBasicBlock *NewMBB) {
383 RS->enterBasicBlock(CurMBB);
384 if (!CurMBB->empty())
385 RS->forward(prior(CurMBB->end()));
386 BitVector RegsLiveAtExit(TRI->getNumRegs());
387 RS->getRegsUsed(RegsLiveAtExit, false);
388 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
389 if (RegsLiveAtExit[i])
390 NewMBB->addLiveIn(i);
394 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
395 /// after it, replacing it with an unconditional branch to NewDest.
396 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
397 MachineBasicBlock *NewDest) {
398 MachineBasicBlock *CurMBB = OldInst->getParent();
400 TII->ReplaceTailWithBranchTo(OldInst, NewDest);
402 // For targets that use the register scavenger, we must maintain LiveIns.
403 MaintainLiveIns(CurMBB, NewDest);
408 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the
409 /// MBB so that the part before the iterator falls into the part starting at the
410 /// iterator. This returns the new MBB.
411 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
412 MachineBasicBlock::iterator BBI1) {
413 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
416 MachineFunction &MF = *CurMBB.getParent();
418 // Create the fall-through block.
419 MachineFunction::iterator MBBI = &CurMBB;
420 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
421 CurMBB.getParent()->insert(++MBBI, NewMBB);
423 // Move all the successors of this block to the specified block.
424 NewMBB->transferSuccessors(&CurMBB);
426 // Add an edge from CurMBB to NewMBB for the fall-through.
427 CurMBB.addSuccessor(NewMBB);
429 // Splice the code over.
430 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
432 // For targets that use the register scavenger, we must maintain LiveIns.
433 MaintainLiveIns(&CurMBB, NewMBB);
438 /// EstimateRuntime - Make a rough estimate for how long it will take to run
439 /// the specified code.
440 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
441 MachineBasicBlock::iterator E) {
443 for (; I != E; ++I) {
444 if (I->isDebugValue())
448 else if (I->mayLoad() || I->mayStore())
456 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
457 // branches temporarily for tail merging). In the case where CurMBB ends
458 // with a conditional branch to the next block, optimize by reversing the
459 // test and conditionally branching to SuccMBB instead.
460 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
461 const TargetInstrInfo *TII) {
462 MachineFunction *MF = CurMBB->getParent();
463 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB));
464 MachineBasicBlock *TBB = 0, *FBB = 0;
465 SmallVector<MachineOperand, 4> Cond;
466 DebugLoc dl; // FIXME: this is nowhere
467 if (I != MF->end() &&
468 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
469 MachineBasicBlock *NextBB = I;
470 if (TBB == NextBB && !Cond.empty() && !FBB) {
471 if (!TII->ReverseBranchCondition(Cond)) {
472 TII->RemoveBranch(*CurMBB);
473 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl);
478 TII->InsertBranch(*CurMBB, SuccBB, NULL,
479 SmallVector<MachineOperand, 0>(), dl);
483 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
484 if (getHash() < o.getHash())
486 else if (getHash() > o.getHash())
488 else if (getBlock()->getNumber() < o.getBlock()->getNumber())
490 else if (getBlock()->getNumber() > o.getBlock()->getNumber())
493 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
494 // an object with itself.
495 #ifndef _GLIBCXX_DEBUG
496 llvm_unreachable("Predecessor appears twice");
503 /// CountTerminators - Count the number of terminators in the given
504 /// block and set I to the position of the first non-terminator, if there
505 /// is one, or MBB->end() otherwise.
506 static unsigned CountTerminators(MachineBasicBlock *MBB,
507 MachineBasicBlock::iterator &I) {
509 unsigned NumTerms = 0;
511 if (I == MBB->begin()) {
516 if (!I->isTerminator()) break;
522 /// ProfitableToMerge - Check if two machine basic blocks have a common tail
523 /// and decide if it would be profitable to merge those tails. Return the
524 /// length of the common tail and iterators to the first common instruction
526 static bool ProfitableToMerge(MachineBasicBlock *MBB1,
527 MachineBasicBlock *MBB2,
528 unsigned minCommonTailLength,
529 unsigned &CommonTailLen,
530 MachineBasicBlock::iterator &I1,
531 MachineBasicBlock::iterator &I2,
532 MachineBasicBlock *SuccBB,
533 MachineBasicBlock *PredBB) {
534 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
535 if (CommonTailLen == 0)
537 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
538 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
541 // It's almost always profitable to merge any number of non-terminator
542 // instructions with the block that falls through into the common successor.
543 if (MBB1 == PredBB || MBB2 == PredBB) {
544 MachineBasicBlock::iterator I;
545 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
546 if (CommonTailLen > NumTerms)
550 // If one of the blocks can be completely merged and happens to be in
551 // a position where the other could fall through into it, merge any number
552 // of instructions, because it can be done without a branch.
553 // TODO: If the blocks are not adjacent, move one of them so that they are?
554 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
556 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
559 // If both blocks have an unconditional branch temporarily stripped out,
560 // count that as an additional common instruction for the following
562 unsigned EffectiveTailLen = CommonTailLen;
563 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
564 !MBB1->back().isBarrier() &&
565 !MBB2->back().isBarrier())
568 // Check if the common tail is long enough to be worthwhile.
569 if (EffectiveTailLen >= minCommonTailLength)
572 // If we are optimizing for code size, 2 instructions in common is enough if
573 // we don't have to split a block. At worst we will be introducing 1 new
574 // branch instruction, which is likely to be smaller than the 2
575 // instructions that would be deleted in the merge.
576 MachineFunction *MF = MBB1->getParent();
577 if (EffectiveTailLen >= 2 &&
578 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
579 (I1 == MBB1->begin() || I2 == MBB2->begin()))
585 /// ComputeSameTails - Look through all the blocks in MergePotentials that have
586 /// hash CurHash (guaranteed to match the last element). Build the vector
587 /// SameTails of all those that have the (same) largest number of instructions
588 /// in common of any pair of these blocks. SameTails entries contain an
589 /// iterator into MergePotentials (from which the MachineBasicBlock can be
590 /// found) and a MachineBasicBlock::iterator into that MBB indicating the
591 /// instruction where the matching code sequence begins.
592 /// Order of elements in SameTails is the reverse of the order in which
593 /// those blocks appear in MergePotentials (where they are not necessarily
595 unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
596 unsigned minCommonTailLength,
597 MachineBasicBlock *SuccBB,
598 MachineBasicBlock *PredBB) {
599 unsigned maxCommonTailLength = 0U;
601 MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
602 MPIterator HighestMPIter = prior(MergePotentials.end());
603 for (MPIterator CurMPIter = prior(MergePotentials.end()),
604 B = MergePotentials.begin();
605 CurMPIter != B && CurMPIter->getHash() == CurHash;
607 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
608 unsigned CommonTailLen;
609 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
611 CommonTailLen, TrialBBI1, TrialBBI2,
613 if (CommonTailLen > maxCommonTailLength) {
615 maxCommonTailLength = CommonTailLen;
616 HighestMPIter = CurMPIter;
617 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
619 if (HighestMPIter == CurMPIter &&
620 CommonTailLen == maxCommonTailLength)
621 SameTails.push_back(SameTailElt(I, TrialBBI2));
627 return maxCommonTailLength;
630 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
631 /// MergePotentials, restoring branches at ends of blocks as appropriate.
632 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
633 MachineBasicBlock *SuccBB,
634 MachineBasicBlock *PredBB) {
635 MPIterator CurMPIter, B;
636 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
637 CurMPIter->getHash() == CurHash;
639 // Put the unconditional branch back, if we need one.
640 MachineBasicBlock *CurMBB = CurMPIter->getBlock();
641 if (SuccBB && CurMBB != PredBB)
642 FixTail(CurMBB, SuccBB, TII);
646 if (CurMPIter->getHash() != CurHash)
648 MergePotentials.erase(CurMPIter, MergePotentials.end());
651 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
652 /// only of the common tail. Create a block that does by splitting one.
653 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
654 unsigned maxCommonTailLength,
655 unsigned &commonTailIndex) {
657 unsigned TimeEstimate = ~0U;
658 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
659 // Use PredBB if possible; that doesn't require a new branch.
660 if (SameTails[i].getBlock() == PredBB) {
664 // Otherwise, make a (fairly bogus) choice based on estimate of
665 // how long it will take the various blocks to execute.
666 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
667 SameTails[i].getTailStartPos());
668 if (t <= TimeEstimate) {
674 MachineBasicBlock::iterator BBI =
675 SameTails[commonTailIndex].getTailStartPos();
676 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
678 // If the common tail includes any debug info we will take it pretty
679 // randomly from one of the inputs. Might be better to remove it?
680 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
681 << maxCommonTailLength);
683 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
685 DEBUG(dbgs() << "... failed!");
689 SameTails[commonTailIndex].setBlock(newMBB);
690 SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
692 // If we split PredBB, newMBB is the new predecessor.
699 // See if any of the blocks in MergePotentials (which all have a common single
700 // successor, or all have no successor) can be tail-merged. If there is a
701 // successor, any blocks in MergePotentials that are not tail-merged and
702 // are not immediately before Succ must have an unconditional branch to
703 // Succ added (but the predecessor/successor lists need no adjustment).
704 // The lone predecessor of Succ that falls through into Succ,
705 // if any, is given in PredBB.
707 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
708 MachineBasicBlock *PredBB) {
709 bool MadeChange = false;
711 // Except for the special cases below, tail-merge if there are at least
712 // this many instructions in common.
713 unsigned minCommonTailLength = TailMergeSize;
715 DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
716 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
717 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
718 << (i == e-1 ? "" : ", ");
721 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
723 dbgs() << " which has fall-through from BB#"
724 << PredBB->getNumber() << "\n";
726 dbgs() << "Looking for common tails of at least "
727 << minCommonTailLength << " instruction"
728 << (minCommonTailLength == 1 ? "" : "s") << '\n';
731 // Sort by hash value so that blocks with identical end sequences sort
733 std::stable_sort(MergePotentials.begin(), MergePotentials.end());
735 // Walk through equivalence sets looking for actual exact matches.
736 while (MergePotentials.size() > 1) {
737 unsigned CurHash = MergePotentials.back().getHash();
739 // Build SameTails, identifying the set of blocks with this hash code
740 // and with the maximum number of instructions in common.
741 unsigned maxCommonTailLength = ComputeSameTails(CurHash,
745 // If we didn't find any pair that has at least minCommonTailLength
746 // instructions in common, remove all blocks with this hash code and retry.
747 if (SameTails.empty()) {
748 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
752 // If one of the blocks is the entire common tail (and not the entry
753 // block, which we can't jump to), we can treat all blocks with this same
754 // tail at once. Use PredBB if that is one of the possibilities, as that
755 // will not introduce any extra branches.
756 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
757 getParent()->begin();
758 unsigned commonTailIndex = SameTails.size();
759 // If there are two blocks, check to see if one can be made to fall through
761 if (SameTails.size() == 2 &&
762 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
763 SameTails[1].tailIsWholeBlock())
765 else if (SameTails.size() == 2 &&
766 SameTails[1].getBlock()->isLayoutSuccessor(
767 SameTails[0].getBlock()) &&
768 SameTails[0].tailIsWholeBlock())
771 // Otherwise just pick one, favoring the fall-through predecessor if
773 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
774 MachineBasicBlock *MBB = SameTails[i].getBlock();
775 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
781 if (SameTails[i].tailIsWholeBlock())
786 if (commonTailIndex == SameTails.size() ||
787 (SameTails[commonTailIndex].getBlock() == PredBB &&
788 !SameTails[commonTailIndex].tailIsWholeBlock())) {
789 // None of the blocks consist entirely of the common tail.
790 // Split a block so that one does.
791 if (!CreateCommonTailOnlyBlock(PredBB,
792 maxCommonTailLength, commonTailIndex)) {
793 RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
798 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
799 // MBB is common tail. Adjust all other BB's to jump to this one.
800 // Traversal must be forwards so erases work.
801 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
803 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
804 if (commonTailIndex == i)
806 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
807 << (i == e-1 ? "" : ", "));
808 // Hack the end off BB i, making it jump to BB commonTailIndex instead.
809 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
810 // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
811 MergePotentials.erase(SameTails[i].getMPIter());
813 DEBUG(dbgs() << "\n");
814 // We leave commonTailIndex in the worklist in case there are other blocks
815 // that match it with a smaller number of instructions.
821 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
823 if (!EnableTailMerge) return false;
825 bool MadeChange = false;
827 // First find blocks with no successors.
828 MergePotentials.clear();
829 for (MachineFunction::iterator I = MF.begin(), E = MF.end();
830 I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
831 if (TriedMerging.count(I))
834 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
837 // If this is a large problem, avoid visiting the same basic blocks
839 if (MergePotentials.size() == TailMergeThreshold)
840 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
841 TriedMerging.insert(MergePotentials[i].getBlock());
842 // See if we can do any tail merging on those.
843 if (MergePotentials.size() >= 2)
844 MadeChange |= TryTailMergeBlocks(NULL, NULL);
846 // Look at blocks (IBB) with multiple predecessors (PBB).
847 // We change each predecessor to a canonical form, by
848 // (1) temporarily removing any unconditional branch from the predecessor
850 // (2) alter conditional branches so they branch to the other block
851 // not IBB; this may require adding back an unconditional branch to IBB
852 // later, where there wasn't one coming in. E.g.
854 // fallthrough to QBB
857 // with a conceptual B to IBB after that, which never actually exists.
858 // With those changes, we see whether the predecessors' tails match,
859 // and merge them if so. We change things out of canonical form and
860 // back to the way they were later in the process. (OptimizeBranches
861 // would undo some of this, but we can't use it, because we'd get into
862 // a compile-time infinite loop repeatedly doing and undoing the same
865 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
867 if (I->pred_size() >= 2) {
868 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
869 MachineBasicBlock *IBB = I;
870 MachineBasicBlock *PredBB = prior(I);
871 MergePotentials.clear();
872 for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
874 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
875 MachineBasicBlock *PBB = *P;
876 if (TriedMerging.count(PBB))
878 // Skip blocks that loop to themselves, can't tail merge these.
881 // Visit each predecessor only once.
882 if (!UniquePreds.insert(PBB))
884 // Skip blocks which may jump to a landing pad. Can't tail merge these.
885 if (PBB->getLandingPadSuccessor())
887 MachineBasicBlock *TBB = 0, *FBB = 0;
888 SmallVector<MachineOperand, 4> Cond;
889 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
890 // Failing case: IBB is the target of a cbr, and
891 // we cannot reverse the branch.
892 SmallVector<MachineOperand, 4> NewCond(Cond);
893 if (!Cond.empty() && TBB == IBB) {
894 if (TII->ReverseBranchCondition(NewCond))
896 // This is the QBB case described above
898 FBB = llvm::next(MachineFunction::iterator(PBB));
900 // Failing case: the only way IBB can be reached from PBB is via
901 // exception handling. Happens for landing pads. Would be nice
902 // to have a bit in the edge so we didn't have to do all this.
903 if (IBB->isLandingPad()) {
904 MachineFunction::iterator IP = PBB; IP++;
905 MachineBasicBlock *PredNextBB = NULL;
909 if (IBB != PredNextBB) // fallthrough
912 if (TBB != IBB && FBB != IBB) // cbr then ubr
914 } else if (Cond.empty()) {
915 if (TBB != IBB) // ubr
918 if (TBB != IBB && IBB != PredNextBB) // cbr
922 // Remove the unconditional branch at the end, if any.
923 if (TBB && (Cond.empty() || FBB)) {
924 DebugLoc dl; // FIXME: this is nowhere
925 TII->RemoveBranch(*PBB);
927 // reinsert conditional branch only, for now
928 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl);
930 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
933 // If this is a large problem, avoid visiting the same basic blocks
935 if (MergePotentials.size() == TailMergeThreshold)
936 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
937 TriedMerging.insert(MergePotentials[i].getBlock());
938 if (MergePotentials.size() >= 2)
939 MadeChange |= TryTailMergeBlocks(IBB, PredBB);
940 // Reinsert an unconditional branch if needed.
941 // The 1 below can occur as a result of removing blocks in
942 // TryTailMergeBlocks.
943 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
944 if (MergePotentials.size() == 1 &&
945 MergePotentials.begin()->getBlock() != PredBB)
946 FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
952 //===----------------------------------------------------------------------===//
953 // Branch Optimization
954 //===----------------------------------------------------------------------===//
956 bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
957 bool MadeChange = false;
959 // Make sure blocks are numbered in order
962 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
964 MachineBasicBlock *MBB = I++;
965 MadeChange |= OptimizeBlock(MBB);
967 // If it is dead, remove it.
968 if (MBB->pred_empty()) {
969 RemoveDeadBlock(MBB);
977 // Blocks should be considered empty if they contain only debug info;
978 // else the debug info would affect codegen.
979 static bool IsEmptyBlock(MachineBasicBlock *MBB) {
982 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
983 MBBI!=MBBE; ++MBBI) {
984 if (!MBBI->isDebugValue())
990 // Blocks with only debug info and branches should be considered the same
991 // as blocks with only branches.
992 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
993 MachineBasicBlock::iterator MBBI, MBBE;
994 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
995 if (!MBBI->isDebugValue())
998 return (MBBI->isBranch());
1001 /// IsBetterFallthrough - Return true if it would be clearly better to
1002 /// fall-through to MBB1 than to fall through into MBB2. This has to return
1003 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1004 /// result in infinite loops.
1005 static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1006 MachineBasicBlock *MBB2) {
1007 // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1008 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1009 // optimize branches that branch to either a return block or an assert block
1010 // into a fallthrough to the return.
1011 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
1013 // If there is a clear successor ordering we make sure that one block
1014 // will fall through to the next
1015 if (MBB1->isSuccessor(MBB2)) return true;
1016 if (MBB2->isSuccessor(MBB1)) return false;
1018 // Neither block consists entirely of debug info (per IsEmptyBlock check),
1019 // so we needn't test for falling off the beginning here.
1020 MachineBasicBlock::iterator MBB1I = --MBB1->end();
1021 while (MBB1I->isDebugValue())
1023 MachineBasicBlock::iterator MBB2I = --MBB2->end();
1024 while (MBB2I->isDebugValue())
1026 return MBB2I->isCall() && !MBB1I->isCall();
1029 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1030 /// instructions on the block. Always use the DebugLoc of the first
1031 /// branching instruction found unless its absent, in which case use the
1032 /// DebugLoc of the second if present.
1033 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1034 MachineBasicBlock::iterator I = MBB.end();
1035 if (I == MBB.begin())
1038 while (I->isDebugValue() && I != MBB.begin())
1041 return I->getDebugLoc();
1045 /// OptimizeBlock - Analyze and optimize control flow related to the specified
1046 /// block. This is never called on the entry block.
1047 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1048 bool MadeChange = false;
1049 MachineFunction &MF = *MBB->getParent();
1052 MachineFunction::iterator FallThrough = MBB;
1055 // If this block is empty, make everyone use its fall-through, not the block
1056 // explicitly. Landing pads should not do this since the landing-pad table
1057 // points to this block. Blocks with their addresses taken shouldn't be
1059 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
1060 // Dead block? Leave for cleanup later.
1061 if (MBB->pred_empty()) return MadeChange;
1063 if (FallThrough == MF.end()) {
1064 // TODO: Simplify preds to not branch here if possible!
1066 // Rewrite all predecessors of the old block to go to the fallthrough
1068 while (!MBB->pred_empty()) {
1069 MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1070 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
1072 // If MBB was the target of a jump table, update jump tables to go to the
1073 // fallthrough instead.
1074 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1075 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
1081 // Check to see if we can simplify the terminator of the block before this
1083 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
1085 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
1086 SmallVector<MachineOperand, 4> PriorCond;
1087 bool PriorUnAnalyzable =
1088 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1089 if (!PriorUnAnalyzable) {
1090 // If the CFG for the prior block has extra edges, remove them.
1091 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1092 !PriorCond.empty());
1094 // If the previous branch is conditional and both conditions go to the same
1095 // destination, remove the branch, replacing it with an unconditional one or
1097 if (PriorTBB && PriorTBB == PriorFBB) {
1098 DebugLoc dl = getBranchDebugLoc(PrevBB);
1099 TII->RemoveBranch(PrevBB);
1101 if (PriorTBB != MBB)
1102 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1105 goto ReoptimizeBlock;
1108 // If the previous block unconditionally falls through to this block and
1109 // this block has no other predecessors, move the contents of this block
1110 // into the prior block. This doesn't usually happen when SimplifyCFG
1111 // has been used, but it can happen if tail merging splits a fall-through
1112 // predecessor of a block.
1113 // This has to check PrevBB->succ_size() because EH edges are ignored by
1115 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1116 PrevBB.succ_size() == 1 &&
1117 !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
1118 DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1119 << "From MBB: " << *MBB);
1120 // Remove redundant DBG_VALUEs first.
1121 if (PrevBB.begin() != PrevBB.end()) {
1122 MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1124 MachineBasicBlock::iterator MBBIter = MBB->begin();
1125 // Check if DBG_VALUE at the end of PrevBB is identical to the
1126 // DBG_VALUE at the beginning of MBB.
1127 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1128 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
1129 if (!MBBIter->isIdenticalTo(PrevBBIter))
1131 MachineInstr *DuplicateDbg = MBBIter;
1132 ++MBBIter; -- PrevBBIter;
1133 DuplicateDbg->eraseFromParent();
1136 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1137 PrevBB.removeSuccessor(PrevBB.succ_begin());
1138 assert(PrevBB.succ_empty());
1139 PrevBB.transferSuccessors(MBB);
1144 // If the previous branch *only* branches to *this* block (conditional or
1145 // not) remove the branch.
1146 if (PriorTBB == MBB && PriorFBB == 0) {
1147 TII->RemoveBranch(PrevBB);
1150 goto ReoptimizeBlock;
1153 // If the prior block branches somewhere else on the condition and here if
1154 // the condition is false, remove the uncond second branch.
1155 if (PriorFBB == MBB) {
1156 DebugLoc dl = getBranchDebugLoc(PrevBB);
1157 TII->RemoveBranch(PrevBB);
1158 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
1161 goto ReoptimizeBlock;
1164 // If the prior block branches here on true and somewhere else on false, and
1165 // if the branch condition is reversible, reverse the branch to create a
1167 if (PriorTBB == MBB) {
1168 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1169 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1170 DebugLoc dl = getBranchDebugLoc(PrevBB);
1171 TII->RemoveBranch(PrevBB);
1172 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl);
1175 goto ReoptimizeBlock;
1179 // If this block has no successors (e.g. it is a return block or ends with
1180 // a call to a no-return function like abort or __cxa_throw) and if the pred
1181 // falls through into this block, and if it would otherwise fall through
1182 // into the block after this, move this block to the end of the function.
1184 // We consider it more likely that execution will stay in the function (e.g.
1185 // due to loops) than it is to exit it. This asserts in loops etc, moving
1186 // the assert condition out of the loop body.
1187 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
1188 MachineFunction::iterator(PriorTBB) == FallThrough &&
1189 !MBB->canFallThrough()) {
1190 bool DoTransform = true;
1192 // We have to be careful that the succs of PredBB aren't both no-successor
1193 // blocks. If neither have successors and if PredBB is the second from
1194 // last block in the function, we'd just keep swapping the two blocks for
1195 // last. Only do the swap if one is clearly better to fall through than
1197 if (FallThrough == --MF.end() &&
1198 !IsBetterFallthrough(PriorTBB, MBB))
1199 DoTransform = false;
1202 // Reverse the branch so we will fall through on the previous true cond.
1203 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1204 if (!TII->ReverseBranchCondition(NewPriorCond)) {
1205 DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1206 << "To make fallthrough to: " << *PriorTBB << "\n");
1208 DebugLoc dl = getBranchDebugLoc(PrevBB);
1209 TII->RemoveBranch(PrevBB);
1210 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl);
1212 // Move this block to the end of the function.
1213 MBB->moveAfter(--MF.end());
1222 // Analyze the branch in the current block.
1223 MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
1224 SmallVector<MachineOperand, 4> CurCond;
1225 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1226 if (!CurUnAnalyzable) {
1227 // If the CFG for the prior block has extra edges, remove them.
1228 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1230 // If this is a two-way branch, and the FBB branches to this block, reverse
1231 // the condition so the single-basic-block loop is faster. Instead of:
1232 // Loop: xxx; jcc Out; jmp Loop
1234 // Loop: xxx; jncc Loop; jmp Out
1235 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1236 SmallVector<MachineOperand, 4> NewCond(CurCond);
1237 if (!TII->ReverseBranchCondition(NewCond)) {
1238 DebugLoc dl = getBranchDebugLoc(*MBB);
1239 TII->RemoveBranch(*MBB);
1240 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1243 goto ReoptimizeBlock;
1247 // If this branch is the only thing in its block, see if we can forward
1248 // other blocks across it.
1249 if (CurTBB && CurCond.empty() && CurFBB == 0 &&
1250 IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1251 !MBB->hasAddressTaken()) {
1252 DebugLoc dl = getBranchDebugLoc(*MBB);
1253 // This block may contain just an unconditional branch. Because there can
1254 // be 'non-branch terminators' in the block, try removing the branch and
1255 // then seeing if the block is empty.
1256 TII->RemoveBranch(*MBB);
1257 // If the only things remaining in the block are debug info, remove these
1258 // as well, so this will behave the same as an empty block in non-debug
1260 if (!MBB->empty()) {
1261 bool NonDebugInfoFound = false;
1262 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1264 if (!I->isDebugValue()) {
1265 NonDebugInfoFound = true;
1269 if (!NonDebugInfoFound)
1270 // Make the block empty, losing the debug info (we could probably
1271 // improve this in some cases.)
1272 MBB->erase(MBB->begin(), MBB->end());
1274 // If this block is just an unconditional branch to CurTBB, we can
1275 // usually completely eliminate the block. The only case we cannot
1276 // completely eliminate the block is when the block before this one
1277 // falls through into MBB and we can't understand the prior block's branch
1280 bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1281 if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1282 !PrevBB.isSuccessor(MBB)) {
1283 // If the prior block falls through into us, turn it into an
1284 // explicit branch to us to make updates simpler.
1285 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1286 PriorTBB != MBB && PriorFBB != MBB) {
1287 if (PriorTBB == 0) {
1288 assert(PriorCond.empty() && PriorFBB == 0 &&
1289 "Bad branch analysis");
1292 assert(PriorFBB == 0 && "Machine CFG out of date!");
1295 DebugLoc pdl = getBranchDebugLoc(PrevBB);
1296 TII->RemoveBranch(PrevBB);
1297 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1300 // Iterate through all the predecessors, revectoring each in-turn.
1302 bool DidChange = false;
1303 bool HasBranchToSelf = false;
1304 while(PI != MBB->pred_size()) {
1305 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1307 // If this block has an uncond branch to itself, leave it.
1309 HasBranchToSelf = true;
1312 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1313 // If this change resulted in PMBB ending in a conditional
1314 // branch where both conditions go to the same destination,
1315 // change this to an unconditional branch (and fix the CFG).
1316 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
1317 SmallVector<MachineOperand, 4> NewCurCond;
1318 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
1319 NewCurFBB, NewCurCond, true);
1320 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1321 DebugLoc pdl = getBranchDebugLoc(*PMBB);
1322 TII->RemoveBranch(*PMBB);
1324 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, pdl);
1327 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
1332 // Change any jumptables to go to the new MBB.
1333 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1334 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1338 if (!HasBranchToSelf) return MadeChange;
1343 // Add the branch back if the block is more than just an uncond branch.
1344 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl);
1348 // If the prior block doesn't fall through into this block, and if this
1349 // block doesn't fall through into some other block, see if we can find a
1350 // place to move this block where a fall-through will happen.
1351 if (!PrevBB.canFallThrough()) {
1353 // Now we know that there was no fall-through into this block, check to
1354 // see if it has a fall-through into its successor.
1355 bool CurFallsThru = MBB->canFallThrough();
1357 if (!MBB->isLandingPad()) {
1358 // Check all the predecessors of this block. If one of them has no fall
1359 // throughs, move this block right after it.
1360 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
1361 E = MBB->pred_end(); PI != E; ++PI) {
1362 // Analyze the branch at the end of the pred.
1363 MachineBasicBlock *PredBB = *PI;
1364 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
1365 MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
1366 SmallVector<MachineOperand, 4> PredCond;
1367 if (PredBB != MBB && !PredBB->canFallThrough() &&
1368 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
1369 && (!CurFallsThru || !CurTBB || !CurFBB)
1370 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1371 // If the current block doesn't fall through, just move it.
1372 // If the current block can fall through and does not end with a
1373 // conditional branch, we need to append an unconditional jump to
1374 // the (current) next block. To avoid a possible compile-time
1375 // infinite loop, move blocks only backward in this case.
1376 // Also, if there are already 2 branches here, we cannot add a third;
1377 // this means we have the case
1382 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
1384 TII->InsertBranch(*MBB, NextBB, 0, CurCond, DebugLoc());
1386 MBB->moveAfter(PredBB);
1388 goto ReoptimizeBlock;
1393 if (!CurFallsThru) {
1394 // Check all successors to see if we can move this block before it.
1395 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
1396 E = MBB->succ_end(); SI != E; ++SI) {
1397 // Analyze the branch at the end of the block before the succ.
1398 MachineBasicBlock *SuccBB = *SI;
1399 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
1401 // If this block doesn't already fall-through to that successor, and if
1402 // the succ doesn't already have a block that can fall through into it,
1403 // and if the successor isn't an EH destination, we can arrange for the
1404 // fallthrough to happen.
1405 if (SuccBB != MBB && &*SuccPrev != MBB &&
1406 !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1407 !SuccBB->isLandingPad()) {
1408 MBB->moveBefore(SuccBB);
1410 goto ReoptimizeBlock;
1414 // Okay, there is no really great place to put this block. If, however,
1415 // the block before this one would be a fall-through if this block were
1416 // removed, move this block to the end of the function.
1417 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
1418 SmallVector<MachineOperand, 4> PrevCond;
1419 if (FallThrough != MF.end() &&
1420 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1421 PrevBB.isSuccessor(FallThrough)) {
1422 MBB->moveAfter(--MF.end());
1432 //===----------------------------------------------------------------------===//
1433 // Hoist Common Code
1434 //===----------------------------------------------------------------------===//
1436 /// HoistCommonCode - Hoist common instruction sequences at the start of basic
1437 /// blocks to their common predecessor.
1438 bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1439 bool MadeChange = false;
1440 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1441 MachineBasicBlock *MBB = I++;
1442 MadeChange |= HoistCommonCodeInSuccs(MBB);
1448 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1449 /// its 'true' successor.
1450 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1451 MachineBasicBlock *TrueBB) {
1452 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1453 E = BB->succ_end(); SI != E; ++SI) {
1454 MachineBasicBlock *SuccBB = *SI;
1455 if (SuccBB != TrueBB)
1461 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1462 /// in successors to. The location is ususally just before the terminator,
1463 /// however if the terminator is a conditional branch and its previous
1464 /// instruction is the flag setting instruction, the previous instruction is
1465 /// the preferred location. This function also gathers uses and defs of the
1466 /// instructions from the insertion point to the end of the block. The data is
1467 /// used by HoistCommonCodeInSuccs to ensure safety.
1469 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1470 const TargetInstrInfo *TII,
1471 const TargetRegisterInfo *TRI,
1472 SmallSet<unsigned,4> &Uses,
1473 SmallSet<unsigned,4> &Defs) {
1474 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1475 if (!TII->isUnpredicatedTerminator(Loc))
1478 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
1479 const MachineOperand &MO = Loc->getOperand(i);
1482 unsigned Reg = MO.getReg();
1487 for (const uint16_t *AS = TRI->getAliasSet(Reg); *AS; ++AS)
1489 } else if (!MO.isDead())
1490 // Don't try to hoist code in the rare case the terminator defines a
1491 // register that is later used.
1497 if (Loc == MBB->begin())
1500 // The terminator is probably a conditional branch, try not to separate the
1501 // branch from condition setting instruction.
1502 MachineBasicBlock::iterator PI = Loc;
1504 while (PI != MBB->begin() && Loc->isDebugValue())
1508 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
1509 const MachineOperand &MO = PI->getOperand(i);
1510 // If PI has a regmask operand, it is probably a call. Separate away.
1513 if (!MO.isReg() || MO.isUse())
1515 unsigned Reg = MO.getReg();
1518 if (Uses.count(Reg))
1522 // The condition setting instruction is not just before the conditional
1526 // Be conservative, don't insert instruction above something that may have
1527 // side-effects. And since it's potentially bad to separate flag setting
1528 // instruction from the conditional branch, just abort the optimization
1530 // Also avoid moving code above predicated instruction since it's hard to
1531 // reason about register liveness with predicated instruction.
1532 bool DontMoveAcrossStore = true;
1533 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) ||
1534 TII->isPredicated(PI))
1538 // Find out what registers are live. Note this routine is ignoring other live
1539 // registers which are only used by instructions in successor blocks.
1540 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
1541 const MachineOperand &MO = PI->getOperand(i);
1544 unsigned Reg = MO.getReg();
1549 for (const uint16_t *AS = TRI->getAliasSet(Reg); *AS; ++AS)
1552 if (Uses.count(Reg)) {
1554 for (const uint16_t *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1555 Uses.erase(*SR); // Use getSubRegisters to be conservative
1558 for (const uint16_t *AS = TRI->getAliasSet(Reg); *AS; ++AS)
1566 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
1567 /// sequence at the start of the function, move the instructions before MBB
1568 /// terminator if it's legal.
1569 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1570 MachineBasicBlock *TBB = 0, *FBB = 0;
1571 SmallVector<MachineOperand, 4> Cond;
1572 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1575 if (!FBB) FBB = findFalseBlock(MBB, TBB);
1577 // Malformed bcc? True and false blocks are the same?
1580 // Restrict the optimization to cases where MBB is the only predecessor,
1581 // it is an obvious win.
1582 if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1585 // Find a suitable position to hoist the common instructions to. Also figure
1586 // out which registers are used or defined by instructions from the insertion
1587 // point to the end of the block.
1588 SmallSet<unsigned, 4> Uses, Defs;
1589 MachineBasicBlock::iterator Loc =
1590 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
1591 if (Loc == MBB->end())
1594 bool HasDups = false;
1595 SmallVector<unsigned, 4> LocalDefs;
1596 SmallSet<unsigned, 4> LocalDefsSet;
1597 MachineBasicBlock::iterator TIB = TBB->begin();
1598 MachineBasicBlock::iterator FIB = FBB->begin();
1599 MachineBasicBlock::iterator TIE = TBB->end();
1600 MachineBasicBlock::iterator FIE = FBB->end();
1601 while (TIB != TIE && FIB != FIE) {
1602 // Skip dbg_value instructions. These do not count.
1603 if (TIB->isDebugValue()) {
1604 while (TIB != TIE && TIB->isDebugValue())
1609 if (FIB->isDebugValue()) {
1610 while (FIB != FIE && FIB->isDebugValue())
1615 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
1618 if (TII->isPredicated(TIB))
1619 // Hard to reason about register liveness with predicated instruction.
1623 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1624 MachineOperand &MO = TIB->getOperand(i);
1625 // Don't attempt to hoist instructions with register masks.
1626 if (MO.isRegMask()) {
1632 unsigned Reg = MO.getReg();
1636 if (Uses.count(Reg)) {
1637 // Avoid clobbering a register that's used by the instruction at
1638 // the point of insertion.
1643 if (Defs.count(Reg) && !MO.isDead()) {
1644 // Don't hoist the instruction if the def would be clobber by the
1645 // instruction at the point insertion. FIXME: This is overly
1646 // conservative. It should be possible to hoist the instructions
1647 // in BB2 in the following example:
1649 // r1, eflag = op1 r2, r3
1658 } else if (!LocalDefsSet.count(Reg)) {
1659 if (Defs.count(Reg)) {
1660 // Use is defined by the instruction at the point of insertion.
1665 if (MO.isKill() && Uses.count(Reg))
1666 // Kills a register that's read by the instruction at the point of
1667 // insertion. Remove the kill marker.
1668 MO.setIsKill(false);
1674 bool DontMoveAcrossStore = true;
1675 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore))
1678 // Remove kills from LocalDefsSet, these registers had short live ranges.
1679 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1680 MachineOperand &MO = TIB->getOperand(i);
1681 if (!MO.isReg() || !MO.isUse() || !MO.isKill())
1683 unsigned Reg = MO.getReg();
1684 if (!Reg || !LocalDefsSet.count(Reg))
1686 for (const uint16_t *OR = TRI->getOverlaps(Reg); *OR; ++OR)
1687 LocalDefsSet.erase(*OR);
1690 // Track local defs so we can update liveins.
1691 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
1692 MachineOperand &MO = TIB->getOperand(i);
1693 if (!MO.isReg() || !MO.isDef() || MO.isDead())
1695 unsigned Reg = MO.getReg();
1698 LocalDefs.push_back(Reg);
1699 for (const uint16_t *OR = TRI->getOverlaps(Reg); *OR; ++OR)
1700 LocalDefsSet.insert(*OR);
1711 MBB->splice(Loc, TBB, TBB->begin(), TIB);
1712 FBB->erase(FBB->begin(), FIB);
1715 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
1716 unsigned Def = LocalDefs[i];
1717 if (LocalDefsSet.count(Def)) {
1718 TBB->addLiveIn(Def);
1719 FBB->addLiveIn(Def);