1 //===-- IfConversion.cpp - Machine code if conversion pass. ---------------===//
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 file implements the machine instruction level if-conversion pass.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "ifcvt"
15 #include "BranchFolding.h"
16 #include "llvm/Function.h"
17 #include "llvm/CodeGen/Passes.h"
18 #include "llvm/CodeGen/MachineModuleInfo.h"
19 #include "llvm/CodeGen/MachineFunctionPass.h"
20 #include "llvm/CodeGen/MachineLoopInfo.h"
21 #include "llvm/Target/TargetInstrInfo.h"
22 #include "llvm/Target/TargetInstrItineraries.h"
23 #include "llvm/Target/TargetLowering.h"
24 #include "llvm/Target/TargetMachine.h"
25 #include "llvm/Target/TargetRegisterInfo.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/raw_ostream.h"
30 #include "llvm/ADT/DepthFirstIterator.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/ADT/STLExtras.h"
35 // Hidden options for help debugging.
36 static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden);
37 static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden);
38 static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden);
39 static cl::opt<bool> DisableSimple("disable-ifcvt-simple",
40 cl::init(false), cl::Hidden);
41 static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false",
42 cl::init(false), cl::Hidden);
43 static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle",
44 cl::init(false), cl::Hidden);
45 static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev",
46 cl::init(false), cl::Hidden);
47 static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false",
48 cl::init(false), cl::Hidden);
49 static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev",
50 cl::init(false), cl::Hidden);
51 static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond",
52 cl::init(false), cl::Hidden);
53 static cl::opt<bool> IfCvtBranchFold("ifcvt-branch-fold",
54 cl::init(true), cl::Hidden);
56 STATISTIC(NumSimple, "Number of simple if-conversions performed");
57 STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed");
58 STATISTIC(NumTriangle, "Number of triangle if-conversions performed");
59 STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed");
60 STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed");
61 STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed");
62 STATISTIC(NumDiamonds, "Number of diamond if-conversions performed");
63 STATISTIC(NumIfConvBBs, "Number of if-converted blocks");
64 STATISTIC(NumDupBBs, "Number of duplicated blocks");
67 class IfConverter : public MachineFunctionPass {
69 ICNotClassfied, // BB data valid, but not classified.
70 ICSimpleFalse, // Same as ICSimple, but on the false path.
71 ICSimple, // BB is entry of an one split, no rejoin sub-CFG.
72 ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition.
73 ICTriangleRev, // Same as ICTriangle, but true path rev condition.
74 ICTriangleFalse, // Same as ICTriangle, but on the false path.
75 ICTriangle, // BB is entry of a triangle sub-CFG.
76 ICDiamond // BB is entry of a diamond sub-CFG.
79 /// BBInfo - One per MachineBasicBlock, this is used to cache the result
80 /// if-conversion feasibility analysis. This includes results from
81 /// TargetInstrInfo::AnalyzeBranch() (i.e. TBB, FBB, and Cond), and its
82 /// classification, and common tail block of its successors (if it's a
83 /// diamond shape), its size, whether it's predicable, and whether any
84 /// instruction can clobber the 'would-be' predicate.
86 /// IsDone - True if BB is not to be considered for ifcvt.
87 /// IsBeingAnalyzed - True if BB is currently being analyzed.
88 /// IsAnalyzed - True if BB has been analyzed (info is still valid).
89 /// IsEnqueued - True if BB has been enqueued to be ifcvt'ed.
90 /// IsBrAnalyzable - True if AnalyzeBranch() returns false.
91 /// HasFallThrough - True if BB may fallthrough to the following BB.
92 /// IsUnpredicable - True if BB is known to be unpredicable.
93 /// ClobbersPred - True if BB could modify predicates (e.g. has
95 /// NonPredSize - Number of non-predicated instructions.
96 /// BB - Corresponding MachineBasicBlock.
97 /// TrueBB / FalseBB- See AnalyzeBranch().
98 /// BrCond - Conditions for end of block conditional branches.
99 /// Predicate - Predicate used in the BB.
102 bool IsBeingAnalyzed : 1;
105 bool IsBrAnalyzable : 1;
106 bool HasFallThrough : 1;
107 bool IsUnpredicable : 1;
108 bool CannotBeCopied : 1;
109 bool ClobbersPred : 1;
110 unsigned NonPredSize;
111 MachineBasicBlock *BB;
112 MachineBasicBlock *TrueBB;
113 MachineBasicBlock *FalseBB;
114 SmallVector<MachineOperand, 4> BrCond;
115 SmallVector<MachineOperand, 4> Predicate;
116 BBInfo() : IsDone(false), IsBeingAnalyzed(false),
117 IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false),
118 HasFallThrough(false), IsUnpredicable(false),
119 CannotBeCopied(false), ClobbersPred(false), NonPredSize(0),
120 BB(0), TrueBB(0), FalseBB(0) {}
123 /// IfcvtToken - Record information about pending if-conversions to attempt:
124 /// BBI - Corresponding BBInfo.
125 /// Kind - Type of block. See IfcvtKind.
126 /// NeedSubsumption - True if the to-be-predicated BB has already been
128 /// NumDups - Number of instructions that would be duplicated due
129 /// to this if-conversion. (For diamonds, the number of
130 /// identical instructions at the beginnings of both
132 /// NumDups2 - For diamonds, the number of identical instructions
133 /// at the ends of both paths.
137 bool NeedSubsumption;
140 IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0)
141 : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {}
144 /// Roots - Basic blocks that do not have successors. These are the starting
145 /// points of Graph traversal.
146 std::vector<MachineBasicBlock*> Roots;
148 /// BBAnalysis - Results of if-conversion feasibility analysis indexed by
149 /// basic block number.
150 std::vector<BBInfo> BBAnalysis;
152 const TargetLowering *TLI;
153 const TargetInstrInfo *TII;
154 const TargetRegisterInfo *TRI;
155 const InstrItineraryData *InstrItins;
156 const MachineLoopInfo *MLI;
161 IfConverter() : MachineFunctionPass(ID), FnNum(-1) {}
163 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
164 AU.addRequired<MachineLoopInfo>();
165 MachineFunctionPass::getAnalysisUsage(AU);
168 virtual bool runOnMachineFunction(MachineFunction &MF);
169 virtual const char *getPassName() const { return "If Converter"; }
172 bool ReverseBranchCondition(BBInfo &BBI);
173 bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups,
174 float Prediction, float Confidence) const;
175 bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
176 bool FalseBranch, unsigned &Dups,
177 float Prediction, float Confidence) const;
178 bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
179 unsigned &Dups1, unsigned &Dups2) const;
180 void ScanInstructions(BBInfo &BBI);
181 BBInfo &AnalyzeBlock(MachineBasicBlock *BB,
182 std::vector<IfcvtToken*> &Tokens);
183 bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl<MachineOperand> &Cond,
184 bool isTriangle = false, bool RevBranch = false);
185 void AnalyzeBlocks(MachineFunction &MF, std::vector<IfcvtToken*> &Tokens);
186 void InvalidatePreds(MachineBasicBlock *BB);
187 void RemoveExtraEdges(BBInfo &BBI);
188 bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind);
189 bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind);
190 bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
191 unsigned NumDups1, unsigned NumDups2);
192 void PredicateBlock(BBInfo &BBI,
193 MachineBasicBlock::iterator E,
194 SmallVectorImpl<MachineOperand> &Cond,
195 SmallSet<unsigned, 4> &Redefs);
196 void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
197 SmallVectorImpl<MachineOperand> &Cond,
198 SmallSet<unsigned, 4> &Redefs,
199 bool IgnoreBr = false);
200 void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true);
202 bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, unsigned Size,
203 float Prediction, float Confidence) const {
204 return Size > 0 && TII->isProfitableToIfCvt(BB, Size,
205 Prediction, Confidence);
208 bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, unsigned TSize,
209 MachineBasicBlock &FBB, unsigned FSize,
210 float Prediction, float Confidence) const {
211 return TSize > 0 && FSize > 0 &&
212 TII->isProfitableToIfCvt(TBB, TSize, FBB, FSize,
213 Prediction, Confidence);
216 // blockAlwaysFallThrough - Block ends without a terminator.
217 bool blockAlwaysFallThrough(BBInfo &BBI) const {
218 return BBI.IsBrAnalyzable && BBI.TrueBB == NULL;
221 // IfcvtTokenCmp - Used to sort if-conversion candidates.
222 static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) {
223 int Incr1 = (C1->Kind == ICDiamond)
224 ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups;
225 int Incr2 = (C2->Kind == ICDiamond)
226 ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups;
229 else if (Incr1 == Incr2) {
230 // Favors subsumption.
231 if (C1->NeedSubsumption == false && C2->NeedSubsumption == true)
233 else if (C1->NeedSubsumption == C2->NeedSubsumption) {
234 // Favors diamond over triangle, etc.
235 if ((unsigned)C1->Kind < (unsigned)C2->Kind)
237 else if (C1->Kind == C2->Kind)
238 return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber();
245 char IfConverter::ID = 0;
248 INITIALIZE_PASS_BEGIN(IfConverter, "if-converter", "If Converter", false, false)
249 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
250 INITIALIZE_PASS_END(IfConverter, "if-converter", "If Converter", false, false)
252 FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); }
254 bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
255 TLI = MF.getTarget().getTargetLowering();
256 TII = MF.getTarget().getInstrInfo();
257 TRI = MF.getTarget().getRegisterInfo();
258 MLI = &getAnalysis<MachineLoopInfo>();
259 InstrItins = MF.getTarget().getInstrItineraryData();
260 if (!TII) return false;
262 // Tail merge tend to expose more if-conversion opportunities.
263 BranchFolder BF(true);
264 bool BFChange = BF.OptimizeFunction(MF, TII,
265 MF.getTarget().getRegisterInfo(),
266 getAnalysisIfAvailable<MachineModuleInfo>());
268 DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'"
269 << MF.getFunction()->getName() << "\'");
271 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) {
272 DEBUG(dbgs() << " skipped\n");
275 DEBUG(dbgs() << "\n");
278 BBAnalysis.resize(MF.getNumBlockIDs());
280 // Look for root nodes, i.e. blocks without successors.
281 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
285 std::vector<IfcvtToken*> Tokens;
287 unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle +
288 NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds;
289 while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) {
290 // Do an initial analysis for each basic block and find all the potential
291 // candidates to perform if-conversion.
293 AnalyzeBlocks(MF, Tokens);
294 while (!Tokens.empty()) {
295 IfcvtToken *Token = Tokens.back();
297 BBInfo &BBI = Token->BBI;
298 IfcvtKind Kind = Token->Kind;
299 unsigned NumDups = Token->NumDups;
300 unsigned NumDups2 = Token->NumDups2;
304 // If the block has been evicted out of the queue or it has already been
305 // marked dead (due to it being predicated), then skip it.
307 BBI.IsEnqueued = false;
311 BBI.IsEnqueued = false;
315 default: assert(false && "Unexpected!");
318 case ICSimpleFalse: {
319 bool isFalse = Kind == ICSimpleFalse;
320 if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break;
321 DEBUG(dbgs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ?
323 << "): BB#" << BBI.BB->getNumber() << " ("
324 << ((Kind == ICSimpleFalse)
325 ? BBI.FalseBB->getNumber()
326 : BBI.TrueBB->getNumber()) << ") ");
327 RetVal = IfConvertSimple(BBI, Kind);
328 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
330 if (isFalse) ++NumSimpleFalse;
337 case ICTriangleFalse:
338 case ICTriangleFRev: {
339 bool isFalse = Kind == ICTriangleFalse;
340 bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev);
341 if (DisableTriangle && !isFalse && !isRev) break;
342 if (DisableTriangleR && !isFalse && isRev) break;
343 if (DisableTriangleF && isFalse && !isRev) break;
344 if (DisableTriangleFR && isFalse && isRev) break;
345 DEBUG(dbgs() << "Ifcvt (Triangle");
347 DEBUG(dbgs() << " false");
349 DEBUG(dbgs() << " rev");
350 DEBUG(dbgs() << "): BB#" << BBI.BB->getNumber() << " (T:"
351 << BBI.TrueBB->getNumber() << ",F:"
352 << BBI.FalseBB->getNumber() << ") ");
353 RetVal = IfConvertTriangle(BBI, Kind);
354 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
357 if (isRev) ++NumTriangleFRev;
358 else ++NumTriangleFalse;
360 if (isRev) ++NumTriangleRev;
367 if (DisableDiamond) break;
368 DEBUG(dbgs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:"
369 << BBI.TrueBB->getNumber() << ",F:"
370 << BBI.FalseBB->getNumber() << ") ");
371 RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2);
372 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
373 if (RetVal) ++NumDiamonds;
380 NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev +
381 NumTriangleFalse + NumTriangleFRev + NumDiamonds;
382 if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit)
388 MadeChange |= Change;
391 // Delete tokens in case of early exit.
392 while (!Tokens.empty()) {
393 IfcvtToken *Token = Tokens.back();
402 if (MadeChange && IfCvtBranchFold) {
403 BranchFolder BF(false);
404 BF.OptimizeFunction(MF, TII,
405 MF.getTarget().getRegisterInfo(),
406 getAnalysisIfAvailable<MachineModuleInfo>());
409 MadeChange |= BFChange;
413 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
414 /// its 'true' successor.
415 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
416 MachineBasicBlock *TrueBB) {
417 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
418 E = BB->succ_end(); SI != E; ++SI) {
419 MachineBasicBlock *SuccBB = *SI;
420 if (SuccBB != TrueBB)
426 /// ReverseBranchCondition - Reverse the condition of the end of the block
427 /// branch. Swap block's 'true' and 'false' successors.
428 bool IfConverter::ReverseBranchCondition(BBInfo &BBI) {
429 DebugLoc dl; // FIXME: this is nowhere
430 if (!TII->ReverseBranchCondition(BBI.BrCond)) {
431 TII->RemoveBranch(*BBI.BB);
432 TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond, dl);
433 std::swap(BBI.TrueBB, BBI.FalseBB);
439 /// getNextBlock - Returns the next block in the function blocks ordering. If
440 /// it is the end, returns NULL.
441 static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) {
442 MachineFunction::iterator I = BB;
443 MachineFunction::iterator E = BB->getParent()->end();
449 /// ValidSimple - Returns true if the 'true' block (along with its
450 /// predecessor) forms a valid simple shape for ifcvt. It also returns the
451 /// number of instructions that the ifcvt would need to duplicate if performed
453 bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups,
454 float Prediction, float Confidence) const {
456 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
459 if (TrueBBI.IsBrAnalyzable)
462 if (TrueBBI.BB->pred_size() > 1) {
463 if (TrueBBI.CannotBeCopied ||
464 !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize,
465 Prediction, Confidence))
467 Dups = TrueBBI.NonPredSize;
473 /// ValidTriangle - Returns true if the 'true' and 'false' blocks (along
474 /// with their common predecessor) forms a valid triangle shape for ifcvt.
475 /// If 'FalseBranch' is true, it checks if 'true' block's false branch
476 /// branches to the 'false' block rather than the other way around. It also
477 /// returns the number of instructions that the ifcvt would need to duplicate
478 /// if performed in 'Dups'.
479 bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
480 bool FalseBranch, unsigned &Dups,
481 float Prediction, float Confidence) const {
483 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
486 if (TrueBBI.BB->pred_size() > 1) {
487 if (TrueBBI.CannotBeCopied)
490 unsigned Size = TrueBBI.NonPredSize;
491 if (TrueBBI.IsBrAnalyzable) {
492 if (TrueBBI.TrueBB && TrueBBI.BrCond.empty())
493 // Ends with an unconditional branch. It will be removed.
496 MachineBasicBlock *FExit = FalseBranch
497 ? TrueBBI.TrueBB : TrueBBI.FalseBB;
499 // Require a conditional branch
503 if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size,
504 Prediction, Confidence))
509 MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB;
510 if (!TExit && blockAlwaysFallThrough(TrueBBI)) {
511 MachineFunction::iterator I = TrueBBI.BB;
512 if (++I == TrueBBI.BB->getParent()->end())
516 return TExit && TExit == FalseBBI.BB;
520 MachineBasicBlock::iterator firstNonBranchInst(MachineBasicBlock *BB,
521 const TargetInstrInfo *TII) {
522 MachineBasicBlock::iterator I = BB->end();
523 while (I != BB->begin()) {
525 if (!I->getDesc().isBranch())
531 /// ValidDiamond - Returns true if the 'true' and 'false' blocks (along
532 /// with their common predecessor) forms a valid diamond shape for ifcvt.
533 bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
534 unsigned &Dups1, unsigned &Dups2) const {
536 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone ||
537 FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone)
540 MachineBasicBlock *TT = TrueBBI.TrueBB;
541 MachineBasicBlock *FT = FalseBBI.TrueBB;
543 if (!TT && blockAlwaysFallThrough(TrueBBI))
544 TT = getNextBlock(TrueBBI.BB);
545 if (!FT && blockAlwaysFallThrough(FalseBBI))
546 FT = getNextBlock(FalseBBI.BB);
549 if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable))
551 if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1)
554 // FIXME: Allow true block to have an early exit?
555 if (TrueBBI.FalseBB || FalseBBI.FalseBB ||
556 (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred))
559 MachineBasicBlock::iterator TI = TrueBBI.BB->begin();
560 MachineBasicBlock::iterator FI = FalseBBI.BB->begin();
561 MachineBasicBlock::iterator TIE = TrueBBI.BB->end();
562 MachineBasicBlock::iterator FIE = FalseBBI.BB->end();
563 // Skip dbg_value instructions
564 while (TI != TIE && TI->isDebugValue())
566 while (FI != FIE && FI->isDebugValue())
568 while (TI != TIE && FI != FIE) {
569 // Skip dbg_value instructions. These do not count.
570 if (TI->isDebugValue()) {
571 while (TI != TIE && TI->isDebugValue())
576 if (FI->isDebugValue()) {
577 while (FI != FIE && FI->isDebugValue())
582 if (!TI->isIdenticalTo(FI))
589 TI = firstNonBranchInst(TrueBBI.BB, TII);
590 FI = firstNonBranchInst(FalseBBI.BB, TII);
591 MachineBasicBlock::iterator TIB = TrueBBI.BB->begin();
592 MachineBasicBlock::iterator FIB = FalseBBI.BB->begin();
593 // Skip dbg_value instructions at end of the bb's.
594 while (TI != TIB && TI->isDebugValue())
596 while (FI != FIB && FI->isDebugValue())
598 while (TI != TIB && FI != FIB) {
599 // Skip dbg_value instructions. These do not count.
600 if (TI->isDebugValue()) {
601 while (TI != TIB && TI->isDebugValue())
606 if (FI->isDebugValue()) {
607 while (FI != FIB && FI->isDebugValue())
612 if (!TI->isIdenticalTo(FI))
622 /// ScanInstructions - Scan all the instructions in the block to determine if
623 /// the block is predicable. In most cases, that means all the instructions
624 /// in the block are isPredicable(). Also checks if the block contains any
625 /// instruction which can clobber a predicate (e.g. condition code register).
626 /// If so, the block is not predicable unless it's the last instruction.
627 void IfConverter::ScanInstructions(BBInfo &BBI) {
631 bool AlreadyPredicated = BBI.Predicate.size() > 0;
632 // First analyze the end of BB branches.
633 BBI.TrueBB = BBI.FalseBB = NULL;
636 !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond);
637 BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL;
639 if (BBI.BrCond.size()) {
640 // No false branch. This BB must end with a conditional branch and a
643 BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB);
645 // Malformed bcc? True and false blocks are the same?
646 BBI.IsUnpredicable = true;
651 // Then scan all the instructions.
653 BBI.ClobbersPred = false;
654 for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end();
656 if (I->isDebugValue())
659 const TargetInstrDesc &TID = I->getDesc();
660 if (TID.isNotDuplicable())
661 BBI.CannotBeCopied = true;
663 bool isPredicated = TII->isPredicated(I);
664 bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch();
668 unsigned NumOps = TII->getNumMicroOps(&*I, InstrItins);
669 BBI.NonPredSize += NumOps;
670 } else if (!AlreadyPredicated) {
671 // FIXME: This instruction is already predicated before the
672 // if-conversion pass. It's probably something like a conditional move.
673 // Mark this block unpredicable for now.
674 BBI.IsUnpredicable = true;
679 if (BBI.ClobbersPred && !isPredicated) {
680 // Predicate modification instruction should end the block (except for
681 // already predicated instructions and end of block branches).
683 // A conditional branch is not predicable, but it may be eliminated.
687 // Predicate may have been modified, the subsequent (currently)
688 // unpredicated instructions cannot be correctly predicated.
689 BBI.IsUnpredicable = true;
693 // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are
694 // still potentially predicable.
695 std::vector<MachineOperand> PredDefs;
696 if (TII->DefinesPredicate(I, PredDefs))
697 BBI.ClobbersPred = true;
699 if (!TII->isPredicable(I)) {
700 BBI.IsUnpredicable = true;
706 /// FeasibilityAnalysis - Determine if the block is a suitable candidate to be
707 /// predicated by the specified predicate.
708 bool IfConverter::FeasibilityAnalysis(BBInfo &BBI,
709 SmallVectorImpl<MachineOperand> &Pred,
710 bool isTriangle, bool RevBranch) {
711 // If the block is dead or unpredicable, then it cannot be predicated.
712 if (BBI.IsDone || BBI.IsUnpredicable)
715 // If it is already predicated, check if its predicate subsumes the new
717 if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred))
720 if (BBI.BrCond.size()) {
724 // Test predicate subsumption.
725 SmallVector<MachineOperand, 4> RevPred(Pred.begin(), Pred.end());
726 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
728 if (TII->ReverseBranchCondition(Cond))
731 if (TII->ReverseBranchCondition(RevPred) ||
732 !TII->SubsumesPredicate(Cond, RevPred))
739 /// AnalyzeBlock - Analyze the structure of the sub-CFG starting from
740 /// the specified block. Record its successors and whether it looks like an
741 /// if-conversion candidate.
742 IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
743 std::vector<IfcvtToken*> &Tokens) {
744 BBInfo &BBI = BBAnalysis[BB->getNumber()];
746 if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed)
750 BBI.IsBeingAnalyzed = true;
752 ScanInstructions(BBI);
754 // Unanalyzable or ends with fallthrough or unconditional branch.
755 if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) {
756 BBI.IsBeingAnalyzed = false;
757 BBI.IsAnalyzed = true;
761 // Do not ifcvt if either path is a back edge to the entry block.
762 if (BBI.TrueBB == BB || BBI.FalseBB == BB) {
763 BBI.IsBeingAnalyzed = false;
764 BBI.IsAnalyzed = true;
768 // Do not ifcvt if true and false fallthrough blocks are the same.
770 BBI.IsBeingAnalyzed = false;
771 BBI.IsAnalyzed = true;
775 BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens);
776 BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens);
778 if (TrueBBI.IsDone && FalseBBI.IsDone) {
779 BBI.IsBeingAnalyzed = false;
780 BBI.IsAnalyzed = true;
784 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end());
785 bool CanRevCond = !TII->ReverseBranchCondition(RevCond);
789 bool TNeedSub = TrueBBI.Predicate.size() > 0;
790 bool FNeedSub = FalseBBI.Predicate.size() > 0;
791 bool Enqueued = false;
793 // Try to predict the branch, using loop info to guide us.
794 // General heuristics are:
795 // - backedge -> 90% taken
796 // - early exit -> 20% taken
797 // - branch predictor confidence -> 90%
798 float Prediction = 0.5f;
799 float Confidence = 0.9f;
800 MachineLoop *Loop = MLI->getLoopFor(BB);
802 if (TrueBBI.BB == Loop->getHeader())
804 else if (FalseBBI.BB == Loop->getHeader())
807 MachineLoop *TrueLoop = MLI->getLoopFor(TrueBBI.BB);
808 MachineLoop *FalseLoop = MLI->getLoopFor(FalseBBI.BB);
809 if (!TrueLoop || TrueLoop->getParentLoop() == Loop)
811 else if (!FalseLoop || FalseLoop->getParentLoop() == Loop)
815 if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) &&
816 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize - (Dups + Dups2),
817 *FalseBBI.BB, FalseBBI.NonPredSize - (Dups + Dups2),
818 Prediction, Confidence) &&
819 FeasibilityAnalysis(TrueBBI, BBI.BrCond) &&
820 FeasibilityAnalysis(FalseBBI, RevCond)) {
828 // Note TailBB can be empty.
829 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups,
834 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction, Confidence) &&
835 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize,
836 Prediction, Confidence) &&
837 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
845 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups));
849 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction, Confidence) &&
850 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize,
851 Prediction, Confidence) &&
852 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) {
853 Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups));
857 if (ValidSimple(TrueBBI, Dups, Prediction, Confidence) &&
858 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize,
859 Prediction, Confidence) &&
860 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) {
861 // Simple (split, no rejoin):
868 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups));
873 // Try the other path...
874 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups,
875 1.0-Prediction, Confidence) &&
876 MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize,
877 1.0-Prediction, Confidence) &&
878 FeasibilityAnalysis(FalseBBI, RevCond, true)) {
879 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups));
883 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups,
884 1.0-Prediction, Confidence) &&
885 MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize,
886 1.0-Prediction, Confidence) &&
887 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) {
888 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups));
892 if (ValidSimple(FalseBBI, Dups, 1.0-Prediction, Confidence) &&
893 MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize,
894 1.0-Prediction, Confidence) &&
895 FeasibilityAnalysis(FalseBBI, RevCond)) {
896 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups));
901 BBI.IsEnqueued = Enqueued;
902 BBI.IsBeingAnalyzed = false;
903 BBI.IsAnalyzed = true;
907 /// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion
909 void IfConverter::AnalyzeBlocks(MachineFunction &MF,
910 std::vector<IfcvtToken*> &Tokens) {
911 std::set<MachineBasicBlock*> Visited;
912 for (unsigned i = 0, e = Roots.size(); i != e; ++i) {
913 for (idf_ext_iterator<MachineBasicBlock*> I=idf_ext_begin(Roots[i],Visited),
914 E = idf_ext_end(Roots[i], Visited); I != E; ++I) {
915 MachineBasicBlock *BB = *I;
916 AnalyzeBlock(BB, Tokens);
920 // Sort to favor more complex ifcvt scheme.
921 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp);
924 /// canFallThroughTo - Returns true either if ToBB is the next block after BB or
925 /// that all the intervening blocks are empty (given BB can fall through to its
927 static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) {
928 MachineFunction::iterator PI = BB;
929 MachineFunction::iterator I = llvm::next(PI);
930 MachineFunction::iterator TI = ToBB;
931 MachineFunction::iterator E = BB->getParent()->end();
933 // Check isSuccessor to avoid case where the next block is empty, but
934 // it's not a successor.
935 if (I == E || !I->empty() || !PI->isSuccessor(I))
942 /// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed
943 /// to determine if it can be if-converted. If predecessor is already enqueued,
945 void IfConverter::InvalidatePreds(MachineBasicBlock *BB) {
946 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(),
947 E = BB->pred_end(); PI != E; ++PI) {
948 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()];
949 if (PBBI.IsDone || PBBI.BB == BB)
951 PBBI.IsAnalyzed = false;
952 PBBI.IsEnqueued = false;
956 /// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB.
958 static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB,
959 const TargetInstrInfo *TII) {
960 DebugLoc dl; // FIXME: this is nowhere
961 SmallVector<MachineOperand, 0> NoCond;
962 TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl);
965 /// RemoveExtraEdges - Remove true / false edges if either / both are no longer
967 void IfConverter::RemoveExtraEdges(BBInfo &BBI) {
968 MachineBasicBlock *TBB = NULL, *FBB = NULL;
969 SmallVector<MachineOperand, 4> Cond;
970 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond))
971 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
974 /// InitPredRedefs / UpdatePredRedefs - Defs by predicated instructions are
975 /// modeled as read + write (sort like two-address instructions). These
976 /// routines track register liveness and add implicit uses to if-converted
977 /// instructions to conform to the model.
978 static void InitPredRedefs(MachineBasicBlock *BB, SmallSet<unsigned,4> &Redefs,
979 const TargetRegisterInfo *TRI) {
980 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(),
981 E = BB->livein_end(); I != E; ++I) {
984 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
986 Redefs.insert(*Subreg);
990 static void UpdatePredRedefs(MachineInstr *MI, SmallSet<unsigned,4> &Redefs,
991 const TargetRegisterInfo *TRI,
992 bool AddImpUse = false) {
993 SmallVector<unsigned, 4> Defs;
994 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
995 const MachineOperand &MO = MI->getOperand(i);
998 unsigned Reg = MO.getReg();
1002 Defs.push_back(Reg);
1003 else if (MO.isKill()) {
1005 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1009 for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1010 unsigned Reg = Defs[i];
1011 if (Redefs.count(Reg)) {
1013 // Treat predicated update as read + write.
1014 MI->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/,
1015 true/*IsImp*/,false/*IsKill*/));
1018 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1024 static void UpdatePredRedefs(MachineBasicBlock::iterator I,
1025 MachineBasicBlock::iterator E,
1026 SmallSet<unsigned,4> &Redefs,
1027 const TargetRegisterInfo *TRI) {
1029 UpdatePredRedefs(I, Redefs, TRI);
1034 /// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG.
1036 bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) {
1037 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
1038 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1039 BBInfo *CvtBBI = &TrueBBI;
1040 BBInfo *NextBBI = &FalseBBI;
1042 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
1043 if (Kind == ICSimpleFalse)
1044 std::swap(CvtBBI, NextBBI);
1046 if (CvtBBI->IsDone ||
1047 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
1048 // Something has changed. It's no longer safe to predicate this block.
1049 BBI.IsAnalyzed = false;
1050 CvtBBI->IsAnalyzed = false;
1054 if (Kind == ICSimpleFalse)
1055 if (TII->ReverseBranchCondition(Cond))
1056 assert(false && "Unable to reverse branch condition!");
1058 // Initialize liveins to the first BB. These are potentiall redefined by
1059 // predicated instructions.
1060 SmallSet<unsigned, 4> Redefs;
1061 InitPredRedefs(CvtBBI->BB, Redefs, TRI);
1062 InitPredRedefs(NextBBI->BB, Redefs, TRI);
1064 if (CvtBBI->BB->pred_size() > 1) {
1065 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1066 // Copy instructions in the true block, predicate them, and add them to
1068 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs);
1070 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs);
1072 // Merge converted block into entry block.
1073 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1074 MergeBlocks(BBI, *CvtBBI);
1077 bool IterIfcvt = true;
1078 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) {
1079 InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
1080 BBI.HasFallThrough = false;
1081 // Now ifcvt'd block will look like this:
1088 // We cannot further ifcvt this block because the unconditional branch
1089 // will have to be predicated on the new condition, that will not be
1090 // available if cmp executes.
1094 RemoveExtraEdges(BBI);
1096 // Update block info. BB can be iteratively if-converted.
1099 InvalidatePreds(BBI.BB);
1100 CvtBBI->IsDone = true;
1102 // FIXME: Must maintain LiveIns.
1106 /// IfConvertTriangle - If convert a triangle sub-CFG.
1108 bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
1109 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
1110 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1111 BBInfo *CvtBBI = &TrueBBI;
1112 BBInfo *NextBBI = &FalseBBI;
1113 DebugLoc dl; // FIXME: this is nowhere
1115 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
1116 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
1117 std::swap(CvtBBI, NextBBI);
1119 if (CvtBBI->IsDone ||
1120 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
1121 // Something has changed. It's no longer safe to predicate this block.
1122 BBI.IsAnalyzed = false;
1123 CvtBBI->IsAnalyzed = false;
1127 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
1128 if (TII->ReverseBranchCondition(Cond))
1129 assert(false && "Unable to reverse branch condition!");
1131 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) {
1132 if (ReverseBranchCondition(*CvtBBI)) {
1133 // BB has been changed, modify its predecessors (except for this
1134 // one) so they don't get ifcvt'ed based on bad intel.
1135 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(),
1136 E = CvtBBI->BB->pred_end(); PI != E; ++PI) {
1137 MachineBasicBlock *PBB = *PI;
1140 BBInfo &PBBI = BBAnalysis[PBB->getNumber()];
1141 if (PBBI.IsEnqueued) {
1142 PBBI.IsAnalyzed = false;
1143 PBBI.IsEnqueued = false;
1149 // Initialize liveins to the first BB. These are potentially redefined by
1150 // predicated instructions.
1151 SmallSet<unsigned, 4> Redefs;
1152 InitPredRedefs(CvtBBI->BB, Redefs, TRI);
1153 InitPredRedefs(NextBBI->BB, Redefs, TRI);
1155 bool HasEarlyExit = CvtBBI->FalseBB != NULL;
1156 if (CvtBBI->BB->pred_size() > 1) {
1157 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1158 // Copy instructions in the true block, predicate them, and add them to
1160 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true);
1162 // Predicate the 'true' block after removing its branch.
1163 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB);
1164 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs);
1166 // Now merge the entry of the triangle with the true block.
1167 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1168 MergeBlocks(BBI, *CvtBBI, false);
1171 // If 'true' block has a 'false' successor, add an exit branch to it.
1173 SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(),
1174 CvtBBI->BrCond.end());
1175 if (TII->ReverseBranchCondition(RevCond))
1176 assert(false && "Unable to reverse branch condition!");
1177 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl);
1178 BBI.BB->addSuccessor(CvtBBI->FalseBB);
1181 // Merge in the 'false' block if the 'false' block has no other
1182 // predecessors. Otherwise, add an unconditional branch to 'false'.
1183 bool FalseBBDead = false;
1184 bool IterIfcvt = true;
1185 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB);
1186 if (!isFallThrough) {
1187 // Only merge them if the true block does not fallthrough to the false
1188 // block. By not merging them, we make it possible to iteratively
1189 // ifcvt the blocks.
1190 if (!HasEarlyExit &&
1191 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) {
1192 MergeBlocks(BBI, *NextBBI);
1195 InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
1196 BBI.HasFallThrough = false;
1198 // Mixed predicated and unpredicated code. This cannot be iteratively
1203 RemoveExtraEdges(BBI);
1205 // Update block info. BB can be iteratively if-converted.
1208 InvalidatePreds(BBI.BB);
1209 CvtBBI->IsDone = true;
1211 NextBBI->IsDone = true;
1213 // FIXME: Must maintain LiveIns.
1217 /// IfConvertDiamond - If convert a diamond sub-CFG.
1219 bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
1220 unsigned NumDups1, unsigned NumDups2) {
1221 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
1222 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1223 MachineBasicBlock *TailBB = TrueBBI.TrueBB;
1224 // True block must fall through or end with an unanalyzable terminator.
1226 if (blockAlwaysFallThrough(TrueBBI))
1227 TailBB = FalseBBI.TrueBB;
1228 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!");
1231 if (TrueBBI.IsDone || FalseBBI.IsDone ||
1232 TrueBBI.BB->pred_size() > 1 ||
1233 FalseBBI.BB->pred_size() > 1) {
1234 // Something has changed. It's no longer safe to predicate these blocks.
1235 BBI.IsAnalyzed = false;
1236 TrueBBI.IsAnalyzed = false;
1237 FalseBBI.IsAnalyzed = false;
1241 // Put the predicated instructions from the 'true' block before the
1242 // instructions from the 'false' block, unless the true block would clobber
1243 // the predicate, in which case, do the opposite.
1244 BBInfo *BBI1 = &TrueBBI;
1245 BBInfo *BBI2 = &FalseBBI;
1246 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end());
1247 if (TII->ReverseBranchCondition(RevCond))
1248 assert(false && "Unable to reverse branch condition!");
1249 SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond;
1250 SmallVector<MachineOperand, 4> *Cond2 = &RevCond;
1252 // Figure out the more profitable ordering.
1253 bool DoSwap = false;
1254 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred)
1256 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) {
1257 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize)
1261 std::swap(BBI1, BBI2);
1262 std::swap(Cond1, Cond2);
1265 // Remove the conditional branch from entry to the blocks.
1266 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1268 // Initialize liveins to the first BB. These are potentially redefined by
1269 // predicated instructions.
1270 SmallSet<unsigned, 4> Redefs;
1271 InitPredRedefs(BBI1->BB, Redefs, TRI);
1273 // Remove the duplicated instructions at the beginnings of both paths.
1274 MachineBasicBlock::iterator DI1 = BBI1->BB->begin();
1275 MachineBasicBlock::iterator DI2 = BBI2->BB->begin();
1276 MachineBasicBlock::iterator DIE1 = BBI1->BB->end();
1277 MachineBasicBlock::iterator DIE2 = BBI2->BB->end();
1278 // Skip dbg_value instructions
1279 while (DI1 != DIE1 && DI1->isDebugValue())
1281 while (DI2 != DIE2 && DI2->isDebugValue())
1283 BBI1->NonPredSize -= NumDups1;
1284 BBI2->NonPredSize -= NumDups1;
1286 // Skip past the dups on each side separately since there may be
1287 // differing dbg_value entries.
1288 for (unsigned i = 0; i < NumDups1; ++DI1) {
1289 if (!DI1->isDebugValue())
1292 while (NumDups1 != 0) {
1294 if (!DI2->isDebugValue())
1298 UpdatePredRedefs(BBI1->BB->begin(), DI1, Redefs, TRI);
1299 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1);
1300 BBI2->BB->erase(BBI2->BB->begin(), DI2);
1302 // Predicate the 'true' block after removing its branch.
1303 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB);
1304 DI1 = BBI1->BB->end();
1305 for (unsigned i = 0; i != NumDups2; ) {
1306 // NumDups2 only counted non-dbg_value instructions, so this won't
1307 // run off the head of the list.
1308 assert (DI1 != BBI1->BB->begin());
1310 // skip dbg_value instructions
1311 if (!DI1->isDebugValue())
1314 BBI1->BB->erase(DI1, BBI1->BB->end());
1315 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs);
1317 // Predicate the 'false' block.
1318 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB);
1319 DI2 = BBI2->BB->end();
1320 while (NumDups2 != 0) {
1321 // NumDups2 only counted non-dbg_value instructions, so this won't
1322 // run off the head of the list.
1323 assert (DI2 != BBI2->BB->begin());
1325 // skip dbg_value instructions
1326 if (!DI2->isDebugValue())
1329 PredicateBlock(*BBI2, DI2, *Cond2, Redefs);
1331 // Merge the true block into the entry of the diamond.
1332 MergeBlocks(BBI, *BBI1, TailBB == 0);
1333 MergeBlocks(BBI, *BBI2, TailBB == 0);
1335 // If the if-converted block falls through or unconditionally branches into
1336 // the tail block, and the tail block does not have other predecessors, then
1337 // fold the tail block in as well. Otherwise, unless it falls through to the
1338 // tail, add a unconditional branch to it.
1340 BBInfo TailBBI = BBAnalysis[TailBB->getNumber()];
1341 bool CanMergeTail = !TailBBI.HasFallThrough;
1342 // There may still be a fall-through edge from BBI1 or BBI2 to TailBB;
1343 // check if there are any other predecessors besides those.
1344 unsigned NumPreds = TailBB->pred_size();
1346 CanMergeTail = false;
1347 else if (NumPreds == 1 && CanMergeTail) {
1348 MachineBasicBlock::pred_iterator PI = TailBB->pred_begin();
1349 if (*PI != BBI1->BB && *PI != BBI2->BB)
1350 CanMergeTail = false;
1353 MergeBlocks(BBI, TailBBI);
1354 TailBBI.IsDone = true;
1356 BBI.BB->addSuccessor(TailBB);
1357 InsertUncondBranch(BBI.BB, TailBB, TII);
1358 BBI.HasFallThrough = false;
1362 // RemoveExtraEdges won't work if the block has an unanalyzable branch,
1363 // which can happen here if TailBB is unanalyzable and is merged, so
1364 // explicitly remove BBI1 and BBI2 as successors.
1365 BBI.BB->removeSuccessor(BBI1->BB);
1366 BBI.BB->removeSuccessor(BBI2->BB);
1367 RemoveExtraEdges(BBI);
1369 // Update block info.
1370 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true;
1371 InvalidatePreds(BBI.BB);
1373 // FIXME: Must maintain LiveIns.
1377 /// PredicateBlock - Predicate instructions from the start of the block to the
1378 /// specified end with the specified condition.
1379 void IfConverter::PredicateBlock(BBInfo &BBI,
1380 MachineBasicBlock::iterator E,
1381 SmallVectorImpl<MachineOperand> &Cond,
1382 SmallSet<unsigned, 4> &Redefs) {
1383 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) {
1384 if (I->isDebugValue() || TII->isPredicated(I))
1386 if (!TII->PredicateInstruction(I, Cond)) {
1388 dbgs() << "Unable to predicate " << *I << "!\n";
1390 llvm_unreachable(0);
1393 // If the predicated instruction now redefines a register as the result of
1394 // if-conversion, add an implicit kill.
1395 UpdatePredRedefs(I, Redefs, TRI, true);
1398 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate));
1400 BBI.IsAnalyzed = false;
1401 BBI.NonPredSize = 0;
1406 /// CopyAndPredicateBlock - Copy and predicate instructions from source BB to
1407 /// the destination block. Skip end of block branches if IgnoreBr is true.
1408 void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
1409 SmallVectorImpl<MachineOperand> &Cond,
1410 SmallSet<unsigned, 4> &Redefs,
1412 MachineFunction &MF = *ToBBI.BB->getParent();
1414 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(),
1415 E = FromBBI.BB->end(); I != E; ++I) {
1416 const TargetInstrDesc &TID = I->getDesc();
1417 // Do not copy the end of the block branches.
1418 if (IgnoreBr && TID.isBranch())
1421 MachineInstr *MI = MF.CloneMachineInstr(I);
1422 ToBBI.BB->insert(ToBBI.BB->end(), MI);
1423 unsigned NumOps = TII->getNumMicroOps(MI, InstrItins);
1424 ToBBI.NonPredSize += NumOps;
1426 if (!TII->isPredicated(I) && !MI->isDebugValue()) {
1427 if (!TII->PredicateInstruction(MI, Cond)) {
1429 dbgs() << "Unable to predicate " << *I << "!\n";
1431 llvm_unreachable(0);
1435 // If the predicated instruction now redefines a register as the result of
1436 // if-conversion, add an implicit kill.
1437 UpdatePredRedefs(MI, Redefs, TRI, true);
1441 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1442 FromBBI.BB->succ_end());
1443 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1444 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1446 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1447 MachineBasicBlock *Succ = Succs[i];
1448 // Fallthrough edge can't be transferred.
1449 if (Succ == FallThrough)
1451 ToBBI.BB->addSuccessor(Succ);
1455 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1456 std::back_inserter(ToBBI.Predicate));
1457 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate));
1459 ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1460 ToBBI.IsAnalyzed = false;
1465 /// MergeBlocks - Move all instructions from FromBB to the end of ToBB.
1466 /// This will leave FromBB as an empty block, so remove all of its
1467 /// successor edges except for the fall-through edge. If AddEdges is true,
1468 /// i.e., when FromBBI's branch is being moved, add those successor edges to
1470 void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) {
1471 ToBBI.BB->splice(ToBBI.BB->end(),
1472 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end());
1474 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1475 FromBBI.BB->succ_end());
1476 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1477 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1479 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1480 MachineBasicBlock *Succ = Succs[i];
1481 // Fallthrough edge can't be transferred.
1482 if (Succ == FallThrough)
1484 FromBBI.BB->removeSuccessor(Succ);
1486 ToBBI.BB->addSuccessor(Succ);
1489 // Now FromBBI always falls through to the next block!
1490 if (NBB && !FromBBI.BB->isSuccessor(NBB))
1491 FromBBI.BB->addSuccessor(NBB);
1493 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1494 std::back_inserter(ToBBI.Predicate));
1495 FromBBI.Predicate.clear();
1497 ToBBI.NonPredSize += FromBBI.NonPredSize;
1498 FromBBI.NonPredSize = 0;
1500 ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1501 ToBBI.HasFallThrough = FromBBI.HasFallThrough;
1502 ToBBI.IsAnalyzed = false;
1503 FromBBI.IsAnalyzed = false;