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/SmallSet.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 /// ExtraCost - Extra cost for multi-cycle instructions.
97 /// ExtraCost2 - Some instructions are slower when predicated
98 /// BB - Corresponding MachineBasicBlock.
99 /// TrueBB / FalseBB- See AnalyzeBranch().
100 /// BrCond - Conditions for end of block conditional branches.
101 /// Predicate - Predicate used in the BB.
104 bool IsBeingAnalyzed : 1;
107 bool IsBrAnalyzable : 1;
108 bool HasFallThrough : 1;
109 bool IsUnpredicable : 1;
110 bool CannotBeCopied : 1;
111 bool ClobbersPred : 1;
112 unsigned NonPredSize;
115 MachineBasicBlock *BB;
116 MachineBasicBlock *TrueBB;
117 MachineBasicBlock *FalseBB;
118 SmallVector<MachineOperand, 4> BrCond;
119 SmallVector<MachineOperand, 4> Predicate;
120 BBInfo() : IsDone(false), IsBeingAnalyzed(false),
121 IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false),
122 HasFallThrough(false), IsUnpredicable(false),
123 CannotBeCopied(false), ClobbersPred(false), NonPredSize(0),
124 ExtraCost(0), ExtraCost2(0), BB(0), TrueBB(0), FalseBB(0) {}
127 /// IfcvtToken - Record information about pending if-conversions to attempt:
128 /// BBI - Corresponding BBInfo.
129 /// Kind - Type of block. See IfcvtKind.
130 /// NeedSubsumption - True if the to-be-predicated BB has already been
132 /// NumDups - Number of instructions that would be duplicated due
133 /// to this if-conversion. (For diamonds, the number of
134 /// identical instructions at the beginnings of both
136 /// NumDups2 - For diamonds, the number of identical instructions
137 /// at the ends of both paths.
141 bool NeedSubsumption;
144 IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0)
145 : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {}
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) {
162 initializeIfConverterPass(*PassRegistry::getPassRegistry());
165 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
166 AU.addRequired<MachineLoopInfo>();
167 MachineFunctionPass::getAnalysisUsage(AU);
170 virtual bool runOnMachineFunction(MachineFunction &MF);
171 virtual const char *getPassName() const { return "If Converter"; }
174 bool ReverseBranchCondition(BBInfo &BBI);
175 bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups,
176 float Prediction, float Confidence) const;
177 bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
178 bool FalseBranch, unsigned &Dups,
179 float Prediction, float Confidence) const;
180 bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
181 unsigned &Dups1, unsigned &Dups2) const;
182 void ScanInstructions(BBInfo &BBI);
183 BBInfo &AnalyzeBlock(MachineBasicBlock *BB,
184 std::vector<IfcvtToken*> &Tokens);
185 bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl<MachineOperand> &Cond,
186 bool isTriangle = false, bool RevBranch = false);
187 void AnalyzeBlocks(MachineFunction &MF, std::vector<IfcvtToken*> &Tokens);
188 void InvalidatePreds(MachineBasicBlock *BB);
189 void RemoveExtraEdges(BBInfo &BBI);
190 bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind);
191 bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind);
192 bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
193 unsigned NumDups1, unsigned NumDups2);
194 void PredicateBlock(BBInfo &BBI,
195 MachineBasicBlock::iterator E,
196 SmallVectorImpl<MachineOperand> &Cond,
197 SmallSet<unsigned, 4> &Redefs);
198 void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
199 SmallVectorImpl<MachineOperand> &Cond,
200 SmallSet<unsigned, 4> &Redefs,
201 bool IgnoreBr = false);
202 void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true);
204 bool MeetIfcvtSizeLimit(MachineBasicBlock &BB,
205 unsigned Cycle, unsigned Extra,
206 float Prediction, float Confidence) const {
207 return Cycle > 0 && TII->isProfitableToIfCvt(BB, Cycle, Extra,
208 Prediction, Confidence);
211 bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB,
212 unsigned TCycle, unsigned TExtra,
213 MachineBasicBlock &FBB,
214 unsigned FCycle, unsigned FExtra,
215 float Prediction, float Confidence) const {
216 return TCycle > 0 && FCycle > 0 &&
217 TII->isProfitableToIfCvt(TBB, TCycle, TExtra, FBB, FCycle, FExtra,
218 Prediction, Confidence);
221 // blockAlwaysFallThrough - Block ends without a terminator.
222 bool blockAlwaysFallThrough(BBInfo &BBI) const {
223 return BBI.IsBrAnalyzable && BBI.TrueBB == NULL;
226 // IfcvtTokenCmp - Used to sort if-conversion candidates.
227 static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) {
228 int Incr1 = (C1->Kind == ICDiamond)
229 ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups;
230 int Incr2 = (C2->Kind == ICDiamond)
231 ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups;
234 else if (Incr1 == Incr2) {
235 // Favors subsumption.
236 if (C1->NeedSubsumption == false && C2->NeedSubsumption == true)
238 else if (C1->NeedSubsumption == C2->NeedSubsumption) {
239 // Favors diamond over triangle, etc.
240 if ((unsigned)C1->Kind < (unsigned)C2->Kind)
242 else if (C1->Kind == C2->Kind)
243 return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber();
250 char IfConverter::ID = 0;
253 INITIALIZE_PASS_BEGIN(IfConverter, "if-converter", "If Converter", false, false)
254 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
255 INITIALIZE_PASS_END(IfConverter, "if-converter", "If Converter", false, false)
257 FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); }
259 bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
260 TLI = MF.getTarget().getTargetLowering();
261 TII = MF.getTarget().getInstrInfo();
262 TRI = MF.getTarget().getRegisterInfo();
263 MLI = &getAnalysis<MachineLoopInfo>();
264 InstrItins = MF.getTarget().getInstrItineraryData();
265 if (!TII) return false;
267 // Tail merge tend to expose more if-conversion opportunities.
268 BranchFolder BF(true, false);
269 bool BFChange = BF.OptimizeFunction(MF, TII,
270 MF.getTarget().getRegisterInfo(),
271 getAnalysisIfAvailable<MachineModuleInfo>());
273 DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'"
274 << MF.getFunction()->getName() << "\'");
276 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) {
277 DEBUG(dbgs() << " skipped\n");
280 DEBUG(dbgs() << "\n");
283 BBAnalysis.resize(MF.getNumBlockIDs());
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();
401 if (MadeChange && IfCvtBranchFold) {
402 BranchFolder BF(false, false);
403 BF.OptimizeFunction(MF, TII,
404 MF.getTarget().getRegisterInfo(),
405 getAnalysisIfAvailable<MachineModuleInfo>());
408 MadeChange |= BFChange;
412 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
413 /// its 'true' successor.
414 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
415 MachineBasicBlock *TrueBB) {
416 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
417 E = BB->succ_end(); SI != E; ++SI) {
418 MachineBasicBlock *SuccBB = *SI;
419 if (SuccBB != TrueBB)
425 /// ReverseBranchCondition - Reverse the condition of the end of the block
426 /// branch. Swap block's 'true' and 'false' successors.
427 bool IfConverter::ReverseBranchCondition(BBInfo &BBI) {
428 DebugLoc dl; // FIXME: this is nowhere
429 if (!TII->ReverseBranchCondition(BBI.BrCond)) {
430 TII->RemoveBranch(*BBI.BB);
431 TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond, dl);
432 std::swap(BBI.TrueBB, BBI.FalseBB);
438 /// getNextBlock - Returns the next block in the function blocks ordering. If
439 /// it is the end, returns NULL.
440 static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) {
441 MachineFunction::iterator I = BB;
442 MachineFunction::iterator E = BB->getParent()->end();
448 /// ValidSimple - Returns true if the 'true' block (along with its
449 /// predecessor) forms a valid simple shape for ifcvt. It also returns the
450 /// number of instructions that the ifcvt would need to duplicate if performed
452 bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups,
453 float Prediction, float Confidence) const {
455 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
458 if (TrueBBI.IsBrAnalyzable)
461 if (TrueBBI.BB->pred_size() > 1) {
462 if (TrueBBI.CannotBeCopied ||
463 !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize,
464 Prediction, Confidence))
466 Dups = TrueBBI.NonPredSize;
472 /// ValidTriangle - Returns true if the 'true' and 'false' blocks (along
473 /// with their common predecessor) forms a valid triangle shape for ifcvt.
474 /// If 'FalseBranch' is true, it checks if 'true' block's false branch
475 /// branches to the 'false' block rather than the other way around. It also
476 /// returns the number of instructions that the ifcvt would need to duplicate
477 /// if performed in 'Dups'.
478 bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
479 bool FalseBranch, unsigned &Dups,
480 float Prediction, float Confidence) const {
482 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
485 if (TrueBBI.BB->pred_size() > 1) {
486 if (TrueBBI.CannotBeCopied)
489 unsigned Size = TrueBBI.NonPredSize;
490 if (TrueBBI.IsBrAnalyzable) {
491 if (TrueBBI.TrueBB && TrueBBI.BrCond.empty())
492 // Ends with an unconditional branch. It will be removed.
495 MachineBasicBlock *FExit = FalseBranch
496 ? TrueBBI.TrueBB : TrueBBI.FalseBB;
498 // Require a conditional branch
502 if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size,
503 Prediction, Confidence))
508 MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB;
509 if (!TExit && blockAlwaysFallThrough(TrueBBI)) {
510 MachineFunction::iterator I = TrueBBI.BB;
511 if (++I == TrueBBI.BB->getParent()->end())
515 return TExit && TExit == FalseBBI.BB;
518 /// ValidDiamond - Returns true if the 'true' and 'false' blocks (along
519 /// with their common predecessor) forms a valid diamond shape for ifcvt.
520 bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
521 unsigned &Dups1, unsigned &Dups2) const {
523 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone ||
524 FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone)
527 MachineBasicBlock *TT = TrueBBI.TrueBB;
528 MachineBasicBlock *FT = FalseBBI.TrueBB;
530 if (!TT && blockAlwaysFallThrough(TrueBBI))
531 TT = getNextBlock(TrueBBI.BB);
532 if (!FT && blockAlwaysFallThrough(FalseBBI))
533 FT = getNextBlock(FalseBBI.BB);
536 if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable))
538 if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1)
541 // FIXME: Allow true block to have an early exit?
542 if (TrueBBI.FalseBB || FalseBBI.FalseBB ||
543 (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred))
546 // Count duplicate instructions at the beginning of the true and false blocks.
547 MachineBasicBlock::iterator TIB = TrueBBI.BB->begin();
548 MachineBasicBlock::iterator FIB = FalseBBI.BB->begin();
549 MachineBasicBlock::iterator TIE = TrueBBI.BB->end();
550 MachineBasicBlock::iterator FIE = FalseBBI.BB->end();
551 while (TIB != TIE && FIB != FIE) {
552 // Skip dbg_value instructions. These do not count.
553 if (TIB->isDebugValue()) {
554 while (TIB != TIE && TIB->isDebugValue())
559 if (FIB->isDebugValue()) {
560 while (FIB != FIE && FIB->isDebugValue())
565 if (!TIB->isIdenticalTo(FIB))
572 // Now, in preparation for counting duplicate instructions at the ends of the
573 // blocks, move the end iterators up past any branch instructions.
576 if (!TIE->getDesc().isBranch())
581 if (!FIE->getDesc().isBranch())
585 // If Dups1 includes all of a block, then don't count duplicate
586 // instructions at the end of the blocks.
587 if (TIB == TIE || FIB == FIE)
590 // Count duplicate instructions at the ends of the blocks.
591 while (TIE != TIB && FIE != FIB) {
592 // Skip dbg_value instructions. These do not count.
593 if (TIE->isDebugValue()) {
594 while (TIE != TIB && TIE->isDebugValue())
599 if (FIE->isDebugValue()) {
600 while (FIE != FIB && FIE->isDebugValue())
605 if (!TIE->isIdenticalTo(FIE))
615 /// ScanInstructions - Scan all the instructions in the block to determine if
616 /// the block is predicable. In most cases, that means all the instructions
617 /// in the block are isPredicable(). Also checks if the block contains any
618 /// instruction which can clobber a predicate (e.g. condition code register).
619 /// If so, the block is not predicable unless it's the last instruction.
620 void IfConverter::ScanInstructions(BBInfo &BBI) {
624 bool AlreadyPredicated = BBI.Predicate.size() > 0;
625 // First analyze the end of BB branches.
626 BBI.TrueBB = BBI.FalseBB = NULL;
629 !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond);
630 BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL;
632 if (BBI.BrCond.size()) {
633 // No false branch. This BB must end with a conditional branch and a
636 BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB);
638 // Malformed bcc? True and false blocks are the same?
639 BBI.IsUnpredicable = true;
644 // Then scan all the instructions.
648 BBI.ClobbersPred = false;
649 for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end();
651 if (I->isDebugValue())
654 const TargetInstrDesc &TID = I->getDesc();
655 if (TID.isNotDuplicable())
656 BBI.CannotBeCopied = true;
658 bool isPredicated = TII->isPredicated(I);
659 bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch();
664 unsigned ExtraPredCost = 0;
665 unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I,
668 BBI.ExtraCost += NumCycles-1;
669 BBI.ExtraCost2 += ExtraPredCost;
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 TrueBBI.ExtraCost), TrueBBI.ExtraCost2,
818 *FalseBBI.BB, (FalseBBI.NonPredSize - (Dups + Dups2) +
819 FalseBBI.ExtraCost),FalseBBI.ExtraCost2,
820 Prediction, Confidence) &&
821 FeasibilityAnalysis(TrueBBI, BBI.BrCond) &&
822 FeasibilityAnalysis(FalseBBI, RevCond)) {
830 // Note TailBB can be empty.
831 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups,
836 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction, Confidence) &&
837 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
838 TrueBBI.ExtraCost2, Prediction, Confidence) &&
839 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
847 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups));
851 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction, Confidence) &&
852 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
853 TrueBBI.ExtraCost2, Prediction, Confidence) &&
854 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) {
855 Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups));
859 if (ValidSimple(TrueBBI, Dups, Prediction, Confidence) &&
860 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
861 TrueBBI.ExtraCost2, Prediction, Confidence) &&
862 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) {
863 // Simple (split, no rejoin):
870 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups));
875 // Try the other path...
876 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups,
877 1.0-Prediction, Confidence) &&
878 MeetIfcvtSizeLimit(*FalseBBI.BB,
879 FalseBBI.NonPredSize + FalseBBI.ExtraCost,
880 FalseBBI.ExtraCost2, 1.0-Prediction, Confidence) &&
881 FeasibilityAnalysis(FalseBBI, RevCond, true)) {
882 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups));
886 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups,
887 1.0-Prediction, Confidence) &&
888 MeetIfcvtSizeLimit(*FalseBBI.BB,
889 FalseBBI.NonPredSize + FalseBBI.ExtraCost,
890 FalseBBI.ExtraCost2, 1.0-Prediction, Confidence) &&
891 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) {
892 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups));
896 if (ValidSimple(FalseBBI, Dups, 1.0-Prediction, Confidence) &&
897 MeetIfcvtSizeLimit(*FalseBBI.BB,
898 FalseBBI.NonPredSize + FalseBBI.ExtraCost,
899 FalseBBI.ExtraCost2, 1.0-Prediction, Confidence) &&
900 FeasibilityAnalysis(FalseBBI, RevCond)) {
901 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups));
906 BBI.IsEnqueued = Enqueued;
907 BBI.IsBeingAnalyzed = false;
908 BBI.IsAnalyzed = true;
912 /// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion
914 void IfConverter::AnalyzeBlocks(MachineFunction &MF,
915 std::vector<IfcvtToken*> &Tokens) {
916 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
917 MachineBasicBlock *BB = I;
918 AnalyzeBlock(BB, Tokens);
921 // Sort to favor more complex ifcvt scheme.
922 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp);
925 /// canFallThroughTo - Returns true either if ToBB is the next block after BB or
926 /// that all the intervening blocks are empty (given BB can fall through to its
928 static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) {
929 MachineFunction::iterator PI = BB;
930 MachineFunction::iterator I = llvm::next(PI);
931 MachineFunction::iterator TI = ToBB;
932 MachineFunction::iterator E = BB->getParent()->end();
934 // Check isSuccessor to avoid case where the next block is empty, but
935 // it's not a successor.
936 if (I == E || !I->empty() || !PI->isSuccessor(I))
943 /// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed
944 /// to determine if it can be if-converted. If predecessor is already enqueued,
946 void IfConverter::InvalidatePreds(MachineBasicBlock *BB) {
947 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(),
948 E = BB->pred_end(); PI != E; ++PI) {
949 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()];
950 if (PBBI.IsDone || PBBI.BB == BB)
952 PBBI.IsAnalyzed = false;
953 PBBI.IsEnqueued = false;
957 /// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB.
959 static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB,
960 const TargetInstrInfo *TII) {
961 DebugLoc dl; // FIXME: this is nowhere
962 SmallVector<MachineOperand, 0> NoCond;
963 TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl);
966 /// RemoveExtraEdges - Remove true / false edges if either / both are no longer
968 void IfConverter::RemoveExtraEdges(BBInfo &BBI) {
969 MachineBasicBlock *TBB = NULL, *FBB = NULL;
970 SmallVector<MachineOperand, 4> Cond;
971 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond))
972 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
975 /// InitPredRedefs / UpdatePredRedefs - Defs by predicated instructions are
976 /// modeled as read + write (sort like two-address instructions). These
977 /// routines track register liveness and add implicit uses to if-converted
978 /// instructions to conform to the model.
979 static void InitPredRedefs(MachineBasicBlock *BB, SmallSet<unsigned,4> &Redefs,
980 const TargetRegisterInfo *TRI) {
981 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(),
982 E = BB->livein_end(); I != E; ++I) {
985 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
987 Redefs.insert(*Subreg);
991 static void UpdatePredRedefs(MachineInstr *MI, SmallSet<unsigned,4> &Redefs,
992 const TargetRegisterInfo *TRI,
993 bool AddImpUse = false) {
994 SmallVector<unsigned, 4> Defs;
995 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
996 const MachineOperand &MO = MI->getOperand(i);
999 unsigned Reg = MO.getReg();
1003 Defs.push_back(Reg);
1004 else if (MO.isKill()) {
1006 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1010 for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1011 unsigned Reg = Defs[i];
1012 if (Redefs.count(Reg)) {
1014 // Treat predicated update as read + write.
1015 MI->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/,
1016 true/*IsImp*/,false/*IsKill*/));
1019 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1025 static void UpdatePredRedefs(MachineBasicBlock::iterator I,
1026 MachineBasicBlock::iterator E,
1027 SmallSet<unsigned,4> &Redefs,
1028 const TargetRegisterInfo *TRI) {
1030 UpdatePredRedefs(I, Redefs, TRI);
1035 /// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG.
1037 bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) {
1038 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
1039 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1040 BBInfo *CvtBBI = &TrueBBI;
1041 BBInfo *NextBBI = &FalseBBI;
1043 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
1044 if (Kind == ICSimpleFalse)
1045 std::swap(CvtBBI, NextBBI);
1047 if (CvtBBI->IsDone ||
1048 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
1049 // Something has changed. It's no longer safe to predicate this block.
1050 BBI.IsAnalyzed = false;
1051 CvtBBI->IsAnalyzed = false;
1055 if (Kind == ICSimpleFalse)
1056 if (TII->ReverseBranchCondition(Cond))
1057 assert(false && "Unable to reverse branch condition!");
1059 // Initialize liveins to the first BB. These are potentiall redefined by
1060 // predicated instructions.
1061 SmallSet<unsigned, 4> Redefs;
1062 InitPredRedefs(CvtBBI->BB, Redefs, TRI);
1063 InitPredRedefs(NextBBI->BB, Redefs, TRI);
1065 if (CvtBBI->BB->pred_size() > 1) {
1066 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1067 // Copy instructions in the true block, predicate them, and add them to
1069 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs);
1071 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs);
1073 // Merge converted block into entry block.
1074 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1075 MergeBlocks(BBI, *CvtBBI);
1078 bool IterIfcvt = true;
1079 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) {
1080 InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
1081 BBI.HasFallThrough = false;
1082 // Now ifcvt'd block will look like this:
1089 // We cannot further ifcvt this block because the unconditional branch
1090 // will have to be predicated on the new condition, that will not be
1091 // available if cmp executes.
1095 RemoveExtraEdges(BBI);
1097 // Update block info. BB can be iteratively if-converted.
1100 InvalidatePreds(BBI.BB);
1101 CvtBBI->IsDone = true;
1103 // FIXME: Must maintain LiveIns.
1107 /// IfConvertTriangle - If convert a triangle sub-CFG.
1109 bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
1110 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
1111 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1112 BBInfo *CvtBBI = &TrueBBI;
1113 BBInfo *NextBBI = &FalseBBI;
1114 DebugLoc dl; // FIXME: this is nowhere
1116 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
1117 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
1118 std::swap(CvtBBI, NextBBI);
1120 if (CvtBBI->IsDone ||
1121 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
1122 // Something has changed. It's no longer safe to predicate this block.
1123 BBI.IsAnalyzed = false;
1124 CvtBBI->IsAnalyzed = false;
1128 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
1129 if (TII->ReverseBranchCondition(Cond))
1130 assert(false && "Unable to reverse branch condition!");
1132 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) {
1133 if (ReverseBranchCondition(*CvtBBI)) {
1134 // BB has been changed, modify its predecessors (except for this
1135 // one) so they don't get ifcvt'ed based on bad intel.
1136 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(),
1137 E = CvtBBI->BB->pred_end(); PI != E; ++PI) {
1138 MachineBasicBlock *PBB = *PI;
1141 BBInfo &PBBI = BBAnalysis[PBB->getNumber()];
1142 if (PBBI.IsEnqueued) {
1143 PBBI.IsAnalyzed = false;
1144 PBBI.IsEnqueued = false;
1150 // Initialize liveins to the first BB. These are potentially redefined by
1151 // predicated instructions.
1152 SmallSet<unsigned, 4> Redefs;
1153 InitPredRedefs(CvtBBI->BB, Redefs, TRI);
1154 InitPredRedefs(NextBBI->BB, Redefs, TRI);
1156 bool HasEarlyExit = CvtBBI->FalseBB != NULL;
1157 if (CvtBBI->BB->pred_size() > 1) {
1158 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1159 // Copy instructions in the true block, predicate them, and add them to
1161 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true);
1163 // Predicate the 'true' block after removing its branch.
1164 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB);
1165 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs);
1167 // Now merge the entry of the triangle with the true block.
1168 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1169 MergeBlocks(BBI, *CvtBBI, false);
1172 // If 'true' block has a 'false' successor, add an exit branch to it.
1174 SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(),
1175 CvtBBI->BrCond.end());
1176 if (TII->ReverseBranchCondition(RevCond))
1177 assert(false && "Unable to reverse branch condition!");
1178 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl);
1179 BBI.BB->addSuccessor(CvtBBI->FalseBB);
1182 // Merge in the 'false' block if the 'false' block has no other
1183 // predecessors. Otherwise, add an unconditional branch to 'false'.
1184 bool FalseBBDead = false;
1185 bool IterIfcvt = true;
1186 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB);
1187 if (!isFallThrough) {
1188 // Only merge them if the true block does not fallthrough to the false
1189 // block. By not merging them, we make it possible to iteratively
1190 // ifcvt the blocks.
1191 if (!HasEarlyExit &&
1192 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) {
1193 MergeBlocks(BBI, *NextBBI);
1196 InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
1197 BBI.HasFallThrough = false;
1199 // Mixed predicated and unpredicated code. This cannot be iteratively
1204 RemoveExtraEdges(BBI);
1206 // Update block info. BB can be iteratively if-converted.
1209 InvalidatePreds(BBI.BB);
1210 CvtBBI->IsDone = true;
1212 NextBBI->IsDone = true;
1214 // FIXME: Must maintain LiveIns.
1218 /// IfConvertDiamond - If convert a diamond sub-CFG.
1220 bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
1221 unsigned NumDups1, unsigned NumDups2) {
1222 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
1223 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1224 MachineBasicBlock *TailBB = TrueBBI.TrueBB;
1225 // True block must fall through or end with an unanalyzable terminator.
1227 if (blockAlwaysFallThrough(TrueBBI))
1228 TailBB = FalseBBI.TrueBB;
1229 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!");
1232 if (TrueBBI.IsDone || FalseBBI.IsDone ||
1233 TrueBBI.BB->pred_size() > 1 ||
1234 FalseBBI.BB->pred_size() > 1) {
1235 // Something has changed. It's no longer safe to predicate these blocks.
1236 BBI.IsAnalyzed = false;
1237 TrueBBI.IsAnalyzed = false;
1238 FalseBBI.IsAnalyzed = false;
1242 // Put the predicated instructions from the 'true' block before the
1243 // instructions from the 'false' block, unless the true block would clobber
1244 // the predicate, in which case, do the opposite.
1245 BBInfo *BBI1 = &TrueBBI;
1246 BBInfo *BBI2 = &FalseBBI;
1247 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end());
1248 if (TII->ReverseBranchCondition(RevCond))
1249 assert(false && "Unable to reverse branch condition!");
1250 SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond;
1251 SmallVector<MachineOperand, 4> *Cond2 = &RevCond;
1253 // Figure out the more profitable ordering.
1254 bool DoSwap = false;
1255 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred)
1257 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) {
1258 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize)
1262 std::swap(BBI1, BBI2);
1263 std::swap(Cond1, Cond2);
1266 // Remove the conditional branch from entry to the blocks.
1267 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1269 // Initialize liveins to the first BB. These are potentially redefined by
1270 // predicated instructions.
1271 SmallSet<unsigned, 4> Redefs;
1272 InitPredRedefs(BBI1->BB, Redefs, TRI);
1274 // Remove the duplicated instructions at the beginnings of both paths.
1275 MachineBasicBlock::iterator DI1 = BBI1->BB->begin();
1276 MachineBasicBlock::iterator DI2 = BBI2->BB->begin();
1277 MachineBasicBlock::iterator DIE1 = BBI1->BB->end();
1278 MachineBasicBlock::iterator DIE2 = BBI2->BB->end();
1279 // Skip dbg_value instructions
1280 while (DI1 != DIE1 && DI1->isDebugValue())
1282 while (DI2 != DIE2 && DI2->isDebugValue())
1284 BBI1->NonPredSize -= NumDups1;
1285 BBI2->NonPredSize -= NumDups1;
1287 // Skip past the dups on each side separately since there may be
1288 // differing dbg_value entries.
1289 for (unsigned i = 0; i < NumDups1; ++DI1) {
1290 if (!DI1->isDebugValue())
1293 while (NumDups1 != 0) {
1295 if (!DI2->isDebugValue())
1299 UpdatePredRedefs(BBI1->BB->begin(), DI1, Redefs, TRI);
1300 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1);
1301 BBI2->BB->erase(BBI2->BB->begin(), DI2);
1303 // Predicate the 'true' block after removing its branch.
1304 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB);
1305 DI1 = BBI1->BB->end();
1306 for (unsigned i = 0; i != NumDups2; ) {
1307 // NumDups2 only counted non-dbg_value instructions, so this won't
1308 // run off the head of the list.
1309 assert (DI1 != BBI1->BB->begin());
1311 // skip dbg_value instructions
1312 if (!DI1->isDebugValue())
1315 BBI1->BB->erase(DI1, BBI1->BB->end());
1316 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs);
1318 // Predicate the 'false' block.
1319 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB);
1320 DI2 = BBI2->BB->end();
1321 while (NumDups2 != 0) {
1322 // NumDups2 only counted non-dbg_value instructions, so this won't
1323 // run off the head of the list.
1324 assert (DI2 != BBI2->BB->begin());
1326 // skip dbg_value instructions
1327 if (!DI2->isDebugValue())
1330 PredicateBlock(*BBI2, DI2, *Cond2, Redefs);
1332 // Merge the true block into the entry of the diamond.
1333 MergeBlocks(BBI, *BBI1, TailBB == 0);
1334 MergeBlocks(BBI, *BBI2, TailBB == 0);
1336 // If the if-converted block falls through or unconditionally branches into
1337 // the tail block, and the tail block does not have other predecessors, then
1338 // fold the tail block in as well. Otherwise, unless it falls through to the
1339 // tail, add a unconditional branch to it.
1341 BBInfo TailBBI = BBAnalysis[TailBB->getNumber()];
1342 bool CanMergeTail = !TailBBI.HasFallThrough;
1343 // There may still be a fall-through edge from BBI1 or BBI2 to TailBB;
1344 // check if there are any other predecessors besides those.
1345 unsigned NumPreds = TailBB->pred_size();
1347 CanMergeTail = false;
1348 else if (NumPreds == 1 && CanMergeTail) {
1349 MachineBasicBlock::pred_iterator PI = TailBB->pred_begin();
1350 if (*PI != BBI1->BB && *PI != BBI2->BB)
1351 CanMergeTail = false;
1354 MergeBlocks(BBI, TailBBI);
1355 TailBBI.IsDone = true;
1357 BBI.BB->addSuccessor(TailBB);
1358 InsertUncondBranch(BBI.BB, TailBB, TII);
1359 BBI.HasFallThrough = false;
1363 // RemoveExtraEdges won't work if the block has an unanalyzable branch,
1364 // which can happen here if TailBB is unanalyzable and is merged, so
1365 // explicitly remove BBI1 and BBI2 as successors.
1366 BBI.BB->removeSuccessor(BBI1->BB);
1367 BBI.BB->removeSuccessor(BBI2->BB);
1368 RemoveExtraEdges(BBI);
1370 // Update block info.
1371 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true;
1372 InvalidatePreds(BBI.BB);
1374 // FIXME: Must maintain LiveIns.
1378 /// PredicateBlock - Predicate instructions from the start of the block to the
1379 /// specified end with the specified condition.
1380 void IfConverter::PredicateBlock(BBInfo &BBI,
1381 MachineBasicBlock::iterator E,
1382 SmallVectorImpl<MachineOperand> &Cond,
1383 SmallSet<unsigned, 4> &Redefs) {
1384 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) {
1385 if (I->isDebugValue() || TII->isPredicated(I))
1387 if (!TII->PredicateInstruction(I, Cond)) {
1389 dbgs() << "Unable to predicate " << *I << "!\n";
1391 llvm_unreachable(0);
1394 // If the predicated instruction now redefines a register as the result of
1395 // if-conversion, add an implicit kill.
1396 UpdatePredRedefs(I, Redefs, TRI, true);
1399 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate));
1401 BBI.IsAnalyzed = false;
1402 BBI.NonPredSize = 0;
1407 /// CopyAndPredicateBlock - Copy and predicate instructions from source BB to
1408 /// the destination block. Skip end of block branches if IgnoreBr is true.
1409 void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
1410 SmallVectorImpl<MachineOperand> &Cond,
1411 SmallSet<unsigned, 4> &Redefs,
1413 MachineFunction &MF = *ToBBI.BB->getParent();
1415 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(),
1416 E = FromBBI.BB->end(); I != E; ++I) {
1417 const TargetInstrDesc &TID = I->getDesc();
1418 // Do not copy the end of the block branches.
1419 if (IgnoreBr && TID.isBranch())
1422 MachineInstr *MI = MF.CloneMachineInstr(I);
1423 ToBBI.BB->insert(ToBBI.BB->end(), MI);
1424 ToBBI.NonPredSize++;
1425 unsigned ExtraPredCost = 0;
1426 unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, &ExtraPredCost);
1428 ToBBI.ExtraCost += NumCycles-1;
1429 ToBBI.ExtraCost2 += ExtraPredCost;
1431 if (!TII->isPredicated(I) && !MI->isDebugValue()) {
1432 if (!TII->PredicateInstruction(MI, Cond)) {
1434 dbgs() << "Unable to predicate " << *I << "!\n";
1436 llvm_unreachable(0);
1440 // If the predicated instruction now redefines a register as the result of
1441 // if-conversion, add an implicit kill.
1442 UpdatePredRedefs(MI, Redefs, TRI, true);
1446 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1447 FromBBI.BB->succ_end());
1448 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1449 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1451 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1452 MachineBasicBlock *Succ = Succs[i];
1453 // Fallthrough edge can't be transferred.
1454 if (Succ == FallThrough)
1456 ToBBI.BB->addSuccessor(Succ);
1460 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1461 std::back_inserter(ToBBI.Predicate));
1462 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate));
1464 ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1465 ToBBI.IsAnalyzed = false;
1470 /// MergeBlocks - Move all instructions from FromBB to the end of ToBB.
1471 /// This will leave FromBB as an empty block, so remove all of its
1472 /// successor edges except for the fall-through edge. If AddEdges is true,
1473 /// i.e., when FromBBI's branch is being moved, add those successor edges to
1475 void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) {
1476 ToBBI.BB->splice(ToBBI.BB->end(),
1477 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end());
1479 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1480 FromBBI.BB->succ_end());
1481 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1482 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1484 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1485 MachineBasicBlock *Succ = Succs[i];
1486 // Fallthrough edge can't be transferred.
1487 if (Succ == FallThrough)
1489 FromBBI.BB->removeSuccessor(Succ);
1491 ToBBI.BB->addSuccessor(Succ);
1494 // Now FromBBI always falls through to the next block!
1495 if (NBB && !FromBBI.BB->isSuccessor(NBB))
1496 FromBBI.BB->addSuccessor(NBB);
1498 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1499 std::back_inserter(ToBBI.Predicate));
1500 FromBBI.Predicate.clear();
1502 ToBBI.NonPredSize += FromBBI.NonPredSize;
1503 ToBBI.ExtraCost += FromBBI.ExtraCost;
1504 ToBBI.ExtraCost2 += FromBBI.ExtraCost2;
1505 FromBBI.NonPredSize = 0;
1506 FromBBI.ExtraCost = 0;
1507 FromBBI.ExtraCost2 = 0;
1509 ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1510 ToBBI.HasFallThrough = FromBBI.HasFallThrough;
1511 ToBBI.IsAnalyzed = false;
1512 FromBBI.IsAnalyzed = false;