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 "llvm/Function.h"
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/CodeGen/MachineModuleInfo.h"
18 #include "llvm/CodeGen/MachineFunctionPass.h"
19 #include "llvm/Target/TargetInstrInfo.h"
20 #include "llvm/Target/TargetLowering.h"
21 #include "llvm/Target/TargetMachine.h"
22 #include "llvm/Support/CommandLine.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/ADT/DepthFirstIterator.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/STLExtras.h"
29 // Hidden options for help debugging.
30 static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden);
31 static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden);
32 static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden);
33 static cl::opt<bool> DisableSimple("disable-ifcvt-simple",
34 cl::init(false), cl::Hidden);
35 static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false",
36 cl::init(false), cl::Hidden);
37 static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle",
38 cl::init(false), cl::Hidden);
39 static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev",
40 cl::init(false), cl::Hidden);
41 static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false",
42 cl::init(false), cl::Hidden);
43 static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev",
44 cl::init(false), cl::Hidden);
45 static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond",
46 cl::init(false), cl::Hidden);
48 STATISTIC(NumSimple, "Number of simple if-conversions performed");
49 STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed");
50 STATISTIC(NumTriangle, "Number of triangle if-conversions performed");
51 STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed");
52 STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed");
53 STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed");
54 STATISTIC(NumDiamonds, "Number of diamond if-conversions performed");
55 STATISTIC(NumIfConvBBs, "Number of if-converted blocks");
56 STATISTIC(NumDupBBs, "Number of duplicated blocks");
59 class VISIBILITY_HIDDEN IfConverter : public MachineFunctionPass {
61 ICNotClassfied, // BB data valid, but not classified.
62 ICSimpleFalse, // Same as ICSimple, but on the false path.
63 ICSimple, // BB is entry of an one split, no rejoin sub-CFG.
64 ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition.
65 ICTriangleRev, // Same as ICTriangle, but true path rev condition.
66 ICTriangleFalse, // Same as ICTriangle, but on the false path.
67 ICTriangle, // BB is entry of a triangle sub-CFG.
68 ICDiamond // BB is entry of a diamond sub-CFG.
71 /// BBInfo - One per MachineBasicBlock, this is used to cache the result
72 /// if-conversion feasibility analysis. This includes results from
73 /// TargetInstrInfo::AnalyzeBranch() (i.e. TBB, FBB, and Cond), and its
74 /// classification, and common tail block of its successors (if it's a
75 /// diamond shape), its size, whether it's predicable, and whether any
76 /// instruction can clobber the 'would-be' predicate.
78 /// IsDone - True if BB is not to be considered for ifcvt.
79 /// IsBeingAnalyzed - True if BB is currently being analyzed.
80 /// IsAnalyzed - True if BB has been analyzed (info is still valid).
81 /// IsEnqueued - True if BB has been enqueued to be ifcvt'ed.
82 /// IsBrAnalyzable - True if AnalyzeBranch() returns false.
83 /// HasFallThrough - True if BB may fallthrough to the following BB.
84 /// IsUnpredicable - True if BB is known to be unpredicable.
85 /// ClobbersPred - True if BB could modify predicates (e.g. has
87 /// NonPredSize - Number of non-predicated instructions.
88 /// BB - Corresponding MachineBasicBlock.
89 /// TrueBB / FalseBB- See AnalyzeBranch().
90 /// BrCond - Conditions for end of block conditional branches.
91 /// Predicate - Predicate used in the BB.
94 bool IsBeingAnalyzed : 1;
97 bool IsBrAnalyzable : 1;
98 bool HasFallThrough : 1;
99 bool IsUnpredicable : 1;
100 bool CannotBeCopied : 1;
101 bool ClobbersPred : 1;
102 unsigned NonPredSize;
103 MachineBasicBlock *BB;
104 MachineBasicBlock *TrueBB;
105 MachineBasicBlock *FalseBB;
106 std::vector<MachineOperand> BrCond;
107 std::vector<MachineOperand> Predicate;
108 BBInfo() : IsDone(false), IsBeingAnalyzed(false),
109 IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false),
110 HasFallThrough(false), IsUnpredicable(false),
111 CannotBeCopied(false), ClobbersPred(false), NonPredSize(0),
112 BB(0), TrueBB(0), FalseBB(0) {}
115 /// IfcvtToken - Record information about pending if-conversions to attemp:
116 /// BBI - Corresponding BBInfo.
117 /// Kind - Type of block. See IfcvtKind.
118 /// NeedSubsumsion - True if the to be predicated BB has already been
120 /// NumDups - Number of instructions that would be duplicated due
121 /// to this if-conversion. (For diamonds, the number of
122 /// identical instructions at the beginnings of both
124 /// NumDups2 - For diamonds, the number of identical instructions
125 /// at the ends of both paths.
132 IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0)
133 : BBI(b), Kind(k), NeedSubsumsion(s), NumDups(d), NumDups2(d2) {}
136 /// Roots - Basic blocks that do not have successors. These are the starting
137 /// points of Graph traversal.
138 std::vector<MachineBasicBlock*> Roots;
140 /// BBAnalysis - Results of if-conversion feasibility analysis indexed by
141 /// basic block number.
142 std::vector<BBInfo> BBAnalysis;
144 const TargetLowering *TLI;
145 const TargetInstrInfo *TII;
149 IfConverter() : MachineFunctionPass((intptr_t)&ID) {}
151 virtual bool runOnMachineFunction(MachineFunction &MF);
152 virtual const char *getPassName() const { return "If Converter"; }
155 bool ReverseBranchCondition(BBInfo &BBI);
156 bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const;
157 bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
158 bool FalseBranch, unsigned &Dups) const;
159 bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
160 unsigned &Dups1, unsigned &Dups2) const;
161 void ScanInstructions(BBInfo &BBI);
162 BBInfo &AnalyzeBlock(MachineBasicBlock *BB,
163 std::vector<IfcvtToken*> &Tokens);
164 bool FeasibilityAnalysis(BBInfo &BBI, std::vector<MachineOperand> &Cond,
165 bool isTriangle = false, bool RevBranch = false);
166 bool AnalyzeBlocks(MachineFunction &MF,
167 std::vector<IfcvtToken*> &Tokens);
168 void InvalidatePreds(MachineBasicBlock *BB);
169 void RemoveExtraEdges(BBInfo &BBI);
170 bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind);
171 bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind);
172 bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
173 unsigned NumDups1, unsigned NumDups2);
174 void PredicateBlock(BBInfo &BBI,
175 MachineBasicBlock::iterator E,
176 std::vector<MachineOperand> &Cond);
177 void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
178 std::vector<MachineOperand> &Cond,
179 bool IgnoreBr = false);
180 void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI);
182 bool MeetIfcvtSizeLimit(unsigned Size) const {
183 return Size > 0 && Size <= TLI->getIfCvtBlockSizeLimit();
186 // blockAlwaysFallThrough - Block ends without a terminator.
187 bool blockAlwaysFallThrough(BBInfo &BBI) const {
188 return BBI.IsBrAnalyzable && BBI.TrueBB == NULL;
191 // IfcvtTokenCmp - Used to sort if-conversion candidates.
192 static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) {
193 int Incr1 = (C1->Kind == ICDiamond)
194 ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups;
195 int Incr2 = (C2->Kind == ICDiamond)
196 ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups;
199 else if (Incr1 == Incr2) {
200 // Favors subsumsion.
201 if (C1->NeedSubsumsion == false && C2->NeedSubsumsion == true)
203 else if (C1->NeedSubsumsion == C2->NeedSubsumsion) {
204 // Favors diamond over triangle, etc.
205 if ((unsigned)C1->Kind < (unsigned)C2->Kind)
207 else if (C1->Kind == C2->Kind)
208 return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber();
215 char IfConverter::ID = 0;
218 static RegisterPass<IfConverter>
219 X("if-converter", "If Converter");
221 FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); }
223 bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
224 TLI = MF.getTarget().getTargetLowering();
225 TII = MF.getTarget().getInstrInfo();
226 if (!TII) return false;
228 static int FnNum = -1;
229 DOUT << "\nIfcvt: function (" << ++FnNum << ") \'"
230 << MF.getFunction()->getName() << "\'";
232 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) {
233 DOUT << " skipped\n";
239 BBAnalysis.resize(MF.getNumBlockIDs());
241 // Look for root nodes, i.e. blocks without successors.
242 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
246 std::vector<IfcvtToken*> Tokens;
248 unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle +
249 NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds;
250 while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) {
251 // Do an intial analysis for each basic block and finding all the potential
252 // candidates to perform if-convesion.
253 bool Change = AnalyzeBlocks(MF, Tokens);
254 while (!Tokens.empty()) {
255 IfcvtToken *Token = Tokens.back();
257 BBInfo &BBI = Token->BBI;
258 IfcvtKind Kind = Token->Kind;
260 // If the block has been evicted out of the queue or it has already been
261 // marked dead (due to it being predicated), then skip it.
263 BBI.IsEnqueued = false;
267 BBI.IsEnqueued = false;
271 default: assert(false && "Unexpected!");
274 case ICSimpleFalse: {
275 bool isFalse = Kind == ICSimpleFalse;
276 if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break;
277 DOUT << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? " false" :"")
278 << "): BB#" << BBI.BB->getNumber() << " ("
279 << ((Kind == ICSimpleFalse)
280 ? BBI.FalseBB->getNumber()
281 : BBI.TrueBB->getNumber()) << ") ";
282 RetVal = IfConvertSimple(BBI, Kind);
283 DOUT << (RetVal ? "succeeded!" : "failed!") << "\n";
285 if (isFalse) NumSimpleFalse++;
292 case ICTriangleFalse:
293 case ICTriangleFRev: {
294 bool isFalse = Kind == ICTriangleFalse;
295 bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev);
296 if (DisableTriangle && !isFalse && !isRev) break;
297 if (DisableTriangleR && !isFalse && isRev) break;
298 if (DisableTriangleF && isFalse && !isRev) break;
299 if (DisableTriangleFR && isFalse && isRev) break;
300 DOUT << "Ifcvt (Triangle";
305 DOUT << "): BB#" << BBI.BB->getNumber() << " (T:"
306 << BBI.TrueBB->getNumber() << ",F:"
307 << BBI.FalseBB->getNumber() << ") ";
308 RetVal = IfConvertTriangle(BBI, Kind);
309 DOUT << (RetVal ? "succeeded!" : "failed!") << "\n";
312 if (isRev) NumTriangleFRev++;
313 else NumTriangleFalse++;
315 if (isRev) NumTriangleRev++;
322 if (DisableDiamond) break;
323 DOUT << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:"
324 << BBI.TrueBB->getNumber() << ",F:"
325 << BBI.FalseBB->getNumber() << ") ";
326 RetVal = IfConvertDiamond(BBI, Kind, Token->NumDups, Token->NumDups2);
327 DOUT << (RetVal ? "succeeded!" : "failed!") << "\n";
328 if (RetVal) NumDiamonds++;
335 NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev +
336 NumTriangleFalse + NumTriangleFRev + NumDiamonds;
337 if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit)
343 MadeChange |= Change;
346 // Delete tokens in case of early exit.
347 while (!Tokens.empty()) {
348 IfcvtToken *Token = Tokens.back();
360 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
361 /// its 'true' successor.
362 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
363 MachineBasicBlock *TrueBB) {
364 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
365 E = BB->succ_end(); SI != E; ++SI) {
366 MachineBasicBlock *SuccBB = *SI;
367 if (SuccBB != TrueBB)
373 /// ReverseBranchCondition - Reverse the condition of the end of the block
374 /// branchs. Swap block's 'true' and 'false' successors.
375 bool IfConverter::ReverseBranchCondition(BBInfo &BBI) {
376 if (!TII->ReverseBranchCondition(BBI.BrCond)) {
377 TII->RemoveBranch(*BBI.BB);
378 TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond);
379 std::swap(BBI.TrueBB, BBI.FalseBB);
385 /// getNextBlock - Returns the next block in the function blocks ordering. If
386 /// it is the end, returns NULL.
387 static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) {
388 MachineFunction::iterator I = BB;
389 MachineFunction::iterator E = BB->getParent()->end();
395 /// ValidSimple - Returns true if the 'true' block (along with its
396 /// predecessor) forms a valid simple shape for ifcvt. It also returns the
397 /// number of instructions that the ifcvt would need to duplicate if performed
399 bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const {
401 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
404 if (TrueBBI.IsBrAnalyzable)
407 if (TrueBBI.BB->pred_size() > 1) {
408 if (TrueBBI.CannotBeCopied ||
409 TrueBBI.NonPredSize > TLI->getIfCvtDupBlockSizeLimit())
411 Dups = TrueBBI.NonPredSize;
417 /// ValidTriangle - Returns true if the 'true' and 'false' blocks (along
418 /// with their common predecessor) forms a valid triangle shape for ifcvt.
419 /// If 'FalseBranch' is true, it checks if 'true' block's false branch
420 /// branches to the false branch rather than the other way around. It also
421 /// returns the number of instructions that the ifcvt would need to duplicate
422 /// if performed in 'Dups'.
423 bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
424 bool FalseBranch, unsigned &Dups) const {
426 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
429 if (TrueBBI.BB->pred_size() > 1) {
430 if (TrueBBI.CannotBeCopied)
433 unsigned Size = TrueBBI.NonPredSize;
434 if (TrueBBI.IsBrAnalyzable) {
435 if (TrueBBI.TrueBB && TrueBBI.BrCond.empty())
436 // End with an unconditional branch. It will be removed.
439 MachineBasicBlock *FExit = FalseBranch
440 ? TrueBBI.TrueBB : TrueBBI.FalseBB;
442 // Require a conditional branch
446 if (Size > TLI->getIfCvtDupBlockSizeLimit())
451 MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB;
452 if (!TExit && blockAlwaysFallThrough(TrueBBI)) {
453 MachineFunction::iterator I = TrueBBI.BB;
454 if (++I == TrueBBI.BB->getParent()->end())
458 return TExit && TExit == FalseBBI.BB;
462 MachineBasicBlock::iterator firstNonBranchInst(MachineBasicBlock *BB,
463 const TargetInstrInfo *TII) {
464 MachineBasicBlock::iterator I = BB->end();
465 while (I != BB->begin()) {
467 if (!I->getDesc().isBranch())
473 /// ValidDiamond - Returns true if the 'true' and 'false' blocks (along
474 /// with their common predecessor) forms a valid diamond shape for ifcvt.
475 bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
476 unsigned &Dups1, unsigned &Dups2) const {
478 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone ||
479 FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone)
482 MachineBasicBlock *TT = TrueBBI.TrueBB;
483 MachineBasicBlock *FT = FalseBBI.TrueBB;
485 if (!TT && blockAlwaysFallThrough(TrueBBI))
486 TT = getNextBlock(TrueBBI.BB);
487 if (!FT && blockAlwaysFallThrough(FalseBBI))
488 FT = getNextBlock(FalseBBI.BB);
491 if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable))
493 if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1)
496 // FIXME: Allow true block to have an early exit?
497 if (TrueBBI.FalseBB || FalseBBI.FalseBB ||
498 (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred))
501 MachineBasicBlock::iterator TI = TrueBBI.BB->begin();
502 MachineBasicBlock::iterator FI = FalseBBI.BB->begin();
503 while (TI != TrueBBI.BB->end() && FI != FalseBBI.BB->end()) {
504 if (!TI->isIdenticalTo(FI))
511 TI = firstNonBranchInst(TrueBBI.BB, TII);
512 FI = firstNonBranchInst(FalseBBI.BB, TII);
513 while (TI != TrueBBI.BB->begin() && FI != FalseBBI.BB->begin()) {
514 if (!TI->isIdenticalTo(FI))
524 /// ScanInstructions - Scan all the instructions in the block to determine if
525 /// the block is predicable. In most cases, that means all the instructions
526 /// in the block are isPredicable(). Also checks if the block contains any
527 /// instruction which can clobber a predicate (e.g. condition code register).
528 /// If so, the block is not predicable unless it's the last instruction.
529 void IfConverter::ScanInstructions(BBInfo &BBI) {
533 bool AlreadyPredicated = BBI.Predicate.size() > 0;
534 // First analyze the end of BB branches.
535 BBI.TrueBB = BBI.FalseBB = NULL;
538 !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond);
539 BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL;
541 if (BBI.BrCond.size()) {
542 // No false branch. This BB must end with a conditional branch and a
545 BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB);
546 assert(BBI.FalseBB && "Expected to find the fallthrough block!");
549 // Then scan all the instructions.
551 BBI.ClobbersPred = false;
552 bool SeenCondBr = false;
553 for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end();
555 const TargetInstrDesc &TID = I->getDesc();
556 if (TID.isNotDuplicable())
557 BBI.CannotBeCopied = true;
559 bool isPredicated = TII->isPredicated(I);
560 bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch();
565 else if (!AlreadyPredicated) {
566 // FIXME: This instruction is already predicated before the
567 // if-conversion pass. It's probably something like a conditional move.
568 // Mark this block unpredicable for now.
569 BBI.IsUnpredicable = true;
575 if (BBI.ClobbersPred && !isPredicated) {
576 // Predicate modification instruction should end the block (except for
577 // already predicated instructions and end of block branches).
581 // Conditional branches is not predicable. But it may be eliminated.
585 // Predicate may have been modified, the subsequent (currently)
586 // unpredicated instructions cannot be correctly predicated.
587 BBI.IsUnpredicable = true;
591 // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are
592 // still potentially predicable.
593 std::vector<MachineOperand> PredDefs;
594 if (TII->DefinesPredicate(I, PredDefs))
595 BBI.ClobbersPred = true;
597 if (!TID.isPredicable()) {
598 BBI.IsUnpredicable = true;
604 /// FeasibilityAnalysis - Determine if the block is a suitable candidate to be
605 /// predicated by the specified predicate.
606 bool IfConverter::FeasibilityAnalysis(BBInfo &BBI,
607 std::vector<MachineOperand> &Pred,
608 bool isTriangle, bool RevBranch) {
609 // If the block is dead or unpredicable, then it cannot be predicated.
610 if (BBI.IsDone || BBI.IsUnpredicable)
613 // If it is already predicated, check if its predicate subsumes the new
615 if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred))
618 if (BBI.BrCond.size()) {
622 // Test predicate subsumsion.
623 std::vector<MachineOperand> RevPred(Pred);
624 std::vector<MachineOperand> Cond(BBI.BrCond);
626 if (TII->ReverseBranchCondition(Cond))
629 if (TII->ReverseBranchCondition(RevPred) ||
630 !TII->SubsumesPredicate(Cond, RevPred))
637 /// AnalyzeBlock - Analyze the structure of the sub-CFG starting from
638 /// the specified block. Record its successors and whether it looks like an
639 /// if-conversion candidate.
640 IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
641 std::vector<IfcvtToken*> &Tokens) {
642 BBInfo &BBI = BBAnalysis[BB->getNumber()];
644 if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed)
648 BBI.IsBeingAnalyzed = true;
650 ScanInstructions(BBI);
652 // Unanalyable or ends with fallthrough or unconditional branch.
653 if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) {
654 BBI.IsBeingAnalyzed = false;
655 BBI.IsAnalyzed = true;
659 // Do not ifcvt if either path is a back edge to the entry block.
660 if (BBI.TrueBB == BB || BBI.FalseBB == BB) {
661 BBI.IsBeingAnalyzed = false;
662 BBI.IsAnalyzed = true;
666 BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens);
667 BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens);
669 if (TrueBBI.IsDone && FalseBBI.IsDone) {
670 BBI.IsBeingAnalyzed = false;
671 BBI.IsAnalyzed = true;
675 std::vector<MachineOperand> RevCond(BBI.BrCond);
676 bool CanRevCond = !TII->ReverseBranchCondition(RevCond);
680 bool TNeedSub = TrueBBI.Predicate.size() > 0;
681 bool FNeedSub = FalseBBI.Predicate.size() > 0;
682 bool Enqueued = false;
683 if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) &&
684 MeetIfcvtSizeLimit(TrueBBI.NonPredSize - (Dups + Dups2)) &&
685 MeetIfcvtSizeLimit(FalseBBI.NonPredSize - (Dups + Dups2)) &&
686 FeasibilityAnalysis(TrueBBI, BBI.BrCond) &&
687 FeasibilityAnalysis(FalseBBI, RevCond)) {
695 // Note TailBB can be empty.
696 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups,
701 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) &&
702 MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
703 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
711 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups));
715 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups) &&
716 MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
717 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) {
718 Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups));
722 if (ValidSimple(TrueBBI, Dups) &&
723 MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
724 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) {
725 // Simple (split, no rejoin):
732 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups));
737 // Try the other path...
738 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups) &&
739 MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
740 FeasibilityAnalysis(FalseBBI, RevCond, true)) {
741 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups));
745 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups) &&
746 MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
747 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) {
748 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups));
752 if (ValidSimple(FalseBBI, Dups) &&
753 MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
754 FeasibilityAnalysis(FalseBBI, RevCond)) {
755 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups));
760 BBI.IsEnqueued = Enqueued;
761 BBI.IsBeingAnalyzed = false;
762 BBI.IsAnalyzed = true;
766 /// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion
767 /// candidates. It returns true if any CFG restructuring is done to expose more
768 /// if-conversion opportunities.
769 bool IfConverter::AnalyzeBlocks(MachineFunction &MF,
770 std::vector<IfcvtToken*> &Tokens) {
772 std::set<MachineBasicBlock*> Visited;
773 for (unsigned i = 0, e = Roots.size(); i != e; ++i) {
774 for (idf_ext_iterator<MachineBasicBlock*> I=idf_ext_begin(Roots[i],Visited),
775 E = idf_ext_end(Roots[i], Visited); I != E; ++I) {
776 MachineBasicBlock *BB = *I;
777 AnalyzeBlock(BB, Tokens);
781 // Sort to favor more complex ifcvt scheme.
782 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp);
787 /// canFallThroughTo - Returns true either if ToBB is the next block after BB or
788 /// that all the intervening blocks are empty (given BB can fall through to its
790 static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) {
791 MachineFunction::iterator I = BB;
792 MachineFunction::iterator TI = ToBB;
793 MachineFunction::iterator E = BB->getParent()->end();
795 if (I == E || !I->empty())
800 /// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed
801 /// to determine if it can be if-converted. If predecessor is already enqueued,
803 void IfConverter::InvalidatePreds(MachineBasicBlock *BB) {
804 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(),
805 E = BB->pred_end(); PI != E; ++PI) {
806 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()];
807 if (PBBI.IsDone || PBBI.BB == BB)
809 PBBI.IsAnalyzed = false;
810 PBBI.IsEnqueued = false;
814 /// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB.
816 static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB,
817 const TargetInstrInfo *TII) {
818 std::vector<MachineOperand> NoCond;
819 TII->InsertBranch(*BB, ToBB, NULL, NoCond);
822 /// RemoveExtraEdges - Remove true / false edges if either / both are no longer
824 void IfConverter::RemoveExtraEdges(BBInfo &BBI) {
825 MachineBasicBlock *TBB = NULL, *FBB = NULL;
826 std::vector<MachineOperand> Cond;
827 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond))
828 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
831 /// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG.
833 bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) {
834 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
835 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
836 BBInfo *CvtBBI = &TrueBBI;
837 BBInfo *NextBBI = &FalseBBI;
839 std::vector<MachineOperand> Cond(BBI.BrCond);
840 if (Kind == ICSimpleFalse)
841 std::swap(CvtBBI, NextBBI);
843 if (CvtBBI->IsDone ||
844 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
845 // Something has changed. It's no longer safe to predicate this block.
846 BBI.IsAnalyzed = false;
847 CvtBBI->IsAnalyzed = false;
851 if (Kind == ICSimpleFalse)
852 TII->ReverseBranchCondition(Cond);
854 if (CvtBBI->BB->pred_size() > 1) {
855 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
856 // Copy instructions in the true block, predicate them add them to
858 CopyAndPredicateBlock(BBI, *CvtBBI, Cond);
860 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond);
862 // Merge converted block into entry block.
863 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
864 MergeBlocks(BBI, *CvtBBI);
867 bool IterIfcvt = true;
868 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) {
869 InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
870 BBI.HasFallThrough = false;
871 // Now ifcvt'd block will look like this:
878 // We cannot further ifcvt this block because the unconditional branch
879 // will have to be predicated on the new condition, that will not be
880 // available if cmp executes.
884 RemoveExtraEdges(BBI);
886 // Update block info. BB can be iteratively if-converted.
889 InvalidatePreds(BBI.BB);
890 CvtBBI->IsDone = true;
892 // FIXME: Must maintain LiveIns.
896 /// IfConvertTriangle - If convert a triangle sub-CFG.
898 bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
899 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
900 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
901 BBInfo *CvtBBI = &TrueBBI;
902 BBInfo *NextBBI = &FalseBBI;
904 std::vector<MachineOperand> Cond(BBI.BrCond);
905 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
906 std::swap(CvtBBI, NextBBI);
908 if (CvtBBI->IsDone ||
909 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
910 // Something has changed. It's no longer safe to predicate this block.
911 BBI.IsAnalyzed = false;
912 CvtBBI->IsAnalyzed = false;
916 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
917 TII->ReverseBranchCondition(Cond);
919 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) {
920 ReverseBranchCondition(*CvtBBI);
921 // BB has been changed, modify its predecessors (except for this
922 // one) so they don't get ifcvt'ed based on bad intel.
923 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(),
924 E = CvtBBI->BB->pred_end(); PI != E; ++PI) {
925 MachineBasicBlock *PBB = *PI;
928 BBInfo &PBBI = BBAnalysis[PBB->getNumber()];
929 if (PBBI.IsEnqueued) {
930 PBBI.IsAnalyzed = false;
931 PBBI.IsEnqueued = false;
936 bool HasEarlyExit = CvtBBI->FalseBB != NULL;
937 bool DupBB = CvtBBI->BB->pred_size() > 1;
939 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
940 // Copy instructions in the true block, predicate them add them to
942 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, true);
944 // Predicate the 'true' block after removing its branch.
945 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB);
946 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond);
950 // Now merge the entry of the triangle with the true block.
951 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
952 MergeBlocks(BBI, *CvtBBI);
955 // If 'true' block has a 'false' successor, add an exit branch to it.
957 std::vector<MachineOperand> RevCond(CvtBBI->BrCond);
958 if (TII->ReverseBranchCondition(RevCond))
959 assert(false && "Unable to reverse branch condition!");
960 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond);
961 BBI.BB->addSuccessor(CvtBBI->FalseBB);
964 // Merge in the 'false' block if the 'false' block has no other
965 // predecessors. Otherwise, add a unconditional branch from to 'false'.
966 bool FalseBBDead = false;
967 bool IterIfcvt = true;
968 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB);
969 if (!isFallThrough) {
970 // Only merge them if the true block does not fallthrough to the false
971 // block. By not merging them, we make it possible to iteratively
974 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) {
975 MergeBlocks(BBI, *NextBBI);
978 InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
979 BBI.HasFallThrough = false;
981 // Mixed predicated and unpredicated code. This cannot be iteratively
986 RemoveExtraEdges(BBI);
988 // Update block info. BB can be iteratively if-converted.
991 InvalidatePreds(BBI.BB);
992 CvtBBI->IsDone = true;
994 NextBBI->IsDone = true;
996 // FIXME: Must maintain LiveIns.
1000 /// IfConvertDiamond - If convert a diamond sub-CFG.
1002 bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
1003 unsigned NumDups1, unsigned NumDups2) {
1004 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
1005 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1006 MachineBasicBlock *TailBB = TrueBBI.TrueBB;
1007 // True block must fall through or ended with unanalyzable terminator.
1009 if (blockAlwaysFallThrough(TrueBBI))
1010 TailBB = FalseBBI.TrueBB;
1011 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!");
1014 if (TrueBBI.IsDone || FalseBBI.IsDone ||
1015 TrueBBI.BB->pred_size() > 1 ||
1016 FalseBBI.BB->pred_size() > 1) {
1017 // Something has changed. It's no longer safe to predicate these blocks.
1018 BBI.IsAnalyzed = false;
1019 TrueBBI.IsAnalyzed = false;
1020 FalseBBI.IsAnalyzed = false;
1024 // Merge the 'true' and 'false' blocks by copying the instructions
1025 // from the 'false' block to the 'true' block. That is, unless the true
1026 // block would clobber the predicate, in that case, do the opposite.
1027 BBInfo *BBI1 = &TrueBBI;
1028 BBInfo *BBI2 = &FalseBBI;
1029 std::vector<MachineOperand> RevCond(BBI.BrCond);
1030 TII->ReverseBranchCondition(RevCond);
1031 std::vector<MachineOperand> *Cond1 = &BBI.BrCond;
1032 std::vector<MachineOperand> *Cond2 = &RevCond;
1034 // Figure out the more profitable ordering.
1035 bool DoSwap = false;
1036 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred)
1038 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) {
1039 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize)
1043 std::swap(BBI1, BBI2);
1044 std::swap(Cond1, Cond2);
1047 // Remove the conditional branch from entry to the blocks.
1048 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1050 // Remove the duplicated instructions at the beginnings of both paths.
1051 MachineBasicBlock::iterator DI1 = BBI1->BB->begin();
1052 MachineBasicBlock::iterator DI2 = BBI2->BB->begin();
1053 BBI1->NonPredSize -= NumDups1;
1054 BBI2->NonPredSize -= NumDups1;
1055 while (NumDups1 != 0) {
1060 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1);
1061 BBI2->BB->erase(BBI2->BB->begin(), DI2);
1063 // Predicate the 'true' block after removing its branch.
1064 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB);
1065 DI1 = BBI1->BB->end();
1066 for (unsigned i = 0; i != NumDups2; ++i)
1068 BBI1->BB->erase(DI1, BBI1->BB->end());
1069 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1);
1071 // Predicate the 'false' block.
1072 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB);
1073 DI2 = BBI2->BB->end();
1074 while (NumDups2 != 0) {
1078 PredicateBlock(*BBI2, DI2, *Cond2);
1080 // Merge the true block into the entry of the diamond.
1081 MergeBlocks(BBI, *BBI1);
1082 MergeBlocks(BBI, *BBI2);
1084 // If the if-converted block fallthrough or unconditionally branch into the
1085 // tail block, and the tail block does not have other predecessors, then
1086 // fold the tail block in as well. Otherwise, unless it falls through to the
1087 // tail, add a unconditional branch to it.
1089 BBInfo TailBBI = BBAnalysis[TailBB->getNumber()];
1090 if (TailBB->pred_size() == 1 && !TailBBI.HasFallThrough) {
1091 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1092 MergeBlocks(BBI, TailBBI);
1093 TailBBI.IsDone = true;
1095 InsertUncondBranch(BBI.BB, TailBB, TII);
1096 BBI.HasFallThrough = false;
1100 RemoveExtraEdges(BBI);
1102 // Update block info.
1103 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true;
1104 InvalidatePreds(BBI.BB);
1106 // FIXME: Must maintain LiveIns.
1110 /// PredicateBlock - Predicate instructions from the start of the block to the
1111 /// specified end with the specified condition.
1112 void IfConverter::PredicateBlock(BBInfo &BBI,
1113 MachineBasicBlock::iterator E,
1114 std::vector<MachineOperand> &Cond) {
1115 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) {
1116 if (TII->isPredicated(I))
1118 if (!TII->PredicateInstruction(I, Cond)) {
1119 cerr << "Unable to predicate " << *I << "!\n";
1124 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate));
1126 BBI.IsAnalyzed = false;
1127 BBI.NonPredSize = 0;
1132 /// CopyAndPredicateBlock - Copy and predicate instructions from source BB to
1133 /// the destination block. Skip end of block branches if IgnoreBr is true.
1134 void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
1135 std::vector<MachineOperand> &Cond,
1137 MachineFunction &MF = *ToBBI.BB->getParent();
1139 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(),
1140 E = FromBBI.BB->end(); I != E; ++I) {
1141 const TargetInstrDesc &TID = I->getDesc();
1142 bool isPredicated = TII->isPredicated(I);
1143 // Do not copy the end of the block branches.
1144 if (IgnoreBr && !isPredicated && TID.isBranch())
1147 MachineInstr *MI = MF.CloneMachineInstr(I);
1148 ToBBI.BB->insert(ToBBI.BB->end(), MI);
1149 ToBBI.NonPredSize++;
1152 if (!TII->PredicateInstruction(MI, Cond)) {
1153 cerr << "Unable to predicate " << *MI << "!\n";
1158 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1159 FromBBI.BB->succ_end());
1160 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1161 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1163 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1164 MachineBasicBlock *Succ = Succs[i];
1165 // Fallthrough edge can't be transferred.
1166 if (Succ == FallThrough)
1168 if (!ToBBI.BB->isSuccessor(Succ))
1169 ToBBI.BB->addSuccessor(Succ);
1172 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1173 std::back_inserter(ToBBI.Predicate));
1174 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate));
1176 ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1177 ToBBI.IsAnalyzed = false;
1182 /// MergeBlocks - Move all instructions from FromBB to the end of ToBB.
1184 void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) {
1185 ToBBI.BB->splice(ToBBI.BB->end(),
1186 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end());
1188 // Redirect all branches to FromBB to ToBB.
1189 std::vector<MachineBasicBlock *> Preds(FromBBI.BB->pred_begin(),
1190 FromBBI.BB->pred_end());
1191 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
1192 MachineBasicBlock *Pred = Preds[i];
1193 if (Pred == ToBBI.BB)
1195 Pred->ReplaceUsesOfBlockWith(FromBBI.BB, ToBBI.BB);
1198 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1199 FromBBI.BB->succ_end());
1200 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1201 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1203 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1204 MachineBasicBlock *Succ = Succs[i];
1205 // Fallthrough edge can't be transferred.
1206 if (Succ == FallThrough)
1208 FromBBI.BB->removeSuccessor(Succ);
1209 if (!ToBBI.BB->isSuccessor(Succ))
1210 ToBBI.BB->addSuccessor(Succ);
1213 // Now FromBBI always fall through to the next block!
1214 if (NBB && !FromBBI.BB->isSuccessor(NBB))
1215 FromBBI.BB->addSuccessor(NBB);
1217 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1218 std::back_inserter(ToBBI.Predicate));
1219 FromBBI.Predicate.clear();
1221 ToBBI.NonPredSize += FromBBI.NonPredSize;
1222 FromBBI.NonPredSize = 0;
1224 ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1225 ToBBI.HasFallThrough = FromBBI.HasFallThrough;
1226 ToBBI.IsAnalyzed = false;
1227 FromBBI.IsAnalyzed = false;