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 FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); }
220 bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
221 TLI = MF.getTarget().getTargetLowering();
222 TII = MF.getTarget().getInstrInfo();
223 if (!TII) return false;
225 static int FnNum = -1;
226 DOUT << "\nIfcvt: function (" << ++FnNum << ") \'"
227 << MF.getFunction()->getName() << "\'";
229 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) {
230 DOUT << " skipped\n";
236 BBAnalysis.resize(MF.getNumBlockIDs());
238 // Look for root nodes, i.e. blocks without successors.
239 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
243 std::vector<IfcvtToken*> Tokens;
245 unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle +
246 NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds;
247 while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) {
248 // Do an intial analysis for each basic block and finding all the potential
249 // candidates to perform if-convesion.
250 bool Change = AnalyzeBlocks(MF, Tokens);
251 while (!Tokens.empty()) {
252 IfcvtToken *Token = Tokens.back();
254 BBInfo &BBI = Token->BBI;
255 IfcvtKind Kind = Token->Kind;
257 // If the block has been evicted out of the queue or it has already been
258 // marked dead (due to it being predicated), then skip it.
260 BBI.IsEnqueued = false;
264 BBI.IsEnqueued = false;
268 default: assert(false && "Unexpected!");
271 case ICSimpleFalse: {
272 bool isFalse = Kind == ICSimpleFalse;
273 if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break;
274 DOUT << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? " false" :"")
275 << "): BB#" << BBI.BB->getNumber() << " ("
276 << ((Kind == ICSimpleFalse)
277 ? BBI.FalseBB->getNumber()
278 : BBI.TrueBB->getNumber()) << ") ";
279 RetVal = IfConvertSimple(BBI, Kind);
280 DOUT << (RetVal ? "succeeded!" : "failed!") << "\n";
282 if (isFalse) NumSimpleFalse++;
289 case ICTriangleFalse:
290 case ICTriangleFRev: {
291 bool isFalse = Kind == ICTriangleFalse;
292 bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev);
293 if (DisableTriangle && !isFalse && !isRev) break;
294 if (DisableTriangleR && !isFalse && isRev) break;
295 if (DisableTriangleF && isFalse && !isRev) break;
296 if (DisableTriangleFR && isFalse && isRev) break;
297 DOUT << "Ifcvt (Triangle";
302 DOUT << "): BB#" << BBI.BB->getNumber() << " (T:"
303 << BBI.TrueBB->getNumber() << ",F:"
304 << BBI.FalseBB->getNumber() << ") ";
305 RetVal = IfConvertTriangle(BBI, Kind);
306 DOUT << (RetVal ? "succeeded!" : "failed!") << "\n";
309 if (isRev) NumTriangleFRev++;
310 else NumTriangleFalse++;
312 if (isRev) NumTriangleRev++;
319 if (DisableDiamond) break;
320 DOUT << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:"
321 << BBI.TrueBB->getNumber() << ",F:"
322 << BBI.FalseBB->getNumber() << ") ";
323 RetVal = IfConvertDiamond(BBI, Kind, Token->NumDups, Token->NumDups2);
324 DOUT << (RetVal ? "succeeded!" : "failed!") << "\n";
325 if (RetVal) NumDiamonds++;
332 NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev +
333 NumTriangleFalse + NumTriangleFRev + NumDiamonds;
334 if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit)
340 MadeChange |= Change;
343 // Delete tokens in case of early exit.
344 while (!Tokens.empty()) {
345 IfcvtToken *Token = Tokens.back();
357 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
358 /// its 'true' successor.
359 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
360 MachineBasicBlock *TrueBB) {
361 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
362 E = BB->succ_end(); SI != E; ++SI) {
363 MachineBasicBlock *SuccBB = *SI;
364 if (SuccBB != TrueBB)
370 /// ReverseBranchCondition - Reverse the condition of the end of the block
371 /// branchs. Swap block's 'true' and 'false' successors.
372 bool IfConverter::ReverseBranchCondition(BBInfo &BBI) {
373 if (!TII->ReverseBranchCondition(BBI.BrCond)) {
374 TII->RemoveBranch(*BBI.BB);
375 TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond);
376 std::swap(BBI.TrueBB, BBI.FalseBB);
382 /// getNextBlock - Returns the next block in the function blocks ordering. If
383 /// it is the end, returns NULL.
384 static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) {
385 MachineFunction::iterator I = BB;
386 MachineFunction::iterator E = BB->getParent()->end();
392 /// ValidSimple - Returns true if the 'true' block (along with its
393 /// predecessor) forms a valid simple shape for ifcvt. It also returns the
394 /// number of instructions that the ifcvt would need to duplicate if performed
396 bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const {
398 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
401 if (TrueBBI.IsBrAnalyzable)
404 if (TrueBBI.BB->pred_size() > 1) {
405 if (TrueBBI.CannotBeCopied ||
406 TrueBBI.NonPredSize > TLI->getIfCvtDupBlockSizeLimit())
408 Dups = TrueBBI.NonPredSize;
414 /// ValidTriangle - Returns true if the 'true' and 'false' blocks (along
415 /// with their common predecessor) forms a valid triangle shape for ifcvt.
416 /// If 'FalseBranch' is true, it checks if 'true' block's false branch
417 /// branches to the false branch rather than the other way around. It also
418 /// returns the number of instructions that the ifcvt would need to duplicate
419 /// if performed in 'Dups'.
420 bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
421 bool FalseBranch, unsigned &Dups) const {
423 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
426 if (TrueBBI.BB->pred_size() > 1) {
427 if (TrueBBI.CannotBeCopied)
430 unsigned Size = TrueBBI.NonPredSize;
431 if (TrueBBI.IsBrAnalyzable) {
432 if (TrueBBI.TrueBB && TrueBBI.BrCond.empty())
433 // End with an unconditional branch. It will be removed.
436 MachineBasicBlock *FExit = FalseBranch
437 ? TrueBBI.TrueBB : TrueBBI.FalseBB;
439 // Require a conditional branch
443 if (Size > TLI->getIfCvtDupBlockSizeLimit())
448 MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB;
449 if (!TExit && blockAlwaysFallThrough(TrueBBI)) {
450 MachineFunction::iterator I = TrueBBI.BB;
451 if (++I == TrueBBI.BB->getParent()->end())
455 return TExit && TExit == FalseBBI.BB;
459 MachineBasicBlock::iterator firstNonBranchInst(MachineBasicBlock *BB,
460 const TargetInstrInfo *TII) {
461 MachineBasicBlock::iterator I = BB->end();
462 while (I != BB->begin()) {
464 if (!I->getDesc().isBranch())
470 /// ValidDiamond - Returns true if the 'true' and 'false' blocks (along
471 /// with their common predecessor) forms a valid diamond shape for ifcvt.
472 bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
473 unsigned &Dups1, unsigned &Dups2) const {
475 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone ||
476 FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone)
479 MachineBasicBlock *TT = TrueBBI.TrueBB;
480 MachineBasicBlock *FT = FalseBBI.TrueBB;
482 if (!TT && blockAlwaysFallThrough(TrueBBI))
483 TT = getNextBlock(TrueBBI.BB);
484 if (!FT && blockAlwaysFallThrough(FalseBBI))
485 FT = getNextBlock(FalseBBI.BB);
488 if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable))
490 if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1)
493 // FIXME: Allow true block to have an early exit?
494 if (TrueBBI.FalseBB || FalseBBI.FalseBB ||
495 (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred))
498 MachineBasicBlock::iterator TI = TrueBBI.BB->begin();
499 MachineBasicBlock::iterator FI = FalseBBI.BB->begin();
500 while (TI != TrueBBI.BB->end() && FI != FalseBBI.BB->end()) {
501 if (!TI->isIdenticalTo(FI))
508 TI = firstNonBranchInst(TrueBBI.BB, TII);
509 FI = firstNonBranchInst(FalseBBI.BB, TII);
510 while (TI != TrueBBI.BB->begin() && FI != FalseBBI.BB->begin()) {
511 if (!TI->isIdenticalTo(FI))
521 /// ScanInstructions - Scan all the instructions in the block to determine if
522 /// the block is predicable. In most cases, that means all the instructions
523 /// in the block are isPredicable(). Also checks if the block contains any
524 /// instruction which can clobber a predicate (e.g. condition code register).
525 /// If so, the block is not predicable unless it's the last instruction.
526 void IfConverter::ScanInstructions(BBInfo &BBI) {
530 bool AlreadyPredicated = BBI.Predicate.size() > 0;
531 // First analyze the end of BB branches.
532 BBI.TrueBB = BBI.FalseBB = NULL;
535 !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond);
536 BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL;
538 if (BBI.BrCond.size()) {
539 // No false branch. This BB must end with a conditional branch and a
542 BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB);
543 assert(BBI.FalseBB && "Expected to find the fallthrough block!");
546 // Then scan all the instructions.
548 BBI.ClobbersPred = false;
549 bool SeenCondBr = false;
550 for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end();
552 const TargetInstrDesc &TID = I->getDesc();
553 if (TID.isNotDuplicable())
554 BBI.CannotBeCopied = true;
556 bool isPredicated = TII->isPredicated(I);
557 bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch();
562 else if (!AlreadyPredicated) {
563 // FIXME: This instruction is already predicated before the
564 // if-conversion pass. It's probably something like a conditional move.
565 // Mark this block unpredicable for now.
566 BBI.IsUnpredicable = true;
572 if (BBI.ClobbersPred && !isPredicated) {
573 // Predicate modification instruction should end the block (except for
574 // already predicated instructions and end of block branches).
578 // Conditional branches is not predicable. But it may be eliminated.
582 // Predicate may have been modified, the subsequent (currently)
583 // unpredicated instructions cannot be correctly predicated.
584 BBI.IsUnpredicable = true;
588 // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are
589 // still potentially predicable.
590 std::vector<MachineOperand> PredDefs;
591 if (TII->DefinesPredicate(I, PredDefs))
592 BBI.ClobbersPred = true;
594 if (!TID.isPredicable()) {
595 BBI.IsUnpredicable = true;
601 /// FeasibilityAnalysis - Determine if the block is a suitable candidate to be
602 /// predicated by the specified predicate.
603 bool IfConverter::FeasibilityAnalysis(BBInfo &BBI,
604 std::vector<MachineOperand> &Pred,
605 bool isTriangle, bool RevBranch) {
606 // If the block is dead or unpredicable, then it cannot be predicated.
607 if (BBI.IsDone || BBI.IsUnpredicable)
610 // If it is already predicated, check if its predicate subsumes the new
612 if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred))
615 if (BBI.BrCond.size()) {
619 // Test predicate subsumsion.
620 std::vector<MachineOperand> RevPred(Pred);
621 std::vector<MachineOperand> Cond(BBI.BrCond);
623 if (TII->ReverseBranchCondition(Cond))
626 if (TII->ReverseBranchCondition(RevPred) ||
627 !TII->SubsumesPredicate(Cond, RevPred))
634 /// AnalyzeBlock - Analyze the structure of the sub-CFG starting from
635 /// the specified block. Record its successors and whether it looks like an
636 /// if-conversion candidate.
637 IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
638 std::vector<IfcvtToken*> &Tokens) {
639 BBInfo &BBI = BBAnalysis[BB->getNumber()];
641 if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed)
645 BBI.IsBeingAnalyzed = true;
647 ScanInstructions(BBI);
649 // Unanalyable or ends with fallthrough or unconditional branch.
650 if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) {
651 BBI.IsBeingAnalyzed = false;
652 BBI.IsAnalyzed = true;
656 // Do not ifcvt if either path is a back edge to the entry block.
657 if (BBI.TrueBB == BB || BBI.FalseBB == BB) {
658 BBI.IsBeingAnalyzed = false;
659 BBI.IsAnalyzed = true;
663 BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens);
664 BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens);
666 if (TrueBBI.IsDone && FalseBBI.IsDone) {
667 BBI.IsBeingAnalyzed = false;
668 BBI.IsAnalyzed = true;
672 std::vector<MachineOperand> RevCond(BBI.BrCond);
673 bool CanRevCond = !TII->ReverseBranchCondition(RevCond);
677 bool TNeedSub = TrueBBI.Predicate.size() > 0;
678 bool FNeedSub = FalseBBI.Predicate.size() > 0;
679 bool Enqueued = false;
680 if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) &&
681 MeetIfcvtSizeLimit(TrueBBI.NonPredSize - (Dups + Dups2)) &&
682 MeetIfcvtSizeLimit(FalseBBI.NonPredSize - (Dups + Dups2)) &&
683 FeasibilityAnalysis(TrueBBI, BBI.BrCond) &&
684 FeasibilityAnalysis(FalseBBI, RevCond)) {
692 // Note TailBB can be empty.
693 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups,
698 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) &&
699 MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
700 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
708 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups));
712 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups) &&
713 MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
714 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) {
715 Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups));
719 if (ValidSimple(TrueBBI, Dups) &&
720 MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
721 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) {
722 // Simple (split, no rejoin):
729 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups));
734 // Try the other path...
735 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups) &&
736 MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
737 FeasibilityAnalysis(FalseBBI, RevCond, true)) {
738 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups));
742 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups) &&
743 MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
744 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) {
745 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups));
749 if (ValidSimple(FalseBBI, Dups) &&
750 MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
751 FeasibilityAnalysis(FalseBBI, RevCond)) {
752 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups));
757 BBI.IsEnqueued = Enqueued;
758 BBI.IsBeingAnalyzed = false;
759 BBI.IsAnalyzed = true;
763 /// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion
764 /// candidates. It returns true if any CFG restructuring is done to expose more
765 /// if-conversion opportunities.
766 bool IfConverter::AnalyzeBlocks(MachineFunction &MF,
767 std::vector<IfcvtToken*> &Tokens) {
769 std::set<MachineBasicBlock*> Visited;
770 for (unsigned i = 0, e = Roots.size(); i != e; ++i) {
771 for (idf_ext_iterator<MachineBasicBlock*> I=idf_ext_begin(Roots[i],Visited),
772 E = idf_ext_end(Roots[i], Visited); I != E; ++I) {
773 MachineBasicBlock *BB = *I;
774 AnalyzeBlock(BB, Tokens);
778 // Sort to favor more complex ifcvt scheme.
779 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp);
784 /// canFallThroughTo - Returns true either if ToBB is the next block after BB or
785 /// that all the intervening blocks are empty (given BB can fall through to its
787 static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) {
788 MachineFunction::iterator I = BB;
789 MachineFunction::iterator TI = ToBB;
790 MachineFunction::iterator E = BB->getParent()->end();
792 if (I == E || !I->empty())
797 /// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed
798 /// to determine if it can be if-converted. If predecessor is already enqueued,
800 void IfConverter::InvalidatePreds(MachineBasicBlock *BB) {
801 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(),
802 E = BB->pred_end(); PI != E; ++PI) {
803 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()];
804 if (PBBI.IsDone || PBBI.BB == BB)
806 PBBI.IsAnalyzed = false;
807 PBBI.IsEnqueued = false;
811 /// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB.
813 static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB,
814 const TargetInstrInfo *TII) {
815 std::vector<MachineOperand> NoCond;
816 TII->InsertBranch(*BB, ToBB, NULL, NoCond);
819 /// RemoveExtraEdges - Remove true / false edges if either / both are no longer
821 void IfConverter::RemoveExtraEdges(BBInfo &BBI) {
822 MachineBasicBlock *TBB = NULL, *FBB = NULL;
823 std::vector<MachineOperand> Cond;
824 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond))
825 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
828 /// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG.
830 bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) {
831 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
832 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
833 BBInfo *CvtBBI = &TrueBBI;
834 BBInfo *NextBBI = &FalseBBI;
836 std::vector<MachineOperand> Cond(BBI.BrCond);
837 if (Kind == ICSimpleFalse)
838 std::swap(CvtBBI, NextBBI);
840 if (CvtBBI->IsDone ||
841 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
842 // Something has changed. It's no longer safe to predicate this block.
843 BBI.IsAnalyzed = false;
844 CvtBBI->IsAnalyzed = false;
848 if (Kind == ICSimpleFalse)
849 TII->ReverseBranchCondition(Cond);
851 if (CvtBBI->BB->pred_size() > 1) {
852 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
853 // Copy instructions in the true block, predicate them add them to
855 CopyAndPredicateBlock(BBI, *CvtBBI, Cond);
857 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond);
859 // Merge converted block into entry block.
860 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
861 MergeBlocks(BBI, *CvtBBI);
864 bool IterIfcvt = true;
865 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) {
866 InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
867 BBI.HasFallThrough = false;
868 // Now ifcvt'd block will look like this:
875 // We cannot further ifcvt this block because the unconditional branch
876 // will have to be predicated on the new condition, that will not be
877 // available if cmp executes.
881 RemoveExtraEdges(BBI);
883 // Update block info. BB can be iteratively if-converted.
886 InvalidatePreds(BBI.BB);
887 CvtBBI->IsDone = true;
889 // FIXME: Must maintain LiveIns.
893 /// IfConvertTriangle - If convert a triangle sub-CFG.
895 bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
896 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
897 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
898 BBInfo *CvtBBI = &TrueBBI;
899 BBInfo *NextBBI = &FalseBBI;
901 std::vector<MachineOperand> Cond(BBI.BrCond);
902 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
903 std::swap(CvtBBI, NextBBI);
905 if (CvtBBI->IsDone ||
906 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
907 // Something has changed. It's no longer safe to predicate this block.
908 BBI.IsAnalyzed = false;
909 CvtBBI->IsAnalyzed = false;
913 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
914 TII->ReverseBranchCondition(Cond);
916 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) {
917 ReverseBranchCondition(*CvtBBI);
918 // BB has been changed, modify its predecessors (except for this
919 // one) so they don't get ifcvt'ed based on bad intel.
920 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(),
921 E = CvtBBI->BB->pred_end(); PI != E; ++PI) {
922 MachineBasicBlock *PBB = *PI;
925 BBInfo &PBBI = BBAnalysis[PBB->getNumber()];
926 if (PBBI.IsEnqueued) {
927 PBBI.IsAnalyzed = false;
928 PBBI.IsEnqueued = false;
933 bool HasEarlyExit = CvtBBI->FalseBB != NULL;
934 bool DupBB = CvtBBI->BB->pred_size() > 1;
936 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
937 // Copy instructions in the true block, predicate them add them to
939 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, true);
941 // Predicate the 'true' block after removing its branch.
942 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB);
943 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond);
947 // Now merge the entry of the triangle with the true block.
948 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
949 MergeBlocks(BBI, *CvtBBI);
952 // If 'true' block has a 'false' successor, add an exit branch to it.
954 std::vector<MachineOperand> RevCond(CvtBBI->BrCond);
955 if (TII->ReverseBranchCondition(RevCond))
956 assert(false && "Unable to reverse branch condition!");
957 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond);
958 BBI.BB->addSuccessor(CvtBBI->FalseBB);
961 // Merge in the 'false' block if the 'false' block has no other
962 // predecessors. Otherwise, add a unconditional branch from to 'false'.
963 bool FalseBBDead = false;
964 bool IterIfcvt = true;
965 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB);
966 if (!isFallThrough) {
967 // Only merge them if the true block does not fallthrough to the false
968 // block. By not merging them, we make it possible to iteratively
971 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) {
972 MergeBlocks(BBI, *NextBBI);
975 InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
976 BBI.HasFallThrough = false;
978 // Mixed predicated and unpredicated code. This cannot be iteratively
983 RemoveExtraEdges(BBI);
985 // Update block info. BB can be iteratively if-converted.
988 InvalidatePreds(BBI.BB);
989 CvtBBI->IsDone = true;
991 NextBBI->IsDone = true;
993 // FIXME: Must maintain LiveIns.
997 /// IfConvertDiamond - If convert a diamond sub-CFG.
999 bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
1000 unsigned NumDups1, unsigned NumDups2) {
1001 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
1002 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1003 MachineBasicBlock *TailBB = TrueBBI.TrueBB;
1004 // True block must fall through or ended with unanalyzable terminator.
1006 if (blockAlwaysFallThrough(TrueBBI))
1007 TailBB = FalseBBI.TrueBB;
1008 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!");
1011 if (TrueBBI.IsDone || FalseBBI.IsDone ||
1012 TrueBBI.BB->pred_size() > 1 ||
1013 FalseBBI.BB->pred_size() > 1) {
1014 // Something has changed. It's no longer safe to predicate these blocks.
1015 BBI.IsAnalyzed = false;
1016 TrueBBI.IsAnalyzed = false;
1017 FalseBBI.IsAnalyzed = false;
1021 // Merge the 'true' and 'false' blocks by copying the instructions
1022 // from the 'false' block to the 'true' block. That is, unless the true
1023 // block would clobber the predicate, in that case, do the opposite.
1024 BBInfo *BBI1 = &TrueBBI;
1025 BBInfo *BBI2 = &FalseBBI;
1026 std::vector<MachineOperand> RevCond(BBI.BrCond);
1027 TII->ReverseBranchCondition(RevCond);
1028 std::vector<MachineOperand> *Cond1 = &BBI.BrCond;
1029 std::vector<MachineOperand> *Cond2 = &RevCond;
1031 // Figure out the more profitable ordering.
1032 bool DoSwap = false;
1033 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred)
1035 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) {
1036 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize)
1040 std::swap(BBI1, BBI2);
1041 std::swap(Cond1, Cond2);
1044 // Remove the conditional branch from entry to the blocks.
1045 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1047 // Remove the duplicated instructions at the beginnings of both paths.
1048 MachineBasicBlock::iterator DI1 = BBI1->BB->begin();
1049 MachineBasicBlock::iterator DI2 = BBI2->BB->begin();
1050 BBI1->NonPredSize -= NumDups1;
1051 BBI2->NonPredSize -= NumDups1;
1052 while (NumDups1 != 0) {
1057 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1);
1058 BBI2->BB->erase(BBI2->BB->begin(), DI2);
1060 // Predicate the 'true' block after removing its branch.
1061 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB);
1062 DI1 = BBI1->BB->end();
1063 for (unsigned i = 0; i != NumDups2; ++i)
1065 BBI1->BB->erase(DI1, BBI1->BB->end());
1066 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1);
1068 // Predicate the 'false' block.
1069 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB);
1070 DI2 = BBI2->BB->end();
1071 while (NumDups2 != 0) {
1075 PredicateBlock(*BBI2, DI2, *Cond2);
1077 // Merge the true block into the entry of the diamond.
1078 MergeBlocks(BBI, *BBI1);
1079 MergeBlocks(BBI, *BBI2);
1081 // If the if-converted block fallthrough or unconditionally branch into the
1082 // tail block, and the tail block does not have other predecessors, then
1083 // fold the tail block in as well. Otherwise, unless it falls through to the
1084 // tail, add a unconditional branch to it.
1086 BBInfo TailBBI = BBAnalysis[TailBB->getNumber()];
1087 if (TailBB->pred_size() == 1 && !TailBBI.HasFallThrough) {
1088 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1089 MergeBlocks(BBI, TailBBI);
1090 TailBBI.IsDone = true;
1092 InsertUncondBranch(BBI.BB, TailBB, TII);
1093 BBI.HasFallThrough = false;
1097 RemoveExtraEdges(BBI);
1099 // Update block info.
1100 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true;
1101 InvalidatePreds(BBI.BB);
1103 // FIXME: Must maintain LiveIns.
1107 /// PredicateBlock - Predicate instructions from the start of the block to the
1108 /// specified end with the specified condition.
1109 void IfConverter::PredicateBlock(BBInfo &BBI,
1110 MachineBasicBlock::iterator E,
1111 std::vector<MachineOperand> &Cond) {
1112 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) {
1113 if (TII->isPredicated(I))
1115 if (!TII->PredicateInstruction(I, Cond)) {
1116 cerr << "Unable to predicate " << *I << "!\n";
1121 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate));
1123 BBI.IsAnalyzed = false;
1124 BBI.NonPredSize = 0;
1129 /// CopyAndPredicateBlock - Copy and predicate instructions from source BB to
1130 /// the destination block. Skip end of block branches if IgnoreBr is true.
1131 void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
1132 std::vector<MachineOperand> &Cond,
1134 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(),
1135 E = FromBBI.BB->end(); I != E; ++I) {
1136 const TargetInstrDesc &TID = I->getDesc();
1137 bool isPredicated = TII->isPredicated(I);
1138 // Do not copy the end of the block branches.
1139 if (IgnoreBr && !isPredicated && TID.isBranch())
1142 MachineInstr *MI = I->clone();
1143 ToBBI.BB->insert(ToBBI.BB->end(), MI);
1144 ToBBI.NonPredSize++;
1147 if (!TII->PredicateInstruction(MI, Cond)) {
1148 cerr << "Unable to predicate " << *MI << "!\n";
1153 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1154 FromBBI.BB->succ_end());
1155 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1156 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1158 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1159 MachineBasicBlock *Succ = Succs[i];
1160 // Fallthrough edge can't be transferred.
1161 if (Succ == FallThrough)
1163 if (!ToBBI.BB->isSuccessor(Succ))
1164 ToBBI.BB->addSuccessor(Succ);
1167 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1168 std::back_inserter(ToBBI.Predicate));
1169 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate));
1171 ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1172 ToBBI.IsAnalyzed = false;
1177 /// MergeBlocks - Move all instructions from FromBB to the end of ToBB.
1179 void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) {
1180 ToBBI.BB->splice(ToBBI.BB->end(),
1181 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end());
1183 // Redirect all branches to FromBB to ToBB.
1184 std::vector<MachineBasicBlock *> Preds(FromBBI.BB->pred_begin(),
1185 FromBBI.BB->pred_end());
1186 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
1187 MachineBasicBlock *Pred = Preds[i];
1188 if (Pred == ToBBI.BB)
1190 Pred->ReplaceUsesOfBlockWith(FromBBI.BB, ToBBI.BB);
1193 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1194 FromBBI.BB->succ_end());
1195 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1196 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1198 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1199 MachineBasicBlock *Succ = Succs[i];
1200 // Fallthrough edge can't be transferred.
1201 if (Succ == FallThrough)
1203 FromBBI.BB->removeSuccessor(Succ);
1204 if (!ToBBI.BB->isSuccessor(Succ))
1205 ToBBI.BB->addSuccessor(Succ);
1208 // Now FromBBI always fall through to the next block!
1209 if (NBB && !FromBBI.BB->isSuccessor(NBB))
1210 FromBBI.BB->addSuccessor(NBB);
1212 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1213 std::back_inserter(ToBBI.Predicate));
1214 FromBBI.Predicate.clear();
1216 ToBBI.NonPredSize += FromBBI.NonPredSize;
1217 FromBBI.NonPredSize = 0;
1219 ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1220 ToBBI.HasFallThrough = FromBBI.HasFallThrough;
1221 ToBBI.IsAnalyzed = false;
1222 FromBBI.IsAnalyzed = false;