IfConversion.cpp revision 081c34b725980f995be9080eaec24cd3dfaaf065 
1//===-- IfConversion.cpp - Machine code if conversion pass. ---------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file implements the machine instruction level if-conversion pass.
11//
12//===----------------------------------------------------------------------===//
13
14#define DEBUG_TYPE "ifcvt"
15#include "BranchFolding.h"
16#include "llvm/Function.h"
17#include "llvm/CodeGen/Passes.h"
18#include "llvm/CodeGen/MachineModuleInfo.h"
19#include "llvm/CodeGen/MachineFunctionPass.h"
20#include "llvm/CodeGen/MachineLoopInfo.h"
21#include "llvm/Target/TargetInstrInfo.h"
22#include "llvm/Target/TargetInstrItineraries.h"
23#include "llvm/Target/TargetLowering.h"
24#include "llvm/Target/TargetMachine.h"
25#include "llvm/Target/TargetRegisterInfo.h"
26#include "llvm/Support/CommandLine.h"
27#include "llvm/Support/Debug.h"
28#include "llvm/Support/ErrorHandling.h"
29#include "llvm/Support/raw_ostream.h"
30#include "llvm/ADT/DepthFirstIterator.h"
31#include "llvm/ADT/Statistic.h"
32#include "llvm/ADT/STLExtras.h"
33using namespace llvm;
34
35// Hidden options for help debugging.
36static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden);
37static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden);
38static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden);
39static cl::opt<bool> DisableSimple("disable-ifcvt-simple",
40                                   cl::init(false), cl::Hidden);
41static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false",
42                                    cl::init(false), cl::Hidden);
43static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle",
44                                     cl::init(false), cl::Hidden);
45static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev",
46                                      cl::init(false), cl::Hidden);
47static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false",
48                                      cl::init(false), cl::Hidden);
49static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev",
50                                       cl::init(false), cl::Hidden);
51static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond",
52                                    cl::init(false), cl::Hidden);
53static cl::opt<bool> IfCvtBranchFold("ifcvt-branch-fold",
54                                     cl::init(true), cl::Hidden);
55
56STATISTIC(NumSimple,       "Number of simple if-conversions performed");
57STATISTIC(NumSimpleFalse,  "Number of simple (F) if-conversions performed");
58STATISTIC(NumTriangle,     "Number of triangle if-conversions performed");
59STATISTIC(NumTriangleRev,  "Number of triangle (R) if-conversions performed");
60STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed");
61STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed");
62STATISTIC(NumDiamonds,     "Number of diamond if-conversions performed");
63STATISTIC(NumIfConvBBs,    "Number of if-converted blocks");
64STATISTIC(NumDupBBs,       "Number of duplicated blocks");
65
66namespace {
67  class IfConverter : public MachineFunctionPass {
68    enum IfcvtKind {
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.
77    };
78
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.
85    ///
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
94    ///                   cmp, call, etc.)
95    /// NonPredSize     - Number of non-predicated instructions.
96    /// BB              - Corresponding MachineBasicBlock.
97    /// TrueBB / FalseBB- See AnalyzeBranch().
98    /// BrCond          - Conditions for end of block conditional branches.
99    /// Predicate       - Predicate used in the BB.
100    struct BBInfo {
101      bool IsDone          : 1;
102      bool IsBeingAnalyzed : 1;
103      bool IsAnalyzed      : 1;
104      bool IsEnqueued      : 1;
105      bool IsBrAnalyzable  : 1;
106      bool HasFallThrough  : 1;
107      bool IsUnpredicable  : 1;
108      bool CannotBeCopied  : 1;
109      bool ClobbersPred    : 1;
110      unsigned NonPredSize;
111      MachineBasicBlock *BB;
112      MachineBasicBlock *TrueBB;
113      MachineBasicBlock *FalseBB;
114      SmallVector<MachineOperand, 4> BrCond;
115      SmallVector<MachineOperand, 4> Predicate;
116      BBInfo() : IsDone(false), IsBeingAnalyzed(false),
117                 IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false),
118                 HasFallThrough(false), IsUnpredicable(false),
119                 CannotBeCopied(false), ClobbersPred(false), NonPredSize(0),
120                 BB(0), TrueBB(0), FalseBB(0) {}
121    };
122
123    /// IfcvtToken - Record information about pending if-conversions to attempt:
124    /// BBI             - Corresponding BBInfo.
125    /// Kind            - Type of block. See IfcvtKind.
126    /// NeedSubsumption - True if the to-be-predicated BB has already been
127    ///                   predicated.
128    /// NumDups      - Number of instructions that would be duplicated due
129    ///                   to this if-conversion. (For diamonds, the number of
130    ///                   identical instructions at the beginnings of both
131    ///                   paths).
132    /// NumDups2     - For diamonds, the number of identical instructions
133    ///                   at the ends of both paths.
134    struct IfcvtToken {
135      BBInfo &BBI;
136      IfcvtKind Kind;
137      bool NeedSubsumption;
138      unsigned NumDups;
139      unsigned NumDups2;
140      IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0)
141        : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {}
142    };
143
144    /// Roots - Basic blocks that do not have successors. These are the starting
145    /// points of Graph traversal.
146    std::vector<MachineBasicBlock*> Roots;
147
148    /// BBAnalysis - Results of if-conversion feasibility analysis indexed by
149    /// basic block number.
150    std::vector<BBInfo> BBAnalysis;
151
152    const TargetLowering *TLI;
153    const TargetInstrInfo *TII;
154    const TargetRegisterInfo *TRI;
155    const InstrItineraryData *InstrItins;
156    const MachineLoopInfo *MLI;
157    bool MadeChange;
158    int FnNum;
159  public:
160    static char ID;
161    IfConverter() : MachineFunctionPass(ID), FnNum(-1) {
162      initializeIfConverterPass(*PassRegistry::getPassRegistry());
163    }
164
165    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
166      AU.addRequired<MachineLoopInfo>();
167      MachineFunctionPass::getAnalysisUsage(AU);
168    }
169
170    virtual bool runOnMachineFunction(MachineFunction &MF);
171    virtual const char *getPassName() const { return "If Converter"; }
172
173  private:
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);
203
204    bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, unsigned Size,
205                            float Prediction, float Confidence) const {
206      return Size > 0 && TII->isProfitableToIfCvt(BB, Size,
207                                                  Prediction, Confidence);
208    }
209
210    bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, unsigned TSize,
211                            MachineBasicBlock &FBB, unsigned FSize,
212                            float Prediction, float Confidence) const {
213      return TSize > 0 && FSize > 0 &&
214        TII->isProfitableToIfCvt(TBB, TSize, FBB, FSize,
215                                 Prediction, Confidence);
216    }
217
218    // blockAlwaysFallThrough - Block ends without a terminator.
219    bool blockAlwaysFallThrough(BBInfo &BBI) const {
220      return BBI.IsBrAnalyzable && BBI.TrueBB == NULL;
221    }
222
223    // IfcvtTokenCmp - Used to sort if-conversion candidates.
224    static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) {
225      int Incr1 = (C1->Kind == ICDiamond)
226        ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups;
227      int Incr2 = (C2->Kind == ICDiamond)
228        ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups;
229      if (Incr1 > Incr2)
230        return true;
231      else if (Incr1 == Incr2) {
232        // Favors subsumption.
233        if (C1->NeedSubsumption == false && C2->NeedSubsumption == true)
234          return true;
235        else if (C1->NeedSubsumption == C2->NeedSubsumption) {
236          // Favors diamond over triangle, etc.
237          if ((unsigned)C1->Kind < (unsigned)C2->Kind)
238            return true;
239          else if (C1->Kind == C2->Kind)
240            return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber();
241        }
242      }
243      return false;
244    }
245  };
246
247  char IfConverter::ID = 0;
248}
249
250INITIALIZE_PASS_BEGIN(IfConverter, "if-converter", "If Converter", false, false)
251INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
252INITIALIZE_PASS_END(IfConverter, "if-converter", "If Converter", false, false)
253
254FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); }
255
256bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
257  TLI = MF.getTarget().getTargetLowering();
258  TII = MF.getTarget().getInstrInfo();
259  TRI = MF.getTarget().getRegisterInfo();
260  MLI = &getAnalysis<MachineLoopInfo>();
261  InstrItins = MF.getTarget().getInstrItineraryData();
262  if (!TII) return false;
263
264  // Tail merge tend to expose more if-conversion opportunities.
265  BranchFolder BF(true);
266  bool BFChange = BF.OptimizeFunction(MF, TII,
267                                   MF.getTarget().getRegisterInfo(),
268                                   getAnalysisIfAvailable<MachineModuleInfo>());
269
270  DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum <<  ") \'"
271               << MF.getFunction()->getName() << "\'");
272
273  if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) {
274    DEBUG(dbgs() << " skipped\n");
275    return false;
276  }
277  DEBUG(dbgs() << "\n");
278
279  MF.RenumberBlocks();
280  BBAnalysis.resize(MF.getNumBlockIDs());
281
282  // Look for root nodes, i.e. blocks without successors.
283  for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
284    if (I->succ_empty())
285      Roots.push_back(I);
286
287  std::vector<IfcvtToken*> Tokens;
288  MadeChange = false;
289  unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle +
290    NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds;
291  while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) {
292    // Do an initial analysis for each basic block and find all the potential
293    // candidates to perform if-conversion.
294    bool Change = false;
295    AnalyzeBlocks(MF, Tokens);
296    while (!Tokens.empty()) {
297      IfcvtToken *Token = Tokens.back();
298      Tokens.pop_back();
299      BBInfo &BBI = Token->BBI;
300      IfcvtKind Kind = Token->Kind;
301      unsigned NumDups = Token->NumDups;
302      unsigned NumDups2 = Token->NumDups2;
303
304      delete Token;
305
306      // If the block has been evicted out of the queue or it has already been
307      // marked dead (due to it being predicated), then skip it.
308      if (BBI.IsDone)
309        BBI.IsEnqueued = false;
310      if (!BBI.IsEnqueued)
311        continue;
312
313      BBI.IsEnqueued = false;
314
315      bool RetVal = false;
316      switch (Kind) {
317      default: assert(false && "Unexpected!");
318        break;
319      case ICSimple:
320      case ICSimpleFalse: {
321        bool isFalse = Kind == ICSimpleFalse;
322        if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break;
323        DEBUG(dbgs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ?
324                                            " false" : "")
325                     << "): BB#" << BBI.BB->getNumber() << " ("
326                     << ((Kind == ICSimpleFalse)
327                         ? BBI.FalseBB->getNumber()
328                         : BBI.TrueBB->getNumber()) << ") ");
329        RetVal = IfConvertSimple(BBI, Kind);
330        DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
331        if (RetVal) {
332          if (isFalse) ++NumSimpleFalse;
333          else         ++NumSimple;
334        }
335       break;
336      }
337      case ICTriangle:
338      case ICTriangleRev:
339      case ICTriangleFalse:
340      case ICTriangleFRev: {
341        bool isFalse = Kind == ICTriangleFalse;
342        bool isRev   = (Kind == ICTriangleRev || Kind == ICTriangleFRev);
343        if (DisableTriangle && !isFalse && !isRev) break;
344        if (DisableTriangleR && !isFalse && isRev) break;
345        if (DisableTriangleF && isFalse && !isRev) break;
346        if (DisableTriangleFR && isFalse && isRev) break;
347        DEBUG(dbgs() << "Ifcvt (Triangle");
348        if (isFalse)
349          DEBUG(dbgs() << " false");
350        if (isRev)
351          DEBUG(dbgs() << " rev");
352        DEBUG(dbgs() << "): BB#" << BBI.BB->getNumber() << " (T:"
353                     << BBI.TrueBB->getNumber() << ",F:"
354                     << BBI.FalseBB->getNumber() << ") ");
355        RetVal = IfConvertTriangle(BBI, Kind);
356        DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
357        if (RetVal) {
358          if (isFalse) {
359            if (isRev) ++NumTriangleFRev;
360            else       ++NumTriangleFalse;
361          } else {
362            if (isRev) ++NumTriangleRev;
363            else       ++NumTriangle;
364          }
365        }
366        break;
367      }
368      case ICDiamond: {
369        if (DisableDiamond) break;
370        DEBUG(dbgs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:"
371                     << BBI.TrueBB->getNumber() << ",F:"
372                     << BBI.FalseBB->getNumber() << ") ");
373        RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2);
374        DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
375        if (RetVal) ++NumDiamonds;
376        break;
377      }
378      }
379
380      Change |= RetVal;
381
382      NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev +
383        NumTriangleFalse + NumTriangleFRev + NumDiamonds;
384      if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit)
385        break;
386    }
387
388    if (!Change)
389      break;
390    MadeChange |= Change;
391  }
392
393  // Delete tokens in case of early exit.
394  while (!Tokens.empty()) {
395    IfcvtToken *Token = Tokens.back();
396    Tokens.pop_back();
397    delete Token;
398  }
399
400  Tokens.clear();
401  Roots.clear();
402  BBAnalysis.clear();
403
404  if (MadeChange && IfCvtBranchFold) {
405    BranchFolder BF(false);
406    BF.OptimizeFunction(MF, TII,
407                        MF.getTarget().getRegisterInfo(),
408                        getAnalysisIfAvailable<MachineModuleInfo>());
409  }
410
411  MadeChange |= BFChange;
412  return MadeChange;
413}
414
415/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
416/// its 'true' successor.
417static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
418                                         MachineBasicBlock *TrueBB) {
419  for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
420         E = BB->succ_end(); SI != E; ++SI) {
421    MachineBasicBlock *SuccBB = *SI;
422    if (SuccBB != TrueBB)
423      return SuccBB;
424  }
425  return NULL;
426}
427
428/// ReverseBranchCondition - Reverse the condition of the end of the block
429/// branch. Swap block's 'true' and 'false' successors.
430bool IfConverter::ReverseBranchCondition(BBInfo &BBI) {
431  DebugLoc dl;  // FIXME: this is nowhere
432  if (!TII->ReverseBranchCondition(BBI.BrCond)) {
433    TII->RemoveBranch(*BBI.BB);
434    TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond, dl);
435    std::swap(BBI.TrueBB, BBI.FalseBB);
436    return true;
437  }
438  return false;
439}
440
441/// getNextBlock - Returns the next block in the function blocks ordering. If
442/// it is the end, returns NULL.
443static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) {
444  MachineFunction::iterator I = BB;
445  MachineFunction::iterator E = BB->getParent()->end();
446  if (++I == E)
447    return NULL;
448  return I;
449}
450
451/// ValidSimple - Returns true if the 'true' block (along with its
452/// predecessor) forms a valid simple shape for ifcvt. It also returns the
453/// number of instructions that the ifcvt would need to duplicate if performed
454/// in Dups.
455bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups,
456                              float Prediction, float Confidence) const {
457  Dups = 0;
458  if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
459    return false;
460
461  if (TrueBBI.IsBrAnalyzable)
462    return false;
463
464  if (TrueBBI.BB->pred_size() > 1) {
465    if (TrueBBI.CannotBeCopied ||
466        !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize,
467                                        Prediction, Confidence))
468      return false;
469    Dups = TrueBBI.NonPredSize;
470  }
471
472  return true;
473}
474
475/// ValidTriangle - Returns true if the 'true' and 'false' blocks (along
476/// with their common predecessor) forms a valid triangle shape for ifcvt.
477/// If 'FalseBranch' is true, it checks if 'true' block's false branch
478/// branches to the 'false' block rather than the other way around. It also
479/// returns the number of instructions that the ifcvt would need to duplicate
480/// if performed in 'Dups'.
481bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
482                                bool FalseBranch, unsigned &Dups,
483                                float Prediction, float Confidence) const {
484  Dups = 0;
485  if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
486    return false;
487
488  if (TrueBBI.BB->pred_size() > 1) {
489    if (TrueBBI.CannotBeCopied)
490      return false;
491
492    unsigned Size = TrueBBI.NonPredSize;
493    if (TrueBBI.IsBrAnalyzable) {
494      if (TrueBBI.TrueBB && TrueBBI.BrCond.empty())
495        // Ends with an unconditional branch. It will be removed.
496        --Size;
497      else {
498        MachineBasicBlock *FExit = FalseBranch
499          ? TrueBBI.TrueBB : TrueBBI.FalseBB;
500        if (FExit)
501          // Require a conditional branch
502          ++Size;
503      }
504    }
505    if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size,
506                                        Prediction, Confidence))
507      return false;
508    Dups = Size;
509  }
510
511  MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB;
512  if (!TExit && blockAlwaysFallThrough(TrueBBI)) {
513    MachineFunction::iterator I = TrueBBI.BB;
514    if (++I == TrueBBI.BB->getParent()->end())
515      return false;
516    TExit = I;
517  }
518  return TExit && TExit == FalseBBI.BB;
519}
520
521static
522MachineBasicBlock::iterator firstNonBranchInst(MachineBasicBlock *BB,
523                                               const TargetInstrInfo *TII) {
524  MachineBasicBlock::iterator I = BB->end();
525  while (I != BB->begin()) {
526    --I;
527    if (!I->getDesc().isBranch())
528      break;
529  }
530  return I;
531}
532
533/// ValidDiamond - Returns true if the 'true' and 'false' blocks (along
534/// with their common predecessor) forms a valid diamond shape for ifcvt.
535bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
536                               unsigned &Dups1, unsigned &Dups2) const {
537  Dups1 = Dups2 = 0;
538  if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone ||
539      FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone)
540    return false;
541
542  MachineBasicBlock *TT = TrueBBI.TrueBB;
543  MachineBasicBlock *FT = FalseBBI.TrueBB;
544
545  if (!TT && blockAlwaysFallThrough(TrueBBI))
546    TT = getNextBlock(TrueBBI.BB);
547  if (!FT && blockAlwaysFallThrough(FalseBBI))
548    FT = getNextBlock(FalseBBI.BB);
549  if (TT != FT)
550    return false;
551  if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable))
552    return false;
553  if  (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1)
554    return false;
555
556  // FIXME: Allow true block to have an early exit?
557  if (TrueBBI.FalseBB || FalseBBI.FalseBB ||
558      (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred))
559    return false;
560
561  MachineBasicBlock::iterator TI = TrueBBI.BB->begin();
562  MachineBasicBlock::iterator FI = FalseBBI.BB->begin();
563  MachineBasicBlock::iterator TIE = TrueBBI.BB->end();
564  MachineBasicBlock::iterator FIE = FalseBBI.BB->end();
565  // Skip dbg_value instructions
566  while (TI != TIE && TI->isDebugValue())
567    ++TI;
568  while (FI != FIE && FI->isDebugValue())
569    ++FI;
570  while (TI != TIE && FI != FIE) {
571    // Skip dbg_value instructions. These do not count.
572    if (TI->isDebugValue()) {
573      while (TI != TIE && TI->isDebugValue())
574        ++TI;
575      if (TI == TIE)
576        break;
577    }
578    if (FI->isDebugValue()) {
579      while (FI != FIE && FI->isDebugValue())
580        ++FI;
581      if (FI == FIE)
582        break;
583    }
584    if (!TI->isIdenticalTo(FI))
585      break;
586    ++Dups1;
587    ++TI;
588    ++FI;
589  }
590
591  TI = firstNonBranchInst(TrueBBI.BB, TII);
592  FI = firstNonBranchInst(FalseBBI.BB, TII);
593  MachineBasicBlock::iterator TIB = TrueBBI.BB->begin();
594  MachineBasicBlock::iterator FIB = FalseBBI.BB->begin();
595  // Skip dbg_value instructions at end of the bb's.
596  while (TI != TIB && TI->isDebugValue())
597    --TI;
598  while (FI != FIB && FI->isDebugValue())
599    --FI;
600  while (TI != TIB && FI != FIB) {
601    // Skip dbg_value instructions. These do not count.
602    if (TI->isDebugValue()) {
603      while (TI != TIB && TI->isDebugValue())
604        --TI;
605      if (TI == TIB)
606        break;
607    }
608    if (FI->isDebugValue()) {
609      while (FI != FIB && FI->isDebugValue())
610        --FI;
611      if (FI == FIB)
612        break;
613    }
614    if (!TI->isIdenticalTo(FI))
615      break;
616    ++Dups2;
617    --TI;
618    --FI;
619  }
620
621  return true;
622}
623
624/// ScanInstructions - Scan all the instructions in the block to determine if
625/// the block is predicable. In most cases, that means all the instructions
626/// in the block are isPredicable(). Also checks if the block contains any
627/// instruction which can clobber a predicate (e.g. condition code register).
628/// If so, the block is not predicable unless it's the last instruction.
629void IfConverter::ScanInstructions(BBInfo &BBI) {
630  if (BBI.IsDone)
631    return;
632
633  bool AlreadyPredicated = BBI.Predicate.size() > 0;
634  // First analyze the end of BB branches.
635  BBI.TrueBB = BBI.FalseBB = NULL;
636  BBI.BrCond.clear();
637  BBI.IsBrAnalyzable =
638    !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond);
639  BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL;
640
641  if (BBI.BrCond.size()) {
642    // No false branch. This BB must end with a conditional branch and a
643    // fallthrough.
644    if (!BBI.FalseBB)
645      BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB);
646    if (!BBI.FalseBB) {
647      // Malformed bcc? True and false blocks are the same?
648      BBI.IsUnpredicable = true;
649      return;
650    }
651  }
652
653  // Then scan all the instructions.
654  BBI.NonPredSize = 0;
655  BBI.ClobbersPred = false;
656  for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end();
657       I != E; ++I) {
658    if (I->isDebugValue())
659      continue;
660
661    const TargetInstrDesc &TID = I->getDesc();
662    if (TID.isNotDuplicable())
663      BBI.CannotBeCopied = true;
664
665    bool isPredicated = TII->isPredicated(I);
666    bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch();
667
668    if (!isCondBr) {
669      if (!isPredicated) {
670        unsigned NumOps = TII->getNumMicroOps(&*I, InstrItins);
671        BBI.NonPredSize += NumOps;
672      } else if (!AlreadyPredicated) {
673        // FIXME: This instruction is already predicated before the
674        // if-conversion pass. It's probably something like a conditional move.
675        // Mark this block unpredicable for now.
676        BBI.IsUnpredicable = true;
677        return;
678      }
679    }
680
681    if (BBI.ClobbersPred && !isPredicated) {
682      // Predicate modification instruction should end the block (except for
683      // already predicated instructions and end of block branches).
684      if (isCondBr) {
685        // A conditional branch is not predicable, but it may be eliminated.
686        continue;
687      }
688
689      // Predicate may have been modified, the subsequent (currently)
690      // unpredicated instructions cannot be correctly predicated.
691      BBI.IsUnpredicable = true;
692      return;
693    }
694
695    // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are
696    // still potentially predicable.
697    std::vector<MachineOperand> PredDefs;
698    if (TII->DefinesPredicate(I, PredDefs))
699      BBI.ClobbersPred = true;
700
701    if (!TII->isPredicable(I)) {
702      BBI.IsUnpredicable = true;
703      return;
704    }
705  }
706}
707
708/// FeasibilityAnalysis - Determine if the block is a suitable candidate to be
709/// predicated by the specified predicate.
710bool IfConverter::FeasibilityAnalysis(BBInfo &BBI,
711                                      SmallVectorImpl<MachineOperand> &Pred,
712                                      bool isTriangle, bool RevBranch) {
713  // If the block is dead or unpredicable, then it cannot be predicated.
714  if (BBI.IsDone || BBI.IsUnpredicable)
715    return false;
716
717  // If it is already predicated, check if its predicate subsumes the new
718  // predicate.
719  if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred))
720    return false;
721
722  if (BBI.BrCond.size()) {
723    if (!isTriangle)
724      return false;
725
726    // Test predicate subsumption.
727    SmallVector<MachineOperand, 4> RevPred(Pred.begin(), Pred.end());
728    SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
729    if (RevBranch) {
730      if (TII->ReverseBranchCondition(Cond))
731        return false;
732    }
733    if (TII->ReverseBranchCondition(RevPred) ||
734        !TII->SubsumesPredicate(Cond, RevPred))
735      return false;
736  }
737
738  return true;
739}
740
741/// AnalyzeBlock - Analyze the structure of the sub-CFG starting from
742/// the specified block. Record its successors and whether it looks like an
743/// if-conversion candidate.
744IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
745                                             std::vector<IfcvtToken*> &Tokens) {
746  BBInfo &BBI = BBAnalysis[BB->getNumber()];
747
748  if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed)
749    return BBI;
750
751  BBI.BB = BB;
752  BBI.IsBeingAnalyzed = true;
753
754  ScanInstructions(BBI);
755
756  // Unanalyzable or ends with fallthrough or unconditional branch.
757  if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) {
758    BBI.IsBeingAnalyzed = false;
759    BBI.IsAnalyzed = true;
760    return BBI;
761  }
762
763  // Do not ifcvt if either path is a back edge to the entry block.
764  if (BBI.TrueBB == BB || BBI.FalseBB == BB) {
765    BBI.IsBeingAnalyzed = false;
766    BBI.IsAnalyzed = true;
767    return BBI;
768  }
769
770  // Do not ifcvt if true and false fallthrough blocks are the same.
771  if (!BBI.FalseBB) {
772    BBI.IsBeingAnalyzed = false;
773    BBI.IsAnalyzed = true;
774    return BBI;
775  }
776
777  BBInfo &TrueBBI  = AnalyzeBlock(BBI.TrueBB, Tokens);
778  BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens);
779
780  if (TrueBBI.IsDone && FalseBBI.IsDone) {
781    BBI.IsBeingAnalyzed = false;
782    BBI.IsAnalyzed = true;
783    return BBI;
784  }
785
786  SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end());
787  bool CanRevCond = !TII->ReverseBranchCondition(RevCond);
788
789  unsigned Dups = 0;
790  unsigned Dups2 = 0;
791  bool TNeedSub = TrueBBI.Predicate.size() > 0;
792  bool FNeedSub = FalseBBI.Predicate.size() > 0;
793  bool Enqueued = false;
794
795  // Try to predict the branch, using loop info to guide us.
796  // General heuristics are:
797  //   - backedge -> 90% taken
798  //   - early exit -> 20% taken
799  //   - branch predictor confidence -> 90%
800  float Prediction = 0.5f;
801  float Confidence = 0.9f;
802  MachineLoop *Loop = MLI->getLoopFor(BB);
803  if (Loop) {
804    if (TrueBBI.BB == Loop->getHeader())
805      Prediction = 0.9f;
806    else if (FalseBBI.BB == Loop->getHeader())
807      Prediction = 0.1f;
808
809    MachineLoop *TrueLoop = MLI->getLoopFor(TrueBBI.BB);
810    MachineLoop *FalseLoop = MLI->getLoopFor(FalseBBI.BB);
811    if (!TrueLoop || TrueLoop->getParentLoop() == Loop)
812      Prediction = 0.2f;
813    else if (!FalseLoop || FalseLoop->getParentLoop() == Loop)
814      Prediction = 0.8f;
815  }
816
817  if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) &&
818      MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize - (Dups + Dups2),
819                         *FalseBBI.BB, FalseBBI.NonPredSize - (Dups + Dups2),
820                         Prediction, Confidence) &&
821      FeasibilityAnalysis(TrueBBI, BBI.BrCond) &&
822      FeasibilityAnalysis(FalseBBI, RevCond)) {
823    // Diamond:
824    //   EBB
825    //   / \_
826    //  |   |
827    // TBB FBB
828    //   \ /
829    //  TailBB
830    // Note TailBB can be empty.
831    Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups,
832                                    Dups2));
833    Enqueued = true;
834  }
835
836  if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction, Confidence) &&
837      MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize,
838                         Prediction, Confidence) &&
839      FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
840    // Triangle:
841    //   EBB
842    //   | \_
843    //   |  |
844    //   | TBB
845    //   |  /
846    //   FBB
847    Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups));
848    Enqueued = true;
849  }
850
851  if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction, Confidence) &&
852      MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize,
853                         Prediction, Confidence) &&
854      FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) {
855    Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups));
856    Enqueued = true;
857  }
858
859  if (ValidSimple(TrueBBI, Dups, Prediction, Confidence) &&
860      MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize,
861                         Prediction, Confidence) &&
862      FeasibilityAnalysis(TrueBBI, BBI.BrCond)) {
863    // Simple (split, no rejoin):
864    //   EBB
865    //   | \_
866    //   |  |
867    //   | TBB---> exit
868    //   |
869    //   FBB
870    Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups));
871    Enqueued = true;
872  }
873
874  if (CanRevCond) {
875    // Try the other path...
876    if (ValidTriangle(FalseBBI, TrueBBI, false, Dups,
877                      1.0-Prediction, Confidence) &&
878        MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize,
879                           1.0-Prediction, Confidence) &&
880        FeasibilityAnalysis(FalseBBI, RevCond, true)) {
881      Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups));
882      Enqueued = true;
883    }
884
885    if (ValidTriangle(FalseBBI, TrueBBI, true, Dups,
886                      1.0-Prediction, Confidence) &&
887        MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize,
888                           1.0-Prediction, Confidence) &&
889        FeasibilityAnalysis(FalseBBI, RevCond, true, true)) {
890      Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups));
891      Enqueued = true;
892    }
893
894    if (ValidSimple(FalseBBI, Dups, 1.0-Prediction, Confidence) &&
895        MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize,
896                           1.0-Prediction, Confidence) &&
897        FeasibilityAnalysis(FalseBBI, RevCond)) {
898      Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups));
899      Enqueued = true;
900    }
901  }
902
903  BBI.IsEnqueued = Enqueued;
904  BBI.IsBeingAnalyzed = false;
905  BBI.IsAnalyzed = true;
906  return BBI;
907}
908
909/// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion
910/// candidates.
911void IfConverter::AnalyzeBlocks(MachineFunction &MF,
912                                std::vector<IfcvtToken*> &Tokens) {
913  std::set<MachineBasicBlock*> Visited;
914  for (unsigned i = 0, e = Roots.size(); i != e; ++i) {
915    for (idf_ext_iterator<MachineBasicBlock*> I=idf_ext_begin(Roots[i],Visited),
916           E = idf_ext_end(Roots[i], Visited); I != E; ++I) {
917      MachineBasicBlock *BB = *I;
918      AnalyzeBlock(BB, Tokens);
919    }
920  }
921
922  // Sort to favor more complex ifcvt scheme.
923  std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp);
924}
925
926/// canFallThroughTo - Returns true either if ToBB is the next block after BB or
927/// that all the intervening blocks are empty (given BB can fall through to its
928/// next block).
929static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) {
930  MachineFunction::iterator PI = BB;
931  MachineFunction::iterator I = llvm::next(PI);
932  MachineFunction::iterator TI = ToBB;
933  MachineFunction::iterator E = BB->getParent()->end();
934  while (I != TI) {
935    // Check isSuccessor to avoid case where the next block is empty, but
936    // it's not a successor.
937    if (I == E || !I->empty() || !PI->isSuccessor(I))
938      return false;
939    PI = I++;
940  }
941  return true;
942}
943
944/// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed
945/// to determine if it can be if-converted. If predecessor is already enqueued,
946/// dequeue it!
947void IfConverter::InvalidatePreds(MachineBasicBlock *BB) {
948  for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(),
949         E = BB->pred_end(); PI != E; ++PI) {
950    BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()];
951    if (PBBI.IsDone || PBBI.BB == BB)
952      continue;
953    PBBI.IsAnalyzed = false;
954    PBBI.IsEnqueued = false;
955  }
956}
957
958/// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB.
959///
960static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB,
961                               const TargetInstrInfo *TII) {
962  DebugLoc dl;  // FIXME: this is nowhere
963  SmallVector<MachineOperand, 0> NoCond;
964  TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl);
965}
966
967/// RemoveExtraEdges - Remove true / false edges if either / both are no longer
968/// successors.
969void IfConverter::RemoveExtraEdges(BBInfo &BBI) {
970  MachineBasicBlock *TBB = NULL, *FBB = NULL;
971  SmallVector<MachineOperand, 4> Cond;
972  if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond))
973    BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
974}
975
976/// InitPredRedefs / UpdatePredRedefs - Defs by predicated instructions are
977/// modeled as read + write (sort like two-address instructions). These
978/// routines track register liveness and add implicit uses to if-converted
979/// instructions to conform to the model.
980static void InitPredRedefs(MachineBasicBlock *BB, SmallSet<unsigned,4> &Redefs,
981                           const TargetRegisterInfo *TRI) {
982  for (MachineBasicBlock::livein_iterator I = BB->livein_begin(),
983         E = BB->livein_end(); I != E; ++I) {
984    unsigned Reg = *I;
985    Redefs.insert(Reg);
986    for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
987         *Subreg; ++Subreg)
988      Redefs.insert(*Subreg);
989  }
990}
991
992static void UpdatePredRedefs(MachineInstr *MI, SmallSet<unsigned,4> &Redefs,
993                             const TargetRegisterInfo *TRI,
994                             bool AddImpUse = false) {
995  SmallVector<unsigned, 4> Defs;
996  for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
997    const MachineOperand &MO = MI->getOperand(i);
998    if (!MO.isReg())
999      continue;
1000    unsigned Reg = MO.getReg();
1001    if (!Reg)
1002      continue;
1003    if (MO.isDef())
1004      Defs.push_back(Reg);
1005    else if (MO.isKill()) {
1006      Redefs.erase(Reg);
1007      for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1008        Redefs.erase(*SR);
1009    }
1010  }
1011  for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1012    unsigned Reg = Defs[i];
1013    if (Redefs.count(Reg)) {
1014      if (AddImpUse)
1015        // Treat predicated update as read + write.
1016        MI->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/,
1017                                                true/*IsImp*/,false/*IsKill*/));
1018    } else {
1019      Redefs.insert(Reg);
1020      for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
1021        Redefs.insert(*SR);
1022    }
1023  }
1024}
1025
1026static void UpdatePredRedefs(MachineBasicBlock::iterator I,
1027                             MachineBasicBlock::iterator E,
1028                             SmallSet<unsigned,4> &Redefs,
1029                             const TargetRegisterInfo *TRI) {
1030  while (I != E) {
1031    UpdatePredRedefs(I, Redefs, TRI);
1032    ++I;
1033  }
1034}
1035
1036/// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG.
1037///
1038bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) {
1039  BBInfo &TrueBBI  = BBAnalysis[BBI.TrueBB->getNumber()];
1040  BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1041  BBInfo *CvtBBI = &TrueBBI;
1042  BBInfo *NextBBI = &FalseBBI;
1043
1044  SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
1045  if (Kind == ICSimpleFalse)
1046    std::swap(CvtBBI, NextBBI);
1047
1048  if (CvtBBI->IsDone ||
1049      (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
1050    // Something has changed. It's no longer safe to predicate this block.
1051    BBI.IsAnalyzed = false;
1052    CvtBBI->IsAnalyzed = false;
1053    return false;
1054  }
1055
1056  if (Kind == ICSimpleFalse)
1057    if (TII->ReverseBranchCondition(Cond))
1058      assert(false && "Unable to reverse branch condition!");
1059
1060  // Initialize liveins to the first BB. These are potentiall redefined by
1061  // predicated instructions.
1062  SmallSet<unsigned, 4> Redefs;
1063  InitPredRedefs(CvtBBI->BB, Redefs, TRI);
1064  InitPredRedefs(NextBBI->BB, Redefs, TRI);
1065
1066  if (CvtBBI->BB->pred_size() > 1) {
1067    BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1068    // Copy instructions in the true block, predicate them, and add them to
1069    // the entry block.
1070    CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs);
1071  } else {
1072    PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs);
1073
1074    // Merge converted block into entry block.
1075    BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1076    MergeBlocks(BBI, *CvtBBI);
1077  }
1078
1079  bool IterIfcvt = true;
1080  if (!canFallThroughTo(BBI.BB, NextBBI->BB)) {
1081    InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
1082    BBI.HasFallThrough = false;
1083    // Now ifcvt'd block will look like this:
1084    // BB:
1085    // ...
1086    // t, f = cmp
1087    // if t op
1088    // b BBf
1089    //
1090    // We cannot further ifcvt this block because the unconditional branch
1091    // will have to be predicated on the new condition, that will not be
1092    // available if cmp executes.
1093    IterIfcvt = false;
1094  }
1095
1096  RemoveExtraEdges(BBI);
1097
1098  // Update block info. BB can be iteratively if-converted.
1099  if (!IterIfcvt)
1100    BBI.IsDone = true;
1101  InvalidatePreds(BBI.BB);
1102  CvtBBI->IsDone = true;
1103
1104  // FIXME: Must maintain LiveIns.
1105  return true;
1106}
1107
1108/// IfConvertTriangle - If convert a triangle sub-CFG.
1109///
1110bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
1111  BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
1112  BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1113  BBInfo *CvtBBI = &TrueBBI;
1114  BBInfo *NextBBI = &FalseBBI;
1115  DebugLoc dl;  // FIXME: this is nowhere
1116
1117  SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
1118  if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
1119    std::swap(CvtBBI, NextBBI);
1120
1121  if (CvtBBI->IsDone ||
1122      (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
1123    // Something has changed. It's no longer safe to predicate this block.
1124    BBI.IsAnalyzed = false;
1125    CvtBBI->IsAnalyzed = false;
1126    return false;
1127  }
1128
1129  if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
1130    if (TII->ReverseBranchCondition(Cond))
1131      assert(false && "Unable to reverse branch condition!");
1132
1133  if (Kind == ICTriangleRev || Kind == ICTriangleFRev) {
1134    if (ReverseBranchCondition(*CvtBBI)) {
1135      // BB has been changed, modify its predecessors (except for this
1136      // one) so they don't get ifcvt'ed based on bad intel.
1137      for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(),
1138             E = CvtBBI->BB->pred_end(); PI != E; ++PI) {
1139        MachineBasicBlock *PBB = *PI;
1140        if (PBB == BBI.BB)
1141          continue;
1142        BBInfo &PBBI = BBAnalysis[PBB->getNumber()];
1143        if (PBBI.IsEnqueued) {
1144          PBBI.IsAnalyzed = false;
1145          PBBI.IsEnqueued = false;
1146        }
1147      }
1148    }
1149  }
1150
1151  // Initialize liveins to the first BB. These are potentially redefined by
1152  // predicated instructions.
1153  SmallSet<unsigned, 4> Redefs;
1154  InitPredRedefs(CvtBBI->BB, Redefs, TRI);
1155  InitPredRedefs(NextBBI->BB, Redefs, TRI);
1156
1157  bool HasEarlyExit = CvtBBI->FalseBB != NULL;
1158  if (CvtBBI->BB->pred_size() > 1) {
1159    BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1160    // Copy instructions in the true block, predicate them, and add them to
1161    // the entry block.
1162    CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true);
1163  } else {
1164    // Predicate the 'true' block after removing its branch.
1165    CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB);
1166    PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs);
1167
1168    // Now merge the entry of the triangle with the true block.
1169    BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1170    MergeBlocks(BBI, *CvtBBI, false);
1171  }
1172
1173  // If 'true' block has a 'false' successor, add an exit branch to it.
1174  if (HasEarlyExit) {
1175    SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(),
1176                                           CvtBBI->BrCond.end());
1177    if (TII->ReverseBranchCondition(RevCond))
1178      assert(false && "Unable to reverse branch condition!");
1179    TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl);
1180    BBI.BB->addSuccessor(CvtBBI->FalseBB);
1181  }
1182
1183  // Merge in the 'false' block if the 'false' block has no other
1184  // predecessors. Otherwise, add an unconditional branch to 'false'.
1185  bool FalseBBDead = false;
1186  bool IterIfcvt = true;
1187  bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB);
1188  if (!isFallThrough) {
1189    // Only merge them if the true block does not fallthrough to the false
1190    // block. By not merging them, we make it possible to iteratively
1191    // ifcvt the blocks.
1192    if (!HasEarlyExit &&
1193        NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) {
1194      MergeBlocks(BBI, *NextBBI);
1195      FalseBBDead = true;
1196    } else {
1197      InsertUncondBranch(BBI.BB, NextBBI->BB, TII);
1198      BBI.HasFallThrough = false;
1199    }
1200    // Mixed predicated and unpredicated code. This cannot be iteratively
1201    // predicated.
1202    IterIfcvt = false;
1203  }
1204
1205  RemoveExtraEdges(BBI);
1206
1207  // Update block info. BB can be iteratively if-converted.
1208  if (!IterIfcvt)
1209    BBI.IsDone = true;
1210  InvalidatePreds(BBI.BB);
1211  CvtBBI->IsDone = true;
1212  if (FalseBBDead)
1213    NextBBI->IsDone = true;
1214
1215  // FIXME: Must maintain LiveIns.
1216  return true;
1217}
1218
1219/// IfConvertDiamond - If convert a diamond sub-CFG.
1220///
1221bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
1222                                   unsigned NumDups1, unsigned NumDups2) {
1223  BBInfo &TrueBBI  = BBAnalysis[BBI.TrueBB->getNumber()];
1224  BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
1225  MachineBasicBlock *TailBB = TrueBBI.TrueBB;
1226  // True block must fall through or end with an unanalyzable terminator.
1227  if (!TailBB) {
1228    if (blockAlwaysFallThrough(TrueBBI))
1229      TailBB = FalseBBI.TrueBB;
1230    assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!");
1231  }
1232
1233  if (TrueBBI.IsDone || FalseBBI.IsDone ||
1234      TrueBBI.BB->pred_size() > 1 ||
1235      FalseBBI.BB->pred_size() > 1) {
1236    // Something has changed. It's no longer safe to predicate these blocks.
1237    BBI.IsAnalyzed = false;
1238    TrueBBI.IsAnalyzed = false;
1239    FalseBBI.IsAnalyzed = false;
1240    return false;
1241  }
1242
1243  // Put the predicated instructions from the 'true' block before the
1244  // instructions from the 'false' block, unless the true block would clobber
1245  // the predicate, in which case, do the opposite.
1246  BBInfo *BBI1 = &TrueBBI;
1247  BBInfo *BBI2 = &FalseBBI;
1248  SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end());
1249  if (TII->ReverseBranchCondition(RevCond))
1250    assert(false && "Unable to reverse branch condition!");
1251  SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond;
1252  SmallVector<MachineOperand, 4> *Cond2 = &RevCond;
1253
1254  // Figure out the more profitable ordering.
1255  bool DoSwap = false;
1256  if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred)
1257    DoSwap = true;
1258  else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) {
1259    if (TrueBBI.NonPredSize > FalseBBI.NonPredSize)
1260      DoSwap = true;
1261  }
1262  if (DoSwap) {
1263    std::swap(BBI1, BBI2);
1264    std::swap(Cond1, Cond2);
1265  }
1266
1267  // Remove the conditional branch from entry to the blocks.
1268  BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
1269
1270  // Initialize liveins to the first BB. These are potentially redefined by
1271  // predicated instructions.
1272  SmallSet<unsigned, 4> Redefs;
1273  InitPredRedefs(BBI1->BB, Redefs, TRI);
1274
1275  // Remove the duplicated instructions at the beginnings of both paths.
1276  MachineBasicBlock::iterator DI1 = BBI1->BB->begin();
1277  MachineBasicBlock::iterator DI2 = BBI2->BB->begin();
1278  MachineBasicBlock::iterator DIE1 = BBI1->BB->end();
1279  MachineBasicBlock::iterator DIE2 = BBI2->BB->end();
1280  // Skip dbg_value instructions
1281  while (DI1 != DIE1 && DI1->isDebugValue())
1282    ++DI1;
1283  while (DI2 != DIE2 && DI2->isDebugValue())
1284    ++DI2;
1285  BBI1->NonPredSize -= NumDups1;
1286  BBI2->NonPredSize -= NumDups1;
1287
1288  // Skip past the dups on each side separately since there may be
1289  // differing dbg_value entries.
1290  for (unsigned i = 0; i < NumDups1; ++DI1) {
1291    if (!DI1->isDebugValue())
1292      ++i;
1293  }
1294  while (NumDups1 != 0) {
1295    ++DI2;
1296    if (!DI2->isDebugValue())
1297      --NumDups1;
1298  }
1299
1300  UpdatePredRedefs(BBI1->BB->begin(), DI1, Redefs, TRI);
1301  BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1);
1302  BBI2->BB->erase(BBI2->BB->begin(), DI2);
1303
1304  // Predicate the 'true' block after removing its branch.
1305  BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB);
1306  DI1 = BBI1->BB->end();
1307  for (unsigned i = 0; i != NumDups2; ) {
1308    // NumDups2 only counted non-dbg_value instructions, so this won't
1309    // run off the head of the list.
1310    assert (DI1 != BBI1->BB->begin());
1311    --DI1;
1312    // skip dbg_value instructions
1313    if (!DI1->isDebugValue())
1314      ++i;
1315  }
1316  BBI1->BB->erase(DI1, BBI1->BB->end());
1317  PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs);
1318
1319  // Predicate the 'false' block.
1320  BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB);
1321  DI2 = BBI2->BB->end();
1322  while (NumDups2 != 0) {
1323    // NumDups2 only counted non-dbg_value instructions, so this won't
1324    // run off the head of the list.
1325    assert (DI2 != BBI2->BB->begin());
1326    --DI2;
1327    // skip dbg_value instructions
1328    if (!DI2->isDebugValue())
1329      --NumDups2;
1330  }
1331  PredicateBlock(*BBI2, DI2, *Cond2, Redefs);
1332
1333  // Merge the true block into the entry of the diamond.
1334  MergeBlocks(BBI, *BBI1, TailBB == 0);
1335  MergeBlocks(BBI, *BBI2, TailBB == 0);
1336
1337  // If the if-converted block falls through or unconditionally branches into
1338  // the tail block, and the tail block does not have other predecessors, then
1339  // fold the tail block in as well. Otherwise, unless it falls through to the
1340  // tail, add a unconditional branch to it.
1341  if (TailBB) {
1342    BBInfo TailBBI = BBAnalysis[TailBB->getNumber()];
1343    bool CanMergeTail = !TailBBI.HasFallThrough;
1344    // There may still be a fall-through edge from BBI1 or BBI2 to TailBB;
1345    // check if there are any other predecessors besides those.
1346    unsigned NumPreds = TailBB->pred_size();
1347    if (NumPreds > 1)
1348      CanMergeTail = false;
1349    else if (NumPreds == 1 && CanMergeTail) {
1350      MachineBasicBlock::pred_iterator PI = TailBB->pred_begin();
1351      if (*PI != BBI1->BB && *PI != BBI2->BB)
1352        CanMergeTail = false;
1353    }
1354    if (CanMergeTail) {
1355      MergeBlocks(BBI, TailBBI);
1356      TailBBI.IsDone = true;
1357    } else {
1358      BBI.BB->addSuccessor(TailBB);
1359      InsertUncondBranch(BBI.BB, TailBB, TII);
1360      BBI.HasFallThrough = false;
1361    }
1362  }
1363
1364  // RemoveExtraEdges won't work if the block has an unanalyzable branch,
1365  // which can happen here if TailBB is unanalyzable and is merged, so
1366  // explicitly remove BBI1 and BBI2 as successors.
1367  BBI.BB->removeSuccessor(BBI1->BB);
1368  BBI.BB->removeSuccessor(BBI2->BB);
1369  RemoveExtraEdges(BBI);
1370
1371  // Update block info.
1372  BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true;
1373  InvalidatePreds(BBI.BB);
1374
1375  // FIXME: Must maintain LiveIns.
1376  return true;
1377}
1378
1379/// PredicateBlock - Predicate instructions from the start of the block to the
1380/// specified end with the specified condition.
1381void IfConverter::PredicateBlock(BBInfo &BBI,
1382                                 MachineBasicBlock::iterator E,
1383                                 SmallVectorImpl<MachineOperand> &Cond,
1384                                 SmallSet<unsigned, 4> &Redefs) {
1385  for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) {
1386    if (I->isDebugValue() || TII->isPredicated(I))
1387      continue;
1388    if (!TII->PredicateInstruction(I, Cond)) {
1389#ifndef NDEBUG
1390      dbgs() << "Unable to predicate " << *I << "!\n";
1391#endif
1392      llvm_unreachable(0);
1393    }
1394
1395    // If the predicated instruction now redefines a register as the result of
1396    // if-conversion, add an implicit kill.
1397    UpdatePredRedefs(I, Redefs, TRI, true);
1398  }
1399
1400  std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate));
1401
1402  BBI.IsAnalyzed = false;
1403  BBI.NonPredSize = 0;
1404
1405  ++NumIfConvBBs;
1406}
1407
1408/// CopyAndPredicateBlock - Copy and predicate instructions from source BB to
1409/// the destination block. Skip end of block branches if IgnoreBr is true.
1410void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
1411                                        SmallVectorImpl<MachineOperand> &Cond,
1412                                        SmallSet<unsigned, 4> &Redefs,
1413                                        bool IgnoreBr) {
1414  MachineFunction &MF = *ToBBI.BB->getParent();
1415
1416  for (MachineBasicBlock::iterator I = FromBBI.BB->begin(),
1417         E = FromBBI.BB->end(); I != E; ++I) {
1418    const TargetInstrDesc &TID = I->getDesc();
1419    // Do not copy the end of the block branches.
1420    if (IgnoreBr && TID.isBranch())
1421      break;
1422
1423    MachineInstr *MI = MF.CloneMachineInstr(I);
1424    ToBBI.BB->insert(ToBBI.BB->end(), MI);
1425    unsigned NumOps = TII->getNumMicroOps(MI, InstrItins);
1426    ToBBI.NonPredSize += NumOps;
1427
1428    if (!TII->isPredicated(I) && !MI->isDebugValue()) {
1429      if (!TII->PredicateInstruction(MI, Cond)) {
1430#ifndef NDEBUG
1431        dbgs() << "Unable to predicate " << *I << "!\n";
1432#endif
1433        llvm_unreachable(0);
1434      }
1435    }
1436
1437    // If the predicated instruction now redefines a register as the result of
1438    // if-conversion, add an implicit kill.
1439    UpdatePredRedefs(MI, Redefs, TRI, true);
1440  }
1441
1442  if (!IgnoreBr) {
1443    std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1444                                           FromBBI.BB->succ_end());
1445    MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1446    MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1447
1448    for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1449      MachineBasicBlock *Succ = Succs[i];
1450      // Fallthrough edge can't be transferred.
1451      if (Succ == FallThrough)
1452        continue;
1453      ToBBI.BB->addSuccessor(Succ);
1454    }
1455  }
1456
1457  std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1458            std::back_inserter(ToBBI.Predicate));
1459  std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate));
1460
1461  ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1462  ToBBI.IsAnalyzed = false;
1463
1464  ++NumDupBBs;
1465}
1466
1467/// MergeBlocks - Move all instructions from FromBB to the end of ToBB.
1468/// This will leave FromBB as an empty block, so remove all of its
1469/// successor edges except for the fall-through edge.  If AddEdges is true,
1470/// i.e., when FromBBI's branch is being moved, add those successor edges to
1471/// ToBBI.
1472void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) {
1473  ToBBI.BB->splice(ToBBI.BB->end(),
1474                   FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end());
1475
1476  std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
1477                                         FromBBI.BB->succ_end());
1478  MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
1479  MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
1480
1481  for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1482    MachineBasicBlock *Succ = Succs[i];
1483    // Fallthrough edge can't be transferred.
1484    if (Succ == FallThrough)
1485      continue;
1486    FromBBI.BB->removeSuccessor(Succ);
1487    if (AddEdges)
1488      ToBBI.BB->addSuccessor(Succ);
1489  }
1490
1491  // Now FromBBI always falls through to the next block!
1492  if (NBB && !FromBBI.BB->isSuccessor(NBB))
1493    FromBBI.BB->addSuccessor(NBB);
1494
1495  std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
1496            std::back_inserter(ToBBI.Predicate));
1497  FromBBI.Predicate.clear();
1498
1499  ToBBI.NonPredSize += FromBBI.NonPredSize;
1500  FromBBI.NonPredSize = 0;
1501
1502  ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
1503  ToBBI.HasFallThrough = FromBBI.HasFallThrough;
1504  ToBBI.IsAnalyzed = false;
1505  FromBBI.IsAnalyzed = false;
1506}
1507