BugReporter.cpp revision 49a246f4fad959888bb0164c624c3c2b03078e91
1// BugReporter.cpp - Generate PathDiagnostics for Bugs ------------*- C++ -*--//
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 defines BugReporter, a utility class for generating
11//  PathDiagnostics.
12//
13//===----------------------------------------------------------------------===//
14
15#define DEBUG_TYPE "BugReporter"
16
17#include "clang/StaticAnalyzer/Core/BugReporter/BugReporter.h"
18#include "clang/AST/ASTContext.h"
19#include "clang/AST/DeclObjC.h"
20#include "clang/AST/Expr.h"
21#include "clang/AST/ParentMap.h"
22#include "clang/AST/StmtObjC.h"
23#include "clang/Analysis/CFG.h"
24#include "clang/Analysis/ProgramPoint.h"
25#include "clang/Basic/SourceManager.h"
26#include "clang/StaticAnalyzer/Core/BugReporter/BugType.h"
27#include "clang/StaticAnalyzer/Core/BugReporter/PathDiagnostic.h"
28#include "clang/StaticAnalyzer/Core/PathSensitive/ExprEngine.h"
29#include "llvm/ADT/DenseMap.h"
30#include "llvm/ADT/IntrusiveRefCntPtr.h"
31#include "llvm/ADT/OwningPtr.h"
32#include "llvm/ADT/STLExtras.h"
33#include "llvm/ADT/SmallString.h"
34#include "llvm/ADT/Statistic.h"
35#include "llvm/Support/raw_ostream.h"
36#include <queue>
37
38using namespace clang;
39using namespace ento;
40
41STATISTIC(MaxBugClassSize,
42          "The maximum number of bug reports in the same equivalence class");
43STATISTIC(MaxValidBugClassSize,
44          "The maximum number of bug reports in the same equivalence class "
45          "where at least one report is valid (not suppressed)");
46
47BugReporterVisitor::~BugReporterVisitor() {}
48
49void BugReporterContext::anchor() {}
50
51//===----------------------------------------------------------------------===//
52// Helper routines for walking the ExplodedGraph and fetching statements.
53//===----------------------------------------------------------------------===//
54
55static const Stmt *GetPreviousStmt(const ExplodedNode *N) {
56  for (N = N->getFirstPred(); N; N = N->getFirstPred())
57    if (const Stmt *S = PathDiagnosticLocation::getStmt(N))
58      return S;
59
60  return 0;
61}
62
63static inline const Stmt*
64GetCurrentOrPreviousStmt(const ExplodedNode *N) {
65  if (const Stmt *S = PathDiagnosticLocation::getStmt(N))
66    return S;
67
68  return GetPreviousStmt(N);
69}
70
71//===----------------------------------------------------------------------===//
72// Diagnostic cleanup.
73//===----------------------------------------------------------------------===//
74
75static PathDiagnosticEventPiece *
76eventsDescribeSameCondition(PathDiagnosticEventPiece *X,
77                            PathDiagnosticEventPiece *Y) {
78  // Prefer diagnostics that come from ConditionBRVisitor over
79  // those that came from TrackConstraintBRVisitor.
80  const void *tagPreferred = ConditionBRVisitor::getTag();
81  const void *tagLesser = TrackConstraintBRVisitor::getTag();
82
83  if (X->getLocation() != Y->getLocation())
84    return 0;
85
86  if (X->getTag() == tagPreferred && Y->getTag() == tagLesser)
87    return X;
88
89  if (Y->getTag() == tagPreferred && X->getTag() == tagLesser)
90    return Y;
91
92  return 0;
93}
94
95/// An optimization pass over PathPieces that removes redundant diagnostics
96/// generated by both ConditionBRVisitor and TrackConstraintBRVisitor.  Both
97/// BugReporterVisitors use different methods to generate diagnostics, with
98/// one capable of emitting diagnostics in some cases but not in others.  This
99/// can lead to redundant diagnostic pieces at the same point in a path.
100static void removeRedundantMsgs(PathPieces &path) {
101  unsigned N = path.size();
102  if (N < 2)
103    return;
104  // NOTE: this loop intentionally is not using an iterator.  Instead, we
105  // are streaming the path and modifying it in place.  This is done by
106  // grabbing the front, processing it, and if we decide to keep it append
107  // it to the end of the path.  The entire path is processed in this way.
108  for (unsigned i = 0; i < N; ++i) {
109    IntrusiveRefCntPtr<PathDiagnosticPiece> piece(path.front());
110    path.pop_front();
111
112    switch (piece->getKind()) {
113      case clang::ento::PathDiagnosticPiece::Call:
114        removeRedundantMsgs(cast<PathDiagnosticCallPiece>(piece)->path);
115        break;
116      case clang::ento::PathDiagnosticPiece::Macro:
117        removeRedundantMsgs(cast<PathDiagnosticMacroPiece>(piece)->subPieces);
118        break;
119      case clang::ento::PathDiagnosticPiece::ControlFlow:
120        break;
121      case clang::ento::PathDiagnosticPiece::Event: {
122        if (i == N-1)
123          break;
124
125        if (PathDiagnosticEventPiece *nextEvent =
126            dyn_cast<PathDiagnosticEventPiece>(path.front().getPtr())) {
127          PathDiagnosticEventPiece *event =
128            cast<PathDiagnosticEventPiece>(piece);
129          // Check to see if we should keep one of the two pieces.  If we
130          // come up with a preference, record which piece to keep, and consume
131          // another piece from the path.
132          if (PathDiagnosticEventPiece *pieceToKeep =
133              eventsDescribeSameCondition(event, nextEvent)) {
134            piece = pieceToKeep;
135            path.pop_front();
136            ++i;
137          }
138        }
139        break;
140      }
141    }
142    path.push_back(piece);
143  }
144}
145
146/// A map from PathDiagnosticPiece to the LocationContext of the inlined
147/// function call it represents.
148typedef llvm::DenseMap<const PathPieces *, const LocationContext *>
149        LocationContextMap;
150
151/// Recursively scan through a path and prune out calls and macros pieces
152/// that aren't needed.  Return true if afterwards the path contains
153/// "interesting stuff" which means it shouldn't be pruned from the parent path.
154static bool removeUnneededCalls(PathPieces &pieces, BugReport *R,
155                                LocationContextMap &LCM) {
156  bool containsSomethingInteresting = false;
157  const unsigned N = pieces.size();
158
159  for (unsigned i = 0 ; i < N ; ++i) {
160    // Remove the front piece from the path.  If it is still something we
161    // want to keep once we are done, we will push it back on the end.
162    IntrusiveRefCntPtr<PathDiagnosticPiece> piece(pieces.front());
163    pieces.pop_front();
164
165    // Throw away pieces with invalid locations. Note that we can't throw away
166    // calls just yet because they might have something interesting inside them.
167    // If so, their locations will be adjusted as necessary later.
168    if (piece->getKind() != PathDiagnosticPiece::Call &&
169        piece->getLocation().asLocation().isInvalid())
170      continue;
171
172    switch (piece->getKind()) {
173      case PathDiagnosticPiece::Call: {
174        PathDiagnosticCallPiece *call = cast<PathDiagnosticCallPiece>(piece);
175        // Check if the location context is interesting.
176        assert(LCM.count(&call->path));
177        if (R->isInteresting(LCM[&call->path])) {
178          containsSomethingInteresting = true;
179          break;
180        }
181
182        if (!removeUnneededCalls(call->path, R, LCM))
183          continue;
184
185        containsSomethingInteresting = true;
186        break;
187      }
188      case PathDiagnosticPiece::Macro: {
189        PathDiagnosticMacroPiece *macro = cast<PathDiagnosticMacroPiece>(piece);
190        if (!removeUnneededCalls(macro->subPieces, R, LCM))
191          continue;
192        containsSomethingInteresting = true;
193        break;
194      }
195      case PathDiagnosticPiece::Event: {
196        PathDiagnosticEventPiece *event = cast<PathDiagnosticEventPiece>(piece);
197
198        // We never throw away an event, but we do throw it away wholesale
199        // as part of a path if we throw the entire path away.
200        containsSomethingInteresting |= !event->isPrunable();
201        break;
202      }
203      case PathDiagnosticPiece::ControlFlow:
204        break;
205    }
206
207    pieces.push_back(piece);
208  }
209
210  return containsSomethingInteresting;
211}
212
213/// Returns true if the given decl has been implicitly given a body, either by
214/// the analyzer or by the compiler proper.
215static bool hasImplicitBody(const Decl *D) {
216  assert(D);
217  return D->isImplicit() || !D->hasBody();
218}
219
220/// Recursively scan through a path and make sure that all call pieces have
221/// valid locations. Note that all other pieces with invalid locations should
222/// have already been pruned out.
223static void adjustCallLocations(PathPieces &Pieces,
224                                PathDiagnosticLocation *LastCallLocation = 0) {
225  for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E; ++I) {
226    PathDiagnosticCallPiece *Call = dyn_cast<PathDiagnosticCallPiece>(*I);
227
228    if (!Call) {
229      assert((*I)->getLocation().asLocation().isValid());
230      continue;
231    }
232
233    if (LastCallLocation) {
234      bool CallerIsImplicit = hasImplicitBody(Call->getCaller());
235      if (CallerIsImplicit || !Call->callEnter.asLocation().isValid())
236        Call->callEnter = *LastCallLocation;
237      if (CallerIsImplicit || !Call->callReturn.asLocation().isValid())
238        Call->callReturn = *LastCallLocation;
239    }
240
241    // Recursively clean out the subclass.  Keep this call around if
242    // it contains any informative diagnostics.
243    PathDiagnosticLocation *ThisCallLocation;
244    if (Call->callEnterWithin.asLocation().isValid() &&
245        !hasImplicitBody(Call->getCallee()))
246      ThisCallLocation = &Call->callEnterWithin;
247    else
248      ThisCallLocation = &Call->callEnter;
249
250    assert(ThisCallLocation && "Outermost call has an invalid location");
251    adjustCallLocations(Call->path, ThisCallLocation);
252  }
253}
254
255//===----------------------------------------------------------------------===//
256// PathDiagnosticBuilder and its associated routines and helper objects.
257//===----------------------------------------------------------------------===//
258
259namespace {
260class NodeMapClosure : public BugReport::NodeResolver {
261  InterExplodedGraphMap &M;
262public:
263  NodeMapClosure(InterExplodedGraphMap &m) : M(m) {}
264
265  const ExplodedNode *getOriginalNode(const ExplodedNode *N) {
266    return M.lookup(N);
267  }
268};
269
270class PathDiagnosticBuilder : public BugReporterContext {
271  BugReport *R;
272  PathDiagnosticConsumer *PDC;
273  NodeMapClosure NMC;
274public:
275  const LocationContext *LC;
276
277  PathDiagnosticBuilder(GRBugReporter &br,
278                        BugReport *r, InterExplodedGraphMap &Backmap,
279                        PathDiagnosticConsumer *pdc)
280    : BugReporterContext(br),
281      R(r), PDC(pdc), NMC(Backmap), LC(r->getErrorNode()->getLocationContext())
282  {}
283
284  PathDiagnosticLocation ExecutionContinues(const ExplodedNode *N);
285
286  PathDiagnosticLocation ExecutionContinues(llvm::raw_string_ostream &os,
287                                            const ExplodedNode *N);
288
289  BugReport *getBugReport() { return R; }
290
291  Decl const &getCodeDecl() { return R->getErrorNode()->getCodeDecl(); }
292
293  ParentMap& getParentMap() { return LC->getParentMap(); }
294
295  const Stmt *getParent(const Stmt *S) {
296    return getParentMap().getParent(S);
297  }
298
299  virtual NodeMapClosure& getNodeResolver() { return NMC; }
300
301  PathDiagnosticLocation getEnclosingStmtLocation(const Stmt *S);
302
303  PathDiagnosticConsumer::PathGenerationScheme getGenerationScheme() const {
304    return PDC ? PDC->getGenerationScheme() : PathDiagnosticConsumer::Extensive;
305  }
306
307  bool supportsLogicalOpControlFlow() const {
308    return PDC ? PDC->supportsLogicalOpControlFlow() : true;
309  }
310};
311} // end anonymous namespace
312
313PathDiagnosticLocation
314PathDiagnosticBuilder::ExecutionContinues(const ExplodedNode *N) {
315  if (const Stmt *S = PathDiagnosticLocation::getNextStmt(N))
316    return PathDiagnosticLocation(S, getSourceManager(), LC);
317
318  return PathDiagnosticLocation::createDeclEnd(N->getLocationContext(),
319                                               getSourceManager());
320}
321
322PathDiagnosticLocation
323PathDiagnosticBuilder::ExecutionContinues(llvm::raw_string_ostream &os,
324                                          const ExplodedNode *N) {
325
326  // Slow, but probably doesn't matter.
327  if (os.str().empty())
328    os << ' ';
329
330  const PathDiagnosticLocation &Loc = ExecutionContinues(N);
331
332  if (Loc.asStmt())
333    os << "Execution continues on line "
334       << getSourceManager().getExpansionLineNumber(Loc.asLocation())
335       << '.';
336  else {
337    os << "Execution jumps to the end of the ";
338    const Decl *D = N->getLocationContext()->getDecl();
339    if (isa<ObjCMethodDecl>(D))
340      os << "method";
341    else if (isa<FunctionDecl>(D))
342      os << "function";
343    else {
344      assert(isa<BlockDecl>(D));
345      os << "anonymous block";
346    }
347    os << '.';
348  }
349
350  return Loc;
351}
352
353static const Stmt *getEnclosingParent(const Stmt *S, const ParentMap &PM) {
354  if (isa<Expr>(S) && PM.isConsumedExpr(cast<Expr>(S)))
355    return PM.getParentIgnoreParens(S);
356
357  const Stmt *Parent = PM.getParentIgnoreParens(S);
358  if (!Parent)
359    return 0;
360
361  switch (Parent->getStmtClass()) {
362  case Stmt::ForStmtClass:
363  case Stmt::DoStmtClass:
364  case Stmt::WhileStmtClass:
365  case Stmt::ObjCForCollectionStmtClass:
366  case Stmt::CXXForRangeStmtClass:
367    return Parent;
368  default:
369    break;
370  }
371
372  return 0;
373}
374
375static PathDiagnosticLocation
376getEnclosingStmtLocation(const Stmt *S, SourceManager &SMgr, const ParentMap &P,
377                         const LocationContext *LC, bool allowNestedContexts) {
378  if (!S)
379    return PathDiagnosticLocation();
380
381  while (const Stmt *Parent = getEnclosingParent(S, P)) {
382    switch (Parent->getStmtClass()) {
383      case Stmt::BinaryOperatorClass: {
384        const BinaryOperator *B = cast<BinaryOperator>(Parent);
385        if (B->isLogicalOp())
386          return PathDiagnosticLocation(allowNestedContexts ? B : S, SMgr, LC);
387        break;
388      }
389      case Stmt::CompoundStmtClass:
390      case Stmt::StmtExprClass:
391        return PathDiagnosticLocation(S, SMgr, LC);
392      case Stmt::ChooseExprClass:
393        // Similar to '?' if we are referring to condition, just have the edge
394        // point to the entire choose expression.
395        if (allowNestedContexts || cast<ChooseExpr>(Parent)->getCond() == S)
396          return PathDiagnosticLocation(Parent, SMgr, LC);
397        else
398          return PathDiagnosticLocation(S, SMgr, LC);
399      case Stmt::BinaryConditionalOperatorClass:
400      case Stmt::ConditionalOperatorClass:
401        // For '?', if we are referring to condition, just have the edge point
402        // to the entire '?' expression.
403        if (allowNestedContexts ||
404            cast<AbstractConditionalOperator>(Parent)->getCond() == S)
405          return PathDiagnosticLocation(Parent, SMgr, LC);
406        else
407          return PathDiagnosticLocation(S, SMgr, LC);
408      case Stmt::CXXForRangeStmtClass:
409        if (cast<CXXForRangeStmt>(Parent)->getBody() == S)
410          return PathDiagnosticLocation(S, SMgr, LC);
411        break;
412      case Stmt::DoStmtClass:
413          return PathDiagnosticLocation(S, SMgr, LC);
414      case Stmt::ForStmtClass:
415        if (cast<ForStmt>(Parent)->getBody() == S)
416          return PathDiagnosticLocation(S, SMgr, LC);
417        break;
418      case Stmt::IfStmtClass:
419        if (cast<IfStmt>(Parent)->getCond() != S)
420          return PathDiagnosticLocation(S, SMgr, LC);
421        break;
422      case Stmt::ObjCForCollectionStmtClass:
423        if (cast<ObjCForCollectionStmt>(Parent)->getBody() == S)
424          return PathDiagnosticLocation(S, SMgr, LC);
425        break;
426      case Stmt::WhileStmtClass:
427        if (cast<WhileStmt>(Parent)->getCond() != S)
428          return PathDiagnosticLocation(S, SMgr, LC);
429        break;
430      default:
431        break;
432    }
433
434    S = Parent;
435  }
436
437  assert(S && "Cannot have null Stmt for PathDiagnosticLocation");
438
439  return PathDiagnosticLocation(S, SMgr, LC);
440}
441
442PathDiagnosticLocation
443PathDiagnosticBuilder::getEnclosingStmtLocation(const Stmt *S) {
444  assert(S && "Null Stmt passed to getEnclosingStmtLocation");
445  return ::getEnclosingStmtLocation(S, getSourceManager(), getParentMap(), LC,
446                                    /*allowNestedContexts=*/false);
447}
448
449//===----------------------------------------------------------------------===//
450// "Visitors only" path diagnostic generation algorithm.
451//===----------------------------------------------------------------------===//
452static bool GenerateVisitorsOnlyPathDiagnostic(PathDiagnostic &PD,
453                                               PathDiagnosticBuilder &PDB,
454                                               const ExplodedNode *N,
455                                      ArrayRef<BugReporterVisitor *> visitors) {
456  // All path generation skips the very first node (the error node).
457  // This is because there is special handling for the end-of-path note.
458  N = N->getFirstPred();
459  if (!N)
460    return true;
461
462  BugReport *R = PDB.getBugReport();
463  while (const ExplodedNode *Pred = N->getFirstPred()) {
464    for (ArrayRef<BugReporterVisitor *>::iterator I = visitors.begin(),
465                                                  E = visitors.end();
466         I != E; ++I) {
467      // Visit all the node pairs, but throw the path pieces away.
468      PathDiagnosticPiece *Piece = (*I)->VisitNode(N, Pred, PDB, *R);
469      delete Piece;
470    }
471
472    N = Pred;
473  }
474
475  return R->isValid();
476}
477
478//===----------------------------------------------------------------------===//
479// "Minimal" path diagnostic generation algorithm.
480//===----------------------------------------------------------------------===//
481typedef std::pair<PathDiagnosticCallPiece*, const ExplodedNode*> StackDiagPair;
482typedef SmallVector<StackDiagPair, 6> StackDiagVector;
483
484static void updateStackPiecesWithMessage(PathDiagnosticPiece *P,
485                                         StackDiagVector &CallStack) {
486  // If the piece contains a special message, add it to all the call
487  // pieces on the active stack.
488  if (PathDiagnosticEventPiece *ep =
489        dyn_cast<PathDiagnosticEventPiece>(P)) {
490
491    if (ep->hasCallStackHint())
492      for (StackDiagVector::iterator I = CallStack.begin(),
493                                     E = CallStack.end(); I != E; ++I) {
494        PathDiagnosticCallPiece *CP = I->first;
495        const ExplodedNode *N = I->second;
496        std::string stackMsg = ep->getCallStackMessage(N);
497
498        // The last message on the path to final bug is the most important
499        // one. Since we traverse the path backwards, do not add the message
500        // if one has been previously added.
501        if  (!CP->hasCallStackMessage())
502          CP->setCallStackMessage(stackMsg);
503      }
504  }
505}
506
507static void CompactPathDiagnostic(PathPieces &path, const SourceManager& SM);
508
509static bool GenerateMinimalPathDiagnostic(PathDiagnostic& PD,
510                                          PathDiagnosticBuilder &PDB,
511                                          const ExplodedNode *N,
512                                          LocationContextMap &LCM,
513                                      ArrayRef<BugReporterVisitor *> visitors) {
514
515  SourceManager& SMgr = PDB.getSourceManager();
516  const LocationContext *LC = PDB.LC;
517  const ExplodedNode *NextNode = N->pred_empty()
518                                        ? NULL : *(N->pred_begin());
519
520  StackDiagVector CallStack;
521
522  while (NextNode) {
523    N = NextNode;
524    PDB.LC = N->getLocationContext();
525    NextNode = N->getFirstPred();
526
527    ProgramPoint P = N->getLocation();
528
529    do {
530      if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
531        PathDiagnosticCallPiece *C =
532            PathDiagnosticCallPiece::construct(N, *CE, SMgr);
533        // Record the mapping from call piece to LocationContext.
534        LCM[&C->path] = CE->getCalleeContext();
535        PD.getActivePath().push_front(C);
536        PD.pushActivePath(&C->path);
537        CallStack.push_back(StackDiagPair(C, N));
538        break;
539      }
540
541      if (Optional<CallEnter> CE = P.getAs<CallEnter>()) {
542        // Flush all locations, and pop the active path.
543        bool VisitedEntireCall = PD.isWithinCall();
544        PD.popActivePath();
545
546        // Either we just added a bunch of stuff to the top-level path, or
547        // we have a previous CallExitEnd.  If the former, it means that the
548        // path terminated within a function call.  We must then take the
549        // current contents of the active path and place it within
550        // a new PathDiagnosticCallPiece.
551        PathDiagnosticCallPiece *C;
552        if (VisitedEntireCall) {
553          C = cast<PathDiagnosticCallPiece>(PD.getActivePath().front());
554        } else {
555          const Decl *Caller = CE->getLocationContext()->getDecl();
556          C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller);
557          // Record the mapping from call piece to LocationContext.
558          LCM[&C->path] = CE->getCalleeContext();
559        }
560
561        C->setCallee(*CE, SMgr);
562        if (!CallStack.empty()) {
563          assert(CallStack.back().first == C);
564          CallStack.pop_back();
565        }
566        break;
567      }
568
569      if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) {
570        const CFGBlock *Src = BE->getSrc();
571        const CFGBlock *Dst = BE->getDst();
572        const Stmt *T = Src->getTerminator();
573
574        if (!T)
575          break;
576
577        PathDiagnosticLocation Start =
578            PathDiagnosticLocation::createBegin(T, SMgr,
579                N->getLocationContext());
580
581        switch (T->getStmtClass()) {
582        default:
583          break;
584
585        case Stmt::GotoStmtClass:
586        case Stmt::IndirectGotoStmtClass: {
587          const Stmt *S = PathDiagnosticLocation::getNextStmt(N);
588
589          if (!S)
590            break;
591
592          std::string sbuf;
593          llvm::raw_string_ostream os(sbuf);
594          const PathDiagnosticLocation &End = PDB.getEnclosingStmtLocation(S);
595
596          os << "Control jumps to line "
597              << End.asLocation().getExpansionLineNumber();
598          PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
599              Start, End, os.str()));
600          break;
601        }
602
603        case Stmt::SwitchStmtClass: {
604          // Figure out what case arm we took.
605          std::string sbuf;
606          llvm::raw_string_ostream os(sbuf);
607
608          if (const Stmt *S = Dst->getLabel()) {
609            PathDiagnosticLocation End(S, SMgr, LC);
610
611            switch (S->getStmtClass()) {
612            default:
613              os << "No cases match in the switch statement. "
614              "Control jumps to line "
615              << End.asLocation().getExpansionLineNumber();
616              break;
617            case Stmt::DefaultStmtClass:
618              os << "Control jumps to the 'default' case at line "
619              << End.asLocation().getExpansionLineNumber();
620              break;
621
622            case Stmt::CaseStmtClass: {
623              os << "Control jumps to 'case ";
624              const CaseStmt *Case = cast<CaseStmt>(S);
625              const Expr *LHS = Case->getLHS()->IgnoreParenCasts();
626
627              // Determine if it is an enum.
628              bool GetRawInt = true;
629
630              if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(LHS)) {
631                // FIXME: Maybe this should be an assertion.  Are there cases
632                // were it is not an EnumConstantDecl?
633                const EnumConstantDecl *D =
634                    dyn_cast<EnumConstantDecl>(DR->getDecl());
635
636                if (D) {
637                  GetRawInt = false;
638                  os << *D;
639                }
640              }
641
642              if (GetRawInt)
643                os << LHS->EvaluateKnownConstInt(PDB.getASTContext());
644
645              os << ":'  at line "
646                  << End.asLocation().getExpansionLineNumber();
647              break;
648            }
649            }
650            PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
651                Start, End, os.str()));
652          }
653          else {
654            os << "'Default' branch taken. ";
655            const PathDiagnosticLocation &End = PDB.ExecutionContinues(os, N);
656            PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
657                Start, End, os.str()));
658          }
659
660          break;
661        }
662
663        case Stmt::BreakStmtClass:
664        case Stmt::ContinueStmtClass: {
665          std::string sbuf;
666          llvm::raw_string_ostream os(sbuf);
667          PathDiagnosticLocation End = PDB.ExecutionContinues(os, N);
668          PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
669              Start, End, os.str()));
670          break;
671        }
672
673        // Determine control-flow for ternary '?'.
674        case Stmt::BinaryConditionalOperatorClass:
675        case Stmt::ConditionalOperatorClass: {
676          std::string sbuf;
677          llvm::raw_string_ostream os(sbuf);
678          os << "'?' condition is ";
679
680          if (*(Src->succ_begin()+1) == Dst)
681            os << "false";
682          else
683            os << "true";
684
685          PathDiagnosticLocation End = PDB.ExecutionContinues(N);
686
687          if (const Stmt *S = End.asStmt())
688            End = PDB.getEnclosingStmtLocation(S);
689
690          PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
691              Start, End, os.str()));
692          break;
693        }
694
695        // Determine control-flow for short-circuited '&&' and '||'.
696        case Stmt::BinaryOperatorClass: {
697          if (!PDB.supportsLogicalOpControlFlow())
698            break;
699
700          const BinaryOperator *B = cast<BinaryOperator>(T);
701          std::string sbuf;
702          llvm::raw_string_ostream os(sbuf);
703          os << "Left side of '";
704
705          if (B->getOpcode() == BO_LAnd) {
706            os << "&&" << "' is ";
707
708            if (*(Src->succ_begin()+1) == Dst) {
709              os << "false";
710              PathDiagnosticLocation End(B->getLHS(), SMgr, LC);
711              PathDiagnosticLocation Start =
712                  PathDiagnosticLocation::createOperatorLoc(B, SMgr);
713              PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
714                  Start, End, os.str()));
715            }
716            else {
717              os << "true";
718              PathDiagnosticLocation Start(B->getLHS(), SMgr, LC);
719              PathDiagnosticLocation End = PDB.ExecutionContinues(N);
720              PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
721                  Start, End, os.str()));
722            }
723          }
724          else {
725            assert(B->getOpcode() == BO_LOr);
726            os << "||" << "' is ";
727
728            if (*(Src->succ_begin()+1) == Dst) {
729              os << "false";
730              PathDiagnosticLocation Start(B->getLHS(), SMgr, LC);
731              PathDiagnosticLocation End = PDB.ExecutionContinues(N);
732              PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
733                  Start, End, os.str()));
734            }
735            else {
736              os << "true";
737              PathDiagnosticLocation End(B->getLHS(), SMgr, LC);
738              PathDiagnosticLocation Start =
739                  PathDiagnosticLocation::createOperatorLoc(B, SMgr);
740              PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
741                  Start, End, os.str()));
742            }
743          }
744
745          break;
746        }
747
748        case Stmt::DoStmtClass:  {
749          if (*(Src->succ_begin()) == Dst) {
750            std::string sbuf;
751            llvm::raw_string_ostream os(sbuf);
752
753            os << "Loop condition is true. ";
754            PathDiagnosticLocation End = PDB.ExecutionContinues(os, N);
755
756            if (const Stmt *S = End.asStmt())
757              End = PDB.getEnclosingStmtLocation(S);
758
759            PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
760                Start, End, os.str()));
761          }
762          else {
763            PathDiagnosticLocation End = PDB.ExecutionContinues(N);
764
765            if (const Stmt *S = End.asStmt())
766              End = PDB.getEnclosingStmtLocation(S);
767
768            PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
769                Start, End, "Loop condition is false.  Exiting loop"));
770          }
771
772          break;
773        }
774
775        case Stmt::WhileStmtClass:
776        case Stmt::ForStmtClass: {
777          if (*(Src->succ_begin()+1) == Dst) {
778            std::string sbuf;
779            llvm::raw_string_ostream os(sbuf);
780
781            os << "Loop condition is false. ";
782            PathDiagnosticLocation End = PDB.ExecutionContinues(os, N);
783            if (const Stmt *S = End.asStmt())
784              End = PDB.getEnclosingStmtLocation(S);
785
786            PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
787                Start, End, os.str()));
788          }
789          else {
790            PathDiagnosticLocation End = PDB.ExecutionContinues(N);
791            if (const Stmt *S = End.asStmt())
792              End = PDB.getEnclosingStmtLocation(S);
793
794            PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
795                Start, End, "Loop condition is true.  Entering loop body"));
796          }
797
798          break;
799        }
800
801        case Stmt::IfStmtClass: {
802          PathDiagnosticLocation End = PDB.ExecutionContinues(N);
803
804          if (const Stmt *S = End.asStmt())
805            End = PDB.getEnclosingStmtLocation(S);
806
807          if (*(Src->succ_begin()+1) == Dst)
808            PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
809                Start, End, "Taking false branch"));
810          else
811            PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
812                Start, End, "Taking true branch"));
813
814          break;
815        }
816        }
817      }
818    } while(0);
819
820    if (NextNode) {
821      // Add diagnostic pieces from custom visitors.
822      BugReport *R = PDB.getBugReport();
823      for (ArrayRef<BugReporterVisitor *>::iterator I = visitors.begin(),
824                                                    E = visitors.end();
825           I != E; ++I) {
826        if (PathDiagnosticPiece *p = (*I)->VisitNode(N, NextNode, PDB, *R)) {
827          PD.getActivePath().push_front(p);
828          updateStackPiecesWithMessage(p, CallStack);
829        }
830      }
831    }
832  }
833
834  if (!PDB.getBugReport()->isValid())
835    return false;
836
837  // After constructing the full PathDiagnostic, do a pass over it to compact
838  // PathDiagnosticPieces that occur within a macro.
839  CompactPathDiagnostic(PD.getMutablePieces(), PDB.getSourceManager());
840  return true;
841}
842
843//===----------------------------------------------------------------------===//
844// "Extensive" PathDiagnostic generation.
845//===----------------------------------------------------------------------===//
846
847static bool IsControlFlowExpr(const Stmt *S) {
848  const Expr *E = dyn_cast<Expr>(S);
849
850  if (!E)
851    return false;
852
853  E = E->IgnoreParenCasts();
854
855  if (isa<AbstractConditionalOperator>(E))
856    return true;
857
858  if (const BinaryOperator *B = dyn_cast<BinaryOperator>(E))
859    if (B->isLogicalOp())
860      return true;
861
862  return false;
863}
864
865namespace {
866class ContextLocation : public PathDiagnosticLocation {
867  bool IsDead;
868public:
869  ContextLocation(const PathDiagnosticLocation &L, bool isdead = false)
870    : PathDiagnosticLocation(L), IsDead(isdead) {}
871
872  void markDead() { IsDead = true; }
873  bool isDead() const { return IsDead; }
874};
875
876static PathDiagnosticLocation cleanUpLocation(PathDiagnosticLocation L,
877                                              const LocationContext *LC,
878                                              bool firstCharOnly = false) {
879  if (const Stmt *S = L.asStmt()) {
880    const Stmt *Original = S;
881    while (1) {
882      // Adjust the location for some expressions that are best referenced
883      // by one of their subexpressions.
884      switch (S->getStmtClass()) {
885        default:
886          break;
887        case Stmt::ParenExprClass:
888        case Stmt::GenericSelectionExprClass:
889          S = cast<Expr>(S)->IgnoreParens();
890          firstCharOnly = true;
891          continue;
892        case Stmt::BinaryConditionalOperatorClass:
893        case Stmt::ConditionalOperatorClass:
894          S = cast<AbstractConditionalOperator>(S)->getCond();
895          firstCharOnly = true;
896          continue;
897        case Stmt::ChooseExprClass:
898          S = cast<ChooseExpr>(S)->getCond();
899          firstCharOnly = true;
900          continue;
901        case Stmt::BinaryOperatorClass:
902          S = cast<BinaryOperator>(S)->getLHS();
903          firstCharOnly = true;
904          continue;
905      }
906
907      break;
908    }
909
910    if (S != Original)
911      L = PathDiagnosticLocation(S, L.getManager(), LC);
912  }
913
914  if (firstCharOnly)
915    L  = PathDiagnosticLocation::createSingleLocation(L);
916
917  return L;
918}
919
920class EdgeBuilder {
921  std::vector<ContextLocation> CLocs;
922  typedef std::vector<ContextLocation>::iterator iterator;
923  PathDiagnostic &PD;
924  PathDiagnosticBuilder &PDB;
925  PathDiagnosticLocation PrevLoc;
926
927  bool IsConsumedExpr(const PathDiagnosticLocation &L);
928
929  bool containsLocation(const PathDiagnosticLocation &Container,
930                        const PathDiagnosticLocation &Containee);
931
932  PathDiagnosticLocation getContextLocation(const PathDiagnosticLocation &L);
933
934
935
936  void popLocation() {
937    if (!CLocs.back().isDead() && CLocs.back().asLocation().isFileID()) {
938      // For contexts, we only one the first character as the range.
939      rawAddEdge(cleanUpLocation(CLocs.back(), PDB.LC, true));
940    }
941    CLocs.pop_back();
942  }
943
944public:
945  EdgeBuilder(PathDiagnostic &pd, PathDiagnosticBuilder &pdb)
946    : PD(pd), PDB(pdb) {
947
948      // If the PathDiagnostic already has pieces, add the enclosing statement
949      // of the first piece as a context as well.
950      if (!PD.path.empty()) {
951        PrevLoc = (*PD.path.begin())->getLocation();
952
953        if (const Stmt *S = PrevLoc.asStmt())
954          addExtendedContext(PDB.getEnclosingStmtLocation(S).asStmt());
955      }
956  }
957
958  ~EdgeBuilder() {
959    while (!CLocs.empty()) popLocation();
960
961    // Finally, add an initial edge from the start location of the first
962    // statement (if it doesn't already exist).
963    PathDiagnosticLocation L = PathDiagnosticLocation::createDeclBegin(
964                                                       PDB.LC,
965                                                       PDB.getSourceManager());
966    if (L.isValid())
967      rawAddEdge(L);
968  }
969
970  void flushLocations() {
971    while (!CLocs.empty())
972      popLocation();
973    PrevLoc = PathDiagnosticLocation();
974  }
975
976  void addEdge(PathDiagnosticLocation NewLoc, bool alwaysAdd = false,
977               bool IsPostJump = false);
978
979  void rawAddEdge(PathDiagnosticLocation NewLoc);
980
981  void addContext(const Stmt *S);
982  void addContext(const PathDiagnosticLocation &L);
983  void addExtendedContext(const Stmt *S);
984};
985} // end anonymous namespace
986
987
988PathDiagnosticLocation
989EdgeBuilder::getContextLocation(const PathDiagnosticLocation &L) {
990  if (const Stmt *S = L.asStmt()) {
991    if (IsControlFlowExpr(S))
992      return L;
993
994    return PDB.getEnclosingStmtLocation(S);
995  }
996
997  return L;
998}
999
1000bool EdgeBuilder::containsLocation(const PathDiagnosticLocation &Container,
1001                                   const PathDiagnosticLocation &Containee) {
1002
1003  if (Container == Containee)
1004    return true;
1005
1006  if (Container.asDecl())
1007    return true;
1008
1009  if (const Stmt *S = Containee.asStmt())
1010    if (const Stmt *ContainerS = Container.asStmt()) {
1011      while (S) {
1012        if (S == ContainerS)
1013          return true;
1014        S = PDB.getParent(S);
1015      }
1016      return false;
1017    }
1018
1019  // Less accurate: compare using source ranges.
1020  SourceRange ContainerR = Container.asRange();
1021  SourceRange ContaineeR = Containee.asRange();
1022
1023  SourceManager &SM = PDB.getSourceManager();
1024  SourceLocation ContainerRBeg = SM.getExpansionLoc(ContainerR.getBegin());
1025  SourceLocation ContainerREnd = SM.getExpansionLoc(ContainerR.getEnd());
1026  SourceLocation ContaineeRBeg = SM.getExpansionLoc(ContaineeR.getBegin());
1027  SourceLocation ContaineeREnd = SM.getExpansionLoc(ContaineeR.getEnd());
1028
1029  unsigned ContainerBegLine = SM.getExpansionLineNumber(ContainerRBeg);
1030  unsigned ContainerEndLine = SM.getExpansionLineNumber(ContainerREnd);
1031  unsigned ContaineeBegLine = SM.getExpansionLineNumber(ContaineeRBeg);
1032  unsigned ContaineeEndLine = SM.getExpansionLineNumber(ContaineeREnd);
1033
1034  assert(ContainerBegLine <= ContainerEndLine);
1035  assert(ContaineeBegLine <= ContaineeEndLine);
1036
1037  return (ContainerBegLine <= ContaineeBegLine &&
1038          ContainerEndLine >= ContaineeEndLine &&
1039          (ContainerBegLine != ContaineeBegLine ||
1040           SM.getExpansionColumnNumber(ContainerRBeg) <=
1041           SM.getExpansionColumnNumber(ContaineeRBeg)) &&
1042          (ContainerEndLine != ContaineeEndLine ||
1043           SM.getExpansionColumnNumber(ContainerREnd) >=
1044           SM.getExpansionColumnNumber(ContaineeREnd)));
1045}
1046
1047void EdgeBuilder::rawAddEdge(PathDiagnosticLocation NewLoc) {
1048  if (!PrevLoc.isValid()) {
1049    PrevLoc = NewLoc;
1050    return;
1051  }
1052
1053  const PathDiagnosticLocation &NewLocClean = cleanUpLocation(NewLoc, PDB.LC);
1054  const PathDiagnosticLocation &PrevLocClean = cleanUpLocation(PrevLoc, PDB.LC);
1055
1056  if (PrevLocClean.asLocation().isInvalid()) {
1057    PrevLoc = NewLoc;
1058    return;
1059  }
1060
1061  if (NewLocClean.asLocation() == PrevLocClean.asLocation())
1062    return;
1063
1064  // FIXME: Ignore intra-macro edges for now.
1065  if (NewLocClean.asLocation().getExpansionLoc() ==
1066      PrevLocClean.asLocation().getExpansionLoc())
1067    return;
1068
1069  PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(NewLocClean, PrevLocClean));
1070  PrevLoc = NewLoc;
1071}
1072
1073void EdgeBuilder::addEdge(PathDiagnosticLocation NewLoc, bool alwaysAdd,
1074                          bool IsPostJump) {
1075
1076  if (!alwaysAdd && NewLoc.asLocation().isMacroID())
1077    return;
1078
1079  const PathDiagnosticLocation &CLoc = getContextLocation(NewLoc);
1080
1081  while (!CLocs.empty()) {
1082    ContextLocation &TopContextLoc = CLocs.back();
1083
1084    // Is the top location context the same as the one for the new location?
1085    if (TopContextLoc == CLoc) {
1086      if (alwaysAdd) {
1087        if (IsConsumedExpr(TopContextLoc))
1088          TopContextLoc.markDead();
1089
1090        rawAddEdge(NewLoc);
1091      }
1092
1093      if (IsPostJump)
1094        TopContextLoc.markDead();
1095      return;
1096    }
1097
1098    if (containsLocation(TopContextLoc, CLoc)) {
1099      if (alwaysAdd) {
1100        rawAddEdge(NewLoc);
1101
1102        if (IsConsumedExpr(CLoc)) {
1103          CLocs.push_back(ContextLocation(CLoc, /*IsDead=*/true));
1104          return;
1105        }
1106      }
1107
1108      CLocs.push_back(ContextLocation(CLoc, /*IsDead=*/IsPostJump));
1109      return;
1110    }
1111
1112    // Context does not contain the location.  Flush it.
1113    popLocation();
1114  }
1115
1116  // If we reach here, there is no enclosing context.  Just add the edge.
1117  rawAddEdge(NewLoc);
1118}
1119
1120bool EdgeBuilder::IsConsumedExpr(const PathDiagnosticLocation &L) {
1121  if (const Expr *X = dyn_cast_or_null<Expr>(L.asStmt()))
1122    return PDB.getParentMap().isConsumedExpr(X) && !IsControlFlowExpr(X);
1123
1124  return false;
1125}
1126
1127void EdgeBuilder::addExtendedContext(const Stmt *S) {
1128  if (!S)
1129    return;
1130
1131  const Stmt *Parent = PDB.getParent(S);
1132  while (Parent) {
1133    if (isa<CompoundStmt>(Parent))
1134      Parent = PDB.getParent(Parent);
1135    else
1136      break;
1137  }
1138
1139  if (Parent) {
1140    switch (Parent->getStmtClass()) {
1141      case Stmt::DoStmtClass:
1142      case Stmt::ObjCAtSynchronizedStmtClass:
1143        addContext(Parent);
1144      default:
1145        break;
1146    }
1147  }
1148
1149  addContext(S);
1150}
1151
1152void EdgeBuilder::addContext(const Stmt *S) {
1153  if (!S)
1154    return;
1155
1156  PathDiagnosticLocation L(S, PDB.getSourceManager(), PDB.LC);
1157  addContext(L);
1158}
1159
1160void EdgeBuilder::addContext(const PathDiagnosticLocation &L) {
1161  while (!CLocs.empty()) {
1162    const PathDiagnosticLocation &TopContextLoc = CLocs.back();
1163
1164    // Is the top location context the same as the one for the new location?
1165    if (TopContextLoc == L)
1166      return;
1167
1168    if (containsLocation(TopContextLoc, L)) {
1169      CLocs.push_back(L);
1170      return;
1171    }
1172
1173    // Context does not contain the location.  Flush it.
1174    popLocation();
1175  }
1176
1177  CLocs.push_back(L);
1178}
1179
1180// Cone-of-influence: support the reverse propagation of "interesting" symbols
1181// and values by tracing interesting calculations backwards through evaluated
1182// expressions along a path.  This is probably overly complicated, but the idea
1183// is that if an expression computed an "interesting" value, the child
1184// expressions are are also likely to be "interesting" as well (which then
1185// propagates to the values they in turn compute).  This reverse propagation
1186// is needed to track interesting correlations across function call boundaries,
1187// where formal arguments bind to actual arguments, etc.  This is also needed
1188// because the constraint solver sometimes simplifies certain symbolic values
1189// into constants when appropriate, and this complicates reasoning about
1190// interesting values.
1191typedef llvm::DenseSet<const Expr *> InterestingExprs;
1192
1193static void reversePropagateIntererstingSymbols(BugReport &R,
1194                                                InterestingExprs &IE,
1195                                                const ProgramState *State,
1196                                                const Expr *Ex,
1197                                                const LocationContext *LCtx) {
1198  SVal V = State->getSVal(Ex, LCtx);
1199  if (!(R.isInteresting(V) || IE.count(Ex)))
1200    return;
1201
1202  switch (Ex->getStmtClass()) {
1203    default:
1204      if (!isa<CastExpr>(Ex))
1205        break;
1206      // Fall through.
1207    case Stmt::BinaryOperatorClass:
1208    case Stmt::UnaryOperatorClass: {
1209      for (Stmt::const_child_iterator CI = Ex->child_begin(),
1210            CE = Ex->child_end();
1211            CI != CE; ++CI) {
1212        if (const Expr *child = dyn_cast_or_null<Expr>(*CI)) {
1213          IE.insert(child);
1214          SVal ChildV = State->getSVal(child, LCtx);
1215          R.markInteresting(ChildV);
1216        }
1217        break;
1218      }
1219    }
1220  }
1221
1222  R.markInteresting(V);
1223}
1224
1225static void reversePropagateInterestingSymbols(BugReport &R,
1226                                               InterestingExprs &IE,
1227                                               const ProgramState *State,
1228                                               const LocationContext *CalleeCtx,
1229                                               const LocationContext *CallerCtx)
1230{
1231  // FIXME: Handle non-CallExpr-based CallEvents.
1232  const StackFrameContext *Callee = CalleeCtx->getCurrentStackFrame();
1233  const Stmt *CallSite = Callee->getCallSite();
1234  if (const CallExpr *CE = dyn_cast_or_null<CallExpr>(CallSite)) {
1235    if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(CalleeCtx->getDecl())) {
1236      FunctionDecl::param_const_iterator PI = FD->param_begin(),
1237                                         PE = FD->param_end();
1238      CallExpr::const_arg_iterator AI = CE->arg_begin(), AE = CE->arg_end();
1239      for (; AI != AE && PI != PE; ++AI, ++PI) {
1240        if (const Expr *ArgE = *AI) {
1241          if (const ParmVarDecl *PD = *PI) {
1242            Loc LV = State->getLValue(PD, CalleeCtx);
1243            if (R.isInteresting(LV) || R.isInteresting(State->getRawSVal(LV)))
1244              IE.insert(ArgE);
1245          }
1246        }
1247      }
1248    }
1249  }
1250}
1251
1252//===----------------------------------------------------------------------===//
1253// Functions for determining if a loop was executed 0 times.
1254//===----------------------------------------------------------------------===//
1255
1256static bool isLoop(const Stmt *Term) {
1257  switch (Term->getStmtClass()) {
1258    case Stmt::ForStmtClass:
1259    case Stmt::WhileStmtClass:
1260    case Stmt::ObjCForCollectionStmtClass:
1261    case Stmt::CXXForRangeStmtClass:
1262      return true;
1263    default:
1264      // Note that we intentionally do not include do..while here.
1265      return false;
1266  }
1267}
1268
1269static bool isJumpToFalseBranch(const BlockEdge *BE) {
1270  const CFGBlock *Src = BE->getSrc();
1271  assert(Src->succ_size() == 2);
1272  return (*(Src->succ_begin()+1) == BE->getDst());
1273}
1274
1275/// Return true if the terminator is a loop and the destination is the
1276/// false branch.
1277static bool isLoopJumpPastBody(const Stmt *Term, const BlockEdge *BE) {
1278  if (!isLoop(Term))
1279    return false;
1280
1281  // Did we take the false branch?
1282  return isJumpToFalseBranch(BE);
1283}
1284
1285static bool isContainedByStmt(ParentMap &PM, const Stmt *S, const Stmt *SubS) {
1286  while (SubS) {
1287    if (SubS == S)
1288      return true;
1289    SubS = PM.getParent(SubS);
1290  }
1291  return false;
1292}
1293
1294static const Stmt *getStmtBeforeCond(ParentMap &PM, const Stmt *Term,
1295                                     const ExplodedNode *N) {
1296  while (N) {
1297    Optional<StmtPoint> SP = N->getLocation().getAs<StmtPoint>();
1298    if (SP) {
1299      const Stmt *S = SP->getStmt();
1300      if (!isContainedByStmt(PM, Term, S))
1301        return S;
1302    }
1303    N = N->getFirstPred();
1304  }
1305  return 0;
1306}
1307
1308static bool isInLoopBody(ParentMap &PM, const Stmt *S, const Stmt *Term) {
1309  const Stmt *LoopBody = 0;
1310  switch (Term->getStmtClass()) {
1311    case Stmt::CXXForRangeStmtClass: {
1312      const CXXForRangeStmt *FR = cast<CXXForRangeStmt>(Term);
1313      if (isContainedByStmt(PM, FR->getInc(), S))
1314        return true;
1315      if (isContainedByStmt(PM, FR->getLoopVarStmt(), S))
1316        return true;
1317      LoopBody = FR->getBody();
1318      break;
1319    }
1320    case Stmt::ForStmtClass: {
1321      const ForStmt *FS = cast<ForStmt>(Term);
1322      if (isContainedByStmt(PM, FS->getInc(), S))
1323        return true;
1324      LoopBody = FS->getBody();
1325      break;
1326    }
1327    case Stmt::ObjCForCollectionStmtClass: {
1328      const ObjCForCollectionStmt *FC = cast<ObjCForCollectionStmt>(Term);
1329      LoopBody = FC->getBody();
1330      break;
1331    }
1332    case Stmt::WhileStmtClass:
1333      LoopBody = cast<WhileStmt>(Term)->getBody();
1334      break;
1335    default:
1336      return false;
1337  }
1338  return isContainedByStmt(PM, LoopBody, S);
1339}
1340
1341//===----------------------------------------------------------------------===//
1342// Top-level logic for generating extensive path diagnostics.
1343//===----------------------------------------------------------------------===//
1344
1345static bool GenerateExtensivePathDiagnostic(PathDiagnostic& PD,
1346                                            PathDiagnosticBuilder &PDB,
1347                                            const ExplodedNode *N,
1348                                            LocationContextMap &LCM,
1349                                      ArrayRef<BugReporterVisitor *> visitors) {
1350  EdgeBuilder EB(PD, PDB);
1351  const SourceManager& SM = PDB.getSourceManager();
1352  StackDiagVector CallStack;
1353  InterestingExprs IE;
1354
1355  const ExplodedNode *NextNode = N->pred_empty() ? NULL : *(N->pred_begin());
1356  while (NextNode) {
1357    N = NextNode;
1358    NextNode = N->getFirstPred();
1359    ProgramPoint P = N->getLocation();
1360
1361    do {
1362      if (Optional<PostStmt> PS = P.getAs<PostStmt>()) {
1363        if (const Expr *Ex = PS->getStmtAs<Expr>())
1364          reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1365                                              N->getState().getPtr(), Ex,
1366                                              N->getLocationContext());
1367      }
1368
1369      if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
1370        const Stmt *S = CE->getCalleeContext()->getCallSite();
1371        if (const Expr *Ex = dyn_cast_or_null<Expr>(S)) {
1372            reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1373                                                N->getState().getPtr(), Ex,
1374                                                N->getLocationContext());
1375        }
1376
1377        PathDiagnosticCallPiece *C =
1378          PathDiagnosticCallPiece::construct(N, *CE, SM);
1379        LCM[&C->path] = CE->getCalleeContext();
1380
1381        EB.addEdge(C->callReturn, /*AlwaysAdd=*/true, /*IsPostJump=*/true);
1382        EB.flushLocations();
1383
1384        PD.getActivePath().push_front(C);
1385        PD.pushActivePath(&C->path);
1386        CallStack.push_back(StackDiagPair(C, N));
1387        break;
1388      }
1389
1390      // Pop the call hierarchy if we are done walking the contents
1391      // of a function call.
1392      if (Optional<CallEnter> CE = P.getAs<CallEnter>()) {
1393        // Add an edge to the start of the function.
1394        const Decl *D = CE->getCalleeContext()->getDecl();
1395        PathDiagnosticLocation pos =
1396          PathDiagnosticLocation::createBegin(D, SM);
1397        EB.addEdge(pos);
1398
1399        // Flush all locations, and pop the active path.
1400        bool VisitedEntireCall = PD.isWithinCall();
1401        EB.flushLocations();
1402        PD.popActivePath();
1403        PDB.LC = N->getLocationContext();
1404
1405        // Either we just added a bunch of stuff to the top-level path, or
1406        // we have a previous CallExitEnd.  If the former, it means that the
1407        // path terminated within a function call.  We must then take the
1408        // current contents of the active path and place it within
1409        // a new PathDiagnosticCallPiece.
1410        PathDiagnosticCallPiece *C;
1411        if (VisitedEntireCall) {
1412          C = cast<PathDiagnosticCallPiece>(PD.getActivePath().front());
1413        } else {
1414          const Decl *Caller = CE->getLocationContext()->getDecl();
1415          C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller);
1416          LCM[&C->path] = CE->getCalleeContext();
1417        }
1418
1419        C->setCallee(*CE, SM);
1420        EB.addContext(C->getLocation());
1421
1422        if (!CallStack.empty()) {
1423          assert(CallStack.back().first == C);
1424          CallStack.pop_back();
1425        }
1426        break;
1427      }
1428
1429      // Note that is important that we update the LocationContext
1430      // after looking at CallExits.  CallExit basically adds an
1431      // edge in the *caller*, so we don't want to update the LocationContext
1432      // too soon.
1433      PDB.LC = N->getLocationContext();
1434
1435      // Block edges.
1436      if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) {
1437        // Does this represent entering a call?  If so, look at propagating
1438        // interesting symbols across call boundaries.
1439        if (NextNode) {
1440          const LocationContext *CallerCtx = NextNode->getLocationContext();
1441          const LocationContext *CalleeCtx = PDB.LC;
1442          if (CallerCtx != CalleeCtx) {
1443            reversePropagateInterestingSymbols(*PDB.getBugReport(), IE,
1444                                               N->getState().getPtr(),
1445                                               CalleeCtx, CallerCtx);
1446          }
1447        }
1448
1449        // Are we jumping to the head of a loop?  Add a special diagnostic.
1450        if (const Stmt *Loop = BE->getSrc()->getLoopTarget()) {
1451          PathDiagnosticLocation L(Loop, SM, PDB.LC);
1452          const CompoundStmt *CS = NULL;
1453
1454          if (const ForStmt *FS = dyn_cast<ForStmt>(Loop))
1455            CS = dyn_cast<CompoundStmt>(FS->getBody());
1456          else if (const WhileStmt *WS = dyn_cast<WhileStmt>(Loop))
1457            CS = dyn_cast<CompoundStmt>(WS->getBody());
1458
1459          PathDiagnosticEventPiece *p =
1460            new PathDiagnosticEventPiece(L,
1461                                        "Looping back to the head of the loop");
1462          p->setPrunable(true);
1463
1464          EB.addEdge(p->getLocation(), true);
1465          PD.getActivePath().push_front(p);
1466
1467          if (CS) {
1468            PathDiagnosticLocation BL =
1469              PathDiagnosticLocation::createEndBrace(CS, SM);
1470            EB.addEdge(BL);
1471          }
1472        }
1473
1474        const CFGBlock *BSrc = BE->getSrc();
1475        ParentMap &PM = PDB.getParentMap();
1476
1477        if (const Stmt *Term = BSrc->getTerminator()) {
1478          // Are we jumping past the loop body without ever executing the
1479          // loop (because the condition was false)?
1480          if (isLoopJumpPastBody(Term, &*BE) &&
1481              !isInLoopBody(PM,
1482                            getStmtBeforeCond(PM,
1483                                              BSrc->getTerminatorCondition(),
1484                                              N),
1485                            Term)) {
1486            PathDiagnosticLocation L(Term, SM, PDB.LC);
1487            PathDiagnosticEventPiece *PE =
1488                new PathDiagnosticEventPiece(L, "Loop body executed 0 times");
1489            PE->setPrunable(true);
1490
1491            EB.addEdge(PE->getLocation(), true);
1492            PD.getActivePath().push_front(PE);
1493          }
1494
1495          // In any case, add the terminator as the current statement
1496          // context for control edges.
1497          EB.addContext(Term);
1498        }
1499
1500        break;
1501      }
1502
1503      if (Optional<BlockEntrance> BE = P.getAs<BlockEntrance>()) {
1504        Optional<CFGElement> First = BE->getFirstElement();
1505        if (Optional<CFGStmt> S = First ? First->getAs<CFGStmt>() : None) {
1506          const Stmt *stmt = S->getStmt();
1507          if (IsControlFlowExpr(stmt)) {
1508            // Add the proper context for '&&', '||', and '?'.
1509            EB.addContext(stmt);
1510          }
1511          else
1512            EB.addExtendedContext(PDB.getEnclosingStmtLocation(stmt).asStmt());
1513        }
1514
1515        break;
1516      }
1517
1518
1519    } while (0);
1520
1521    if (!NextNode)
1522      continue;
1523
1524    // Add pieces from custom visitors.
1525    BugReport *R = PDB.getBugReport();
1526    for (ArrayRef<BugReporterVisitor *>::iterator I = visitors.begin(),
1527                                                  E = visitors.end();
1528         I != E; ++I) {
1529      if (PathDiagnosticPiece *p = (*I)->VisitNode(N, NextNode, PDB, *R)) {
1530        const PathDiagnosticLocation &Loc = p->getLocation();
1531        EB.addEdge(Loc, true);
1532        PD.getActivePath().push_front(p);
1533        updateStackPiecesWithMessage(p, CallStack);
1534
1535        if (const Stmt *S = Loc.asStmt())
1536          EB.addExtendedContext(PDB.getEnclosingStmtLocation(S).asStmt());
1537      }
1538    }
1539  }
1540
1541  return PDB.getBugReport()->isValid();
1542}
1543
1544/// \brief Adds a sanitized control-flow diagnostic edge to a path.
1545static void addEdgeToPath(PathPieces &path,
1546                          PathDiagnosticLocation &PrevLoc,
1547                          PathDiagnosticLocation NewLoc,
1548                          const LocationContext *LC) {
1549  if (!NewLoc.isValid())
1550    return;
1551
1552  SourceLocation NewLocL = NewLoc.asLocation();
1553  if (NewLocL.isInvalid() || NewLocL.isMacroID())
1554    return;
1555
1556  if (!PrevLoc.isValid() || !PrevLoc.asLocation().isValid()) {
1557    PrevLoc = NewLoc;
1558    return;
1559  }
1560
1561  // Ignore self-edges, which occur when there are multiple nodes at the same
1562  // statement.
1563  if (NewLoc.asStmt() && NewLoc.asStmt() == PrevLoc.asStmt())
1564    return;
1565
1566  path.push_front(new PathDiagnosticControlFlowPiece(NewLoc,
1567                                                     PrevLoc));
1568  PrevLoc = NewLoc;
1569}
1570
1571/// A customized wrapper for CFGBlock::getTerminatorCondition()
1572/// which returns the element for ObjCForCollectionStmts.
1573static const Stmt *getTerminatorCondition(const CFGBlock *B) {
1574  const Stmt *S = B->getTerminatorCondition();
1575  if (const ObjCForCollectionStmt *FS =
1576      dyn_cast_or_null<ObjCForCollectionStmt>(S))
1577    return FS->getElement();
1578  return S;
1579}
1580
1581static const char *StrEnteringLoop = "Entering loop body";
1582static const char *StrLoopBodyZero = "Loop body executed 0 times";
1583
1584static bool
1585GenerateAlternateExtensivePathDiagnostic(PathDiagnostic& PD,
1586                                         PathDiagnosticBuilder &PDB,
1587                                         const ExplodedNode *N,
1588                                         LocationContextMap &LCM,
1589                                      ArrayRef<BugReporterVisitor *> visitors) {
1590
1591  BugReport *report = PDB.getBugReport();
1592  const SourceManager& SM = PDB.getSourceManager();
1593  StackDiagVector CallStack;
1594  InterestingExprs IE;
1595
1596  PathDiagnosticLocation PrevLoc = PD.getLocation();
1597
1598  const ExplodedNode *NextNode = N->getFirstPred();
1599  while (NextNode) {
1600    N = NextNode;
1601    NextNode = N->getFirstPred();
1602    ProgramPoint P = N->getLocation();
1603
1604    do {
1605      // Have we encountered an entrance to a call?  It may be
1606      // the case that we have not encountered a matching
1607      // call exit before this point.  This means that the path
1608      // terminated within the call itself.
1609      if (Optional<CallEnter> CE = P.getAs<CallEnter>()) {
1610        // Add an edge to the start of the function.
1611        const StackFrameContext *CalleeLC = CE->getCalleeContext();
1612        const Decl *D = CalleeLC->getDecl();
1613        addEdgeToPath(PD.getActivePath(), PrevLoc,
1614                      PathDiagnosticLocation::createBegin(D, SM),
1615                      CalleeLC);
1616
1617        // Did we visit an entire call?
1618        bool VisitedEntireCall = PD.isWithinCall();
1619        PD.popActivePath();
1620
1621        PathDiagnosticCallPiece *C;
1622        if (VisitedEntireCall) {
1623          PathDiagnosticPiece *P = PD.getActivePath().front().getPtr();
1624          C = cast<PathDiagnosticCallPiece>(P);
1625        } else {
1626          const Decl *Caller = CE->getLocationContext()->getDecl();
1627          C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller);
1628
1629          // Since we just transferred the path over to the call piece,
1630          // reset the mapping from active to location context.
1631          assert(PD.getActivePath().size() == 1 &&
1632                 PD.getActivePath().front() == C);
1633          LCM[&PD.getActivePath()] = 0;
1634
1635          // Record the location context mapping for the path within
1636          // the call.
1637          assert(LCM[&C->path] == 0 ||
1638                 LCM[&C->path] == CE->getCalleeContext());
1639          LCM[&C->path] = CE->getCalleeContext();
1640
1641          // If this is the first item in the active path, record
1642          // the new mapping from active path to location context.
1643          const LocationContext *&NewLC = LCM[&PD.getActivePath()];
1644          if (!NewLC)
1645            NewLC = N->getLocationContext();
1646
1647          PDB.LC = NewLC;
1648        }
1649        C->setCallee(*CE, SM);
1650
1651        // Update the previous location in the active path.
1652        PrevLoc = C->getLocation();
1653
1654        if (!CallStack.empty()) {
1655          assert(CallStack.back().first == C);
1656          CallStack.pop_back();
1657        }
1658        break;
1659      }
1660
1661      // Query the location context here and the previous location
1662      // as processing CallEnter may change the active path.
1663      PDB.LC = N->getLocationContext();
1664
1665      // Record the mapping from the active path to the location
1666      // context.
1667      assert(!LCM[&PD.getActivePath()] ||
1668             LCM[&PD.getActivePath()] == PDB.LC);
1669      LCM[&PD.getActivePath()] = PDB.LC;
1670
1671      // Have we encountered an exit from a function call?
1672      if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
1673        const Stmt *S = CE->getCalleeContext()->getCallSite();
1674        // Propagate the interesting symbols accordingly.
1675        if (const Expr *Ex = dyn_cast_or_null<Expr>(S)) {
1676          reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1677                                              N->getState().getPtr(), Ex,
1678                                              N->getLocationContext());
1679        }
1680
1681        // We are descending into a call (backwards).  Construct
1682        // a new call piece to contain the path pieces for that call.
1683        PathDiagnosticCallPiece *C =
1684          PathDiagnosticCallPiece::construct(N, *CE, SM);
1685
1686        // Record the location context for this call piece.
1687        LCM[&C->path] = CE->getCalleeContext();
1688
1689        // Add the edge to the return site.
1690        addEdgeToPath(PD.getActivePath(), PrevLoc, C->callReturn, PDB.LC);
1691        PD.getActivePath().push_front(C);
1692        PrevLoc.invalidate();
1693
1694        // Make the contents of the call the active path for now.
1695        PD.pushActivePath(&C->path);
1696        CallStack.push_back(StackDiagPair(C, N));
1697        break;
1698      }
1699
1700      if (Optional<PostStmt> PS = P.getAs<PostStmt>()) {
1701        // For expressions, make sure we propagate the
1702        // interesting symbols correctly.
1703        if (const Expr *Ex = PS->getStmtAs<Expr>())
1704          reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1705                                              N->getState().getPtr(), Ex,
1706                                              N->getLocationContext());
1707
1708        // Add an edge.  If this is an ObjCForCollectionStmt do
1709        // not add an edge here as it appears in the CFG both
1710        // as a terminator and as a terminator condition.
1711        if (!isa<ObjCForCollectionStmt>(PS->getStmt())) {
1712          PathDiagnosticLocation L =
1713            PathDiagnosticLocation(PS->getStmt(), SM, PDB.LC);
1714          addEdgeToPath(PD.getActivePath(), PrevLoc, L, PDB.LC);
1715        }
1716        break;
1717      }
1718
1719      // Block edges.
1720      if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) {
1721        // Does this represent entering a call?  If so, look at propagating
1722        // interesting symbols across call boundaries.
1723        if (NextNode) {
1724          const LocationContext *CallerCtx = NextNode->getLocationContext();
1725          const LocationContext *CalleeCtx = PDB.LC;
1726          if (CallerCtx != CalleeCtx) {
1727            reversePropagateInterestingSymbols(*PDB.getBugReport(), IE,
1728                                               N->getState().getPtr(),
1729                                               CalleeCtx, CallerCtx);
1730          }
1731        }
1732
1733        // Are we jumping to the head of a loop?  Add a special diagnostic.
1734        if (const Stmt *Loop = BE->getSrc()->getLoopTarget()) {
1735          PathDiagnosticLocation L(Loop, SM, PDB.LC);
1736          const Stmt *Body = NULL;
1737
1738          if (const ForStmt *FS = dyn_cast<ForStmt>(Loop))
1739            Body = FS->getBody();
1740          else if (const WhileStmt *WS = dyn_cast<WhileStmt>(Loop))
1741            Body = WS->getBody();
1742          else if (const ObjCForCollectionStmt *OFS =
1743                     dyn_cast<ObjCForCollectionStmt>(Loop)) {
1744            Body = OFS->getBody();
1745          } else if (const CXXForRangeStmt *FRS =
1746                       dyn_cast<CXXForRangeStmt>(Loop)) {
1747            Body = FRS->getBody();
1748          }
1749          // do-while statements are explicitly excluded here
1750
1751          PathDiagnosticEventPiece *p =
1752            new PathDiagnosticEventPiece(L, "Looping back to the head "
1753                                            "of the loop");
1754          p->setPrunable(true);
1755
1756          addEdgeToPath(PD.getActivePath(), PrevLoc, p->getLocation(), PDB.LC);
1757          PD.getActivePath().push_front(p);
1758
1759          if (const CompoundStmt *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
1760            addEdgeToPath(PD.getActivePath(), PrevLoc,
1761                          PathDiagnosticLocation::createEndBrace(CS, SM),
1762                          PDB.LC);
1763          }
1764        }
1765
1766        const CFGBlock *BSrc = BE->getSrc();
1767        ParentMap &PM = PDB.getParentMap();
1768
1769        if (const Stmt *Term = BSrc->getTerminator()) {
1770          // Are we jumping past the loop body without ever executing the
1771          // loop (because the condition was false)?
1772          if (isLoop(Term)) {
1773            const Stmt *TermCond = getTerminatorCondition(BSrc);
1774            bool IsInLoopBody =
1775              isInLoopBody(PM, getStmtBeforeCond(PM, TermCond, N), Term);
1776
1777            const char *str = 0;
1778
1779            if (isJumpToFalseBranch(&*BE)) {
1780              if (!IsInLoopBody) {
1781                str = StrLoopBodyZero;
1782              }
1783            }
1784            else {
1785              str = StrEnteringLoop;
1786            }
1787
1788            if (str) {
1789              PathDiagnosticLocation L(TermCond ? TermCond : Term, SM, PDB.LC);
1790              PathDiagnosticEventPiece *PE =
1791                new PathDiagnosticEventPiece(L, str);
1792              PE->setPrunable(true);
1793              addEdgeToPath(PD.getActivePath(), PrevLoc,
1794                            PE->getLocation(), PDB.LC);
1795              PD.getActivePath().push_front(PE);
1796            }
1797          }
1798          else if (isa<BreakStmt>(Term) || isa<ContinueStmt>(Term) ||
1799                   isa<GotoStmt>(Term)) {
1800            PathDiagnosticLocation L(Term, SM, PDB.LC);
1801            addEdgeToPath(PD.getActivePath(), PrevLoc, L, PDB.LC);
1802          }
1803        }
1804        break;
1805      }
1806    } while (0);
1807
1808    if (!NextNode)
1809      continue;
1810
1811    // Add pieces from custom visitors.
1812    for (ArrayRef<BugReporterVisitor *>::iterator I = visitors.begin(),
1813         E = visitors.end();
1814         I != E; ++I) {
1815      if (PathDiagnosticPiece *p = (*I)->VisitNode(N, NextNode, PDB, *report)) {
1816        addEdgeToPath(PD.getActivePath(), PrevLoc, p->getLocation(), PDB.LC);
1817        PD.getActivePath().push_front(p);
1818        updateStackPiecesWithMessage(p, CallStack);
1819      }
1820    }
1821  }
1822
1823  // Add an edge to the start of the function.
1824  // We'll prune it out later, but it helps make diagnostics more uniform.
1825  const StackFrameContext *CalleeLC = PDB.LC->getCurrentStackFrame();
1826  const Decl *D = CalleeLC->getDecl();
1827  addEdgeToPath(PD.getActivePath(), PrevLoc,
1828                PathDiagnosticLocation::createBegin(D, SM),
1829                CalleeLC);
1830
1831  return report->isValid();
1832}
1833
1834static const Stmt *getLocStmt(PathDiagnosticLocation L) {
1835  if (!L.isValid())
1836    return 0;
1837  return L.asStmt();
1838}
1839
1840static const Stmt *getStmtParent(const Stmt *S, const ParentMap &PM) {
1841  if (!S)
1842    return 0;
1843
1844  while (true) {
1845    S = PM.getParentIgnoreParens(S);
1846
1847    if (!S)
1848      break;
1849
1850    if (isa<ExprWithCleanups>(S) ||
1851        isa<CXXBindTemporaryExpr>(S) ||
1852        isa<SubstNonTypeTemplateParmExpr>(S))
1853      continue;
1854
1855    break;
1856  }
1857
1858  return S;
1859}
1860
1861static bool isConditionForTerminator(const Stmt *S, const Stmt *Cond) {
1862  switch (S->getStmtClass()) {
1863    case Stmt::BinaryOperatorClass: {
1864      const BinaryOperator *BO = cast<BinaryOperator>(S);
1865      if (!BO->isLogicalOp())
1866        return false;
1867      return BO->getLHS() == Cond || BO->getRHS() == Cond;
1868    }
1869    case Stmt::IfStmtClass:
1870      return cast<IfStmt>(S)->getCond() == Cond;
1871    case Stmt::ForStmtClass:
1872      return cast<ForStmt>(S)->getCond() == Cond;
1873    case Stmt::WhileStmtClass:
1874      return cast<WhileStmt>(S)->getCond() == Cond;
1875    case Stmt::DoStmtClass:
1876      return cast<DoStmt>(S)->getCond() == Cond;
1877    case Stmt::ChooseExprClass:
1878      return cast<ChooseExpr>(S)->getCond() == Cond;
1879    case Stmt::IndirectGotoStmtClass:
1880      return cast<IndirectGotoStmt>(S)->getTarget() == Cond;
1881    case Stmt::SwitchStmtClass:
1882      return cast<SwitchStmt>(S)->getCond() == Cond;
1883    case Stmt::BinaryConditionalOperatorClass:
1884      return cast<BinaryConditionalOperator>(S)->getCond() == Cond;
1885    case Stmt::ConditionalOperatorClass: {
1886      const ConditionalOperator *CO = cast<ConditionalOperator>(S);
1887      return CO->getCond() == Cond ||
1888             CO->getLHS() == Cond ||
1889             CO->getRHS() == Cond;
1890    }
1891    case Stmt::ObjCForCollectionStmtClass:
1892      return cast<ObjCForCollectionStmt>(S)->getElement() == Cond;
1893    case Stmt::CXXForRangeStmtClass: {
1894      const CXXForRangeStmt *FRS = cast<CXXForRangeStmt>(S);
1895      return FRS->getCond() == Cond || FRS->getRangeInit() == Cond;
1896    }
1897    default:
1898      return false;
1899  }
1900}
1901
1902static bool isIncrementOrInitInForLoop(const Stmt *S, const Stmt *FL) {
1903  if (const ForStmt *FS = dyn_cast<ForStmt>(FL))
1904    return FS->getInc() == S || FS->getInit() == S;
1905  if (const CXXForRangeStmt *FRS = dyn_cast<CXXForRangeStmt>(FL))
1906    return FRS->getInc() == S || FRS->getRangeStmt() == S ||
1907           FRS->getLoopVarStmt() || FRS->getRangeInit() == S;
1908  return false;
1909}
1910
1911typedef llvm::DenseSet<const PathDiagnosticCallPiece *>
1912        OptimizedCallsSet;
1913
1914/// Adds synthetic edges from top-level statements to their subexpressions.
1915///
1916/// This avoids a "swoosh" effect, where an edge from a top-level statement A
1917/// points to a sub-expression B.1 that's not at the start of B. In these cases,
1918/// we'd like to see an edge from A to B, then another one from B to B.1.
1919static void addContextEdges(PathPieces &pieces, SourceManager &SM,
1920                            const ParentMap &PM, const LocationContext *LCtx) {
1921  PathPieces::iterator Prev = pieces.end();
1922  for (PathPieces::iterator I = pieces.begin(), E = Prev; I != E;
1923       Prev = I, ++I) {
1924    PathDiagnosticControlFlowPiece *Piece =
1925      dyn_cast<PathDiagnosticControlFlowPiece>(*I);
1926
1927    if (!Piece)
1928      continue;
1929
1930    PathDiagnosticLocation SrcLoc = Piece->getStartLocation();
1931    SmallVector<PathDiagnosticLocation, 4> SrcContexts;
1932
1933    PathDiagnosticLocation NextSrcContext = SrcLoc;
1934    const Stmt *InnerStmt = 0;
1935    while (NextSrcContext.isValid() && NextSrcContext.asStmt() != InnerStmt) {
1936      SrcContexts.push_back(NextSrcContext);
1937      InnerStmt = NextSrcContext.asStmt();
1938      NextSrcContext = getEnclosingStmtLocation(InnerStmt, SM, PM, LCtx,
1939                                                /*allowNested=*/true);
1940    }
1941
1942    // Repeatedly split the edge as necessary.
1943    // This is important for nested logical expressions (||, &&, ?:) where we
1944    // want to show all the levels of context.
1945    while (true) {
1946      const Stmt *Dst = getLocStmt(Piece->getEndLocation());
1947
1948      // We are looking at an edge. Is the destination within a larger
1949      // expression?
1950      PathDiagnosticLocation DstContext =
1951        getEnclosingStmtLocation(Dst, SM, PM, LCtx, /*allowNested=*/true);
1952      if (!DstContext.isValid() || DstContext.asStmt() == Dst)
1953        break;
1954
1955      // If the source is in the same context, we're already good.
1956      if (std::find(SrcContexts.begin(), SrcContexts.end(), DstContext) !=
1957          SrcContexts.end())
1958        break;
1959
1960      // Update the subexpression node to point to the context edge.
1961      Piece->setStartLocation(DstContext);
1962
1963      // Try to extend the previous edge if it's at the same level as the source
1964      // context.
1965      if (Prev != E) {
1966        PathDiagnosticControlFlowPiece *PrevPiece =
1967          dyn_cast<PathDiagnosticControlFlowPiece>(*Prev);
1968
1969        if (PrevPiece) {
1970          if (const Stmt *PrevSrc = getLocStmt(PrevPiece->getStartLocation())) {
1971            const Stmt *PrevSrcParent = getStmtParent(PrevSrc, PM);
1972            if (PrevSrcParent == getStmtParent(getLocStmt(DstContext), PM)) {
1973              PrevPiece->setEndLocation(DstContext);
1974              break;
1975            }
1976          }
1977        }
1978      }
1979
1980      // Otherwise, split the current edge into a context edge and a
1981      // subexpression edge. Note that the context statement may itself have
1982      // context.
1983      Piece = new PathDiagnosticControlFlowPiece(SrcLoc, DstContext);
1984      I = pieces.insert(I, Piece);
1985    }
1986  }
1987}
1988
1989/// \brief Move edges from a branch condition to a branch target
1990///        when the condition is simple.
1991///
1992/// This restructures some of the work of addContextEdges.  That function
1993/// creates edges this may destroy, but they work together to create a more
1994/// aesthetically set of edges around branches.  After the call to
1995/// addContextEdges, we may have (1) an edge to the branch, (2) an edge from
1996/// the branch to the branch condition, and (3) an edge from the branch
1997/// condition to the branch target.  We keep (1), but may wish to remove (2)
1998/// and move the source of (3) to the branch if the branch condition is simple.
1999///
2000static void simplifySimpleBranches(PathPieces &pieces) {
2001  for (PathPieces::iterator I = pieces.begin(), E = pieces.end(); I != E; ++I) {
2002
2003    PathDiagnosticControlFlowPiece *PieceI =
2004      dyn_cast<PathDiagnosticControlFlowPiece>(*I);
2005
2006    if (!PieceI)
2007      continue;
2008
2009    const Stmt *s1Start = getLocStmt(PieceI->getStartLocation());
2010    const Stmt *s1End   = getLocStmt(PieceI->getEndLocation());
2011
2012    if (!s1Start || !s1End)
2013      continue;
2014
2015    PathPieces::iterator NextI = I; ++NextI;
2016    if (NextI == E)
2017      break;
2018
2019    PathDiagnosticControlFlowPiece *PieceNextI = 0;
2020
2021    while (true) {
2022      if (NextI == E)
2023        break;
2024
2025      PathDiagnosticEventPiece *EV = dyn_cast<PathDiagnosticEventPiece>(*NextI);
2026      if (EV) {
2027        StringRef S = EV->getString();
2028        if (S == StrEnteringLoop || S == StrLoopBodyZero) {
2029          ++NextI;
2030          continue;
2031        }
2032        break;
2033      }
2034
2035      PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(*NextI);
2036      break;
2037    }
2038
2039    if (!PieceNextI)
2040      continue;
2041
2042    const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation());
2043    const Stmt *s2End   = getLocStmt(PieceNextI->getEndLocation());
2044
2045    if (!s2Start || !s2End || s1End != s2Start)
2046      continue;
2047
2048    // We only perform this transformation for specific branch kinds.
2049    // We don't want to do this for do..while, for example.
2050    if (!(isa<ForStmt>(s1Start) || isa<WhileStmt>(s1Start) ||
2051          isa<IfStmt>(s1Start) || isa<ObjCForCollectionStmt>(s1Start) ||
2052          isa<CXXForRangeStmt>(s1Start)))
2053      continue;
2054
2055    // Is s1End the branch condition?
2056    if (!isConditionForTerminator(s1Start, s1End))
2057      continue;
2058
2059    // Perform the hoisting by eliminating (2) and changing the start
2060    // location of (3).
2061    PieceNextI->setStartLocation(PieceI->getStartLocation());
2062    I = pieces.erase(I);
2063  }
2064}
2065
2066/// Returns the number of bytes in the given (character-based) SourceRange.
2067///
2068/// If the locations in the range are not on the same line, returns None.
2069///
2070/// Note that this does not do a precise user-visible character or column count.
2071static Optional<size_t> getLengthOnSingleLine(SourceManager &SM,
2072                                              SourceRange Range) {
2073  SourceRange ExpansionRange(SM.getExpansionLoc(Range.getBegin()),
2074                             SM.getExpansionRange(Range.getEnd()).second);
2075
2076  FileID FID = SM.getFileID(ExpansionRange.getBegin());
2077  if (FID != SM.getFileID(ExpansionRange.getEnd()))
2078    return None;
2079
2080  bool Invalid;
2081  const llvm::MemoryBuffer *Buffer = SM.getBuffer(FID, &Invalid);
2082  if (Invalid)
2083    return None;
2084
2085  unsigned BeginOffset = SM.getFileOffset(ExpansionRange.getBegin());
2086  unsigned EndOffset = SM.getFileOffset(ExpansionRange.getEnd());
2087  StringRef Snippet = Buffer->getBuffer().slice(BeginOffset, EndOffset);
2088
2089  // We're searching the raw bytes of the buffer here, which might include
2090  // escaped newlines and such. That's okay; we're trying to decide whether the
2091  // SourceRange is covering a large or small amount of space in the user's
2092  // editor.
2093  if (Snippet.find_first_of("\r\n") != StringRef::npos)
2094    return None;
2095
2096  // This isn't Unicode-aware, but it doesn't need to be.
2097  return Snippet.size();
2098}
2099
2100/// \sa getLengthOnSingleLine(SourceManager, SourceRange)
2101static Optional<size_t> getLengthOnSingleLine(SourceManager &SM,
2102                                              const Stmt *S) {
2103  return getLengthOnSingleLine(SM, S->getSourceRange());
2104}
2105
2106/// Eliminate two-edge cycles created by addContextEdges().
2107///
2108/// Once all the context edges are in place, there are plenty of cases where
2109/// there's a single edge from a top-level statement to a subexpression,
2110/// followed by a single path note, and then a reverse edge to get back out to
2111/// the top level. If the statement is simple enough, the subexpression edges
2112/// just add noise and make it harder to understand what's going on.
2113///
2114/// This function only removes edges in pairs, because removing only one edge
2115/// might leave other edges dangling.
2116///
2117/// This will not remove edges in more complicated situations:
2118/// - if there is more than one "hop" leading to or from a subexpression.
2119/// - if there is an inlined call between the edges instead of a single event.
2120/// - if the whole statement is large enough that having subexpression arrows
2121///   might be helpful.
2122static void removeContextCycles(PathPieces &Path, SourceManager &SM,
2123                                ParentMap &PM) {
2124  for (PathPieces::iterator I = Path.begin(), E = Path.end(); I != E; ) {
2125    // Pattern match the current piece and its successor.
2126    PathDiagnosticControlFlowPiece *PieceI =
2127      dyn_cast<PathDiagnosticControlFlowPiece>(*I);
2128
2129    if (!PieceI) {
2130      ++I;
2131      continue;
2132    }
2133
2134    const Stmt *s1Start = getLocStmt(PieceI->getStartLocation());
2135    const Stmt *s1End   = getLocStmt(PieceI->getEndLocation());
2136
2137    PathPieces::iterator NextI = I; ++NextI;
2138    if (NextI == E)
2139      break;
2140
2141    PathDiagnosticControlFlowPiece *PieceNextI =
2142      dyn_cast<PathDiagnosticControlFlowPiece>(*NextI);
2143
2144    if (!PieceNextI) {
2145      if (isa<PathDiagnosticEventPiece>(*NextI)) {
2146        ++NextI;
2147        if (NextI == E)
2148          break;
2149        PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(*NextI);
2150      }
2151
2152      if (!PieceNextI) {
2153        ++I;
2154        continue;
2155      }
2156    }
2157
2158    const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation());
2159    const Stmt *s2End   = getLocStmt(PieceNextI->getEndLocation());
2160
2161    if (s1Start && s2Start && s1Start == s2End && s2Start == s1End) {
2162      const size_t MAX_SHORT_LINE_LENGTH = 80;
2163      Optional<size_t> s1Length = getLengthOnSingleLine(SM, s1Start);
2164      if (s1Length && *s1Length <= MAX_SHORT_LINE_LENGTH) {
2165        Optional<size_t> s2Length = getLengthOnSingleLine(SM, s2Start);
2166        if (s2Length && *s2Length <= MAX_SHORT_LINE_LENGTH) {
2167          Path.erase(I);
2168          I = Path.erase(NextI);
2169          continue;
2170        }
2171      }
2172    }
2173
2174    ++I;
2175  }
2176}
2177
2178/// \brief Return true if X is contained by Y.
2179static bool lexicalContains(ParentMap &PM,
2180                            const Stmt *X,
2181                            const Stmt *Y) {
2182  while (X) {
2183    if (X == Y)
2184      return true;
2185    X = PM.getParent(X);
2186  }
2187  return false;
2188}
2189
2190// Remove short edges on the same line less than 3 columns in difference.
2191static void removePunyEdges(PathPieces &path,
2192                            SourceManager &SM,
2193                            ParentMap &PM) {
2194
2195  bool erased = false;
2196
2197  for (PathPieces::iterator I = path.begin(), E = path.end(); I != E;
2198       erased ? I : ++I) {
2199
2200    erased = false;
2201
2202    PathDiagnosticControlFlowPiece *PieceI =
2203      dyn_cast<PathDiagnosticControlFlowPiece>(*I);
2204
2205    if (!PieceI)
2206      continue;
2207
2208    const Stmt *start = getLocStmt(PieceI->getStartLocation());
2209    const Stmt *end   = getLocStmt(PieceI->getEndLocation());
2210
2211    if (!start || !end)
2212      continue;
2213
2214    const Stmt *endParent = PM.getParent(end);
2215    if (!endParent)
2216      continue;
2217
2218    if (isConditionForTerminator(end, endParent))
2219      continue;
2220
2221    SourceLocation FirstLoc = start->getLocStart();
2222    SourceLocation SecondLoc = end->getLocStart();
2223
2224    if (!SM.isFromSameFile(FirstLoc, SecondLoc))
2225      continue;
2226    if (SM.isBeforeInTranslationUnit(SecondLoc, FirstLoc))
2227      std::swap(SecondLoc, FirstLoc);
2228
2229    SourceRange EdgeRange(FirstLoc, SecondLoc);
2230    Optional<size_t> ByteWidth = getLengthOnSingleLine(SM, EdgeRange);
2231
2232    // If the statements are on different lines, continue.
2233    if (!ByteWidth)
2234      continue;
2235
2236    const size_t MAX_PUNY_EDGE_LENGTH = 2;
2237    if (*ByteWidth <= MAX_PUNY_EDGE_LENGTH) {
2238      // FIXME: There are enough /bytes/ between the endpoints of the edge, but
2239      // there might not be enough /columns/. A proper user-visible column count
2240      // is probably too expensive, though.
2241      I = path.erase(I);
2242      erased = true;
2243      continue;
2244    }
2245  }
2246}
2247
2248static void removeIdenticalEvents(PathPieces &path) {
2249  for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ++I) {
2250    PathDiagnosticEventPiece *PieceI =
2251      dyn_cast<PathDiagnosticEventPiece>(*I);
2252
2253    if (!PieceI)
2254      continue;
2255
2256    PathPieces::iterator NextI = I; ++NextI;
2257    if (NextI == E)
2258      return;
2259
2260    PathDiagnosticEventPiece *PieceNextI =
2261      dyn_cast<PathDiagnosticEventPiece>(*NextI);
2262
2263    if (!PieceNextI)
2264      continue;
2265
2266    // Erase the second piece if it has the same exact message text.
2267    if (PieceI->getString() == PieceNextI->getString()) {
2268      path.erase(NextI);
2269    }
2270  }
2271}
2272
2273static bool optimizeEdges(PathPieces &path, SourceManager &SM,
2274                          OptimizedCallsSet &OCS,
2275                          LocationContextMap &LCM) {
2276  bool hasChanges = false;
2277  const LocationContext *LC = LCM[&path];
2278  assert(LC);
2279  ParentMap &PM = LC->getParentMap();
2280
2281  for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ) {
2282    // Optimize subpaths.
2283    if (PathDiagnosticCallPiece *CallI = dyn_cast<PathDiagnosticCallPiece>(*I)){
2284      // Record the fact that a call has been optimized so we only do the
2285      // effort once.
2286      if (!OCS.count(CallI)) {
2287        while (optimizeEdges(CallI->path, SM, OCS, LCM)) {}
2288        OCS.insert(CallI);
2289      }
2290      ++I;
2291      continue;
2292    }
2293
2294    // Pattern match the current piece and its successor.
2295    PathDiagnosticControlFlowPiece *PieceI =
2296      dyn_cast<PathDiagnosticControlFlowPiece>(*I);
2297
2298    if (!PieceI) {
2299      ++I;
2300      continue;
2301    }
2302
2303    const Stmt *s1Start = getLocStmt(PieceI->getStartLocation());
2304    const Stmt *s1End   = getLocStmt(PieceI->getEndLocation());
2305    const Stmt *level1 = getStmtParent(s1Start, PM);
2306    const Stmt *level2 = getStmtParent(s1End, PM);
2307
2308    PathPieces::iterator NextI = I; ++NextI;
2309    if (NextI == E)
2310      break;
2311
2312    PathDiagnosticControlFlowPiece *PieceNextI =
2313      dyn_cast<PathDiagnosticControlFlowPiece>(*NextI);
2314
2315    if (!PieceNextI) {
2316      ++I;
2317      continue;
2318    }
2319
2320    const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation());
2321    const Stmt *s2End   = getLocStmt(PieceNextI->getEndLocation());
2322    const Stmt *level3 = getStmtParent(s2Start, PM);
2323    const Stmt *level4 = getStmtParent(s2End, PM);
2324
2325    // Rule I.
2326    //
2327    // If we have two consecutive control edges whose end/begin locations
2328    // are at the same level (e.g. statements or top-level expressions within
2329    // a compound statement, or siblings share a single ancestor expression),
2330    // then merge them if they have no interesting intermediate event.
2331    //
2332    // For example:
2333    //
2334    // (1.1 -> 1.2) -> (1.2 -> 1.3) becomes (1.1 -> 1.3) because the common
2335    // parent is '1'.  Here 'x.y.z' represents the hierarchy of statements.
2336    //
2337    // NOTE: this will be limited later in cases where we add barriers
2338    // to prevent this optimization.
2339    //
2340    if (level1 && level1 == level2 && level1 == level3 && level1 == level4) {
2341      PieceI->setEndLocation(PieceNextI->getEndLocation());
2342      path.erase(NextI);
2343      hasChanges = true;
2344      continue;
2345    }
2346
2347    // Rule II.
2348    //
2349    // Eliminate edges between subexpressions and parent expressions
2350    // when the subexpression is consumed.
2351    //
2352    // NOTE: this will be limited later in cases where we add barriers
2353    // to prevent this optimization.
2354    //
2355    if (s1End && s1End == s2Start && level2) {
2356      bool removeEdge = false;
2357      // Remove edges into the increment or initialization of a
2358      // loop that have no interleaving event.  This means that
2359      // they aren't interesting.
2360      if (isIncrementOrInitInForLoop(s1End, level2))
2361        removeEdge = true;
2362      // Next only consider edges that are not anchored on
2363      // the condition of a terminator.  This are intermediate edges
2364      // that we might want to trim.
2365      else if (!isConditionForTerminator(level2, s1End)) {
2366        // Trim edges on expressions that are consumed by
2367        // the parent expression.
2368        if (isa<Expr>(s1End) && PM.isConsumedExpr(cast<Expr>(s1End))) {
2369          removeEdge = true;
2370        }
2371        // Trim edges where a lexical containment doesn't exist.
2372        // For example:
2373        //
2374        //  X -> Y -> Z
2375        //
2376        // If 'Z' lexically contains Y (it is an ancestor) and
2377        // 'X' does not lexically contain Y (it is a descendant OR
2378        // it has no lexical relationship at all) then trim.
2379        //
2380        // This can eliminate edges where we dive into a subexpression
2381        // and then pop back out, etc.
2382        else if (s1Start && s2End &&
2383                 lexicalContains(PM, s2Start, s2End) &&
2384                 !lexicalContains(PM, s1End, s1Start)) {
2385          removeEdge = true;
2386        }
2387        // Trim edges from a subexpression back to the top level if the
2388        // subexpression is on a different line.
2389        //
2390        // A.1 -> A -> B
2391        // becomes
2392        // A.1 -> B
2393        //
2394        // These edges just look ugly and don't usually add anything.
2395        else if (s1Start && s2End &&
2396                 lexicalContains(PM, s1Start, s1End)) {
2397          SourceRange EdgeRange(PieceI->getEndLocation().asLocation(),
2398                                PieceI->getStartLocation().asLocation());
2399          if (!getLengthOnSingleLine(SM, EdgeRange).hasValue())
2400            removeEdge = true;
2401        }
2402      }
2403
2404      if (removeEdge) {
2405        PieceI->setEndLocation(PieceNextI->getEndLocation());
2406        path.erase(NextI);
2407        hasChanges = true;
2408        continue;
2409      }
2410    }
2411
2412    // Optimize edges for ObjC fast-enumeration loops.
2413    //
2414    // (X -> collection) -> (collection -> element)
2415    //
2416    // becomes:
2417    //
2418    // (X -> element)
2419    if (s1End == s2Start) {
2420      const ObjCForCollectionStmt *FS =
2421        dyn_cast_or_null<ObjCForCollectionStmt>(level3);
2422      if (FS && FS->getCollection()->IgnoreParens() == s2Start &&
2423          s2End == FS->getElement()) {
2424        PieceI->setEndLocation(PieceNextI->getEndLocation());
2425        path.erase(NextI);
2426        hasChanges = true;
2427        continue;
2428      }
2429    }
2430
2431    // No changes at this index?  Move to the next one.
2432    ++I;
2433  }
2434
2435  if (!hasChanges) {
2436    // Adjust edges into subexpressions to make them more uniform
2437    // and aesthetically pleasing.
2438    addContextEdges(path, SM, PM, LC);
2439    // Remove "cyclical" edges that include one or more context edges.
2440    removeContextCycles(path, SM, PM);
2441    // Hoist edges originating from branch conditions to branches
2442    // for simple branches.
2443    simplifySimpleBranches(path);
2444    // Remove any puny edges left over after primary optimization pass.
2445    removePunyEdges(path, SM, PM);
2446    // Remove identical events.
2447    removeIdenticalEvents(path);
2448  }
2449
2450  return hasChanges;
2451}
2452
2453/// Drop the very first edge in a path, which should be a function entry edge.
2454///
2455/// If the first edge is not a function entry edge (say, because the first
2456/// statement had an invalid source location), this function does nothing.
2457// FIXME: We should just generate invalid edges anyway and have the optimizer
2458// deal with them.
2459static void dropFunctionEntryEdge(PathPieces &Path,
2460                                  LocationContextMap &LCM,
2461                                  SourceManager &SM) {
2462  const PathDiagnosticControlFlowPiece *FirstEdge =
2463    dyn_cast<PathDiagnosticControlFlowPiece>(Path.front());
2464  if (!FirstEdge)
2465    return;
2466
2467  const Decl *D = LCM[&Path]->getDecl();
2468  PathDiagnosticLocation EntryLoc = PathDiagnosticLocation::createBegin(D, SM);
2469  if (FirstEdge->getStartLocation() != EntryLoc)
2470    return;
2471
2472  Path.pop_front();
2473}
2474
2475
2476//===----------------------------------------------------------------------===//
2477// Methods for BugType and subclasses.
2478//===----------------------------------------------------------------------===//
2479BugType::~BugType() { }
2480
2481void BugType::FlushReports(BugReporter &BR) {}
2482
2483void BuiltinBug::anchor() {}
2484
2485//===----------------------------------------------------------------------===//
2486// Methods for BugReport and subclasses.
2487//===----------------------------------------------------------------------===//
2488
2489void BugReport::NodeResolver::anchor() {}
2490
2491void BugReport::addVisitor(BugReporterVisitor* visitor) {
2492  if (!visitor)
2493    return;
2494
2495  llvm::FoldingSetNodeID ID;
2496  visitor->Profile(ID);
2497  void *InsertPos;
2498
2499  if (CallbacksSet.FindNodeOrInsertPos(ID, InsertPos)) {
2500    delete visitor;
2501    return;
2502  }
2503
2504  CallbacksSet.InsertNode(visitor, InsertPos);
2505  Callbacks.push_back(visitor);
2506  ++ConfigurationChangeToken;
2507}
2508
2509BugReport::~BugReport() {
2510  for (visitor_iterator I = visitor_begin(), E = visitor_end(); I != E; ++I) {
2511    delete *I;
2512  }
2513  while (!interestingSymbols.empty()) {
2514    popInterestingSymbolsAndRegions();
2515  }
2516}
2517
2518const Decl *BugReport::getDeclWithIssue() const {
2519  if (DeclWithIssue)
2520    return DeclWithIssue;
2521
2522  const ExplodedNode *N = getErrorNode();
2523  if (!N)
2524    return 0;
2525
2526  const LocationContext *LC = N->getLocationContext();
2527  return LC->getCurrentStackFrame()->getDecl();
2528}
2529
2530void BugReport::Profile(llvm::FoldingSetNodeID& hash) const {
2531  hash.AddPointer(&BT);
2532  hash.AddString(Description);
2533  PathDiagnosticLocation UL = getUniqueingLocation();
2534  if (UL.isValid()) {
2535    UL.Profile(hash);
2536  } else if (Location.isValid()) {
2537    Location.Profile(hash);
2538  } else {
2539    assert(ErrorNode);
2540    hash.AddPointer(GetCurrentOrPreviousStmt(ErrorNode));
2541  }
2542
2543  for (SmallVectorImpl<SourceRange>::const_iterator I =
2544      Ranges.begin(), E = Ranges.end(); I != E; ++I) {
2545    const SourceRange range = *I;
2546    if (!range.isValid())
2547      continue;
2548    hash.AddInteger(range.getBegin().getRawEncoding());
2549    hash.AddInteger(range.getEnd().getRawEncoding());
2550  }
2551}
2552
2553void BugReport::markInteresting(SymbolRef sym) {
2554  if (!sym)
2555    return;
2556
2557  // If the symbol wasn't already in our set, note a configuration change.
2558  if (getInterestingSymbols().insert(sym).second)
2559    ++ConfigurationChangeToken;
2560
2561  if (const SymbolMetadata *meta = dyn_cast<SymbolMetadata>(sym))
2562    getInterestingRegions().insert(meta->getRegion());
2563}
2564
2565void BugReport::markInteresting(const MemRegion *R) {
2566  if (!R)
2567    return;
2568
2569  // If the base region wasn't already in our set, note a configuration change.
2570  R = R->getBaseRegion();
2571  if (getInterestingRegions().insert(R).second)
2572    ++ConfigurationChangeToken;
2573
2574  if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R))
2575    getInterestingSymbols().insert(SR->getSymbol());
2576}
2577
2578void BugReport::markInteresting(SVal V) {
2579  markInteresting(V.getAsRegion());
2580  markInteresting(V.getAsSymbol());
2581}
2582
2583void BugReport::markInteresting(const LocationContext *LC) {
2584  if (!LC)
2585    return;
2586  InterestingLocationContexts.insert(LC);
2587}
2588
2589bool BugReport::isInteresting(SVal V) {
2590  return isInteresting(V.getAsRegion()) || isInteresting(V.getAsSymbol());
2591}
2592
2593bool BugReport::isInteresting(SymbolRef sym) {
2594  if (!sym)
2595    return false;
2596  // We don't currently consider metadata symbols to be interesting
2597  // even if we know their region is interesting. Is that correct behavior?
2598  return getInterestingSymbols().count(sym);
2599}
2600
2601bool BugReport::isInteresting(const MemRegion *R) {
2602  if (!R)
2603    return false;
2604  R = R->getBaseRegion();
2605  bool b = getInterestingRegions().count(R);
2606  if (b)
2607    return true;
2608  if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R))
2609    return getInterestingSymbols().count(SR->getSymbol());
2610  return false;
2611}
2612
2613bool BugReport::isInteresting(const LocationContext *LC) {
2614  if (!LC)
2615    return false;
2616  return InterestingLocationContexts.count(LC);
2617}
2618
2619void BugReport::lazyInitializeInterestingSets() {
2620  if (interestingSymbols.empty()) {
2621    interestingSymbols.push_back(new Symbols());
2622    interestingRegions.push_back(new Regions());
2623  }
2624}
2625
2626BugReport::Symbols &BugReport::getInterestingSymbols() {
2627  lazyInitializeInterestingSets();
2628  return *interestingSymbols.back();
2629}
2630
2631BugReport::Regions &BugReport::getInterestingRegions() {
2632  lazyInitializeInterestingSets();
2633  return *interestingRegions.back();
2634}
2635
2636void BugReport::pushInterestingSymbolsAndRegions() {
2637  interestingSymbols.push_back(new Symbols(getInterestingSymbols()));
2638  interestingRegions.push_back(new Regions(getInterestingRegions()));
2639}
2640
2641void BugReport::popInterestingSymbolsAndRegions() {
2642  delete interestingSymbols.back();
2643  interestingSymbols.pop_back();
2644  delete interestingRegions.back();
2645  interestingRegions.pop_back();
2646}
2647
2648const Stmt *BugReport::getStmt() const {
2649  if (!ErrorNode)
2650    return 0;
2651
2652  ProgramPoint ProgP = ErrorNode->getLocation();
2653  const Stmt *S = NULL;
2654
2655  if (Optional<BlockEntrance> BE = ProgP.getAs<BlockEntrance>()) {
2656    CFGBlock &Exit = ProgP.getLocationContext()->getCFG()->getExit();
2657    if (BE->getBlock() == &Exit)
2658      S = GetPreviousStmt(ErrorNode);
2659  }
2660  if (!S)
2661    S = PathDiagnosticLocation::getStmt(ErrorNode);
2662
2663  return S;
2664}
2665
2666std::pair<BugReport::ranges_iterator, BugReport::ranges_iterator>
2667BugReport::getRanges() {
2668    // If no custom ranges, add the range of the statement corresponding to
2669    // the error node.
2670    if (Ranges.empty()) {
2671      if (const Expr *E = dyn_cast_or_null<Expr>(getStmt()))
2672        addRange(E->getSourceRange());
2673      else
2674        return std::make_pair(ranges_iterator(), ranges_iterator());
2675    }
2676
2677    // User-specified absence of range info.
2678    if (Ranges.size() == 1 && !Ranges.begin()->isValid())
2679      return std::make_pair(ranges_iterator(), ranges_iterator());
2680
2681    return std::make_pair(Ranges.begin(), Ranges.end());
2682}
2683
2684PathDiagnosticLocation BugReport::getLocation(const SourceManager &SM) const {
2685  if (ErrorNode) {
2686    assert(!Location.isValid() &&
2687     "Either Location or ErrorNode should be specified but not both.");
2688    return PathDiagnosticLocation::createEndOfPath(ErrorNode, SM);
2689  } else {
2690    assert(Location.isValid());
2691    return Location;
2692  }
2693
2694  return PathDiagnosticLocation();
2695}
2696
2697//===----------------------------------------------------------------------===//
2698// Methods for BugReporter and subclasses.
2699//===----------------------------------------------------------------------===//
2700
2701BugReportEquivClass::~BugReportEquivClass() { }
2702GRBugReporter::~GRBugReporter() { }
2703BugReporterData::~BugReporterData() {}
2704
2705ExplodedGraph &GRBugReporter::getGraph() { return Eng.getGraph(); }
2706
2707ProgramStateManager&
2708GRBugReporter::getStateManager() { return Eng.getStateManager(); }
2709
2710BugReporter::~BugReporter() {
2711  FlushReports();
2712
2713  // Free the bug reports we are tracking.
2714  typedef std::vector<BugReportEquivClass *> ContTy;
2715  for (ContTy::iterator I = EQClassesVector.begin(), E = EQClassesVector.end();
2716       I != E; ++I) {
2717    delete *I;
2718  }
2719}
2720
2721void BugReporter::FlushReports() {
2722  if (BugTypes.isEmpty())
2723    return;
2724
2725  // First flush the warnings for each BugType.  This may end up creating new
2726  // warnings and new BugTypes.
2727  // FIXME: Only NSErrorChecker needs BugType's FlushReports.
2728  // Turn NSErrorChecker into a proper checker and remove this.
2729  SmallVector<const BugType*, 16> bugTypes;
2730  for (BugTypesTy::iterator I=BugTypes.begin(), E=BugTypes.end(); I!=E; ++I)
2731    bugTypes.push_back(*I);
2732  for (SmallVector<const BugType*, 16>::iterator
2733         I = bugTypes.begin(), E = bugTypes.end(); I != E; ++I)
2734    const_cast<BugType*>(*I)->FlushReports(*this);
2735
2736  // We need to flush reports in deterministic order to ensure the order
2737  // of the reports is consistent between runs.
2738  typedef std::vector<BugReportEquivClass *> ContVecTy;
2739  for (ContVecTy::iterator EI=EQClassesVector.begin(), EE=EQClassesVector.end();
2740       EI != EE; ++EI){
2741    BugReportEquivClass& EQ = **EI;
2742    FlushReport(EQ);
2743  }
2744
2745  // BugReporter owns and deletes only BugTypes created implicitly through
2746  // EmitBasicReport.
2747  // FIXME: There are leaks from checkers that assume that the BugTypes they
2748  // create will be destroyed by the BugReporter.
2749  for (llvm::StringMap<BugType*>::iterator
2750         I = StrBugTypes.begin(), E = StrBugTypes.end(); I != E; ++I)
2751    delete I->second;
2752
2753  // Remove all references to the BugType objects.
2754  BugTypes = F.getEmptySet();
2755}
2756
2757//===----------------------------------------------------------------------===//
2758// PathDiagnostics generation.
2759//===----------------------------------------------------------------------===//
2760
2761namespace {
2762/// A wrapper around a report graph, which contains only a single path, and its
2763/// node maps.
2764class ReportGraph {
2765public:
2766  InterExplodedGraphMap BackMap;
2767  OwningPtr<ExplodedGraph> Graph;
2768  const ExplodedNode *ErrorNode;
2769  size_t Index;
2770};
2771
2772/// A wrapper around a trimmed graph and its node maps.
2773class TrimmedGraph {
2774  InterExplodedGraphMap InverseMap;
2775
2776  typedef llvm::DenseMap<const ExplodedNode *, unsigned> PriorityMapTy;
2777  PriorityMapTy PriorityMap;
2778
2779  typedef std::pair<const ExplodedNode *, size_t> NodeIndexPair;
2780  SmallVector<NodeIndexPair, 32> ReportNodes;
2781
2782  OwningPtr<ExplodedGraph> G;
2783
2784  /// A helper class for sorting ExplodedNodes by priority.
2785  template <bool Descending>
2786  class PriorityCompare {
2787    const PriorityMapTy &PriorityMap;
2788
2789  public:
2790    PriorityCompare(const PriorityMapTy &M) : PriorityMap(M) {}
2791
2792    bool operator()(const ExplodedNode *LHS, const ExplodedNode *RHS) const {
2793      PriorityMapTy::const_iterator LI = PriorityMap.find(LHS);
2794      PriorityMapTy::const_iterator RI = PriorityMap.find(RHS);
2795      PriorityMapTy::const_iterator E = PriorityMap.end();
2796
2797      if (LI == E)
2798        return Descending;
2799      if (RI == E)
2800        return !Descending;
2801
2802      return Descending ? LI->second > RI->second
2803                        : LI->second < RI->second;
2804    }
2805
2806    bool operator()(const NodeIndexPair &LHS, const NodeIndexPair &RHS) const {
2807      return (*this)(LHS.first, RHS.first);
2808    }
2809  };
2810
2811public:
2812  TrimmedGraph(const ExplodedGraph *OriginalGraph,
2813               ArrayRef<const ExplodedNode *> Nodes);
2814
2815  bool popNextReportGraph(ReportGraph &GraphWrapper);
2816};
2817}
2818
2819TrimmedGraph::TrimmedGraph(const ExplodedGraph *OriginalGraph,
2820                           ArrayRef<const ExplodedNode *> Nodes) {
2821  // The trimmed graph is created in the body of the constructor to ensure
2822  // that the DenseMaps have been initialized already.
2823  InterExplodedGraphMap ForwardMap;
2824  G.reset(OriginalGraph->trim(Nodes, &ForwardMap, &InverseMap));
2825
2826  // Find the (first) error node in the trimmed graph.  We just need to consult
2827  // the node map which maps from nodes in the original graph to nodes
2828  // in the new graph.
2829  llvm::SmallPtrSet<const ExplodedNode *, 32> RemainingNodes;
2830
2831  for (unsigned i = 0, count = Nodes.size(); i < count; ++i) {
2832    if (const ExplodedNode *NewNode = ForwardMap.lookup(Nodes[i])) {
2833      ReportNodes.push_back(std::make_pair(NewNode, i));
2834      RemainingNodes.insert(NewNode);
2835    }
2836  }
2837
2838  assert(!RemainingNodes.empty() && "No error node found in the trimmed graph");
2839
2840  // Perform a forward BFS to find all the shortest paths.
2841  std::queue<const ExplodedNode *> WS;
2842
2843  assert(G->num_roots() == 1);
2844  WS.push(*G->roots_begin());
2845  unsigned Priority = 0;
2846
2847  while (!WS.empty()) {
2848    const ExplodedNode *Node = WS.front();
2849    WS.pop();
2850
2851    PriorityMapTy::iterator PriorityEntry;
2852    bool IsNew;
2853    llvm::tie(PriorityEntry, IsNew) =
2854      PriorityMap.insert(std::make_pair(Node, Priority));
2855    ++Priority;
2856
2857    if (!IsNew) {
2858      assert(PriorityEntry->second <= Priority);
2859      continue;
2860    }
2861
2862    if (RemainingNodes.erase(Node))
2863      if (RemainingNodes.empty())
2864        break;
2865
2866    for (ExplodedNode::const_pred_iterator I = Node->succ_begin(),
2867                                           E = Node->succ_end();
2868         I != E; ++I)
2869      WS.push(*I);
2870  }
2871
2872  // Sort the error paths from longest to shortest.
2873  std::sort(ReportNodes.begin(), ReportNodes.end(),
2874            PriorityCompare<true>(PriorityMap));
2875}
2876
2877bool TrimmedGraph::popNextReportGraph(ReportGraph &GraphWrapper) {
2878  if (ReportNodes.empty())
2879    return false;
2880
2881  const ExplodedNode *OrigN;
2882  llvm::tie(OrigN, GraphWrapper.Index) = ReportNodes.pop_back_val();
2883  assert(PriorityMap.find(OrigN) != PriorityMap.end() &&
2884         "error node not accessible from root");
2885
2886  // Create a new graph with a single path.  This is the graph
2887  // that will be returned to the caller.
2888  ExplodedGraph *GNew = new ExplodedGraph();
2889  GraphWrapper.Graph.reset(GNew);
2890  GraphWrapper.BackMap.clear();
2891
2892  // Now walk from the error node up the BFS path, always taking the
2893  // predeccessor with the lowest number.
2894  ExplodedNode *Succ = 0;
2895  while (true) {
2896    // Create the equivalent node in the new graph with the same state
2897    // and location.
2898    ExplodedNode *NewN = GNew->getNode(OrigN->getLocation(), OrigN->getState(),
2899                                       OrigN->isSink());
2900
2901    // Store the mapping to the original node.
2902    InterExplodedGraphMap::const_iterator IMitr = InverseMap.find(OrigN);
2903    assert(IMitr != InverseMap.end() && "No mapping to original node.");
2904    GraphWrapper.BackMap[NewN] = IMitr->second;
2905
2906    // Link up the new node with the previous node.
2907    if (Succ)
2908      Succ->addPredecessor(NewN, *GNew);
2909    else
2910      GraphWrapper.ErrorNode = NewN;
2911
2912    Succ = NewN;
2913
2914    // Are we at the final node?
2915    if (OrigN->pred_empty()) {
2916      GNew->addRoot(NewN);
2917      break;
2918    }
2919
2920    // Find the next predeccessor node.  We choose the node that is marked
2921    // with the lowest BFS number.
2922    OrigN = *std::min_element(OrigN->pred_begin(), OrigN->pred_end(),
2923                          PriorityCompare<false>(PriorityMap));
2924  }
2925
2926  return true;
2927}
2928
2929
2930/// CompactPathDiagnostic - This function postprocesses a PathDiagnostic object
2931///  and collapses PathDiagosticPieces that are expanded by macros.
2932static void CompactPathDiagnostic(PathPieces &path, const SourceManager& SM) {
2933  typedef std::vector<std::pair<IntrusiveRefCntPtr<PathDiagnosticMacroPiece>,
2934                                SourceLocation> > MacroStackTy;
2935
2936  typedef std::vector<IntrusiveRefCntPtr<PathDiagnosticPiece> >
2937          PiecesTy;
2938
2939  MacroStackTy MacroStack;
2940  PiecesTy Pieces;
2941
2942  for (PathPieces::const_iterator I = path.begin(), E = path.end();
2943       I!=E; ++I) {
2944
2945    PathDiagnosticPiece *piece = I->getPtr();
2946
2947    // Recursively compact calls.
2948    if (PathDiagnosticCallPiece *call=dyn_cast<PathDiagnosticCallPiece>(piece)){
2949      CompactPathDiagnostic(call->path, SM);
2950    }
2951
2952    // Get the location of the PathDiagnosticPiece.
2953    const FullSourceLoc Loc = piece->getLocation().asLocation();
2954
2955    // Determine the instantiation location, which is the location we group
2956    // related PathDiagnosticPieces.
2957    SourceLocation InstantiationLoc = Loc.isMacroID() ?
2958                                      SM.getExpansionLoc(Loc) :
2959                                      SourceLocation();
2960
2961    if (Loc.isFileID()) {
2962      MacroStack.clear();
2963      Pieces.push_back(piece);
2964      continue;
2965    }
2966
2967    assert(Loc.isMacroID());
2968
2969    // Is the PathDiagnosticPiece within the same macro group?
2970    if (!MacroStack.empty() && InstantiationLoc == MacroStack.back().second) {
2971      MacroStack.back().first->subPieces.push_back(piece);
2972      continue;
2973    }
2974
2975    // We aren't in the same group.  Are we descending into a new macro
2976    // or are part of an old one?
2977    IntrusiveRefCntPtr<PathDiagnosticMacroPiece> MacroGroup;
2978
2979    SourceLocation ParentInstantiationLoc = InstantiationLoc.isMacroID() ?
2980                                          SM.getExpansionLoc(Loc) :
2981                                          SourceLocation();
2982
2983    // Walk the entire macro stack.
2984    while (!MacroStack.empty()) {
2985      if (InstantiationLoc == MacroStack.back().second) {
2986        MacroGroup = MacroStack.back().first;
2987        break;
2988      }
2989
2990      if (ParentInstantiationLoc == MacroStack.back().second) {
2991        MacroGroup = MacroStack.back().first;
2992        break;
2993      }
2994
2995      MacroStack.pop_back();
2996    }
2997
2998    if (!MacroGroup || ParentInstantiationLoc == MacroStack.back().second) {
2999      // Create a new macro group and add it to the stack.
3000      PathDiagnosticMacroPiece *NewGroup =
3001        new PathDiagnosticMacroPiece(
3002          PathDiagnosticLocation::createSingleLocation(piece->getLocation()));
3003
3004      if (MacroGroup)
3005        MacroGroup->subPieces.push_back(NewGroup);
3006      else {
3007        assert(InstantiationLoc.isFileID());
3008        Pieces.push_back(NewGroup);
3009      }
3010
3011      MacroGroup = NewGroup;
3012      MacroStack.push_back(std::make_pair(MacroGroup, InstantiationLoc));
3013    }
3014
3015    // Finally, add the PathDiagnosticPiece to the group.
3016    MacroGroup->subPieces.push_back(piece);
3017  }
3018
3019  // Now take the pieces and construct a new PathDiagnostic.
3020  path.clear();
3021
3022  for (PiecesTy::iterator I=Pieces.begin(), E=Pieces.end(); I!=E; ++I)
3023    path.push_back(*I);
3024}
3025
3026bool GRBugReporter::generatePathDiagnostic(PathDiagnostic& PD,
3027                                           PathDiagnosticConsumer &PC,
3028                                           ArrayRef<BugReport *> &bugReports) {
3029  assert(!bugReports.empty());
3030
3031  bool HasValid = false;
3032  bool HasInvalid = false;
3033  SmallVector<const ExplodedNode *, 32> errorNodes;
3034  for (ArrayRef<BugReport*>::iterator I = bugReports.begin(),
3035                                      E = bugReports.end(); I != E; ++I) {
3036    if ((*I)->isValid()) {
3037      HasValid = true;
3038      errorNodes.push_back((*I)->getErrorNode());
3039    } else {
3040      // Keep the errorNodes list in sync with the bugReports list.
3041      HasInvalid = true;
3042      errorNodes.push_back(0);
3043    }
3044  }
3045
3046  // If all the reports have been marked invalid by a previous path generation,
3047  // we're done.
3048  if (!HasValid)
3049    return false;
3050
3051  typedef PathDiagnosticConsumer::PathGenerationScheme PathGenerationScheme;
3052  PathGenerationScheme ActiveScheme = PC.getGenerationScheme();
3053
3054  if (ActiveScheme == PathDiagnosticConsumer::Extensive) {
3055    AnalyzerOptions &options = getAnalyzerOptions();
3056    if (options.getBooleanOption("path-diagnostics-alternate", true)) {
3057      ActiveScheme = PathDiagnosticConsumer::AlternateExtensive;
3058    }
3059  }
3060
3061  TrimmedGraph TrimG(&getGraph(), errorNodes);
3062  ReportGraph ErrorGraph;
3063
3064  while (TrimG.popNextReportGraph(ErrorGraph)) {
3065    // Find the BugReport with the original location.
3066    assert(ErrorGraph.Index < bugReports.size());
3067    BugReport *R = bugReports[ErrorGraph.Index];
3068    assert(R && "No original report found for sliced graph.");
3069    assert(R->isValid() && "Report selected by trimmed graph marked invalid.");
3070
3071    // Start building the path diagnostic...
3072    PathDiagnosticBuilder PDB(*this, R, ErrorGraph.BackMap, &PC);
3073    const ExplodedNode *N = ErrorGraph.ErrorNode;
3074
3075    // Register additional node visitors.
3076    R->addVisitor(new NilReceiverBRVisitor());
3077    R->addVisitor(new ConditionBRVisitor());
3078    R->addVisitor(new LikelyFalsePositiveSuppressionBRVisitor());
3079
3080    BugReport::VisitorList visitors;
3081    unsigned origReportConfigToken, finalReportConfigToken;
3082    LocationContextMap LCM;
3083
3084    // While generating diagnostics, it's possible the visitors will decide
3085    // new symbols and regions are interesting, or add other visitors based on
3086    // the information they find. If they do, we need to regenerate the path
3087    // based on our new report configuration.
3088    do {
3089      // Get a clean copy of all the visitors.
3090      for (BugReport::visitor_iterator I = R->visitor_begin(),
3091                                       E = R->visitor_end(); I != E; ++I)
3092        visitors.push_back((*I)->clone());
3093
3094      // Clear out the active path from any previous work.
3095      PD.resetPath();
3096      origReportConfigToken = R->getConfigurationChangeToken();
3097
3098      // Generate the very last diagnostic piece - the piece is visible before
3099      // the trace is expanded.
3100      PathDiagnosticPiece *LastPiece = 0;
3101      for (BugReport::visitor_iterator I = visitors.begin(), E = visitors.end();
3102          I != E; ++I) {
3103        if (PathDiagnosticPiece *Piece = (*I)->getEndPath(PDB, N, *R)) {
3104          assert (!LastPiece &&
3105              "There can only be one final piece in a diagnostic.");
3106          LastPiece = Piece;
3107        }
3108      }
3109
3110      if (ActiveScheme != PathDiagnosticConsumer::None) {
3111        if (!LastPiece)
3112          LastPiece = BugReporterVisitor::getDefaultEndPath(PDB, N, *R);
3113        assert(LastPiece);
3114        PD.setEndOfPath(LastPiece);
3115      }
3116
3117      // Make sure we get a clean location context map so we don't
3118      // hold onto old mappings.
3119      LCM.clear();
3120
3121      switch (ActiveScheme) {
3122      case PathDiagnosticConsumer::AlternateExtensive:
3123        GenerateAlternateExtensivePathDiagnostic(PD, PDB, N, LCM, visitors);
3124        break;
3125      case PathDiagnosticConsumer::Extensive:
3126        GenerateExtensivePathDiagnostic(PD, PDB, N, LCM, visitors);
3127        break;
3128      case PathDiagnosticConsumer::Minimal:
3129        GenerateMinimalPathDiagnostic(PD, PDB, N, LCM, visitors);
3130        break;
3131      case PathDiagnosticConsumer::None:
3132        GenerateVisitorsOnlyPathDiagnostic(PD, PDB, N, visitors);
3133        break;
3134      }
3135
3136      // Clean up the visitors we used.
3137      llvm::DeleteContainerPointers(visitors);
3138
3139      // Did anything change while generating this path?
3140      finalReportConfigToken = R->getConfigurationChangeToken();
3141    } while (finalReportConfigToken != origReportConfigToken);
3142
3143    if (!R->isValid())
3144      continue;
3145
3146    // Finally, prune the diagnostic path of uninteresting stuff.
3147    if (!PD.path.empty()) {
3148      if (R->shouldPrunePath() && getAnalyzerOptions().shouldPrunePaths()) {
3149        bool stillHasNotes = removeUnneededCalls(PD.getMutablePieces(), R, LCM);
3150        assert(stillHasNotes);
3151        (void)stillHasNotes;
3152      }
3153
3154      adjustCallLocations(PD.getMutablePieces());
3155
3156      if (ActiveScheme == PathDiagnosticConsumer::AlternateExtensive) {
3157        SourceManager &SM = getSourceManager();
3158
3159        // Reduce the number of edges from a very conservative set
3160        // to an aesthetically pleasing subset that conveys the
3161        // necessary information.
3162        OptimizedCallsSet OCS;
3163        while (optimizeEdges(PD.getMutablePieces(), SM, OCS, LCM)) {}
3164
3165        // Drop the very first function-entry edge. It's not really necessary
3166        // for top-level functions.
3167        dropFunctionEntryEdge(PD.getMutablePieces(), LCM, SM);
3168      }
3169
3170      // Remove messages that are basically the same.
3171      // We have to do this after edge optimization in the Extensive mode.
3172      removeRedundantMsgs(PD.getMutablePieces());
3173    }
3174
3175    // We found a report and didn't suppress it.
3176    return true;
3177  }
3178
3179  // We suppressed all the reports in this equivalence class.
3180  assert(!HasInvalid && "Inconsistent suppression");
3181  (void)HasInvalid;
3182  return false;
3183}
3184
3185void BugReporter::Register(BugType *BT) {
3186  BugTypes = F.add(BugTypes, BT);
3187}
3188
3189void BugReporter::emitReport(BugReport* R) {
3190  // Compute the bug report's hash to determine its equivalence class.
3191  llvm::FoldingSetNodeID ID;
3192  R->Profile(ID);
3193
3194  // Lookup the equivance class.  If there isn't one, create it.
3195  BugType& BT = R->getBugType();
3196  Register(&BT);
3197  void *InsertPos;
3198  BugReportEquivClass* EQ = EQClasses.FindNodeOrInsertPos(ID, InsertPos);
3199
3200  if (!EQ) {
3201    EQ = new BugReportEquivClass(R);
3202    EQClasses.InsertNode(EQ, InsertPos);
3203    EQClassesVector.push_back(EQ);
3204  }
3205  else
3206    EQ->AddReport(R);
3207}
3208
3209
3210//===----------------------------------------------------------------------===//
3211// Emitting reports in equivalence classes.
3212//===----------------------------------------------------------------------===//
3213
3214namespace {
3215struct FRIEC_WLItem {
3216  const ExplodedNode *N;
3217  ExplodedNode::const_succ_iterator I, E;
3218
3219  FRIEC_WLItem(const ExplodedNode *n)
3220  : N(n), I(N->succ_begin()), E(N->succ_end()) {}
3221};
3222}
3223
3224static BugReport *
3225FindReportInEquivalenceClass(BugReportEquivClass& EQ,
3226                             SmallVectorImpl<BugReport*> &bugReports) {
3227
3228  BugReportEquivClass::iterator I = EQ.begin(), E = EQ.end();
3229  assert(I != E);
3230  BugType& BT = I->getBugType();
3231
3232  // If we don't need to suppress any of the nodes because they are
3233  // post-dominated by a sink, simply add all the nodes in the equivalence class
3234  // to 'Nodes'.  Any of the reports will serve as a "representative" report.
3235  if (!BT.isSuppressOnSink()) {
3236    BugReport *R = I;
3237    for (BugReportEquivClass::iterator I=EQ.begin(), E=EQ.end(); I!=E; ++I) {
3238      const ExplodedNode *N = I->getErrorNode();
3239      if (N) {
3240        R = I;
3241        bugReports.push_back(R);
3242      }
3243    }
3244    return R;
3245  }
3246
3247  // For bug reports that should be suppressed when all paths are post-dominated
3248  // by a sink node, iterate through the reports in the equivalence class
3249  // until we find one that isn't post-dominated (if one exists).  We use a
3250  // DFS traversal of the ExplodedGraph to find a non-sink node.  We could write
3251  // this as a recursive function, but we don't want to risk blowing out the
3252  // stack for very long paths.
3253  BugReport *exampleReport = 0;
3254
3255  for (; I != E; ++I) {
3256    const ExplodedNode *errorNode = I->getErrorNode();
3257
3258    if (!errorNode)
3259      continue;
3260    if (errorNode->isSink()) {
3261      llvm_unreachable(
3262           "BugType::isSuppressSink() should not be 'true' for sink end nodes");
3263    }
3264    // No successors?  By definition this nodes isn't post-dominated by a sink.
3265    if (errorNode->succ_empty()) {
3266      bugReports.push_back(I);
3267      if (!exampleReport)
3268        exampleReport = I;
3269      continue;
3270    }
3271
3272    // At this point we know that 'N' is not a sink and it has at least one
3273    // successor.  Use a DFS worklist to find a non-sink end-of-path node.
3274    typedef FRIEC_WLItem WLItem;
3275    typedef SmallVector<WLItem, 10> DFSWorkList;
3276    llvm::DenseMap<const ExplodedNode *, unsigned> Visited;
3277
3278    DFSWorkList WL;
3279    WL.push_back(errorNode);
3280    Visited[errorNode] = 1;
3281
3282    while (!WL.empty()) {
3283      WLItem &WI = WL.back();
3284      assert(!WI.N->succ_empty());
3285
3286      for (; WI.I != WI.E; ++WI.I) {
3287        const ExplodedNode *Succ = *WI.I;
3288        // End-of-path node?
3289        if (Succ->succ_empty()) {
3290          // If we found an end-of-path node that is not a sink.
3291          if (!Succ->isSink()) {
3292            bugReports.push_back(I);
3293            if (!exampleReport)
3294              exampleReport = I;
3295            WL.clear();
3296            break;
3297          }
3298          // Found a sink?  Continue on to the next successor.
3299          continue;
3300        }
3301        // Mark the successor as visited.  If it hasn't been explored,
3302        // enqueue it to the DFS worklist.
3303        unsigned &mark = Visited[Succ];
3304        if (!mark) {
3305          mark = 1;
3306          WL.push_back(Succ);
3307          break;
3308        }
3309      }
3310
3311      // The worklist may have been cleared at this point.  First
3312      // check if it is empty before checking the last item.
3313      if (!WL.empty() && &WL.back() == &WI)
3314        WL.pop_back();
3315    }
3316  }
3317
3318  // ExampleReport will be NULL if all the nodes in the equivalence class
3319  // were post-dominated by sinks.
3320  return exampleReport;
3321}
3322
3323void BugReporter::FlushReport(BugReportEquivClass& EQ) {
3324  SmallVector<BugReport*, 10> bugReports;
3325  BugReport *exampleReport = FindReportInEquivalenceClass(EQ, bugReports);
3326  if (exampleReport) {
3327    const PathDiagnosticConsumers &C = getPathDiagnosticConsumers();
3328    for (PathDiagnosticConsumers::const_iterator I=C.begin(),
3329                                                 E=C.end(); I != E; ++I) {
3330      FlushReport(exampleReport, **I, bugReports);
3331    }
3332  }
3333}
3334
3335void BugReporter::FlushReport(BugReport *exampleReport,
3336                              PathDiagnosticConsumer &PD,
3337                              ArrayRef<BugReport*> bugReports) {
3338
3339  // FIXME: Make sure we use the 'R' for the path that was actually used.
3340  // Probably doesn't make a difference in practice.
3341  BugType& BT = exampleReport->getBugType();
3342
3343  OwningPtr<PathDiagnostic>
3344    D(new PathDiagnostic(exampleReport->getDeclWithIssue(),
3345                         exampleReport->getBugType().getName(),
3346                         exampleReport->getDescription(),
3347                         exampleReport->getShortDescription(/*Fallback=*/false),
3348                         BT.getCategory(),
3349                         exampleReport->getUniqueingLocation(),
3350                         exampleReport->getUniqueingDecl()));
3351
3352  MaxBugClassSize = std::max(bugReports.size(),
3353                             static_cast<size_t>(MaxBugClassSize));
3354
3355  // Generate the full path diagnostic, using the generation scheme
3356  // specified by the PathDiagnosticConsumer. Note that we have to generate
3357  // path diagnostics even for consumers which do not support paths, because
3358  // the BugReporterVisitors may mark this bug as a false positive.
3359  if (!bugReports.empty())
3360    if (!generatePathDiagnostic(*D.get(), PD, bugReports))
3361      return;
3362
3363  MaxValidBugClassSize = std::max(bugReports.size(),
3364                                  static_cast<size_t>(MaxValidBugClassSize));
3365
3366  // Examine the report and see if the last piece is in a header. Reset the
3367  // report location to the last piece in the main source file.
3368  AnalyzerOptions& Opts = getAnalyzerOptions();
3369  if (Opts.shouldReportIssuesInMainSourceFile() && !Opts.AnalyzeAll)
3370    D->resetDiagnosticLocationToMainFile();
3371
3372  // If the path is empty, generate a single step path with the location
3373  // of the issue.
3374  if (D->path.empty()) {
3375    PathDiagnosticLocation L = exampleReport->getLocation(getSourceManager());
3376    PathDiagnosticPiece *piece =
3377      new PathDiagnosticEventPiece(L, exampleReport->getDescription());
3378    BugReport::ranges_iterator Beg, End;
3379    llvm::tie(Beg, End) = exampleReport->getRanges();
3380    for ( ; Beg != End; ++Beg)
3381      piece->addRange(*Beg);
3382    D->setEndOfPath(piece);
3383  }
3384
3385  // Get the meta data.
3386  const BugReport::ExtraTextList &Meta = exampleReport->getExtraText();
3387  for (BugReport::ExtraTextList::const_iterator i = Meta.begin(),
3388                                                e = Meta.end(); i != e; ++i) {
3389    D->addMeta(*i);
3390  }
3391
3392  PD.HandlePathDiagnostic(D.take());
3393}
3394
3395void BugReporter::EmitBasicReport(const Decl *DeclWithIssue,
3396                                  StringRef name,
3397                                  StringRef category,
3398                                  StringRef str, PathDiagnosticLocation Loc,
3399                                  SourceRange* RBeg, unsigned NumRanges) {
3400
3401  // 'BT' is owned by BugReporter.
3402  BugType *BT = getBugTypeForName(name, category);
3403  BugReport *R = new BugReport(*BT, str, Loc);
3404  R->setDeclWithIssue(DeclWithIssue);
3405  for ( ; NumRanges > 0 ; --NumRanges, ++RBeg) R->addRange(*RBeg);
3406  emitReport(R);
3407}
3408
3409BugType *BugReporter::getBugTypeForName(StringRef name,
3410                                        StringRef category) {
3411  SmallString<136> fullDesc;
3412  llvm::raw_svector_ostream(fullDesc) << name << ":" << category;
3413  llvm::StringMapEntry<BugType *> &
3414      entry = StrBugTypes.GetOrCreateValue(fullDesc);
3415  BugType *BT = entry.getValue();
3416  if (!BT) {
3417    BT = new BugType(name, category);
3418    entry.setValue(BT);
3419  }
3420  return BT;
3421}
3422
3423
3424void PathPieces::dump() const {
3425  unsigned index = 0;
3426  for (PathPieces::const_iterator I = begin(), E = end(); I != E; ++I) {
3427    llvm::errs() << "[" << index++ << "]  ";
3428    (*I)->dump();
3429    llvm::errs() << "\n";
3430  }
3431}
3432
3433void PathDiagnosticCallPiece::dump() const {
3434  llvm::errs() << "CALL\n--------------\n";
3435
3436  if (const Stmt *SLoc = getLocStmt(getLocation()))
3437    SLoc->dump();
3438  else if (const NamedDecl *ND = dyn_cast<NamedDecl>(getCallee()))
3439    llvm::errs() << *ND << "\n";
3440  else
3441    getLocation().dump();
3442}
3443
3444void PathDiagnosticEventPiece::dump() const {
3445  llvm::errs() << "EVENT\n--------------\n";
3446  llvm::errs() << getString() << "\n";
3447  llvm::errs() << " ---- at ----\n";
3448  getLocation().dump();
3449}
3450
3451void PathDiagnosticControlFlowPiece::dump() const {
3452  llvm::errs() << "CONTROL\n--------------\n";
3453  getStartLocation().dump();
3454  llvm::errs() << " ---- to ----\n";
3455  getEndLocation().dump();
3456}
3457
3458void PathDiagnosticMacroPiece::dump() const {
3459  llvm::errs() << "MACRO\n--------------\n";
3460  // FIXME: Print which macro is being invoked.
3461}
3462
3463void PathDiagnosticLocation::dump() const {
3464  if (!isValid()) {
3465    llvm::errs() << "<INVALID>\n";
3466    return;
3467  }
3468
3469  switch (K) {
3470  case RangeK:
3471    // FIXME: actually print the range.
3472    llvm::errs() << "<range>\n";
3473    break;
3474  case SingleLocK:
3475    asLocation().dump();
3476    llvm::errs() << "\n";
3477    break;
3478  case StmtK:
3479    if (S)
3480      S->dump();
3481    else
3482      llvm::errs() << "<NULL STMT>\n";
3483    break;
3484  case DeclK:
3485    if (const NamedDecl *ND = dyn_cast_or_null<NamedDecl>(D))
3486      llvm::errs() << *ND << "\n";
3487    else if (isa<BlockDecl>(D))
3488      // FIXME: Make this nicer.
3489      llvm::errs() << "<block>\n";
3490    else if (D)
3491      llvm::errs() << "<unknown decl>\n";
3492    else
3493      llvm::errs() << "<NULL DECL>\n";
3494    break;
3495  }
3496}
3497