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 #include "clang/StaticAnalyzer/Core/BugReporter/BugReporter.h"
16 #include "clang/AST/ASTContext.h"
17 #include "clang/AST/DeclObjC.h"
18 #include "clang/AST/Expr.h"
19 #include "clang/AST/ExprCXX.h"
20 #include "clang/AST/ParentMap.h"
21 #include "clang/AST/StmtCXX.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/STLExtras.h"
32 #include "llvm/ADT/SmallString.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include <memory>
36 #include <queue>
37
38 using namespace clang;
39 using namespace ento;
40
41 #define DEBUG_TYPE "BugReporter"
42
43 STATISTIC(MaxBugClassSize,
44 "The maximum number of bug reports in the same equivalence class");
45 STATISTIC(MaxValidBugClassSize,
46 "The maximum number of bug reports in the same equivalence class "
47 "where at least one report is valid (not suppressed)");
48
~BugReporterVisitor()49 BugReporterVisitor::~BugReporterVisitor() {}
50
anchor()51 void BugReporterContext::anchor() {}
52
53 //===----------------------------------------------------------------------===//
54 // Helper routines for walking the ExplodedGraph and fetching statements.
55 //===----------------------------------------------------------------------===//
56
GetPreviousStmt(const ExplodedNode * N)57 static const Stmt *GetPreviousStmt(const ExplodedNode *N) {
58 for (N = N->getFirstPred(); N; N = N->getFirstPred())
59 if (const Stmt *S = PathDiagnosticLocation::getStmt(N))
60 return S;
61
62 return nullptr;
63 }
64
65 static inline const Stmt*
GetCurrentOrPreviousStmt(const ExplodedNode * N)66 GetCurrentOrPreviousStmt(const ExplodedNode *N) {
67 if (const Stmt *S = PathDiagnosticLocation::getStmt(N))
68 return S;
69
70 return GetPreviousStmt(N);
71 }
72
73 //===----------------------------------------------------------------------===//
74 // Diagnostic cleanup.
75 //===----------------------------------------------------------------------===//
76
77 static PathDiagnosticEventPiece *
eventsDescribeSameCondition(PathDiagnosticEventPiece * X,PathDiagnosticEventPiece * Y)78 eventsDescribeSameCondition(PathDiagnosticEventPiece *X,
79 PathDiagnosticEventPiece *Y) {
80 // Prefer diagnostics that come from ConditionBRVisitor over
81 // those that came from TrackConstraintBRVisitor.
82 const void *tagPreferred = ConditionBRVisitor::getTag();
83 const void *tagLesser = TrackConstraintBRVisitor::getTag();
84
85 if (X->getLocation() != Y->getLocation())
86 return nullptr;
87
88 if (X->getTag() == tagPreferred && Y->getTag() == tagLesser)
89 return X;
90
91 if (Y->getTag() == tagPreferred && X->getTag() == tagLesser)
92 return Y;
93
94 return nullptr;
95 }
96
97 /// An optimization pass over PathPieces that removes redundant diagnostics
98 /// generated by both ConditionBRVisitor and TrackConstraintBRVisitor. Both
99 /// BugReporterVisitors use different methods to generate diagnostics, with
100 /// one capable of emitting diagnostics in some cases but not in others. This
101 /// can lead to redundant diagnostic pieces at the same point in a path.
removeRedundantMsgs(PathPieces & path)102 static void removeRedundantMsgs(PathPieces &path) {
103 unsigned N = path.size();
104 if (N < 2)
105 return;
106 // NOTE: this loop intentionally is not using an iterator. Instead, we
107 // are streaming the path and modifying it in place. This is done by
108 // grabbing the front, processing it, and if we decide to keep it append
109 // it to the end of the path. The entire path is processed in this way.
110 for (unsigned i = 0; i < N; ++i) {
111 IntrusiveRefCntPtr<PathDiagnosticPiece> piece(path.front());
112 path.pop_front();
113
114 switch (piece->getKind()) {
115 case clang::ento::PathDiagnosticPiece::Call:
116 removeRedundantMsgs(cast<PathDiagnosticCallPiece>(piece)->path);
117 break;
118 case clang::ento::PathDiagnosticPiece::Macro:
119 removeRedundantMsgs(cast<PathDiagnosticMacroPiece>(piece)->subPieces);
120 break;
121 case clang::ento::PathDiagnosticPiece::ControlFlow:
122 break;
123 case clang::ento::PathDiagnosticPiece::Event: {
124 if (i == N-1)
125 break;
126
127 if (PathDiagnosticEventPiece *nextEvent =
128 dyn_cast<PathDiagnosticEventPiece>(path.front().get())) {
129 PathDiagnosticEventPiece *event =
130 cast<PathDiagnosticEventPiece>(piece);
131 // Check to see if we should keep one of the two pieces. If we
132 // come up with a preference, record which piece to keep, and consume
133 // another piece from the path.
134 if (PathDiagnosticEventPiece *pieceToKeep =
135 eventsDescribeSameCondition(event, nextEvent)) {
136 piece = pieceToKeep;
137 path.pop_front();
138 ++i;
139 }
140 }
141 break;
142 }
143 }
144 path.push_back(piece);
145 }
146 }
147
148 /// A map from PathDiagnosticPiece to the LocationContext of the inlined
149 /// function call it represents.
150 typedef llvm::DenseMap<const PathPieces *, const LocationContext *>
151 LocationContextMap;
152
153 /// Recursively scan through a path and prune out calls and macros pieces
154 /// that aren't needed. Return true if afterwards the path contains
155 /// "interesting stuff" which means it shouldn't be pruned from the parent path.
removeUnneededCalls(PathPieces & pieces,BugReport * R,LocationContextMap & LCM)156 static bool removeUnneededCalls(PathPieces &pieces, BugReport *R,
157 LocationContextMap &LCM) {
158 bool containsSomethingInteresting = false;
159 const unsigned N = pieces.size();
160
161 for (unsigned i = 0 ; i < N ; ++i) {
162 // Remove the front piece from the path. If it is still something we
163 // want to keep once we are done, we will push it back on the end.
164 IntrusiveRefCntPtr<PathDiagnosticPiece> piece(pieces.front());
165 pieces.pop_front();
166
167 switch (piece->getKind()) {
168 case PathDiagnosticPiece::Call: {
169 PathDiagnosticCallPiece *call = cast<PathDiagnosticCallPiece>(piece);
170 // Check if the location context is interesting.
171 assert(LCM.count(&call->path));
172 if (R->isInteresting(LCM[&call->path])) {
173 containsSomethingInteresting = true;
174 break;
175 }
176
177 if (!removeUnneededCalls(call->path, R, LCM))
178 continue;
179
180 containsSomethingInteresting = true;
181 break;
182 }
183 case PathDiagnosticPiece::Macro: {
184 PathDiagnosticMacroPiece *macro = cast<PathDiagnosticMacroPiece>(piece);
185 if (!removeUnneededCalls(macro->subPieces, R, LCM))
186 continue;
187 containsSomethingInteresting = true;
188 break;
189 }
190 case PathDiagnosticPiece::Event: {
191 PathDiagnosticEventPiece *event = cast<PathDiagnosticEventPiece>(piece);
192
193 // We never throw away an event, but we do throw it away wholesale
194 // as part of a path if we throw the entire path away.
195 containsSomethingInteresting |= !event->isPrunable();
196 break;
197 }
198 case PathDiagnosticPiece::ControlFlow:
199 break;
200 }
201
202 pieces.push_back(piece);
203 }
204
205 return containsSomethingInteresting;
206 }
207
208 /// Returns true if the given decl has been implicitly given a body, either by
209 /// the analyzer or by the compiler proper.
hasImplicitBody(const Decl * D)210 static bool hasImplicitBody(const Decl *D) {
211 assert(D);
212 return D->isImplicit() || !D->hasBody();
213 }
214
215 /// Recursively scan through a path and make sure that all call pieces have
216 /// valid locations.
217 static void
adjustCallLocations(PathPieces & Pieces,PathDiagnosticLocation * LastCallLocation=nullptr)218 adjustCallLocations(PathPieces &Pieces,
219 PathDiagnosticLocation *LastCallLocation = nullptr) {
220 for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E; ++I) {
221 PathDiagnosticCallPiece *Call = dyn_cast<PathDiagnosticCallPiece>(*I);
222
223 if (!Call) {
224 assert((*I)->getLocation().asLocation().isValid());
225 continue;
226 }
227
228 if (LastCallLocation) {
229 bool CallerIsImplicit = hasImplicitBody(Call->getCaller());
230 if (CallerIsImplicit || !Call->callEnter.asLocation().isValid())
231 Call->callEnter = *LastCallLocation;
232 if (CallerIsImplicit || !Call->callReturn.asLocation().isValid())
233 Call->callReturn = *LastCallLocation;
234 }
235
236 // Recursively clean out the subclass. Keep this call around if
237 // it contains any informative diagnostics.
238 PathDiagnosticLocation *ThisCallLocation;
239 if (Call->callEnterWithin.asLocation().isValid() &&
240 !hasImplicitBody(Call->getCallee()))
241 ThisCallLocation = &Call->callEnterWithin;
242 else
243 ThisCallLocation = &Call->callEnter;
244
245 assert(ThisCallLocation && "Outermost call has an invalid location");
246 adjustCallLocations(Call->path, ThisCallLocation);
247 }
248 }
249
250 /// Remove edges in and out of C++ default initializer expressions. These are
251 /// for fields that have in-class initializers, as opposed to being initialized
252 /// explicitly in a constructor or braced list.
removeEdgesToDefaultInitializers(PathPieces & Pieces)253 static void removeEdgesToDefaultInitializers(PathPieces &Pieces) {
254 for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E;) {
255 if (PathDiagnosticCallPiece *C = dyn_cast<PathDiagnosticCallPiece>(*I))
256 removeEdgesToDefaultInitializers(C->path);
257
258 if (PathDiagnosticMacroPiece *M = dyn_cast<PathDiagnosticMacroPiece>(*I))
259 removeEdgesToDefaultInitializers(M->subPieces);
260
261 if (PathDiagnosticControlFlowPiece *CF =
262 dyn_cast<PathDiagnosticControlFlowPiece>(*I)) {
263 const Stmt *Start = CF->getStartLocation().asStmt();
264 const Stmt *End = CF->getEndLocation().asStmt();
265 if (Start && isa<CXXDefaultInitExpr>(Start)) {
266 I = Pieces.erase(I);
267 continue;
268 } else if (End && isa<CXXDefaultInitExpr>(End)) {
269 PathPieces::iterator Next = std::next(I);
270 if (Next != E) {
271 if (PathDiagnosticControlFlowPiece *NextCF =
272 dyn_cast<PathDiagnosticControlFlowPiece>(*Next)) {
273 NextCF->setStartLocation(CF->getStartLocation());
274 }
275 }
276 I = Pieces.erase(I);
277 continue;
278 }
279 }
280
281 I++;
282 }
283 }
284
285 /// Remove all pieces with invalid locations as these cannot be serialized.
286 /// We might have pieces with invalid locations as a result of inlining Body
287 /// Farm generated functions.
removePiecesWithInvalidLocations(PathPieces & Pieces)288 static void removePiecesWithInvalidLocations(PathPieces &Pieces) {
289 for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E;) {
290 if (PathDiagnosticCallPiece *C = dyn_cast<PathDiagnosticCallPiece>(*I))
291 removePiecesWithInvalidLocations(C->path);
292
293 if (PathDiagnosticMacroPiece *M = dyn_cast<PathDiagnosticMacroPiece>(*I))
294 removePiecesWithInvalidLocations(M->subPieces);
295
296 if (!(*I)->getLocation().isValid() ||
297 !(*I)->getLocation().asLocation().isValid()) {
298 I = Pieces.erase(I);
299 continue;
300 }
301 I++;
302 }
303 }
304
305 //===----------------------------------------------------------------------===//
306 // PathDiagnosticBuilder and its associated routines and helper objects.
307 //===----------------------------------------------------------------------===//
308
309 namespace {
310 class NodeMapClosure : public BugReport::NodeResolver {
311 InterExplodedGraphMap &M;
312 public:
NodeMapClosure(InterExplodedGraphMap & m)313 NodeMapClosure(InterExplodedGraphMap &m) : M(m) {}
314
getOriginalNode(const ExplodedNode * N)315 const ExplodedNode *getOriginalNode(const ExplodedNode *N) override {
316 return M.lookup(N);
317 }
318 };
319
320 class PathDiagnosticBuilder : public BugReporterContext {
321 BugReport *R;
322 PathDiagnosticConsumer *PDC;
323 NodeMapClosure NMC;
324 public:
325 const LocationContext *LC;
326
PathDiagnosticBuilder(GRBugReporter & br,BugReport * r,InterExplodedGraphMap & Backmap,PathDiagnosticConsumer * pdc)327 PathDiagnosticBuilder(GRBugReporter &br,
328 BugReport *r, InterExplodedGraphMap &Backmap,
329 PathDiagnosticConsumer *pdc)
330 : BugReporterContext(br),
331 R(r), PDC(pdc), NMC(Backmap), LC(r->getErrorNode()->getLocationContext())
332 {}
333
334 PathDiagnosticLocation ExecutionContinues(const ExplodedNode *N);
335
336 PathDiagnosticLocation ExecutionContinues(llvm::raw_string_ostream &os,
337 const ExplodedNode *N);
338
getBugReport()339 BugReport *getBugReport() { return R; }
340
getCodeDecl()341 Decl const &getCodeDecl() { return R->getErrorNode()->getCodeDecl(); }
342
getParentMap()343 ParentMap& getParentMap() { return LC->getParentMap(); }
344
getParent(const Stmt * S)345 const Stmt *getParent(const Stmt *S) {
346 return getParentMap().getParent(S);
347 }
348
getNodeResolver()349 NodeMapClosure& getNodeResolver() override { return NMC; }
350
351 PathDiagnosticLocation getEnclosingStmtLocation(const Stmt *S);
352
getGenerationScheme() const353 PathDiagnosticConsumer::PathGenerationScheme getGenerationScheme() const {
354 return PDC ? PDC->getGenerationScheme() : PathDiagnosticConsumer::Extensive;
355 }
356
supportsLogicalOpControlFlow() const357 bool supportsLogicalOpControlFlow() const {
358 return PDC ? PDC->supportsLogicalOpControlFlow() : true;
359 }
360 };
361 } // end anonymous namespace
362
363 PathDiagnosticLocation
ExecutionContinues(const ExplodedNode * N)364 PathDiagnosticBuilder::ExecutionContinues(const ExplodedNode *N) {
365 if (const Stmt *S = PathDiagnosticLocation::getNextStmt(N))
366 return PathDiagnosticLocation(S, getSourceManager(), LC);
367
368 return PathDiagnosticLocation::createDeclEnd(N->getLocationContext(),
369 getSourceManager());
370 }
371
372 PathDiagnosticLocation
ExecutionContinues(llvm::raw_string_ostream & os,const ExplodedNode * N)373 PathDiagnosticBuilder::ExecutionContinues(llvm::raw_string_ostream &os,
374 const ExplodedNode *N) {
375
376 // Slow, but probably doesn't matter.
377 if (os.str().empty())
378 os << ' ';
379
380 const PathDiagnosticLocation &Loc = ExecutionContinues(N);
381
382 if (Loc.asStmt())
383 os << "Execution continues on line "
384 << getSourceManager().getExpansionLineNumber(Loc.asLocation())
385 << '.';
386 else {
387 os << "Execution jumps to the end of the ";
388 const Decl *D = N->getLocationContext()->getDecl();
389 if (isa<ObjCMethodDecl>(D))
390 os << "method";
391 else if (isa<FunctionDecl>(D))
392 os << "function";
393 else {
394 assert(isa<BlockDecl>(D));
395 os << "anonymous block";
396 }
397 os << '.';
398 }
399
400 return Loc;
401 }
402
getEnclosingParent(const Stmt * S,const ParentMap & PM)403 static const Stmt *getEnclosingParent(const Stmt *S, const ParentMap &PM) {
404 if (isa<Expr>(S) && PM.isConsumedExpr(cast<Expr>(S)))
405 return PM.getParentIgnoreParens(S);
406
407 const Stmt *Parent = PM.getParentIgnoreParens(S);
408 if (!Parent)
409 return nullptr;
410
411 switch (Parent->getStmtClass()) {
412 case Stmt::ForStmtClass:
413 case Stmt::DoStmtClass:
414 case Stmt::WhileStmtClass:
415 case Stmt::ObjCForCollectionStmtClass:
416 case Stmt::CXXForRangeStmtClass:
417 return Parent;
418 default:
419 break;
420 }
421
422 return nullptr;
423 }
424
425 static PathDiagnosticLocation
getEnclosingStmtLocation(const Stmt * S,SourceManager & SMgr,const ParentMap & P,const LocationContext * LC,bool allowNestedContexts)426 getEnclosingStmtLocation(const Stmt *S, SourceManager &SMgr, const ParentMap &P,
427 const LocationContext *LC, bool allowNestedContexts) {
428 if (!S)
429 return PathDiagnosticLocation();
430
431 while (const Stmt *Parent = getEnclosingParent(S, P)) {
432 switch (Parent->getStmtClass()) {
433 case Stmt::BinaryOperatorClass: {
434 const BinaryOperator *B = cast<BinaryOperator>(Parent);
435 if (B->isLogicalOp())
436 return PathDiagnosticLocation(allowNestedContexts ? B : S, SMgr, LC);
437 break;
438 }
439 case Stmt::CompoundStmtClass:
440 case Stmt::StmtExprClass:
441 return PathDiagnosticLocation(S, SMgr, LC);
442 case Stmt::ChooseExprClass:
443 // Similar to '?' if we are referring to condition, just have the edge
444 // point to the entire choose expression.
445 if (allowNestedContexts || cast<ChooseExpr>(Parent)->getCond() == S)
446 return PathDiagnosticLocation(Parent, SMgr, LC);
447 else
448 return PathDiagnosticLocation(S, SMgr, LC);
449 case Stmt::BinaryConditionalOperatorClass:
450 case Stmt::ConditionalOperatorClass:
451 // For '?', if we are referring to condition, just have the edge point
452 // to the entire '?' expression.
453 if (allowNestedContexts ||
454 cast<AbstractConditionalOperator>(Parent)->getCond() == S)
455 return PathDiagnosticLocation(Parent, SMgr, LC);
456 else
457 return PathDiagnosticLocation(S, SMgr, LC);
458 case Stmt::CXXForRangeStmtClass:
459 if (cast<CXXForRangeStmt>(Parent)->getBody() == S)
460 return PathDiagnosticLocation(S, SMgr, LC);
461 break;
462 case Stmt::DoStmtClass:
463 return PathDiagnosticLocation(S, SMgr, LC);
464 case Stmt::ForStmtClass:
465 if (cast<ForStmt>(Parent)->getBody() == S)
466 return PathDiagnosticLocation(S, SMgr, LC);
467 break;
468 case Stmt::IfStmtClass:
469 if (cast<IfStmt>(Parent)->getCond() != S)
470 return PathDiagnosticLocation(S, SMgr, LC);
471 break;
472 case Stmt::ObjCForCollectionStmtClass:
473 if (cast<ObjCForCollectionStmt>(Parent)->getBody() == S)
474 return PathDiagnosticLocation(S, SMgr, LC);
475 break;
476 case Stmt::WhileStmtClass:
477 if (cast<WhileStmt>(Parent)->getCond() != S)
478 return PathDiagnosticLocation(S, SMgr, LC);
479 break;
480 default:
481 break;
482 }
483
484 S = Parent;
485 }
486
487 assert(S && "Cannot have null Stmt for PathDiagnosticLocation");
488
489 return PathDiagnosticLocation(S, SMgr, LC);
490 }
491
492 PathDiagnosticLocation
getEnclosingStmtLocation(const Stmt * S)493 PathDiagnosticBuilder::getEnclosingStmtLocation(const Stmt *S) {
494 assert(S && "Null Stmt passed to getEnclosingStmtLocation");
495 return ::getEnclosingStmtLocation(S, getSourceManager(), getParentMap(), LC,
496 /*allowNestedContexts=*/false);
497 }
498
499 //===----------------------------------------------------------------------===//
500 // "Visitors only" path diagnostic generation algorithm.
501 //===----------------------------------------------------------------------===//
GenerateVisitorsOnlyPathDiagnostic(PathDiagnostic & PD,PathDiagnosticBuilder & PDB,const ExplodedNode * N,ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors)502 static bool GenerateVisitorsOnlyPathDiagnostic(
503 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N,
504 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) {
505 // All path generation skips the very first node (the error node).
506 // This is because there is special handling for the end-of-path note.
507 N = N->getFirstPred();
508 if (!N)
509 return true;
510
511 BugReport *R = PDB.getBugReport();
512 while (const ExplodedNode *Pred = N->getFirstPred()) {
513 for (auto &V : visitors) {
514 // Visit all the node pairs, but throw the path pieces away.
515 PathDiagnosticPiece *Piece = V->VisitNode(N, Pred, PDB, *R);
516 delete Piece;
517 }
518
519 N = Pred;
520 }
521
522 return R->isValid();
523 }
524
525 //===----------------------------------------------------------------------===//
526 // "Minimal" path diagnostic generation algorithm.
527 //===----------------------------------------------------------------------===//
528 typedef std::pair<PathDiagnosticCallPiece*, const ExplodedNode*> StackDiagPair;
529 typedef SmallVector<StackDiagPair, 6> StackDiagVector;
530
updateStackPiecesWithMessage(PathDiagnosticPiece * P,StackDiagVector & CallStack)531 static void updateStackPiecesWithMessage(PathDiagnosticPiece *P,
532 StackDiagVector &CallStack) {
533 // If the piece contains a special message, add it to all the call
534 // pieces on the active stack.
535 if (PathDiagnosticEventPiece *ep =
536 dyn_cast<PathDiagnosticEventPiece>(P)) {
537
538 if (ep->hasCallStackHint())
539 for (StackDiagVector::iterator I = CallStack.begin(),
540 E = CallStack.end(); I != E; ++I) {
541 PathDiagnosticCallPiece *CP = I->first;
542 const ExplodedNode *N = I->second;
543 std::string stackMsg = ep->getCallStackMessage(N);
544
545 // The last message on the path to final bug is the most important
546 // one. Since we traverse the path backwards, do not add the message
547 // if one has been previously added.
548 if (!CP->hasCallStackMessage())
549 CP->setCallStackMessage(stackMsg);
550 }
551 }
552 }
553
554 static void CompactPathDiagnostic(PathPieces &path, const SourceManager& SM);
555
GenerateMinimalPathDiagnostic(PathDiagnostic & PD,PathDiagnosticBuilder & PDB,const ExplodedNode * N,LocationContextMap & LCM,ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors)556 static bool GenerateMinimalPathDiagnostic(
557 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N,
558 LocationContextMap &LCM,
559 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) {
560
561 SourceManager& SMgr = PDB.getSourceManager();
562 const LocationContext *LC = PDB.LC;
563 const ExplodedNode *NextNode = N->pred_empty()
564 ? nullptr : *(N->pred_begin());
565
566 StackDiagVector CallStack;
567
568 while (NextNode) {
569 N = NextNode;
570 PDB.LC = N->getLocationContext();
571 NextNode = N->getFirstPred();
572
573 ProgramPoint P = N->getLocation();
574
575 do {
576 if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
577 PathDiagnosticCallPiece *C =
578 PathDiagnosticCallPiece::construct(N, *CE, SMgr);
579 // Record the mapping from call piece to LocationContext.
580 LCM[&C->path] = CE->getCalleeContext();
581 PD.getActivePath().push_front(C);
582 PD.pushActivePath(&C->path);
583 CallStack.push_back(StackDiagPair(C, N));
584 break;
585 }
586
587 if (Optional<CallEnter> CE = P.getAs<CallEnter>()) {
588 // Flush all locations, and pop the active path.
589 bool VisitedEntireCall = PD.isWithinCall();
590 PD.popActivePath();
591
592 // Either we just added a bunch of stuff to the top-level path, or
593 // we have a previous CallExitEnd. If the former, it means that the
594 // path terminated within a function call. We must then take the
595 // current contents of the active path and place it within
596 // a new PathDiagnosticCallPiece.
597 PathDiagnosticCallPiece *C;
598 if (VisitedEntireCall) {
599 C = cast<PathDiagnosticCallPiece>(PD.getActivePath().front());
600 } else {
601 const Decl *Caller = CE->getLocationContext()->getDecl();
602 C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller);
603 // Record the mapping from call piece to LocationContext.
604 LCM[&C->path] = CE->getCalleeContext();
605 }
606
607 C->setCallee(*CE, SMgr);
608 if (!CallStack.empty()) {
609 assert(CallStack.back().first == C);
610 CallStack.pop_back();
611 }
612 break;
613 }
614
615 if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) {
616 const CFGBlock *Src = BE->getSrc();
617 const CFGBlock *Dst = BE->getDst();
618 const Stmt *T = Src->getTerminator();
619
620 if (!T)
621 break;
622
623 PathDiagnosticLocation Start =
624 PathDiagnosticLocation::createBegin(T, SMgr,
625 N->getLocationContext());
626
627 switch (T->getStmtClass()) {
628 default:
629 break;
630
631 case Stmt::GotoStmtClass:
632 case Stmt::IndirectGotoStmtClass: {
633 const Stmt *S = PathDiagnosticLocation::getNextStmt(N);
634
635 if (!S)
636 break;
637
638 std::string sbuf;
639 llvm::raw_string_ostream os(sbuf);
640 const PathDiagnosticLocation &End = PDB.getEnclosingStmtLocation(S);
641
642 os << "Control jumps to line "
643 << End.asLocation().getExpansionLineNumber();
644 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
645 Start, End, os.str()));
646 break;
647 }
648
649 case Stmt::SwitchStmtClass: {
650 // Figure out what case arm we took.
651 std::string sbuf;
652 llvm::raw_string_ostream os(sbuf);
653
654 if (const Stmt *S = Dst->getLabel()) {
655 PathDiagnosticLocation End(S, SMgr, LC);
656
657 switch (S->getStmtClass()) {
658 default:
659 os << "No cases match in the switch statement. "
660 "Control jumps to line "
661 << End.asLocation().getExpansionLineNumber();
662 break;
663 case Stmt::DefaultStmtClass:
664 os << "Control jumps to the 'default' case at line "
665 << End.asLocation().getExpansionLineNumber();
666 break;
667
668 case Stmt::CaseStmtClass: {
669 os << "Control jumps to 'case ";
670 const CaseStmt *Case = cast<CaseStmt>(S);
671 const Expr *LHS = Case->getLHS()->IgnoreParenCasts();
672
673 // Determine if it is an enum.
674 bool GetRawInt = true;
675
676 if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(LHS)) {
677 // FIXME: Maybe this should be an assertion. Are there cases
678 // were it is not an EnumConstantDecl?
679 const EnumConstantDecl *D =
680 dyn_cast<EnumConstantDecl>(DR->getDecl());
681
682 if (D) {
683 GetRawInt = false;
684 os << *D;
685 }
686 }
687
688 if (GetRawInt)
689 os << LHS->EvaluateKnownConstInt(PDB.getASTContext());
690
691 os << ":' at line "
692 << End.asLocation().getExpansionLineNumber();
693 break;
694 }
695 }
696 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
697 Start, End, os.str()));
698 }
699 else {
700 os << "'Default' branch taken. ";
701 const PathDiagnosticLocation &End = PDB.ExecutionContinues(os, N);
702 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
703 Start, End, os.str()));
704 }
705
706 break;
707 }
708
709 case Stmt::BreakStmtClass:
710 case Stmt::ContinueStmtClass: {
711 std::string sbuf;
712 llvm::raw_string_ostream os(sbuf);
713 PathDiagnosticLocation End = PDB.ExecutionContinues(os, N);
714 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
715 Start, End, os.str()));
716 break;
717 }
718
719 // Determine control-flow for ternary '?'.
720 case Stmt::BinaryConditionalOperatorClass:
721 case Stmt::ConditionalOperatorClass: {
722 std::string sbuf;
723 llvm::raw_string_ostream os(sbuf);
724 os << "'?' condition is ";
725
726 if (*(Src->succ_begin()+1) == Dst)
727 os << "false";
728 else
729 os << "true";
730
731 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
732
733 if (const Stmt *S = End.asStmt())
734 End = PDB.getEnclosingStmtLocation(S);
735
736 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
737 Start, End, os.str()));
738 break;
739 }
740
741 // Determine control-flow for short-circuited '&&' and '||'.
742 case Stmt::BinaryOperatorClass: {
743 if (!PDB.supportsLogicalOpControlFlow())
744 break;
745
746 const BinaryOperator *B = cast<BinaryOperator>(T);
747 std::string sbuf;
748 llvm::raw_string_ostream os(sbuf);
749 os << "Left side of '";
750
751 if (B->getOpcode() == BO_LAnd) {
752 os << "&&" << "' is ";
753
754 if (*(Src->succ_begin()+1) == Dst) {
755 os << "false";
756 PathDiagnosticLocation End(B->getLHS(), SMgr, LC);
757 PathDiagnosticLocation Start =
758 PathDiagnosticLocation::createOperatorLoc(B, SMgr);
759 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
760 Start, End, os.str()));
761 }
762 else {
763 os << "true";
764 PathDiagnosticLocation Start(B->getLHS(), SMgr, LC);
765 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
766 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
767 Start, End, os.str()));
768 }
769 }
770 else {
771 assert(B->getOpcode() == BO_LOr);
772 os << "||" << "' is ";
773
774 if (*(Src->succ_begin()+1) == Dst) {
775 os << "false";
776 PathDiagnosticLocation Start(B->getLHS(), SMgr, LC);
777 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
778 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
779 Start, End, os.str()));
780 }
781 else {
782 os << "true";
783 PathDiagnosticLocation End(B->getLHS(), SMgr, LC);
784 PathDiagnosticLocation Start =
785 PathDiagnosticLocation::createOperatorLoc(B, SMgr);
786 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
787 Start, End, os.str()));
788 }
789 }
790
791 break;
792 }
793
794 case Stmt::DoStmtClass: {
795 if (*(Src->succ_begin()) == Dst) {
796 std::string sbuf;
797 llvm::raw_string_ostream os(sbuf);
798
799 os << "Loop condition is true. ";
800 PathDiagnosticLocation End = PDB.ExecutionContinues(os, N);
801
802 if (const Stmt *S = End.asStmt())
803 End = PDB.getEnclosingStmtLocation(S);
804
805 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
806 Start, End, os.str()));
807 }
808 else {
809 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
810
811 if (const Stmt *S = End.asStmt())
812 End = PDB.getEnclosingStmtLocation(S);
813
814 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
815 Start, End, "Loop condition is false. Exiting loop"));
816 }
817
818 break;
819 }
820
821 case Stmt::WhileStmtClass:
822 case Stmt::ForStmtClass: {
823 if (*(Src->succ_begin()+1) == Dst) {
824 std::string sbuf;
825 llvm::raw_string_ostream os(sbuf);
826
827 os << "Loop condition is false. ";
828 PathDiagnosticLocation End = PDB.ExecutionContinues(os, N);
829 if (const Stmt *S = End.asStmt())
830 End = PDB.getEnclosingStmtLocation(S);
831
832 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
833 Start, End, os.str()));
834 }
835 else {
836 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
837 if (const Stmt *S = End.asStmt())
838 End = PDB.getEnclosingStmtLocation(S);
839
840 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
841 Start, End, "Loop condition is true. Entering loop body"));
842 }
843
844 break;
845 }
846
847 case Stmt::IfStmtClass: {
848 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
849
850 if (const Stmt *S = End.asStmt())
851 End = PDB.getEnclosingStmtLocation(S);
852
853 if (*(Src->succ_begin()+1) == Dst)
854 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
855 Start, End, "Taking false branch"));
856 else
857 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(
858 Start, End, "Taking true branch"));
859
860 break;
861 }
862 }
863 }
864 } while(0);
865
866 if (NextNode) {
867 // Add diagnostic pieces from custom visitors.
868 BugReport *R = PDB.getBugReport();
869 for (auto &V : visitors) {
870 if (PathDiagnosticPiece *p = V->VisitNode(N, NextNode, PDB, *R)) {
871 PD.getActivePath().push_front(p);
872 updateStackPiecesWithMessage(p, CallStack);
873 }
874 }
875 }
876 }
877
878 if (!PDB.getBugReport()->isValid())
879 return false;
880
881 // After constructing the full PathDiagnostic, do a pass over it to compact
882 // PathDiagnosticPieces that occur within a macro.
883 CompactPathDiagnostic(PD.getMutablePieces(), PDB.getSourceManager());
884 return true;
885 }
886
887 //===----------------------------------------------------------------------===//
888 // "Extensive" PathDiagnostic generation.
889 //===----------------------------------------------------------------------===//
890
IsControlFlowExpr(const Stmt * S)891 static bool IsControlFlowExpr(const Stmt *S) {
892 const Expr *E = dyn_cast<Expr>(S);
893
894 if (!E)
895 return false;
896
897 E = E->IgnoreParenCasts();
898
899 if (isa<AbstractConditionalOperator>(E))
900 return true;
901
902 if (const BinaryOperator *B = dyn_cast<BinaryOperator>(E))
903 if (B->isLogicalOp())
904 return true;
905
906 return false;
907 }
908
909 namespace {
910 class ContextLocation : public PathDiagnosticLocation {
911 bool IsDead;
912 public:
ContextLocation(const PathDiagnosticLocation & L,bool isdead=false)913 ContextLocation(const PathDiagnosticLocation &L, bool isdead = false)
914 : PathDiagnosticLocation(L), IsDead(isdead) {}
915
markDead()916 void markDead() { IsDead = true; }
isDead() const917 bool isDead() const { return IsDead; }
918 };
919
cleanUpLocation(PathDiagnosticLocation L,const LocationContext * LC,bool firstCharOnly=false)920 static PathDiagnosticLocation cleanUpLocation(PathDiagnosticLocation L,
921 const LocationContext *LC,
922 bool firstCharOnly = false) {
923 if (const Stmt *S = L.asStmt()) {
924 const Stmt *Original = S;
925 while (1) {
926 // Adjust the location for some expressions that are best referenced
927 // by one of their subexpressions.
928 switch (S->getStmtClass()) {
929 default:
930 break;
931 case Stmt::ParenExprClass:
932 case Stmt::GenericSelectionExprClass:
933 S = cast<Expr>(S)->IgnoreParens();
934 firstCharOnly = true;
935 continue;
936 case Stmt::BinaryConditionalOperatorClass:
937 case Stmt::ConditionalOperatorClass:
938 S = cast<AbstractConditionalOperator>(S)->getCond();
939 firstCharOnly = true;
940 continue;
941 case Stmt::ChooseExprClass:
942 S = cast<ChooseExpr>(S)->getCond();
943 firstCharOnly = true;
944 continue;
945 case Stmt::BinaryOperatorClass:
946 S = cast<BinaryOperator>(S)->getLHS();
947 firstCharOnly = true;
948 continue;
949 }
950
951 break;
952 }
953
954 if (S != Original)
955 L = PathDiagnosticLocation(S, L.getManager(), LC);
956 }
957
958 if (firstCharOnly)
959 L = PathDiagnosticLocation::createSingleLocation(L);
960
961 return L;
962 }
963
964 class EdgeBuilder {
965 std::vector<ContextLocation> CLocs;
966 typedef std::vector<ContextLocation>::iterator iterator;
967 PathDiagnostic &PD;
968 PathDiagnosticBuilder &PDB;
969 PathDiagnosticLocation PrevLoc;
970
971 bool IsConsumedExpr(const PathDiagnosticLocation &L);
972
973 bool containsLocation(const PathDiagnosticLocation &Container,
974 const PathDiagnosticLocation &Containee);
975
976 PathDiagnosticLocation getContextLocation(const PathDiagnosticLocation &L);
977
978
979
popLocation()980 void popLocation() {
981 if (!CLocs.back().isDead() && CLocs.back().asLocation().isFileID()) {
982 // For contexts, we only one the first character as the range.
983 rawAddEdge(cleanUpLocation(CLocs.back(), PDB.LC, true));
984 }
985 CLocs.pop_back();
986 }
987
988 public:
EdgeBuilder(PathDiagnostic & pd,PathDiagnosticBuilder & pdb)989 EdgeBuilder(PathDiagnostic &pd, PathDiagnosticBuilder &pdb)
990 : PD(pd), PDB(pdb) {
991
992 // If the PathDiagnostic already has pieces, add the enclosing statement
993 // of the first piece as a context as well.
994 if (!PD.path.empty()) {
995 PrevLoc = (*PD.path.begin())->getLocation();
996
997 if (const Stmt *S = PrevLoc.asStmt())
998 addExtendedContext(PDB.getEnclosingStmtLocation(S).asStmt());
999 }
1000 }
1001
~EdgeBuilder()1002 ~EdgeBuilder() {
1003 while (!CLocs.empty()) popLocation();
1004
1005 // Finally, add an initial edge from the start location of the first
1006 // statement (if it doesn't already exist).
1007 PathDiagnosticLocation L = PathDiagnosticLocation::createDeclBegin(
1008 PDB.LC,
1009 PDB.getSourceManager());
1010 if (L.isValid())
1011 rawAddEdge(L);
1012 }
1013
flushLocations()1014 void flushLocations() {
1015 while (!CLocs.empty())
1016 popLocation();
1017 PrevLoc = PathDiagnosticLocation();
1018 }
1019
1020 void addEdge(PathDiagnosticLocation NewLoc, bool alwaysAdd = false,
1021 bool IsPostJump = false);
1022
1023 void rawAddEdge(PathDiagnosticLocation NewLoc);
1024
1025 void addContext(const Stmt *S);
1026 void addContext(const PathDiagnosticLocation &L);
1027 void addExtendedContext(const Stmt *S);
1028 };
1029 } // end anonymous namespace
1030
1031
1032 PathDiagnosticLocation
getContextLocation(const PathDiagnosticLocation & L)1033 EdgeBuilder::getContextLocation(const PathDiagnosticLocation &L) {
1034 if (const Stmt *S = L.asStmt()) {
1035 if (IsControlFlowExpr(S))
1036 return L;
1037
1038 return PDB.getEnclosingStmtLocation(S);
1039 }
1040
1041 return L;
1042 }
1043
containsLocation(const PathDiagnosticLocation & Container,const PathDiagnosticLocation & Containee)1044 bool EdgeBuilder::containsLocation(const PathDiagnosticLocation &Container,
1045 const PathDiagnosticLocation &Containee) {
1046
1047 if (Container == Containee)
1048 return true;
1049
1050 if (Container.asDecl())
1051 return true;
1052
1053 if (const Stmt *S = Containee.asStmt())
1054 if (const Stmt *ContainerS = Container.asStmt()) {
1055 while (S) {
1056 if (S == ContainerS)
1057 return true;
1058 S = PDB.getParent(S);
1059 }
1060 return false;
1061 }
1062
1063 // Less accurate: compare using source ranges.
1064 SourceRange ContainerR = Container.asRange();
1065 SourceRange ContaineeR = Containee.asRange();
1066
1067 SourceManager &SM = PDB.getSourceManager();
1068 SourceLocation ContainerRBeg = SM.getExpansionLoc(ContainerR.getBegin());
1069 SourceLocation ContainerREnd = SM.getExpansionLoc(ContainerR.getEnd());
1070 SourceLocation ContaineeRBeg = SM.getExpansionLoc(ContaineeR.getBegin());
1071 SourceLocation ContaineeREnd = SM.getExpansionLoc(ContaineeR.getEnd());
1072
1073 unsigned ContainerBegLine = SM.getExpansionLineNumber(ContainerRBeg);
1074 unsigned ContainerEndLine = SM.getExpansionLineNumber(ContainerREnd);
1075 unsigned ContaineeBegLine = SM.getExpansionLineNumber(ContaineeRBeg);
1076 unsigned ContaineeEndLine = SM.getExpansionLineNumber(ContaineeREnd);
1077
1078 assert(ContainerBegLine <= ContainerEndLine);
1079 assert(ContaineeBegLine <= ContaineeEndLine);
1080
1081 return (ContainerBegLine <= ContaineeBegLine &&
1082 ContainerEndLine >= ContaineeEndLine &&
1083 (ContainerBegLine != ContaineeBegLine ||
1084 SM.getExpansionColumnNumber(ContainerRBeg) <=
1085 SM.getExpansionColumnNumber(ContaineeRBeg)) &&
1086 (ContainerEndLine != ContaineeEndLine ||
1087 SM.getExpansionColumnNumber(ContainerREnd) >=
1088 SM.getExpansionColumnNumber(ContaineeREnd)));
1089 }
1090
rawAddEdge(PathDiagnosticLocation NewLoc)1091 void EdgeBuilder::rawAddEdge(PathDiagnosticLocation NewLoc) {
1092 if (!PrevLoc.isValid()) {
1093 PrevLoc = NewLoc;
1094 return;
1095 }
1096
1097 const PathDiagnosticLocation &NewLocClean = cleanUpLocation(NewLoc, PDB.LC);
1098 const PathDiagnosticLocation &PrevLocClean = cleanUpLocation(PrevLoc, PDB.LC);
1099
1100 if (PrevLocClean.asLocation().isInvalid()) {
1101 PrevLoc = NewLoc;
1102 return;
1103 }
1104
1105 if (NewLocClean.asLocation() == PrevLocClean.asLocation())
1106 return;
1107
1108 // FIXME: Ignore intra-macro edges for now.
1109 if (NewLocClean.asLocation().getExpansionLoc() ==
1110 PrevLocClean.asLocation().getExpansionLoc())
1111 return;
1112
1113 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(NewLocClean, PrevLocClean));
1114 PrevLoc = NewLoc;
1115 }
1116
addEdge(PathDiagnosticLocation NewLoc,bool alwaysAdd,bool IsPostJump)1117 void EdgeBuilder::addEdge(PathDiagnosticLocation NewLoc, bool alwaysAdd,
1118 bool IsPostJump) {
1119
1120 if (!alwaysAdd && NewLoc.asLocation().isMacroID())
1121 return;
1122
1123 const PathDiagnosticLocation &CLoc = getContextLocation(NewLoc);
1124
1125 while (!CLocs.empty()) {
1126 ContextLocation &TopContextLoc = CLocs.back();
1127
1128 // Is the top location context the same as the one for the new location?
1129 if (TopContextLoc == CLoc) {
1130 if (alwaysAdd) {
1131 if (IsConsumedExpr(TopContextLoc))
1132 TopContextLoc.markDead();
1133
1134 rawAddEdge(NewLoc);
1135 }
1136
1137 if (IsPostJump)
1138 TopContextLoc.markDead();
1139 return;
1140 }
1141
1142 if (containsLocation(TopContextLoc, CLoc)) {
1143 if (alwaysAdd) {
1144 rawAddEdge(NewLoc);
1145
1146 if (IsConsumedExpr(CLoc)) {
1147 CLocs.push_back(ContextLocation(CLoc, /*IsDead=*/true));
1148 return;
1149 }
1150 }
1151
1152 CLocs.push_back(ContextLocation(CLoc, /*IsDead=*/IsPostJump));
1153 return;
1154 }
1155
1156 // Context does not contain the location. Flush it.
1157 popLocation();
1158 }
1159
1160 // If we reach here, there is no enclosing context. Just add the edge.
1161 rawAddEdge(NewLoc);
1162 }
1163
IsConsumedExpr(const PathDiagnosticLocation & L)1164 bool EdgeBuilder::IsConsumedExpr(const PathDiagnosticLocation &L) {
1165 if (const Expr *X = dyn_cast_or_null<Expr>(L.asStmt()))
1166 return PDB.getParentMap().isConsumedExpr(X) && !IsControlFlowExpr(X);
1167
1168 return false;
1169 }
1170
addExtendedContext(const Stmt * S)1171 void EdgeBuilder::addExtendedContext(const Stmt *S) {
1172 if (!S)
1173 return;
1174
1175 const Stmt *Parent = PDB.getParent(S);
1176 while (Parent) {
1177 if (isa<CompoundStmt>(Parent))
1178 Parent = PDB.getParent(Parent);
1179 else
1180 break;
1181 }
1182
1183 if (Parent) {
1184 switch (Parent->getStmtClass()) {
1185 case Stmt::DoStmtClass:
1186 case Stmt::ObjCAtSynchronizedStmtClass:
1187 addContext(Parent);
1188 default:
1189 break;
1190 }
1191 }
1192
1193 addContext(S);
1194 }
1195
addContext(const Stmt * S)1196 void EdgeBuilder::addContext(const Stmt *S) {
1197 if (!S)
1198 return;
1199
1200 PathDiagnosticLocation L(S, PDB.getSourceManager(), PDB.LC);
1201 addContext(L);
1202 }
1203
addContext(const PathDiagnosticLocation & L)1204 void EdgeBuilder::addContext(const PathDiagnosticLocation &L) {
1205 while (!CLocs.empty()) {
1206 const PathDiagnosticLocation &TopContextLoc = CLocs.back();
1207
1208 // Is the top location context the same as the one for the new location?
1209 if (TopContextLoc == L)
1210 return;
1211
1212 if (containsLocation(TopContextLoc, L)) {
1213 CLocs.push_back(L);
1214 return;
1215 }
1216
1217 // Context does not contain the location. Flush it.
1218 popLocation();
1219 }
1220
1221 CLocs.push_back(L);
1222 }
1223
1224 // Cone-of-influence: support the reverse propagation of "interesting" symbols
1225 // and values by tracing interesting calculations backwards through evaluated
1226 // expressions along a path. This is probably overly complicated, but the idea
1227 // is that if an expression computed an "interesting" value, the child
1228 // expressions are are also likely to be "interesting" as well (which then
1229 // propagates to the values they in turn compute). This reverse propagation
1230 // is needed to track interesting correlations across function call boundaries,
1231 // where formal arguments bind to actual arguments, etc. This is also needed
1232 // because the constraint solver sometimes simplifies certain symbolic values
1233 // into constants when appropriate, and this complicates reasoning about
1234 // interesting values.
1235 typedef llvm::DenseSet<const Expr *> InterestingExprs;
1236
reversePropagateIntererstingSymbols(BugReport & R,InterestingExprs & IE,const ProgramState * State,const Expr * Ex,const LocationContext * LCtx)1237 static void reversePropagateIntererstingSymbols(BugReport &R,
1238 InterestingExprs &IE,
1239 const ProgramState *State,
1240 const Expr *Ex,
1241 const LocationContext *LCtx) {
1242 SVal V = State->getSVal(Ex, LCtx);
1243 if (!(R.isInteresting(V) || IE.count(Ex)))
1244 return;
1245
1246 switch (Ex->getStmtClass()) {
1247 default:
1248 if (!isa<CastExpr>(Ex))
1249 break;
1250 // Fall through.
1251 case Stmt::BinaryOperatorClass:
1252 case Stmt::UnaryOperatorClass: {
1253 for (Stmt::const_child_iterator CI = Ex->child_begin(),
1254 CE = Ex->child_end();
1255 CI != CE; ++CI) {
1256 if (const Expr *child = dyn_cast_or_null<Expr>(*CI)) {
1257 IE.insert(child);
1258 SVal ChildV = State->getSVal(child, LCtx);
1259 R.markInteresting(ChildV);
1260 }
1261 }
1262 break;
1263 }
1264 }
1265
1266 R.markInteresting(V);
1267 }
1268
reversePropagateInterestingSymbols(BugReport & R,InterestingExprs & IE,const ProgramState * State,const LocationContext * CalleeCtx,const LocationContext * CallerCtx)1269 static void reversePropagateInterestingSymbols(BugReport &R,
1270 InterestingExprs &IE,
1271 const ProgramState *State,
1272 const LocationContext *CalleeCtx,
1273 const LocationContext *CallerCtx)
1274 {
1275 // FIXME: Handle non-CallExpr-based CallEvents.
1276 const StackFrameContext *Callee = CalleeCtx->getCurrentStackFrame();
1277 const Stmt *CallSite = Callee->getCallSite();
1278 if (const CallExpr *CE = dyn_cast_or_null<CallExpr>(CallSite)) {
1279 if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(CalleeCtx->getDecl())) {
1280 FunctionDecl::param_const_iterator PI = FD->param_begin(),
1281 PE = FD->param_end();
1282 CallExpr::const_arg_iterator AI = CE->arg_begin(), AE = CE->arg_end();
1283 for (; AI != AE && PI != PE; ++AI, ++PI) {
1284 if (const Expr *ArgE = *AI) {
1285 if (const ParmVarDecl *PD = *PI) {
1286 Loc LV = State->getLValue(PD, CalleeCtx);
1287 if (R.isInteresting(LV) || R.isInteresting(State->getRawSVal(LV)))
1288 IE.insert(ArgE);
1289 }
1290 }
1291 }
1292 }
1293 }
1294 }
1295
1296 //===----------------------------------------------------------------------===//
1297 // Functions for determining if a loop was executed 0 times.
1298 //===----------------------------------------------------------------------===//
1299
isLoop(const Stmt * Term)1300 static bool isLoop(const Stmt *Term) {
1301 switch (Term->getStmtClass()) {
1302 case Stmt::ForStmtClass:
1303 case Stmt::WhileStmtClass:
1304 case Stmt::ObjCForCollectionStmtClass:
1305 case Stmt::CXXForRangeStmtClass:
1306 return true;
1307 default:
1308 // Note that we intentionally do not include do..while here.
1309 return false;
1310 }
1311 }
1312
isJumpToFalseBranch(const BlockEdge * BE)1313 static bool isJumpToFalseBranch(const BlockEdge *BE) {
1314 const CFGBlock *Src = BE->getSrc();
1315 assert(Src->succ_size() == 2);
1316 return (*(Src->succ_begin()+1) == BE->getDst());
1317 }
1318
1319 /// Return true if the terminator is a loop and the destination is the
1320 /// false branch.
isLoopJumpPastBody(const Stmt * Term,const BlockEdge * BE)1321 static bool isLoopJumpPastBody(const Stmt *Term, const BlockEdge *BE) {
1322 if (!isLoop(Term))
1323 return false;
1324
1325 // Did we take the false branch?
1326 return isJumpToFalseBranch(BE);
1327 }
1328
isContainedByStmt(ParentMap & PM,const Stmt * S,const Stmt * SubS)1329 static bool isContainedByStmt(ParentMap &PM, const Stmt *S, const Stmt *SubS) {
1330 while (SubS) {
1331 if (SubS == S)
1332 return true;
1333 SubS = PM.getParent(SubS);
1334 }
1335 return false;
1336 }
1337
getStmtBeforeCond(ParentMap & PM,const Stmt * Term,const ExplodedNode * N)1338 static const Stmt *getStmtBeforeCond(ParentMap &PM, const Stmt *Term,
1339 const ExplodedNode *N) {
1340 while (N) {
1341 Optional<StmtPoint> SP = N->getLocation().getAs<StmtPoint>();
1342 if (SP) {
1343 const Stmt *S = SP->getStmt();
1344 if (!isContainedByStmt(PM, Term, S))
1345 return S;
1346 }
1347 N = N->getFirstPred();
1348 }
1349 return nullptr;
1350 }
1351
isInLoopBody(ParentMap & PM,const Stmt * S,const Stmt * Term)1352 static bool isInLoopBody(ParentMap &PM, const Stmt *S, const Stmt *Term) {
1353 const Stmt *LoopBody = nullptr;
1354 switch (Term->getStmtClass()) {
1355 case Stmt::CXXForRangeStmtClass: {
1356 const CXXForRangeStmt *FR = cast<CXXForRangeStmt>(Term);
1357 if (isContainedByStmt(PM, FR->getInc(), S))
1358 return true;
1359 if (isContainedByStmt(PM, FR->getLoopVarStmt(), S))
1360 return true;
1361 LoopBody = FR->getBody();
1362 break;
1363 }
1364 case Stmt::ForStmtClass: {
1365 const ForStmt *FS = cast<ForStmt>(Term);
1366 if (isContainedByStmt(PM, FS->getInc(), S))
1367 return true;
1368 LoopBody = FS->getBody();
1369 break;
1370 }
1371 case Stmt::ObjCForCollectionStmtClass: {
1372 const ObjCForCollectionStmt *FC = cast<ObjCForCollectionStmt>(Term);
1373 LoopBody = FC->getBody();
1374 break;
1375 }
1376 case Stmt::WhileStmtClass:
1377 LoopBody = cast<WhileStmt>(Term)->getBody();
1378 break;
1379 default:
1380 return false;
1381 }
1382 return isContainedByStmt(PM, LoopBody, S);
1383 }
1384
1385 //===----------------------------------------------------------------------===//
1386 // Top-level logic for generating extensive path diagnostics.
1387 //===----------------------------------------------------------------------===//
1388
GenerateExtensivePathDiagnostic(PathDiagnostic & PD,PathDiagnosticBuilder & PDB,const ExplodedNode * N,LocationContextMap & LCM,ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors)1389 static bool GenerateExtensivePathDiagnostic(
1390 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N,
1391 LocationContextMap &LCM,
1392 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) {
1393 EdgeBuilder EB(PD, PDB);
1394 const SourceManager& SM = PDB.getSourceManager();
1395 StackDiagVector CallStack;
1396 InterestingExprs IE;
1397
1398 const ExplodedNode *NextNode = N->pred_empty() ? nullptr : *(N->pred_begin());
1399 while (NextNode) {
1400 N = NextNode;
1401 NextNode = N->getFirstPred();
1402 ProgramPoint P = N->getLocation();
1403
1404 do {
1405 if (Optional<PostStmt> PS = P.getAs<PostStmt>()) {
1406 if (const Expr *Ex = PS->getStmtAs<Expr>())
1407 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1408 N->getState().get(), Ex,
1409 N->getLocationContext());
1410 }
1411
1412 if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
1413 const Stmt *S = CE->getCalleeContext()->getCallSite();
1414 if (const Expr *Ex = dyn_cast_or_null<Expr>(S)) {
1415 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1416 N->getState().get(), Ex,
1417 N->getLocationContext());
1418 }
1419
1420 PathDiagnosticCallPiece *C =
1421 PathDiagnosticCallPiece::construct(N, *CE, SM);
1422 LCM[&C->path] = CE->getCalleeContext();
1423
1424 EB.addEdge(C->callReturn, /*AlwaysAdd=*/true, /*IsPostJump=*/true);
1425 EB.flushLocations();
1426
1427 PD.getActivePath().push_front(C);
1428 PD.pushActivePath(&C->path);
1429 CallStack.push_back(StackDiagPair(C, N));
1430 break;
1431 }
1432
1433 // Pop the call hierarchy if we are done walking the contents
1434 // of a function call.
1435 if (Optional<CallEnter> CE = P.getAs<CallEnter>()) {
1436 // Add an edge to the start of the function.
1437 const Decl *D = CE->getCalleeContext()->getDecl();
1438 PathDiagnosticLocation pos =
1439 PathDiagnosticLocation::createBegin(D, SM);
1440 EB.addEdge(pos);
1441
1442 // Flush all locations, and pop the active path.
1443 bool VisitedEntireCall = PD.isWithinCall();
1444 EB.flushLocations();
1445 PD.popActivePath();
1446 PDB.LC = N->getLocationContext();
1447
1448 // Either we just added a bunch of stuff to the top-level path, or
1449 // we have a previous CallExitEnd. If the former, it means that the
1450 // path terminated within a function call. We must then take the
1451 // current contents of the active path and place it within
1452 // a new PathDiagnosticCallPiece.
1453 PathDiagnosticCallPiece *C;
1454 if (VisitedEntireCall) {
1455 C = cast<PathDiagnosticCallPiece>(PD.getActivePath().front());
1456 } else {
1457 const Decl *Caller = CE->getLocationContext()->getDecl();
1458 C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller);
1459 LCM[&C->path] = CE->getCalleeContext();
1460 }
1461
1462 C->setCallee(*CE, SM);
1463 EB.addContext(C->getLocation());
1464
1465 if (!CallStack.empty()) {
1466 assert(CallStack.back().first == C);
1467 CallStack.pop_back();
1468 }
1469 break;
1470 }
1471
1472 // Note that is important that we update the LocationContext
1473 // after looking at CallExits. CallExit basically adds an
1474 // edge in the *caller*, so we don't want to update the LocationContext
1475 // too soon.
1476 PDB.LC = N->getLocationContext();
1477
1478 // Block edges.
1479 if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) {
1480 // Does this represent entering a call? If so, look at propagating
1481 // interesting symbols across call boundaries.
1482 if (NextNode) {
1483 const LocationContext *CallerCtx = NextNode->getLocationContext();
1484 const LocationContext *CalleeCtx = PDB.LC;
1485 if (CallerCtx != CalleeCtx) {
1486 reversePropagateInterestingSymbols(*PDB.getBugReport(), IE,
1487 N->getState().get(),
1488 CalleeCtx, CallerCtx);
1489 }
1490 }
1491
1492 // Are we jumping to the head of a loop? Add a special diagnostic.
1493 if (const Stmt *Loop = BE->getSrc()->getLoopTarget()) {
1494 PathDiagnosticLocation L(Loop, SM, PDB.LC);
1495 const CompoundStmt *CS = nullptr;
1496
1497 if (const ForStmt *FS = dyn_cast<ForStmt>(Loop))
1498 CS = dyn_cast<CompoundStmt>(FS->getBody());
1499 else if (const WhileStmt *WS = dyn_cast<WhileStmt>(Loop))
1500 CS = dyn_cast<CompoundStmt>(WS->getBody());
1501
1502 PathDiagnosticEventPiece *p =
1503 new PathDiagnosticEventPiece(L,
1504 "Looping back to the head of the loop");
1505 p->setPrunable(true);
1506
1507 EB.addEdge(p->getLocation(), true);
1508 PD.getActivePath().push_front(p);
1509
1510 if (CS) {
1511 PathDiagnosticLocation BL =
1512 PathDiagnosticLocation::createEndBrace(CS, SM);
1513 EB.addEdge(BL);
1514 }
1515 }
1516
1517 const CFGBlock *BSrc = BE->getSrc();
1518 ParentMap &PM = PDB.getParentMap();
1519
1520 if (const Stmt *Term = BSrc->getTerminator()) {
1521 // Are we jumping past the loop body without ever executing the
1522 // loop (because the condition was false)?
1523 if (isLoopJumpPastBody(Term, &*BE) &&
1524 !isInLoopBody(PM,
1525 getStmtBeforeCond(PM,
1526 BSrc->getTerminatorCondition(),
1527 N),
1528 Term)) {
1529 PathDiagnosticLocation L(Term, SM, PDB.LC);
1530 PathDiagnosticEventPiece *PE =
1531 new PathDiagnosticEventPiece(L, "Loop body executed 0 times");
1532 PE->setPrunable(true);
1533
1534 EB.addEdge(PE->getLocation(), true);
1535 PD.getActivePath().push_front(PE);
1536 }
1537
1538 // In any case, add the terminator as the current statement
1539 // context for control edges.
1540 EB.addContext(Term);
1541 }
1542
1543 break;
1544 }
1545
1546 if (Optional<BlockEntrance> BE = P.getAs<BlockEntrance>()) {
1547 Optional<CFGElement> First = BE->getFirstElement();
1548 if (Optional<CFGStmt> S = First ? First->getAs<CFGStmt>() : None) {
1549 const Stmt *stmt = S->getStmt();
1550 if (IsControlFlowExpr(stmt)) {
1551 // Add the proper context for '&&', '||', and '?'.
1552 EB.addContext(stmt);
1553 }
1554 else
1555 EB.addExtendedContext(PDB.getEnclosingStmtLocation(stmt).asStmt());
1556 }
1557
1558 break;
1559 }
1560
1561
1562 } while (0);
1563
1564 if (!NextNode)
1565 continue;
1566
1567 // Add pieces from custom visitors.
1568 BugReport *R = PDB.getBugReport();
1569 for (auto &V : visitors) {
1570 if (PathDiagnosticPiece *p = V->VisitNode(N, NextNode, PDB, *R)) {
1571 const PathDiagnosticLocation &Loc = p->getLocation();
1572 EB.addEdge(Loc, true);
1573 PD.getActivePath().push_front(p);
1574 updateStackPiecesWithMessage(p, CallStack);
1575
1576 if (const Stmt *S = Loc.asStmt())
1577 EB.addExtendedContext(PDB.getEnclosingStmtLocation(S).asStmt());
1578 }
1579 }
1580 }
1581
1582 return PDB.getBugReport()->isValid();
1583 }
1584
1585 /// \brief Adds a sanitized control-flow diagnostic edge to a path.
addEdgeToPath(PathPieces & path,PathDiagnosticLocation & PrevLoc,PathDiagnosticLocation NewLoc,const LocationContext * LC)1586 static void addEdgeToPath(PathPieces &path,
1587 PathDiagnosticLocation &PrevLoc,
1588 PathDiagnosticLocation NewLoc,
1589 const LocationContext *LC) {
1590 if (!NewLoc.isValid())
1591 return;
1592
1593 SourceLocation NewLocL = NewLoc.asLocation();
1594 if (NewLocL.isInvalid())
1595 return;
1596
1597 if (!PrevLoc.isValid() || !PrevLoc.asLocation().isValid()) {
1598 PrevLoc = NewLoc;
1599 return;
1600 }
1601
1602 // Ignore self-edges, which occur when there are multiple nodes at the same
1603 // statement.
1604 if (NewLoc.asStmt() && NewLoc.asStmt() == PrevLoc.asStmt())
1605 return;
1606
1607 path.push_front(new PathDiagnosticControlFlowPiece(NewLoc,
1608 PrevLoc));
1609 PrevLoc = NewLoc;
1610 }
1611
1612 /// A customized wrapper for CFGBlock::getTerminatorCondition()
1613 /// which returns the element for ObjCForCollectionStmts.
getTerminatorCondition(const CFGBlock * B)1614 static const Stmt *getTerminatorCondition(const CFGBlock *B) {
1615 const Stmt *S = B->getTerminatorCondition();
1616 if (const ObjCForCollectionStmt *FS =
1617 dyn_cast_or_null<ObjCForCollectionStmt>(S))
1618 return FS->getElement();
1619 return S;
1620 }
1621
1622 static const char StrEnteringLoop[] = "Entering loop body";
1623 static const char StrLoopBodyZero[] = "Loop body executed 0 times";
1624 static const char StrLoopRangeEmpty[] =
1625 "Loop body skipped when range is empty";
1626 static const char StrLoopCollectionEmpty[] =
1627 "Loop body skipped when collection is empty";
1628
GenerateAlternateExtensivePathDiagnostic(PathDiagnostic & PD,PathDiagnosticBuilder & PDB,const ExplodedNode * N,LocationContextMap & LCM,ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors)1629 static bool GenerateAlternateExtensivePathDiagnostic(
1630 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N,
1631 LocationContextMap &LCM,
1632 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) {
1633
1634 BugReport *report = PDB.getBugReport();
1635 const SourceManager& SM = PDB.getSourceManager();
1636 StackDiagVector CallStack;
1637 InterestingExprs IE;
1638
1639 PathDiagnosticLocation PrevLoc = PD.getLocation();
1640
1641 const ExplodedNode *NextNode = N->getFirstPred();
1642 while (NextNode) {
1643 N = NextNode;
1644 NextNode = N->getFirstPred();
1645 ProgramPoint P = N->getLocation();
1646
1647 do {
1648 // Have we encountered an entrance to a call? It may be
1649 // the case that we have not encountered a matching
1650 // call exit before this point. This means that the path
1651 // terminated within the call itself.
1652 if (Optional<CallEnter> CE = P.getAs<CallEnter>()) {
1653 // Add an edge to the start of the function.
1654 const StackFrameContext *CalleeLC = CE->getCalleeContext();
1655 const Decl *D = CalleeLC->getDecl();
1656 addEdgeToPath(PD.getActivePath(), PrevLoc,
1657 PathDiagnosticLocation::createBegin(D, SM),
1658 CalleeLC);
1659
1660 // Did we visit an entire call?
1661 bool VisitedEntireCall = PD.isWithinCall();
1662 PD.popActivePath();
1663
1664 PathDiagnosticCallPiece *C;
1665 if (VisitedEntireCall) {
1666 PathDiagnosticPiece *P = PD.getActivePath().front().get();
1667 C = cast<PathDiagnosticCallPiece>(P);
1668 } else {
1669 const Decl *Caller = CE->getLocationContext()->getDecl();
1670 C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller);
1671
1672 // Since we just transferred the path over to the call piece,
1673 // reset the mapping from active to location context.
1674 assert(PD.getActivePath().size() == 1 &&
1675 PD.getActivePath().front() == C);
1676 LCM[&PD.getActivePath()] = nullptr;
1677
1678 // Record the location context mapping for the path within
1679 // the call.
1680 assert(LCM[&C->path] == nullptr ||
1681 LCM[&C->path] == CE->getCalleeContext());
1682 LCM[&C->path] = CE->getCalleeContext();
1683
1684 // If this is the first item in the active path, record
1685 // the new mapping from active path to location context.
1686 const LocationContext *&NewLC = LCM[&PD.getActivePath()];
1687 if (!NewLC)
1688 NewLC = N->getLocationContext();
1689
1690 PDB.LC = NewLC;
1691 }
1692 C->setCallee(*CE, SM);
1693
1694 // Update the previous location in the active path.
1695 PrevLoc = C->getLocation();
1696
1697 if (!CallStack.empty()) {
1698 assert(CallStack.back().first == C);
1699 CallStack.pop_back();
1700 }
1701 break;
1702 }
1703
1704 // Query the location context here and the previous location
1705 // as processing CallEnter may change the active path.
1706 PDB.LC = N->getLocationContext();
1707
1708 // Record the mapping from the active path to the location
1709 // context.
1710 assert(!LCM[&PD.getActivePath()] ||
1711 LCM[&PD.getActivePath()] == PDB.LC);
1712 LCM[&PD.getActivePath()] = PDB.LC;
1713
1714 // Have we encountered an exit from a function call?
1715 if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
1716 const Stmt *S = CE->getCalleeContext()->getCallSite();
1717 // Propagate the interesting symbols accordingly.
1718 if (const Expr *Ex = dyn_cast_or_null<Expr>(S)) {
1719 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1720 N->getState().get(), Ex,
1721 N->getLocationContext());
1722 }
1723
1724 // We are descending into a call (backwards). Construct
1725 // a new call piece to contain the path pieces for that call.
1726 PathDiagnosticCallPiece *C =
1727 PathDiagnosticCallPiece::construct(N, *CE, SM);
1728
1729 // Record the location context for this call piece.
1730 LCM[&C->path] = CE->getCalleeContext();
1731
1732 // Add the edge to the return site.
1733 addEdgeToPath(PD.getActivePath(), PrevLoc, C->callReturn, PDB.LC);
1734 PD.getActivePath().push_front(C);
1735 PrevLoc.invalidate();
1736
1737 // Make the contents of the call the active path for now.
1738 PD.pushActivePath(&C->path);
1739 CallStack.push_back(StackDiagPair(C, N));
1740 break;
1741 }
1742
1743 if (Optional<PostStmt> PS = P.getAs<PostStmt>()) {
1744 // For expressions, make sure we propagate the
1745 // interesting symbols correctly.
1746 if (const Expr *Ex = PS->getStmtAs<Expr>())
1747 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1748 N->getState().get(), Ex,
1749 N->getLocationContext());
1750
1751 // Add an edge. If this is an ObjCForCollectionStmt do
1752 // not add an edge here as it appears in the CFG both
1753 // as a terminator and as a terminator condition.
1754 if (!isa<ObjCForCollectionStmt>(PS->getStmt())) {
1755 PathDiagnosticLocation L =
1756 PathDiagnosticLocation(PS->getStmt(), SM, PDB.LC);
1757 addEdgeToPath(PD.getActivePath(), PrevLoc, L, PDB.LC);
1758 }
1759 break;
1760 }
1761
1762 // Block edges.
1763 if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) {
1764 // Does this represent entering a call? If so, look at propagating
1765 // interesting symbols across call boundaries.
1766 if (NextNode) {
1767 const LocationContext *CallerCtx = NextNode->getLocationContext();
1768 const LocationContext *CalleeCtx = PDB.LC;
1769 if (CallerCtx != CalleeCtx) {
1770 reversePropagateInterestingSymbols(*PDB.getBugReport(), IE,
1771 N->getState().get(),
1772 CalleeCtx, CallerCtx);
1773 }
1774 }
1775
1776 // Are we jumping to the head of a loop? Add a special diagnostic.
1777 if (const Stmt *Loop = BE->getSrc()->getLoopTarget()) {
1778 PathDiagnosticLocation L(Loop, SM, PDB.LC);
1779 const Stmt *Body = nullptr;
1780
1781 if (const ForStmt *FS = dyn_cast<ForStmt>(Loop))
1782 Body = FS->getBody();
1783 else if (const WhileStmt *WS = dyn_cast<WhileStmt>(Loop))
1784 Body = WS->getBody();
1785 else if (const ObjCForCollectionStmt *OFS =
1786 dyn_cast<ObjCForCollectionStmt>(Loop)) {
1787 Body = OFS->getBody();
1788 } else if (const CXXForRangeStmt *FRS =
1789 dyn_cast<CXXForRangeStmt>(Loop)) {
1790 Body = FRS->getBody();
1791 }
1792 // do-while statements are explicitly excluded here
1793
1794 PathDiagnosticEventPiece *p =
1795 new PathDiagnosticEventPiece(L, "Looping back to the head "
1796 "of the loop");
1797 p->setPrunable(true);
1798
1799 addEdgeToPath(PD.getActivePath(), PrevLoc, p->getLocation(), PDB.LC);
1800 PD.getActivePath().push_front(p);
1801
1802 if (const CompoundStmt *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
1803 addEdgeToPath(PD.getActivePath(), PrevLoc,
1804 PathDiagnosticLocation::createEndBrace(CS, SM),
1805 PDB.LC);
1806 }
1807 }
1808
1809 const CFGBlock *BSrc = BE->getSrc();
1810 ParentMap &PM = PDB.getParentMap();
1811
1812 if (const Stmt *Term = BSrc->getTerminator()) {
1813 // Are we jumping past the loop body without ever executing the
1814 // loop (because the condition was false)?
1815 if (isLoop(Term)) {
1816 const Stmt *TermCond = getTerminatorCondition(BSrc);
1817 bool IsInLoopBody =
1818 isInLoopBody(PM, getStmtBeforeCond(PM, TermCond, N), Term);
1819
1820 const char *str = nullptr;
1821
1822 if (isJumpToFalseBranch(&*BE)) {
1823 if (!IsInLoopBody) {
1824 if (isa<ObjCForCollectionStmt>(Term)) {
1825 str = StrLoopCollectionEmpty;
1826 } else if (isa<CXXForRangeStmt>(Term)) {
1827 str = StrLoopRangeEmpty;
1828 } else {
1829 str = StrLoopBodyZero;
1830 }
1831 }
1832 } else {
1833 str = StrEnteringLoop;
1834 }
1835
1836 if (str) {
1837 PathDiagnosticLocation L(TermCond ? TermCond : Term, SM, PDB.LC);
1838 PathDiagnosticEventPiece *PE =
1839 new PathDiagnosticEventPiece(L, str);
1840 PE->setPrunable(true);
1841 addEdgeToPath(PD.getActivePath(), PrevLoc,
1842 PE->getLocation(), PDB.LC);
1843 PD.getActivePath().push_front(PE);
1844 }
1845 } else if (isa<BreakStmt>(Term) || isa<ContinueStmt>(Term) ||
1846 isa<GotoStmt>(Term)) {
1847 PathDiagnosticLocation L(Term, SM, PDB.LC);
1848 addEdgeToPath(PD.getActivePath(), PrevLoc, L, PDB.LC);
1849 }
1850 }
1851 break;
1852 }
1853 } while (0);
1854
1855 if (!NextNode)
1856 continue;
1857
1858 // Add pieces from custom visitors.
1859 for (auto &V : visitors) {
1860 if (PathDiagnosticPiece *p = V->VisitNode(N, NextNode, PDB, *report)) {
1861 addEdgeToPath(PD.getActivePath(), PrevLoc, p->getLocation(), PDB.LC);
1862 PD.getActivePath().push_front(p);
1863 updateStackPiecesWithMessage(p, CallStack);
1864 }
1865 }
1866 }
1867
1868 // Add an edge to the start of the function.
1869 // We'll prune it out later, but it helps make diagnostics more uniform.
1870 const StackFrameContext *CalleeLC = PDB.LC->getCurrentStackFrame();
1871 const Decl *D = CalleeLC->getDecl();
1872 addEdgeToPath(PD.getActivePath(), PrevLoc,
1873 PathDiagnosticLocation::createBegin(D, SM),
1874 CalleeLC);
1875
1876 return report->isValid();
1877 }
1878
getLocStmt(PathDiagnosticLocation L)1879 static const Stmt *getLocStmt(PathDiagnosticLocation L) {
1880 if (!L.isValid())
1881 return nullptr;
1882 return L.asStmt();
1883 }
1884
getStmtParent(const Stmt * S,const ParentMap & PM)1885 static const Stmt *getStmtParent(const Stmt *S, const ParentMap &PM) {
1886 if (!S)
1887 return nullptr;
1888
1889 while (true) {
1890 S = PM.getParentIgnoreParens(S);
1891
1892 if (!S)
1893 break;
1894
1895 if (isa<ExprWithCleanups>(S) ||
1896 isa<CXXBindTemporaryExpr>(S) ||
1897 isa<SubstNonTypeTemplateParmExpr>(S))
1898 continue;
1899
1900 break;
1901 }
1902
1903 return S;
1904 }
1905
isConditionForTerminator(const Stmt * S,const Stmt * Cond)1906 static bool isConditionForTerminator(const Stmt *S, const Stmt *Cond) {
1907 switch (S->getStmtClass()) {
1908 case Stmt::BinaryOperatorClass: {
1909 const BinaryOperator *BO = cast<BinaryOperator>(S);
1910 if (!BO->isLogicalOp())
1911 return false;
1912 return BO->getLHS() == Cond || BO->getRHS() == Cond;
1913 }
1914 case Stmt::IfStmtClass:
1915 return cast<IfStmt>(S)->getCond() == Cond;
1916 case Stmt::ForStmtClass:
1917 return cast<ForStmt>(S)->getCond() == Cond;
1918 case Stmt::WhileStmtClass:
1919 return cast<WhileStmt>(S)->getCond() == Cond;
1920 case Stmt::DoStmtClass:
1921 return cast<DoStmt>(S)->getCond() == Cond;
1922 case Stmt::ChooseExprClass:
1923 return cast<ChooseExpr>(S)->getCond() == Cond;
1924 case Stmt::IndirectGotoStmtClass:
1925 return cast<IndirectGotoStmt>(S)->getTarget() == Cond;
1926 case Stmt::SwitchStmtClass:
1927 return cast<SwitchStmt>(S)->getCond() == Cond;
1928 case Stmt::BinaryConditionalOperatorClass:
1929 return cast<BinaryConditionalOperator>(S)->getCond() == Cond;
1930 case Stmt::ConditionalOperatorClass: {
1931 const ConditionalOperator *CO = cast<ConditionalOperator>(S);
1932 return CO->getCond() == Cond ||
1933 CO->getLHS() == Cond ||
1934 CO->getRHS() == Cond;
1935 }
1936 case Stmt::ObjCForCollectionStmtClass:
1937 return cast<ObjCForCollectionStmt>(S)->getElement() == Cond;
1938 case Stmt::CXXForRangeStmtClass: {
1939 const CXXForRangeStmt *FRS = cast<CXXForRangeStmt>(S);
1940 return FRS->getCond() == Cond || FRS->getRangeInit() == Cond;
1941 }
1942 default:
1943 return false;
1944 }
1945 }
1946
isIncrementOrInitInForLoop(const Stmt * S,const Stmt * FL)1947 static bool isIncrementOrInitInForLoop(const Stmt *S, const Stmt *FL) {
1948 if (const ForStmt *FS = dyn_cast<ForStmt>(FL))
1949 return FS->getInc() == S || FS->getInit() == S;
1950 if (const CXXForRangeStmt *FRS = dyn_cast<CXXForRangeStmt>(FL))
1951 return FRS->getInc() == S || FRS->getRangeStmt() == S ||
1952 FRS->getLoopVarStmt() || FRS->getRangeInit() == S;
1953 return false;
1954 }
1955
1956 typedef llvm::DenseSet<const PathDiagnosticCallPiece *>
1957 OptimizedCallsSet;
1958
1959 /// Adds synthetic edges from top-level statements to their subexpressions.
1960 ///
1961 /// This avoids a "swoosh" effect, where an edge from a top-level statement A
1962 /// points to a sub-expression B.1 that's not at the start of B. In these cases,
1963 /// we'd like to see an edge from A to B, then another one from B to B.1.
addContextEdges(PathPieces & pieces,SourceManager & SM,const ParentMap & PM,const LocationContext * LCtx)1964 static void addContextEdges(PathPieces &pieces, SourceManager &SM,
1965 const ParentMap &PM, const LocationContext *LCtx) {
1966 PathPieces::iterator Prev = pieces.end();
1967 for (PathPieces::iterator I = pieces.begin(), E = Prev; I != E;
1968 Prev = I, ++I) {
1969 PathDiagnosticControlFlowPiece *Piece =
1970 dyn_cast<PathDiagnosticControlFlowPiece>(*I);
1971
1972 if (!Piece)
1973 continue;
1974
1975 PathDiagnosticLocation SrcLoc = Piece->getStartLocation();
1976 SmallVector<PathDiagnosticLocation, 4> SrcContexts;
1977
1978 PathDiagnosticLocation NextSrcContext = SrcLoc;
1979 const Stmt *InnerStmt = nullptr;
1980 while (NextSrcContext.isValid() && NextSrcContext.asStmt() != InnerStmt) {
1981 SrcContexts.push_back(NextSrcContext);
1982 InnerStmt = NextSrcContext.asStmt();
1983 NextSrcContext = getEnclosingStmtLocation(InnerStmt, SM, PM, LCtx,
1984 /*allowNested=*/true);
1985 }
1986
1987 // Repeatedly split the edge as necessary.
1988 // This is important for nested logical expressions (||, &&, ?:) where we
1989 // want to show all the levels of context.
1990 while (true) {
1991 const Stmt *Dst = getLocStmt(Piece->getEndLocation());
1992
1993 // We are looking at an edge. Is the destination within a larger
1994 // expression?
1995 PathDiagnosticLocation DstContext =
1996 getEnclosingStmtLocation(Dst, SM, PM, LCtx, /*allowNested=*/true);
1997 if (!DstContext.isValid() || DstContext.asStmt() == Dst)
1998 break;
1999
2000 // If the source is in the same context, we're already good.
2001 if (std::find(SrcContexts.begin(), SrcContexts.end(), DstContext) !=
2002 SrcContexts.end())
2003 break;
2004
2005 // Update the subexpression node to point to the context edge.
2006 Piece->setStartLocation(DstContext);
2007
2008 // Try to extend the previous edge if it's at the same level as the source
2009 // context.
2010 if (Prev != E) {
2011 PathDiagnosticControlFlowPiece *PrevPiece =
2012 dyn_cast<PathDiagnosticControlFlowPiece>(*Prev);
2013
2014 if (PrevPiece) {
2015 if (const Stmt *PrevSrc = getLocStmt(PrevPiece->getStartLocation())) {
2016 const Stmt *PrevSrcParent = getStmtParent(PrevSrc, PM);
2017 if (PrevSrcParent == getStmtParent(getLocStmt(DstContext), PM)) {
2018 PrevPiece->setEndLocation(DstContext);
2019 break;
2020 }
2021 }
2022 }
2023 }
2024
2025 // Otherwise, split the current edge into a context edge and a
2026 // subexpression edge. Note that the context statement may itself have
2027 // context.
2028 Piece = new PathDiagnosticControlFlowPiece(SrcLoc, DstContext);
2029 I = pieces.insert(I, Piece);
2030 }
2031 }
2032 }
2033
2034 /// \brief Move edges from a branch condition to a branch target
2035 /// when the condition is simple.
2036 ///
2037 /// This restructures some of the work of addContextEdges. That function
2038 /// creates edges this may destroy, but they work together to create a more
2039 /// aesthetically set of edges around branches. After the call to
2040 /// addContextEdges, we may have (1) an edge to the branch, (2) an edge from
2041 /// the branch to the branch condition, and (3) an edge from the branch
2042 /// condition to the branch target. We keep (1), but may wish to remove (2)
2043 /// and move the source of (3) to the branch if the branch condition is simple.
2044 ///
simplifySimpleBranches(PathPieces & pieces)2045 static void simplifySimpleBranches(PathPieces &pieces) {
2046 for (PathPieces::iterator I = pieces.begin(), E = pieces.end(); I != E; ++I) {
2047
2048 PathDiagnosticControlFlowPiece *PieceI =
2049 dyn_cast<PathDiagnosticControlFlowPiece>(*I);
2050
2051 if (!PieceI)
2052 continue;
2053
2054 const Stmt *s1Start = getLocStmt(PieceI->getStartLocation());
2055 const Stmt *s1End = getLocStmt(PieceI->getEndLocation());
2056
2057 if (!s1Start || !s1End)
2058 continue;
2059
2060 PathPieces::iterator NextI = I; ++NextI;
2061 if (NextI == E)
2062 break;
2063
2064 PathDiagnosticControlFlowPiece *PieceNextI = nullptr;
2065
2066 while (true) {
2067 if (NextI == E)
2068 break;
2069
2070 PathDiagnosticEventPiece *EV = dyn_cast<PathDiagnosticEventPiece>(*NextI);
2071 if (EV) {
2072 StringRef S = EV->getString();
2073 if (S == StrEnteringLoop || S == StrLoopBodyZero ||
2074 S == StrLoopCollectionEmpty || S == StrLoopRangeEmpty) {
2075 ++NextI;
2076 continue;
2077 }
2078 break;
2079 }
2080
2081 PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(*NextI);
2082 break;
2083 }
2084
2085 if (!PieceNextI)
2086 continue;
2087
2088 const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation());
2089 const Stmt *s2End = getLocStmt(PieceNextI->getEndLocation());
2090
2091 if (!s2Start || !s2End || s1End != s2Start)
2092 continue;
2093
2094 // We only perform this transformation for specific branch kinds.
2095 // We don't want to do this for do..while, for example.
2096 if (!(isa<ForStmt>(s1Start) || isa<WhileStmt>(s1Start) ||
2097 isa<IfStmt>(s1Start) || isa<ObjCForCollectionStmt>(s1Start) ||
2098 isa<CXXForRangeStmt>(s1Start)))
2099 continue;
2100
2101 // Is s1End the branch condition?
2102 if (!isConditionForTerminator(s1Start, s1End))
2103 continue;
2104
2105 // Perform the hoisting by eliminating (2) and changing the start
2106 // location of (3).
2107 PieceNextI->setStartLocation(PieceI->getStartLocation());
2108 I = pieces.erase(I);
2109 }
2110 }
2111
2112 /// Returns the number of bytes in the given (character-based) SourceRange.
2113 ///
2114 /// If the locations in the range are not on the same line, returns None.
2115 ///
2116 /// Note that this does not do a precise user-visible character or column count.
getLengthOnSingleLine(SourceManager & SM,SourceRange Range)2117 static Optional<size_t> getLengthOnSingleLine(SourceManager &SM,
2118 SourceRange Range) {
2119 SourceRange ExpansionRange(SM.getExpansionLoc(Range.getBegin()),
2120 SM.getExpansionRange(Range.getEnd()).second);
2121
2122 FileID FID = SM.getFileID(ExpansionRange.getBegin());
2123 if (FID != SM.getFileID(ExpansionRange.getEnd()))
2124 return None;
2125
2126 bool Invalid;
2127 const llvm::MemoryBuffer *Buffer = SM.getBuffer(FID, &Invalid);
2128 if (Invalid)
2129 return None;
2130
2131 unsigned BeginOffset = SM.getFileOffset(ExpansionRange.getBegin());
2132 unsigned EndOffset = SM.getFileOffset(ExpansionRange.getEnd());
2133 StringRef Snippet = Buffer->getBuffer().slice(BeginOffset, EndOffset);
2134
2135 // We're searching the raw bytes of the buffer here, which might include
2136 // escaped newlines and such. That's okay; we're trying to decide whether the
2137 // SourceRange is covering a large or small amount of space in the user's
2138 // editor.
2139 if (Snippet.find_first_of("\r\n") != StringRef::npos)
2140 return None;
2141
2142 // This isn't Unicode-aware, but it doesn't need to be.
2143 return Snippet.size();
2144 }
2145
2146 /// \sa getLengthOnSingleLine(SourceManager, SourceRange)
getLengthOnSingleLine(SourceManager & SM,const Stmt * S)2147 static Optional<size_t> getLengthOnSingleLine(SourceManager &SM,
2148 const Stmt *S) {
2149 return getLengthOnSingleLine(SM, S->getSourceRange());
2150 }
2151
2152 /// Eliminate two-edge cycles created by addContextEdges().
2153 ///
2154 /// Once all the context edges are in place, there are plenty of cases where
2155 /// there's a single edge from a top-level statement to a subexpression,
2156 /// followed by a single path note, and then a reverse edge to get back out to
2157 /// the top level. If the statement is simple enough, the subexpression edges
2158 /// just add noise and make it harder to understand what's going on.
2159 ///
2160 /// This function only removes edges in pairs, because removing only one edge
2161 /// might leave other edges dangling.
2162 ///
2163 /// This will not remove edges in more complicated situations:
2164 /// - if there is more than one "hop" leading to or from a subexpression.
2165 /// - if there is an inlined call between the edges instead of a single event.
2166 /// - if the whole statement is large enough that having subexpression arrows
2167 /// might be helpful.
removeContextCycles(PathPieces & Path,SourceManager & SM,ParentMap & PM)2168 static void removeContextCycles(PathPieces &Path, SourceManager &SM,
2169 ParentMap &PM) {
2170 for (PathPieces::iterator I = Path.begin(), E = Path.end(); I != E; ) {
2171 // Pattern match the current piece and its successor.
2172 PathDiagnosticControlFlowPiece *PieceI =
2173 dyn_cast<PathDiagnosticControlFlowPiece>(*I);
2174
2175 if (!PieceI) {
2176 ++I;
2177 continue;
2178 }
2179
2180 const Stmt *s1Start = getLocStmt(PieceI->getStartLocation());
2181 const Stmt *s1End = getLocStmt(PieceI->getEndLocation());
2182
2183 PathPieces::iterator NextI = I; ++NextI;
2184 if (NextI == E)
2185 break;
2186
2187 PathDiagnosticControlFlowPiece *PieceNextI =
2188 dyn_cast<PathDiagnosticControlFlowPiece>(*NextI);
2189
2190 if (!PieceNextI) {
2191 if (isa<PathDiagnosticEventPiece>(*NextI)) {
2192 ++NextI;
2193 if (NextI == E)
2194 break;
2195 PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(*NextI);
2196 }
2197
2198 if (!PieceNextI) {
2199 ++I;
2200 continue;
2201 }
2202 }
2203
2204 const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation());
2205 const Stmt *s2End = getLocStmt(PieceNextI->getEndLocation());
2206
2207 if (s1Start && s2Start && s1Start == s2End && s2Start == s1End) {
2208 const size_t MAX_SHORT_LINE_LENGTH = 80;
2209 Optional<size_t> s1Length = getLengthOnSingleLine(SM, s1Start);
2210 if (s1Length && *s1Length <= MAX_SHORT_LINE_LENGTH) {
2211 Optional<size_t> s2Length = getLengthOnSingleLine(SM, s2Start);
2212 if (s2Length && *s2Length <= MAX_SHORT_LINE_LENGTH) {
2213 Path.erase(I);
2214 I = Path.erase(NextI);
2215 continue;
2216 }
2217 }
2218 }
2219
2220 ++I;
2221 }
2222 }
2223
2224 /// \brief Return true if X is contained by Y.
lexicalContains(ParentMap & PM,const Stmt * X,const Stmt * Y)2225 static bool lexicalContains(ParentMap &PM,
2226 const Stmt *X,
2227 const Stmt *Y) {
2228 while (X) {
2229 if (X == Y)
2230 return true;
2231 X = PM.getParent(X);
2232 }
2233 return false;
2234 }
2235
2236 // Remove short edges on the same line less than 3 columns in difference.
removePunyEdges(PathPieces & path,SourceManager & SM,ParentMap & PM)2237 static void removePunyEdges(PathPieces &path,
2238 SourceManager &SM,
2239 ParentMap &PM) {
2240
2241 bool erased = false;
2242
2243 for (PathPieces::iterator I = path.begin(), E = path.end(); I != E;
2244 erased ? I : ++I) {
2245
2246 erased = false;
2247
2248 PathDiagnosticControlFlowPiece *PieceI =
2249 dyn_cast<PathDiagnosticControlFlowPiece>(*I);
2250
2251 if (!PieceI)
2252 continue;
2253
2254 const Stmt *start = getLocStmt(PieceI->getStartLocation());
2255 const Stmt *end = getLocStmt(PieceI->getEndLocation());
2256
2257 if (!start || !end)
2258 continue;
2259
2260 const Stmt *endParent = PM.getParent(end);
2261 if (!endParent)
2262 continue;
2263
2264 if (isConditionForTerminator(end, endParent))
2265 continue;
2266
2267 SourceLocation FirstLoc = start->getLocStart();
2268 SourceLocation SecondLoc = end->getLocStart();
2269
2270 if (!SM.isWrittenInSameFile(FirstLoc, SecondLoc))
2271 continue;
2272 if (SM.isBeforeInTranslationUnit(SecondLoc, FirstLoc))
2273 std::swap(SecondLoc, FirstLoc);
2274
2275 SourceRange EdgeRange(FirstLoc, SecondLoc);
2276 Optional<size_t> ByteWidth = getLengthOnSingleLine(SM, EdgeRange);
2277
2278 // If the statements are on different lines, continue.
2279 if (!ByteWidth)
2280 continue;
2281
2282 const size_t MAX_PUNY_EDGE_LENGTH = 2;
2283 if (*ByteWidth <= MAX_PUNY_EDGE_LENGTH) {
2284 // FIXME: There are enough /bytes/ between the endpoints of the edge, but
2285 // there might not be enough /columns/. A proper user-visible column count
2286 // is probably too expensive, though.
2287 I = path.erase(I);
2288 erased = true;
2289 continue;
2290 }
2291 }
2292 }
2293
removeIdenticalEvents(PathPieces & path)2294 static void removeIdenticalEvents(PathPieces &path) {
2295 for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ++I) {
2296 PathDiagnosticEventPiece *PieceI =
2297 dyn_cast<PathDiagnosticEventPiece>(*I);
2298
2299 if (!PieceI)
2300 continue;
2301
2302 PathPieces::iterator NextI = I; ++NextI;
2303 if (NextI == E)
2304 return;
2305
2306 PathDiagnosticEventPiece *PieceNextI =
2307 dyn_cast<PathDiagnosticEventPiece>(*NextI);
2308
2309 if (!PieceNextI)
2310 continue;
2311
2312 // Erase the second piece if it has the same exact message text.
2313 if (PieceI->getString() == PieceNextI->getString()) {
2314 path.erase(NextI);
2315 }
2316 }
2317 }
2318
optimizeEdges(PathPieces & path,SourceManager & SM,OptimizedCallsSet & OCS,LocationContextMap & LCM)2319 static bool optimizeEdges(PathPieces &path, SourceManager &SM,
2320 OptimizedCallsSet &OCS,
2321 LocationContextMap &LCM) {
2322 bool hasChanges = false;
2323 const LocationContext *LC = LCM[&path];
2324 assert(LC);
2325 ParentMap &PM = LC->getParentMap();
2326
2327 for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ) {
2328 // Optimize subpaths.
2329 if (PathDiagnosticCallPiece *CallI = dyn_cast<PathDiagnosticCallPiece>(*I)){
2330 // Record the fact that a call has been optimized so we only do the
2331 // effort once.
2332 if (!OCS.count(CallI)) {
2333 while (optimizeEdges(CallI->path, SM, OCS, LCM)) {}
2334 OCS.insert(CallI);
2335 }
2336 ++I;
2337 continue;
2338 }
2339
2340 // Pattern match the current piece and its successor.
2341 PathDiagnosticControlFlowPiece *PieceI =
2342 dyn_cast<PathDiagnosticControlFlowPiece>(*I);
2343
2344 if (!PieceI) {
2345 ++I;
2346 continue;
2347 }
2348
2349 const Stmt *s1Start = getLocStmt(PieceI->getStartLocation());
2350 const Stmt *s1End = getLocStmt(PieceI->getEndLocation());
2351 const Stmt *level1 = getStmtParent(s1Start, PM);
2352 const Stmt *level2 = getStmtParent(s1End, PM);
2353
2354 PathPieces::iterator NextI = I; ++NextI;
2355 if (NextI == E)
2356 break;
2357
2358 PathDiagnosticControlFlowPiece *PieceNextI =
2359 dyn_cast<PathDiagnosticControlFlowPiece>(*NextI);
2360
2361 if (!PieceNextI) {
2362 ++I;
2363 continue;
2364 }
2365
2366 const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation());
2367 const Stmt *s2End = getLocStmt(PieceNextI->getEndLocation());
2368 const Stmt *level3 = getStmtParent(s2Start, PM);
2369 const Stmt *level4 = getStmtParent(s2End, PM);
2370
2371 // Rule I.
2372 //
2373 // If we have two consecutive control edges whose end/begin locations
2374 // are at the same level (e.g. statements or top-level expressions within
2375 // a compound statement, or siblings share a single ancestor expression),
2376 // then merge them if they have no interesting intermediate event.
2377 //
2378 // For example:
2379 //
2380 // (1.1 -> 1.2) -> (1.2 -> 1.3) becomes (1.1 -> 1.3) because the common
2381 // parent is '1'. Here 'x.y.z' represents the hierarchy of statements.
2382 //
2383 // NOTE: this will be limited later in cases where we add barriers
2384 // to prevent this optimization.
2385 //
2386 if (level1 && level1 == level2 && level1 == level3 && level1 == level4) {
2387 PieceI->setEndLocation(PieceNextI->getEndLocation());
2388 path.erase(NextI);
2389 hasChanges = true;
2390 continue;
2391 }
2392
2393 // Rule II.
2394 //
2395 // Eliminate edges between subexpressions and parent expressions
2396 // when the subexpression is consumed.
2397 //
2398 // NOTE: this will be limited later in cases where we add barriers
2399 // to prevent this optimization.
2400 //
2401 if (s1End && s1End == s2Start && level2) {
2402 bool removeEdge = false;
2403 // Remove edges into the increment or initialization of a
2404 // loop that have no interleaving event. This means that
2405 // they aren't interesting.
2406 if (isIncrementOrInitInForLoop(s1End, level2))
2407 removeEdge = true;
2408 // Next only consider edges that are not anchored on
2409 // the condition of a terminator. This are intermediate edges
2410 // that we might want to trim.
2411 else if (!isConditionForTerminator(level2, s1End)) {
2412 // Trim edges on expressions that are consumed by
2413 // the parent expression.
2414 if (isa<Expr>(s1End) && PM.isConsumedExpr(cast<Expr>(s1End))) {
2415 removeEdge = true;
2416 }
2417 // Trim edges where a lexical containment doesn't exist.
2418 // For example:
2419 //
2420 // X -> Y -> Z
2421 //
2422 // If 'Z' lexically contains Y (it is an ancestor) and
2423 // 'X' does not lexically contain Y (it is a descendant OR
2424 // it has no lexical relationship at all) then trim.
2425 //
2426 // This can eliminate edges where we dive into a subexpression
2427 // and then pop back out, etc.
2428 else if (s1Start && s2End &&
2429 lexicalContains(PM, s2Start, s2End) &&
2430 !lexicalContains(PM, s1End, s1Start)) {
2431 removeEdge = true;
2432 }
2433 // Trim edges from a subexpression back to the top level if the
2434 // subexpression is on a different line.
2435 //
2436 // A.1 -> A -> B
2437 // becomes
2438 // A.1 -> B
2439 //
2440 // These edges just look ugly and don't usually add anything.
2441 else if (s1Start && s2End &&
2442 lexicalContains(PM, s1Start, s1End)) {
2443 SourceRange EdgeRange(PieceI->getEndLocation().asLocation(),
2444 PieceI->getStartLocation().asLocation());
2445 if (!getLengthOnSingleLine(SM, EdgeRange).hasValue())
2446 removeEdge = true;
2447 }
2448 }
2449
2450 if (removeEdge) {
2451 PieceI->setEndLocation(PieceNextI->getEndLocation());
2452 path.erase(NextI);
2453 hasChanges = true;
2454 continue;
2455 }
2456 }
2457
2458 // Optimize edges for ObjC fast-enumeration loops.
2459 //
2460 // (X -> collection) -> (collection -> element)
2461 //
2462 // becomes:
2463 //
2464 // (X -> element)
2465 if (s1End == s2Start) {
2466 const ObjCForCollectionStmt *FS =
2467 dyn_cast_or_null<ObjCForCollectionStmt>(level3);
2468 if (FS && FS->getCollection()->IgnoreParens() == s2Start &&
2469 s2End == FS->getElement()) {
2470 PieceI->setEndLocation(PieceNextI->getEndLocation());
2471 path.erase(NextI);
2472 hasChanges = true;
2473 continue;
2474 }
2475 }
2476
2477 // No changes at this index? Move to the next one.
2478 ++I;
2479 }
2480
2481 if (!hasChanges) {
2482 // Adjust edges into subexpressions to make them more uniform
2483 // and aesthetically pleasing.
2484 addContextEdges(path, SM, PM, LC);
2485 // Remove "cyclical" edges that include one or more context edges.
2486 removeContextCycles(path, SM, PM);
2487 // Hoist edges originating from branch conditions to branches
2488 // for simple branches.
2489 simplifySimpleBranches(path);
2490 // Remove any puny edges left over after primary optimization pass.
2491 removePunyEdges(path, SM, PM);
2492 // Remove identical events.
2493 removeIdenticalEvents(path);
2494 }
2495
2496 return hasChanges;
2497 }
2498
2499 /// Drop the very first edge in a path, which should be a function entry edge.
2500 ///
2501 /// If the first edge is not a function entry edge (say, because the first
2502 /// statement had an invalid source location), this function does nothing.
2503 // FIXME: We should just generate invalid edges anyway and have the optimizer
2504 // deal with them.
dropFunctionEntryEdge(PathPieces & Path,LocationContextMap & LCM,SourceManager & SM)2505 static void dropFunctionEntryEdge(PathPieces &Path,
2506 LocationContextMap &LCM,
2507 SourceManager &SM) {
2508 const PathDiagnosticControlFlowPiece *FirstEdge =
2509 dyn_cast<PathDiagnosticControlFlowPiece>(Path.front());
2510 if (!FirstEdge)
2511 return;
2512
2513 const Decl *D = LCM[&Path]->getDecl();
2514 PathDiagnosticLocation EntryLoc = PathDiagnosticLocation::createBegin(D, SM);
2515 if (FirstEdge->getStartLocation() != EntryLoc)
2516 return;
2517
2518 Path.pop_front();
2519 }
2520
2521
2522 //===----------------------------------------------------------------------===//
2523 // Methods for BugType and subclasses.
2524 //===----------------------------------------------------------------------===//
anchor()2525 void BugType::anchor() { }
2526
FlushReports(BugReporter & BR)2527 void BugType::FlushReports(BugReporter &BR) {}
2528
anchor()2529 void BuiltinBug::anchor() {}
2530
2531 //===----------------------------------------------------------------------===//
2532 // Methods for BugReport and subclasses.
2533 //===----------------------------------------------------------------------===//
2534
anchor()2535 void BugReport::NodeResolver::anchor() {}
2536
addVisitor(std::unique_ptr<BugReporterVisitor> visitor)2537 void BugReport::addVisitor(std::unique_ptr<BugReporterVisitor> visitor) {
2538 if (!visitor)
2539 return;
2540
2541 llvm::FoldingSetNodeID ID;
2542 visitor->Profile(ID);
2543 void *InsertPos;
2544
2545 if (CallbacksSet.FindNodeOrInsertPos(ID, InsertPos))
2546 return;
2547
2548 CallbacksSet.InsertNode(visitor.get(), InsertPos);
2549 Callbacks.push_back(std::move(visitor));
2550 ++ConfigurationChangeToken;
2551 }
2552
~BugReport()2553 BugReport::~BugReport() {
2554 while (!interestingSymbols.empty()) {
2555 popInterestingSymbolsAndRegions();
2556 }
2557 }
2558
getDeclWithIssue() const2559 const Decl *BugReport::getDeclWithIssue() const {
2560 if (DeclWithIssue)
2561 return DeclWithIssue;
2562
2563 const ExplodedNode *N = getErrorNode();
2564 if (!N)
2565 return nullptr;
2566
2567 const LocationContext *LC = N->getLocationContext();
2568 return LC->getCurrentStackFrame()->getDecl();
2569 }
2570
Profile(llvm::FoldingSetNodeID & hash) const2571 void BugReport::Profile(llvm::FoldingSetNodeID& hash) const {
2572 hash.AddPointer(&BT);
2573 hash.AddString(Description);
2574 PathDiagnosticLocation UL = getUniqueingLocation();
2575 if (UL.isValid()) {
2576 UL.Profile(hash);
2577 } else if (Location.isValid()) {
2578 Location.Profile(hash);
2579 } else {
2580 assert(ErrorNode);
2581 hash.AddPointer(GetCurrentOrPreviousStmt(ErrorNode));
2582 }
2583
2584 for (SmallVectorImpl<SourceRange>::const_iterator I =
2585 Ranges.begin(), E = Ranges.end(); I != E; ++I) {
2586 const SourceRange range = *I;
2587 if (!range.isValid())
2588 continue;
2589 hash.AddInteger(range.getBegin().getRawEncoding());
2590 hash.AddInteger(range.getEnd().getRawEncoding());
2591 }
2592 }
2593
markInteresting(SymbolRef sym)2594 void BugReport::markInteresting(SymbolRef sym) {
2595 if (!sym)
2596 return;
2597
2598 // If the symbol wasn't already in our set, note a configuration change.
2599 if (getInterestingSymbols().insert(sym).second)
2600 ++ConfigurationChangeToken;
2601
2602 if (const SymbolMetadata *meta = dyn_cast<SymbolMetadata>(sym))
2603 getInterestingRegions().insert(meta->getRegion());
2604 }
2605
markInteresting(const MemRegion * R)2606 void BugReport::markInteresting(const MemRegion *R) {
2607 if (!R)
2608 return;
2609
2610 // If the base region wasn't already in our set, note a configuration change.
2611 R = R->getBaseRegion();
2612 if (getInterestingRegions().insert(R).second)
2613 ++ConfigurationChangeToken;
2614
2615 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R))
2616 getInterestingSymbols().insert(SR->getSymbol());
2617 }
2618
markInteresting(SVal V)2619 void BugReport::markInteresting(SVal V) {
2620 markInteresting(V.getAsRegion());
2621 markInteresting(V.getAsSymbol());
2622 }
2623
markInteresting(const LocationContext * LC)2624 void BugReport::markInteresting(const LocationContext *LC) {
2625 if (!LC)
2626 return;
2627 InterestingLocationContexts.insert(LC);
2628 }
2629
isInteresting(SVal V)2630 bool BugReport::isInteresting(SVal V) {
2631 return isInteresting(V.getAsRegion()) || isInteresting(V.getAsSymbol());
2632 }
2633
isInteresting(SymbolRef sym)2634 bool BugReport::isInteresting(SymbolRef sym) {
2635 if (!sym)
2636 return false;
2637 // We don't currently consider metadata symbols to be interesting
2638 // even if we know their region is interesting. Is that correct behavior?
2639 return getInterestingSymbols().count(sym);
2640 }
2641
isInteresting(const MemRegion * R)2642 bool BugReport::isInteresting(const MemRegion *R) {
2643 if (!R)
2644 return false;
2645 R = R->getBaseRegion();
2646 bool b = getInterestingRegions().count(R);
2647 if (b)
2648 return true;
2649 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R))
2650 return getInterestingSymbols().count(SR->getSymbol());
2651 return false;
2652 }
2653
isInteresting(const LocationContext * LC)2654 bool BugReport::isInteresting(const LocationContext *LC) {
2655 if (!LC)
2656 return false;
2657 return InterestingLocationContexts.count(LC);
2658 }
2659
lazyInitializeInterestingSets()2660 void BugReport::lazyInitializeInterestingSets() {
2661 if (interestingSymbols.empty()) {
2662 interestingSymbols.push_back(new Symbols());
2663 interestingRegions.push_back(new Regions());
2664 }
2665 }
2666
getInterestingSymbols()2667 BugReport::Symbols &BugReport::getInterestingSymbols() {
2668 lazyInitializeInterestingSets();
2669 return *interestingSymbols.back();
2670 }
2671
getInterestingRegions()2672 BugReport::Regions &BugReport::getInterestingRegions() {
2673 lazyInitializeInterestingSets();
2674 return *interestingRegions.back();
2675 }
2676
pushInterestingSymbolsAndRegions()2677 void BugReport::pushInterestingSymbolsAndRegions() {
2678 interestingSymbols.push_back(new Symbols(getInterestingSymbols()));
2679 interestingRegions.push_back(new Regions(getInterestingRegions()));
2680 }
2681
popInterestingSymbolsAndRegions()2682 void BugReport::popInterestingSymbolsAndRegions() {
2683 delete interestingSymbols.pop_back_val();
2684 delete interestingRegions.pop_back_val();
2685 }
2686
getStmt() const2687 const Stmt *BugReport::getStmt() const {
2688 if (!ErrorNode)
2689 return nullptr;
2690
2691 ProgramPoint ProgP = ErrorNode->getLocation();
2692 const Stmt *S = nullptr;
2693
2694 if (Optional<BlockEntrance> BE = ProgP.getAs<BlockEntrance>()) {
2695 CFGBlock &Exit = ProgP.getLocationContext()->getCFG()->getExit();
2696 if (BE->getBlock() == &Exit)
2697 S = GetPreviousStmt(ErrorNode);
2698 }
2699 if (!S)
2700 S = PathDiagnosticLocation::getStmt(ErrorNode);
2701
2702 return S;
2703 }
2704
getRanges()2705 llvm::iterator_range<BugReport::ranges_iterator> BugReport::getRanges() {
2706 // If no custom ranges, add the range of the statement corresponding to
2707 // the error node.
2708 if (Ranges.empty()) {
2709 if (const Expr *E = dyn_cast_or_null<Expr>(getStmt()))
2710 addRange(E->getSourceRange());
2711 else
2712 return llvm::make_range(ranges_iterator(), ranges_iterator());
2713 }
2714
2715 // User-specified absence of range info.
2716 if (Ranges.size() == 1 && !Ranges.begin()->isValid())
2717 return llvm::make_range(ranges_iterator(), ranges_iterator());
2718
2719 return llvm::iterator_range<BugReport::ranges_iterator>(Ranges.begin(),
2720 Ranges.end());
2721 }
2722
getLocation(const SourceManager & SM) const2723 PathDiagnosticLocation BugReport::getLocation(const SourceManager &SM) const {
2724 if (ErrorNode) {
2725 assert(!Location.isValid() &&
2726 "Either Location or ErrorNode should be specified but not both.");
2727 return PathDiagnosticLocation::createEndOfPath(ErrorNode, SM);
2728 }
2729
2730 assert(Location.isValid());
2731 return Location;
2732 }
2733
2734 //===----------------------------------------------------------------------===//
2735 // Methods for BugReporter and subclasses.
2736 //===----------------------------------------------------------------------===//
2737
~BugReportEquivClass()2738 BugReportEquivClass::~BugReportEquivClass() { }
~GRBugReporter()2739 GRBugReporter::~GRBugReporter() { }
~BugReporterData()2740 BugReporterData::~BugReporterData() {}
2741
getGraph()2742 ExplodedGraph &GRBugReporter::getGraph() { return Eng.getGraph(); }
2743
2744 ProgramStateManager&
getStateManager()2745 GRBugReporter::getStateManager() { return Eng.getStateManager(); }
2746
~BugReporter()2747 BugReporter::~BugReporter() {
2748 FlushReports();
2749
2750 // Free the bug reports we are tracking.
2751 typedef std::vector<BugReportEquivClass *> ContTy;
2752 for (ContTy::iterator I = EQClassesVector.begin(), E = EQClassesVector.end();
2753 I != E; ++I) {
2754 delete *I;
2755 }
2756 }
2757
FlushReports()2758 void BugReporter::FlushReports() {
2759 if (BugTypes.isEmpty())
2760 return;
2761
2762 // First flush the warnings for each BugType. This may end up creating new
2763 // warnings and new BugTypes.
2764 // FIXME: Only NSErrorChecker needs BugType's FlushReports.
2765 // Turn NSErrorChecker into a proper checker and remove this.
2766 SmallVector<const BugType *, 16> bugTypes(BugTypes.begin(), BugTypes.end());
2767 for (SmallVectorImpl<const BugType *>::iterator
2768 I = bugTypes.begin(), E = bugTypes.end(); I != E; ++I)
2769 const_cast<BugType*>(*I)->FlushReports(*this);
2770
2771 // We need to flush reports in deterministic order to ensure the order
2772 // of the reports is consistent between runs.
2773 typedef std::vector<BugReportEquivClass *> ContVecTy;
2774 for (ContVecTy::iterator EI=EQClassesVector.begin(), EE=EQClassesVector.end();
2775 EI != EE; ++EI){
2776 BugReportEquivClass& EQ = **EI;
2777 FlushReport(EQ);
2778 }
2779
2780 // BugReporter owns and deletes only BugTypes created implicitly through
2781 // EmitBasicReport.
2782 // FIXME: There are leaks from checkers that assume that the BugTypes they
2783 // create will be destroyed by the BugReporter.
2784 llvm::DeleteContainerSeconds(StrBugTypes);
2785
2786 // Remove all references to the BugType objects.
2787 BugTypes = F.getEmptySet();
2788 }
2789
2790 //===----------------------------------------------------------------------===//
2791 // PathDiagnostics generation.
2792 //===----------------------------------------------------------------------===//
2793
2794 namespace {
2795 /// A wrapper around a report graph, which contains only a single path, and its
2796 /// node maps.
2797 class ReportGraph {
2798 public:
2799 InterExplodedGraphMap BackMap;
2800 std::unique_ptr<ExplodedGraph> Graph;
2801 const ExplodedNode *ErrorNode;
2802 size_t Index;
2803 };
2804
2805 /// A wrapper around a trimmed graph and its node maps.
2806 class TrimmedGraph {
2807 InterExplodedGraphMap InverseMap;
2808
2809 typedef llvm::DenseMap<const ExplodedNode *, unsigned> PriorityMapTy;
2810 PriorityMapTy PriorityMap;
2811
2812 typedef std::pair<const ExplodedNode *, size_t> NodeIndexPair;
2813 SmallVector<NodeIndexPair, 32> ReportNodes;
2814
2815 std::unique_ptr<ExplodedGraph> G;
2816
2817 /// A helper class for sorting ExplodedNodes by priority.
2818 template <bool Descending>
2819 class PriorityCompare {
2820 const PriorityMapTy &PriorityMap;
2821
2822 public:
PriorityCompare(const PriorityMapTy & M)2823 PriorityCompare(const PriorityMapTy &M) : PriorityMap(M) {}
2824
operator ()(const ExplodedNode * LHS,const ExplodedNode * RHS) const2825 bool operator()(const ExplodedNode *LHS, const ExplodedNode *RHS) const {
2826 PriorityMapTy::const_iterator LI = PriorityMap.find(LHS);
2827 PriorityMapTy::const_iterator RI = PriorityMap.find(RHS);
2828 PriorityMapTy::const_iterator E = PriorityMap.end();
2829
2830 if (LI == E)
2831 return Descending;
2832 if (RI == E)
2833 return !Descending;
2834
2835 return Descending ? LI->second > RI->second
2836 : LI->second < RI->second;
2837 }
2838
operator ()(const NodeIndexPair & LHS,const NodeIndexPair & RHS) const2839 bool operator()(const NodeIndexPair &LHS, const NodeIndexPair &RHS) const {
2840 return (*this)(LHS.first, RHS.first);
2841 }
2842 };
2843
2844 public:
2845 TrimmedGraph(const ExplodedGraph *OriginalGraph,
2846 ArrayRef<const ExplodedNode *> Nodes);
2847
2848 bool popNextReportGraph(ReportGraph &GraphWrapper);
2849 };
2850 }
2851
TrimmedGraph(const ExplodedGraph * OriginalGraph,ArrayRef<const ExplodedNode * > Nodes)2852 TrimmedGraph::TrimmedGraph(const ExplodedGraph *OriginalGraph,
2853 ArrayRef<const ExplodedNode *> Nodes) {
2854 // The trimmed graph is created in the body of the constructor to ensure
2855 // that the DenseMaps have been initialized already.
2856 InterExplodedGraphMap ForwardMap;
2857 G = OriginalGraph->trim(Nodes, &ForwardMap, &InverseMap);
2858
2859 // Find the (first) error node in the trimmed graph. We just need to consult
2860 // the node map which maps from nodes in the original graph to nodes
2861 // in the new graph.
2862 llvm::SmallPtrSet<const ExplodedNode *, 32> RemainingNodes;
2863
2864 for (unsigned i = 0, count = Nodes.size(); i < count; ++i) {
2865 if (const ExplodedNode *NewNode = ForwardMap.lookup(Nodes[i])) {
2866 ReportNodes.push_back(std::make_pair(NewNode, i));
2867 RemainingNodes.insert(NewNode);
2868 }
2869 }
2870
2871 assert(!RemainingNodes.empty() && "No error node found in the trimmed graph");
2872
2873 // Perform a forward BFS to find all the shortest paths.
2874 std::queue<const ExplodedNode *> WS;
2875
2876 assert(G->num_roots() == 1);
2877 WS.push(*G->roots_begin());
2878 unsigned Priority = 0;
2879
2880 while (!WS.empty()) {
2881 const ExplodedNode *Node = WS.front();
2882 WS.pop();
2883
2884 PriorityMapTy::iterator PriorityEntry;
2885 bool IsNew;
2886 std::tie(PriorityEntry, IsNew) =
2887 PriorityMap.insert(std::make_pair(Node, Priority));
2888 ++Priority;
2889
2890 if (!IsNew) {
2891 assert(PriorityEntry->second <= Priority);
2892 continue;
2893 }
2894
2895 if (RemainingNodes.erase(Node))
2896 if (RemainingNodes.empty())
2897 break;
2898
2899 for (ExplodedNode::const_pred_iterator I = Node->succ_begin(),
2900 E = Node->succ_end();
2901 I != E; ++I)
2902 WS.push(*I);
2903 }
2904
2905 // Sort the error paths from longest to shortest.
2906 std::sort(ReportNodes.begin(), ReportNodes.end(),
2907 PriorityCompare<true>(PriorityMap));
2908 }
2909
popNextReportGraph(ReportGraph & GraphWrapper)2910 bool TrimmedGraph::popNextReportGraph(ReportGraph &GraphWrapper) {
2911 if (ReportNodes.empty())
2912 return false;
2913
2914 const ExplodedNode *OrigN;
2915 std::tie(OrigN, GraphWrapper.Index) = ReportNodes.pop_back_val();
2916 assert(PriorityMap.find(OrigN) != PriorityMap.end() &&
2917 "error node not accessible from root");
2918
2919 // Create a new graph with a single path. This is the graph
2920 // that will be returned to the caller.
2921 auto GNew = llvm::make_unique<ExplodedGraph>();
2922 GraphWrapper.BackMap.clear();
2923
2924 // Now walk from the error node up the BFS path, always taking the
2925 // predeccessor with the lowest number.
2926 ExplodedNode *Succ = nullptr;
2927 while (true) {
2928 // Create the equivalent node in the new graph with the same state
2929 // and location.
2930 ExplodedNode *NewN = GNew->getNode(OrigN->getLocation(), OrigN->getState(),
2931 OrigN->isSink());
2932
2933 // Store the mapping to the original node.
2934 InterExplodedGraphMap::const_iterator IMitr = InverseMap.find(OrigN);
2935 assert(IMitr != InverseMap.end() && "No mapping to original node.");
2936 GraphWrapper.BackMap[NewN] = IMitr->second;
2937
2938 // Link up the new node with the previous node.
2939 if (Succ)
2940 Succ->addPredecessor(NewN, *GNew);
2941 else
2942 GraphWrapper.ErrorNode = NewN;
2943
2944 Succ = NewN;
2945
2946 // Are we at the final node?
2947 if (OrigN->pred_empty()) {
2948 GNew->addRoot(NewN);
2949 break;
2950 }
2951
2952 // Find the next predeccessor node. We choose the node that is marked
2953 // with the lowest BFS number.
2954 OrigN = *std::min_element(OrigN->pred_begin(), OrigN->pred_end(),
2955 PriorityCompare<false>(PriorityMap));
2956 }
2957
2958 GraphWrapper.Graph = std::move(GNew);
2959
2960 return true;
2961 }
2962
2963
2964 /// CompactPathDiagnostic - This function postprocesses a PathDiagnostic object
2965 /// and collapses PathDiagosticPieces that are expanded by macros.
CompactPathDiagnostic(PathPieces & path,const SourceManager & SM)2966 static void CompactPathDiagnostic(PathPieces &path, const SourceManager& SM) {
2967 typedef std::vector<std::pair<IntrusiveRefCntPtr<PathDiagnosticMacroPiece>,
2968 SourceLocation> > MacroStackTy;
2969
2970 typedef std::vector<IntrusiveRefCntPtr<PathDiagnosticPiece> >
2971 PiecesTy;
2972
2973 MacroStackTy MacroStack;
2974 PiecesTy Pieces;
2975
2976 for (PathPieces::const_iterator I = path.begin(), E = path.end();
2977 I!=E; ++I) {
2978
2979 PathDiagnosticPiece *piece = I->get();
2980
2981 // Recursively compact calls.
2982 if (PathDiagnosticCallPiece *call=dyn_cast<PathDiagnosticCallPiece>(piece)){
2983 CompactPathDiagnostic(call->path, SM);
2984 }
2985
2986 // Get the location of the PathDiagnosticPiece.
2987 const FullSourceLoc Loc = piece->getLocation().asLocation();
2988
2989 // Determine the instantiation location, which is the location we group
2990 // related PathDiagnosticPieces.
2991 SourceLocation InstantiationLoc = Loc.isMacroID() ?
2992 SM.getExpansionLoc(Loc) :
2993 SourceLocation();
2994
2995 if (Loc.isFileID()) {
2996 MacroStack.clear();
2997 Pieces.push_back(piece);
2998 continue;
2999 }
3000
3001 assert(Loc.isMacroID());
3002
3003 // Is the PathDiagnosticPiece within the same macro group?
3004 if (!MacroStack.empty() && InstantiationLoc == MacroStack.back().second) {
3005 MacroStack.back().first->subPieces.push_back(piece);
3006 continue;
3007 }
3008
3009 // We aren't in the same group. Are we descending into a new macro
3010 // or are part of an old one?
3011 IntrusiveRefCntPtr<PathDiagnosticMacroPiece> MacroGroup;
3012
3013 SourceLocation ParentInstantiationLoc = InstantiationLoc.isMacroID() ?
3014 SM.getExpansionLoc(Loc) :
3015 SourceLocation();
3016
3017 // Walk the entire macro stack.
3018 while (!MacroStack.empty()) {
3019 if (InstantiationLoc == MacroStack.back().second) {
3020 MacroGroup = MacroStack.back().first;
3021 break;
3022 }
3023
3024 if (ParentInstantiationLoc == MacroStack.back().second) {
3025 MacroGroup = MacroStack.back().first;
3026 break;
3027 }
3028
3029 MacroStack.pop_back();
3030 }
3031
3032 if (!MacroGroup || ParentInstantiationLoc == MacroStack.back().second) {
3033 // Create a new macro group and add it to the stack.
3034 PathDiagnosticMacroPiece *NewGroup =
3035 new PathDiagnosticMacroPiece(
3036 PathDiagnosticLocation::createSingleLocation(piece->getLocation()));
3037
3038 if (MacroGroup)
3039 MacroGroup->subPieces.push_back(NewGroup);
3040 else {
3041 assert(InstantiationLoc.isFileID());
3042 Pieces.push_back(NewGroup);
3043 }
3044
3045 MacroGroup = NewGroup;
3046 MacroStack.push_back(std::make_pair(MacroGroup, InstantiationLoc));
3047 }
3048
3049 // Finally, add the PathDiagnosticPiece to the group.
3050 MacroGroup->subPieces.push_back(piece);
3051 }
3052
3053 // Now take the pieces and construct a new PathDiagnostic.
3054 path.clear();
3055
3056 path.insert(path.end(), Pieces.begin(), Pieces.end());
3057 }
3058
generatePathDiagnostic(PathDiagnostic & PD,PathDiagnosticConsumer & PC,ArrayRef<BugReport * > & bugReports)3059 bool GRBugReporter::generatePathDiagnostic(PathDiagnostic& PD,
3060 PathDiagnosticConsumer &PC,
3061 ArrayRef<BugReport *> &bugReports) {
3062 assert(!bugReports.empty());
3063
3064 bool HasValid = false;
3065 bool HasInvalid = false;
3066 SmallVector<const ExplodedNode *, 32> errorNodes;
3067 for (ArrayRef<BugReport*>::iterator I = bugReports.begin(),
3068 E = bugReports.end(); I != E; ++I) {
3069 if ((*I)->isValid()) {
3070 HasValid = true;
3071 errorNodes.push_back((*I)->getErrorNode());
3072 } else {
3073 // Keep the errorNodes list in sync with the bugReports list.
3074 HasInvalid = true;
3075 errorNodes.push_back(nullptr);
3076 }
3077 }
3078
3079 // If all the reports have been marked invalid by a previous path generation,
3080 // we're done.
3081 if (!HasValid)
3082 return false;
3083
3084 typedef PathDiagnosticConsumer::PathGenerationScheme PathGenerationScheme;
3085 PathGenerationScheme ActiveScheme = PC.getGenerationScheme();
3086
3087 if (ActiveScheme == PathDiagnosticConsumer::Extensive) {
3088 AnalyzerOptions &options = getAnalyzerOptions();
3089 if (options.getBooleanOption("path-diagnostics-alternate", true)) {
3090 ActiveScheme = PathDiagnosticConsumer::AlternateExtensive;
3091 }
3092 }
3093
3094 TrimmedGraph TrimG(&getGraph(), errorNodes);
3095 ReportGraph ErrorGraph;
3096
3097 while (TrimG.popNextReportGraph(ErrorGraph)) {
3098 // Find the BugReport with the original location.
3099 assert(ErrorGraph.Index < bugReports.size());
3100 BugReport *R = bugReports[ErrorGraph.Index];
3101 assert(R && "No original report found for sliced graph.");
3102 assert(R->isValid() && "Report selected by trimmed graph marked invalid.");
3103
3104 // Start building the path diagnostic...
3105 PathDiagnosticBuilder PDB(*this, R, ErrorGraph.BackMap, &PC);
3106 const ExplodedNode *N = ErrorGraph.ErrorNode;
3107
3108 // Register additional node visitors.
3109 R->addVisitor(llvm::make_unique<NilReceiverBRVisitor>());
3110 R->addVisitor(llvm::make_unique<ConditionBRVisitor>());
3111 R->addVisitor(llvm::make_unique<LikelyFalsePositiveSuppressionBRVisitor>());
3112
3113 BugReport::VisitorList visitors;
3114 unsigned origReportConfigToken, finalReportConfigToken;
3115 LocationContextMap LCM;
3116
3117 // While generating diagnostics, it's possible the visitors will decide
3118 // new symbols and regions are interesting, or add other visitors based on
3119 // the information they find. If they do, we need to regenerate the path
3120 // based on our new report configuration.
3121 do {
3122 // Get a clean copy of all the visitors.
3123 for (BugReport::visitor_iterator I = R->visitor_begin(),
3124 E = R->visitor_end(); I != E; ++I)
3125 visitors.push_back((*I)->clone());
3126
3127 // Clear out the active path from any previous work.
3128 PD.resetPath();
3129 origReportConfigToken = R->getConfigurationChangeToken();
3130
3131 // Generate the very last diagnostic piece - the piece is visible before
3132 // the trace is expanded.
3133 std::unique_ptr<PathDiagnosticPiece> LastPiece;
3134 for (BugReport::visitor_iterator I = visitors.begin(), E = visitors.end();
3135 I != E; ++I) {
3136 if (std::unique_ptr<PathDiagnosticPiece> Piece =
3137 (*I)->getEndPath(PDB, N, *R)) {
3138 assert (!LastPiece &&
3139 "There can only be one final piece in a diagnostic.");
3140 LastPiece = std::move(Piece);
3141 }
3142 }
3143
3144 if (ActiveScheme != PathDiagnosticConsumer::None) {
3145 if (!LastPiece)
3146 LastPiece = BugReporterVisitor::getDefaultEndPath(PDB, N, *R);
3147 assert(LastPiece);
3148 PD.setEndOfPath(std::move(LastPiece));
3149 }
3150
3151 // Make sure we get a clean location context map so we don't
3152 // hold onto old mappings.
3153 LCM.clear();
3154
3155 switch (ActiveScheme) {
3156 case PathDiagnosticConsumer::AlternateExtensive:
3157 GenerateAlternateExtensivePathDiagnostic(PD, PDB, N, LCM, visitors);
3158 break;
3159 case PathDiagnosticConsumer::Extensive:
3160 GenerateExtensivePathDiagnostic(PD, PDB, N, LCM, visitors);
3161 break;
3162 case PathDiagnosticConsumer::Minimal:
3163 GenerateMinimalPathDiagnostic(PD, PDB, N, LCM, visitors);
3164 break;
3165 case PathDiagnosticConsumer::None:
3166 GenerateVisitorsOnlyPathDiagnostic(PD, PDB, N, visitors);
3167 break;
3168 }
3169
3170 // Clean up the visitors we used.
3171 visitors.clear();
3172
3173 // Did anything change while generating this path?
3174 finalReportConfigToken = R->getConfigurationChangeToken();
3175 } while (finalReportConfigToken != origReportConfigToken);
3176
3177 if (!R->isValid())
3178 continue;
3179
3180 // Finally, prune the diagnostic path of uninteresting stuff.
3181 if (!PD.path.empty()) {
3182 if (R->shouldPrunePath() && getAnalyzerOptions().shouldPrunePaths()) {
3183 bool stillHasNotes = removeUnneededCalls(PD.getMutablePieces(), R, LCM);
3184 assert(stillHasNotes);
3185 (void)stillHasNotes;
3186 }
3187
3188 // Redirect all call pieces to have valid locations.
3189 adjustCallLocations(PD.getMutablePieces());
3190 removePiecesWithInvalidLocations(PD.getMutablePieces());
3191
3192 if (ActiveScheme == PathDiagnosticConsumer::AlternateExtensive) {
3193 SourceManager &SM = getSourceManager();
3194
3195 // Reduce the number of edges from a very conservative set
3196 // to an aesthetically pleasing subset that conveys the
3197 // necessary information.
3198 OptimizedCallsSet OCS;
3199 while (optimizeEdges(PD.getMutablePieces(), SM, OCS, LCM)) {}
3200
3201 // Drop the very first function-entry edge. It's not really necessary
3202 // for top-level functions.
3203 dropFunctionEntryEdge(PD.getMutablePieces(), LCM, SM);
3204 }
3205
3206 // Remove messages that are basically the same, and edges that may not
3207 // make sense.
3208 // We have to do this after edge optimization in the Extensive mode.
3209 removeRedundantMsgs(PD.getMutablePieces());
3210 removeEdgesToDefaultInitializers(PD.getMutablePieces());
3211 }
3212
3213 // We found a report and didn't suppress it.
3214 return true;
3215 }
3216
3217 // We suppressed all the reports in this equivalence class.
3218 assert(!HasInvalid && "Inconsistent suppression");
3219 (void)HasInvalid;
3220 return false;
3221 }
3222
Register(BugType * BT)3223 void BugReporter::Register(BugType *BT) {
3224 BugTypes = F.add(BugTypes, BT);
3225 }
3226
emitReport(BugReport * R)3227 void BugReporter::emitReport(BugReport* R) {
3228 // To guarantee memory release.
3229 std::unique_ptr<BugReport> UniqueR(R);
3230
3231 if (const ExplodedNode *E = R->getErrorNode()) {
3232 const AnalysisDeclContext *DeclCtx =
3233 E->getLocationContext()->getAnalysisDeclContext();
3234 // The source of autosynthesized body can be handcrafted AST or a model
3235 // file. The locations from handcrafted ASTs have no valid source locations
3236 // and have to be discarded. Locations from model files should be preserved
3237 // for processing and reporting.
3238 if (DeclCtx->isBodyAutosynthesized() &&
3239 !DeclCtx->isBodyAutosynthesizedFromModelFile())
3240 return;
3241 }
3242
3243 bool ValidSourceLoc = R->getLocation(getSourceManager()).isValid();
3244 assert(ValidSourceLoc);
3245 // If we mess up in a release build, we'd still prefer to just drop the bug
3246 // instead of trying to go on.
3247 if (!ValidSourceLoc)
3248 return;
3249
3250 // Compute the bug report's hash to determine its equivalence class.
3251 llvm::FoldingSetNodeID ID;
3252 R->Profile(ID);
3253
3254 // Lookup the equivance class. If there isn't one, create it.
3255 BugType& BT = R->getBugType();
3256 Register(&BT);
3257 void *InsertPos;
3258 BugReportEquivClass* EQ = EQClasses.FindNodeOrInsertPos(ID, InsertPos);
3259
3260 if (!EQ) {
3261 EQ = new BugReportEquivClass(std::move(UniqueR));
3262 EQClasses.InsertNode(EQ, InsertPos);
3263 EQClassesVector.push_back(EQ);
3264 } else
3265 EQ->AddReport(std::move(UniqueR));
3266 }
3267
3268
3269 //===----------------------------------------------------------------------===//
3270 // Emitting reports in equivalence classes.
3271 //===----------------------------------------------------------------------===//
3272
3273 namespace {
3274 struct FRIEC_WLItem {
3275 const ExplodedNode *N;
3276 ExplodedNode::const_succ_iterator I, E;
3277
FRIEC_WLItem__anon889b26420411::FRIEC_WLItem3278 FRIEC_WLItem(const ExplodedNode *n)
3279 : N(n), I(N->succ_begin()), E(N->succ_end()) {}
3280 };
3281 }
3282
3283 static BugReport *
FindReportInEquivalenceClass(BugReportEquivClass & EQ,SmallVectorImpl<BugReport * > & bugReports)3284 FindReportInEquivalenceClass(BugReportEquivClass& EQ,
3285 SmallVectorImpl<BugReport*> &bugReports) {
3286
3287 BugReportEquivClass::iterator I = EQ.begin(), E = EQ.end();
3288 assert(I != E);
3289 BugType& BT = I->getBugType();
3290
3291 // If we don't need to suppress any of the nodes because they are
3292 // post-dominated by a sink, simply add all the nodes in the equivalence class
3293 // to 'Nodes'. Any of the reports will serve as a "representative" report.
3294 if (!BT.isSuppressOnSink()) {
3295 BugReport *R = I;
3296 for (BugReportEquivClass::iterator I=EQ.begin(), E=EQ.end(); I!=E; ++I) {
3297 const ExplodedNode *N = I->getErrorNode();
3298 if (N) {
3299 R = I;
3300 bugReports.push_back(R);
3301 }
3302 }
3303 return R;
3304 }
3305
3306 // For bug reports that should be suppressed when all paths are post-dominated
3307 // by a sink node, iterate through the reports in the equivalence class
3308 // until we find one that isn't post-dominated (if one exists). We use a
3309 // DFS traversal of the ExplodedGraph to find a non-sink node. We could write
3310 // this as a recursive function, but we don't want to risk blowing out the
3311 // stack for very long paths.
3312 BugReport *exampleReport = nullptr;
3313
3314 for (; I != E; ++I) {
3315 const ExplodedNode *errorNode = I->getErrorNode();
3316
3317 if (!errorNode)
3318 continue;
3319 if (errorNode->isSink()) {
3320 llvm_unreachable(
3321 "BugType::isSuppressSink() should not be 'true' for sink end nodes");
3322 }
3323 // No successors? By definition this nodes isn't post-dominated by a sink.
3324 if (errorNode->succ_empty()) {
3325 bugReports.push_back(I);
3326 if (!exampleReport)
3327 exampleReport = I;
3328 continue;
3329 }
3330
3331 // At this point we know that 'N' is not a sink and it has at least one
3332 // successor. Use a DFS worklist to find a non-sink end-of-path node.
3333 typedef FRIEC_WLItem WLItem;
3334 typedef SmallVector<WLItem, 10> DFSWorkList;
3335 llvm::DenseMap<const ExplodedNode *, unsigned> Visited;
3336
3337 DFSWorkList WL;
3338 WL.push_back(errorNode);
3339 Visited[errorNode] = 1;
3340
3341 while (!WL.empty()) {
3342 WLItem &WI = WL.back();
3343 assert(!WI.N->succ_empty());
3344
3345 for (; WI.I != WI.E; ++WI.I) {
3346 const ExplodedNode *Succ = *WI.I;
3347 // End-of-path node?
3348 if (Succ->succ_empty()) {
3349 // If we found an end-of-path node that is not a sink.
3350 if (!Succ->isSink()) {
3351 bugReports.push_back(I);
3352 if (!exampleReport)
3353 exampleReport = I;
3354 WL.clear();
3355 break;
3356 }
3357 // Found a sink? Continue on to the next successor.
3358 continue;
3359 }
3360 // Mark the successor as visited. If it hasn't been explored,
3361 // enqueue it to the DFS worklist.
3362 unsigned &mark = Visited[Succ];
3363 if (!mark) {
3364 mark = 1;
3365 WL.push_back(Succ);
3366 break;
3367 }
3368 }
3369
3370 // The worklist may have been cleared at this point. First
3371 // check if it is empty before checking the last item.
3372 if (!WL.empty() && &WL.back() == &WI)
3373 WL.pop_back();
3374 }
3375 }
3376
3377 // ExampleReport will be NULL if all the nodes in the equivalence class
3378 // were post-dominated by sinks.
3379 return exampleReport;
3380 }
3381
FlushReport(BugReportEquivClass & EQ)3382 void BugReporter::FlushReport(BugReportEquivClass& EQ) {
3383 SmallVector<BugReport*, 10> bugReports;
3384 BugReport *exampleReport = FindReportInEquivalenceClass(EQ, bugReports);
3385 if (exampleReport) {
3386 for (PathDiagnosticConsumer *PDC : getPathDiagnosticConsumers()) {
3387 FlushReport(exampleReport, *PDC, bugReports);
3388 }
3389 }
3390 }
3391
FlushReport(BugReport * exampleReport,PathDiagnosticConsumer & PD,ArrayRef<BugReport * > bugReports)3392 void BugReporter::FlushReport(BugReport *exampleReport,
3393 PathDiagnosticConsumer &PD,
3394 ArrayRef<BugReport*> bugReports) {
3395
3396 // FIXME: Make sure we use the 'R' for the path that was actually used.
3397 // Probably doesn't make a difference in practice.
3398 BugType& BT = exampleReport->getBugType();
3399
3400 std::unique_ptr<PathDiagnostic> D(new PathDiagnostic(
3401 exampleReport->getBugType().getCheckName(),
3402 exampleReport->getDeclWithIssue(), exampleReport->getBugType().getName(),
3403 exampleReport->getDescription(),
3404 exampleReport->getShortDescription(/*Fallback=*/false), BT.getCategory(),
3405 exampleReport->getUniqueingLocation(),
3406 exampleReport->getUniqueingDecl()));
3407
3408 MaxBugClassSize = std::max(bugReports.size(),
3409 static_cast<size_t>(MaxBugClassSize));
3410
3411 // Generate the full path diagnostic, using the generation scheme
3412 // specified by the PathDiagnosticConsumer. Note that we have to generate
3413 // path diagnostics even for consumers which do not support paths, because
3414 // the BugReporterVisitors may mark this bug as a false positive.
3415 if (!bugReports.empty())
3416 if (!generatePathDiagnostic(*D.get(), PD, bugReports))
3417 return;
3418
3419 MaxValidBugClassSize = std::max(bugReports.size(),
3420 static_cast<size_t>(MaxValidBugClassSize));
3421
3422 // Examine the report and see if the last piece is in a header. Reset the
3423 // report location to the last piece in the main source file.
3424 AnalyzerOptions& Opts = getAnalyzerOptions();
3425 if (Opts.shouldReportIssuesInMainSourceFile() && !Opts.AnalyzeAll)
3426 D->resetDiagnosticLocationToMainFile();
3427
3428 // If the path is empty, generate a single step path with the location
3429 // of the issue.
3430 if (D->path.empty()) {
3431 PathDiagnosticLocation L = exampleReport->getLocation(getSourceManager());
3432 auto piece = llvm::make_unique<PathDiagnosticEventPiece>(
3433 L, exampleReport->getDescription());
3434 for (const SourceRange &Range : exampleReport->getRanges())
3435 piece->addRange(Range);
3436 D->setEndOfPath(std::move(piece));
3437 }
3438
3439 // Get the meta data.
3440 const BugReport::ExtraTextList &Meta = exampleReport->getExtraText();
3441 for (BugReport::ExtraTextList::const_iterator i = Meta.begin(),
3442 e = Meta.end(); i != e; ++i) {
3443 D->addMeta(*i);
3444 }
3445
3446 PD.HandlePathDiagnostic(std::move(D));
3447 }
3448
EmitBasicReport(const Decl * DeclWithIssue,const CheckerBase * Checker,StringRef Name,StringRef Category,StringRef Str,PathDiagnosticLocation Loc,ArrayRef<SourceRange> Ranges)3449 void BugReporter::EmitBasicReport(const Decl *DeclWithIssue,
3450 const CheckerBase *Checker,
3451 StringRef Name, StringRef Category,
3452 StringRef Str, PathDiagnosticLocation Loc,
3453 ArrayRef<SourceRange> Ranges) {
3454 EmitBasicReport(DeclWithIssue, Checker->getCheckName(), Name, Category, Str,
3455 Loc, Ranges);
3456 }
EmitBasicReport(const Decl * DeclWithIssue,CheckName CheckName,StringRef name,StringRef category,StringRef str,PathDiagnosticLocation Loc,ArrayRef<SourceRange> Ranges)3457 void BugReporter::EmitBasicReport(const Decl *DeclWithIssue,
3458 CheckName CheckName,
3459 StringRef name, StringRef category,
3460 StringRef str, PathDiagnosticLocation Loc,
3461 ArrayRef<SourceRange> Ranges) {
3462
3463 // 'BT' is owned by BugReporter.
3464 BugType *BT = getBugTypeForName(CheckName, name, category);
3465 BugReport *R = new BugReport(*BT, str, Loc);
3466 R->setDeclWithIssue(DeclWithIssue);
3467 for (ArrayRef<SourceRange>::iterator I = Ranges.begin(), E = Ranges.end();
3468 I != E; ++I)
3469 R->addRange(*I);
3470 emitReport(R);
3471 }
3472
getBugTypeForName(CheckName CheckName,StringRef name,StringRef category)3473 BugType *BugReporter::getBugTypeForName(CheckName CheckName, StringRef name,
3474 StringRef category) {
3475 SmallString<136> fullDesc;
3476 llvm::raw_svector_ostream(fullDesc) << CheckName.getName() << ":" << name
3477 << ":" << category;
3478 BugType *&BT = StrBugTypes[fullDesc];
3479 if (!BT)
3480 BT = new BugType(CheckName, name, category);
3481 return BT;
3482 }
3483
dump() const3484 LLVM_DUMP_METHOD void PathPieces::dump() const {
3485 unsigned index = 0;
3486 for (PathPieces::const_iterator I = begin(), E = end(); I != E; ++I) {
3487 llvm::errs() << "[" << index++ << "] ";
3488 (*I)->dump();
3489 llvm::errs() << "\n";
3490 }
3491 }
3492
dump() const3493 void PathDiagnosticCallPiece::dump() const {
3494 llvm::errs() << "CALL\n--------------\n";
3495
3496 if (const Stmt *SLoc = getLocStmt(getLocation()))
3497 SLoc->dump();
3498 else if (const NamedDecl *ND = dyn_cast<NamedDecl>(getCallee()))
3499 llvm::errs() << *ND << "\n";
3500 else
3501 getLocation().dump();
3502 }
3503
dump() const3504 void PathDiagnosticEventPiece::dump() const {
3505 llvm::errs() << "EVENT\n--------------\n";
3506 llvm::errs() << getString() << "\n";
3507 llvm::errs() << " ---- at ----\n";
3508 getLocation().dump();
3509 }
3510
dump() const3511 void PathDiagnosticControlFlowPiece::dump() const {
3512 llvm::errs() << "CONTROL\n--------------\n";
3513 getStartLocation().dump();
3514 llvm::errs() << " ---- to ----\n";
3515 getEndLocation().dump();
3516 }
3517
dump() const3518 void PathDiagnosticMacroPiece::dump() const {
3519 llvm::errs() << "MACRO\n--------------\n";
3520 // FIXME: Print which macro is being invoked.
3521 }
3522
dump() const3523 void PathDiagnosticLocation::dump() const {
3524 if (!isValid()) {
3525 llvm::errs() << "<INVALID>\n";
3526 return;
3527 }
3528
3529 switch (K) {
3530 case RangeK:
3531 // FIXME: actually print the range.
3532 llvm::errs() << "<range>\n";
3533 break;
3534 case SingleLocK:
3535 asLocation().dump();
3536 llvm::errs() << "\n";
3537 break;
3538 case StmtK:
3539 if (S)
3540 S->dump();
3541 else
3542 llvm::errs() << "<NULL STMT>\n";
3543 break;
3544 case DeclK:
3545 if (const NamedDecl *ND = dyn_cast_or_null<NamedDecl>(D))
3546 llvm::errs() << *ND << "\n";
3547 else if (isa<BlockDecl>(D))
3548 // FIXME: Make this nicer.
3549 llvm::errs() << "<block>\n";
3550 else if (D)
3551 llvm::errs() << "<unknown decl>\n";
3552 else
3553 llvm::errs() << "<NULL DECL>\n";
3554 break;
3555 }
3556 }
3557