1 // BugReporter.cpp - Generate PathDiagnostics for Bugs ------------*- C++ -*--//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines BugReporter, a utility class for generating
13 //===----------------------------------------------------------------------===//
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/CFGStmtMap.h"
25 #include "clang/Analysis/ProgramPoint.h"
26 #include "clang/Basic/SourceManager.h"
27 #include "clang/StaticAnalyzer/Core/BugReporter/BugType.h"
28 #include "clang/StaticAnalyzer/Core/BugReporter/PathDiagnostic.h"
29 #include "clang/StaticAnalyzer/Core/PathSensitive/ExprEngine.h"
30 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/IntrusiveRefCntPtr.h"
32 #include "llvm/ADT/STLExtras.h"
33 #include "llvm/ADT/SmallString.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Support/raw_ostream.h"
39 using namespace clang;
42 #define DEBUG_TYPE "BugReporter"
44 STATISTIC(MaxBugClassSize,
45 "The maximum number of bug reports in the same equivalence class");
46 STATISTIC(MaxValidBugClassSize,
47 "The maximum number of bug reports in the same equivalence class "
48 "where at least one report is valid (not suppressed)");
50 BugReporterVisitor::~BugReporterVisitor() {}
52 void BugReporterContext::anchor() {}
54 //===----------------------------------------------------------------------===//
55 // Helper routines for walking the ExplodedGraph and fetching statements.
56 //===----------------------------------------------------------------------===//
58 static const Stmt *GetPreviousStmt(const ExplodedNode *N) {
59 for (N = N->getFirstPred(); N; N = N->getFirstPred())
60 if (const Stmt *S = PathDiagnosticLocation::getStmt(N))
66 static inline const Stmt*
67 GetCurrentOrPreviousStmt(const ExplodedNode *N) {
68 if (const Stmt *S = PathDiagnosticLocation::getStmt(N))
71 return GetPreviousStmt(N);
74 //===----------------------------------------------------------------------===//
75 // Diagnostic cleanup.
76 //===----------------------------------------------------------------------===//
78 static PathDiagnosticEventPiece *
79 eventsDescribeSameCondition(PathDiagnosticEventPiece *X,
80 PathDiagnosticEventPiece *Y) {
81 // Prefer diagnostics that come from ConditionBRVisitor over
82 // those that came from TrackConstraintBRVisitor,
83 // unless the one from ConditionBRVisitor is
84 // its generic fallback diagnostic.
85 const void *tagPreferred = ConditionBRVisitor::getTag();
86 const void *tagLesser = TrackConstraintBRVisitor::getTag();
88 if (X->getLocation() != Y->getLocation())
91 if (X->getTag() == tagPreferred && Y->getTag() == tagLesser)
92 return ConditionBRVisitor::isPieceMessageGeneric(X) ? Y : X;
94 if (Y->getTag() == tagPreferred && X->getTag() == tagLesser)
95 return ConditionBRVisitor::isPieceMessageGeneric(Y) ? X : Y;
100 /// An optimization pass over PathPieces that removes redundant diagnostics
101 /// generated by both ConditionBRVisitor and TrackConstraintBRVisitor. Both
102 /// BugReporterVisitors use different methods to generate diagnostics, with
103 /// one capable of emitting diagnostics in some cases but not in others. This
104 /// can lead to redundant diagnostic pieces at the same point in a path.
105 static void removeRedundantMsgs(PathPieces &path) {
106 unsigned N = path.size();
109 // NOTE: this loop intentionally is not using an iterator. Instead, we
110 // are streaming the path and modifying it in place. This is done by
111 // grabbing the front, processing it, and if we decide to keep it append
112 // it to the end of the path. The entire path is processed in this way.
113 for (unsigned i = 0; i < N; ++i) {
114 auto piece = std::move(path.front());
117 switch (piece->getKind()) {
118 case PathDiagnosticPiece::Call:
119 removeRedundantMsgs(cast<PathDiagnosticCallPiece>(*piece).path);
121 case PathDiagnosticPiece::Macro:
122 removeRedundantMsgs(cast<PathDiagnosticMacroPiece>(*piece).subPieces);
124 case PathDiagnosticPiece::ControlFlow:
126 case PathDiagnosticPiece::Event: {
130 if (PathDiagnosticEventPiece *nextEvent =
131 dyn_cast<PathDiagnosticEventPiece>(path.front().get())) {
132 PathDiagnosticEventPiece *event =
133 cast<PathDiagnosticEventPiece>(piece.get());
134 // Check to see if we should keep one of the two pieces. If we
135 // come up with a preference, record which piece to keep, and consume
136 // another piece from the path.
137 if (auto *pieceToKeep =
138 eventsDescribeSameCondition(event, nextEvent)) {
139 piece = std::move(pieceToKeep == event ? piece : path.front());
146 case PathDiagnosticPiece::Note:
149 path.push_back(std::move(piece));
153 /// A map from PathDiagnosticPiece to the LocationContext of the inlined
154 /// function call it represents.
155 typedef llvm::DenseMap<const PathPieces *, const LocationContext *>
158 /// Recursively scan through a path and prune out calls and macros pieces
159 /// that aren't needed. Return true if afterwards the path contains
160 /// "interesting stuff" which means it shouldn't be pruned from the parent path.
161 static bool removeUnneededCalls(PathPieces &pieces, BugReport *R,
162 LocationContextMap &LCM) {
163 bool containsSomethingInteresting = false;
164 const unsigned N = pieces.size();
166 for (unsigned i = 0 ; i < N ; ++i) {
167 // Remove the front piece from the path. If it is still something we
168 // want to keep once we are done, we will push it back on the end.
169 auto piece = std::move(pieces.front());
172 switch (piece->getKind()) {
173 case PathDiagnosticPiece::Call: {
174 auto &call = cast<PathDiagnosticCallPiece>(*piece);
175 // Check if the location context is interesting.
176 assert(LCM.count(&call.path));
177 if (R->isInteresting(LCM[&call.path])) {
178 containsSomethingInteresting = true;
182 if (!removeUnneededCalls(call.path, R, LCM))
185 containsSomethingInteresting = true;
188 case PathDiagnosticPiece::Macro: {
189 auto ¯o = cast<PathDiagnosticMacroPiece>(*piece);
190 if (!removeUnneededCalls(macro.subPieces, R, LCM))
192 containsSomethingInteresting = true;
195 case PathDiagnosticPiece::Event: {
196 auto &event = cast<PathDiagnosticEventPiece>(*piece);
198 // We never throw away an event, but we do throw it away wholesale
199 // as part of a path if we throw the entire path away.
200 containsSomethingInteresting |= !event.isPrunable();
203 case PathDiagnosticPiece::ControlFlow:
206 case PathDiagnosticPiece::Note:
210 pieces.push_back(std::move(piece));
213 return containsSomethingInteresting;
216 /// Returns true if the given decl has been implicitly given a body, either by
217 /// the analyzer or by the compiler proper.
218 static bool hasImplicitBody(const Decl *D) {
220 return D->isImplicit() || !D->hasBody();
223 /// Recursively scan through a path and make sure that all call pieces have
226 adjustCallLocations(PathPieces &Pieces,
227 PathDiagnosticLocation *LastCallLocation = nullptr) {
228 for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E; ++I) {
229 PathDiagnosticCallPiece *Call = dyn_cast<PathDiagnosticCallPiece>(I->get());
232 assert((*I)->getLocation().asLocation().isValid());
236 if (LastCallLocation) {
237 bool CallerIsImplicit = hasImplicitBody(Call->getCaller());
238 if (CallerIsImplicit || !Call->callEnter.asLocation().isValid())
239 Call->callEnter = *LastCallLocation;
240 if (CallerIsImplicit || !Call->callReturn.asLocation().isValid())
241 Call->callReturn = *LastCallLocation;
244 // Recursively clean out the subclass. Keep this call around if
245 // it contains any informative diagnostics.
246 PathDiagnosticLocation *ThisCallLocation;
247 if (Call->callEnterWithin.asLocation().isValid() &&
248 !hasImplicitBody(Call->getCallee()))
249 ThisCallLocation = &Call->callEnterWithin;
251 ThisCallLocation = &Call->callEnter;
253 assert(ThisCallLocation && "Outermost call has an invalid location");
254 adjustCallLocations(Call->path, ThisCallLocation);
258 /// Remove edges in and out of C++ default initializer expressions. These are
259 /// for fields that have in-class initializers, as opposed to being initialized
260 /// explicitly in a constructor or braced list.
261 static void removeEdgesToDefaultInitializers(PathPieces &Pieces) {
262 for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E;) {
263 if (auto *C = dyn_cast<PathDiagnosticCallPiece>(I->get()))
264 removeEdgesToDefaultInitializers(C->path);
266 if (auto *M = dyn_cast<PathDiagnosticMacroPiece>(I->get()))
267 removeEdgesToDefaultInitializers(M->subPieces);
269 if (auto *CF = dyn_cast<PathDiagnosticControlFlowPiece>(I->get())) {
270 const Stmt *Start = CF->getStartLocation().asStmt();
271 const Stmt *End = CF->getEndLocation().asStmt();
272 if (Start && isa<CXXDefaultInitExpr>(Start)) {
275 } else if (End && isa<CXXDefaultInitExpr>(End)) {
276 PathPieces::iterator Next = std::next(I);
279 dyn_cast<PathDiagnosticControlFlowPiece>(Next->get())) {
280 NextCF->setStartLocation(CF->getStartLocation());
292 /// Remove all pieces with invalid locations as these cannot be serialized.
293 /// We might have pieces with invalid locations as a result of inlining Body
294 /// Farm generated functions.
295 static void removePiecesWithInvalidLocations(PathPieces &Pieces) {
296 for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E;) {
297 if (auto *C = dyn_cast<PathDiagnosticCallPiece>(I->get()))
298 removePiecesWithInvalidLocations(C->path);
300 if (auto *M = dyn_cast<PathDiagnosticMacroPiece>(I->get()))
301 removePiecesWithInvalidLocations(M->subPieces);
303 if (!(*I)->getLocation().isValid() ||
304 !(*I)->getLocation().asLocation().isValid()) {
312 //===----------------------------------------------------------------------===//
313 // PathDiagnosticBuilder and its associated routines and helper objects.
314 //===----------------------------------------------------------------------===//
317 class NodeMapClosure : public BugReport::NodeResolver {
318 InterExplodedGraphMap &M;
320 NodeMapClosure(InterExplodedGraphMap &m) : M(m) {}
322 const ExplodedNode *getOriginalNode(const ExplodedNode *N) override {
327 class PathDiagnosticBuilder : public BugReporterContext {
329 PathDiagnosticConsumer *PDC;
332 const LocationContext *LC;
334 PathDiagnosticBuilder(GRBugReporter &br,
335 BugReport *r, InterExplodedGraphMap &Backmap,
336 PathDiagnosticConsumer *pdc)
337 : BugReporterContext(br),
338 R(r), PDC(pdc), NMC(Backmap), LC(r->getErrorNode()->getLocationContext())
341 PathDiagnosticLocation ExecutionContinues(const ExplodedNode *N);
343 PathDiagnosticLocation ExecutionContinues(llvm::raw_string_ostream &os,
344 const ExplodedNode *N);
346 BugReport *getBugReport() { return R; }
348 Decl const &getCodeDecl() { return R->getErrorNode()->getCodeDecl(); }
350 ParentMap& getParentMap() { return LC->getParentMap(); }
352 const Stmt *getParent(const Stmt *S) {
353 return getParentMap().getParent(S);
356 NodeMapClosure& getNodeResolver() override { return NMC; }
358 PathDiagnosticLocation getEnclosingStmtLocation(const Stmt *S);
360 PathDiagnosticConsumer::PathGenerationScheme getGenerationScheme() const {
361 return PDC ? PDC->getGenerationScheme() : PathDiagnosticConsumer::Extensive;
364 bool supportsLogicalOpControlFlow() const {
365 return PDC ? PDC->supportsLogicalOpControlFlow() : true;
368 } // end anonymous namespace
370 PathDiagnosticLocation
371 PathDiagnosticBuilder::ExecutionContinues(const ExplodedNode *N) {
372 if (const Stmt *S = PathDiagnosticLocation::getNextStmt(N))
373 return PathDiagnosticLocation(S, getSourceManager(), LC);
375 return PathDiagnosticLocation::createDeclEnd(N->getLocationContext(),
379 PathDiagnosticLocation
380 PathDiagnosticBuilder::ExecutionContinues(llvm::raw_string_ostream &os,
381 const ExplodedNode *N) {
383 // Slow, but probably doesn't matter.
384 if (os.str().empty())
387 const PathDiagnosticLocation &Loc = ExecutionContinues(N);
390 os << "Execution continues on line "
391 << getSourceManager().getExpansionLineNumber(Loc.asLocation())
394 os << "Execution jumps to the end of the ";
395 const Decl *D = N->getLocationContext()->getDecl();
396 if (isa<ObjCMethodDecl>(D))
398 else if (isa<FunctionDecl>(D))
401 assert(isa<BlockDecl>(D));
402 os << "anonymous block";
410 static const Stmt *getEnclosingParent(const Stmt *S, const ParentMap &PM) {
411 if (isa<Expr>(S) && PM.isConsumedExpr(cast<Expr>(S)))
412 return PM.getParentIgnoreParens(S);
414 const Stmt *Parent = PM.getParentIgnoreParens(S);
418 switch (Parent->getStmtClass()) {
419 case Stmt::ForStmtClass:
420 case Stmt::DoStmtClass:
421 case Stmt::WhileStmtClass:
422 case Stmt::ObjCForCollectionStmtClass:
423 case Stmt::CXXForRangeStmtClass:
432 static PathDiagnosticLocation
433 getEnclosingStmtLocation(const Stmt *S, SourceManager &SMgr, const ParentMap &P,
434 const LocationContext *LC, bool allowNestedContexts) {
436 return PathDiagnosticLocation();
438 while (const Stmt *Parent = getEnclosingParent(S, P)) {
439 switch (Parent->getStmtClass()) {
440 case Stmt::BinaryOperatorClass: {
441 const BinaryOperator *B = cast<BinaryOperator>(Parent);
442 if (B->isLogicalOp())
443 return PathDiagnosticLocation(allowNestedContexts ? B : S, SMgr, LC);
446 case Stmt::CompoundStmtClass:
447 case Stmt::StmtExprClass:
448 return PathDiagnosticLocation(S, SMgr, LC);
449 case Stmt::ChooseExprClass:
450 // Similar to '?' if we are referring to condition, just have the edge
451 // point to the entire choose expression.
452 if (allowNestedContexts || cast<ChooseExpr>(Parent)->getCond() == S)
453 return PathDiagnosticLocation(Parent, SMgr, LC);
455 return PathDiagnosticLocation(S, SMgr, LC);
456 case Stmt::BinaryConditionalOperatorClass:
457 case Stmt::ConditionalOperatorClass:
458 // For '?', if we are referring to condition, just have the edge point
459 // to the entire '?' expression.
460 if (allowNestedContexts ||
461 cast<AbstractConditionalOperator>(Parent)->getCond() == S)
462 return PathDiagnosticLocation(Parent, SMgr, LC);
464 return PathDiagnosticLocation(S, SMgr, LC);
465 case Stmt::CXXForRangeStmtClass:
466 if (cast<CXXForRangeStmt>(Parent)->getBody() == S)
467 return PathDiagnosticLocation(S, SMgr, LC);
469 case Stmt::DoStmtClass:
470 return PathDiagnosticLocation(S, SMgr, LC);
471 case Stmt::ForStmtClass:
472 if (cast<ForStmt>(Parent)->getBody() == S)
473 return PathDiagnosticLocation(S, SMgr, LC);
475 case Stmt::IfStmtClass:
476 if (cast<IfStmt>(Parent)->getCond() != S)
477 return PathDiagnosticLocation(S, SMgr, LC);
479 case Stmt::ObjCForCollectionStmtClass:
480 if (cast<ObjCForCollectionStmt>(Parent)->getBody() == S)
481 return PathDiagnosticLocation(S, SMgr, LC);
483 case Stmt::WhileStmtClass:
484 if (cast<WhileStmt>(Parent)->getCond() != S)
485 return PathDiagnosticLocation(S, SMgr, LC);
494 assert(S && "Cannot have null Stmt for PathDiagnosticLocation");
496 return PathDiagnosticLocation(S, SMgr, LC);
499 PathDiagnosticLocation
500 PathDiagnosticBuilder::getEnclosingStmtLocation(const Stmt *S) {
501 assert(S && "Null Stmt passed to getEnclosingStmtLocation");
502 return ::getEnclosingStmtLocation(S, getSourceManager(), getParentMap(), LC,
503 /*allowNestedContexts=*/false);
506 //===----------------------------------------------------------------------===//
507 // "Visitors only" path diagnostic generation algorithm.
508 //===----------------------------------------------------------------------===//
509 static bool GenerateVisitorsOnlyPathDiagnostic(
510 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N,
511 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) {
512 // All path generation skips the very first node (the error node).
513 // This is because there is special handling for the end-of-path note.
514 N = N->getFirstPred();
518 BugReport *R = PDB.getBugReport();
519 while (const ExplodedNode *Pred = N->getFirstPred()) {
520 for (auto &V : visitors)
521 // Visit all the node pairs, but throw the path pieces away.
522 V->VisitNode(N, Pred, PDB, *R);
530 //===----------------------------------------------------------------------===//
531 // "Minimal" path diagnostic generation algorithm.
532 //===----------------------------------------------------------------------===//
533 typedef std::pair<PathDiagnosticCallPiece*, const ExplodedNode*> StackDiagPair;
534 typedef SmallVector<StackDiagPair, 6> StackDiagVector;
536 static void updateStackPiecesWithMessage(PathDiagnosticPiece &P,
537 StackDiagVector &CallStack) {
538 // If the piece contains a special message, add it to all the call
539 // pieces on the active stack.
540 if (PathDiagnosticEventPiece *ep = dyn_cast<PathDiagnosticEventPiece>(&P)) {
542 if (ep->hasCallStackHint())
543 for (StackDiagVector::iterator I = CallStack.begin(),
544 E = CallStack.end(); I != E; ++I) {
545 PathDiagnosticCallPiece *CP = I->first;
546 const ExplodedNode *N = I->second;
547 std::string stackMsg = ep->getCallStackMessage(N);
549 // The last message on the path to final bug is the most important
550 // one. Since we traverse the path backwards, do not add the message
551 // if one has been previously added.
552 if (!CP->hasCallStackMessage())
553 CP->setCallStackMessage(stackMsg);
558 static void CompactPathDiagnostic(PathPieces &path, const SourceManager& SM);
560 static bool GenerateMinimalPathDiagnostic(
561 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N,
562 LocationContextMap &LCM,
563 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) {
565 SourceManager& SMgr = PDB.getSourceManager();
566 const LocationContext *LC = PDB.LC;
567 const ExplodedNode *NextNode = N->pred_empty()
568 ? nullptr : *(N->pred_begin());
570 StackDiagVector CallStack;
574 PDB.LC = N->getLocationContext();
575 NextNode = N->getFirstPred();
577 ProgramPoint P = N->getLocation();
580 if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
581 auto C = PathDiagnosticCallPiece::construct(N, *CE, SMgr);
582 // Record the mapping from call piece to LocationContext.
583 LCM[&C->path] = CE->getCalleeContext();
585 PD.getActivePath().push_front(std::move(C));
586 PD.pushActivePath(&P->path);
587 CallStack.push_back(StackDiagPair(P, N));
591 if (Optional<CallEnter> CE = P.getAs<CallEnter>()) {
592 // Flush all locations, and pop the active path.
593 bool VisitedEntireCall = PD.isWithinCall();
596 // Either we just added a bunch of stuff to the top-level path, or
597 // we have a previous CallExitEnd. If the former, it means that the
598 // path terminated within a function call. We must then take the
599 // current contents of the active path and place it within
600 // a new PathDiagnosticCallPiece.
601 PathDiagnosticCallPiece *C;
602 if (VisitedEntireCall) {
603 C = cast<PathDiagnosticCallPiece>(PD.getActivePath().front().get());
605 const Decl *Caller = CE->getLocationContext()->getDecl();
606 C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller);
607 // Record the mapping from call piece to LocationContext.
608 LCM[&C->path] = CE->getCalleeContext();
611 C->setCallee(*CE, SMgr);
612 if (!CallStack.empty()) {
613 assert(CallStack.back().first == C);
614 CallStack.pop_back();
619 if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) {
620 const CFGBlock *Src = BE->getSrc();
621 const CFGBlock *Dst = BE->getDst();
622 const Stmt *T = Src->getTerminator();
627 PathDiagnosticLocation Start =
628 PathDiagnosticLocation::createBegin(T, SMgr,
629 N->getLocationContext());
631 switch (T->getStmtClass()) {
635 case Stmt::GotoStmtClass:
636 case Stmt::IndirectGotoStmtClass: {
637 const Stmt *S = PathDiagnosticLocation::getNextStmt(N);
643 llvm::raw_string_ostream os(sbuf);
644 const PathDiagnosticLocation &End = PDB.getEnclosingStmtLocation(S);
646 os << "Control jumps to line "
647 << End.asLocation().getExpansionLineNumber();
648 PD.getActivePath().push_front(
649 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
654 case Stmt::SwitchStmtClass: {
655 // Figure out what case arm we took.
657 llvm::raw_string_ostream os(sbuf);
659 if (const Stmt *S = Dst->getLabel()) {
660 PathDiagnosticLocation End(S, SMgr, LC);
662 switch (S->getStmtClass()) {
664 os << "No cases match in the switch statement. "
665 "Control jumps to line "
666 << End.asLocation().getExpansionLineNumber();
668 case Stmt::DefaultStmtClass:
669 os << "Control jumps to the 'default' case at line "
670 << End.asLocation().getExpansionLineNumber();
673 case Stmt::CaseStmtClass: {
674 os << "Control jumps to 'case ";
675 const CaseStmt *Case = cast<CaseStmt>(S);
676 const Expr *LHS = Case->getLHS()->IgnoreParenCasts();
678 // Determine if it is an enum.
679 bool GetRawInt = true;
681 if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(LHS)) {
682 // FIXME: Maybe this should be an assertion. Are there cases
683 // were it is not an EnumConstantDecl?
684 const EnumConstantDecl *D =
685 dyn_cast<EnumConstantDecl>(DR->getDecl());
694 os << LHS->EvaluateKnownConstInt(PDB.getASTContext());
697 << End.asLocation().getExpansionLineNumber();
701 PD.getActivePath().push_front(
702 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
706 os << "'Default' branch taken. ";
707 const PathDiagnosticLocation &End = PDB.ExecutionContinues(os, N);
708 PD.getActivePath().push_front(
709 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
716 case Stmt::BreakStmtClass:
717 case Stmt::ContinueStmtClass: {
719 llvm::raw_string_ostream os(sbuf);
720 PathDiagnosticLocation End = PDB.ExecutionContinues(os, N);
721 PD.getActivePath().push_front(
722 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
727 // Determine control-flow for ternary '?'.
728 case Stmt::BinaryConditionalOperatorClass:
729 case Stmt::ConditionalOperatorClass: {
731 llvm::raw_string_ostream os(sbuf);
732 os << "'?' condition is ";
734 if (*(Src->succ_begin()+1) == Dst)
739 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
741 if (const Stmt *S = End.asStmt())
742 End = PDB.getEnclosingStmtLocation(S);
744 PD.getActivePath().push_front(
745 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
750 // Determine control-flow for short-circuited '&&' and '||'.
751 case Stmt::BinaryOperatorClass: {
752 if (!PDB.supportsLogicalOpControlFlow())
755 const BinaryOperator *B = cast<BinaryOperator>(T);
757 llvm::raw_string_ostream os(sbuf);
758 os << "Left side of '";
760 if (B->getOpcode() == BO_LAnd) {
761 os << "&&" << "' is ";
763 if (*(Src->succ_begin()+1) == Dst) {
765 PathDiagnosticLocation End(B->getLHS(), SMgr, LC);
766 PathDiagnosticLocation Start =
767 PathDiagnosticLocation::createOperatorLoc(B, SMgr);
768 PD.getActivePath().push_front(
769 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
774 PathDiagnosticLocation Start(B->getLHS(), SMgr, LC);
775 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
776 PD.getActivePath().push_front(
777 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
782 assert(B->getOpcode() == BO_LOr);
783 os << "||" << "' is ";
785 if (*(Src->succ_begin()+1) == Dst) {
787 PathDiagnosticLocation Start(B->getLHS(), SMgr, LC);
788 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
789 PD.getActivePath().push_front(
790 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
795 PathDiagnosticLocation End(B->getLHS(), SMgr, LC);
796 PathDiagnosticLocation Start =
797 PathDiagnosticLocation::createOperatorLoc(B, SMgr);
798 PD.getActivePath().push_front(
799 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
807 case Stmt::DoStmtClass: {
808 if (*(Src->succ_begin()) == Dst) {
810 llvm::raw_string_ostream os(sbuf);
812 os << "Loop condition is true. ";
813 PathDiagnosticLocation End = PDB.ExecutionContinues(os, N);
815 if (const Stmt *S = End.asStmt())
816 End = PDB.getEnclosingStmtLocation(S);
818 PD.getActivePath().push_front(
819 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
823 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
825 if (const Stmt *S = End.asStmt())
826 End = PDB.getEnclosingStmtLocation(S);
828 PD.getActivePath().push_front(
829 std::make_shared<PathDiagnosticControlFlowPiece>(
830 Start, End, "Loop condition is false. Exiting loop"));
836 case Stmt::WhileStmtClass:
837 case Stmt::ForStmtClass: {
838 if (*(Src->succ_begin()+1) == Dst) {
840 llvm::raw_string_ostream os(sbuf);
842 os << "Loop condition is false. ";
843 PathDiagnosticLocation End = PDB.ExecutionContinues(os, N);
844 if (const Stmt *S = End.asStmt())
845 End = PDB.getEnclosingStmtLocation(S);
847 PD.getActivePath().push_front(
848 std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
852 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
853 if (const Stmt *S = End.asStmt())
854 End = PDB.getEnclosingStmtLocation(S);
856 PD.getActivePath().push_front(
857 std::make_shared<PathDiagnosticControlFlowPiece>(
858 Start, End, "Loop condition is true. Entering loop body"));
864 case Stmt::IfStmtClass: {
865 PathDiagnosticLocation End = PDB.ExecutionContinues(N);
867 if (const Stmt *S = End.asStmt())
868 End = PDB.getEnclosingStmtLocation(S);
870 if (*(Src->succ_begin()+1) == Dst)
871 PD.getActivePath().push_front(
872 std::make_shared<PathDiagnosticControlFlowPiece>(
873 Start, End, "Taking false branch"));
875 PD.getActivePath().push_front(
876 std::make_shared<PathDiagnosticControlFlowPiece>(
877 Start, End, "Taking true branch"));
886 // Add diagnostic pieces from custom visitors.
887 BugReport *R = PDB.getBugReport();
888 for (auto &V : visitors) {
889 if (auto p = V->VisitNode(N, NextNode, PDB, *R)) {
890 updateStackPiecesWithMessage(*p, CallStack);
891 PD.getActivePath().push_front(std::move(p));
897 if (!PDB.getBugReport()->isValid())
900 // After constructing the full PathDiagnostic, do a pass over it to compact
901 // PathDiagnosticPieces that occur within a macro.
902 CompactPathDiagnostic(PD.getMutablePieces(), PDB.getSourceManager());
906 //===----------------------------------------------------------------------===//
907 // "Extensive" PathDiagnostic generation.
908 //===----------------------------------------------------------------------===//
910 static bool IsControlFlowExpr(const Stmt *S) {
911 const Expr *E = dyn_cast<Expr>(S);
916 E = E->IgnoreParenCasts();
918 if (isa<AbstractConditionalOperator>(E))
921 if (const BinaryOperator *B = dyn_cast<BinaryOperator>(E))
922 if (B->isLogicalOp())
929 class ContextLocation : public PathDiagnosticLocation {
932 ContextLocation(const PathDiagnosticLocation &L, bool isdead = false)
933 : PathDiagnosticLocation(L), IsDead(isdead) {}
935 void markDead() { IsDead = true; }
936 bool isDead() const { return IsDead; }
939 static PathDiagnosticLocation cleanUpLocation(PathDiagnosticLocation L,
940 const LocationContext *LC,
941 bool firstCharOnly = false) {
942 if (const Stmt *S = L.asStmt()) {
943 const Stmt *Original = S;
945 // Adjust the location for some expressions that are best referenced
946 // by one of their subexpressions.
947 switch (S->getStmtClass()) {
950 case Stmt::ParenExprClass:
951 case Stmt::GenericSelectionExprClass:
952 S = cast<Expr>(S)->IgnoreParens();
953 firstCharOnly = true;
955 case Stmt::BinaryConditionalOperatorClass:
956 case Stmt::ConditionalOperatorClass:
957 S = cast<AbstractConditionalOperator>(S)->getCond();
958 firstCharOnly = true;
960 case Stmt::ChooseExprClass:
961 S = cast<ChooseExpr>(S)->getCond();
962 firstCharOnly = true;
964 case Stmt::BinaryOperatorClass:
965 S = cast<BinaryOperator>(S)->getLHS();
966 firstCharOnly = true;
974 L = PathDiagnosticLocation(S, L.getManager(), LC);
978 L = PathDiagnosticLocation::createSingleLocation(L);
984 std::vector<ContextLocation> CLocs;
985 typedef std::vector<ContextLocation>::iterator iterator;
987 PathDiagnosticBuilder &PDB;
988 PathDiagnosticLocation PrevLoc;
990 bool IsConsumedExpr(const PathDiagnosticLocation &L);
992 bool containsLocation(const PathDiagnosticLocation &Container,
993 const PathDiagnosticLocation &Containee);
995 PathDiagnosticLocation getContextLocation(const PathDiagnosticLocation &L);
1000 if (!CLocs.back().isDead() && CLocs.back().asLocation().isFileID()) {
1001 // For contexts, we only one the first character as the range.
1002 rawAddEdge(cleanUpLocation(CLocs.back(), PDB.LC, true));
1008 EdgeBuilder(PathDiagnostic &pd, PathDiagnosticBuilder &pdb)
1009 : PD(pd), PDB(pdb) {
1011 // If the PathDiagnostic already has pieces, add the enclosing statement
1012 // of the first piece as a context as well.
1013 if (!PD.path.empty()) {
1014 PrevLoc = (*PD.path.begin())->getLocation();
1016 if (const Stmt *S = PrevLoc.asStmt())
1017 addExtendedContext(PDB.getEnclosingStmtLocation(S).asStmt());
1022 while (!CLocs.empty()) popLocation();
1024 // Finally, add an initial edge from the start location of the first
1025 // statement (if it doesn't already exist).
1026 PathDiagnosticLocation L = PathDiagnosticLocation::createDeclBegin(
1028 PDB.getSourceManager());
1033 void flushLocations() {
1034 while (!CLocs.empty())
1036 PrevLoc = PathDiagnosticLocation();
1039 void addEdge(PathDiagnosticLocation NewLoc, bool alwaysAdd = false,
1040 bool IsPostJump = false);
1042 void rawAddEdge(PathDiagnosticLocation NewLoc);
1044 void addContext(const Stmt *S);
1045 void addContext(const PathDiagnosticLocation &L);
1046 void addExtendedContext(const Stmt *S);
1048 } // end anonymous namespace
1051 PathDiagnosticLocation
1052 EdgeBuilder::getContextLocation(const PathDiagnosticLocation &L) {
1053 if (const Stmt *S = L.asStmt()) {
1054 if (IsControlFlowExpr(S))
1057 return PDB.getEnclosingStmtLocation(S);
1063 bool EdgeBuilder::containsLocation(const PathDiagnosticLocation &Container,
1064 const PathDiagnosticLocation &Containee) {
1066 if (Container == Containee)
1069 if (Container.asDecl())
1072 if (const Stmt *S = Containee.asStmt())
1073 if (const Stmt *ContainerS = Container.asStmt()) {
1075 if (S == ContainerS)
1077 S = PDB.getParent(S);
1082 // Less accurate: compare using source ranges.
1083 SourceRange ContainerR = Container.asRange();
1084 SourceRange ContaineeR = Containee.asRange();
1086 SourceManager &SM = PDB.getSourceManager();
1087 SourceLocation ContainerRBeg = SM.getExpansionLoc(ContainerR.getBegin());
1088 SourceLocation ContainerREnd = SM.getExpansionLoc(ContainerR.getEnd());
1089 SourceLocation ContaineeRBeg = SM.getExpansionLoc(ContaineeR.getBegin());
1090 SourceLocation ContaineeREnd = SM.getExpansionLoc(ContaineeR.getEnd());
1092 unsigned ContainerBegLine = SM.getExpansionLineNumber(ContainerRBeg);
1093 unsigned ContainerEndLine = SM.getExpansionLineNumber(ContainerREnd);
1094 unsigned ContaineeBegLine = SM.getExpansionLineNumber(ContaineeRBeg);
1095 unsigned ContaineeEndLine = SM.getExpansionLineNumber(ContaineeREnd);
1097 assert(ContainerBegLine <= ContainerEndLine);
1098 assert(ContaineeBegLine <= ContaineeEndLine);
1100 return (ContainerBegLine <= ContaineeBegLine &&
1101 ContainerEndLine >= ContaineeEndLine &&
1102 (ContainerBegLine != ContaineeBegLine ||
1103 SM.getExpansionColumnNumber(ContainerRBeg) <=
1104 SM.getExpansionColumnNumber(ContaineeRBeg)) &&
1105 (ContainerEndLine != ContaineeEndLine ||
1106 SM.getExpansionColumnNumber(ContainerREnd) >=
1107 SM.getExpansionColumnNumber(ContaineeREnd)));
1110 void EdgeBuilder::rawAddEdge(PathDiagnosticLocation NewLoc) {
1111 if (!PrevLoc.isValid()) {
1116 const PathDiagnosticLocation &NewLocClean = cleanUpLocation(NewLoc, PDB.LC);
1117 const PathDiagnosticLocation &PrevLocClean = cleanUpLocation(PrevLoc, PDB.LC);
1119 if (PrevLocClean.asLocation().isInvalid()) {
1124 if (NewLocClean.asLocation() == PrevLocClean.asLocation())
1127 // FIXME: Ignore intra-macro edges for now.
1128 if (NewLocClean.asLocation().getExpansionLoc() ==
1129 PrevLocClean.asLocation().getExpansionLoc())
1132 PD.getActivePath().push_front(
1133 std::make_shared<PathDiagnosticControlFlowPiece>(NewLocClean,
1138 void EdgeBuilder::addEdge(PathDiagnosticLocation NewLoc, bool alwaysAdd,
1141 if (!alwaysAdd && NewLoc.asLocation().isMacroID())
1144 const PathDiagnosticLocation &CLoc = getContextLocation(NewLoc);
1146 while (!CLocs.empty()) {
1147 ContextLocation &TopContextLoc = CLocs.back();
1149 // Is the top location context the same as the one for the new location?
1150 if (TopContextLoc == CLoc) {
1152 if (IsConsumedExpr(TopContextLoc))
1153 TopContextLoc.markDead();
1159 TopContextLoc.markDead();
1163 if (containsLocation(TopContextLoc, CLoc)) {
1167 if (IsConsumedExpr(CLoc)) {
1168 CLocs.push_back(ContextLocation(CLoc, /*IsDead=*/true));
1173 CLocs.push_back(ContextLocation(CLoc, /*IsDead=*/IsPostJump));
1177 // Context does not contain the location. Flush it.
1181 // If we reach here, there is no enclosing context. Just add the edge.
1185 bool EdgeBuilder::IsConsumedExpr(const PathDiagnosticLocation &L) {
1186 if (const Expr *X = dyn_cast_or_null<Expr>(L.asStmt()))
1187 return PDB.getParentMap().isConsumedExpr(X) && !IsControlFlowExpr(X);
1192 void EdgeBuilder::addExtendedContext(const Stmt *S) {
1196 const Stmt *Parent = PDB.getParent(S);
1198 if (isa<CompoundStmt>(Parent))
1199 Parent = PDB.getParent(Parent);
1205 switch (Parent->getStmtClass()) {
1206 case Stmt::DoStmtClass:
1207 case Stmt::ObjCAtSynchronizedStmtClass:
1217 void EdgeBuilder::addContext(const Stmt *S) {
1221 PathDiagnosticLocation L(S, PDB.getSourceManager(), PDB.LC);
1225 void EdgeBuilder::addContext(const PathDiagnosticLocation &L) {
1226 while (!CLocs.empty()) {
1227 const PathDiagnosticLocation &TopContextLoc = CLocs.back();
1229 // Is the top location context the same as the one for the new location?
1230 if (TopContextLoc == L)
1233 if (containsLocation(TopContextLoc, L)) {
1238 // Context does not contain the location. Flush it.
1245 // Cone-of-influence: support the reverse propagation of "interesting" symbols
1246 // and values by tracing interesting calculations backwards through evaluated
1247 // expressions along a path. This is probably overly complicated, but the idea
1248 // is that if an expression computed an "interesting" value, the child
1249 // expressions are are also likely to be "interesting" as well (which then
1250 // propagates to the values they in turn compute). This reverse propagation
1251 // is needed to track interesting correlations across function call boundaries,
1252 // where formal arguments bind to actual arguments, etc. This is also needed
1253 // because the constraint solver sometimes simplifies certain symbolic values
1254 // into constants when appropriate, and this complicates reasoning about
1255 // interesting values.
1256 typedef llvm::DenseSet<const Expr *> InterestingExprs;
1258 static void reversePropagateIntererstingSymbols(BugReport &R,
1259 InterestingExprs &IE,
1260 const ProgramState *State,
1262 const LocationContext *LCtx) {
1263 SVal V = State->getSVal(Ex, LCtx);
1264 if (!(R.isInteresting(V) || IE.count(Ex)))
1267 switch (Ex->getStmtClass()) {
1269 if (!isa<CastExpr>(Ex))
1272 case Stmt::BinaryOperatorClass:
1273 case Stmt::UnaryOperatorClass: {
1274 for (const Stmt *SubStmt : Ex->children()) {
1275 if (const Expr *child = dyn_cast_or_null<Expr>(SubStmt)) {
1277 SVal ChildV = State->getSVal(child, LCtx);
1278 R.markInteresting(ChildV);
1285 R.markInteresting(V);
1288 static void reversePropagateInterestingSymbols(BugReport &R,
1289 InterestingExprs &IE,
1290 const ProgramState *State,
1291 const LocationContext *CalleeCtx,
1292 const LocationContext *CallerCtx)
1294 // FIXME: Handle non-CallExpr-based CallEvents.
1295 const StackFrameContext *Callee = CalleeCtx->getCurrentStackFrame();
1296 const Stmt *CallSite = Callee->getCallSite();
1297 if (const CallExpr *CE = dyn_cast_or_null<CallExpr>(CallSite)) {
1298 if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(CalleeCtx->getDecl())) {
1299 FunctionDecl::param_const_iterator PI = FD->param_begin(),
1300 PE = FD->param_end();
1301 CallExpr::const_arg_iterator AI = CE->arg_begin(), AE = CE->arg_end();
1302 for (; AI != AE && PI != PE; ++AI, ++PI) {
1303 if (const Expr *ArgE = *AI) {
1304 if (const ParmVarDecl *PD = *PI) {
1305 Loc LV = State->getLValue(PD, CalleeCtx);
1306 if (R.isInteresting(LV) || R.isInteresting(State->getRawSVal(LV)))
1315 //===----------------------------------------------------------------------===//
1316 // Functions for determining if a loop was executed 0 times.
1317 //===----------------------------------------------------------------------===//
1319 static bool isLoop(const Stmt *Term) {
1320 switch (Term->getStmtClass()) {
1321 case Stmt::ForStmtClass:
1322 case Stmt::WhileStmtClass:
1323 case Stmt::ObjCForCollectionStmtClass:
1324 case Stmt::CXXForRangeStmtClass:
1327 // Note that we intentionally do not include do..while here.
1332 static bool isJumpToFalseBranch(const BlockEdge *BE) {
1333 const CFGBlock *Src = BE->getSrc();
1334 assert(Src->succ_size() == 2);
1335 return (*(Src->succ_begin()+1) == BE->getDst());
1338 /// Return true if the terminator is a loop and the destination is the
1340 static bool isLoopJumpPastBody(const Stmt *Term, const BlockEdge *BE) {
1344 // Did we take the false branch?
1345 return isJumpToFalseBranch(BE);
1348 static bool isContainedByStmt(ParentMap &PM, const Stmt *S, const Stmt *SubS) {
1352 SubS = PM.getParent(SubS);
1357 static const Stmt *getStmtBeforeCond(ParentMap &PM, const Stmt *Term,
1358 const ExplodedNode *N) {
1360 Optional<StmtPoint> SP = N->getLocation().getAs<StmtPoint>();
1362 const Stmt *S = SP->getStmt();
1363 if (!isContainedByStmt(PM, Term, S))
1366 N = N->getFirstPred();
1371 static bool isInLoopBody(ParentMap &PM, const Stmt *S, const Stmt *Term) {
1372 const Stmt *LoopBody = nullptr;
1373 switch (Term->getStmtClass()) {
1374 case Stmt::CXXForRangeStmtClass: {
1375 const CXXForRangeStmt *FR = cast<CXXForRangeStmt>(Term);
1376 if (isContainedByStmt(PM, FR->getInc(), S))
1378 if (isContainedByStmt(PM, FR->getLoopVarStmt(), S))
1380 LoopBody = FR->getBody();
1383 case Stmt::ForStmtClass: {
1384 const ForStmt *FS = cast<ForStmt>(Term);
1385 if (isContainedByStmt(PM, FS->getInc(), S))
1387 LoopBody = FS->getBody();
1390 case Stmt::ObjCForCollectionStmtClass: {
1391 const ObjCForCollectionStmt *FC = cast<ObjCForCollectionStmt>(Term);
1392 LoopBody = FC->getBody();
1395 case Stmt::WhileStmtClass:
1396 LoopBody = cast<WhileStmt>(Term)->getBody();
1401 return isContainedByStmt(PM, LoopBody, S);
1404 //===----------------------------------------------------------------------===//
1405 // Top-level logic for generating extensive path diagnostics.
1406 //===----------------------------------------------------------------------===//
1408 static bool GenerateExtensivePathDiagnostic(
1409 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N,
1410 LocationContextMap &LCM,
1411 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) {
1412 EdgeBuilder EB(PD, PDB);
1413 const SourceManager& SM = PDB.getSourceManager();
1414 StackDiagVector CallStack;
1415 InterestingExprs IE;
1417 const ExplodedNode *NextNode = N->pred_empty() ? nullptr : *(N->pred_begin());
1420 NextNode = N->getFirstPred();
1421 ProgramPoint P = N->getLocation();
1424 if (Optional<PostStmt> PS = P.getAs<PostStmt>()) {
1425 if (const Expr *Ex = PS->getStmtAs<Expr>())
1426 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1427 N->getState().get(), Ex,
1428 N->getLocationContext());
1431 if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
1432 const Stmt *S = CE->getCalleeContext()->getCallSite();
1433 if (const Expr *Ex = dyn_cast_or_null<Expr>(S)) {
1434 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1435 N->getState().get(), Ex,
1436 N->getLocationContext());
1439 auto C = PathDiagnosticCallPiece::construct(N, *CE, SM);
1440 LCM[&C->path] = CE->getCalleeContext();
1442 EB.addEdge(C->callReturn, /*AlwaysAdd=*/true, /*IsPostJump=*/true);
1443 EB.flushLocations();
1446 PD.getActivePath().push_front(std::move(C));
1447 PD.pushActivePath(&P->path);
1448 CallStack.push_back(StackDiagPair(P, N));
1452 // Pop the call hierarchy if we are done walking the contents
1453 // of a function call.
1454 if (Optional<CallEnter> CE = P.getAs<CallEnter>()) {
1455 // Add an edge to the start of the function.
1456 const Decl *D = CE->getCalleeContext()->getDecl();
1457 PathDiagnosticLocation pos =
1458 PathDiagnosticLocation::createBegin(D, SM);
1461 // Flush all locations, and pop the active path.
1462 bool VisitedEntireCall = PD.isWithinCall();
1463 EB.flushLocations();
1465 PDB.LC = N->getLocationContext();
1467 // Either we just added a bunch of stuff to the top-level path, or
1468 // we have a previous CallExitEnd. If the former, it means that the
1469 // path terminated within a function call. We must then take the
1470 // current contents of the active path and place it within
1471 // a new PathDiagnosticCallPiece.
1472 PathDiagnosticCallPiece *C;
1473 if (VisitedEntireCall) {
1474 C = cast<PathDiagnosticCallPiece>(PD.getActivePath().front().get());
1476 const Decl *Caller = CE->getLocationContext()->getDecl();
1477 C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller);
1478 LCM[&C->path] = CE->getCalleeContext();
1481 C->setCallee(*CE, SM);
1482 EB.addContext(C->getLocation());
1484 if (!CallStack.empty()) {
1485 assert(CallStack.back().first == C);
1486 CallStack.pop_back();
1491 // Note that is important that we update the LocationContext
1492 // after looking at CallExits. CallExit basically adds an
1493 // edge in the *caller*, so we don't want to update the LocationContext
1495 PDB.LC = N->getLocationContext();
1498 if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) {
1499 // Does this represent entering a call? If so, look at propagating
1500 // interesting symbols across call boundaries.
1502 const LocationContext *CallerCtx = NextNode->getLocationContext();
1503 const LocationContext *CalleeCtx = PDB.LC;
1504 if (CallerCtx != CalleeCtx) {
1505 reversePropagateInterestingSymbols(*PDB.getBugReport(), IE,
1506 N->getState().get(),
1507 CalleeCtx, CallerCtx);
1511 // Are we jumping to the head of a loop? Add a special diagnostic.
1512 if (const Stmt *Loop = BE->getSrc()->getLoopTarget()) {
1513 PathDiagnosticLocation L(Loop, SM, PDB.LC);
1514 const CompoundStmt *CS = nullptr;
1516 if (const ForStmt *FS = dyn_cast<ForStmt>(Loop))
1517 CS = dyn_cast<CompoundStmt>(FS->getBody());
1518 else if (const WhileStmt *WS = dyn_cast<WhileStmt>(Loop))
1519 CS = dyn_cast<CompoundStmt>(WS->getBody());
1521 auto p = std::make_shared<PathDiagnosticEventPiece>(
1522 L, "Looping back to the head of the loop");
1523 p->setPrunable(true);
1525 EB.addEdge(p->getLocation(), true);
1526 PD.getActivePath().push_front(std::move(p));
1529 PathDiagnosticLocation BL =
1530 PathDiagnosticLocation::createEndBrace(CS, SM);
1535 const CFGBlock *BSrc = BE->getSrc();
1536 ParentMap &PM = PDB.getParentMap();
1538 if (const Stmt *Term = BSrc->getTerminator()) {
1539 // Are we jumping past the loop body without ever executing the
1540 // loop (because the condition was false)?
1541 if (isLoopJumpPastBody(Term, &*BE) &&
1543 getStmtBeforeCond(PM,
1544 BSrc->getTerminatorCondition(),
1547 PathDiagnosticLocation L(Term, SM, PDB.LC);
1548 auto PE = std::make_shared<PathDiagnosticEventPiece>(
1549 L, "Loop body executed 0 times");
1550 PE->setPrunable(true);
1552 EB.addEdge(PE->getLocation(), true);
1553 PD.getActivePath().push_front(std::move(PE));
1556 // In any case, add the terminator as the current statement
1557 // context for control edges.
1558 EB.addContext(Term);
1564 if (Optional<BlockEntrance> BE = P.getAs<BlockEntrance>()) {
1565 Optional<CFGElement> First = BE->getFirstElement();
1566 if (Optional<CFGStmt> S = First ? First->getAs<CFGStmt>() : None) {
1567 const Stmt *stmt = S->getStmt();
1568 if (IsControlFlowExpr(stmt)) {
1569 // Add the proper context for '&&', '||', and '?'.
1570 EB.addContext(stmt);
1573 EB.addExtendedContext(PDB.getEnclosingStmtLocation(stmt).asStmt());
1585 // Add pieces from custom visitors.
1586 BugReport *R = PDB.getBugReport();
1587 for (auto &V : visitors) {
1588 if (auto p = V->VisitNode(N, NextNode, PDB, *R)) {
1589 const PathDiagnosticLocation &Loc = p->getLocation();
1590 EB.addEdge(Loc, true);
1591 updateStackPiecesWithMessage(*p, CallStack);
1592 PD.getActivePath().push_front(std::move(p));
1594 if (const Stmt *S = Loc.asStmt())
1595 EB.addExtendedContext(PDB.getEnclosingStmtLocation(S).asStmt());
1600 return PDB.getBugReport()->isValid();
1603 /// \brief Adds a sanitized control-flow diagnostic edge to a path.
1604 static void addEdgeToPath(PathPieces &path,
1605 PathDiagnosticLocation &PrevLoc,
1606 PathDiagnosticLocation NewLoc,
1607 const LocationContext *LC) {
1608 if (!NewLoc.isValid())
1611 SourceLocation NewLocL = NewLoc.asLocation();
1612 if (NewLocL.isInvalid())
1615 if (!PrevLoc.isValid() || !PrevLoc.asLocation().isValid()) {
1620 // Ignore self-edges, which occur when there are multiple nodes at the same
1622 if (NewLoc.asStmt() && NewLoc.asStmt() == PrevLoc.asStmt())
1626 std::make_shared<PathDiagnosticControlFlowPiece>(NewLoc, PrevLoc));
1630 /// A customized wrapper for CFGBlock::getTerminatorCondition()
1631 /// which returns the element for ObjCForCollectionStmts.
1632 static const Stmt *getTerminatorCondition(const CFGBlock *B) {
1633 const Stmt *S = B->getTerminatorCondition();
1634 if (const ObjCForCollectionStmt *FS =
1635 dyn_cast_or_null<ObjCForCollectionStmt>(S))
1636 return FS->getElement();
1640 static const char StrEnteringLoop[] = "Entering loop body";
1641 static const char StrLoopBodyZero[] = "Loop body executed 0 times";
1642 static const char StrLoopRangeEmpty[] =
1643 "Loop body skipped when range is empty";
1644 static const char StrLoopCollectionEmpty[] =
1645 "Loop body skipped when collection is empty";
1647 static bool GenerateAlternateExtensivePathDiagnostic(
1648 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N,
1649 LocationContextMap &LCM,
1650 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) {
1652 BugReport *report = PDB.getBugReport();
1653 const SourceManager& SM = PDB.getSourceManager();
1654 StackDiagVector CallStack;
1655 InterestingExprs IE;
1657 PathDiagnosticLocation PrevLoc = PD.getLocation();
1659 const ExplodedNode *NextNode = N->getFirstPred();
1662 NextNode = N->getFirstPred();
1663 ProgramPoint P = N->getLocation();
1666 // Have we encountered an entrance to a call? It may be
1667 // the case that we have not encountered a matching
1668 // call exit before this point. This means that the path
1669 // terminated within the call itself.
1670 if (Optional<CallEnter> CE = P.getAs<CallEnter>()) {
1671 // Add an edge to the start of the function.
1672 const StackFrameContext *CalleeLC = CE->getCalleeContext();
1673 const Decl *D = CalleeLC->getDecl();
1674 // Add the edge only when the callee has body. We jump to the beginning
1675 // of the *declaration*, however we expect it to be followed by the
1676 // body. This isn't the case for autosynthesized property accessors in
1677 // Objective-C. No need for a similar extra check for CallExit points
1678 // because the exit edge comes from a statement (i.e. return),
1679 // not from declaration.
1681 addEdgeToPath(PD.getActivePath(), PrevLoc,
1682 PathDiagnosticLocation::createBegin(D, SM), CalleeLC);
1684 // Did we visit an entire call?
1685 bool VisitedEntireCall = PD.isWithinCall();
1688 PathDiagnosticCallPiece *C;
1689 if (VisitedEntireCall) {
1690 PathDiagnosticPiece *P = PD.getActivePath().front().get();
1691 C = cast<PathDiagnosticCallPiece>(P);
1693 const Decl *Caller = CE->getLocationContext()->getDecl();
1694 C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller);
1696 // Since we just transferred the path over to the call piece,
1697 // reset the mapping from active to location context.
1698 assert(PD.getActivePath().size() == 1 &&
1699 PD.getActivePath().front().get() == C);
1700 LCM[&PD.getActivePath()] = nullptr;
1702 // Record the location context mapping for the path within
1704 assert(LCM[&C->path] == nullptr ||
1705 LCM[&C->path] == CE->getCalleeContext());
1706 LCM[&C->path] = CE->getCalleeContext();
1708 // If this is the first item in the active path, record
1709 // the new mapping from active path to location context.
1710 const LocationContext *&NewLC = LCM[&PD.getActivePath()];
1712 NewLC = N->getLocationContext();
1716 C->setCallee(*CE, SM);
1718 // Update the previous location in the active path.
1719 PrevLoc = C->getLocation();
1721 if (!CallStack.empty()) {
1722 assert(CallStack.back().first == C);
1723 CallStack.pop_back();
1728 // Query the location context here and the previous location
1729 // as processing CallEnter may change the active path.
1730 PDB.LC = N->getLocationContext();
1732 // Record the mapping from the active path to the location
1734 assert(!LCM[&PD.getActivePath()] ||
1735 LCM[&PD.getActivePath()] == PDB.LC);
1736 LCM[&PD.getActivePath()] = PDB.LC;
1738 // Have we encountered an exit from a function call?
1739 if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
1740 const Stmt *S = CE->getCalleeContext()->getCallSite();
1741 // Propagate the interesting symbols accordingly.
1742 if (const Expr *Ex = dyn_cast_or_null<Expr>(S)) {
1743 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1744 N->getState().get(), Ex,
1745 N->getLocationContext());
1748 // We are descending into a call (backwards). Construct
1749 // a new call piece to contain the path pieces for that call.
1750 auto C = PathDiagnosticCallPiece::construct(N, *CE, SM);
1752 // Record the location context for this call piece.
1753 LCM[&C->path] = CE->getCalleeContext();
1755 // Add the edge to the return site.
1756 addEdgeToPath(PD.getActivePath(), PrevLoc, C->callReturn, PDB.LC);
1758 PD.getActivePath().push_front(std::move(C));
1759 PrevLoc.invalidate();
1761 // Make the contents of the call the active path for now.
1762 PD.pushActivePath(&P->path);
1763 CallStack.push_back(StackDiagPair(P, N));
1767 if (Optional<PostStmt> PS = P.getAs<PostStmt>()) {
1768 // For expressions, make sure we propagate the
1769 // interesting symbols correctly.
1770 if (const Expr *Ex = PS->getStmtAs<Expr>())
1771 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE,
1772 N->getState().get(), Ex,
1773 N->getLocationContext());
1775 // Add an edge. If this is an ObjCForCollectionStmt do
1776 // not add an edge here as it appears in the CFG both
1777 // as a terminator and as a terminator condition.
1778 if (!isa<ObjCForCollectionStmt>(PS->getStmt())) {
1779 PathDiagnosticLocation L =
1780 PathDiagnosticLocation(PS->getStmt(), SM, PDB.LC);
1781 addEdgeToPath(PD.getActivePath(), PrevLoc, L, PDB.LC);
1787 if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) {
1788 // Does this represent entering a call? If so, look at propagating
1789 // interesting symbols across call boundaries.
1791 const LocationContext *CallerCtx = NextNode->getLocationContext();
1792 const LocationContext *CalleeCtx = PDB.LC;
1793 if (CallerCtx != CalleeCtx) {
1794 reversePropagateInterestingSymbols(*PDB.getBugReport(), IE,
1795 N->getState().get(),
1796 CalleeCtx, CallerCtx);
1800 // Are we jumping to the head of a loop? Add a special diagnostic.
1801 if (const Stmt *Loop = BE->getSrc()->getLoopTarget()) {
1802 PathDiagnosticLocation L(Loop, SM, PDB.LC);
1803 const Stmt *Body = nullptr;
1805 if (const ForStmt *FS = dyn_cast<ForStmt>(Loop))
1806 Body = FS->getBody();
1807 else if (const WhileStmt *WS = dyn_cast<WhileStmt>(Loop))
1808 Body = WS->getBody();
1809 else if (const ObjCForCollectionStmt *OFS =
1810 dyn_cast<ObjCForCollectionStmt>(Loop)) {
1811 Body = OFS->getBody();
1812 } else if (const CXXForRangeStmt *FRS =
1813 dyn_cast<CXXForRangeStmt>(Loop)) {
1814 Body = FRS->getBody();
1816 // do-while statements are explicitly excluded here
1818 auto p = std::make_shared<PathDiagnosticEventPiece>(
1819 L, "Looping back to the head "
1821 p->setPrunable(true);
1823 addEdgeToPath(PD.getActivePath(), PrevLoc, p->getLocation(), PDB.LC);
1824 PD.getActivePath().push_front(std::move(p));
1826 if (const CompoundStmt *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
1827 addEdgeToPath(PD.getActivePath(), PrevLoc,
1828 PathDiagnosticLocation::createEndBrace(CS, SM),
1833 const CFGBlock *BSrc = BE->getSrc();
1834 ParentMap &PM = PDB.getParentMap();
1836 if (const Stmt *Term = BSrc->getTerminator()) {
1837 // Are we jumping past the loop body without ever executing the
1838 // loop (because the condition was false)?
1840 const Stmt *TermCond = getTerminatorCondition(BSrc);
1842 isInLoopBody(PM, getStmtBeforeCond(PM, TermCond, N), Term);
1844 const char *str = nullptr;
1846 if (isJumpToFalseBranch(&*BE)) {
1847 if (!IsInLoopBody) {
1848 if (isa<ObjCForCollectionStmt>(Term)) {
1849 str = StrLoopCollectionEmpty;
1850 } else if (isa<CXXForRangeStmt>(Term)) {
1851 str = StrLoopRangeEmpty;
1853 str = StrLoopBodyZero;
1857 str = StrEnteringLoop;
1861 PathDiagnosticLocation L(TermCond ? TermCond : Term, SM, PDB.LC);
1862 auto PE = std::make_shared<PathDiagnosticEventPiece>(L, str);
1863 PE->setPrunable(true);
1864 addEdgeToPath(PD.getActivePath(), PrevLoc,
1865 PE->getLocation(), PDB.LC);
1866 PD.getActivePath().push_front(std::move(PE));
1868 } else if (isa<BreakStmt>(Term) || isa<ContinueStmt>(Term) ||
1869 isa<GotoStmt>(Term)) {
1870 PathDiagnosticLocation L(Term, SM, PDB.LC);
1871 addEdgeToPath(PD.getActivePath(), PrevLoc, L, PDB.LC);
1881 // Add pieces from custom visitors.
1882 for (auto &V : visitors) {
1883 if (auto p = V->VisitNode(N, NextNode, PDB, *report)) {
1884 addEdgeToPath(PD.getActivePath(), PrevLoc, p->getLocation(), PDB.LC);
1885 updateStackPiecesWithMessage(*p, CallStack);
1886 PD.getActivePath().push_front(std::move(p));
1891 // Add an edge to the start of the function.
1892 // We'll prune it out later, but it helps make diagnostics more uniform.
1893 const StackFrameContext *CalleeLC = PDB.LC->getCurrentStackFrame();
1894 const Decl *D = CalleeLC->getDecl();
1895 addEdgeToPath(PD.getActivePath(), PrevLoc,
1896 PathDiagnosticLocation::createBegin(D, SM),
1899 return report->isValid();
1902 static const Stmt *getLocStmt(PathDiagnosticLocation L) {
1908 static const Stmt *getStmtParent(const Stmt *S, const ParentMap &PM) {
1913 S = PM.getParentIgnoreParens(S);
1918 if (isa<ExprWithCleanups>(S) ||
1919 isa<CXXBindTemporaryExpr>(S) ||
1920 isa<SubstNonTypeTemplateParmExpr>(S))
1929 static bool isConditionForTerminator(const Stmt *S, const Stmt *Cond) {
1930 switch (S->getStmtClass()) {
1931 case Stmt::BinaryOperatorClass: {
1932 const BinaryOperator *BO = cast<BinaryOperator>(S);
1933 if (!BO->isLogicalOp())
1935 return BO->getLHS() == Cond || BO->getRHS() == Cond;
1937 case Stmt::IfStmtClass:
1938 return cast<IfStmt>(S)->getCond() == Cond;
1939 case Stmt::ForStmtClass:
1940 return cast<ForStmt>(S)->getCond() == Cond;
1941 case Stmt::WhileStmtClass:
1942 return cast<WhileStmt>(S)->getCond() == Cond;
1943 case Stmt::DoStmtClass:
1944 return cast<DoStmt>(S)->getCond() == Cond;
1945 case Stmt::ChooseExprClass:
1946 return cast<ChooseExpr>(S)->getCond() == Cond;
1947 case Stmt::IndirectGotoStmtClass:
1948 return cast<IndirectGotoStmt>(S)->getTarget() == Cond;
1949 case Stmt::SwitchStmtClass:
1950 return cast<SwitchStmt>(S)->getCond() == Cond;
1951 case Stmt::BinaryConditionalOperatorClass:
1952 return cast<BinaryConditionalOperator>(S)->getCond() == Cond;
1953 case Stmt::ConditionalOperatorClass: {
1954 const ConditionalOperator *CO = cast<ConditionalOperator>(S);
1955 return CO->getCond() == Cond ||
1956 CO->getLHS() == Cond ||
1957 CO->getRHS() == Cond;
1959 case Stmt::ObjCForCollectionStmtClass:
1960 return cast<ObjCForCollectionStmt>(S)->getElement() == Cond;
1961 case Stmt::CXXForRangeStmtClass: {
1962 const CXXForRangeStmt *FRS = cast<CXXForRangeStmt>(S);
1963 return FRS->getCond() == Cond || FRS->getRangeInit() == Cond;
1970 static bool isIncrementOrInitInForLoop(const Stmt *S, const Stmt *FL) {
1971 if (const ForStmt *FS = dyn_cast<ForStmt>(FL))
1972 return FS->getInc() == S || FS->getInit() == S;
1973 if (const CXXForRangeStmt *FRS = dyn_cast<CXXForRangeStmt>(FL))
1974 return FRS->getInc() == S || FRS->getRangeStmt() == S ||
1975 FRS->getLoopVarStmt() || FRS->getRangeInit() == S;
1979 typedef llvm::DenseSet<const PathDiagnosticCallPiece *>
1982 /// Adds synthetic edges from top-level statements to their subexpressions.
1984 /// This avoids a "swoosh" effect, where an edge from a top-level statement A
1985 /// points to a sub-expression B.1 that's not at the start of B. In these cases,
1986 /// we'd like to see an edge from A to B, then another one from B to B.1.
1987 static void addContextEdges(PathPieces &pieces, SourceManager &SM,
1988 const ParentMap &PM, const LocationContext *LCtx) {
1989 PathPieces::iterator Prev = pieces.end();
1990 for (PathPieces::iterator I = pieces.begin(), E = Prev; I != E;
1992 PathDiagnosticControlFlowPiece *Piece =
1993 dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
1998 PathDiagnosticLocation SrcLoc = Piece->getStartLocation();
1999 SmallVector<PathDiagnosticLocation, 4> SrcContexts;
2001 PathDiagnosticLocation NextSrcContext = SrcLoc;
2002 const Stmt *InnerStmt = nullptr;
2003 while (NextSrcContext.isValid() && NextSrcContext.asStmt() != InnerStmt) {
2004 SrcContexts.push_back(NextSrcContext);
2005 InnerStmt = NextSrcContext.asStmt();
2006 NextSrcContext = getEnclosingStmtLocation(InnerStmt, SM, PM, LCtx,
2007 /*allowNested=*/true);
2010 // Repeatedly split the edge as necessary.
2011 // This is important for nested logical expressions (||, &&, ?:) where we
2012 // want to show all the levels of context.
2014 const Stmt *Dst = getLocStmt(Piece->getEndLocation());
2016 // We are looking at an edge. Is the destination within a larger
2018 PathDiagnosticLocation DstContext =
2019 getEnclosingStmtLocation(Dst, SM, PM, LCtx, /*allowNested=*/true);
2020 if (!DstContext.isValid() || DstContext.asStmt() == Dst)
2023 // If the source is in the same context, we're already good.
2024 if (std::find(SrcContexts.begin(), SrcContexts.end(), DstContext) !=
2028 // Update the subexpression node to point to the context edge.
2029 Piece->setStartLocation(DstContext);
2031 // Try to extend the previous edge if it's at the same level as the source
2034 auto *PrevPiece = dyn_cast<PathDiagnosticControlFlowPiece>(Prev->get());
2037 if (const Stmt *PrevSrc = getLocStmt(PrevPiece->getStartLocation())) {
2038 const Stmt *PrevSrcParent = getStmtParent(PrevSrc, PM);
2039 if (PrevSrcParent == getStmtParent(getLocStmt(DstContext), PM)) {
2040 PrevPiece->setEndLocation(DstContext);
2047 // Otherwise, split the current edge into a context edge and a
2048 // subexpression edge. Note that the context statement may itself have
2051 std::make_shared<PathDiagnosticControlFlowPiece>(SrcLoc, DstContext);
2053 I = pieces.insert(I, std::move(P));
2058 /// \brief Move edges from a branch condition to a branch target
2059 /// when the condition is simple.
2061 /// This restructures some of the work of addContextEdges. That function
2062 /// creates edges this may destroy, but they work together to create a more
2063 /// aesthetically set of edges around branches. After the call to
2064 /// addContextEdges, we may have (1) an edge to the branch, (2) an edge from
2065 /// the branch to the branch condition, and (3) an edge from the branch
2066 /// condition to the branch target. We keep (1), but may wish to remove (2)
2067 /// and move the source of (3) to the branch if the branch condition is simple.
2069 static void simplifySimpleBranches(PathPieces &pieces) {
2070 for (PathPieces::iterator I = pieces.begin(), E = pieces.end(); I != E; ++I) {
2072 auto *PieceI = dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
2077 const Stmt *s1Start = getLocStmt(PieceI->getStartLocation());
2078 const Stmt *s1End = getLocStmt(PieceI->getEndLocation());
2080 if (!s1Start || !s1End)
2083 PathPieces::iterator NextI = I; ++NextI;
2087 PathDiagnosticControlFlowPiece *PieceNextI = nullptr;
2093 auto *EV = dyn_cast<PathDiagnosticEventPiece>(NextI->get());
2095 StringRef S = EV->getString();
2096 if (S == StrEnteringLoop || S == StrLoopBodyZero ||
2097 S == StrLoopCollectionEmpty || S == StrLoopRangeEmpty) {
2104 PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(NextI->get());
2111 const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation());
2112 const Stmt *s2End = getLocStmt(PieceNextI->getEndLocation());
2114 if (!s2Start || !s2End || s1End != s2Start)
2117 // We only perform this transformation for specific branch kinds.
2118 // We don't want to do this for do..while, for example.
2119 if (!(isa<ForStmt>(s1Start) || isa<WhileStmt>(s1Start) ||
2120 isa<IfStmt>(s1Start) || isa<ObjCForCollectionStmt>(s1Start) ||
2121 isa<CXXForRangeStmt>(s1Start)))
2124 // Is s1End the branch condition?
2125 if (!isConditionForTerminator(s1Start, s1End))
2128 // Perform the hoisting by eliminating (2) and changing the start
2130 PieceNextI->setStartLocation(PieceI->getStartLocation());
2131 I = pieces.erase(I);
2135 /// Returns the number of bytes in the given (character-based) SourceRange.
2137 /// If the locations in the range are not on the same line, returns None.
2139 /// Note that this does not do a precise user-visible character or column count.
2140 static Optional<size_t> getLengthOnSingleLine(SourceManager &SM,
2141 SourceRange Range) {
2142 SourceRange ExpansionRange(SM.getExpansionLoc(Range.getBegin()),
2143 SM.getExpansionRange(Range.getEnd()).second);
2145 FileID FID = SM.getFileID(ExpansionRange.getBegin());
2146 if (FID != SM.getFileID(ExpansionRange.getEnd()))
2150 const llvm::MemoryBuffer *Buffer = SM.getBuffer(FID, &Invalid);
2154 unsigned BeginOffset = SM.getFileOffset(ExpansionRange.getBegin());
2155 unsigned EndOffset = SM.getFileOffset(ExpansionRange.getEnd());
2156 StringRef Snippet = Buffer->getBuffer().slice(BeginOffset, EndOffset);
2158 // We're searching the raw bytes of the buffer here, which might include
2159 // escaped newlines and such. That's okay; we're trying to decide whether the
2160 // SourceRange is covering a large or small amount of space in the user's
2162 if (Snippet.find_first_of("\r\n") != StringRef::npos)
2165 // This isn't Unicode-aware, but it doesn't need to be.
2166 return Snippet.size();
2169 /// \sa getLengthOnSingleLine(SourceManager, SourceRange)
2170 static Optional<size_t> getLengthOnSingleLine(SourceManager &SM,
2172 return getLengthOnSingleLine(SM, S->getSourceRange());
2175 /// Eliminate two-edge cycles created by addContextEdges().
2177 /// Once all the context edges are in place, there are plenty of cases where
2178 /// there's a single edge from a top-level statement to a subexpression,
2179 /// followed by a single path note, and then a reverse edge to get back out to
2180 /// the top level. If the statement is simple enough, the subexpression edges
2181 /// just add noise and make it harder to understand what's going on.
2183 /// This function only removes edges in pairs, because removing only one edge
2184 /// might leave other edges dangling.
2186 /// This will not remove edges in more complicated situations:
2187 /// - if there is more than one "hop" leading to or from a subexpression.
2188 /// - if there is an inlined call between the edges instead of a single event.
2189 /// - if the whole statement is large enough that having subexpression arrows
2190 /// might be helpful.
2191 static void removeContextCycles(PathPieces &Path, SourceManager &SM,
2193 for (PathPieces::iterator I = Path.begin(), E = Path.end(); I != E; ) {
2194 // Pattern match the current piece and its successor.
2195 PathDiagnosticControlFlowPiece *PieceI =
2196 dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
2203 const Stmt *s1Start = getLocStmt(PieceI->getStartLocation());
2204 const Stmt *s1End = getLocStmt(PieceI->getEndLocation());
2206 PathPieces::iterator NextI = I; ++NextI;
2210 PathDiagnosticControlFlowPiece *PieceNextI =
2211 dyn_cast<PathDiagnosticControlFlowPiece>(NextI->get());
2214 if (isa<PathDiagnosticEventPiece>(NextI->get())) {
2218 PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(NextI->get());
2227 const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation());
2228 const Stmt *s2End = getLocStmt(PieceNextI->getEndLocation());
2230 if (s1Start && s2Start && s1Start == s2End && s2Start == s1End) {
2231 const size_t MAX_SHORT_LINE_LENGTH = 80;
2232 Optional<size_t> s1Length = getLengthOnSingleLine(SM, s1Start);
2233 if (s1Length && *s1Length <= MAX_SHORT_LINE_LENGTH) {
2234 Optional<size_t> s2Length = getLengthOnSingleLine(SM, s2Start);
2235 if (s2Length && *s2Length <= MAX_SHORT_LINE_LENGTH) {
2237 I = Path.erase(NextI);
2247 /// \brief Return true if X is contained by Y.
2248 static bool lexicalContains(ParentMap &PM,
2254 X = PM.getParent(X);
2259 // Remove short edges on the same line less than 3 columns in difference.
2260 static void removePunyEdges(PathPieces &path,
2264 bool erased = false;
2266 for (PathPieces::iterator I = path.begin(), E = path.end(); I != E;
2271 auto *PieceI = dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
2276 const Stmt *start = getLocStmt(PieceI->getStartLocation());
2277 const Stmt *end = getLocStmt(PieceI->getEndLocation());
2282 const Stmt *endParent = PM.getParent(end);
2286 if (isConditionForTerminator(end, endParent))
2289 SourceLocation FirstLoc = start->getLocStart();
2290 SourceLocation SecondLoc = end->getLocStart();
2292 if (!SM.isWrittenInSameFile(FirstLoc, SecondLoc))
2294 if (SM.isBeforeInTranslationUnit(SecondLoc, FirstLoc))
2295 std::swap(SecondLoc, FirstLoc);
2297 SourceRange EdgeRange(FirstLoc, SecondLoc);
2298 Optional<size_t> ByteWidth = getLengthOnSingleLine(SM, EdgeRange);
2300 // If the statements are on different lines, continue.
2304 const size_t MAX_PUNY_EDGE_LENGTH = 2;
2305 if (*ByteWidth <= MAX_PUNY_EDGE_LENGTH) {
2306 // FIXME: There are enough /bytes/ between the endpoints of the edge, but
2307 // there might not be enough /columns/. A proper user-visible column count
2308 // is probably too expensive, though.
2316 static void removeIdenticalEvents(PathPieces &path) {
2317 for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ++I) {
2318 auto *PieceI = dyn_cast<PathDiagnosticEventPiece>(I->get());
2323 PathPieces::iterator NextI = I; ++NextI;
2327 auto *PieceNextI = dyn_cast<PathDiagnosticEventPiece>(NextI->get());
2332 // Erase the second piece if it has the same exact message text.
2333 if (PieceI->getString() == PieceNextI->getString()) {
2339 static bool optimizeEdges(PathPieces &path, SourceManager &SM,
2340 OptimizedCallsSet &OCS,
2341 LocationContextMap &LCM) {
2342 bool hasChanges = false;
2343 const LocationContext *LC = LCM[&path];
2345 ParentMap &PM = LC->getParentMap();
2347 for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ) {
2348 // Optimize subpaths.
2349 if (auto *CallI = dyn_cast<PathDiagnosticCallPiece>(I->get())) {
2350 // Record the fact that a call has been optimized so we only do the
2352 if (!OCS.count(CallI)) {
2353 while (optimizeEdges(CallI->path, SM, OCS, LCM)) {}
2360 // Pattern match the current piece and its successor.
2361 auto *PieceI = dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
2368 const Stmt *s1Start = getLocStmt(PieceI->getStartLocation());
2369 const Stmt *s1End = getLocStmt(PieceI->getEndLocation());
2370 const Stmt *level1 = getStmtParent(s1Start, PM);
2371 const Stmt *level2 = getStmtParent(s1End, PM);
2373 PathPieces::iterator NextI = I; ++NextI;
2377 auto *PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(NextI->get());
2384 const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation());
2385 const Stmt *s2End = getLocStmt(PieceNextI->getEndLocation());
2386 const Stmt *level3 = getStmtParent(s2Start, PM);
2387 const Stmt *level4 = getStmtParent(s2End, PM);
2391 // If we have two consecutive control edges whose end/begin locations
2392 // are at the same level (e.g. statements or top-level expressions within
2393 // a compound statement, or siblings share a single ancestor expression),
2394 // then merge them if they have no interesting intermediate event.
2398 // (1.1 -> 1.2) -> (1.2 -> 1.3) becomes (1.1 -> 1.3) because the common
2399 // parent is '1'. Here 'x.y.z' represents the hierarchy of statements.
2401 // NOTE: this will be limited later in cases where we add barriers
2402 // to prevent this optimization.
2404 if (level1 && level1 == level2 && level1 == level3 && level1 == level4) {
2405 PieceI->setEndLocation(PieceNextI->getEndLocation());
2413 // Eliminate edges between subexpressions and parent expressions
2414 // when the subexpression is consumed.
2416 // NOTE: this will be limited later in cases where we add barriers
2417 // to prevent this optimization.
2419 if (s1End && s1End == s2Start && level2) {
2420 bool removeEdge = false;
2421 // Remove edges into the increment or initialization of a
2422 // loop that have no interleaving event. This means that
2423 // they aren't interesting.
2424 if (isIncrementOrInitInForLoop(s1End, level2))
2426 // Next only consider edges that are not anchored on
2427 // the condition of a terminator. This are intermediate edges
2428 // that we might want to trim.
2429 else if (!isConditionForTerminator(level2, s1End)) {
2430 // Trim edges on expressions that are consumed by
2431 // the parent expression.
2432 if (isa<Expr>(s1End) && PM.isConsumedExpr(cast<Expr>(s1End))) {
2435 // Trim edges where a lexical containment doesn't exist.
2440 // If 'Z' lexically contains Y (it is an ancestor) and
2441 // 'X' does not lexically contain Y (it is a descendant OR
2442 // it has no lexical relationship at all) then trim.
2444 // This can eliminate edges where we dive into a subexpression
2445 // and then pop back out, etc.
2446 else if (s1Start && s2End &&
2447 lexicalContains(PM, s2Start, s2End) &&
2448 !lexicalContains(PM, s1End, s1Start)) {
2451 // Trim edges from a subexpression back to the top level if the
2452 // subexpression is on a different line.
2458 // These edges just look ugly and don't usually add anything.
2459 else if (s1Start && s2End &&
2460 lexicalContains(PM, s1Start, s1End)) {
2461 SourceRange EdgeRange(PieceI->getEndLocation().asLocation(),
2462 PieceI->getStartLocation().asLocation());
2463 if (!getLengthOnSingleLine(SM, EdgeRange).hasValue())
2469 PieceI->setEndLocation(PieceNextI->getEndLocation());
2476 // Optimize edges for ObjC fast-enumeration loops.
2478 // (X -> collection) -> (collection -> element)
2483 if (s1End == s2Start) {
2484 const ObjCForCollectionStmt *FS =
2485 dyn_cast_or_null<ObjCForCollectionStmt>(level3);
2486 if (FS && FS->getCollection()->IgnoreParens() == s2Start &&
2487 s2End == FS->getElement()) {
2488 PieceI->setEndLocation(PieceNextI->getEndLocation());
2495 // No changes at this index? Move to the next one.
2500 // Adjust edges into subexpressions to make them more uniform
2501 // and aesthetically pleasing.
2502 addContextEdges(path, SM, PM, LC);
2503 // Remove "cyclical" edges that include one or more context edges.
2504 removeContextCycles(path, SM, PM);
2505 // Hoist edges originating from branch conditions to branches
2506 // for simple branches.
2507 simplifySimpleBranches(path);
2508 // Remove any puny edges left over after primary optimization pass.
2509 removePunyEdges(path, SM, PM);
2510 // Remove identical events.
2511 removeIdenticalEvents(path);
2517 /// Drop the very first edge in a path, which should be a function entry edge.
2519 /// If the first edge is not a function entry edge (say, because the first
2520 /// statement had an invalid source location), this function does nothing.
2521 // FIXME: We should just generate invalid edges anyway and have the optimizer
2523 static void dropFunctionEntryEdge(PathPieces &Path,
2524 LocationContextMap &LCM,
2525 SourceManager &SM) {
2526 const auto *FirstEdge =
2527 dyn_cast<PathDiagnosticControlFlowPiece>(Path.front().get());
2531 const Decl *D = LCM[&Path]->getDecl();
2532 PathDiagnosticLocation EntryLoc = PathDiagnosticLocation::createBegin(D, SM);
2533 if (FirstEdge->getStartLocation() != EntryLoc)
2540 //===----------------------------------------------------------------------===//
2541 // Methods for BugType and subclasses.
2542 //===----------------------------------------------------------------------===//
2543 void BugType::anchor() { }
2545 void BugType::FlushReports(BugReporter &BR) {}
2547 void BuiltinBug::anchor() {}
2549 //===----------------------------------------------------------------------===//
2550 // Methods for BugReport and subclasses.
2551 //===----------------------------------------------------------------------===//
2553 void BugReport::NodeResolver::anchor() {}
2555 void BugReport::addVisitor(std::unique_ptr<BugReporterVisitor> visitor) {
2559 llvm::FoldingSetNodeID ID;
2560 visitor->Profile(ID);
2563 if (CallbacksSet.FindNodeOrInsertPos(ID, InsertPos))
2566 CallbacksSet.InsertNode(visitor.get(), InsertPos);
2567 Callbacks.push_back(std::move(visitor));
2568 ++ConfigurationChangeToken;
2571 BugReport::~BugReport() {
2572 while (!interestingSymbols.empty()) {
2573 popInterestingSymbolsAndRegions();
2577 const Decl *BugReport::getDeclWithIssue() const {
2579 return DeclWithIssue;
2581 const ExplodedNode *N = getErrorNode();
2585 const LocationContext *LC = N->getLocationContext();
2586 return LC->getCurrentStackFrame()->getDecl();
2589 void BugReport::Profile(llvm::FoldingSetNodeID& hash) const {
2590 hash.AddPointer(&BT);
2591 hash.AddString(Description);
2592 PathDiagnosticLocation UL = getUniqueingLocation();
2595 } else if (Location.isValid()) {
2596 Location.Profile(hash);
2599 hash.AddPointer(GetCurrentOrPreviousStmt(ErrorNode));
2602 for (SourceRange range : Ranges) {
2603 if (!range.isValid())
2605 hash.AddInteger(range.getBegin().getRawEncoding());
2606 hash.AddInteger(range.getEnd().getRawEncoding());
2610 void BugReport::markInteresting(SymbolRef sym) {
2614 // If the symbol wasn't already in our set, note a configuration change.
2615 if (getInterestingSymbols().insert(sym).second)
2616 ++ConfigurationChangeToken;
2618 if (const SymbolMetadata *meta = dyn_cast<SymbolMetadata>(sym))
2619 getInterestingRegions().insert(meta->getRegion());
2622 void BugReport::markInteresting(const MemRegion *R) {
2626 // If the base region wasn't already in our set, note a configuration change.
2627 R = R->getBaseRegion();
2628 if (getInterestingRegions().insert(R).second)
2629 ++ConfigurationChangeToken;
2631 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R))
2632 getInterestingSymbols().insert(SR->getSymbol());
2635 void BugReport::markInteresting(SVal V) {
2636 markInteresting(V.getAsRegion());
2637 markInteresting(V.getAsSymbol());
2640 void BugReport::markInteresting(const LocationContext *LC) {
2643 InterestingLocationContexts.insert(LC);
2646 bool BugReport::isInteresting(SVal V) {
2647 return isInteresting(V.getAsRegion()) || isInteresting(V.getAsSymbol());
2650 bool BugReport::isInteresting(SymbolRef sym) {
2653 // We don't currently consider metadata symbols to be interesting
2654 // even if we know their region is interesting. Is that correct behavior?
2655 return getInterestingSymbols().count(sym);
2658 bool BugReport::isInteresting(const MemRegion *R) {
2661 R = R->getBaseRegion();
2662 bool b = getInterestingRegions().count(R);
2665 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R))
2666 return getInterestingSymbols().count(SR->getSymbol());
2670 bool BugReport::isInteresting(const LocationContext *LC) {
2673 return InterestingLocationContexts.count(LC);
2676 void BugReport::lazyInitializeInterestingSets() {
2677 if (interestingSymbols.empty()) {
2678 interestingSymbols.push_back(new Symbols());
2679 interestingRegions.push_back(new Regions());
2683 BugReport::Symbols &BugReport::getInterestingSymbols() {
2684 lazyInitializeInterestingSets();
2685 return *interestingSymbols.back();
2688 BugReport::Regions &BugReport::getInterestingRegions() {
2689 lazyInitializeInterestingSets();
2690 return *interestingRegions.back();
2693 void BugReport::pushInterestingSymbolsAndRegions() {
2694 interestingSymbols.push_back(new Symbols(getInterestingSymbols()));
2695 interestingRegions.push_back(new Regions(getInterestingRegions()));
2698 void BugReport::popInterestingSymbolsAndRegions() {
2699 delete interestingSymbols.pop_back_val();
2700 delete interestingRegions.pop_back_val();
2703 const Stmt *BugReport::getStmt() const {
2707 ProgramPoint ProgP = ErrorNode->getLocation();
2708 const Stmt *S = nullptr;
2710 if (Optional<BlockEntrance> BE = ProgP.getAs<BlockEntrance>()) {
2711 CFGBlock &Exit = ProgP.getLocationContext()->getCFG()->getExit();
2712 if (BE->getBlock() == &Exit)
2713 S = GetPreviousStmt(ErrorNode);
2716 S = PathDiagnosticLocation::getStmt(ErrorNode);
2721 llvm::iterator_range<BugReport::ranges_iterator> BugReport::getRanges() {
2722 // If no custom ranges, add the range of the statement corresponding to
2724 if (Ranges.empty()) {
2725 if (const Expr *E = dyn_cast_or_null<Expr>(getStmt()))
2726 addRange(E->getSourceRange());
2728 return llvm::make_range(ranges_iterator(), ranges_iterator());
2731 // User-specified absence of range info.
2732 if (Ranges.size() == 1 && !Ranges.begin()->isValid())
2733 return llvm::make_range(ranges_iterator(), ranges_iterator());
2735 return llvm::make_range(Ranges.begin(), Ranges.end());
2738 PathDiagnosticLocation BugReport::getLocation(const SourceManager &SM) const {
2740 assert(!Location.isValid() &&
2741 "Either Location or ErrorNode should be specified but not both.");
2742 return PathDiagnosticLocation::createEndOfPath(ErrorNode, SM);
2745 assert(Location.isValid());
2749 //===----------------------------------------------------------------------===//
2750 // Methods for BugReporter and subclasses.
2751 //===----------------------------------------------------------------------===//
2753 BugReportEquivClass::~BugReportEquivClass() { }
2754 GRBugReporter::~GRBugReporter() { }
2755 BugReporterData::~BugReporterData() {}
2757 ExplodedGraph &GRBugReporter::getGraph() { return Eng.getGraph(); }
2759 ProgramStateManager&
2760 GRBugReporter::getStateManager() { return Eng.getStateManager(); }
2762 BugReporter::~BugReporter() {
2765 // Free the bug reports we are tracking.
2766 typedef std::vector<BugReportEquivClass *> ContTy;
2767 for (ContTy::iterator I = EQClassesVector.begin(), E = EQClassesVector.end();
2773 void BugReporter::FlushReports() {
2774 if (BugTypes.isEmpty())
2777 // First flush the warnings for each BugType. This may end up creating new
2778 // warnings and new BugTypes.
2779 // FIXME: Only NSErrorChecker needs BugType's FlushReports.
2780 // Turn NSErrorChecker into a proper checker and remove this.
2781 SmallVector<const BugType *, 16> bugTypes(BugTypes.begin(), BugTypes.end());
2782 for (SmallVectorImpl<const BugType *>::iterator
2783 I = bugTypes.begin(), E = bugTypes.end(); I != E; ++I)
2784 const_cast<BugType*>(*I)->FlushReports(*this);
2786 // We need to flush reports in deterministic order to ensure the order
2787 // of the reports is consistent between runs.
2788 typedef std::vector<BugReportEquivClass *> ContVecTy;
2789 for (ContVecTy::iterator EI=EQClassesVector.begin(), EE=EQClassesVector.end();
2791 BugReportEquivClass& EQ = **EI;
2795 // BugReporter owns and deletes only BugTypes created implicitly through
2797 // FIXME: There are leaks from checkers that assume that the BugTypes they
2798 // create will be destroyed by the BugReporter.
2799 llvm::DeleteContainerSeconds(StrBugTypes);
2801 // Remove all references to the BugType objects.
2802 BugTypes = F.getEmptySet();
2805 //===----------------------------------------------------------------------===//
2806 // PathDiagnostics generation.
2807 //===----------------------------------------------------------------------===//
2810 /// A wrapper around a report graph, which contains only a single path, and its
2814 InterExplodedGraphMap BackMap;
2815 std::unique_ptr<ExplodedGraph> Graph;
2816 const ExplodedNode *ErrorNode;
2820 /// A wrapper around a trimmed graph and its node maps.
2821 class TrimmedGraph {
2822 InterExplodedGraphMap InverseMap;
2824 typedef llvm::DenseMap<const ExplodedNode *, unsigned> PriorityMapTy;
2825 PriorityMapTy PriorityMap;
2827 typedef std::pair<const ExplodedNode *, size_t> NodeIndexPair;
2828 SmallVector<NodeIndexPair, 32> ReportNodes;
2830 std::unique_ptr<ExplodedGraph> G;
2832 /// A helper class for sorting ExplodedNodes by priority.
2833 template <bool Descending>
2834 class PriorityCompare {
2835 const PriorityMapTy &PriorityMap;
2838 PriorityCompare(const PriorityMapTy &M) : PriorityMap(M) {}
2840 bool operator()(const ExplodedNode *LHS, const ExplodedNode *RHS) const {
2841 PriorityMapTy::const_iterator LI = PriorityMap.find(LHS);
2842 PriorityMapTy::const_iterator RI = PriorityMap.find(RHS);
2843 PriorityMapTy::const_iterator E = PriorityMap.end();
2850 return Descending ? LI->second > RI->second
2851 : LI->second < RI->second;
2854 bool operator()(const NodeIndexPair &LHS, const NodeIndexPair &RHS) const {
2855 return (*this)(LHS.first, RHS.first);
2860 TrimmedGraph(const ExplodedGraph *OriginalGraph,
2861 ArrayRef<const ExplodedNode *> Nodes);
2863 bool popNextReportGraph(ReportGraph &GraphWrapper);
2867 TrimmedGraph::TrimmedGraph(const ExplodedGraph *OriginalGraph,
2868 ArrayRef<const ExplodedNode *> Nodes) {
2869 // The trimmed graph is created in the body of the constructor to ensure
2870 // that the DenseMaps have been initialized already.
2871 InterExplodedGraphMap ForwardMap;
2872 G = OriginalGraph->trim(Nodes, &ForwardMap, &InverseMap);
2874 // Find the (first) error node in the trimmed graph. We just need to consult
2875 // the node map which maps from nodes in the original graph to nodes
2876 // in the new graph.
2877 llvm::SmallPtrSet<const ExplodedNode *, 32> RemainingNodes;
2879 for (unsigned i = 0, count = Nodes.size(); i < count; ++i) {
2880 if (const ExplodedNode *NewNode = ForwardMap.lookup(Nodes[i])) {
2881 ReportNodes.push_back(std::make_pair(NewNode, i));
2882 RemainingNodes.insert(NewNode);
2886 assert(!RemainingNodes.empty() && "No error node found in the trimmed graph");
2888 // Perform a forward BFS to find all the shortest paths.
2889 std::queue<const ExplodedNode *> WS;
2891 assert(G->num_roots() == 1);
2892 WS.push(*G->roots_begin());
2893 unsigned Priority = 0;
2895 while (!WS.empty()) {
2896 const ExplodedNode *Node = WS.front();
2899 PriorityMapTy::iterator PriorityEntry;
2901 std::tie(PriorityEntry, IsNew) =
2902 PriorityMap.insert(std::make_pair(Node, Priority));
2906 assert(PriorityEntry->second <= Priority);
2910 if (RemainingNodes.erase(Node))
2911 if (RemainingNodes.empty())
2914 for (ExplodedNode::const_pred_iterator I = Node->succ_begin(),
2915 E = Node->succ_end();
2920 // Sort the error paths from longest to shortest.
2921 std::sort(ReportNodes.begin(), ReportNodes.end(),
2922 PriorityCompare<true>(PriorityMap));
2925 bool TrimmedGraph::popNextReportGraph(ReportGraph &GraphWrapper) {
2926 if (ReportNodes.empty())
2929 const ExplodedNode *OrigN;
2930 std::tie(OrigN, GraphWrapper.Index) = ReportNodes.pop_back_val();
2931 assert(PriorityMap.find(OrigN) != PriorityMap.end() &&
2932 "error node not accessible from root");
2934 // Create a new graph with a single path. This is the graph
2935 // that will be returned to the caller.
2936 auto GNew = llvm::make_unique<ExplodedGraph>();
2937 GraphWrapper.BackMap.clear();
2939 // Now walk from the error node up the BFS path, always taking the
2940 // predeccessor with the lowest number.
2941 ExplodedNode *Succ = nullptr;
2943 // Create the equivalent node in the new graph with the same state
2945 ExplodedNode *NewN = GNew->createUncachedNode(OrigN->getLocation(), OrigN->getState(),
2948 // Store the mapping to the original node.
2949 InterExplodedGraphMap::const_iterator IMitr = InverseMap.find(OrigN);
2950 assert(IMitr != InverseMap.end() && "No mapping to original node.");
2951 GraphWrapper.BackMap[NewN] = IMitr->second;
2953 // Link up the new node with the previous node.
2955 Succ->addPredecessor(NewN, *GNew);
2957 GraphWrapper.ErrorNode = NewN;
2961 // Are we at the final node?
2962 if (OrigN->pred_empty()) {
2963 GNew->addRoot(NewN);
2967 // Find the next predeccessor node. We choose the node that is marked
2968 // with the lowest BFS number.
2969 OrigN = *std::min_element(OrigN->pred_begin(), OrigN->pred_end(),
2970 PriorityCompare<false>(PriorityMap));
2973 GraphWrapper.Graph = std::move(GNew);
2979 /// CompactPathDiagnostic - This function postprocesses a PathDiagnostic object
2980 /// and collapses PathDiagosticPieces that are expanded by macros.
2981 static void CompactPathDiagnostic(PathPieces &path, const SourceManager& SM) {
2982 typedef std::vector<
2983 std::pair<std::shared_ptr<PathDiagnosticMacroPiece>, SourceLocation>>
2986 typedef std::vector<std::shared_ptr<PathDiagnosticPiece>> PiecesTy;
2988 MacroStackTy MacroStack;
2991 for (PathPieces::const_iterator I = path.begin(), E = path.end();
2996 // Recursively compact calls.
2997 if (auto *call = dyn_cast<PathDiagnosticCallPiece>(&*piece)) {
2998 CompactPathDiagnostic(call->path, SM);
3001 // Get the location of the PathDiagnosticPiece.
3002 const FullSourceLoc Loc = piece->getLocation().asLocation();
3004 // Determine the instantiation location, which is the location we group
3005 // related PathDiagnosticPieces.
3006 SourceLocation InstantiationLoc = Loc.isMacroID() ?
3007 SM.getExpansionLoc(Loc) :
3010 if (Loc.isFileID()) {
3012 Pieces.push_back(piece);
3016 assert(Loc.isMacroID());
3018 // Is the PathDiagnosticPiece within the same macro group?
3019 if (!MacroStack.empty() && InstantiationLoc == MacroStack.back().second) {
3020 MacroStack.back().first->subPieces.push_back(piece);
3024 // We aren't in the same group. Are we descending into a new macro
3025 // or are part of an old one?
3026 std::shared_ptr<PathDiagnosticMacroPiece> MacroGroup;
3028 SourceLocation ParentInstantiationLoc = InstantiationLoc.isMacroID() ?
3029 SM.getExpansionLoc(Loc) :
3032 // Walk the entire macro stack.
3033 while (!MacroStack.empty()) {
3034 if (InstantiationLoc == MacroStack.back().second) {
3035 MacroGroup = MacroStack.back().first;
3039 if (ParentInstantiationLoc == MacroStack.back().second) {
3040 MacroGroup = MacroStack.back().first;
3044 MacroStack.pop_back();
3047 if (!MacroGroup || ParentInstantiationLoc == MacroStack.back().second) {
3048 // Create a new macro group and add it to the stack.
3049 auto NewGroup = std::make_shared<PathDiagnosticMacroPiece>(
3050 PathDiagnosticLocation::createSingleLocation(piece->getLocation()));
3053 MacroGroup->subPieces.push_back(NewGroup);
3055 assert(InstantiationLoc.isFileID());
3056 Pieces.push_back(NewGroup);
3059 MacroGroup = NewGroup;
3060 MacroStack.push_back(std::make_pair(MacroGroup, InstantiationLoc));
3063 // Finally, add the PathDiagnosticPiece to the group.
3064 MacroGroup->subPieces.push_back(piece);
3067 // Now take the pieces and construct a new PathDiagnostic.
3070 path.insert(path.end(), Pieces.begin(), Pieces.end());
3073 bool GRBugReporter::generatePathDiagnostic(PathDiagnostic& PD,
3074 PathDiagnosticConsumer &PC,
3075 ArrayRef<BugReport *> &bugReports) {
3076 assert(!bugReports.empty());
3078 bool HasValid = false;
3079 bool HasInvalid = false;
3080 SmallVector<const ExplodedNode *, 32> errorNodes;
3081 for (ArrayRef<BugReport*>::iterator I = bugReports.begin(),
3082 E = bugReports.end(); I != E; ++I) {
3083 if ((*I)->isValid()) {
3085 errorNodes.push_back((*I)->getErrorNode());
3087 // Keep the errorNodes list in sync with the bugReports list.
3089 errorNodes.push_back(nullptr);
3093 // If all the reports have been marked invalid by a previous path generation,
3098 typedef PathDiagnosticConsumer::PathGenerationScheme PathGenerationScheme;
3099 PathGenerationScheme ActiveScheme = PC.getGenerationScheme();
3101 if (ActiveScheme == PathDiagnosticConsumer::Extensive) {
3102 AnalyzerOptions &options = getAnalyzerOptions();
3103 if (options.getBooleanOption("path-diagnostics-alternate", true)) {
3104 ActiveScheme = PathDiagnosticConsumer::AlternateExtensive;
3108 TrimmedGraph TrimG(&getGraph(), errorNodes);
3109 ReportGraph ErrorGraph;
3111 while (TrimG.popNextReportGraph(ErrorGraph)) {
3112 // Find the BugReport with the original location.
3113 assert(ErrorGraph.Index < bugReports.size());
3114 BugReport *R = bugReports[ErrorGraph.Index];
3115 assert(R && "No original report found for sliced graph.");
3116 assert(R->isValid() && "Report selected by trimmed graph marked invalid.");
3118 // Start building the path diagnostic...
3119 PathDiagnosticBuilder PDB(*this, R, ErrorGraph.BackMap, &PC);
3120 const ExplodedNode *N = ErrorGraph.ErrorNode;
3122 // Register additional node visitors.
3123 R->addVisitor(llvm::make_unique<NilReceiverBRVisitor>());
3124 R->addVisitor(llvm::make_unique<ConditionBRVisitor>());
3125 R->addVisitor(llvm::make_unique<LikelyFalsePositiveSuppressionBRVisitor>());
3126 R->addVisitor(llvm::make_unique<CXXSelfAssignmentBRVisitor>());
3128 BugReport::VisitorList visitors;
3129 unsigned origReportConfigToken, finalReportConfigToken;
3130 LocationContextMap LCM;
3132 // While generating diagnostics, it's possible the visitors will decide
3133 // new symbols and regions are interesting, or add other visitors based on
3134 // the information they find. If they do, we need to regenerate the path
3135 // based on our new report configuration.
3137 // Get a clean copy of all the visitors.
3138 for (BugReport::visitor_iterator I = R->visitor_begin(),
3139 E = R->visitor_end(); I != E; ++I)
3140 visitors.push_back((*I)->clone());
3142 // Clear out the active path from any previous work.
3144 origReportConfigToken = R->getConfigurationChangeToken();
3146 // Generate the very last diagnostic piece - the piece is visible before
3147 // the trace is expanded.
3148 std::unique_ptr<PathDiagnosticPiece> LastPiece;
3149 for (BugReport::visitor_iterator I = visitors.begin(), E = visitors.end();
3151 if (std::unique_ptr<PathDiagnosticPiece> Piece =
3152 (*I)->getEndPath(PDB, N, *R)) {
3153 assert (!LastPiece &&
3154 "There can only be one final piece in a diagnostic.");
3155 LastPiece = std::move(Piece);
3159 if (ActiveScheme != PathDiagnosticConsumer::None) {
3161 LastPiece = BugReporterVisitor::getDefaultEndPath(PDB, N, *R);
3163 PD.setEndOfPath(std::move(LastPiece));
3166 // Make sure we get a clean location context map so we don't
3167 // hold onto old mappings.
3170 switch (ActiveScheme) {
3171 case PathDiagnosticConsumer::AlternateExtensive:
3172 GenerateAlternateExtensivePathDiagnostic(PD, PDB, N, LCM, visitors);
3174 case PathDiagnosticConsumer::Extensive:
3175 GenerateExtensivePathDiagnostic(PD, PDB, N, LCM, visitors);
3177 case PathDiagnosticConsumer::Minimal:
3178 GenerateMinimalPathDiagnostic(PD, PDB, N, LCM, visitors);
3180 case PathDiagnosticConsumer::None:
3181 GenerateVisitorsOnlyPathDiagnostic(PD, PDB, N, visitors);
3185 // Clean up the visitors we used.
3188 // Did anything change while generating this path?
3189 finalReportConfigToken = R->getConfigurationChangeToken();
3190 } while (finalReportConfigToken != origReportConfigToken);
3195 // Finally, prune the diagnostic path of uninteresting stuff.
3196 if (!PD.path.empty()) {
3197 if (R->shouldPrunePath() && getAnalyzerOptions().shouldPrunePaths()) {
3198 bool stillHasNotes = removeUnneededCalls(PD.getMutablePieces(), R, LCM);
3199 assert(stillHasNotes);
3200 (void)stillHasNotes;
3203 // Redirect all call pieces to have valid locations.
3204 adjustCallLocations(PD.getMutablePieces());
3205 removePiecesWithInvalidLocations(PD.getMutablePieces());
3207 if (ActiveScheme == PathDiagnosticConsumer::AlternateExtensive) {
3208 SourceManager &SM = getSourceManager();
3210 // Reduce the number of edges from a very conservative set
3211 // to an aesthetically pleasing subset that conveys the
3212 // necessary information.
3213 OptimizedCallsSet OCS;
3214 while (optimizeEdges(PD.getMutablePieces(), SM, OCS, LCM)) {}
3216 // Drop the very first function-entry edge. It's not really necessary
3217 // for top-level functions.
3218 dropFunctionEntryEdge(PD.getMutablePieces(), LCM, SM);
3221 // Remove messages that are basically the same, and edges that may not
3223 // We have to do this after edge optimization in the Extensive mode.
3224 removeRedundantMsgs(PD.getMutablePieces());
3225 removeEdgesToDefaultInitializers(PD.getMutablePieces());
3228 // We found a report and didn't suppress it.
3232 // We suppressed all the reports in this equivalence class.
3233 assert(!HasInvalid && "Inconsistent suppression");
3238 void BugReporter::Register(BugType *BT) {
3239 BugTypes = F.add(BugTypes, BT);
3242 void BugReporter::emitReport(std::unique_ptr<BugReport> R) {
3243 if (const ExplodedNode *E = R->getErrorNode()) {
3244 // An error node must either be a sink or have a tag, otherwise
3245 // it could get reclaimed before the path diagnostic is created.
3246 assert((E->isSink() || E->getLocation().getTag()) &&
3247 "Error node must either be a sink or have a tag");
3249 const AnalysisDeclContext *DeclCtx =
3250 E->getLocationContext()->getAnalysisDeclContext();
3251 // The source of autosynthesized body can be handcrafted AST or a model
3252 // file. The locations from handcrafted ASTs have no valid source locations
3253 // and have to be discarded. Locations from model files should be preserved
3254 // for processing and reporting.
3255 if (DeclCtx->isBodyAutosynthesized() &&
3256 !DeclCtx->isBodyAutosynthesizedFromModelFile())
3260 bool ValidSourceLoc = R->getLocation(getSourceManager()).isValid();
3261 assert(ValidSourceLoc);
3262 // If we mess up in a release build, we'd still prefer to just drop the bug
3263 // instead of trying to go on.
3264 if (!ValidSourceLoc)
3267 // Compute the bug report's hash to determine its equivalence class.
3268 llvm::FoldingSetNodeID ID;
3271 // Lookup the equivance class. If there isn't one, create it.
3272 BugType& BT = R->getBugType();
3275 BugReportEquivClass* EQ = EQClasses.FindNodeOrInsertPos(ID, InsertPos);
3278 EQ = new BugReportEquivClass(std::move(R));
3279 EQClasses.InsertNode(EQ, InsertPos);
3280 EQClassesVector.push_back(EQ);
3282 EQ->AddReport(std::move(R));
3286 //===----------------------------------------------------------------------===//
3287 // Emitting reports in equivalence classes.
3288 //===----------------------------------------------------------------------===//
3291 struct FRIEC_WLItem {
3292 const ExplodedNode *N;
3293 ExplodedNode::const_succ_iterator I, E;
3295 FRIEC_WLItem(const ExplodedNode *n)
3296 : N(n), I(N->succ_begin()), E(N->succ_end()) {}
3300 static const CFGBlock *findBlockForNode(const ExplodedNode *N) {
3301 ProgramPoint P = N->getLocation();
3302 if (auto BEP = P.getAs<BlockEntrance>())
3303 return BEP->getBlock();
3305 // Find the node's current statement in the CFG.
3306 if (const Stmt *S = PathDiagnosticLocation::getStmt(N))
3307 return N->getLocationContext()->getAnalysisDeclContext()
3308 ->getCFGStmtMap()->getBlock(S);
3313 // Returns true if by simply looking at the block, we can be sure that it
3314 // results in a sink during analysis. This is useful to know when the analysis
3315 // was interrupted, and we try to figure out if it would sink eventually.
3316 // There may be many more reasons why a sink would appear during analysis
3317 // (eg. checkers may generate sinks arbitrarily), but here we only consider
3318 // sinks that would be obvious by looking at the CFG.
3319 static bool isImmediateSinkBlock(const CFGBlock *Blk) {
3320 if (Blk->hasNoReturnElement())
3323 // FIXME: Throw-expressions are currently generating sinks during analysis:
3324 // they're not supported yet, and also often used for actually terminating
3325 // the program. So we should treat them as sinks in this analysis as well,
3326 // at least for now, but once we have better support for exceptions,
3327 // we'd need to carefully handle the case when the throw is being
3328 // immediately caught.
3329 if (std::any_of(Blk->begin(), Blk->end(), [](const CFGElement &Elm) {
3330 if (Optional<CFGStmt> StmtElm = Elm.getAs<CFGStmt>())
3331 if (isa<CXXThrowExpr>(StmtElm->getStmt()))
3340 // Returns true if by looking at the CFG surrounding the node's program
3341 // point, we can be sure that any analysis starting from this point would
3342 // eventually end with a sink. We scan the child CFG blocks in a depth-first
3343 // manner and see if all paths eventually end up in an immediate sink block.
3344 static bool isInevitablySinking(const ExplodedNode *N) {
3345 const CFG &Cfg = N->getCFG();
3347 const CFGBlock *StartBlk = findBlockForNode(N);
3350 if (isImmediateSinkBlock(StartBlk))
3353 llvm::SmallVector<const CFGBlock *, 32> DFSWorkList;
3354 llvm::SmallPtrSet<const CFGBlock *, 32> Visited;
3356 DFSWorkList.push_back(StartBlk);
3357 while (!DFSWorkList.empty()) {
3358 const CFGBlock *Blk = DFSWorkList.back();
3359 DFSWorkList.pop_back();
3360 Visited.insert(Blk);
3362 for (const auto &Succ : Blk->succs()) {
3363 if (const CFGBlock *SuccBlk = Succ.getReachableBlock()) {
3364 if (SuccBlk == &Cfg.getExit()) {
3365 // If at least one path reaches the CFG exit, it means that control is
3366 // returned to the caller. For now, say that we are not sure what
3367 // happens next. If necessary, this can be improved to analyze
3368 // the parent StackFrameContext's call site in a similar manner.
3372 if (!isImmediateSinkBlock(SuccBlk) && !Visited.count(SuccBlk)) {
3373 // If the block has reachable child blocks that aren't no-return,
3374 // add them to the worklist.
3375 DFSWorkList.push_back(SuccBlk);
3381 // Nothing reached the exit. It can only mean one thing: there's no return.
3386 FindReportInEquivalenceClass(BugReportEquivClass& EQ,
3387 SmallVectorImpl<BugReport*> &bugReports) {
3389 BugReportEquivClass::iterator I = EQ.begin(), E = EQ.end();
3391 BugType& BT = I->getBugType();
3393 // If we don't need to suppress any of the nodes because they are
3394 // post-dominated by a sink, simply add all the nodes in the equivalence class
3395 // to 'Nodes'. Any of the reports will serve as a "representative" report.
3396 if (!BT.isSuppressOnSink()) {
3398 for (BugReportEquivClass::iterator I=EQ.begin(), E=EQ.end(); I!=E; ++I) {
3399 const ExplodedNode *N = I->getErrorNode();
3402 bugReports.push_back(R);
3408 // For bug reports that should be suppressed when all paths are post-dominated
3409 // by a sink node, iterate through the reports in the equivalence class
3410 // until we find one that isn't post-dominated (if one exists). We use a
3411 // DFS traversal of the ExplodedGraph to find a non-sink node. We could write
3412 // this as a recursive function, but we don't want to risk blowing out the
3413 // stack for very long paths.
3414 BugReport *exampleReport = nullptr;
3416 for (; I != E; ++I) {
3417 const ExplodedNode *errorNode = I->getErrorNode();
3421 if (errorNode->isSink()) {
3423 "BugType::isSuppressSink() should not be 'true' for sink end nodes");
3425 // No successors? By definition this nodes isn't post-dominated by a sink.
3426 if (errorNode->succ_empty()) {
3427 bugReports.push_back(&*I);
3429 exampleReport = &*I;
3433 // See if we are in a no-return CFG block. If so, treat this similarly
3434 // to being post-dominated by a sink. This works better when the analysis
3435 // is incomplete and we have never reached the no-return function call(s)
3436 // that we'd inevitably bump into on this path.
3437 if (isInevitablySinking(errorNode))
3440 // At this point we know that 'N' is not a sink and it has at least one
3441 // successor. Use a DFS worklist to find a non-sink end-of-path node.
3442 typedef FRIEC_WLItem WLItem;
3443 typedef SmallVector<WLItem, 10> DFSWorkList;
3444 llvm::DenseMap<const ExplodedNode *, unsigned> Visited;
3447 WL.push_back(errorNode);
3448 Visited[errorNode] = 1;
3450 while (!WL.empty()) {
3451 WLItem &WI = WL.back();
3452 assert(!WI.N->succ_empty());
3454 for (; WI.I != WI.E; ++WI.I) {
3455 const ExplodedNode *Succ = *WI.I;
3456 // End-of-path node?
3457 if (Succ->succ_empty()) {
3458 // If we found an end-of-path node that is not a sink.
3459 if (!Succ->isSink()) {
3460 bugReports.push_back(&*I);
3462 exampleReport = &*I;
3466 // Found a sink? Continue on to the next successor.
3469 // Mark the successor as visited. If it hasn't been explored,
3470 // enqueue it to the DFS worklist.
3471 unsigned &mark = Visited[Succ];
3479 // The worklist may have been cleared at this point. First
3480 // check if it is empty before checking the last item.
3481 if (!WL.empty() && &WL.back() == &WI)
3486 // ExampleReport will be NULL if all the nodes in the equivalence class
3487 // were post-dominated by sinks.
3488 return exampleReport;
3491 void BugReporter::FlushReport(BugReportEquivClass& EQ) {
3492 SmallVector<BugReport*, 10> bugReports;
3493 BugReport *exampleReport = FindReportInEquivalenceClass(EQ, bugReports);
3494 if (exampleReport) {
3495 for (PathDiagnosticConsumer *PDC : getPathDiagnosticConsumers()) {
3496 FlushReport(exampleReport, *PDC, bugReports);
3501 void BugReporter::FlushReport(BugReport *exampleReport,
3502 PathDiagnosticConsumer &PD,
3503 ArrayRef<BugReport*> bugReports) {
3505 // FIXME: Make sure we use the 'R' for the path that was actually used.
3506 // Probably doesn't make a difference in practice.
3507 BugType& BT = exampleReport->getBugType();
3509 std::unique_ptr<PathDiagnostic> D(new PathDiagnostic(
3510 exampleReport->getBugType().getCheckName(),
3511 exampleReport->getDeclWithIssue(), exampleReport->getBugType().getName(),
3512 exampleReport->getDescription(),
3513 exampleReport->getShortDescription(/*Fallback=*/false), BT.getCategory(),
3514 exampleReport->getUniqueingLocation(),
3515 exampleReport->getUniqueingDecl()));
3517 if (exampleReport->isPathSensitive()) {
3518 // Generate the full path diagnostic, using the generation scheme
3519 // specified by the PathDiagnosticConsumer. Note that we have to generate
3520 // path diagnostics even for consumers which do not support paths, because
3521 // the BugReporterVisitors may mark this bug as a false positive.
3522 assert(!bugReports.empty());
3524 MaxBugClassSize.updateMax(bugReports.size());
3526 if (!generatePathDiagnostic(*D.get(), PD, bugReports))
3529 MaxValidBugClassSize.updateMax(bugReports.size());
3531 // Examine the report and see if the last piece is in a header. Reset the
3532 // report location to the last piece in the main source file.
3533 AnalyzerOptions &Opts = getAnalyzerOptions();
3534 if (Opts.shouldReportIssuesInMainSourceFile() && !Opts.AnalyzeAll)
3535 D->resetDiagnosticLocationToMainFile();
3538 // If the path is empty, generate a single step path with the location
3540 if (D->path.empty()) {
3541 PathDiagnosticLocation L = exampleReport->getLocation(getSourceManager());
3542 auto piece = llvm::make_unique<PathDiagnosticEventPiece>(
3543 L, exampleReport->getDescription());
3544 for (SourceRange Range : exampleReport->getRanges())
3545 piece->addRange(Range);
3546 D->setEndOfPath(std::move(piece));
3549 PathPieces &Pieces = D->getMutablePieces();
3550 if (getAnalyzerOptions().shouldDisplayNotesAsEvents()) {
3551 // For path diagnostic consumers that don't support extra notes,
3552 // we may optionally convert those to path notes.
3553 for (auto I = exampleReport->getNotes().rbegin(),
3554 E = exampleReport->getNotes().rend(); I != E; ++I) {
3555 PathDiagnosticNotePiece *Piece = I->get();
3556 auto ConvertedPiece = std::make_shared<PathDiagnosticEventPiece>(
3557 Piece->getLocation(), Piece->getString());
3558 for (const auto &R: Piece->getRanges())
3559 ConvertedPiece->addRange(R);
3561 Pieces.push_front(std::move(ConvertedPiece));
3564 for (auto I = exampleReport->getNotes().rbegin(),
3565 E = exampleReport->getNotes().rend(); I != E; ++I)
3566 Pieces.push_front(*I);
3569 // Get the meta data.
3570 const BugReport::ExtraTextList &Meta = exampleReport->getExtraText();
3571 for (BugReport::ExtraTextList::const_iterator i = Meta.begin(),
3572 e = Meta.end(); i != e; ++i) {
3576 PD.HandlePathDiagnostic(std::move(D));
3579 void BugReporter::EmitBasicReport(const Decl *DeclWithIssue,
3580 const CheckerBase *Checker,
3581 StringRef Name, StringRef Category,
3582 StringRef Str, PathDiagnosticLocation Loc,
3583 ArrayRef<SourceRange> Ranges) {
3584 EmitBasicReport(DeclWithIssue, Checker->getCheckName(), Name, Category, Str,
3587 void BugReporter::EmitBasicReport(const Decl *DeclWithIssue,
3588 CheckName CheckName,
3589 StringRef name, StringRef category,
3590 StringRef str, PathDiagnosticLocation Loc,
3591 ArrayRef<SourceRange> Ranges) {
3593 // 'BT' is owned by BugReporter.
3594 BugType *BT = getBugTypeForName(CheckName, name, category);
3595 auto R = llvm::make_unique<BugReport>(*BT, str, Loc);
3596 R->setDeclWithIssue(DeclWithIssue);
3597 for (ArrayRef<SourceRange>::iterator I = Ranges.begin(), E = Ranges.end();
3600 emitReport(std::move(R));
3603 BugType *BugReporter::getBugTypeForName(CheckName CheckName, StringRef name,
3604 StringRef category) {
3605 SmallString<136> fullDesc;
3606 llvm::raw_svector_ostream(fullDesc) << CheckName.getName() << ":" << name
3608 BugType *&BT = StrBugTypes[fullDesc];
3610 BT = new BugType(CheckName, name, category);
3614 LLVM_DUMP_METHOD void PathPieces::dump() const {
3616 for (PathPieces::const_iterator I = begin(), E = end(); I != E; ++I) {
3617 llvm::errs() << "[" << index++ << "] ";
3619 llvm::errs() << "\n";
3623 LLVM_DUMP_METHOD void PathDiagnosticCallPiece::dump() const {
3624 llvm::errs() << "CALL\n--------------\n";
3626 if (const Stmt *SLoc = getLocStmt(getLocation()))
3628 else if (const NamedDecl *ND = dyn_cast<NamedDecl>(getCallee()))
3629 llvm::errs() << *ND << "\n";
3631 getLocation().dump();
3634 LLVM_DUMP_METHOD void PathDiagnosticEventPiece::dump() const {
3635 llvm::errs() << "EVENT\n--------------\n";
3636 llvm::errs() << getString() << "\n";
3637 llvm::errs() << " ---- at ----\n";
3638 getLocation().dump();
3641 LLVM_DUMP_METHOD void PathDiagnosticControlFlowPiece::dump() const {
3642 llvm::errs() << "CONTROL\n--------------\n";
3643 getStartLocation().dump();
3644 llvm::errs() << " ---- to ----\n";
3645 getEndLocation().dump();
3648 LLVM_DUMP_METHOD void PathDiagnosticMacroPiece::dump() const {
3649 llvm::errs() << "MACRO\n--------------\n";
3650 // FIXME: Print which macro is being invoked.
3653 LLVM_DUMP_METHOD void PathDiagnosticNotePiece::dump() const {
3654 llvm::errs() << "NOTE\n--------------\n";
3655 llvm::errs() << getString() << "\n";
3656 llvm::errs() << " ---- at ----\n";
3657 getLocation().dump();
3660 LLVM_DUMP_METHOD void PathDiagnosticLocation::dump() const {
3662 llvm::errs() << "<INVALID>\n";
3668 // FIXME: actually print the range.
3669 llvm::errs() << "<range>\n";
3672 asLocation().dump();
3673 llvm::errs() << "\n";
3679 llvm::errs() << "<NULL STMT>\n";
3682 if (const NamedDecl *ND = dyn_cast_or_null<NamedDecl>(D))
3683 llvm::errs() << *ND << "\n";
3684 else if (isa<BlockDecl>(D))
3685 // FIXME: Make this nicer.
3686 llvm::errs() << "<block>\n";
3688 llvm::errs() << "<unknown decl>\n";
3690 llvm::errs() << "<NULL DECL>\n";