1 // BugReporter.cpp - Generate PathDiagnostics for Bugs ------------*- C++ -*--// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines BugReporter, a utility class for generating 11 // PathDiagnostics. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "clang/StaticAnalyzer/Core/BugReporter/BugReporter.h" 16 #include "clang/AST/ASTContext.h" 17 #include "clang/AST/DeclObjC.h" 18 #include "clang/AST/Expr.h" 19 #include "clang/AST/ExprCXX.h" 20 #include "clang/AST/ParentMap.h" 21 #include "clang/AST/StmtCXX.h" 22 #include "clang/AST/StmtObjC.h" 23 #include "clang/Analysis/CFG.h" 24 #include "clang/Analysis/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" 36 #include <memory> 37 #include <queue> 38 39 using namespace clang; 40 using namespace ento; 41 42 #define DEBUG_TYPE "BugReporter" 43 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)"); 49 50 BugReporterVisitor::~BugReporterVisitor() {} 51 52 void BugReporterContext::anchor() {} 53 54 //===----------------------------------------------------------------------===// 55 // Helper routines for walking the ExplodedGraph and fetching statements. 56 //===----------------------------------------------------------------------===// 57 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)) 61 return S; 62 63 return nullptr; 64 } 65 66 static inline const Stmt* 67 GetCurrentOrPreviousStmt(const ExplodedNode *N) { 68 if (const Stmt *S = PathDiagnosticLocation::getStmt(N)) 69 return S; 70 71 return GetPreviousStmt(N); 72 } 73 74 //===----------------------------------------------------------------------===// 75 // Diagnostic cleanup. 76 //===----------------------------------------------------------------------===// 77 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(); 87 88 if (X->getLocation() != Y->getLocation()) 89 return nullptr; 90 91 if (X->getTag() == tagPreferred && Y->getTag() == tagLesser) 92 return ConditionBRVisitor::isPieceMessageGeneric(X) ? Y : X; 93 94 if (Y->getTag() == tagPreferred && X->getTag() == tagLesser) 95 return ConditionBRVisitor::isPieceMessageGeneric(Y) ? X : Y; 96 97 return nullptr; 98 } 99 100 /// 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(); 107 if (N < 2) 108 return; 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 IntrusiveRefCntPtr<PathDiagnosticPiece> piece(path.front()); 115 path.pop_front(); 116 117 switch (piece->getKind()) { 118 case PathDiagnosticPiece::Call: 119 removeRedundantMsgs(cast<PathDiagnosticCallPiece>(piece)->path); 120 break; 121 case PathDiagnosticPiece::Macro: 122 removeRedundantMsgs(cast<PathDiagnosticMacroPiece>(piece)->subPieces); 123 break; 124 case PathDiagnosticPiece::ControlFlow: 125 break; 126 case PathDiagnosticPiece::Event: { 127 if (i == N-1) 128 break; 129 130 if (PathDiagnosticEventPiece *nextEvent = 131 dyn_cast<PathDiagnosticEventPiece>(path.front().get())) { 132 PathDiagnosticEventPiece *event = 133 cast<PathDiagnosticEventPiece>(piece); 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 (PathDiagnosticEventPiece *pieceToKeep = 138 eventsDescribeSameCondition(event, nextEvent)) { 139 piece = pieceToKeep; 140 path.pop_front(); 141 ++i; 142 } 143 } 144 break; 145 } 146 case PathDiagnosticPiece::Note: 147 break; 148 } 149 path.push_back(piece); 150 } 151 } 152 153 /// A map from PathDiagnosticPiece to the LocationContext of the inlined 154 /// function call it represents. 155 typedef llvm::DenseMap<const PathPieces *, const LocationContext *> 156 LocationContextMap; 157 158 /// 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(); 165 166 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 IntrusiveRefCntPtr<PathDiagnosticPiece> piece(pieces.front()); 170 pieces.pop_front(); 171 172 switch (piece->getKind()) { 173 case PathDiagnosticPiece::Call: { 174 PathDiagnosticCallPiece *call = cast<PathDiagnosticCallPiece>(piece); 175 // Check if the location context is interesting. 176 assert(LCM.count(&call->path)); 177 if (R->isInteresting(LCM[&call->path])) { 178 containsSomethingInteresting = true; 179 break; 180 } 181 182 if (!removeUnneededCalls(call->path, R, LCM)) 183 continue; 184 185 containsSomethingInteresting = true; 186 break; 187 } 188 case PathDiagnosticPiece::Macro: { 189 PathDiagnosticMacroPiece *macro = cast<PathDiagnosticMacroPiece>(piece); 190 if (!removeUnneededCalls(macro->subPieces, R, LCM)) 191 continue; 192 containsSomethingInteresting = true; 193 break; 194 } 195 case PathDiagnosticPiece::Event: { 196 PathDiagnosticEventPiece *event = cast<PathDiagnosticEventPiece>(piece); 197 198 // We never throw away an event, but we do throw it away wholesale 199 // as part of a path if we throw the entire path away. 200 containsSomethingInteresting |= !event->isPrunable(); 201 break; 202 } 203 case PathDiagnosticPiece::ControlFlow: 204 break; 205 206 case PathDiagnosticPiece::Note: 207 break; 208 } 209 210 pieces.push_back(piece); 211 } 212 213 return containsSomethingInteresting; 214 } 215 216 /// 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) { 219 assert(D); 220 return D->isImplicit() || !D->hasBody(); 221 } 222 223 /// Recursively scan through a path and make sure that all call pieces have 224 /// valid locations. 225 static void 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); 230 231 if (!Call) { 232 assert((*I)->getLocation().asLocation().isValid()); 233 continue; 234 } 235 236 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; 242 } 243 244 // 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; 250 else 251 ThisCallLocation = &Call->callEnter; 252 253 assert(ThisCallLocation && "Outermost call has an invalid location"); 254 adjustCallLocations(Call->path, ThisCallLocation); 255 } 256 } 257 258 /// 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 (PathDiagnosticCallPiece *C = dyn_cast<PathDiagnosticCallPiece>(*I)) 264 removeEdgesToDefaultInitializers(C->path); 265 266 if (PathDiagnosticMacroPiece *M = dyn_cast<PathDiagnosticMacroPiece>(*I)) 267 removeEdgesToDefaultInitializers(M->subPieces); 268 269 if (PathDiagnosticControlFlowPiece *CF = 270 dyn_cast<PathDiagnosticControlFlowPiece>(*I)) { 271 const Stmt *Start = CF->getStartLocation().asStmt(); 272 const Stmt *End = CF->getEndLocation().asStmt(); 273 if (Start && isa<CXXDefaultInitExpr>(Start)) { 274 I = Pieces.erase(I); 275 continue; 276 } else if (End && isa<CXXDefaultInitExpr>(End)) { 277 PathPieces::iterator Next = std::next(I); 278 if (Next != E) { 279 if (PathDiagnosticControlFlowPiece *NextCF = 280 dyn_cast<PathDiagnosticControlFlowPiece>(*Next)) { 281 NextCF->setStartLocation(CF->getStartLocation()); 282 } 283 } 284 I = Pieces.erase(I); 285 continue; 286 } 287 } 288 289 I++; 290 } 291 } 292 293 /// Remove all pieces with invalid locations as these cannot be serialized. 294 /// We might have pieces with invalid locations as a result of inlining Body 295 /// Farm generated functions. 296 static void removePiecesWithInvalidLocations(PathPieces &Pieces) { 297 for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E;) { 298 if (PathDiagnosticCallPiece *C = dyn_cast<PathDiagnosticCallPiece>(*I)) 299 removePiecesWithInvalidLocations(C->path); 300 301 if (PathDiagnosticMacroPiece *M = dyn_cast<PathDiagnosticMacroPiece>(*I)) 302 removePiecesWithInvalidLocations(M->subPieces); 303 304 if (!(*I)->getLocation().isValid() || 305 !(*I)->getLocation().asLocation().isValid()) { 306 I = Pieces.erase(I); 307 continue; 308 } 309 I++; 310 } 311 } 312 313 //===----------------------------------------------------------------------===// 314 // PathDiagnosticBuilder and its associated routines and helper objects. 315 //===----------------------------------------------------------------------===// 316 317 namespace { 318 class NodeMapClosure : public BugReport::NodeResolver { 319 InterExplodedGraphMap &M; 320 public: 321 NodeMapClosure(InterExplodedGraphMap &m) : M(m) {} 322 323 const ExplodedNode *getOriginalNode(const ExplodedNode *N) override { 324 return M.lookup(N); 325 } 326 }; 327 328 class PathDiagnosticBuilder : public BugReporterContext { 329 BugReport *R; 330 PathDiagnosticConsumer *PDC; 331 NodeMapClosure NMC; 332 public: 333 const LocationContext *LC; 334 335 PathDiagnosticBuilder(GRBugReporter &br, 336 BugReport *r, InterExplodedGraphMap &Backmap, 337 PathDiagnosticConsumer *pdc) 338 : BugReporterContext(br), 339 R(r), PDC(pdc), NMC(Backmap), LC(r->getErrorNode()->getLocationContext()) 340 {} 341 342 PathDiagnosticLocation ExecutionContinues(const ExplodedNode *N); 343 344 PathDiagnosticLocation ExecutionContinues(llvm::raw_string_ostream &os, 345 const ExplodedNode *N); 346 347 BugReport *getBugReport() { return R; } 348 349 Decl const &getCodeDecl() { return R->getErrorNode()->getCodeDecl(); } 350 351 ParentMap& getParentMap() { return LC->getParentMap(); } 352 353 const Stmt *getParent(const Stmt *S) { 354 return getParentMap().getParent(S); 355 } 356 357 NodeMapClosure& getNodeResolver() override { return NMC; } 358 359 PathDiagnosticLocation getEnclosingStmtLocation(const Stmt *S); 360 361 PathDiagnosticConsumer::PathGenerationScheme getGenerationScheme() const { 362 return PDC ? PDC->getGenerationScheme() : PathDiagnosticConsumer::Extensive; 363 } 364 365 bool supportsLogicalOpControlFlow() const { 366 return PDC ? PDC->supportsLogicalOpControlFlow() : true; 367 } 368 }; 369 } // end anonymous namespace 370 371 PathDiagnosticLocation 372 PathDiagnosticBuilder::ExecutionContinues(const ExplodedNode *N) { 373 if (const Stmt *S = PathDiagnosticLocation::getNextStmt(N)) 374 return PathDiagnosticLocation(S, getSourceManager(), LC); 375 376 return PathDiagnosticLocation::createDeclEnd(N->getLocationContext(), 377 getSourceManager()); 378 } 379 380 PathDiagnosticLocation 381 PathDiagnosticBuilder::ExecutionContinues(llvm::raw_string_ostream &os, 382 const ExplodedNode *N) { 383 384 // Slow, but probably doesn't matter. 385 if (os.str().empty()) 386 os << ' '; 387 388 const PathDiagnosticLocation &Loc = ExecutionContinues(N); 389 390 if (Loc.asStmt()) 391 os << "Execution continues on line " 392 << getSourceManager().getExpansionLineNumber(Loc.asLocation()) 393 << '.'; 394 else { 395 os << "Execution jumps to the end of the "; 396 const Decl *D = N->getLocationContext()->getDecl(); 397 if (isa<ObjCMethodDecl>(D)) 398 os << "method"; 399 else if (isa<FunctionDecl>(D)) 400 os << "function"; 401 else { 402 assert(isa<BlockDecl>(D)); 403 os << "anonymous block"; 404 } 405 os << '.'; 406 } 407 408 return Loc; 409 } 410 411 static const Stmt *getEnclosingParent(const Stmt *S, const ParentMap &PM) { 412 if (isa<Expr>(S) && PM.isConsumedExpr(cast<Expr>(S))) 413 return PM.getParentIgnoreParens(S); 414 415 const Stmt *Parent = PM.getParentIgnoreParens(S); 416 if (!Parent) 417 return nullptr; 418 419 switch (Parent->getStmtClass()) { 420 case Stmt::ForStmtClass: 421 case Stmt::DoStmtClass: 422 case Stmt::WhileStmtClass: 423 case Stmt::ObjCForCollectionStmtClass: 424 case Stmt::CXXForRangeStmtClass: 425 return Parent; 426 default: 427 break; 428 } 429 430 return nullptr; 431 } 432 433 static PathDiagnosticLocation 434 getEnclosingStmtLocation(const Stmt *S, SourceManager &SMgr, const ParentMap &P, 435 const LocationContext *LC, bool allowNestedContexts) { 436 if (!S) 437 return PathDiagnosticLocation(); 438 439 while (const Stmt *Parent = getEnclosingParent(S, P)) { 440 switch (Parent->getStmtClass()) { 441 case Stmt::BinaryOperatorClass: { 442 const BinaryOperator *B = cast<BinaryOperator>(Parent); 443 if (B->isLogicalOp()) 444 return PathDiagnosticLocation(allowNestedContexts ? B : S, SMgr, LC); 445 break; 446 } 447 case Stmt::CompoundStmtClass: 448 case Stmt::StmtExprClass: 449 return PathDiagnosticLocation(S, SMgr, LC); 450 case Stmt::ChooseExprClass: 451 // Similar to '?' if we are referring to condition, just have the edge 452 // point to the entire choose expression. 453 if (allowNestedContexts || cast<ChooseExpr>(Parent)->getCond() == S) 454 return PathDiagnosticLocation(Parent, SMgr, LC); 455 else 456 return PathDiagnosticLocation(S, SMgr, LC); 457 case Stmt::BinaryConditionalOperatorClass: 458 case Stmt::ConditionalOperatorClass: 459 // For '?', if we are referring to condition, just have the edge point 460 // to the entire '?' expression. 461 if (allowNestedContexts || 462 cast<AbstractConditionalOperator>(Parent)->getCond() == S) 463 return PathDiagnosticLocation(Parent, SMgr, LC); 464 else 465 return PathDiagnosticLocation(S, SMgr, LC); 466 case Stmt::CXXForRangeStmtClass: 467 if (cast<CXXForRangeStmt>(Parent)->getBody() == S) 468 return PathDiagnosticLocation(S, SMgr, LC); 469 break; 470 case Stmt::DoStmtClass: 471 return PathDiagnosticLocation(S, SMgr, LC); 472 case Stmt::ForStmtClass: 473 if (cast<ForStmt>(Parent)->getBody() == S) 474 return PathDiagnosticLocation(S, SMgr, LC); 475 break; 476 case Stmt::IfStmtClass: 477 if (cast<IfStmt>(Parent)->getCond() != S) 478 return PathDiagnosticLocation(S, SMgr, LC); 479 break; 480 case Stmt::ObjCForCollectionStmtClass: 481 if (cast<ObjCForCollectionStmt>(Parent)->getBody() == S) 482 return PathDiagnosticLocation(S, SMgr, LC); 483 break; 484 case Stmt::WhileStmtClass: 485 if (cast<WhileStmt>(Parent)->getCond() != S) 486 return PathDiagnosticLocation(S, SMgr, LC); 487 break; 488 default: 489 break; 490 } 491 492 S = Parent; 493 } 494 495 assert(S && "Cannot have null Stmt for PathDiagnosticLocation"); 496 497 return PathDiagnosticLocation(S, SMgr, LC); 498 } 499 500 PathDiagnosticLocation 501 PathDiagnosticBuilder::getEnclosingStmtLocation(const Stmt *S) { 502 assert(S && "Null Stmt passed to getEnclosingStmtLocation"); 503 return ::getEnclosingStmtLocation(S, getSourceManager(), getParentMap(), LC, 504 /*allowNestedContexts=*/false); 505 } 506 507 //===----------------------------------------------------------------------===// 508 // "Visitors only" path diagnostic generation algorithm. 509 //===----------------------------------------------------------------------===// 510 static bool GenerateVisitorsOnlyPathDiagnostic( 511 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N, 512 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) { 513 // All path generation skips the very first node (the error node). 514 // This is because there is special handling for the end-of-path note. 515 N = N->getFirstPred(); 516 if (!N) 517 return true; 518 519 BugReport *R = PDB.getBugReport(); 520 while (const ExplodedNode *Pred = N->getFirstPred()) { 521 for (auto &V : visitors) { 522 // Visit all the node pairs, but throw the path pieces away. 523 PathDiagnosticPiece *Piece = V->VisitNode(N, Pred, PDB, *R); 524 delete Piece; 525 } 526 527 N = Pred; 528 } 529 530 return R->isValid(); 531 } 532 533 //===----------------------------------------------------------------------===// 534 // "Minimal" path diagnostic generation algorithm. 535 //===----------------------------------------------------------------------===// 536 typedef std::pair<PathDiagnosticCallPiece*, const ExplodedNode*> StackDiagPair; 537 typedef SmallVector<StackDiagPair, 6> StackDiagVector; 538 539 static void updateStackPiecesWithMessage(PathDiagnosticPiece *P, 540 StackDiagVector &CallStack) { 541 // If the piece contains a special message, add it to all the call 542 // pieces on the active stack. 543 if (PathDiagnosticEventPiece *ep = 544 dyn_cast<PathDiagnosticEventPiece>(P)) { 545 546 if (ep->hasCallStackHint()) 547 for (StackDiagVector::iterator I = CallStack.begin(), 548 E = CallStack.end(); I != E; ++I) { 549 PathDiagnosticCallPiece *CP = I->first; 550 const ExplodedNode *N = I->second; 551 std::string stackMsg = ep->getCallStackMessage(N); 552 553 // The last message on the path to final bug is the most important 554 // one. Since we traverse the path backwards, do not add the message 555 // if one has been previously added. 556 if (!CP->hasCallStackMessage()) 557 CP->setCallStackMessage(stackMsg); 558 } 559 } 560 } 561 562 static void CompactPathDiagnostic(PathPieces &path, const SourceManager& SM); 563 564 static bool GenerateMinimalPathDiagnostic( 565 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N, 566 LocationContextMap &LCM, 567 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) { 568 569 SourceManager& SMgr = PDB.getSourceManager(); 570 const LocationContext *LC = PDB.LC; 571 const ExplodedNode *NextNode = N->pred_empty() 572 ? nullptr : *(N->pred_begin()); 573 574 StackDiagVector CallStack; 575 576 while (NextNode) { 577 N = NextNode; 578 PDB.LC = N->getLocationContext(); 579 NextNode = N->getFirstPred(); 580 581 ProgramPoint P = N->getLocation(); 582 583 do { 584 if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) { 585 PathDiagnosticCallPiece *C = 586 PathDiagnosticCallPiece::construct(N, *CE, SMgr); 587 // Record the mapping from call piece to LocationContext. 588 LCM[&C->path] = CE->getCalleeContext(); 589 PD.getActivePath().push_front(C); 590 PD.pushActivePath(&C->path); 591 CallStack.push_back(StackDiagPair(C, N)); 592 break; 593 } 594 595 if (Optional<CallEnter> CE = P.getAs<CallEnter>()) { 596 // Flush all locations, and pop the active path. 597 bool VisitedEntireCall = PD.isWithinCall(); 598 PD.popActivePath(); 599 600 // Either we just added a bunch of stuff to the top-level path, or 601 // we have a previous CallExitEnd. If the former, it means that the 602 // path terminated within a function call. We must then take the 603 // current contents of the active path and place it within 604 // a new PathDiagnosticCallPiece. 605 PathDiagnosticCallPiece *C; 606 if (VisitedEntireCall) { 607 C = cast<PathDiagnosticCallPiece>(PD.getActivePath().front()); 608 } else { 609 const Decl *Caller = CE->getLocationContext()->getDecl(); 610 C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller); 611 // Record the mapping from call piece to LocationContext. 612 LCM[&C->path] = CE->getCalleeContext(); 613 } 614 615 C->setCallee(*CE, SMgr); 616 if (!CallStack.empty()) { 617 assert(CallStack.back().first == C); 618 CallStack.pop_back(); 619 } 620 break; 621 } 622 623 if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) { 624 const CFGBlock *Src = BE->getSrc(); 625 const CFGBlock *Dst = BE->getDst(); 626 const Stmt *T = Src->getTerminator(); 627 628 if (!T) 629 break; 630 631 PathDiagnosticLocation Start = 632 PathDiagnosticLocation::createBegin(T, SMgr, 633 N->getLocationContext()); 634 635 switch (T->getStmtClass()) { 636 default: 637 break; 638 639 case Stmt::GotoStmtClass: 640 case Stmt::IndirectGotoStmtClass: { 641 const Stmt *S = PathDiagnosticLocation::getNextStmt(N); 642 643 if (!S) 644 break; 645 646 std::string sbuf; 647 llvm::raw_string_ostream os(sbuf); 648 const PathDiagnosticLocation &End = PDB.getEnclosingStmtLocation(S); 649 650 os << "Control jumps to line " 651 << End.asLocation().getExpansionLineNumber(); 652 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 653 Start, End, os.str())); 654 break; 655 } 656 657 case Stmt::SwitchStmtClass: { 658 // Figure out what case arm we took. 659 std::string sbuf; 660 llvm::raw_string_ostream os(sbuf); 661 662 if (const Stmt *S = Dst->getLabel()) { 663 PathDiagnosticLocation End(S, SMgr, LC); 664 665 switch (S->getStmtClass()) { 666 default: 667 os << "No cases match in the switch statement. " 668 "Control jumps to line " 669 << End.asLocation().getExpansionLineNumber(); 670 break; 671 case Stmt::DefaultStmtClass: 672 os << "Control jumps to the 'default' case at line " 673 << End.asLocation().getExpansionLineNumber(); 674 break; 675 676 case Stmt::CaseStmtClass: { 677 os << "Control jumps to 'case "; 678 const CaseStmt *Case = cast<CaseStmt>(S); 679 const Expr *LHS = Case->getLHS()->IgnoreParenCasts(); 680 681 // Determine if it is an enum. 682 bool GetRawInt = true; 683 684 if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(LHS)) { 685 // FIXME: Maybe this should be an assertion. Are there cases 686 // were it is not an EnumConstantDecl? 687 const EnumConstantDecl *D = 688 dyn_cast<EnumConstantDecl>(DR->getDecl()); 689 690 if (D) { 691 GetRawInt = false; 692 os << *D; 693 } 694 } 695 696 if (GetRawInt) 697 os << LHS->EvaluateKnownConstInt(PDB.getASTContext()); 698 699 os << ":' at line " 700 << End.asLocation().getExpansionLineNumber(); 701 break; 702 } 703 } 704 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 705 Start, End, os.str())); 706 } 707 else { 708 os << "'Default' branch taken. "; 709 const PathDiagnosticLocation &End = PDB.ExecutionContinues(os, N); 710 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 711 Start, End, os.str())); 712 } 713 714 break; 715 } 716 717 case Stmt::BreakStmtClass: 718 case Stmt::ContinueStmtClass: { 719 std::string sbuf; 720 llvm::raw_string_ostream os(sbuf); 721 PathDiagnosticLocation End = PDB.ExecutionContinues(os, N); 722 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 723 Start, End, os.str())); 724 break; 725 } 726 727 // Determine control-flow for ternary '?'. 728 case Stmt::BinaryConditionalOperatorClass: 729 case Stmt::ConditionalOperatorClass: { 730 std::string sbuf; 731 llvm::raw_string_ostream os(sbuf); 732 os << "'?' condition is "; 733 734 if (*(Src->succ_begin()+1) == Dst) 735 os << "false"; 736 else 737 os << "true"; 738 739 PathDiagnosticLocation End = PDB.ExecutionContinues(N); 740 741 if (const Stmt *S = End.asStmt()) 742 End = PDB.getEnclosingStmtLocation(S); 743 744 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 745 Start, End, os.str())); 746 break; 747 } 748 749 // Determine control-flow for short-circuited '&&' and '||'. 750 case Stmt::BinaryOperatorClass: { 751 if (!PDB.supportsLogicalOpControlFlow()) 752 break; 753 754 const BinaryOperator *B = cast<BinaryOperator>(T); 755 std::string sbuf; 756 llvm::raw_string_ostream os(sbuf); 757 os << "Left side of '"; 758 759 if (B->getOpcode() == BO_LAnd) { 760 os << "&&" << "' is "; 761 762 if (*(Src->succ_begin()+1) == Dst) { 763 os << "false"; 764 PathDiagnosticLocation End(B->getLHS(), SMgr, LC); 765 PathDiagnosticLocation Start = 766 PathDiagnosticLocation::createOperatorLoc(B, SMgr); 767 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 768 Start, End, os.str())); 769 } 770 else { 771 os << "true"; 772 PathDiagnosticLocation Start(B->getLHS(), SMgr, LC); 773 PathDiagnosticLocation End = PDB.ExecutionContinues(N); 774 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 775 Start, End, os.str())); 776 } 777 } 778 else { 779 assert(B->getOpcode() == BO_LOr); 780 os << "||" << "' is "; 781 782 if (*(Src->succ_begin()+1) == Dst) { 783 os << "false"; 784 PathDiagnosticLocation Start(B->getLHS(), SMgr, LC); 785 PathDiagnosticLocation End = PDB.ExecutionContinues(N); 786 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 787 Start, End, os.str())); 788 } 789 else { 790 os << "true"; 791 PathDiagnosticLocation End(B->getLHS(), SMgr, LC); 792 PathDiagnosticLocation Start = 793 PathDiagnosticLocation::createOperatorLoc(B, SMgr); 794 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 795 Start, End, os.str())); 796 } 797 } 798 799 break; 800 } 801 802 case Stmt::DoStmtClass: { 803 if (*(Src->succ_begin()) == Dst) { 804 std::string sbuf; 805 llvm::raw_string_ostream os(sbuf); 806 807 os << "Loop condition is true. "; 808 PathDiagnosticLocation End = PDB.ExecutionContinues(os, N); 809 810 if (const Stmt *S = End.asStmt()) 811 End = PDB.getEnclosingStmtLocation(S); 812 813 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 814 Start, End, os.str())); 815 } 816 else { 817 PathDiagnosticLocation End = PDB.ExecutionContinues(N); 818 819 if (const Stmt *S = End.asStmt()) 820 End = PDB.getEnclosingStmtLocation(S); 821 822 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 823 Start, End, "Loop condition is false. Exiting loop")); 824 } 825 826 break; 827 } 828 829 case Stmt::WhileStmtClass: 830 case Stmt::ForStmtClass: { 831 if (*(Src->succ_begin()+1) == Dst) { 832 std::string sbuf; 833 llvm::raw_string_ostream os(sbuf); 834 835 os << "Loop condition is false. "; 836 PathDiagnosticLocation End = PDB.ExecutionContinues(os, N); 837 if (const Stmt *S = End.asStmt()) 838 End = PDB.getEnclosingStmtLocation(S); 839 840 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 841 Start, End, os.str())); 842 } 843 else { 844 PathDiagnosticLocation End = PDB.ExecutionContinues(N); 845 if (const Stmt *S = End.asStmt()) 846 End = PDB.getEnclosingStmtLocation(S); 847 848 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 849 Start, End, "Loop condition is true. Entering loop body")); 850 } 851 852 break; 853 } 854 855 case Stmt::IfStmtClass: { 856 PathDiagnosticLocation End = PDB.ExecutionContinues(N); 857 858 if (const Stmt *S = End.asStmt()) 859 End = PDB.getEnclosingStmtLocation(S); 860 861 if (*(Src->succ_begin()+1) == Dst) 862 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 863 Start, End, "Taking false branch")); 864 else 865 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece( 866 Start, End, "Taking true branch")); 867 868 break; 869 } 870 } 871 } 872 } while(0); 873 874 if (NextNode) { 875 // Add diagnostic pieces from custom visitors. 876 BugReport *R = PDB.getBugReport(); 877 for (auto &V : visitors) { 878 if (PathDiagnosticPiece *p = V->VisitNode(N, NextNode, PDB, *R)) { 879 PD.getActivePath().push_front(p); 880 updateStackPiecesWithMessage(p, CallStack); 881 } 882 } 883 } 884 } 885 886 if (!PDB.getBugReport()->isValid()) 887 return false; 888 889 // After constructing the full PathDiagnostic, do a pass over it to compact 890 // PathDiagnosticPieces that occur within a macro. 891 CompactPathDiagnostic(PD.getMutablePieces(), PDB.getSourceManager()); 892 return true; 893 } 894 895 //===----------------------------------------------------------------------===// 896 // "Extensive" PathDiagnostic generation. 897 //===----------------------------------------------------------------------===// 898 899 static bool IsControlFlowExpr(const Stmt *S) { 900 const Expr *E = dyn_cast<Expr>(S); 901 902 if (!E) 903 return false; 904 905 E = E->IgnoreParenCasts(); 906 907 if (isa<AbstractConditionalOperator>(E)) 908 return true; 909 910 if (const BinaryOperator *B = dyn_cast<BinaryOperator>(E)) 911 if (B->isLogicalOp()) 912 return true; 913 914 return false; 915 } 916 917 namespace { 918 class ContextLocation : public PathDiagnosticLocation { 919 bool IsDead; 920 public: 921 ContextLocation(const PathDiagnosticLocation &L, bool isdead = false) 922 : PathDiagnosticLocation(L), IsDead(isdead) {} 923 924 void markDead() { IsDead = true; } 925 bool isDead() const { return IsDead; } 926 }; 927 928 static PathDiagnosticLocation cleanUpLocation(PathDiagnosticLocation L, 929 const LocationContext *LC, 930 bool firstCharOnly = false) { 931 if (const Stmt *S = L.asStmt()) { 932 const Stmt *Original = S; 933 while (1) { 934 // Adjust the location for some expressions that are best referenced 935 // by one of their subexpressions. 936 switch (S->getStmtClass()) { 937 default: 938 break; 939 case Stmt::ParenExprClass: 940 case Stmt::GenericSelectionExprClass: 941 S = cast<Expr>(S)->IgnoreParens(); 942 firstCharOnly = true; 943 continue; 944 case Stmt::BinaryConditionalOperatorClass: 945 case Stmt::ConditionalOperatorClass: 946 S = cast<AbstractConditionalOperator>(S)->getCond(); 947 firstCharOnly = true; 948 continue; 949 case Stmt::ChooseExprClass: 950 S = cast<ChooseExpr>(S)->getCond(); 951 firstCharOnly = true; 952 continue; 953 case Stmt::BinaryOperatorClass: 954 S = cast<BinaryOperator>(S)->getLHS(); 955 firstCharOnly = true; 956 continue; 957 } 958 959 break; 960 } 961 962 if (S != Original) 963 L = PathDiagnosticLocation(S, L.getManager(), LC); 964 } 965 966 if (firstCharOnly) 967 L = PathDiagnosticLocation::createSingleLocation(L); 968 969 return L; 970 } 971 972 class EdgeBuilder { 973 std::vector<ContextLocation> CLocs; 974 typedef std::vector<ContextLocation>::iterator iterator; 975 PathDiagnostic &PD; 976 PathDiagnosticBuilder &PDB; 977 PathDiagnosticLocation PrevLoc; 978 979 bool IsConsumedExpr(const PathDiagnosticLocation &L); 980 981 bool containsLocation(const PathDiagnosticLocation &Container, 982 const PathDiagnosticLocation &Containee); 983 984 PathDiagnosticLocation getContextLocation(const PathDiagnosticLocation &L); 985 986 987 988 void popLocation() { 989 if (!CLocs.back().isDead() && CLocs.back().asLocation().isFileID()) { 990 // For contexts, we only one the first character as the range. 991 rawAddEdge(cleanUpLocation(CLocs.back(), PDB.LC, true)); 992 } 993 CLocs.pop_back(); 994 } 995 996 public: 997 EdgeBuilder(PathDiagnostic &pd, PathDiagnosticBuilder &pdb) 998 : PD(pd), PDB(pdb) { 999 1000 // If the PathDiagnostic already has pieces, add the enclosing statement 1001 // of the first piece as a context as well. 1002 if (!PD.path.empty()) { 1003 PrevLoc = (*PD.path.begin())->getLocation(); 1004 1005 if (const Stmt *S = PrevLoc.asStmt()) 1006 addExtendedContext(PDB.getEnclosingStmtLocation(S).asStmt()); 1007 } 1008 } 1009 1010 ~EdgeBuilder() { 1011 while (!CLocs.empty()) popLocation(); 1012 1013 // Finally, add an initial edge from the start location of the first 1014 // statement (if it doesn't already exist). 1015 PathDiagnosticLocation L = PathDiagnosticLocation::createDeclBegin( 1016 PDB.LC, 1017 PDB.getSourceManager()); 1018 if (L.isValid()) 1019 rawAddEdge(L); 1020 } 1021 1022 void flushLocations() { 1023 while (!CLocs.empty()) 1024 popLocation(); 1025 PrevLoc = PathDiagnosticLocation(); 1026 } 1027 1028 void addEdge(PathDiagnosticLocation NewLoc, bool alwaysAdd = false, 1029 bool IsPostJump = false); 1030 1031 void rawAddEdge(PathDiagnosticLocation NewLoc); 1032 1033 void addContext(const Stmt *S); 1034 void addContext(const PathDiagnosticLocation &L); 1035 void addExtendedContext(const Stmt *S); 1036 }; 1037 } // end anonymous namespace 1038 1039 1040 PathDiagnosticLocation 1041 EdgeBuilder::getContextLocation(const PathDiagnosticLocation &L) { 1042 if (const Stmt *S = L.asStmt()) { 1043 if (IsControlFlowExpr(S)) 1044 return L; 1045 1046 return PDB.getEnclosingStmtLocation(S); 1047 } 1048 1049 return L; 1050 } 1051 1052 bool EdgeBuilder::containsLocation(const PathDiagnosticLocation &Container, 1053 const PathDiagnosticLocation &Containee) { 1054 1055 if (Container == Containee) 1056 return true; 1057 1058 if (Container.asDecl()) 1059 return true; 1060 1061 if (const Stmt *S = Containee.asStmt()) 1062 if (const Stmt *ContainerS = Container.asStmt()) { 1063 while (S) { 1064 if (S == ContainerS) 1065 return true; 1066 S = PDB.getParent(S); 1067 } 1068 return false; 1069 } 1070 1071 // Less accurate: compare using source ranges. 1072 SourceRange ContainerR = Container.asRange(); 1073 SourceRange ContaineeR = Containee.asRange(); 1074 1075 SourceManager &SM = PDB.getSourceManager(); 1076 SourceLocation ContainerRBeg = SM.getExpansionLoc(ContainerR.getBegin()); 1077 SourceLocation ContainerREnd = SM.getExpansionLoc(ContainerR.getEnd()); 1078 SourceLocation ContaineeRBeg = SM.getExpansionLoc(ContaineeR.getBegin()); 1079 SourceLocation ContaineeREnd = SM.getExpansionLoc(ContaineeR.getEnd()); 1080 1081 unsigned ContainerBegLine = SM.getExpansionLineNumber(ContainerRBeg); 1082 unsigned ContainerEndLine = SM.getExpansionLineNumber(ContainerREnd); 1083 unsigned ContaineeBegLine = SM.getExpansionLineNumber(ContaineeRBeg); 1084 unsigned ContaineeEndLine = SM.getExpansionLineNumber(ContaineeREnd); 1085 1086 assert(ContainerBegLine <= ContainerEndLine); 1087 assert(ContaineeBegLine <= ContaineeEndLine); 1088 1089 return (ContainerBegLine <= ContaineeBegLine && 1090 ContainerEndLine >= ContaineeEndLine && 1091 (ContainerBegLine != ContaineeBegLine || 1092 SM.getExpansionColumnNumber(ContainerRBeg) <= 1093 SM.getExpansionColumnNumber(ContaineeRBeg)) && 1094 (ContainerEndLine != ContaineeEndLine || 1095 SM.getExpansionColumnNumber(ContainerREnd) >= 1096 SM.getExpansionColumnNumber(ContaineeREnd))); 1097 } 1098 1099 void EdgeBuilder::rawAddEdge(PathDiagnosticLocation NewLoc) { 1100 if (!PrevLoc.isValid()) { 1101 PrevLoc = NewLoc; 1102 return; 1103 } 1104 1105 const PathDiagnosticLocation &NewLocClean = cleanUpLocation(NewLoc, PDB.LC); 1106 const PathDiagnosticLocation &PrevLocClean = cleanUpLocation(PrevLoc, PDB.LC); 1107 1108 if (PrevLocClean.asLocation().isInvalid()) { 1109 PrevLoc = NewLoc; 1110 return; 1111 } 1112 1113 if (NewLocClean.asLocation() == PrevLocClean.asLocation()) 1114 return; 1115 1116 // FIXME: Ignore intra-macro edges for now. 1117 if (NewLocClean.asLocation().getExpansionLoc() == 1118 PrevLocClean.asLocation().getExpansionLoc()) 1119 return; 1120 1121 PD.getActivePath().push_front(new PathDiagnosticControlFlowPiece(NewLocClean, PrevLocClean)); 1122 PrevLoc = NewLoc; 1123 } 1124 1125 void EdgeBuilder::addEdge(PathDiagnosticLocation NewLoc, bool alwaysAdd, 1126 bool IsPostJump) { 1127 1128 if (!alwaysAdd && NewLoc.asLocation().isMacroID()) 1129 return; 1130 1131 const PathDiagnosticLocation &CLoc = getContextLocation(NewLoc); 1132 1133 while (!CLocs.empty()) { 1134 ContextLocation &TopContextLoc = CLocs.back(); 1135 1136 // Is the top location context the same as the one for the new location? 1137 if (TopContextLoc == CLoc) { 1138 if (alwaysAdd) { 1139 if (IsConsumedExpr(TopContextLoc)) 1140 TopContextLoc.markDead(); 1141 1142 rawAddEdge(NewLoc); 1143 } 1144 1145 if (IsPostJump) 1146 TopContextLoc.markDead(); 1147 return; 1148 } 1149 1150 if (containsLocation(TopContextLoc, CLoc)) { 1151 if (alwaysAdd) { 1152 rawAddEdge(NewLoc); 1153 1154 if (IsConsumedExpr(CLoc)) { 1155 CLocs.push_back(ContextLocation(CLoc, /*IsDead=*/true)); 1156 return; 1157 } 1158 } 1159 1160 CLocs.push_back(ContextLocation(CLoc, /*IsDead=*/IsPostJump)); 1161 return; 1162 } 1163 1164 // Context does not contain the location. Flush it. 1165 popLocation(); 1166 } 1167 1168 // If we reach here, there is no enclosing context. Just add the edge. 1169 rawAddEdge(NewLoc); 1170 } 1171 1172 bool EdgeBuilder::IsConsumedExpr(const PathDiagnosticLocation &L) { 1173 if (const Expr *X = dyn_cast_or_null<Expr>(L.asStmt())) 1174 return PDB.getParentMap().isConsumedExpr(X) && !IsControlFlowExpr(X); 1175 1176 return false; 1177 } 1178 1179 void EdgeBuilder::addExtendedContext(const Stmt *S) { 1180 if (!S) 1181 return; 1182 1183 const Stmt *Parent = PDB.getParent(S); 1184 while (Parent) { 1185 if (isa<CompoundStmt>(Parent)) 1186 Parent = PDB.getParent(Parent); 1187 else 1188 break; 1189 } 1190 1191 if (Parent) { 1192 switch (Parent->getStmtClass()) { 1193 case Stmt::DoStmtClass: 1194 case Stmt::ObjCAtSynchronizedStmtClass: 1195 addContext(Parent); 1196 default: 1197 break; 1198 } 1199 } 1200 1201 addContext(S); 1202 } 1203 1204 void EdgeBuilder::addContext(const Stmt *S) { 1205 if (!S) 1206 return; 1207 1208 PathDiagnosticLocation L(S, PDB.getSourceManager(), PDB.LC); 1209 addContext(L); 1210 } 1211 1212 void EdgeBuilder::addContext(const PathDiagnosticLocation &L) { 1213 while (!CLocs.empty()) { 1214 const PathDiagnosticLocation &TopContextLoc = CLocs.back(); 1215 1216 // Is the top location context the same as the one for the new location? 1217 if (TopContextLoc == L) 1218 return; 1219 1220 if (containsLocation(TopContextLoc, L)) { 1221 CLocs.push_back(L); 1222 return; 1223 } 1224 1225 // Context does not contain the location. Flush it. 1226 popLocation(); 1227 } 1228 1229 CLocs.push_back(L); 1230 } 1231 1232 // Cone-of-influence: support the reverse propagation of "interesting" symbols 1233 // and values by tracing interesting calculations backwards through evaluated 1234 // expressions along a path. This is probably overly complicated, but the idea 1235 // is that if an expression computed an "interesting" value, the child 1236 // expressions are are also likely to be "interesting" as well (which then 1237 // propagates to the values they in turn compute). This reverse propagation 1238 // is needed to track interesting correlations across function call boundaries, 1239 // where formal arguments bind to actual arguments, etc. This is also needed 1240 // because the constraint solver sometimes simplifies certain symbolic values 1241 // into constants when appropriate, and this complicates reasoning about 1242 // interesting values. 1243 typedef llvm::DenseSet<const Expr *> InterestingExprs; 1244 1245 static void reversePropagateIntererstingSymbols(BugReport &R, 1246 InterestingExprs &IE, 1247 const ProgramState *State, 1248 const Expr *Ex, 1249 const LocationContext *LCtx) { 1250 SVal V = State->getSVal(Ex, LCtx); 1251 if (!(R.isInteresting(V) || IE.count(Ex))) 1252 return; 1253 1254 switch (Ex->getStmtClass()) { 1255 default: 1256 if (!isa<CastExpr>(Ex)) 1257 break; 1258 // Fall through. 1259 case Stmt::BinaryOperatorClass: 1260 case Stmt::UnaryOperatorClass: { 1261 for (const Stmt *SubStmt : Ex->children()) { 1262 if (const Expr *child = dyn_cast_or_null<Expr>(SubStmt)) { 1263 IE.insert(child); 1264 SVal ChildV = State->getSVal(child, LCtx); 1265 R.markInteresting(ChildV); 1266 } 1267 } 1268 break; 1269 } 1270 } 1271 1272 R.markInteresting(V); 1273 } 1274 1275 static void reversePropagateInterestingSymbols(BugReport &R, 1276 InterestingExprs &IE, 1277 const ProgramState *State, 1278 const LocationContext *CalleeCtx, 1279 const LocationContext *CallerCtx) 1280 { 1281 // FIXME: Handle non-CallExpr-based CallEvents. 1282 const StackFrameContext *Callee = CalleeCtx->getCurrentStackFrame(); 1283 const Stmt *CallSite = Callee->getCallSite(); 1284 if (const CallExpr *CE = dyn_cast_or_null<CallExpr>(CallSite)) { 1285 if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(CalleeCtx->getDecl())) { 1286 FunctionDecl::param_const_iterator PI = FD->param_begin(), 1287 PE = FD->param_end(); 1288 CallExpr::const_arg_iterator AI = CE->arg_begin(), AE = CE->arg_end(); 1289 for (; AI != AE && PI != PE; ++AI, ++PI) { 1290 if (const Expr *ArgE = *AI) { 1291 if (const ParmVarDecl *PD = *PI) { 1292 Loc LV = State->getLValue(PD, CalleeCtx); 1293 if (R.isInteresting(LV) || R.isInteresting(State->getRawSVal(LV))) 1294 IE.insert(ArgE); 1295 } 1296 } 1297 } 1298 } 1299 } 1300 } 1301 1302 //===----------------------------------------------------------------------===// 1303 // Functions for determining if a loop was executed 0 times. 1304 //===----------------------------------------------------------------------===// 1305 1306 static bool isLoop(const Stmt *Term) { 1307 switch (Term->getStmtClass()) { 1308 case Stmt::ForStmtClass: 1309 case Stmt::WhileStmtClass: 1310 case Stmt::ObjCForCollectionStmtClass: 1311 case Stmt::CXXForRangeStmtClass: 1312 return true; 1313 default: 1314 // Note that we intentionally do not include do..while here. 1315 return false; 1316 } 1317 } 1318 1319 static bool isJumpToFalseBranch(const BlockEdge *BE) { 1320 const CFGBlock *Src = BE->getSrc(); 1321 assert(Src->succ_size() == 2); 1322 return (*(Src->succ_begin()+1) == BE->getDst()); 1323 } 1324 1325 /// Return true if the terminator is a loop and the destination is the 1326 /// false branch. 1327 static bool isLoopJumpPastBody(const Stmt *Term, const BlockEdge *BE) { 1328 if (!isLoop(Term)) 1329 return false; 1330 1331 // Did we take the false branch? 1332 return isJumpToFalseBranch(BE); 1333 } 1334 1335 static bool isContainedByStmt(ParentMap &PM, const Stmt *S, const Stmt *SubS) { 1336 while (SubS) { 1337 if (SubS == S) 1338 return true; 1339 SubS = PM.getParent(SubS); 1340 } 1341 return false; 1342 } 1343 1344 static const Stmt *getStmtBeforeCond(ParentMap &PM, const Stmt *Term, 1345 const ExplodedNode *N) { 1346 while (N) { 1347 Optional<StmtPoint> SP = N->getLocation().getAs<StmtPoint>(); 1348 if (SP) { 1349 const Stmt *S = SP->getStmt(); 1350 if (!isContainedByStmt(PM, Term, S)) 1351 return S; 1352 } 1353 N = N->getFirstPred(); 1354 } 1355 return nullptr; 1356 } 1357 1358 static bool isInLoopBody(ParentMap &PM, const Stmt *S, const Stmt *Term) { 1359 const Stmt *LoopBody = nullptr; 1360 switch (Term->getStmtClass()) { 1361 case Stmt::CXXForRangeStmtClass: { 1362 const CXXForRangeStmt *FR = cast<CXXForRangeStmt>(Term); 1363 if (isContainedByStmt(PM, FR->getInc(), S)) 1364 return true; 1365 if (isContainedByStmt(PM, FR->getLoopVarStmt(), S)) 1366 return true; 1367 LoopBody = FR->getBody(); 1368 break; 1369 } 1370 case Stmt::ForStmtClass: { 1371 const ForStmt *FS = cast<ForStmt>(Term); 1372 if (isContainedByStmt(PM, FS->getInc(), S)) 1373 return true; 1374 LoopBody = FS->getBody(); 1375 break; 1376 } 1377 case Stmt::ObjCForCollectionStmtClass: { 1378 const ObjCForCollectionStmt *FC = cast<ObjCForCollectionStmt>(Term); 1379 LoopBody = FC->getBody(); 1380 break; 1381 } 1382 case Stmt::WhileStmtClass: 1383 LoopBody = cast<WhileStmt>(Term)->getBody(); 1384 break; 1385 default: 1386 return false; 1387 } 1388 return isContainedByStmt(PM, LoopBody, S); 1389 } 1390 1391 //===----------------------------------------------------------------------===// 1392 // Top-level logic for generating extensive path diagnostics. 1393 //===----------------------------------------------------------------------===// 1394 1395 static bool GenerateExtensivePathDiagnostic( 1396 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N, 1397 LocationContextMap &LCM, 1398 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) { 1399 EdgeBuilder EB(PD, PDB); 1400 const SourceManager& SM = PDB.getSourceManager(); 1401 StackDiagVector CallStack; 1402 InterestingExprs IE; 1403 1404 const ExplodedNode *NextNode = N->pred_empty() ? nullptr : *(N->pred_begin()); 1405 while (NextNode) { 1406 N = NextNode; 1407 NextNode = N->getFirstPred(); 1408 ProgramPoint P = N->getLocation(); 1409 1410 do { 1411 if (Optional<PostStmt> PS = P.getAs<PostStmt>()) { 1412 if (const Expr *Ex = PS->getStmtAs<Expr>()) 1413 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE, 1414 N->getState().get(), Ex, 1415 N->getLocationContext()); 1416 } 1417 1418 if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) { 1419 const Stmt *S = CE->getCalleeContext()->getCallSite(); 1420 if (const Expr *Ex = dyn_cast_or_null<Expr>(S)) { 1421 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE, 1422 N->getState().get(), Ex, 1423 N->getLocationContext()); 1424 } 1425 1426 PathDiagnosticCallPiece *C = 1427 PathDiagnosticCallPiece::construct(N, *CE, SM); 1428 LCM[&C->path] = CE->getCalleeContext(); 1429 1430 EB.addEdge(C->callReturn, /*AlwaysAdd=*/true, /*IsPostJump=*/true); 1431 EB.flushLocations(); 1432 1433 PD.getActivePath().push_front(C); 1434 PD.pushActivePath(&C->path); 1435 CallStack.push_back(StackDiagPair(C, N)); 1436 break; 1437 } 1438 1439 // Pop the call hierarchy if we are done walking the contents 1440 // of a function call. 1441 if (Optional<CallEnter> CE = P.getAs<CallEnter>()) { 1442 // Add an edge to the start of the function. 1443 const Decl *D = CE->getCalleeContext()->getDecl(); 1444 PathDiagnosticLocation pos = 1445 PathDiagnosticLocation::createBegin(D, SM); 1446 EB.addEdge(pos); 1447 1448 // Flush all locations, and pop the active path. 1449 bool VisitedEntireCall = PD.isWithinCall(); 1450 EB.flushLocations(); 1451 PD.popActivePath(); 1452 PDB.LC = N->getLocationContext(); 1453 1454 // Either we just added a bunch of stuff to the top-level path, or 1455 // we have a previous CallExitEnd. If the former, it means that the 1456 // path terminated within a function call. We must then take the 1457 // current contents of the active path and place it within 1458 // a new PathDiagnosticCallPiece. 1459 PathDiagnosticCallPiece *C; 1460 if (VisitedEntireCall) { 1461 C = cast<PathDiagnosticCallPiece>(PD.getActivePath().front()); 1462 } else { 1463 const Decl *Caller = CE->getLocationContext()->getDecl(); 1464 C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller); 1465 LCM[&C->path] = CE->getCalleeContext(); 1466 } 1467 1468 C->setCallee(*CE, SM); 1469 EB.addContext(C->getLocation()); 1470 1471 if (!CallStack.empty()) { 1472 assert(CallStack.back().first == C); 1473 CallStack.pop_back(); 1474 } 1475 break; 1476 } 1477 1478 // Note that is important that we update the LocationContext 1479 // after looking at CallExits. CallExit basically adds an 1480 // edge in the *caller*, so we don't want to update the LocationContext 1481 // too soon. 1482 PDB.LC = N->getLocationContext(); 1483 1484 // Block edges. 1485 if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) { 1486 // Does this represent entering a call? If so, look at propagating 1487 // interesting symbols across call boundaries. 1488 if (NextNode) { 1489 const LocationContext *CallerCtx = NextNode->getLocationContext(); 1490 const LocationContext *CalleeCtx = PDB.LC; 1491 if (CallerCtx != CalleeCtx) { 1492 reversePropagateInterestingSymbols(*PDB.getBugReport(), IE, 1493 N->getState().get(), 1494 CalleeCtx, CallerCtx); 1495 } 1496 } 1497 1498 // Are we jumping to the head of a loop? Add a special diagnostic. 1499 if (const Stmt *Loop = BE->getSrc()->getLoopTarget()) { 1500 PathDiagnosticLocation L(Loop, SM, PDB.LC); 1501 const CompoundStmt *CS = nullptr; 1502 1503 if (const ForStmt *FS = dyn_cast<ForStmt>(Loop)) 1504 CS = dyn_cast<CompoundStmt>(FS->getBody()); 1505 else if (const WhileStmt *WS = dyn_cast<WhileStmt>(Loop)) 1506 CS = dyn_cast<CompoundStmt>(WS->getBody()); 1507 1508 PathDiagnosticEventPiece *p = 1509 new PathDiagnosticEventPiece(L, 1510 "Looping back to the head of the loop"); 1511 p->setPrunable(true); 1512 1513 EB.addEdge(p->getLocation(), true); 1514 PD.getActivePath().push_front(p); 1515 1516 if (CS) { 1517 PathDiagnosticLocation BL = 1518 PathDiagnosticLocation::createEndBrace(CS, SM); 1519 EB.addEdge(BL); 1520 } 1521 } 1522 1523 const CFGBlock *BSrc = BE->getSrc(); 1524 ParentMap &PM = PDB.getParentMap(); 1525 1526 if (const Stmt *Term = BSrc->getTerminator()) { 1527 // Are we jumping past the loop body without ever executing the 1528 // loop (because the condition was false)? 1529 if (isLoopJumpPastBody(Term, &*BE) && 1530 !isInLoopBody(PM, 1531 getStmtBeforeCond(PM, 1532 BSrc->getTerminatorCondition(), 1533 N), 1534 Term)) { 1535 PathDiagnosticLocation L(Term, SM, PDB.LC); 1536 PathDiagnosticEventPiece *PE = 1537 new PathDiagnosticEventPiece(L, "Loop body executed 0 times"); 1538 PE->setPrunable(true); 1539 1540 EB.addEdge(PE->getLocation(), true); 1541 PD.getActivePath().push_front(PE); 1542 } 1543 1544 // In any case, add the terminator as the current statement 1545 // context for control edges. 1546 EB.addContext(Term); 1547 } 1548 1549 break; 1550 } 1551 1552 if (Optional<BlockEntrance> BE = P.getAs<BlockEntrance>()) { 1553 Optional<CFGElement> First = BE->getFirstElement(); 1554 if (Optional<CFGStmt> S = First ? First->getAs<CFGStmt>() : None) { 1555 const Stmt *stmt = S->getStmt(); 1556 if (IsControlFlowExpr(stmt)) { 1557 // Add the proper context for '&&', '||', and '?'. 1558 EB.addContext(stmt); 1559 } 1560 else 1561 EB.addExtendedContext(PDB.getEnclosingStmtLocation(stmt).asStmt()); 1562 } 1563 1564 break; 1565 } 1566 1567 1568 } while (0); 1569 1570 if (!NextNode) 1571 continue; 1572 1573 // Add pieces from custom visitors. 1574 BugReport *R = PDB.getBugReport(); 1575 for (auto &V : visitors) { 1576 if (PathDiagnosticPiece *p = V->VisitNode(N, NextNode, PDB, *R)) { 1577 const PathDiagnosticLocation &Loc = p->getLocation(); 1578 EB.addEdge(Loc, true); 1579 PD.getActivePath().push_front(p); 1580 updateStackPiecesWithMessage(p, CallStack); 1581 1582 if (const Stmt *S = Loc.asStmt()) 1583 EB.addExtendedContext(PDB.getEnclosingStmtLocation(S).asStmt()); 1584 } 1585 } 1586 } 1587 1588 return PDB.getBugReport()->isValid(); 1589 } 1590 1591 /// \brief Adds a sanitized control-flow diagnostic edge to a path. 1592 static void addEdgeToPath(PathPieces &path, 1593 PathDiagnosticLocation &PrevLoc, 1594 PathDiagnosticLocation NewLoc, 1595 const LocationContext *LC) { 1596 if (!NewLoc.isValid()) 1597 return; 1598 1599 SourceLocation NewLocL = NewLoc.asLocation(); 1600 if (NewLocL.isInvalid()) 1601 return; 1602 1603 if (!PrevLoc.isValid() || !PrevLoc.asLocation().isValid()) { 1604 PrevLoc = NewLoc; 1605 return; 1606 } 1607 1608 // Ignore self-edges, which occur when there are multiple nodes at the same 1609 // statement. 1610 if (NewLoc.asStmt() && NewLoc.asStmt() == PrevLoc.asStmt()) 1611 return; 1612 1613 path.push_front(new PathDiagnosticControlFlowPiece(NewLoc, 1614 PrevLoc)); 1615 PrevLoc = NewLoc; 1616 } 1617 1618 /// A customized wrapper for CFGBlock::getTerminatorCondition() 1619 /// which returns the element for ObjCForCollectionStmts. 1620 static const Stmt *getTerminatorCondition(const CFGBlock *B) { 1621 const Stmt *S = B->getTerminatorCondition(); 1622 if (const ObjCForCollectionStmt *FS = 1623 dyn_cast_or_null<ObjCForCollectionStmt>(S)) 1624 return FS->getElement(); 1625 return S; 1626 } 1627 1628 static const char StrEnteringLoop[] = "Entering loop body"; 1629 static const char StrLoopBodyZero[] = "Loop body executed 0 times"; 1630 static const char StrLoopRangeEmpty[] = 1631 "Loop body skipped when range is empty"; 1632 static const char StrLoopCollectionEmpty[] = 1633 "Loop body skipped when collection is empty"; 1634 1635 static bool GenerateAlternateExtensivePathDiagnostic( 1636 PathDiagnostic &PD, PathDiagnosticBuilder &PDB, const ExplodedNode *N, 1637 LocationContextMap &LCM, 1638 ArrayRef<std::unique_ptr<BugReporterVisitor>> visitors) { 1639 1640 BugReport *report = PDB.getBugReport(); 1641 const SourceManager& SM = PDB.getSourceManager(); 1642 StackDiagVector CallStack; 1643 InterestingExprs IE; 1644 1645 PathDiagnosticLocation PrevLoc = PD.getLocation(); 1646 1647 const ExplodedNode *NextNode = N->getFirstPred(); 1648 while (NextNode) { 1649 N = NextNode; 1650 NextNode = N->getFirstPred(); 1651 ProgramPoint P = N->getLocation(); 1652 1653 do { 1654 // Have we encountered an entrance to a call? It may be 1655 // the case that we have not encountered a matching 1656 // call exit before this point. This means that the path 1657 // terminated within the call itself. 1658 if (Optional<CallEnter> CE = P.getAs<CallEnter>()) { 1659 // Add an edge to the start of the function. 1660 const StackFrameContext *CalleeLC = CE->getCalleeContext(); 1661 const Decl *D = CalleeLC->getDecl(); 1662 addEdgeToPath(PD.getActivePath(), PrevLoc, 1663 PathDiagnosticLocation::createBegin(D, SM), 1664 CalleeLC); 1665 1666 // Did we visit an entire call? 1667 bool VisitedEntireCall = PD.isWithinCall(); 1668 PD.popActivePath(); 1669 1670 PathDiagnosticCallPiece *C; 1671 if (VisitedEntireCall) { 1672 PathDiagnosticPiece *P = PD.getActivePath().front().get(); 1673 C = cast<PathDiagnosticCallPiece>(P); 1674 } else { 1675 const Decl *Caller = CE->getLocationContext()->getDecl(); 1676 C = PathDiagnosticCallPiece::construct(PD.getActivePath(), Caller); 1677 1678 // Since we just transferred the path over to the call piece, 1679 // reset the mapping from active to location context. 1680 assert(PD.getActivePath().size() == 1 && 1681 PD.getActivePath().front() == C); 1682 LCM[&PD.getActivePath()] = nullptr; 1683 1684 // Record the location context mapping for the path within 1685 // the call. 1686 assert(LCM[&C->path] == nullptr || 1687 LCM[&C->path] == CE->getCalleeContext()); 1688 LCM[&C->path] = CE->getCalleeContext(); 1689 1690 // If this is the first item in the active path, record 1691 // the new mapping from active path to location context. 1692 const LocationContext *&NewLC = LCM[&PD.getActivePath()]; 1693 if (!NewLC) 1694 NewLC = N->getLocationContext(); 1695 1696 PDB.LC = NewLC; 1697 } 1698 C->setCallee(*CE, SM); 1699 1700 // Update the previous location in the active path. 1701 PrevLoc = C->getLocation(); 1702 1703 if (!CallStack.empty()) { 1704 assert(CallStack.back().first == C); 1705 CallStack.pop_back(); 1706 } 1707 break; 1708 } 1709 1710 // Query the location context here and the previous location 1711 // as processing CallEnter may change the active path. 1712 PDB.LC = N->getLocationContext(); 1713 1714 // Record the mapping from the active path to the location 1715 // context. 1716 assert(!LCM[&PD.getActivePath()] || 1717 LCM[&PD.getActivePath()] == PDB.LC); 1718 LCM[&PD.getActivePath()] = PDB.LC; 1719 1720 // Have we encountered an exit from a function call? 1721 if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) { 1722 const Stmt *S = CE->getCalleeContext()->getCallSite(); 1723 // Propagate the interesting symbols accordingly. 1724 if (const Expr *Ex = dyn_cast_or_null<Expr>(S)) { 1725 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE, 1726 N->getState().get(), Ex, 1727 N->getLocationContext()); 1728 } 1729 1730 // We are descending into a call (backwards). Construct 1731 // a new call piece to contain the path pieces for that call. 1732 PathDiagnosticCallPiece *C = 1733 PathDiagnosticCallPiece::construct(N, *CE, SM); 1734 1735 // Record the location context for this call piece. 1736 LCM[&C->path] = CE->getCalleeContext(); 1737 1738 // Add the edge to the return site. 1739 addEdgeToPath(PD.getActivePath(), PrevLoc, C->callReturn, PDB.LC); 1740 PD.getActivePath().push_front(C); 1741 PrevLoc.invalidate(); 1742 1743 // Make the contents of the call the active path for now. 1744 PD.pushActivePath(&C->path); 1745 CallStack.push_back(StackDiagPair(C, N)); 1746 break; 1747 } 1748 1749 if (Optional<PostStmt> PS = P.getAs<PostStmt>()) { 1750 // For expressions, make sure we propagate the 1751 // interesting symbols correctly. 1752 if (const Expr *Ex = PS->getStmtAs<Expr>()) 1753 reversePropagateIntererstingSymbols(*PDB.getBugReport(), IE, 1754 N->getState().get(), Ex, 1755 N->getLocationContext()); 1756 1757 // Add an edge. If this is an ObjCForCollectionStmt do 1758 // not add an edge here as it appears in the CFG both 1759 // as a terminator and as a terminator condition. 1760 if (!isa<ObjCForCollectionStmt>(PS->getStmt())) { 1761 PathDiagnosticLocation L = 1762 PathDiagnosticLocation(PS->getStmt(), SM, PDB.LC); 1763 addEdgeToPath(PD.getActivePath(), PrevLoc, L, PDB.LC); 1764 } 1765 break; 1766 } 1767 1768 // Block edges. 1769 if (Optional<BlockEdge> BE = P.getAs<BlockEdge>()) { 1770 // Does this represent entering a call? If so, look at propagating 1771 // interesting symbols across call boundaries. 1772 if (NextNode) { 1773 const LocationContext *CallerCtx = NextNode->getLocationContext(); 1774 const LocationContext *CalleeCtx = PDB.LC; 1775 if (CallerCtx != CalleeCtx) { 1776 reversePropagateInterestingSymbols(*PDB.getBugReport(), IE, 1777 N->getState().get(), 1778 CalleeCtx, CallerCtx); 1779 } 1780 } 1781 1782 // Are we jumping to the head of a loop? Add a special diagnostic. 1783 if (const Stmt *Loop = BE->getSrc()->getLoopTarget()) { 1784 PathDiagnosticLocation L(Loop, SM, PDB.LC); 1785 const Stmt *Body = nullptr; 1786 1787 if (const ForStmt *FS = dyn_cast<ForStmt>(Loop)) 1788 Body = FS->getBody(); 1789 else if (const WhileStmt *WS = dyn_cast<WhileStmt>(Loop)) 1790 Body = WS->getBody(); 1791 else if (const ObjCForCollectionStmt *OFS = 1792 dyn_cast<ObjCForCollectionStmt>(Loop)) { 1793 Body = OFS->getBody(); 1794 } else if (const CXXForRangeStmt *FRS = 1795 dyn_cast<CXXForRangeStmt>(Loop)) { 1796 Body = FRS->getBody(); 1797 } 1798 // do-while statements are explicitly excluded here 1799 1800 PathDiagnosticEventPiece *p = 1801 new PathDiagnosticEventPiece(L, "Looping back to the head " 1802 "of the loop"); 1803 p->setPrunable(true); 1804 1805 addEdgeToPath(PD.getActivePath(), PrevLoc, p->getLocation(), PDB.LC); 1806 PD.getActivePath().push_front(p); 1807 1808 if (const CompoundStmt *CS = dyn_cast_or_null<CompoundStmt>(Body)) { 1809 addEdgeToPath(PD.getActivePath(), PrevLoc, 1810 PathDiagnosticLocation::createEndBrace(CS, SM), 1811 PDB.LC); 1812 } 1813 } 1814 1815 const CFGBlock *BSrc = BE->getSrc(); 1816 ParentMap &PM = PDB.getParentMap(); 1817 1818 if (const Stmt *Term = BSrc->getTerminator()) { 1819 // Are we jumping past the loop body without ever executing the 1820 // loop (because the condition was false)? 1821 if (isLoop(Term)) { 1822 const Stmt *TermCond = getTerminatorCondition(BSrc); 1823 bool IsInLoopBody = 1824 isInLoopBody(PM, getStmtBeforeCond(PM, TermCond, N), Term); 1825 1826 const char *str = nullptr; 1827 1828 if (isJumpToFalseBranch(&*BE)) { 1829 if (!IsInLoopBody) { 1830 if (isa<ObjCForCollectionStmt>(Term)) { 1831 str = StrLoopCollectionEmpty; 1832 } else if (isa<CXXForRangeStmt>(Term)) { 1833 str = StrLoopRangeEmpty; 1834 } else { 1835 str = StrLoopBodyZero; 1836 } 1837 } 1838 } else { 1839 str = StrEnteringLoop; 1840 } 1841 1842 if (str) { 1843 PathDiagnosticLocation L(TermCond ? TermCond : Term, SM, PDB.LC); 1844 PathDiagnosticEventPiece *PE = 1845 new PathDiagnosticEventPiece(L, str); 1846 PE->setPrunable(true); 1847 addEdgeToPath(PD.getActivePath(), PrevLoc, 1848 PE->getLocation(), PDB.LC); 1849 PD.getActivePath().push_front(PE); 1850 } 1851 } else if (isa<BreakStmt>(Term) || isa<ContinueStmt>(Term) || 1852 isa<GotoStmt>(Term)) { 1853 PathDiagnosticLocation L(Term, SM, PDB.LC); 1854 addEdgeToPath(PD.getActivePath(), PrevLoc, L, PDB.LC); 1855 } 1856 } 1857 break; 1858 } 1859 } while (0); 1860 1861 if (!NextNode) 1862 continue; 1863 1864 // Add pieces from custom visitors. 1865 for (auto &V : visitors) { 1866 if (PathDiagnosticPiece *p = V->VisitNode(N, NextNode, PDB, *report)) { 1867 addEdgeToPath(PD.getActivePath(), PrevLoc, p->getLocation(), PDB.LC); 1868 PD.getActivePath().push_front(p); 1869 updateStackPiecesWithMessage(p, CallStack); 1870 } 1871 } 1872 } 1873 1874 // Add an edge to the start of the function. 1875 // We'll prune it out later, but it helps make diagnostics more uniform. 1876 const StackFrameContext *CalleeLC = PDB.LC->getCurrentStackFrame(); 1877 const Decl *D = CalleeLC->getDecl(); 1878 addEdgeToPath(PD.getActivePath(), PrevLoc, 1879 PathDiagnosticLocation::createBegin(D, SM), 1880 CalleeLC); 1881 1882 return report->isValid(); 1883 } 1884 1885 static const Stmt *getLocStmt(PathDiagnosticLocation L) { 1886 if (!L.isValid()) 1887 return nullptr; 1888 return L.asStmt(); 1889 } 1890 1891 static const Stmt *getStmtParent(const Stmt *S, const ParentMap &PM) { 1892 if (!S) 1893 return nullptr; 1894 1895 while (true) { 1896 S = PM.getParentIgnoreParens(S); 1897 1898 if (!S) 1899 break; 1900 1901 if (isa<ExprWithCleanups>(S) || 1902 isa<CXXBindTemporaryExpr>(S) || 1903 isa<SubstNonTypeTemplateParmExpr>(S)) 1904 continue; 1905 1906 break; 1907 } 1908 1909 return S; 1910 } 1911 1912 static bool isConditionForTerminator(const Stmt *S, const Stmt *Cond) { 1913 switch (S->getStmtClass()) { 1914 case Stmt::BinaryOperatorClass: { 1915 const BinaryOperator *BO = cast<BinaryOperator>(S); 1916 if (!BO->isLogicalOp()) 1917 return false; 1918 return BO->getLHS() == Cond || BO->getRHS() == Cond; 1919 } 1920 case Stmt::IfStmtClass: 1921 return cast<IfStmt>(S)->getCond() == Cond; 1922 case Stmt::ForStmtClass: 1923 return cast<ForStmt>(S)->getCond() == Cond; 1924 case Stmt::WhileStmtClass: 1925 return cast<WhileStmt>(S)->getCond() == Cond; 1926 case Stmt::DoStmtClass: 1927 return cast<DoStmt>(S)->getCond() == Cond; 1928 case Stmt::ChooseExprClass: 1929 return cast<ChooseExpr>(S)->getCond() == Cond; 1930 case Stmt::IndirectGotoStmtClass: 1931 return cast<IndirectGotoStmt>(S)->getTarget() == Cond; 1932 case Stmt::SwitchStmtClass: 1933 return cast<SwitchStmt>(S)->getCond() == Cond; 1934 case Stmt::BinaryConditionalOperatorClass: 1935 return cast<BinaryConditionalOperator>(S)->getCond() == Cond; 1936 case Stmt::ConditionalOperatorClass: { 1937 const ConditionalOperator *CO = cast<ConditionalOperator>(S); 1938 return CO->getCond() == Cond || 1939 CO->getLHS() == Cond || 1940 CO->getRHS() == Cond; 1941 } 1942 case Stmt::ObjCForCollectionStmtClass: 1943 return cast<ObjCForCollectionStmt>(S)->getElement() == Cond; 1944 case Stmt::CXXForRangeStmtClass: { 1945 const CXXForRangeStmt *FRS = cast<CXXForRangeStmt>(S); 1946 return FRS->getCond() == Cond || FRS->getRangeInit() == Cond; 1947 } 1948 default: 1949 return false; 1950 } 1951 } 1952 1953 static bool isIncrementOrInitInForLoop(const Stmt *S, const Stmt *FL) { 1954 if (const ForStmt *FS = dyn_cast<ForStmt>(FL)) 1955 return FS->getInc() == S || FS->getInit() == S; 1956 if (const CXXForRangeStmt *FRS = dyn_cast<CXXForRangeStmt>(FL)) 1957 return FRS->getInc() == S || FRS->getRangeStmt() == S || 1958 FRS->getLoopVarStmt() || FRS->getRangeInit() == S; 1959 return false; 1960 } 1961 1962 typedef llvm::DenseSet<const PathDiagnosticCallPiece *> 1963 OptimizedCallsSet; 1964 1965 /// Adds synthetic edges from top-level statements to their subexpressions. 1966 /// 1967 /// This avoids a "swoosh" effect, where an edge from a top-level statement A 1968 /// points to a sub-expression B.1 that's not at the start of B. In these cases, 1969 /// we'd like to see an edge from A to B, then another one from B to B.1. 1970 static void addContextEdges(PathPieces &pieces, SourceManager &SM, 1971 const ParentMap &PM, const LocationContext *LCtx) { 1972 PathPieces::iterator Prev = pieces.end(); 1973 for (PathPieces::iterator I = pieces.begin(), E = Prev; I != E; 1974 Prev = I, ++I) { 1975 PathDiagnosticControlFlowPiece *Piece = 1976 dyn_cast<PathDiagnosticControlFlowPiece>(*I); 1977 1978 if (!Piece) 1979 continue; 1980 1981 PathDiagnosticLocation SrcLoc = Piece->getStartLocation(); 1982 SmallVector<PathDiagnosticLocation, 4> SrcContexts; 1983 1984 PathDiagnosticLocation NextSrcContext = SrcLoc; 1985 const Stmt *InnerStmt = nullptr; 1986 while (NextSrcContext.isValid() && NextSrcContext.asStmt() != InnerStmt) { 1987 SrcContexts.push_back(NextSrcContext); 1988 InnerStmt = NextSrcContext.asStmt(); 1989 NextSrcContext = getEnclosingStmtLocation(InnerStmt, SM, PM, LCtx, 1990 /*allowNested=*/true); 1991 } 1992 1993 // Repeatedly split the edge as necessary. 1994 // This is important for nested logical expressions (||, &&, ?:) where we 1995 // want to show all the levels of context. 1996 while (true) { 1997 const Stmt *Dst = getLocStmt(Piece->getEndLocation()); 1998 1999 // We are looking at an edge. Is the destination within a larger 2000 // expression? 2001 PathDiagnosticLocation DstContext = 2002 getEnclosingStmtLocation(Dst, SM, PM, LCtx, /*allowNested=*/true); 2003 if (!DstContext.isValid() || DstContext.asStmt() == Dst) 2004 break; 2005 2006 // If the source is in the same context, we're already good. 2007 if (std::find(SrcContexts.begin(), SrcContexts.end(), DstContext) != 2008 SrcContexts.end()) 2009 break; 2010 2011 // Update the subexpression node to point to the context edge. 2012 Piece->setStartLocation(DstContext); 2013 2014 // Try to extend the previous edge if it's at the same level as the source 2015 // context. 2016 if (Prev != E) { 2017 PathDiagnosticControlFlowPiece *PrevPiece = 2018 dyn_cast<PathDiagnosticControlFlowPiece>(*Prev); 2019 2020 if (PrevPiece) { 2021 if (const Stmt *PrevSrc = getLocStmt(PrevPiece->getStartLocation())) { 2022 const Stmt *PrevSrcParent = getStmtParent(PrevSrc, PM); 2023 if (PrevSrcParent == getStmtParent(getLocStmt(DstContext), PM)) { 2024 PrevPiece->setEndLocation(DstContext); 2025 break; 2026 } 2027 } 2028 } 2029 } 2030 2031 // Otherwise, split the current edge into a context edge and a 2032 // subexpression edge. Note that the context statement may itself have 2033 // context. 2034 Piece = new PathDiagnosticControlFlowPiece(SrcLoc, DstContext); 2035 I = pieces.insert(I, Piece); 2036 } 2037 } 2038 } 2039 2040 /// \brief Move edges from a branch condition to a branch target 2041 /// when the condition is simple. 2042 /// 2043 /// This restructures some of the work of addContextEdges. That function 2044 /// creates edges this may destroy, but they work together to create a more 2045 /// aesthetically set of edges around branches. After the call to 2046 /// addContextEdges, we may have (1) an edge to the branch, (2) an edge from 2047 /// the branch to the branch condition, and (3) an edge from the branch 2048 /// condition to the branch target. We keep (1), but may wish to remove (2) 2049 /// and move the source of (3) to the branch if the branch condition is simple. 2050 /// 2051 static void simplifySimpleBranches(PathPieces &pieces) { 2052 for (PathPieces::iterator I = pieces.begin(), E = pieces.end(); I != E; ++I) { 2053 2054 PathDiagnosticControlFlowPiece *PieceI = 2055 dyn_cast<PathDiagnosticControlFlowPiece>(*I); 2056 2057 if (!PieceI) 2058 continue; 2059 2060 const Stmt *s1Start = getLocStmt(PieceI->getStartLocation()); 2061 const Stmt *s1End = getLocStmt(PieceI->getEndLocation()); 2062 2063 if (!s1Start || !s1End) 2064 continue; 2065 2066 PathPieces::iterator NextI = I; ++NextI; 2067 if (NextI == E) 2068 break; 2069 2070 PathDiagnosticControlFlowPiece *PieceNextI = nullptr; 2071 2072 while (true) { 2073 if (NextI == E) 2074 break; 2075 2076 PathDiagnosticEventPiece *EV = dyn_cast<PathDiagnosticEventPiece>(*NextI); 2077 if (EV) { 2078 StringRef S = EV->getString(); 2079 if (S == StrEnteringLoop || S == StrLoopBodyZero || 2080 S == StrLoopCollectionEmpty || S == StrLoopRangeEmpty) { 2081 ++NextI; 2082 continue; 2083 } 2084 break; 2085 } 2086 2087 PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(*NextI); 2088 break; 2089 } 2090 2091 if (!PieceNextI) 2092 continue; 2093 2094 const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation()); 2095 const Stmt *s2End = getLocStmt(PieceNextI->getEndLocation()); 2096 2097 if (!s2Start || !s2End || s1End != s2Start) 2098 continue; 2099 2100 // We only perform this transformation for specific branch kinds. 2101 // We don't want to do this for do..while, for example. 2102 if (!(isa<ForStmt>(s1Start) || isa<WhileStmt>(s1Start) || 2103 isa<IfStmt>(s1Start) || isa<ObjCForCollectionStmt>(s1Start) || 2104 isa<CXXForRangeStmt>(s1Start))) 2105 continue; 2106 2107 // Is s1End the branch condition? 2108 if (!isConditionForTerminator(s1Start, s1End)) 2109 continue; 2110 2111 // Perform the hoisting by eliminating (2) and changing the start 2112 // location of (3). 2113 PieceNextI->setStartLocation(PieceI->getStartLocation()); 2114 I = pieces.erase(I); 2115 } 2116 } 2117 2118 /// Returns the number of bytes in the given (character-based) SourceRange. 2119 /// 2120 /// If the locations in the range are not on the same line, returns None. 2121 /// 2122 /// Note that this does not do a precise user-visible character or column count. 2123 static Optional<size_t> getLengthOnSingleLine(SourceManager &SM, 2124 SourceRange Range) { 2125 SourceRange ExpansionRange(SM.getExpansionLoc(Range.getBegin()), 2126 SM.getExpansionRange(Range.getEnd()).second); 2127 2128 FileID FID = SM.getFileID(ExpansionRange.getBegin()); 2129 if (FID != SM.getFileID(ExpansionRange.getEnd())) 2130 return None; 2131 2132 bool Invalid; 2133 const llvm::MemoryBuffer *Buffer = SM.getBuffer(FID, &Invalid); 2134 if (Invalid) 2135 return None; 2136 2137 unsigned BeginOffset = SM.getFileOffset(ExpansionRange.getBegin()); 2138 unsigned EndOffset = SM.getFileOffset(ExpansionRange.getEnd()); 2139 StringRef Snippet = Buffer->getBuffer().slice(BeginOffset, EndOffset); 2140 2141 // We're searching the raw bytes of the buffer here, which might include 2142 // escaped newlines and such. That's okay; we're trying to decide whether the 2143 // SourceRange is covering a large or small amount of space in the user's 2144 // editor. 2145 if (Snippet.find_first_of("\r\n") != StringRef::npos) 2146 return None; 2147 2148 // This isn't Unicode-aware, but it doesn't need to be. 2149 return Snippet.size(); 2150 } 2151 2152 /// \sa getLengthOnSingleLine(SourceManager, SourceRange) 2153 static Optional<size_t> getLengthOnSingleLine(SourceManager &SM, 2154 const Stmt *S) { 2155 return getLengthOnSingleLine(SM, S->getSourceRange()); 2156 } 2157 2158 /// Eliminate two-edge cycles created by addContextEdges(). 2159 /// 2160 /// Once all the context edges are in place, there are plenty of cases where 2161 /// there's a single edge from a top-level statement to a subexpression, 2162 /// followed by a single path note, and then a reverse edge to get back out to 2163 /// the top level. If the statement is simple enough, the subexpression edges 2164 /// just add noise and make it harder to understand what's going on. 2165 /// 2166 /// This function only removes edges in pairs, because removing only one edge 2167 /// might leave other edges dangling. 2168 /// 2169 /// This will not remove edges in more complicated situations: 2170 /// - if there is more than one "hop" leading to or from a subexpression. 2171 /// - if there is an inlined call between the edges instead of a single event. 2172 /// - if the whole statement is large enough that having subexpression arrows 2173 /// might be helpful. 2174 static void removeContextCycles(PathPieces &Path, SourceManager &SM, 2175 ParentMap &PM) { 2176 for (PathPieces::iterator I = Path.begin(), E = Path.end(); I != E; ) { 2177 // Pattern match the current piece and its successor. 2178 PathDiagnosticControlFlowPiece *PieceI = 2179 dyn_cast<PathDiagnosticControlFlowPiece>(*I); 2180 2181 if (!PieceI) { 2182 ++I; 2183 continue; 2184 } 2185 2186 const Stmt *s1Start = getLocStmt(PieceI->getStartLocation()); 2187 const Stmt *s1End = getLocStmt(PieceI->getEndLocation()); 2188 2189 PathPieces::iterator NextI = I; ++NextI; 2190 if (NextI == E) 2191 break; 2192 2193 PathDiagnosticControlFlowPiece *PieceNextI = 2194 dyn_cast<PathDiagnosticControlFlowPiece>(*NextI); 2195 2196 if (!PieceNextI) { 2197 if (isa<PathDiagnosticEventPiece>(*NextI)) { 2198 ++NextI; 2199 if (NextI == E) 2200 break; 2201 PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(*NextI); 2202 } 2203 2204 if (!PieceNextI) { 2205 ++I; 2206 continue; 2207 } 2208 } 2209 2210 const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation()); 2211 const Stmt *s2End = getLocStmt(PieceNextI->getEndLocation()); 2212 2213 if (s1Start && s2Start && s1Start == s2End && s2Start == s1End) { 2214 const size_t MAX_SHORT_LINE_LENGTH = 80; 2215 Optional<size_t> s1Length = getLengthOnSingleLine(SM, s1Start); 2216 if (s1Length && *s1Length <= MAX_SHORT_LINE_LENGTH) { 2217 Optional<size_t> s2Length = getLengthOnSingleLine(SM, s2Start); 2218 if (s2Length && *s2Length <= MAX_SHORT_LINE_LENGTH) { 2219 Path.erase(I); 2220 I = Path.erase(NextI); 2221 continue; 2222 } 2223 } 2224 } 2225 2226 ++I; 2227 } 2228 } 2229 2230 /// \brief Return true if X is contained by Y. 2231 static bool lexicalContains(ParentMap &PM, 2232 const Stmt *X, 2233 const Stmt *Y) { 2234 while (X) { 2235 if (X == Y) 2236 return true; 2237 X = PM.getParent(X); 2238 } 2239 return false; 2240 } 2241 2242 // Remove short edges on the same line less than 3 columns in difference. 2243 static void removePunyEdges(PathPieces &path, 2244 SourceManager &SM, 2245 ParentMap &PM) { 2246 2247 bool erased = false; 2248 2249 for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; 2250 erased ? I : ++I) { 2251 2252 erased = false; 2253 2254 PathDiagnosticControlFlowPiece *PieceI = 2255 dyn_cast<PathDiagnosticControlFlowPiece>(*I); 2256 2257 if (!PieceI) 2258 continue; 2259 2260 const Stmt *start = getLocStmt(PieceI->getStartLocation()); 2261 const Stmt *end = getLocStmt(PieceI->getEndLocation()); 2262 2263 if (!start || !end) 2264 continue; 2265 2266 const Stmt *endParent = PM.getParent(end); 2267 if (!endParent) 2268 continue; 2269 2270 if (isConditionForTerminator(end, endParent)) 2271 continue; 2272 2273 SourceLocation FirstLoc = start->getLocStart(); 2274 SourceLocation SecondLoc = end->getLocStart(); 2275 2276 if (!SM.isWrittenInSameFile(FirstLoc, SecondLoc)) 2277 continue; 2278 if (SM.isBeforeInTranslationUnit(SecondLoc, FirstLoc)) 2279 std::swap(SecondLoc, FirstLoc); 2280 2281 SourceRange EdgeRange(FirstLoc, SecondLoc); 2282 Optional<size_t> ByteWidth = getLengthOnSingleLine(SM, EdgeRange); 2283 2284 // If the statements are on different lines, continue. 2285 if (!ByteWidth) 2286 continue; 2287 2288 const size_t MAX_PUNY_EDGE_LENGTH = 2; 2289 if (*ByteWidth <= MAX_PUNY_EDGE_LENGTH) { 2290 // FIXME: There are enough /bytes/ between the endpoints of the edge, but 2291 // there might not be enough /columns/. A proper user-visible column count 2292 // is probably too expensive, though. 2293 I = path.erase(I); 2294 erased = true; 2295 continue; 2296 } 2297 } 2298 } 2299 2300 static void removeIdenticalEvents(PathPieces &path) { 2301 for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ++I) { 2302 PathDiagnosticEventPiece *PieceI = 2303 dyn_cast<PathDiagnosticEventPiece>(*I); 2304 2305 if (!PieceI) 2306 continue; 2307 2308 PathPieces::iterator NextI = I; ++NextI; 2309 if (NextI == E) 2310 return; 2311 2312 PathDiagnosticEventPiece *PieceNextI = 2313 dyn_cast<PathDiagnosticEventPiece>(*NextI); 2314 2315 if (!PieceNextI) 2316 continue; 2317 2318 // Erase the second piece if it has the same exact message text. 2319 if (PieceI->getString() == PieceNextI->getString()) { 2320 path.erase(NextI); 2321 } 2322 } 2323 } 2324 2325 static bool optimizeEdges(PathPieces &path, SourceManager &SM, 2326 OptimizedCallsSet &OCS, 2327 LocationContextMap &LCM) { 2328 bool hasChanges = false; 2329 const LocationContext *LC = LCM[&path]; 2330 assert(LC); 2331 ParentMap &PM = LC->getParentMap(); 2332 2333 for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ) { 2334 // Optimize subpaths. 2335 if (PathDiagnosticCallPiece *CallI = dyn_cast<PathDiagnosticCallPiece>(*I)){ 2336 // Record the fact that a call has been optimized so we only do the 2337 // effort once. 2338 if (!OCS.count(CallI)) { 2339 while (optimizeEdges(CallI->path, SM, OCS, LCM)) {} 2340 OCS.insert(CallI); 2341 } 2342 ++I; 2343 continue; 2344 } 2345 2346 // Pattern match the current piece and its successor. 2347 PathDiagnosticControlFlowPiece *PieceI = 2348 dyn_cast<PathDiagnosticControlFlowPiece>(*I); 2349 2350 if (!PieceI) { 2351 ++I; 2352 continue; 2353 } 2354 2355 const Stmt *s1Start = getLocStmt(PieceI->getStartLocation()); 2356 const Stmt *s1End = getLocStmt(PieceI->getEndLocation()); 2357 const Stmt *level1 = getStmtParent(s1Start, PM); 2358 const Stmt *level2 = getStmtParent(s1End, PM); 2359 2360 PathPieces::iterator NextI = I; ++NextI; 2361 if (NextI == E) 2362 break; 2363 2364 PathDiagnosticControlFlowPiece *PieceNextI = 2365 dyn_cast<PathDiagnosticControlFlowPiece>(*NextI); 2366 2367 if (!PieceNextI) { 2368 ++I; 2369 continue; 2370 } 2371 2372 const Stmt *s2Start = getLocStmt(PieceNextI->getStartLocation()); 2373 const Stmt *s2End = getLocStmt(PieceNextI->getEndLocation()); 2374 const Stmt *level3 = getStmtParent(s2Start, PM); 2375 const Stmt *level4 = getStmtParent(s2End, PM); 2376 2377 // Rule I. 2378 // 2379 // If we have two consecutive control edges whose end/begin locations 2380 // are at the same level (e.g. statements or top-level expressions within 2381 // a compound statement, or siblings share a single ancestor expression), 2382 // then merge them if they have no interesting intermediate event. 2383 // 2384 // For example: 2385 // 2386 // (1.1 -> 1.2) -> (1.2 -> 1.3) becomes (1.1 -> 1.3) because the common 2387 // parent is '1'. Here 'x.y.z' represents the hierarchy of statements. 2388 // 2389 // NOTE: this will be limited later in cases where we add barriers 2390 // to prevent this optimization. 2391 // 2392 if (level1 && level1 == level2 && level1 == level3 && level1 == level4) { 2393 PieceI->setEndLocation(PieceNextI->getEndLocation()); 2394 path.erase(NextI); 2395 hasChanges = true; 2396 continue; 2397 } 2398 2399 // Rule II. 2400 // 2401 // Eliminate edges between subexpressions and parent expressions 2402 // when the subexpression is consumed. 2403 // 2404 // NOTE: this will be limited later in cases where we add barriers 2405 // to prevent this optimization. 2406 // 2407 if (s1End && s1End == s2Start && level2) { 2408 bool removeEdge = false; 2409 // Remove edges into the increment or initialization of a 2410 // loop that have no interleaving event. This means that 2411 // they aren't interesting. 2412 if (isIncrementOrInitInForLoop(s1End, level2)) 2413 removeEdge = true; 2414 // Next only consider edges that are not anchored on 2415 // the condition of a terminator. This are intermediate edges 2416 // that we might want to trim. 2417 else if (!isConditionForTerminator(level2, s1End)) { 2418 // Trim edges on expressions that are consumed by 2419 // the parent expression. 2420 if (isa<Expr>(s1End) && PM.isConsumedExpr(cast<Expr>(s1End))) { 2421 removeEdge = true; 2422 } 2423 // Trim edges where a lexical containment doesn't exist. 2424 // For example: 2425 // 2426 // X -> Y -> Z 2427 // 2428 // If 'Z' lexically contains Y (it is an ancestor) and 2429 // 'X' does not lexically contain Y (it is a descendant OR 2430 // it has no lexical relationship at all) then trim. 2431 // 2432 // This can eliminate edges where we dive into a subexpression 2433 // and then pop back out, etc. 2434 else if (s1Start && s2End && 2435 lexicalContains(PM, s2Start, s2End) && 2436 !lexicalContains(PM, s1End, s1Start)) { 2437 removeEdge = true; 2438 } 2439 // Trim edges from a subexpression back to the top level if the 2440 // subexpression is on a different line. 2441 // 2442 // A.1 -> A -> B 2443 // becomes 2444 // A.1 -> B 2445 // 2446 // These edges just look ugly and don't usually add anything. 2447 else if (s1Start && s2End && 2448 lexicalContains(PM, s1Start, s1End)) { 2449 SourceRange EdgeRange(PieceI->getEndLocation().asLocation(), 2450 PieceI->getStartLocation().asLocation()); 2451 if (!getLengthOnSingleLine(SM, EdgeRange).hasValue()) 2452 removeEdge = true; 2453 } 2454 } 2455 2456 if (removeEdge) { 2457 PieceI->setEndLocation(PieceNextI->getEndLocation()); 2458 path.erase(NextI); 2459 hasChanges = true; 2460 continue; 2461 } 2462 } 2463 2464 // Optimize edges for ObjC fast-enumeration loops. 2465 // 2466 // (X -> collection) -> (collection -> element) 2467 // 2468 // becomes: 2469 // 2470 // (X -> element) 2471 if (s1End == s2Start) { 2472 const ObjCForCollectionStmt *FS = 2473 dyn_cast_or_null<ObjCForCollectionStmt>(level3); 2474 if (FS && FS->getCollection()->IgnoreParens() == s2Start && 2475 s2End == FS->getElement()) { 2476 PieceI->setEndLocation(PieceNextI->getEndLocation()); 2477 path.erase(NextI); 2478 hasChanges = true; 2479 continue; 2480 } 2481 } 2482 2483 // No changes at this index? Move to the next one. 2484 ++I; 2485 } 2486 2487 if (!hasChanges) { 2488 // Adjust edges into subexpressions to make them more uniform 2489 // and aesthetically pleasing. 2490 addContextEdges(path, SM, PM, LC); 2491 // Remove "cyclical" edges that include one or more context edges. 2492 removeContextCycles(path, SM, PM); 2493 // Hoist edges originating from branch conditions to branches 2494 // for simple branches. 2495 simplifySimpleBranches(path); 2496 // Remove any puny edges left over after primary optimization pass. 2497 removePunyEdges(path, SM, PM); 2498 // Remove identical events. 2499 removeIdenticalEvents(path); 2500 } 2501 2502 return hasChanges; 2503 } 2504 2505 /// Drop the very first edge in a path, which should be a function entry edge. 2506 /// 2507 /// If the first edge is not a function entry edge (say, because the first 2508 /// statement had an invalid source location), this function does nothing. 2509 // FIXME: We should just generate invalid edges anyway and have the optimizer 2510 // deal with them. 2511 static void dropFunctionEntryEdge(PathPieces &Path, 2512 LocationContextMap &LCM, 2513 SourceManager &SM) { 2514 const PathDiagnosticControlFlowPiece *FirstEdge = 2515 dyn_cast<PathDiagnosticControlFlowPiece>(Path.front()); 2516 if (!FirstEdge) 2517 return; 2518 2519 const Decl *D = LCM[&Path]->getDecl(); 2520 PathDiagnosticLocation EntryLoc = PathDiagnosticLocation::createBegin(D, SM); 2521 if (FirstEdge->getStartLocation() != EntryLoc) 2522 return; 2523 2524 Path.pop_front(); 2525 } 2526 2527 2528 //===----------------------------------------------------------------------===// 2529 // Methods for BugType and subclasses. 2530 //===----------------------------------------------------------------------===// 2531 void BugType::anchor() { } 2532 2533 void BugType::FlushReports(BugReporter &BR) {} 2534 2535 void BuiltinBug::anchor() {} 2536 2537 //===----------------------------------------------------------------------===// 2538 // Methods for BugReport and subclasses. 2539 //===----------------------------------------------------------------------===// 2540 2541 void BugReport::NodeResolver::anchor() {} 2542 2543 void BugReport::addVisitor(std::unique_ptr<BugReporterVisitor> visitor) { 2544 if (!visitor) 2545 return; 2546 2547 llvm::FoldingSetNodeID ID; 2548 visitor->Profile(ID); 2549 void *InsertPos; 2550 2551 if (CallbacksSet.FindNodeOrInsertPos(ID, InsertPos)) 2552 return; 2553 2554 CallbacksSet.InsertNode(visitor.get(), InsertPos); 2555 Callbacks.push_back(std::move(visitor)); 2556 ++ConfigurationChangeToken; 2557 } 2558 2559 BugReport::~BugReport() { 2560 while (!interestingSymbols.empty()) { 2561 popInterestingSymbolsAndRegions(); 2562 } 2563 } 2564 2565 const Decl *BugReport::getDeclWithIssue() const { 2566 if (DeclWithIssue) 2567 return DeclWithIssue; 2568 2569 const ExplodedNode *N = getErrorNode(); 2570 if (!N) 2571 return nullptr; 2572 2573 const LocationContext *LC = N->getLocationContext(); 2574 return LC->getCurrentStackFrame()->getDecl(); 2575 } 2576 2577 void BugReport::Profile(llvm::FoldingSetNodeID& hash) const { 2578 hash.AddPointer(&BT); 2579 hash.AddString(Description); 2580 PathDiagnosticLocation UL = getUniqueingLocation(); 2581 if (UL.isValid()) { 2582 UL.Profile(hash); 2583 } else if (Location.isValid()) { 2584 Location.Profile(hash); 2585 } else { 2586 assert(ErrorNode); 2587 hash.AddPointer(GetCurrentOrPreviousStmt(ErrorNode)); 2588 } 2589 2590 for (SourceRange range : Ranges) { 2591 if (!range.isValid()) 2592 continue; 2593 hash.AddInteger(range.getBegin().getRawEncoding()); 2594 hash.AddInteger(range.getEnd().getRawEncoding()); 2595 } 2596 } 2597 2598 void BugReport::markInteresting(SymbolRef sym) { 2599 if (!sym) 2600 return; 2601 2602 // If the symbol wasn't already in our set, note a configuration change. 2603 if (getInterestingSymbols().insert(sym).second) 2604 ++ConfigurationChangeToken; 2605 2606 if (const SymbolMetadata *meta = dyn_cast<SymbolMetadata>(sym)) 2607 getInterestingRegions().insert(meta->getRegion()); 2608 } 2609 2610 void BugReport::markInteresting(const MemRegion *R) { 2611 if (!R) 2612 return; 2613 2614 // If the base region wasn't already in our set, note a configuration change. 2615 R = R->getBaseRegion(); 2616 if (getInterestingRegions().insert(R).second) 2617 ++ConfigurationChangeToken; 2618 2619 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) 2620 getInterestingSymbols().insert(SR->getSymbol()); 2621 } 2622 2623 void BugReport::markInteresting(SVal V) { 2624 markInteresting(V.getAsRegion()); 2625 markInteresting(V.getAsSymbol()); 2626 } 2627 2628 void BugReport::markInteresting(const LocationContext *LC) { 2629 if (!LC) 2630 return; 2631 InterestingLocationContexts.insert(LC); 2632 } 2633 2634 bool BugReport::isInteresting(SVal V) { 2635 return isInteresting(V.getAsRegion()) || isInteresting(V.getAsSymbol()); 2636 } 2637 2638 bool BugReport::isInteresting(SymbolRef sym) { 2639 if (!sym) 2640 return false; 2641 // We don't currently consider metadata symbols to be interesting 2642 // even if we know their region is interesting. Is that correct behavior? 2643 return getInterestingSymbols().count(sym); 2644 } 2645 2646 bool BugReport::isInteresting(const MemRegion *R) { 2647 if (!R) 2648 return false; 2649 R = R->getBaseRegion(); 2650 bool b = getInterestingRegions().count(R); 2651 if (b) 2652 return true; 2653 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) 2654 return getInterestingSymbols().count(SR->getSymbol()); 2655 return false; 2656 } 2657 2658 bool BugReport::isInteresting(const LocationContext *LC) { 2659 if (!LC) 2660 return false; 2661 return InterestingLocationContexts.count(LC); 2662 } 2663 2664 void BugReport::lazyInitializeInterestingSets() { 2665 if (interestingSymbols.empty()) { 2666 interestingSymbols.push_back(new Symbols()); 2667 interestingRegions.push_back(new Regions()); 2668 } 2669 } 2670 2671 BugReport::Symbols &BugReport::getInterestingSymbols() { 2672 lazyInitializeInterestingSets(); 2673 return *interestingSymbols.back(); 2674 } 2675 2676 BugReport::Regions &BugReport::getInterestingRegions() { 2677 lazyInitializeInterestingSets(); 2678 return *interestingRegions.back(); 2679 } 2680 2681 void BugReport::pushInterestingSymbolsAndRegions() { 2682 interestingSymbols.push_back(new Symbols(getInterestingSymbols())); 2683 interestingRegions.push_back(new Regions(getInterestingRegions())); 2684 } 2685 2686 void BugReport::popInterestingSymbolsAndRegions() { 2687 delete interestingSymbols.pop_back_val(); 2688 delete interestingRegions.pop_back_val(); 2689 } 2690 2691 const Stmt *BugReport::getStmt() const { 2692 if (!ErrorNode) 2693 return nullptr; 2694 2695 ProgramPoint ProgP = ErrorNode->getLocation(); 2696 const Stmt *S = nullptr; 2697 2698 if (Optional<BlockEntrance> BE = ProgP.getAs<BlockEntrance>()) { 2699 CFGBlock &Exit = ProgP.getLocationContext()->getCFG()->getExit(); 2700 if (BE->getBlock() == &Exit) 2701 S = GetPreviousStmt(ErrorNode); 2702 } 2703 if (!S) 2704 S = PathDiagnosticLocation::getStmt(ErrorNode); 2705 2706 return S; 2707 } 2708 2709 llvm::iterator_range<BugReport::ranges_iterator> BugReport::getRanges() { 2710 // If no custom ranges, add the range of the statement corresponding to 2711 // the error node. 2712 if (Ranges.empty()) { 2713 if (const Expr *E = dyn_cast_or_null<Expr>(getStmt())) 2714 addRange(E->getSourceRange()); 2715 else 2716 return llvm::make_range(ranges_iterator(), ranges_iterator()); 2717 } 2718 2719 // User-specified absence of range info. 2720 if (Ranges.size() == 1 && !Ranges.begin()->isValid()) 2721 return llvm::make_range(ranges_iterator(), ranges_iterator()); 2722 2723 return llvm::make_range(Ranges.begin(), Ranges.end()); 2724 } 2725 2726 PathDiagnosticLocation BugReport::getLocation(const SourceManager &SM) const { 2727 if (ErrorNode) { 2728 assert(!Location.isValid() && 2729 "Either Location or ErrorNode should be specified but not both."); 2730 return PathDiagnosticLocation::createEndOfPath(ErrorNode, SM); 2731 } 2732 2733 assert(Location.isValid()); 2734 return Location; 2735 } 2736 2737 //===----------------------------------------------------------------------===// 2738 // Methods for BugReporter and subclasses. 2739 //===----------------------------------------------------------------------===// 2740 2741 BugReportEquivClass::~BugReportEquivClass() { } 2742 GRBugReporter::~GRBugReporter() { } 2743 BugReporterData::~BugReporterData() {} 2744 2745 ExplodedGraph &GRBugReporter::getGraph() { return Eng.getGraph(); } 2746 2747 ProgramStateManager& 2748 GRBugReporter::getStateManager() { return Eng.getStateManager(); } 2749 2750 BugReporter::~BugReporter() { 2751 FlushReports(); 2752 2753 // Free the bug reports we are tracking. 2754 typedef std::vector<BugReportEquivClass *> ContTy; 2755 for (ContTy::iterator I = EQClassesVector.begin(), E = EQClassesVector.end(); 2756 I != E; ++I) { 2757 delete *I; 2758 } 2759 } 2760 2761 void BugReporter::FlushReports() { 2762 if (BugTypes.isEmpty()) 2763 return; 2764 2765 // First flush the warnings for each BugType. This may end up creating new 2766 // warnings and new BugTypes. 2767 // FIXME: Only NSErrorChecker needs BugType's FlushReports. 2768 // Turn NSErrorChecker into a proper checker and remove this. 2769 SmallVector<const BugType *, 16> bugTypes(BugTypes.begin(), BugTypes.end()); 2770 for (SmallVectorImpl<const BugType *>::iterator 2771 I = bugTypes.begin(), E = bugTypes.end(); I != E; ++I) 2772 const_cast<BugType*>(*I)->FlushReports(*this); 2773 2774 // We need to flush reports in deterministic order to ensure the order 2775 // of the reports is consistent between runs. 2776 typedef std::vector<BugReportEquivClass *> ContVecTy; 2777 for (ContVecTy::iterator EI=EQClassesVector.begin(), EE=EQClassesVector.end(); 2778 EI != EE; ++EI){ 2779 BugReportEquivClass& EQ = **EI; 2780 FlushReport(EQ); 2781 } 2782 2783 // BugReporter owns and deletes only BugTypes created implicitly through 2784 // EmitBasicReport. 2785 // FIXME: There are leaks from checkers that assume that the BugTypes they 2786 // create will be destroyed by the BugReporter. 2787 llvm::DeleteContainerSeconds(StrBugTypes); 2788 2789 // Remove all references to the BugType objects. 2790 BugTypes = F.getEmptySet(); 2791 } 2792 2793 //===----------------------------------------------------------------------===// 2794 // PathDiagnostics generation. 2795 //===----------------------------------------------------------------------===// 2796 2797 namespace { 2798 /// A wrapper around a report graph, which contains only a single path, and its 2799 /// node maps. 2800 class ReportGraph { 2801 public: 2802 InterExplodedGraphMap BackMap; 2803 std::unique_ptr<ExplodedGraph> Graph; 2804 const ExplodedNode *ErrorNode; 2805 size_t Index; 2806 }; 2807 2808 /// A wrapper around a trimmed graph and its node maps. 2809 class TrimmedGraph { 2810 InterExplodedGraphMap InverseMap; 2811 2812 typedef llvm::DenseMap<const ExplodedNode *, unsigned> PriorityMapTy; 2813 PriorityMapTy PriorityMap; 2814 2815 typedef std::pair<const ExplodedNode *, size_t> NodeIndexPair; 2816 SmallVector<NodeIndexPair, 32> ReportNodes; 2817 2818 std::unique_ptr<ExplodedGraph> G; 2819 2820 /// A helper class for sorting ExplodedNodes by priority. 2821 template <bool Descending> 2822 class PriorityCompare { 2823 const PriorityMapTy &PriorityMap; 2824 2825 public: 2826 PriorityCompare(const PriorityMapTy &M) : PriorityMap(M) {} 2827 2828 bool operator()(const ExplodedNode *LHS, const ExplodedNode *RHS) const { 2829 PriorityMapTy::const_iterator LI = PriorityMap.find(LHS); 2830 PriorityMapTy::const_iterator RI = PriorityMap.find(RHS); 2831 PriorityMapTy::const_iterator E = PriorityMap.end(); 2832 2833 if (LI == E) 2834 return Descending; 2835 if (RI == E) 2836 return !Descending; 2837 2838 return Descending ? LI->second > RI->second 2839 : LI->second < RI->second; 2840 } 2841 2842 bool operator()(const NodeIndexPair &LHS, const NodeIndexPair &RHS) const { 2843 return (*this)(LHS.first, RHS.first); 2844 } 2845 }; 2846 2847 public: 2848 TrimmedGraph(const ExplodedGraph *OriginalGraph, 2849 ArrayRef<const ExplodedNode *> Nodes); 2850 2851 bool popNextReportGraph(ReportGraph &GraphWrapper); 2852 }; 2853 } 2854 2855 TrimmedGraph::TrimmedGraph(const ExplodedGraph *OriginalGraph, 2856 ArrayRef<const ExplodedNode *> Nodes) { 2857 // The trimmed graph is created in the body of the constructor to ensure 2858 // that the DenseMaps have been initialized already. 2859 InterExplodedGraphMap ForwardMap; 2860 G = OriginalGraph->trim(Nodes, &ForwardMap, &InverseMap); 2861 2862 // Find the (first) error node in the trimmed graph. We just need to consult 2863 // the node map which maps from nodes in the original graph to nodes 2864 // in the new graph. 2865 llvm::SmallPtrSet<const ExplodedNode *, 32> RemainingNodes; 2866 2867 for (unsigned i = 0, count = Nodes.size(); i < count; ++i) { 2868 if (const ExplodedNode *NewNode = ForwardMap.lookup(Nodes[i])) { 2869 ReportNodes.push_back(std::make_pair(NewNode, i)); 2870 RemainingNodes.insert(NewNode); 2871 } 2872 } 2873 2874 assert(!RemainingNodes.empty() && "No error node found in the trimmed graph"); 2875 2876 // Perform a forward BFS to find all the shortest paths. 2877 std::queue<const ExplodedNode *> WS; 2878 2879 assert(G->num_roots() == 1); 2880 WS.push(*G->roots_begin()); 2881 unsigned Priority = 0; 2882 2883 while (!WS.empty()) { 2884 const ExplodedNode *Node = WS.front(); 2885 WS.pop(); 2886 2887 PriorityMapTy::iterator PriorityEntry; 2888 bool IsNew; 2889 std::tie(PriorityEntry, IsNew) = 2890 PriorityMap.insert(std::make_pair(Node, Priority)); 2891 ++Priority; 2892 2893 if (!IsNew) { 2894 assert(PriorityEntry->second <= Priority); 2895 continue; 2896 } 2897 2898 if (RemainingNodes.erase(Node)) 2899 if (RemainingNodes.empty()) 2900 break; 2901 2902 for (ExplodedNode::const_pred_iterator I = Node->succ_begin(), 2903 E = Node->succ_end(); 2904 I != E; ++I) 2905 WS.push(*I); 2906 } 2907 2908 // Sort the error paths from longest to shortest. 2909 std::sort(ReportNodes.begin(), ReportNodes.end(), 2910 PriorityCompare<true>(PriorityMap)); 2911 } 2912 2913 bool TrimmedGraph::popNextReportGraph(ReportGraph &GraphWrapper) { 2914 if (ReportNodes.empty()) 2915 return false; 2916 2917 const ExplodedNode *OrigN; 2918 std::tie(OrigN, GraphWrapper.Index) = ReportNodes.pop_back_val(); 2919 assert(PriorityMap.find(OrigN) != PriorityMap.end() && 2920 "error node not accessible from root"); 2921 2922 // Create a new graph with a single path. This is the graph 2923 // that will be returned to the caller. 2924 auto GNew = llvm::make_unique<ExplodedGraph>(); 2925 GraphWrapper.BackMap.clear(); 2926 2927 // Now walk from the error node up the BFS path, always taking the 2928 // predeccessor with the lowest number. 2929 ExplodedNode *Succ = nullptr; 2930 while (true) { 2931 // Create the equivalent node in the new graph with the same state 2932 // and location. 2933 ExplodedNode *NewN = GNew->createUncachedNode(OrigN->getLocation(), OrigN->getState(), 2934 OrigN->isSink()); 2935 2936 // Store the mapping to the original node. 2937 InterExplodedGraphMap::const_iterator IMitr = InverseMap.find(OrigN); 2938 assert(IMitr != InverseMap.end() && "No mapping to original node."); 2939 GraphWrapper.BackMap[NewN] = IMitr->second; 2940 2941 // Link up the new node with the previous node. 2942 if (Succ) 2943 Succ->addPredecessor(NewN, *GNew); 2944 else 2945 GraphWrapper.ErrorNode = NewN; 2946 2947 Succ = NewN; 2948 2949 // Are we at the final node? 2950 if (OrigN->pred_empty()) { 2951 GNew->addRoot(NewN); 2952 break; 2953 } 2954 2955 // Find the next predeccessor node. We choose the node that is marked 2956 // with the lowest BFS number. 2957 OrigN = *std::min_element(OrigN->pred_begin(), OrigN->pred_end(), 2958 PriorityCompare<false>(PriorityMap)); 2959 } 2960 2961 GraphWrapper.Graph = std::move(GNew); 2962 2963 return true; 2964 } 2965 2966 2967 /// CompactPathDiagnostic - This function postprocesses a PathDiagnostic object 2968 /// and collapses PathDiagosticPieces that are expanded by macros. 2969 static void CompactPathDiagnostic(PathPieces &path, const SourceManager& SM) { 2970 typedef std::vector<std::pair<IntrusiveRefCntPtr<PathDiagnosticMacroPiece>, 2971 SourceLocation> > MacroStackTy; 2972 2973 typedef std::vector<IntrusiveRefCntPtr<PathDiagnosticPiece> > 2974 PiecesTy; 2975 2976 MacroStackTy MacroStack; 2977 PiecesTy Pieces; 2978 2979 for (PathPieces::const_iterator I = path.begin(), E = path.end(); 2980 I!=E; ++I) { 2981 2982 PathDiagnosticPiece *piece = I->get(); 2983 2984 // Recursively compact calls. 2985 if (PathDiagnosticCallPiece *call=dyn_cast<PathDiagnosticCallPiece>(piece)){ 2986 CompactPathDiagnostic(call->path, SM); 2987 } 2988 2989 // Get the location of the PathDiagnosticPiece. 2990 const FullSourceLoc Loc = piece->getLocation().asLocation(); 2991 2992 // Determine the instantiation location, which is the location we group 2993 // related PathDiagnosticPieces. 2994 SourceLocation InstantiationLoc = Loc.isMacroID() ? 2995 SM.getExpansionLoc(Loc) : 2996 SourceLocation(); 2997 2998 if (Loc.isFileID()) { 2999 MacroStack.clear(); 3000 Pieces.push_back(piece); 3001 continue; 3002 } 3003 3004 assert(Loc.isMacroID()); 3005 3006 // Is the PathDiagnosticPiece within the same macro group? 3007 if (!MacroStack.empty() && InstantiationLoc == MacroStack.back().second) { 3008 MacroStack.back().first->subPieces.push_back(piece); 3009 continue; 3010 } 3011 3012 // We aren't in the same group. Are we descending into a new macro 3013 // or are part of an old one? 3014 IntrusiveRefCntPtr<PathDiagnosticMacroPiece> MacroGroup; 3015 3016 SourceLocation ParentInstantiationLoc = InstantiationLoc.isMacroID() ? 3017 SM.getExpansionLoc(Loc) : 3018 SourceLocation(); 3019 3020 // Walk the entire macro stack. 3021 while (!MacroStack.empty()) { 3022 if (InstantiationLoc == MacroStack.back().second) { 3023 MacroGroup = MacroStack.back().first; 3024 break; 3025 } 3026 3027 if (ParentInstantiationLoc == MacroStack.back().second) { 3028 MacroGroup = MacroStack.back().first; 3029 break; 3030 } 3031 3032 MacroStack.pop_back(); 3033 } 3034 3035 if (!MacroGroup || ParentInstantiationLoc == MacroStack.back().second) { 3036 // Create a new macro group and add it to the stack. 3037 PathDiagnosticMacroPiece *NewGroup = 3038 new PathDiagnosticMacroPiece( 3039 PathDiagnosticLocation::createSingleLocation(piece->getLocation())); 3040 3041 if (MacroGroup) 3042 MacroGroup->subPieces.push_back(NewGroup); 3043 else { 3044 assert(InstantiationLoc.isFileID()); 3045 Pieces.push_back(NewGroup); 3046 } 3047 3048 MacroGroup = NewGroup; 3049 MacroStack.push_back(std::make_pair(MacroGroup, InstantiationLoc)); 3050 } 3051 3052 // Finally, add the PathDiagnosticPiece to the group. 3053 MacroGroup->subPieces.push_back(piece); 3054 } 3055 3056 // Now take the pieces and construct a new PathDiagnostic. 3057 path.clear(); 3058 3059 path.insert(path.end(), Pieces.begin(), Pieces.end()); 3060 } 3061 3062 bool GRBugReporter::generatePathDiagnostic(PathDiagnostic& PD, 3063 PathDiagnosticConsumer &PC, 3064 ArrayRef<BugReport *> &bugReports) { 3065 assert(!bugReports.empty()); 3066 3067 bool HasValid = false; 3068 bool HasInvalid = false; 3069 SmallVector<const ExplodedNode *, 32> errorNodes; 3070 for (ArrayRef<BugReport*>::iterator I = bugReports.begin(), 3071 E = bugReports.end(); I != E; ++I) { 3072 if ((*I)->isValid()) { 3073 HasValid = true; 3074 errorNodes.push_back((*I)->getErrorNode()); 3075 } else { 3076 // Keep the errorNodes list in sync with the bugReports list. 3077 HasInvalid = true; 3078 errorNodes.push_back(nullptr); 3079 } 3080 } 3081 3082 // If all the reports have been marked invalid by a previous path generation, 3083 // we're done. 3084 if (!HasValid) 3085 return false; 3086 3087 typedef PathDiagnosticConsumer::PathGenerationScheme PathGenerationScheme; 3088 PathGenerationScheme ActiveScheme = PC.getGenerationScheme(); 3089 3090 if (ActiveScheme == PathDiagnosticConsumer::Extensive) { 3091 AnalyzerOptions &options = getAnalyzerOptions(); 3092 if (options.getBooleanOption("path-diagnostics-alternate", true)) { 3093 ActiveScheme = PathDiagnosticConsumer::AlternateExtensive; 3094 } 3095 } 3096 3097 TrimmedGraph TrimG(&getGraph(), errorNodes); 3098 ReportGraph ErrorGraph; 3099 3100 while (TrimG.popNextReportGraph(ErrorGraph)) { 3101 // Find the BugReport with the original location. 3102 assert(ErrorGraph.Index < bugReports.size()); 3103 BugReport *R = bugReports[ErrorGraph.Index]; 3104 assert(R && "No original report found for sliced graph."); 3105 assert(R->isValid() && "Report selected by trimmed graph marked invalid."); 3106 3107 // Start building the path diagnostic... 3108 PathDiagnosticBuilder PDB(*this, R, ErrorGraph.BackMap, &PC); 3109 const ExplodedNode *N = ErrorGraph.ErrorNode; 3110 3111 // Register additional node visitors. 3112 R->addVisitor(llvm::make_unique<NilReceiverBRVisitor>()); 3113 R->addVisitor(llvm::make_unique<ConditionBRVisitor>()); 3114 R->addVisitor(llvm::make_unique<LikelyFalsePositiveSuppressionBRVisitor>()); 3115 R->addVisitor(llvm::make_unique<CXXSelfAssignmentBRVisitor>()); 3116 3117 BugReport::VisitorList visitors; 3118 unsigned origReportConfigToken, finalReportConfigToken; 3119 LocationContextMap LCM; 3120 3121 // While generating diagnostics, it's possible the visitors will decide 3122 // new symbols and regions are interesting, or add other visitors based on 3123 // the information they find. If they do, we need to regenerate the path 3124 // based on our new report configuration. 3125 do { 3126 // Get a clean copy of all the visitors. 3127 for (BugReport::visitor_iterator I = R->visitor_begin(), 3128 E = R->visitor_end(); I != E; ++I) 3129 visitors.push_back((*I)->clone()); 3130 3131 // Clear out the active path from any previous work. 3132 PD.resetPath(); 3133 origReportConfigToken = R->getConfigurationChangeToken(); 3134 3135 // Generate the very last diagnostic piece - the piece is visible before 3136 // the trace is expanded. 3137 std::unique_ptr<PathDiagnosticPiece> LastPiece; 3138 for (BugReport::visitor_iterator I = visitors.begin(), E = visitors.end(); 3139 I != E; ++I) { 3140 if (std::unique_ptr<PathDiagnosticPiece> Piece = 3141 (*I)->getEndPath(PDB, N, *R)) { 3142 assert (!LastPiece && 3143 "There can only be one final piece in a diagnostic."); 3144 LastPiece = std::move(Piece); 3145 } 3146 } 3147 3148 if (ActiveScheme != PathDiagnosticConsumer::None) { 3149 if (!LastPiece) 3150 LastPiece = BugReporterVisitor::getDefaultEndPath(PDB, N, *R); 3151 assert(LastPiece); 3152 PD.setEndOfPath(std::move(LastPiece)); 3153 } 3154 3155 // Make sure we get a clean location context map so we don't 3156 // hold onto old mappings. 3157 LCM.clear(); 3158 3159 switch (ActiveScheme) { 3160 case PathDiagnosticConsumer::AlternateExtensive: 3161 GenerateAlternateExtensivePathDiagnostic(PD, PDB, N, LCM, visitors); 3162 break; 3163 case PathDiagnosticConsumer::Extensive: 3164 GenerateExtensivePathDiagnostic(PD, PDB, N, LCM, visitors); 3165 break; 3166 case PathDiagnosticConsumer::Minimal: 3167 GenerateMinimalPathDiagnostic(PD, PDB, N, LCM, visitors); 3168 break; 3169 case PathDiagnosticConsumer::None: 3170 GenerateVisitorsOnlyPathDiagnostic(PD, PDB, N, visitors); 3171 break; 3172 } 3173 3174 // Clean up the visitors we used. 3175 visitors.clear(); 3176 3177 // Did anything change while generating this path? 3178 finalReportConfigToken = R->getConfigurationChangeToken(); 3179 } while (finalReportConfigToken != origReportConfigToken); 3180 3181 if (!R->isValid()) 3182 continue; 3183 3184 // Finally, prune the diagnostic path of uninteresting stuff. 3185 if (!PD.path.empty()) { 3186 if (R->shouldPrunePath() && getAnalyzerOptions().shouldPrunePaths()) { 3187 bool stillHasNotes = removeUnneededCalls(PD.getMutablePieces(), R, LCM); 3188 assert(stillHasNotes); 3189 (void)stillHasNotes; 3190 } 3191 3192 // Redirect all call pieces to have valid locations. 3193 adjustCallLocations(PD.getMutablePieces()); 3194 removePiecesWithInvalidLocations(PD.getMutablePieces()); 3195 3196 if (ActiveScheme == PathDiagnosticConsumer::AlternateExtensive) { 3197 SourceManager &SM = getSourceManager(); 3198 3199 // Reduce the number of edges from a very conservative set 3200 // to an aesthetically pleasing subset that conveys the 3201 // necessary information. 3202 OptimizedCallsSet OCS; 3203 while (optimizeEdges(PD.getMutablePieces(), SM, OCS, LCM)) {} 3204 3205 // Drop the very first function-entry edge. It's not really necessary 3206 // for top-level functions. 3207 dropFunctionEntryEdge(PD.getMutablePieces(), LCM, SM); 3208 } 3209 3210 // Remove messages that are basically the same, and edges that may not 3211 // make sense. 3212 // We have to do this after edge optimization in the Extensive mode. 3213 removeRedundantMsgs(PD.getMutablePieces()); 3214 removeEdgesToDefaultInitializers(PD.getMutablePieces()); 3215 } 3216 3217 // We found a report and didn't suppress it. 3218 return true; 3219 } 3220 3221 // We suppressed all the reports in this equivalence class. 3222 assert(!HasInvalid && "Inconsistent suppression"); 3223 (void)HasInvalid; 3224 return false; 3225 } 3226 3227 void BugReporter::Register(BugType *BT) { 3228 BugTypes = F.add(BugTypes, BT); 3229 } 3230 3231 void BugReporter::emitReport(std::unique_ptr<BugReport> R) { 3232 if (const ExplodedNode *E = R->getErrorNode()) { 3233 // An error node must either be a sink or have a tag, otherwise 3234 // it could get reclaimed before the path diagnostic is created. 3235 assert((E->isSink() || E->getLocation().getTag()) && 3236 "Error node must either be a sink or have a tag"); 3237 3238 const AnalysisDeclContext *DeclCtx = 3239 E->getLocationContext()->getAnalysisDeclContext(); 3240 // The source of autosynthesized body can be handcrafted AST or a model 3241 // file. The locations from handcrafted ASTs have no valid source locations 3242 // and have to be discarded. Locations from model files should be preserved 3243 // for processing and reporting. 3244 if (DeclCtx->isBodyAutosynthesized() && 3245 !DeclCtx->isBodyAutosynthesizedFromModelFile()) 3246 return; 3247 } 3248 3249 bool ValidSourceLoc = R->getLocation(getSourceManager()).isValid(); 3250 assert(ValidSourceLoc); 3251 // If we mess up in a release build, we'd still prefer to just drop the bug 3252 // instead of trying to go on. 3253 if (!ValidSourceLoc) 3254 return; 3255 3256 // Compute the bug report's hash to determine its equivalence class. 3257 llvm::FoldingSetNodeID ID; 3258 R->Profile(ID); 3259 3260 // Lookup the equivance class. If there isn't one, create it. 3261 BugType& BT = R->getBugType(); 3262 Register(&BT); 3263 void *InsertPos; 3264 BugReportEquivClass* EQ = EQClasses.FindNodeOrInsertPos(ID, InsertPos); 3265 3266 if (!EQ) { 3267 EQ = new BugReportEquivClass(std::move(R)); 3268 EQClasses.InsertNode(EQ, InsertPos); 3269 EQClassesVector.push_back(EQ); 3270 } else 3271 EQ->AddReport(std::move(R)); 3272 } 3273 3274 3275 //===----------------------------------------------------------------------===// 3276 // Emitting reports in equivalence classes. 3277 //===----------------------------------------------------------------------===// 3278 3279 namespace { 3280 struct FRIEC_WLItem { 3281 const ExplodedNode *N; 3282 ExplodedNode::const_succ_iterator I, E; 3283 3284 FRIEC_WLItem(const ExplodedNode *n) 3285 : N(n), I(N->succ_begin()), E(N->succ_end()) {} 3286 }; 3287 } 3288 3289 static const CFGBlock *findBlockForNode(const ExplodedNode *N) { 3290 ProgramPoint P = N->getLocation(); 3291 if (auto BEP = P.getAs<BlockEntrance>()) 3292 return BEP->getBlock(); 3293 3294 // Find the node's current statement in the CFG. 3295 if (const Stmt *S = PathDiagnosticLocation::getStmt(N)) 3296 return N->getLocationContext()->getAnalysisDeclContext() 3297 ->getCFGStmtMap()->getBlock(S); 3298 3299 return nullptr; 3300 } 3301 3302 static BugReport * 3303 FindReportInEquivalenceClass(BugReportEquivClass& EQ, 3304 SmallVectorImpl<BugReport*> &bugReports) { 3305 3306 BugReportEquivClass::iterator I = EQ.begin(), E = EQ.end(); 3307 assert(I != E); 3308 BugType& BT = I->getBugType(); 3309 3310 // If we don't need to suppress any of the nodes because they are 3311 // post-dominated by a sink, simply add all the nodes in the equivalence class 3312 // to 'Nodes'. Any of the reports will serve as a "representative" report. 3313 if (!BT.isSuppressOnSink()) { 3314 BugReport *R = &*I; 3315 for (BugReportEquivClass::iterator I=EQ.begin(), E=EQ.end(); I!=E; ++I) { 3316 const ExplodedNode *N = I->getErrorNode(); 3317 if (N) { 3318 R = &*I; 3319 bugReports.push_back(R); 3320 } 3321 } 3322 return R; 3323 } 3324 3325 // For bug reports that should be suppressed when all paths are post-dominated 3326 // by a sink node, iterate through the reports in the equivalence class 3327 // until we find one that isn't post-dominated (if one exists). We use a 3328 // DFS traversal of the ExplodedGraph to find a non-sink node. We could write 3329 // this as a recursive function, but we don't want to risk blowing out the 3330 // stack for very long paths. 3331 BugReport *exampleReport = nullptr; 3332 3333 for (; I != E; ++I) { 3334 const ExplodedNode *errorNode = I->getErrorNode(); 3335 3336 if (!errorNode) 3337 continue; 3338 if (errorNode->isSink()) { 3339 llvm_unreachable( 3340 "BugType::isSuppressSink() should not be 'true' for sink end nodes"); 3341 } 3342 // No successors? By definition this nodes isn't post-dominated by a sink. 3343 if (errorNode->succ_empty()) { 3344 bugReports.push_back(&*I); 3345 if (!exampleReport) 3346 exampleReport = &*I; 3347 continue; 3348 } 3349 3350 // See if we are in a no-return CFG block. If so, treat this similarly 3351 // to being post-dominated by a sink. This works better when the analysis 3352 // is incomplete and we have never reached a no-return function 3353 // we're post-dominated by. 3354 // This is not quite enough to handle the incomplete analysis case. 3355 // We may be post-dominated in subsequent blocks, or even 3356 // inter-procedurally. However, it is not clear if more complicated 3357 // cases are generally worth suppressing. 3358 if (const CFGBlock *B = findBlockForNode(errorNode)) 3359 if (B->hasNoReturnElement()) 3360 continue; 3361 3362 // At this point we know that 'N' is not a sink and it has at least one 3363 // successor. Use a DFS worklist to find a non-sink end-of-path node. 3364 typedef FRIEC_WLItem WLItem; 3365 typedef SmallVector<WLItem, 10> DFSWorkList; 3366 llvm::DenseMap<const ExplodedNode *, unsigned> Visited; 3367 3368 DFSWorkList WL; 3369 WL.push_back(errorNode); 3370 Visited[errorNode] = 1; 3371 3372 while (!WL.empty()) { 3373 WLItem &WI = WL.back(); 3374 assert(!WI.N->succ_empty()); 3375 3376 for (; WI.I != WI.E; ++WI.I) { 3377 const ExplodedNode *Succ = *WI.I; 3378 // End-of-path node? 3379 if (Succ->succ_empty()) { 3380 // If we found an end-of-path node that is not a sink. 3381 if (!Succ->isSink()) { 3382 bugReports.push_back(&*I); 3383 if (!exampleReport) 3384 exampleReport = &*I; 3385 WL.clear(); 3386 break; 3387 } 3388 // Found a sink? Continue on to the next successor. 3389 continue; 3390 } 3391 // Mark the successor as visited. If it hasn't been explored, 3392 // enqueue it to the DFS worklist. 3393 unsigned &mark = Visited[Succ]; 3394 if (!mark) { 3395 mark = 1; 3396 WL.push_back(Succ); 3397 break; 3398 } 3399 } 3400 3401 // The worklist may have been cleared at this point. First 3402 // check if it is empty before checking the last item. 3403 if (!WL.empty() && &WL.back() == &WI) 3404 WL.pop_back(); 3405 } 3406 } 3407 3408 // ExampleReport will be NULL if all the nodes in the equivalence class 3409 // were post-dominated by sinks. 3410 return exampleReport; 3411 } 3412 3413 void BugReporter::FlushReport(BugReportEquivClass& EQ) { 3414 SmallVector<BugReport*, 10> bugReports; 3415 BugReport *exampleReport = FindReportInEquivalenceClass(EQ, bugReports); 3416 if (exampleReport) { 3417 for (PathDiagnosticConsumer *PDC : getPathDiagnosticConsumers()) { 3418 FlushReport(exampleReport, *PDC, bugReports); 3419 } 3420 } 3421 } 3422 3423 void BugReporter::FlushReport(BugReport *exampleReport, 3424 PathDiagnosticConsumer &PD, 3425 ArrayRef<BugReport*> bugReports) { 3426 3427 // FIXME: Make sure we use the 'R' for the path that was actually used. 3428 // Probably doesn't make a difference in practice. 3429 BugType& BT = exampleReport->getBugType(); 3430 3431 std::unique_ptr<PathDiagnostic> D(new PathDiagnostic( 3432 exampleReport->getBugType().getCheckName(), 3433 exampleReport->getDeclWithIssue(), exampleReport->getBugType().getName(), 3434 exampleReport->getDescription(), 3435 exampleReport->getShortDescription(/*Fallback=*/false), BT.getCategory(), 3436 exampleReport->getUniqueingLocation(), 3437 exampleReport->getUniqueingDecl())); 3438 3439 if (exampleReport->isPathSensitive()) { 3440 // Generate the full path diagnostic, using the generation scheme 3441 // specified by the PathDiagnosticConsumer. Note that we have to generate 3442 // path diagnostics even for consumers which do not support paths, because 3443 // the BugReporterVisitors may mark this bug as a false positive. 3444 assert(!bugReports.empty()); 3445 3446 MaxBugClassSize = 3447 std::max(bugReports.size(), static_cast<size_t>(MaxBugClassSize)); 3448 3449 if (!generatePathDiagnostic(*D.get(), PD, bugReports)) 3450 return; 3451 3452 MaxValidBugClassSize = 3453 std::max(bugReports.size(), static_cast<size_t>(MaxValidBugClassSize)); 3454 3455 // Examine the report and see if the last piece is in a header. Reset the 3456 // report location to the last piece in the main source file. 3457 AnalyzerOptions &Opts = getAnalyzerOptions(); 3458 if (Opts.shouldReportIssuesInMainSourceFile() && !Opts.AnalyzeAll) 3459 D->resetDiagnosticLocationToMainFile(); 3460 } 3461 3462 // If the path is empty, generate a single step path with the location 3463 // of the issue. 3464 if (D->path.empty()) { 3465 PathDiagnosticLocation L = exampleReport->getLocation(getSourceManager()); 3466 auto piece = llvm::make_unique<PathDiagnosticEventPiece>( 3467 L, exampleReport->getDescription()); 3468 for (SourceRange Range : exampleReport->getRanges()) 3469 piece->addRange(Range); 3470 D->setEndOfPath(std::move(piece)); 3471 } 3472 3473 PathPieces &Pieces = D->getMutablePieces(); 3474 if (getAnalyzerOptions().shouldDisplayNotesAsEvents()) { 3475 // For path diagnostic consumers that don't support extra notes, 3476 // we may optionally convert those to path notes. 3477 for (auto I = exampleReport->getNotes().rbegin(), 3478 E = exampleReport->getNotes().rend(); I != E; ++I) { 3479 PathDiagnosticNotePiece *Piece = I->get(); 3480 PathDiagnosticEventPiece *ConvertedPiece = 3481 new PathDiagnosticEventPiece(Piece->getLocation(), 3482 Piece->getString()); 3483 for (const auto &R: Piece->getRanges()) 3484 ConvertedPiece->addRange(R); 3485 3486 Pieces.push_front(ConvertedPiece); 3487 } 3488 } else { 3489 for (auto I = exampleReport->getNotes().rbegin(), 3490 E = exampleReport->getNotes().rend(); I != E; ++I) 3491 Pieces.push_front(*I); 3492 } 3493 3494 // Get the meta data. 3495 const BugReport::ExtraTextList &Meta = exampleReport->getExtraText(); 3496 for (BugReport::ExtraTextList::const_iterator i = Meta.begin(), 3497 e = Meta.end(); i != e; ++i) { 3498 D->addMeta(*i); 3499 } 3500 3501 PD.HandlePathDiagnostic(std::move(D)); 3502 } 3503 3504 void BugReporter::EmitBasicReport(const Decl *DeclWithIssue, 3505 const CheckerBase *Checker, 3506 StringRef Name, StringRef Category, 3507 StringRef Str, PathDiagnosticLocation Loc, 3508 ArrayRef<SourceRange> Ranges) { 3509 EmitBasicReport(DeclWithIssue, Checker->getCheckName(), Name, Category, Str, 3510 Loc, Ranges); 3511 } 3512 void BugReporter::EmitBasicReport(const Decl *DeclWithIssue, 3513 CheckName CheckName, 3514 StringRef name, StringRef category, 3515 StringRef str, PathDiagnosticLocation Loc, 3516 ArrayRef<SourceRange> Ranges) { 3517 3518 // 'BT' is owned by BugReporter. 3519 BugType *BT = getBugTypeForName(CheckName, name, category); 3520 auto R = llvm::make_unique<BugReport>(*BT, str, Loc); 3521 R->setDeclWithIssue(DeclWithIssue); 3522 for (ArrayRef<SourceRange>::iterator I = Ranges.begin(), E = Ranges.end(); 3523 I != E; ++I) 3524 R->addRange(*I); 3525 emitReport(std::move(R)); 3526 } 3527 3528 BugType *BugReporter::getBugTypeForName(CheckName CheckName, StringRef name, 3529 StringRef category) { 3530 SmallString<136> fullDesc; 3531 llvm::raw_svector_ostream(fullDesc) << CheckName.getName() << ":" << name 3532 << ":" << category; 3533 BugType *&BT = StrBugTypes[fullDesc]; 3534 if (!BT) 3535 BT = new BugType(CheckName, name, category); 3536 return BT; 3537 } 3538 3539 LLVM_DUMP_METHOD void PathPieces::dump() const { 3540 unsigned index = 0; 3541 for (PathPieces::const_iterator I = begin(), E = end(); I != E; ++I) { 3542 llvm::errs() << "[" << index++ << "] "; 3543 (*I)->dump(); 3544 llvm::errs() << "\n"; 3545 } 3546 } 3547 3548 LLVM_DUMP_METHOD void PathDiagnosticCallPiece::dump() const { 3549 llvm::errs() << "CALL\n--------------\n"; 3550 3551 if (const Stmt *SLoc = getLocStmt(getLocation())) 3552 SLoc->dump(); 3553 else if (const NamedDecl *ND = dyn_cast<NamedDecl>(getCallee())) 3554 llvm::errs() << *ND << "\n"; 3555 else 3556 getLocation().dump(); 3557 } 3558 3559 LLVM_DUMP_METHOD void PathDiagnosticEventPiece::dump() const { 3560 llvm::errs() << "EVENT\n--------------\n"; 3561 llvm::errs() << getString() << "\n"; 3562 llvm::errs() << " ---- at ----\n"; 3563 getLocation().dump(); 3564 } 3565 3566 LLVM_DUMP_METHOD void PathDiagnosticControlFlowPiece::dump() const { 3567 llvm::errs() << "CONTROL\n--------------\n"; 3568 getStartLocation().dump(); 3569 llvm::errs() << " ---- to ----\n"; 3570 getEndLocation().dump(); 3571 } 3572 3573 LLVM_DUMP_METHOD void PathDiagnosticMacroPiece::dump() const { 3574 llvm::errs() << "MACRO\n--------------\n"; 3575 // FIXME: Print which macro is being invoked. 3576 } 3577 3578 LLVM_DUMP_METHOD void PathDiagnosticNotePiece::dump() const { 3579 llvm::errs() << "NOTE\n--------------\n"; 3580 llvm::errs() << getString() << "\n"; 3581 llvm::errs() << " ---- at ----\n"; 3582 getLocation().dump(); 3583 } 3584 3585 LLVM_DUMP_METHOD void PathDiagnosticLocation::dump() const { 3586 if (!isValid()) { 3587 llvm::errs() << "<INVALID>\n"; 3588 return; 3589 } 3590 3591 switch (K) { 3592 case RangeK: 3593 // FIXME: actually print the range. 3594 llvm::errs() << "<range>\n"; 3595 break; 3596 case SingleLocK: 3597 asLocation().dump(); 3598 llvm::errs() << "\n"; 3599 break; 3600 case StmtK: 3601 if (S) 3602 S->dump(); 3603 else 3604 llvm::errs() << "<NULL STMT>\n"; 3605 break; 3606 case DeclK: 3607 if (const NamedDecl *ND = dyn_cast_or_null<NamedDecl>(D)) 3608 llvm::errs() << *ND << "\n"; 3609 else if (isa<BlockDecl>(D)) 3610 // FIXME: Make this nicer. 3611 llvm::errs() << "<block>\n"; 3612 else if (D) 3613 llvm::errs() << "<unknown decl>\n"; 3614 else 3615 llvm::errs() << "<NULL DECL>\n"; 3616 break; 3617 } 3618 } 3619