1//===- DynamicTypePropagation.cpp ------------------------------*- C++ -*--===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file contains two checkers. One helps the static analyzer core to track
10// types, the other does type inference on Obj-C generics and report type
11// errors.
12//
13// Dynamic Type Propagation:
14// This checker defines the rules for dynamic type gathering and propagation.
15//
16// Generics Checker for Objective-C:
17// This checker tries to find type errors that the compiler is not able to catch
18// due to the implicit conversions that were introduced for backward
19// compatibility.
20//
21//===----------------------------------------------------------------------===//
22
23#include "clang/AST/ParentMap.h"
24#include "clang/AST/RecursiveASTVisitor.h"
25#include "clang/Basic/Builtins.h"
26#include "clang/StaticAnalyzer/Checkers/BuiltinCheckerRegistration.h"
27#include "clang/StaticAnalyzer/Core/BugReporter/BugType.h"
28#include "clang/StaticAnalyzer/Core/Checker.h"
29#include "clang/StaticAnalyzer/Core/CheckerManager.h"
30#include "clang/StaticAnalyzer/Core/PathSensitive/CallEvent.h"
31#include "clang/StaticAnalyzer/Core/PathSensitive/CheckerContext.h"
32#include "clang/StaticAnalyzer/Core/PathSensitive/DynamicType.h"
33#include "clang/StaticAnalyzer/Core/PathSensitive/ProgramStateTrait.h"
34
35using namespace clang;
36using namespace ento;
37
38// ProgramState trait - The type inflation is tracked by DynamicTypeMap. This is
39// an auxiliary map that tracks more information about generic types, because in
40// some cases the most derived type is not the most informative one about the
41// type parameters. This types that are stored for each symbol in this map must
42// be specialized.
43// TODO: In some case the type stored in this map is exactly the same that is
44// stored in DynamicTypeMap. We should no store duplicated information in those
45// cases.
46REGISTER_MAP_WITH_PROGRAMSTATE(MostSpecializedTypeArgsMap, SymbolRef,
47                               const ObjCObjectPointerType *)
48
49namespace {
50class DynamicTypePropagation:
51    public Checker< check::PreCall,
52                    check::PostCall,
53                    check::DeadSymbols,
54                    check::PostStmt<CastExpr>,
55                    check::PostStmt<CXXNewExpr>,
56                    check::PreObjCMessage,
57                    check::PostObjCMessage > {
58
59  const ObjCObjectType *getObjectTypeForAllocAndNew(const ObjCMessageExpr *MsgE,
60                                                    CheckerContext &C) const;
61
62  /// Return a better dynamic type if one can be derived from the cast.
63  const ObjCObjectPointerType *getBetterObjCType(const Expr *CastE,
64                                                 CheckerContext &C) const;
65
66  ExplodedNode *dynamicTypePropagationOnCasts(const CastExpr *CE,
67                                              ProgramStateRef &State,
68                                              CheckerContext &C) const;
69
70  mutable std::unique_ptr<BugType> ObjCGenericsBugType;
71  void initBugType() const {
72    if (!ObjCGenericsBugType)
73      ObjCGenericsBugType.reset(new BugType(
74          GenericCheckName, "Generics", categories::CoreFoundationObjectiveC));
75  }
76
77  class GenericsBugVisitor : public BugReporterVisitor {
78  public:
79    GenericsBugVisitor(SymbolRef S) : Sym(S) {}
80
81    void Profile(llvm::FoldingSetNodeID &ID) const override {
82      static int X = 0;
83      ID.AddPointer(&X);
84      ID.AddPointer(Sym);
85    }
86
87    PathDiagnosticPieceRef VisitNode(const ExplodedNode *N,
88                                     BugReporterContext &BRC,
89                                     PathSensitiveBugReport &BR) override;
90
91  private:
92    // The tracked symbol.
93    SymbolRef Sym;
94  };
95
96  void reportGenericsBug(const ObjCObjectPointerType *From,
97                         const ObjCObjectPointerType *To, ExplodedNode *N,
98                         SymbolRef Sym, CheckerContext &C,
99                         const Stmt *ReportedNode = nullptr) const;
100
101public:
102  void checkPreCall(const CallEvent &Call, CheckerContext &C) const;
103  void checkPostCall(const CallEvent &Call, CheckerContext &C) const;
104  void checkPostStmt(const CastExpr *CastE, CheckerContext &C) const;
105  void checkPostStmt(const CXXNewExpr *NewE, CheckerContext &C) const;
106  void checkDeadSymbols(SymbolReaper &SR, CheckerContext &C) const;
107  void checkPreObjCMessage(const ObjCMethodCall &M, CheckerContext &C) const;
108  void checkPostObjCMessage(const ObjCMethodCall &M, CheckerContext &C) const;
109
110  /// This value is set to true, when the Generics checker is turned on.
111  DefaultBool CheckGenerics;
112  CheckerNameRef GenericCheckName;
113};
114
115bool isObjCClassType(QualType Type) {
116  if (const auto *PointerType = dyn_cast<ObjCObjectPointerType>(Type)) {
117    return PointerType->getObjectType()->isObjCClass();
118  }
119  return false;
120}
121
122struct RuntimeType {
123  const ObjCObjectType *Type = nullptr;
124  bool Precise = false;
125
126  operator bool() const { return Type != nullptr; }
127};
128
129RuntimeType inferReceiverType(const ObjCMethodCall &Message,
130                              CheckerContext &C) {
131  const ObjCMessageExpr *MessageExpr = Message.getOriginExpr();
132
133  // Check if we can statically infer the actual type precisely.
134  //
135  // 1. Class is written directly in the message:
136  // \code
137  //   [ActualClass classMethod];
138  // \endcode
139  if (MessageExpr->getReceiverKind() == ObjCMessageExpr::Class) {
140    return {MessageExpr->getClassReceiver()->getAs<ObjCObjectType>(),
141            /*Precise=*/true};
142  }
143
144  // 2. Receiver is 'super' from a class method (a.k.a 'super' is a
145  //    class object).
146  // \code
147  //   [super classMethod];
148  // \endcode
149  if (MessageExpr->getReceiverKind() == ObjCMessageExpr::SuperClass) {
150    return {MessageExpr->getSuperType()->getAs<ObjCObjectType>(),
151            /*Precise=*/true};
152  }
153
154  // 3. Receiver is 'super' from an instance method (a.k.a 'super' is an
155  //    instance of a super class).
156  // \code
157  //   [super instanceMethod];
158  // \encode
159  if (MessageExpr->getReceiverKind() == ObjCMessageExpr::SuperInstance) {
160    if (const auto *ObjTy =
161            MessageExpr->getSuperType()->getAs<ObjCObjectPointerType>())
162      return {ObjTy->getObjectType(), /*Precise=*/true};
163  }
164
165  const Expr *RecE = MessageExpr->getInstanceReceiver();
166
167  if (!RecE)
168    return {};
169
170  // Otherwise, let's try to get type information from our estimations of
171  // runtime types.
172  QualType InferredType;
173  SVal ReceiverSVal = C.getSVal(RecE);
174  ProgramStateRef State = C.getState();
175
176  if (const MemRegion *ReceiverRegion = ReceiverSVal.getAsRegion()) {
177    if (DynamicTypeInfo DTI = getDynamicTypeInfo(State, ReceiverRegion)) {
178      InferredType = DTI.getType().getCanonicalType();
179    }
180  }
181
182  if (SymbolRef ReceiverSymbol = ReceiverSVal.getAsSymbol()) {
183    if (InferredType.isNull()) {
184      InferredType = ReceiverSymbol->getType();
185    }
186
187    // If receiver is a Class object, we want to figure out the type it
188    // represents.
189    if (isObjCClassType(InferredType)) {
190      // We actually might have some info on what type is contained in there.
191      if (DynamicTypeInfo DTI =
192              getClassObjectDynamicTypeInfo(State, ReceiverSymbol)) {
193
194        // Types in Class objects can be ONLY Objective-C types
195        return {cast<ObjCObjectType>(DTI.getType()), !DTI.canBeASubClass()};
196      }
197
198      SVal SelfSVal = State->getSelfSVal(C.getLocationContext());
199
200      // Another way we can guess what is in Class object, is when it is a
201      // 'self' variable of the current class method.
202      if (ReceiverSVal == SelfSVal) {
203        // In this case, we should return the type of the enclosing class
204        // declaration.
205        if (const ObjCMethodDecl *MD =
206                dyn_cast<ObjCMethodDecl>(C.getStackFrame()->getDecl()))
207          if (const ObjCObjectType *ObjTy = dyn_cast<ObjCObjectType>(
208                  MD->getClassInterface()->getTypeForDecl()))
209            return {ObjTy};
210      }
211    }
212  }
213
214  // Unfortunately, it seems like we have no idea what that type is.
215  if (InferredType.isNull()) {
216    return {};
217  }
218
219  // We can end up here if we got some dynamic type info and the
220  // receiver is not one of the known Class objects.
221  if (const auto *ReceiverInferredType =
222          dyn_cast<ObjCObjectPointerType>(InferredType)) {
223    return {ReceiverInferredType->getObjectType()};
224  }
225
226  // Any other type (like 'Class') is not really useful at this point.
227  return {};
228}
229} // end anonymous namespace
230
231void DynamicTypePropagation::checkDeadSymbols(SymbolReaper &SR,
232                                              CheckerContext &C) const {
233  ProgramStateRef State = removeDeadTypes(C.getState(), SR);
234  State = removeDeadClassObjectTypes(State, SR);
235
236  MostSpecializedTypeArgsMapTy TyArgMap =
237      State->get<MostSpecializedTypeArgsMap>();
238  for (MostSpecializedTypeArgsMapTy::iterator I = TyArgMap.begin(),
239                                              E = TyArgMap.end();
240       I != E; ++I) {
241    if (SR.isDead(I->first)) {
242      State = State->remove<MostSpecializedTypeArgsMap>(I->first);
243    }
244  }
245
246  C.addTransition(State);
247}
248
249static void recordFixedType(const MemRegion *Region, const CXXMethodDecl *MD,
250                            CheckerContext &C) {
251  assert(Region);
252  assert(MD);
253
254  ASTContext &Ctx = C.getASTContext();
255  QualType Ty = Ctx.getPointerType(Ctx.getRecordType(MD->getParent()));
256
257  ProgramStateRef State = C.getState();
258  State = setDynamicTypeInfo(State, Region, Ty, /*CanBeSubClassed=*/false);
259  C.addTransition(State);
260}
261
262void DynamicTypePropagation::checkPreCall(const CallEvent &Call,
263                                          CheckerContext &C) const {
264  if (const CXXConstructorCall *Ctor = dyn_cast<CXXConstructorCall>(&Call)) {
265    // C++11 [class.cdtor]p4: When a virtual function is called directly or
266    //   indirectly from a constructor or from a destructor, including during
267    //   the construction or destruction of the class's non-static data members,
268    //   and the object to which the call applies is the object under
269    //   construction or destruction, the function called is the final overrider
270    //   in the constructor's or destructor's class and not one overriding it in
271    //   a more-derived class.
272
273    switch (Ctor->getOriginExpr()->getConstructionKind()) {
274    case CXXConstructExpr::CK_Complete:
275    case CXXConstructExpr::CK_Delegating:
276      // No additional type info necessary.
277      return;
278    case CXXConstructExpr::CK_NonVirtualBase:
279    case CXXConstructExpr::CK_VirtualBase:
280      if (const MemRegion *Target = Ctor->getCXXThisVal().getAsRegion())
281        recordFixedType(Target, Ctor->getDecl(), C);
282      return;
283    }
284
285    return;
286  }
287
288  if (const CXXDestructorCall *Dtor = dyn_cast<CXXDestructorCall>(&Call)) {
289    // C++11 [class.cdtor]p4 (see above)
290    if (!Dtor->isBaseDestructor())
291      return;
292
293    const MemRegion *Target = Dtor->getCXXThisVal().getAsRegion();
294    if (!Target)
295      return;
296
297    const Decl *D = Dtor->getDecl();
298    if (!D)
299      return;
300
301    recordFixedType(Target, cast<CXXDestructorDecl>(D), C);
302    return;
303  }
304}
305
306void DynamicTypePropagation::checkPostCall(const CallEvent &Call,
307                                           CheckerContext &C) const {
308  // We can obtain perfect type info for return values from some calls.
309  if (const ObjCMethodCall *Msg = dyn_cast<ObjCMethodCall>(&Call)) {
310
311    // Get the returned value if it's a region.
312    const MemRegion *RetReg = Call.getReturnValue().getAsRegion();
313    if (!RetReg)
314      return;
315
316    ProgramStateRef State = C.getState();
317    const ObjCMethodDecl *D = Msg->getDecl();
318
319    if (D && D->hasRelatedResultType()) {
320      switch (Msg->getMethodFamily()) {
321      default:
322        break;
323
324      // We assume that the type of the object returned by alloc and new are the
325      // pointer to the object of the class specified in the receiver of the
326      // message.
327      case OMF_alloc:
328      case OMF_new: {
329        // Get the type of object that will get created.
330        RuntimeType ObjTy = inferReceiverType(*Msg, C);
331
332        if (!ObjTy)
333          return;
334
335        QualType DynResTy =
336            C.getASTContext().getObjCObjectPointerType(QualType(ObjTy.Type, 0));
337        // We used to assume that whatever type we got from inferring the
338        // type is actually precise (and it is not exactly correct).
339        // A big portion of the existing behavior depends on that assumption
340        // (e.g. certain inlining won't take place). For this reason, we don't
341        // use ObjTy.Precise flag here.
342        //
343        // TODO: We should mitigate this problem some time in the future
344        // and replace hardcoded 'false' with '!ObjTy.Precise'.
345        C.addTransition(setDynamicTypeInfo(State, RetReg, DynResTy, false));
346        break;
347      }
348      case OMF_init: {
349        // Assume, the result of the init method has the same dynamic type as
350        // the receiver and propagate the dynamic type info.
351        const MemRegion *RecReg = Msg->getReceiverSVal().getAsRegion();
352        if (!RecReg)
353          return;
354        DynamicTypeInfo RecDynType = getDynamicTypeInfo(State, RecReg);
355        C.addTransition(setDynamicTypeInfo(State, RetReg, RecDynType));
356        break;
357      }
358      }
359    }
360    return;
361  }
362
363  if (const CXXConstructorCall *Ctor = dyn_cast<CXXConstructorCall>(&Call)) {
364    // We may need to undo the effects of our pre-call check.
365    switch (Ctor->getOriginExpr()->getConstructionKind()) {
366    case CXXConstructExpr::CK_Complete:
367    case CXXConstructExpr::CK_Delegating:
368      // No additional work necessary.
369      // Note: This will leave behind the actual type of the object for
370      // complete constructors, but arguably that's a good thing, since it
371      // means the dynamic type info will be correct even for objects
372      // constructed with operator new.
373      return;
374    case CXXConstructExpr::CK_NonVirtualBase:
375    case CXXConstructExpr::CK_VirtualBase:
376      if (const MemRegion *Target = Ctor->getCXXThisVal().getAsRegion()) {
377        // We just finished a base constructor. Now we can use the subclass's
378        // type when resolving virtual calls.
379        const LocationContext *LCtx = C.getLocationContext();
380
381        // FIXME: In C++17 classes with non-virtual bases may be treated as
382        // aggregates, and in such case no top-frame constructor will be called.
383        // Figure out if we need to do anything in this case.
384        // FIXME: Instead of relying on the ParentMap, we should have the
385        // trigger-statement (InitListExpr in this case) available in this
386        // callback, ideally as part of CallEvent.
387        if (dyn_cast_or_null<InitListExpr>(
388                LCtx->getParentMap().getParent(Ctor->getOriginExpr())))
389          return;
390
391        recordFixedType(Target, cast<CXXConstructorDecl>(LCtx->getDecl()), C);
392      }
393      return;
394    }
395  }
396}
397
398/// TODO: Handle explicit casts.
399///       Handle C++ casts.
400///
401/// Precondition: the cast is between ObjCObjectPointers.
402ExplodedNode *DynamicTypePropagation::dynamicTypePropagationOnCasts(
403    const CastExpr *CE, ProgramStateRef &State, CheckerContext &C) const {
404  // We only track type info for regions.
405  const MemRegion *ToR = C.getSVal(CE).getAsRegion();
406  if (!ToR)
407    return C.getPredecessor();
408
409  if (isa<ExplicitCastExpr>(CE))
410    return C.getPredecessor();
411
412  if (const Type *NewTy = getBetterObjCType(CE, C)) {
413    State = setDynamicTypeInfo(State, ToR, QualType(NewTy, 0));
414    return C.addTransition(State);
415  }
416  return C.getPredecessor();
417}
418
419void DynamicTypePropagation::checkPostStmt(const CXXNewExpr *NewE,
420                                           CheckerContext &C) const {
421  if (NewE->isArray())
422    return;
423
424  // We only track dynamic type info for regions.
425  const MemRegion *MR = C.getSVal(NewE).getAsRegion();
426  if (!MR)
427    return;
428
429  C.addTransition(setDynamicTypeInfo(C.getState(), MR, NewE->getType(),
430                                     /*CanBeSubClassed=*/false));
431}
432
433// Return a better dynamic type if one can be derived from the cast.
434// Compare the current dynamic type of the region and the new type to which we
435// are casting. If the new type is lower in the inheritance hierarchy, pick it.
436const ObjCObjectPointerType *
437DynamicTypePropagation::getBetterObjCType(const Expr *CastE,
438                                          CheckerContext &C) const {
439  const MemRegion *ToR = C.getSVal(CastE).getAsRegion();
440  assert(ToR);
441
442  // Get the old and new types.
443  const ObjCObjectPointerType *NewTy =
444      CastE->getType()->getAs<ObjCObjectPointerType>();
445  if (!NewTy)
446    return nullptr;
447  QualType OldDTy = getDynamicTypeInfo(C.getState(), ToR).getType();
448  if (OldDTy.isNull()) {
449    return NewTy;
450  }
451  const ObjCObjectPointerType *OldTy =
452    OldDTy->getAs<ObjCObjectPointerType>();
453  if (!OldTy)
454    return nullptr;
455
456  // Id the old type is 'id', the new one is more precise.
457  if (OldTy->isObjCIdType() && !NewTy->isObjCIdType())
458    return NewTy;
459
460  // Return new if it's a subclass of old.
461  const ObjCInterfaceDecl *ToI = NewTy->getInterfaceDecl();
462  const ObjCInterfaceDecl *FromI = OldTy->getInterfaceDecl();
463  if (ToI && FromI && FromI->isSuperClassOf(ToI))
464    return NewTy;
465
466  return nullptr;
467}
468
469static const ObjCObjectPointerType *getMostInformativeDerivedClassImpl(
470    const ObjCObjectPointerType *From, const ObjCObjectPointerType *To,
471    const ObjCObjectPointerType *MostInformativeCandidate, ASTContext &C) {
472  // Checking if from and to are the same classes modulo specialization.
473  if (From->getInterfaceDecl()->getCanonicalDecl() ==
474      To->getInterfaceDecl()->getCanonicalDecl()) {
475    if (To->isSpecialized()) {
476      assert(MostInformativeCandidate->isSpecialized());
477      return MostInformativeCandidate;
478    }
479    return From;
480  }
481
482  if (To->getObjectType()->getSuperClassType().isNull()) {
483    // If To has no super class and From and To aren't the same then
484    // To was not actually a descendent of From. In this case the best we can
485    // do is 'From'.
486    return From;
487  }
488
489  const auto *SuperOfTo =
490      To->getObjectType()->getSuperClassType()->castAs<ObjCObjectType>();
491  assert(SuperOfTo);
492  QualType SuperPtrOfToQual =
493      C.getObjCObjectPointerType(QualType(SuperOfTo, 0));
494  const auto *SuperPtrOfTo = SuperPtrOfToQual->castAs<ObjCObjectPointerType>();
495  if (To->isUnspecialized())
496    return getMostInformativeDerivedClassImpl(From, SuperPtrOfTo, SuperPtrOfTo,
497                                              C);
498  else
499    return getMostInformativeDerivedClassImpl(From, SuperPtrOfTo,
500                                              MostInformativeCandidate, C);
501}
502
503/// A downcast may loose specialization information. E. g.:
504///   MutableMap<T, U> : Map
505/// The downcast to MutableMap looses the information about the types of the
506/// Map (due to the type parameters are not being forwarded to Map), and in
507/// general there is no way to recover that information from the
508/// declaration. In order to have to most information, lets find the most
509/// derived type that has all the type parameters forwarded.
510///
511/// Get the a subclass of \p From (which has a lower bound \p To) that do not
512/// loose information about type parameters. \p To has to be a subclass of
513/// \p From. From has to be specialized.
514static const ObjCObjectPointerType *
515getMostInformativeDerivedClass(const ObjCObjectPointerType *From,
516                               const ObjCObjectPointerType *To, ASTContext &C) {
517  return getMostInformativeDerivedClassImpl(From, To, To, C);
518}
519
520/// Inputs:
521///   \param StaticLowerBound Static lower bound for a symbol. The dynamic lower
522///   bound might be the subclass of this type.
523///   \param StaticUpperBound A static upper bound for a symbol.
524///   \p StaticLowerBound expected to be the subclass of \p StaticUpperBound.
525///   \param Current The type that was inferred for a symbol in a previous
526///   context. Might be null when this is the first time that inference happens.
527/// Precondition:
528///   \p StaticLowerBound or \p StaticUpperBound is specialized. If \p Current
529///   is not null, it is specialized.
530/// Possible cases:
531///   (1) The \p Current is null and \p StaticLowerBound <: \p StaticUpperBound
532///   (2) \p StaticLowerBound <: \p Current <: \p StaticUpperBound
533///   (3) \p Current <: \p StaticLowerBound <: \p StaticUpperBound
534///   (4) \p StaticLowerBound <: \p StaticUpperBound <: \p Current
535/// Effect:
536///   Use getMostInformativeDerivedClass with the upper and lower bound of the
537///   set {\p StaticLowerBound, \p Current, \p StaticUpperBound}. The computed
538///   lower bound must be specialized. If the result differs from \p Current or
539///   \p Current is null, store the result.
540static bool
541storeWhenMoreInformative(ProgramStateRef &State, SymbolRef Sym,
542                         const ObjCObjectPointerType *const *Current,
543                         const ObjCObjectPointerType *StaticLowerBound,
544                         const ObjCObjectPointerType *StaticUpperBound,
545                         ASTContext &C) {
546  // TODO: The above 4 cases are not exhaustive. In particular, it is possible
547  // for Current to be incomparable with StaticLowerBound, StaticUpperBound,
548  // or both.
549  //
550  // For example, suppose Foo<T> and Bar<T> are unrelated types.
551  //
552  //  Foo<T> *f = ...
553  //  Bar<T> *b = ...
554  //
555  //  id t1 = b;
556  //  f = t1;
557  //  id t2 = f; // StaticLowerBound is Foo<T>, Current is Bar<T>
558  //
559  // We should either constrain the callers of this function so that the stated
560  // preconditions hold (and assert it) or rewrite the function to expicitly
561  // handle the additional cases.
562
563  // Precondition
564  assert(StaticUpperBound->isSpecialized() ||
565         StaticLowerBound->isSpecialized());
566  assert(!Current || (*Current)->isSpecialized());
567
568  // Case (1)
569  if (!Current) {
570    if (StaticUpperBound->isUnspecialized()) {
571      State = State->set<MostSpecializedTypeArgsMap>(Sym, StaticLowerBound);
572      return true;
573    }
574    // Upper bound is specialized.
575    const ObjCObjectPointerType *WithMostInfo =
576        getMostInformativeDerivedClass(StaticUpperBound, StaticLowerBound, C);
577    State = State->set<MostSpecializedTypeArgsMap>(Sym, WithMostInfo);
578    return true;
579  }
580
581  // Case (3)
582  if (C.canAssignObjCInterfaces(StaticLowerBound, *Current)) {
583    return false;
584  }
585
586  // Case (4)
587  if (C.canAssignObjCInterfaces(*Current, StaticUpperBound)) {
588    // The type arguments might not be forwarded at any point of inheritance.
589    const ObjCObjectPointerType *WithMostInfo =
590        getMostInformativeDerivedClass(*Current, StaticUpperBound, C);
591    WithMostInfo =
592        getMostInformativeDerivedClass(WithMostInfo, StaticLowerBound, C);
593    if (WithMostInfo == *Current)
594      return false;
595    State = State->set<MostSpecializedTypeArgsMap>(Sym, WithMostInfo);
596    return true;
597  }
598
599  // Case (2)
600  const ObjCObjectPointerType *WithMostInfo =
601      getMostInformativeDerivedClass(*Current, StaticLowerBound, C);
602  if (WithMostInfo != *Current) {
603    State = State->set<MostSpecializedTypeArgsMap>(Sym, WithMostInfo);
604    return true;
605  }
606
607  return false;
608}
609
610/// Type inference based on static type information that is available for the
611/// cast and the tracked type information for the given symbol. When the tracked
612/// symbol and the destination type of the cast are unrelated, report an error.
613void DynamicTypePropagation::checkPostStmt(const CastExpr *CE,
614                                           CheckerContext &C) const {
615  if (CE->getCastKind() != CK_BitCast)
616    return;
617
618  QualType OriginType = CE->getSubExpr()->getType();
619  QualType DestType = CE->getType();
620
621  const auto *OrigObjectPtrType = OriginType->getAs<ObjCObjectPointerType>();
622  const auto *DestObjectPtrType = DestType->getAs<ObjCObjectPointerType>();
623
624  if (!OrigObjectPtrType || !DestObjectPtrType)
625    return;
626
627  ProgramStateRef State = C.getState();
628  ExplodedNode *AfterTypeProp = dynamicTypePropagationOnCasts(CE, State, C);
629
630  ASTContext &ASTCtxt = C.getASTContext();
631
632  // This checker detects the subtyping relationships using the assignment
633  // rules. In order to be able to do this the kindofness must be stripped
634  // first. The checker treats every type as kindof type anyways: when the
635  // tracked type is the subtype of the static type it tries to look up the
636  // methods in the tracked type first.
637  OrigObjectPtrType = OrigObjectPtrType->stripObjCKindOfTypeAndQuals(ASTCtxt);
638  DestObjectPtrType = DestObjectPtrType->stripObjCKindOfTypeAndQuals(ASTCtxt);
639
640  if (OrigObjectPtrType->isUnspecialized() &&
641      DestObjectPtrType->isUnspecialized())
642    return;
643
644  SymbolRef Sym = C.getSVal(CE).getAsSymbol();
645  if (!Sym)
646    return;
647
648  const ObjCObjectPointerType *const *TrackedType =
649      State->get<MostSpecializedTypeArgsMap>(Sym);
650
651  if (isa<ExplicitCastExpr>(CE)) {
652    // Treat explicit casts as an indication from the programmer that the
653    // Objective-C type system is not rich enough to express the needed
654    // invariant. In such cases, forget any existing information inferred
655    // about the type arguments. We don't assume the casted-to specialized
656    // type here because the invariant the programmer specifies in the cast
657    // may only hold at this particular program point and not later ones.
658    // We don't want a suppressing cast to require a cascade of casts down the
659    // line.
660    if (TrackedType) {
661      State = State->remove<MostSpecializedTypeArgsMap>(Sym);
662      C.addTransition(State, AfterTypeProp);
663    }
664    return;
665  }
666
667  // Check which assignments are legal.
668  bool OrigToDest =
669      ASTCtxt.canAssignObjCInterfaces(DestObjectPtrType, OrigObjectPtrType);
670  bool DestToOrig =
671      ASTCtxt.canAssignObjCInterfaces(OrigObjectPtrType, DestObjectPtrType);
672
673  // The tracked type should be the sub or super class of the static destination
674  // type. When an (implicit) upcast or a downcast happens according to static
675  // types, and there is no subtyping relationship between the tracked and the
676  // static destination types, it indicates an error.
677  if (TrackedType &&
678      !ASTCtxt.canAssignObjCInterfaces(DestObjectPtrType, *TrackedType) &&
679      !ASTCtxt.canAssignObjCInterfaces(*TrackedType, DestObjectPtrType)) {
680    static CheckerProgramPointTag IllegalConv(this, "IllegalConversion");
681    ExplodedNode *N = C.addTransition(State, AfterTypeProp, &IllegalConv);
682    reportGenericsBug(*TrackedType, DestObjectPtrType, N, Sym, C);
683    return;
684  }
685
686  // Handle downcasts and upcasts.
687
688  const ObjCObjectPointerType *LowerBound = DestObjectPtrType;
689  const ObjCObjectPointerType *UpperBound = OrigObjectPtrType;
690  if (OrigToDest && !DestToOrig)
691    std::swap(LowerBound, UpperBound);
692
693  // The id type is not a real bound. Eliminate it.
694  LowerBound = LowerBound->isObjCIdType() ? UpperBound : LowerBound;
695  UpperBound = UpperBound->isObjCIdType() ? LowerBound : UpperBound;
696
697  if (storeWhenMoreInformative(State, Sym, TrackedType, LowerBound, UpperBound,
698                               ASTCtxt)) {
699    C.addTransition(State, AfterTypeProp);
700  }
701}
702
703static const Expr *stripCastsAndSugar(const Expr *E) {
704  E = E->IgnoreParenImpCasts();
705  if (const PseudoObjectExpr *POE = dyn_cast<PseudoObjectExpr>(E))
706    E = POE->getSyntacticForm()->IgnoreParenImpCasts();
707  if (const OpaqueValueExpr *OVE = dyn_cast<OpaqueValueExpr>(E))
708    E = OVE->getSourceExpr()->IgnoreParenImpCasts();
709  return E;
710}
711
712static bool isObjCTypeParamDependent(QualType Type) {
713  // It is illegal to typedef parameterized types inside an interface. Therefore
714  // an Objective-C type can only be dependent on a type parameter when the type
715  // parameter structurally present in the type itself.
716  class IsObjCTypeParamDependentTypeVisitor
717      : public RecursiveASTVisitor<IsObjCTypeParamDependentTypeVisitor> {
718  public:
719    IsObjCTypeParamDependentTypeVisitor() : Result(false) {}
720    bool VisitObjCTypeParamType(const ObjCTypeParamType *Type) {
721      if (isa<ObjCTypeParamDecl>(Type->getDecl())) {
722        Result = true;
723        return false;
724      }
725      return true;
726    }
727
728    bool Result;
729  };
730
731  IsObjCTypeParamDependentTypeVisitor Visitor;
732  Visitor.TraverseType(Type);
733  return Visitor.Result;
734}
735
736/// A method might not be available in the interface indicated by the static
737/// type. However it might be available in the tracked type. In order to
738/// properly substitute the type parameters we need the declaration context of
739/// the method. The more specialized the enclosing class of the method is, the
740/// more likely that the parameter substitution will be successful.
741static const ObjCMethodDecl *
742findMethodDecl(const ObjCMessageExpr *MessageExpr,
743               const ObjCObjectPointerType *TrackedType, ASTContext &ASTCtxt) {
744  const ObjCMethodDecl *Method = nullptr;
745
746  QualType ReceiverType = MessageExpr->getReceiverType();
747  const auto *ReceiverObjectPtrType =
748      ReceiverType->getAs<ObjCObjectPointerType>();
749
750  // Do this "devirtualization" on instance and class methods only. Trust the
751  // static type on super and super class calls.
752  if (MessageExpr->getReceiverKind() == ObjCMessageExpr::Instance ||
753      MessageExpr->getReceiverKind() == ObjCMessageExpr::Class) {
754    // When the receiver type is id, Class, or some super class of the tracked
755    // type, look up the method in the tracked type, not in the receiver type.
756    // This way we preserve more information.
757    if (ReceiverType->isObjCIdType() || ReceiverType->isObjCClassType() ||
758        ASTCtxt.canAssignObjCInterfaces(ReceiverObjectPtrType, TrackedType)) {
759      const ObjCInterfaceDecl *InterfaceDecl = TrackedType->getInterfaceDecl();
760      // The method might not be found.
761      Selector Sel = MessageExpr->getSelector();
762      Method = InterfaceDecl->lookupInstanceMethod(Sel);
763      if (!Method)
764        Method = InterfaceDecl->lookupClassMethod(Sel);
765    }
766  }
767
768  // Fallback to statick method lookup when the one based on the tracked type
769  // failed.
770  return Method ? Method : MessageExpr->getMethodDecl();
771}
772
773/// Get the returned ObjCObjectPointerType by a method based on the tracked type
774/// information, or null pointer when the returned type is not an
775/// ObjCObjectPointerType.
776static QualType getReturnTypeForMethod(
777    const ObjCMethodDecl *Method, ArrayRef<QualType> TypeArgs,
778    const ObjCObjectPointerType *SelfType, ASTContext &C) {
779  QualType StaticResultType = Method->getReturnType();
780
781  // Is the return type declared as instance type?
782  if (StaticResultType == C.getObjCInstanceType())
783    return QualType(SelfType, 0);
784
785  // Check whether the result type depends on a type parameter.
786  if (!isObjCTypeParamDependent(StaticResultType))
787    return QualType();
788
789  QualType ResultType = StaticResultType.substObjCTypeArgs(
790      C, TypeArgs, ObjCSubstitutionContext::Result);
791
792  return ResultType;
793}
794
795/// When the receiver has a tracked type, use that type to validate the
796/// argumments of the message expression and the return value.
797void DynamicTypePropagation::checkPreObjCMessage(const ObjCMethodCall &M,
798                                                 CheckerContext &C) const {
799  ProgramStateRef State = C.getState();
800  SymbolRef Sym = M.getReceiverSVal().getAsSymbol();
801  if (!Sym)
802    return;
803
804  const ObjCObjectPointerType *const *TrackedType =
805      State->get<MostSpecializedTypeArgsMap>(Sym);
806  if (!TrackedType)
807    return;
808
809  // Get the type arguments from tracked type and substitute type arguments
810  // before do the semantic check.
811
812  ASTContext &ASTCtxt = C.getASTContext();
813  const ObjCMessageExpr *MessageExpr = M.getOriginExpr();
814  const ObjCMethodDecl *Method =
815      findMethodDecl(MessageExpr, *TrackedType, ASTCtxt);
816
817  // It is possible to call non-existent methods in Obj-C.
818  if (!Method)
819    return;
820
821  // If the method is declared on a class that has a non-invariant
822  // type parameter, don't warn about parameter mismatches after performing
823  // substitution. This prevents warning when the programmer has purposely
824  // casted the receiver to a super type or unspecialized type but the analyzer
825  // has a more precise tracked type than the programmer intends at the call
826  // site.
827  //
828  // For example, consider NSArray (which has a covariant type parameter)
829  // and NSMutableArray (a subclass of NSArray where the type parameter is
830  // invariant):
831  // NSMutableArray *a = [[NSMutableArray<NSString *> alloc] init;
832  //
833  // [a containsObject:number]; // Safe: -containsObject is defined on NSArray.
834  // NSArray<NSObject *> *other = [a arrayByAddingObject:number]  // Safe
835  //
836  // [a addObject:number] // Unsafe: -addObject: is defined on NSMutableArray
837  //
838
839  const ObjCInterfaceDecl *Interface = Method->getClassInterface();
840  if (!Interface)
841    return;
842
843  ObjCTypeParamList *TypeParams = Interface->getTypeParamList();
844  if (!TypeParams)
845    return;
846
847  for (ObjCTypeParamDecl *TypeParam : *TypeParams) {
848    if (TypeParam->getVariance() != ObjCTypeParamVariance::Invariant)
849      return;
850  }
851
852  Optional<ArrayRef<QualType>> TypeArgs =
853      (*TrackedType)->getObjCSubstitutions(Method->getDeclContext());
854  // This case might happen when there is an unspecialized override of a
855  // specialized method.
856  if (!TypeArgs)
857    return;
858
859  for (unsigned i = 0; i < Method->param_size(); i++) {
860    const Expr *Arg = MessageExpr->getArg(i);
861    const ParmVarDecl *Param = Method->parameters()[i];
862
863    QualType OrigParamType = Param->getType();
864    if (!isObjCTypeParamDependent(OrigParamType))
865      continue;
866
867    QualType ParamType = OrigParamType.substObjCTypeArgs(
868        ASTCtxt, *TypeArgs, ObjCSubstitutionContext::Parameter);
869    // Check if it can be assigned
870    const auto *ParamObjectPtrType = ParamType->getAs<ObjCObjectPointerType>();
871    const auto *ArgObjectPtrType =
872        stripCastsAndSugar(Arg)->getType()->getAs<ObjCObjectPointerType>();
873    if (!ParamObjectPtrType || !ArgObjectPtrType)
874      continue;
875
876    // Check if we have more concrete tracked type that is not a super type of
877    // the static argument type.
878    SVal ArgSVal = M.getArgSVal(i);
879    SymbolRef ArgSym = ArgSVal.getAsSymbol();
880    if (ArgSym) {
881      const ObjCObjectPointerType *const *TrackedArgType =
882          State->get<MostSpecializedTypeArgsMap>(ArgSym);
883      if (TrackedArgType &&
884          ASTCtxt.canAssignObjCInterfaces(ArgObjectPtrType, *TrackedArgType)) {
885        ArgObjectPtrType = *TrackedArgType;
886      }
887    }
888
889    // Warn when argument is incompatible with the parameter.
890    if (!ASTCtxt.canAssignObjCInterfaces(ParamObjectPtrType,
891                                         ArgObjectPtrType)) {
892      static CheckerProgramPointTag Tag(this, "ArgTypeMismatch");
893      ExplodedNode *N = C.addTransition(State, &Tag);
894      reportGenericsBug(ArgObjectPtrType, ParamObjectPtrType, N, Sym, C, Arg);
895      return;
896    }
897  }
898}
899
900/// This callback is used to infer the types for Class variables. This info is
901/// used later to validate messages that sent to classes. Class variables are
902/// initialized with by invoking the 'class' method on a class.
903/// This method is also used to infer the type information for the return
904/// types.
905// TODO: right now it only tracks generic types. Extend this to track every
906// type in the DynamicTypeMap and diagnose type errors!
907void DynamicTypePropagation::checkPostObjCMessage(const ObjCMethodCall &M,
908                                                  CheckerContext &C) const {
909  const ObjCMessageExpr *MessageExpr = M.getOriginExpr();
910
911  SymbolRef RetSym = M.getReturnValue().getAsSymbol();
912  if (!RetSym)
913    return;
914
915  Selector Sel = MessageExpr->getSelector();
916  ProgramStateRef State = C.getState();
917
918  // Here we try to propagate information on Class objects.
919  if (Sel.getAsString() == "class") {
920    // We try to figure out the type from the receiver of the 'class' message.
921    if (RuntimeType ReceiverRuntimeType = inferReceiverType(M, C)) {
922
923      ReceiverRuntimeType.Type->getSuperClassType();
924      QualType ReceiverClassType(ReceiverRuntimeType.Type, 0);
925
926      // We want to consider only precise information on generics.
927      if (ReceiverRuntimeType.Type->isSpecialized() &&
928          ReceiverRuntimeType.Precise) {
929        QualType ReceiverClassPointerType =
930            C.getASTContext().getObjCObjectPointerType(ReceiverClassType);
931        const auto *InferredType =
932            ReceiverClassPointerType->castAs<ObjCObjectPointerType>();
933        State = State->set<MostSpecializedTypeArgsMap>(RetSym, InferredType);
934      }
935
936      // Constrain the resulting class object to the inferred type.
937      State = setClassObjectDynamicTypeInfo(State, RetSym, ReceiverClassType,
938                                            !ReceiverRuntimeType.Precise);
939
940      C.addTransition(State);
941      return;
942    }
943  }
944
945  if (Sel.getAsString() == "superclass") {
946    // We try to figure out the type from the receiver of the 'superclass'
947    // message.
948    if (RuntimeType ReceiverRuntimeType = inferReceiverType(M, C)) {
949
950      // Result type would be a super class of the receiver's type.
951      QualType ReceiversSuperClass =
952          ReceiverRuntimeType.Type->getSuperClassType();
953
954      // Check if it really had super class.
955      //
956      // TODO: we can probably pay closer attention to cases when the class
957      // object can be 'nil' as the result of such message.
958      if (!ReceiversSuperClass.isNull()) {
959        // Constrain the resulting class object to the inferred type.
960        State = setClassObjectDynamicTypeInfo(
961            State, RetSym, ReceiversSuperClass, !ReceiverRuntimeType.Precise);
962
963        C.addTransition(State);
964      }
965      return;
966    }
967  }
968
969  // Tracking for return types.
970  SymbolRef RecSym = M.getReceiverSVal().getAsSymbol();
971  if (!RecSym)
972    return;
973
974  const ObjCObjectPointerType *const *TrackedType =
975      State->get<MostSpecializedTypeArgsMap>(RecSym);
976  if (!TrackedType)
977    return;
978
979  ASTContext &ASTCtxt = C.getASTContext();
980  const ObjCMethodDecl *Method =
981      findMethodDecl(MessageExpr, *TrackedType, ASTCtxt);
982  if (!Method)
983    return;
984
985  Optional<ArrayRef<QualType>> TypeArgs =
986      (*TrackedType)->getObjCSubstitutions(Method->getDeclContext());
987  if (!TypeArgs)
988    return;
989
990  QualType ResultType =
991      getReturnTypeForMethod(Method, *TypeArgs, *TrackedType, ASTCtxt);
992  // The static type is the same as the deduced type.
993  if (ResultType.isNull())
994    return;
995
996  const MemRegion *RetRegion = M.getReturnValue().getAsRegion();
997  ExplodedNode *Pred = C.getPredecessor();
998  // When there is an entry available for the return symbol in DynamicTypeMap,
999  // the call was inlined, and the information in the DynamicTypeMap is should
1000  // be precise.
1001  if (RetRegion && !getRawDynamicTypeInfo(State, RetRegion)) {
1002    // TODO: we have duplicated information in DynamicTypeMap and
1003    // MostSpecializedTypeArgsMap. We should only store anything in the later if
1004    // the stored data differs from the one stored in the former.
1005    State = setDynamicTypeInfo(State, RetRegion, ResultType,
1006                               /*CanBeSubClassed=*/true);
1007    Pred = C.addTransition(State);
1008  }
1009
1010  const auto *ResultPtrType = ResultType->getAs<ObjCObjectPointerType>();
1011
1012  if (!ResultPtrType || ResultPtrType->isUnspecialized())
1013    return;
1014
1015  // When the result is a specialized type and it is not tracked yet, track it
1016  // for the result symbol.
1017  if (!State->get<MostSpecializedTypeArgsMap>(RetSym)) {
1018    State = State->set<MostSpecializedTypeArgsMap>(RetSym, ResultPtrType);
1019    C.addTransition(State, Pred);
1020  }
1021}
1022
1023void DynamicTypePropagation::reportGenericsBug(
1024    const ObjCObjectPointerType *From, const ObjCObjectPointerType *To,
1025    ExplodedNode *N, SymbolRef Sym, CheckerContext &C,
1026    const Stmt *ReportedNode) const {
1027  if (!CheckGenerics)
1028    return;
1029
1030  initBugType();
1031  SmallString<192> Buf;
1032  llvm::raw_svector_ostream OS(Buf);
1033  OS << "Conversion from value of type '";
1034  QualType::print(From, Qualifiers(), OS, C.getLangOpts(), llvm::Twine());
1035  OS << "' to incompatible type '";
1036  QualType::print(To, Qualifiers(), OS, C.getLangOpts(), llvm::Twine());
1037  OS << "'";
1038  auto R = std::make_unique<PathSensitiveBugReport>(*ObjCGenericsBugType,
1039                                                    OS.str(), N);
1040  R->markInteresting(Sym);
1041  R->addVisitor(std::make_unique<GenericsBugVisitor>(Sym));
1042  if (ReportedNode)
1043    R->addRange(ReportedNode->getSourceRange());
1044  C.emitReport(std::move(R));
1045}
1046
1047PathDiagnosticPieceRef DynamicTypePropagation::GenericsBugVisitor::VisitNode(
1048    const ExplodedNode *N, BugReporterContext &BRC,
1049    PathSensitiveBugReport &BR) {
1050  ProgramStateRef state = N->getState();
1051  ProgramStateRef statePrev = N->getFirstPred()->getState();
1052
1053  const ObjCObjectPointerType *const *TrackedType =
1054      state->get<MostSpecializedTypeArgsMap>(Sym);
1055  const ObjCObjectPointerType *const *TrackedTypePrev =
1056      statePrev->get<MostSpecializedTypeArgsMap>(Sym);
1057  if (!TrackedType)
1058    return nullptr;
1059
1060  if (TrackedTypePrev && *TrackedTypePrev == *TrackedType)
1061    return nullptr;
1062
1063  // Retrieve the associated statement.
1064  const Stmt *S = N->getStmtForDiagnostics();
1065  if (!S)
1066    return nullptr;
1067
1068  const LangOptions &LangOpts = BRC.getASTContext().getLangOpts();
1069
1070  SmallString<256> Buf;
1071  llvm::raw_svector_ostream OS(Buf);
1072  OS << "Type '";
1073  QualType::print(*TrackedType, Qualifiers(), OS, LangOpts, llvm::Twine());
1074  OS << "' is inferred from ";
1075
1076  if (const auto *ExplicitCast = dyn_cast<ExplicitCastExpr>(S)) {
1077    OS << "explicit cast (from '";
1078    QualType::print(ExplicitCast->getSubExpr()->getType().getTypePtr(),
1079                    Qualifiers(), OS, LangOpts, llvm::Twine());
1080    OS << "' to '";
1081    QualType::print(ExplicitCast->getType().getTypePtr(), Qualifiers(), OS,
1082                    LangOpts, llvm::Twine());
1083    OS << "')";
1084  } else if (const auto *ImplicitCast = dyn_cast<ImplicitCastExpr>(S)) {
1085    OS << "implicit cast (from '";
1086    QualType::print(ImplicitCast->getSubExpr()->getType().getTypePtr(),
1087                    Qualifiers(), OS, LangOpts, llvm::Twine());
1088    OS << "' to '";
1089    QualType::print(ImplicitCast->getType().getTypePtr(), Qualifiers(), OS,
1090                    LangOpts, llvm::Twine());
1091    OS << "')";
1092  } else {
1093    OS << "this context";
1094  }
1095
1096  // Generate the extra diagnostic.
1097  PathDiagnosticLocation Pos(S, BRC.getSourceManager(),
1098                             N->getLocationContext());
1099  return std::make_shared<PathDiagnosticEventPiece>(Pos, OS.str(), true);
1100}
1101
1102/// Register checkers.
1103void ento::registerObjCGenericsChecker(CheckerManager &mgr) {
1104  DynamicTypePropagation *checker = mgr.getChecker<DynamicTypePropagation>();
1105  checker->CheckGenerics = true;
1106  checker->GenericCheckName = mgr.getCurrentCheckerName();
1107}
1108
1109bool ento::shouldRegisterObjCGenericsChecker(const CheckerManager &mgr) {
1110  return true;
1111}
1112
1113void ento::registerDynamicTypePropagation(CheckerManager &mgr) {
1114  mgr.registerChecker<DynamicTypePropagation>();
1115}
1116
1117bool ento::shouldRegisterDynamicTypePropagation(const CheckerManager &mgr) {
1118  return true;
1119}
1120