1 //== Store.cpp - Interface for maps from Locations to Values ----*- 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 defined the types Store and StoreManager.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "clang/StaticAnalyzer/Core/PathSensitive/Store.h"
15 #include "clang/AST/CXXInheritance.h"
16 #include "clang/AST/CharUnits.h"
17 #include "clang/AST/DeclObjC.h"
18 #include "clang/StaticAnalyzer/Core/PathSensitive/CallEvent.h"
19 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h"
20
21 using namespace clang;
22 using namespace ento;
23
StoreManager(ProgramStateManager & stateMgr)24 StoreManager::StoreManager(ProgramStateManager &stateMgr)
25 : svalBuilder(stateMgr.getSValBuilder()), StateMgr(stateMgr),
26 MRMgr(svalBuilder.getRegionManager()), Ctx(stateMgr.getContext()) {}
27
enterStackFrame(Store OldStore,const CallEvent & Call,const StackFrameContext * LCtx)28 StoreRef StoreManager::enterStackFrame(Store OldStore,
29 const CallEvent &Call,
30 const StackFrameContext *LCtx) {
31 StoreRef Store = StoreRef(OldStore, *this);
32
33 SmallVector<CallEvent::FrameBindingTy, 16> InitialBindings;
34 Call.getInitialStackFrameContents(LCtx, InitialBindings);
35
36 for (CallEvent::BindingsTy::iterator I = InitialBindings.begin(),
37 E = InitialBindings.end();
38 I != E; ++I) {
39 Store = Bind(Store.getStore(), I->first, I->second);
40 }
41
42 return Store;
43 }
44
MakeElementRegion(const MemRegion * Base,QualType EleTy,uint64_t index)45 const MemRegion *StoreManager::MakeElementRegion(const MemRegion *Base,
46 QualType EleTy, uint64_t index) {
47 NonLoc idx = svalBuilder.makeArrayIndex(index);
48 return MRMgr.getElementRegion(EleTy, idx, Base, svalBuilder.getContext());
49 }
50
51 // FIXME: Merge with the implementation of the same method in MemRegion.cpp
IsCompleteType(ASTContext & Ctx,QualType Ty)52 static bool IsCompleteType(ASTContext &Ctx, QualType Ty) {
53 if (const RecordType *RT = Ty->getAs<RecordType>()) {
54 const RecordDecl *D = RT->getDecl();
55 if (!D->getDefinition())
56 return false;
57 }
58
59 return true;
60 }
61
BindDefault(Store store,const MemRegion * R,SVal V)62 StoreRef StoreManager::BindDefault(Store store, const MemRegion *R, SVal V) {
63 return StoreRef(store, *this);
64 }
65
GetElementZeroRegion(const MemRegion * R,QualType T)66 const ElementRegion *StoreManager::GetElementZeroRegion(const MemRegion *R,
67 QualType T) {
68 NonLoc idx = svalBuilder.makeZeroArrayIndex();
69 assert(!T.isNull());
70 return MRMgr.getElementRegion(T, idx, R, Ctx);
71 }
72
castRegion(const MemRegion * R,QualType CastToTy)73 const MemRegion *StoreManager::castRegion(const MemRegion *R, QualType CastToTy) {
74
75 ASTContext &Ctx = StateMgr.getContext();
76
77 // Handle casts to Objective-C objects.
78 if (CastToTy->isObjCObjectPointerType())
79 return R->StripCasts();
80
81 if (CastToTy->isBlockPointerType()) {
82 // FIXME: We may need different solutions, depending on the symbol
83 // involved. Blocks can be casted to/from 'id', as they can be treated
84 // as Objective-C objects. This could possibly be handled by enhancing
85 // our reasoning of downcasts of symbolic objects.
86 if (isa<CodeTextRegion>(R) || isa<SymbolicRegion>(R))
87 return R;
88
89 // We don't know what to make of it. Return a NULL region, which
90 // will be interpretted as UnknownVal.
91 return nullptr;
92 }
93
94 // Now assume we are casting from pointer to pointer. Other cases should
95 // already be handled.
96 QualType PointeeTy = CastToTy->getPointeeType();
97 QualType CanonPointeeTy = Ctx.getCanonicalType(PointeeTy);
98
99 // Handle casts to void*. We just pass the region through.
100 if (CanonPointeeTy.getLocalUnqualifiedType() == Ctx.VoidTy)
101 return R;
102
103 // Handle casts from compatible types.
104 if (R->isBoundable())
105 if (const TypedValueRegion *TR = dyn_cast<TypedValueRegion>(R)) {
106 QualType ObjTy = Ctx.getCanonicalType(TR->getValueType());
107 if (CanonPointeeTy == ObjTy)
108 return R;
109 }
110
111 // Process region cast according to the kind of the region being cast.
112 switch (R->getKind()) {
113 case MemRegion::CXXThisRegionKind:
114 case MemRegion::GenericMemSpaceRegionKind:
115 case MemRegion::StackLocalsSpaceRegionKind:
116 case MemRegion::StackArgumentsSpaceRegionKind:
117 case MemRegion::HeapSpaceRegionKind:
118 case MemRegion::UnknownSpaceRegionKind:
119 case MemRegion::StaticGlobalSpaceRegionKind:
120 case MemRegion::GlobalInternalSpaceRegionKind:
121 case MemRegion::GlobalSystemSpaceRegionKind:
122 case MemRegion::GlobalImmutableSpaceRegionKind: {
123 llvm_unreachable("Invalid region cast");
124 }
125
126 case MemRegion::FunctionTextRegionKind:
127 case MemRegion::BlockTextRegionKind:
128 case MemRegion::BlockDataRegionKind:
129 case MemRegion::StringRegionKind:
130 // FIXME: Need to handle arbitrary downcasts.
131 case MemRegion::SymbolicRegionKind:
132 case MemRegion::AllocaRegionKind:
133 case MemRegion::CompoundLiteralRegionKind:
134 case MemRegion::FieldRegionKind:
135 case MemRegion::ObjCIvarRegionKind:
136 case MemRegion::ObjCStringRegionKind:
137 case MemRegion::VarRegionKind:
138 case MemRegion::CXXTempObjectRegionKind:
139 case MemRegion::CXXBaseObjectRegionKind:
140 return MakeElementRegion(R, PointeeTy);
141
142 case MemRegion::ElementRegionKind: {
143 // If we are casting from an ElementRegion to another type, the
144 // algorithm is as follows:
145 //
146 // (1) Compute the "raw offset" of the ElementRegion from the
147 // base region. This is done by calling 'getAsRawOffset()'.
148 //
149 // (2a) If we get a 'RegionRawOffset' after calling
150 // 'getAsRawOffset()', determine if the absolute offset
151 // can be exactly divided into chunks of the size of the
152 // casted-pointee type. If so, create a new ElementRegion with
153 // the pointee-cast type as the new ElementType and the index
154 // being the offset divded by the chunk size. If not, create
155 // a new ElementRegion at offset 0 off the raw offset region.
156 //
157 // (2b) If we don't a get a 'RegionRawOffset' after calling
158 // 'getAsRawOffset()', it means that we are at offset 0.
159 //
160 // FIXME: Handle symbolic raw offsets.
161
162 const ElementRegion *elementR = cast<ElementRegion>(R);
163 const RegionRawOffset &rawOff = elementR->getAsArrayOffset();
164 const MemRegion *baseR = rawOff.getRegion();
165
166 // If we cannot compute a raw offset, throw up our hands and return
167 // a NULL MemRegion*.
168 if (!baseR)
169 return nullptr;
170
171 CharUnits off = rawOff.getOffset();
172
173 if (off.isZero()) {
174 // Edge case: we are at 0 bytes off the beginning of baseR. We
175 // check to see if type we are casting to is the same as the base
176 // region. If so, just return the base region.
177 if (const TypedValueRegion *TR = dyn_cast<TypedValueRegion>(baseR)) {
178 QualType ObjTy = Ctx.getCanonicalType(TR->getValueType());
179 QualType CanonPointeeTy = Ctx.getCanonicalType(PointeeTy);
180 if (CanonPointeeTy == ObjTy)
181 return baseR;
182 }
183
184 // Otherwise, create a new ElementRegion at offset 0.
185 return MakeElementRegion(baseR, PointeeTy);
186 }
187
188 // We have a non-zero offset from the base region. We want to determine
189 // if the offset can be evenly divided by sizeof(PointeeTy). If so,
190 // we create an ElementRegion whose index is that value. Otherwise, we
191 // create two ElementRegions, one that reflects a raw offset and the other
192 // that reflects the cast.
193
194 // Compute the index for the new ElementRegion.
195 int64_t newIndex = 0;
196 const MemRegion *newSuperR = nullptr;
197
198 // We can only compute sizeof(PointeeTy) if it is a complete type.
199 if (IsCompleteType(Ctx, PointeeTy)) {
200 // Compute the size in **bytes**.
201 CharUnits pointeeTySize = Ctx.getTypeSizeInChars(PointeeTy);
202 if (!pointeeTySize.isZero()) {
203 // Is the offset a multiple of the size? If so, we can layer the
204 // ElementRegion (with elementType == PointeeTy) directly on top of
205 // the base region.
206 if (off % pointeeTySize == 0) {
207 newIndex = off / pointeeTySize;
208 newSuperR = baseR;
209 }
210 }
211 }
212
213 if (!newSuperR) {
214 // Create an intermediate ElementRegion to represent the raw byte.
215 // This will be the super region of the final ElementRegion.
216 newSuperR = MakeElementRegion(baseR, Ctx.CharTy, off.getQuantity());
217 }
218
219 return MakeElementRegion(newSuperR, PointeeTy, newIndex);
220 }
221 }
222
223 llvm_unreachable("unreachable");
224 }
225
regionMatchesCXXRecordType(SVal V,QualType Ty)226 static bool regionMatchesCXXRecordType(SVal V, QualType Ty) {
227 const MemRegion *MR = V.getAsRegion();
228 if (!MR)
229 return true;
230
231 const TypedValueRegion *TVR = dyn_cast<TypedValueRegion>(MR);
232 if (!TVR)
233 return true;
234
235 const CXXRecordDecl *RD = TVR->getValueType()->getAsCXXRecordDecl();
236 if (!RD)
237 return true;
238
239 const CXXRecordDecl *Expected = Ty->getPointeeCXXRecordDecl();
240 if (!Expected)
241 Expected = Ty->getAsCXXRecordDecl();
242
243 return Expected->getCanonicalDecl() == RD->getCanonicalDecl();
244 }
245
evalDerivedToBase(SVal Derived,const CastExpr * Cast)246 SVal StoreManager::evalDerivedToBase(SVal Derived, const CastExpr *Cast) {
247 // Sanity check to avoid doing the wrong thing in the face of
248 // reinterpret_cast.
249 if (!regionMatchesCXXRecordType(Derived, Cast->getSubExpr()->getType()))
250 return UnknownVal();
251
252 // Walk through the cast path to create nested CXXBaseRegions.
253 SVal Result = Derived;
254 for (CastExpr::path_const_iterator I = Cast->path_begin(),
255 E = Cast->path_end();
256 I != E; ++I) {
257 Result = evalDerivedToBase(Result, (*I)->getType(), (*I)->isVirtual());
258 }
259 return Result;
260 }
261
evalDerivedToBase(SVal Derived,const CXXBasePath & Path)262 SVal StoreManager::evalDerivedToBase(SVal Derived, const CXXBasePath &Path) {
263 // Walk through the path to create nested CXXBaseRegions.
264 SVal Result = Derived;
265 for (CXXBasePath::const_iterator I = Path.begin(), E = Path.end();
266 I != E; ++I) {
267 Result = evalDerivedToBase(Result, I->Base->getType(),
268 I->Base->isVirtual());
269 }
270 return Result;
271 }
272
evalDerivedToBase(SVal Derived,QualType BaseType,bool IsVirtual)273 SVal StoreManager::evalDerivedToBase(SVal Derived, QualType BaseType,
274 bool IsVirtual) {
275 Optional<loc::MemRegionVal> DerivedRegVal =
276 Derived.getAs<loc::MemRegionVal>();
277 if (!DerivedRegVal)
278 return Derived;
279
280 const CXXRecordDecl *BaseDecl = BaseType->getPointeeCXXRecordDecl();
281 if (!BaseDecl)
282 BaseDecl = BaseType->getAsCXXRecordDecl();
283 assert(BaseDecl && "not a C++ object?");
284
285 const MemRegion *BaseReg =
286 MRMgr.getCXXBaseObjectRegion(BaseDecl, DerivedRegVal->getRegion(),
287 IsVirtual);
288
289 return loc::MemRegionVal(BaseReg);
290 }
291
292 /// Returns the static type of the given region, if it represents a C++ class
293 /// object.
294 ///
295 /// This handles both fully-typed regions, where the dynamic type is known, and
296 /// symbolic regions, where the dynamic type is merely bounded (and even then,
297 /// only ostensibly!), but does not take advantage of any dynamic type info.
getCXXRecordType(const MemRegion * MR)298 static const CXXRecordDecl *getCXXRecordType(const MemRegion *MR) {
299 if (const TypedValueRegion *TVR = dyn_cast<TypedValueRegion>(MR))
300 return TVR->getValueType()->getAsCXXRecordDecl();
301 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(MR))
302 return SR->getSymbol()->getType()->getPointeeCXXRecordDecl();
303 return nullptr;
304 }
305
evalDynamicCast(SVal Base,QualType TargetType,bool & Failed)306 SVal StoreManager::evalDynamicCast(SVal Base, QualType TargetType,
307 bool &Failed) {
308 Failed = false;
309
310 const MemRegion *MR = Base.getAsRegion();
311 if (!MR)
312 return UnknownVal();
313
314 // Assume the derived class is a pointer or a reference to a CXX record.
315 TargetType = TargetType->getPointeeType();
316 assert(!TargetType.isNull());
317 const CXXRecordDecl *TargetClass = TargetType->getAsCXXRecordDecl();
318 if (!TargetClass && !TargetType->isVoidType())
319 return UnknownVal();
320
321 // Drill down the CXXBaseObject chains, which represent upcasts (casts from
322 // derived to base).
323 while (const CXXRecordDecl *MRClass = getCXXRecordType(MR)) {
324 // If found the derived class, the cast succeeds.
325 if (MRClass == TargetClass)
326 return loc::MemRegionVal(MR);
327
328 // We skip over incomplete types. They must be the result of an earlier
329 // reinterpret_cast, as one can only dynamic_cast between types in the same
330 // class hierarchy.
331 if (!TargetType->isVoidType() && MRClass->hasDefinition()) {
332 // Static upcasts are marked as DerivedToBase casts by Sema, so this will
333 // only happen when multiple or virtual inheritance is involved.
334 CXXBasePaths Paths(/*FindAmbiguities=*/false, /*RecordPaths=*/true,
335 /*DetectVirtual=*/false);
336 if (MRClass->isDerivedFrom(TargetClass, Paths))
337 return evalDerivedToBase(loc::MemRegionVal(MR), Paths.front());
338 }
339
340 if (const CXXBaseObjectRegion *BaseR = dyn_cast<CXXBaseObjectRegion>(MR)) {
341 // Drill down the chain to get the derived classes.
342 MR = BaseR->getSuperRegion();
343 continue;
344 }
345
346 // If this is a cast to void*, return the region.
347 if (TargetType->isVoidType())
348 return loc::MemRegionVal(MR);
349
350 // Strange use of reinterpret_cast can give us paths we don't reason
351 // about well, by putting in ElementRegions where we'd expect
352 // CXXBaseObjectRegions. If it's a valid reinterpret_cast (i.e. if the
353 // derived class has a zero offset from the base class), then it's safe
354 // to strip the cast; if it's invalid, -Wreinterpret-base-class should
355 // catch it. In the interest of performance, the analyzer will silently
356 // do the wrong thing in the invalid case (because offsets for subregions
357 // will be wrong).
358 const MemRegion *Uncasted = MR->StripCasts(/*IncludeBaseCasts=*/false);
359 if (Uncasted == MR) {
360 // We reached the bottom of the hierarchy and did not find the derived
361 // class. We we must be casting the base to derived, so the cast should
362 // fail.
363 break;
364 }
365
366 MR = Uncasted;
367 }
368
369 // We failed if the region we ended up with has perfect type info.
370 Failed = isa<TypedValueRegion>(MR);
371 return UnknownVal();
372 }
373
374
375 /// CastRetrievedVal - Used by subclasses of StoreManager to implement
376 /// implicit casts that arise from loads from regions that are reinterpreted
377 /// as another region.
CastRetrievedVal(SVal V,const TypedValueRegion * R,QualType castTy,bool performTestOnly)378 SVal StoreManager::CastRetrievedVal(SVal V, const TypedValueRegion *R,
379 QualType castTy, bool performTestOnly) {
380
381 if (castTy.isNull() || V.isUnknownOrUndef())
382 return V;
383
384 ASTContext &Ctx = svalBuilder.getContext();
385
386 if (performTestOnly) {
387 // Automatically translate references to pointers.
388 QualType T = R->getValueType();
389 if (const ReferenceType *RT = T->getAs<ReferenceType>())
390 T = Ctx.getPointerType(RT->getPointeeType());
391
392 assert(svalBuilder.getContext().hasSameUnqualifiedType(castTy, T));
393 return V;
394 }
395
396 return svalBuilder.dispatchCast(V, castTy);
397 }
398
getLValueFieldOrIvar(const Decl * D,SVal Base)399 SVal StoreManager::getLValueFieldOrIvar(const Decl *D, SVal Base) {
400 if (Base.isUnknownOrUndef())
401 return Base;
402
403 Loc BaseL = Base.castAs<Loc>();
404 const MemRegion* BaseR = nullptr;
405
406 switch (BaseL.getSubKind()) {
407 case loc::MemRegionKind:
408 BaseR = BaseL.castAs<loc::MemRegionVal>().getRegion();
409 break;
410
411 case loc::GotoLabelKind:
412 // These are anormal cases. Flag an undefined value.
413 return UndefinedVal();
414
415 case loc::ConcreteIntKind:
416 // While these seem funny, this can happen through casts.
417 // FIXME: What we should return is the field offset. For example,
418 // add the field offset to the integer value. That way funny things
419 // like this work properly: &(((struct foo *) 0xa)->f)
420 return Base;
421
422 default:
423 llvm_unreachable("Unhandled Base.");
424 }
425
426 // NOTE: We must have this check first because ObjCIvarDecl is a subclass
427 // of FieldDecl.
428 if (const ObjCIvarDecl *ID = dyn_cast<ObjCIvarDecl>(D))
429 return loc::MemRegionVal(MRMgr.getObjCIvarRegion(ID, BaseR));
430
431 return loc::MemRegionVal(MRMgr.getFieldRegion(cast<FieldDecl>(D), BaseR));
432 }
433
getLValueIvar(const ObjCIvarDecl * decl,SVal base)434 SVal StoreManager::getLValueIvar(const ObjCIvarDecl *decl, SVal base) {
435 return getLValueFieldOrIvar(decl, base);
436 }
437
getLValueElement(QualType elementType,NonLoc Offset,SVal Base)438 SVal StoreManager::getLValueElement(QualType elementType, NonLoc Offset,
439 SVal Base) {
440
441 // If the base is an unknown or undefined value, just return it back.
442 // FIXME: For absolute pointer addresses, we just return that value back as
443 // well, although in reality we should return the offset added to that
444 // value.
445 if (Base.isUnknownOrUndef() || Base.getAs<loc::ConcreteInt>())
446 return Base;
447
448 const MemRegion* BaseRegion = Base.castAs<loc::MemRegionVal>().getRegion();
449
450 // Pointer of any type can be cast and used as array base.
451 const ElementRegion *ElemR = dyn_cast<ElementRegion>(BaseRegion);
452
453 // Convert the offset to the appropriate size and signedness.
454 Offset = svalBuilder.convertToArrayIndex(Offset).castAs<NonLoc>();
455
456 if (!ElemR) {
457 //
458 // If the base region is not an ElementRegion, create one.
459 // This can happen in the following example:
460 //
461 // char *p = __builtin_alloc(10);
462 // p[1] = 8;
463 //
464 // Observe that 'p' binds to an AllocaRegion.
465 //
466 return loc::MemRegionVal(MRMgr.getElementRegion(elementType, Offset,
467 BaseRegion, Ctx));
468 }
469
470 SVal BaseIdx = ElemR->getIndex();
471
472 if (!BaseIdx.getAs<nonloc::ConcreteInt>())
473 return UnknownVal();
474
475 const llvm::APSInt &BaseIdxI =
476 BaseIdx.castAs<nonloc::ConcreteInt>().getValue();
477
478 // Only allow non-integer offsets if the base region has no offset itself.
479 // FIXME: This is a somewhat arbitrary restriction. We should be using
480 // SValBuilder here to add the two offsets without checking their types.
481 if (!Offset.getAs<nonloc::ConcreteInt>()) {
482 if (isa<ElementRegion>(BaseRegion->StripCasts()))
483 return UnknownVal();
484
485 return loc::MemRegionVal(MRMgr.getElementRegion(elementType, Offset,
486 ElemR->getSuperRegion(),
487 Ctx));
488 }
489
490 const llvm::APSInt& OffI = Offset.castAs<nonloc::ConcreteInt>().getValue();
491 assert(BaseIdxI.isSigned());
492
493 // Compute the new index.
494 nonloc::ConcreteInt NewIdx(svalBuilder.getBasicValueFactory().getValue(BaseIdxI +
495 OffI));
496
497 // Construct the new ElementRegion.
498 const MemRegion *ArrayR = ElemR->getSuperRegion();
499 return loc::MemRegionVal(MRMgr.getElementRegion(elementType, NewIdx, ArrayR,
500 Ctx));
501 }
502
~BindingsHandler()503 StoreManager::BindingsHandler::~BindingsHandler() {}
504
HandleBinding(StoreManager & SMgr,Store store,const MemRegion * R,SVal val)505 bool StoreManager::FindUniqueBinding::HandleBinding(StoreManager& SMgr,
506 Store store,
507 const MemRegion* R,
508 SVal val) {
509 SymbolRef SymV = val.getAsLocSymbol();
510 if (!SymV || SymV != Sym)
511 return true;
512
513 if (Binding) {
514 First = false;
515 return false;
516 }
517 else
518 Binding = R;
519
520 return true;
521 }
522