1 //===- Store.cpp - Interface for maps from Locations to Values ------------===//
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 defined the types Store and StoreManager.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "clang/StaticAnalyzer/Core/PathSensitive/Store.h"
14 #include "clang/AST/ASTContext.h"
15 #include "clang/AST/CXXInheritance.h"
16 #include "clang/AST/CharUnits.h"
17 #include "clang/AST/Decl.h"
18 #include "clang/AST/DeclCXX.h"
19 #include "clang/AST/DeclObjC.h"
20 #include "clang/AST/Expr.h"
21 #include "clang/AST/Type.h"
22 #include "clang/Basic/LLVM.h"
23 #include "clang/StaticAnalyzer/Core/PathSensitive/BasicValueFactory.h"
24 #include "clang/StaticAnalyzer/Core/PathSensitive/CallEvent.h"
25 #include "clang/StaticAnalyzer/Core/PathSensitive/MemRegion.h"
26 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h"
27 #include "clang/StaticAnalyzer/Core/PathSensitive/SValBuilder.h"
28 #include "clang/StaticAnalyzer/Core/PathSensitive/SVals.h"
29 #include "clang/StaticAnalyzer/Core/PathSensitive/StoreRef.h"
30 #include "clang/StaticAnalyzer/Core/PathSensitive/SymExpr.h"
31 #include "llvm/ADT/APSInt.h"
32 #include "llvm/ADT/Optional.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/Support/Casting.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include <cassert>
37 #include <cstdint>
38
39 using namespace clang;
40 using namespace ento;
41
StoreManager(ProgramStateManager & stateMgr)42 StoreManager::StoreManager(ProgramStateManager &stateMgr)
43 : svalBuilder(stateMgr.getSValBuilder()), StateMgr(stateMgr),
44 MRMgr(svalBuilder.getRegionManager()), Ctx(stateMgr.getContext()) {}
45
enterStackFrame(Store OldStore,const CallEvent & Call,const StackFrameContext * LCtx)46 StoreRef StoreManager::enterStackFrame(Store OldStore,
47 const CallEvent &Call,
48 const StackFrameContext *LCtx) {
49 StoreRef Store = StoreRef(OldStore, *this);
50
51 SmallVector<CallEvent::FrameBindingTy, 16> InitialBindings;
52 Call.getInitialStackFrameContents(LCtx, InitialBindings);
53
54 for (const auto &I : InitialBindings)
55 Store = Bind(Store.getStore(), I.first.castAs<Loc>(), I.second);
56
57 return Store;
58 }
59
MakeElementRegion(const SubRegion * Base,QualType EleTy,uint64_t index)60 const ElementRegion *StoreManager::MakeElementRegion(const SubRegion *Base,
61 QualType EleTy,
62 uint64_t index) {
63 NonLoc idx = svalBuilder.makeArrayIndex(index);
64 return MRMgr.getElementRegion(EleTy, idx, Base, svalBuilder.getContext());
65 }
66
GetElementZeroRegion(const SubRegion * R,QualType T)67 const ElementRegion *StoreManager::GetElementZeroRegion(const SubRegion *R,
68 QualType T) {
69 NonLoc idx = svalBuilder.makeZeroArrayIndex();
70 assert(!T.isNull());
71 return MRMgr.getElementRegion(T, idx, R, Ctx);
72 }
73
castRegion(const MemRegion * R,QualType CastToTy)74 const MemRegion *StoreManager::castRegion(const MemRegion *R, QualType CastToTy) {
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 interpreted 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 auto *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::CodeSpaceRegionKind:
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::FunctionCodeRegionKind:
127 case MemRegion::BlockCodeRegionKind:
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::NonParamVarRegionKind:
138 case MemRegion::ParamVarRegionKind:
139 case MemRegion::CXXTempObjectRegionKind:
140 case MemRegion::CXXBaseObjectRegionKind:
141 case MemRegion::CXXDerivedObjectRegionKind:
142 return MakeElementRegion(cast<SubRegion>(R), PointeeTy);
143
144 case MemRegion::ElementRegionKind: {
145 // If we are casting from an ElementRegion to another type, the
146 // algorithm is as follows:
147 //
148 // (1) Compute the "raw offset" of the ElementRegion from the
149 // base region. This is done by calling 'getAsRawOffset()'.
150 //
151 // (2a) If we get a 'RegionRawOffset' after calling
152 // 'getAsRawOffset()', determine if the absolute offset
153 // can be exactly divided into chunks of the size of the
154 // casted-pointee type. If so, create a new ElementRegion with
155 // the pointee-cast type as the new ElementType and the index
156 // being the offset divded by the chunk size. If not, create
157 // a new ElementRegion at offset 0 off the raw offset region.
158 //
159 // (2b) If we don't a get a 'RegionRawOffset' after calling
160 // 'getAsRawOffset()', it means that we are at offset 0.
161 //
162 // FIXME: Handle symbolic raw offsets.
163
164 const ElementRegion *elementR = cast<ElementRegion>(R);
165 const RegionRawOffset &rawOff = elementR->getAsArrayOffset();
166 const MemRegion *baseR = rawOff.getRegion();
167
168 // If we cannot compute a raw offset, throw up our hands and return
169 // a NULL MemRegion*.
170 if (!baseR)
171 return nullptr;
172
173 CharUnits off = rawOff.getOffset();
174
175 if (off.isZero()) {
176 // Edge case: we are at 0 bytes off the beginning of baseR. We
177 // check to see if type we are casting to is the same as the base
178 // region. If so, just return the base region.
179 if (const auto *TR = dyn_cast<TypedValueRegion>(baseR)) {
180 QualType ObjTy = Ctx.getCanonicalType(TR->getValueType());
181 QualType CanonPointeeTy = Ctx.getCanonicalType(PointeeTy);
182 if (CanonPointeeTy == ObjTy)
183 return baseR;
184 }
185
186 // Otherwise, create a new ElementRegion at offset 0.
187 return MakeElementRegion(cast<SubRegion>(baseR), PointeeTy);
188 }
189
190 // We have a non-zero offset from the base region. We want to determine
191 // if the offset can be evenly divided by sizeof(PointeeTy). If so,
192 // we create an ElementRegion whose index is that value. Otherwise, we
193 // create two ElementRegions, one that reflects a raw offset and the other
194 // that reflects the cast.
195
196 // Compute the index for the new ElementRegion.
197 int64_t newIndex = 0;
198 const MemRegion *newSuperR = nullptr;
199
200 // We can only compute sizeof(PointeeTy) if it is a complete type.
201 if (!PointeeTy->isIncompleteType()) {
202 // Compute the size in **bytes**.
203 CharUnits pointeeTySize = Ctx.getTypeSizeInChars(PointeeTy);
204 if (!pointeeTySize.isZero()) {
205 // Is the offset a multiple of the size? If so, we can layer the
206 // ElementRegion (with elementType == PointeeTy) directly on top of
207 // the base region.
208 if (off % pointeeTySize == 0) {
209 newIndex = off / pointeeTySize;
210 newSuperR = baseR;
211 }
212 }
213 }
214
215 if (!newSuperR) {
216 // Create an intermediate ElementRegion to represent the raw byte.
217 // This will be the super region of the final ElementRegion.
218 newSuperR = MakeElementRegion(cast<SubRegion>(baseR), Ctx.CharTy,
219 off.getQuantity());
220 }
221
222 return MakeElementRegion(cast<SubRegion>(newSuperR), PointeeTy, newIndex);
223 }
224 }
225
226 llvm_unreachable("unreachable");
227 }
228
regionMatchesCXXRecordType(SVal V,QualType Ty)229 static bool regionMatchesCXXRecordType(SVal V, QualType Ty) {
230 const MemRegion *MR = V.getAsRegion();
231 if (!MR)
232 return true;
233
234 const auto *TVR = dyn_cast<TypedValueRegion>(MR);
235 if (!TVR)
236 return true;
237
238 const CXXRecordDecl *RD = TVR->getValueType()->getAsCXXRecordDecl();
239 if (!RD)
240 return true;
241
242 const CXXRecordDecl *Expected = Ty->getPointeeCXXRecordDecl();
243 if (!Expected)
244 Expected = Ty->getAsCXXRecordDecl();
245
246 return Expected->getCanonicalDecl() == RD->getCanonicalDecl();
247 }
248
evalDerivedToBase(SVal Derived,const CastExpr * Cast)249 SVal StoreManager::evalDerivedToBase(SVal Derived, const CastExpr *Cast) {
250 // Sanity check to avoid doing the wrong thing in the face of
251 // reinterpret_cast.
252 if (!regionMatchesCXXRecordType(Derived, Cast->getSubExpr()->getType()))
253 return UnknownVal();
254
255 // Walk through the cast path to create nested CXXBaseRegions.
256 SVal Result = Derived;
257 for (CastExpr::path_const_iterator I = Cast->path_begin(),
258 E = Cast->path_end();
259 I != E; ++I) {
260 Result = evalDerivedToBase(Result, (*I)->getType(), (*I)->isVirtual());
261 }
262 return Result;
263 }
264
evalDerivedToBase(SVal Derived,const CXXBasePath & Path)265 SVal StoreManager::evalDerivedToBase(SVal Derived, const CXXBasePath &Path) {
266 // Walk through the path to create nested CXXBaseRegions.
267 SVal Result = Derived;
268 for (const auto &I : Path)
269 Result = evalDerivedToBase(Result, I.Base->getType(),
270 I.Base->isVirtual());
271 return Result;
272 }
273
evalDerivedToBase(SVal Derived,QualType BaseType,bool IsVirtual)274 SVal StoreManager::evalDerivedToBase(SVal Derived, QualType BaseType,
275 bool IsVirtual) {
276 const MemRegion *DerivedReg = Derived.getAsRegion();
277 if (!DerivedReg)
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 if (const auto *AlreadyDerivedReg =
286 dyn_cast<CXXDerivedObjectRegion>(DerivedReg)) {
287 if (const auto *SR =
288 dyn_cast<SymbolicRegion>(AlreadyDerivedReg->getSuperRegion()))
289 if (SR->getSymbol()->getType()->getPointeeCXXRecordDecl() == BaseDecl)
290 return loc::MemRegionVal(SR);
291
292 DerivedReg = AlreadyDerivedReg->getSuperRegion();
293 }
294
295 const MemRegion *BaseReg = MRMgr.getCXXBaseObjectRegion(
296 BaseDecl, cast<SubRegion>(DerivedReg), IsVirtual);
297
298 return loc::MemRegionVal(BaseReg);
299 }
300
301 /// Returns the static type of the given region, if it represents a C++ class
302 /// object.
303 ///
304 /// This handles both fully-typed regions, where the dynamic type is known, and
305 /// symbolic regions, where the dynamic type is merely bounded (and even then,
306 /// only ostensibly!), but does not take advantage of any dynamic type info.
getCXXRecordType(const MemRegion * MR)307 static const CXXRecordDecl *getCXXRecordType(const MemRegion *MR) {
308 if (const auto *TVR = dyn_cast<TypedValueRegion>(MR))
309 return TVR->getValueType()->getAsCXXRecordDecl();
310 if (const auto *SR = dyn_cast<SymbolicRegion>(MR))
311 return SR->getSymbol()->getType()->getPointeeCXXRecordDecl();
312 return nullptr;
313 }
314
attemptDownCast(SVal Base,QualType TargetType,bool & Failed)315 SVal StoreManager::attemptDownCast(SVal Base, QualType TargetType,
316 bool &Failed) {
317 Failed = false;
318
319 const MemRegion *MR = Base.getAsRegion();
320 if (!MR)
321 return UnknownVal();
322
323 // Assume the derived class is a pointer or a reference to a CXX record.
324 TargetType = TargetType->getPointeeType();
325 assert(!TargetType.isNull());
326 const CXXRecordDecl *TargetClass = TargetType->getAsCXXRecordDecl();
327 if (!TargetClass && !TargetType->isVoidType())
328 return UnknownVal();
329
330 // Drill down the CXXBaseObject chains, which represent upcasts (casts from
331 // derived to base).
332 while (const CXXRecordDecl *MRClass = getCXXRecordType(MR)) {
333 // If found the derived class, the cast succeeds.
334 if (MRClass == TargetClass)
335 return loc::MemRegionVal(MR);
336
337 // We skip over incomplete types. They must be the result of an earlier
338 // reinterpret_cast, as one can only dynamic_cast between types in the same
339 // class hierarchy.
340 if (!TargetType->isVoidType() && MRClass->hasDefinition()) {
341 // Static upcasts are marked as DerivedToBase casts by Sema, so this will
342 // only happen when multiple or virtual inheritance is involved.
343 CXXBasePaths Paths(/*FindAmbiguities=*/false, /*RecordPaths=*/true,
344 /*DetectVirtual=*/false);
345 if (MRClass->isDerivedFrom(TargetClass, Paths))
346 return evalDerivedToBase(loc::MemRegionVal(MR), Paths.front());
347 }
348
349 if (const auto *BaseR = dyn_cast<CXXBaseObjectRegion>(MR)) {
350 // Drill down the chain to get the derived classes.
351 MR = BaseR->getSuperRegion();
352 continue;
353 }
354
355 // If this is a cast to void*, return the region.
356 if (TargetType->isVoidType())
357 return loc::MemRegionVal(MR);
358
359 // Strange use of reinterpret_cast can give us paths we don't reason
360 // about well, by putting in ElementRegions where we'd expect
361 // CXXBaseObjectRegions. If it's a valid reinterpret_cast (i.e. if the
362 // derived class has a zero offset from the base class), then it's safe
363 // to strip the cast; if it's invalid, -Wreinterpret-base-class should
364 // catch it. In the interest of performance, the analyzer will silently
365 // do the wrong thing in the invalid case (because offsets for subregions
366 // will be wrong).
367 const MemRegion *Uncasted = MR->StripCasts(/*IncludeBaseCasts=*/false);
368 if (Uncasted == MR) {
369 // We reached the bottom of the hierarchy and did not find the derived
370 // class. We must be casting the base to derived, so the cast should
371 // fail.
372 break;
373 }
374
375 MR = Uncasted;
376 }
377
378 // If we're casting a symbolic base pointer to a derived class, use
379 // CXXDerivedObjectRegion to represent the cast. If it's a pointer to an
380 // unrelated type, it must be a weird reinterpret_cast and we have to
381 // be fine with ElementRegion. TODO: Should we instead make
382 // Derived{TargetClass, Element{SourceClass, SR}}?
383 if (const auto *SR = dyn_cast<SymbolicRegion>(MR)) {
384 QualType T = SR->getSymbol()->getType();
385 const CXXRecordDecl *SourceClass = T->getPointeeCXXRecordDecl();
386 if (TargetClass && SourceClass && TargetClass->isDerivedFrom(SourceClass))
387 return loc::MemRegionVal(
388 MRMgr.getCXXDerivedObjectRegion(TargetClass, SR));
389 return loc::MemRegionVal(GetElementZeroRegion(SR, TargetType));
390 }
391
392 // We failed if the region we ended up with has perfect type info.
393 Failed = isa<TypedValueRegion>(MR);
394 return UnknownVal();
395 }
396
getLValueFieldOrIvar(const Decl * D,SVal Base)397 SVal StoreManager::getLValueFieldOrIvar(const Decl *D, SVal Base) {
398 if (Base.isUnknownOrUndef())
399 return Base;
400
401 Loc BaseL = Base.castAs<Loc>();
402 const SubRegion* BaseR = nullptr;
403
404 switch (BaseL.getSubKind()) {
405 case loc::MemRegionValKind:
406 BaseR = cast<SubRegion>(BaseL.castAs<loc::MemRegionVal>().getRegion());
407 break;
408
409 case loc::GotoLabelKind:
410 // These are anormal cases. Flag an undefined value.
411 return UndefinedVal();
412
413 case loc::ConcreteIntKind:
414 // While these seem funny, this can happen through casts.
415 // FIXME: What we should return is the field offset, not base. For example,
416 // add the field offset to the integer value. That way things
417 // like this work properly: &(((struct foo *) 0xa)->f)
418 // However, that's not easy to fix without reducing our abilities
419 // to catch null pointer dereference. Eg., ((struct foo *)0x0)->f = 7
420 // is a null dereference even though we're dereferencing offset of f
421 // rather than null. Coming up with an approach that computes offsets
422 // over null pointers properly while still being able to catch null
423 // dereferences might be worth it.
424 return Base;
425
426 default:
427 llvm_unreachable("Unhandled Base.");
428 }
429
430 // NOTE: We must have this check first because ObjCIvarDecl is a subclass
431 // of FieldDecl.
432 if (const auto *ID = dyn_cast<ObjCIvarDecl>(D))
433 return loc::MemRegionVal(MRMgr.getObjCIvarRegion(ID, BaseR));
434
435 return loc::MemRegionVal(MRMgr.getFieldRegion(cast<FieldDecl>(D), BaseR));
436 }
437
getLValueIvar(const ObjCIvarDecl * decl,SVal base)438 SVal StoreManager::getLValueIvar(const ObjCIvarDecl *decl, SVal base) {
439 return getLValueFieldOrIvar(decl, base);
440 }
441
getLValueElement(QualType elementType,NonLoc Offset,SVal Base)442 SVal StoreManager::getLValueElement(QualType elementType, NonLoc Offset,
443 SVal Base) {
444 // If the base is an unknown or undefined value, just return it back.
445 // FIXME: For absolute pointer addresses, we just return that value back as
446 // well, although in reality we should return the offset added to that
447 // value. See also the similar FIXME in getLValueFieldOrIvar().
448 if (Base.isUnknownOrUndef() || Base.getAs<loc::ConcreteInt>())
449 return Base;
450
451 if (Base.getAs<loc::GotoLabel>())
452 return UnknownVal();
453
454 const SubRegion *BaseRegion =
455 Base.castAs<loc::MemRegionVal>().getRegionAs<SubRegion>();
456
457 // Pointer of any type can be cast and used as array base.
458 const auto *ElemR = dyn_cast<ElementRegion>(BaseRegion);
459
460 // Convert the offset to the appropriate size and signedness.
461 Offset = svalBuilder.convertToArrayIndex(Offset).castAs<NonLoc>();
462
463 if (!ElemR) {
464 // If the base region is not an ElementRegion, create one.
465 // This can happen in the following example:
466 //
467 // char *p = __builtin_alloc(10);
468 // p[1] = 8;
469 //
470 // Observe that 'p' binds to an AllocaRegion.
471 return loc::MemRegionVal(MRMgr.getElementRegion(elementType, Offset,
472 BaseRegion, Ctx));
473 }
474
475 SVal BaseIdx = ElemR->getIndex();
476
477 if (!BaseIdx.getAs<nonloc::ConcreteInt>())
478 return UnknownVal();
479
480 const llvm::APSInt &BaseIdxI =
481 BaseIdx.castAs<nonloc::ConcreteInt>().getValue();
482
483 // Only allow non-integer offsets if the base region has no offset itself.
484 // FIXME: This is a somewhat arbitrary restriction. We should be using
485 // SValBuilder here to add the two offsets without checking their types.
486 if (!Offset.getAs<nonloc::ConcreteInt>()) {
487 if (isa<ElementRegion>(BaseRegion->StripCasts()))
488 return UnknownVal();
489
490 return loc::MemRegionVal(MRMgr.getElementRegion(
491 elementType, Offset, cast<SubRegion>(ElemR->getSuperRegion()), Ctx));
492 }
493
494 const llvm::APSInt& OffI = Offset.castAs<nonloc::ConcreteInt>().getValue();
495 assert(BaseIdxI.isSigned());
496
497 // Compute the new index.
498 nonloc::ConcreteInt NewIdx(svalBuilder.getBasicValueFactory().getValue(BaseIdxI +
499 OffI));
500
501 // Construct the new ElementRegion.
502 const SubRegion *ArrayR = cast<SubRegion>(ElemR->getSuperRegion());
503 return loc::MemRegionVal(MRMgr.getElementRegion(elementType, NewIdx, ArrayR,
504 Ctx));
505 }
506
507 StoreManager::BindingsHandler::~BindingsHandler() = default;
508
HandleBinding(StoreManager & SMgr,Store store,const MemRegion * R,SVal val)509 bool StoreManager::FindUniqueBinding::HandleBinding(StoreManager& SMgr,
510 Store store,
511 const MemRegion* R,
512 SVal val) {
513 SymbolRef SymV = val.getAsLocSymbol();
514 if (!SymV || SymV != Sym)
515 return true;
516
517 if (Binding) {
518 First = false;
519 return false;
520 }
521 else
522 Binding = R;
523
524 return true;
525 }
526