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/SmallVector.h"
33 #include "llvm/Support/Casting.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include <cassert>
36 #include <cstdint>
37 #include <optional>
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 std::optional<const MemRegion *> StoreManager::castRegion(const MemRegion *R,
75 QualType CastToTy) {
76 ASTContext &Ctx = StateMgr.getContext();
77
78 // Handle casts to Objective-C objects.
79 if (CastToTy->isObjCObjectPointerType())
80 return R->StripCasts();
81
82 if (CastToTy->isBlockPointerType()) {
83 // FIXME: We may need different solutions, depending on the symbol
84 // involved. Blocks can be casted to/from 'id', as they can be treated
85 // as Objective-C objects. This could possibly be handled by enhancing
86 // our reasoning of downcasts of symbolic objects.
87 if (isa<CodeTextRegion, SymbolicRegion>(R))
88 return R;
89
90 // We don't know what to make of it. Return a NULL region, which
91 // will be interpreted as UnknownVal.
92 return std::nullopt;
93 }
94
95 // Now assume we are casting from pointer to pointer. Other cases should
96 // already be handled.
97 QualType PointeeTy = CastToTy->getPointeeType();
98 QualType CanonPointeeTy = Ctx.getCanonicalType(PointeeTy);
99 CanonPointeeTy = CanonPointeeTy.getLocalUnqualifiedType();
100
101 // Handle casts to void*. We just pass the region through.
102 if (CanonPointeeTy == Ctx.VoidTy)
103 return R;
104
105 const auto IsSameRegionType = [&Ctx](const MemRegion *R, QualType OtherTy) {
106 if (const auto *TR = dyn_cast<TypedValueRegion>(R)) {
107 QualType ObjTy = Ctx.getCanonicalType(TR->getValueType());
108 if (OtherTy == ObjTy.getLocalUnqualifiedType())
109 return true;
110 }
111 return false;
112 };
113
114 // Handle casts from compatible types.
115 if (R->isBoundable() && IsSameRegionType(R, CanonPointeeTy))
116 return R;
117
118 // Process region cast according to the kind of the region being cast.
119 switch (R->getKind()) {
120 case MemRegion::CXXThisRegionKind:
121 case MemRegion::CodeSpaceRegionKind:
122 case MemRegion::StackLocalsSpaceRegionKind:
123 case MemRegion::StackArgumentsSpaceRegionKind:
124 case MemRegion::HeapSpaceRegionKind:
125 case MemRegion::UnknownSpaceRegionKind:
126 case MemRegion::StaticGlobalSpaceRegionKind:
127 case MemRegion::GlobalInternalSpaceRegionKind:
128 case MemRegion::GlobalSystemSpaceRegionKind:
129 case MemRegion::GlobalImmutableSpaceRegionKind: {
130 llvm_unreachable("Invalid region cast");
131 }
132
133 case MemRegion::FunctionCodeRegionKind:
134 case MemRegion::BlockCodeRegionKind:
135 case MemRegion::BlockDataRegionKind:
136 case MemRegion::StringRegionKind:
137 // FIXME: Need to handle arbitrary downcasts.
138 case MemRegion::SymbolicRegionKind:
139 case MemRegion::AllocaRegionKind:
140 case MemRegion::CompoundLiteralRegionKind:
141 case MemRegion::FieldRegionKind:
142 case MemRegion::ObjCIvarRegionKind:
143 case MemRegion::ObjCStringRegionKind:
144 case MemRegion::NonParamVarRegionKind:
145 case MemRegion::ParamVarRegionKind:
146 case MemRegion::CXXTempObjectRegionKind:
147 case MemRegion::CXXLifetimeExtendedObjectRegionKind:
148 case MemRegion::CXXBaseObjectRegionKind:
149 case MemRegion::CXXDerivedObjectRegionKind:
150 return MakeElementRegion(cast<SubRegion>(R), PointeeTy);
151
152 case MemRegion::ElementRegionKind: {
153 // If we are casting from an ElementRegion to another type, the
154 // algorithm is as follows:
155 //
156 // (1) Compute the "raw offset" of the ElementRegion from the
157 // base region. This is done by calling 'getAsRawOffset()'.
158 //
159 // (2a) If we get a 'RegionRawOffset' after calling
160 // 'getAsRawOffset()', determine if the absolute offset
161 // can be exactly divided into chunks of the size of the
162 // casted-pointee type. If so, create a new ElementRegion with
163 // the pointee-cast type as the new ElementType and the index
164 // being the offset divded by the chunk size. If not, create
165 // a new ElementRegion at offset 0 off the raw offset region.
166 //
167 // (2b) If we don't a get a 'RegionRawOffset' after calling
168 // 'getAsRawOffset()', it means that we are at offset 0.
169 //
170 // FIXME: Handle symbolic raw offsets.
171
172 const ElementRegion *elementR = cast<ElementRegion>(R);
173 const RegionRawOffset &rawOff = elementR->getAsArrayOffset();
174 const MemRegion *baseR = rawOff.getRegion();
175
176 // If we cannot compute a raw offset, throw up our hands and return
177 // a NULL MemRegion*.
178 if (!baseR)
179 return std::nullopt;
180
181 CharUnits off = rawOff.getOffset();
182
183 if (off.isZero()) {
184 // Edge case: we are at 0 bytes off the beginning of baseR. We check to
185 // see if the type we are casting to is the same as the type of the base
186 // region. If so, just return the base region.
187 if (IsSameRegionType(baseR, CanonPointeeTy))
188 return baseR;
189 // Otherwise, create a new ElementRegion at offset 0.
190 return MakeElementRegion(cast<SubRegion>(baseR), PointeeTy);
191 }
192
193 // We have a non-zero offset from the base region. We want to determine
194 // if the offset can be evenly divided by sizeof(PointeeTy). If so,
195 // we create an ElementRegion whose index is that value. Otherwise, we
196 // create two ElementRegions, one that reflects a raw offset and the other
197 // that reflects the cast.
198
199 // Compute the index for the new ElementRegion.
200 int64_t newIndex = 0;
201 const MemRegion *newSuperR = nullptr;
202
203 // We can only compute sizeof(PointeeTy) if it is a complete type.
204 if (!PointeeTy->isIncompleteType()) {
205 // Compute the size in **bytes**.
206 CharUnits pointeeTySize = Ctx.getTypeSizeInChars(PointeeTy);
207 if (!pointeeTySize.isZero()) {
208 // Is the offset a multiple of the size? If so, we can layer the
209 // ElementRegion (with elementType == PointeeTy) directly on top of
210 // the base region.
211 if (off % pointeeTySize == 0) {
212 newIndex = off / pointeeTySize;
213 newSuperR = baseR;
214 }
215 }
216 }
217
218 if (!newSuperR) {
219 // Create an intermediate ElementRegion to represent the raw byte.
220 // This will be the super region of the final ElementRegion.
221 newSuperR = MakeElementRegion(cast<SubRegion>(baseR), Ctx.CharTy,
222 off.getQuantity());
223 }
224
225 return MakeElementRegion(cast<SubRegion>(newSuperR), PointeeTy, newIndex);
226 }
227 }
228
229 llvm_unreachable("unreachable");
230 }
231
regionMatchesCXXRecordType(SVal V,QualType Ty)232 static bool regionMatchesCXXRecordType(SVal V, QualType Ty) {
233 const MemRegion *MR = V.getAsRegion();
234 if (!MR)
235 return true;
236
237 const auto *TVR = dyn_cast<TypedValueRegion>(MR);
238 if (!TVR)
239 return true;
240
241 const CXXRecordDecl *RD = TVR->getValueType()->getAsCXXRecordDecl();
242 if (!RD)
243 return true;
244
245 const CXXRecordDecl *Expected = Ty->getPointeeCXXRecordDecl();
246 if (!Expected)
247 Expected = Ty->getAsCXXRecordDecl();
248
249 return Expected->getCanonicalDecl() == RD->getCanonicalDecl();
250 }
251
evalDerivedToBase(SVal Derived,const CastExpr * Cast)252 SVal StoreManager::evalDerivedToBase(SVal Derived, const CastExpr *Cast) {
253 // Early return to avoid doing the wrong thing in the face of
254 // reinterpret_cast.
255 if (!regionMatchesCXXRecordType(Derived, Cast->getSubExpr()->getType()))
256 return UnknownVal();
257
258 // Walk through the cast path to create nested CXXBaseRegions.
259 SVal Result = Derived;
260 for (const CXXBaseSpecifier *Base : Cast->path()) {
261 Result = evalDerivedToBase(Result, Base->getType(), Base->isVirtual());
262 }
263 return Result;
264 }
265
evalDerivedToBase(SVal Derived,const CXXBasePath & Path)266 SVal StoreManager::evalDerivedToBase(SVal Derived, const CXXBasePath &Path) {
267 // Walk through the path to create nested CXXBaseRegions.
268 SVal Result = Derived;
269 for (const auto &I : Path)
270 Result = evalDerivedToBase(Result, I.Base->getType(),
271 I.Base->isVirtual());
272 return Result;
273 }
274
evalDerivedToBase(SVal Derived,QualType BaseType,bool IsVirtual)275 SVal StoreManager::evalDerivedToBase(SVal Derived, QualType BaseType,
276 bool IsVirtual) {
277 const MemRegion *DerivedReg = Derived.getAsRegion();
278 if (!DerivedReg)
279 return Derived;
280
281 const CXXRecordDecl *BaseDecl = BaseType->getPointeeCXXRecordDecl();
282 if (!BaseDecl)
283 BaseDecl = BaseType->getAsCXXRecordDecl();
284 assert(BaseDecl && "not a C++ object?");
285
286 if (const auto *AlreadyDerivedReg =
287 dyn_cast<CXXDerivedObjectRegion>(DerivedReg)) {
288 if (const auto *SR =
289 dyn_cast<SymbolicRegion>(AlreadyDerivedReg->getSuperRegion()))
290 if (SR->getSymbol()->getType()->getPointeeCXXRecordDecl() == BaseDecl)
291 return loc::MemRegionVal(SR);
292
293 DerivedReg = AlreadyDerivedReg->getSuperRegion();
294 }
295
296 const MemRegion *BaseReg = MRMgr.getCXXBaseObjectRegion(
297 BaseDecl, cast<SubRegion>(DerivedReg), IsVirtual);
298
299 return loc::MemRegionVal(BaseReg);
300 }
301
302 /// Returns the static type of the given region, if it represents a C++ class
303 /// object.
304 ///
305 /// This handles both fully-typed regions, where the dynamic type is known, and
306 /// symbolic regions, where the dynamic type is merely bounded (and even then,
307 /// only ostensibly!), but does not take advantage of any dynamic type info.
getCXXRecordType(const MemRegion * MR)308 static const CXXRecordDecl *getCXXRecordType(const MemRegion *MR) {
309 if (const auto *TVR = dyn_cast<TypedValueRegion>(MR))
310 return TVR->getValueType()->getAsCXXRecordDecl();
311 if (const auto *SR = dyn_cast<SymbolicRegion>(MR))
312 return SR->getSymbol()->getType()->getPointeeCXXRecordDecl();
313 return nullptr;
314 }
315
evalBaseToDerived(SVal Base,QualType TargetType)316 std::optional<SVal> StoreManager::evalBaseToDerived(SVal Base,
317 QualType TargetType) {
318 const MemRegion *MR = Base.getAsRegion();
319 if (!MR)
320 return UnknownVal();
321
322 // Assume the derived class is a pointer or a reference to a CXX record.
323 TargetType = TargetType->getPointeeType();
324 assert(!TargetType.isNull());
325 const CXXRecordDecl *TargetClass = TargetType->getAsCXXRecordDecl();
326 if (!TargetClass && !TargetType->isVoidType())
327 return UnknownVal();
328
329 // Drill down the CXXBaseObject chains, which represent upcasts (casts from
330 // derived to base).
331 while (const CXXRecordDecl *MRClass = getCXXRecordType(MR)) {
332 // If found the derived class, the cast succeeds.
333 if (MRClass == TargetClass)
334 return loc::MemRegionVal(MR);
335
336 // We skip over incomplete types. They must be the result of an earlier
337 // reinterpret_cast, as one can only dynamic_cast between types in the same
338 // class hierarchy.
339 if (!TargetType->isVoidType() && MRClass->hasDefinition()) {
340 // Static upcasts are marked as DerivedToBase casts by Sema, so this will
341 // only happen when multiple or virtual inheritance is involved.
342 CXXBasePaths Paths(/*FindAmbiguities=*/false, /*RecordPaths=*/true,
343 /*DetectVirtual=*/false);
344 if (MRClass->isDerivedFrom(TargetClass, Paths))
345 return evalDerivedToBase(loc::MemRegionVal(MR), Paths.front());
346 }
347
348 if (const auto *BaseR = dyn_cast<CXXBaseObjectRegion>(MR)) {
349 // Drill down the chain to get the derived classes.
350 MR = BaseR->getSuperRegion();
351 continue;
352 }
353
354 // If this is a cast to void*, return the region.
355 if (TargetType->isVoidType())
356 return loc::MemRegionVal(MR);
357
358 // Strange use of reinterpret_cast can give us paths we don't reason
359 // about well, by putting in ElementRegions where we'd expect
360 // CXXBaseObjectRegions. If it's a valid reinterpret_cast (i.e. if the
361 // derived class has a zero offset from the base class), then it's safe
362 // to strip the cast; if it's invalid, -Wreinterpret-base-class should
363 // catch it. In the interest of performance, the analyzer will silently
364 // do the wrong thing in the invalid case (because offsets for subregions
365 // will be wrong).
366 const MemRegion *Uncasted = MR->StripCasts(/*IncludeBaseCasts=*/false);
367 if (Uncasted == MR) {
368 // We reached the bottom of the hierarchy and did not find the derived
369 // class. We must be casting the base to derived, so the cast should
370 // fail.
371 break;
372 }
373
374 MR = Uncasted;
375 }
376
377 // If we're casting a symbolic base pointer to a derived class, use
378 // CXXDerivedObjectRegion to represent the cast. If it's a pointer to an
379 // unrelated type, it must be a weird reinterpret_cast and we have to
380 // be fine with ElementRegion. TODO: Should we instead make
381 // Derived{TargetClass, Element{SourceClass, SR}}?
382 if (const auto *SR = dyn_cast<SymbolicRegion>(MR)) {
383 QualType T = SR->getSymbol()->getType();
384 const CXXRecordDecl *SourceClass = T->getPointeeCXXRecordDecl();
385 if (TargetClass && SourceClass && TargetClass->isDerivedFrom(SourceClass))
386 return loc::MemRegionVal(
387 MRMgr.getCXXDerivedObjectRegion(TargetClass, SR));
388 return loc::MemRegionVal(GetElementZeroRegion(SR, TargetType));
389 }
390
391 // We failed if the region we ended up with has perfect type info.
392 if (isa<TypedValueRegion>(MR))
393 return std::nullopt;
394
395 return UnknownVal();
396 }
397
getLValueFieldOrIvar(const Decl * D,SVal Base)398 SVal StoreManager::getLValueFieldOrIvar(const Decl *D, SVal Base) {
399 if (Base.isUnknownOrUndef())
400 return Base;
401
402 Loc BaseL = Base.castAs<Loc>();
403 const SubRegion* BaseR = nullptr;
404
405 switch (BaseL.getKind()) {
406 case loc::MemRegionValKind:
407 BaseR = cast<SubRegion>(BaseL.castAs<loc::MemRegionVal>().getRegion());
408 break;
409
410 case loc::GotoLabelKind:
411 // These are anormal cases. Flag an undefined value.
412 return UndefinedVal();
413
414 case loc::ConcreteIntKind:
415 // While these seem funny, this can happen through casts.
416 // FIXME: What we should return is the field offset, not base. For example,
417 // add the field offset to the integer value. That way things
418 // like this work properly: &(((struct foo *) 0xa)->f)
419 // However, that's not easy to fix without reducing our abilities
420 // to catch null pointer dereference. Eg., ((struct foo *)0x0)->f = 7
421 // is a null dereference even though we're dereferencing offset of f
422 // rather than null. Coming up with an approach that computes offsets
423 // over null pointers properly while still being able to catch null
424 // dereferences might be worth it.
425 return Base;
426
427 default:
428 llvm_unreachable("Unhandled Base.");
429 }
430
431 // NOTE: We must have this check first because ObjCIvarDecl is a subclass
432 // of FieldDecl.
433 if (const auto *ID = dyn_cast<ObjCIvarDecl>(D))
434 return loc::MemRegionVal(MRMgr.getObjCIvarRegion(ID, BaseR));
435
436 return loc::MemRegionVal(MRMgr.getFieldRegion(cast<FieldDecl>(D), BaseR));
437 }
438
getLValueIvar(const ObjCIvarDecl * decl,SVal base)439 SVal StoreManager::getLValueIvar(const ObjCIvarDecl *decl, SVal base) {
440 return getLValueFieldOrIvar(decl, base);
441 }
442
getLValueElement(QualType elementType,NonLoc Offset,SVal Base)443 SVal StoreManager::getLValueElement(QualType elementType, NonLoc Offset,
444 SVal Base) {
445
446 // Special case, if index is 0, return the same type as if
447 // this was not an array dereference.
448 if (Offset.isZeroConstant()) {
449 QualType BT = Base.getType(this->Ctx);
450 if (!BT.isNull() && !elementType.isNull()) {
451 QualType PointeeTy = BT->getPointeeType();
452 if (!PointeeTy.isNull() &&
453 PointeeTy.getCanonicalType() == elementType.getCanonicalType())
454 return Base;
455 }
456 }
457
458 // If the base is an unknown or undefined value, just return it back.
459 // FIXME: For absolute pointer addresses, we just return that value back as
460 // well, although in reality we should return the offset added to that
461 // value. See also the similar FIXME in getLValueFieldOrIvar().
462 if (Base.isUnknownOrUndef() || isa<loc::ConcreteInt>(Base))
463 return Base;
464
465 if (isa<loc::GotoLabel>(Base))
466 return UnknownVal();
467
468 const SubRegion *BaseRegion =
469 Base.castAs<loc::MemRegionVal>().getRegionAs<SubRegion>();
470
471 // Pointer of any type can be cast and used as array base.
472 const auto *ElemR = dyn_cast<ElementRegion>(BaseRegion);
473
474 // Convert the offset to the appropriate size and signedness.
475 auto Off = svalBuilder.convertToArrayIndex(Offset).getAs<NonLoc>();
476 if (!Off) {
477 // Handle cases when LazyCompoundVal is used for an array index.
478 // Such case is possible if code does:
479 // char b[4];
480 // a[__builtin_bitcast(int, b)];
481 // Return UnknownVal, since we cannot model it.
482 return UnknownVal();
483 }
484
485 Offset = Off.value();
486
487 if (!ElemR) {
488 // If the base region is not an ElementRegion, create one.
489 // This can happen in the following example:
490 //
491 // char *p = __builtin_alloc(10);
492 // p[1] = 8;
493 //
494 // Observe that 'p' binds to an AllocaRegion.
495 return loc::MemRegionVal(MRMgr.getElementRegion(elementType, Offset,
496 BaseRegion, Ctx));
497 }
498
499 SVal BaseIdx = ElemR->getIndex();
500
501 if (!isa<nonloc::ConcreteInt>(BaseIdx))
502 return UnknownVal();
503
504 const llvm::APSInt &BaseIdxI =
505 BaseIdx.castAs<nonloc::ConcreteInt>().getValue();
506
507 // Only allow non-integer offsets if the base region has no offset itself.
508 // FIXME: This is a somewhat arbitrary restriction. We should be using
509 // SValBuilder here to add the two offsets without checking their types.
510 if (!isa<nonloc::ConcreteInt>(Offset)) {
511 if (isa<ElementRegion>(BaseRegion->StripCasts()))
512 return UnknownVal();
513
514 return loc::MemRegionVal(MRMgr.getElementRegion(
515 elementType, Offset, cast<SubRegion>(ElemR->getSuperRegion()), Ctx));
516 }
517
518 const llvm::APSInt& OffI = Offset.castAs<nonloc::ConcreteInt>().getValue();
519 assert(BaseIdxI.isSigned());
520
521 // Compute the new index.
522 nonloc::ConcreteInt NewIdx(svalBuilder.getBasicValueFactory().getValue(BaseIdxI +
523 OffI));
524
525 // Construct the new ElementRegion.
526 const SubRegion *ArrayR = cast<SubRegion>(ElemR->getSuperRegion());
527 return loc::MemRegionVal(MRMgr.getElementRegion(elementType, NewIdx, ArrayR,
528 Ctx));
529 }
530
531 StoreManager::BindingsHandler::~BindingsHandler() = default;
532
HandleBinding(StoreManager & SMgr,Store store,const MemRegion * R,SVal val)533 bool StoreManager::FindUniqueBinding::HandleBinding(StoreManager& SMgr,
534 Store store,
535 const MemRegion* R,
536 SVal val) {
537 SymbolRef SymV = val.getAsLocSymbol();
538 if (!SymV || SymV != Sym)
539 return true;
540
541 if (Binding) {
542 First = false;
543 return false;
544 }
545 else
546 Binding = R;
547
548 return true;
549 }
550