1 //===- Store.cpp - Interface for maps from Locations to Values ------------===//
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
7 //===----------------------------------------------------------------------===//
9 // This file defined the types Store and StoreManager.
11 //===----------------------------------------------------------------------===//
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"
39 using namespace clang;
42 StoreManager::StoreManager(ProgramStateManager &stateMgr)
43 : svalBuilder(stateMgr.getSValBuilder()), StateMgr(stateMgr),
44 MRMgr(svalBuilder.getRegionManager()), Ctx(stateMgr.getContext()) {}
46 StoreRef StoreManager::enterStackFrame(Store OldStore,
47 const CallEvent &Call,
48 const StackFrameContext *LCtx) {
49 StoreRef Store = StoreRef(OldStore, *this);
51 SmallVector<CallEvent::FrameBindingTy, 16> InitialBindings;
52 Call.getInitialStackFrameContents(LCtx, InitialBindings);
54 for (const auto &I : InitialBindings)
55 Store = Bind(Store.getStore(), I.first.castAs<Loc>(), I.second);
60 const ElementRegion *StoreManager::MakeElementRegion(const SubRegion *Base,
63 NonLoc idx = svalBuilder.makeArrayIndex(index);
64 return MRMgr.getElementRegion(EleTy, idx, Base, svalBuilder.getContext());
67 const ElementRegion *StoreManager::GetElementZeroRegion(const SubRegion *R,
69 NonLoc idx = svalBuilder.makeZeroArrayIndex();
71 return MRMgr.getElementRegion(T, idx, R, Ctx);
74 const MemRegion *StoreManager::castRegion(const MemRegion *R, QualType CastToTy) {
75 ASTContext &Ctx = StateMgr.getContext();
77 // Handle casts to Objective-C objects.
78 if (CastToTy->isObjCObjectPointerType())
79 return R->StripCasts();
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))
89 // We don't know what to make of it. Return a NULL region, which
90 // will be interpreted as UnknownVal.
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);
99 // Handle casts to void*. We just pass the region through.
100 if (CanonPointeeTy.getLocalUnqualifiedType() == Ctx.VoidTy)
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)
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");
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::VarRegionKind:
138 case MemRegion::CXXTempObjectRegionKind:
139 case MemRegion::CXXBaseObjectRegionKind:
140 case MemRegion::CXXDerivedObjectRegionKind:
141 return MakeElementRegion(cast<SubRegion>(R), PointeeTy);
143 case MemRegion::ElementRegionKind: {
144 // If we are casting from an ElementRegion to another type, the
145 // algorithm is as follows:
147 // (1) Compute the "raw offset" of the ElementRegion from the
148 // base region. This is done by calling 'getAsRawOffset()'.
150 // (2a) If we get a 'RegionRawOffset' after calling
151 // 'getAsRawOffset()', determine if the absolute offset
152 // can be exactly divided into chunks of the size of the
153 // casted-pointee type. If so, create a new ElementRegion with
154 // the pointee-cast type as the new ElementType and the index
155 // being the offset divded by the chunk size. If not, create
156 // a new ElementRegion at offset 0 off the raw offset region.
158 // (2b) If we don't a get a 'RegionRawOffset' after calling
159 // 'getAsRawOffset()', it means that we are at offset 0.
161 // FIXME: Handle symbolic raw offsets.
163 const ElementRegion *elementR = cast<ElementRegion>(R);
164 const RegionRawOffset &rawOff = elementR->getAsArrayOffset();
165 const MemRegion *baseR = rawOff.getRegion();
167 // If we cannot compute a raw offset, throw up our hands and return
168 // a NULL MemRegion*.
172 CharUnits off = rawOff.getOffset();
175 // Edge case: we are at 0 bytes off the beginning of baseR. We
176 // check to see if type we are casting to is the same as the base
177 // region. If so, just return the base region.
178 if (const auto *TR = dyn_cast<TypedValueRegion>(baseR)) {
179 QualType ObjTy = Ctx.getCanonicalType(TR->getValueType());
180 QualType CanonPointeeTy = Ctx.getCanonicalType(PointeeTy);
181 if (CanonPointeeTy == ObjTy)
185 // Otherwise, create a new ElementRegion at offset 0.
186 return MakeElementRegion(cast<SubRegion>(baseR), PointeeTy);
189 // We have a non-zero offset from the base region. We want to determine
190 // if the offset can be evenly divided by sizeof(PointeeTy). If so,
191 // we create an ElementRegion whose index is that value. Otherwise, we
192 // create two ElementRegions, one that reflects a raw offset and the other
193 // that reflects the cast.
195 // Compute the index for the new ElementRegion.
196 int64_t newIndex = 0;
197 const MemRegion *newSuperR = nullptr;
199 // We can only compute sizeof(PointeeTy) if it is a complete type.
200 if (!PointeeTy->isIncompleteType()) {
201 // Compute the size in **bytes**.
202 CharUnits pointeeTySize = Ctx.getTypeSizeInChars(PointeeTy);
203 if (!pointeeTySize.isZero()) {
204 // Is the offset a multiple of the size? If so, we can layer the
205 // ElementRegion (with elementType == PointeeTy) directly on top of
207 if (off % pointeeTySize == 0) {
208 newIndex = off / pointeeTySize;
215 // Create an intermediate ElementRegion to represent the raw byte.
216 // This will be the super region of the final ElementRegion.
217 newSuperR = MakeElementRegion(cast<SubRegion>(baseR), Ctx.CharTy,
221 return MakeElementRegion(cast<SubRegion>(newSuperR), PointeeTy, newIndex);
225 llvm_unreachable("unreachable");
228 static bool regionMatchesCXXRecordType(SVal V, QualType Ty) {
229 const MemRegion *MR = V.getAsRegion();
233 const auto *TVR = dyn_cast<TypedValueRegion>(MR);
237 const CXXRecordDecl *RD = TVR->getValueType()->getAsCXXRecordDecl();
241 const CXXRecordDecl *Expected = Ty->getPointeeCXXRecordDecl();
243 Expected = Ty->getAsCXXRecordDecl();
245 return Expected->getCanonicalDecl() == RD->getCanonicalDecl();
248 SVal StoreManager::evalDerivedToBase(SVal Derived, const CastExpr *Cast) {
249 // Sanity check to avoid doing the wrong thing in the face of
251 if (!regionMatchesCXXRecordType(Derived, Cast->getSubExpr()->getType()))
254 // Walk through the cast path to create nested CXXBaseRegions.
255 SVal Result = Derived;
256 for (CastExpr::path_const_iterator I = Cast->path_begin(),
257 E = Cast->path_end();
259 Result = evalDerivedToBase(Result, (*I)->getType(), (*I)->isVirtual());
264 SVal StoreManager::evalDerivedToBase(SVal Derived, const CXXBasePath &Path) {
265 // Walk through the path to create nested CXXBaseRegions.
266 SVal Result = Derived;
267 for (const auto &I : Path)
268 Result = evalDerivedToBase(Result, I.Base->getType(),
269 I.Base->isVirtual());
273 SVal StoreManager::evalDerivedToBase(SVal Derived, QualType BaseType,
275 const MemRegion *DerivedReg = Derived.getAsRegion();
279 const CXXRecordDecl *BaseDecl = BaseType->getPointeeCXXRecordDecl();
281 BaseDecl = BaseType->getAsCXXRecordDecl();
282 assert(BaseDecl && "not a C++ object?");
284 if (const auto *AlreadyDerivedReg =
285 dyn_cast<CXXDerivedObjectRegion>(DerivedReg)) {
287 dyn_cast<SymbolicRegion>(AlreadyDerivedReg->getSuperRegion()))
288 if (SR->getSymbol()->getType()->getPointeeCXXRecordDecl() == BaseDecl)
289 return loc::MemRegionVal(SR);
291 DerivedReg = AlreadyDerivedReg->getSuperRegion();
294 const MemRegion *BaseReg = MRMgr.getCXXBaseObjectRegion(
295 BaseDecl, cast<SubRegion>(DerivedReg), IsVirtual);
297 return loc::MemRegionVal(BaseReg);
300 /// Returns the static type of the given region, if it represents a C++ class
303 /// This handles both fully-typed regions, where the dynamic type is known, and
304 /// symbolic regions, where the dynamic type is merely bounded (and even then,
305 /// only ostensibly!), but does not take advantage of any dynamic type info.
306 static const CXXRecordDecl *getCXXRecordType(const MemRegion *MR) {
307 if (const auto *TVR = dyn_cast<TypedValueRegion>(MR))
308 return TVR->getValueType()->getAsCXXRecordDecl();
309 if (const auto *SR = dyn_cast<SymbolicRegion>(MR))
310 return SR->getSymbol()->getType()->getPointeeCXXRecordDecl();
314 SVal StoreManager::attemptDownCast(SVal Base, QualType TargetType,
318 const MemRegion *MR = Base.getAsRegion();
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())
329 // Drill down the CXXBaseObject chains, which represent upcasts (casts from
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);
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
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());
348 if (const auto *BaseR = dyn_cast<CXXBaseObjectRegion>(MR)) {
349 // Drill down the chain to get the derived classes.
350 MR = BaseR->getSuperRegion();
354 // If this is a cast to void*, return the region.
355 if (TargetType->isVoidType())
356 return loc::MemRegionVal(MR);
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
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
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));
391 // We failed if the region we ended up with has perfect type info.
392 Failed = isa<TypedValueRegion>(MR);
396 static bool hasSameUnqualifiedPointeeType(QualType ty1, QualType ty2) {
397 return ty1->getPointeeType().getCanonicalType().getTypePtr() ==
398 ty2->getPointeeType().getCanonicalType().getTypePtr();
401 /// CastRetrievedVal - Used by subclasses of StoreManager to implement
402 /// implicit casts that arise from loads from regions that are reinterpreted
403 /// as another region.
404 SVal StoreManager::CastRetrievedVal(SVal V, const TypedValueRegion *R,
406 if (castTy.isNull() || V.isUnknownOrUndef())
409 // The dispatchCast() call below would convert the int into a float.
410 // What we want, however, is a bit-by-bit reinterpretation of the int
411 // as a float, which usually yields nothing garbage. For now skip casts
412 // from ints to floats.
413 // TODO: What other combinations of types are affected?
414 if (castTy->isFloatingType()) {
415 SymbolRef Sym = V.getAsSymbol();
416 if (Sym && !Sym->getType()->isFloatingType())
420 // When retrieving symbolic pointer and expecting a non-void pointer,
421 // wrap them into element regions of the expected type if necessary.
422 // SValBuilder::dispatchCast() doesn't do that, but it is necessary to
423 // make sure that the retrieved value makes sense, because there's no other
424 // cast in the AST that would tell us to cast it to the correct pointer type.
425 // We might need to do that for non-void pointers as well.
426 // FIXME: We really need a single good function to perform casts for us
427 // correctly every time we need it.
428 if (castTy->isPointerType() && !castTy->isVoidPointerType())
429 if (const auto *SR = dyn_cast_or_null<SymbolicRegion>(V.getAsRegion())) {
430 QualType sr = SR->getSymbol()->getType();
431 if (!hasSameUnqualifiedPointeeType(sr, castTy))
432 return loc::MemRegionVal(castRegion(SR, castTy));
435 return svalBuilder.dispatchCast(V, castTy);
438 SVal StoreManager::getLValueFieldOrIvar(const Decl *D, SVal Base) {
439 if (Base.isUnknownOrUndef())
442 Loc BaseL = Base.castAs<Loc>();
443 const SubRegion* BaseR = nullptr;
445 switch (BaseL.getSubKind()) {
446 case loc::MemRegionValKind:
447 BaseR = cast<SubRegion>(BaseL.castAs<loc::MemRegionVal>().getRegion());
450 case loc::GotoLabelKind:
451 // These are anormal cases. Flag an undefined value.
452 return UndefinedVal();
454 case loc::ConcreteIntKind:
455 // While these seem funny, this can happen through casts.
456 // FIXME: What we should return is the field offset, not base. For example,
457 // add the field offset to the integer value. That way things
458 // like this work properly: &(((struct foo *) 0xa)->f)
459 // However, that's not easy to fix without reducing our abilities
460 // to catch null pointer dereference. Eg., ((struct foo *)0x0)->f = 7
461 // is a null dereference even though we're dereferencing offset of f
462 // rather than null. Coming up with an approach that computes offsets
463 // over null pointers properly while still being able to catch null
464 // dereferences might be worth it.
468 llvm_unreachable("Unhandled Base.");
471 // NOTE: We must have this check first because ObjCIvarDecl is a subclass
473 if (const auto *ID = dyn_cast<ObjCIvarDecl>(D))
474 return loc::MemRegionVal(MRMgr.getObjCIvarRegion(ID, BaseR));
476 return loc::MemRegionVal(MRMgr.getFieldRegion(cast<FieldDecl>(D), BaseR));
479 SVal StoreManager::getLValueIvar(const ObjCIvarDecl *decl, SVal base) {
480 return getLValueFieldOrIvar(decl, base);
483 SVal StoreManager::getLValueElement(QualType elementType, NonLoc Offset,
485 // If the base is an unknown or undefined value, just return it back.
486 // FIXME: For absolute pointer addresses, we just return that value back as
487 // well, although in reality we should return the offset added to that
488 // value. See also the similar FIXME in getLValueFieldOrIvar().
489 if (Base.isUnknownOrUndef() || Base.getAs<loc::ConcreteInt>())
492 if (Base.getAs<loc::GotoLabel>())
495 const SubRegion *BaseRegion =
496 Base.castAs<loc::MemRegionVal>().getRegionAs<SubRegion>();
498 // Pointer of any type can be cast and used as array base.
499 const auto *ElemR = dyn_cast<ElementRegion>(BaseRegion);
501 // Convert the offset to the appropriate size and signedness.
502 Offset = svalBuilder.convertToArrayIndex(Offset).castAs<NonLoc>();
505 // If the base region is not an ElementRegion, create one.
506 // This can happen in the following example:
508 // char *p = __builtin_alloc(10);
511 // Observe that 'p' binds to an AllocaRegion.
512 return loc::MemRegionVal(MRMgr.getElementRegion(elementType, Offset,
516 SVal BaseIdx = ElemR->getIndex();
518 if (!BaseIdx.getAs<nonloc::ConcreteInt>())
521 const llvm::APSInt &BaseIdxI =
522 BaseIdx.castAs<nonloc::ConcreteInt>().getValue();
524 // Only allow non-integer offsets if the base region has no offset itself.
525 // FIXME: This is a somewhat arbitrary restriction. We should be using
526 // SValBuilder here to add the two offsets without checking their types.
527 if (!Offset.getAs<nonloc::ConcreteInt>()) {
528 if (isa<ElementRegion>(BaseRegion->StripCasts()))
531 return loc::MemRegionVal(MRMgr.getElementRegion(
532 elementType, Offset, cast<SubRegion>(ElemR->getSuperRegion()), Ctx));
535 const llvm::APSInt& OffI = Offset.castAs<nonloc::ConcreteInt>().getValue();
536 assert(BaseIdxI.isSigned());
538 // Compute the new index.
539 nonloc::ConcreteInt NewIdx(svalBuilder.getBasicValueFactory().getValue(BaseIdxI +
542 // Construct the new ElementRegion.
543 const SubRegion *ArrayR = cast<SubRegion>(ElemR->getSuperRegion());
544 return loc::MemRegionVal(MRMgr.getElementRegion(elementType, NewIdx, ArrayR,
548 StoreManager::BindingsHandler::~BindingsHandler() = default;
550 bool StoreManager::FindUniqueBinding::HandleBinding(StoreManager& SMgr,
554 SymbolRef SymV = val.getAsLocSymbol();
555 if (!SymV || SymV != Sym)