Ted Kremenek, kremenek at apple
Zhongxing Xu, xuzhongzhing at gmail
The path-sensitive analysis engine in libAnalysis employs an extensible API for abstractly modeling the memory of an analyzed program. This API employs the concept of "memory regions" to abstractly model chunks of program memory such as program variables and dynamically allocated memory such as those returned from 'malloc' and 'alloca'. Regions are hierarchical, with subregions modeling subtyping relationships, field and array offsets into larger chunks of memory, and so on.
The region API consists of two components:
Symbolic stores, which can be thought of as representing the relation regions -> values, are implemented by subclasses of the StoreManager class (Store.h). A particular StoreManager implementation has complete flexibility concerning the following:
Together, both points allow different StoreManagers to tradeoff between different levels of analysis precision and scalability concerning the reasoning of program memory. Meanwhile, the core path-sensitive engine makes no assumptions about either points, and queries a StoreManager about the bindings to a memory region through a generic interface that all StoreManagers share. If a particular StoreManager cannot reason about the potential bindings of a given memory region (e.g., 'BasicStoreManager' does not reason about fields of structures) then the StoreManager can simply return 'unknown' (represented by 'UnknownVal') for a particular region-binding. This separation of concerns not only isolates the core analysis engine from the details of reasoning about program memory but also facilities the option of a client of the path-sensitive engine to easily swap in different StoreManager implementations that internally reason about program memory in very different ways.
The rest of this document is divided into two parts. We first discuss region taxonomy and the semantics of regions. We then discuss the StoreManager interface, and details of how the currently available StoreManager classes implement region bindings.
Before talking about the memory regions, we would talk about the pointers since memory regions are essentially used to represent pointer values.
The pointer is a type of values. Pointer values have two semantic aspects. One is its physical value, which is an address or location. The other is the type of the memory object residing in the address.
Memory regions are designed to abstract these two properties of the pointer. The physical value of a pointer is represented by MemRegion pointers. The rvalue type of the region corresponds to the type of the pointee object.
One complication is that we could have different view regions on the same memory chunk. They represent the same memory location, but have different abstract location, i.e., MemRegion pointers. Thus we need to canonicalize the abstract locations to get a unique abstract location for one physical location.
Furthermore, these different view regions may or may not represent memory objects of different types. Some different types are semantically the same, for example, 'struct s' and 'my_type' are the same type.
struct s; typedef struct s my_type;
But char and int are not the same type in the code below:
void *p; int *q = (int*) p; char *r = (char*) p;
Thus we need to canonicalize the MemRegion which is used in binding and retrieving.
Region is the entity used to model pointer values. A Region has the following properties:
A symbolic region is a map of the concept of symbolic values into the domain of regions. It is the way that we represent symbolic pointers. Whenever a symbolic pointer value is needed, a symbolic region is created to represent it.
A symbolic region has no type. It wraps a SymbolData. But sometimes we have type information associated with a symbolic region. For this case, a TypedViewRegion is created to layer the type information on top of the symbolic region. The reason we do not carry type information with the symbolic region is that the symbolic regions can have no type. To be consistent, we don't let them to carry type information.
Like a symbolic pointer, a symbolic region may be NULL, has unknown extent, and represents a generic chunk of memory.
NOTE: We plan not to use loc::SymbolVal in RegionStore and remove it gradually.
Symbolic regions get their rvalue types through the following ways:
void f(struct s* p) {
...
}
The symbolic region pointed to by p has type struct s.
void f(void* p) {
struct s* q = (struct s*) p;
...
}
We attach the type information to the symbolic region lazily. For the first case above, we create the TypedViewRegion only when the pointer is actually used to access the pointee memory object, that is when the element or field region is created. For the cast case, the TypedViewRegion is created when visiting the CastExpr.
The reason for doing lazy typing is that symbolic regions are sometimes only used to do location comparison.
Pointer casts allow people to impose different 'views' onto a chunk of memory.
Usually we have two kinds of casts. One kind of casts cast down with in the type hierarchy. It imposes more specific views onto more generic memory regions. The other kind of casts cast up with in the type hierarchy. It strips away more specific views on top of the more generic memory regions.
We simulate the down casts by layering another TypedViewRegion on top of the original region. We simulate the up casts by striping away the top TypedViewRegion. Down casts is usually simple. For up casts, if the there is no TypedViewRegion to be stripped, we return the original region. If the underlying region is of the different type than the cast-to type, we flag an error state.
For toll-free bridging casts, we return the original region.
We can set up a partial order for pointer types, with the most general type void* at the top. The partial order forms a tree with void* as its root node.
Every MemRegion has a root position in the type tree. For example, the pointee region of void *p has its root position at the root node of the tree. VarRegion of int x has its root position at the 'int type' node.
TypedViewRegion is used to move the region down or up in the tree. Moving down in the tree adds a TypedViewRegion. Moving up in the tree removes a TypedViewRegion.
Do we want to allow moving up beyond the root position? This happens when:
int x; void *p = &x;
The region of x has its root position at 'int*' node. the cast to void* moves that region up to the 'void*' node. I propose to not allow such casts, and assign the region of x for p.
Another non-ideal case is that people might cast to a non-generic pointer from another non-generic pointer instead of first casting it back to the generic pointer. Direct handling of this case would result in multiple layers of TypedViewRegions. This enforces an incorrect semantic view to the region, because we can only have one typed view on a region at a time. To avoid this inconsistency, before casting the region, we strip the TypedViewRegion, then do the cast. In summary, we only allow one layer of TypedViewRegion.
The following region kinds are boundable: VarRegion, CompoundLiteralRegion, StringRegion, ElementRegion, FieldRegion, and ObjCIvarRegion.
When binding regions, we perform canonicalization on element regions and field regions. This is because we can have different views on the same region, some of which are essentially the same view with different sugar type names.
To canonicalize a region, we get the canonical types for all TypedViewRegions along the way up to the root region, and make new TypedViewRegions with those canonical types.
For Objective-C and C++, perhaps another canonicalization rule should be added: for FieldRegion, the least derived class that has the field is used as the type of the super region of the FieldRegion.
All bindings and retrievings are done on the canonicalized regions.
Canonicalization is transparent outside the region store manager, and more specifically, unaware outside the Bind() and Retrieve() method. We don't need to consider region canonicalization when doing pointer cast.
The constraint manager reasons about the abstract location of memory objects. We can have different views on a region, but none of these views changes the location of that object. Thus we should get the same abstract location for those regions.