Archive
Using and Abusing Unions
The C union type is one of those features that is generally frowned on by those who set programming standards for critical systems, yet is quite often used. MISRA C 2004 rule 18.4 bans them (“unions shall not be used”) on the grounds that there is a risk that the data may be misinterpreted. However, it goes on to say that deviations are acceptable for packing and unpacking of data, and for implementing variant records provided that the variants are differentiated by a common field. Read more…
Using constrained types in C
When writing critical software, one of the advantages cited for using Ada rather than C is that Ada lets you define constrained types, like this:
type Percentage is Integer range 0 .. 100;
Reasoning about null-terminated strings in C/C++
In my last post I described how ArC supports reasoning about array access, by allowing you to refer to the bounds of an array in a specification. If the code itself needs to know the size of an array, then the size is provided by other means, for example by passing it as an extra parameter. However, when using arrays of characters, standard practice in C is not to pass the number of elements, but to use a null character to indicate the end. Read more…
The Taming of the Pointer – Part 2
In last Wednesday’s post I mentioned three ways in which pointers are troublesome in C/C++, and I introduced the null ArC keyword to mitigate one of them. Now I’ll turn to the second issue: the fact that given (say) a variable or parameter of type int*, the type does not allow us to determine whether it refers to a single int, or to an array of ints – nor, if it refers to an array, can we find how many elements the array contains. Read more…
Using strongly-typed Booleans in C and C++
One of the glaring omissions from the original C language was provision of a Boolean type. Booleans in C are represented by integers, with false being represented as zero and true being represented as 1. When an integer value is used as a condition, all nonzero values are intrepreted as true.
Strong typing is a valuable asset when writing code – whether critical or not – because type checks can and do uncover errors. So how can we use strongly-typed Booleans in C and C++? Read more…
Taming Pointers in C/C++
When doing verification or deep static analysis of C and C++, pointers are troublesome in several ways:
- Zero (i.e. NULL) is an allowed value of every pointer type in the language. Occasionally we want to allow null pointers, for example in the link field of the last element of a linked list. More usually, we don’t want to allow null. Verification requires that anywhere we use the * or [ ] operator on a pointer, we can be sure that it is not null.
- C and C++ do not distinguish between pointers to single variables and pointers to arrays. So, where we have a parameter or variable of type T*, we can’t tell whether it is supposed to point to a variable or an array. If it points to a single variable, then we mustn’t do pointer arithmetic or indexing on it. The verifier must be able to check this.
- Array parameters in C/C++ are passed as pointers. Aside from the problem that we can’t distinguish array pointers from pointers to single variables, we also have the problem that there is no size information contained in an array pointer.
- Anywhere we use pointers to mutable data, there is the possibility of aliasing. In other words, there may be more than one pointer to the same data. The verifier needs to take account of the fact that changes to data made through one pointer may affect the values subsequently read through another pointer.
Although pointers are less troublesome in Ada, the aliasing problem still exists. The SPARK verifiable subset of Ada handles this by banning pointers altogether. Unfortunately, this isn’t an option in a C/C++ subset for critical systems, because pointers are the only mechanism for passing parameters by reference.
I’ll deal with the issue of unwanted nullability first. One solution is to add invariants stating that particular variables of pointer type cannot be NULL. Similarly, where a function takes parameters of pointer type, we can write preconditions that these parameters cannot be NULL. Here are a couple of examples:
struct Status1 {
const char* message;
...
invariant(message != NULL)
}
void sum(int *arr, int size)
pre(arr != NULL)
{ ... }
The problem with this approach is that you need a lot of these invariants and preconditions because, more often than not, it makes no sense to allow NULL. So in ArC we take the opposite approach. We assume that pointers are not allowed to be NULL except where you say otherwise. In the above examples, you can leave out the precondition and invariant if you don’t want to allow either pointer to be NULL.
To tell ArC that a pointer is allowed to be NULL, you flag it with the null attribute, like this:
struct Status2 {
const char* null message;
...
}
void setMessage(const char * null msg) { ... }
This greatly reduces the amount of annotation needed, because the null annotation is more concise than a precondition or invariant, and it is needed less often. As you might expect, null is another macro defined in arc.h that expands to emptiness when you compile the code. Syntactically, it behaves like const or volatile.
That’s all for today – I’ll discuss how we handle the other problems with pointers later.
Invariants for C/C++ Classes and Structs
In yesterday’s post, I proposed the use of simple C++ classes in critical software. I pointed out that classes are better than C structs, because they offer encapsulation and make it easier to avoid using objects that are not completely initialized. Now I’m going to point out another advantage of classes over structs, which is that they make it easier to enforce invariants.
Consider the following C code:
typedef struct _Limits {
int minValue;
int maxValue; // must always be >= minValue
} Limits;
The comment is an example of an invariant, i.e. a condition on the values of the members that we always expect to be true. During testing, we might want to do runtime checks to report any violation of the invariant. We would also like to do static analysis to make sure it always holds.
The problem with enforcing this invariant is that minValue and maxValue are public. This means that any piece if code that uses a variable of type Limits can break the invariant by assigning a new valie to minValue or maxValue. If we want to check the invariant at runtime, we must add a runtime check everywhere that the code assigns a value to either of these fields. Likewise, a static analyser must consider whether the invariant is broken at every place where one of these fields is assigned.
Let’s look at how we would define the Limits type using a C++ class instead:
class Limits {
int _minValue;
int _maxValue; // must always be >= minValue
public:
int minValue() const { return _minValue; }
int maxValue() const { return _maxValue; }
Limits(int n, int x)
: _minValue(n), _maxValue(x) {}
}
I’ve made the data private, and I’ve added a couple of functions to allow the min and max values to be read, but not written (don’t worry about whether this is efficient – any reasonable C++ compiler will inline calls to these functions). I’ve also added a constructor so that we can create values of type Limits. Using this new declaration of Limits, the only way that anyone can break the invariant is by calling the constructor with n > x. So there is just one place where we need to insert a runtime check to catch every instance where this invariant might be broken.
Finally, let’s look at what you need to do to get ArC to verify statically that the invariant always holds:
#include "arc.h"
class Limits {
int _minValue;
int _maxValue;
invariant(_maxValue >= _minValue)
public:
int minValue() const { return _minValue; }
int maxValue() const { return _maxValue; }
Limits(int n, int x)
: _minValue(n), _maxValue(x) pre(x >= n) {}
}
Instead of expressing the invariant as a comment, we have expressed it using the invariant keyword. We #include “arc.h” at the start so that when you are compiling the file using a normal C++ compiler, invariant(…) is defined as a macro that expands to nothing. This makes the invariant invisible to the compiler. But when ArC sees the invariant, it know that it needs to prove that the invariant holds anywhere that we create or modify a value of type Limits.
Since the invariant only depends on private data, ArC only has to worry about breaking the invariant within the class’s own constructors and members. In order to prove that the Limits constructor satisfies the invariant, we need to ensure x >= n whenever it is called. That’s why I added the pre(x >= n) clause in the constructor. This clause tells ArC to assume x >= n when it verifies the constructor, and to verify x >= n whenever we call the constructor. pre is another ArC keyword – it stands for precondition.
Incidentally, although Microsoft’s Vcc doesn’t support any C++ (unlike ArC), it does allow you to declare invariants on structures. But when you want to initialize or modify such a structure, you’ll generally need to add some more annotations to “unwrap” and “wrap” it. That’s the price of not having encapsulation.