Initialisation of Pointer Variables

[gray]

In Oberon, as in many (most?) languages, variables are not initialised by the compiler. Eiffel would be a counter example. I am not sure if general automatic initialisation is a Good Thing, as I prefer if the program text is as explicit as possible.

However, there's one exception: pointer variables.

I propose considering to automatically initialise pointers to NIL. In my book, the comprehensive use of precondition checks is a Good Thing, both for procedure parameters as well as other assumptions as regards the validity of a procedure (eg. task state in a scheduler). If a procedure's correctness depends on specific values, or ranges thereof, of its actual parameters, it can easily be ensured using ASSERT statements. Astrobe also implements the ASSERTs in an efficient fashion, so there's little run-time overhead for the much improved safety of the code.

Pointers are different beasts compared to all other types, though. For example, if a module defines (and exports) a pointer type, and a set of related procedures, the latter will require a NEW-initialised pointer parameter. However, the procedure has no way of ASSERTing if the pointer is valid, which is, obviously, crucial for its correctness. If the non-initialised pointer's arbitrary value happens to point into the heap or the stack memory regions, obscure defects can ensue. (There's no run-time error when dereferencing a NIL pointer, probably for performance reasons, given the usual frequency of access.)

One can argue that non-initialised variables are a basic problem, and the programmer's mistake. However, parameters of other types can be checked for validity, pointers cannot. NIL auto-initialisation would remedy that. (I don't know the practical impact of such a change on the current compiler design and implementation of the procedure prologue, module initialisation code, and NEW).

Thoughts?

(The Report does not say anything about this issue, N.W.'s ARM compiler paper says (p. 25): "The value of a pointer variable is the address of the referenced record, or it is NIL, when no record is referenced.")

[cfbsoftware]

The implementation effort (and subsequent runtime overheads) required for such are scheme are not insignificant. Keep in mind that pointers can exist as global variables, local variables, register variables, fields of records, dynamic lists and elements of arrays (both global and local). The location in memory of most of these are not known at compile time so it would be necessary to generate code to initialise many of these pointers at runtime. Procedure variables are similar to pointers with respect to safety concerns so may also need to be considered.

This would all unnecessarily penalise the more careful programmers who need greater control of resources and manage the risks themselves instead of assuming that the system is going to prevent them from shooting themselves in the foot. Hence it would need to be an optional feature.

The consequence of all of this is that the feature is low down on our current list of priorities.

If you are interested in investigating further, you can find the most recent discussion related to this topic on the ETH Oberon Mailing List here:

Should one automatically initalize local pointer or procedure variables (safety precaution)?

If your requirements are limited to scalar global pointer variables, then you might be able to implement a simple solution yourself. You could use SYSTEM.PUT and the LinkOption variables to zero the area of RAM that is allocated to global variables in your application. You would need to do this in the initialisation section of the first module that is loaded. Be aware however that would initialise ALL variables - contrary to what many programmers might think, that can have the effect of hiding problems rather than preventing them.

[gray]

Thanks for your explanation. Looks (sounds) like much added complexity without sufficient benefits indeed.