ACATS c380004

[ed schonberg]

Duncan Sands wrote:

>Hi Richard, thanks for the informative reply.  I think I am convinced.
>
>  
>
>Some comments on this: protected objects and tasks differ from the above record type in that
>they are limited types: they can't be copied.  I guess one reason for the compiler choosing to
>use the maximal possible size for a record with a default discriminant is that copying is
>simplified: just do a fixed size memory copy.  This reason does not apply to protected and task
>types since they can't be copied.  However another reason for allocating the maximum size springs
>to mind: it simplifies the creation of arrays (eg an array of tasks), since all members of the
>array have the same size.  Having arrays where each component has a different size is too
>horrible to contemplate.  On the other hand, I suppose arrays of protected objects and tasks could
>be implemented as arrays of pointers to the actual tasks/protected objects, invisibly to the
>programmer.  Anyway, after thinking a bit about the problems involved I now have a lot more sympathy
>for ACT's position!  In fact I'm surprised their protest was rejected - why was it rejected?
>  
>
An entry family is most naturally implemented with an array of queues, 
and the task discriminants are mapped into the discriminants of the 
enclosing record, which is why we use
the same implementation strategy for regular records and for task 
records. So far so good.
However, in the case of protected types, an entry family can be 
implemented with a single
queue, and pending requests on it can be kept ordered by index (if the 
queue is long the
system is thrashing anyway!). In this case, the  array implementation 
suggested by the "family" is bogus, and the problem disappears. The ARG 
feels that this is a preferable implementation anyway, and we will 
probably adopt it when we revisit the implementation of protected types 
for Ada2005.

>In the dark old days, the GNAT compiler didn't allocate the maximum possible size for records with default discriminants.  They changed it a some point, for efficiency reasons IIRC.
>  
>
as far back as I can recall, we always allocated the maximum possible 
size for mutable
discriminated records. That's basic: no under-the-covers dynamic 
allocation.

Ed Schonberg