Re: 'address on a subprogram parameter

[Niklas Holsti <niklas.holsti@tidorum.invalid>]

On 2021-04-24 4:25, Randy Brukardt wrote:
> "Vincent Marciante" <vincent.marciante@l3harris.com> wrote in message
> news:07a090e1-85b4-4ebb-a093-dcda57b56579n@googlegroups.com...
>> Got it (and "no way" that the big record that is in the real code would be
>> passed by copy
>> but I'll see if specifying it explicitly would not be disruptive).  Thanks.
> 
> Careful about assumptions, if you want the code to be portable.


Seconded.

Long ago I was working with early "public" GNAT versions ("gnatp") on a 
certain program. After we upgraded to the next version of gnatp (I think 
around version 3.12) our program started failing with Storage_Error. 
Turned out that the GNAT developers had decided that this gnatp version 
would pass all record-type parameters by value, "to be more like C" 
(IIRC Robert Dewar was said to be the origin of that idea, but that may 
be a calumny). And our program had some large record types.

We ended up shipping the program with the advice to use a 10 MiB stack 
allocation -- but it worked (it did not have any real-time requirements).

As a positive, the value-passing gnatp version made us discover a couple 
of unintended parameter-vs-global aliasing issues that had been hidden 
when all records were passed by reference.

The very next gnatp version returned to pass-by-reference for 
record-type objects. It seems there had been some negative user 
feed-back :-)


> At one point, Janus/Ada would pass pretty anything by copy to 
> parameters of certain generic types. (We got rid of that in the Ada
> 95 version because of the need to deal with "aliased" components, but
> the cost is that one needs to use a thunk to access anything of most
> generic formal types in a generic body. [Janus/Ada still uses
> universal generic sharing.])
>                                     Randy.