Re: GNAT: Unbounded_Strings

[Florian Weimer <fw@deneb.enyo.de>]

"Maxim Reznik" <max1@mbank.com.ua> writes:

> Modifying of a shared string is synchronized by reference counter.
> While this counter >1 any change of shared string can't be done.

Yes, that's necessary for correct operation, and it might be
sufficient if the counter cannot raise after the comparison.

> One thing I could lose sight of is synchronization of counter itself.
> 
> Are following changes enough ?

I'm not sure, I think I would have to see the whole source code.

> I think operators
>++++
>           if Source.Share.Count.Value = 1 then
>                Change_Somehow( Source.Share.Space );
>           else
>                Make_A_Copy
>                Decrement_Count (Source.Share);
>           end if;
>++++

> are Ok, because of Count.Value=1 means that there isn't any other
> references to string and nobody can change neither Share.Count nor
> Share.Space.

I think this agrees with the letter of the ARM.  I'm still not sure if
it breaks any of my expectations regarding string types, but I

> One bad consequence of this solution is 40 bytes overhead for each strings.

You could use one global lock for all reference counter operations.
Contention on this lock isn't a problem when compared to the original
GNAT implementation because all operations which require copying a
string already involve the global allocator lock.