Re: concurrent access to containers (was: GNAT for MS Visual Studio)

[Georg Bauhaus <rm.tsoh+bauhaus@maps.futureapps.de>]

On Fri, 2007-10-05 at 06:25 -0700, Maciej Sobczak wrote:
> On 4 Pa , 22:21, Pascal Obry <pas...@obry.net> wrote:
> 
> > >http://www.adaic.org/standards/05aarm/html/AA-A.html
> >
> > > 3/2:
> >
> > > "The implementation shall ensure that each language- defined
> > > subprogram is reentrant  in the sense that concurrent calls on the
> > > same subprogram perform as specified, so long as all parameters that
> > > could be passed by reference denote nonoverlapping objects."
> >
> > Fine but we are speaking about "reading" only the parameters. This does
> > not apply.
> 
> I understand you, but I don't see this exclusion in the paragraph
> above.
> 
> What I see is "subprogram", "calling", "parameter" and "overlapping".
> I don't see *any* distinction between reading and writing.
> 
> I other words, the obvious guarantee [...]

I don't think it is obvious that something of this sort
should be guaranteed (read-only access from multiple tasks is
possible, and trouble free). For example, algorithms used
by search engines or by distributed databases do not
necessarily shy away from copying shared data. So nothing is
physically shared in these systems and algorithms at the level
will be free of concurrent access even though the system
is concurrent. It still works on exactly one virtually shared
set of data...

Should standard container design assume it makes sense to
restrict implementers by assuming one specific computer
architecture like shared memory with multiple CPUs?

Or should standard container design assume that multiple
copies of data will be ubiquitous given current hardware options?

Can an Ada standard library assume that there is a compare
and swap operation available on typical 8bit or 16bit CPUs?
What if there isn't?