[dmd-concurrency] real

[Benjamin Shropshire]

Andrei Alexandrescu wrote:
> It looks like we're converging on this:
>
> - shared pointer read/assignment works
>
> - shared numeric data read/assignment works, EXCEPT for real
>
> Real is an odd beast - it has no guaranteed width and is mostly 
> system-dependent. Should we guarantee atomic assignment to 
> shared(real), leave it to the platform, or just disallow it entirely?
>

As food for thought, I see 3 options

- real can't be shared, end of story
- real can be shared but only on one machine (or maybe architecture)
- real can be shared anywhere and it gets translated as needed.

The first would be anoying.
The second would cause odd issues in some cases (WTF doesn't it work 
here?!).
The last would cause some interesting perf issues. OTOH, translation 
issue will need to be dealt with as soon as big/little endian issue come up.

What does MPI do?

OTOH

>
> Andrei
> _______________________________________________
> dmd-concurrency mailing list
> dmd-concurrency at puremagic.com
> http://lists.puremagic.com/mailman/listinfo/dmd-concurrency
>