[J3] Question about coarray implementations

[Thomas König]

On 19.05.23 17:33, Damian Rouson via J3 wrote:
> A student has also worked on a threads-based parallel runtime for 
> gfortran that was nearly complete a couple of years ago but never quite 
> made it into a release.

I was/am also involved in it, so I can relate a bit first-hand.

This implementation is process-based, using fork after setting up a
shared memory segment.  This decision was taken to avoid having to
do anything with the library, specifically I/O.  It was also felt that
using pthreads could introduce hard-to-trace and subtle bugs in
interaction with OpenMP and asynchronous I/O.

One of the issues we faced was that, while Linux allows growing a
shared memory segement, MacOS does not (and neither does Windows).
This means that one has to reserve a "big enough" memory chunk on
startup. Inelegant, but not fatal on a virtual memory system.

Another issue was the question if it was possible to pass a pointer
to an entity on another image to some unsuspecting code which knows
nothing about coarrays.  This would have made things extremely hairy
because it would have required a different ABI for coarray-aware code.
This problem more or less stopped development.

Now, it _seems_, after some internal discussions, that the limitations
on ultimate pointer components are designed to disallow this.
Therefore, a question to this list: Is that indeed the case?  The
language of the standard is opaque enough that I cannot really answer
that question.

And we haven't even started thinking about teams.

> There has been some renewed interest in 
> finishing it recently so possibly it will appear in gfortran 14 next year.

Right now, the stutent is busy working on his Master's, and work hasn't
started.

Best regards

	Thomas