Hi Tobias,

That surely is a lot of janitorial work :-) LGTM.

Thanks for pushing forward on this. Do I take it that you will implement
device_type(host) in gcc or is somebody else onto it?

Best regards

Paul


On Wed, 26 Nov 2025 at 12:08, Tobias Burnus <[email protected]> wrote:

> I started looking into this based on some unified-/managed-memory
> discussions. At the end, I ended up doing the following where only
> the 'local' clause is directly related to USM - albeit 'groupprivate'
> implies 'declare target' with 'local' clause, which is the link.
>
> Note: The following is only the Fortran parser work, i.e. new features
> will print a sorry.
>
> - some janitorial changes to 'declare target' (missing diagnostic,
>    cleanup). In particular, check that with 'link' there is no
>    'device_type(nohost)'.
> - add the 'local' clause to 'declare target' (+ sorry)
> - add the 'groupprivate' directive (+ sorry)
> - add the 'dyn_groupprivate' clause to 'target' (+ sorry)
>
> Comments before I commit it?
>
> Tobias
>
> PS: Eventually, GCC needs to handle 'device_type(host)'
> (for variables and functions and enter/link/local),
> besides improving/supporting the 'device_type(nohost)'
> behavior. And, of course, actually implementing 'local'
> and (dyn_)groupprivate also has eventually to be done ...
>

Reply via email to