[PATCH] [WIP] openmp: Add OpenMP 5.0 task detach clause support

[Jakub Jelinek]

On Thu, Dec 10, 2020 at 03:38:40PM +0100, Jakub Jelinek via Gcc-patches wrote:
> I don't understand this.  My reading of:
> "The event-handle will be considered as if it was specified on a
> firstprivate clause. The use of a variable in a detach clause expression of a task
> construct causes an implicit reference to the variable in all enclosing
> constructs."
> is that we should do:
>       case OMP_CLAUSE_DETACH:
> 	decl = OMP_CLAUSE_DECL (c);
> 	goto do_notice;
> which does the second sentence, and for the first sentence I believe it
> talks about the task construct rather than about the outer construct.
> So (again, something for testsuite):
> void
> foo (void)
> {
>   omp_event_handle_t t;
>   #pragma omp parallel master default (none) /* { dg-error "..." } */
>   {
>     #pragma omp task detach (t)
>     ;
>   }
> }
> The dg-error should be the usual error about t being referenced in the
> construct but not specified in the data sharing clauses on parallel.
> And then
> void
> bar (void)
> {
>   omp_event_handle_t t;
>   #pragma omp task detach (t) default (none)
>   omp_fullfill_event (t); // This should be ok, above first sentence says
> // that it is as if firstprivate (t)
> }
> 
> But I think it is actually even stronger than that,
>   #pragma omp task detach (t) firstprivate (t)
> and
>   #pragma omp task detach (t) shared (t)
> etc. should be invalid too (at least in pedantic reading).
> I guess we should ask on omp-lang.  If it actually works as firstprivate
> (t), perhaps we should handle it that way already in the FEs.

Asked and Alex said that both should be invalid.  Though, if we implement
detach as passing address of the variable to GOMP_task, if we implicitly
add firstprivate clause it would copy the value from before it has been
initialized.  One way to handle that would be not add firstprivate clause
next to detach, but treat detach like a firstprivate clause in most places,
and just for the passing pass it specially (let parent of task pass address
of the variable and let the receiving side recieve the value instead,
which would force task_cpyfn, or handle it more like we handle the bounds
of a taskloop - force the omp_eventhandler_t to be the first variable in the
structure and let GOMP_task write the address not just to *detach, but also
to the first element in the structure.

	Jakub