Re: [Gomp-discuss] sentinels

[Steven Bosscher]

Op di 04-02-2003, om 09:43 schreef Lars Segerlund:
> 
>   Ok so all agrees that we can use the existing frontends ?
> 
>   Does this mean that we will use g95 for FORTRAN and the new C/C++ parser ?
> 
>   Another issue that has emerged is where we keep the code ? My 
> suggestion is that if possible we should live in the tree-ssa branch, 
> which raises another even bigger question, how do we organize our code ?
> 
>   Since we will be doing modifications to the other frontends and some 
> of the optimizers, do we keep our own copies of the modified files or 
> how do we proceed ?

I think the Fortran stuff can wait.  G95 is nowhere near complete and
like I told you in a private discussion, the g95 parser is basically a
template matcher that is to brittle that I would rather not touch it for
a while.

So IMHO it would be best to start with the C front end.  The sooner we
can show off our Great Effort to the GCC community, the more support
we're likely to get, so I think we should be in GCC CVS.  Some
sub-branch from the tree-ssa branch to 1) implement the OMP C pragmas
and 2) Define the tree codes to go with that?

>   A reasonable first goal would be to get the sentinels, ( '!omp' and 
> '#pragma omp' ) to the middle end.

Agree.  We can make the middle end just ignore the extra information for
now, that's always legal.
 
>   What about the distinction of 'parallell' and 'distributed' regions, 
> does everyone think this is a good idea ?

The keyword is "concurrent".  We can leave it to the middle-end to
decide what that means for the generated code.
 
>   As for where to start with openMP, there is only one place, we need 
> parallell regions to start with, and at least the private/shared attributes.

Yup.
 
>   Further more I suggest we start with the SECTIONS and SECTION 
> directives, if we can do these and the suncronisation constructs we 
> should have the most of the functionality in place.
> 
>   Any thoughts ?

Scott said he would have a look at the differences between Fortran on
the one hand and C/C++ on the other.  I suggest we wait for him to draw
conclusions and then decide which directives to implement first.

Greetz
Steven