This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [gomp] More IL work
On Tue, Jul 19, 2005 at 06:06:53PM +0200, Sebastian Pop wrote:
> Why too early? Their development is orthogonal to OpenMP. It will
> just use the existing infrastructure for generating parallel code.
>
Nope, it's very much related. They will likely need to call into
the IL infrastructure for generating the OpenMP directives. They
won't be able to inject a #pragma in the code (we cannot do
source-to-source easily). They will either want to generate
GOMP_PARALLEL_* IL or call the APIs directly. All of which is
still very much in development.
In a sense, it may be an advantage to the development of the
general framework. But the current objective is OpenMP
code generation, not a general parallelization framework.
> The aim is to split a loop nest into chunks and to distribute these
> into several threads, so the target will be data parallel code with
> calls to libgomp.
>
Exactly (see above).
> I was thinking to put their contributions into autovect only because I
> will work on the loop fission there, and that we have some more bits
> on the data dependence tester. But for what they want to do the
> existing data dep of mainline/gomp is ok.
>
The data dependency analysis, maybe. The rest will be changing
under their feet every couple of weeks. If they can live with
that, then OK.
> > How about a sub-branch?
>
> Their changes won't be too big, one or two new files: a new pass.
> Then with time I think they will be interested in having their pass
> triggered only for some loops and then they will probably contribute
> also to the general goal of gomp. So I won't scare them away by
> asking them to contribute to general infrastructure.
>
They *will* need new general infrastructure. We've got nothing
right now. But, I guess if they can live with a shifting
framework, then I don't mind too much.
Diego.