Re: [Gomp-discuss] Frontend ..

[Scott Robert Ladd]

My apologies for being scarce recently; between hurricanes and other 
distractions like paying bills, I haven;t had much chance to focus on 
OpenMP in the last week or so.

Lars Segerlund wrote:
> From the discusions lately I have understood that basicly we would
> like to bring the OMP stuff into tree-ssa whithout much modification,
> simply doing a syntax and context check but not generating something
> interesting.

That seems sensible to me.

> Diego's dissertation ( I think it was a dissertation ) seem's like a
> strong argument to handle the parallellisation in the 'middle end'
> aka. tree-ssa.
>
> I have also looked a bit at intels omp compiler, and the question
> remains if our approach of 'doing everything in the lib' will be able
> to match the kind of overhead their solution gives.

Implementing parallelism in the library ensures portability; generating 
inline code provides performance. My thought would be to build a 
portable OpenMP based on a library implementation, while keeping our 
options open for generating inline parallel code for specific popular 
platforms.

> So how do we progress ?

I'm studying NPTL and NUMA elements of the Linux kernel for other 
projects, and this meshes well with our initial target.

> I was thinking of generating a small set of patches agains the
> fortran frontend and the C parser, but it will be hard to get started
> until we know where to do most of the processing. My opinion at the
> moment is to get it all through the frontend, and don't give a damn
> about anything else since this is a primary problem to solve.

Agreed.

> This doesn't exclude that work can be done on the lib, and other
> phases it only relates to the info. However this totally leaves out
> ANY implementation details, which are postponed to a later stage,
> thus there is nothing to hinder us in progressing.
>
> So if anything we should make some testcases for the #pragmas and
> !$OMP directives for C and fortran.

I'm working on more test cases.

..Scott

--
Scott Robert Ladd
site: http://www.coyotegulch.com
blog: http://chaoticcoyote.blogspot.com