Re: Automatic Parallelization & Graphite - future plans

[Albert Cohen <Albert dot Cohen at inria dot fr>]

Steven Bosscher wrote:
> On Wed, Mar 18, 2009 at 8:17 PM, Albert Cohen <Albert.Cohen@inria.fr> wrote:
> > Antoniu Pop wrote:
> > (...)
> > > The multiple backends compilation is not directly related, so you
> > > should use a separate branch. It makes sense to go in that direction.
> > Indeed.
>
> Work has been going on for years in this direction, but it has never
> gone very far.
>
>
> > There has been some work in the area, using different approaches. I've been
> > involved in one attempt, for the Cell, with Cupertino Miranda in CC.
> >
> > Cupertino: could the URL where to find documentation on your experiments,
> > and the (old) patch to GCC and the (old) Cell SDK for that purpose?
>
> What approach was taken in these experiments?

Cupertino will send you the documentation and reference to the old patch 
sent on the gcc-patches list.

In brief, there is no hotswapping, just attributes to let the MIDDLE-END 
decide, right before expanding to RTL, which part of the trees to keep 
and which to drop. Selection is done at the function level, and at the 
variable declaration level.

You still need multiple runs of cc1, but they could be hidden behind a 
single run of the driver. It may not be a good idea, though, since 
different optimization flags may be relevant for the different backends 
(and even for the different functions, but this is a distinct issue).

The point, for the Cell, was to perform whole-program analysis across 
ISA boundaries. E.g., looking at IPCP or specialization and inlining. 
Another example is to be able to assess the profitability of a 
transformation on a function that compiles to target X, but internally 
depends (calls) a function on target Y. You would have to assess the 
side-effects, cost, and pass static and dynamic profile data across the 
boundaries again. This was only a target direction, we did not do 
anything there.

We struggled a lot with the data types and API that differ widely and 
non-consistently between the Cell PPE and SPE... as if IBM did not think 
until very late that single-source-multiple-backend compilation was 
relevant!

Albert