Re: what will get backported?

[Janne Blomqvist <jblomqvi at cc dot hut dot fi>]

On Mon, Mar 13, 2006 at 08:46:25PM -0500, Jack Howarth wrote:
> Jerry,
>    Actually I am far more concerned about changes where optimization
> improvements are being made by wiring in the backend optimizers to
> gfortran. Specifically thing like...
> 
> http://gcc.gnu.org/ml/fortran/2006-03/msg00174.html
> 
> which result in significant perform increases in the generated code

No, except for nf it doesn't result in any significant performance
improvements.

Looking at the numbers, it's quite clear that something else was
running in the background during the first set. Look at the user times
instead of the wallclock.

> while not actually implementing new features or touching things outside
> of the gfortran code. I would really hate to see those sort of things
> put off until gcc 4.2.

Although I agree with you in principle, there is the issue that 4.1 is
a release branch, the first one where gfortran is generally usable to
boot. Introducing regressions there is a very bad thing, so one needs
to be extra careful.

One thing that was deemed significant enough to backport to 4.1 was
Paul Thomas patch inlining dot_product. It's the one responsible for
the huge improvement in induct which you can see at
http://www.suse.de/~rguenther/polyhedron/ .

-- 
Janne Blomqvist

Attachment: pgp00000.pgp
Description: PGP signature