Re: Polyhedron shootout between g95 and gfortran

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

On Thu, Feb 23, 2006 at 11:45:21AM +0100, François-Xavier Coudert wrote:
> I was just thinking about the dot_product timings issues (and the
> cross-over problem), and I can't think of a fundamental reason why we
> couldn't have all the speed of the library in an inline version?

I guess the fundamental reason is that the optimizer won't be clever
enough to outsmart a tuned library version. Then again, the current
library version is anything but a tuned implementation, not to mention
a set of arch-specific tuned implementations.

IMHO a good solution would be to always inline (thus getting rid of
dot_product in libgfortran), and have an option to use an optimized
BLAS version for large arrays (with a user-selectable cross-over
switch to override the default) if applicable (see PR26025).

One reason to keep the libgfortran library version would be if we'd
like to avoid inlining for -Os, but I suspect that issue won't matter
anyway considering how Fortran is used in practice.

> (at
> the cost of adding maybe a little extra info to help the optimizer)

Such as?

-- 
Janne Blomqvist

Attachment: pgp00000.pgp
Description: PGP signature