induct regression in gcc 4.3.1

[Dominique Dhumieres]

Jack,

I think the cost model for Intel Core is still wrong in 4.4.
The difference between 4.3 and 4.4 is that in 4.4 the inner loops
(array ops in f95) are unrolled and not vectorized, the vectorization
occuring on a length 9 loop, leading to an almost factor 2 speed-up
(~9/5 to be more precise).  I doubt there are plans to backport that
to 4.3, but who knows!

Dominique