[Bug fortran/36599] [4.3 regression] induct.f90 polyhedron benchmark in 4.3.1 on Intel

["dominiq at lps dot ens dot fr" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #16 from dominiq at lps dot ens dot fr  2008-09-08 09:00 -------
A few personal comments.

> 2) The problem doesn't occur on powerpc-apple-darwin9.

This is normal. REAL(8) are not vectorized on ppc since they are not part of
altivec. IBM has preferred to add a second FPU. Although the gcc scheduler
needs some improvements for a better use of it, this is an other issue.

Concerning the problem on Intel Core, I think there is very little that anythin
can be done for gfortran 4.3. The easiest "fix" would be to include a CAVEAT in
the gfortran documentation stating that short dot-products should be unrolled
by hand in order to allow the vectorization.

Indeed this can be done at the level of the gfortran front-end, but I don't
this it is worth the work.
Note that this should be restricted to 4.3, otherwise it will give a
performance regression on 4.4 (see below).

A last possibility is to convince the middle-end people to tweak the vector
cost in the middle-end (good luck!).

I think the early unrolling improved the benchmark results, but introduced a
performance regression in one of my avatars of the induct test. I have tried to
reduced the problem in pr36099, but this led to a wrong direction. I am
planning to do a better reduction, but I did not find the time so far.

To conclude I think this pr should be closed as WON'T FIX with or without the
caveat since it is not a gfortran front-end problem.




-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36599