This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: What is acceptable for -ffast-math? A numerical viewpoint


Jonathan Thornburg <jthorn@galileo.thp.univie.ac.at>:
<<I normally refrain from posting "me too"s, but in this case
Wolfgang has expressed my thoughts so precisely that I'd like to
chime in.  I too do number-crunching for a living, with large C++
programs running for hours to weeks on a mix of workstations and
supercomputers, using a mix of simple and complex data structures.>>

"me too" but with Ada 95 :).

Following the thread there seem to be two camps that have quite
incompatible viewpoints: (1) experts, (2) other people. I suggest
the following scheme:

* -O2 and -O2 -mieee stay with their current meaning, are
"conservative" and all optimizations need expert group approval before
going in.

* -ffast-math will not be used by experts, and what goes in will be
decided by group (2) (where there seem to be consensus on use and
expectations BTW). Group (2) people will report on all new
optimisations to the GCC list on their own code but from performance
and result change standpoint and then decision will be made by group
(2) alone. Experts are welcomed to show problems with new
optimizations, but if they don't happen in practice and there is
performance gain I expect the optimization will go in.

Note that if some people in the expert group feel like they want their
own -ffast-math I don't mind having a -fsafe-fast-math or
-funsafe-fast-math as long as group (2) has a flag to play with
without much expert interference.

I feel that if we don't make the split, people with potential to do
some "aggressive" FP optimizations will be scared off, and that's bad
for some GNU users and so to the project as well.

Once GNAT is GCC CVS I will be able and willing to report on MIPS,
PowerPC, SPARC and a wide range of x86 processors on our numerical
code (equity derivative pricing at work).

Do people think this proposal is ok? Could people feeling they belong
to group (2) post if they're ready to commit to this procedure, if
they have improvements  to suggest, and what platform they have access
too?

-- 
Laurent Guerby <guerby@acm.org>


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]