This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: g77; numerical accuracy; 80 bit spills
- To: egcs at egcs dot cygnus dot com
- Subject: Re: g77; numerical accuracy; 80 bit spills
- From: Dave Love <d dot love at dl dot ac dot uk>
- Date: 11 Jul 1999 20:42:51 +0100
- References: <37867A8B.BD3DA613@cls.usask.ca>
The `Floating point precision' node of the manual was written by a
physicist/programmer, though perhaps it won't be left in long. Do you
actually have a problem following its advice? I've heard of only one
real case where it apparently didn't solve the x86 extended precision
problem (because of the remaining extended exponent range for reasons
I don't recall).
I also supplied the references in `Floating-point Errors' (after
reading them and before they appeared elsewhere and became fashionable
to quote). See particularly the new appendix to the Goldberg paper.
[Yes, in _some_ cases a 50% speed penalty might be significant to me
(my users). But note that sometimes you can pay an arbitrary price
for IEEE compliance. I have seen the order of magnitude sort of
slowdown mentioned in the Sun numerical computation guide (should be
at http://docs.sun.com) in real life. Life's hard and these things
aren't black and white in my book after embarrassingly long experience
which often seems to be wasted.]
BTW, anyone who thinks that there are unique issues with g77 in such
regards, and that writers of potentially-ported code don't have to be
careful, hasn't had to support as many systems as I have (often
without actual access to them or proper docs). That was my main
motivation for trying to contribute to g77, regardless of GNU
sympathies.