Re: g77 & f2c serious problems
Douglas Bates scripsit:
|Are the differences in floating point results relative to other
|machines based on Intel processors? Remember that the floating point
|registers on Intel processors use an 80 bit IEEE extended
|representation. Other architectures such as SPARC use the 64 bit IEEE
|representation. Differences in the results are not necessarily
|because of compiler flaws.
Yes, they are indeed, I was not aware of the difference between Intel
and other architectures. I guess I should have imagined it, but the
NAG F90 compiler doesn't seem to have such problems (of course, being
commercial and expensive one would think they'd be careful about
it...)
|You can emulate the behaviour of other IEEE machines by using the
|-ffloat-store compiler option.
[...]
|Using this option is a bit of a sledge-hammer approach. It purposely
|slows down the calculation and makes it less accurate all for the sake
|of getting the "standard" result.
Doesn't seem to work at all! It makes no difference to the results,
does one have to turn off optimisation? Neither g77 nor f2c+gcc seem
to notice.
|In the alpha-test group for a statistical package called `R' we
|recently encountered a situation where the results under Debian Linux
|were different than those under other architectures. Turns out that
|the code was a little sloppy about some floating point comparisons,
|not that the Linux compilers were wrong.
Fair enough, it simply sounds strange that we are seeing this problem
only with f2c and g77...
Thanks,
Arrigo
--
TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word "unsubscribe" to
debian-user-REQUEST@lists.debian.org . Trouble? e-mail to Bruce@Pixar.com
Reply to: