Re: crash in 1.0.3a fortran compilation

[Dave Love <d dot love at dl dot ac dot uk>]

>>>>> "Greg" == Greg Lindahl <lindahl@cs.virginia.edu> writes:

 >> Possibly because these codes are large and most people want to
 >> download them along with the compiler.

 Greg> That sounds like a social problem, not a technical one.

I, for one, simply don't have enough space to do some of the testing I
think should be done on significant packages with multiple compiler
versions, even though I can NFS-mount Netlib stuff, and probably not
even the cycles.  I think that's the basic reason it hasn't been done
for previous g77 releases, at least on x86.  I don't know if that's
technical or social...  

If people can take this on for Fortran, it would be very useful,
especially if they can check performance too, but not for double
precision on x86 or with broken benchmarks, of course :-/.

 >> Unfortunately, by the time I'm done running these tests,

 Greg> ... which shouldn't take long, since they should be automated.

You're obviously blessed with more resources than some :-(.

Probably the best candidate for `realistic' tests if people can do it
is Octave, since it's a GNU package with Fortran and C++ and it
exercises a reasonable amount of the Fortran linear algegra.  There's
been at least one g77 problem with it in the past.

The problem is that its test framework is based on the released
dejagnu and isn't compatible with the egcs `infrastructure' version.
I can't remember the problem and don't know enough about dejagnu
anyhow, but when I did it for the last (?) g77 release I had to
(un)install the dejagnu versions in turn, clobbering something in the
process.  I think the default Octave configuration only uses -O for
g77, which might usefully be bumped to -O3 -funroll-loops for testing,
and I think you have to configure specially to exercise PIC, which we
should.