This is the mail archive of the
fortran@gcc.gnu.org
mailing list for the GNU Fortran project.
Re: [RFC] Merge criteria for tree-ssa
- From: Steve Kargl <sgk at troutmask dot apl dot washington dot edu>
- To: Joost VandeVondele <jv244 at hermes dot cam dot ac dot uk>
- Cc: gcc at gcc dot gnu dot org, fortran at gcc dot gnu dot org, dnovillo at redhat dot com
- Date: Thu, 5 Feb 2004 12:58:40 -0800
- Subject: Re: [RFC] Merge criteria for tree-ssa
- References: <Pine.SOL.4.44.0402051525300.523-100000@yellow.csi.cam.ac.uk>
On Thu, Feb 05, 2004 at 04:09:26PM +0000, Joost VandeVondele wrote:
>
> I think that it would therefore be appropriate to have some kind of
> quality criterium for gfortran, to ensure that a functional fortran
> compiler will remain part of mainline gcc (useful for some part
> of the scientific community). Running the g77 testsuite without new
> regressions seems like asking a bit too much
g77 defines a boat load of extensions to the Fortran 77 standard.
Supporting the g77 language is much different than supporting F77.
> (also because gfortran intends to be a fortran 95 compiler),
A Fortran 77 conforming program is a Fortran 95 conforming program.
> but compiling e.g. lapack and running the testsuite without new
> regressions could be more reasonable (this is what g95.sourceforge.net
> is presently able to do, it is still strongly f77 biased, but I think
> mostly legal f95).
For testing conformance to the Fortran 77 standard, you should run
gfortran with the NIST testsuite. There are 4 main problems:
(1) characters in COMMON or EQUIVALENCE statements.
(2) arrays of CHARACTER strings
(3) non-integral DO LOOP variables
(4) ICE on ENTRY statement
> The fortran parts of the spec benchmarks seem also good candidates,
> since several people have already filed detailed bug reports. Ideally,
> compiling some larger f9X programs correctly would be nice.
I run 225000 lines of code through gfortran several times a week.
This only checks for compile regressions because of several gfc_todo
failures prevent creation of my math library [See (1) and (2) above].
On the bright, a 27 Jan 04 gfortran produced the following benchmark
numbers on one of may codes:
Time
6.92 gfortran -x f95 -o zz -O0 -static shell.f
3.74 gfortran -x f95 -o zz -O1 -static shell.f
3.12 gfortran -x f95 -o zz -O2 -static shell.f
2.83 f95 -o zz -O2 -dusty -Bstatic shell.f
6.76 gfortran -x f95 -o zz -march=athlon -O0 -static shell.f
3.16 gfortran -x f95 -o zz -march=athlon -O1 -static shell.f
2.74 gfortran -x f95 -o zz -march=athlon -O2 -static shell.f
2.82 f95 -Wc,-match=athlon -o zz -O2 -dusty -Bstatic shell.f
Time is in seconds and the average of 3 runs. f95 is a commercially
available compiler for the FreeBSD platform. Finally, gfortran
produced the correct answer. :-)
--
Steve