IRIX6.4 REGRESSION TEST EGCS-1.0.3A VS EGCS-19980803 -Re

[tprince@cat.e-mail.com]

          -Reply

I recall having a short discussion about adding MAC support with
Jeff Law.  There's probably more work to be done in
implementing sub-architecture options like -mcpu= so that we
don't lose backward compatibility; since gas isn't supported on
Irix6 there wouldn't be an issue there as there is on hpux.

I suppose more people would be interested if I get a working C
version of Livermore Kernels; I imagine I'd want to do quite a bit
of work on one of the standard ones to get it into what I would
consider reasonable shape.  I have doubts that I could just f2c
the one I have, but I imagine the official one isn't a lot better than
that.  I guess I'd rather wait until I hear from someone that my
maintaining a C version of LK would motivate them to do the job.
In case people have forgotten, I've been using my version of LK
which modernizes some of the syntax and is more
compiler-friendly, while performing the standard operations.

The other issue, which I've discussed by e-mail with a couple of
people, is how much work people want to do for an architecture
whose field of application is likely to be vastly restricted a couple
of years from now if  the IA64 act gets together. (Both HP and SGI
say already that they will not continue architectures which
compete directly with Intel).  I've been amazed how much egcs
work is being done with Irix and at the quality of the results.


> Livermore Kernel tests also show progress.  All kernels show
> expected double precision accuracy, and 7 of the 24 run at the
same
> speed as with SGI's best compiler.  Performance differences
are due
> mainly to the lack of implementation in egcs of the floating
point
> multiply-add sequence instructions.

Maybe you can write a canonical testcase in C, show us the
output that
egcs gets, and the output by SGI's best ocmpiler, and find
someone
willing to give it a try.  For pointers on who to implement this in
gcc, people can look for mac in a29k, h8300, sh, and sparc.

Get way to get your feet wet.


           To:                                              INTERNET - IBMMAIL
                                                            N0531271 - IBMMAIL