Re: Pgcc in Deb
David Starner wrote:
> > I disagree. From experience, I know that up to %50 speedup can be gained
> > in number crunching stuff. I'd suspect %20 could be pretty normal for
> > most CPU hungry apps, and the overall speedup would be significant.
>
> Compared to gcc 2.95 with -march=i686 (or i586, as the case may be)? That
> sounds a lot higher from what I've heard, even from those who would boost
> it (the Intel engineers, for instance.)
Okay, I'm mistaken. I'm talking about the whole speedup from -march=i386,
which is I believe to be the default when I issue debian/rules binary (?).
Excuse my ignorance in that. However, there is something that I do not fully
understand. There doesn't seem to be march=i586 optimizations in all of
the debian packages, right? Or is the i386 distribution already pentium
optimized like hell? I'd be very glad if the developer tools that I use ran a
bit slicker. As a programmer, I feel that might be the case when I compare the
performance of my system to a Windows box.
For random, I picked jed sources, and when it was running the ./configure
script I came across the following lines: <<
You need to edit src/Makefile if the following are not correct:
CC = gcc
CFLAGS = -g -O2 -fno-strength-reduce
>>
I suppose that means no pentium optimizations turned on for jed. Now that I
have enjoyed using the pentium-builder package, I do think that the pentium
specific optimizations should be considered seriously for the i386 arch. I'm
referring to -march=i586 or -march=i686 throughout the discussion, not
-mcpu=i586 or -mcpu=i686. Once potato is released, a pentium optimized single
disc distro may be attempted.
>
> Be that as it may, you didn't answer the rest of my claims, or the arguments
> that's too unstable to be used. Nobody's going to complain if someone puts
> together a pgcc deb, but a pgcc/i686 distribution is going to take up manpower
> and archive space that the developers* don't want to support.
Now that I'm trying to correct my point, do you think that enough resources
have been put to optimize the distribution for any architecture? Is there
any policy on optimizations? Or is it just that the i386 distribution *should*
work on i386/i486/i586 without any problems, otherwise letting developers turn
on any optimization flags as they see fit? How about such specific optimizations
on *other* architectures, or do we always take the safest path? [I guess there
might be packages that would be broken when built with march=pentium-pro,
that could be an ultimate test for Intel and AMD anyway! :) ]
Regards,
--
++++-+++-+++-++-++-++--+---+----+----- --- -- - -
+ Eray "eXa" Ozkural . . . . . .
+ CS, Bilkent University, Ankara ^ . o . .
| mail: erayo@cs.bilkent.edu.tr . ^ . .
Reply to: