Public SourceMage Discussion List

[Eric Sandall <eric AT sandall.us>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Sun, 22 Jan 2006, David Brown wrote:
> I would like to see some additions and changes to the archspec files.
<snip>
> == Proposal ==
> I would like to see minimal archspec files for each archspec that we
> add optimizations that are optional to what's needed to define the
> archspec. These are not replacements but additions to the archspecs,
> the fully optimized archspecs would still be there. For example, the
> pentium4 archspecs would get redused to just CFLAGS="-march=pentium4".
> Along with the optional custom CFLAGS users can add back in what
> optimizations they would really like and remove the optimizations they
> don't want.  The x86-64 archspecs would probably look like
> CFLAGS="-march=x86-64 -fPIC -DPIC" because the prelink flags are
> required to make the system work, but the other stuff wouldn't need to
> be there.

Would there be two entries for the arch (e.g. pentium4 and
pentium4-ricer) or just one pentium4 with the minimum
(-march=pentium4) and then a separate menu for the additional archs?
If the latter, the separate menu could be generated by adding an
OPT_CFLAGS in the archspec file for the optional CFLAGS that we have
removed from the minimal.

e.g. for pentium4:
CPUNAME="Intel Pentium IV"
if use_gcc2; then
 CFLAGS="-march=i686"
else
 CFLAGS="-march=pentium4 -mmmx -mfpmath=sse -msse -msse2"
 OPT_CFLAGS="-mmmx -mfpmath=sse -msse -msse2"
fi
   HOST="i686-pc-linux-gnu"
   FAST="-O3"
  SMALL="-Os

The optional archspec menu would the parse the space separated list of
OPT_CFLAGS (if it exists) and have checkboxes for each. e.g.
  [ ] -mmmx
  [ ] -mfpmath=sse
  [ ] -msse
  [ ] -msse2

Or we can just keep it simple (KISS) and have the minimal
specifications in the archspecs and let the users use the custom
CFLAGS area to add their own.

Any way we do it I saw we only test our spells against the new minimal
set of archspecs and if a bug comes up caused by custom archspecs we
can put in code to remove them, but let users know that this is a
courteousy and by no means required. :)

> Also I was messing around with the automake config.sub and I noticed
> something different about what it returns and what we define as the
> HOST for the x86-64 family of archspecs:
>
> dmlb2000@dmlb2004:~$ /usr/share/automake/config.sub x86_64-linux
> x86_64-pc-linux-gnu
> dmlb2000@dmlb2004:~$
>
> I was wondering if the HOST for the x86-64 family of archspecs could
> be changed to what config.sub returns.  The only problem I've noticed
> by changing the archspec to this was that glibc checks for
> x86_64-unknown-linux-gnu instead of x86_64-*-linux-gnu, not quite sure
> why, maybe Seth can shed some light on that.
>
> If I don't hear any chickens running around with their heads cut off,
> I'll put the additions into perforce in a week, in case people don't
> quite "get what I'm proposing"
>
> - David Brown

What would we gain from setting the full explicit name rather than
letting ./configure figure it out?

- -sandalle

- --
Eric Sandall                     |  Source Mage GNU/Linux Developer
eric AT sandall.us                  |  http://www.sourcemage.org/
http://eric.sandall.us/          |  SysAdmin @ Inst. Shock Physics @ WSU
http://counter.li.org/  #196285  |  http://www.shock.wsu.edu/
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.1 (GNU/Linux)

iD8DBQFD1V6YHXt9dKjv3WERAuWdAJ0fo6AJ2owO6qzzJYe8E7tJaeLsowCfVuiX
j3xaSDazb3ZZnPvZK6Xc98k=
=3+Qo
-----END PGP SIGNATURE-----