geode-optimized glibc routines

[Rob Savoye]

Marcelo Tosatti wrote:

> I did some testing by replacing the current glibc string functions we
> use with the Geode optimized ones, and the results are good:
> 
> http://dev.laptop.org/~marcelo/glibcperf/optimizations.html

  Yep, that's the changes I was using as my code base.

> Hard part of this is to get it merged upstream. As far as I can grasp,
> there is no framework on glibc for providing optimized string functions
> for given x86 processors versions.

  Well, sort. They can go in glibc/sysdeps. I created a single new
directory to hold the improved assembler versions.

> I heard John Zulauf tried to merge those patches when he first wrote
> them, but conversation with Ulrich was not very productive. I'm unaware
> of the technical details.

  I'd love to get a patch if you already have them building. I'm almost
done tweaking my Binutils/GCC changes that add an OLPC configuration, so
I've restarted on the glibc changes since people seem to want them. I
can see about getting the changes migrated upstream into
GCC/Binutils/Glibc once we're all happy with them. I know many of the
maintainers personally since we were all co-workers at Cygnus, which
makes things much easier.

  One issue is what to use for the config triplet. I mentioned this
before and got no response, but I'm using i585-olpc-linux, which uses
olpc as the vendor field, which I believe is the correct way to do this.
Then these changes will work when cross compiling, instead of only for
native builds. This is what would trigger the optimized versions being
configured into the build.

	- rob -