geode-optimized glibc routines
rob at welcomehome.org
Sat Apr 14 20:30:25 EDT 2007
Jordan Crouse wrote:
> On 09/04/07 11:34 -0700, Andrew Burgess wrote:
> I thought thats what the middle tuple is for. This machine has an i586
> instruction set - make no mistake about it. The difference comes in the
> slightly different timings and cache behavior on the Geode as opposed to
> a reference Pentium machine.
That was why I thought using the vendor field was most appropriate. I
just got a new set of cross tools built with the Geode optimizations
enabled (GCC 4.3.x cvs), so I restarted adding the glibc optimizations.
This version isn't substantially different than the previous snapshot
other than this has support for a few addition assembler instructions
kicked out by the GCC code generator that turned out to not be
supported. Lack of those kept any i386 cpu type from building for the
last week or so. This patch is not checked in yet, but it's now in GCC
bugzilla. Works for me... :-)
Anyway, in looking at geode-perf, I notice there are multiple copies
of memcpy, strlen(), etc... in each directory. I've been digging through
the docs and the csv result files, but was wondering if there was a
simple document someplace that just specifies which version of the each
optimized routine was the best one for an OLPC ? I'll figure this out
eventually, but would appreciate saving the time. I've got the config
changes worked out, but everything depends on which files I use. I
figure I'll get glibc rebuilt and add it to the cross tools before
making a new snapshot available.
Then after a little testing by more people than myself, I'll see about
getting the changes migrated upstream in GCC and Glibc.
- rob -
More information about the Devel