Features/ArchitectureSupport - changing what we build for

Gregory Maxwell gmaxwell at gmail.com
Tue Feb 3 06:01:28 UTC 2009


2009/2/2 King InuYasha <ngompa13 at gmail.com>:
[snip]
> Honestly, I think Fedora needs to stop
> bottlenecking its performance by not optimizing its code for i586 or i686. I
> would prefer the code to be optimized to i686 when built as RPMs.

-march != -mtune.  The code is already tuned for more modern CPUs.
Allow me to repeat this: (loudly, not to be mean but because people
seem to miss it) FEDORA x86 IS ALREADY OPTIMIZED FOR MODERN CPUS.

mtune should result in the overwhelming majority of the performance
difference with none of the compatibility issues.  -march=i686 adds a
single instruction of disputable value, cmov. In exchange for getting
this one additional instruction it breaks compatibility with pentium,
VIA C3 (Nehemiah has cmov, but eden does not; I have an eden600 board
here which is only ~3 years old!) , AMD K6, etc and potential
performance problems on Pentium 4.


Do you have benchmarks that show given a constant -mtune that
-march=i686 makes a material difference for any significant userspace
apps vs -march=i586,  if not why are you being so insistent and
damning of the current compatible behavior?

We know that -march=i686 will break compatibility for a small but
measurable number of users (and potentially a larger unmeasurable
number), do we know if we'll even get a measurable gain in exchange?


We would see much more substantial gains from things like -msse2 &
fpmath=sse but, unfortunately, unlike i586 there are a *lot* of
systems out there (and still being sold) which do not have all the
fancy instruction set extensions.

Repeating myself here… When it comes down to it, we already have a
"performance compiled" distribution: x86_64 which has every i686 knob
turned on and then some. If you care about whatever really minute gain
you'd get out of arch=i686, then you really should be looking for an
x86_64 system. (The higher end atoms are quite attractive I hear…)




More information about the devel mailing list