Am 07.11.2015 um 20:36 schrieb drago01:
On Sat, Nov 7, 2015 at 7:40 PM, Reindl Harald
<h.reindl(a)thelounge.net> wrote:
> the point is compile a single application with new features won#t gain that
> muc 8until you do the same with most libraries used by the software) but
> having the whole distribution is a summary with a completly different
> behavior than a single test of software xyz
Uh no you just have to compile the software any libraries used by the
hot paths of the software you are trying to test.
for which one?
that don't show the impact on a complete distribution anyways
> if there would be no difference kernel upstream won't invest
that much time
> for runtime-cpu-detection (look at the bootlog on different hardware)
>
> here are examples
http://www.phoronix.com/scan.php?page=news_item&px=MTU0MTY
>
http://www.phoronix.com/scan.php?page=news_item&px=MTI1Njc
You missed the point completely.
1. AVX != (S)SSE3
i know that
There are lots of workloads that would benefit from AVX but SSE3
doesn't add that much (enabling SSE2 on i686 would gain you more for
instance).
enable SSE3 don't diable SSE2
so "gain you more" is impossible
2. Runtime detection does not have the cost of dropping support for
specific hardware
runtime detection needs to be implemented in every relevant software at
it's own and hence has a *high cost*
add -msse3 to the default falgs have no costs at all but only an impact
of a unknown amount f *very* outdated hardware which is unlikely running
a bleeding edge distribution
frankly whatever somebody has run on a 10 years old machine can be
easily virtualized and i doubt that many people have a 10 years old
computer as their only device, as far there is something with a core2 or
newer in the house you can virtualize the other machine and save a lot
of energy