Re: [lmi] Benchmarking: gcc-8 beats gcc-10 soundly?

[Vadim Zeitlin]

On Sun, 20 Sep 2020 16:54:06 +0000 Greg Chicares <gchicares@sbcglobal.net> 
wrote:

GC> On 2020-09-20 12:21, Vadim Zeitlin wrote:
GC> > On Sun, 20 Sep 2020 09:59:31 +0000 Greg Chicares 
<gchicares@sbcglobal.net> wrote:
[...]
GC> >  Anyhow, I think it's still worth comparing the binaries produced with 8.3
GC> > and 10.2 under native MSW. If there is no difference there, we probably
GC> > should just stop there and chalk all this up to Wine. If there is still a
GC> > difference even there, we need to continue with investigating it as
GC> > discussed before.
GC> 
GC> We're in full agreement.

 We've built the fardels (not without running into some problems on the
way, but this will be in a separate message) and can confirm that there is
no significant difference between the binaries produced using 8.3 and 10.1
under (two different) native MSW systems. There still seems to be a slight
advantage to 8.3, but it's of order of 1-2% and the maximum I've seen is
less than 3%, while there are differences of order of 3% between different
runs of the same binaries, so I'm not even sure if it's significant. Even
if it is, it's definitely nowhere close to 34% in your initial results.
BTW, the absolute numbers are surprisingly close to the ones you get under
Wine 4: I get ~35ms for the 3rd scenario and ~10ms for the fourth one
(compared to 37/13 for you). Of course, we're testing under completely
different machines, but this test runs on a single CPU core only, so it
looks like the peak/turbo performance must be relatively close.

 Anyhow, I think there is no need to do anything else here, there is no
important performance regression in the code generated by gcc 10.1. As for
Wine, I really don't know what to do about it, the last time I tried
debugging Wine performance problem I spent a lot of time on it without
anything to show for it, unfortunately.

 Regards,
VZ

pgpThdiSgaMfS.pgp
Description: PGP signature