Speed issues taken seriously?

[Jan Hubicka]

> On Wed, 15 Oct 2003, Lars Gullik Bj?nnes wrote:
> > a = g++ (GCC) 3.3.2 20031015 (prerelease)
> > b = g++ (GCC) 3.4 20031014 (experimental)
> >
> >             | user time  | file size | size dec | size text | size data |
> > ------------+------------+-----------+----------+-----------+-----------|
> > -O      | a | 19m44.730s |  4788930  | 3759350  |  3684558  |   19884   |
> >         | b | 22m45.080s |  5582322  | 4025379  |  3950263  |   19980   |
> > --------+---+------------+-----------+----------+-----------+-----------|
> > -O2     | a | 22m39.550s |  5052961  | 4024174  |  3949410  |   19884   |
> >         | b | 24m57.820s |  5084835  | 4160489  |  4085829  |   19860   |
> > --------+---+------------+-----------+----------+-----------+-----------|
> > -O3     | a | 23m07.770s |  5037109  | 3992125  |  3917397  |   19880   |
> >         | b | 26m30.820s |  5332135  | 4408351  |  4333699  |   19856   |
> > --------+---+------------+-----------+----------+-----------+-----------|
> 
> Hmm, that's bad.  C++ in GCC 3.4 shouldn't be slower, but faster, at least
> at -O2 and -O3 according to previous tests. Jan?

It seems to me that in this particular (i've missed the begining of
thread) case we are not hitting inlining heuristics problems in 3.3.2
and thus 3.4 is not better (the produced code sizes are mostly
comparable when taking into the account the increase of inlining
limits). What is the runtime performance?

3.4 seems to be overall somewhat slower than 3.3 again when optimizing
for sources where inlining does not make important difference.
I plan to do some oprofiling and analysis of SPEC data we have
accumulated during stage3 to figure out what are the main reasons for
this.  (I am experimenting with a tool automatically looking for changes
of performance in Andrea's SPEC testers and logging changelogs so we at
least can identify offending patches more easilly.  Very first results
are at http://www.ucw.cz/~hubicka/analysis.html
I don't have currently much time to improve my tool but I will try to do
so after hunting all the bugs in my TODO list.

Honza
> 
> Or file a PR with one preprocessed file that shows the slowdown? (Something
> like PR8361.)
> 
> > -O2 (1) | b | 24m25.590s |  4860852  | 3884278  |  3809882  |   19888   |
> > --------+---+------------+-----------+----------+-----------+-----------|
> > -O3 (2) | b | 25m45.220s |  5025193  | 4047663  |  3973255  |   19896   |
> > -------------------------------------------------------------------------
> >
> > (1) -O2 --param max-inline-insns-single=100
> > (2) -O3 --param max-inline-insns-single=100 --param max-inline-insns-auto=100
> 
> Gerald
> -- 
> Gerald Pfeifer (Jerry)   gerald@pfeifer.com   http://www.pfeifer.com/gerald/