Re: Memory requirement regressions

[Richard Guenther <rguenth at tat dot physik dot uni-tuebingen dot de>]

Steven Bosscher wrote:
> On Wednesday 07 July 2004 19:43, Richard Guenther wrote:
>
> > Steven Bosscher wrote:
> >
> > > On Wednesday 07 July 2004 17:48, Richard Guenther wrote:
> > >
> > > > The automated POOMA/tramp3d tester at
> > > > http://www.tat.physik.uni-tuebingen.de/~rguenth/gcc/monitor-summary.html
> > > > now tracks the maximum used virtual memory of cc1plus of mainline, too.
> > > > As a first "hint", today it revealed that mainline needs about 1GB of
> > > > ram compared to 3.4.1 which needs only 625MB.  This is on ia64 for the
> > > > tramp3d-v3.cpp testcase at -O2 -funroll-loops -ffast-math with the
> > > > leafify patch included.
> > >
> > > I get the attached timings and memory used on i686 (with memory use
> > > "measured" with 'top') on a machine with 3GB RAM, gcc configured with
> >
> > The results do not disagree with mine.  You need to do a comparison run
> > with 3.4.1, that may show a difference.  I tested on ia64 and that
> > obviously seems to use a lot more memory (btw. on a 16GB machine).

On ia32 (1GB ram) with -O2 it's 540MB (3.5.0 20040630) vs. 350MB 
(3.4.1).  Compared to the ia64 results this is a factor of 1.54 instead 
of 1.6 (so, the memory regression is comparable).  Your numbers (498MB 
vs. 372MB) are a factor of 1.34 apart.  I can only guess the difference
is either due to different mainline version, different amount of memory 
in the machine or not using leafify for 3.4.1 (though that would 
certainly make it "better" results).

Anyway - I'm now tracking memory-usage daily, so we'll see 
regressions/improvements.

Thanks,
Richard.