This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: [tree-ssa] POOMA compile time / memory requirement comparison
On Tue, 4 May 2004, Steven Bosscher wrote:
> On May 04, 2004 12:22 PM, Richard Guenther <rguenth@tat.physik.uni-tuebingen.de> wrote:
>
> > Here is more data for the merge-criteria of tree-ssa compared to mainline.
> > I compiled the tramp-v3.cpp testcase on a 1GB ram ia32 machine with
> > gcc-3.5 (GCC) 3.5.0 20040430 (experimental)
> > and
> > gcc-ssa (GCC) 3.5-tree-ssa 20040504 (merged 20040428)
> > with leafify attribute disabled and enabled.
> >
> > The summary is
> > mem user sys wall
> > 3.5 387MB 168.98 3.85 181.07
> > 3.5 w/leafify 452MB 266.71 4.37 284.90
> > ssa 493MB 226.71 5.35 245.18
> > ssa w/leafify 575MB 377.43 8.16 412.24
> >
> > So tree-ssa memory requirement is 127% of mainline, compile time
> > is 134% of mainline (without leafify) and 126% of mainline (with leafify).
>
> I still think this merge requirement is simply unreasonable. There are RTL optimizers
> that we know to be comile time consumers, and which are totally ineffective on
> tree-ssa. For example jump threading, big parts of GCSE, and big chunks of CSE.
> Disable those, and then see how tree-ssa compile and execution times compare
> to mainline.
Yes, I only wanted to provide some data. I think that the regressions
shown above are reasonable, but of course I expect these to be fixed until
3.5 release (as far as possible).
> I don't know enough about the RTL part of the middle end, but I do know that
> there are some things hidden behind flag_expensive_optimizations that increase
> memory usage, but have not a whole lot of effect either, with tree-ssa enabled.
> Obviously there's still work to be done to decrease the amount of memory that
> tree-ssa needs.
>
> How did you measure these memory requirements?
I had top running besides and remember the maximum virtual memory
requirement.
> > Details (from the first entries you see, mainline is better at optimizing gcc than tree-ssa):
>
> Where do you see that?
If you f.i. compare parser times (which should be identical):
parser 14.04 14.62 parser 17.63 17.79
but I just noticed that mainline was compiled with -march=athlon, while
tree-ssa without it. So I'll redo the summary measurement in a second.
Richard.
--
Richard Guenther <richard dot guenther at uni-tuebingen dot de>
WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/