This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Please try out new inlining heuristics
- From: Jan Hubicka <hubicka at ucw dot cz>
- To: "S. Bosscher" <S dot Bosscher at student dot tudelft dot nl>
- Cc: 'Jan Hubicka ' <hubicka at ucw dot cz>,'Gerald Pfeifer ' <pfeifer at dbai dot tuwien dot ac dot at>,"'gcc at gcc dot gnu dot org '" <gcc at gcc dot gnu dot org>
- Date: Mon, 30 Jun 2003 00:27:54 +0200
- Subject: Re: Please try out new inlining heuristics
- References: <4195D82C2DB1D211B9910008C7C9B06F01F3735B@lr0nt3.lr.tudelft.nl>
> > > > I now also have new # of instructions approximation scheme, so hope
> > > > to get scores better now.
> > >
> > > That sounds exciting.
> > >
> > And it seems to work somehow... I've stripped down your testsuite so it
> > runs only 18 minutes and tested all three heuristics I've implemented.
> > Interestingly enought the first one is still a winner.
>
> Do you have a comparison of the "accuracy" of the old and your new # of
> insns approximation? Have you tried -O2 without unit-at-a-time but with the
> new approximation?
>
> > The compile time savings are not that high tought. Instead compilation
> > time comes form 8:30 to 7:20 but I hope I will be able to tune them a
> > bit better.
> >
> > I will send you what I have once tests finish.
>
> Can you just send it to the list please? I'm sure everybody's interested
> ;-)
OK, I split it out from the rest of my patch and send it to patches
list.
In the case you do have machinery to dump the differences in between
esitmates and sizes, it would be nice to try it and figure out functions
where the ratio is off by largest amount and improve the function
itself.
I plan to do something like this in future, but I don't have time for
that this week as I am leaving to US at friday.
Honza
>
> Gr.
> Steven