This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: PCH merge breakage on PA
- From: Daniel Berlin <dberlin at dberlin dot org>
- To: Geoff Keating <geoffk at redhat dot com>
- Cc: janis187 at us dot ibm dot com, <law at redhat dot com>, <dave at hiauly1 dot hia dot nrc dot ca>,<gcc-patches at gcc dot gnu dot org>
- Date: Wed, 5 Jun 2002 22:46:36 -0400 (EDT)
- Subject: Re: PCH merge breakage on PA
On Wed, 5 Jun 2002, Geoff Keating wrote:
> > Date: Wed, 5 Jun 2002 12:47:40 -0700
> > From: Janis Johnson <janis187@us.ibm.com>
> > Cc: Geoff Keating <geoffk@redhat.com>, dave@hiauly1.hia.nrc.ca,
> > gcc-patches@gcc.gnu.org
> > Content-Disposition: inline
> > User-Agent: Mutt/1.2.5i
> >
> > On Wed, Jun 05, 2002 at 01:06:55PM -0600, law@redhat.com wrote:
> > > Seems to be working. I got a clean 3-stage of the PA port.
> > >
> > > Note it seems like the compiler just got a hell of a lot slower. I don't
> > > know if it's the PCH code, the tracer code or something else, but things
> > > sure feel like they went to hell recently.
> > >
> > > jeff
> > >
> >
> > I just started looking into why the compiler I built today on ia64-linux
> > takes 3-4 times as long to compile as the compiler I built yesterday.
> > The information from -ftime-report shows the same distribution across
> > passes. In both cases I did a simple make of C only starting with
> > GCC 3.0.4.
> >
> > Is it possible that additional checking got turned on by default?
>
> I did add some additional checking to the C++ compiler. It really
> shouldn't have cost that much, though. Could you try running a
> profile and seeing where the time is going?
>
> I am a bit worried about what's going on with performance. Even on
> x86, reports indicate a 10% slowdown. Since all of the GC marking
> takes less than 20% of the total time, that means that either the GC
> code got 50% slower, which didn't happen when I was testing (I was
> expecting less than 10% slowdown of the 10%, so about a 1% performance
> impact)---or something else was changed.
>
As I mentioned, something changes from after June 3rd, to now that caused
bootstrap times to go from 1800 to 1850 seconds (roughly, it's 49 seconds,
the exact numbers can be gotten from the spec95 pages, which tracks
bootstrap times now).
Jan's patch is what got it to 1800 (from 1770), not what got it to 1850.
It also does not increase the number of objects marked.
--Dan