rguenth at tat.physik.uni-tuebingen.de
Fri Mar 7 21:19:59 UTC 2003
On Fri, 7 Mar 2003, Matt Austern wrote:
> On Friday, March 7, 2003, at 01:12 PM, Richard Guenther wrote:
> > Hi!
> > GCC is not very smart at inlining, and with not tuning -finline-limit
> > you
> > get (due to top-down inlining) trivial methods like FillLocStorageImpl
> > stuff and even ~Loc<1> not inlined which hurts performance a lot. GCC
> > as
> > of version 3.2 (I think) has the ability to mark functions as "always
> > inline". This unfortunately cannot be turned on by sth like #define
> > inline
> > __attibute__((always_inline)), but needs to be manually done.
> > Candidates
> > would be all Domain related stuff and possibly Mesh and Engine stuff as
> > well (yes - I get non-inlined Engine::read()s, too...).
> > Seems like a mess to me (note: intel icpc isnt any better here). I hope
> > we can sneak something like the -fobey-inline patch from Matt Austern
> > (see http://gcc.gnu.org/ml/gcc/2003-02/msg01666.html for discussion,
> > havent found the actual patch, Matt?) into gcc 3.3/3.4... Mark?
> The reason you haven't found the patch is that I didn't submit it.
> Stuart Hastings (also from Apple) submitted it. So far it's neither
> been approved nor rejected.
Ok, found it - its in
I'll try to apply it to 3.3 branch and report benchmark numbers if I
Richard Guenther <richard.guenther at uni-tuebingen.de>
More information about the pooma-dev