Re: [PATCH] more loop.c fixes for -fprofile-loop-arrays

[Janis Johnson <janis187 at us dot ibm dot com>]

On Thu, May 16, 2002 at 06:42:57PM +0200, Jan Hubicka wrote:
> > On Wed, May 15, 2002 at 05:27:50PM -0700, Richard Henderson wrote:
> > > On Wed, May 15, 2002 at 02:46:38PM -0700, Janis Johnson wrote:
> > > >    -  For conditional code, skip the prefetch if PREFETCH_CONDITIONAL
> > > >       is zero (this was originally intended), but preserve the current
> > > >       behavior by defining PREFETCH_CONDITIONAL to one if it has not
> > > >       already been defined.
> > > 
> > > Is there some benefit to preserving the old behaviour?
> > 
> > On large benchmarks (SPEC int) it doesn't make much difference on ia64.
> > On small artificial benchmarks written just to check out this behavior,
> > prefetching even for accesses that are seldom executed helps on ia64.
> > For this and the other symbols controlling which heuristics are used
> > I'd like to do more experimenting so that the defines in loop.c are good
> > for the common case, and they can be overridden for each target (and
> > documented, of course).
> 
> While developing the patch I didn't make much testing about tunning the
> heuristics, unfortunately.  It is quite well possible, that the old behaviour
> is losing at 386 as well.  My approach has been to be conservative at start and
> try tunning later, unforutnately I didn't get into the second stage quite yet.
> 
> Definitly there are two (if my memory serves well) benchmarks in specs that
> lose due to -fprefetch-loop-arrays.  It may be interesting to investigate the
> purposes.  I hope to do so next month, after finals.

I've got some more changes in the works, and then plan to run SPEC
benchmarks with lots of different tweaks to the tuning parameters to see
what works best on ia64 and perhaps a couple of other platforms, and
take a closer look at the code generated for the benchmarks in order to
improve the heuristics in general.  We can coordinate our investigations.

Janis