This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch] Improve prefetch heuristics
- From: Christian Borntraeger <borntraeger at de dot ibm dot com>
- To: gcc-patches at gcc dot gnu dot org
- Cc: "Fang, Changpeng" <Changpeng dot Fang at amd dot com>, "rguenther at suse dot de" <rguenther at suse dot de>, "sebpop at gmail dot com" <sebpop at gmail dot com>
- Date: Thu, 29 Apr 2010 12:59:20 +0200
- Subject: Re: [patch] Improve prefetch heuristics
- References: <1C13CD442679CE45A2E80AE9251D7EF9220E4107@SAUSEXMBP01.amd.com>
Am Dienstag 27 April 2010 20:53:40 schrieb Fang, Changpeng:
> Hi,
> The attached patches enhanced the current gcc software prefetch heuristics.
> We have done some performance evaluations on cpu2006 on a 2.6GHz phenom
> System with ubuntu 9.10.
These patches look sane and seem to produce good code on s390 as well.
I am currently looking at tuning the current prefetch code for s390 and I had to override some of the parameters in our backend(e.g. our l1-cache-line-size is 256) to improve the behaviour of the prefetch code.
I dont think that I really have to change your new parameters for s390, but making them configurable like most of the other ones in tree-ssa-loop-prefetch.c might be a good idea.
So, what do you think about changing patch 6 and 7 so that
#define TRIP_COUNT_TO_AHEAD_RATIO 2
and
#define PREFETCH_MIN_LOOP_TRIP_COUNT 5
looks more like the other parameters in prefetch, e.g.
#ifndef PREFETCH_TRIP_COUNT_TO_AHEAD_RATIO
#define PREFETCH_TRIP_COUNT_TO_AHEAD_RATIO 2
#endif
#ifndef PREFETCH_MIN_LOOP_TRIP_COUNT
#define PREFETCH_MIN_LOOP_TRIP_COUNT 5
#endif
Christian