This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH v3] powerpc: strstr optimization
- From: OndÅej BÃlka <neleai at seznam dot cz>
- To: Carlos O'Donell <carlos at redhat dot com>
- Cc: Tulio Magno Quites Machado Filho <tuliom at linux dot vnet dot ibm dot com>, GNU C Library <libc-alpha at sourceware dot org>, Steve Munroe <sjmunroe at us dot ibm dot com>, Rajalakshmi Srinivasaraghavan <raji at linux dot vnet dot ibm dot com>
- Date: Sat, 25 Jul 2015 07:28:24 +0200
- Subject: Re: [PATCH v3] powerpc: strstr optimization
- Authentication-results: sourceware.org; auth=none
- References: <558A5642 dot 5020107 at linux dot vnet dot ibm dot com> <558A5761 dot 2000409 at linux dot vnet dot ibm dot com> <87oajpm8nc dot fsf at totoro dot br dot ibm dot com> <871tgijuri dot fsf at linux dot vnet dot ibm dot com> <55A6FE3F dot 6090701 at redhat dot com> <55A70B70 dot 6090607 at redhat dot com> <20150716195538 dot GA5140 at domone> <55A8110C dot 7000209 at redhat dot com>
On Thu, Jul 16, 2015 at 04:16:12PM -0400, Carlos O'Donell wrote:
> On 07/16/2015 03:55 PM, OndÅej BÃlka wrote:
> > On Wed, Jul 15, 2015 at 09:40:00PM -0400, Carlos O'Donell wrote:
> >> On 07/15/2015 08:43 PM, Carlos O'Donell wrote:
> >>>> May I proceed with this commit?
> >>>
> >>> Yes, please commit this for 2.22.
> >>
> >> For the record I trust IBM to make sure these patches make incremental
> >> improvements in performance even if they are not the best possible
> >> performance as pointed out by Ondrej Bilka.
> >>
> > Sorry Carlos, your trust is misplaced. This patch wasn't reviewed at
> > all. I did that as test how much we could test IBM to verify patches.
> >
> > I pointed out that it could have possibly quadratic behaviour which
> > still does. So please don't accept unreviewed patches next time.
>
> They showed cases for which the code does go faster and objectively
> so using the microbenchmark, and that's a win for now. Please continue
> to work with IBM to remove the quadratic worst case.
>
Carlos, that benchmark lack consensus. When I wrote it at thread it was
explicitly mentioned as counterexample why it shouldn't be used. Florian
raised objection that strstr isn't used on random inputs.
Also I don't think that its 'objective' benchmark at all, for example it
never measures case when we find string could happen relatively often
when user knows that string is there.