This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] Fix rawmemchr regression on bulldozer
- From: Liubov Dmitrieva <liubov dot dmitrieva at gmail dot com>
- To: Andreas Jaeger <aj at suse dot com>
- Cc: Ondřej Bílka <neleai at seznam dot cz>, GNU C Library <libc-alpha at sourceware dot org>
- Date: Tue, 27 Aug 2013 21:10:52 +0400
- Subject: Re: [PATCH] Fix rawmemchr regression on bulldozer
- Authentication-results: sourceware.org; auth=none
- References: <20130805122624 dot GA4682 at domone dot kolej dot mff dot cuni dot cz> <521CDBF1 dot 7020005 at suse dot com>
This patch is ok. We don't use SSE42 version for any Intel Processor,
so there is no changes here impacting Intel.
This is a good clean up.
--
Liubov Dmitrieva
On Tue, Aug 27, 2013 at 9:03 PM, Andreas Jaeger <aj@suse.com> wrote:
> On 08/05/2013 02:26 PM, OndÅej BÃlka wrote:
>> Hi,
>>
>> We now have a huge performance regression of rawmemchr on bulldozer.
>> We select __rawmemchr_sse42 which is about five times slower than
>> alternative __rawmemchr_sse2. Our benchtests can capture asymptotic
>> speed relatively well and their output on AMD FX(tm)-8150 is below.
>>
>>
>> On intel a gap between __rawmemchr_sse42 and __rawmemchr_sse2 is similar
>> but a bit_Prefer_PMINUB_for_stringop flag selects __rawmemchr_sse2.
>>
>> After this a bit_Prefer_PMINUB_for_stringop will no longer be needed
>> and can be removed by separate cleanup.
>>
>> OK to commit?
>>
>> simple_rawmemchr __rawmemchr_sse42 __rawmemchr_sse2
>> Length 32, alignment 0: 264 108 75
>> Length 64, alignment 1: 431 193 95
>> Length 32, alignment 0: 254 138 62
>> Length 64, alignment 1: 431 193 69
>> Length 64, alignment 0: 419 193 85
>> Length 64, alignment 2: 419 177 77
>> Length 64, alignment 0: 419 177 85
>> Length 64, alignment 2: 419 164 69
>> Length 128, alignment 0: 748 298 93
>> Length 64, alignment 3: 419 164 92
>> Length 128, alignment 0: 751 298 118
>> Length 64, alignment 3: 422 180 59
>> Length 256, alignment 0: 1407 628 131
>> Length 64, alignment 4: 419 192 72
>> Length 256, alignment 0: 1409 642 131
>> Length 64, alignment 4: 422 192 56
>> Length 512, alignment 0: 2723 1111 183
>> Length 64, alignment 5: 419 193 59
>> Length 512, alignment 0: 2725 1111 190
>> Length 64, alignment 5: 422 193 84
>> Length 1024, alignment 0: 5356 2026 417
>> Length 64, alignment 6: 419 192 84
>> Length 1024, alignment 0: 5359 2026 388
>>
>>
>>
>> * sysdeps/x86_64/multiarch/rawmemchr.S: Delete.
>
> This needs changes in ./sysdeps/x86_64/multiarch/ifunc-impl-list.c as well.
>
> Dmitrieva, could you comment on this patch, please?
>
> Andreas
> --
> Andreas Jaeger aj@{suse.com,opensuse.org} Twitter/Identica: jaegerandi
> SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 NÃrnberg, Germany
> GF: Jeff Hawn,Jennifer Guild,Felix ImendÃrffer,HRB16746 (AG NÃrnberg)
> GPG fingerprint = 93A3 365E CE47 B889 DF7F FED1 389A 563C C272 A126