This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] faster strlen implementation
- From: Dmitrieva Liubov <liubov dot dmitrieva at gmail dot com>
- To: neleai at seznam dot cz, libc-alpha at sourceware dot org
- Date: Wed, 11 Jul 2012 17:56:44 +0400
- Subject: Re: [PATCH] faster strlen implementation
> How important today is handle SLOW_BSF? Google search showed
> that K8 had slow bsf.
At the moment all versions marked as "for SLOW_BSF" are using just for
atom. But I care about Atom performance and if you don't have atom
machine I can help with regression testing.
> I also added sligthly faster version that uses sse4.1 instruction ptest instead pmovmskb
> testl pair.
Ptest has almost the same performance as pmovmskb + test
(Core i7:
ptest - latency 2, throughput 1
pmovmskb - latency 1, throughput 1
test - latency 1, throughput 0.33)
It seems to me better to remove IFUNC wrapper at all and use just sse2
version if there are no regressions for atom without SLOW_BSF version
because we can eliminate IFUNC wrapper overhead.
--
Liubov Dmitrieva
Intel Corporation