This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Faster strchr implementation.


I have some comments on the results getting from here:

http://kam.mff.cuni.cz/~ondra/benchmark_string/strcmp_profile.html
http://kam.mff.cuni.cz/~ondra/benchmark_string/strchr_profile.html
http://kam.mff.cuni.cz/~ondra/benchmark_string/strrchr_profile.html

* strrchr: As far as I understood you are suggesting to make total
clean up for strrchr removing all IFUNCs and use your new version
everywhere.
I see _no_bsf version is still better for atom for some cases. Are you
sure that cases where your version is better are more important for
atom users?
http://kam.mff.cuni.cz/~ondra/benchmark_string/atom/strrchr_profile/results_rand_L3/result.html
You links to documentation don't work and I can't to get more details
about the tests until reading the test codes.
http://kam.mff.cuni.cz/~ondra/benchmark_string/atom/strrchr_profile/results_rand_L3/doc/properties.html

* strchr: Are you suggesting to set ON slow_BSF flag for Silvermont
and use IFUNC for switching between slow_BSF and the new version,
remove sse42 version from the glibc sources?
With this, I have a concern with Haswell degradation showing by the plots:
http://kam.mff.cuni.cz/~ondra/benchmark_string/haswell/strchr_profile/results_rand_noicache/result.html
And switching slow_BSF flag on for Silvermont will impact other
functions, how are you going to resolve the dispatching here?

* strcmp: I can't agree that the new version is always better for atom:
http://kam.mff.cuni.cz/~ondra/benchmark_string/atom/strcmp_profile/results_rand_L3/result.html
http://kam.mff.cuni.cz/~ondra/benchmark_string/atom/strcmp_profile/results_rand_L2/result.html
If you are going to make total clean up and replace everything with
the new version, this probably can introduce some performance
regressions for atom customers. It is ok to replace for Haswell and
Silvermont but we can't harm atom performance.

--
Liubov


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]