[Patch] Speedup locale::operator==

jlquinn@optonline.net jlquinn@optonline.net
Wed Apr 14 18:37:00 GMT 2004


I was going to ask why construct the name string before I went and really read the underlying code.

Upon further reflection, the real question is why the compiler can't hoist the first test above the string constructor.  Are we required to run it by the language, or is it just a limitation in the current optimizer?  And if so, will tree-ssa be able to optimize it?

BTW, nice little improvement :-)

Jerry Quinn

----- Original Message -----
From: Paolo Carlini <pcarlini@suse.de>
Date: Wednesday, April 14, 2004 12:47 pm
Subject: [Patch] Speedup locale::operator==

> Hi,
> 
> I noticed some time ago that we (*) weren't exploting the ref-counted
> nature of locale in operator==.
> 
> For this stupid testcase (P4-2400, -O2):
> 
>  std::locale loc1("fr_FR");
>  std::locale loc2 = loc1;
>  bool test = false;
>  for (int i = 0; i < 10000000; ++i)
>    if (loc2 == loc1)
>      test = true;
> 
> current mainline
> ----------------
> 5.190u 0.000s 0:05.21 99.6%     0+0k 0+0io 245pf+0w
> 
> patched
> -------
> 0.150u 0.000s 0:00.15 100.0%    0+0k 0+0io 242pf+0w
> 
> Regtested x86-linux, will commit later.
> 
> Paolo.
> 
> (*) At variance with our famous competitor ;) ...
> 
> //////////////
> 
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: patch_locale
URL: <http://gcc.gnu.org/pipermail/libstdc++/attachments/20040414/37a27e82/attachment.ksh>


More information about the Libstdc++ mailing list