Re: [Mingw-users] Default cpu tuning for mingw gcc and binutils toolset

[Luke D. <cod...@ho...>]

>From: Wu Yongwei <ad...@ne...>
>To: min...@li...
>Subject: Re: [Mingw-users] Default cpu tuning for mingw gcc and binutils 
>toolset
>Date: Mon, 07 Apr 2003 17:32:18 +0800
>
>I tested two more programs, one is an implementation of memmem, the other 
>test of std::map<...>.  The former still shows "-O2 -mcpu=i586 -march=i586" 
>is better than "-O2 -mcpu=i686 -march=i686", and the latter does not show 
>significant performance difference.  So I have at hand two out of three 
>show i586 is better, and zero shows i686 is.
>
>No, I don't think it a "bug" in i686 optimization code.  Optimization is so 
>complex a job that on different platforms and using different applications 
>you are nearly sure to have different results.  And, after all, Celeron is 
>not a true Pentium Pro,

I thought that the only difference was the level 2 cache size, which 
shouldn't affect optimisation much, but I haven't checked the Intel 
documentation so I could be completely off.

>and Thinkpad hardly a good benchmarking platform.
>
>What I really want is some real test results to persuade me that i686 is in 
>general a better setting.  My test results are against it, but I won't 
>insist on my results if others have more satisfactory figures (better on 
>typical Pentium III boxes; I already know that pentium4 optimization really 
>does a great job on Pentium 4 boxes).
>
>Best regards,
>
>Wu Yongwei

If the test programs aren't too big could you post them here so that other 
people can try them on their hardware? We won't learn a great deal if we all 
use different code for benchmarking.


Luke

>
>--- Original Message from Danny Smith ---
>
>Yes I did notice your wording.  It started off as "One of my test
>programmes".  "One" of anything is hardly a representative sample unless 
>the population has only one element.
>
>However, if you think that your prime number example exposes a fault in the 
>i686 optimisation code, you should report it to gcc list.
>
>i686 probably gets more testing (across all ix86 platforms) than does i586. 
>i586 is no longer considered a "modern" cpu and so tends to receive little 
>interest when people look at long term development.  So in a sense, i686 is 
>the conservative choice (more in the mainstream of development).
>
>Danny


_________________________________________________________________
Hotmail now available on Australian mobile phones. Go to  
http://ninemsn.com.au/mobilecentral/hotmail_mobile.asp