Re: -O2 vs. -O3

On Tuesday 22 April 2003 9:07 am, James Richard Tyrer wrote:
> I had this odd thought last week: does -O3 optimization really
> make an executable run faster on a system with a L3 cache?
>
> IIUC, -O3 optimization unrolls loops, but caching is based on the
> theory that the same code will be run more than once.

How large is L3 cache? I would imagine that unrolling loops would provide a 
speed improvement where the number of instructions performed per iteration is 
too large to fit in the cache, but slow things down where it isn't. Whether 
the former is a realistic scenario of course depends on whether L3 cache is a 
few bytes or a couple of K.

GCC seems to be quite good at knowing the quirks of different processors - I'd 
be surprised if specific optimisations such as unrolling loops remain enabled 
even in cases where they'd prove detrimental to performance. It's just a 
shame that GCC's interpretation of what constitutes valid C code changes 
seemingly at random with each release.

-- 
dan

"Were i susceptible to having my mind changed, it would have changed my mind" 
- Matthew Parris

Kde-cafe mailing list - [email protected]
http://ofb.biz/lists/listinfo.cgi/kde-cafe

DISCLAIMER: The views expressed on this mailinglist are the personal
opinions of the author and do not represent OfB.biz: Open for Business, KDE or the author's employer.