Re: performance testing

Miguel Sousa Filipe wrote:
> 
> A good CFLAGS would be something not very agressive, something like:
> -march=<cpuType> -O3 or -O2 and at most -fomit-frame-pointer.
> (Scientific workloads can speedup considerably with: -ffast-math)

You must be careful on cpuflags. -ffast-math should never be used 
globally. Applications that benefit from it specify it themselves. The 
most realistic is -O2.

The remark about processors being optimized for i386 is a bit off beat. 
Actually what -march does is enable additional instructions to be used 
that are faster (and available) on those processors. Especially mmx{,2}, 
sse{,2,3} are very beneficial to certain kinds of applications, and can 
give big speedups. Considering that, however, many applications that are 
sensitive to this such as mplayer can do run-time cpu detection and use 
different code (often written in assembler) depending on the cpu at runtime.

In general it is not wise to use -O3. Most distributions compile with 
-O2. -O3 therefore is less well tested and may expose bugs in both the 
compiler and (more importantly) the applications.

Besides the instructions that are available to the compiler (influenced 
by -march) there is then the issue of instruction scheduling. To put it 
simply, all modern processors have the possibility to process multiple 
instructions in parallel when they do not conflict. The precise 
architectures of this differ and the ability to parallelize depends 
quite much on the scheduling of the instructions. This is influenced by 
the -mtune flag (set implicitly by -march). This is an optimization that 
is most likely also present in a binary distribution. Gentoo can however 
optimize to the precise processor used.

Paul

-- 
[email protected] mailing list