On Tue, 2003-06-10 at 08:59, Christos E. Chrisostomidis wrote: > Dear all, > > Most of the RPM packages shipped with RH9 (and older), come for a i386 > target. > > I am having a pentium 4 box and I am currently using rpmbiuld > --rebuild --target i686 to rebuild > the SRPMs in order to optimize performance. > > I am trying to optimize packages such as sendmail, samba, squid etc. First of all before you go into a frenzy of package rebuilding how about MEASURING the performance of a CPU intensive benchmark both using RedHat default parms (who BTW target the i686 for optim purposes) and what you get when compiler is using the same parms than an rpm built with --target i686? And once you have done that what about EVALUATING if it is worth the trouble? In case you still go for it there are a couple of caveats: 1) Gcc 3.2 does not optimize for the Pentium 4. It only knows about the i686 family (PPro to PIII) who is very different for optimization purposes. In fact you could try to play with other -mcpu= parms since it could happen than another CPU has a timetable closer to the PIV than the PIII family. It _could_ happen that your P4 were faster with say, -mcpu=i386 or -mcpu=athlon than with -mcpu=i686. 2) Whenever the program enters glibc or kernel you will be gaining nothing since RedHat already ships optimized RPMs for these. 3) RedHat does not invoke gcc with parms for MMX/SSE instructions since the i686 packages are supposed to run on PPro. I expect them giving a real boost on some programs. However I also expect you run into some interesting bugs since this is relatively uncharted territory. But since the only way to debug GCC's MMX/SSE code generation is for people using it I strongly encorage you to boldly go where no geek has ventured before. :-) > > I am wondering if I need to specify any gcc/g++ compiler optimization > flags, perhaps in the form of an rpmrc file. > > If this is the case, does anybody knows a good reference to look at so > I can find the appropriate optflags > to use ? > > > Best Regards > Christos > > -- Jean Francois Martinez <jfm512@xxxxxxx>
