On Sun, Nov 30, 2003 at 03:41:39PM -0500, Rakesh Patel wrote: > So the Fedora Core 1 Kernel for x86 uses NPTL but the glibc for x86 does > not? If you install glibc-2.3.2*.i386.rpm, then it does not use NPTL. This is the library which should be installed on all < i686 CPUs (where i686 in this context means i686 CPUs with CMOV instructions). But, on almost all current x86 boxes you want glibc-2.3.2*.i686.rpm installed (FC1 installer or up2date will certainly prefer this package if the hardware supports it) and that has NPTL support in it. Similarly it supports NPTL if you build a .athlon.rpm package (not included in the distribution, because the gains are not even measurable). > make any difference [using athlon-mp and -O4]. Just seems odd not to Why specifically -O4? -O3 or -O2147483647 do exactly the same. Also, it would surprise me if you see a significant difference with -march=athlon-mp vs. -march=i686 compiled glibc. Athlon's reorder the instructions themselves a lot and so are much less sensitive to instruction scheduling. Using -mfpmath=sse for glibc is a bad idea, since it could make the math library less accurate (given the assumptions some routines do). The most important instructions which actually matter for performance (CMOV*; worth on average something like 1%-2%) are already used by glibc-*.i686.rpm. And there are no Athlon optimized assembly string operations in glibc. This is something which would actually be worth doing, so if somebody is looking for a project he can try to do something and benchmark it. Jakub
