First of all , thanks for all your experiences ;-)
Forgive me , i didn't response to all your mail.
So let me do a summary , after collection all you mails:
I) generating CPU featured assemble takes longer time , also risks in
creating wrong code
II) customize compiler flag on OpenBSD doesn't values so much as what we
costs to achieve it.
Thanks again !
Ivan Nudzik wrote:
Aaron,
I built 3.7 base + OpenVPN for VPN gateway with custom
CFLAGS/LDFLAGS and it worked well for years, even with such aggressive
flags as CFLAGS="-O3 -march=c3 -mmmx -m3dnow -fomit-frame-pointer" and
LDFLAGS="-Wl,-z,combreloc -Wl,-O2 -Wl,--sort-common
-Wl,--enable-new-dtags". But build was broken and I must to fix many
things in OpenBSD makefiles, to get binaries done. Such LDFLAGS
significantly speedups application loading, cause it's doing the same
things as prelink utility, but it link time. CFLAGS helps me speedup
C3 500MHz router device as about 20-30%, but it was synthetic
measurement on nbench built with and without CLFAGS. Kernel build was
also forced for such CFLAGS and no stability issue and subjectively
more speedy.
I've read about that , e.g kernel avoid intel MMX instructions , cause
it may cause interrupts (handle lots of register together to combine a
bigger register)
It was few years ago, when I've had mood to compile everything
with all CPU features on. ;-)
Yeah , i'm so interesting with CPU features on my Linux right now , like
what you did a few years ago ;-)
Looking at this concrete OpenBSD build my summary is that it took me
more than 20h to build it and speedups was not such significant,
comparing that it should took an about 1h to do the same functionality
with release binaries only bit slow. And two other things to keep on
mind: nowadays x86 CPUs has very heuristic instruction
handling/prefetching so it will not speedup as much with custom
CFLAGS; OpenBSD still use gcc 3.3.5 which can't do as good
optimization for new CPUs as to loose time with custom flags build.
And many makefiles in OpenBSD base don't handle CFLAGS/LDFLAGS
correctly. The last time I tried it (about 1/2y. ago), I made only
kernel with custom flags (for kernel it is COPTS variable if I
remember right). Rest of system didn't build and I didn't had a time
to bother with broken makefiles.
So if you don't want to do HPC, don't loose time with custom
flags. I if you want to do HPC, choose other operating system, cause
OpenBSD strengths are other.
I take your suggestion , ain't gonna waste time playing gcc's flags.
--
Best Regards,
Aaron Lewis - PGP: 0x4A6D32A0
FingerPrint EA63 26B2 6C52 72EA A4A5 EB6B BDFE 35B0 4A6D 32A0
irc: A4r0n on freenode
