Experiment: using -O3 in the toolchain

 Last month, I was querying the use of -O3 in glibc with x86_64 on
lfs-dev : turned out my problems on that one machine are down to
buying cheap consumer-grade hardware (it works, mostly) :)  But I
then got to thinking about using -O3 for the rest of the toolchain.
I've now completed test builds (CFLAGS and CXXFLAGS = -O2 everywhere,
and with -O3 in chapter 5 pass 2 and chapter 6 binutils, gmp, mpc,
mpfr,gcc).  Note that some packages, particularly in BLFS, ignore
CFLAGS.  The toolchain testsuite results are no different (same
failures in gcc, and with static libs which I do not install).

 This is x86_64 on an i3 (hyperthreaded dual processor, the kernel
thinks it is 4 processors) with 4GB of RAM.  I built these in
directories because they aren't intended to be used, and I didn't
build the kernel (that would be another 4 or 5 minutes!).

 Apart from that, I built everything that is currently part of my
normal desktop, and using -j3 for builds (but not installs!) except
where that gives problems.  I try to reserve some cpu so that cron
jobs, particularly backups, can run, or  so that I can browse or
listen to music if I wish to, in both cases without impacting the
build time or my listening/browsing experience.

 It's nice to now have hardware that is capable of building with -O3
on current gcc (my previous single processor machines were too slow
and too lacking in RAM), but I don't think I'll bother again.

 I used my standard scripts, so timing for each package is from
configure to the end of the install, to the nearest second.  The
times for the total build were 6 hours and 6 hours 3 minutes with
the -O3 toolchain.  The -O3 build used slightly more disk space in
/usr/bin and /usr/lib, but the difference was minimal.

 Obviously, building the toolchain with -O3 took a bit longer (20
seconds in LFS chapter 5, a bit over 3 minutes 12 sec in chapter
6), but then I saw *minor* gains in many BLFS packages, at least
until I got to QT (1294 seconds in my configuration, instead of 1256
using -O2).  I also lost half a minute in building the gnome packages
I use, but the gnome script takes about 90 minutes, mostly in webkit
with -j1, so the differences are immaterial and lost in the noise.

 If I was to do it again, I wouldn't be suprised to see a bigger
variation from doing identical builds - this isn't a hard realtime
OS.

 So, I wondered how using -O3 in the toolchain would affect the
build, and now I know.  Back when I first arrived here,
optimisations were in vogue among some people to speed up our
compiles.  I think I can happily avoid investigating optimisations
again. :)

ĸen
-- 
das eine Mal als Tragödie, das andere Mal als Farce
-- 
http://linuxfromscratch.org/mailman/listinfo/lfs-chat
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page