This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
effect of -fmove-all-movables and -freduce-all-givs
- From: Laszlo Ladanyi <ladanyi at us dot ibm dot com>
- To: gcc at gcc dot gnu dot org
- Date: Mon, 8 Jul 2002 23:52:36 -0400 (EDT)
- Subject: effect of -fmove-all-movables and -freduce-all-givs
Hello,
In the gcc docs I saw that you'd need info on when these optimizations hurt. I
decided to giev them a try on the freebench.org test set, and on two types of
benchmarks these option made a marked difference, once for good and once for
bad. Details are below. I hope this helps in your work (and thanks a lot for
gcc!!! :-)
--Laci
The new results are with the flags:
distray : Old score of 4.258511 better than current 3.872644
neural : Updated result to 5.973763 (old was 5.232598)
I'm using 'gcc version 3.1.1 20020606 (Debian prerelease)' on Linux, kernel
level 2.4.19-pre10-ac, The machine is an Athlon XP 2000+, using 512MB
PC2700 CAS2 memory. The rest of the optimization flags I used are:
COMMON =
COMMON += -O9
COMMON += -march=athlon-xp
COMMON += -mfpmath=387
COMMON += -m3dnow
COMMON += -foptimize-sibling-calls
COMMON += -fomit-frame-pointer
COMMON += -falign-loops=4
COMMON += -falign-functions=4
COMMON += -fexpensive-optimizations
COMMON += -fschedule-insns
COMMON += -fschedule-insns2
COMMON += -funroll-loops
COMMON += -fstrength-reduce
COMMON += -frerun-cse-after-loop
COMMON += -fprefetch-loop-arrays
COMMON += -fmove-all-movables
COMMON += -freduce-all-givs