gcc 3.2 slows down athlon?

[samwichse]

Ok, I know little about compilation, but I was still running the old core I compiled for ecc2 under redhat 7.3 (gcc 2.96) with the defaults (I don't know how to change them or what to change them to). It was getting about 930kit/s.

So I just realized, hey, gcc 3.x is supposed to compile better for athlons, I recompiled it (still with defaults) and now I get 665 kit/s.

HELP! (note, I failed to back up the old core). How do I optimize this better???

Sam

[pbharris]

hi sam,
what flags did you pass in? gcc3x slowed down pure C a bit but sped up C++ progies, in either case dropping that much is much much worse than can be accounted for by going from 2.96 to 3.2

[samwichse]

Um... flags?

[pbharris]

i use these:

export CFLAGS=$'-O2 -march=athlon'
export CXXFLAGS=$'-O2 -march=athlon'

w/o out them your compiler will used default glibc i think, which is i386 under rh

[samwichse]

So you just type the exports before compiling?

Where does this add/change things?

Is there a way to edit this directly with vim/kate/gedit or something?

Sam

make -C ./src client
make[1]: Entering directory `/home/samwich/Desktop/ecc2/src'
nasm -felf mult109_mmx.asm
gcc -Wall -Wno-unused -O3 -funroll-loops -fomit-frame-pointer -fexpensive-optimizations -I../include -o client client.c ecc2-109.c sq109_mmx.c mult109_mmx.o
client.c:41:15: warning: multi-line string literals are deprecated
mv client ..
make[1]: Leaving directory `/home/samwich/Desktop/ecc2/src'

That's what I get when I try it.

EDIT: it's also just as slow.

[BFlurie]

Put/edit these lines in /etc/profile

export CFLAGS="02 -march=athlon"
export CXXFLAGS=$CFLAGS

then further down the file, make sure those 2 from above are in the gereral export line:
export CFLAGS CXXFLAGS PATH MANPATH........etc

gotta reboot then. You can use more flags than the above, but those'll give you most of the benefits.

[samwichse]

Ok, after adding the code, this is what it looks like:
But, it didn't seem to change anything when I did a make on the core. The file size of the executible was identical, and the speed is the same.

# /etc/profile

# System wide environment and startup programs, for login setup
# Functions and aliases go in /etc/bashrc

export CFLAGS="02 -march=athlon"
export CXXFLAGS=$CFLAGS

pathmunge () {
	if ! echo $PATH | /bin/egrep -q "(^|:)$1($|:)" ; then
	   if [ "$2" = "after" ] ; then
	      PATH=$PATH:$1
	   else
	      PATH=$1:$PATH
	   fi
	fi
}

# Path manipulation
if [ `id -u` = 0 ]; then
	pathmunge /sbin
	pathmunge /usr/sbin
	pathmunge /usr/local/sbin
fi

pathmunge /usr/X11R6/bin after

unset pathmunge

# No core files by default
ulimit -S -c 0 > /dev/null 2>&1

USER="`id -un`"
LOGNAME=$USER
MAIL="/var/spool/mail/$USER"

HOSTNAME=`/bin/hostname`
HISTSIZE=1000

if [ -z "$INPUTRC" -a ! -f "$HOME/.inputrc" ]; then
    INPUTRC=/etc/inputrc
fi

export CFLAGS CXXFLAGS PATH USER LOGNAME MAIL HOSTNAME HISTSIZE INPUTRC

for i in /etc/profile.d/*.sh ; do
    if [ -r "$i" ]; then
    	. $i
    fi
done

unset i

[BFlurie]

Sorry, I didn't realize it was the kernel. Add -march=athlon -malign-functions=4 -mpreferred-stack-boundary=2

to the file /usr/src/linux..../Makefile, on this line near the top:

HOSTCFLAGS = -Wall -Wstrict-prototypes.....

[samwichse]

No, it's not the kernel, it's the ecc2-109 core I'm trying to compile here.

Can I add those options to /etc/profile ?

[pbharris]

i just type them when i need to, if ya want them everywhere you can add them there. what were the orginal ones made with?

[samwichse]

The original core I was using was made with all the defaults in RedHat 7.3 with gcc 2.96, it got 930-940k it/s.

But when I compile now (even after adding all those to the profile, or just doing the "export" thing... I still get the same problem... about 665k it/s.

Any other ideas?

[BFlurie]

While compiling, are you seeing the:
-02 -march=athlon -mpreferred-stack-boundary=2......
in the compilation lines as they pass by? Also, -fomit-frame-pointer should appear by default for a -02 level or above compilation.

Also, do you "make clean" before recompiling to start fresh?

[samwichse]

No, I'm seeing this exactly (this is a very small program):

make -C ./src client
make[1]: Entering directory `/home/samwich/Desktop/ecc2/src'
nasm -felf mult109_mmx.asm
gcc -Wall -Wno-unused -O3 -funroll-loops -fomit-frame-pointer -fexpensive-optimizations -I../include -o client client.c ecc2-109.c sq109_mmx.c mult109_mmx.o
client.c:41:15: warning: multi-line string literals are deprecated
mv client ..
make[1]: Leaving directory `/home/samwich/Desktop/ecc2/src'

And no, no "make clean" I'm just deleting the whole source and re-unpacking it.

[BFlurie]

Then for some reason GCC isn't importing your profile settings. Compiling as root?

Try adding the flags:

-march=athlon -malign-functions=4 -mpreferred-stack-boundary=2

to the appropriate line in the install's makefile before running configure. Check it again before running "make" to be sure it didn't change.

Personally, the -fexpensive-optimizations gains very little -- I'd drop that. I've also heard a fair bit of warning on using GCC 3xx, tho the newest stuff prb'ly requires it.

[strangerstill]

Yep - I've tried various optimisations and have comprehensively failed to push it over 650 kIt/s (gcc 3.2, Athlon XP 1800) whereas the gcc 2.96 core gets 880 kIt/s.

Get a gcc 2.96 compiled core from Sven Geier's page - http://www.its.caltech.edu/~sgeier/code/ecc2.html

(Sven is in Ars Technica Team Vodka Martini, and has written a couple of cool shells for ecc2)

[samwichse]

Oooooh... so this isn't just my stupidity perhaps?

Oh well, downloading the precompiled one put it back up to about 910k it/s on a 1700+ (running at 2000+), so I guess the simplest solution is the best in this case.

Thanks to everyone for their help though, I learned something anyway

Sam

[Scott Tiger]

Yeah, but I can't help but be disappointed here to know that the latest version of GCC puts out slower code than the one that came before it.. Or is this just an anomoly? (I know very little of compilers and there affects on performance).