How to resolve -march=pentium4 stability issues.

gubbs · n00b Joined: 28 Feb 2003 Posts: 35

Hi, just before I start .. This was not my idea. Infodragon (rreich) came up with this after a week or so of pretty extensive testing on his new p4 3ghz box.

He's a busy chap so I'm posting on his behalf.

Section 1 - Instability

If you search around these forums for posts regarding gcc 3.2.2 and the -march=pentium4 flag you will find plenty of hear say and buffer overflows.

It is a well documented issue with gcc 3.2.2 that this flag does result in some adverse affects. Its a fact that if you compile your whole system in this way you will experience problems.

This is (according to us obviously) chiefly because of glibc being unstable with the -march=pentium4 flag. Infodragon discovered this through trial and error, he's a software engineer and so I trust him if he says he's compiled the whole system 50, 000 times on 5-6 sets of CFLAGS and narrowed it down to glibc and certain flag combos.

Section 2 - The Fix (as we see it)

One you have installed your system or are experiencing some issues with an already problematic install having used -march=pentium4 on everything you can sort things out without having to use -march=pentium3 (which we all agree as p4 users is way too much of a peformance compromise).

Emerge glibc with these flags CFLAGS="-march=pentium3 -mcpu=pentium4 -O2 -pipe -mmmx -msse -msse2 -mfpmath=sse -fforce-addr -fomit-frame-pointer -funroll-loops -frerun-loop-opt -frerun-cse-after-loop -falign-functions=4 -fexpensive-optimizations" in your make.conf.

Obviously these are pretty aggressive, they do work fine, but you will probably want to change them in some way. The only really important part is -march=pentium3 -mcpu=pentium4 that will result in a stable but mildly p4 optimized glibc. HOWEVER, we tried all sorts of CFLAGS combos and variations and ONLY reccommend those we have pasted here. If you use other sets of flags and start to burn, thats not US. Its you.

The rest of your system can now be compiled CFLAGS="-march=pentium4 -O2 -pipe -mmmx -msse -msse2 -mfpmath=sse -fforce-addr -fomit-frame-pointer -funroll-loops -frerun-loop-opt -frerun-cse-after-loop -falign-functions=4 -fexpensive-optimizations"
for maximum uber-geek p4 optimization.

Right now portage is a nasty fecker and wont let you compile glibc, then emerge world -eup without recompiling glibc with the new (and dogey for glibc) flags. So kerframil wrote this little script that you can run to rebuild everything EXCEPT glibc. Good eh? Just open up glibcscript.sh and paste in:

for p in `emerge -eup world | grep -v "glibc" | egrep -o " [a-z\-].+\/[^\-]+" | egrep -o "[^ ]+"`; do emerge ${p}; done

......... and save it.

Now you should consult your /var/cache/edb/world file to comment out anything you dont want to rebuild (i.e. stuff you know is already optimized to the hilt or just dont want to sit around waiting for - altho this is kinda not the point but KDE does take a looong time and you may be happy with -march=pentium3 peformance).

Now run the script as root and sit back and relax while your whole system gets recompiled against the SAFE and STABLE glibc for maximum (as it presently is capable) p4 peformance.

NOTE: A lot of 'important' ebuilds contain instructions to strip CFLAGS and go with a safe proven set ie. to swap -march=pentium4 for -march=pentium3 and/or to include debugging etc. Xfree ebuilds all do this. So you will still need to have a look into these ebuilds to make sure you are using the p4 opts! I personally edit the xfree ebuid to use -march=pentium4 get rid of debugging and to strip symbols the normal way and it works fine with the CFLAGS I posted above. Fine and FAST!

Cheers.

Gubbs.

PS: Infodragon and myself (Gubbs) have been running our systems compiled in this fashion for a couple of weeks now. So far we've not experienced any adverse problems and have thoroughly enjoyed having uber fast bins!

That said .. We're both nuts. So use the information above with your own discretion to avoid excessive flood.

rojaro · l33t Joined: 06 May 2002 Posts: 732

any reason for not modifying the glibc ebuild to contain the CFLAGS? that way you wont have to edit the world file and you also won't need the mentioned script ...
i also just read in the weekly news that there is a gcc 3.3 ebuild ... any experiences with that on the same topic?
_________________
A mathematician is a machine for turning coffee into theorems. ~ Alfred Renyi (*1921 - †1970)

jesterspet · Posted: Wed May 07, 2003 3:53 am Post subject:

I must pipe up and say that I can vouch for the glibc stability issue. I had posted before asking someone to verify my findings, but I think that request got lost in the mix.

I had originally compiled my system with:

floam · Posted: Wed May 07, 2003 4:19 am Post subject:

or switch to gcc 3.3

https://forums.gentoo.org/viewtopic.php?t=50679

kilimanjaro · n00b Joined: 29 Jan 2003 Posts: 37 Location: Jakarta

how do you set the CFLAGS when you emerge xfree so it doesnt default itself to march=pentium3 ? anyone ?

antipop · Tux's lil' helper Joined: 04 Oct 2002 Posts: 84

TeddY-T · n00b Joined: 08 May 2003 Posts: 10 Location: Australia

I have just started using the below use flags as you have specified.

Beetle B. · Guru Joined: 01 Mar 2003 Posts: 524

I tried the information in the first post and everything is running smoothly.

Some "important" ebuilds intentionally replace pentium4 with pentium3 for CFLAGS. I did a search to see what I could find and came up with only:

xfree-4.3.0-r2
mozilla-1.2.1-r5
recode-3.6

There were other ebuilds where pentium4 was replaced with i686 or something other than pentium3. I'm assuming that those were not related to the gcc bug (anyone know different?).

I compiled XFree 4.3 with pentium4 (commented out the replace line in the ebuild) and it works fine (note: I did this only today so I can't guarantee complete success - it may die later on).

I tried mozilla with pentium4 and it wouldn't install. I did -march=pentium3 -mcpu=pentium4 and it installed.

No idea what recode is...

Just thought I'd let you all know that Mozilla wouldn't compile with pentium4 (yes, I know it's been mentioned elsewhere, but I think most people read this thread to resolve the pentium4 problem...).

Hopefully, this will soon be a non-issue as it seems the latest gcc doesn't have this problem. I don't know if people have compiled Mozilla with it.
_________________
Beetle B.

Please update the table of equivalents.
A Firefox guide.

zojas · Posted: Tue May 20, 2003 11:45 pm Post subject:

I simply use this as my CFLAGS:

zojas · Posted: Tue May 20, 2003 11:49 pm Post subject:

here's the forum post where I read about the sse2 problems:

https://forums.gentoo.org/viewtopic.php?t=43373&highlight=sse2+pentium4

Chris Finch · Posted: Sun Jul 06, 2003 9:47 pm Post subject:

One question:

If glibc manages to make gcc produce wrong instructions when tuned for pentium4, how can you be so sure that none of the other programs in portage can? Until anyone is able to explain that, I won't use the sse2 flag (or anything that implies it) for any program.

zojas · Posted: Sun Jul 06, 2003 10:31 pm Post subject:

I turn off all sse2 in my CFLAGS now (-mno-sse2). and yes, it does cause problems with all kinds of things. I saw weird problems in gnumeric & python both, I'm sure there are others too.

all this convinces me never to buy a pentium 4 system. This machine is a work computer the company bought. go AMD! :wink:

_________________
http://www.desertsol.com/~kevin/ppc

Chris Finch · Posted: Mon Jul 07, 2003 5:29 am Post subject:

zojas: Yes, I was addressing those who suggest using -march=pentium4 for everything except glibc.

About the problem with python: I seem to remember that that problem occurs only if you compile glibc with -march=pentium4, regardless of how you compile python itself. I would not know of any example of pentium4 - breakage except for glibc. But that is not a proof that it never happens, either. Or does anyone know first hand that -march=pentium4 breaks other packages as well?

About AMD vs. Intel: I fail to see how a bug in gcc is an argument for either company.

ps:
while the gentoo forums definitely make a great part of gentoo's success, it can be a disadvantage sometimes when people try to help without having first hand experience. Particularly in the pentium4 threads I find it hard to distinguish those who have a solid backgroud from those who just repeat hearsay. This is a general observation, not targeted at anyone particularly. I do not exclude myself from this criticism, by the way.

Chris Finch · Posted: Mon Jul 07, 2003 12:04 pm Post subject:

zojas · Posted: Mon Jul 07, 2003 4:17 pm Post subject:

in the thread I posted above, you want to read the second post by vazagi. his post has the real info.

the gist is you need to hack the flag-o-matic.eclass file in /usr/portage/eclass directory to allow -mno-sse2. (just add it to the ALLOWED_FLAGS variable). then set your CFLAGS to

Garbz · Posted: Wed Jul 09, 2003 2:34 am Post subject:

I'd just like to say some of the experiences with the ever elusive stability with the pentium4 flag. Please note these are person experiences and by no means a professional oppinion! The flags represent what i have read, and my solutions are an act of desperation to get my system up and running.

When i installed my gentoo system on my home computer i have always been plagued with problems.

I started by leaving a very generous set of CFLAGS for bootstrap.sh I didn't do much at all in the way of optimisations and definatly not add the pentium4 flag.

After bootstrapping was done a quick gcc -v showed that i was running version 3.2.2.something, definatly not the version which was desirable considering all the problems. So a quick upgrade via emerge -u gcc. THis placed me with gcc-3.2.3 20030422. Apparently this version had none of the problems.
I then proceeded to install the system with these flags: