Making full use of cpu registers in CFLAGS

ghetto · Guru Joined: 10 Jul 2002 Posts: 369 Location: BC, Canada

Ok thanks, thats good to know. Its pretty painful to recompile an entire system on a 1ghz cpu. :evil:

_________________
Blizzard you suck.

taskara · Posted: Mon May 05, 2003 10:08 pm Post subject:

thanks gnufsh

I think I read on the other post that someone showed how adding -mmmx and -msse etc actually caused them to become -mnommx and -mnosse because they were already enabled in -march=athlon-xp.

so leaving them in won't hurt.. sweet as. ta..
_________________
Kororaa install method - have Gentoo up and running quickly and easily, fully automated with an installer!

Gnufsh · Guru Joined: 28 Dec 2002 Posts: 400 Location: Portland, OR

-mno-mmx, -mno-sse, and -mno-3dnow do show up on the "options enabled" part, but the macros that actually impliment the mmx, sse, and 3dnow support are still enabled. And, sse, mmx, and 3dnow are still enabled.

taskara · Posted: Tue May 06, 2003 10:37 pm Post subject:

hmm ok..

so adding -mmmx, etc to -march=athlon-xp makes them look disabled, but they are in fact enabled still ?

crazy

we need to get someone from gnu to clarify optimum settings!

_________________
Kororaa install method - have Gentoo up and running quickly and easily, fully automated with an installer!

wrc1944 · Posted: Wed May 07, 2003 12:43 am Post subject:

I agree- if some clarification from THE definitive source was forthcoming, that would be great, although it's hard to recognize what that source is. I guess this is what makes Linux interesting, but it would be nice if what seemingly was a straightforward procedure wasn't so ambiguous. Realizing different hardware is involved in each case, it would seem one could find out exactly what gcc optimizations actually do, or how they interact- but that knowledge is so far, elusive. Maybe it's just that nobody really knows, or the gcc manual has a few errors itself, and leads us astray? We can get conflicting reports, all from sources who obviously know more than we do, but what are we to make of it?

As one who has spent much time trying to optimize to the max, I now realize I don't know what to take as gospel, though I surely appreciate all the help I've gotten. At this point, I'm realizing I simply have to educate myself, and then do it myself, and see what happens. Apparently, when you're on the edge with your own hardware, there is no other way.

wrc1944
_________________
Main box- AsRock x370 Gaming K4
Ryzen 7 3700x, 3.6GHz, 16GB GSkill Flare DDR4 3200mhz
Samsung SATA 1000GB, Radeon HD R7 350 2GB DDR5
OpenRC Gentoo ~amd64 plasma, glibc-2.36-r7, gcc-13.2.1_p20230304
kernel-6.8.4 USE=experimental python3_11

drdabbles · n00b Joined: 08 May 2003 Posts: 31 Location: NH, USA

Optimizations are like...well...I can't think of anything they are like. I can only say that optimizations are like recursion. "To understand recursion, you must first understand recursion".

One suggestion I would give to anyone with about 10 minutes of ree time, would be to grab the GCC source from GNU, unpack it, and grep the resulting directory for "i586", or some other optimization flag. Heck, you could even grep for "march" if you wanted to.

This will allow you to find out what file contains the optimization interpretation code. You can then open it in your favorite viewer/editor, and peruse to your hearts content. It is pretty straight-forward. You will be able to see what flags are enabled for which CPUs, etc.

With regard to getting anything from developers...most of them are just as confused as we are. There is no clear, concise documentation that discusses the effects of optimization flags...anywhere. There are people's opinions, and people that look at the macros, etc. I choose to believe the latter of the two, but again, they are just opinions as to the effects.

Also, I would like to say that it might be difficult to get an official response from the developers that work(ed) on the GCC compiler. They are an open source community, and many times it is difficult to pin down the specific person or group responsible for a particular segment of code. However, perhaps someone could try anyway.

Good luck all,
Thomas Cameron
CEI Systems, Inc.

taskara · Posted: Thu May 08, 2003 3:59 am Post subject:

this is a pretty simple issue tho, surely someone who knows programming can test it and work it out ?

we aren't asking what optimisations do, we simply want to know if "-march=athlon-xp" enables "-mmmx, -m3dnow, and -msse" by default.

if it doesn't then we need to put them in our CFLAGS.

if it does, then we don't need to put them in our CFLAGS.

some people have said, put them in anyway. but others have said if you do that and -march DOES automatically put them in, then you will actually DISable them ...

it's all too confusing a situation, but surely there is a clear answer, either -march=athlon-xp DOES autmatically incorporate those flags or it doesn't.

surely someone knows that?
_________________
Kororaa install method - have Gentoo up and running quickly and easily, fully automated with an installer!

ghetto · Guru Joined: 10 Jul 2002 Posts: 369 Location: BC, Canada

/me goes and writes stallman a letter..

Dear Mr Stallman,
I know your a pretty busy guy with your.. uh.. hmm.. well whatever it is that
your doing, but could you please explain to us how gcc works?
Particularly I would like to know how I can make gcc read my email.. oh wait.. thats Emacs, well then, uh.. could you please explain how to make gcc do proper optimizations? Its very important to me that I run the absolutly fastest binary code imaginable and if I have even the slightest doubt that my code is not completely optimized I break out in a horrid rash. Im sure you understand where Im coming from.
Any help would be appreciated.

Sincerely
Gnu/Ghetto
_________________
Blizzard you suck.

taskara · Posted: Thu May 08, 2003 6:47 am Post subject:

ghetto · Guru Joined: 10 Jul 2002 Posts: 369 Location: BC, Canada

ok so maybe i was a touch on the sarcastic side with that last comment.

i was only having fun, I honestly think this is pretty important as well.. but I just couldnt help but try to joke because in a way its kind of funny and my little imaginary email is true, some people really do break out in rashes if they think their code might not be optimized to the absolute degree..

_________________
Blizzard you suck.

taskara · Posted: Thu May 08, 2003 8:15 am Post subject:

gr00vy man... I didn't wanna come downl on you hash

seriously tho.. an answer to the problem woudl be cool!

hehe
_________________
Kororaa install method - have Gentoo up and running quickly and easily, fully automated with an installer!

ghetto · Guru Joined: 10 Jul 2002 Posts: 369 Location: BC, Canada

Why dont we send some real email then?
_________________
Blizzard you suck.

taskara · Posted: Thu May 08, 2003 11:00 am Post subject:

cause don't wanna bother them

they have more important things to do ehe

_________________
Kororaa install method - have Gentoo up and running quickly and easily, fully automated with an installer!

nico-- · n00b Joined: 29 Jul 2002 Posts: 59

m00dawg · Posted: Fri May 09, 2003 8:10 pm Post subject: mfpmath results

I tried -mfpmath=sse,387 and then ran Pov-Ray. The results showed no real change in the results - two of the three results were even slightly slower. I don't know how this might impact other benchmarks, but for Pov-Ray it seems to make no difference. This is unfortuante since you would think the additional registers would help - if nothing else, temporary data could be placed there.

m00dawg · Posted: Fri May 09, 2003 8:21 pm Post subject: A solution?

Perhaps it would be a good idea if someone were to post up some benchmarks using different flags? I have already commented on Pov-Ray benchmarks vs using mfpmath=sse,387, but it would be interesting to see others.

An easy benchmark to run is timing a kernel build. It easy, reasonably fast, and the kernel is already there for you to play with

_Edulix · n00b Joined: 09 Dec 2002 Posts: 68

Hi all!

I've read the whole thread, and I have somethings to say.

There' some people that uses very large CFLAGS, but don't know really what does their
options really do. I have been one of them yesterday when compiled my new gentoo system, but now I'm going to change my flags for some reasons.

Thanks you defconfoo, you have been very helper for me! 'Ive read in http://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html the functions which are not included by -Ox, and there some of them about we must read before adding them to our CFLAGS:

wrc1944 · Posted: Tue Jun 10, 2003 10:52 am Post subject:

After reading man gcc many times, and going over all the info on the Gentoo forum's Cflags Central thread (about 20 pages, and very informative), and every other forum and info source I could find for months, I finally settled on what I thought was the best set of optflags for athlon-xp platforms. Here they are:

optflags: athlon -O3 -fomit-frame-pointer -pipe -march=athlon-xp -mmmx -msse -m3dnow -falign-functions=16 -falign-labels=1 -falign-loops=16 -falign-jumps=16 -fprefetch-loop-arrays -mfpmath=sse,387 -ffast-math -fforce-addr

I won't go into details about why I included, or removed specific flags here. I did try these on XFree86, and they built and installed fine, with only a few warnings and no errors, and after five days, no apparent problems have surfaced, and fine performance. Curiously, the XFree86 compile dropped the -ffast-math flag, but other packages keep it.

For what it's worth, this was done on Mandrake 9.1, as It's almost impossible to really utilize Gentoo's advantages on a 56k dialup connection without at least using the wvdial "resuming downloads" function (I must share the one phone line with others, so I never get more than 1-2 hours at a time). By the time I downloaded the equivalent of emerge world on dialup without the option of leaving my system on overnight, it would be obsolete.

wrc1944
_________________
Main box- AsRock x370 Gaming K4
Ryzen 7 3700x, 3.6GHz, 16GB GSkill Flare DDR4 3200mhz
Samsung SATA 1000GB, Radeon HD R7 350 2GB DDR5
OpenRC Gentoo ~amd64 plasma, glibc-2.36-r7, gcc-13.2.1_p20230304
kernel-6.8.4 USE=experimental python3_11

xedx · Tux's lil' helper Joined: 23 May 2003 Posts: 93

i read it is advised to add -falign-functions=64 to your CFLAGS if you have an athlon/duron, does it make any difference on a pentium4?

btw how 'bout hyperthreading. I have a pentium4 with [ht] and another one without. ICC does have some optimizations on [ht] IIRC
Any thoughts
_________________
--+//+

TheCoop · Posted: Wed Jun 11, 2003 8:34 am Post subject:

wrc1944 · Posted: Wed Jun 11, 2003 12:56 pm Post subject:

TheCoop,

Glad to here those flags work on Gentoo also. In all my reading, I ran accross many conflicting opinions about whether to include --mmmx, -m3dnow, and -msse. Some implied -march=athlon-xp did not in all cases automatically activate those opts. I decided to err on the side of caution, and add them, even if they aren't really needed with -march=athlon-xp.

I'd sure like to know one way or the other, but many persons who obviously knew more than I did felt you should include them.

wrc1944
_________________
Main box- AsRock x370 Gaming K4
Ryzen 7 3700x, 3.6GHz, 16GB GSkill Flare DDR4 3200mhz
Samsung SATA 1000GB, Radeon HD R7 350 2GB DDR5
OpenRC Gentoo ~amd64 plasma, glibc-2.36-r7, gcc-13.2.1_p20230304
kernel-6.8.4 USE=experimental python3_11

bkeating · Posted: Thu Jun 12, 2003 4:06 am Post subject:

I don't quite understand the construction of this line... Im running a Pentium 4 (3.06Ghz) and these are the flags it gives me;

puddpunk · l33t Joined: 20 Jul 2002 Posts: 681 Location: New Zealand

bsolar · Bodhisattva Joined: 12 Jan 2003 Posts: 2764

dgrant · Posted: Fri Jun 13, 2003 2:46 pm Post subject:

from freehackers.org: