Good CFLAGS for Intel Core 2

karp · n00b Joined: 20 May 2002 Posts: 38 Location: Champaign, IL

I'm trying to figure out how to produce good code with gcc using an Intel Core 2 CPU. Since there isn't anything official in gcc that applies to Core 2, I thought I would do some experiments with custom CFLAGS. (gcc currently doesn't have an architecture flag for the Core 2, and there won't be one in gcc 4.2.0 either.) I'm running a strictly 64-bit system here, so thats the mode I'm testing here. I'm using Acovea to sort out what flags are important.

First off, the march/mtune flags. The only ones that gcc will accept in 64-bit are 'opteron' and 'nocona', all others produce "CPU you selected does not support x86-64 instruction set". Neither one is really appropriate for Core 2... in fact, I've found its actually best to leave march/mtune unspecified. MMX, SSE, and SSE2 are built into the 64bit specification, so they don't need to be specified.

Secondly, optimization level. Supposedly Apple does all their compilation using -Os (optimizing for small size), but I have yet to personally see -Os beating -O1 or -O2, and quite often -Os lags by 10-20%. My guess this is because Core 2 processors have pretty large caches, so cache misses aren't as much of a concern as they would be on other CPUs. -O2 is a tad better than -O1, I haven't done enough testing to say for sure though. -O3 doesn't seem to provide much benefit, takes more time to compile, makes much larger binaries, and sometimes produces incorrect results, so I'm not using it.

Then there are the flags that determine how floating-point math is done: -mfpmath={sse|sse,387|387}. For some reason, its fastest to leave this unspecified, which is strange because the documentation says -mfpmath=sse is default for x86-64, yet the actual default is faster than -mfpmath=sse. !?

One interesting thing I've found is that you can actually get a slight speedup from disabling MMX: -mno-mmx. Perhaps disabling MMX allows programs to avoid overhead? Or maybe its a gcc bug?

Flags I won't be testing: -ffast-math and friends. Not suitable for system-wide CFLAGS, because they break several algorithms.

So, for my own machine, the CFLAGS are set to "-01 -pipe".

I'll be working more on getting some hard numbers from Acovea, and I'll post the results here.

desultory · Administrator Joined: 04 Nov 2005 Posts: 4648

Actual testing, nice. Interesting results too.

Some relevant links:
The Intel Core 2 Solo/Duo (Allendale, Conroe, Merom) entry in the Safe CFLAGS Wiki article.
Topic Core 2 Duo - Merom.

kernelOfTruth · Posted: Sat Nov 18, 2006 4:06 pm Post subject:

Interesting thread, karp

please keep us updated !

_________________
Unofficial minimal livecd x86/amd64 w/reiser4+truecrypt
gcc version 4.4.0 (Gentoo Hardened 4.4.0-r4 p1.1, espf-0.2.9)
watching "shit" scroll by since 2004 =)

lplatypus · n00b Joined: 26 Mar 2004 Posts: 16

karp · n00b Joined: 20 May 2002 Posts: 38 Location: Champaign, IL

SSE3 isn't enabled by default. But gcc doesn't seem to take advantage of it very well, because the acovea results I'm getting indicate it hurts performance when enabled, somehow.

Last I heard, SSSE3, which Core 2 supports, won't make it until 4.3:
http://gcc.gnu.org/ml/gcc-patches/2006-09/msg01285.html

Oh, and since my first post I have come across several benchmarks that benefit from -O2 over -O1, and none that are hurt. So my current CFLAGS is "-O2 -pipe".

karp · n00b Joined: 20 May 2002 Posts: 38 Location: Champaign, IL

Okay, here's my first result. I tweaked the gcc 4.0 Opteron config file from the Acovea website, and added gcc 4.1 and Core 2 flags. The benchmark is based on a simple raytracer from http://www.ffconsultancy.com/free/ray_tracer/languages.html. I chose a raytracer because it has a good mix of indirection, mathematics, and branching.

karp · n00b Joined: 20 May 2002 Posts: 38 Location: Champaign, IL

The Huffman encoding benchmark included with Acovea:

schmobag · Posted: Tue Nov 21, 2006 12:00 pm Post subject:

I'm a little confused about -Os.

You original post says that you've never seen -Os beat -O1.

The gcc man page defines -Os like this:

batistuta · Veteran Joined: 29 Jul 2005 Posts: 1384 Location: Aachen

schmobag · Posted: Tue Nov 21, 2006 12:55 pm Post subject:

karp · n00b Joined: 20 May 2002 Posts: 38 Location: Champaign, IL

The best test for -Os would probably be to recompile an entire system with it, and then measure overall system performance compared to an entirely -O2 system. Having only a benchmark give up some code size for The Good of the Cache doesn't mean much when the kernel, glibc, etc are still large. But if they *all* made a sacrifice, it could be worth it.

That said, I don't know if I feel up to recompiling my entire system multiple times to test this. Besides, I'm not even sure if there is a good overall-system-performance benchmark... maybe kernel-compilation?

batistuta · Veteran Joined: 29 Jul 2005 Posts: 1384 Location: Aachen

kernel compilation is always a special case. So special that it has its own set of flags, some are ignored even if you pass them. So I don't think kernel is a good candidate for it... :roll:

karp · n00b Joined: 20 May 2002 Posts: 38 Location: Champaign, IL

I think you misunderstand... I wouldn't be changing the flags used inside of the kernel build process. I'd be comparing how toolchains compiled with various settings perform on a standard task: compiling the kernel.

karp · n00b Joined: 20 May 2002 Posts: 38 Location: Champaign, IL

batistuta · Veteran Joined: 29 Jul 2005 Posts: 1384 Location: Aachen

anli · Tux's lil' helper Joined: 08 Sep 2006 Posts: 78

karp,

Thanks for your information!

Just want to clarify one moment. After these months which arch cfg you treat as the best - 'nocona' or without any arch flag at all?

Keruskerfuerst · Veteran Joined: 01 Feb 2006 Posts: 1717

Has anyone done a CPU test for AMD Opteron or AMD Athlon64 as discribed here: http://gentoo-wiki.com/TIP_Acovea

peaceful · Apprentice Joined: 06 Jun 2003 Posts: 285 Location: Utah

Element Dave · n00b Joined: 10 Nov 2006 Posts: 54

nxsty · Veteran Joined: 23 Jun 2004 Posts: 1556 Location: .se

likewhoa · Posted: Sat Feb 03, 2007 4:45 pm Post subject:

yaneurabeya · Posted: Sun Mar 25, 2007 12:15 am Post subject:

Interesting thread. It may be because of my CLI resolution compared to the size of the FB resolution on the suse livecd, but it appears that things are trolling along a bit slower on my Core 2 Duo box. I only have 2GB of ram in the box currently and need to add the other 2 (running it in 64-bit mode).

Current CFLAGS: -O2 -pipe -fno-strict-aliasing -funroll-loops -msse -msse2 -msse3 <- guess I don't need -msse3 any more, do I?

If there's anything I can help test out, I'd be more than happy to.