Are extreme global CFLAGS harmful?

vipernicus · Last edited by vipernicus on Tue Apr 04, 2006 11:11 am; edited 2 times in total

I love options and all, but I am starting to believe that higher level optimization with global CFLAGS is bad.

I can understand CFLAGS="-march=x -Ox -pipe -fomit-frame-pointer", but anything beyond that globally is beginning to no longer make sense.

While some CFLAGS help some, they hurt others, so adding any CFLAGS beyond those I've already mentioned, creates the issue, "Would I prefer more performance here, and less performance here?"

What if someone could create a syncable database for per-package CFLAGS, and create test-bed for that?

What do you think would be a good idea for fixing this? Performance and stability being the goal.
_________________
Viper-Sources Maintainer || nesl247 Projects || vipernicus.org blog

XenoTerraCide · Veteran Joined: 18 Jan 2004 Posts: 1418 Location: MI, USA

yeah I thought it would it would be good to be able to do cflags like we can now do use flags on a per package basis. because even -fomit-frame-pointer has been known to break certain packages noatun comes to mind. and breaks debugging on many, in which case it would be nice to disable it for that package like I can use flags.
_________________
I don't hang out here anymore, try asking on http://unix.stackexchange.com/ if you want my help.

vipernicus · Posted: Mon Apr 03, 2006 6:17 pm Post subject:

I wouldn't know if it were possible, but I think there should be a movement in testing most used apps with various levels of optimization and automatically setting those optimizations for the user if they set a USE flag of "auto-optimize". Make them "Gentoo-Dev approved" maybe. I think that would be ideal for performance, if I could set my make.conf with an march=x option, an option for default_opt=-Ox. Then each ebuild could pull that package's tested set of performance flags to tag on to a user's set of defaults. This could help bugs.gentoo.org, and the user I believe.
_________________
Viper-Sources Maintainer || nesl247 Projects || vipernicus.org blog

XenoTerraCide · Veteran Joined: 18 Jan 2004 Posts: 1418 Location: MI, USA

I think most of the time people try to use stupid code breaking "ricer" optimization and that's bad for most software, however -fomit-frame-pointer is usually stable but in certain cases it would be nice to disable for only individual packages.
_________________
I don't hang out here anymore, try asking on http://unix.stackexchange.com/ if you want my help.

drwook · Veteran Joined: 30 Mar 2005 Posts: 1324 Location: London

I think the 'problem' is that the only mechanism generally is stripping users flags in ebuilds, rather than adding known-safe flags. E.g. setiathome could do better than '-O2 -march=x', so why not set every CFLAG you can find on the web and then bug the devs for 'broken ebuilds' when the uber-crazy ones aren't filtered?

vipernicus · Posted: Mon Apr 03, 2006 7:05 pm Post subject:

vipernicus · Posted: Mon Apr 03, 2006 7:34 pm Post subject:

I see votes, but where are the comments?
_________________
Viper-Sources Maintainer || nesl247 Projects || vipernicus.org blog

Glorandar · Posted: Mon Apr 03, 2006 8:35 pm Post subject:

Taking away folks' options, due to the inexperience of some, is, IMO, a bad idea.

Sure CFLAGS can be abused. But so can plenty of other things.

I agree entirely, that it seems to be a best practice to keep the CFLAGS simple.

Yet I suspect that some would disagree with that statement.

IMO fundamentally, Gentoo is all about choice, a theme that I fully support. Hence I support the continued existence of CFLAGS within /etc/make.conf
_________________
----- Glorandar

vipernicus · Posted: Mon Apr 03, 2006 8:43 pm Post subject:

slycordinator · Advocate Joined: 31 Jan 2004 Posts: 3065 Location: Korea

codergeek42 · Posted: Tue Apr 04, 2006 5:08 am Post subject:

The wording of the question in your poll is ambiguous. Global CFLAGS are not inherently dangerous (such as the unoffically-recommended "-O2 -march=whatever -pipe"). It's when people go overboard that it becomes very harmful: things like "-O3 -march=whatever -fforce-instructions-to-align-blah -fdo-other-things" etc.

My $0.02...
_________________
~~ Peter: Programmer, Mathematician, STEM & Free Software Advocate, Enlightened Agent, Transhumanist, Fedora contributor
Who am I? :: EFF & FSF

ciaranm · Posted: Tue Apr 04, 2006 5:52 am Post subject:

If we had a slightly better profiles structure, we could remove the need for users to tinker with CFLAGS (and those pesky CPU feature USE flags that everyone gets wrong). Instead, one could just select, say, a pentium4 profile and have perfect CFLAGS from the profile.

bjweeks · n00b Joined: 06 Jan 2006 Posts: 10

ciaranm · Posted: Tue Apr 04, 2006 7:10 am Post subject:

mikegpitt · Advocate Joined: 22 May 2004 Posts: 3224

Wow I cast my vote and it was tied 50:50 (15 votes vs. 15 votes).

I said that it is not harmful to use extreme CFLAGS, but you need to know what you are doing. For a while I was using a string of handpicked CFLAGS and things ran great. I kind of got bored tinkering with them though, so I am not just running the regular gentoo opts.

There are CFLAGS out there that will mess you up, or more likely make your machine slower, but a lot of what people say in the forums is often untrue. One thing that I hate seeing is people saying -funroll-loops will make your binarys slower, when in fact loop unrolling allows for more optimizations to be applied to the code. (Yes yes, In some cases however you will not get a benefit.)

So I think bottom line, is it is ok to use "extreme" CFLAGS, but unless you know something about compiler optimizations, or read the gcc man page in depth, you could screw yourself up.

vipernicus · Posted: Wed Apr 05, 2006 4:03 am Post subject:

Tekara · n00b Joined: 01 Feb 2006 Posts: 56 Location: UofI, Moscow

Seems that most packages filter out the extreme flags anyway, kde, openoffice and a good number of the small applications that I've emerged have stripped most of the cflags that I have set (I'm trying out the jackass! installation).
_________________
The danger from computers is not that they will eventually get as smart as men, but that we will agree to meet them halfway. - Bernard Avishai
Computers are a lot like air conditioners - they both work great until you open windows.

ciaranm · Posted: Wed Apr 05, 2006 6:03 am Post subject:

drwook · Veteran Joined: 30 Mar 2005 Posts: 1324 Location: London

I'm just confused as to the value of "-O3 -ftracer -ffastmath -fmudflap -fearly-inlining -fforce-addr...." when you're compiling bash or udev.

Specific programs might benefit from some of these 'crazy' flags, setiathome or games maybe, but do I want a 30MB shell binary because I've got all these set globally? (exaggeration I know

)

plus weigh it up with the cost of breakage. a broken udev or shell, or even editor, could cause you a hell of a lot of hassle if you aren't careful and especially for some of the n00bs who get sucked in by cflags and screw up.

Tekara · n00b Joined: 01 Feb 2006 Posts: 56 Location: UofI, Moscow

However though, when you have a fixed set of known hardware, it's wouldn't be horribly dangerous to intelligently add to the cflags on top of the generic ones that the gentoo manual suggests. A common one would be -fweb (included in -O3 wich many users use in lieu) and of course the cxxflag -fvisibility-inlines-hidden which are both reccomended frequently and I have yet to see anyone comment about them causing issues. Well that's at least with gcc3.4, I've seen a few mentions about fweb causing slowdowns in 4.0, but that's something that I haven't been following.

I do agree with you though, drwook, that if a person doesn't do their homework with the flags their liable to break something and of I wouldn't use anything but the most conservative flags on a system that I expected some high level of reliability from.
_________________
The danger from computers is not that they will eventually get as smart as men, but that we will agree to meet them halfway. - Bernard Avishai
Computers are a lot like air conditioners - they both work great until you open windows.

vipernicus · Posted: Tue Apr 18, 2006 7:47 pm Post subject:

Some good reading supporting my theory:

http://www.redhat.com/magazine/011sep05/features/gcc/
_________________
Viper-Sources Maintainer || nesl247 Projects || vipernicus.org blog

Tagx · n00b Joined: 18 Apr 2006 Posts: 61

I think portage should have an /etc/portage/package.cflags file to specify per package cflags just as package.use allows us to specify per package use flags.

layout...

apps-foo/foo -O1
apps-foo/foo3 -O2 -fomit-frame-pointer -ftree-vectorize

Feel free to modify this if u want.

vipernicus · Posted: Tue Apr 18, 2006 8:24 pm Post subject: Re: Portage Per Package CFlags

Fitzzy · n00b Joined: 07 Nov 2005 Posts: 17 Location: Fitzville

The problem is that some specific packages may benefit from certain CFLAGS while other will pay dearly for the use.

For global use, use a minimal set of flags, they will not give you much anyway. Then if there are certain packages that you know can benefit from a certain flag, then add it for that package only.

If you don't know the implications of setting a flag, don't!