Re: [gentoo-dev] Replacing cpu-feature USE flags

[Martin Schlemmer]

On Fri, 2006-07-07 at 05:31 -0700, Brian Harring wrote:
> On Fri, Jul 07, 2006 at 02:24:49PM +0200, Martin Schlemmer wrote:
> > On Fri, 2006-07-07 at 02:08 +0200, Diego 'Flameeyes' Pettenò wrote:
> > > On Friday 07 July 2006 01:54, Ciaran McCreesh wrote:
> > > > | No, we never spent years telling them not to use your so-called
> > > > | "CFLAGS hacks" that are rather a proper usage of what the compiler
> > > > | gives you.
> > > > Wrong. We did.
> > > Then you were wrong. I could have spent time explaining them when they make
> > > sense and why they don't in their usecases. If you did, well, then you really
> > > need to know better what you do because you seem to me pretty confused
> > > yourself, and I feel pity for you.
> > >
> >
> > Yes, we did.  Were we wrong?  Out of a purest point of view ... maybe.
> > The problem was though that earlier gcc's was very bad at generating
> > sse/sse2, and sometimes mmx code.
> >
> > Users being what they are though (ricers should say it all), they
> > enabled every flag that sounded like it could make their old box two
> > times faster.  This included -msse, -msse2, etc.  Which quite frankly
> > produced bad code in many cases.  So we told the users to not add any
> > -m* flags, and let gcc do its job with the proper -march.
> >
> > So yeah, I can see that general use flags for cpu features might become
> > more tedious with the many new modules of processors out there, but to
> > say handle it by adding -msse, etc to CFLAGS, will surely if not on
> > gcc4, but then on gcc3 systems just ask for trouble.
> >
> > And yes, I know you are saying that that is not exactly what you are
> > proposing, but the users will see it as a clear passport to stick all
> > those nice sounding flags just right back in, and then it will be too
> > late to tell them its not proper thing to do when the bugs start
> > flooding in.
>
> Dumb question, but what really blocks them from doing that now for
> x86 (for example)?
>
> Yes, can't enable certain flags for non x86/x86_64 arches, but the con
> you're pointing at exists now for the most part.
>

I thought it was obvious, but apparently I overrated my writing
skills :/

Anyhow, because now we can say 'don't do that!', or just close the bug
as INVALID.  If not, you can still try it, but the user might say we
told him to enable sse/whatever like that.

Also, as Luca stated, USE=mmx/sse/sse2/etc means that you enable
tailored mmx/sse/whatever code, that should be working fine, as it was
not gcc doing some shot in the dark at optimising, where if its
enveloped with the CFLAGS, you cannot disable broken gcc optimisations,
but enabled mmx/sse/whatever that should work on those older gcc's.


-- 
Martin Schlemmer