On Sat, Apr 26, 2014, Dongsheng Song wrote:
> On Sat, Apr 26, 2014 at 3:22 PM, Adrien Nader <adr...@notk.org> wrote:
> > On Sat, Apr 26, 2014, Dongsheng Song wrote:
> >> On Sat, Apr 26, 2014 at 12:31 AM, Adrien Nader <adr...@notk.org> wrote:
> >> > I believe --with-arch=core2 is a big issue for generic toolchains. It
> >> > will create troubles which will be very annoying to pinpoint.
> >> >
> >> > Any x86_64/amd64/EM64T (I love how Intel always makes up awful names)
> >> > already has SSE2 and there is little value in restricting this except in
> >> > very specific situation which are better dealt on a case-by-case basis.
> >> >
> >> > Regards,
> >> > Adrien Nader
> >> >
> >>
> >> Do you really meet this issue ? I can not image someone still running
> >> 64 bit Windows on such old CPU.
> >>
> >> Intel core2 release in July 2006, shutdown in January 2010.
> >> http://en.wikipedia.org/wiki/Intel_Core_2
> >
> > I have troubles believing people still run XP but many do. :P
> >
> > The issue is rather when you look at the AMD CPUs.
> > First models with SSE3 are from the end of 2007 (meaning you would still
> > easily see machines with them at the end of 2008, even beginning of
> > 2009).
> > The real issue is with SSSE3 (one more 'S') since the first mobile CPUs
> > with it have been released beginning of 2011 and the first desktop CPUs
> > with it have been released end of 2011 are are still easily found.
> >
> > I have a machine without SSE3 in my living room (albeit it needs a new
> > PSU), along with a machine without SSSE3 which is definitely running
> > strong (in particular since the rate of CPU speed improvement has
> > dramatically slowed down in the last few years).
> >
> > I don't think there is a point in making core2 the default; I don't
> > think it will bring any improvements except when building multimedia
> > stuff and even then it's not unlikely they don't provide hand-tuned
> > assembly but even then, -mtune should do be able to bring the
> > performance benefits on the newer CPUs while still running on the older
> > ones.
> >
> I agree with you the performance view.
> But from gcc view, Intel 64 cpu is: nocona, core2 or later. In my
> memory, I never hear someone can run 64 bit windows on nocona without
> problem. I know it's unfair for AMD CPUs, but both without -march and
> with -march=<amd cpu type> give 3DNow defined, it's not acceptable for
> Intel CPUs.

"Intel 64" is a name that doesn't exist. It's either AMD64 (AMD
parlance), EM64T (Intel parlance), x64 (Microsoft parlance) or x86_64
(parlance of anyone not interested in marketing and propaganda). The
first CPUs handling x86_64 date from 2003 and were server-class CPUs.

Using "--with-arch=core2" means that there many CPUs sold during pretty
much *10* years will not be able to run the programs compiled with these
toolchains and will crash at surprising times in surprising ways.

Right now I have a Windows 8 x86_64 VM running on a CPU from 2009 or
2010 without SSSE3. Actually it maybe doesn't have SSE3 either. I've
never had issues with x86_64 stuff on it.
I've also grabbed my XP 64 box and there's a cute logo of a Pentium 4.
While everyone will agree XP 64 was far from perfect, I think it's a
good indication that it worked at least for some people.

As for not specifying any arch, I wasn't able to quickly find a
reference or documentation on the matter. However, Linux distributions
are a good example however: they run on all x86_64 CPUs and don't set
anything specific and that's what a generic toolchain should do too.

In any case, if not setting --with-arch makes code that cannot run on
the triplet specified during the build of GCC (which I do not believe)
then this is an issue in GCC and should be dealt in GCC and not
worked-around (but again, I believe it is not the case).

Adrien Nader

Start Your Social Network Today - Download eXo Platform
Build your Enterprise Intranet with eXo Platform Software
Java Based Open Source Intranet - Social, Extensible, Cloud Ready
Get Started Now And Turn Your Intranet Into A Collaboration Platform
Mingw-w64-public mailing list

Reply via email to