On Mon, Jun 4, 2018 at 11:01 AM, Adam Jackson <ajax@redhat.com> wrote:
On Mon, 2018-06-04 at 08:59 -0500, Ian Pilcher wrote:
> On 06/04/2018 04:28 AM, Florian Weimer wrote:
> > It should, because -march=x86-64 implies just SSE2 and FXSR, and Xeon MP
> > supports both.  But the intent is what the subject says: i686 binaries
> > are for running legacy software on x86-64 systems, and nothing more.
> So the 32-bit x86 SIG would be required to rebuild all of the userspace
> packages to run on actual 32-bit hardware, right?  (What would those be
> called, since i686 is taken?)

No. The 32-bit ABI level being targeted here is basically the Pentium
4, which was a 32-bit part. On a P3 or older, yes, you'd need a
rebuild, but the P4 was 2001 so we're talking about hardware old enough
to vote here.

If we did want to allow for pre-P4 32-bit userspace, what I'd probably
do is (mostly) change %_libdir to /usr/lib/sse2 for i686, and leave it
as /usr/lib for a (reintroduced) i586 architecture.

If there is truly a pressing need to enable SSE2 for *all* i686 binaries, I'd support this as a compromise.