Doug S (foo.delete@this.bar.bar) on May 23, 2022 8:08 am wrote:
> -.- (blarg.delete@this.mailinator.com) on May 23, 2022 4:23 am wrote:
> > Doug S (foo.delete@this.bar.bar) on May 22, 2022 10:57 am wrote:
> > > I would assume the first step in deprecating NEON (if that was the plan) would
> > > be to require SVE2 or some successor spec. SVE2 remains optional in ARMv9.
> >
> > On GCC and Clang, -march=armv9-a is an alias for -march=armv8.5-a+sve2, so assuming they're correct,
> > it sounds like SVE2 is as "optional" on ARMv9 as NEON is "optional" on ARMv8. (i.e. "opt-out optional")
> >
> > ARM have been less afraid to drop support for old ISAs/extensions (like AArch32), unlike x86. Given NEON's
> > prevalence, it'll likely stay around for quite some time, and given the markets ARM is now in (or entering
> > into), they probably can't drop backwards compatibility as much as they used to. But SVE was designed with
> > no reliance on NEON, so it's possible the latter will slowly be phased out over many years.
>
>
> I don't know why they've chosen to do that, but according to the latest ARMv9 documentation
> (which covers through ARMv9.3) FEAT_SVE2 is still listed as optional for ARMv9. AFAIK,
> NEON is required and not something you can opt out on - if you have a link to some
> documentation from ARM showing otherwise I'd be interested to see it.
>
> See page B.1.1.2 on page 26: https://developer.arm.com/documentation/ddi0608/latest
>
> It also notes "Although FEAT_SVE2 is OPTIONAL, standard Armv9-A software platforms support
> FEAT_SVE2" which I guess refers to the ARMv9 cores ARM has designed so far. If Apple releases
> an ARMv9 core that does not include SVE2, gcc & clang will have to change that default.

Section D1.20.3 of the ARMv8 atchiteture reference manual talks about "Implementations not including Advanced SIMD and floating-point instruction" - but says it is something exceptional for specialized markets.