On Mon, Nov 22, 2021 at 06:10:25PM +0000, Marc Zyngier wrote:

Mark Brown [off-list ref] wrote:

quoted

While we're on the subject of potential future work we might in future
want to not disable SVE on every syscall if (as seems likely) it turns
out that that's more performant for small vector lengths

How are you going to retrofit that into userspace? This would be an
ABI change, and I'm not sure how you'd want to deal with that
transition...

We don't need to change the ABI, the ABI just says we zero the registers
that aren't shared with FPSIMD.  Instead of doing that on taking a SVE
access trap to reenable SVE after having disabled TIF_SVE we could do
that during the syscall, userspace can't tell the difference other than
via the different formats we use to report the SVE register set via
ptrace if it single steps over a syscall.  Even then I'm struggling to
think of a scenario where userspace would be relying on that.

You could also implement a similar optimisation by forcing on TIF_SVE
whenever we return to userspace but that would create a cost for
userspace tasks that don't use SVE on SVE capable hardware so doesn't
seem as good.  In any case it's not an issue for now since anything here
will need benchmarking on a reasonable range of hardware.