On Wed, Nov 22, 2017 at 12:04:58PM +0000, Mark Rutland wrote:
> On Wed, Nov 22, 2017 at 10:12:17AM +0000, Ard Biesheuvel wrote:
> > On 22 November 2017 at 10:05, Alex Matveev <alxmtvv@xxxxxxxxx> wrote:
> > > This is better than my simple fix, thank you.
> > >
> > > Out of curiosity, why doesn't NEON code use barrier() to prevent
> > > reordering?
> >
> > Because barrier() affects ordering of memory accesses, not register accesses.
>
> My understanding was that barrier() also hazards with any volatile asm
> (and that in general, volatile asm block couldn't be re-ordered w.r.t.
> one another). I'm sure we rely on this, even if it's not true. :/
>From a quick scan, certainly in tls_thread_flush(),
compat_arm_syscall(), elsewhere where we have isb() and other helpers
like {read,write}_sysreg().
Mark.
