On Tue, 2013-06-04 at 13:40 -0400, Tom Lane wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
> > On Tue, Jun 4, 2013 at 11:48 AM, Pavel Raiskup <praiskup@redhat.com> wrote:
> >> Oh, I see now it was already consulted here:
> >> http://www.postgresql.org/message-id/1368448758.23422.12.camel@t520.redhat.com
>
> > I think we should go ahead and commit this patch, or some variant of
> > it. Having a buildfarm machine would be good... but I don't think
> > that should be a prerequisite for this sort of support. We certainly
> > have spinlock support for other platforms for which we don't have
> > buildfarm machines.
>
> We got no response to the question of whether it couldn't be merged with
> the existing ARM32 code block. I'd prefer not to have essentially
> duplicate sections in s_lock.h if it's not necessary.
Of course it could be. I don't think it should be. Aarch64 is a
completely new architecture with faint resemblance to 32bit arm. I went
back and read the previous thread concerning the problems with builtin
gcc atomics with certain tool/hardware combinations. Would it be okay to
add a default implementation based on gcc builtins to be used if no
arch-specific definitions exist? If so, I could work up a patch for
review.
--Mark