Spinlocks and semaphores in 9.2 and 9.3

This rabbit hole keeps getting deeper and deeper :-(

I realized a couple days ago that it had been close to three years since
I last tried building the further-back branches on my ancient HPPA box.
Not so surprisingly, things were broken: commits 37de8de9e33606a0 et al
had introduced use of memory barriers into 9.2 and 9.3, which up to that
moment had had none, so we'd never bothered to make barrier.h actually
work in those branches. I back-patched some fixes in barrier.h, which
got things to compile, and then commit 44cd47c1d49655c5, which got things
to actually work ... on that box anyway. I now see from the buildfarm
(and can confirm locally) that 44cd47c1d49655c5 breaks builds with
--disable-spinlock, because it introduces an initialization of a spinlock
before the underlying PGSemaphore infrastructure can possibly get
initialized.

In 9.4 and up, this works because of daa7527afc227443, which decoupled
spinlocks from semaphores enough that you can do s_init_lock_sema()
long before you can actually use the spinlock. Come to think of it,
that commit also means that you can get away with using a spinlock
you've never done initialization on at all, which is not so good.

So at this point I'm not sure what to do. I could back out the back-patch
of 44cd47c1d49655c5, which would mean accepting that 9.2/9.3 are broken
and will never be fixed for HPPA, as well as any other architectures that
use the same fallback memory barrier implementation. The lack of
complaints from the field suggests that nobody would care. Or I could
push forward by back-patching daa7527afc227443 (and a couple of minor
follow-on cleanups). That doesn't seem particularly risky, now that
9.4's been out for awhile, but it's kind of a large back-patch to benefit
architectures that apparently no actual users care about.

Thoughts?

Independently of that, I think we should fix the --disable-spinlock
support so that a spinlock containing zero is rejected as not properly
initialized. A large part of the mess here comes from the fact that HPPA
is our only architecture in which zero is not the correct initialization
state for a spinlock. I'd like to have some more portable method of
catching failure to initialize a spinlock. I only propose doing this
part in HEAD, though.

regards, tom lane