Re: Requiring 32 bit atomics

Thomas Munro <thomas(dot)munro(at)gmail(dot)com> writes:
> We have fallback code for computers that don't have 32 bit atomic ops.
> Of course all modern ISAs have 32 bit atomics, but various comments
> imagine that a new architecture might be born that we don't have
> support for yet, so the fallback provides a way to bring a new system
> up by implementing only the spinlock operations and emulating the
> rest. This seems pretty strange to me: by the time someone brings an
> SMP kernel up on a hypothetical new architecture and gets around to
> porting relational databases, it's hard to imagine that the compiler
> builtins and C11 atomic support wouldn't be working.

Fair point. Another point you could make is that we no longer have
any test coverage for machines without 32-bit atomic ops.

But wait, you say, what about mamba-nee-gaur, my HPPA dinosaur?
The only actual hardware support there is equivalent to TAS();
nonetheless, if you read mamba's configure report you'll see it
claims to have atomic ops. I wondered if NetBSD was implementing
that by using kernel calls to disable interrupts, or something
equally badly-performing. Turns out they have a pretty cute
workaround for it, on HPPA and a couple of other atomics-less
arches they still support. They've written short sequences that
have the effect of CAS and are designed to store to memory only
at the end. To make them atomic, libc asks the kernel "pretty
please, if you happen to notice that I've been interrupted in
the PC range from here to here, would you reset the PC to the
start of that before returning?". At least on HPPA, this is
implemented for 8-bit, 16-bit, and 32-bit CAS and then all the
other standard atomics are implemented on top of that, so that
the kernel doesn't spend too much time checking for these
address ranges when it takes an interrupt.

Of course this only works on single-CPU machines. On multi-CPU
there's a completely different implementation that I've not spent
time looking at ... but I assume the performance is a lot worse.

Anyway, I think the big picture here is that nowadays we could
assume that the platform offers this feature.

regards, tom lane