optimize atomic exchanges

On Fri, Nov 10, 2023 at 08:55:29PM -0600, Nathan Bossart wrote:
> On Fri, Nov 10, 2023 at 06:48:39PM -0800, Andres Freund wrote:
>> Yes. We should optimize pg_atomic_exchange_u32() one of these days - it can be
>> done *far* faster than a cmpxchg. When I was adding the atomic abstraction
>> there was concern with utilizing too many different atomic instructions. I
>> didn't really agree back then, but these days I really don't see a reason to
>> not use a few more intrinsics.
>
> I might give this a try, if for no other reason than it'd force me to
> improve my mental model of this stuff. :)

Here's a first draft. I haven't attempted to add implementations for
PowerPC, and I only added the __atomic version for gcc since
__sync_lock_test_and_set() only supports setting the value to 1 on some
platforms. Otherwise, I tried to add specialized atomic exchange
implementations wherever there existed other specialized atomic
implementations.

I haven't done any sort of performance testing on this yet. Some
preliminary web searches suggest that there is unlikely to be much
difference between cmpxchg and xchg, but presumably there's some difference
between xchg and doing cmpxchg in a while loop (as is done in
atomics/generic.h today). I'll report back once I've had a chance to do
some testing...

--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com