Re: Fix performance of generic atomics

On 2017-09-06 15:25:20 -0400, Tom Lane wrote:
> Andres Freund <andres(at)anarazel(dot)de> writes:
> > On 2017-09-06 15:12:13 -0400, Tom Lane wrote:
> >> It looks to me like two of the three implementations promise no such
> >> thing.
>
> > They're volatile vars, so why not?
>
> Yeah, but so are the caller's variables. That is, in
>
> pg_atomic_exchange_u64_impl(volatile pg_atomic_uint64 *ptr, uint64 xchg_)
> {
> uint64 old;
> old = ptr->value;
>
> ISTM that the compiler is required to actually fetch ptr->value, not
> rely on some previous read of it. I do not think that (the first
> version of) pg_atomic_read_u64_impl is adding any guarantee that wasn't
> there already.
>
> >> Even if they somehow do, it hardly matters given that the cmpxchg loop
> >> would be self-correcting.
>
> > Well, in this one instance maybe, hardly in others.
>
> All the functions involved use nigh-identical cmpxchg loops.
>
> > What are you suggesting as an alternative?
>
> I think we can just use "old = ptr->value" to set up for the cmpxchg
> loop in every generic.h function that uses such a loop.

I think we might have been talking past each other - I thought you were
talking about changing the pg_atomic_read_u64_impl implementation for
external users.

Greetings,

Andres Freund