Speedup to our barrier code

Hi,

This is more a note for the future, than something I'm planning to
pursue right now. Turns out our x86 full memory barrier could be sped
up noticably on larger systems with a trivial change.

Just changing
__asm__ __volatile__ ("lock; addl $0,0(%%rsp)" : : : "memory", "cc")
into something like
__asm__ __volatile__ ("lock; addl $0,-8(%%rsp)" : : : "memory", "cc")

makes the barrier faster because there's no dependency slowing down
other uses of %rsp. Which are, rsp being the stack pointer, not rare.
Obviously that requires to have a few bytes below the stack pointer, but
I can't see that ever being a problem on x86.

More details, among others, is available at:
https://shipilev.net/blog/2014/on-the-fence-with-dependencies/

In the past we'd run into barriers being relevant for performance both
around the latch code and shm_mq. So it's probably worth trying the
above in a benchmark exercising either heavily.

Greetings,

Andres Freund