Horton

Re: Fix performance of generic atomics

Tom Lane <tgl@sss.pgh.pa.us>

View thread View on postgr.es
From: Tom Lane <tgl@sss.pgh.pa.us>
To: Andres Freund <andres@anarazel.de>
Cc: Sokolov Yura <funny.falcon@postgrespro.ru>, Robert Haas <robertmhaas@gmail.com>, Simon Riggs <simon@2ndquadrant.com>, Jeff Janes <jeff.janes@gmail.com>, Jesper Pedersen <jesper.pedersen@redhat.com>, pgsql-hackers <pgsql-hackers@postgresql.org>
Date: 2017-09-06T19:12:13Z
Lists: pgsql-hackers 
Andres Freund <andres@anarazel.de> writes:
> On 2017-09-06 14:31:26 -0400, Tom Lane wrote:
>> However, if that's the reasoning, why don't we make all of these
>> use simple reads?  It seems unlikely that a locked read is free.

> We don't really use locked reads? All the _atomic_ wrapper forces is an
> actual read from memory rather than a register.

It looks to me like two of the three implementations promise no such
thing.  Even if they somehow do, it hardly matters given that the cmpxchg
loop would be self-correcting.  Mostly, though, I'm looking at the
fallback pg_atomic_read_u64_impl implementation (with a CAS), which
seems far more expensive than can be justified for this.

			regards, tom lane

Commits

  1. Further marginal hacking on generic atomic ops.

  2. Use more of gcc's __sync_fetch_and_xxx builtin functions for atomic ops.

  3. Remove duplicate reads from the inner loops in generic atomic ops.