On Wed, Jul 6, 2022 at 12:27 PM Andres Freund <andres@anarazel.de> wrote:
> I think my proposal of introducing a version of double that is marked to be 8
> byte aligned should do the trick as well, and doesn't have the problem of
> changing the meaning of 'double' references in external headers. In fact, we
> already have float8 as a type, so we could just add it there.
Yeah, but how easily will it be to know whether we've used that in
every relevant place?
Could we insist on 8-byte alignment even on 32-bit platforms? I think
we have a few of those in the buildfarm, so maybe that would help us
spot problems. Although I'm not sure how, exactly.
> The problem with having a lot more alignment values is that it adds a bunch of
> overhead to very performance critical paths. We don't want to add more
> branches to att_align_nominal() if we can avoid it.
Fair.
> I'm fairly certain that we're going to add a lot more 64bit ints to catalogs
> in the next few years, so this will become a bigger issue over time...
Absolutely.
> Outside of the catalogs I still think that we should work towards not aligning
> byval values (and instead memcpy-ing the values to deal with alignment
> sensitive platforms), so we don't waste so much space. And for catalogs we've
> been talking about giving up the struct mapping as well, in the thread about
> variable length names. In which case we could the cost of handling more
> alignment values wouldn't be incurred as frequently.
+1. Aligning stuff on disk appears to have few redeeming properties
for the amount of pain it causes.
--
Robert Haas
EDB: http://www.enterprisedb.com