I wrote:
> A post-pass is not out of the question, but it's enough unlike
> everything else the parser does that I'm not too thrilled about it.
On the other hand ... one thing that's been bothering me is that
select_common_collation assumes that "explicit" collation derivation
doesn't bubble up in the tree, ie a COLLATE is only a forcing function
for the immediate parent expression node. It's not at all clear to me
that that's a correct reading of the spec. If it's not, the only way
we could make it work correctly in the current design is to keep
*two* additional fields, both the collation OID and an explicit/implicit
derivation flag. Which would be well past the level of annoying.
But in a post-pass implementation it would be no great trouble to do
either one, and we'd not be looking at a forced initdb to change our
minds either.
Maybe a post-pass, with only collation-to-apply fields actually stored
in the tree, is the way to go.
Comments?
regards, tom lane