Re: BUG #17486: [pg_restore] Restoring a view fails if this view contains an attribute without alias name.

I wrote:
> Hmm ... it's a very easy code change, but it results in a lot of
> changes in the regression tests (and I've only tried the core tests
> so far). Given the lack of prior complaints, I wonder if it's going
> to be worth this much behavioral churn.

> It'd be better if we could do this only when the name is actually
> referenced somewhere, but I don't think that's an easy thing to
> determine.

I thought of a compromise position that's not hard to implement:
change the behavior only if the SELECT output column name is *possibly*
visible elsewhere, which it is not in (for example) an EXISTS subquery.
This is easy to keep track of while descending the parse tree.
The attached quick-hack draft results in only one place changing in
the regression tests, and that's a place where a view's visible
column name is already "?column?", so whoever wrote that test case
didn't give a fig for prettiness anyway. This seems like it might be
an acceptable amount of behavioral churn.

Thoughts?

regards, tom lane