Re: Two constraints with the same name not always allowed

I wrote:
> I think the code supposes that checking for duplicate relation name
> is sufficient; but of course it is not if we want a table's constraints
> to have distinct names, since they may not all correspond to indexes.
> I do not think we can back-patch a change here --- it might break
> databases that are working satisfactorily today. But it seems like
> we could tighten this up in HEAD and maybe v11.

Attached is a draft patchset for this.

0001 replaces the existing index with a unique one and makes necessary
backend code adjustments. Said adjustments could have been as simple
as s/ConstraintRelidIndexId/ConstraintRelidTypidNameIndexId/g -- I tried
that, and it passed regression tests -- but I couldn't resist the
temptation to fix a few places that could make better use of the
redesigned index.

0002 adds user-friendliness by installing a nicer error message for
the complained-of case and by improving ChooseIndexName to avoid
autogenerating index names that will conflict with existing constraints.

I didn't look for possible documentation changes yet, but I think the
code changes are OK.

regards, tom lane