I wrote:
> A fairly common solution to that is to return NULL for bad input,
> but in this case we could just have it return the OID unchanged.
After sleeping on it, I concluded that was a bad idea and we'd
be best off returning NULL for invalid type OIDs. So this is
just about back to Steve's original proposal, except for being
a bit more bulletproof against races with DROP TYPE.
Pushed that way.
regards, tom lane