I wrote:
> I think the correct thing is to do nothing, and assume the expanded
> expression must have the right type already, if the function is declared
> to return a pseudotype. The only pseudotypes allowed as SQL-function
> results are RECORD, VOID, and polymorphic, and this seems OK and maybe
> even required in each case. But having gotten this wrong once already,
> maybe I better call for comments...
Make that 0 for 2 :-(. On closer inspection the correct patch seems to
be just to use "result_type", ie the result type the function call node
was already labeled with, not funcform->prorettype (the function's
declared result type). This can be seen to be correct from two
arguments:
1. The whole point of the RelabelType insertion is to ensure that the
exposed type of the expression tree (as reported by exprType say)
remains the same as before. And "result_type" is exactly what it was
before.
2. result_type, not prorettype, is in fact what check_sql_fn_retval()
was checking against. That Assert was intended to back up that we were
in sync with check_sql_fn_retval(), but we weren't.
So this is just a pure thinko in the previous patch. Sigh.
regards, tom lane
