cache invalidation for PL/pgsql functions

The following behavior surprised me, and a few other people at
EnterpriseDB, and one of our customers:

rhaas=# create table foo (a int);
CREATE TABLE
rhaas=# create or replace function test (x foo) returns int as $$begin
return x.b; end$$ language plpgsql;
CREATE FUNCTION
rhaas=# alter table foo add column b int;
ALTER TABLE
rhaas=# select test(null::foo);
ERROR: record "x" has no field "b"
LINE 1: SELECT x.b
^
QUERY: SELECT x.b
CONTEXT: PL/pgSQL function test(foo) line 1 at RETURN
rhaas=# \c
You are now connected to database "rhaas" as user "rhaas".
rhaas=# select test(null::foo);
test
------

(1 row)

I hate to use the term "bug" for what somebody's probably going to
tell me is acceptable behavior, but that seems like a bug. I guess
the root of the problem is that PL/plgsql's cache invalidation logic
only considers the pg_proc row's TID and xmin when deciding whether to
recompile. For base types that's probably OK, but for composite
types, not so much.

Thoughts?

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company