Making proiscachable believable

I am just about to commit Bernard Frankpitt's
constant-expression-reducing code, along with some of the improvements
I suggested a couple days ago.  In particular, it will not try to reduce
any op/func not marked "proiscachable" in pg_proc.  This renders it a
tad ineffective with the current contents of pg_proc :-( ...

The only functions marked cachable in 6.5.2 are

play=> select proname from pg_proc where proiscachable;
proname
---------
version
hashsel
hashnpage
gistsel
gistnpage
(5 rows)

and to add insult to injury, I believe all five of these markings are
wrong!  Functions whose outputs can vary for the same inputs must not
be marked cacheable --- and all of these use data other than their
arguments.

I have been working on modifying pg_proc.h to have believable
cacheability information.  What I did was to mark everything cacheable
and then go through and unmark the stuff that shouldn't be
constant-foldable: basically, stuff that fetches data from tables,
stuff that involves datetime conversion, and a few special cases like
nextval() and oidrand().

I am worried that I may have missed some things, and am seeking advice
on how I can check my work.

One thing I did not realize at first was that *none* of the datetime,
date, abstime, timespan, or tinterval operators can safely be marked
cachable.  The reason: these datatypes have special data values that
mean "now" (this has nothing to do with what the conversion to/from
text form yields, BTW).  Thus, for example, datetimeeq might say
one thing today and another tomorrow for the same input values,
if one of them is "now" and the other is a constant time.  Short of
inserting knowledge about these special values into the constant-
folder, we have to mark all the operations on the datatype non-foldable.

Are there any other gotchas like that one?
        regards, tom lane