Re: Postgres 7.3.5 and count('x')

I said:
> ramirez@idconcepts.org (Edwin S. Ramirez) writes:
> [ "select count('x') from some_table" fails ]

> The most direct fix is probably to make any_in() return some random
> value (may as well be ((Datum) 0)) instead of producing an error.

After further investigation I have decided that this is really just an
issue of incorrect ordering of tests in parse_coerce.c.  The attached
proposed patch arranges for the ANY/ANYELEMENT/ANYARRAY short-circuit
path to be taken before we check for UNKNOWN-type literals and Params.
The net effect is that coercion to ANY/ANYELEMENT/ANYARRAY will have
*no effect* on an UNKNOWN literal.  This seems reasonable, since in
these contexts we don't really know any more than we did before about
the type of the literal; it's really still UNKNOWN.

One side effect besides fixing Edwin's gripe is that where before,

regression=# select 'x'::unknown::anyelement;
ERROR:  cannot accept a value of type anyelement

now you get a result:

regression=# select 'x'::unknown::anyelement;anyelement
------------x
(1 row)

I think this is reasonable, since coercion of other objects besides
unknown literals to anyelement is a no-op:

regression=# select 'x'::text::anyelement;anyelement
------------x
(1 row)

regression=# select 2::anyelement;anyelement
------------         2
(1 row)

Comments?  Any objections to the patch?
        regards, tom lane


*** src/backend/parser/parse_coerce.c~    Mon Dec  8 15:37:30 2003
--- src/backend/parser/parse_coerce.c    Sat Dec 13 13:21:41 2003
***************
*** 153,158 ****
--- 153,166 ----         /* no conversion needed */         return node;     }
+     if (targetTypeId == ANYOID ||
+         targetTypeId == ANYARRAYOID ||
+         targetTypeId == ANYELEMENTOID)
+     {
+         /* assume can_coerce_type verified that implicit coercion is okay */
+         /* NB: we do NOT want a RelabelType here */
+         return node;
+     }     if (inputTypeId == UNKNOWNOID && IsA(node, Const))     {         /*
***************
*** 260,273 ****         param->paramtype = targetTypeId;         return (Node *) param;     }
-     if (targetTypeId == ANYOID ||
-         targetTypeId == ANYARRAYOID ||
-         targetTypeId == ANYELEMENTOID)
-     {
-         /* assume can_coerce_type verified that implicit coercion is okay */
-         /* NB: we do NOT want a RelabelType here */
-         return node;
-     }     if (find_coercion_pathway(targetTypeId, inputTypeId, ccontext,                               &funcId))
{
--- 268,273 ----
***************
*** 372,388 ****         if (!typeidIsValid(targetTypeId))             return false; 
-         /*
-          * If input is an untyped string constant, assume we can convert
-          * it to anything except a class type.
-          */
-         if (inputTypeId == UNKNOWNOID)
-         {
-             if (ISCOMPLEX(targetTypeId))
-                 return false;
-             continue;
-         }
-          /* accept if target is ANY */         if (targetTypeId == ANYOID)             continue;
--- 372,377 ----
***************
*** 396,401 ****
--- 385,401 ----         }          /*
+          * If input is an untyped string constant, assume we can convert
+          * it to anything except a class type.
+          */
+         if (inputTypeId == UNKNOWNOID)
+         {
+             if (ISCOMPLEX(targetTypeId))
+                 return false;
+             continue;
+         }
+ 
+         /*          * If pg_cast shows that we can coerce, accept.  This test now          * covers both
binary-compatibleand coercion-function cases.          */