Making OFF unreserved

OFF is a reserved keyword. It's not a reserved keyword in the SQL spec,
and it's not hard to see people using off as a variable or column name,
so it would be nice to relax that. To make things worse, OFFSET is also
a reserved keyword, which would be the other natural name for a variable
or column that stores an offset of some sort.

I bumped into this because we have a test case in the EDB regression
suite that uses 'off' as a PL/pgSQL variable name. It used to work
before 9.0, because PL/pgSQL variable names were replaced with $n-style
parameter markers before handing off the query to the backend parser.
It's a problem with all keywords in general, but 'off' seems like a
likely variable name in real applications, and there was no ambiguity
with it.

Looking at the grammar, OFF is only used here:

> opt_boolean:
> TRUE_P { $$ = "true"; }
> | FALSE_P { $$ = "false"; }
> | ON { $$ = "on"; }
> | OFF { $$ = "off"; }
> ;

And opt_boolean in turn is used in the following places:

> var_value: opt_boolean
> { $$ = makeStringConst($1, @1); }
> | ColId_or_Sconst
> { $$ = makeStringConst($1, @1); }
> | NumericOnly
> { $$ = makeAConst($1, @1); }
> ;
> ...
> copy_generic_opt_arg:
> opt_boolean { $$ = (Node *) makeString($1); }
> | ColId_or_Sconst { $$ = (Node *) makeString($1); }
> ...
> copy_generic_opt_arg_list_item:
> opt_boolean { $$ = (Node *) makeString($1); }
> | ColId_or_Sconst { $$ = (Node *) makeString($1); }
> ;
> ...
> explain_option_arg:
> opt_boolean { $$ = (Node *) makeString($1); }
> | ColId_or_Sconst { $$ = (Node *) makeString($1); }

Note that ColId is also accepted alongside opt_boolean in all of those
with the same action, so if we just remove OFF from opt_boolean rule and
make it unreserved, nothing changes.

ECPG uses OFF as a keyword in its "SET autocommit = [ON | OFF]" rule, so
we have to retain it as an unreserved keyword, or make it an
ecpg-specific keyword in the ecpg grammar. But I don't know how to do
that, and it feels like a good idea to keep it in the unreserved keyword
list anyway, so I propose the attached patch.

Any objections? Any objections to backpatching to 9.0, where the
PL/pgSQL variable handling was changed?

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com