Re: BUG #15200: Support ANSI OFFSET .. FETCH syntax with bind variables

>>>>> "PG" == PG Bug reporting form <noreply(at)postgresql(dot)org> writes:

PG> And as shown in this Stack Overflow question [2], it can be shown
PG> that the standard syntax doesn't work with anything but constant
PG> literals, including bind variables (which to me, are a kind of
PG> constant literal).

This does seem to be a defect in the code - the comments in the parser
say that the spec only allows constants, which is clearly wrong (what
the spec actually allows is <simple value specification>, which is
either a literal or a parameter/host variable/embedded variable).

PG> This is regrettable, the workaround when using this syntax from
PG> Java is to write:

PG> OFFSET (?) ROWS FETCH FIRST (?) ROWS ONLY

Having to use a workaround rather defeats the point of supporting
standard syntax in the first place. :-(

PG> This is also inconsistent with OFFSET .. LIMIT. The following works
PG> just fine:

PG> OFFSET ? LIMIT ?

This syntax doesn't have trailing ROW or ROWS keywords to contend with,
so it can allow arbitrary expressions without confusing the parser.

PG> I suggest relaxing this syntactic limitation and allowing for at
PG> least constant literals AND bind variables in this syntax

I think this is fixable with a fairly simple tweak to gram.y. The tricky
part is that FETCH FIRST allows the expression to be omitted entirely,
which causes shift/reduce conflicts (if we see ROW, is that the first
token in the expression, or an empty expression?). But since ONLY is a
fully reserved word, I think this can be solved by splitting into two
productions (one with the expression and one without), since by the time
we need to decide whether to reduce ROW/ROWS to anything, we can see
whether ONLY is the lookahead token.

Going to poke at this a bit to see if I can make it work right.

--
Andrew (irc:RhodiumToad)