I wrote:
> I'm not seeing any terribly pleasing ways to fix this. Aside from
> the option of doing nothing, it seems like these are the choices:
> 1. We could hack base_yylex() to reduce NOT LIKE to a single token
> which could be given the same precedence as LIKE. Ditto for the other
> four cases. This would result in nice behavior for these cases, but as
> Robert has correctly pointed out, hacks in base_yylex() risk creating
> other poorly-understood behaviors.
> 2. We could change the precedence levels so that LIKE/ILIKE/SIMILAR,
> BETWEEN, and IN all have precedence just above NOT, which would pretty
> much mask the problem since there would be no other tokens with in-between
> precedence levels. In the context of the operator precedence changes
> I proposed earlier, this would mean inserting the IS tests and comparison
> operators between IN_P and POSTFIXOP rather than where the
> single-character comparison ops live now. We would likely also have to
> flatten LIKE/BETWEEN/IN into a single %nonassoc precedence level to avoid
> having weird interactions among them. This isn't terribly attractive
> because it would risk larger behavioral changes than the previous
> proposal.
I thought of another possibility:
3. Leave everything as-is but mark the NOT-operator productions as having
the precedence of NOT rather than of LIKE etc. This would change the
behavior only for the NOT-LIKE-followed-by-< example, and would make the
two cases for NOT LIKE consistent though they'd remain inconsistent with
LIKE. This behavior seems at least somewhat explainable/documentable
("NOT-foo operators have the precedence of NOT"), whereas what we have
seems about impossible to justify.
regards, tom lane