Laurenz Albe <laurenz.a...@cybertec.at> writes:
> On Sun, 2022-09-18 at 18:24 +0300, Евгений Плискин wrote:
>> So why not use this index for this query?
> Because the conditions are different:
> SELECT NULL = TRUE, NULL IS TRUE;
> ?column? │ ?column?
> ══════════╪══════════
> │ f
> (1 row)
> The first result is NULL.
Nonetheless, indxpath.c knows it can transform between "bool = true"
and "bool IS TRUE" for the purpose of making an index search
qualification, so it seems a bit odd that we fail to do the equivalent
transformation when attempting to prove an index predicate.
It'd be possible to improve this by adding some proof rules to
predicate_implied_by_simple_clause: I think both "x => x IS TRUE"
and the converse would be valid per the proof requirements, and
if you wanted to gild the lily it'd likely be possible to handle
some related cases like "x => x IS NOT FALSE".
Whether it's worth the cycles isn't too clear to me, but we could
argue about that if somebody submitted a patch.
regards, tom lane
