Re: POC, WIP: OR-clause support for indexes

On 1/13/25 01:39, Alexander Korotkov wrote:
> The revised patch is attached. Most notably it revises
> group_similar_or_args() to have the same notion of const-ness as
> others. In that function we split potential index key and constant
> early to save time on enumerating all possible index keys. But it
> appears to be possible to split by relids bitmapsets: index key should
> use our relid, while const shouldn't. Other that that, comments,
> commit message and naming are revised.
Hmm, I would say we should carefully review this code.
Curiously, this patch has activated the dormant problem of duplicated
clauses in joinorclauses. Look:

EXPLAIN (COSTS OFF)
SELECT * FROM bitmap_split_or t1, bitmap_split_or t2
WHERE t1.a=t2.b OR t1.a=1;

Nested Loop
-> Seq Scan on bitmap_split_or t2
-> Bitmap Heap Scan on bitmap_split_or t1
Recheck Cond: (((a = t2.b) OR (a = 1)) AND
((a = t2.b) OR (a = 1)))
-> Bitmap Index Scan on t_a_b_idx
Index Cond: ((a = ANY (ARRAY[t2.b, 1])) AND
(a = ANY (ARRAY[t2.b, 1])))

It can be resolved with a single-line change (see attached). But I need
some time to ponder over the changing behaviour when a clause may match
an index and be in joinorclauses.

--
regards, Andrei Lepikhov