Hi Hackers,
relation_has_unique_index_for() checks whether join clause equality
operators belong to the index's opfamily via mergeopfamilies. Since
mergeopfamilies only lists btree opfamilies, this check always fails
for GiST-backed unique indexes such as those created by PRIMARY KEY
with WITHOUT OVERLAPS, preventing self-join elimination.
Fix by falling back to op_in_opfamily() when the mergeopfamilies check
fails. The clause is already known to be a mergejoinable equality, so
confirming the operator is registered in the index's opfamily is
sufficient to prove that the index's uniqueness guarantee applies.
Attached a patch to fix this and added corresponding tests.
Thanks,
Satya
