By this I mean reading an index (via full scan) and not finding a rowid which is known to belong to the table should tell the rdms that the indexed column/expression is null for a particular row.
This could be used to count nulls of a columm:
select count(*) from table_a where col1 is null
Could be treated as
Select (select count(*) from table_a)-(select count(col1) from table_a) from dual
Or, more involved: for a table with columns A to Z, all individually indexed and nullable. The query
select b from table_a where a=:a
could transform to
select b.b
from table_a a
left join
(select rowid rowid_b, b
from table_a
where b is is not null) b
On a.rowid = b.rowid_b
where a.a = :a