attribute distinct values estimate

We need to estimate the number of distinct values of an attribute from
inside optimizer, for an extension we are working on.
Should we use stacommonfrac from pg_statistic, or attdisbursion, or what
else?