I'm not sure what I think about Ashutosh's ideas about doing this somewhere else than eval_const_expressions. I do not buy the argument that it's interesting to do this separately for each child partition. Child partitions that have attnotnull constraints different from their parent's are at best a tiny minority use-case, if indeed we allow them at all (I tend to think we shouldn't).
I agree about partitions. But, IMO, a child having constraints different from that of a parent is more common in inheritance trees.
Another point I raised in my mail was about constraint exclusion. Why aren't these clauses constant-folded by constraint exclusion? Sorry, I haven't looked at the constraint exclusion code myself for this.
As a not incidental example, consider
select ... from t1 left join t2 on (...) where t2.x is not null;
reduce_outer_joins will realize that the left join can be reduced to a plain join, whereupon (if t2.x is attnotnull) the WHERE clause really is constant-true --- and this seems like a poster-child case for it being useful to optimize away the WHERE clause. But we won't be able to detect that if we apply the optimization during eval_const_expressions. So maybe that's a good reason to do it somewhere later.
