Re: remove_useless_groupby_columns is too enthusiastic

I tried the attached quick-hack patch that just prevents
remove_useless_groupby_columns from removing anything that
appears in ORDER BY.  That successfully fixes the complained-of
case, and it doesn't change any existing regression test results.
I'm not sure whether there are any cases that it makes significantly
worse.

If there happens to be an index on t2.othercol, t2.x, the patch would
definitely win since we can perform a GroupAggregate with no explicit
sort. If there is no such index, considering the redundant sorting work
due to the excess columns, do we still win?

I'm testing with the query below:

create table t (a int primary key, b int, c int);
insert into t select i, i%1000, i from generate_series(1,1000000)i;
analyze t;
create index t_b_a_idx on t (b, a);

select sum(c) from t group by a, b order by b limit 10;

If we have index 't_b_a_idx', there would be big performance
improvement. Without the index, I can see some performance drop (I'm not

(I also kind of wonder if the fundamental problem here is that
remove_useless_groupby_columns is being done at the wrong time,
and we ought to do it later when we have more semantic info.