Fix planner failure in some cases of sorting by an aggregate.
authorTom Lane <tgl@sss.pgh.pa.us>
Tue, 20 Apr 2021 15:32:02 +0000 (11:32 -0400)
committerTom Lane <tgl@sss.pgh.pa.us>
Tue, 20 Apr 2021 15:32:02 +0000 (11:32 -0400)
commit375398244168add84a884347625d14581a421e71
tree2d4e63dd53b05c82c2847a04918a9c82f4088b54
parent95c3a1956ec9eac686c1b69b033dd79211b72343
Fix planner failure in some cases of sorting by an aggregate.

An oversight introduced by the incremental-sort patches caused
"could not find pathkey item to sort" errors in some situations
where a sort key involves an aggregate or window function.

The basic problem here is that find_em_expr_usable_for_sorting_rel
isn't properly modeling what prepare_sort_from_pathkeys will do
later.  Rather than hoping we can keep those functions in sync,
let's refactor so that they actually share the code for
identifying a suitable sort expression.

With this refactoring, tlist.c's tlist_member_ignore_relabel
is unused.  I removed it in HEAD but left it in place in v13,
in case any extensions are using it.

Per report from Luc Vlaming.  Back-patch to v13 where the
problem arose.

James Coleman and Tom Lane

Discussion: https://postgr.es/m/91f3ec99-85a4-fa55-ea74-33f85a5c651f@swarm64.com
src/backend/optimizer/path/equivclass.c
src/backend/optimizer/plan/createplan.c
src/backend/optimizer/util/tlist.c
src/include/optimizer/paths.h
src/include/optimizer/tlist.h
src/test/regress/expected/incremental_sort.out
src/test/regress/sql/incremental_sort.sql