"David G. Johnston" <david.g.johnston@gmail.com> writes: > On Tue, Jan 24, 2023 at 12:31 PM Tom Lane <tgl@sss.pgh.pa.us> wrote: >> select ... from t1 left join t2 on (t1.x = t2.y and t1.x = 1); >> >> If we turn the generic equivclass.c logic loose on these clauses, >> it will deduce t2.y = 1, which is good, and then apply t2.y = 1 at >> the scan of t2, which is even better (since we might be able to turn >> that into an indexscan qual). However, it will also try to apply >> t1.x = 1 at the scan of t1, and that's just wrong, because that >> will eliminate t1 rows that should come through with null extension.
> Is there a particular comment or README where that last conclusion is > explained so that it makes sense.
Hm? It's a LEFT JOIN, so it must not eliminate any rows from t1. A row that doesn't have t1.x = 1 will appear in the output with null columns for t2 ... but it must still appear, so we cannot filter on t1.x = 1 in the scan of t1.
Ran some queries, figured it out. Sorry for the noise. I had turned the behavior of the RHS side appearing in the ON clause into a personal general rule then tried to apply it to the LHS (left join mental model) without working through the rules from first principles.