Hi,
Around
https://www.postgresql.org/message-id/20230224015417.75yimxbksejpffh3%40awork3.anarazel.de
I suggested that we should evaluate the arguments of correlated SubPlans as
part of the expression referencing the subplan.
Here's a patch for that.
Ended up simpler than I'd thought. I see small, consistent, speedups and
reductions in memory usage.
I think individual arguments are mainly (always?) Var nodes. By evaluating
them as part of the containing expression we avoid the increased memory usage,
and the increased dispatch of going through another layer of
ExprState. Because the arguments are a single Var, which end up with a
slot_getattr() via ExecJust*Var, we also elide redundant slot_getattr()
checks. I think we already avoided redundant tuple deforming, because the
parent ExprState will have done that already.
Greetings,
Andres Freund