I wrote:
> Peter Geoghegan <pg@heroku.com> writes:
>> This is because the ArrayExpr jumble case jumbles any ArrayExpr's list
>> of elements recursively. In this case it's a list of Const nodes, and
>> the fingerprinting logic jumbles those nodes indifferently.
> I think this is a vastly oversimplified explanation of the problem.
> In particular, because the planner will flatten an ArrayExpr containing
> only Const nodes to an array constant (see eval_const_expressions),
> I don't believe the case ever arises in exactly the form you posit here.
Um ... disregard that. For some reason I was thinking that
pg_stat_statements looks at plan trees, but of course what it looks at
is the query tree immediately post-parse-analysis. So the behavior of
the const-folding pass is not relevant.
The heuristics in transformAExprIn() are still relevant, though, and
I suspect that the question of whether Params should be considered
the same as Consts is also highly relevant.
I wonder whether we could improve this by arranging things so that both
Consts and Params contribute zero to the jumble hash, and a list of these
things also contributes zero, regardless of the length of the list.
regards, tom lane