Andres Freund <and...@anarazel.de> writes:
> On 2017-03-16 11:15:16 -0400, Tom Lane wrote:
>> The thing that actually made the ExecEvalCase code into a bug was that
>> we were using ExprContext-level fields to store the current caseValue,
>> allowing aliasing to occur across nested CASEs. I think that in
>> this implementation, it ought to be possible to get rid of
>> ExprContext.caseValue_datum et al altogether, in favor of some storage
>> location that's private to each CASE expression. I'm a bit disappointed
>> that that doesn't seem to have happened.
> ... Unfortunately
> CaseTest/CoerceToDomainValue are reused outside of domain / case
> expressions in a bunch of places (plpgsql uses CaseTest for casts
> evaluation, validateDomainConstraint/domain_check_input evaluate domain
> constraints without a CoerceToDomain node).
Yeah, there's various stuff that did that for expediency.
> I'd like to get rid of those usages, but that'd recurse into rewriting
> plpgsql casts and other random pieces of code into a different approach
> - something I'd like to avoid doing at the same as this already large
Agreed, maybe we should just plan to clean that up later.
> I've been pondering if we can't entirely get rid of CaseTest etc, the
> amount of hackery required seems not like a good thing. One way I'd
> prototyped was to replace them with PARAM_EXEC nodes - then the whole
> issue of them potentially having different values at different parts of
> an expression vanishes because the aliasing is removed.
Yes, replacing all of that with Param slots had occurred to me too.
We might want to keep the special parse node types for convenience in
reverse-listing, but having them act just like PARAM_EXEC for execution
purposes seems promising.
>> Eventually, I would also like to find a way to remove the restriction
>> imposed by the other part of f0c7b789a, ie that we can't inline a SQL
>> function when that would result in intermixing two levels of CASE
> That seems to suggest my PARAM_EXEC idea isn't necessarily perfect -
> inlining would cause aliasing again, but it'd also not hard to fix that
Right, inlining would probably require some parameter-renumbering to avoid
aliasing. But at least we'd have a clear framework for how to handle it,
whereas the way things are now, it's just impossible.
regards, tom lane
Sent via pgsql-hackers mailing list (firstname.lastname@example.org)
To make changes to your subscription: