Re: Prefered Types - Mailing list pgsql-hackers

2011/5/6 Tom Lane <tgl@sss.pgh.pa.us>:
> Robert Haas <robertmhaas@gmail.com> writes:
>> 2011/5/4 Tom Lane <tgl@sss.pgh.pa.us>:
>>> Perhaps it would be adequate to allow automatic resolution of an
>>> overloading conflict only when one of the available alternatives
>>> dominates all others, ie, none of the argument positions requires a
>>> "longer distance" cast than is used in that position by any other
>>> available alternative.  I'm just throwing that out as a possibility,
>>> I haven't tried it.
>
>> That works OK for most things, but there's one case where I think we
>> might need a better solution - suppose A is a subtype of B.  It's
>> fairly common to define a function or operator f(A,A) and f(B,B), and
>> to want f(A,B) or f(B,A) to be interpreted as a the latter rather than
>> the former.  For example, let A=int2, B=int4, f=+.  Now, we can (and
>> currently do) handle that by just defining all the combinations
>> explicitly, but people don't always want to do that.
>
> That case still works as long as downcasts (int4 -> int2) are either not
> allowed to be invoked implicitly at all, or heavily penalized in the
> distance assignments.

Not at all works, but heavily penalized doesn't.  Suppose A->B has
distance 1 and B->A has distance 1000.  Then f(A,B) can match f(A,A)
with distances (0,1000) or f(B,B) with distances (1,0).  If you add up
the *total* distance it's easy to say that the latter wins, but if you
compare position-by-position as you proposed (and, generally, I agree
that's the better route, BTW) then each candidate is superior to the
other in one of the two available positions.

>>> BTW, not to rain on the parade or anything, but I'll bet that
>>> rejiggering anything at all here will result in whining that puts the
>>> 8.3-era removal of a few implicit casts to shame.
>
>> Yeah, I share that fear, which is why I think the idea of generalizing
>> typispreferred to an integer has more than no merit: it's less likely
>> to break in ways we can't anticipate.
>
> Well, if you change it to an int and then don't change any of the values
> from what they were before, I agree.  But then there's no point.
> Presumably, the reason we are doing this is so that we can assign some
> other preferredness values besides 0/1, and that will change the
> behavior.  We'd better be damn sure that the new behavior is really
> better.  Which is why it seems a bit premature to be working on an
> implementation when we don't have even a suggestion as to what the
> behavioral changes ought to be.

Well, sure, to some degree.  But if you keep the currently preferred
types as having the highest level of preferred-ness in their same
categories, then the only effect (I think) will be to make some cases
work that don't now; and that's unlikely to break anything too badly.
Going to some whole new system will almost inevitably involve more
breakage.

> Which is why it seems a bit premature to be working on an
> implementation when we don't have even a suggestion as to what the
> behavioral changes ought to be.

I'm in complete agreement on this point.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company