Defaults for multi-argument dispatch

10 views
Skip to first unread message

Konrad Hinsen

unread,
Mar 20, 2009, 10:56:49 AM3/20/09
to Clojure
Providing a :default implementation for multimethods is a very common
and useful technique, but it is really useful only for multimethods
that dispatch on a single argument. What I am looking for is an
equivalent technique for multiple-argument dispatch.

Suppose you have a multimethod + of two arguments, and you want to
dispatch on both of them:

(defmulti + (fn [x y] [(type x) (type y)]))

You can then write implementations such as

(defmethod + [java.lang.Integer java.lang.Double] ...)

You can also provide a default implementation, of course:

(defmethod + :default ...)

But suppose you want to provide a default for one argument only?
Something like

(defmethod + [java.lang.Integer ::any] ...)

i.e. a multimethod that matches all invocations in which the first
argument is an integer. I don't currently see a simple way to do
this. For types in the Java class hierarchy, you can use Object as
the parent of all types, but there is nothing equivalent in Clojure's
ad-hoc hierarchies.

Would it be a good idea to provide the possiblity to add a universal
parent to hierarchies? Or would that create any problems? Is there
another solution for the situation I described?

Konrad.

Rich Hickey

unread,
Mar 20, 2009, 11:18:22 AM3/20/09
to Clojure


On Mar 20, 10:56 am, Konrad Hinsen <konrad.hin...@laposte.net> wrote:
> Providing a :default implementation for multimethods is a very common
> and useful technique, but it is really useful only for multimethods
> that dispatch on a single argument.

I disagree about that. No dispatch value, composite or not, is still a
valid concept.

> What I am looking for is an
> equivalent technique for multiple-argument dispatch.
>
> Suppose you have a multimethod + of two arguments, and you want to
> dispatch on both of them:
>
> (defmulti + (fn [x y] [(type x) (type y)]))
>
> You can then write implementations such as
>
> (defmethod + [java.lang.Integer java.lang.Double] ...)
>
> You can also provide a default implementation, of course:
>
> (defmethod + :default ...)
>
> But suppose you want to provide a default for one argument only?
> Something like
>
> (defmethod + [java.lang.Integer ::any] ...)
>

I think it is best to think about this differently than :default, it's
more about a universal parent than about a missing dispatch value.

> i.e. a multimethod that matches all invocations in which the first
> argument is an integer. I don't currently see a simple way to do
> this. For types in the Java class hierarchy, you can use Object as
> the parent of all types, but there is nothing equivalent in Clojure's
> ad-hoc hierarchies.
>
> Would it be a good idea to provide the possiblity to add a universal
> parent to hierarchies? Or would that create any problems? Is there
> another solution for the situation I described?
>

Yes, don't know, and no. I briefly looked at this but only got as far
as to decide Object couldn't be the universal parent. I think you have
to reserve a value that will never otherwise be used.

Rich

Paul Stadig

unread,
Mar 20, 2009, 11:35:12 AM3/20/09
to clo...@googlegroups.com
You could use multiple multi-methods:

user=> (defmulti plus-int (fn [x y] (type y)))
#'user/plus-int
user=> (defmethod plus-int :default [x y] (println "the first is an int"))
#<MultiFn clojure.lang.MultiFn@11992cc>
user=> (defmethod plus-int java.lang.Double [x y] (println "one of each"))
#<MultiFn clojure.lang.MultiFn@11992cc>
user=> (defmulti plus (fn [x y] (type x)))
#'user/plus
user=> (defmethod plus java.lang.Integer [x y] (plus-int x y))
#<MultiFn clojure.lang.MultiFn@40f603>
user=> (plus 1 5)
the first is an int
nil
user=> (plus 1 3.0)
one of each
nil
user=>

Not very pretty. (Written backwards courtesy of the REPL ;))


Paul

David Nolen

unread,
Mar 20, 2009, 11:38:19 AM3/20/09
to clo...@googlegroups.com
This works well as long as you don't mind the perf hit for the second dispatch :)

Paul Stadig

unread,
Mar 20, 2009, 12:15:36 PM3/20/09
to clo...@googlegroups.com
It's also not as maintainable as using a single multi-method.

Like I said, not pretty, but it works.


Paul

Konrad Hinsen

unread,
Mar 20, 2009, 1:31:17 PM3/20/09
to clo...@googlegroups.com
On Mar 20, 2009, at 16:35, Paul Stadig wrote:

> You could use multiple multi-methods:

...

Not pretty, as you said, but also not quite the same in behaviour as
a single multimethod dispatching on both arguments. Multiple dispatch
can be made symmetric in the arguments, whereas a chain of
multimethods always implies a dispatch hierarchy. In your example,
the type of the first argument always dominates.

Konrad.

Konrad Hinsen

unread,
Mar 20, 2009, 1:35:37 PM3/20/09
to clo...@googlegroups.com
On Mar 20, 2009, at 16:18, Rich Hickey wrote:

>> Providing a :default implementation for multimethods is a very common
>> and useful technique, but it is really useful only for multimethods
>> that dispatch on a single argument.
>
> I disagree about that. No dispatch value, composite or not, is still a
> valid concept.

True, for reasons other than a default in a type hierarchy.

>> But suppose you want to provide a default for one argument only?
>> Something like
>>
>> (defmethod + [java.lang.Integer ::any] ...)
>>
>
> I think it is best to think about this differently than :default, it's
> more about a universal parent than about a missing dispatch value.

It could be seen from both points of view (universal parent in the
hierarchy, or a partial match with a default), but I agree that the
universal parent point of view makes more sense.

>> Would it be a good idea to provide the possiblity to add a universal
>> parent to hierarchies? Or would that create any problems? Is there
>> another solution for the situation I described?
>
> Yes, don't know, and no. I briefly looked at this but only got as far
> as to decide Object couldn't be the universal parent. I think you have
> to reserve a value that will never otherwise be used.

Object would indeed not work, the universal parent would have to be
even above Object.

I just looked at the implementation of hierarchies and I have the
impression that this should be rather simple to implement. I will try
and see how it works out in practice.

Konrad.

Reply all
Reply to author
Forward
0 new messages