Lisp Corner Cases: Method Combinations
In the process of writing Declt, I had to deepen my knowledge of some Lisp corner cases, notably in the area of introspection. As you know, Lisp has in fact more than 2 namespaces. Sometimes, introspecting a potential symbol definition in one namespace is trivial. COMPILER-MACRO-FUNCTION is one example. The "functional" namespace is heterogeneous but you can still make your way out of macros, regular or generic functions very easily. In the same vein, distinguishing constants, special variables and symbol macros in the "variables" namespace is more complicated because there is no standard way to access that information, but with the help of some vendor-specific machinery (e.g. SB-INT:INFO), it's still doable.
At some point, I tackled the case of method combinations, and to my greatest surprise, found out that introspecting a method combination by name is not simple. In fact, it's not even doable. The reason is that method combinations don't have a namespace proper. Let me explain why.
You define a new method combination of some NAME with DEFINE-METHOD-COMBINATION, and you use it with the :METHOD-COMBINATION option to DEFGENERIC. But how do you introspect a NAME for a potential method combination definition? Well, you can't do exactly that. First, there is no standard way to do so. Next, let's look at what the MOP provides. One would expect something along the lines of FIND-CLASS...
There is indeed something called FIND-METHOD-COMBINATION, but either I'm missing something, or it is really, and I mean really badly designed. The arguments to this function are: a generic function meta-object, a method combination name and some method combination options. So if this function is supposed to be the equivalent of FIND-CLASS for method combinations, what in the hell are the 1st and 3rd arguments for? In fact, it's not supposed to do what its name suggests. According to the AMOP, p.191, the purpose of this function is to "determine the method combination object used by a generic function". In practice however (at least in SBCL), it's not doing that either (see below), and anyway, determining the method combination object used by a generic function is better achieved by using the GENERIC-FUNCTION-METHOD-COMBINATION accessor.
So this protocol doesn't seem to make any sense. In order to understand what to make of all this, I looked at how SBCL handles method combinations (I don't know what other vendors do) and here is what I found. When you call DEFINE-METHOD-COMBINATION for some NAME, SBCL creates a new method for FIND-METHOD-COMBINATION, eql-specialized on that NAME. This method is encapsulated in a closure containing the method combination's definition, ignores its first argument (the generic function meta-object; that's why I said that it's not really doing what the AMOP says it does), recreates and returns a new method combination object on the fly every time it is called.
This has several consequences. First, the notion of "method combination named NAME" doesn't really make sense. Method combinations don't have a proper namespace. Every time you create a new generic function, the method combination essentially becomes local to that function. Redefining a method combination has no effect on existing generic functions using a method combination of the same NAME. To put it differently, you can end up with several generic functions using different yet equally named method combinations.
In Declt, I'm now assuming that programmers behave and only define method combinations once. Which brings us to the second consequence. With that assumption in mind, the "proper" way to introspect a NAME for a method combination definition (at least in SBCL) is to first look for the existence of a FIND-METHOD-COMBINATION method eql-specialized on that NAME, and then call it to retrieve the actual definition.
I haven't thought about it a lot yet, but it would be interesting to investigate the idea of cleaning this up a bit. I mean, establishing a real global namespace for method combinations, creating an alternate FIND-METHOD-COMBINATION protocol more along the lines of FIND-CLASS. And then, it could also be interesting to investigate on the potential applications of allowing method combination redefinition to affect live generic functions, just like class redefinition affects live instances...