Monday, December 31, 2012

The wide ranging impact of design decisions - Free Functions

I talked about this problem a little bit in my last post, but wanted to elaborate on it a bit since it's an interesting lesson learned. The issue in question is the general desire to allow free functions to satisfy interface requirements in Tangent. Say you had a third party library that defines a simple connection interface:
Connection => class {
  Open => void;
  Close => void;
  Read => IEnumerable<byte>;
};
But you wanted to use it with a different third party library that provided a connectionless version:
Reader => class {
  Read => IEnumerable<byte>;
};
Sadly, you can't just inherit from Reader. It sits off in some 3rd party library and you don't have control over how it gets instantiated. In these cases, you'll end up having to use some icky wrapper to make stuff work. Tangent makes this better, but still a lot of boilerplate:
adapt (r: Reader) to Connection => Connection with class {
  Open => void {}
  Close => void {}
  Read => IEnumerable<byte> { return r.Read; }
};
So since the language allows free functions, this should work:
(r: Reader).Open => void {}
(r: Reader).Close => void {}
Since Reader can now satisfy the interface needed by Connection everything should be good. For this (perhaps poor) example, it's not a clear win for the free function version. But consider cases with multiple parameters as is often the case with infix operators. It quickly becomes less clear how to do the adaptation for each of the different parameters involved; as well as who should own the extension. Why doesn't this just work? Because the Open and Close extensions might not be visible in certain scopes. Once that happens, the type checker can be incorrect; verifying that Reader satisfies Connection in one scope and then once it's passed into a function, it suddenly doesn't. Beyond that, this general idea that free functions can satisfy interfaces had a much larger implication I had missed. It means the subtyping check cannot simply accept two types as its inputs anymore. Even if the types have a list of functions that they require, those functions aren't the only ones that can satisfy an interface. Indeed, any function that works with the type can satisfy an interface we're checking. This is nifty and powerful, but means the compiler actually has to do it. But that's not all. Consider:
A: class {
  foo => void;
};

B: class {
  bar => void;
};

C: class {};

(b: B).foo => void {}
(c: C).bar => void {}
Is C a subtype of A? Sure. It ends up satisfying those constraints, but the sub-type function actually needs to handle this case, including the sort of loops it can get into because to see if a type applies, it needs to see if a method applies, which causes it to see if a type applies, which... So when people ask why aren't user defined operators more common, I think of this sort of rabbit hole that a design decision leads to and cannot help think that people far smarter than I knew this decades ago which led to languages (by and large) not mixing infix operators and objects.