| > I imagine shoehorning all numbers into doubles makes for slower integer code. Maybe, though in a bytecode-compiled, dynamically typed language, my hunch is that using doubles for everything is actually faster. If I had separate integer and float types, all of the arithmetic operators would have do runtime type tests to handle the different combinations and conversions. I could be wrong, though. Either way, only having doubles definitely simplifies the VM. > Also a shame you included pervasive null. Yeah, it doesn't come up too often in practice. Unlike, say, JS, Wren considers it a runtime error to call a function without enough arguments. I'm not a big fan of null, but in a dynamically typed language it's sort of inevitable. It's the result type of a function with no explicit return. > Wren looks straightforward enough that you could have a static type system without losing any expressiveness. Have you considered that? Yes! I'm interested in trying to layer an optional type system on top of it, and I tried to design most of the language and core library such that it's amenable to that. However, designing a type system is really hard, so I don't know how well it'll pan out. Now that the language itself is pretty far along and you can write useful programs in it, I'm hoping to start tinkering with this and see what I can come up with. You can see the first baby steps here: https://github.com/munificent/wrenalyzer |
Perhaps. I suppose this dovetails with my preference for static types with optional dynamic behaviour, and not the contrary default which seems to be your position.
> I'm interested in trying to layer an optional type system on top of it, and I tried to design most of the language and core library such that it's amenable to that.
Cool. I think Wren is simple enough that basic type inference should be straightforward and complete, and anything more difficult just infers a top/'dynamic' type. A trivially correct type inference algorithm then just infers type 'dynamic' everywhere, which recovers Wren's current semantics. Then you incrementally add more specific rules where you can, ie. simple data flow from constants will generate more efficient string, floating point and integer code.
That would be enough for a decent speedup on most programs. If you want to type check arbitrary objects (or at least avoid method-not-understood), you can use something like "Type Inference for First-Class Messages with Match-Functions" [1] or the even more general first-class labels [2]. I believe first-class labels are more general than Wren's semantics.
[1] http://lambda-the-ultimate.org/node/4808 [2] http://lambda-the-ultimate.org/node/174