Why build your own type system?
In Ceylon 1.2 we've factored out the type system of Ceylon
as an independent module, with minimal dependencies and a
clean API. The
- an extensible object-oriented model of the type system in Ceylon,
- algorithms for reasoning about types at compile time—or even at runtime in a system with reified generics—and
- a framework for model loading, that is, for loading a model representing the types in a program or module from binary artifacts, source artifacts, or anything else—this is a necessary feature for incremental compilation and modularity.
Why might this be interesting to someone outside the Ceylon
team? Well, if you're thinking of creating a programming
language with nominal typing, you could consider reusing
ceylon-model in your own language, saving yourself the
problem of implementing all the hairy typing algorithms that
we've already spent years getting right. That way, you can
concentrate on the stuff that's really different about your
language, that is, the syntax, the compiler backend, the
runtime semantics, and the basic language types.
ceylon-model implements an extremely complete type
system, including all the following features:
- nominal typing with single-inheritance of classes and multiple-inheritance of interfaces,
- bounded parametric polymorphism (generics with constraints),
- sum types,
- union and intersection types, with sophisticated reasoning about principal types and canonicalization,
- reasoning about subtyping,
- reasoning about coverage and disjoint types,
- principal instantiation inheritance (more powerful than the single-instantiation inheritance found in other languages with subtyping and generics),
- declaration-site and use-site variance,
- type constructor polymorphism (higher-kinded types),
- higher-rank / impredicative polymorphism,
- type aliases including generic type aliases,
- self types, and
- some support for overloading.
It doesn't include any functionality for type inference, flow-sensitive typing, or pattern matching, since in Ceylon that stuff is implemented within the typechecker, and anyway your language probably has different rules to ours.
It seems to me that unless your type system is based on
structural typing or dependent types,
ceylon-model is an
awesome starting point. It might already have everything you
need, and, even if it doesn't, it provides a framework that
would let you incorporate other features into the type
system. For example, I imagine that primitively-defined
function, array, tuple, or record types would be easy to
We're already reusing this model between the typechecker, two compiler backends, two IDE projects, and the runtime, so we know its API is reasonable.
The only stumbling block is that the project is implemented in Java, so if you want to use it in a compiler written in a different language, you'll probably want to translate it to that language first. The code is pretty clean and straightforward, so that's probably not a massive hurdle to overcome.
The license for
ceylon-model is Apache 2.0, so you can
basically do what you like with it. But if you decide to
use it in a project, please drop us a note, because it makes
us happy to see people reuse our work.