The last two posts introduced the fundamental elements underlying Cone’s type system. They also distinguished between nominal vs. structural type equivalence. However, they were largely silent about subtyping relationships between types. Given how important subtyping is to Cone’s type versatility, I thought it might also be valuable to offer a detailed treatment of these mechanisms. My approach will be practical, rather than formal (alas! no subsumptions nor lattices), distilling lessons learned from implementing a rich, but type-safe collection of subtyping mechanisms.
The prior post described the semantics underlying Cone’s fundamental types, for all types except the all-important pointer and reference types. The presence of versatile pointers/references as an explicit and distinctly-separate kind of type is a distinguishing characteristic of systems programming languages. I choose to cover them here as a separate post, since the safety guarantees of Cone’s references make them semantically rich and complex. In harmony with the subatomic theme of the prior post, reference types are themselves composed of “quarks”: pointers, regions, permissions, and lifetimes.
When I began work on the Cone programming language, I believed I had a confident handle on which primitive types this modern, safe systems language required: various sizes of atomic number types, fixed-size arrays, programmer-defined structs, raw pointers, safe references, discriminated (and raw) unions, generics, and limited support for void and tuples. My confidence was misplaced. Once or twice a year, implementation challenges and further research forced me to re-think aspects of the type system, each time deepening my understanding of what it required.
When the clock ticks over to a new decade, it is customary to look back, to reflect on how much we have accomplished, and then look forward, to sort out where we want to go. Ten years is long enough that substantive progress should be visible in the glacially-slow evolutionary pace of programming languages. One can see this by noticing how many now-influential languages had no notable marketplace presence only ten years ago: Rust, Go, Swift, Kotlin, Dart, and Julia.
Type theory has many kinds of typing mechanisms: product, sum, top, bottom, recursive, universal, existential, subtypes, linear, refinement, dependent, and so on. I have run across something different from all those, which I am calling “infectious typing”. If anyone knows a formal name and/or treatment for this mechanism, I would love to hear about it. Type composition allows us to create new types out of existing ones. A struct or class product types is composed of several fields, each with its own type.
In 2001, Trevor Jim (AT&T Research) and Greg Morrisett (Cornell) launched a joint project to develop a safe dialect of the C programming language, an outgrowth of earlier work on Typed Assembly Language. After five years of hard work and some published papers, the team (including Dan Grossman, Michael Hicks, Nik Swamy, and others) released Cyclone 1.0. And then the developers moved on to other things. Few have heard of Cyclone and almost no one has used it.
The prior post, When Sum Types Inherit, shows that we can (and should) use the magic of inheritance to enrich sum types to offer as powerful a capability as traditional classes. Graph-based, variant node data structures benefit from variant nodes being able to support common fields, methods, and virtual dispatch. That result, however, raises a new question: do we really need two separate language abstractions that offer very similar capabilities?
Inheritance makes it easier than any other mechanism (e.g. generics, macros, composition/delegation) to define a type that reuses the state and some methods of other types. After reading my inheritance posts, I hope you are convinced that simplifying inheritance to a namespace-based mechanism ensures we obtain this convenient reuse capability, while avoiding most of the complexity and coupling dangers of traditional inheritance. However, you might still wonder whether real-world code needs inheritance’s reuse capability.
After removing the interface, inversion of control, and protected access capabilities from traditional inheritance, what do we have left (besides composition)? This is what we have: placing a few extra tokens on a derived class causes all named fields and methods of one or more base classes to be absorbed as if explicitly incorporated. Further, certain inherited methods can be customized (overridden) with their own implementation. The primary selling point for inheritance has always been this sort of code reuse.