Type directed array disambiguation

rfcs/type_directed_array_disambiguation.md

@@ -0,0 +1,68 @@
+# Type-directed Array Disambiguation
+
+This RFC describes a way we can use literal syntax for array-like structures
+other than the built-in `array` type, hooking into the existing support for
+type-directed disambiguation of constructors.
+
+## Motivation
+
+OCaml currently supports two different array-like structures, `array`, and
+`floatarray`. There are discussions about adding [immutable
+arrays](https://github.com/ocaml/ocaml/pull/13097), and I believe several people
+have expressed a desire for compiler support for uniform arrays (which do not
+participate in the float-array opimization). (This is just hearsay; I do not
+have a reference.) Yet only `array` has first-class syntax, available both for
+construction and for pattern-matching. This RFC allows us to use this nice
+syntax for these other structures, making it easier to adopt these other types
+in code.
+
+We could also imagine extending this syntax to cover `string` and `bytes`, but
+this proposal as written does not do so.
+
+## Proposal
+
+During type-checking, when encountering array-literal syntax, in expressions or
+in patterns, check to see whether the expected type is known (principally known,
+in `-principal` mode). Then:
+
+* If the type is not known, the literal has type `'a array`, as now.
+* If the type is known to be ...
+  * ... `_ array`, the literal is treated as it is now.
+  * ... `floatarray`, the literal is treated as a `floatarray`.
+  * any other type, error.
+
+If the language begins to support immutable or uniform arrays (or immutable
+`floatarray`s or immutable uniform arrays), then disambiguation would work for
+these structures, too.
+
+## Discussion
+
+* This disambiguation does *not* extend to lists: list literals are just syntax
+for nested calls to `(::)` and `[]`, which can be disambiguated through the
+usual mechanisms for constructors.
+
+* This allows us to add new array-like structures without worrying about
+  concrete syntax.
+
+* The `[| ... |]` syntax behaves very much like a constructor, available both in
+  expressions and in patterns. Extending type-directed disambiguation in this
+  way seems like a natural step.
+
+* This disambiguation does *not* extend to array access operators `.()` and
+  `.()<-`. The former is currently an abbreviation for `Array.get` (whichever
+  `Array.get` is in scope) and the latter is an abbreviation for `Array.set`
+  (whichever `Array.set` is in scope). In addition, the language already
+  supports binding new array-access operators, with (mostly) arbitrary symbols
+  between the `.` and the brackets. These existing mechanisms serve well to
+  control the meaning of access operators, and thus this RFC does not add to
+  them.
+
+## Implementation
+
+There is currently no implementation of this feature (even in prototype). Jane
+Street will either implement or will fund the implementation, if this RFC is
+accepted. If implementation / testing reveals unexpected interactions /
+complications, we would naturally expect to re-engage with the community for
+discussion of whether this is, in practice, a good idea. (That is, we propose
+that acceptance of this RFC is conditional on having a smooth implementation
+experience.)