MEMORY: Protect against coercion ("gradual typing")

[HenrikBengtsson]

Wish

> x <- integer(10)
> allow_coercion(x) <- FALSE
> x[1] <- 2
Error: Cannot assign value of type 'double' to an object of type 'integer' without coercion.

Ideas

Hmm... I need to think about this one. It might be that this has to be done deep down in core R;

• tracemem() will detect it, or more specifically, it'll output text to stdout that carry information/traces of coercions, but I don't see how to capture such output automatically.

• It's semi doable by trace():ing [ but that is of course not a general solution and it won't detect intermediate coercions occurring internally.

  • EDIT 2018-02-10: though trace() could run a temporary tracemem() with capture.output() upon entering the function, meaning it should be able to detect also intermediate coercions.
  • EDIT 2018-02-15: trace() cannot be stacked/nested, i.e. using it may introduce unwanted side effects.

• Coercion in core R takes place in https://github.com/wch/r-source/blob/trunk/src/main/coerce.c

Related

• gradual typing, e.g. mental programming model going from dynamic typing to static typing as code matures/settles. (http://willcrichton.net/notes/gradual-programming/)

[hammer]

Jan Vitek has given talks on this topic at Directions in Statistical Computing (DSC):

• DSC 2016: This is not a type: Gradual typing for R
• DSC 2017: @:annotations

Within Jan's group, aviralg seems to be implementing these ideas in annotatr, contractr, and profilr. He's also making use of PRL-PRG/typeR, which looks nice.

[lionel-]

It's possible to do this with classes. The vctrs package developed a class of vectors with stronger coercion semantics. This is all runtime checking of types though, so a different idea than gradual typing.