Since this is internal to the compiler it can easily be improved later. But why not, in the interest of generality, provide a module Targetint with the same interface as Nativeint (augmented with your word_size and of_int_exn and other needed utilities) and defined by

module type TARGETINT = sig ... end
module Targetint : TARGETINT = struct
include
  (val
    (match Sys.word_size with
    | 32 -> (module Int32)
    | 64 -> (module Int64)
    | _ -> assert false) : T)
end

This would make it easy to replace Nativeint by Targetint in the native code backends.

I was planning to have such an "int" module, but haven't had time. If someone wants to make a pull request against my branch I'll happily include it.

I fully agree with @nojb 's proposal: in order for this to be reusable for cross-compilation, we need a compilation unit called Targetint (the name Address doesn't make much sense), obtained by selection between first-class modules corresponding to Int32 and Int64 enriched with whatever extra operations debugger support needs. The type Targetint.t should be abstract, not a sum of int32 and int64.

Let's try to work as a group to produce this, rather than pushing the strict minimum needed for debugging support.

@xavierleroy : My original idea was that Target_system would have (at least) two submodules: one Int and one Address. The rationale for that was to cope with any cases where the target's addressing range might be substantively different from the range of machine integers. I am now thinking that it's maybe a mistake two have two such types, and instead, in the "target integer" module have some values that define e.g. the maximum representable address. What do you think?

@mshinwell : I'm not sure what you have in mind with "the target's addressing range might be substantively different from the range of machine integers". Since the birth of OCaml we've assumed that pointers/addresses, tagged integers, and native integers have the same bit size (32 or 64). I don't think we could make good use of a platform like x32 where pointers are 32 bits and integers are 64 bits. So, I fell that Targetint is enough and we don't need Targetaddress in addition.

@xavierleroy The x32 case was the sort of thing I was thinking of. Will comment on the other GPR and close this one.