cxx

cxx copied to clipboard

WIP, but potentially working already; needs judicious tests. Addresses part of #682.

Based on the sketch in #682, the to-do list is roughly

• [X] Centralize atom spotting into Types
  • Remaining: parse_repr_attribute, parse_int_suffix, implicit_impl, and centralized spotting; all seem legit
• [X] Only spot atoms that don't already have a user-defined resolution
• [ ] Test a bunch of cases to ensure that it works

If anyone would like to pick this up, feel free to do so (and/or PR tests onto my branch)!

Note that this does not handle builtins (that is, Box, UniquePtr, SharedPtr, WeakPtr, Vec, CxxVector, and str), which are different from atoms (that is, bool, c_char, uNN, iNN, fNN, CxxString, String). Those will take more thorough refactoring to weaken similarly.

Thanks for working on this!

The difference between an atom and a builtin is that an atom is a non-generic type, and a builtin is a generic type^[1]. As such, builtins are parsed specially, and are unique kinds of Types.

As such, my proposal for how to handle builtin shadowing is a bit interesting: don't! Or rather,

• Don't prohibit use of builtin names in check_reserved_name anymore.
• After type collection, spot every use of builtin names as idents, along with spotting atoms.
• If a builtin name is used as an ident but not defined, instead of an "undefined type" error, emit a helpful "did you mean the builtin" error.
• If a builtin name is used as an ident and defined, scan for any uses of the actual generic builtin type, and if found, emit an error indicating the name collision.

The key insight here is that cxx doesn't actually mind if we define types with the reserved builtin names (as far as I understand it); the issue comes for the user juggling that UniquePtr<T> and UniquePtr are disparate types. As such, we can "just" allow both types to be defined, and then emit an error that this is weird, rather than building a more sophisticated system supporting more complicated cases.

[1]: str isn't. I say just leave str as a reserved name. Unfortunately bailing out here... means that CStr is also put in a similar spot. I think the full solution might be to not use Type::Str, but instead Type::Ref to an atom str, and the ref-atom is handled specially for str/CStr atoms.