Typedef chains via __ symbols cause errors

[adetaylor]

Build the s2 example with #438 and no generate! directives.

See:

unsupported type: pid_t", "unsupported type: FILE", "unsupported type: fpos_t", "unsupported type: clock_t", "unsupported type: time_t", "unsupported type: wint_t", "unsupported type: wctype_t", "unsupported type: mbstate_t", "unsupported type: ssize_t", "unsupported type: id_t", "unsupported type: siginfo_t", "unsupported type: dev_t", "unsupported type: mode_t", "unsupported type: pthread_attr_t", "unsupported type: pthread_t", "unsupported type: pthread_cond_t", "unsupported type: pthread_condattr_t", "unsupported type: pthread_mutex_t", "unsupported type: pthread_key_t", "unsupported type: pthread_mutexattr_t", "unsupported type: pthread_rwlock_t", "unsupported type: pthread_rwlockattr_t", "unsupported type: mach_port_t", "unsupported type: sigset_t", "unsupported type: wctrans_t", "unsupported type: nl_catd", "unsupported type: u128"

Rather than have a blocklist, I think we should ignore functions that refer to any type which isn't in known_types or specified in bindgen.

[adetaylor]

Example function:

 pub fn wait(arg1: *mut ::std::os::raw::c_int) -> root::pid_t;

[dtolnay]

Relevant cxx design sketch for supporting things like pid_t: dtolnay/cxx#529.

[adetaylor]

This turns out to be nothing to do with anything above, and it's an interaction between #437 and #220.

Types undergo conversion from a bindgen format (e.g. root::std::unique_ptr<root::std::string>) to a cxx format (e.g. UniquePtr<CxxString>). This is done in a struct called TypeConverter.

The TypeConverter struct stores typedefs, as a map from type name -> syn::Type. The syn::Type is converted at insertion time. (Sometimes the syn::Type is another Type::Path, which could refer to another typedef.)

Then, when we encounter a function that uses a type which has a typedef name, it gets resolved recursively until we end up with something which isn't a typedef. That's what ends up in #[cxx::bridge].

Something similar happens with fields of POD structs, but I can't remember when this resolution occurs.

So far so... grudgingly OK.

The problem happens because in #437 TypeConverter now rejects types which contain __ because cxx doesn't like them. This needs to happen only for types which are actually going to be listed in the #[cxx::bridge] mod. That means the final thing to which the chain of typedefs actually resolves. In the case of the test in #445, pid_t is fine, and so is int to which it eventually resolves. Intermediate types are things like __int32_t or __darwin_pid_t which are not fine.

The specific issue is fixed here: https://github.com/google/autocxx/compare/double-underscore-typedefs-wip but this needs significant rework:

• The API logging stuff needs to be cleaned up and abstracted to its own PR, hopefully with a way to autogenerate the enum variant descriptors.
• There are now enough parameters to the various type conversion functions that I think we should introduce a TypeConversionOptions struct.
• I'm not sure this fix is exactly right. There might be cases when we do need to check the __-compliance of idents in typedefs under some circumstances. Needs more thought.
• TypeConverter needs nuking from orbit per Isolate TypeConverter into an analysis phase #220.

[adetaylor]

Further problems encountered in #452.