Problem

Bindgen supports different enum representations which are either marked repr(transparent) or repr($ty) where $ty is the representation used for the enum (an integer type or an alias for one). However, for cross-language CFI to work correctly, it is sometimes desirable to use repr(C) instead as it causes the compiler to emit the right type ID for cross-language CFI to work.

Example:

enum handler_return {
    INT_NO_RESCHEDULE = 0,
    INT_RESCHEDULE,
};

gives us the following bindings (using --rustified-enum handler_return):

// in context of cross-language CFI, the return type is `core::ffi::c_uint` but
// the expected type is `enum handler_return`) thus an indirect function
// call using this type will abort
#[repr(u32)]
#[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
pub enum handler_return {
    INT_NO_RESCHEDULE = 0,
    INT_RESCHEDULE = 1,
}

The following two variations work:

// this works but isn't supported by bindgen as far as I can tell
// (e.g. can't pass `repr(c)` using `--with-attribute-custom-enum ` flag)
#[repr(C)]
#[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
pub enum handler_return {
    INT_NO_RESCHEDULE = 0,
    INT_RESCHEDULE = 1,
}

// this also works but requires the user to know about the 
// C++ Itanium ABI/cross-lang CFI internals - not ideal...
#[repr(u32)]
#[cfi_encoding = "14handler_return"]
#[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
pub enum handler_return {
    INT_NO_RESCHEDULE = 0,
    INT_RESCHEDULE = 1,
}

Using --newtype-enum handler_return fails the CFI check too:

// fails cfi check
#[repr(transparent)]
#[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
pub struct handler_return(pub core::ffi::c_uint);

and repr(C) or cfi_encoding = "14handler_return" attributes make the CFI check pass again.

Potential Solutions

1. Add new flags to request repr(C) for enums, e.g. --rustified-enum-repr, --newtype-enum-repr, etc.
2. Allow --with-attribute-custom-enum REGEX=#[repr(C)]
3. No action; require the user to use #[cfi_encoding] or change the LLVM cross-language CFI implementation.

The first solution complicates the bindgen CLI but the functionality is easily discoverable. The second solution makes the inverse tradeoff: more can be done with the existing flags but few people will discover that this is the case.

I have prototyped the second approach and can put up a draft PR unless there's a feeling it is not the right way to go. Also very interested in hearing about potentially better solutions I just haven't thought of.

cc @rcvalle @maurer

(Edited to fix mistake pointed out by @emilio in this comment.)