`[[no_unique_address]]` on an empty member is laid out as a 1-byte field, disagreeing with clang

[lwz23]

Input C/C++ Header

struct Empty {};
struct S {
  [[no_unique_address]] Empty e;
  int x;
};
int get_x(const S *s);

int get_x(const S *s) { return s->x; }

Bindgen Invocation

$ bindgen input.h --enable-cxx-namespaces --output bindings.rs -- -x c++ -std=c++20

and the same with --no-layout-tests.

Actual Results

clang++ layout (same compiler bindgen uses):

sizeof(S)=4  offsetof(S, x)=0  sizeof(Empty)=1

Setting s.x = 7 and calling get_x(&s) from C++ returns 7.

bindgen emits Empty as a 1-byte struct and then a normal repr(C) pair of fields:

pub struct Empty { pub _address: u8 }
pub struct S {
    pub e: root::Empty,
    pub x: ::std::os::raw::c_int,
}

rustc lays that out as sizeof=8, offsetof(x)=4. The generated layout tests still assert clang's numbers (sizeof == 4, offsetof(x) == 0), so the default output fails to compile:

error[E0080]: index out of bounds: the length is 1 but the index is 4
["Size of S"][size_of::<S>() - 4usize]

With --no-layout-tests the bindings compile. A Rust caller that does s.x = 7 and calls get_x(&s) (the C++ function) gets 0, because C++ reads offset 0 and the int in Rust lives at offset 4.

Expected Results

The generated Rust type should have the same size and x offset as clang (4 and 0), so that field access and FFI agree. Empty bases are already omitted in similar cases; [[no_unique_address]] on an empty member is not.

If overlapping fields cannot be expressed in repr(C), skipping the empty member (or making S a newtype over c_int) would match clang better than a 1-byte field plus padding.

Environment

bindgen:  0.72.1
clang++:  Apple clang 21.0.0
rustc:    1.97.1
target:   aarch64-apple-darwin

Reproduced with -std=c++20. The mismatch is the generated field list vs clang's layout, not a target-specific calling convention.