On Tue, Aug 15, 2023 at 8:54 AM Andrea Righi <andrea.righi@xxxxxxxxxxxxx> wrote:
>
> Commit df8fc4e934c1 ("kbuild: Enable -fstrict-flex-arrays=3") enabled
> '-fstrict-flex-arrays=3' globally, but bindgen does not recognized this
It may be more accurate to say libclang here (bindgen forwards the options).
Also, df8fc4e934c1 did it so only conditionally (if the C compiler
supports it). This explains what you are seeing: if I am right, you
are compiling with a modern enough GCC, which enables the option, but
with an old enough Clang.
> compiler option, triggering the following build error:
>
> error: unknown argument: '-fstrict-flex-arrays=3', err: true
This should only be true with libclang < 16, since Clang 16
implemented the option, right?
In fact, Clang 15 seems to work too -- it seems the compiler does not
error if the option is not within [0,3] (unlike GCC).
Kees: this should only affect `__builtin_object_size` and not `sizeof`, right?
>From a quick test across Clang 14/15/16 and different levels for the
flag (0/1/2/3/4 and no flag), bindgen seems to generate the same
output in all cases: `__IncompleteArrayField` is always generated for
`[]` and `[0]`; and never for `[1]`, i.e. regardless of the flag. The
`sizeof`s agree with the C side, and they are all the same (as
expected because this only changes BOS).
So I think the patch contents (i.e. ignoring the flag for bindgen)
should be fine. Except we could have somewhere a
`__builtin_object_size` influencing a layout. Since GCC seems to
always treat it as not a constexpr (unlike Clang), I assume nobody is
using it like that (since we compile the kernel with and without the
flag). But it is still possible in theory -- it would require
something like:
struct X {
int x[__is_constexpr(__builtin_object_size(s->c, 1)) ? 1 : 2];
};
to make GCC and Clang (as well as Clang `=3` vs. Clang without the
flag) to compile but disagree on the size:
https://godbolt.org/z/6TqPjGK3d
Cheers,
Miguel
