Re: [PATCH 3/5] Drivers: hv: kvp: Fix the recent regression caused by incorrect clean-up

[Miguel Ojeda]

+On Wed, Oct 17, 2018 at 8:25 AM Dan Carpenter <dan.carpenter@xxxxxxxxxx> wrote:
>
> It's not common at all.  It should be wrapped in a macro and put into
> compiler.h.
>
> But I hope it does become adopted.  It's better than randomly grepping
> for non-standard comments.

Using an attribute is indeed better whenever possible. In C++17 it is
an standard attribute and there have been proposals to include some of
them for C as well since 2016 at least, e.g. the latest for
fallthrough at:

  http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2268.pdf

I have taken a quick look into supporting it (typing it here to save
it on the mailing list :-), and we have:

  * gcc >= 7.1 supports -Wimplicit-fallthrough with
__attribute__((fallthrough)), the comment syntax and the C++
[[fallthrough]] syntax.
  * gcc < 7.1 complains about empty declarations (it does not know
about attributes for null statements) and also
-Wdeclaration-after-statement.
  * clang 7 supports -Wimplicit-fallthrough (not enabled in
-Wall/extra/pedantic like gcc, though) but *only* in C++ mode and with
the C++ syntax [[fallthrough]]. In other words, in C mode, no syntax
works and no diagnostics are emitted. It complains about
Wmissing-declarations. [IMO they should allow the __attribute__ syntax
for fallthrough (and enable it on C mode) to be compatible with gcc.
Maybe they are simply waiting for the C2x attributes... :-)]
  * icc 19 does not know about -Wimplicit-fallthrough at all (but
seems to allow [[fallthrough]] on C++17 mode to comply with the
standard).

Therefore, the only improvement we could do right now is starting to
use the attribute for gcc > 7.1, and a comment for everybody else.
However, even if that was worth the trouble of changing the 2500+
instances of fall through markings that we have, comments are replaced
before the preprocessor stage, so we would need some (probably
non-portable) macro magic.

So, I would say, let's revisit this again in a few years. Possibly,
when we move the minimum version to gcc 7.1, clang and icc may both
support the fallthrough warning in C mode; so that we can always use
the __attribute__((fallthrough)) unconditionally under a __fallthrough
#define.

Cheers,
Miguel