Thanks for the great explanation, I agree with this advice. I'm going to
investigate writing a new patch to do this.
On Fri, Nov 18, 2016 at 1:59 PM, Arthur O'Dwyer <arthur.j.odw...@gmail.com>
wrote:
> On Thu, Nov 17, 2016 at 2:14 PM, Sasha Bermeister <sas...@chromium.org>
> wrote:
>
>> Although I agree with your philosophical discussion and suggestions, the
>> reality is that MSVC's behavior is not a bug and compilers are free to
>> interpret enum bitfields with no explicit underlying type in any way they
>> want (see spec reference in GCC bug link), with a signed interpretation
>> being a valid one. I'd say it's undefined behavior in C/C++ to store an
>> enum in a bitfield without specifying an underlying type, since the
>> compiler is free to interpret this bitfield in any way it wants -- in
>> general, if you haven't specified an underlying type you should probably be
>> warned when trying to store it in a bitfield because the compiler may not
>> do what you expect.
>>
>
> Incorrect. The following program has perfectly well defined behavior:
>
> enum E { e = 0 };
> struct S { E bf : 4; } s;
> int main() { s.bf = e; }
>
> No compiler in the world should produce a warning on the above program.
>
> Also, once you've got a struct type containing an offending bit-field,
> *any* use of that bit-field is subject to the implementation-defined
> behavior you noticed on MSVC. It's not just limited to
> assignment-expressions of constants. That's why it's important to produce a
> warning on the *declaration* of the bit-field, not on each subsequent
> expression that refers to that bit-field.
>
> enum E2 { e = 0, f = 1, g = 2, h = 3 };
> struct S2 { E2 bf : 2; } s; // this line should trigger a diagnostic
> int main() { s.bf = e; }
>
> Also, the current patch's diagnostic wording suggests to "consider giving
> the enum E an unsigned underlying type", which would be very bad advice in
> this situation (because it only works in C++11, and because it triggers a
> GCC bug, and because it has non-local effects on the program's semantics).
> The correct advice is to "consider giving the bit-field bf a width of 3
> bits instead of 2."
>
> HTH,
> –Arthur
>
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
