Possible __VA_OPT__ bug

[Jędrzej Dudkiewicz]

On Thu, Oct 31, 2019 at 9:28 PM Edward Diener
<eldlistmailingz@tropicsoft.com> wrote:
>
> On 10/31/2019 4:20 PM, Segher Boessenkool wrote:
> > On Thu, Oct 31, 2019 at 02:50:07PM -0400, Edward Diener wrote:
> >> On 10/31/2019 1:14 PM, Segher Boessenkool wrote:
> >>> On Thu, Oct 31, 2019 at 12:42:54PM -0400, Edward Diener wrote:
> >>>> Given:
> >>>>
> >>>> #define NO_DATA
> >>>> #define TRY_VA_OPT(...)  __VA_OPT__ (0) 1
> >>>>
> >>>> TRY_VA_OPT() -> expands to 1 as expected
> >>>> TRY_VA_OPT(NO_DATA) -> expands to 0 1 which is not expected
> >>>>
> >>>> when compiled with gcc-9.2 with -std=c++2a.
> >>>
> >>> Why is that not expected?  The variadic macro TRY_VA_OPT does get tokens
> >>> in its variable argument (namely, NO_DATA), so __VA_OPT__ expands to its
> >>> argument (which is 0).
> >>>
> >>> https://gcc.gnu.org/onlinedocs/cpp/Variadic-Macros.html
> >>
> >> The wording of 15.6.1 paragraph 3 is confusing to me. I had thought that
> >> the arguments to the variadic ... parameter were completely macro
> >> expanded before considering whether the __VA_OPT__ ( pp-tokens ) would
> >> expand to 'pp_tokens' or a single placemarker token. Evidently you are
> >> saying that the correct interpretation with the __VA_OPT__ construct is
> >> that the variadic ... arguments are not macro expanded before
> >> consideration of the __VA_OPT__ construct processing . This seems to me
> >> very odd because the arguments to the variadic ... parameter are always
> >> completely macro expanded before being replaced by any __VA_ARGS__
> >> parameter in the replacement list. I wonder why the C++ standard
> >> committee decided to treat __VA_OPT__ differently from __VA_ARGS__ in
> >> this regard ?
> >
> > I don't know.  I don't know if the GCC implementation is correct, either.
>
> Clang in its C++20 mode treats __VA_OPT__ as I thought it should be,
> where the determination of whether __VA_OPT__ ( pp-tokens ) expands to a
> placeholder token or pp-tokens is determined by whether the arguments
> for the variadic data parameter ... are empty or not after the arguments
> have been fully macro replaced. In other words clang expands to 1 in
> both my examples above.

#define T(a) #a
T(NO_DATA)

Expands to "NO_DATA", which means that it is:

T(NO_DATA) -> #NO_DATA -> "NO_DATA"

and not:

T(NO_DATA) -> T() -> <error, no argument>

Which explains why __VA_OPT__ behaves this way. According to godbolt
T(NO_DATA) also expands to "NO_DATA", which would mean that difference
between compilers boils down to checking if list is empty - gcc checks
it before macro expansion, and clang after expansion?
-- 
Jędrzej Dudkiewicz

I really hate this damn machine, I wish that they would sell it.
It never does just what I want, but only what I tell it.