[cfe-dev] [PATCH] New syntax and functionality for __has_attribute

[Aaron Ballman]

On Sat, Jan 11, 2014 at 2:31 PM, Alp Toker <alp at nuanti.com> wrote:
>
> On 10/01/2014 22:07, Aaron Ballman wrote:
>>
>> The __has_attribute feature macro is fantastic in many respects, but
>> is lacking the ability to determine whether a specific attribute
>> syntax is available or not. Instead, it currently checks whether the
>> attribute is known within the compilation target, and nothing more.
>> This can cause problems because not all attributes are applied in the
>> same way.
>>
>> Consider dllexport as a contrived example:
>>
>> #if __has_attribute(dllexport)
>>    void foo(void) __attribute__((dllexport));
>> #endif
>>
>> This code looks fine, but is actually broken because clang only
>> supports __declspec(dllexport) and not __attribute__((dllexport)), and
>> __declspec must precede the declaration.
>>
>> The attached patch implements new syntax for __has_attribute while
>> retaining backwards compatibility. It allows you to specify exactly
>> which attribute syntax you desire. If no specific syntax is specified,
>> it behaves as it always has.
>>
>> The supported forms are:
>>
>> __has_attribute(__attribute__((ident))) // GNU-style
>> __has_attribute(__declspec(ident)) // MS-style
>> __has_attribute([[ident]])  // C++11-style
>> __has_attribute([[scope::ident]]) // C++11-style
>> __has_attribute(ident) // Keywords, or "don't care"
>>
>> Note that attribute arguments are not supported by design -- they
>> really don't make any sense in the context of a feature macro.
>
>
> Hi Aaron,
>
> This is a step forward with some long-standing problems so certainly would
> be a step forward. The syntax is unconventional but not unreasonable.
>
> Have you confirmed that the new __has_attribute() syntax can still be
> defined to an empty expansion? That pattern is important to provide source
> compatibility with gcc / MSVC. The latter in particular has fairly different
> expansion rules to watch out for -- I've got a feeling it'll be OK as long
> as no commas appear in the argument list (which was a problem with the other
> proposed "cxx, ..." syntax) but it's worth double checking.

There's currently a test in the test suite which I think covers this case:

// CHECK: has_has_attribute
#ifdef __has_attribute
int has_has_attribute();
#endif

Is that what you are referring to? If so, then yes, this patch does
still meet that need.

Thanks!

~Aaron