Although rarely desirable it has always been legal in C++ to pass nothing, aka no preprocessor tokens, as an argument when invoking a macro, whether the equivalent parameter be a regular parameter or a variadic one.

      #define SOME_MACRO(Parameter1,Parameter2) macro expansion using Parameter1 and Parameter2
      #define SOME_VARIADIC_MACRO(Parameter1,...) macro expansion using Parameter1 and __VA_ARGS__
      
      SOME_MACRO(a,b) // Normal
      SOME_MACRO(a,)  // Legal, second argument is empty
      SOME_MACRO(,b)  // Legal, first argument is empty
      SOME_MACRO(a)   // Preprocessor error, passing the wrong number of arguments
      
      SOME_VARIADIC_MACRO(a,b,c,d) // Normal
      SOME_VARIADIC_MACRO(a,)      // Legal, variadic argument is empty
      SOME_VARIADIC_MACRO(,b,c,d)  // Legal, first argument is empty
      SOME_VARIADIC_MACRO(a)       /* Preprocessor error in standard below C++20 level,
                                      but in C++20 exactly equivalent to SOME_VARIADIC_MACRO(a,) */

Given certain arguments a macro might expand to nothing, aka no preprocessor tokens. This may happen more than in the previous case of an argument to a macro being nothing because the expansion of a macro is often used to initialize some C++ construct, and C++ has some places where a part of a compile-time construct can be empty. However a macro which expands to nothing rarely occurs when that macro's expansion is used as an argument to another macro because we would again have a macro where we are passing nothing as an argument.

      #define ANOTHER_MACRO(Parameter1,Parameter2) /* expands to nothing when Parameter1 and Parameter2
                                                      are numbers, otherwise expands to some preprocessing
                                                      token, such as '1' */
      
      int another_int = { ANOTHER_MACRO(x,y) }; // ANOTHER_MACRO Expands to 1
      int some_int = { ANOTHER_MACRO(1,2) };    // ANOTHER_MACRO Expands to nothing
      SOME_MACRO(ANOTHER_MACRO(x,y),z)          // Normal, ANOTHER_MACRO Expands to 1
      SOME_MACRO(ANOTHER_MACRO(1,2),z)          // Legal, first argument is empty as ANOTHER_MACRO Expands to nothing

Passing nothing as a macro argument or a macro expanding to nothing I term as 'emptiness', as 'nothing' is too amorphous a term which can be used in too many other contexts for my liking. In the vast majority of cases when designing a macro for use emptiness is not a part of such a design, and passing emptiness as an argument or expanding to emptiness is not anything that someone writing a macro takes into account when he explains to other programmers how a macro should be used.

Other than the fact that macros are generally created so that some actual preprocessor data of a particular kind needs to be passed as arguments or gets generated as part of macro expansion when a macro is invoked, there is another very good reason why working with emptiness is not part of a macro's design: there has been no perfectly fail-safe way to test for emptiness during macro expansion, whether it be in creating macros using just the facilities of the C++ standard or using a 3rd party library, such as this Boost preprocessor library. When I say 'fail-safe' I mean that there has always been some argument input, no matter how small the number of potential cases, where a macro designed to test whether or not the preprocessor data passed to it as an argument when the macro is invoked is actually empty fails in some way, with the failure normally occurring as a preprocessor error.

Of course this does not mean that the best macro designed to test for emptiness will not work correctly the vast majority of the time. It only means that there has been no guarantee that such a macro will work correctly all 100% of the time. Nonetheless there have been uses of testing for emptiness, when a macro documents what a particular argument should generally consist of, even if the test is not guaranteed to work 100% of the time if particular unexpected argument data does get passed.

The C++ standard committee recognized, in the upcoming specification for the C++20 standard, that a way of testing whether variadic data is empty or not in the expansion of a variadic macro would be very useful when designing certain types of macros. Because of this the C++20 standard added a preprocessor construct which could do this in a certain way for variadic data in the expansion of a variadic macro. The construct is called __VA_OPT__, as in '__VA_OPT__ ( prepocessing tokens )' specified in the replacement list of a variadic macro.

The way that the __VA_OPT__ constructs works is that if the variadic arguments to the variadic macro are empty or expand to emptiness then the __VA_OPT__ construct and its enclosed preprocessing token data expands to nothing, or in C++ terms "a single placemarker preprocessing token". Otherwise the __VA_OPT__ construct expands to its enclosed preprocessing tokens. A further, possibly unintended, upshot of adding the __VA_OPT__ construct to C++20 is that it is now possible to create a variadic macro which is 100% reliable in testing for emptiness whenever a compiler supports the __VA_OPT__ construct in its compilation of preprocessor code.

For such a macro to always work which tests for emptiness the code must know when the __VA_OPT__ construct is available. It is not enough to know that a compiler is working at the C++20 level, since as all C++ programmers know an adherence to a C++ standard level never guarantees that a particular compiler supports every aspect of that level. Happily there is a way to test whether a compiler supports the __VA_OPT__ construct as long as the compiler supports variadic macros, and that way has been openly published on the Internet, although the actual macro code would not have been hard to create even if it had not publicly appeared. This library uses that code to test for __VA_OPT__ as a necessary prelude for creating a variadic macro which is 100% reliable in testing for emptiness.

The Boost Preprocessor macro for testing whether the __VA_OPT__ construct is supported during compilation is called BOOST_PP_VARIADIC_HAS_OPT, which is a function-like macro taking no parameters and returning 1 if the __VA_OPT__ construct is supported and 0 if it is not. The macro only returns 1 when the compiler is at the C++20 level, and when the __VA_OPT__ construct can be used according to the C++20 standard. In particular the macro needs the compiler to be working at the C++20 level despite the fact that at least one major compiler supports the __VA_OPT__ construct in some of its latest releases even when the compiler is being used at a C++ standard level below that of C++20. The reason this Boost preprocessor library requires the C++20 level is because that same major compiler can produce a warning, or even an error, when it even sees a macro using the __VA_OPT__ construct at a level below C++20, even though it supports it, if other compiler options requiring strict adherence to the level of the C++ standard being used are passed on the command line. So taking a conservative approach the BOOST_PP_VARIADIC_HAS_OPT macros requires compilation at the C++20 level, along with the testing code expanding to 1, in order to specify that __VA_OPT__ is supported.

The actual Boost Preprocessor library for testing for emptiness in C++20 mode is called BOOST_PP_CHECK_EMPTY. The macro is a variadic macro with a single variadic parameter. The macro only exists if our previous macro for testing for __VA_OPT__, called BOOST_PP_VARIADIC_HAS_OPT, expands to 1 when invoked as BOOST_PP_VARIADIC_HAS_OPT(). If BOOST_PP_VARIADIC_HAS_OPT() expands to 0 the BOOST_PP_CHECK_EMPTY macro does not exist at all in this library. The input to the BOOST_PP_CHECK_EMPTY macro can be any variadic data. If the data passed to the macro is empty, or if the data passed to the macro is not empty but when the data itself is expanded it is empty, the macro returns 1, otherwise it returns 0. The macro works 100% of the time and is completely reliable no matter what preprocessor data is passed to it. But of course it only works when compiling at the C++20 level with the __VA_OPT__ construct supported by the compiler. It solves an old problem that it has never been possible, prior to C++20, to provide a 100% reliable implementation of a macro which tests for emptiness in C++.

Along with the valuable BOOST_PP_CHECK_EMPTY macro the Boost Preprocessor library has also added a more flexible, if slightly verbose, alternative to the __VA_OPT__ construct, which works by using the ability of BOOST_PP_CHECK_EMPTY to reliably test for emptiness. This macro is called BOOST_PP_VA_OPT and allows the programmer to specify preprocessing tokens for expansion both when the variadic data is not empty and when the variadic data is empty. This improves on the __VA_OPT__ construct's ability to specify preprocessing tokens for expansion only when the variadic data is not empty. Like BOOST_PP_CHECK_EMPTY, which it uses, the BOOST_PP_VA_OPT macro only exists when BOOST_PP_VARIADIC_HAS_OPT() expands to 1. You can read further about how this macro works as an alternative to the C++20 __VA_OPT__ construct in the documentation for the macro itself.

Eventually more C++ compilers will support C++20 and the __VA_OPT__ construct and more programmers will use compilers at the C++20 level. At that point the macro BOOST_PP_CHECK_EMPTY can be used reliably for testing emptiness in preprocessor data in macro code by all those programmers. The BOOST_PP_VA_OPT macro serves as a useful example of such use. This does not mean that designing macros with emptiness in mind needs to be done, much less considered, but that the possibility of doing so with complete reliability will be there if needed by the macro programmer. Along with the __VA_OPT__ construct as mandated by the C++20 standard the BOOST_PP_CHECK_EMPTY and BOOST_PP_VA_OPT macros add three more tools in the arsenal of macro programming, which is a good thing, while programmers who wanted to ignore any dealing with emptiness in macro code can continue to do so.

© Copyright Edward Diener 2019

Distributed under the Boost Software License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at www.boost.org/LICENSE_1_0.txt)