Deconstructing function pointers in a C++ template, vexing variadics

Last time, we taught our little traits class about noexcept functions. One of many oddball cases in the world of function pointers is that of the variadic function, a classic example of which is printf.

Recall that we had this:

template<typename R, typename... Args>
struct FunctionTraitsBase
{
   using RetType = R;
   using ArgTypes = std::tuple<Args...>;
   static constexpr std::size_t ArgCount = sizeof...(Args);
   template<std::size_t N>
   using NthArg = std::tuple_element_t<N, ArgTypes>;
};

template<typename F> struct FunctionTraits;

template<typename R, typename... Args>
struct FunctionTraits<R(*)(Args...)>
    : FunctionTraitsBase<R, Args...>
{
  using Pointer = R(*)(Args...);
  constexpr static bool IsNoexcept = false;
};

template<typename R, typename... Args>
struct FunctionTraits<R(*)(Args...) noexcept>
    : FunctionTraitsBase<R, Args...>
{
  using Pointer = R(*)(Args...);
  constexpr static bool IsNoexcept = true;
};

But it falls apart when we give it a function like printf because none of our specializations handle that case. Let’s fix that.

template<typename R, typename... Args>
struct FunctionTraitsBase
{
  using RetType = R;
  using ArgTypes = std::tuple<Args...>;
  static constexpr std::size_t ArgCount = sizeof...(Args);
  template<std::size_t N>
  using NthArg = std::tuple_element_t<N, ArgTypes>;
};

template<typename R, typename... Args>
struct FunctionTraits<R(*)(Args...)>
    : FunctionTraitsBase<R, Args...>
{
  using Pointer = R(*)(Args...);
  constexpr static bool IsNoexcept = false;
  static constexpr bool IsVariadic = false;
};

template<typename R, typename... Args>
struct FunctionTraits<R(*)(Args..., ...)> // variadic
    : FunctionTraitsBase<R, Args...>
{
  using Pointer = R(*)(Args..., ...); // variadic
  static constexpr bool IsNoexcept = false;
  static constexpr bool IsVariadic = true;
};

We have to repeat this for the noexcept version, so things are getting pretty repetitive now that we’re up to four flavors of this structure. We saw with noexcept that the varieties cannot be deduced via templating, but we’ll reduce the repetitiveness with a temporary macro.

#define MAKE_TRAITS(Noexcept, Variadic, ...)                \
template<typename R, typename... Args>                      \
struct FunctionTraits<R(*)(__VA_ARGS__) noexcept(Noexcept)> \
    : FunctionTraitsBase<R, Args...>                        \
{                                                           \
  using Pointer = R(*)(__VA_ARGS__) noexcept(Noexcept);     \
  constexpr static bool IsNoexcept = Noexcept;              \
  constexpr static bool IsVariadic = Variadic;              \
}

MAKE_TRAITS(false, false, Args...);
MAKE_TRAITS(false,  true, Args..., ...);
MAKE_TRAITS( true, false, Args...);
MAKE_TRAITS( true,  true, Args..., ...);

#undef MAKE_TRAITS

The MAKE_TRAITS macro takes three-ish parameters.

• Noexcept is true to create the noexcept version, or false to create the regular potentially-throwing version.
• Variadic is true to mark the result as variadic, or false if not.
• The “third” parameter is either Args... to create the non-variadic version or Args..., ... to create the variadic version.

The treatment of Noexcept takes advantage of the optional argument to the noexcept specifier, discussed earlier.

The way we interpret the “third” parameter is a workaround for preprocessor limitations.

Passing a comma in a macro parameter is complicated, because it is normally interpreted as a parameter separator. In order to protect it, you need to enclose the comma in parentheses, as we discussed some time ago.

The hack is to use a variadic macro. The remaining parameters are all captured into the pseudo-parameter __VA_ARGS__, and you can spit them back out in the macro expansion.

If you fail to pass anything for the variadic parameter, then the result is emptiness, but that creates a problem, because we want to include a comma prior to the ... if it is present, but omit it if we are generating the non-variadic version. C++20 adds the __VA_OPT__ pseudo-macro which expands its argument only if the __VA_ARGS__ is nonempty. I’m trying to stick with C++17 for now, so I can’t use that.

Instead, I make the caller pass the prior parameter Args... as well, so that the comma is “baked into” the __VA_ARGS__.

I can simplify the above macro a bit, by inferring whether the result is variadic by comparing the function pointer against an explicitly non-variadic version.

#define MAKE_TRAITS(Noexcept, ...)                          \
template<typename R, typename... Args>                      \
struct FunctionTraits<R(*)(__VA_ARGS__) noexcept(Noexcept)> \
    : FunctionTraitsBase<R, Args...>                        \
{                                                           \
  using Pointer = R(*)(__VA_ARGS__) noexcept(Noexcept);     \
  constexpr static bool IsNoexcept = Noexcept;              \
  constexpr static bool IsVariadic =                        \
    !std::is_same_v<void(*)(__VA_ARGS__), void(*)(Args...)>;\
}

MAKE_TRAITS(false, Args...);
MAKE_TRAITS(false, Args..., ...);
MAKE_TRAITS( true, Args...);
MAKE_TRAITS( true, Args..., ...);

#undef MAKE_TRAITS

I could also avoid the whole __VA_ARGS__ nonsense by requiring that the argument list be parenthesized:

#define MAKE_TRAITS(Noexcept, ArgsList)                \
template<typename R, typename... Args>                 \
struct FunctionTraits<R(*)ArgsList noexcept(Noexcept)> \
    : FunctionTraitsBase<R, Args...>                   \
{                                                      \
  using Pointer = R(*)ArgsList noexcept(Noexcept);     \
  constexpr static bool IsNoexcept = Noexcept;         \
  constexpr static bool IsVariadic =                   \
    !std::is_same_v<void(*)ArgsList, void(*)(Args...)>;\
}

MAKE_TRAITS(false, (Args...));
MAKE_TRAITS(false, (Args..., ...));
MAKE_TRAITS( true, (Args...));
MAKE_TRAITS( true, (Args..., ...));

#undef MAKE_TRAITS

Are we done?

Nope, there’s still more that needs to be done to cover function pointers. We’ll look at another complication next time.

In case you were wondering: I’m not talking about abominable functions. Those things are evil, and I’m going to pretend they simply don’t exist. There’s nothing you can do with them anyway.