Why can't I use std::function as a std::set or std::unordered_set value type?

Question

Why can't I have a std::set or std::unordered_set of std::functions?

Is there any way to get it to work anyway?

Answer 1

Why can't I have a std::set or std::unordered_set of std::functions?

std::set relies on a comparator, which is used to determine if one element is less than the other.

It uses std::less by default, and std::less doesn't work with std::functions.
(Because there is no way to properly compare std::functions.)

Similarly, std::unordered_set relies on std::hash and std::equal_to (or custom replacements for them), which also don't operate on std::functions.

---

Is there any way to get it to work anyway?

You can write a wrapper around (or a replacement for) std::function that works with std::less, std::equal_to and/or std::hash.

Via power of type erasure, you can forward std::less/std::equal_to/std::hash to objects stored in your wrapper.

Here's a proof-of-concept for such a wrapper.

Notes:

• You can specify whether or not you want the class FancyFunction to work with std::less, std::equal_to and std::hash separetely, by adjusting a template argument.
  If some of those is enabled, you'll be able to apply them to FancyFunction.

  Naturally, you'll be able to construct FancyFunction from a type only if they can be applied to that type.

  There is a static assertion that fires when a type fails to provide std::hash if it's needed.
  It seems to be impossible to SFINAE on availability of std::less and std::equal_to, so I couldn't make similar assertions for those.

• In theory, you could support types that don't work with std::less, std::equal_to and/or std::hash by always considering all instances of one type equivalent, and using typeid(T).hash_code() as a hash.

  I'm not sure if that behavior is desirable or not, implementing it is left as an exercise to the reader.
  (Lack of SFINAE for std::less and std::equal_to will make it harder to implement properly.)

• Specifying custom replacements for std::less, std::equal_to and std::hash is not supported, implementing that is also left as an exercise to the reader.

  (This means that this implementation can only be used to put lambdas into a regular std::set, not std::unordered_set.)

• When applied to FancyFunction, std::less and std::equal_to will first compare types of stored functors.

  If types are identical, they'll resort to calling std::less/std::equal_to on underlying instances.

  (Thus, for two arbitrary different functor types, std::less will always consider instances of one of them less than instances of the other one. The order is not stable between program invocations.)

Example usage:

// With `std::set`:

#include <iostream>
#include <set>

struct AddN
{
    int n;
    int operator()(int x) const {return n + x;}
    friend bool operator<(AddN a, AddN b) {return a.n < b.n;}
};

int main()
{   
    using func_t = FancyFunction<int(int), FunctionFlags::comparable_less>;

    // Note that `std::less` can operate on stateless lambdas by converting them to function pointers first. Otherwise this wouldn't work.
    auto square = [](int x){return x*x;};
    auto cube = [](int x){return x*x*x;};

    std::set<func_t> set;
    set.insert(square);
    set.insert(square); // Dupe.
    set.insert(cube);
    set.insert(AddN{100});
    set.insert(AddN{200});
    set.insert(AddN{200}); // Dupe.

    for (const auto &it : set)
        std::cout << "2 -> " << it(2) << '\n';
    std::cout << '\n';
    /* Prints:
     * 2 -> 4   // `square`, note that it appears only once.
     * 2 -> 8   // `cube`
     * 2 -> 102 // `AddN{100}`
     * 2 -> 202 // `AddN{200}`, also appears once.
     */

    set.erase(set.find(cube));
    set.erase(set.find(AddN{100}));

    for (const auto &it : set)
        std::cout << "2 -> " << it(2) << '\n';
    std::cout << '\n';
    /* Prints:
     * 2 -> 4   // `square`
     * 2 -> 202 // `AddN{200}`
     * `cube` and `AddN{100}` were removed.
     */
}


// With `std::unordered_set`:

#include <iostream>
#include <unordered_set>

struct AddN
{
    int n;
    int operator()(int x) const {return n + x;}
    friend bool operator==(AddN a, AddN b) {return a.n == b.n;}
};

struct MulByN
{
    int n;
    int operator()(int x) const {return n * x;}
    friend bool operator==(MulByN a, MulByN b) {return a.n == b.n;}
};

namespace std
{
    template <> struct hash<AddN>
    {
        using argument_type = AddN;
        using result_type = std::size_t;
        size_t operator()(AddN f) const {return f.n;}
    };

    template <> struct hash<MulByN>
    {
        using argument_type = MulByN;
        using result_type = std::size_t;
        size_t operator()(MulByN f) const {return f.n;}
    };
}

int main()
{   
    using hashable_func_t = FancyFunction<int(int), FunctionFlags::hashable | FunctionFlags::comparable_eq>;
    std::unordered_set<hashable_func_t> set;
    set.insert(AddN{100});
    set.insert(AddN{100}); // Dupe.
    set.insert(AddN{200});
    set.insert(MulByN{10});
    set.insert(MulByN{20});
    set.insert(MulByN{20}); // Dupe.

    for (const auto &it : set)
        std::cout << "2 -> " << it(2) << '\n';
    std::cout << '\n';
    /* Prints:
     * 2 -> 40  // `MulByN{20}`
     * 2 -> 20  // `MulByN{10}`
     * 2 -> 102 // `AddN{100}`
     * 2 -> 202 // `AddN{200}`
     */

    set.erase(set.find(AddN{100}));
    set.erase(set.find(MulByN{20}));

    for (const auto &it : set)
        std::cout << "2 -> " << it(2) << '\n';
    std::cout << '\n';
    /* Prints:
     * 2 -> 20  // `MulByN{10}`
     * 2 -> 202 // `AddN{200}`
     * `MulByN{20}` and `AddN{100}` were removed.
     */
}

Implementation:

#include <cstddef>
#include <functional>
#include <experimental/type_traits>
#include <utility>

enum class FunctionFlags
{
    none            = 0,
    comparable_less = 0b1,
    comparable_eq   = 0b10,
    hashable        = 0b100,
};
constexpr FunctionFlags operator|(FunctionFlags a, FunctionFlags b) {return FunctionFlags(int(a) | int(b));}
constexpr FunctionFlags operator&(FunctionFlags a, FunctionFlags b) {return FunctionFlags(int(a) & int(b));}


template <typename T> using detect_hashable = decltype(std::hash<T>{}(std::declval<const T &>()));


template <typename T, FunctionFlags Flags = FunctionFlags::none>
class FancyFunction;

template <typename ReturnType, typename ...ParamTypes, FunctionFlags Flags>
class FancyFunction<ReturnType(ParamTypes...), Flags>
{
    struct TypeDetails
    {
        int index = 0;
        bool (*less)(const void *, const void *) = 0;
        bool (*eq)(const void *, const void *) = 0;
        std::size_t (*hash)(const void *) = 0;

        inline static int index_counter = 0;
    };

    template <typename T> const TypeDetails *GetDetails()
    {
        static TypeDetails ret = []()
        {
            using type = std::remove_cv_t<std::remove_reference_t<T>>;

            TypeDetails d;

            d.index = TypeDetails::index_counter++;

            if constexpr (comparable_less)
            {
                // We can't SFINAE on `std::less`.
                d.less = [](const void *a_ptr, const void *b_ptr) -> bool
                {
                    const type &a = *static_cast<const FancyFunction *>(a_ptr)->func.template target<type>();
                    const type &b = *static_cast<const FancyFunction *>(b_ptr)->func.template target<type>();
                    return std::less<type>{}(a, b);
                };
            }

            if constexpr (comparable_eq)
            {
                // We can't SFINAE on `std::equal_to`.
                d.eq = [](const void *a_ptr, const void *b_ptr) -> bool
                {
                    const type &a = *static_cast<const FancyFunction *>(a_ptr)->func.template target<type>();
                    const type &b = *static_cast<const FancyFunction *>(b_ptr)->func.template target<type>();
                    return std::equal_to<type>{}(a, b);
                };
            }

            if constexpr (hashable)
            {
                static_assert(std::experimental::is_detected_v<detect_hashable, type>, "This type is not hashable.");
                d.hash = [](const void *a_ptr) -> std::size_t
                {
                    const type &a = *static_cast<const FancyFunction *>(a_ptr)->func.template target<type>();
                    return std::hash<type>(a);
                };
            }

            return d;
        }();
        return &ret;
    }

    std::function<ReturnType(ParamTypes...)> func;
    const TypeDetails *details = 0;

  public:
    inline static constexpr bool
        comparable_less = bool(Flags & FunctionFlags::comparable_less),
        comparable_eq   = bool(Flags & FunctionFlags::comparable_eq),
        hashable        = bool(Flags & FunctionFlags::hashable);

    FancyFunction(decltype(nullptr) = nullptr) {}

    template <typename T>
    FancyFunction(T &&obj)
    {
        func = std::forward<T>(obj);    
        details = GetDetails<T>();
    }

    explicit operator bool() const
    {
        return bool(func);
    }

    ReturnType operator()(ParamTypes ... params) const
    {
        return ReturnType(func(std::forward<ParamTypes>(params)...));
    }

    bool less(const FancyFunction &other) const
    {
        static_assert(comparable_less, "This function is disabled.");
        if (int delta = bool(details) - bool(other.details)) return delta < 0;
        if (!details) return 0;
        if (int delta = details->index - other.details->index) return delta < 0;
        return details->less(this, &other);
    }

    bool equal_to(const FancyFunction &other) const
    {
        static_assert(comparable_eq, "This function is disabled.");
        if (bool(details) != bool(other.details)) return 0;
        if (!details) return 1;
        if (details->index != other.details->index) return 0;
        return details->eq(this, &other);
    }

    std::size_t hash() const
    {
        static_assert(hashable, "This function is disabled.");
        if (!details) return 0;
        return details->hash(this);
    }

    friend bool operator<(const FancyFunction &a, const FancyFunction &b) {return a.less(b);}
    friend bool operator>(const FancyFunction &a, const FancyFunction &b) {return b.less(a);}
    friend bool operator<=(const FancyFunction &a, const FancyFunction &b) {return !b.less(a);}
    friend bool operator>=(const FancyFunction &a, const FancyFunction &b) {return !a.less(b);}
    friend bool operator==(const FancyFunction &a, const FancyFunction &b) {return a.equal_to(b);}
    friend bool operator!=(const FancyFunction &a, const FancyFunction &b) {return !a.equal_to(b);}
};

namespace std
{
    template <typename T, FunctionFlags Flags> struct hash<FancyFunction<T, Flags>>
    {
        using argument_type = FancyFunction<T, Flags>;
        using result_type = std::size_t;
        size_t operator()(const FancyFunction<T, Flags> &f) const
        {
            return f.hash();
        }
    };
}

Answer 2

Well, you can only check function-pointers for (in-)equality, not order. And whether two functions with the same behavior must compare differently is not quite as cut-and-dry as you might perhaps hope for.

Next, you might not only store function-pointers, but also other callables. There is no guarantee any random user-defined class has a strict weak ordering. As an example, lambdas don't.

And finally, how would you order callables of different types?

You can make the same Argument for hashing (needed for unordered containers) as for ordering (needed for ordered containers). Even the equality-comparison needed for unordered containers might not exist.

Answer 3

There is no meaningful equality operation for the general function, held by a std::function.

• C++ functions aren't mathematical functions. What if the "function" holds state? And that state is different for different instances?
• To your suggestion of using the "entry point address": Again, consider state. A std::function can hold a bind of some function/method to some parameters. What is the "entry point address" of that? Ans: Some function/method in the "bind" package. And does then that "entry point address" uniquely identify that function? Ans: No.
• Suppose you have two different functions (by "entry point address") that in fact are identical in the sense that they produce the same result for every parameter? They may even be the same source code - or machine code. Are those functions equal, or not? (If not, why not, if they're behaviorally identical and can't be distinguished by any caller?)

Your particular use case (for sticking std::function in a set) may not be impacted by the above issues. In that case simply wrap the std::function instance in a small struct of your own (either via direct containment or via indirection) (forwarding calls to the contained function object) and put those things in your set.

Answer 4

You can very well create an std::set of functions. The problem is that sets require an absolute order to exist between the values of its elements. This order is defined by a comparator that is then used to sort the elements of a set, to check if an element already exists, and to find a specific element back.

Unfortunately, there doesn't exist an order between functions. Suppose, that you have two functions f1() and f2(), what would be the meaning of f1 < f2 ?

Also equality is not really defined. For example, if you have

int fun1(int) { return 1; }
int fun2(int) { return 1; }
function<int(int)> f1=fun1, f2=fun2; 

Should f1 and f2 be the same value if you'd insert them in a set (because it's always the same result), or is it something different (because it's different functions even though they have the same body) ?

Of course, you could trick the compiler in letting it believe that you have defined an order:

struct Comp {
    using T = function<int(int)>;
    bool operator()(const T &lhs, const T &rhs) const 
    {
        return &lhs < &rhs;
    }
};

set <function<int(int)>,Comp> s; 

You could then insert functions in the set. But this will not work very well, because you take the address of the element, and if the same elements are swapped, the order is different.

I think that the best way to proceed, would be to use a wrapper with a member string that defines an id and use this id to sort the elements in the set (or to do the hashing in case of an unordered_set)