51

I developed a scripting engine that has many built-in functions, so to call any function, my code just went into an if .. else if .. else if wall checking the name but I would like to develop a more efficient solution.

Should I use a hashmap with strings as keys and pointers as values? How could I do it by using an STL map?

EDIT: Another point that came into my mind: of course using a map will force the compiler not to inline functions, but my inefficient approach didn't have any overhead generated by the necessity of function calls, it just executes code.

So I wonder if the overhead generated by the function call will be any better than having an if..else chain.. otherwise I could minimize the number of comparisons by checking a character at runtime (will be longer but faster).

RustyTheBoyRobot
  • 5,891
  • 4
  • 36
  • 55
Jack
  • 131,802
  • 30
  • 241
  • 343

6 Answers6

66

Whatever your function signatures are:

typedef void (*ScriptFunction)(void); // function pointer type
typedef std::unordered_map<std::string, ScriptFunction> script_map;

// ...

void some_function()
{
}

// ...

script_map m;
m.emplace("blah", &some_function);

// ...

void call_script(const std::string& pFunction)
{
    auto iter = m.find(pFunction);
    if (iter == m.end())
    {
        // not found
    }

    (*iter->second)();
}

Note that the ScriptFunction type could be generalized to std::function</* whatever*/> so you can support any callable thing, not just exactly function pointers.

GManNickG
  • 494,350
  • 52
  • 494
  • 543
  • 1
    Also there really is no need to use a real hash table like `unordered_map`. There won't be that many elements that a hash table would bring performance advantages, I even wouldn't be surprised if `map` were faster in this case. – sth Jan 26 '10 at 01:41
  • 3
    Actually, I've done some similar stuff and `unordered_map` was *much* faster. I had only about 10,000 things in it, and I profiled both `map` and `unordered_map`. – GManNickG Jan 26 '10 at 01:46
  • 1
    I'd expect `"many builtin functions" << 10.000`. Hasmap in the OP's case has the clear advantage of being "true O(1)" since it doesn't have to grow, and a collision-free hash could be constructed for the strings. I doubt that it makes a *significant* difference compared to a `map` for even a few 100 items. – peterchen Jan 26 '10 at 08:35
  • Actually "many builtin functions" is like ~100. Of course they can grow with time but undoubtely they'll reach 1000. I'll try with the map. Also because I didn't use Boost so far and I would avoid that (just because I really finished everything apart from some optimizations). – Jack Jan 26 '10 at 14:19
  • Does this actually work? Don't you still need to retrieve the function pointer out of the map? i.e. the second last line should be (*(iter->second))(); or something. – Ben Farmer May 07 '14 at 03:30
  • @BenFarmer: Yup, thanks for pointing that out. Such an old answer, too! – GManNickG May 08 '14 at 12:23
  • This is now out-of-date, because there is an `unordered_map` as part of the standard library. – Nic Aug 29 '16 at 15:14
  • @QPaysTaxes: Thanks for pinging! Fixed. – GManNickG Aug 29 '16 at 17:44
25

In C++11 you can do something like this : This Interface needs only the return type and it takes care of everything else from the caller side.

#include <string>
#include <iostream>
#include <map>
#include <vector>
#include <typeinfo>
#include <typeindex>
#include <cassert>

void fun1(void){
    std::cout<<"inside fun1\n";
}

int fun2(){
    std::cout<<"inside fun2\n";
    return 2;
}

int fun3(int a){
    std::cout<<"inside fun3\n";
    return a;
}

std::vector<int> fun4(){
    std::cout<<"inside fun4\n";
    std::vector<int> v(4,100);
    return v;
}

// every function pointer will be stored as this type
typedef void (*voidFunctionType)(void); 

struct Interface{

    std::map<std::string,std::pair<voidFunctionType,std::type_index>> m1;

    template<typename T>
    void insert(std::string s1, T f1){
        auto tt = std::type_index(typeid(f1));
        m1.insert(std::make_pair(s1,
                        std::make_pair((voidFunctionType)f1,tt)));
    }

    template<typename T,typename... Args>
    T searchAndCall(std::string s1, Args&&... args){
        auto mapIter = m1.find(s1);
        /*chk if not end*/
        auto mapVal = mapIter->second;

        // auto typeCastedFun = reinterpret_cast<T(*)(Args ...)>(mapVal.first); 
        auto typeCastedFun = (T(*)(Args ...))(mapVal.first); 

        //compare the types is equal or not
        assert(mapVal.second == std::type_index(typeid(typeCastedFun)));
        return typeCastedFun(std::forward<Args>(args)...);
    }
};

int main(){
    Interface a1;
    a1.insert("fun1",fun1);
    a1.insert("fun2",fun2);
    a1.insert("fun3",fun3);
    a1.insert("fun4",fun4);

    a1.searchAndCall<void>("fun1");
    int retVal = a1.searchAndCall<int>("fun3",2);
    a1.searchAndCall<int>("fun2");
    auto temp = a1.searchAndCall<std::vector<int>>("fun4");

    return 0;
}
Mohit
  • 1,859
  • 1
  • 16
  • 25
  • 3
    This is gold. Is it possible to add member functions to the mix ? Maybe by casting it to a non-member type at some point ? Thanks – Manspider Nov 25 '17 at 06:00
  • How does the call look like if it is a pointer to member function? I would like to do the same thing. If I have this: `typedef int(ObjectT::*Command)();` And calling throws an error. `int result = (*itr->second)();` Invalid use of unary * on pointer to member. – rem45acp Oct 10 '18 at 15:32
  • Will the `std::type_index` provide same result across DLLs and compilers? I know `typeid` does not guarantee consistency. – User 10482 Jun 16 '23 at 20:35
17

You can also use Boost.Function and Boost.Bind what even allows you, to some degree, to have map of heterogeneous functions:

typedef boost::function<void, void> fun_t;
typedef std::map<std::string, fun_t> funs_t;
funs_t f;

void foo() {}
void goo(std::string& p) {}
void bar(int& p) {}

f["foo"] = foo;
f["goo"] = boost::bind(goo, "I am goo");
f["bar"] = boost::bind(bar, int(17));

It can be a map of functions of compatible prototypes as well, of course.

mloskot
  • 37,086
  • 11
  • 109
  • 136
  • This didnt' work for me. i got a compiler error. 'boost::function' : too many template arguments – excray Mar 22 '11 at 11:35
  • @vivek-g, there are many problems possible, compiler version, missing include, etc. It does compile and run for me as well as for codepad: http://codepad.org/ciKTrh2r – mloskot Mar 22 '11 at 15:59
7

Above answers seem to give a complete overview, this regards only your second question:

Map element retrieval by key has O(log n) complexity. Hashmap retrieval by key has O(1) complexity + a little stuff on the side in case of collisions. So if theres a good hash function for your function names, use it. Your implementation will have a standard one. It should be fine.

But be aware, that anything below a hundred elements will not benefit all too much.

The only downside of a hash map is collision. In your case, the hashmap will be relatively static. You know the function names you support. So I advise you to create a simple test case, where you call unordered_map<...>::hash_function with all your keys to make sure that nothing collides. After that, you can forget about it.

A quick google for potential improvements on hash functions got me there:

A fiew good hash functions

Maybe, depending on your naming conventions, you can improve on some aspects of the function.

AndreasT
  • 9,417
  • 11
  • 46
  • 60
6

Well, you can use any_map to store functions with different signatures (but calling it will be messy) and you can use int_map to call functions with a specific signature (looks nicer).

int FuncA()
{
    return 1;
}

float FuncB()
{
    return 2;
}


int main()
{
    // Int map
    map<string,int(*)()> int_map;
    int_map["A"] = FuncA;
    // Call it
    cout<<int_map["A"]()<<endl;

    // Add it to your map
    map<string, void(*)> any_map;
    any_map["A"] = FuncA;
    any_map["B"] = FuncB;

    // Call
    cout<<reinterpret_cast<float(*)()>(any_map["B"])()<<endl;
}
Abhinav Gauniyal
  • 7,034
  • 7
  • 50
  • 93
Jacob
  • 34,255
  • 14
  • 110
  • 165
  • 1
    Actually, I find this very useful. You can basically write your own functions that wrap up the reinterpreting (i.e. float my_b(){ return reinterpret.....}. – Ben J Oct 17 '12 at 15:38
  • 3
    Did you really just write `void main` in a C++ program? – LB-- Oct 01 '13 at 18:48
3

I've managed to modify the example from Mohit to work on member function pointers:

#include <string>
#include <iostream>
#include <map>
#include <vector>
#include <typeinfo>
#include <typeindex>
#include <cassert>


template <typename A>
using voidFunctionType = void (A::*)(void);

template <typename A>
struct Interface{

    std::map<std::string,std::pair<voidFunctionType<A>,std::type_index>> m1;

    template<typename T>
    void insert(std::string s1, T f1){
        auto tt = std::type_index(typeid(f1));
        m1.insert(std::make_pair(s1,
                        std::make_pair((voidFunctionType<A>)f1,tt)));
    }

    template<typename T,typename... Args>
    T searchAndCall(A a, std::string s1, Args&&... args){
        auto mapIter = m1.find(s1);
        auto mapVal = mapIter->second;  

        auto typeCastedFun = (T(A::*)(Args ...))(mapVal.first); 

        assert(mapVal.second == std::type_index(typeid(typeCastedFun)));
        return (a.*typeCastedFun)(std::forward<Args>(args)...);
    }
};

class someclass {
    public:
        void fun1(void);
        int fun2();
        int fun3(int a);
        std::vector<int> fun4();
};

void someclass::fun1(void){
    std::cout<<"inside fun1\n";
}

int someclass::fun2(){
    std::cout<<"inside fun2\n";
    return 2;
}

int someclass::fun3(int a){
    std::cout<<"inside fun3\n";
    return a;
}

std::vector<int> someclass::fun4(){
    std::cout<<"inside fun4\n";
    std::vector<int> v(4,100);
    return v;
}

int main(){
    Interface<someclass> a1;
    a1.insert("fun3",&someclass::fun3);
     someclass s;
    int retVal = a1.searchAndCall<int>(s, "fun3", 3);
    return 0;
}
Ia Rigby
  • 41
  • 3