切换语句可变参数模板扩展

时间:2015-08-26 20:11:45

标签: c++ templates c++11 variadic-templates

请允许我考虑以下合成示例:

inline int fun2(int x) {
    return x;
}
inline int fun2(double x) {
    return 0;   
}
inline int fun2(float x) {
    return -1;   
}

int fun(const std::tuple<int,double,float>& t, std::size_t i) {
    switch(i) {
        case 0: return fun2(std::get<0>(t));
        case 1: return fun2(std::get<1>(t));
        case 2: return fun2(std::get<2>(t));
    }    
}

问题是我应该如何将其扩展到一般情况

template<class... Args> int fun(const std::tuple<Args...>& t, std::size_t i) {
// ?
}

保证

  1. fun2可以内联到有趣的
  2. 搜索复杂度不低于O(log(i))(对于大i)。
  3. 众所周知,当足够大的交换机扩展时,优化器通常使用查找跳转表或编译时二进制搜索树。所以,我想保持这个属性影响大量项目的性能。

    更新#3:我使用统一随机索引值重新衡量了效果:

                          1       10      20      100
    @TartanLlama
        gcc               ~0      42.9235 44.7900 46.5233
        clang             10.2046 38.7656 40.4316 41.7557
    @chris-beck
        gcc               ~0      37.564  51.3653 81.552
        clang             ~0      38.0361 51.6968 83.7704
    naive tail recursion
        gcc                3.0798 40.6061 48.6744 118.171
        clang             11.5907 40.6197 42.8172 137.066
    manual switch statement
        gcc                       41.7236 
        clang                      7.3768 
    

    更新#2:似乎clang能够在@TartanLlama解决方案中内联函数,而gcc总是生成函数调用。

2 个答案:

答案 0 :(得分:7)

好的,我改写了我的回答。这给了TartanLlama以及我之前建议的不同方法。这符合您的复杂性要求,并且不使用函数指针,因此一切都是可内联的。

编辑:非常感谢Yakk在评论中指出了一个非常重要的优化(需要编译时模板递归深度)

基本上我使用模板制作类型/函数处理程序的二叉树,并手动实现二进制搜索。

使用mpl或boost :: fusion可以更干净地完成此操作,但无论如何这个实现都是自包含的。

它绝对符合您的要求,函数是可内联的,运行时查找是元组中类型数量的O(log n)。

以下是完整列表:

#include <cassert>
#include <cstdint>
#include <tuple>
#include <iostream>

using std::size_t;

// Basic typelist object
template<typename... TL>
struct TypeList{
   static const int size = sizeof...(TL);
};

// Metafunction Concat: Concatenate two typelists
template<typename L, typename R>
struct Concat;

template<typename... TL, typename... TR>
struct Concat <TypeList<TL...>, TypeList<TR...>> {
    typedef TypeList<TL..., TR...> type;
};

template<typename L, typename R>
using Concat_t = typename Concat<L,R>::type;

// Metafunction First: Get first type from a typelist
template<typename T>
struct First;

template<typename T, typename... TL>
struct First <TypeList<T, TL...>> {
    typedef T type;
};

template<typename T>
using First_t = typename First<T>::type;


// Metafunction Split: Split a typelist at a particular index
template<int i, typename TL>
struct Split;

template<int k, typename... TL>
struct Split<k, TypeList<TL...>> {
private:
    typedef Split<k/2, TypeList<TL...>> FirstSplit;
    typedef Split<k-k/2, typename FirstSplit::R> SecondSplit;
public:
    typedef Concat_t<typename FirstSplit::L, typename SecondSplit::L> L;
    typedef typename SecondSplit::R R;
};

template<typename T, typename... TL>
struct Split<0, TypeList<T, TL...>> {
    typedef TypeList<> L;
    typedef TypeList<T, TL...> R;
};

template<typename T, typename... TL>
struct Split<1, TypeList<T, TL...>> {
    typedef TypeList<T> L;
    typedef TypeList<TL...> R;
};

template<int k>
struct Split<k, TypeList<>> {
    typedef TypeList<> L;
    typedef TypeList<> R;
};


// Metafunction Subdivide: Split a typelist into two roughly equal typelists
template<typename TL>
struct Subdivide : Split<TL::size / 2, TL> {};

// Metafunction MakeTree: Make a tree from a typelist
template<typename T>
struct MakeTree;

/*
template<>
struct MakeTree<TypeList<>> {
    typedef TypeList<> L;
    typedef TypeList<> R;
    static const int size = 0;
};*/

template<typename T>
struct MakeTree<TypeList<T>> {
    typedef TypeList<> L;
    typedef TypeList<T> R;
    static const int size = R::size;
};

template<typename T1, typename T2, typename... TL>
struct MakeTree<TypeList<T1, T2, TL...>> {
private:
    typedef TypeList<T1, T2, TL...> MyList;
    typedef Subdivide<MyList> MySubdivide;
public:
    typedef MakeTree<typename MySubdivide::L> L;
    typedef MakeTree<typename MySubdivide::R> R;
    static const int size = L::size + R::size;
};

// Typehandler: What our lists will be made of
template<typename T>
struct type_handler_helper {
    typedef int result_type;
    typedef T input_type;
    typedef result_type (*func_ptr_type)(const input_type &);
};

template<typename T, typename type_handler_helper<T>::func_ptr_type me>
struct type_handler {
    typedef type_handler_helper<T> base;
    typedef typename base::func_ptr_type func_ptr_type;
    typedef typename base::result_type result_type;
    typedef typename base::input_type input_type;

    static constexpr func_ptr_type my_func = me;
    static result_type apply(const input_type & t) {
        return me(t);
    }
};

// Binary search implementation
template <typename T, bool b = (T::L::size != 0)>
struct apply_helper;

template <typename T>
struct apply_helper<T, false> {
    template<typename V>
    static int apply(const V & v, size_t index) {
        assert(index == 0);
        return First_t<typename T::R>::apply(v);
    }
};

template <typename T>
struct apply_helper<T, true> {
    template<typename V>
    static int apply(const V & v, size_t index) {
        if( index >= T::L::size ) {
            return apply_helper<typename T::R>::apply(v, index - T::L::size);
        } else {
            return apply_helper<typename T::L>::apply(v, index);
        }
    }
};

// Original functions

inline int fun2(int x) {
    return x;
}
inline int fun2(double x) {
    return 0;   
}
inline int fun2(float x) {
    return -1;   
}

// Adapted functions
typedef std::tuple<int, double, float> tup;

inline int g0(const tup & t) { return fun2(std::get<0>(t)); }
inline int g1(const tup & t) { return fun2(std::get<1>(t)); }
inline int g2(const tup & t) { return fun2(std::get<2>(t)); }

// Registry

typedef TypeList<
   type_handler<tup, &g0>,
   type_handler<tup, &g1>,
   type_handler<tup, &g2>
> registry;

typedef MakeTree<registry> jump_table;

int apply(const tup & t, size_t index) {
    return apply_helper<jump_table>::apply(t, index);
}

// Demo

int main() {
    {
        tup t{5, 1.5, 15.5f};

        std::cout << apply(t, 0) << std::endl;
        std::cout << apply(t, 1) << std::endl;
        std::cout << apply(t, 2) << std::endl;
    }

    {
        tup t{10, 1.5, 15.5f};

        std::cout << apply(t, 0) << std::endl;
        std::cout << apply(t, 1) << std::endl;
        std::cout << apply(t, 2) << std::endl;
    }

    {
        tup t{15, 1.5, 15.5f};

        std::cout << apply(t, 0) << std::endl;
        std::cout << apply(t, 1) << std::endl;
        std::cout << apply(t, 2) << std::endl;
    }

    {
        tup t{20, 1.5, 15.5f};

        std::cout << apply(t, 0) << std::endl;
        std::cout << apply(t, 1) << std::endl;
        std::cout << apply(t, 2) << std::endl;
    }
}

住在Coliru: http://coliru.stacked-crooked.com/a/3cfbd4d9ebd3bb3a

答案 1 :(得分:5)

如果您将fun2设置为超载operator()的类:

struct fun2 {
    inline int operator()(int x) {
        return x;
    }
    inline int operator()(double x) {
        return 0;   
    }
    inline int operator()(float x) {
        return -1;   
    }
};

然后我们可以修改来自here的dyp答案,为我们工作。

请注意,这在C ++ 14中看起来会更加简洁,因为我们可以推导出所有返回类型并使用std::index_sequence

//call the function with the tuple element at the given index
template<class Ret, int N, class T, class Func>
auto apply_one(T&& p, Func func) -> Ret
{
    return func( std::get<N>(std::forward<T>(p)) );
}

//call with runtime index
template<class Ret, class T, class Func, int... Is>
auto apply(T&& p, int index, Func func, seq<Is...>) -> Ret
{
    using FT = Ret(T&&, Func);
    //build up a constexpr array of function pointers to index
    static constexpr FT* arr[] = { &apply_one<Ret, Is, T&&, Func>... };
    //call the function pointer at the specified index
    return arr[index](std::forward<T>(p), func);
}

//tag dispatcher
template<class Ret, class T, class Func>
auto apply(T&& p, int index, Func func) -> Ret
{
    return apply<Ret>(std::forward<T>(p), index, func, 
                      gen_seq<std::tuple_size<typename std::decay<T>::type>::value>{});
}

然后我们调用apply并将返回类型作为模板参数传递(您可以使用decltype或C ++ 14推断它):

auto t = std::make_tuple(1,1.0,1.0f);
std::cout << apply<int>(t, 0, fun2{}) << std::endl;
std::cout << apply<int>(t, 1, fun2{}) << std::endl;
std::cout << apply<int>(t, 2, fun2{}) << std::endl;

Live Demo

由于使用了函数指针,我不确定这是否能完全满足您的要求,但编译器可以非常积极地优化这种事情。搜索将是O(1),因为指针数组只是构建一次然后直接索引,这是非常好的。我会尝试一下,测量一下,看看它是否适合你。