为什么通过weak_ptr打电话这么慢

Why calling via weak_ptr is so slow?

本文关键字：打电话 ptr weak 为什么      更新时间：2023-10-16

我已经阅读了问题weak_ptr的性能损失是多少？ 但我自己的测试显示了不同的结果。

我正在用智能指针制作代表。下面的简单代码显示了 weak_ptr 的性能问题。谁能告诉我为什么？

#include <chrono>
#include <functional>
#include <iostream>
#include <memory>
#include <stdint.h>
#include <string>
#include <utility>
struct Foo
{
    Foo() : counter(0) { incrStep = 1;}
    void bar()
    {
        counter += incrStep;
    }
    virtual ~Foo()
    {
        std::cout << "End " << counter << std::endl;
    }
private:
    uint64_t counter;
    uint64_t incrStep;
};
void pf(const std::string &md, const std::function<void()> &g)
{
    const auto st = std::chrono::high_resolution_clock::now();
    g();
    const auto ft = std::chrono::high_resolution_clock::now();
    const auto del = std::chrono::duration_cast<std::chrono::milliseconds>(ft - st);
    std::cout << md << " t: t" << del.count() << std::endl;
}

和测试：

int main(int , char** )
{
    volatile size_t l = 1000000000ULL;
    size_t maxCounter = l;
    auto a = std::make_shared<Foo>();
    std::weak_ptr<Foo> wp = a;
    pf("call via raw ptr        ", [=](){
        for (size_t i = 0; i < maxCounter; ++i)
        {
            auto p = a.get();
            if (p)
            {
                p->bar();
            }
        }
    });
    pf("call via shared_ptr      ", [=](){
        for (size_t i = 0; i < maxCounter; ++i)
        {
            if (a)
            {
                a->bar();
            }
        }
    });
    pf("call via weak_ptr       ", [=](){
        std::shared_ptr<Foo> p;
        for (size_t i = 0; i < maxCounter; ++i)
        {
            p = wp.lock();
            if (p)
            {
                p->bar();
            }
        }
    });
    pf("call via shared_ptr copy", [=](){
        volatile std::shared_ptr<Foo> p1 = a;
        std::shared_ptr<Foo> p;
        for (size_t i = 0; i < maxCounter; ++i)
        {
            p = const_cast<std::shared_ptr<Foo>& >(p1);
            if (p)
            {
                p->bar();
            }
        }
    });
    pf("call via mem_fn         ", [=](){
        auto fff = std::mem_fn(&Foo::bar);
        for (size_t i = 0; i < maxCounter; ++i)
        {
            fff(a.get());
        }
    });
    return 0;
}

结果：

$ ./test
call via raw ptr            :   369
call via shared_ptr         :   302
call via weak_ptr           :   22663
call via shared_ptr copy    :   2171
call via mem_fn             :   2124
End 5000000000

如您所见，weak_ptr比使用复制和std::mem_fn shared_ptr慢 10 倍，比使用原始 ptr 或shared_ptr.get()慢 60 倍

在尝试重现您的测试时，我意识到优化器可能会消除超出应有的数量。我所做的是利用随机数来击败过度优化，这些结果看起来很现实，std::weak_ptr比std::shared_ptr或其原始指针慢近三倍。

我在每个测试中计算一个校验和，以确保它们都在做相同的工作：

#include <chrono>
#include <memory>
#include <random>
#include <vector>
#include <iomanip>
#include <iostream>
#define OUT(m) do{std::cout << m << 'n';}while(0)
class Timer
{
    using clock = std::chrono::steady_clock;
    using microseconds = std::chrono::microseconds;
    clock::time_point tsb;
    clock::time_point tse;
public:
    void start() { tsb = clock::now(); }
    void stop()  { tse = clock::now(); }
    void clear() { tsb = tse; }
    friend std::ostream& operator<<(std::ostream& o, const Timer& timer)
    {
        return o << timer.secs();
    }
    // return time difference in seconds
    double secs() const
    {
        if(tse <= tsb)
            return 0.0;
        auto d = std::chrono::duration_cast<microseconds>(tse - tsb);
        return double(d.count()) / 1000000.0;
    }
};
constexpr auto N = 100000000U;
int main()
{
    std::mt19937 rnd{std::random_device{}()};
    std::uniform_int_distribution<int> pick{0, 100};
    std::vector<int> random_ints;
    for(auto i = 0U; i < 1024; ++i)
        random_ints.push_back(pick(rnd));
    std::shared_ptr<int> sptr = std::make_shared<int>(std::rand() % 100);
    int* rptr = sptr.get();
    std::weak_ptr<int> wptr = sptr;
    Timer timer;
    unsigned sum = 0;
    sum = 0;
    timer.start();
    for(auto i = 0U; i < N; ++i)
    {
        sum += random_ints[i % random_ints.size()] * *sptr;
    }
    timer.stop();
    OUT("sptr: " << sum << " " << timer);
    sum = 0;
    timer.start();
    for(auto i = 0U; i < N; ++i)
    {
        sum += random_ints[i % random_ints.size()] * *rptr;
    }
    timer.stop();
    OUT("rptr: " << sum << " " << timer);
    sum = 0;
    timer.start();
    for(auto i = 0U; i < N; ++i)
    {
        sum += random_ints[i % random_ints.size()] * *wptr.lock();
    }
    timer.stop();
    OUT("wptr: " << sum << " " << timer);
}

编译器标志：

g++ -std=c++14 -O3 -g0 -D NDEBUG -o bin/timecpp src/timecpp.cpp

示例输出：

sptr: 1367265700 1.26869 // shared pointer
rptr: 1367265700 1.26435 // raw pointer
wptr: 1367265700 2.99008 // weak pointer
相关文章：