Why calling via weak_ptr is so slow?

Question

I have read the question What's the performance penalty of weak_ptr? but my own tests show different results.

I'm making delegates with smart pointers. The simple code below shows reproduces the performance issues with weak_ptr. Can anybody tell me why?

#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;
}

And the test:

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;
}

Results:

$ ./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

As you can see, weak_ptr is 10 times slower than shared_ptr with copying and std::mem_fn and 60 times slower than using raw ptr or shared_ptr.get()

Answer 1

In trying to reproduce your test I realised that the optimizer might be eliminating more than it should. What I did was to utilize random numbers to defeat over-optimization and these results seem realistic with std::weak_ptr being nearly three times slower than the std::shared_ptr or its raw pointer.

I calculate a checksum in each test to ensure they are all doing the same work:

#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);
}

Compiler flags:

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

Example Output:

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