Cost of virtuality and inheritance for non virtual members?

Question

The virtuality can have a double overhead :

memory (because of vptr and vtable)
runtime speed

Because of the memory overhead, I use some CRTP techniques to have a kind of static virtuality, when I need very high memory optimization.

But I wonder about the cost of virtuality for the runtime speed for non-virtual members :

#include <iostream>

class Base
{
    public:
        Base() {;}
        virtual ~Base() {;}

    public:
        virtual void f1() {std::cout<<"f1 : Base"<<std::endl; /* FUNCTION BODY */}
        void f2() {std::cout<<"f2 : Base"<<std::endl; /* FUNCTION BODY */}
        void f3() {f1();}
};

class Derived : public Base
{
    public:
        Derived() {;}
        virtual ~Derived() {;}

    public:
        virtual void f1() {std::cout<<"f1 : Derived"<<std::endl; /* FUNCTION BODY */}
};

And the main :

int main()
{
    Base b;
    Derived d;
    Base* ptr = new Derived();

    std::cout<<std::endl;
    b.f1(); // 1a
    b.f2(); // 1b
    b.f3(); // 1c
    std::cout<<std::endl;
    d.f1(); // 2a
    d.f2(); // 2b
    d.f3(); // 2c
    std::cout<<std::endl;
    ptr->f1(); // 3a
    ptr->f2(); // 3b
    ptr->f3(); // 3c
    std::cout<<std::endl;

    return 0;
}

For each case : 1a, 1b ... 3c, where do I have a runtime overhead (increase of execution time) due to inheritance+virtuality compared to the case where Base and Derived are two completely independant classes with no inheritance ?

And particularly, is there any runtime overhead for the f2 function ?

Note : std::cout is just an example. /* FUNCTION BODY */ can be 1k lines of code...

"For each case...where do I have...more CPU instructions?" Your compiler will produce an assembly listing if you [ask it politely](http://stackoverflow.com/questions/840321/how-can-i-see-the-assembly-code-for-a-c-program). — James McNellis, Oct 06 '12 at 04:18
Just counting the number of CPU instructions will provide a false metric. Loop-unrolling by its very nature increases the number of instructions, yet 'paradoxically' it increases run-time speed. The overhead of using cout rather than printf is orders of magnitude slower than any examination of the vtable and the subsequent re-direction. (cout used as example - I understand it not to be the essence of the question) — enhzflep, Oct 06 '12 at 04:27
@enhzflep : I edited my question taking in account your remarks. — Vincent, Oct 06 '12 at 04:32
See here: http://stackoverflow.com/questions/6749621/high-resolution-timer-in-linux for something perhaps more suitable.. — enhzflep, Oct 06 '12 at 10:33

score 4 · Accepted Answer · answered Oct 06 '12 at 05:11

Why not just time it? It's an entirely trivial exercise..

First, some results

100 million instances of b.f1() = 0.774852 secs.
100 million instances of b.f2() = 0.78162 secs.
100 million instances of b.f3() = 1.85278 secs.

100 million instances of d.f1() = 0.773115 secs.
100 million instances of d.f2() = 0.886528 secs.
100 million instances of d.f3() = 1.88562 secs.

100 million instances of ptr->f1() = 1.02043 secs.
100 million instances of ptr->f2() = 0.778072 secs.
100 million instances of ptr->f3() = 1.72503 secs.

Assuming win32, (QueryPerformanceXXXXX & LARGE_INTEGER) you could use the following:

#include <windows.h>
#include <iostream>
using namespace std;

class Base
{
    public:
        Base() {;}
        virtual ~Base() {;}

    public:
        virtual void f1() {};//std::cout<<"f1 : Base"<<std::endl; /* FUNCTION BODY */}
        void f2() {}; //std::cout<<"f2 : Base"<<std::endl; /* FUNCTION BODY */}
        void f3() {f1();}
};

class Derived : public Base
{
    public:
        Derived() {;}
        virtual ~Derived() {;}

    public:
        virtual void f1() {};//std::cout<<"f1 : Derived"<<std::endl; /* FUNCTION BODY */}
};


LARGE_INTEGER clockFreq;

LARGE_INTEGER getTicks()
{
    LARGE_INTEGER result;
    QueryPerformanceCounter(&result);
    return result;
}

double elapsedSecs(LARGE_INTEGER tStart, LARGE_INTEGER tEnd)
{
    long ticksElapsed = tEnd.QuadPart - tStart.QuadPart;
    double timePeriod = (double)ticksElapsed / (double)clockFreq.QuadPart;
    return timePeriod;
}


int main()
{
    LARGE_INTEGER tStart, tEnd;
    Base b;
    Derived d;
    long i, max=100000000;

    Base* ptr = new Derived();

    // find how fast the clock ticks
    QueryPerformanceFrequency(&clockFreq);


/*====================================================================================================
    Test for access using b
    b.f1()
    b.f2()
    b.f3()
====================================================================================================*/
    std::cout<<std::endl;
    tStart = getTicks();
    for (i=0; i<max; i++)
    {
        b.f1(); // 1a
    }
    tEnd = getTicks();
    double elapsed = elapsedSecs(tStart, tEnd);
    cout << "100 million instances of b.f1() = " << elapsed << " secs." << endl;


    std::cout<<std::endl;
    tStart = getTicks();
    for (i=0; i<max; i++)
    {
        b.f2(); // 1a
    }
    tEnd = getTicks();
    elapsed = elapsedSecs(tStart, tEnd);
    cout << "100 million instances of b.f2() = " << elapsed << " secs." << endl;

    std::cout<<std::endl;
    tStart = getTicks();
    for (i=0; i<max; i++)
    {
        b.f3(); // 1a
    }
    tEnd = getTicks();
    elapsed = elapsedSecs(tStart, tEnd);
    cout << "100 million instances of b.f3() = " << elapsed << " secs." << endl;


/*====================================================================================================
    Test for access using d
    d.f1()
    d.f2()
    d.f3()
====================================================================================================*/
    std::cout<<std::endl;
    tStart = getTicks();
    for (i=0; i<max; i++)
    {
        d.f1(); // 1a
    }
    tEnd = getTicks();
    elapsed = elapsedSecs(tStart, tEnd);
    cout << "100 million instances of d.f1() = " << elapsed << " secs." << endl;


    std::cout<<std::endl;
    tStart = getTicks();
    for (i=0; i<max; i++)
    {
        d.f2(); // 1a
    }
    tEnd = getTicks();
    elapsed = elapsedSecs(tStart, tEnd);
    cout << "100 million instances of d.f2() = " << elapsed << " secs." << endl;

    std::cout<<std::endl;
    tStart = getTicks();
    for (i=0; i<max; i++)
    {
        d.f3(); // 1a
    }
    tEnd = getTicks();
    elapsed = elapsedSecs(tStart, tEnd);
    cout << "100 million instances of d.f3() = " << elapsed << " secs." << endl;

/*====================================================================================================
    Test for access using ptr
    ptr->f1()
    ptr->f2()
    ptr->f3()
====================================================================================================*/
    std::cout<<std::endl;
    tStart = getTicks();
    for (i=0; i<max; i++)
    {
        ptr->f1(); // 1a
    }
    tEnd = getTicks();
    elapsed = elapsedSecs(tStart, tEnd);
    cout << "100 million instances of ptr->f1() = " << elapsed << " secs." << endl;


    std::cout<<std::endl;
    tStart = getTicks();
    for (i=0; i<max; i++)
    {
        ptr->f2(); // 1a
    }
    tEnd = getTicks();
    elapsed = elapsedSecs(tStart, tEnd);
    cout << "100 million instances of ptr->f2() = " << elapsed << " secs." << endl;

    std::cout<<std::endl;
    tStart = getTicks();
    for (i=0; i<max; i++)
    {
        ptr->f3(); // 1a
    }
    tEnd = getTicks();
    elapsed = elapsedSecs(tStart, tEnd);
    cout << "100 million instances of ptr->f3() = " << elapsed << " secs." << endl;

    return 0;
}

Cost of virtuality and inheritance for non virtual members?

1 Answers1