C++ abstract base class constructors/destructors - general correctness

问题

Recently I have dumb as a developer, so I took the plunge, got a C++ book and learning how to do things properly. In my head, I know what I would like to do. I effectively want an Interface that when inherited, must be overridden (if this is possible?). So far, I have the following:

class ICommand{

public:
    //  Virtual constructor. Needs to take a name as parameter
    //virtual ICommand(char*) =0;
    //  Virtual destructor, prevents memory leaks by forcing clean up on derived classes?
    //virtual ~ICommand() =0; 
    virtual void CallMe() =0;
    virtual void CallMe2() =0;
};

class MyCommand : public ICommand
{
public:
    // Is this correct?
    MyCommand(char* Name) { /* do stuff */ }
    virtual void CallMe() {}
    virtual void CallMe2() {}
};

I have purposely left how I think the constructor/destructor's should be implemented in ICommand. I know if I remove the comments, it will not compile. Please could someone:

1. Show me how to declare the constructor/destructor's in ICommand and how they are meant to be used in MyCommand
2. Have I set things up correctly in ICommand so that MyCommand must override CallMe and CallMe2.

I hope I haven't missed something really simple...

回答1:

C++ does not allow for virtual constructors. A simple implementation (without the virtual constructor) would look something like this:

class ICommand {
public:
    virtual ~ICommand() = 0;
    virtual void callMe() = 0;
    virtual void callMe2() = 0;
};

ICommand::~ICommand() { } // all destructors must exist

Note that even a pure virtual destructor must be defined.

A concrete implementation would look exactly like your example:

class MyCommand : public ICommand {
public:
    virtual void callMe() { }
    virtual void callMe2() { }
};

You have a couple of options for the constructor. One option is to disable the default constructor for ICommand, so that subclasses will have to implement a constructor that calls your ICommand constructor:

#include <string>

class ICommand {
private:
    const std::string name;
    ICommand();
public:
    ICommand(const std::string& name) : name(name) { }
    virtual ~ICommand() = 0;
    virtual void callMe() = 0;
    virtual void callMe2() = 0;
};

ICommand::~ICommand() { } // all destructors must exist

A concrete implementation would now look something like this:

class MyCommand : public ICommand {
public:
    MyCommand(const std::string& name) : ICommand(name) { }
    virtual void callMe() { }
    virtual void callMe2() { }
};

回答2:

I know this one is old, but it is still my first hit on this issue. This is how I would do it.

Interface header foo.h:

#pragma once
#include <memory>

enum class Implementations {Simple, Fancy};

class Foo
{
public:
    using Ptr = std::unique_ptr<Foo>;
    virtual ~Foo() = default;
    virtual void do_it() = 0;
};

Foo::Ptr create_foo(Implementations impl); // factory

Yes I know that "pragma once" is strictly speaking not standard, but it works for me.

Note that nothing is implemented here. There is no constructor: an abstract class can not be instantiated. You get a pointer to the interface through the factory. For the virtual function calls to work, they must be called through a pointer. The virtual destructor is defaulted because it doesn't have to do anything special except polymorphing to the implementation. The factory is a free function. No need to try to make it a static member or something like that. This is not java.

Interface foo.cpp:

#include "foo.h"
#include "foo_impl.h"

Foo::Ptr create_foo(Implementations impl)
{
    switch (impl)
    {
    case Implementations::Simple:
        return std::make_unique<Simple_foo>();
    case Implementations::Fancy:
        return std::make_unique<Fancy_foo>();
    default:
        return nullptr;
    }
}

Here the factory is implemented. Notice that the factory has to know the implementation(s). That is why we don't implement it inline: if it was inline, the interface header would have to include the implementation header, and through it, knowledge of the implementation would "leak out" to the callsite.

The implementation header foo_impl.h:

#pragma once
#include "foo.h"

class Simple_foo : public Foo
{
    void do_it() override;
};

class Fancy_foo : public Foo
{
    void do_it() override;
};

Nothing special, just override the virtual functions of the interface. Because this exaple is simple, I put both implementations in the same files. In real applications that will be different.

The implementation foo_impl.cpp:

#include "foo_impl.h"
#include <iostream>

void Simple_foo::do_it()
{
    std::cout << "simple foo\n";
}

void Fancy_foo::do_it()
{
    std::cout << "fancy foo\n";
}

Just implement the functions.

The main.cpp:

#include "foo.h"

int main()
{
    auto sf = create_foo(Implementations::Simple);
    sf->do_it();
    auto ff = create_foo(Implementations::Fancy);
    ff->do_it();
    return 0;
}

Through the enum we can select the implementation we want. The pointers are of type Foo::Ptr, an alias for std::unique_ptr<Foo>. The callsite has no knowledge of the implementation at all, only the interface.

The output will be as expected:

simple foo
fancy foo

来源：https://stackoverflow.com/questions/8513408/c-abstract-base-class-constructors-destructors-general-correctness