C++ Core Guidelines: Rules for Templates and Hierarchies
This is a cross-post from www.ModernesCpp.com.
Due to the C++ core guidelines "Templates are the backbone of C++’s support for generic programming and class hierarchies the backbone of its support for object-oriented programming. The two language mechanisms can be used effectively in combination, but a few design pitfalls must be avoided." Let me see what this means.
This section consists of five rules.
- T.80: Do not naively templatize a class hierarchy
- T.81: Do not mix hierarchies and arrays
- T.82: Linearize a hierarchy when virtual functions are undesirable
- T.83: Do not declare a member function template virtual
- T.84: Use a non-template core implementation to provide an ABI-stable interface
The rule T.81 is too loosely related to templates and hierarchies, and the rule T.82 is empty; therefore, my post boils down to the tree remaining rules.
I will write about the rules T.80 and T.84 together because T.84 continues the story of T.80.
T.80: Do not naively templatize a class hierarchy and T.84: Use a non-template core implementation to provide an ABI-stable interface
Here is as an example a naively templatized class hierarchy from the guidelines:
template<typename T>
struct Container { // an interfacevirtual T* get(int i);
virtual T* first();
virtual T* next();
virtual void sort();
};
template<typename T>
class Vector : public Container<T> {
public:
// ...
};
Vector<int> vi;
Vector<string> vs;
Why is this due to the guidelines naively? This is in particular naively because the base class Container has many virtual functions. The presented designed introduces code bloat. Virtual functions are instantiated every time in a class template. In contrast, non-virtual functions are only instantiated if they are used.
A simple test with CppInsight proves my point.
Non-virtual functions
The following program uses a std::vector<int> and a std::vector<std::string>.
CppInsight shows it. No method of std::vector is instantiated.
Virtual functions
Here is the simplified program of the C++ core guidelines including the virtual function sort.
Now, the virtual function sort is instantiated. Here is only the output of CppInsight which shows the instantiation of the class template Container.
In total, I get 100 lines of code in the case of the virtual function.
The note to the guidelines gives a hint on how to overcome this code bloat. Often you can provide a stable interface by not parameterising a base. This brings me to the related rule T.84: Use a non-template core implementation to provide an ABI-stable interface.
Okay, I already have written about this technique in the post C++ Core Guidelines: Template Definitions. The rule T.84 mentions an alternative way to address a stable interface: Pimpl.
Pimpl
Pimpl stands for "pointer to implementation" and means to remove implementation details of a class by placing them in a separate class, accessed through a pointer. This technique should be in the toolbox of each serious C++ programmer. Pimpl is often also called compilation firewall because this technique breaks the dependency between the implementation and the users of the class interface. This means the implementation can be changed without recompiling the user code.
Here is the general structure from Herb Sutters Blog: GotW #100: Compilation Firewalls.
// in header file
class widget {
public:
widget();
~widget();
private:
class impl; // (1)unique_ptr<impl> pimpl;
};
// in implementation file // (2)
class widget::impl {
// :::
};
widget::widget() : pimpl{ new impl{ /*...*/ } } { }
widget::~widget() { }
This are the points of this idiom.
- put all private non-virtual members into impl (1)
- forward declare impl
- define impl in the corresponding implementation file (2)
Okay. Let me show a full example based on the one from cppreference.com.
// pimpl.cpp
#include <iostream>
#include <memory>
// interface (widget.h)
class widget {
class impl;
std::unique_ptr<impl> pImpl;
public:
void draw();
bool shown() const { return true; }
widget(int);
~widget();
widget(widget&&) = default;
widget(const widget&) = delete;
widget& operator=(widget&&);
widget& operator=(const widget&) = delete;
};
// implementation (widget.cpp)
class widget::impl {
int n; // private datapublic:
void draw(const widget& w) { // (1)if(w.shown())
std::cout << "drawing a widget " << n << '\n';
}
impl(int n) : n(n) {}
};
void widget::draw() { pImpl->draw(*this); } // (2)
widget::widget(int n) : pImpl{std::make_unique<impl>(n)} {}
widget::~widget() = default; // (3)
widget& widget::operator=(widget&&) = default;
// user (main.cpp)
int main()
{
std::cout << std::endl;
widget w(7);
w.draw();
std::cout << std::endl;
}
The draw call of the class impl uses a back-reference to widget (lines (1) and (2)). The destructor and the move assignment operator (line (3)) must be user-defined in the implementation file base because std::unique_ptr requires that the pointed-to type is a complete type.
The program behaves as expected:
Besides its pros, the pimpl idiom has two cons: there is an additional pointer indirection, and you have to store the pointer.
The last guideline for this post is about a typical misconception.
T.83: Do not declare a member function template virtual
Let me try to use a virtual member function template.
// virtualMember.cpp
class Shape {
template<class T>
virtual bool intersect(T* p);
};
int main(){
Shape shape;
}
The error message from my GCC 8.2 compiler is crystal-clear:
What's next?
The next guidelines and, therefore, my next post is about variadic templates. Variadic templates are templates which can accept an arbitrary number of arguments. The rules in the guidelines to variadic templates as many rules to templates consist only of the headings. This means I write in my next post more general about variadic templates.