Curiously Recurring Template Pattern(CRTP) in C++ is definitely a powerful technique & static alternative to virtual functions. But at the same time, learning it may seem a bit weird at first. If you are like me who struggled to grasp anything in one go. Then this article might help you to provide a thought process on where CRTP fits in day-to-day coding. And, if you are an Embedded Programmer, you may run into CRTP more often. Although, std::variant
+ std::visit
will also help but 90% of the compilers for embedded processors are either not up to date with standard or dumb.
/!\: This article has been originally published on my blog. If you are interested in receiving my latest articles, please sign up to my newsletter.
There is various material effectively accessible for "How" and "What" on CRTP. So, I won't centre there rather address "Where" part that CRTP Applicability.
CRTP and Static Polymorphism In C++
template<typename specific_animal>
struct animal {
void who() { static_cast<specific_animal*>(this)->who(); }
};
struct dog : animal<dog> {
void who() { cout << "dog" << endl; }
};
struct cat : animal<cat> {
void who() { cout << "cat" << endl; }
};
template<typename specific_animal>
void who_am_i(animal<specific_animal> &animal) {
animal.who();
}
cat c;
who_am_i(c); // prints `cat`
dog d;
who_am_i(d); // prints `dog`
- Curiously Recurring Template Pattern widely employed for static polymorphism without bearing the cost of virtual dispatch mechanism. Consider the above code, we haven't used virtual keyword & still achieved the functionality of polymorphism.
- How it works is not the topic of this article. So, I am leaving it to you to figure out.
Limiting Object Count with CRTP
- There are times when you have to manage the critical resource with single or predefined object count. And we have Singleton & Monotone Design Patterns for this. But this works as long as your object counts are smaller in number.
- When you want to limit the arbitrary type to be limited with an arbitrary number of instances. CRTP will come to rescue:
template <class ToBeLimited, uint32_t maxInstance>
struct LimitNoOfInstances {
static atomic<uint32_t> cnt;
LimitNoOfInstances() {
if (cnt >= maxInstance)
throw logic_error{"Too Many Instances"};
++cnt;
}
~LimitNoOfInstances() { --cnt; }
}; // Copy, move & other sanity checks to be complete
struct One : LimitNoOfInstances<One, 1> {};
struct Two : LimitNoOfInstances<Two, 2> {};
template <class T, uint32_t maxNoOfInstace>
atomic<uint32_t> LimitNoOfInstances<T, maxNoOfInstace>::cnt(0);
void use_case() {
Two _2_0, _2_1;
try {
One _1_0, _1_1;
} catch (exception &e) {
cout << e.what() << endl;
}
}
- You might be wondering that what is the point of the template parameter
ToBeLimited
, if it isn't used. In that case, you should have brush up your C++ Template fundamentals or use cppinsights.io. As it isn't useless.
CRTP to Avoid Code Duplication
- Let say you have a set of containers that support the functions
begin()
&end()
. But, the standard library's requirements for containers require more functionalities likefront()
,back()
,size()
, etc. - We can design such functionalities with a CRTP base class that provides common utilities solely based on derived class member function i.e.
begin()
&end()
in our cases:
template <typename T>
class Container {
T &actual() { return *static_cast<T *>(this); }
T const &actual() const { return *static_cast<T const *>(this); }
public:
decltype(auto) front() { return *actual().begin(); }
decltype(auto) back() { return *std::prev(actual().end()); }
decltype(auto) size() const { return std::distance(actual().begin(), actual().end()); }
decltype(auto) operator[](size_t i) { return *std::next(actual().begin(), i); }
};
- The above class provides the functions
front()
,back()
,size()
andoperator[ ]
for any subclass that hasbegin()
&end()
. - For example, subclass could be a simple dynamically allocated array as:
template <typename T>
class DynArray : public Container<DynArray<T>> {
size_t m_size;
unique_ptr<T[]> m_data;
public:
DynArray(size_t s) : m_size{s}, m_data{make_unique<T[]>(s)} {}
T *begin() { return m_data.get(); }
const T *begin() const { return m_data.get(); }
T *end() { return m_data.get() + m_size; }
const T *end() const { return m_data.get() + m_size; }
};
DynArray<int> arr(10);
arr.front() = 2;
arr[2] = 5;
asssert(arr.size() == 10);
Modern C++ Composite Design Pattern Leveraging CRTP
- Composite Design Pattern states that we should treat the group of objects in the same manner as a single object. And to implement such pattern we can leverage the CRTP.
- For example, as a part of machine learning, we have to deal with
Neuron
which for simplicity defined as:
struct Neuron {
vector<Neuron*> in, out; // Stores the input-output connnections to other Neurons
uint32_t id;
Neuron() {
static int id = 1;
this->id = id++;
}
void connect_to(Neuron &other) {
out.push_back(&other);
other.in.push_back(this);
}
friend ostream &operator<<(ostream &os, const Neuron &obj) {
for (Neuron *n : obj.in)
os << n->id << "\t-->\t[" << obj.id << "]" << endl;
for (Neuron *n : obj.out)
os << "[" << obj.id << "]\t-->\t" << n->id << endl;
return os;
}
};
Neuron n1, n2;
n1.connect_to(n2);
cout << n1 << n2 << endl;
/* Output
[1] --> 2
1 --> [2]
*/
- And there is also a
NeuronLayer
i.e. collection ofNeuron
which for simplicity defined as:
struct NeuronLayer : vector<Neuron> {
NeuronLayer(int count) {
while (count --> 0)
emplace_back(Neuron{});
}
friend ostream &operator<<(ostream &os, NeuronLayer &obj) {
for (auto &n : obj)
os << n;
return os;
}
};
- Now, if you want to connect the
Neuron
withNeuronLayer
and vice-versa. You're going to have a total of four different functions as follows:
Neuron::connect_to(Neuron&)
Neuron::connect_to(NeuronLayer&)
NeuronLayer::connect_to(NeuronLayer&)
NeuronLayer::connect_to(Neuron&)
- You see this is state-space explosion(permutation in layman terms) problem and it's not good. Because we want a single function that enumerable both the layer as well as individual neurons. CRTP comes handy here as:
template <typename Self>
struct SomeNeurons {
template <typename T>
void connect_to(T &other);
};
struct Neuron : SomeNeurons<Neuron> {
vector<Neuron*> in, out;
uint32_t id;
Neuron() {
static int id = 1;
this->id = id++;
}
Neuron* begin() { return this; }
Neuron* end() { return this + 1; }
};
struct NeuronLayer : vector<Neuron>, SomeNeurons<NeuronLayer> {
NeuronLayer(int count) {
while (count-- > 0)
emplace_back(Neuron{});
}
};
/* ----------------------------------------------------------------------- */
template <typename Self>
template <typename T>
void SomeNeurons<Self>::connect_to(T &other) {
for (Neuron &from : *static_cast<Self *>(this)) {
for (Neuron &to : other) {
from.out.push_back(&to);
to.in.push_back(&from);
}
}
}
/* ----------------------------------------------------------------------- */
template <typename Self>
ostream &operator<<(ostream &os, SomeNeurons<Self> &object) {
for (Neuron &obj : *static_cast<Self *>(&object)) {
for (Neuron *n : obj.in)
os << n->id << "\t-->\t[" << obj.id << "]" << endl;
for (Neuron *n : obj.out)
os << "[" << obj.id << "]\t-->\t" << n->id << endl;
}
return os;
}
int main() {
Neuron n1, n2;
NeuronLayer l1{1}, l2{2};
n1.connect_to(l1); // Scenario 1: Neuron connects to Layer
l2.connect_to(n2); // Scenario 2: Layer connects to Neuron
l1.connect_to(l2); // Scenario 3: Layer connects to Layer
n1.connect_to(n2); // Scenario 4: Neuron connects to Neuron
cout << "Neuron " << n1.id << endl << n1 << endl;
cout << "Neuron " << n2.id << endl << n2 << endl;
cout << "Layer " << endl << l1 << endl;
cout << "Layer " << endl << l2 << endl;
return EXIT_SUCCESS;
}
/* Output
Neuron 1
[1] --> 3
[1] --> 2
Neuron 2
4 --> [2]
5 --> [2]
1 --> [2]
Layer
1 --> [3]
[3] --> 4
[3] --> 5
Layer
3 --> [4]
[4] --> 2
3 --> [5]
[5] --> 2
*/
- As you can see we have covered all four different permutation scenarios using a single
SomeNeurons::connect_to
method. And bothNeuron
&NeuronLayer
conforms to this interface via self templatization.
C++20 Spaceship Operator With the Help of CRTP
Problem
struct obj_type_1 {
bool operator<(const value &rhs) const { return m_x < rhs.m_x; }
// bool operator==(const value &rhs) const;
// bool operator!=(const value &rhs) const;
// List goes on. . . . . . . . . . . . . . . . . . . .
private:
// data members to compare
};
struct obj_type_2 {
bool operator<(const value &rhs) const { return m_x < rhs.m_x; }
// bool operator==(const value &rhs) const;
// bool operator!=(const value &rhs) const;
// List goes on. . . . . . . . . . . . . . . . . . . .
private:
// data members to compare
};
struct obj_type_3 { ...
struct obj_type_4 { ...
// List goes on. . . . . . . . . . . . . . . . . . . .
- For each comparable objects, you need to define respective comparison operators. This is redundant because if we have an
operator <
, we can overload other operators on the basis of it. - Thus,
operator <
is the only one operator having type information, other operators can be made type independent for reusability purpose.
Solution till C++17 with CRTP
template <class derived>
struct compare {};
struct value : compare<value> {
int m_x;
value(int x) : m_x(x) {}
bool operator < (const value &rhs) const { return m_x < rhs.m_x; }
};
template <class derived>
bool operator > (const compare<derived> &lhs, const compare<derived> &rhs) {
// static_assert(std::is_base_of_v<compare<derived>, derived>); // Compile time safety measures
return (static_cast<const derived&>(rhs) < static_cast<const derived&>(lhs));
}
/* Same goes with other operators
== :: returns !(lhs < rhs) and !(rhs < lhs)
!= :: returns !(lhs == rhs)
>= :: returns (rhs < lhs) or (rhs == lhs)
<= :: returns (lhs < rhs) or (rhs == lhs)
*/
int main() {
value v1{5}, v2{10};
cout << boolalpha << "v1 > v2: " << (v1 > v2) << '\n';
return EXIT_SUCCESS;
}
// Now no need to write comparator operators for all the classes,
// Write only type dependent `operator <` & inherit with `compare<T>`
C++20 Solution : Spaceship Operator
struct value{
int m_x;
value(int x) : m_x(x) {}
auto operator<=>(const value &rhs) const = default;
};
// Defaulted equality comparisons
// More Info: https://en.cppreference.com/w/cpp/language/default_comparisons
Enabling Polymorphic Method Chaining
- Method Chaining is a common syntax for invoking multiple methods on a single object back to back. That too, in a single statement without requiring variables to store the intermediate results. For example:
class Printer {
ostream &m_stream;
public:
Printer(ostream &s) : m_stream(s) { }
Printer &print(auto &&t) {
m_stream << t;
return *this;
}
Printer &println(auto &&t) {
m_stream << t << endl;
return *this;
}
};
Printer{cout}.println("hello").println(500); // Method Chaining
- But, when method chaining applied to an object hierarchy, things can go wrong. For example:
struct ColorPrinter : Printer {
enum Color{red, blue, green};
ColorPrinter(ostream &s) : Printer(s) {}
ColorPrinter &SetConsoleColor(Color c) {
// ...
return *this;
}
};
ColorPrinter(cout).print("Hello ").SetConsoleColor(ColorPrinter::Color::red).println("Printer!"); // Not OK
- Compiling above code prompt you with the following error:
error: 'class Printer' has no member named 'SetConsoleColor'
ColorPrinter(cout).print("Hello ").SetConsoleColor(ColorPrinter::Color::red).println("Printer!");
^
|______________ We have a 'Printer' here, not a 'ColorPrinter'
- This happens because we "lose" the concrete class as soon as we invoke a function of the base class.
- The CRTP can be useful to avoid such problem and to enable Polymorphic Method Chaining.
template <typename ConcretePrinter>
class Printer {
ostream &m_stream;
public:
Printer(ostream &s) : m_stream(s) { }
ConcretePrinter &print(auto &&t) {
m_stream << t;
return static_cast<ConcretePrinter &>(*this);
}
ConcretePrinter &println(auto &&t) {
m_stream << t << endl;
return static_cast<ConcretePrinter &>(*this);
}
};
struct ColorPrinter : Printer<ColorPrinter> {
enum Color { red, blue, green };
ColorPrinter(ostream &s) : Printer(s) {}
ColorPrinter &SetConsoleColor(Color c) {
// ...
return *this;
}
};
int main() {
ColorPrinter(cout).print("Hello ").SetConsoleColor(ColorPrinter::Color::red).println("Printer!");
return EXIT_SUCCESS;
}
Enabling Polymorphic Copy Construction in C++ with CRTP
Problem
- C++ has the support of polymorphic object destruction using it's base class's virtual destructor. But, equivalent support for creation and copying of objects is missing as С++ doesn't support virtual constructor/copy-constructors.
- Moreover, you can't create an object unless you know its static type, because the compiler must know the amount of space it needs to allocate. For the same reason, copy of an object also requires its type to known at compile-time.
struct animal { virtual ~animal(){ cout << "~animal\n"; } };
struct dog : animal { ~dog(){ cout << "~dog\n"; } };
struct cat : animal { ~cat(){ cout << "~cat\n"; } };
void who_am_i(animal *who) { // not sure whether `dog` would be passed here or `cat`
// How to `copy` object of the same type i.e. pointed by who?
delete who; // you can delete object pointed by who
}
Solution 1 : Dynamic Polymorphism
- As the name suggests, we will use virtual methods to delegate the act of copying(and/or creation) of the object as below:
struct animal {
virtual unique_ptr<animal> clone() = 0;
};
struct dog : animal {
unique_ptr<animal> clone() override { return make_unique<dog>(*this); }
};
struct cat : animal {
unique_ptr<animal> clone() override { return make_unique<cat>(*this); }
};
void who_am_i(animal *who) {
auto duplicate_who = who->clone(); // `copy` object of same type i.e. pointed by who ?
}
Solution 2 : Static Polymorphism
- Same thing can be accomplished with CRTP as below:
template <class specific>
struct animal {
unique_ptr<animal> clone() {
return make_unique<specific>(static_cast<specific &>(*this));
}
protected: // Forcing animal class to be inherited
animal(const animal &) = default;
};
struct dog : animal<dog> {
dog(const dog &) { cout << "copied dog" << endl; }
};
struct cat : animal<cat> {
cat(const cat &) { cout << "copied cat" << endl; }
};
template <class specific>
void who_am_i(animal<specific> *who) {
auto duplicate_who = who->clone(); // `copy` object of same type i.e. pointed by who ?
}
Closing Words
Everything comes with its own price. And CRTP is no exception. For example, if you are using CRTP with run time object creation, your code may behave weird. Moreover,
- As the base class is templated, you can not point derived class object with the base class pointer.
- Also, you can not create generic container like
std::vector<animal*>
becauseanimal
is not a class, but a template needing specialization. A container defined asstd::vector<animal<dog>*>
can only storedog
s, notcat
s. This is because each of the classes derived from the CRTP base classanimal
is a unique type. A common solution to this problem is to add one more layer of indirection i.e. abstract class with a virtual destructor, like theabstract_animal
& inheritanimal
class, allowing for the creation of astd::vector<abstract_animal*>
.
There are other useful application of CRTP as well. If you think I am missing any major one & have any suggestion you can always reach me here.