CS100 Lecture 26

Templates II

Contents

• Template specialization
• Variadic templates: an example
• Curiously Recurring Template Pattern (CRTP)
• Introduction to template metaprogramming

Template specialization

Templates are for generic programming, but some things need special treatments.

Specialization for a function template

template <typename T>
int compare(T const &lhs, T const &rhs) {
  if (lhs < rhs) return -1;
  else if (rhs < lhs) return 1;
  else return 0;
}

What happens for C-style strings?

const char *a = "hello", *b = "world";
auto x = compare(a, b);

This is comparing two pointers, instead of comparing the strings!

Specialization for a function template

template <typename T>
int compare(T const &lhs, T const &rhs) {
  if (lhs < rhs) return -1;
  else if (rhs < lhs) return 1;
  else return 0;
}
template <> // specialized version for T = const char *
int compare<const char *>(const char *const &lhs, const char *const &rhs) {
  return std::strcmp(lhs, rhs);
}

Write a specialized version of that function with the template parameters taking a certain group of values.

The type const T & with T = const char * is const char *const &: A reference bound to a const pointer which points to const char.

Specialization for a function template

It is also allowed to omit <const char *> following the name:

template <typename T>
int compare(T const &lhs, T const &rhs) {
  if (lhs < rhs) return -1;
  else if (rhs < lhs) return 1;
  else return 0;
}
template <>
int compare(const char *const &lhs, const char *const &rhs) {
  return std::strcmp(lhs, rhs);
}

Specialization for a function template

Is this a specialization?

template <typename T>
int compare(T const &lhs, T const &rhs);
template <typename T>
int compare(const std::vector<T> &lhs, const std::vector<T> &rhs);

No! These functions constitute overloading (allowed).

Specialization for a function template

Is this a specialization?

template <typename T>
int compare(T const &lhs, T const &rhs);
template <typename T>
int compare<std::vector<T>>(const std::vector<T> &lhs,
                            const std::vector<T> &rhs);

• Since we write int compare<std::vector<T>>(...), this is a specialization.

• However, such specialization is a partial specialization: The specialized function is still a function template.

  • Partial specialization for function templates is not allowed.

Specialization for a class template

It is allowed to write a specialization for class templates.

template <typename T>
struct Dynarray { /* ... */ };
template <> // specialization for T = bool
struct Dynarray<bool> { /* ... */ };

Partial specialization is also allowed:

template <typename T, typename Alloc>
class vector { /* ... */ };
// specialization for T = bool, while Alloc remains a template parameter.
template <typename Alloc>
class vector<bool, Alloc> { /* ... */ };

Variadic templates: an example

A print function

template <typename First, typename... Rest>
void print(std::ostream &os, const First &first, const Rest &...rest) {
  os << first;
  if (/* `rest` is not empty */) // How to test this?
    print(os, rest...);
}
int i = 42; double d = 3.14; std::string s = "hello";
print(std::cout, s, d, i);

First, understand the different meanings of ....

• typename... Rest indicates that Rest is a template parameter pack.
• const Rest &...rest indicates that rest is a function parameter pack.
• rest... in print(os, rest...) is pack expansion.

Compile-time recurrence

template <typename First, typename... Rest>
void print(std::ostream &os, const First &first, const Rest &...rest) {
  os << first;
  if (/* `rest` is not empty */) // How to test this?
    print(os, rest...);
}
std::string s = "hello"; double d = 3.14; int i = 42;
print(std::cout, s, d, i);

print(std::cout, s, d, i) leads to the instantiation of the following functions:

void print(std::ostream &os, const std::string &first, const double &rest0,
           const int &rest1);
void print(std::ostream &os, const double &first, const int &rest0);
void print(std::ostream &os, const int &first);

Note: first is not a parameter pack, so print must have at least two arguments.

sizeof...(pack)

How many arguments are there in a pack? Use the sizeof... operator, which is evaluated at compile-time.

template <typename First, typename... Rest>
void print(std::ostream &os, const First &first, const Rest &...rest) {
  os << first;
  if (sizeof...(Rest) > 0)
    print(os, rest...);
}
std::string s = "hello"; double d = 3.14; int i = 42;
print(std::cout, s, d, i);

Looks good ... But a compile-error?

sizeof...(pack)

Looks good ... But a compile-error?

b.cpp: In instantiation of ‘void print(std::ostream&, const First&, const Rest& ...)
[with First = int; Rest = {}; std::ostream = std::basic_ostream<char>]’:
b.cpp:11:8:   required from here
b.cpp:7:10: error: no matching function for call to ‘print(std::ostream&)’
    7 |     print(os, rest...);

It says that when First = int, Rest = {}, we are trying to call print(os) (with nothing to print).

Compile-time if

Let's see what the function looks like when Rest = {}:

template <typename First>
void print(std::ostream &os, const First &first) {
  os << first;
  if (false)   // sizeof... (Rest) == 0
    print(os); // Ooops! `print` needs at least two arguments!
}

The problem is that if is a run-time control flow statement! The statements must compile even if the condition is 100% false!

We need a compile-time if.

Compile-time if: if constexpr

if constexpr (condition)
  statement1

if constexpr (condition)
  statement1
else
  statement2

condition must be a compile-time constant.

Only when condition is true will statement1 be compiled.

Only when condition is false will statement2 be compiled.

Use if constexpr

template <typename First, typename... Rest>
void print(std::ostream &os, const First &first, const Rest &...rest) {
  os << first;
  if constexpr (sizeof...(Rest) > 0)
    print(os, rest...);
}

Solution without if constexpr: overloading.

template <typename T>
void print(std::ostream &os, const T &x) { os << x; }
template <typename First, typename... Rest>
void print(std::ostream &os, const First &first, const Rest &...rest) {
  print(os, first);
  print(os, rest...);
}

Curiously Recurring Template Pattern (CRTP)

Example 1: Uncopyable

We have seen this Uncopyable in Homework 6:

class Uncopyable {
  Uncopyable(const Uncopyable &) = delete;
  Uncopyable &operator=(const Uncopyable &) = delete;

public:
  Uncopyable() = default;
};

class ComplexDevice : public Uncopyable { /* ... */ };

A class can be made uncopyable by inheriting Uncopyable.

Example 1: Uncopyable

But if two classes inherit from Uncopyable publicly, odd things may happen ...

class Uncopyable {
  Uncopyable(const Uncopyable &) = delete;
  Uncopyable &operator=(const Uncopyable &) = delete;

public:
  Uncopyable() = default;
};

class Airplane : public Uncopyable {}; // Copying an airplane is too costly.
class MonaLisa : public Uncopyable {}; // An artwork is not copyable.

Uncopyable *foo1 = new Airplane();
Uncopyable *foo2 = new MonaLisa();

Ooops ... A Uncopyable* can point to two things that are totally unrelated to each other!

Example 1: Uncopyable

template <typename Derived>
class Uncopyable {
  Uncopyable(const Uncopyable &) = delete;
  Uncopyable &operator=(const Uncopyable &) = delete;

public:
  Uncopyable() = default;
};

class Airplane : public Uncopyable<Airplane> {};
class MonaLisa : public Uncopyable<MonaLisa> {};

Now Airplane and MonaLisa inherit from different bases: Uncopyable<Airplane> and Uncopyable<MonaLisa> are different types.

Example 2: Incrementable

template <typename T>
class Iterator {
  T *cur;

public:
  auto &operator++() {
    ++cur;
    return *this;
  }
  auto operator++(int) {
    auto tmp = *this;
    ++*this;
    return tmp;
  }
};

class Rational {
  int num;
  unsigned denom;

public:
  auto &operator++() {
    num += denom;
    return *this;
  }
  auto operator++(int) {
    auto tmp = *this;
    ++*this;
    return tmp;
  }
};

class AtomicCounter {
  int cnt;
  std::mutex m;

public:
  auto &operator++() {
    std::lock_guard l(m);
    ++cnt;
    return *this;
  }
  auto operator++(int) {
    auto tmp = *this;
    ++*this;
    return tmp;
  }
};

Example 2: Incrementable

With the prefix incrementation operator operator++ defined, the postfix version is always defined as follows:

auto operator++(int) {
  auto tmp = *this;
  ++*this;
  return tmp;
}

How can we avoid repeating ourselves?

Example 2: Incrementable

template <typename Derived>
class Incrementable {
public:
  auto operator++(int) {
    // Since we are sure that the dynamic type of `*this` is `Derived`,
    // we can use `static_cast` here to perform the downcasting.
    auto real_this = static_cast<Derived *>(this);
    auto tmp = *real_this;
    ++*real_this;
    return tmp;
  }
};
class A : public Incrementable<A> {
public:
  A &operator++() { /* ... */ }
  // The operator++(int) is inherited from Incrementable<A>.
};

Curiously Recurring Template Pattern

By writing the common parts of X, Y, Z, ... in a base class Base,

• we can avoid repeating ourselves.
• However, X, Y and Z have a common base (which may lead to weird things), and Base does not know who is inheriting from it.

By letting X, Y, Z, ... inherit from Base<X>, Base<Y>, Base<Z>, ... respectively,

• each class inherits from a unique base class, and
• the base class knows what the derived class is, so a safe downcast can be performed.

CRTP idiom adopted in the standard library: std::enable_shared_from_this.

Introduction to template metaprogramming

Know whether two types are the same?

template <typename T, typename U>
struct is_same {
  static const bool result = false;
};
template <typename T> // specialization for U = T
struct is_same<T, T> {
  static const bool result = true;
};

• is_same<int, double>::result is false.
• is_same<int, int>::result is true.
• Are int and signed the same type? Let is_same tell you!

Know whether a type is a pointer?

template <typename T>
struct is_pointer {
  static const bool result = false;
};
template <typename T>
struct is_pointer<T *> { // specialization for <T *> for some T.
  static const bool result = true;
};

• is_pointer<int *>::result is true.
• is_pointer<int>::result is false.
• Is std::vector<int>::iterator actually a pointer? Is int[10] the same thing as int *? Consult these "functions"!

<type_traits>

std::is_same, std::is_pointer, as well as a whole bunch of other "functions": Go to this standard library.

This is part of the metaprogramming library.

Compute in compile-time?

template <unsigned N>
struct Factorial {
  static const unsigned long long value = N * Factorial<N - 1>::value;
};
template <>
struct Factorial<0u> {
  static const unsigned long long value = 1;
};

int main() {
  int a[Factorial<5>::value]; // 120, which is a compile-time constant.
}

Check whether an integer is a prime in compile-time?

template <unsigned N, unsigned Div> struct PrimeTest {
  static const bool result = (N % Div != 0) && PrimeTest<N, Div + 1>::result;
};
template <unsigned N> struct PrimeTest<N, N> { // end
  static const bool result = true;
};
template <unsigned N> struct IsPrime {
  static const bool result = PrimeTest<N, 2>::result;
};
template <> struct IsPrime<1u> {
  static const bool result = false;
};

static_assert(IsPrime<197>::result); // 197 is a prime
static_assert(!IsPrime<42>::result); // 42 is not

Seven basic quantities in physics

When performing computations in physics, the correctness in dimensions is important.

double mass = getMass();
double acceleration = getAcc();
double force = mass + acceleration; // Ooops! A mistake here!

Can we avoid such mistakes in compile-time? That is, to make mistakes in dimensions a compile error.

Seven basic quantities in physics

Each of the seven basic quantities corresponds to a template parameter:

template <int mass, int length, int time, int charge,
          int temperature, int intensity, int amount_of_substance>
struct quantity { /* ... */ };

using mass = quantity<1, 0, 0, 0, 0, 0, 0>;
using force = quantity<1, 1, -2, 0, 0, 0, 0>;
using pressure = quantity<1, -1, -2, 0, 0, 0, 0>;
using acceleration = quantity<0, 1, -2, 0, 0, 0, 0, 0>;

mass m = getMass();
acceleration a = getAcc();
force f = m + a; // Error! No match operator+ for 'mass' and 'acceleration'!
force f = m * a; // Correct.

If the arithmetic operations of different quantitys are defined correctly, we can avoid dimension mistakes in compile-time!

Template metaprogramming

Template metaprogramming is a very special and powerful technique that makes use of the compile-time computation of C++ compilers. (It is Turing-complete and pure functional programming.)

Learn a little bit more in recitations.

In modern C++, there are many more things that facilitate compile-time computations: constexpr, consteval, constinit, concept, requires, ...

Summary

Template specialization

• Specializes the template function / class when the template argument satisfies certain properties.
• Partial specialization: The specialization still has template parameters.
• Full specialization: The specialization no longer has no template parameters.
• Function templates cannot have partial specializations.

Summary

Variadic template example: A print function.

• pack...: pack expansion.
• sizeof...(pack) returns the number of arguments in a parameter pack. It is compile-time evaluated.
• if constexpr: compile-time if: compile statements conditioned on a compile-time boolean expression.

Summary

Curiously Recurring Template Pattern (CRTP)

• Let X inherit from Base<X>.
• Each class inherits from a unique base class.
• The base class knows what the derived class is, so a safe downcast can be performed.

Template metaprogramming (TMP)

• TMP can shift work from runtime to compile-time, thus enabling earlier error detection and higher runtime performance.