pinIntroduction to C++ Templates

Overload Journal #45 - Oct 2001 + Programming Topics   Author: Anthony Williams

C++ templates are a powerful mechanism for code reuse, as they enable the programmer to write code that behaves the same for data of any type. Suppose you write a function printData:

void printData(int value) {
   std::cout << "The value is " << value << std::endl;

If you later decide you also want to print double values, or std::string values, then you have to overload the function:

void printData(double value) {
   std::cout << "The value is " << value << std::endl; 
void printData(std::string value) {
   std::cout << "The value is " << value << std::endl; 

The actual code written for the function is identical in each case; it is just the type of the variable value that changes, yet we have to duplicate the function for each distinct type. This is where templates come in - they enable the user to write the function once for any type of the variable value.

template<typename T> void printData(T value) {
   std::cout << "The value is " << value << std::endl; 

That's all there is to it1 - we can now use printData for any data that can be written to a std::ostream. Here, the template<typename T> part tells the compiler that what follows is a template, and that T is a template parameter that identifies a type. Then, anywhere in the function where T appears, it is replaced with whatever type the function is instantiated for - e.g.:

int i=3; 
double d=4.75; 
std::string s("hello"); 
bool b=false; 
printData(i); // T is int 
printData(d); // T is double 
printData(s); // T is std::string 
printData(b); // T is bool

It is possible to write class templates as well as function templates like printData. A common example is std::vector - you can specify a vector of integers, or of strings, or of some user-defined class, simply by specifying the template parameter:

class MyClass{}; 
std::vectorint vi; // contains ints 
std::vectordouble vd; // contains doubles 
std::vectorstd::string vs; // contains std::strings 
std::vectorMyClass vmc; // contains MyClass objects 

Defining Templates

A Template Definition starts with the keyword template, followed by a list of Template Parameters. What follows is then either a class definition, or a function definition, defining a class template or a function template respectively.

The template parameters introduce names into the scope of the definition, which can be types, values or templates. These names can be used just like any other name of the same kind. Then, when the template is instantiated, real types, values or templates are substituted in place of these names, and the code compiled.

Template Parameters

Templates can have one or more template parameters, which can be:

Two template instantiations refer to the same template if their parameters are all the same, irrespective of any typedefs that may apply. Therefore vec1, vec2 and vec3 in the following example are all the same type.

typedef std::string MyString; typedef std::vectorstd::string T1; typedef std::vectorMyString T2; T1 vec1; T2 vec2; std::vectorstd::string vec3;

Multiple parameters may be specified, separated by commas in the template parameter list:

templatetypename T1,typename T2 class MyClass{}; 

MyClass is thus a class template with two template type parameters, T1 and T2.

Every time a template is referenced with distinct template arguments, then the template is instantiated with the arguments substituted as explained in the following sections. If the resultant code is not valid, then a compilation error will occur.

Template Type Parameters

Template Type Parameters are template parameters that refer to a type; they are the most common form of template parameters. The template parameter for printData in the introduction [1] is an example of a template type parameter. The syntax is simple:

typename name


class name

Both alternatives are identical in meaning, and the choice is merely a matter of style. name is any valid C++ symbol name. Once name has been introduced in the template parameter list, then any reference to name within the body of the template automatically refers to the type of the corresponding template argument for each instantiation, and can be used anywhere a type can normally be used. e.g.

template<typename T> void func(T value) { const T& ref=value; T* p=new T; T temp(23); }

The code generated for func<int> is then identical to:

void func(int value) { const int& ref=value; int* p=new int; int temp(23); }

If, however, a reference was made to func<std::string>, then a compilation error would result, as the statement

std::string temp(23);

is not valid.

Template Non-Type Parameters

Non-Type Template Parameters are template parameters that are values rather than types. They can be any value that is a compile-time constant[2].Their syntax is akin to a variable declaration, e.g.:

template<int i> class A{}; template<double* dp> class B{}; template<void (*func)(int)> void c() {}

Class template A has a non-type template parameter which is an integer, class template B has a non-type template parameter which is a pointer-to-double, and function template c has a non-type template parameter which is a pointer to a function returning void, with a single int parameter. Examples of uses are:

A3 a3; A<sizeof(std::string)> as; double d; // at global scope B<&d> bpd; B<NULL> bn; void myfunc(int); struct MyClass { static void staticFunc(int); }; int main() { c<&myfunc>(); c<&MyClass::staticFunc>(); }

Within the definition of a template, the names of a non-type template parameter refers to a constant of the appropriate type, so given

template<int i> void func() { std::cout<<i<<std::endl; }

func<3>() will print 3, and func<999>() will print 999.

Template Template Parameters

Template Template Parameters enable a template to be parameterized by the name of another template. Say, for example, that you have a class which contains a couple of collections of items, some strings, some integers, and you want the users of your class to choose what type of collection to use (vector, list, stack, etc.). The natural thought is to make the collection type a template parameter. However, a collection of strings is a different type from a collection of integers, so the user would have to specify both individually if Template Type Parameters are used. The solution is a Template Template Parameter:

template<template<typename T> class ContainerType> class MyClass { ContainerType<int> intContainer; ContainerType<std::string> stringContainer; // rest of class };

ContainerType is a Template Template Parameter that refers to a template with a single Template Type Parameter. You can thus say MyClass<vector> or MyClass<list> to have vectors or lists respectively[3].

Within the template definition, the Template Template Parameters can be used just like any other template.

Default Template Parameters

Just as functions can have default values for their arguments, so can templates - indeed, this facility works in pretty much the same way. If a template parameter has a default specified, then all subsequent template parameters must also have a default specified. When referencing a template, parameters with default values can be omitted; if a template parameter is omitted, all subsequent template parameters must also be omitted. e.g.

template<class T1, class T2=int, int i=23> class MyClass{}; // specify all parameters MyClass<double,std::string,46> mc1; // omit "i" MyClass<std::string,double> mc2; // same as above MyClass<std::string,double,23> mc3; // all default MyClassint mc4; // we must specify "T2" if we wish to // specify "i" MyClass<int,int,0> mc5; 

The syntax for declaring a default value for a template parameter is simple just add "= default-value" to the parameter declaration, as shown in the example. If a template parameter is omitted when referencing the template, then the default value is substituted instead.

Dependent Names

A Dependent Name is any name within a template definition that depends on one or more of the template parameters. This includes the template parameters themselves and other templates instantiated with template arguments that are dependent names in the current expansion. Dependent names are important, because they are only resolved when the template is instantiated. As a consequence of this, dependent names that are members of types that are themselves dependent names are always assumed to name objects or functions rather than types, unless preceded with the typename keyword. e.g.

struct X { int x; typedef double Z; }; struct Y { typedef int x; double Z; }; template<typename T> struct ZZ { T::Z z1; // 1 typename T::Z z2; // 2 void func(T& t) { t.x=4; // 3 } typedef typename std::vector<T>::iterator VecIt; // 4 }; int main() { X x; Y y; ZZX zzx; // 5 ZZY zzy; // 6 zzx.func(x); // 7 zzy.func(y); // 8 } 

The line marked 1 is illegal, as T::Z is assumed to refer to an object or function rather than a type. Line 2 is the correct way of doing things. At line 3, t.x is a dependent name, but it refers to an object so this is OK. At line 4, std::vector<T>::iterator is a dependent name that refers to a type, so we need the typename keyword. Lines 5 and 6 instantiate the template for the types X and Y. Line 5 is OK, because X::Z is a type, but line 6 is not, as Y::Z is an object. Lines 7 and 8 demonstrate the opposite - X::x is OK because it is an object, but Y::x is a type, so the instantiation of ZZ<Y>::func will fail.

Using Templates

Class templates and function templates can be used anywhere normal classes and functions can be, as well as Template Template Parameters to other templates. However, the compiler needs to know what to use for the template parameters.

For class templates, the template name must be followed by a template argument list, specifying the parameters. This is a comma-separated list of expressions between angle brackets (<>). For Template Type Parameters, the corresponding expression must name a type, for Template Non-Type Parameters, the expression must evaluate to a compile-time constant of the appropriate type, and for Template Template Parameters, the expression must name a template with the correct signature. e.g.:

template<typename T,unsigned i> struct FixedArray { T data[i]; }; FixedArray<int,3> a; // array of 3 integers FixedArray<int,1+6/3> b; // array of 3 integers template<template<typename T, typename Allocator> class Container> struct ContainerPair { Container<int,std::allocator<int> intContainer; Container<std::string,std:: allocator<std::string> > stringContainer; }; ContainerPair<std::deque> deqCont; // two std::deques ContainerPair<std::vector> vecCont; // two std::vectors

The second example demonstrates two things. Firstly, the template signature of the Template Template Parameters must exactly match the signature of the template passed as the argument, and standard containers have two template parameters. Secondly, if the last argument of a template parameter list is a template reference, as for Container<int,std::allocator<int> > , then the two closing angle brackets (>) must be separated by a space, to avoid being interpreted as the shift right operator (>>).

For function templates, there are two options for specifying the template parameters. Firstly, the function name can be followed by a template argument list as for class templates, e.g.:

template<typename T> void func() {} int main() { func<int>(); func<double>(); }

The second alternative is for function template where one or more of the template parameters are used in the function parameter list. In this case, it is possible to omit those template parameters from the template argument list, as if they had a default value specified (Using Templates). e.g.:

template<typename T> void func(T value) {} template<typename T,typename U> T func2(U value) { return T(value); } int main() { // T=int func(3); // T=double func(3.5); // T=int, U=double func2<int>(3.5); // T=std::vector<std::string>, U=int func2<std::vector<std::string> >(5); // specify both T and U // T=std::vector<std::string>, U=int func2<std::vector<std::string>,int>(5.7); }

This Template Argument Deduction can be used to make life easier for the user of a function template, the same way that function overloading makes life easier for the user of a normal function - the correct function instantiation is automatically called based on the function arguments provided.

In some circumstances, Template Argument Deduction will fail, because there is an inconsistency caused during deduction - if two parameters are declared to be the same type, and different types are passed, then the compiler cannot deduce which to use. This often occurs with std::max:std::max<double>(double,double)

int i=std::max(3,4.5);

The compiler cannot deduce whether to instantiate std::max<int>(int,int) or std::max<double>(double,double), so it generates an error. The solution is to provide an explicit template argument list, or cast one of the arguments so it is the same type as the other.

int i=std::max(static_cast<double>(3),4.5); // T=double int j=std::max<int>(3,4.5); // T=int

Note also that the return type is not considered when deducing the template arguments in this way.

The Standard C++ Library contains a large number of templates, which is partly what makes it so powerful - the Standard Library code can be successfully used with classes that didn't exist when the Standard was written.

Template Requirements and Concepts

Templates implicitly impose requirements on their parameters, particularly Template Type Parameters and Template Template Parameters. These requirements generally take the form of operations that must work on objects of the appropriate type, or class members that must exist and refer to objects or types or functions (with the implicit extra requirement that the type referred to is a class). A Concept is a set of requirements that describe a useful feature of a type. For example, the C++ Standard Library makes reference to things being Assignable or Copy-Constructible; these are Concepts that make the following requirements:

Given a type T, that type is Copy-Constructible if the expression T a(b); is defined, where b is an expression of type T, and the resultant object a has a value equivalent to the value of the expression b.

That type is Assignable if the expression a=b; is defined, where a and b are expressions of type T, and the value of the object referred to by the expression a is equivalent to the value of the object referred to by the expression b after the assignment.

All built-in types are both Copy-Constructible and Assignable. For classes, these requirements translate into requirements on member functions:

A class T is Copy-Constructible if it has a constructor which can be called with one argument of type const-reference-to-T. This may be a constructor with one argument, or it may be a constructor with more than one argument, with default values provided for the remaining arguments. The object thus constructed should have a value equivalent to that of the argument[4].

A class T is Assignable if it defines a copy-assignment operator (operator=()) which has an argument of type T, or const-reference-to-T. The object to which the member function belongs should have a value equivalent to that of the argument after the completion of such a member function[5].

As a consequence, the std::auto_ptr template does not fulfil these requirements, as it exhibits transfer-of-ownership semantics on copy-construction and assignment, and consequently the copy-constructor and copy-assignment arguments are of type reference-to-T rather than const-reference-to-T. It is possible to write Concept-checkers, templates which verify that a given type does indeed fulfil all the syntactic requirements of a particular Concept, even if the current template doesn't require all of them in its current implementation. This also makes tracking down errors easier - the compilation error generated by a failure in such a Concept-Checker is more obviously a failure of the parameter type to fulfil the concept requirements than a failure elsewhere in the template definition.

It is possible that a class template may support different operations, depending on which of several concepts a parameter type fulfils the requirements for. This is made possible by a feature of C++ class templates - member functions of class templates are only instantiated if they are referenced. This means that a class template can have a member function which only compiles if the template parameters fulfil a particular concept, but the program will compile even if they don't fulfil the concept, provided that the function is not referenced for that set of parameters. e.g.:

template<typename T> class MyClass { public: T* makeCopy(T* p) { return p-clone(); } }; MyClass<int> mci; double d; MyClass<double> mcd; double* pd=mcd.makeCopy(&d); 

mci is fine; as no reference is made to the makeCopy member, it doesn't matter that int isn't a class. However, the reference to mcd.makeCopy causes an error, as you cannot call member functions on a double.

Template Specialization and Overloading

Template Specialization allows you to decide that for a specific set of template parameters, the code instantiated for a template should be different to the general case. Say, for example, you wish to use a template with a new class as a Template Type Parameters, and though it has the same semantic behaviour as the classes the template was designed for, the syntax is different, so the original template won't compile with the new class as a parameter. The solution to this problem is specialization.

Consider the following template function definition:

templatetypename T void func(T value) { value.add(3); }

This function assumes that the type substituted for the Template Type Parameter T is a class with a member function add that takes a single parameter of a type that can be constructed from an int. Thus the following classes would be OK:

class T1 { public: void add(int i,double d=0.0); }; class T2 { public: std::string add(double d); };

However, the class T3 below is not OK, because Add is has a capital A.

class T3 { public: void Add(int i); };

This means that we cannot use our original function template func for T3.

Explicit Specialization

What we can do, however, is Explicitly Specialize the template for T3:

template<> void func<T3>(T3 value) { value.Add(3); }

There are two distinctive things about this declaration. The first is the empty template parameter list after the template keyword, and the second is the presence of the template argument list after the name of the function. This tells the compiler that (a) a template is being specialized, and (b) what set of template parameters apply for this specialization. Then, whenever the compiler needs to instantiate the template with this set of template parameters, it uses the specialization instead of the general template.

Explicit Specialization applies both to class templates and function templates. Here is an example for a class template, in which the constructor for the general template accepts its parameter by value, but the constructor for the std::string instantiation accepts its parameter by const-reference, for efficiency:

template<typename T> class X { X(T x); }; template<> class X<std::string> { X(const std::string& s); };

Explicit Specialization can be used with recursive class templates to perform compile-time computations. e.g.:

template<unsigned i> struct Fibonacci { static const unsigned result=Fibonacci<i-1> ::result+Fibonaccii-2::result; }; template<> struct Fibonacci<0> { static const unsigned result=1; }; template struct Fibonacci<1> { static const unsigned result=1; }; int main() { std::cout<<Fibonacci<5>::result<<std::endl; }

The class template Fibonacci<i> defines a constant result to be the i-th number in the Fibonacci Sequence (1,1,2,3,5,8,13,. . . ) where every element is the sum of the two previous ones. This can be seen in the definition of the general template - the value of result is set to the sum of the results of the previous two elements. However, by itself, this is not enough, as each of those is the sum of the previous two, and so on. We solve this by introducing the two explicit specializations for Fibonacci<0> and Fibonacci<1>, to define the first two elements in the series. This way the expansion of the template for any value of i will eventually terminate, so the sample main function will output 8.

In addition to Explicit Specialization, there is Partial Specialization for class templates, and Function Template Overloading for function templates.

Partial Specialization

Partial Specialization allows you to specialize a class template for a subset of the possible template parameters, where the definition of the template should be the same for the whole subset, but different to the general case. e.g.:

template<typename T,typename U> struct SameType { static const bool result=false; }; template<typename T> struct SameType<T,T> { static const bool result=true; }; int main() { std::cout << "Is int the same type as double?" <<(SameType<int,double>::result ? "Yes" : "No") << std::endl; std::cout << "Is std::string the same type " << "as std::string?" <<(SameType<std::string, std::string>::result ? "Yes" : "No") << std::endl; }

The partial specialization of the class template SameType looks a bit like a normal template definition (the Template Parameter List is not empty), and a bit like an explicit specialization (there is a Template Argument List after the template name). The compiler knows it is a partial specialization, because both these things are present. The partial specialization applies to all instantations of the class template where the full template argument list supplied for the general definition can be written down in the form for the partial specialization.

In this example, the partial template specialization applies if both types are the same, so if the types are the same, the value of result is true, whereas if they are different, the general definition applies, and the value of result is false. Therefore the example main function will output:

Is int the same type as double?No Is std::string the same type as std::string?Yes

Partial Specialization can be used for very fine control over what template definition is used. For example, it is possible to specialize on the const-ness of a template type parameter, or for template type parameters which are templates:

template<typename T> struct IsConst { static const bool result=false; }; template<typename T> struct IsConstconst T { static const bool result=true; }; template<typename T> struct IsVector { static const bool result=false; }; template<typename T> struct IsVector<std::vector<T> > { static const bool result=true; };

Thus the result member of IsConst is only true if the template parameter is a const type, and the result member of IsVector is only true if the template parameter is an instantation of the std::vector template.

Function Template Overloading

You cannot partially specialize a function template, but function templates can be overloaded, much the same as functions can be overloaded. This means you can define multiple function templates with the same name, but different template parameters, or a different function signature. You can use this to much the same effect as you can use partial specialization of class templates. Consider the function template swap:

template<typename T> void swap(T& lhs,T& rhs) { T temp(lhs); lhs=rhs; rhs=temp; }

This is generally sufficient to swap two values of any type that is copy-constructible, and assignable. However, it creates a new object (temp), and copies data three times. If this template is instantiated for a class that is expensive to copy, then it could be very inefficient. To ease this, we add a member function swap to our class which swaps the value with that of another object as efficiently as possible:

class ExpensiveToCopy { public: void swap(ExpensiveToCopy& other); };

To call this, we now say a.swap(b) rather than swap(a,b), which is inconsistent. We can solve this by specializing the function template swap for ExpensiveToCopy:

template<> void swap<ExpensiveToCopy> (ExpensiveToCopy& lhs, ExpensiveToCopy& rhs) { lhs.swap(rhs); } 

All is well and good, but what if ExpensiveToCopy was a class template, such as std::vector; we don't want to have to specialize swap for every possible instantiation of the template. The solution to this is Function Template Overloading - we just write a new template function which overloads the first, e.g.:

template<typename T> void swap(std::vector<T>& lhs,std::vector<T>& rhs) { lhs.swap(rhs); }

The compiler chooses which overloaded template to instantiate for each call using a mechanism called Partial Ordering in addition to the normal overload resolution. This is quite complicated, but essentially results in more specific overloads, like that of swap for std::vector being preferred to the more general equivalent, if the actual parameters are compatible with the more specific one.

If Function Template Overloading does not suit your needs, and you really do need Partial Specialization (4.2), the only option is to write a helper class template, with a single static member function that does all the work, and have your function template forward to this function. You can then partially specialize the helper class template to change the body of the static member. e.g.:

template<typename T> class MyClass { public: void write(std::ostream& os) const; }; template<typename T> struct PrintDataHelper { static void doPrint(T value) { std::cout << "The data is " << value << std::endl; } }; template<typename T> struct PrintDataHelperMyClassT { static void doPrint(const MyClassT& value) { std::cout << "The data is "; value.write(std::cout); std::cout<<std::endl; } }; template<typename T> void printData(T value) { PrintDataHelper<T>::doPrint(value); }


Hopefully, this article has given a glimpse into the power of templates, and some insight into how they work and how they can be utilised to make the task of writing and maintaining software easier. They provide immense scope for writing generic and customizable classes and functions, and form an essential C++ language feature, without which much of the Standard C++ Library couldn't exist, or would be greatly restricted. Just about every part of the library involves templates, from std::string (a typedef for std::basic_string<char, std::char_traits<char>, std::allocator<char> >) to the I/O-streams classes (e.g. std::ostream is a typedef for std::basic_ostream<char, std::char_traits<char> > ), to the containers and algorithms that form the Standard Template Library portion, to the classes that implement the Standard facets of the localization/internationalization classes (e.g. std::ctype<char>).

[1] If we were really writing a generic function like this, we would probably pass the parameter as a const reference, rather than as a value parameter to avoid copying large objects - the same applies to void printData(std::string value) above, it would be more usual to write void printData(const std::string& value)

[2] A compile-time constant is something that can be evaluated at compile-time, such as an integer, or the address of a global variable. There are restrictions on what can be a compile-time constant, e.g. floating point values can not be compile-time constants.

[3] If you wish to use std::vector or std::list with a Template Type Parameter, then you have to take account of the additional Allocator template parameter, even though it is very rarely specified explicitly.

[4] If a class does not define any copy-constructors (constructors which can be called with a single argument of type reference-to-T or const-reference-to-T), then the compiler will generate one automatically.

[5] If a class does not define any copy-assignment operators which can be called with a single argument of type T, reference-to-T or const-reference-to-T, then the compiler will generate one automatically.

Overload Journal #45 - Oct 2001 + Programming Topics