Value Categories in C++



In C++11, with the introducing of move semantics, C++ extended its value categories. Hence, it’s possibly not a good idea to roughly identify expression’s value categories to be lvalue and rvalue. This post is to dump some notes about “Value Category” of expressions in C++, together with some interesting history. Just some basic knowledge about C++ is required.


An expression is a sequence of operators and their operands, that specifies a computation. (– from cppreference) A program is just a combination of expressions, for example

int a = 1;
int b = 2;
a = a + b;
printf("%d", a);

We assign literal 1 to a, then assign literal 2 to b, then assign a+b to a, and then print the result of a. (Note: all steps are expressions. function printf is an expression as well, which returns the length of the string. In this case, it will return 1, the length of “3”, but we did not assign it to anything).


Usually, we characterized an expression in two dimensions, type and value category. In early days, the taxonomy of value categories was first introduced in CPL (Combined Programming Language, the ancestor of BCPL, Basic Combined Programming Language). In CPL, expressions can be evaluated in two modes, known as left-hand (LH) and right-hand (RH) modes. However, only certain kinds of expression are meaningful in LH mode. Take this factorial computing function for example:

function Fact2[x] = result of
  § real f  = 1
    until x = 0 do
      f, x := xf, x -- 1
    result := f §

The CPL looks like some kinds of pseudocode, so here, we will not give too much details about its grammars. In the expression f, x := xf, x – 1 (xf is similar to what we do xf in C/C++), the assignment command := is to tell the computing system to compute the RH value obtained from right-hand of the symbol and then assign the RH value to the address (LH value) that obtained from left-hand of the assignment symbol (:=). Hence, this is the original version taxonomy of value categories and that is what generally in our minds about lvalue and rvalue. For further features about CPL, I recommend “The main features of CPL”.

Then, in C, Dennis Ritchie (or dmr, father of C programming language and co-author of Unix system) characterized expressions into lvalues and others. lvalue is a “locator value”. This is very similar to the taxonomy in CPL.

In early C++, Bjarne Stroustrup did the same as C, but restored the term “rvalue” for “non-lvalue expressions”. Besides, in C++98, functions are made into lvalues, references could be bound to lvalues, and only reference to const value could be bound to rvalue (we do not need to think too much about reference to const that bound to rvalue, we will come to it later). A small example:

int global_val = 0;

int& func() {
  return global_val;

int main() {
  func() = 2;
  return 0;

In this example, function func return the reference of global_val which is a lvalue (function that returns a lvalue reference is a lvalue), and we could assign literal 2 to the reference returned by func.

In C++11, just like what we mentioned before, expressions have two properties: has identity and “can be moved from”. If an expression “has identity”, we could tell if two copies are identical. For example, int a = 2; int b = 2; we know that a and b are identical, because they are two variables and have different addresses in memory, while values of literals like (char) ‘a’, (int) 12 “do not have identity”. The property “can be moved from” denotes whether the resource managed by this object can be transferred to another object. This is related to move semantics in C++11, which we will discuss in next section. Based on these two independent properties, expressions could be categorized into four groups, See Bjarne Stroustrup post “New” Value Terminology

However, the last one is not existed in C++ or most of programming languages. Hence, the relationships between im, iM, Im can be described in one diagram:

(iM)    (im)    (Im)       (lvalue)  (xvalue)  (prvalue)
   \   /    \   /      or      \      /    \     /
    (i)      (m)              (glvalue)    (rvalue)

Note: glvalue means generalized lvalue, xvalue means eXpiring value and prvalue means pure rvalue.

Here, we just give the names of value categories in C++11, for their details, we have to give them after explaining the move semantics in next section.

Move Semantics& RValue Reference

Move Semantics was introduced in C++11, and was designed to lower the cost when transferring the ownership of resources. For example, we have a big object who owns huge external resources. So, if we want to transferring the ownership of resources, in C++, we could simply make the new owner’s resource pointer point to the same memory area, rather than make a copy of the resource and then delete the original one.

#include <iostream>

class BigObject {
  int *_resource;
  size_t _size;
  BigObject(size_t size)
    : _size(size), _resource(new int[size]) {
    printf("create an object with %zu resources\n", _size);

  BigObject(const BigObject& another)
    : _size(another._size), _resource(new int[another._size]) {
    for (int i = 0; i < another._size; ++ i)
      _resource[i] = another._resource[i];
    printf("copy an object with %zu resources\n", _size);

  ~ BigObject() {
    delete [] _resource;
    printf("destory an object with %zu resources\n", _size);


BigObject createBigObject(size_t size) {
  BigObject tmp(size);
  return tmp;

int main() {
  BigObject bigObj(createBigObject(100));
  return 0;

In this example, we have created a BigObject and it has a dynamic int array. Note: most of modern compilers will inspect our return values, and will do some kinds of optimizations called RVO – (Return Value Optimization) and NRVO – (Named Return Value Optimization), this could help prevent constructing useless temporary variables. In this case, BigObject tmp is a temporary variable returned by function createBigObject(), and tmp was created by BigObject(). So, if we compile the code simply with

clang++ -std=c++11 && ./a.out

compiler will just pass the original data to bigObj and then we could get the printed messages:

create an object with 100 resources
destory an object with 100 resources

Adding the flag “-fno-elide-constructors” could tell the compiler not to perform the optimizations. Here, we will not dig too much about RVO or NRVO.

clang++ -std=c++11 -fno-elide-constructors && ./a.out

then we will get

create an object with 100 resources
copy an object with 100 resources
destory an object with 100 resources
copy an object with 100 resources
destory an object with 100 resources
destory an object with 100 resources

As you see, our data was copied two times and it is not difficult to understand. BigObject tmp(100); created the original data, then original data was copied to tmp, and then tmp was copied to bigObj. Because function createBigObject returned rvalue and when the code run out of the scope, it will disappear, and what we could do is to copy it again. How to solve this problem? One simple solution is to return the instance of object by pointer. However, in C++0x, it provides us a new feature called rvalue reference. It means we could pass a rvalue reference out of its scope. Note: BigObject&& means rvalue reference not reference to reference

BigObject(BigObject&& another) // move constructor
  : _size(another._size), _resource(new int[another._size]) {
  _resource = another._resource;
  another._resource = nullptr;
  another._size = 0;
  printf("move an object with %zu resources\n", _size);

With this, our compiler will automatically using this move constructor to pass the ownership without copying the resources again and again. This will print:

create an object with 100 resources
move an object with 100 resources
destory an object with 0 resources
move an object with 100 resources
destory an object with 0 resources
destory an object with 100 resources

This is a simple example of how moving semantics help improve the performance of codes. Actually, the property “can be moved from” is derived from this. C++ STL provides us a useful function called std::move(), which could help us transfer ownership efficiently. Note: the transfering ownership process is not happening when std::move() is applied, it is happening when move constructor is called. std::move() is defined as

template <typename T>
decltype(auto) move (T&& param) {
  using ReturnType = remove_reference_t<T>&&;
  return static_cast<ReturnType> (param);

std::move() is to help casting lvalue to xvalue

string a0 = "hello, world";
string a1 = std::move(a0);

Now, we could discuss the taxonomy.


All these value categories are designed to help us construct high performance and robust codes, help compiler whether, when and where to perform optimizations. Last thing to remember, moving action is not happening when std::move is called but move constructor is called.

Further Reading

comments powered by Disqus