C++ value categories and decltype demystified

David Mazières

June, 2021


Quick C++ quiz:

The bad functions return references to variables that go out of scope when the function returns. As of this writing, gcc and clang warn about different subsets of the bad functions.

Now obviously all of the above functions, whether buggy or not, should have been written to return auto instead of decltype(auto). So is the lesson just to avoid decltype because it adds reference types in unintuitive ways? Unfortunately not, because there are other places in the language where you can’t avoid the “decltype logic,” notably in C++20 requires expressions. Consider the following function:

#include <concepts>

// Attempt to add integer to arbitrary type TA
template<typename TA, typename TB> auto
add(TA a, TB b)
  requires requires {
    // ok:
    { a + b } -> std::same_as<TA>;
    // incorrect (should be std::same_as<int&>):
    { b } -> std::same_as<int>;
  return a += b;

Why is it that add(1,2) is ill-formed (though as of this writing accepted by clang)? add uses a requires expression to try to limit the types of its arguments. The first occurrence of the requires keyword introduces a requires clause to restrict what types TA and TB are acceptable. The second requires introduces a requires expression inside the requires clause. A requires expression can include several forms of requirement, and here we are using compound requirements which state that a particular expression must have a type fulfilling a particular concept. Specifically, we are requiring that decltype((a+b)) be the same type as TA and that decltype((b)) be the same type as int. That’s true in the former case but not the latter, because decltype((b)) is actually the type int&.1

Examples aside, the real issue here is that C++’s rules for inferring reference types are fairly unintuitive and hard to learn. I attribute the problem to three main factors:

  1. Out of reluctance to introduce new keywords, the C++ language committee gave decltype two entirely different purposes that are related enough to be confusing and lead to bad typos.

  2. Every expression in C++ has both a type and a value category, and the two are fiendishly non-orthogonal.

  3. C++ references are analogous to symbolic links in a file system… except, instead of making symbolic links transparently look like files, C++ makes files transparently look like symbolic links. This isn’t a natural way to think of things.

In practice, #1 and #3 are aren’t particularly hard to learn, it’s just a question of rewiring your intuition to match the twisted logic of the C++ language specification.

#2 is a different story. Every expression has one of three value categories: lvalue, prvalue, or xvalue. The result of decltype on expressions is defined to depend on the value category of the expression. Unfortunately, the official definition of value categories is a tangled mess in the language specification. For instance, an informative note in the section introducing value categories optimistically reads:

The discussion of each built-in operator in [expr.compound] indicates the category of the value it yields and the value categories of the operands it expects. For example, the built-in assignment operators expect that the left operand is an lvalue and that the right operand is a prvalue and yield an lvalue as the result.

From the quote, you might expect to be able to slog through 30 pages of [expr.compound] and find an unambiguous specification for the value category of each operand and result of a built-in operator. Or, since life’s not perfect, you might at least hope to find a normative requirement that “the built-in assignment operators expect… the right operand is a prvalue.” Well if so, you’d be disappointed.2

The upshot is that I’ve spent way too much time staring at rules on the Value categories page of cppreference.com, then trying to substantiate them by clicking through from [expr.prop] to other parts of the standard. I found things organized in a way that was difficult to internalize and remember. And I’ve hit bugs around value category in both clang and gcc, so I don’t just “trust the compiler”—when these bugs are fixed, code will break.

Fortunately, almost every expression has what I’ll call an “expression decltype” that unambiguously implies its value category.3 Moreover, we can reformulate the rules for expression decltype in a way that just subsumes value categories.

There are several reasons expression decltypes are easier to think about than value categories. First, in some cases (such as casts), the expression decltype is explicit and maybe even textually part of the expression. Second, the expression decltypes of built-in operators are analogous to the return types of overloaded operators. Thus, programmers accustomed to operator overloading should find them intuitive (except maybe for the conditional ternary operator E1 ? E2 : E3). Finally, while lvalue, prvalue, and xvalue are kind of abstract concepts, you can use the compiler to check the decltype of an expression.

This blog post is my attempt to refactor expression decltype and value categories in a way that is easier to remember and understand. First, I’ll go over the official definition of decltype. Next, I’ll explain and motivate value categories at a high level and explain why they matter. Finally, I’ll propose my new and (I hope) easier to internalize algorithm for determining expression decltype.

decltype overview

Before getting into value categories, let’s review what decltype does. Though there’s only one keyword, the syntax decltype(E) performs one of two entirely different type calculations depending on E:

  1. If E is an unparenthesized id-expression (e.g., x, s.field, S::field), then decltype(E) returns the exact type with which the variable, field, or non-type template parameter was declared, including an lvalue or rvalue reference if and only if the variable or field was declared as one. This is a bit like the lstat(2) system call, which is one of the few ways to differentiate between files and symbolic links in the file system.

    Let’s call this first calculation variable decltype, since it gives us the type with which a variable (or field) was declared.

  2. If E is anything else, including a parenthesized id-expression (e.g., (x), (s.field)), then C++ makes any reference in E’s type completely transparent and undetectable (think stat(2), not lstat). So decltype(E) takes the underlying, non-reference type T of E and decides whether to make it a reference as follows: If E is a prvalue, then decltype(E) is just T; if E is an lvalue, then decltype(E) is T&; and if E is an xvalue, then decltype(E) is T&&.

    Let’s call this second calculation expression decltype, or, to coin a clunky abbreviation, exprtype. Later on, I’ll provide an equivalent formulation that does not depend on value categories, in which case we can run the above rule backwards and say an expression E is a prvalue if decltype((E)) is a non-reference type T, an lvalue if decltype((E)) is T&, and an xvalue if decltype((E)) is T&&.

The two decltype rules are why fn_A above (with return i) is safe, while fn_B (with return (i)) is not. The unparenthesized return expression i in fn_A indicates the use of variable decltype (“lstat”), so the return type is i’s exact type, namely int. In fn_B, by contrast, the return expression (i) indicates expression decltype, so C++ says, “I must hide whether or not i was declared as a reference, and since I can convert the expression to int&, I’ll do so regardless of how i was actually declared.”

There is a rationale for this logic: prvalue expressions such as a literal int (0) or a function call returning int (getpid()) cannot be converted to an lvalue reference such as int&. Prvalues also cannot be used in certain contexts where an lvalue reference would be valid (e.g., getpid() = 5 [wrong], or int *p = &5 [wrong]). By contrast, an lvalue or xvalue can be converted to a prvalue if necessary. In a sense, providing an lvalue reference whenever possible gives you the most powerful applicable type for a given expression. But it’s certainly unintuitive until you’ve learned the rule.

Towards safer use of decltype

If the dual-purpose decltype keyword seems error-prone and you are concerned about bugs from accidentally using the wrong type calculation, a solution may be to program in a stylized way that protects you from simple errors. There’s ample precedent for such an approach. For example, it’s now considered bad practice to write “if (x = y)”; we write “if ((x = y))” to show that, yes, we really intended to do an assignment inside the conditional. When gcc first started warning about assignments in conditionals in the 1990s, I found it offensively paternalistic. In retrospect, the feature has painlessly caught some typos in my code that would have taken a lot more work to debug at runtime.

One easy thing to do would be to define an exprtype macro and always use it when you want expression decltype:

#define exprtype(E) decltype((E))

This at least makes programmer intent explicit: if you aren’t calling exprtype, you want variable decltype, not expression decltype. Unfortunately, the dangerous case is generally that you wanted variable decltype and got expression decltype. So what can you do when you don’t want expression decltype? Here are a couple of ideas depending on what you really want.

In some cases, you may actually want neither variable nor expression decltype. In particular, you may want the non-reference type of a variable regardless of whether the variable itself was declared as a reference. The type inference for auto variables, auto return types, and non-reference function-template arguments works this way. The <type_traits> header has a decay template that emulates the parameter-passing type transformation, essentially removing references and const/volatile qualifications from a type and converting arrays to pointers. To get these rules, you could define:

#define autotype(v) std::decay_t<decltype(v)>

On the other hand, if you really want variable decltype, you could define a macro specifically for the purpose, like this:

#define IGNORE(x)  // causes error if invoked with 2 arguments
#define vdecltype(v) APPLY_IGNORE(PARENTHESIZED_TO_COMMA v) decltype(v)

The vdecltype macro causes a compilation error by calling the single-argument IGNORE macro with two arguments if the argument v is parenthesized.4 Unfortunately, this won’t catch cases such as vdecltype(++v), which would compile and be a reference type, but at least the use of vdecltype makes intent clear and would be amenable to static checking.

I don’t know the best answer, but it stands to reason that some sort of convention around how to invoke decltype for its different purposes could improve the robustness of C++ code.

Value categories

Expression category taxonomy

C++ organizes value categories into the hierarchy depicted above [fig:basic.lval]. I’ll define and motivate the categories at a high level, but I won’t reproduce the categorization rules from the standard because I think it’s better to think of value categories as synonymous with reference qualifiers on expression decltypes:

expression decltype value category
non-reference T prvalue
lvalue reference T& lvalue
rvalue reference T&& xvalue

Nonetheless, understanding value categories at a high level, and understanding the role they play in the language, will help motivate the reference qualifiers on expression decltypes.

Why value categories matter

Because a prvalue must be materialized to initialize its result object, it must be a complete type and not an undefined forward-declared struct. Similarly, a prvalue cannot be a pure virtual class, since such classes cannot be constructed other than as superclasses. Also, only class and array prvalues can meaningfully have const or volatile qualification. There wouldn’t be much point in a const int prvalue, as it would initialize an int lvalue just well as a plain int prvalue could. By contrast, a glvalue need not be a complete type and can have const and volatile qualification for more than just class and array types.

Built-in operators expect particular value categories for their operands and have particular value categories for their results. For example, an arithmetic operator such as binary + expects prvalues for both operands and produces a prvalue result. This isn’t surprising, since you can supply an integer literal as the argument to +, and you also can’t assign to or take the address of the result. Built-in = expects an lvalue on the left and a prvalue on the right, and results in an lvalue.

Of course, even though + has prvalue operands, x + 1 is still a valid expression when x is an lvalue. The reason is that a glvalue other than a function or array can be converted to a prvalue of the same type through a process confusingly named lvalue-to-rvalue conversion. If you recall, a prvalue is really a recipe for initializing an object. When converting from non-class glvalues such as ints, this recipe is to initialize the new int from the value of the old one. When converting class types, the recipe is to copy- or move-initialize a new instance of the class from a particular existing glvalue. (This is presumably why class-type prvalues can have const and volatile qualifiers—they reflect the type of the old glvalue from which the new one should be initialized.)

An expression’s value category determines what references the expression may initialize. Specifically, if T is a non-reference type, then:

Note that when binding a prvalue to a reference, it must be materialized into a temporary object. Generally, a temporary object is destroyed at the end of the full expression, which would leave a dangling reference. To avoid this, C++ extends the lifetime of temporary objects that are bound to references, so that they survive until the reference goes out of scope.

The reference binding rules mostly explain which functions can be selected in overload resolution for arguments of what value category. For instance, you can’t pass prvalue 5 to a function expecting an int&, but you can pass it to a function expecting a const int& or an int&&. However, there’s one rule that doesn’t fit the logic, namely a function template argument seemingly expecting a non-const, non-volatile rvalue reference to a simple typename function template parameter:

template<typename T> decltype(auto)
f(T&& t)
  return g(std::forward<T>(t));

A function parameter such as T&& t is known as a forwarding reference. It matches arguments of any value category, making t an lvalue reference if the supplied argument was an lvalue or an rvalue reference if the supplied argument was an rvalue. If U is t’s underlying non-reference type (namely std::remove_reference_t<decltype(t)>), then T will be inferred as U& for an lvalue argument and U for an rvalue. (Through reference collapsing, if T is U&, then T&& is also U&.) Regardless of t’s variable decltype, its expression decltype is always an lvalue reference; that’s why you always need to provide an explicit template argument to std::forward.

Note that in the example, f actually demonstrates an appropriate use of decltype(auto) return type to preserve the value category of g’s result (including prvalue). Note also that except for initializer lists, auto bindings use the same type deduction rules as function templates. Hence, “auto &&x = f()” is another form of forwarding reference.

Now if we treat value categories as synonymous with reference qualification on expression decltypes, then there’s a much simpler way to describe the value category of built-in operators. We can say + and = behave as if there were built-in functions declared like the following (even though these are obviously not valid code):

    int operator+(int, int);
    int& int::operator=(int);

Given that most C++ programmers already understand operator overloading, wouldn’t it be clearer to express the value category rules for built-in operators using the same vocabulary as user-defined functions? Expression decltype gets us pretty close to this.

Testing with the compiler

Since value category is just synonymous with the reference qualification on expression decltype, we can get the compiler to tell us the value categories of expressions. All we need to do is invoke expression decltype decltype((E)), and use this as a template argument to a template variable that is specialized on reference types. Using this technique, we can figure out which of the functions at the top of this blog post are bad. The function calls that are prvalues are safe, while the lvalues and xvalue are bad, since they are returning references to values that have gone out of scope.

template<typename T> constexpr const char *category = "prvalue";
template<typename T> constexpr const char *category<T&> = "lvalue";
template<typename T> constexpr const char *category<T&&> = "xvalue";

#define SHOW(E) std::cout << #E << ": " << category<decltype((E))> << std::endl



fn_A(0): prvalue
fn_B(0): lvalue
fn_C(0): prvalue
fn_D(0): prvalue
fn_E(0): lvalue
fn_F(0): prvalue
fn_G(0,1): lvalue
fn_H(): prvalue
fn_I(): xvalue

Here’s the code if you want to play with it.

Recall that forwarding references are only for function template arguments, so the specialization category<T&&> is not a forwarding reference.

Simplified rules for expression decltype

Here is my recipe for determining decltype((E)), the “exprtype” of expression E. For the purposes of this blog post, we care less about decltype((E))’s underlying non-reference type T, and more about whether decltype((E)) is T, T&, or T&&. Generally T itself is either obvious or the result of complicated implicit conversion rules that will have to be the subject of a future blog post.

At a high-level, there are three cases to consider: special-cases, named values, and unnamed values. The special cases are string literals and functions, whose exprtype is always the corresponding lvalue reference T&. Named values consist of variables, data members, and array elements. These have exprtype T& unless they would be destroyed by destroying some containing struct or array that has non-lvalue-reference exprtype, in which case they have exprtype T&&. Finally, unnamed values have the same exprtype as their normal C++ type. Unnamed values can be broken down into two subcases: expressions with an obvious explicit type (e.g., casts and function calls), and ones where you have to think about what a built-in operator does.

I hope the intuition from the previous paragraph is already enough to figure out the vast majority of exprtypes in your head, but since this is all a bit unintuitive and there are some hard cases (ternary operator), let me spell it out in a lot more detail with some added rationale:


C++ is notoriously hard to learn in part because it has so many ad-hoc rules that can’t be rederived from first principles. Nonetheless, at least some of the complexity arises from poor organization of the language specification, rather than the language itself. This can be fixed if we explain the language differently, particularly when the language itself is improved, as happened with prvalues in C++17.

In this post, I argued that value categories and expression decltype are two concepts that really should be one. I hope that presenting them in a unified way makes them more intuitive and easier to learn by leveraging the understanding most C++ programmers already of operator overloading and function types.