1 Explanation and rationale for the approach

The pace of innovation in the standardization of C++ makes long-term stability of implementations unlikely. Features are added to the language because programmers want to use those features. Features are added to (the working draft of) the standard as the features become well-specified. In many cases a feature is added to an implementation well before or well after the standard officially introducing it is approved.

This process makes it difficult for programmers who want to use a feature to know whether it is available in any given implementation. Implementations rarely leap from one formal revision of the standard directly to the next; the implementation process generally proceeds by smaller steps. As a result, testing for a specific revision of the standard (e.g. by examining the value of the __cplusplus macro) often gives the wrong answer. Implementers generally don’t want to appear to be claiming full conformance to a standard revision until all of its features are implemented. That leaves programmers with no portable way to determine which features are actually available to them.

It is often possible for a program to determine, in a manner specific to a single implementation, what features are supported by that implementation; but the means are often poorly documented and ad hoc, and sometimes complex – especially when the availability of a feature is controlled by an invocation option. To make this determination for a variety of implementations in a single source base is complex and error-prone.

1.1 Status quo before feature-test macros

Here is some code that attempts to determine whether rvalue references are available in the implementation in use:

#ifndef __USE_RVALUE_REFERENCES
  #if (__GNUC__ > 4 || __GNUC__ == 4 && __GNUC_MINOR__ >= 3) || \
      _MSC_VER >= 1600
    #if __EDG_VERSION__ > 0
      #define __USE_RVALUE_REFERENCES (__EDG_VERSION__ >= 410)
    #else
      #define __USE_RVALUE_REFERENCES 1
    #endif
  #elif __clang__
    #define __USE_RVALUE_REFERENCES __has_feature(cxx_rvalue_references)
  #else
    #define __USE_RVALUE_REFERENCES 0
  #endif
#endif

First, the GNU and Microsoft version numbers are checked to see if they are high enough. But then a check is made of the EDG version number, since that front end also has compatibility modes for both those compilers, and defines macros indicating (claimed) compatibility with them. If the feature wasn’t implemented in the indicated EDG version, it is assumed that the feature is not available – even though it is possible for a customer of EDG to implement a feature before EDG does.

Fortunately Clang has ways to test specifically for the presence of specific features. But unfortunately, the function-call-like syntax used for such tests won’t work with a standard preprocessor, so this fine new feature winds up adding its own flavor of complexity to the mix.

Also note that this code is only the beginning of a real-world solution. A complete solution would need to take into account more compilers, and also command-line option settings specific to various compilers.

1.2 Characteristics of the proposed solution

To preserve implementers’ freedom to add features in the order that makes the most sense for themselves and their customers, implementers should indicate the availability of each separate feature by adding a definition of a macro with the name corresponding to that feature.

Important note: By recommending the use of these macros, WG21 is not making any feature optional; the absence of a definition for the relevant feature-test macro does not make an implementation that lacks a feature conform to a standard that requires the feature. However, if implementers and programmers follow these recommendations, portability of code between real-world implementations should be improved.

To a first approximation, a feature is identified by the WG21 paper in which it is specified, and by which it is introduced into the working draft of the standard. Not every paper introduces a new feature worth a feature-test macro, but every paper that is not just a collection of issue resolutions is considered a candidate; exceptions are explicitly justified.

For C++14, the feature-test macro name generally consists of some combination of words from the title of the paper. In the future, it is hoped that every paper will include its own recommendations concerning feature-test macro names.

The value specified for a feature-test macro is based on the year and month in which the feature is voted into the working draft. In a case where a feature is subsequently changed in a significant way, but arguably remains the same feature, the value of the macro is changed to indicate the “revision level” of the specification of the feature. However, in most cases it is expected that the presence of a feature can be determined by the presence of any non-zero macro value; for example:

template<typename T>
struct use_empty_base_opt :
    std::integral_constant<bool,
        std::is_empty<T>::value
#if __cpp_lib_is_final
        && !std::is_final<T>::value
#endif
    >
{ };

To avoid the user’s namespace, names of macros for language features are prefixed by __cpp_; for library features, by __cpp_lib_. A library feature that doesn’t introduce a new header is expected to be defined by the header(s) that implement the feature.

1.3 Examples

Selecting a more efficient compile-time implementation based on the availability of a feature:

#if __cpp_variadic_using
// can use the compile-time efficient, flat inheritance
template<typename ...T> struct Callable : T... {
  using T::operator() ...;
};
#else
// fall-back to linear recursion for older compilers
template<typename ...T> struct Callable;
template<typename T, typename ...U>
struct Callable<T, U...> : T, Callable<U...> {
  using T::operator();
  using Callable<U...>::operator();
};
template<typename T> struct Callable<T> : T {
  using T::operator();
};
template<> struct Callable<> {};
#endif

Likewise

#if __cpp_fold_expressions
template<typename... T>
  auto sum(T... args)  { return (args + ...); }
#else
auto sum() { return 0; }
template<typename T>
  auto sum(T t) { return t; }
template(typename T, typename... Ts)
  auto sum(T t, Ts... ts) { return t + sum(ts...); }
#endif

Selecting a more efficient run-time implementation based on the availability of a feature:

void update(std::set<X>& set, const X& elem, int val)
{
    auto pos = set.find(elem);
    if (pos == set.end())
        return;
#if __cpp_lib_node_extract
    auto next = std::next(pos);
    auto x = set.extract(pos);
    x.value().update(val);
    set.insert(next, std::move(x));
#else
    X tmp = *pos;
    pos = set.erase(pos);
    tmp.update(val);
    set.insert(pos, std::move(tmp));
#endif
}

In some cases, the value of a feature-test macro can change over time as the underlying feature changes. To make it easier to follow the evolution of each feature, the tables in this document are grouped by macro name - with a row for each possible value and the proposal(s) associated with it.

Conditionally implementing a feature, based on __cpp_static_assert.

#if __cpp_static_assert
#  if __cpp_static_assert > 201400
#    define Static_Assert(cond) static_assert(cond)
#  else
#    define Static_Assert(cond) static_assert(cond, #cond)
#  endif
#  define Static_Assert_Msg(cond, msg) static_assert(cond, msg)
#else
#  define Static_Assert(cond)
#  define Static_Assert_Msg(cond, msg)
#endif

Attributes can also change semantics over time, which is why the __has_cpp_attribute facility described below evaluates to a value rather than simply 1 or 0. This allows a user to conditionally provide a version of nodiscard based on __has_cpp_attribute(nodiscard):

#if __has_cpp_attribute(nodiscard) >= 201907
   // nodiscard has a reason and can
   // be applied to constructors
#  define NODISCARD(msg)      [[nodiscard(msg)]]
#  define NODISCARD_CTOR(msg) [[nodiscard(msg)]]
#elif __has_cpp_attribute(nodiscard) >= 201603
   // nodiscard doesn't have a reason, nor can
#  define NODISCARD(msg)      [[nodiscard]]
#  define NODISCARD_CTOR(msg)
#else
   // nodiscard doesn't exist at all yet
#  define NODISCARD(msg)
#  define NODISCARD_CTOR(msg)
#endif

2 Recommendations

2.1 Introduction

For the sake of improved portability between partial implementations of various C++ standards, WG21 (the ISO technical committee for the C++ programming language) recommends that implementers and programmers follow the guidelines in this document concerning feature-test macros.

Implementers who provide a new standard feature should define a macro with the recommended name and value, in the same circumstances under which the feature is available (for example, taking into account relevant command-line options), to indicate the presence of support for that feature.

Programmers who wish to determine whether a feature is available in an implementation should base that determination on the state of the macro with the recommended name. (The absence of a tested feature may result in a program with decreased functionality, or the relevant functionality may be provided in a different way. A program that strictly depends on support for a feature can just try to use the feature unconditionally; presumably, on an implementation lacking necessary support, translation will fail. Therefore, if the most useful purpose for a feature-test macro would be to control the inclusion of a #error directive if the feature is unavailable, that is considered inadequate justification for the macro. Note that the usefulness of a test macro for a feature is completely independent of the usefulness of the feature itself.)

2.2 Testing for the presence of a header: __has_include

It is impossible for a C++ program to directly, reliably, and portably determine whether or not a library header is available for inclusion. Conditionally including a header requires the use of a configuration macro, whose setting can be determined by a configuration-test process at build time (reliable, but less portable), or by some other means (often not reliable or portable).

To solve this general problem, WG21 recommends that programmers use the __has_include feature.

2.2.1 Syntax

h-preprocessing-token:
    any preprocessing-token other than >

h-pp-tokens:
    h-preprocessing-token
    h-pp-tokens h-preprocessing-token

has-include-expression:
    __has_include ( header-name )
    __has_include ( string-literal )
    __has_include ( < h-pp-tokens > )

2.2.2 Semantics

In the first form of the has-include-expression, the parenthesized header-name token is not subject to macro expansion. The second and third forms are considered only if the first form does not match, and the preprocessing tokens are processed just as in normal text.

A has-include-expression shall appear only in the controlling constant expression of a #if or #elif directive ([cpp.cond] 16.1). Prior to the evaluation of such an expression, the source file identified by the parenthesized preprocessing token sequence in each contained has-include-expression is searched for as if that preprocessing token sequence were the pp-tokens in a #include directive, except that no further macro expansion is performed. If such a directive would not satisfy the syntactic requirements of a #include directive, the program is ill-formed. The has-include-expression is replaced by the pp-number 1 if the search for the source file succeeds, and by the pp-number 0 if the search fails.

The #ifdef and #ifndef directives, and the defined conditional inclusion operator, shall treat __has_include as if it were the name of a defined macro. The identifier __has_include shall not appear in any context not mentioned in this section.

2.2.3 Example

This demonstrates a way to use a library optional facility only if it is available. Note that having __has_include(<optional>) succeed is insufficient since on many toolchains, headers may exist in installations but have their contents guarded based on compile flags. For example, the following:

#ifdef __has_include
#if __has_include(<optional>)
#include <optional>
std::optional<int> o;
#endif
#endif
int main(){ }

will still fail to compile with g++ -std=c++14 (using libstdc++).

Hence, we need to do:

#ifdef __has_include
#  if __has_include(<optional>)
#    include <optional>
#    if __cpp_lib_optional >= 201606
#      define have_optional 1
#    endif
#  elif __has_include(<experimental/optional>)
#    include <experimental/optional>
#    if __cpp_lib_experimental_optional >= 201400
#      define have_optional 1
#      define experimental_optional 1
#    endif
#endif

#ifndef have_optional
#    define have_optional 0
#endif

Additionally, the <version> header [P0754R2] is a light-weight header that defines all the standard library feature-test macros. An alternate implementation could be:

#ifndef __has_include
#  define __has_include(x) 0
#endif

#if __has_include(<version>)
#  include <version>
#elif __has_include(<optional>)
#  include <optional>
#endif
#if __cpp_lib_optional >= 201606
#  define have_optional 1
#else
#  define have_optional 0
#endif

2.3 Testing for the presence of an attribute: __has_cpp_attribute

A C++ program cannot directly, reliably, and portably determine whether or not a standard or vendor-specific attribute is available for use. Testing for attribute support generally requires complex macro logic, as illustrated above for language features in general.

To solve this general problem, WG21 recommends that programmers use the __has_cpp_attribute feature.

2.3.1 Syntax

has-attribute-expression:
    __has_cpp_attribute ( attribute-token )

2.3.2 Semantics

A has-attribute-expression shall appear only in the controlling constant expression of a #if or #elif directive ([cpp.cond] 16.1). The has-attribute-expression is replaced by a non-zero pp-number if the implementation supports an attribute with the specified name, and by the pp-number 0 otherwise.

For a standard attribute, the value of the __has_cpp_attribute macro is based on the year and month in which the attribute was voted into the working draft. In the case where the attribute is vendor-specific, the value is implementation-defined. However, in most cases it is expected that the availability of an attribute can be detected by any non-zero result.

The #ifdef and #ifndef directives, and the defined conditional inclusion operator, shall treat __has_cpp_attribute as if it were the name of a defined macro. The identifier __has_cpp_attribute shall not appear in any context not mentioned in this section.

2.3.3 Example

This demonstrates a way to use the attribute [[deprecated]] only if it is available.

#ifndef __has_cpp_attribute
# define __has_cpp_attribute(x) 0
#endif
#if __has_cpp_attribute(deprecated)
# define ATTR_DEPRECATED(msg) [[deprecated(msg)]]
#else
# define ATTR_DEPRECATED(msg)
#endif

3 Policies

SG-10 has adopted a number of policies related to our standard practices for determining and naming macros.

3.1 constexpr

For the language, we will have a single __cpp_constexpr macro. It will be bumped every time we extend constexpr in the language. For the library, we will add a specific feature-test macros for significant, special features. Otherwise, for those cases where we are just adding constexpr to more things in the library, we will have a dedicated test macro per header and just bump that header-specific macro on each change. That macro will be named __cpp_lib_constexpr_HEADER (with the exception of a few preexisting macros for array and algorithm which have slightly different names).

From [P1902R1].

3.2 Language Features with Library Components

In some cases a feature requires two macros, one for the language and one for the library. For example, the library does not want to define its three-way comparison operations unless the compiler supports the feature.

For end-users, it is suggested that they test only the library macro, as that will only be true if the language macro is also true. As a result, the language macros contain “impl” to distinguish them from the more general version that is expected to be set by the library.

Note that originally SG10 suggested that the library version of the macro not include the usual _lib part, but LWG was not comfortable with the inconsistency of having a library macro (which requires a header before it can be used) that does not contain _lib.

Also note that SG10 originally proposed that the core feature tests include _lang, but LWG wanted something that more clearly implied that the the macro was for a core feature and not intended to be used by end-users. They sugggested that _impl be used instead.

From [P1353R0].

4 Table of Feature-Test Macros

The following are all the feature-test macros in the standard, broken up by language macros, attribute macros, and library macros, then sorted by name. Each macro will contain its list of possible values and the papers necessary to be implemented before an implementation should define that macro with that particular value.

Note that a given paper may introduce or update multiple feature-test macros. A given value may require multiple papers. A paper may also remove a feature-test macro, in which case its value will be specified as deleted.

4.1 Language Feature-Test Macros

All of the language macros are predefined (i.e. no header needs to be included before doing checks).

In some cases, a feature requires two macros: one for the language and one for the library. For example, the library does not want to define its three-way comparison operators unless the compiler supports the feature. In these cases, it is suggested for end-users that they only test the library macro. Those core language feature-test macros that are intended to be checked by the library are spelled __cpp_impl_*.

4.2 Attribute Feature-Test Macros

All of the following macros are predefined.

4.3 Library Feature-Test Macros

All of the following macros are defined after inclusion of the header <version> or one of the corresponding headers specified below.

5 References