Coding Guidelines for Integral Constant Expressions

Integral Constant Expressions are used in many places in C++; as array bounds, as bit-field lengths, as enumerator initialisers, and as arguments to non-type template parameters. However many compilers have problems handling integral constant expressions; as a result of this, programming using non-type template parameters in particular can be fraught with difficulty, often leading to the incorrect assumption that non-type template parameters are unsupported by a particular compiler. This short article is designed to provide a set of guidelines and workarounds that, if followed, will allow integral constant expressions to be used in a manner portable to all the compilers currently supported by boost. Although this article is mainly targeted at boost library authors, it may also be useful for users who want to understand why boost code is written in a particular way, or who want to write portable code themselves.

What is an Integral Constant Expression?

Integral constant expressions are described in section 5.19 of the standard, and are sometimes referred to as "compile time constants". An integral constant expression can be one of the following:

A literal integral value, for example 0u or 3L.
An enumerator value.
Global integral constants, for example: const int my_INTEGRAL_CONSTANT = 3;
Static member constants, for example:
struct myclass { static const int value = 0; };
Member enumerator values, for example:
struct myclass { enum{ value = 0 }; };
Non-type template parameters of integral or enumerator type.
The result of a sizeof expression, for example:
sizeof(foo(a, b, c))
The result of a static_cast, where the target type is an integral or enumerator type, and the argument is either another integral constant expression, or a floating-point literal.
The result of applying a binary operator to two integral constant expressions:
INTEGRAL_CONSTANT1 op INTEGRAL_CONSTANT2 provided that the operator is not an assignment operator, or comma operator.
The result of applying a unary operator to an integral constant expression:
op INTEGRAL_CONSTANT1provided that the operator is not the increment or decrement operator.

Coding Guidelines

The following guidelines are declared in no particular order (in other words you need to obey all of them - sorry!), and may also be incomplete, more guidelines may be added as compilers change and/or more problems are encountered.

When declaring constants that are class members always use the macro BOOST_STATIC_CONSTANT.

template <class T>
struct myclass
{
   BOOST_STATIC_CONSTANT(int, value = sizeof(T));
};

Rationale: not all compilers support inline initialisation of member constants, others treat member enumerators in strange ways (they're not always treated as integral constant expressions). The BOOST_STATIC_CONSTANT macro uses the most appropriate method for the compiler in question.

Don't declare integral constant expressions whose type is wider than int.

Rationale: while in theory all integral types are usable in integral constant expressions, in practice many compilers limit integral constant expressions to types no wider than int.

Don't use logical operators in integral constant expressions; use template meta-programming instead.

The header <boost/type_traits/ice.hpp> contains a number of workaround templates, that fulfil the role of logical operators, for example instead of:

INTEGRAL_CONSTANT1 || INTEGRAL_CONSTANT2

Use:

::boost::type_traits::ice_or<INTEGRAL_CONSTANT1,INTEGRAL_CONSTANT2>::value

Rationale: A number of compilers (particularly the Borland and Microsoft compilers), tend to not to recognise integral constant expressions involving logical operators as genuine integral constant expressions. The problem generally only shows up when the integral constant expression is nested deep inside template code, and is hard to reproduce and diagnose.

Don't use any operators in an integral constant expression used as a non-type template parameter

Rather than:

typedef myclass<INTEGRAL_CONSTANT1 == INTEGRAL_CONSTANT2> mytypedef;

Use:

typedef myclass< some_symbol> mytypedef;

Where some_symbol is the symbolic name of a an integral constant expression whose value is (INTEGRAL_CONSTANT1 == INTEGRAL_CONSTANT2).

Rationale: the older EDG based compilers (some of which are used in the most recent version of that platform's compiler), don't recognise expressions containing operators as non-type template parameters, even though such expressions can be used as integral constant expressions elsewhere.

Always use a fully qualified name to refer to an integral constant expression.

For example:

typedef myclass< ::boost::is_integral<some_type>::value> mytypedef;

Rationale: at least one compiler (Borland's), doesn't recognise the name of a constant as an integral constant expression unless the name is fully qualified (which is to say it starts with ::).

Always leave a space after a '<' and before '::'

For example:

typedef myclass< ::boost::is_integral<some_type>::value> mytypedef;
                ^
                ensure there is space here!

Rationale: <: is a legal digraph in it's own right, so <:: is interpreted as the same as [:.

Don't use local names as integral constant expressions

Example:

template <class T>
struct foobar
{
   BOOST_STATIC_CONSTANT(int, temp = computed_value);
   typedef myclass<temp> mytypedef;  // error
};

Rationale: At least one compiler (Borland's) doesn't accept this.

Although it is possible to fix this by using:

template <class T>
struct foobar
{
   BOOST_STATIC_CONSTANT(int, temp = computed_value);
   typedef foobar self_type;
   typedef myclass<(self_type::temp)> mytypedef;  // OK
};

This breaks at least one other compiler (VC6), it is better to move the integral constant expression computation out into a separate traits class:

template <class T>
struct foobar_helper
{
   BOOST_STATIC_CONSTANT(int, temp = computed_value);
};

template <class T>
struct foobar
{
   typedef myclass< ::foobar_helper<T>::value> mytypedef;  // OK
};

Don't use dependent default parameters for non-type template parameters.

For example:

template <class T, int I = ::boost::is_integral<T>::value>  // Error can't deduce value of I in some cases.
struct foobar;

Rationale: this kind of usage fails for Borland C++. Note that this is only an issue where the default value is dependent upon a previous template parameter, for example the following is fine:

template <class T, int I = 3>  // OK, default value is not dependent
struct foobar;

Unresolved Issues

The following issues are either unresolved or have fixes that are compiler specific, and/or break one or more of the coding guidelines.

Be careful of numeric_limits

There are three issues here:

The header <limits> may be absent - it is recommended that you never include <limits> directly but use <boost/pending/limits.hpp> instead. This header includes the "real" <limits> header if it is available, otherwise it supplies it's own std::numeric_limits definition. Boost also defines the macro BOOST_NO_LIMITS if <limits> is absent.
The implementation of std::numeric_limits may be defined in such a way that its static-const members may not be usable as integral constant expressions. This contradicts the standard but seems to be a bug that affects at least two standard library vendors; boost defines BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS in <boost/config.hpp> when this is the case.
There is a strange bug in VC6, where the members of std::numeric_limits can be "prematurely evaluated" in template code, for example:

template <class T>
struct limits_test
{
   BOOST_STATIC_ASSERT(::std::numeric_limits<T>::is_specialized);
};

This code fails to compile with VC6 even though no instances of the template are ever created; for some bizarre reason ::std::numeric_limits<T>::is_specializedalways evaluates to false, irrespective of what the template parameter T is. The problem seems to be confined to expressions which depend on std::numeric_limts: for example if you replace ::std::numeric_limits<T>::is_specialized with ::boost::is_arithmetic<T>::value, then everything is fine. The following workaround also works but conflicts with the coding guidelines:

template <class T>
struct limits_test
{
   BOOST_STATIC_CONSTANT(bool, check = ::std::numeric_limits<T>::is_specialized);
   BOOST_STATIC_ASSERT(check);
};

So it is probably best to resort to something like this:

template <class T>
struct limits_test
{
#ifdef BOOST_MSVC
   BOOST_STATIC_CONSTANT(bool, check = ::std::numeric_limits<T>::is_specialized);
   BOOST_STATIC_ASSERT(check);
#else
   BOOST_STATIC_ASSERT(::std::numeric_limits<T>::is_specialized);
#endif
};

Be careful how you use the sizeof operator

As far as I can tell, all compilers treat sizeof expressions correctly when the argument is the name of a type (or a template-id), however problems can occur if:

The argument is the name of a member-variable, or a local variable (code may not compile with VC6).
The argument is an expression which involves the creation of a temporary (code will not compile with Borland C++).
The argument is an expression involving an overloaded function call (code compiles but the result is a garbage value with Metroworks C++).

Don't use boost::is_convertible unless you have to

Since is_convertible is implemented in terms of the sizeof operator, it consistently gives the wrong value when used with the Metroworks compiler, and may not compile with the Borland's compiler (depending upon the template arguments used).