...one of the most highly
regarded and expertly designed C++ library projects in the
world.
— Herb Sutter and Andrei
Alexandrescu, C++
Coding Standards
Besides the functionality of the TTI library which queries whether some inner element of a given name within a type exists, the library also includes functionality for generating a nested type if it exists, else a marker type if it does not exist. By marker type is meant a type either internally created by the library, with no functionality, or designated by the end-user to represent the same idea.
First I will explain the syntax and use of this functionality and then the reason it exists in the library.
The functionality is a metafunction created by the macro BOOST_TTI_MEMBER_TYPE
.
The macro takes a single parameter, which is the name of a nested type. We
will call this our 'named type'. The macro generates a metafunction called
member_type_'named_type'
which, passed an enclosing type, returns the named type if it exists, else
a marker type if it does not.
As with our other macros we can use the alternative form of the macro BOOST_TTI_TRAIT_MEMBER_TYPE
to
pass first the name of the metafunction to be generated and then the name of
the 'named type'. After that the functionality of our resulting metafunction
is exactly the same.
Its general explanation is given as:
Table 1.3. TTI Nested Type Macro Metafunction
Inner Element |
Macro |
Template |
Specific Header File |
---|---|---|---|
Type |
|
class T = enclosing type class U = (optional) marker type returns = the type of 'name' if it exists, else a marker type, as the typedef 'type'. The invoked metafunction also holds the marker type as the typedef 'boost_tti_marker_type'. This is done for convenience so that the marker type does not have to be remembered. |
The marker type is purely optional. If not specified a type internal to the TTI library, which has no functionality, is used. Unless there is a specific reason for the end-user to provide his own marker type, he should let the TTI library use its own internal marker type.
A simple example of this functionality would be:
#include <boost/tti/member_type.hpp> BOOST_TTI_MEMBER_TYPE(ANamedType) typedef typename member_type_ANamedType<EnclosingType>::type AType;
If type 'ANamedType' is a nested type of 'EnclosingType' then AType is the same type as 'ANamedType', otherwise AType is a marker type internal to the TTI library.
Now that we have explained the syntax of BOOST_TTI_MEMBER_TYPE we can now answer the question of why this functionality to create a 'type' exists when looking for a nested type of an enclosing type.
The metafunctions generated by the TTI macros all work with various types, whether in specifying an enclosing type or in specifying the type of some inner element, which may also involve types in the signature of that element, such as a parameter or return type of a function. The C++ notation for a nested type, given an enclosing type 'T' and an inner type 'InnerType', is 'T::InnerType'. If either the enclosing type 'T' does not exist, or the inner type 'InnerType' does not exist within 'T', the expression 'T::InnerType' will give a compiler error if we attempt to use it in our template instantiation of one of TTI's macro metafunctions.
This is a problem if we want to be able to introspect for the existence of inner elements to an enclosing type without producing compiler errors. Of course if we absolutely know what types we have and that a nested type exists, and these declarations are within our scope, we can always use an expression like 'T::InnerType' without compiler error. But this is often not the case when doing template programming since the type being passed to us at compile-time in a class or function template is chosen at instantiation time and is created by the user of a template.
One solution to this is afforded by the library itself. Given an enclosing
type 'T' which we know must exist, either because it is a top-level type we
know about or it is passed to us in some template as a 'class T' or 'typename
T', and given an inner type named 'InnerType' whose existence we would like
ascertain, we can use a BOOST_TTI_HAS_TYPE(InnerType)
macro and it's related has_type_InnerType
metafunction to determine if the nested type 'InnerType' exists. This solution
is perfectly valid, and in conjunction with Boost MPL's selection metafunctions,
we can do compile-time selection to generate the correct template code.
However this does not scale that well syntactically if we need to drill down
further from a top-level enclosing type to a deeply nested type, or even to
look for some deeply nested type's inner elements. We are going to be generating
a great deal of boost::mpl::if_
and/or boost::mpl::eval_if
type selection statements to get to some final condition where we know we can
generate the compile-time code which we want.
The solution given by BOOST_TTI_MEMBER_TYPE is that we can create a type as the return from our metafunction, which is the same type as a nested type if it exists or some other marker type if it does not, and then work with that returned type without producing a compiler error. If we had to use the 'T::InnerType' syntax to specify our type, where 'T' represents out enclosing type and 'InnerType' our nested type, and there was no nested type 'InnerType' within the enclosing type 'T, the compiler would give us an error immediately.
By using BOOST_TTI_MEMBER_TYPE we have a type to work with even when such a type really does not exist. Naturally if the type does not exist, the type which we have to work with, being a marker type, will generally not fulfill any other further functionality we want from it. This is good and will normally produce the correct results in further uses of the type when doing metafunction programming. Occasionally the TTI produced marker type, when our nested type does not exist, is not sufficient for further metafunction programming. In that rare case the end-user can produce his own marker type to be used if the nested type does not exist. In any case, whether the nested type exists, whether the TTI default supplied marker type is used, or whether an end-user marker type is used, template metaprogramming can continue without a compilation problem. Furthermore this scales better than having to constant check for nested type existence via BOOST_TTI_HAS_TYPE in complicated template metaprogramming code.
Once we use BOOST_TTI_MEMBER_TYPE to generate a nested type if it exists we will normally use that type in further metafunction programming. Occasionally, given the type we generate, we will want to ask if the type is really our nested type or the marker type instead. Essentially we are asking if the type generated is the marker type or not. If it is the marker type, then the type generated is not the nested type we had hoped for. If it is not the marker type, then the type generated is the nested type we had hoped for. This is easy enough to do for the template metaprogrammer but TTI makes it easier by providing either of two metafunctions to do this calculation. These two metafunctions are 'boost::tti::valid_member_type' and 'boost::tti::valid_member_metafunction':
Table 1.4. TTI Nested Type Macro Metafunction Existence
Inner Element |
Macro |
Template |
Specific Header File |
---|---|---|---|
Type |
None |
class T = a type class U = (optional) marker type returns = true if the type exists, false if it does not. 'Existence' is determined by whether the type does not equal the marker type of BOOST_TTI_MEMBER_TYPE. |
|
Type |
None |
class T = a metafunction type returns = true if the return 'type' of the metafunction exists, false if it does not.'Existence' is determined by whether the return 'type' does not equal the marker type of BOOST_TTI_MEMBER_TYPE. |
In our first metafunction, 'boost::tti::valid_member_type', the first parameter is the return 'type' from invoking the metafunction generated by BOOST_TTI_MEMBER_TYPE. If when the metafunction was invoked a user-defined marker type had been specified, then the second optional parameter is that marker type, else it is not necessary to specify the optional second template parameter. Since the marker type is saved as the nested type boost::tti::marker_type once we invoke the metafunction generated by BOOST_TTI_MEMBER_TYPE we can always use that as our second template parameter to 'boost::tti::valid_member_type' if we like.
The second metafunction, boost::tti::valid_member_metafunction, makes the process of passing our nested 'type' and our marker type a bit easier. Here the single template parameter is the invoked metafunction generated by BOOST_TTI_MEMBER_TYPE itself. It then picks out from the invoked metafunction both the return 'type' and the nested boost::tti::marker_type to do the correct calculation.
A simple example of this functionality would be:
#include <boost/tti/member_type.hpp> struct UDMarkerType { }; BOOST_TTI_MEMBER_TYPE(ANamedType) typedef member_type_ANamedType<EnclosingType> IMType; typedef member_type_ANamedType<EnclosingType,UDMarkerType> IMTypeWithMarkerType;
then
boost::tti::valid_member_type<IMType::type>::value boost::tti::valid_member_type<IMTypeWithMarkerType::type,IMTypeWithMarkerType::boost_tti_marker_type>::value
or
boost::tti::valid_member_metafunction<IMType>::value boost::tti::valid_member_metafunction<IMTypeWithMarkerType>::value
gives us our compile-time result.
As an extended example, given a type T, let us create a metafunction where there is a nested type FindType whose enclosing type is eventually T, as represented by the following structure:
struct T { struct AType { struct BType { struct CType { struct FindType { }; } }; }; };
In our TTI code we first create a series of member type macros for each of our nested types:
BOOST_TTI_MEMBER_TYPE(AType) BOOST_TTI_MEMBER_TYPE(BType) BOOST_TTI_MEMBER_TYPE(CType) BOOST_TTI_MEMBER_TYPE(FindType)
Next we can create a typedef to reflect a nested type called FindType which
has the relationship as specified above by instantiating our macro metafunctions.
We have to do this in the reverse order of our hypothetical 'struct T' above
since the metafunction BOOST_TTI_MEMBER_TYPE
takes its enclosing type as its template parameter.
typedef typename member_type_FindType < typename member_type_CType < typename member_type_BType < typename member_type_AType < T >::type >::type >::type >::type MyFindType;
We can use the above typedef to pass the type as FindType to one of our macro metafunctions. FindType may not actually exist but we will not generate a compiler error when we use it. It will only generate, if it does not exist, an eventual failure by having whatever metafunction uses such a type return a false value at compile-time.
As one example, let's ask whether FindType has a static member data called MyData of type 'int'. We add:
BOOST_TTI_HAS_STATIC_MEMBER_DATA(MyData)
Next we create our metafunction:
has_static_member_data_MyData < MyFindType, int >
and use this in our metaprogramming code. Our metafunction now tells us whether the nested type FindType has a static member data called MyData of type 'int', even if FindType does not actually exist as we have specified it as a type. If we had tried to do this using normal C++ nested type notation our metafunction code above would be:
has_static_member_data_MyData < typename T::AType::BType::CType::FindType, int >
But this fails with a compiler error if there is no such nested type, and that is exactly what we do not want in our compile-time metaprogramming code.
In the above metafunction we are asking whether or not FindType has a static member data element called 'MyData', and the result will be 'false' if either FindType does not exist or if it does exist but does not have a static member data of type 'int' called 'MyData'. In neither situation will we produce a compiler error.
We may also be interested in ascertaining whether the deeply nested type 'FindType' actually exists. Our metafunction, using BOOST_TTI_MEMBER_TYPE and repeating our macros from above, could be:
BOOST_TTI_MEMBER_TYPE(FindType) BOOST_TTI_MEMBER_TYPE(AType) BOOST_TTI_MEMBER_TYPE(BType) BOOST_TTI_MEMBER_TYPE(CType) BOOST_TTI_HAS_TYPE(FindType) has_type_FindType < typename member_type_CType < typename member_type_BType < typename member_type_AType < T >::type >::type >::type >
But this duplicates much of our code when we generated the 'MyFindType' typedef. Instead we use the functionality already provided by 'boost::tti::valid_member_type'. Using this functionality with our 'MyFindType' type above we create the nullary metafunction:
boost::tti::valid_member_type < MyFindType >
directly instead of replicating the same functionality with our 'has_type_FindType' metafunction.