Tuple Library : design decisions rationale

About namespaces

There was a discussion about whether tuples should be in a separate namespace or directly in the boost namespace. The common principle is that domain libraries (like graph, python) should be on a separate subnamespace, while utility like libraries directly in the boost namespace. Tuples are somewhere in between, as the tuple template is clearly a general utility, but the library introduces quite a lot of names in addition to just the tuple template. Tuples were originally under a subnamespace. As a result of the discussion, tuple definitions were moved directly under the boost namespace. As a result of a continued discussion, the subnamespace was reintroduced. The final (I truly hope so) solution is now to have all definitions in namespace ::boost::tuples, and the most common names in the ::boost namespace as well. This is accomplished with using declarations (suggested by Dave Abrahams):

namespace boost {
  namespace tuples {
      ...      
    // All library code
      ...
  }
  using tuples::tuple; 
  using tuples::make_tuple;
  using tuples::tie;
  using tuples::get;
}

With this arrangement, tuple creation with direct constructor calls, make_tuple or tie functions do not need the namespace qualifier. Further, all functions that manipulate tuples are found with Koenig-lookup. The only exceptions are the get<N> functions, which are always called with an explicitly qualified template argument, and thus Koenig-lookup does not apply. Therefore, get is lifted to ::boost namespace with a using declaration. Hence, the interface for an application programmer is in practice under the namespace ::boost.

The other names, forming an interface for library writers (cons lists, metafunctions manipulating cons lists, ...) remain in the subnamespace ::boost::tuples. Note, that the names ignore, set_open, set_close and set_delimiter are considered to be part of the application programmer's interface, but are still not under boost namespace. The reason being the danger for name clashes for these common names. Further, the usage of these features is probably not very frequent.

For those who are really interested in namespaces

The subnamespace name tuples raised some discussion. The rationale for not using the most natural name 'tuple' is to avoid having an identical name with the tuple template. Namespace names are, however, not generally in plural form in boost libraries. First, no real trouble was reported for using the same name for a namespace and a class and we considered changing the name 'tuples' to 'tuple'. But we found some trouble after all. Both gcc and edg compilers reject using declarations where the namespace and class names are identical:

namespace boost {
  namespace tuple {
    ... tie(...);
    class tuple; 
      ...
  }
  using tuple::tie; // ok
  using tuple::tuple; // error
    ...
}

Note, however, that a corresponding using declaration in the global namespace seems to be ok:

using boost::tuple::tuple; // ok;

The end mark of the cons list (nil, null_type, ...)

Tuples are internally represented as cons lists:

tuple<int, int>

inherits from

cons<int, cons<int, null_type> >

null_type is the end mark of the list. Original proposition was nil, but the name is used in MacOS, and might have caused problems, so null_type was chosen instead. Other names considered were null_t and unit (the empty tuple type in SML).

Note that null_type is the internal representation of an empty tuple: tuple<> inherits from null_type.

Element indexing

Whether to use 0- or 1-based indexing was discussed more than thoroughly, and the following observations were made:

• 0-based indexing is 'the C++ way' and used with arrays etc.
• 1-based 'name like' indexing exists as well, eg. bind1st, bind2nd, pair::first, etc.

Tuple access with the syntax get<N>(a), or a.get<N>() (where a is a tuple and N an index), was considered to be of the first category, hence, the index of the first element in a tuple is 0.

A suggestion to provide 1-based 'name like' indexing with constants like _1st, _2nd, _3rd, ... was made. By suitably chosen constant types, this would allow alternative syntaxes:

a.get<0>() == a.get(_1st) == a[_1st] == a(_1st);

We chose not to provide more than one indexing method for the following reasons:

• 0-based indexing might not please everyone, but once its fixed, it is less confusing than having two different methods (would anyone want such constants for arrays?).
• Adding the other indexing scheme doesn't really provide anything new (like a new feature) to the user of the library.
• C++ variable and constant naming rules don't give many possibilities for defining short and nice index constants (like _1st, ...). Let the binding and lambda libraries use these for a better purpose.
• The access syntax a[_1st] (or a(_1st)) is appealing, and almost made us add the index constants after all. However, 0-based subscripting is so deep in C++, that we had a fear for confusion.
• Such constants are easy to add.

Tuple comparison

The comparison operator implements lexicographical order. Other orderings were considered, mainly dominance (a < b iff for each i a(i) < b(i)). Our belief is, that lexicographical ordering, though not mathematically the most natural one, is the most frequently needed ordering in everyday programming.

Streaming

The characters specified with tuple stream manipulators are stored within the space allocated by ios_base::xalloc, which allocates storage for long type objects. static_cast is used in casting between long and the stream's character type. Streams that have character types not convertible back and forth to long thus fail to compile.

This may be revisited at some point. The two possible solutions are:

• Allow only plain char types as the tuple delimiters and use widen and narrow to convert between the real character type of the stream. This would always compile, but some calls to set manipulators might result in a different character than expected (some default character).
• Allocate enough space to hold the real character type of the stream. This means memory for holding the delimiter characters must be allocated separately, and that pointers to this memory are stored in the space allocated with ios_base::xalloc. Any volunteers?

© Copyright Jaakko Järvi 2001.