...one of the most highly
regarded and expertly designed C++ library projects in the
world.
— Herb Sutter and Andrei
Alexandrescu, C++
Coding Standards
The header <boost/operators.hpp> supplies
several sets of class templates (in namespace boost
). These
templates define operators at namespace scope in terms of a minimal
number of fundamental operators provided by the class.
Overloaded operators for class types typically occur in groups. If you
can write x + y
, you probably also want to be able
to write x += y
. If you can write x < y,
you
also want x > y, x >= y,
and x <= y
.
Moreover, unless your class has really surprising behavior, some of these
related operators can be defined in terms of others (e.g. x >= y
<=> !(x < y)
). Replicating this boilerplate for multiple
classes is both tedious and errorprone. The boost/operators.hpp templates help
by generating operators for you at namespace scope based on other
operators you've defined in your class.
If, for example, you declare a class like this:
class MyInt : boost::operators<MyInt> { bool operator<(const MyInt& x) const; bool operator==(const MyInt& x) const; MyInt& operator+=(const MyInt& x); MyInt& operator=(const MyInt& x); MyInt& operator*=(const MyInt& x); MyInt& operator/=(const MyInt& x); MyInt& operator%=(const MyInt& x); MyInt& operator=(const MyInt& x); MyInt& operator&=(const MyInt& x); MyInt& operator^=(const MyInt& x); MyInt& operator++(); MyInt& operator(); };
then the operators<>
template adds more than a dozen additional operators, such as
operator>
, <=
, >=
, and
(binary) +
. Twoargument forms of the
templates are also provided to allow interaction with other types.
addable
template requires operator+=(T
const&)
and in turn supplies operator+(T const&, T
const&)
.The discussed concepts are not necessarily the standard library's
concepts (CopyConstructible, etc.), although some of them could
be; they are what we call concepts with a small 'c'. In
particular, they are different from the former ones in that they do
not describe precise semantics of the operators they require to be
defined, except the requirements that (a) the semantics of the operators
grouped in one concept should be consistent (e.g. effects of
evaluating of a += b
and
a = a + b
expressions should be the
same), and (b) that the return types of the operators should follow
semantics of return types of corresponding operators for builtin types
(e.g. operator<
should return a type convertible
to bool
, and T::operator=
should return type
convertible to T
). Such "loose" requirements make operators
library applicable to broader set of target classes from different
domains, i.e. eventually more useful.
The arguments to a binary operator commonly have identical types, but
it is not unusual to want to define operators which combine different
types. For example, one might want to multiply a
mathematical vector by a scalar. The twoargument template forms of the
arithmetic operator templates are supplied for this purpose. When
applying the twoargument form of a template, the desired return type of
the operators typically determines which of the two types in question
should be derived from the operator template. For example, if the result
of T + U
is of type T
, then
T
(not U
) should be derived from addable<T, U>
. The comparison templates
(less_than_comparable<T,
U>
, equality_comparable<T, U>
,
equivalent<T, U>
, and
partially_ordered<T,
U>
) are exceptions to this guideline, since the return type
of the operators they define is bool
.
On compilers which do not support partial specialization, the
twoargument forms must be specified by using the names shown below with
the trailing '2'
. The singleargument forms with the
trailing '1'
are provided for symmetry and to enable certain
applications of the base class chaining
technique.
Another application of the twoargument template forms is for mixed
arithmetics between a type T
and a type U
that
is convertible to T
. In this case there are two ways where
the twoargument template forms are helpful: one is to provide the
respective signatures for operator overloading, the second is
performance.
With respect to the operator overloading assume e.g. that
U
is int
, that T
is an
userdefined unlimited integer type, and that double
operator(double, const T&)
exists. If one wants to compute
int  T
and does not provide T operator(int, const
T&)
, the compiler will consider double operator(double,
const T&)
to be a better match than T operator(const
T&, const T&)
, which will probably be different from the
user's intention. To define a complete set of operator signatures,
additional 'left' forms of the twoargument template forms are provided
(subtractable2_left<T,
U>
, dividable2_left<T,
U>
, modable2_left<T,
U>
) that define the signatures for noncommutative
operators where U
appears on the left hand side
(operator(const U&, const T&)
,
operator/(const U&, const T&)
, operator%(const
U&, const T&)
).
With respect to the performance observe that when one uses the single
type binary operator for mixed type arithmetics, the type U
argument has to be converted to type T
. In practice,
however, there are often more efficient implementations of, say
T::operator=(const U&)
that avoid unnecessary
conversions from U
to T
. The twoargument
template forms of the arithmetic operator create additional operator
interfaces that use these more efficient implementations. There is,
however, no performance gain in the 'left' forms: they still need a
conversion from U
to T
and have an
implementation equivalent to the code that would be automatically created
by the compiler if it considered the single type binary operator to be
the best match.
Every operator class template, except the arithmetic examples and the iterator
helpers, has an additional, but optional, template type parameter
B
. This parameter will be a publiclyderived base class of
the instantiated template. This means it must be a class type. It can be
used to avoid the bloating of object sizes that is commonly associated
with multipleinheritance from several empty base classes (see the note for users of older versions for more
details). To provide support for a group of operators, use the
B
parameter to chain operator templates into a singlebase
class hierarchy, demostrated in the usage example.
The technique is also used by the composite operator templates to group
operator definitions. If a chain becomes too long for the compiler to
support, try replacing some of the operator templates with a single
grouped operator template that chains the old templates together; the
length limit only applies to the number of templates directly in the
chain, not those hidden in group templates.
Caveat: to chain to a base class which is
not a Boost operator template when using the singleargument form of a Boost operator template, you
must specify the operator template with the trailing '1'
in
its name. Otherwise the library will assume you mean to define a binary
operation combining the class you intend to use as a base class and the
class you're deriving.
On some compilers (e.g. Borland, GCC) even singleinheritance seems to cause an increase in object size in some cases. If you are not defining a class template, you may get better objectsize performance by avoiding derivation altogether, and instead explicitly instantiating the operator template as follows:
class myclass // lose the inheritance... { //... }; // explicitly instantiate the operators I need. template struct less_than_comparable<myclass>; template struct equality_comparable<myclass>; template struct incrementable<myclass>; template struct decrementable<myclass>; template struct addable<myclass,long>; template struct subtractable<myclass,long>;
Note that some operator templates cannot use this workaround and must be a base class of their primary operand type. Those templates define operators which must be member functions, and the workaround needs the operators to be independent friend functions. The relevant templates are:
dereferenceable<>
indexable<>
As Daniel Krügler pointed out, this technique violates 14.6.5/2
and is thus nonportable. The reasoning is, that the operators injected
by the instantiation of e.g.
less_than_comparable<myclass>
can not be found
by ADL according to the rules given by 3.4.2/2, since myclass is
not an associated class of
less_than_comparable<myclass>
.
Thus only use this technique if all else fails.
Many compilers (e.g. MSVC 6.3, GCC 2.95.2) will not enforce the
requirements in the operator template tables unless the operations which
depend on them are actually used. This is not standardconforming
behavior. In particular, although it would be convenient to derive all
your classes which need binary operators from the operators<>
and operators2<>
templates, regardless of
whether they implement all the requirements of those templates, this
shortcut is not portable. Even if this currently works with your
compiler, it may not work later.
This example shows how some of the arithmetic operator templates can be used with a geometric point class (template).
template <class T> class point // note: private inheritance is OK here! : boost::addable< point<T> // point + point , boost::subtractable< point<T> // point  point , boost::dividable2< point<T>, T // point / T , boost::multipliable2< point<T>, T // point * T, T * point > > > > { public: point(T, T); T x() const; T y() const; point operator+=(const point&); // point operator+(point, const point&) automatically // generated by addable. point operator=(const point&); // point operator(point, const point&) automatically // generated by subtractable. point operator*=(T); // point operator*(point, const T&) and // point operator*(const T&, point) autogenerated // by multipliable. point operator/=(T); // point operator/(point, const T&) autogenerated // by dividable. private: T x_; T y_; }; // now use the point<> class: template <class T> T length(const point<T> p) { return sqrt(p.x()*p.x() + p.y()*p.y()); } const point<float> right(0, 1); const point<float> up(1, 0); const point<float> pi_over_4 = up + right; const point<float> pi_over_4_normalized = pi_over_4 / length(pi_over_4);
The arithmetic operator templates ease the task of creating a custom numeric type. Given a core set of operators, the templates add related operators to the numeric class. These operations are like the ones the standard arithmetic types have, and may include comparisons, adding, incrementing, logical and bitwise manipulations, etc. Further, since most numeric types need more than one of these operators, some templates are provided to combine several of the basic operator templates in one declaration.
The requirements for the types used to instantiate the simple operator templates are specified in terms of expressions which must be valid and the expression's return type. The composite operator templates only list what other templates they use. The supplied operations and requirements of the composite operator templates can be inferred from the operations and requirements of the listed components.
These templates are "simple" since they provide operators based on a
single operation the base type has to provide. They have an additional
optional template parameter B
, which is not shown, for the
base class chaining technique.


Template  Supplied Operations  Requirements  

less_than_comparable<T> less_than_comparable1<T> 
bool operator>(const T&, const T&) bool operator<=(const T&, const T&) bool operator>=(const T&, const T&) 
t < t1 .Return convertible to bool . See the Ordering Note. 

less_than_comparable<T,
U> less_than_comparable2<T, U> 
bool operator<=(const T&, const U&) bool operator>=(const T&, const U&) bool operator>(const U&, const T&) bool operator<(const U&, const T&) bool operator<=(const U&, const T&) bool operator>=(const U&, const T&) 
t < u . t > u .Returns convertible to bool . See the Ordering Note. 

equality_comparable<T> equality_comparable1<T> 
bool operator!=(const T&, const T&) 
t == t1 .Return convertible to bool . 

equality_comparable<T,
U> equality_comparable2<T, U> 
bool operator==(const U&, const T&) bool operator!=(const U&, const T&) bool operator!=(const T&, const U&) 
t == u .Return convertible to bool . 

addable<T> addable1<T> 
T operator+(const T&, const T&) 
T temp(t); temp += t1 .Return convertible to T . See the Symmetry Note. 

addable<T, U> addable2<T, U> 
T operator+(const T&, const U&) T operator+(const U&, const T& ) 
T temp(t); temp += u .Return convertible to T . See the Symmetry Note. 

subtractable<T> subtractable1<T> 
T operator(const T&, const T&) 
T temp(t); temp = t1 .Return convertible to T . See the Symmetry Note. 

subtractable<T,
U> subtractable2<T, U> 
T operator(const T&, const U&) 
T temp(t); temp = u .Return convertible to T . See the Symmetry Note. 

subtractable2_left<T,
U> 
T operator(const U&, const T&) 
T temp(u); temp = t .Return convertible to T . 

multipliable<T> multipliable1<T> 
T operator*(const T&, const T&) 
T temp(t); temp *= t1 .Return convertible to T . See the Symmetry Note. 

multipliable<T,
U> multipliable2<T, U> 
T operator*(const T&, const U&) T operator*(const U&, const T&) 
T temp(t); temp *= u .Return convertible to T . See the Symmetry Note. 

dividable<T> dividable1<T> 
T operator/(const T&, const T&) 
T temp(t); temp /= t1 .Return convertible to T . See the Symmetry Note. 

dividable<T, U> dividable2<T, U> 
T operator/(const T&, const U&) 
T temp(t); temp /= u .Return convertible to T . See the Symmetry Note. 

dividable2_left<T,
U> 
T operator/(const U&, const T&) 
T temp(u); temp /= t .Return convertible to T . 

modable<T> modable1<T> 
T operator%(const T&, const T&) 
T temp(t); temp %= t1 .Return convertible to T . See the Symmetry Note. 

modable<T, U> modable2<T, U> 
T operator%(const T&, const U&) 
T temp(t); temp %= u .Return convertible to T . See the Symmetry Note. 

modable2_left<T,
U> 
T operator%(const U&, const T&) 
T temp(u); temp %= t .Return convertible to T . 

orable<T> orable1<T> 
T operator(const T&, const T&) 
T temp(t); temp = t1 .Return convertible to T . See the Symmetry Note. 

orable<T, U> orable2<T, U> 
T operator(const T&, const U&) T operator(const U&, const T&) 
T temp(t); temp = u .Return convertible to T . See the Symmetry Note. 

andable<T> andable1<T> 
T operator&(const T&, const T&) 
T temp(t); temp &= t1 .Return convertible to T . See the Symmetry Note. 

andable<T, U> andable2<T, U> 
T operator&(const T&, const U&) T operator&(const U&, const T&) 
T temp(t); temp &= u .Return convertible to T . See the Symmetry Note. 

xorable<T> xorable1<T> 
T operator^(const T&, const T&) 
T temp(t); temp ^= t1 .Return convertible to T . See the Symmetry Note. 

xorable<T, U> xorable2<T, U> 
T operator^(const T&, const U&) T operator^(const U&, const T&) 
T temp(t); temp ^= u .Return convertible to T . See the Symmetry Note. 

incrementable<T> 
T operator++(T&, int) 
T temp(t); ++t Return convertible to T . 

decrementable<T> 
T operator(T&, int) 
T temp(t); t; Return convertible to T . 

left_shiftable<T> left_shiftable1<T> 
T operator<<(const T&, const T&) 
T temp(t); temp <<= t1 .Return convertible to T . See the Symmetry Note. 

left_shiftable<T,
U> left_shiftable2<T, U> 
T operator<<(const T&, const U&) 
T temp(t); temp <<= u .Return convertible to T . See the Symmetry Note. 

right_shiftable<T> right_shiftable1<T> 
T operator>>(const T&, const T&) 
T temp(t); temp >>= t1 .Return convertible to T . See the Symmetry Note. 

right_shiftable<T,
U> right_shiftable2<T, U> 
T operator>>(const T&, const U&) 
T temp(t); temp >>= u .Return convertible to T . See the Symmetry Note. 

equivalent<T> equivalent1<T> 
bool operator==(const T&, const T&) 
t < t1 .Return convertible to bool . See the Ordering Note. 

equivalent<T, U> equivalent2<T, U> 
bool operator==(const T&, const U&) 
t < u . t > u .Returns convertible to bool . See the Ordering Note. 

partially_ordered<T> partially_ordered1<T> 
bool operator>(const T&, const T&) bool operator<=(const T&, const T&) bool operator>=(const T&, const T&) 
t < t1 . t == t1 .Returns convertible to bool . See the Ordering Note. 

partially_ordered<T,
U> partially_ordered2<T, U> 
bool operator<=(const T&, const U&) bool operator>=(const T&, const U&) bool operator>(const U&, const T&) bool operator<(const U&, const T&) bool operator<=(const U&, const T&) bool operator>=(const U&, const T&) 
t < u . t > u . t ==
u .Returns convertible to bool . See the Ordering Note. 
The less_than_comparable<T>
and
partially_ordered<T>
templates provide the same set of operations. However, the workings of
less_than_comparable<T>
assume
that all values of type T
can be placed in a total order. If
that is not true (e.g. NotaNumber values in IEEE floating point
arithmetic), then partially_ordered<T>
should be
used. The partially_ordered<T>
template can
be used for a totallyordered type, but it is not as efficient as
less_than_comparable<T>
. This
rule also applies for less_than_comparable<T, U>
and
partially_ordered<T,
U>
with respect to the ordering of all T
and
U
values, and for both versions of equivalent<>
. The solution for equivalent<>
is to write a custom
operator==
for the target class.
Before talking about symmetry, we need to talk about optimizations to
understand the reasons for the different implementation styles of
operators. Let's have a look at operator+
for a class
T
as an example:
T operator+( const T& lhs, const T& rhs ) { return T( lhs ) += rhs; }This would be a normal implementation of
operator+
, but it
is not an efficient one. An unnamed local copy of lhs
is
created, operator+=
is called on it and it is copied to the
function return value (which is another unnamed object of type
T
). The standard doesn't generally allow the intermediate
object to be optimized away:
3.7.2/2: Automatic storage durationThe reference to 12.8 is important for us:
If a named automatic object has initialization or a destructor with side effects, it shall not be destroyed before the end of its block, nor shall it be eliminated as an optimization even if it appears to be unused, except that a class object or its copy may be eliminated as specified in 12.8.
12.8/15: Copying class objectsThis optimization is known as the named return value optimization (NRVO), which leads us to the following implementation for
...
For a function with a class return type, if the expression in the return statement is the name of a local object, and the cvunqualified type of the local object is the same as the function return type, an implementation is permitted to omit creating the temporary object to hold the function return value, even if the class copy constructor or destructor has side effects.
operator+
:
T operator+( const T& lhs, const T& rhs ) { T nrv( lhs ); nrv += rhs; return nrv; }Given this implementation, the compiler is allowed to remove the intermediate object. Sadly, not all compiler implement the NRVO, some even implement it in an incorrect way which makes it useless here. Without the NRVO, the NRVOfriendly code is no worse than the original code showed above, but there is another possible implementation, which has some very special properties:
T operator+( T lhs, const T& rhs ) { return lhs += rhs; }The difference to the first implementation is that
lhs
is
not taken as a constant reference used to create a copy; instead,
lhs
is a byvalue parameter, thus it is already the copy
needed. This allows another optimization (12.2/2) for some cases.
Consider a + b + c
where the result of
a + b
is not copied when used as lhs
when adding c
. This is more efficient than the original
code, but not as efficient as a compiler using the NRVO. For most people,
it is still preferable for compilers that don't implement the NRVO, but
the operator+
now has a different function signature. Also,
the number of objects created differs for
(a + b ) + c
and
a + ( b + c )
. Most probably,
this won't be a problem for you, but if your code relies on the function
signature or a strict symmetric behaviour, you should set
BOOST_FORCE_SYMMETRIC_OPERATORS
in your userconfig. This
will force the NRVOfriendly implementation to be used even for compilers
that don't implement the NRVO. The following templates provide common groups of related operations.
For example, since a type which is addable is usually also subractable,
the additive
template provides the
combined operators of both. The grouped operator templates have an
additional optional template parameter B
, which is not
shown, for the base class chaining technique.


Template  Component Operator Templates  

totally_ordered<T> totally_ordered1<T> 

totally_ordered<T,
U> totally_ordered2<T, U> 

additive<T> additive1<T> 

additive<T, U> additive2<T, U> 

multiplicative<T> multiplicative1<T> 

multiplicative<T,
U> multiplicative2<T, U> 

integer_multiplicative<T> integer_multiplicative1<T> 

integer_multiplicative<T,
U> integer_multiplicative2<T, U> 

arithmetic<T> arithmetic1<T> 

arithmetic<T, U> arithmetic2<T, U> 

integer_arithmetic<T> integer_arithmetic1<T> 

integer_arithmetic<T,
U> integer_arithmetic2<T, U> 

bitwise<T> bitwise1<T> 

bitwise<T, U> bitwise2<T, U> 

unit_steppable<T> 

shiftable<T> shiftable1<T> 

shiftable<T, U> shiftable2<T, U> 

ring_operators<T> ring_operators1<T> 

ring_operators<T,
U> ring_operators2<T, U> 

ordered_ring_operators<T> ordered_ring_operators1<T> 

ordered_ring_operators<T,
U> ordered_ring_operators2<T, U> 

field_operators<T> field_operators1<T> 

field_operators<T,
U> field_operators2<T, U> 

ordered_field_operators<T> ordered_field_operators1<T> 

ordered_field_operators<T,
U> ordered_field_operators2<T, U> 

euclidian_ring_operators<T> euclidian_ring_operators1<T> 

euclidian_ring_operators<T,
U> euclidian_ring_operators2<T, U> 

ordered_euclidian_ring_operators<T> ordered_euclidian_ring_operators1<T> 

ordered_euclidian_ring_operators<T,
U> ordered_euclidian_ring_operators2<T, U> 
The arithmetic operator class templates operators<>
and operators2<>
are examples of
nonextensible operator grouping classes. These legacy class templates,
from previous versions of the header, cannot be used for base class chaining.


Template  Component Operator Templates  

operators<T> 

operators<T, U> operators2<T, U> 
The operators_test.cpp program demonstrates the use of the arithmetic operator templates, and can also be used to verify correct operation. Check the compiler status report for the test results with selected platforms.
The iterator helper templates ease the task of creating a custom iterator. Similar to arithmetic types, a complete iterator has many operators that are "redundant" and can be implemented in terms of the core set of operators.
The dereference operators were motivated by
the iterator helpers, but are often useful in
noniterator contexts as well. Many of the redundant iterator operators
are also arithmetic operators, so the iterator helper classes borrow many
of the operators defined above. In fact, only two new operators need to
be defined (the pointertomember operator>
and the
subscript operator[]
)!
The requirements for the types used to instantiate the dereference operators are specified in terms of expressions which must be valid and their return type. The composite operator templates list their component templates, which the instantiating type must support, and possibly other requirements.
All the dereference operator templates in this table accept an optional template parameter (not shown) to be used for base class chaining.


Template  Supplied Operations  Requirements  

dereferenceable<T,
P> 
P operator>() const 
(&*i) . Return convertible to
P . 

indexable<T, D,
R> 
R operator[](D n) const 
*(i + n) . Return of type
R . 
There are five iterator operator class templates, each for a different
category of iterator. The following table shows the operator groups for
any category that a custom iterator could define. These class templates
have an additional optional template parameter B
, which is
not shown, to support base class chaining.


Template  Component Operator Templates  

input_iteratable<T,
P> 

output_iteratable<T> 

forward_iteratable<T,
P> 

bidirectional_iteratable<T,
P> 

random_access_iteratable<T, P, D,
R> 
There are also five iterator helper class templates, each
corresponding to a different iterator category. These classes cannot be
used for base class chaining. The following
summaries show that these class templates supply both the iterator
operators from the iterator operator class
templates and the iterator typedef's required by the C++ standard
(iterator_category
, value_type
,
etc.).


Template  Operations & Requirements  

input_iterator_helper<T,
V, D, P, R> 
Supports the operations and has the requirements of  
output_iterator_helper<T> 
Supports the operations and has the requirements of See also [1], [2].  
forward_iterator_helper<T, V, D, P,
R> 
Supports the operations and has the requirements of  
bidirectional_iterator_helper<T,
V, D, P, R> 
Supports the operations and has the requirements of  
random_access_iterator_helper<T,
V, D, P, R> 
Supports the operations and has the requirements of
To satisfy RandomAccessIterator,
x1  x2 with return convertible to D is
also required.

[1] Unlike other iterator helpers templates,
output_iterator_helper
takes only one template parameter 
the type of its target class. Although to some it might seem like an
unnecessary restriction, the standard requires
difference_type
and value_type
of any output
iterator to be void
(24.3.1 [lib.iterator.traits]), and
output_iterator_helper
template respects this requirement.
Also, output iterators in the standard have void pointer
and
reference
types, so the output_iterator_helper
does the same.
[2] As selfproxying is the easiest and most common
way to implement output iterators (see, for example, insert [24.4.2] and
stream iterators [24.5] in the standard library),
output_iterator_helper
supports the idiom by defining
operator*
and operator++
member functions which
just return a nonconst reference to the iterator itself. Support for
selfproxying allows us, in many cases, to reduce the task of writing an
output iterator to writing just two member functions  an appropriate
constructor and a copyassignment operator. For example, here is a
possible implementation of boost::function_output_iterator
adaptor:
template<class UnaryFunction> struct function_output_iterator : boost::output_iterator_helper< function_output_iterator<UnaryFunction> > { explicit function_output_iterator(UnaryFunction const& f = UnaryFunction()) : func(f) {} template<typename T> function_output_iterator& operator=(T const& value) { this>func(value); return *this; } private: UnaryFunction func; };
Note that support for selfproxying does not prevent you from using
output_iterator_helper
to ease any other, different kind of
output iterator's implementation. If
output_iterator_helper
's target type provides its own
definition of operator*
or/and operator++
, then
these operators will get used and the ones supplied by
output_iterator_helper
will never be instantiated.
The iterators_test.cpp program demonstrates the use of the iterator templates, and can also be used to verify correct operation. The following is the custom iterator defined in the test program. It demonstrates a correct (though trivial) implementation of the core operations that must be defined in order for the iterator helpers to "fill in" the rest of the iterator operations.
template <class T, class R, class P> struct test_iter : public boost::random_access_iterator_helper< test_iter<T,R,P>, T, std::ptrdiff_t, P, R> { typedef test_iter self; typedef R Reference; typedef std::ptrdiff_t Distance; public: explicit test_iter(T* i =0); test_iter(const self& x); self& operator=(const self& x); Reference operator*() const; self& operator++(); self& operator(); self& operator+=(Distance n); self& operator=(Distance n); bool operator==(const self& x) const; bool operator<(const self& x) const; friend Distance operator(const self& x, const self& y); };
Check the compiler status report for the test results with selected platforms.
The changes in the library interface and
recommended usage were motivated by some practical issues described
below. The new version of the library is still backwardcompatible with
the former one (so you're not forced change any existing code),
but the old usage is deprecated. Though it was arguably simpler and more
intuitive than using base class chaining, it has
been discovered that the old practice of deriving from multiple operator
templates can cause the resulting classes to be much larger than they
should be. Most modern C++ compilers significantly bloat the size of
classes derived from multiple empty base classes, even though the base
classes themselves have no state. For instance, the size of
point<int>
from the example
above was 1224 bytes on various compilers for the Win32 platform,
instead of the expected 8 bytes.
Strictly speaking, it was not the library's faultthe language rules allow the compiler to apply the empty base class optimization in that situation. In principle an arbitrary number of empty base classes can be allocated at the same offset, provided that none of them have a common ancestor (see section 10.5 [class.derived] paragraph 5 of the standard). But the language definition also doesn't require implementations to do the optimization, and few if any of today's compilers implement it when multiple inheritance is involved. What's worse, it is very unlikely that implementors will adopt it as a future enhancement to existing compilers, because it would break binary compatibility between code generated by two different versions of the same compiler. As Matt Austern said, "One of the few times when you have the freedom to do this sort of thing is when you're targeting a new architecture...". On the other hand, many common compilers will use the empty base optimization for single inheritance hierarchies.
Given the importance of the issue for the users of the library (which
aims to be useful for writing lightweight classes like
MyInt
or point<>
), and the forces
described above, we decided to change the library interface so that the
object size bloat could be eliminated even on compilers that support only
the simplest form of the empty base class optimization. The current
library interface is the result of those changes. Though the new usage is
a bit more complicated than the old one, we think it's worth it to make
the library more useful in real world. Alexy Gurtovoy contributed the
code which supports the new usage idiom while allowing the library remain
backwardcompatible.
Revised: 03 Dec 2003
Copyright © Beman Dawes, David Abrahams, 19992001.
Copyright © Daniel Frey, 20022003.
Use, modification, and distribution is subject to the Boost Software License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at www.boost.org/LICENSE_1_0.txt)