Mixin-Based Programming in C++1

Yannis Smaragdakis Don Batory

College of Computing Department of Computer Sciences
Georgia Institute of Technology The University of Texas at Austin
Atlanta, GA 30332 Austin, Texas 78712
yannis@cc.gatech.edu batory@cs.utexas.edu

Abstract. Combinations of C++ features, like inheritance, templates, and class

nesting, allow for the expression of powerful component patterns. In particular,
research has demonstrated that, using C++ mixin classes, one can express lay-
ered component-based designs concisely with efficient implementations. In this
paper, we discuss pragmatic issues related to component-based programming
using C++ mixins. We explain surprising interactions of C++ features and poli-
cies that sometimes complicate mixin implementations, while other times enable
additional functionality without extra effort.

1 Introduction
Large software artifacts are arguably among the most complex products of human
intellect. The complexity of software has led to implementation methodologies that
divide a problem into manageable parts and compose the parts to form the final prod-
uct. Several research efforts have argued that C++ templates (a powerful parameteriza-
tion mechanism) can be used to perform this division elegantly.

In particular, the work of VanHilst and Notkin [29][30][31] showed how one can
implement collaboration-based (or role-based) designs using a certain templatized
class pattern, known as a mixin class (or just mixin). Compared to other techniques
(e.g., a straightforward use of application frameworks [17]) the VanHilst and Notkin
method yields less redundancy and reusable components that reflect the structure of
the design. At the same time, unnecessary dynamic binding can be eliminated, result-
ing into more efficient implementations. Unfortunately, this method resulted in very
complex parameterizations, causing its inventors to question its scalability.

The mixin layers technique was invented to address these concerns. Mixin layers are
mixin classes nested in a pattern such that the parameter (superclass) of the outer
mixin determines the parameters (superclasses) of inner mixins. In previous work
[4][24][25], we showed how mixin layers solve the scalability problems of the Van-
Hilst and Notkin method and result into elegant implementations of collaboration-
based designs.

1. We gratefully acknowledge the sponsorship of Microsoft Research, the Defense Advanced

Research Projects Agency (Cooperative Agreement F30602-96-2-0226), and the University
of Texas at Austin Applied Research Laboratories.
This paper discusses practical issues related to mixin-based programming. We adopt a
viewpoint oriented towards C++ implementations, but our discussion is not geared
towards the C++ expert. Instead, we aim to document common problems and solutions
in C++ mixin writing for the casual programmer. Additionally, we highlight issues that
pertain to language design in general (e.g., to Java parameterization or to the design of
future languages). Most of the issues clearly arise from the interaction of C++ features
with the constructs under study. The discussion mainly stems from actual experience
with C++ mixin-based implementations but a few points are a result of close examina-
tion of the C++ standard, since they refer to features that no compiler we have encoun-
tered implements. Even though we present an introduction to mixins, mixin layers, and
their uses, the primary purpose of this paper is not to convince readers of the value of
these constructs. (The reader should consult [4], [24], [25], [26], or [29] for that.)

We believe that the information presented here represents a valuable step towards mov-
ing some powerful programming techniques into the mainstream. We found that the
mixin programming style is quite practical, as long as one is aware of the possible
interactions with C++ idiosyncrasies. As C++ compilers move closer to sophisticated
template support (e.g., some compilers already support separate template compilation)
the utility of such techniques will increase rapidly.

2 Background (Mixins and Mixin Layers)

The term mixin class (or just mixin) has been overloaded in several occasions. Mixins
were originally explored in the Lisp language with object systems like Flavors [20] and
CLOS [18]. In these systems, mixins are an idiom for specifying a class and allowing
its superclass to be determined by linearization of multiple inheritance. In C++, the
term has been used to describe classes in a particular (multiple) inheritance arrange-
ment: as superclasses of a single class that themselves have a common virtual base
class (see [28], p.402). (This is not the meaning that we will use in this paper.) Both of
these mechanisms are approximations of a general concept described by Bracha and
Cook [6]. The idea is simple: we would like to specify an extension without pre-deter-
mining what exactly it can extend. This is equivalent to specifying a subclass while
leaving its superclass as a parameter to be determined later. The benefit is that a single
class can be used to express an incremental extension, valid for a variety of classes.

Mixins can be implemented using parameterized inheritance. The superclass of a class

is left as a parameter to be specified at instantiation time. In C++ we can write this as:

template <class Super>

class Mixin : public Super {
... /* mixin body */
};

To give an example, consider a mixin implementing operation counting for a graph.

Operation counting means keeping track of how many nodes and edges have been vis-
ited during the execution of a graph algorithm. (This simple example is one of the non-
algorithmic refinements to algorithm functionality discussed in [33]). The mixin could
have the form:

template <class Graph>

class Counting: public Graph {
int nodes_visited, edges_visited;
public:
Counting() : nodes_visited(0), edges_visited(0), Graph() { }
node succ_node (node v) {
nodes_visited++;
return Graph::succ_node(v);
}
edge succ_edge (edge e) {
edges_visited++;
return Graph::succ_edge(e);
}
...
};

By expressing operation counting as a mixin we ensure that it is applicable to many

classes that have the same interface (i.e., many different kinds of graphs). We can have,
for instance, two different compositions:
Counting< Ugraph > counted_ugraph;
and
Counting< Dgraph > counted_dgraph;
for undirected and directed graphs. (We omit parameters to the graph classes for sim-
plicity.) Note that the behavior of the composition is exactly what one would expect:
any methods not affecting the counting process are exported (inherited from the graph
classes). The methods that do need to increase the counts are “wrapped” in the mixin.

VanHilst and Notkin demonstrated that mixins are beneficial for a general class of
object-oriented designs [29]. They used a mixin-based approach to implement collabo-
ration-based (a.k.a. role-based) designs [5][15][16][21][29]. These designs are based
on the view that objects are composed of different roles that they play in their interac-
tion with other objects. The fundamental unit of functionality is a protocol for this
interaction, called a collaboration. The mixin-based approach of VanHilst and Notkin
results in efficient implementations of role-based designs with no redundancy. Some-
times, however, the resulting parameterization code is quite complicated—many mix-
ins need to be composed with others in a complex fashion. This introduces scalability
problems (namely, extensions that instantiate template parameters can be of length
exponential to the number of mixins composed—see [24]). To make the approach
more practical by reducing its complexity, mixin layers were introduced. Because
mixin layers are an incremental improvement of the VanHilst and Notkin method, we
only discuss implementing collaboration-based designs using mixin layers.

Mixin layers [24][25][26] are a particular form of mixins. They are designed with the
purpose of encapsulating refinements for multiple classes. Mixin layers are nested
mixins such that the parameter of an outer mixin determines the parameters of inner
mixins. The general form of a mixin layer in C++ is:

template <class NextLayer>

class ThisLayer : public NextLayer {
public:
class Mixin1 : public NextLayer::Mixin1 { ... };
class Mixin2 : public NextLayer::Mixin2 { ... };
...
};

Mixin layers are a result of the observation that a conceptual unit of functionality is
usually neither one object nor parts of an object—a unit of functionality may span sev-
eral different objects and specify refinements (extensions) to all of them. All such
refinements can be encapsulated in a single mixin layer and the standard inheritance
mechanism can be used for composing extensions.

This property of mixin layers makes them particularly attractive for implementing col-
laboration-based designs. Each layer captures a single collaboration. Roles for all
classes participating in a collaboration are represented by inner classes of the layer.
Inheritance works at two different levels. First, a layer can inherit entire classes from
its superclass (i.e., the parameter of the layer). Second, inner classes inherit members
(variables, methods, or even other classes) from the corresponding inner classes in the
superclass layer. This dual application of inheritance simplifies the implementation of
collaboration-based designs, while preserving the benefits of the VanHilst and Notkin
method. An important source of simplifications is that inner classes of a mixin layer
can refer unambiguously to other inner classes—the layer acts as a namespace.

We illustrate our point with an example (presented in detail in [24]) of a collaboration-

based design and its mixin layers implementation. (Full source code is available, upon
request.) This example presents a graph traversal application and was examined ini-
tially by Holland [16] and subsequently by VanHilst and Notkin [29]. This application
defines three different algorithms on an undirected graph, all implemented using a
depth-first traversal: Vertex Numbering numbers all nodes in the graph in depth-first
order, Cycle Checking examines whether the graph is cyclic, and Connected Regions
classifies graph nodes into connected graph regions. The application has three distinct
classes: Graph, Vertex, and Workspace. The Graph class describes a container of nodes
with the usual graph properties. Each node is an instance of the Vertex class. Finally,
the Workspace class includes the application part that is specific to each graph opera-
tion. For the VertexNumbering operation, for instance, a Workspace object holds the
value of the last number assigned to a vertex as well as the methods to update this
number.

As shown in Fig 1, we can decompose this application into five independent collabora-
tions—one encompassing the functionality of an undirected graph, another encoding
depth-first traversals, and three containing the specifics of each graph algorithm (ver-
tex numbering, cycle checking, and connected regions). Note that each collaboration
 Object Classes

Graph Vertex Workspace

Undirected Graph- VertexWith-

Graph Undirected Adjacencies
Collaborations (Layers)

Depth First GraphDFT VertexDFT

Traversal
Vertex VertexNumber Workspace-
Numbering Number

Cycle GraphCycle VertexCycle Workspace-

Checking Cycle

Connected Graph- Vertex- Workspace-

Connected Connected Connected
Region

Fig 1: Collaboration decomposition of the example application: A depth-first traversal of

an undirected graph is specialized to yield three different graph operations. Ovals rep-
resent collaborations, rectangles represent classes, their intersections represent roles.

captures a distinct aspect of the application and each object may participate in several
aspects. That is to say, each object may play several roles. For instance, the role of a
Graph object in the “Undirected Graph” collaboration supports storing and retrieving
a set of vertices. The role of the same object in the “Depth First Traversal” collabora-
tion implements a part of the actual depth-first traversal algorithm.

By implementing collaborations as mixin layers, the modular design of Fig 1 can be

maintained at the implementation level. For instance, the “Vertex Numbering” collabo-
ration can be implemented using a layer of the general form:

template <class Next>

class NUMBER : public Next {
public:
class Workspace : public Next::Workspace {
... // Workspace role members
};
class Vertex : public Next::Vertex {
... // Vertex role members
};
};

Note that no role (nested class) is prescribed for Graph. A Graph class is inherited
from the superclass of Number (the class denoted by parameter Next).

As shown in [24], such components are flexible and can be reused and interchanged.
For instance, the following composition builds Graph, Vertex, and WorkSpace
classes nested inside class CycleC that implement vertex numbering of undirected
graphs using a depth-first traversal:2

typedef DFT < NUMBER < DEFAULTW < UGRAPH > > > CycleC;

By replacing NUMBER with other mixin layers we get the other two graph algorithms
discussed. Many more combinations are possible. We can use the templates to create
classes that implement more than one algorithm. For instance, we can have an applica-
tion supporting both vertex numbering and cycle checking on the same graph by refin-
ing two depth-first traversals in order:

typedef DFT < NUMBER < DEFAULTW < UGRAPH > > > NumberC;
typedef DFT < CYCLE < NumberC > > CycleC;

Furthermore, all the characteristics of an undirected graph are captured by the UGRAPH
mixin layer. Hence, it is straightforward to apply the same algorithms to a directed
graph (mixin layer DGRAPH interchanged for UGRAPH):3

typedef DFT < NUMBER < DEFAULTW < DGRAPH > > > NumberC;

This technique (of composing source components in a large number of combinations)

underlies the scalable libraries [3] design approach for source code plug-and-play
components.

3 Programming with C++ Mixins: Pragmatic Considerations

Since little has been written about the pragmatics of doing component programming
using C++ mixins (mixin classes or mixin layers), we feel it is necessary to discuss
some pertinent issues. Most of the points raised below concern fine interactions
between the mixin approach and C++ idiosyncrasies. Others are implementation sug-
gestions. They are all useful knowledge before one embarks into a development effort
using C++ mixins and could serve to guide design choices for future parameterization
mechanisms in programming languages. The C++ aspects we discuss are well-docu-
mented and other C++ programmers have probably also made some of our observa-
tions. Nevertheless, we believe that most of them are non-obvious and many only arise
in the context of component programming—that is, when a mixin is designed and used
in complete isolation from other components of the system.

Lack of template type-checking. Templates do not correspond to types in the C++

language. Thus, they are not type-checked until instantiation time (that is, composition
time for mixins). Furthermore, methods of templatized classes are themselves consid-

2. The DEFAULTW mixin layer is an implementation detail, borrowed from the VanHilst and Not-
kin implementation [29]. It contains an empty WorkSpace class and its purpose is to avoid
dynamic binding by changing the order of composition.
3. This is under the assumption that the algorithms are still valid for directed graphs as is the
case with the original code for this example [16].
ered function templates.4 Function templates in C++ are instantiated automatically and
only when needed. Thus, even after mixins are composed, not all their methods will be
type-checked (code will only be produced for methods actually referenced in the
object code). This means that certain errors (including type mismatches and references
to undeclared methods) can only be detected with the right template instantiations and
method calls. Consider the following example:

template <class Super>

class ErrorMixin : public Super {
public:
...
void sort(FOO foo) {
Super::srot(foo); // misspelled
}
};

If client code never calls method sort, the compiler will not catch the misspelled
identifier above. This is true even if the ErrorMixin template is used to create
classes, and methods other than sort are invoked on objects of those classes.

Delaying the instantiation of methods in template classes can be used to advantage, as

we will see later. Nevertheless, many common designs are such that all member meth-
ods of a template class should be valid for all instantiations. It is not straightforward to
enforce the latter part (“for all instantiations”) but for most practical purposes check-
ing all methods for a single instantiation is enough. This can be done by explicit
instantiation of the template class, which forces the instantiation of all its members.
The idiom for explicit instantiation applied to our above example is:

template class ErrorMixin<SomeFoo>;

When “subtype of” does not mean “substitutable for”. There are two instances
where inheritance may not behave the way one might expect in C++. First, constructor
methods are not inherited. Ellis and Stroustrup ([13], p.264) present valid reasons for
this design choice: the constructor of a superclass does not suffice for initializing data
members added by a subclass. Often, however, a mixin class may be used only to
enrich or adapt the method interface of its superclasses without adding data members.
In this case it would be quite reasonable to inherit a constructor, which, unfortunately,
is not possible. The practical consequence of this policy is that the only constructors
that are visible in the result of a mixin composition are the ones present in the outer-
most mixin (bottom-most class in the resulting inheritance hierarchy). To make matters
worse, constructor initialization lists (e.g.,
constr() : init1(1,2), init2(3) {} )
can only be used to initialize direct parent classes. In other words, all classes need to
know the interface for the constructor of their direct superclass (if they are to use con-

4. This wording, although used by the father of C++—see [28], p.330—is not absolutely accu-
rate since there is no automatic type inference.
structor initialization lists). Recall, however, that a desirable property for mixins is that
they be able to act as components: a mixin should be implementable in isolation from
other parts of the system in which it is used. Thus a single mixin class should be usable
with several distinct superclasses and should have as few dependencies as possible. In
practice, several mixin components are just thin wrappers adapting their superclass’s
interface.

A possible workaround for this problem is to use a standardized construction interface.

A way to do this is by creating a construction class encoding the union of all possible
arguments to constructors in a hierarchy. Then a mixin “knows” little about its direct
superclass, but has dependencies on the union of the construction interfaces for all its
possible parent classes. (Of course, another workaround is to circumvent constructors
altogether by having separate initialization methods. This, however, requires a disci-
plined coding style to ensure that methods are always called after object construction.)
As a side-note, destructors for base classes are called automatically so they should not
be replicated.

Synonyms for compositions. In the past sections we have used typedefs to intro-
duce synonyms for complicated mixin compositions—e.g.,
typedef A < B < C > > Synonym;
Another reasonable approach would be to introduce an empty subclass:

class Synonym : public A < B < C > > { };

The first form has the advantage of preserving constructors of component A in the syn-
onym. The second idiom is cleanly integrated into the language (e.g., can be templa-
tized, compilers create short link names for the synonym, etc.). Additionally, it can
solve a common problem with C++ template-based programming: generated names
(template instantiations) can be extremely long, causing compiler messages to be
incomprehensible.

Designating virtual methods. Sometimes C++ policies have pleasant side-effects

when used in conjunction with mixins. An interesting case is that of a mixin used to
create classes where a certain method can be virtual or not, depending on the concrete
class used to instantiate the mixin. This is due to the C++ policy of letting a superclass
declare whether a method is virtual, while the subclass does not need to specify this
explicitly. Consider a regular mixin and two concrete classes instantiating it (a C++
struct is a class whose members are public by default):

template <class Super>

struct MixinA : public Super {
void virtual_or_not(FOO foo) { ... }
};

struct Base1 {
virtual void virtual_or_not(FOO foo) { ... }
 ... // methods using “virtual_or_not”
};

struct Base2 {
void virtual_or_not(FOO foo) { ... }
};

The composition MixinA<Base1> designates a class in which the method

virtual_or_not is virtual. Conversely, the same method is not virtual in the compo-
sition MixinA<Base2>. Hence, calls to virtual_or_not in Base1 will call the
method supplied by the mixin in the former case but not in the latter.

In the general case, this phenomenon allows for interesting mixin configurations.
Classes at an intermediate layer may specify methods and let the inner-most layer
decide whether they are virtual or not.

As we recently found out, this technique was described first in [12].

Single mixin for multiple uses. The lack of template type-checking in C++ can actu-
ally be beneficial in some cases. Consider two classes Base1 and Base2 with very
similar interfaces (except for a few methods):

struct Base1 {
void regular() { ... }
...
};
struct Base2 {
void weird() { ... }
... // otherwise same interface as Base1
};

Because of the similarities between Base1 and Base2, it makes sense to use a single
mixin to adapt both. Such a mixin may need to have methods calling either of the
methods specific to one of the two base classes. This is perfectly feasible. A mixin can
be specified so that it calls either regular or weird:

template <class Super>

class Mixin : public Super {
...
public:
void meth1() { Super::regular(); }
void meth2() { Super::weird(); }
};

This is a correct definition and it will do the right thing for both composition
Mixin<Base1> and Mixin<Base2>! What is remarkable is that part of Mixin seems
invalid (calls an undefined method), no matter which composition we decide to per-
form. But, since methods of class templates are treated as function templates, no error
will be signalled unless the program actually uses the wrong method (which may be
meth1 or meth2 depending on the composition). That is, an error will be signalled
only if the program is indeed wrong. We have used this technique to provide uniform,
componentized extensions to data structures supporting slightly different interfaces (in
particular, the red-black tree and hash table of the SGI implementation of the Standard
Template Library [22]).

Propagating type information. An interesting practical technique (also applicable to

languages other than C++) can be used to propagate type information from a subclass
to a superclass, when both are created from instantiating mixins. This is a common
problem in object-oriented programming. It was, for instance, identified in the design
of the P++ language [23] (an extension of C++ with constructs for component-based
programming) and solved with the addition of the forward keyword. The same prob-
lem is addressed in other programming languages (e.g., Beta [19]) with the concept of
virtual types.

Consider a mixin layer encapsulating the functionality of an allocator. This component

needs to have type information propagated to it from its subclasses (literally, the sub-
classes of the class it will create when instantiated) so that it knows what kind of data
to allocate. (We also discussed this example in detail in [25] but we believe that the
solution presented here is the most practical way to address the problem.) The reason
this propagation is necessary is that subclasses may need to add data members to a
class used by the allocator. One can solve the problem by adding an extra parameter to
the mixin that will be instantiated with the final product of the composition itself. In
essence, we are reducing a conceptual cycle in the parameterization to a single self-ref-
erence (which is well-supported in C++). This is shown in the following code frag-
ment:

template <class EleType, class FINAL>

class ALLOC {
public:
class Node {
EleType element; // stored data type
public:
... // methods using stored data
};

class Container {
protected:
FINAL::Node* node_alloc() {
return new FINAL::Node();
}
... // Other allocation methods
};
};

template <class Super>

class BINTREE : public Super {
public:
 class Node : public Super::Node {
Node* parent_link, left_link, right_link ;
public:
... // Node interface
};

class Container : public Super::Container {

Node* header; // Container data members
public:
... // Interface methods
};
};

class Comp : public BINTREE < ALLOC <int, Comp> > {/* empty */};

Note what is happening in this code fragment (which is abbreviated but preserves the
structure of actual code that we have used). A binary tree data structure is created by
composing a BINTREE mixin layer with an ALLOC mixin layer. The data structure
stores integer (int) elements. Nevertheless, the actual type of the element stored is not
int but a type describing the node of a binary tree (i.e., an integer together with three
pointers for the parent, and the two children of the node). This is the type of element
that the allocator should reserve memory for.

The problem is solved by passing the final product of the composition as a parameter
to the allocator mixin. This is done through the self-referential (or recursive) declara-
tion of class Comp. (Theoretically-inclined readers will recognize this as a fixpoint con-
struction.) Note that Comp is just a synonym for the composition and it has to use the
synonym pattern introducing a class (i.e., the typedef synonym idiom discussed ear-
lier would not work as it does not support recursion).

It should be noted that the above recursive construction has been often used in the liter-
ature. In the C++ world, the technique was introduced by Barton and Nackman [2] and
popularized by Coplien [9]. Nevertheless, the technique is not mixin-specific or even
C++-specific. For instance, it was used by Wadler, Odersky and the first author [32] in
Generic Java [7] (an extension of Java with parametric polymorphism). The origins of
the technique reach back at least to the development of F-bounded polymorphism [8].

Hygienic templates in the C++ standard. The C++ standard ([1], section 14.6)
imposes several rules for name resolution of identifiers that occur inside templates.
The extent to which current compilers implement these rules varies, but full conform-
ance is the best approach to future compatibility for user code.

Although the exact rules are complicated, one can summarize them (at loss of some
detail) as “templates cannot contain code that refers to ‘nonlocal’ variables or meth-
ods”. Intuitively, “nonlocal” denotes variables or methods that do not depend on a tem-
plate parameter and are not in scope at the global point closest to the template
definition. This rule prevents template instantiations from capturing arbitrary names
from their instantiation context, which could lead to behavior not predicted by the tem-
plate author.

A specific rule applies to mixin-based programming. To quote the C++ standard

(14.6.2), “if a base class is a dependent type, a member of that class cannot hide a
name declared within a template, or a name from the templates enclosing scopes”.
Consider the example of a mixin calling a method defined in its parameter (i.e., the
superclass of the class it will create when instantiated):

struct Base {
void foo() { ... }
};

void foo() { }

template <class Super>

struct Mixin : public Super {
void which_one() { foo(); } // ::foo
};

Mixin < Base > test;

That is, the call to foo from method which_one will refer to the global foo, not the
foo method of the Base superclass.

The main implication of these name resolution rules is on the way template-based pro-
grams should be developed. In particular, imagine changing a correct class definition
into a mixin definition (by turning the superclass into a template parameter). Even if
the mixin is instantiated with its superclass in the original code, the new program is not
guaranteed to work identically to the original, because symbols may now be resolved
differently. This may surprise programmers who work by creating concrete classes and
turning them into templates when the need for abstraction arises. To avoid the potential
for insidious bugs, it is a good practice to explicitly qualify references to superclass
methods (e.g., Super::foo instead of just foo).

Compiler support. Most C++ compilers now have good support for parameterized
inheritance (the technique we used for mixins) and nested classes. We have encoun-
tered few problems and mostly with older compilers when programming with C++
mixins. In fact, most of the compiler dependencies are not particular to mixin-based
programming but concern all template-based C++ programs. These include limitations
on the debugging support, error checking, etc. We will not discuss such issues as they
are time-dependent and have been presented before (e.g., [10]). Note, however, that
mixin-based programming is not more complex than regular template instantiation and
typically does not exercise any of the “advanced” features of C++ templates (type
inference, higher-order templates, etc.). Overall, the compiler support issues involved
in mixin-based programming are about the same as those arising in implementing the
C++ Standard Template Library [27].
4 Related Work
Various pieces of related work have been presented in the previous sections. We cannot
exhaustively reference all C++ template-based programming techniques, but we will
discuss two approaches that are distinct from ours but seem to follow parallel courses.

The most prominent example is the various implementations of the STL. Such imple-
mentations often exercise the limits of template support and reveal interactions of C++
policies with template-based programming. Nevertheless, parameterized inheritance is
not a part of STL implementations. Hence, the observations of this paper are mostly
distinct from the conclusions drawn from STL implementation efforts.

Czarnecki and Eisenecker’s generative programming techniques [10][11] were used in

the Generative Matrix Computation Library (GMCL). Their approach is a representa-
tive of techniques using C++ templates as a programming language (that is, to perform
arbitrary computation at template instantiation time). What sets their method apart
from other template meta-programming techniques is that it has similar goals to mixin-
based programming. In particular, Czarnecki and Eisenecker try to develop compo-
nents which can be composed in multiple ways to yield a variety of implementations.
Several of the remarks in this paper are applicable to their method, even though their
use of mixins is different (for instance, they do not use mixin layers).

5 Conclusions
We presented some pragmatic issues pertaining to mixin-based programming in C++.
We believe that mixin-based techniques are valuable and will become much more
widespread in the future. Mixin-based programming promises to provide reusable soft-
ware components that result into flexible and efficient implementations.

Previous papers have argued for the value of mixin-based software components and
their advantages compared to application frameworks. In this paper we tried to make
explicit the engineering considerations specific to mixin-based programming in C++.
Our purpose is to inform programmers of the issues involved in order to help move
mixin-based programming into the mainstream.

References
[1] ANSI / ISO Standard: Programming Languages—C++, ISO/IEC 14882, 1998.
[2] J. Barton and L.R. Nackman, Scientific and Engineering C++: An Introduction
with Advanced Techniques and Applications, Addison-Wesley, 1994.
[3] D. Batory, V. Singhal, M. Sirkin, and J. Thomas, “Scalable Software Libraries”,
ACM SIGSOFT 1993.
[4] D. Batory, R. Cardone, and Y. Smaragdakis, “Object-Oriented Frameworks and
Product-Lines”, 1st Software Product-Line Conference, Denver, Colorado,
August 1999.
[5] K. Beck and W. Cunningham, “A Laboratory for Teaching Object-Oriented
Thinking”, OOPSLA 1989, 1-6.
[6] G. Bracha and W. Cook, “Mixin-Based Inheritance”, ECOOP/OOPSLA 90, 303-
311.
[7] G. Bracha, M. Odersky, D. Stoutamire and P. Wadler, “Making the future safe for
the past: Adding Genericity to the Java Programming Language”, OOPSLA 98.
[8] P. Canning, W. Cook, W. Hill, W. Olthoff, and J. C. Mitchell, “F-bounded
Polymorphism for Object-Oriented Programming”, in Proc. Conf. on Functional
Programming Languages and Computer Architecture, 1989, 273-280.
[9] J. Coplien, “Curiously Recurring Template Patterns”, C++ Report, 7(2):24-27,
Feb. 1995.
[10] K. Czarnecki and U. Eisenecker. Generative Programming: Methods, Tools, and
Applications. Addison-Wesley, 2000.
[11] K. Czarnecki and U. Eisenecker, “Synthesizing Objects”, ECOOP 1999, 18-42.
[12] U. Eisenecker, “Generative Programming in C++”, in Proc. Joint Modular
Languages Conference (JMLC’97), LNCS 1204, Springer, 1997, 351-365.
[13] M.A. Ellis and B. Stroustrup, The Annotated C++ Reference Manual, Addison-
Wesley, 1990.
[14] E. Gamma, R. Helm, R. Johnson, and J. Vlissides, Design Patterns: Elements of
Reusable Object-Oriented Software. Addison-Wesley, 1994.
[15] R. Helm, I. Holland, and D. Gangopadhyay, “Contracts: Specifying Behavioral
Compositions in Object-Oriented Systems”. OOPSLA 1990, 169-180.
[16] I. Holland, “Specifying Reusable Components Using Contracts”, ECOOP 1992,
287-308.
[17] R. Johnson and B. Foote, “Designing Reusable Classes”, J. of Object-Oriented
Programming, 1(2): June/July 1988, 22-35.
[18] G. Kiczales, J. des Rivieres, and D. G. Bobrow, The Art of the Metaobject
Protocol. MIT Press, 1991.
[19] O.L. Madsen, B. Møller-Pedersen, and K. Nygaard, Object-Oriented
Programming in the BETA Programming Language. Addison-Wesley, 1993.
[20] D.A. Moon, “Object-Oriented Programming with Flavors”, OOPSLA 1986.
[21] T. Reenskaug, E. Anderson, A. Berre, A. Hurlen, A. Landmark, O. Lehne, E.
Nordhagen, E. Ness-Ulseth, G. Oftedal, A. Skaar, and P. Stenslet, “OORASS:
Seamless Support for the Creation and Maintenance of Object-Oriented
Systems”, J. of Object-Oriented Programming, 5(6): October 1992, 27-41.
[22] Silicon Graphics Computer Systems Inc., STL Programmer’s Guide. See:
http://www.sgi.com/Technology/STL/ .
[23] V. Singhal, A Programming Language for Writing Domain-Specific Software
System Generators, Ph.D. Dissertation, Dep. of Computer Sciences, University
of Texas at Austin, August 1996.
[24] Y. Smaragdakis and D. Batory, “Implementing Reusable Object-Oriented
Components”. In the 5th Int. Conf. on Software Reuse (ICSR 98).
[25] Y. Smaragdakis and D. Batory, “Implementing Layered Designs with Mixin
Layers”. In ECOOP 98.
[26] Y. Smaragdakis, “Implementing Large-Scale Object-Oriented Components”,
Ph.D. Dissertation, Department of Computer Sciences, University of Texas at
Austin, December 1999.
[27] A. Stepanov and M. Lee, “The Standard Template Library”. Incorporated in
ANSI/ISO Committee C++ Standard.
[28] B. Stroustrup, The C++ Programming Language, 3rd Ed., Addison-Wesley,
1997.
[29] M. VanHilst and D. Notkin, “Using C++ Templates to Implement Role-Based
Designs”. JSSST International Symposium on Object Technologies for Advanced
Software, Springer-Verlag, 1996, 22-37.
[30] M. VanHilst and D. Notkin, “Using Role Components to Implement
Collaboration-Based Designs”. OOPSLA 1996.
[31] M. VanHilst and D. Notkin, “Decoupling Change From Design”, SIGSOFT 96.
[32] P. Wadler, M. Odersky and Y. Smaragdakis, “Do Parametric Types Beat Virtual
Types?”, unpublished manuscript, posted in October 1998 in the Java Genericity
mailing list (java-genericity@cs.rice.edu).
[33] K. Weihe, “A Software Engineering Perspective on Algorithmics”, available at
http://www.informatik.uni-konstanz.de/Preprints/ .