Section: Scientific Foundations

Adaptation Methodologies

Although industry has long recognized the need to develop adaptable programs, methodologies to develop them are still at the research stage. We have presented preliminary results in this area with a detailed study of the applicability of program specialization to various software architectures  [31] . Our latest contributions in this area span from a revolutionary approach based on the definition of programming languages, dedicated to a specific problem family, to a direct exploitation of specialization opportunities generated by a conventional programming methodology.

Domain-Specific languages

DSLs represent a promising approach to modeling a problem family. Yet, this approach currently suffers from the lack of methodology to design and implement DSLs. To address this basic need, we have introduced the Sprint methodology for DSL development  [23] . This methodology bridges the gap between semantics-based approaches to developing general-purpose languages and software engineering. Sprint is a complete software development process starting from the identification of the need for a DSL to its efficient implementation. It uses the denotational framework to formalize the basic components of a DSL. The semantic definition is structured so as to stage design decisions and to smoothly integrate implementation concerns.

Declaring adaptation

A less drastic strategy to developing efficient adaptable programs consists of making specific issues of adaptation explicit via a declarative approach. To do so, we enrich Java classes with declarations, named adaptation classes, aimed to express adaptive behaviors  [20] . As such, this approach allows the programmer to separate the concerns between the basic features of the application and its adaptation aspects. A dedicated compiler automatically generates Java code that implements the adaptive features.

Declaring specialization

When developing components, programmers often hesitate to make them highly generic and configurable. Indeed, genericity and configurability systematically introduce overheads in the resulting component. However, the causes of these overheads are usually well-known by the programmers and their removal could often be automated, if only they could be declared to guide an optimizing tool. The Compose group has worked towards solving this problem.

We introduced a declaration language which enables a component developer to express the configurability of a component. The declarations consist of a collection of specialization scenarios that precisely identify what program constructs are of interest for specialization. The scenarios of a component do not clutter the component code; they are defined aside in a specialization module  [26] , [27] , [25] , [28] .

This work was done in the context of C and declarations were intended to drive our C specializer.

Specializing design patterns

A natural approach to systematically applying program specialization is to exploit opportunities offered by a programming methodology. We have studied a development methodology for object-oriented languages, called design patterns. Design patterns encapsulate knowledge about the design and implementation of highly adaptable software. However, adaptability is obtained at the expense of overheads introduced in the finished program. These overheads can be identified for each design pattern. Our work consisted in using knowledge derived from design patterns to eliminate these overheads in a systematic way. To do so, we analyzed the specialization opportunities provided by specific uses of design patterns, and determined how to eliminate these overheads using program specialization. These opportunities were documented in declarations, called specialization patterns, and were associated with specific design patterns  [39] . The specialization of a program composed of design patterns was then driven by the corresponding declarations. This work was presented in the context of Java and uses our Java specializer  [38] .

Specializing software architectures

The sources of inefficiency in software architectures can be identified in the data and control integration of components, because flexibility is present not only at the design level but also in the implementation. We proposed the use of program specialization in software engineering as a systematic way to improve performance and, in some cases, to reduce program size. We studied several representative, flexible mechanisms found in software architectures: selective broadcast, pattern matching, interpreters, layers and generic libraries. We showed how program specialization can be applied systematically to optimize these mechanisms  [30] , [31] .