Techniques in Active and Generic Software Libraries

Reusing code from software libraries can reduce the time and effort to construct software systems and also enable the development of larger systems. However, the benefits that come from the use of software libraries may not be realized due to limitations in the way that traditional software libraries are constructed. Libraries come equipped with application programming interfaces (API) that help enforce the correct use of the abstractions in those libraries. Writing new components and adapting existing ones to conform to library APIs may require substantial amounts of "glue" code that potentially affects software's efficiency, robustness, and ease-of-maintenance. If, as a result, the idea of reusing functionality from a software library is rejected, no benefits of reuse will be realized. This dissertation explores and develops techniques that support the construction of software libraries with abstraction layers that do not impede efficiency. In many situations, glue code can be expected to have very low (or zero) performance overhead. In particular, we describe advances in the design and development of active libraries

• software libraries that take an active role in the compilation of the user's code. Common to the presented techniques is that they may "break" a library API (in a controlled manner) to adapt the functionality of the library for a particular use case. The concrete contributions of this dissertation are: a library API that supports iterator selection in the Standard Template Library, allowing generic algorithms to find the most suitable traversal through a container, allowing (in one case) a 30-fold improvement in performance; the development of techniques, idioms, and best practices for concepts and concept maps in C++, allowing the construction of algorithms for one domain entirely in terms of formalisms from a second domain; the construction of generic algorithms for algorithmic differentiation, implemented as an active library in Spad, language of the Open Axiom computer algebra system, allowing algorithmic differentiation to be applied to the appropriate mathematical object and not just concrete data-types; and the description of a static analysis framework to describe the generic programming notion of local specialization within Spad, allowing more sophisticated (value-based) control over algorithm selection and specialization in categories and domains. We will find that active libraries simultaneously increase the expressivity of the underlying language and the performance of software using those libraries.