Less Is More : Generic Programming Theory and Practice

Abstract

Abstraction is ubiquitous in computer programming. The work of this thesis focuses on one specific form of abstraction. Computer programs manipulate data, which can either be primitive machine data (such as integer or fractional numbers) or programmer-defined data (such as lists, trees, matrices, images, etc.). There is only a small number of primitive datatypes, but a potentially infinite number of programmer-defined data. The structure of the latter data depends on the problem at hand, and while some structures appear very often (such as sequences of values), others are truly specific to a particular problem. Some kind of functionality is generally desired for all types of data. Reading and storing files to the disk, for instance, is as important for machine integers as it is for complex healthcare databases, or genealogy trees. And not just reading and writing files: testing for equality, sorting, traversing, computing the length, all are examples of functionality that is often desired for all kinds of data. Most programming languages allow defining complex datatypes as a form of abstraction, but few provide good support for defining behaviour that is generic over data. As such, programmers are forced to specify this behaviour over and over again, once for each new type of data, and also to adapt this code whenever the structure of their data changes. This is a tedious task, and can quickly become time-consuming, leading some programmers to write programs to generate this type of functionality automatically from the structure of data. We think that a programming language should allow programmers to define generic programs, which specify behaviour that is generic over the type of data. Moreover, it should automatically provide generic behaviour for new data, eliminating the need for repeated writing and rewriting of trivial code that just specialises general behaviour to a particular type of data. It should do so in a convenient way for the programmer, leading to more abstract and concise programs, while remaining clear and efficient. This leads us to the two research questions we set out to answer: 1) There are many different approaches to generic programming, varying in complexity and expressiveness. How can we better understand each of the approaches, and the way they relate to each other? 2) Poor runtime efficiency, insufficient datatype support, and lack of proper language integration are often pointed out as deficiencies in generic programming implementations. How can we best address these concerns? We answer the first question in the first part of this thesis. We start by picking a number of generic programming approaches and define a concise model for each of them. We then use this model to formally express how to embed the structural representation of data of one approach into another, allowing us to better understand the relation between different approaches. The second part of this thesis deals with answering the second question, devoting one chapter to analysing and mitigating each of the practical concerns