Dynamic Linking of Software Components

I n recent years operating systems have again begun to link software libraries to client programs dynamically at execution time. The concepts underlying dynamic linking date back at least to the Multics system 1 of the mid-1960s and have now been reintroduced in some of the most popular workstation and PC operating systems. Operating systems based on modular languages such as Mesa, 2 Modula-2, and Oberon 3 have offered similar capabilities for well over a decade. Indeed, the concept of dynamic linking reached maturity through these modular operating systems, which largely laid the groundwork for dynamic linking's resurgence. Modular systems are written in languages that blur the distinction between libraries and application programs. In these languages, all software consists of modules arranged in a (usually acyclic) dependence hierarchy. Individual modules serve as libraries when they are referenced by clients higher up in the hierarchy and as clients when they reference modules farther down. Consequently, as Figure 1 shows, the hierarchy's inner nodes serve simultaneously as libraries and as clients. In such a system, the operating system is itself represented as a collection of modules at the bottom of the hierarchy. Operating system and application program modules have no intrinsic difference, and the boundary between them is purely conceptual. Because only those parts of the operating system that the application requires at the moment need be present in memory, dynamic linking lets us build powerful systems that consume relatively little storage. Already powerful by itself, dynamic linking is even more useful when it is combined with a language that directly supports software system exten-sibility. An extensible system can evolve without requiring changes in any of its original parts. Combining extensibility with dynamic linking leads to highly flexible systems that use computing resources efficiently because they don't duplicate functionality. Niklaus Wirth recently presented a compelling case for such runtime-extensible software systems. 4 While the central idea behind dynamic linking is quite straightforward, it can be implemented through a surprising variety of strategies. In this article I contrast three simple dynamic linking schemes with two much more elaborate strategies. The latter represent a workload shift from the compiler to the dynamic linker, just as dynamic linking itself represents a shift that moves the functions of a separate linker into the loader. Because the new techniques I describe promise the profound additional benefit of cross-platform portability, they will most likely displace the currently popular …