Practical synthesis from real-world oracles

As software systems become increasingly heterogeneous, the ability of compilers to reason about an entire system has decreased. When components of a system are not implemented as traditional programs, but rather as specialised hardware, optimised architecture-specific libraries, or network services, the compiler is unable to cross these abstraction barriers and analyse the system as a whole. If these components could be modelled or understood as programs, then the compiler would be able to reason about their behaviour without concern for their internal implementation details: a homogeneous view of the entire system would be afforded. However, it is not often the case that such components ever corresponded to an original program. This means that to facilitate this homogenenous analysis, programmatic models of component behaviour must be learned or constructed automatically. Constructing these models is an inductive program synthesis problem, albeit a challenging one that is largely beyond the ability of existing implementations. In order for the problem to be made tractable, information provided by the underlying context (i.e. the real component behaviour to be matched) must be integrated. This thesis presents three program synthesis approaches that integrate contextual information to synthesise programmatic models for real, existing components. The first, Annote, exploits informally-encoded information about a component's interface (e.g. from documentation) by weaving that information into an extended type-and-attribute system for component interfaces. The second, Presyn, learns a pair of cooperating probabilistic models from prior syntheses, that aim to predict likely program structure based on a component's interface. Finally, Haze uses observations of common side-effects of component executions to bias the search for programs. These approaches are each evaluated against comparable synthesisers from the literature, on a set of benchmark problems derived from real components. Learning models for component behaviour is only a partial solution; the compiler must also have some mechanism to use those models for program analysis and transformation. This thesis additionally proposes a novel mechanism for context-sensitive automatic API migration based on synthesised programmatic models, and evaluates the effectiveness of doing so on real application code. In summary, this thesis proposes a new framing for program synthesis problems that target the behaviour of real components, and demonstrates three different potential approaches to synthesis in this spirit. The success of these approaches is evaluated against implementations from the literature, and their results used to drive a novel API migration technique