Micro Virtual Machines: A Solid Foundation for Managed Language Implementation

Today new programming languages proliferate, but many of them suffer from poor performance and inscrutable semantics. We assert that the root of many of the performance and semantic problems of today's languages is that language implementation is extremely difficult. This thesis addresses the fundamental challenges of efficiently developing high-level managed languages. Modern high-level languages provide abstractions over execution, memory management and concurrency. It requires enormous intellectual capability and engineering effort to properly manage these concerns. Lacking such resources, developers usually choose naive implementation approaches in the early stages of language design, a strategy which too often has long-term consequences, hindering the future development of the language. Existing language development platforms have failed to provide the right level of abstraction, and forced implementers to reinvent low-level mechanisms in order to obtain performance. My thesis is that the introduction of micro virtual machines will allow the development of higher-quality, high-performance managed languages. The first contribution of this thesis is the design of Mu, with the specification of Mu as the main outcome. Mu is the first micro virtual machine, a robust, performant, and light-weight abstraction over just three concerns: execution, concurrency and garbage collection. Such a foundation attacks three of the most fundamental and challenging issues that face existing language designs and implementations, leaving the language implementers free to focus on the higher levels of their language design. The second contribution is an in-depth analysis of on-stack replacement and its efficient implementation. This low-level mechanism underpins run-time feedback-directed optimisation, which is key to the efficient implementation of dynamic languages. The third contribution is demonstrating the viability of Mu through RPython, a real-world non-trivial language implementation. We also did some preliminary research of GHC as a Mu client. We have created the Mu specification and its reference implementation, both of which are open-source. We show that that Mu's on-stack replacement API can gracefully support dynamic languages such as JavaScript, and it is implementable on concrete hardware. Our RPython client has been able to translate and execute non-trivial RPython programs, and can run the RPySOM interpreter and the core of the PyPy interpreter. With micro virtual machines providing a low-level substrate, language developers now have the option to build their next language on a micro virtual machine. We believe that the quality of programming languages will be improved as a result