I have investigated two examples of resource-aware languages. My doctoral work was on the theory and applications of multi-stage languages (such as MetaML and MetaOCaml) and I have continued to work in this area ever since. More recently, I have become also interested in resource-bounded languages (such as RT-FRP and E-FRP). The following describes my work in each of these areas in more detail.

Software is rapidly becoming a critical component of almost every appliance and device around us, including cars, microwaves, defense systems, and medical equipment. As a building material, software's remarkable advantage is that it can be altered and duplicated at virtually no cost. But it is also remarkably fragile: An unintended setting for one bit - one indivisible unit of information - can alter the operation of the largest machine-controlled physical system. While research on high-level languages has produced a wide range of innovative abstraction mechanisms that make programs easier to understand, such mechanisms either have an associated runtime cost, require the programmer to relinquish control over resources, or both. Embedded software components running inside physical devices must run for long periods of time with limited resources. These strict constraints have resulted in embedded software development depending on very low level tools and notation that make programs difficult to understand, check, and manage. A wide range of practical factors, including the constant influx of new devices and the ever increasing demand for interoperability make addressing this problem an urgent matter. The goal of my research is to show that a novel class of languages called Resource-aware Programming (RAP) languages can be an effective approach to addressing this problem. A RAP languages provides two key mechanisms: 1. Allowing the programmer to write programs where part of the computation is performed on the development platform, and the rest is performed on the embedded platform. This facility can allow the programmer to use almost any abstraction mechanism without paying a runtime penalty on the embedded platform. While McIlroy's work on macros in 1960 marked the start of efforts to support this otherwise intuitive functionality, realizing this support has proven surprisingly challenging. A promising technique for achieving this goal is multi-stage programming (MSP), proposed in my doctoral dissertation and extensively developed since. Most recently I proposed the notion of environment classifiers, and showed that it subsumes various previously incompatible approaches to typing MSP languages. I also used MSP to resolve two long-standing open research questions. The first is how to achieve "optimal specialization" in typed languages. This question was posed by the partial evaluation pioneer Neil Jones and had remained open for over a decade despite several efforts before ours. The second is how to define an expressive, type-safe, macro system and to define its behavior rigorously and concisely. Along side these efforts on foundations, I have led the development and implementation of an MSP language called MetaOCaml. This implementation has proven invaluable for experimental work by our research group and by several other groups around the world.

2. Providing static analyses that ensure - ahead of time - that any computation remaining for the embedded platform will be resource-bounded. Most of our early work in this area focused on ensuring safety in the embedded platform. Our first result on resource-boundedness showed that RAP languages can statically guarantee that generated (embedded platform) programs are heap-bounded. As an in-depth case study, we showed that a RAP language can express the Fast Fourier Transform (FFT) with source code only marginally larger than the text-book definition of the Cooley-Tuckey recurrence. This RAP program generates a circuit with an optimal number of multiplications and additions. Further work led to a new insight about FFT itself. Our generator, which uses only a small number optimizations, can generate circuits with arithmatic operation counts that precisely match those of two well-established benchmarks (Split-Radix and FFTW). Furthermore, which benchmark is matched depends only on how multiplication of complex numbers is implemented. Viewed as software, the FFT implementations we generate run within a small factor of the FFTW system.

Current activity is focused on applying RAP techniques to the domain of device driver development. In addition to dealing with the basic resources of time and space, this domain requires addressing event-driven computation as well as concurrency. We also continue to explore hardware circuit design as the next challenging domain for our case studies. We expect that incorporating layout and routing information into a RAP language is possible, and that such a language would provide hardware designers with a tool that addresses some of the most pressing challenges in that domain.

Theses