Generating Juno-2 Figures as Java Bytecodes

James Leifer, Cambridge University

I am two years into my Ph.D. at the Cambridge University Computing Lab (in the UK) where I work with Robin Milner on process algebras (pi-calculus and action calculi). My work is funded by the U.S. National Science Foundation and the British Goverment. I did an undergraduate degree at Oxford ("the Other Place" as they say at Cambridge) where I worked under Bernard Sufrin and C. A. R Hoare. Despite all that, I'm a New Yorker (or at least can pretend to be one).

I decided to apply for a SRC internship because of SRC's reputation for combining systems and theory work, which suited my interests exactly. From Cambridge, I discussed possible projects with Greg Nelson and we agreed to do work involving some "honest programming" - - -- something that I had not done very much of while pursuing my Ph.D.

The goal of the "honest programming'' was to extend Juno ---a dual-view graphics editor built by Greg Nelson and Allan Heydon at SRC--- to allow it to emit Java applets that render Juno graphics on the Web. Juno has a programming language based on Dijkstra's guarded commands (extended with geometrical constraint solving) and its approach to graphics and animation is to create a program that renders the graphics as a side effect of execution. Internally, Juno compiles programs to a bytecode and runs them on a VM.

This summer I built a compiler that translates a Juno program into a Java class file, and I constructed a library of Java classes that provides run-time support. The compiler translates modules to classes, global variables and constants to fields, and procedures to methods. The compiler handles almost all Juno constructions, including modules and imports, global variables and constants, procedures, functions, predicates, control structures, and expressions. The run-time system provides mathematics and PostScript drawing support. Both the compiler and run-time system do not yet handle closures and constraint solving, nor do the graphics libraries do double-buffering, which is necessary for flicker-free animation, but static graphics are working well.

There were several differences between Juno and Java which influenced the design of the translator.

  1. Juno programs are dynamically typed ---values range over numbers, strings, pairs, etc., and this was mirrored in Java by creating a class JV (for "Juno value") and subclasses of JV corresponding to Juno's value types.

  2. Juno's procedures can return multiple out-parameters but Java's methods cannot. In our implementation a method returns a reference to an array containing the out-parameters.

  3. In Juno certain expressions are not strict, that is, they are defined even when subexpressions are not.

This doesn't fit well with Java's notion of exceptions, because when a Java exception is thrown, the entire stack is deleted, not just the items placed on the stack in the surrounding try-catch block. We got around this by creating a global bit that is set as certain parts of an expression are evaluated and tested so as to obtain the correct semantics.

This project was the largest engineering artifact I have ever constructed and I learned much from working with Greg and Allan on how to structure a complicated system and how to test it. I also had many lively talks with others in the lab who gave me good advice and discussed with me their research interests. And, of course, I thoroughly enjoyed the California weather and food!


[Validate me.]