State-transition machines, revisited

In the autumn of 1978, Neil Jones and Steve Muchnick, working at the University of Kansas, were studying compiler synthesis from Scott-Strachey denotational-semantics definitions; I was Neil’s student. Neil read intently John Reynolds’s 1972 paper, Definitional Interpreters for Higher-Order Programming Languages [14], and applied Reynolds’s continuation-passing and defunctionalization transformations to lambda-calculus-coded denotational-semantics definitions, using the transformed definitions as templates for syntax-directed translation. Neil dubbed the translated source programs, “State-Transition Machines” (STMs), because an object program was a set of equationally defined functions that looked like the transition rules of a finite-state machine. Our initial efforts were spent on transforming denotational definitions of block-structured, imperative languages into compiling schemes that generated STMs that looked like ordinary assembly code. In the summer of 1979, Steve moved to the University of California, Berkeley, and Neil and I left for the University of Aarhus, Denmark, where we continued the research project. A summary of the work was eventually published as the paper, Compiler generation from denotational semantics [10]. The continuation-passing and defunctionalization transforms were tedious, and I suggested to Neil that one could do better by writing a translator from lambda-calculus into STMs and then constructing a compiler by composing a denotational definition with the lambdacalculus translator. It was unclear whether this tactic would generate better target code than that generated by Neil’s smart transformations, but Neil agreed that it was worth a try. After several false starts, I formulated a translator from a call-byvalue lambda-calculus to STMs written in a variant of Landin’s SECDmachine, which later appeared in [16] as the “VEC-machine.” (Neil preferred a call-by-value lambda-calculus metalanguage.) At the same time, I was reading Chris Wadsworth’s paper, The relation between computational and denotational properties for Scott’s models of the lambda-calculus [18], and I was fascinated by Wadsworth’s use