This thesis concerns the problem of complexity in operational semantics definitions. The appeal of modern operational semantics is the simplicity of their metatheories, which can be regarded as theories of deduction about certain shapes of operational judgments. However, when applied to real programming languages they produce bulky definitions that are cumbersome to reason about. The theory of interacting deductions is a richer metatheory which simplifies operational judgments and admits new proof techniques.
An interacting deduction is a pair (F, I), where F is a forest of inference trees and I is a set of interaction links (a symmetric set of pairs of formula occurrences of F), which has been built from interacting inference rules (sequences of standard inference rules, or rule atoms). This setting allows one to decompose operational judgments. For instance, for a simple imperative language, one rule atom might concern a program transition, and another a store transition. Program judgments only interact with store judgments when necessary: so stores do not have to be propagated by every inference rule. A deduction in such a semantics would have two inference trees: one for programs and one for stores.
This introduces a natural notion of modularity in proofs about semantics. The proof fragmentation theorem shows that one need only consider the rule atoms relevant to the property being proved. To illustrate, I give the semantics for a simple process calculus, compare it with standard semantics and prove three simple properties: nondivergence, store correctness and an equivalence between the two semantics.
Typically evaluation semantics provide simpler definitions and proofs than transition semantics. However, it turns out that evaluation semantics cannot be easily expressed using interacting deductions: they require a notion of sequentiality. The sequential deductions contain this extra structure. I compare the utility of evaluation and transition semantics in the interacting case by proving a simple translation correctness example. This proof in turn depends on proof-theoretic concerns which can be abstracted using dangling interactions. This gives rise to the techniques of breaking and assembling interaction links. Again I get the proof fragmentation theorem, and also the proof assembly theorem, which allow respectively both the isolation and composition of modules in proofs about semantics. For illustration, I prove a simple type-checking result (in evaluation semantics) and another nondivergence result (in transition semantics).
I apply these results to a bigger language, CSP, to show how the results scale up. Introducing a special scoping side-condition permits a number of linguistic extensions including nested parallelism, mutual exclusion, dynamic process creation and recursive procedures. Then, as an experiment I apply the theory of interacting deductions to present and prove sound a compositional proof system for the partial correctness of CSP programs.
Finally, I show that a deduction corresponds to CCS-like process evaluation, justifying philosophically my use of the theory to give operational semantics. A simple corollary is the undecidability of interacting-deducibility. Practically, the result also indicates how one can build prototype interpreters for definitions.ECS-LFCS-96-351.
This report is available in the following formats: