Projection-Based Program Analysis

Projection-based program analysis techniques are remarkable for their ability to give highly detailed and useful information not obtainable by other methods. The first proposed projection-based analysis techniques were those of Wadler and Hughes for strictness analysis, and Launchbury for binding-time analysis; both techniques are restricted to analysis of first-order monomorphic languages. Hughes and Launchbury generalised the strictness analysis technique, and Launchbury the binding-time analysis technique, to handle polymorphic languages, again restricted to first order. Other than a general approach to higher-order analysis suggested by Hughes, and an ad hoc implementation of higher-order binding-time analysis by Mogensen, neither of which had any formal notion of correctness, there has been no successful generalisation to higher-order analysis. We present a complete redevelopment of monomorphic projection-based program analysis from first principles, starting by considering the analysis of functions (rather than programs) to establish bounds on the intrinsic power of projection-based analysis, showing also that projection-based analysis can capture interesting termination properties. The development of program analysis proceeds in two distinct steps: first for first-order, then higher order. Throughout we maintain a rigorous notion of correctness and prove that our techniques satisfy their correctness conditions. Our higher-order strictness analysis technique is able to capture various so-called data-structure-strictness properties such as head strictness-the fact that a function may be safely assumed to evaluate the head of every cons cell in a list for which it evaluates the cons cell. Our technique, and Hunt's PER-based technique (originally proposed at about the same time as ours), are the first techniques of any kind to capture such properties at higher order. Both the first-order and higher-order techniques are the first projection-based techniques to capture joint strictness properties-for example, the fact that a function may be safely assumed to evaluate at least one of several arguments. The first-order binding-time analysis technique is essentially the same as Launchbury's; the higher-order technique is the first such formally-based higher-order generalisation. Ours are the first projection-based termination analysis techniques, and are the first techniques of any kind that are able to detect termination properties such as head termination-the fact that termination of a cons cell implies termination of the head. A notable feature of the development is the method by which the first-order analysis semantics are generalised to higher-order: except for the fixed-point constant the higher-order semantics are all instances of a higher-order semantics parameterised by the constants defining the various first-order semantics