The Semantics of Rank Polymorphism

TL;DR

This paper introduces Remora, a core language for rank-polymorphic array operations, with formal semantics and a sound static type system that prevents shape errors and supports dynamic array shapes.

Contribution

It develops a formal, dynamic and static semantics for rank polymorphism, including a dependent type system that ensures shape safety and supports dynamic array shapes.

Findings

Proves soundness of the type system with progress and preservation

Shows static shape safety prevents runtime errors

Demonstrates alignment between fully-typed and partially erased semantics

Abstract

Iverson's APL and its descendants (such as J, K and FISh) are examples of the family of "rank-polymorphic" programming languages. The principal control mechanism of such languages is the general lifting of functions that operate on arrays of rank (or dimension) $r$ to operate on arrays of any higher rank $r' > r$. We present a core, functional language, Remora, that captures this mechanism, and develop both a formal, dynamic semantics for the language, and an accompanying static, rank-polymorphic type system for the language. Critically, the static semantics captures the shape-based lifting mechanism of the language. We establish the usual progress and preservation properties for the type system, showing that it is sound, which means that "array shape" errors cannot occur at run time in a well-typed program. Our type system uses dependent types, including an existential type abstraction which permits programs to operate on arrays whose shape or rank is computed dynamically; however, it is restricted enough to permit static type checking. The rank-polymorphic computational paradigm is unusual in that the types of arguments affect the dynamic execution of the program -- they are what drive the rank-polymorphic distribution of a function across arrays of higher rank. To highlight this property, we additionally present a dynamic semantics for a partially erased variant of the fully-typed language and show that a computation performed with a fully-typed term stays in lock step with the computation performed with its partially erased term. The residual types thus precisely characterise the type information that is needed by the dynamic semantics, a property useful for the (eventual) construction of efficient compilers for rank-polymorphic languages.