Efficient Gradual Typing

TL;DR

This paper introduces Grift, a compiler for gradual typing that uses ahead-of-time compilation and efficient coercions, achieving performance comparable to statically typed languages and eliminating slowdowns in partially typed code.

Contribution

The paper presents a novel ahead-of-time compilation approach for gradual typing with space-efficient coercions, improving performance and eliminating slowdowns.

Findings

Achieves OCaml-like performance on statically typed programs.

Reduces slowdown in partially typed code to between 0.42x and 2.36x of untyped Racket.

Eliminates overhead in statically typed code using monotonic references.

Abstract

Gradual typing combines static and dynamic typing in the same program. One would hope that the performance in a gradually typed language would range between that of a dynamically typed language and a statically typed language. Existing implementations of gradually typed languages have not achieved this goal due to overheads associated with runtime casts. Takikawa et al. (2016) report up to 100$\times$ slowdowns for partially typed programs. In this paper we present a compiler, named Grift, for evaluating implementation techniques for gradual typing. We take a straightforward but surprisingly unexplored implementation approach for gradual typing, that is, ahead-of-time compilation to native assembly code with carefully chosen runtime representations and space-efficient coercions. Our experiments show that this approach achieves performance on par with OCaml on statically typed programs and performance between that of Gambit and Racket on untyped programs. On partially typed code, the geometric mean ranges from 0.42$\times$ to 2.36$\times$ that of (untyped) Racket across the benchmarks. We implement casts using the coercions of Siek, Thiemann, and Wadler (2015). This technique eliminates all catastrophic slowdowns without introducing significant overhead. Across the benchmarks, coercions range from 15% slower (fft) to almost 2$\times$ faster (matmult) than regular casts. We also implement the monotonic references of Siek et al. (2015). Monotonic references eliminate all overhead in statically typed code, and for partially typed code, they are faster than proxied references, sometimes up to 1.48$\times$.