Generic Go to Go: Dictionary-Passing, Monomorphisation, and Hybrid

TL;DR

This paper introduces a new call-site based dictionary-passing translation for Go generics, aiming to reduce code bloat and improve compilation speed compared to existing approaches like monomorphisation and hybrid.

Contribution

It formalizes a novel non-specialising call-site based translation for Go generics and proves its correctness using bisimulation techniques.

Findings

The new translation reduces code size and compilation time.

Benchmark results compare five different generics translation approaches.

Insights suggest improvements for Go 1.18's generics implementation.

Abstract

Go is a popular statically-typed industrial programming language. To aid the type safe reuse of code, the recent Go release (Go 1.18) published on 15th March 2022 includes bounded parametric polymorphism via generic types. Go 1.18 implements generic types using combination of monomorphisation and call-graph based dictionary-passing called hybrid. This hybrid approach can be viewed as an optimised form of monomorphisation that statically generates specialised methods and types based on possible instantiations. A monolithic dictionary supplements information lost during monomorphisation, and it is structured according to the program's call graph. Unfortunately, the hybrid approach still suffers from code bloat, poor compilation speed, and limited code coverage. In this paper we propose and formalise a new non-specialising call-site based dictionary-passing translation. Our call-site based translation creates individual dictionaries for each type parameter, with dictionary construction occurring in place of instantiation, overcoming the limitations of hybrid. We prove it correct using a novel and general bisimulation up to technique. To better understand how different generics translations approaches work in practice, we benchmark five translators, Go 1.18, two existing monomorphisation translators, our dictionary-passing translator, and erasure translator. Our findings reveal several suggestions for improvements for Go 1.18 -- specifically how to overcome the expressiveness limitations of generic Go, and improve compile time and compiled code size performance of Go 1.18.