Fast and precise hybrid type inference for JavaScript

  • Authors:
  • Brian Hackett;Shu-yu Guo

  • Affiliations:
  • Mozilla, Mountain View, CA, USA;Mozilla & University of California, Los Angeles, Mountain View, CA, USA

  • Venue:
  • Proceedings of the 33rd ACM SIGPLAN conference on Programming Language Design and Implementation
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

JavaScript performance is often bound by its dynamically typed nature. Compilers do not have access to static type information, making generation of efficient, type-specialized machine code difficult. We seek to solve this problem by inferring types. In this paper we present a hybrid type inference algorithm for JavaScript based on points-to analysis. Our algorithm is fast, in that it pays for itself in the optimizations it enables. Our algorithm is also precise, generating information that closely reflects the program's actual behavior even when analyzing polymorphic code, by augmenting static analysis with run-time type barriers. We showcase an implementation for Mozilla Firefox's JavaScript engine, demonstrating both performance gains and viability. Through integration with the just-in-time (JIT) compiler in Firefox, we have improved performance on major benchmarks and JavaScript-heavy websites by up to 50%. Inference-enabled compilation is the default compilation mode as of Firefox 9.