Practical LR Parser Generation

TL;DR

This paper introduces a novel approach to automatically generate efficient LR parsers for a broad class of programming languages, overcoming traditional limitations and demonstrating practical performance improvements over hand-written parsers.

Contribution

It presents new algorithms and extensions, including automata optimization, grammar transformation, and the XLR extension, enabling automatic parser generation for a wide range of languages.

Findings

Generated parsers are 1.2x faster than hand-written Golang parser.

Generated parsers are 4.3x faster than CPython parser.

The approach supports a broad class of practical grammars.

Abstract

Parsing is a fundamental building block in modern compilers, and for industrial programming languages, it is a surprisingly involved task. There are known approaches to generate parsers automatically, but the prevailing consensus is that automatic parser generation is not practical for real programming languages: LR/LALR parsers are considered to be far too restrictive in the grammars they support, and LR parsers are often considered too inefficient in practice. As a result, virtually all modern languages use recursive-descent parsers written by hand, a lengthy and error-prone process that dramatically increases the barrier to new programming language development. In this work we demonstrate that, contrary to the prevailing consensus, we can have the best of both worlds: for a very general, practical class of grammars -- a strict superset of Knuth's canonical LR -- we can generate parsers automatically, and the resulting parser code, as well as the generation procedure itself, is highly efficient. This advance relies on several new ideas, including novel automata optimization procedures; a new grammar transformation ("CPS"); per-symbol attributes; recursive-descent actions; and an extension of canonical LR parsing, which we refer to as XLR, which endows shift/reduce parsers with the power of bounded nondeterministic choice. With these ingredients, we can automatically generate efficient parsers for virtually all programming languages that are intuitively easy to parse -- a claim we support experimentally, by implementing the new algorithms in a new software tool called langcc, and running them on syntax specifications for Golang 1.17.8 and Python 3.9.12. The tool handles both languages automatically, and the generated code, when run on standard codebases, is 1.2x faster than the corresponding hand-written parser for Golang, and 4.3x faster than the CPython parser, respectively.