Compiler performance optimizations

Summary

Make rustc faster through sustained, incremental, profile-driven optimization. This work will improve performance in the new trait solver and borrow checker (Polonius) and revisit optimization opportunities across the compiler with improved tooling.

Motivation

There are two basic ways to speed up the Rust compiler.

• Big improvements. Large projects like pipelined compilation and the parallel backend have given large speed-ups, e.g. 10-50% across a wide range of benchmarks. However, such projects can be difficult to complete. For example, the parallel front-end was begun in 2018 and is still in progress, having gone through multiple rounds of stasis and reanimation.

• Small improvements. This is steady, incremental, profile-driven work. Each improvement typically results in single-digit percentage improvements, often on smaller ranges of benchmarks, as described in the long-running “How to speed up the Rust compiler” blog series. It is less glamorous work but it adds up over time and it historically has been the primary driver of speed, particularly over the period 2016-2023 when compiler speed increased by roughly 3x. It has slowed down in the past couple of years, however, due to diminishing returns.

The big improvements are still worth pursuing, but the small improvements path has been re-energized by recent developments.

• The new trait solver and new borrow checker (Polonius) are close to shipping. Both are multi-year projects that make Rust more expressive and capable. Both currently cause significant compile-time regressions and require effort to get their performance acceptable. Small improvements will be the solution here, as was the case when the current borrow checker (NLL) was introduced in 2018.

• New tooling, in the form of LLMs, has greatly improved. The recent huge increases in security vulnerability discovery are well known across many projects. We have seen evidence that LLMs are similarly helpful for performance work; an expert human using an LLM can analyze profiles and find and implement optimizations better than an expert human can alone. Previously tapped-out seams of optimization will likely reopen all across the compiler. This LLM usage aligns with the Rust project’s values because the analysis provides all the value; LLM code generation is not required.

This work requires experience with a variety of profiling tools and optimization techniques.

Work items over the next year

Team asks

Funding

Funding will support Nicholas Nethercote’s profiling, benchmarking, and optimization work. Additionally, extra funding can support 1-2 additional engineers to form a small team. Contact Hexcat to fund this goal.

Frequently asked questions