How to speed up the Rust compiler in September 2026

In September 2026, the Rust compiler saw significant improvements in its performance. According to measurements from July 29, 2026, to September 28, 2026, the mean wall-time reduction was 4.57%, a remarkable improvement compared to the previous two-month period. Out of 629 benchmark measurements, 555 showed enhancements, while only 74 experienced regressions. A large number of benchmarks achieved double-digit percentage reductions, earning the term "a sea of green."

Noah Lev made substantial speed gains for rustdoc, which he detailed in a subsequent post. Jakub Beránek enabled Profile Guided Optimization (PGO) for Clippy, resulting in notable improvements across most Clippy benchmarks, with a best-case scenario of an 18% wall-time reduction. Nikita Popov upgraded the LLVM version to 23, leading to a mean wall-time reduction of 1.2% across all benchmarks, which is impressive for a single PR.

The new borrow checker, Polonius Alpha, was enabled on Nightly. It proved more precise than the existing checker and accepted some valid programs rejected by the old checker. Jack Huey addressed the increased work required by Polonius Alpha, resulting in liveness computations being made lazy, which reduced instruction counts for the serde crate by 3-5%, and for other benchmarks by less than 1%.

The new trait solver, Penelope Hammertime, was also enabled on Nightly. Like Polonius Alpha, it underperformed in a few cases. However, a number of PRs made significant contributions to its performance, including reducing compile times for certain outlier crates by up to 50% and making improvements using static dispatch instead of dynamic dispatch.

Xmakro continued their streak of good improvements, optimizing impl handling, reducing instruction counts in the loading of incremental compilation data, and avoiding allocations in a hot obligations processing path. Chris Denton increased the default stack size used by the compiler, allowing the removal of ensure_sufficient_stack, which in turn led to a reduced instruction count across many benchmarks.