Where the figures come from. Each engine repository runs its engines, its platform's competitors and V8 on one runner every night and publishes the run as data; this page gathers the latest run of each.

Foreign runs. .NET: its image ghcr.io/corvus-dotnet/corvus-regex-ecma262-dotnet could not be pulled; Rust: its image ghcr.io/corvus-dotnet/corvus-regex-ecma262-rs could not be pulled. A foreign run's figures come from its repository's own nightly run on another runner on another night, so a figure that crosses to one is a guide, not a measurement.

Compliance beside an engine's name is its result on the ECMA-262 test data, whose answers are V8's: a check mark when it gave no wrong answer (the counts are in its tooltip), the count of wrong answers when it gave some, "the oracle" for V8, and "not ECMA-262" for an engine that implements another dialect, whose figures are a reference for speed and not for answers.

V8 appears once, from the .NET run; the V8 figures of the Rust run are not shown.

.NET

From the published run of corvus-dotnet/Corvus.Regex.Ecma262, a foreign run at abedf55.

Each figure is one Count over the whole text: every match found, as the benchmark of dotnet/performance measures it. The Corvus rows run under the u flag, which .NET has no mode for; counts agree on 53 of the 57 rows, and the four apart are $ and . at CRLF, where ECMA-262 counts CR as a line terminator, and \w, which .NET takes as Unicode letters and digits.

Where this library is slower, and why. Four kinds of row read above 1.0, each the price of an ECMA-262 rule .NET does not apply: . and $ at CRLF, where ECMA-262's . stops at CR as well as LF, so .* finds 36,491 matches to .NET's 26,105 at parity per match; case-insensitive patterns under the u flag, whose search must find the long s and the Kelvin sign as forms of s and k, which .NET's search ignores; a class that admits a supplementary character (two units) before a literal at a fixed distance, so the distance is not fixed ([a-q][^u-z]{13}x); and non-ASCII literals over UTF-8, where every Cyrillic or Chinese character is two or three bytes with no rare byte among them, against UTF-16's one rare unit per character. The other rows above 1.0 are within a few percent of parity and read either side of it from run to run.

The engines are .NET's interpreter (RegexOptions.None), .NET's source generator ([GeneratedRegex]), this library's interpreter over UTF-16, this library's source generator over UTF-16 and over UTF-8, and V8 (Node 24's RegExp, the oracle of the test data) over the same texts with the same flags as the Corvus rows, its matches counted by a test loop that advances lastIndex as matchAll does. The UTF-8 rows match the same text as bytes, with nothing transcoded.

The bar is the mean of BenchmarkDotNet's default job (one launch, warm-ups and iterations as it settles them, fifteen or more) and the whisker one standard deviation of its iterations; a row's ratio is read against its neighbours' rather than to the last few percent. Allocated is the bytes the benchmark allocated per operation; the library allocates nothing per match once its working memory has grown.

The host is a GitHub-hosted ubuntu-latest runner, by the daily benchmarks workflow: every engine in one job on one runner, so the figures of a run are comparable with each other; nothing is pinned, so a run is not comparable with the pinned record to the last few percent. .NET 10. Commit abedf55, 2026-10-20.

Rust

From the published run of corvus-dotnet/corvus-regex-ecma262-rs, a foreign run at e8d2998.

Each figure is one count of every match a global search finds over the whole text, after an empty match going on one character later. This crate and regress run with the workload's flags and u, regex with the pattern as it is and the flags i, m and s, V8 with gu added.

Overall

Head to head

How to read this