Benchmarks

A measurement run over 41,000 files of real .NET source, timing three tools on identical copies of each repository.

The dnf whitespace column is dotnet format whitespace — the whitespace-only pass, not the full style-and-analyser run. curb is compared against its closest equivalent: the tool that does the same job.

curb is 5–25× faster than dotnet format whitespace across the corpus, on every repository, while doing more: dotnet format whitespace does not reflow long lines, does not sort using directives, and does not apply syntax-level code style. Beating it on raw speed while covering a strict superset of its output is the headline result.

dotnet format whitespace loads the full Roslyn workspace per project and is not designed to be fast.

All three tools above are timed cold — repository and formatter cache cleared before each run. Cold is the right baseline for comparing tools. Warm is where curb pulls further ahead.

curb's subsequent runs are protected by MSBuild incremental support: if nothing changed in a project, no process starts at all — not even the CLI. For projects that do have changes, a local cache means only the files that changed since the last build are reformatted, not the whole project.

The cache clears with dotnet clean and autoinvalidates after one week. curb is FAST cold, FASTER warm.

See Churn for what churn means, how much to expect on first adoption, and a worked example. The short version: roslyn (17,167 files, already exactly dotnet format-clean) is curb's sharpest case — 9,191 files rewritten, zero not-fixed-point. All of it is ground dotnet format doesn't cover: BOM handling, using-directive sorting, chain breaking, comment alignment.

Shallow-clone the twelve repositories into a single directory, then:

./build.sh compare --corpus /path/to/repos
		

The results table above is written back into this file automatically. The run takes roughly 40 minutes wall clock, dominated by dotnet format whitespace on roslyn and efcore.

Twelve repositories, cloned shallow, each formatted from a pristine copy. curb is the published native-AOT binary, not the JIT build. All three tools are timed cold — repository and formatter cache cleared before each run. After formatting, dotnet format whitespace was run over curb's output to measure agreement; curb was also run a second time over its own output to verify idempotency. Those per-repository counts are in Churn.

Two repositories have no .editorconfig at all (Newtonsoft.Json, ServiceStack), which is the onboarding case for a repository that has never configured anything.