Performance
Measured on a 6.5 MB corpus (1,196 files), as CPU time:
| Roslyn parse + full red tree — the floor for any tool | ~300 ms |
| Curb | ~350 ms |
dotnet format whitespace |
~12,000 ms |
Parsing is about 2.5% of the budget. Roughly 97% of what a formatter costs is its own work. Curb adds almost nothing on top of the parse: it is within measurement noise of the floor.
The twelve-repository comparison has wall-clock numbers across real repositories; on roslyn (17,167 files) Curb takes 7 s against dotnet format whitespace's 34 s.
The document IR lives in a pooled struct arena — structs, not objects, reusing memory across files. Curb loads no workspace and resolves no symbols. The conditional round-trip reparse (see Safety) means most files pay for one parse, not two.
It ships as a native-AOT binary per platform, so there is no JIT warm-up: about 10 ms to start.
The build integration runs Curb before CoreCompile. If formatting costs more than parse time, the build slows down. Because Curb adds almost nothing beyond the parse, a build with Curb is a build without it — measured rather than assumed.
On unchanged projects the MSBuild stamp means no process starts at all. The performance number is for the case where work actually happens.