Benchmarking a Rust diff/3-way-merge engine against jj-lib, libgit2, imara-diff, diffy (in-process, 4 real repos)

[kroo1]

I've been building a line-diff/3-way-merge engine in Rust and wanted to share an in-process benchmark against a few well-known implementations: jj-lib (ContentDiff, files::merge), libgit2/xdiff (git2::Patch, git2::merge_file), imara-diff (Histogram algorithm), and diffy.

Methodology: all engines run in-process (no subprocess spawning) on the same samples, pulled from real commit history across 4 repos (helix, vscode, rustc, linux) — ~5300 diff samples, ~5200 merge samples. Diff correctness is checked by reconstructing the target content byte-for-byte from each engine's own output and comparing to the real content.

Diff (2-way), total across all 4 repos:

Engine	Time (ms)	vs. mine
jj-lib	33671	0.11x
libgit2 (xdiff)	20513	0.18x
imara-diff (Histogram)	4363	0.85x
mine	3720	1.00x

All 4 engines reconstructed the target content byte-for-byte correctly on every sample.

3-way merge, total across all 4 repos (real base content + synthetic scattered non-overlapping edits on each side):

Engine	Time (ms)	vs. mine
jj-lib	62787	0.08x
libgit2 (merge_file)	14399	0.33x
diffy	10083	0.47x
mine	4757	1.00x

Happy to share more details on the approach or the benchmark setup if useful.