The LSP currently checks full documents synchronously and shares the resulting symbol index between open files in a checked package. Measure the edit-to-diagnostics delay and request round trips before introducing incremental analysis.
Reproduce the measurements
Build the release compiler, then run from the repository root:
cargo build --locked --release
python3 scripts/bench_lsp.py target/release/dodo --features
python3 scripts/bench_lsp.py target/release/dodo --features --functions 200 --documents 5
python3 scripts/bench_lsp.py target/release/dodo --features --functions 20 --stdlibThe script starts a real dodo lsp process, opens unsaved buffers, warms up with
five changes, and measures 30 further changes by default. Each timed change ends
with a request barrier after all diagnostics have been published. It includes
serialization, pipe transport, checking all open buffers, and diagnostic delivery.
The script rejects unexpected diagnostics. Request timings measure completion,
definition, signature help, and hover against the refreshed analysis. It prints
JSON lines with p50, p95, and maximum milliseconds. Use --samples for longer runs.
Run the same script without --features against an older compiler that supports
only diagnostics and hover. Compare the same build profile, hardware, fixtures,
and open-document counts; these are observations, not CI latency thresholds.
Measurements
Measured on 2026-09-14 on Linux x86-64, an Intel Core Ultra 5 125U, Rust 1.95.0, LLVM
22.1.8, and Cargo’s release profile. These runs used --samples 100, with builds and tests finished before measuring.
The baseline is commit
b3aad8cd29725942611ab09a67472eaec66e3c32, before the additional editor features.
The fixture contains small functions with a typed parameter and an arithmetic
return, plus a caller. The std/math case also loads and checks that bundled
dependency. The five-buffer case uses file checking mode.
| Fixture | Baseline change p50 / p95 | Expanded LSP change p50 / p95 | Completion p95 |
|---|---|---|---|
| 20 functions, 933 bytes | 0.23 / 0.36 ms | 0.48 / 0.59 ms | 0.49 ms |
| 200 functions, 9,233 bytes | 2.09 / 2.20 ms | 3.83 / 4.15 ms | 1.60 ms |
| 1,000 functions, 46,833 bytes | 11.59 / 11.86 ms | 23.69 / 28.07 ms | 8.76 ms |
| Five open files, 200 functions each | 6.17 / 9.78 ms | 13.41 / 17.41 ms | 1.73 ms |
20 functions plus std/math |
3.07 / 3.20 ms | 4.57 / 5.81 ms | 0.59 ms |
Hover, definition, and signature-help p95 round trips were below 0.5 ms in these expanded-server runs. Symbol indexing and diagnostic recovery add work, but these fixtures do not establish a need for incremental analysis. Full-document checking is retained. Large individual bodies, error-heavy files, larger import graphs, and more open buffers need separate measurements before drawing conclusions about those workloads. The benchmark can be rerun as the implementation evolves.