ron-bench

Benchmark and conformance harnesses for the ron parser, kept out of ron's own tree because they pull in serde_json, pest, and two ron versions at once — dependencies that don't belong in the library's graph.

Three independent crates, one workspace:

Every crate parses with the same ron, set once in the root Cargo.toml:

[workspace.dependencies]
ron = { git = "https://github.com/enomado/ron", branch = "perf/scalar-parser-hotpath" }

To A/B two parsers, point that one line at each and re-run:

• baseline — git = "https://github.com/ron-rs/ron", branch = "master"
• improved — git = "https://github.com/enomado/ron", branch = "perf/scalar-parser-hotpath" (upstream master + the scalar hot-path optimizations, ron-rs/ron#616)

Or [patch] it at a local checkout.

Parses payloads that are pure floats / ints / strings / mixed, so the worst RON/JSON ratio points straight at the culprit token type.

cargo run --release -p ronprof -- diff

The deterministic A/B signal is instruction count, not wall-clock MB/s (a busy machine makes MB/s wander between runs). Pin a core and count instructions per mode:

taskset -c 2 perf stat -e instructions ./target/release/ronprof intloop 400
taskset -c 2 perf stat -e instructions ./target/release/ronprof ronloop 200   # mixed

Measured this way, perf/scalar-parser-hotpath vs upstream master:

ron.pest is a PEG transcribed from ron's grammar. The differential parses a battery of inputs (and, optionally, every *.ron under given dirs) with both the pest reference and ron's parser and reports where they disagree — each disagreement is a parser bug or a grammar/doc bug.

cargo run --release -p ron-conformance --bin ron_conformance          # hand battery
cargo run --release -p ron-conformance --bin ron_conformance -- path/ # + sweep *.ron

ron_perf_fuzz generates valid RON from the grammar's repeatable units and adversarial scan-in-loop shapes, then runs the scaling oracle to hunt for quadratic parse behaviour on forms a human wouldn't hand-write. It links ron (fixed) and stock ron 0.12.2 (quadratic bugs live) side by side, so the oracle is non-circular.

cargo run --release -p ron-conformance --bin ron_perf_fuzz

A #[track_caller] assert_linear that doubles the input across sizes and asserts the per-doubling time ratio stays near 2, not 4. Strong where a naive 2-point ratio flakes: min-of-N per size, sustained verdict across every doubling, setup kept out of the timer, and a loud guard when the smallest measurement is too small to trust. The same method found two O(N²) bugs in ron (#607, #609).

cargo test --release -p scaling_oracle

Some inline comments in these crates are still in Russian (the crate-level docs and this README are English). Translation is a pending cleanup, not a blocker to reading the code.