perf(linter): speed up hot rules on large codebases by anonrig · Pull Request #23829 · oxc-project/oxc

anonrig · 2026-06-26T16:18:13Z

Summary

Speeds up lint rules that were prominent in --debug timings on a large monorepo (x-web). Optimizations focus on cheaper gates, fewer AST dispatches (run_once over unresolved references instead of every identifier), and avoiding work when rules are not applicable.

Benchmark (release oxlint on ~/workspace/x-web, 5763 files, 260 rules, --silent --threads=1 --debug timings, three runs averaged; baseline = timings from before this PR):

Rule	Before (ms)	After avg (ms)	Calls (after)	Speedup
`oxc/bad-array-method-on-arguments`	6.5	0.21	5.8k (was ~384k)	~31×
`eslint/no-restricted-globals`	7.1	0.36	5.8k (was ~384k)	~20×
`oxc/no-this-in-exported-function`	15.5	2.1	13.7k	~7.5×
`vitest/no-commented-out-tests`	13.4	2.1	5.8k	~6.4×
`eslint/no-useless-call`	10.6	2.6	150k	~4.0×
`import/extensions`	13.2	3.7	195k	~3.5×
`jsx_a11y/interactive-supports-focus`	7.7	2.4	38k	~3.2×
`unicorn/prefer-number-properties`	12.8	6.0	335k (was ~674k)	~2.1×
`unicorn/no-thenable`	7.6	3.9	210k	~1.9×
`eslint/no-nonoctal-decimal-escape`	6.5	3.6	179k	~1.8×
`react/no-array-index-key`	6.7	4.3	188k	~1.6×
`eslint/no-obj-calls`	6.5	4.7	156k	~1.4×
`react/exhaustive-deps`	14.7	11.5	150k	~1.3×
`eslint/no-extend-native`	7.5	5.7	5.8k	~1.3×
`vitest/no-standalone-expect`	7.1	6.0	5.8k	~1.2×
`vitest/valid-expect`	7.5	8.5	5.8k	~noise / shared jest collect cost
`eslint/no-unused-vars`	15.7	19.0	5.7k	~noise (export set still helps structure; wall time dominated by `has_usages`)

Wall clock for the timed run was ~16.7s on 5763 files with 1 thread (machine-dependent). Per-rule ms can vary ±20% across runs; large call-count drops (no-restricted-globals, bad-array-method-on-arguments, prefer-number-properties) are the durable wins.

Changes (high level)

Identifier / unresolved-reference paths

eslint/no-restricted-globals, oxc/bad-array-method-on-arguments: run_once over only relevant root_unresolved_references (not every IdentifierReference). Member access via window/globalThis handled from those refs when checkGlobalObject is set. Do not implement both run and run_once — that forces NODE_TYPES = None and visits every AST node.
unicorn/prefer-number-properties: bare NaN / Infinity / method shorthand via run_once + unresolved names; run only for calls / global object members (drops IdentifierReference from NODE_TYPES).
eslint/no-unused-vars: precompute exported local names once per file for O(1) export checks; skip has_usages for exported symbols.

Call / JSX / comment hot paths

react/exhaustive-deps: gate on use* / React.use* then reactive-hook classification before ancestor walks.
eslint/no-useless-call: fast path for plain StaticMemberExpression .call / .apply.
eslint/no-obj-calls: skip binding resolution for direct non-matching globals.
jsx_a11y/interactive-supports-focus: require interactive handler + role before element-type / a11y helpers.
react/no-array-index-key: skip ancestor walk without key=; method names via match.
vitest/jest no-commented-out-tests: hand-rolled multiline scanner (no regex on every comment) with cheap keyword prefilter.
import/extensions: no-op fast path for empty/default config; cheaper extension parsing (Cow for written extensions).
oxc/no-this-in-exported-function: skip ThisExpressionFinder when body text has no this.
unicorn/no-thenable: fast path for static then keys.
eslint/no-extend-native: ASCII-uppercase gate before is_ecma_script_global.
eslint/no-nonoctal-decimal-escape: byte scan for \8/\9 prefilter.
valid-expect / no-standalone-expect: drop extra sort; reserve node map (remaining cost is largely shared collect_possible_jest_call_node).

Test plan

cargo test -p oxc_linter --lib (1162 passed)
cargo lint / CI Lint job (-D warnings) after clippy single_match_else fix on no-useless-call
RUSTDOCFLAGS='-D warnings' cargo doc / CI Doc job
CI Test Linux, Conformance, Spell Check, Linter changes
Release build + oxlint . --silent --threads=1 --debug timings on ~/workspace/x-web (3 runs, table above)

AI-assisted implementation; human-reviewed before submit.

Optimize several rules that showed up in x-web --debug timings: - eslint/no-unused-vars: precompute exported local names once per file - oxc/no-this-in-exported-function: skip AST visit when body has no "this" - react/exhaustive-deps: classify hooks before ancestor scope walks - vitest/jest no-commented-out-tests: substring prefilter before regex - import/extensions: avoid String allocs for common lowercase extensions - unicorn/prefer-number-properties: name-filter identifiers before global lookup - eslint/no-useless-call: early reject non-call/apply members Benchmarked on ~/workspace/x-web with --threads=1 --debug timings.

Second pass on rules still prominent in large-repo timings: - prefer-number-properties: move bare identifiers to run_once via root_unresolved_references (drops IdentifierReference from NODE_TYPES) - no-commented-out-tests: replace regex with a hand-rolled multiline scanner - exhaustive-deps: gate on use*/React.use* before name/hook classification - import/extensions: no-op fast path for empty/default config - no-useless-call: fast path for plain StaticMemberExpression callees - no-unused-vars: skip has_usages walk for exported symbols Also regenerate rule_runner_impls for prefer-number-properties.

- no-restricted-globals / bad-array-method-on-arguments: walk only relevant unresolved references (run_once) instead of every identifier - no-obj-calls: skip binding resolution for non-matching direct globals - no-nonoctal-decimal-escape: byte scan for \8/\9 prefilter - interactive-supports-focus: cheap handler/role gates before element type - no-array-index-key: skip ancestor walk without key=; match method names - no-extend-native: ASCII uppercase gate before is_ecma_script_global - no-thenable: fast path for static `then` object keys - valid-expect / no-standalone-expect: avoid extra sort; reserve hash map

Use only `run_once` (not `run`+`run_once`) so lintgen keeps RUN_FUNCTIONS as RunOnce and does not visit every AST node. Handle window/globalThis member access from unresolved global-object references. Dedupe global_objects when scanning.

codspeed-hq · 2026-06-26T16:55:12Z

Merging this PR will not alter performance

✅ 5 untouched benchmarks
⏩ 66 skipped benchmarks¹

_{Comparing anonrig:perf/lint-rules-hot-paths (90d1e68) with main (2a2d3b9)²}

66 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩
No successful run was found on main (5ac725d) during the generation of this report, so 2a2d3b9 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

camc314 · 2026-06-26T18:59:34Z

This optimization makes assumptions abount semantic that are incorrect. Hence this change is blocked by #23844

(no need to action this, I need to have a proper look next week)

- Split from #23829; original implementation by Yagiz Nizipli, retained as a commit coauthor. - Skip `ThisExpressionFinder` traversal when an exported function body does not contain the `this` substring. Original large-monorepo timings reported in #23829: 15.5 ms before, 2.1 ms after (~7.5x). Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

…as bytes (#23850) - Split from #23829; original implementation by Yagiz Nizipli, retained as a commit coauthor. - Scan string literal bytes for non-octal decimal escape sequences instead of decoding UTF-8 characters. - Use `array_windows` to keep the adjacent-byte scan clear and readable. Original large-monorepo timings reported in #23829: 6.5 ms before, 3.6 ms after (~1.8x). Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

…#23851) - Split from #23829; original implementation by Yagiz Nizipli, retained as a commit coauthor. - Skip lowercase identifiers before checking the ECMAScript global tables. - Preserve the existing Unicode-aware name filtering semantics while avoiding unnecessary global lookups. Original large-monorepo timings reported in #23829: 7.5 ms before, 5.7 ms after (~1.3x). Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

…ons first (#23854) - Split from #23829; original implementation by Yagiz Nizipli, retained as a commit coauthor. - Return early for JSX elements without interactive handlers or a static `role`. - Delay element-type and accessibility settings lookups until the cheaper checks pass. Original large-monorepo timings reported in #23829: 7.7 ms before, 2.4 ms after (~3.2x). Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

#23856) - Split from #23829. - Replace separate second- and third-index method arrays with a direct `match` that selects the callback index argument position. **Stack** - #23857 - #23856 👈 this PR! - #23855 Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

) Split from #23829. Delay `find_index_param_symbol_id` until after finding a relevant JSX `key` attribute or `React.cloneElement` key property. Use lazy, peekable iterators so the symbol lookup runs once only when needed while still checking every matching key. Add regression coverage for duplicate JSX attributes and object properties. **Stack** - #23857 👈 this PR! - #23856 - #23855 Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

Split from #23829. Adds a fast path when the comment definietly contains no test APIs Replaces the regex with a manual impl to improve perf Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

Split from #23829. Ports the shared valid-expect collection optimization from #23829 to avoid extra sorting work. Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

Split from #23829. Ports the shared no-standalone-expect collection optimization from #23829 to avoid extra sorting work. Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

…ences (#23890) Split from #23829. Ports the no-restricted-globals run_once optimization from #23829 by scanning relevant unresolved references instead of every identifier. Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

camc314 · 2026-07-02T15:24:13Z

Thanks for putting this together! Theres a ton of useful perf work in here!

I split the pieces that were straightforward to validate into smaller PRs so each rule change had its own review surface, tests, and performance/correctness tradeoff. In retrospect this also makes the history much easier to bisect or revert if one rule-specific optimization behaves differently on a real codebase.

Pieces from this PR that have shipped separately:

perf(linter): skip traversal without this expressions #23845 - oxc/no-this-in-exported-function: skip traversal when the body cannot contain this
perf(linter/eslint/no-nonoctal-decimal-escape): scan decimal escapes as bytes #23850 - eslint/no-nonoctal-decimal-escape: byte-scan decimal escapes
perf(linter/eslint/no-extend-native): skip lowercase references early #23851 - eslint/no-extend-native: skip lowercase references early
perf(linter/jsx-a11y/interactive-supports-focus): check cheap conditions first #23854 - jsx-a11y/interactive-supports-focus: check cheap conditions first
perf(linter/react/no-array-index-key): match callback methods directly #23856 - react/no-array-index-key: match callback methods directly
perf(linter/react/no-array-index-key): delay index symbol lookup #23857 - react/no-array-index-key: delay index symbol lookup
perf(linter/jest-vitest): speed up no-commented-out-tests #23864 - jest / vitest no-commented-out-tests: faster scanner
perf(linter/unicorn/prefer-number-properties): speed up global checks #23880 - unicorn/prefer-number-properties: speed up global checks
perf(linter/eslint/no-unused-vars): precompute exported bindings #23881 - eslint/no-unused-vars: precompute exported bindings
perf(linter/react/exhaustive-deps): skip non-reactive calls early #23882 - react/exhaustive-deps: skip non-reactive calls early
perf(linter/jest-vitest): speed up no-standalone-expect #23883 - jest / vitest no-standalone-expect: reserve the node map up front
perf(linter/eslint/no-restricted-globals): only scan unresolved references #23890 - eslint/no-restricted-globals: scan unresolved references once

I am splitting the remaining worthwhile pieces into separate follow-up PRs as well (no-useless-call, import/extensions, and no-obj-calls). I am going to leave out a couple of the smaller/riskier bits where the perf win did not look worth the complexity or where preserving deterministic behavior matters more.

Again, thank you for the benchmarking and implementaiton work here! The smaller PRs are mostly just to make review and maintenance easier, not because the overall direction was wrong.

Hope to see you around!

## Summary - replace the small non-callable global list lookup with a direct `matches!` helper - defer `resolve_global_binding` scoping/node setup until after its existing global-reference check Credit to Yagiz Nizipli for the original work in #23829. Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

…ns (#24075) ## Summary - split out the remaining safe `import/extensions` performance slice from #23829 for easier review and bisectability - skip the rule work early when the config is effectively empty/default and there is nothing to enforce - parse written extensions as `Cow<str>` so common lowercase specifiers stay borrowed - preserve the existing written-extension behavior for cases like written `.js` resolving to `.ts` Credit to Yagiz Nizipli for the original work in #23829. Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>

## Summary - Split the `eslint/no-useless-call` fast path out from #23829 for easier review and bisectability. - Add a direct static-member callee classifier for `foo.call(...)` / `foo.apply(...)`, so common non-`.call` / non-`.apply` static calls return before the generic member-expression fallback. - Keep the fallback for parenthesized, chained, computed, and optional-chain cases, with parenthesized regression coverage. Credit to Yagiz Nizipli for the original work in #23829. Co-authored-by: Yagiz Nizipli <yagiz@nizipli.com>