perf(html): speed up the experimental HTML parser and reduce its memory usage#21130
Merged
Conversation
walkHtmlTokens: replace the per-code-point ASCII predicate chains with a single packed Uint8Array char-class lookup table, hoist the two closure-local predicates to module scope, and slice the named-character-reference candidate run from the input once instead of re-slicing per prefix length. buildHtmlAst: move per-token tag-name membership tests off freshly-allocated array literals (Array#includes) onto shared module-level Sets, hoist the per-call Sets (implied-end-tags, cell close, whitespace) to module scope, collapse inTableScope array arguments to a string/Set fast path, guard the CR and NULL text rewrites behind a cheap presence check, and dispatch token end-offset tracking on type instead of the megamorphic in operator.
The tree builder allocated one token object (plus a nested pos) per
tokenizer callback and one {parent,beforeNode} per inserted node, and
the token union's varying shapes made the per-token t.* reads
megamorphic. Funnel every callback through one reused MutableToken with a
fixed shape (pos reused too) and return insertion places via one shared
object; both are consumed synchronously and never retained, and the only
buffered tokens (inTableText) are snapshotted into fresh objects. Cuts
minor GCs by ~24% on a tag-heavy document with no behaviour change
(full html5lib suite still green).
The data / RCDATA / RAWTEXT / script-data / PLAINTEXT and quoted attribute-value states advanced one code point at a time, re-entering the 80-case state switch for every ordinary character. Fast-forward over the run of insignificant code points in a tight inner loop that stops at the state's delimiters (NULL included so per-character error reporting is preserved). ~45% faster tokenizing on text-heavy input; no behaviour change (full html5lib suite green).
…osure walkHtmlTokens recorded the last open tag's lowercased name on every start tag, but it is only consulted by the RAWTEXT/RCDATA/script end-tag states. Match the content mode against the raw tag-name range (case-insensitive, no slice) and materialize the lowercased lastOpenTagName only when a special content mode is actually entered, so ordinary tags allocate nothing. buildHtmlAst's attribute callback now dedupes with a plain loop instead of an Array#some closure allocated per attribute. ~16% faster tokenizing on attribute-heavy input.
Initialize templateContent on every element (only <template> fills it) and serializedName on every attribute (only foreign content fills it) at creation instead of adding the property later, so each keeps a single monomorphic hidden class for the open-stack/scope walks and the AST consumers. No behaviour change.
Cut intermediate objects that were built and immediately discarded while constructing the AST (the output nodes themselves are irreducible): - insertCharacters merges a run into the adjacent text sibling by appending the string directly, instead of always allocating a text node and letting insertAtPlace discard it on merge. - sameAttrs compares attribute lists with a nested scan instead of building a Map (+ array) per formatting-element comparison. - adjustForeignAttrs / adjustMathmlAttrs fork the attribute array lazily and reuse the original objects when nothing needs adjusting, instead of mapping to a fresh array + object per attribute on every foreign element (~11% faster building SVG-heavy input). No behaviour change; full html5lib suite green.
…ments Most elements have no attributes, yet each was given its own empty attributes array (plus a fresh empty pendingAttrs buffer per tag). Only <html>/<body> ever receive merged attributes and are always built with their own mutable array, so every other attributeless element can share one frozen EMPTY_ATTRS. The tokenizer callbacks now reuse the empty pendingAttrs buffer instead of reallocating it, and synthesized elements pass EMPTY_ATTRS. ~12% fewer minor GCs on attributeless-heavy input (tables/lists/formatted text). No behaviour change; html5lib + html configCases green.
…lookup process() ran modes[mode](t) per token — a megamorphic keyed load over the ~21 insertion-mode strings. Route the four dispatch sites through a runMode() switch (cases ordered by frequency, default falling back to the keyed load), turning the hot per-token dispatch into monomorphic direct calls. ~3-4% faster building a large realistic document; no behaviour change (full html5lib + html configCases green).
contentModeAfterOpenTag ran the isForeign callback (which calls adjustedCurrent) after every open tag, but isForeign only ever vetoes a switch *into* a special content mode — it can't turn a data-state tag into a special one. Resolve the (allocation-free) tag-name range first and only consult isForeign when the tag would actually enter RAWTEXT/RCDATA/script, so ordinary tags skip the callback. Behaviour identical; full html5lib suite green.
🦋 Changeset detectedLatest commit: 3af9ea1 The changes in this PR will be included in the next version bump. This PR includes changesets to release 1 package
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
Contributor
|
This PR is packaged and the instant preview is available (cd45931). Install it locally:
npm i -D webpack@https://pkg.pr.new/webpack@cd45931
yarn add -D webpack@https://pkg.pr.new/webpack@cd45931
pnpm add -D webpack@https://pkg.pr.new/webpack@cd45931 |
Codecov Report❌ Patch coverage is
❌ Your patch check has failed because the patch coverage (84.07%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage. Additional details and impacted files@@ Coverage Diff @@
## main #21130 +/- ##
==========================================
+ Coverage 92.17% 92.32% +0.14%
==========================================
Files 581 581
Lines 62946 63179 +233
Branches 17422 17467 +45
==========================================
+ Hits 58023 58331 +308
+ Misses 4923 4848 -75
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Object.freeze([]) is readonly never[], which TS won't narrow directly to the mutable HtmlAttribute[]; cast through unknown (lint:types only runs in CI, not the pre-commit hooks).
Merging this PR will improve performance by 97.69%
|
| Mode | Benchmark | BASE |
HEAD |
Efficiency | |
|---|---|---|---|---|---|
| ⚡ | Memory | benchmark "asset-modules-bytes", scenario '{"name":"mode-development-rebuild","mode":"development","watch":true}' |
858.9 KB | 320.5 KB | ×2.7 |
| ⚡ | Memory | benchmark "react", scenario '{"name":"mode-development-rebuild","mode":"development","watch":true}' |
332.6 KB | 156.8 KB | ×2.1 |
| ⚡ | Memory | benchmark "side-effects-reexport", scenario '{"name":"mode-development-rebuild","mode":"development","watch":true}' |
1,186.9 KB | 873.1 KB | +35.95% |
Tip
Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.
Comparing perf/html-parser-optimizations (3af9ea1) with main (d39efba)
Footnotes
-
18 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. ↩
Contributor
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
The experimental HTML parser (
walkHtmlTokens+buildHtmlAst, introduced in #21116) ends up on per-module hot paths, so its constant factors matter for build time and peak heap. This PR applies the same kind of low-level work the CSS tokenizer recently got, with no change to parsing behaviour:isForeigncheck for tags that can't switch content mode.switch-based insertion-mode dispatch in place of a megamorphic keyed lookup.Measured on a ~3 MB realistic document: tokenizing ~18% faster, full parse ~28% faster, ~37% fewer minor GCs; on text/prose-heavy input up to ~44–45% faster. No linked issue.
What kind of change does this PR introduce?
perf
Did you add tests for your changes?
No new tests — these are behaviour-preserving performance changes, verified to be byte-for-byte equivalent against the existing suites: the full html5lib tokenizer + tree-construction corpus (
test/html5lib.spectest.js, 15k+ cases),test/walkHtmlTokens.unittest.js,test/buildHtmlAst.unittest.js,test/HtmlParser.unittest.js, and the HTMLconfigCases.Does this PR introduce a breaking change?
No.
If relevant, what needs to be documented once your changes are merged or what have you already documented?
n/a
Use of AI
Yes — these changes were written with Claude Code: it profiled the parser, proposed and implemented the optimizations, and measured before/after. Every change was validated against the full html5lib corpus and the HTML test suites, and candidates that did not show a measurable, regression-free win were discarded. All output was reviewed.
Generated by Claude Code