[fedifazzy]

Motivation

collect() is O(n²) in the number of top-level rules.

It moves every top-level rule of the parsed stylesheet into critical or other with append. For a node that still has a parent, append removes it with splice. The rule being moved is always at index 0, so every move shifts all remaining nodes.

On a 1.6 MB application stylesheet (~13.5k top-level rules), one call takes ~0.2 s on Node 20, and seconds on a busy server while incremental marking is active. collect() is synchronous, so the event loop is blocked for the whole call on every SSR request. The implementation is the same in 6.3.0 and 8.2.0.

While testing, two more bugs showed up:

• other loses the non-critical rules of an at-rule, and repeats its critical ones, whenever a non-critical rule comes first. For example, @media (min-width: 1px) { .n {} .c {} } with class="c" gives @media (min-width: 1px) { .c {} } in other.
• blockedClasses lets every second matching rule into critical. The regexp has the g flag, so test keeps lastIndex between selectors.

Summary

• Rules are collected into arrays first. Then the stylesheet is detached at once with removeAll(), and the nodes are appended, so nothing is spliced out of the big root. Handled top-level at-rules are no longer removed one by one either. Keyframes are looked up only where they used to stay: in top-level at-rules that were not handled.
• handleAtRule filters both clones of the at-rule with isCritical directly, instead of the index - removedNodesFromOther arithmetic.
• The class regexps no longer use the g flag, and the classes from the HTML are deduplicated.
• Whitespace between rules now follows the source stylesheet; previously the removals glued the rules together. Neither output starts with whitespace.
• The @linaria/server tests were not running. The package had no test script, and since prettier 3 made format async, every prettified snapshot was compared with a pending Promise: 16 of 22 failed on master. The first commit awaits the standalone prettier with the postcss plugin, which also works without --experimental-vm-modules, and adds the test script.

One collect() call after a warm-up, on N rules .c{i}{color:red}:

Top-level rules	Node 20, before	Node 20, after	Node 20 --stress-incremental-marking, before	after
8,000	22 ms	12 ms	235 ms	66 ms
16,000	123 ms	26 ms	749 ms	164 ms
32,000	445 ms	77 ms	2,570 ms	454 ms

On a real 1.6 MB stylesheet with the HTML of six pages, a call takes 67–92 ms instead of 164–185 ms on Node 20. critical is identical apart from whitespace.

Most of the remaining time goes to parsing the stylesheet and building other on every call. An API that accepts a prepared stylesheet, or lets callers skip other, could remove that, but it is out of scope here.

Test plan

• pnpm --filter @linaria/server test: 23 tests, 22 snapshots.

• New tests cover:

  • splitting an at-rule between critical and other;
  • repeated blockedClasses;
  • rule order in a 20,000-rule stylesheet.

  The first two fail on master.

• The snapshot changes are whitespace only: blank lines between rules, as in the source. One more snapshot is re-serialized without escaped quotes.

• pnpm --filter @linaria/server typecheck, eslint packages/server and prettier are clean.

• Changeset: @linaria/server patch.

🤖 Generated with Claude Code

[changeset-bot]

🦋 Changeset detected

Latest commit: e591765

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 9 packages

Name	Type
@linaria/server	Patch
@linaria/atomic	Patch
@linaria/core	Patch
@linaria/babel-plugin-interop	Patch
linaria	Patch
@linaria/postcss-linaria	Patch
@linaria/react	Patch
@linaria/stylelint-config-standard-linaria	Patch
@linaria/stylelint	Patch

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

[Anber]

Thanks! There was a small bug, I've fixed it. We'll release this soon.