Mirrors/eepp
1
0
Fork 0
mirror of https://github.com/SpartanJ/eepp.git synced 2026-10-02 03:00:23 +03:00
Code Issues Packages Projects Releases Wiki Activity

Optimize CSS selector matching with cached hashes

Browse Source 
Cache tag, ID, and class hashes in selector rules and widgets to avoid
  repeated string comparisons during selector matching.

  Store widget class hashes in a single-element SmallVector so the common
  zero and one-class cases remain allocation-free. Reorder UINode and
  UIWidget members to recover padding, keeping UIWidget at its original
  1144-byte size.

  Add selector matching benchmarks and regression coverage for class
  mutation, compound selectors, and hash cache synchronization. Document
  the measured Phase 5 performance and memory results.
This commit is contained in:
Martín Lucas Golini
2026-07-10 17:13:13 -03:00
parent bb9a733008
commit 3e29da6a2e
8 changed files with 220 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.agent/plans/eepp_css_selector_optimization_plan.md
+30
View File
@@ -461,6 +461,36 @@ Use these counters to verify each later phase.
# Phase 5: Add Hash-Based Tag/ID/Class Matching
**Status: Implemented**
## Implementation State
- Selector rules cache tag, ID, and sorted unique class hashes.
- `UIWidget` caches its tag hash and maintains sorted unique class hashes across every class
mutation API.
- Widget class hashes use `SmallVector<String::HashType, 1>`. The zero/one-class cases remain
allocation-free, while widgets with multiple classes use indirect storage.
- `StyleSheetSelectorRule::matches()` compares tag and ID hashes and checks all required class
hashes after decoding widget hash storage once per rule match.
- `UINode` and `UIWidget` members were reordered to reclaim alignment holes introduced or exposed
during the cache work.
- A focused selector benchmark and class-mutation regression test cover the implementation.
## Measured Result
- The release unit-test workload observed 30,431 destroyed widgets: 74.5% had no classes, 19.4%
had one, 5.9% had two, and approximately 0.1% had three or more.
- `sizeof( UIWidget )` is 1144 bytes, equal to the pre-Phase-5 size. `sizeof( UINode )` decreased
from 856 to 848 bytes.
- The focused release selector benchmark measured a median near 111.4 ms versus the Phase 3
baseline near 115.1 ms, approximately a 3.2% matching improvement without net `UIWidget`
growth.
- The full release suite passes 734 tests with one skipped test.
Future work should investigate atomized/interned class names as a separate architectural change.
That could replace both per-widget class strings and the hash cache, but requires explicit
ownership, collision, API, and cross-scene lifetime design.
## Motivation
`StyleSheetSelectorRule::matches()` currently compares strings repeatedly and class matching performs a nested linear search: