This branch includes a benchmark to test the optimization of the html_entity_decode function. The previous version is still available as html_entity_decode_old. The benchmark code is in html_entity_decode_benchmark.php.
Just build PHP as usual and run:
./sapi/cli/php html_entity_decode_benchmark.php------------------------------------------------------------------------------
| Test | old avg(ns) | new avg(ns) | diff(%) |
------------------------------------------------------------------------------
| 4k & | 5949 | 21115 | -71.98% |
------------------------------------------------------------------------------
| only entities | 8279 | 10202 | -18.80% |
------------------------------------------------------------------------------
| 400 valid entities | 6439 | 5861 | 7.80% |
------------------------------------------------------------------------------
| 200 valid entities | 4891 | 3178 | 38.12% |
------------------------------------------------------------------------------
| 200 invalid entity | 4777 | 3181 | 37.29% |
------------------------------------------------------------------------------
| 200 ampersand | 4809 | 1221 | 198.35% |
------------------------------------------------------------------------------
| 100 valid entities | 4188 | 1777 | 124.49% |
------------------------------------------------------------------------------
| 50 valid entities | 2885 | 979 | 193.50% |
------------------------------------------------------------------------------
| String ends with & | 2428 | 176 | 1221.69% |
------------------------------------------------------------------------------
As you can see, the speedup depends on the number of entities and & characters in the string — the fewer there are, the more noticeable the performance improvement.
In edge cases, where the string consists entirely of & characters or valid HTML entities, performance actually worsens. However, I don't think this is a common scenario.
Either way, I plan to continue optimizing and implement & scanning using SIMD instructions, which should significantly improve performance even in high-entity-density cases.