[php-src] PR #24142: ext/standard: Stop at the first non-name byte in html_entity_decode()

Pull Request: https://github.com/php/php-src/pull/24142 Author: ArtUkrainskiy # ext/standard: Stop at the first non-name byte in html_entity_decode() Follow-up to #18092. That PR made html_entity_decode() a lot faster on normal text, but strings with many & became slower, as I noted in its description. The reason is the ; search: for every & we call memchr() to look for a ; up to 32 bytes ahead and then hash whatever is in between, even when it is clearly not an entity (&&, & , &foo&bar). @bukka suggested in the review of #18092 that a plain inline loop could beat memchr() here, since entity names are short. I didn't try it back then. Now I did, and it works: scan the name while it is letters and digits and stop at the first other byte. If that byte is not a ;, it is not an entity and there is nothing to hash. I also skip the memchr() for & when the next byte already is one. html_entity_decode(), 4 KB inputs, master vs this PR: | input | | |---|---| | only & | 3.5× faster | | only entities | 1.1× faster | | 1–25% of the bytes in entities | same to 1.06× faster | | jQuery source | 1.1–1.4× faster | | jQuery minified | 1.4–1.7× faster | | HTML pages, Markdown | same | | &param=value URLs in text | up to 1.07× slower | The string of & is back to where it was before #18092. The one case that gets a bit slower is raw URL parameters in text (&logoColor=white): the name is scanned byte by byte where a single memchr() call used to be enough. I think that is an acceptable trade. The output does not change: I compared it with master on 50k random inputs with different flags and charsets, and ext/standard/tests/strings passes. htmlspecialchars_decode() goes through the same code.