[patrick-immometrica]

Without the fast_escape feature, escape_html wrote its output one byte at a time. Since it always gets a &mut dyn Write (it's stored as Tera::escape_fn), that's a dynamic write_all call for every byte of every autoescaped string. This change writes the plain runs between special characters in one call, so a string without anything to escape is now a single write.

Output is the same. I added a small unit test for escape_html that covers the edge cases (empty input, specials at the start and end, adjacent specials, multibyte chars). It passes with and without fast_escape.

Numbers from my machine (release build, default features):

	before	after
render {{ text }}, 128 KB string, .html	765–862 µs	150–188 µs
escape_html via fn pointer, 1 KB plain text	7.2–9.0 µs	0.7–1.1 µs
escape_html via fn pointer, 1 KB HTML-heavy text	7.6–11.4 µs	1.8–2.0 µs
escape_html via fn pointer, 60 B plain text	0.42–1.3 µs	0.05–0.08 µs

The existing escape bench barely changes, because it calls escape_html directly and the compiler inlines it and devirtualizes the writer. During a render the function is called through the pointer, which is what the rows above measure.

[Keats]

Thanks