i18n: keep fold search, isolate wrapping and locale fallbacks linear

- Folded::matches restarted str::find one character past every occurrence
  turned down at a combining mark, so overlapping occurrences of a long
  needle cost haystack x needle. It now finds every occurrence, overlapping
  ones included, in one Knuth-Morris-Pratt pass and checks the boundaries
  as before.
- wrap searched the whole open-control stack for an isolate at every PDI;
  it now keeps a count of open isolates.
- candidates made one joined prefix per subtag of a locale tag of any
  length; a tag longer than 128 bytes is not a locale and has none.

crates/i18n/src/fold.rs

@@ -191,6 +191,23 @@ fn continues_syllable(prev: HangulSyllableType, next: HangulSyllableType) -> boo
     )
 }
 
+/// For each prefix of `n`, the length of its longest proper prefix that is also its suffix: the
+/// Knuth-Morris-Pratt failure links.
+fn prefix_links(n: &[u8]) -> Vec<usize> {
+    let mut links = vec![0; n.len()];
+    let mut k = 0;
+    for i in 1..n.len() {
+        while k > 0 && n[k] != n[i] {
+            k = links[k - 1];
+        }
+        if n[k] == n[i] {
+            k += 1;
+        }
+        links[i] = k;
+    }
+    links
+}
+
 /// Folded text and where each part came from.
 #[derive(Clone, Debug, PartialEq, Eq)]
 pub struct Folded {
@@ -244,8 +261,13 @@ impl Folded {
             let inside = self.starts.binary_search_by_key(&at, |&(f, _)| f).is_err();
             inside && self.text[..at].chars().next_back().is_some_and(|p| continues_syllable(hangul.get(p), hangul.get(c)))
         };
-        let n = needle.text.as_str();
-        let mut from = 0;
+        // Every occurrence, overlapping ones included, is found in one left-to-right pass, so a
+        // run of occurrences turned down at a mark doesn't search the needle again from each
+        // one. Both texts are UTF-8, so a byte match starts and ends on character boundaries.
+        let n = needle.text.as_bytes();
+        let hay = self.text.as_bytes();
+        let links = prefix_links(n);
+        let (mut at, mut k) = (0, 0);
         let mut done = false;
         std::iter::from_fn(move || {
             if done {
@@ -255,14 +277,22 @@ impl Folded {
                 done = true;
                 return Some(0..0);
             }
-            while let Some(i) = self.text.get(from..).and_then(|t| t.find(n)) {
-                let start = from + i;
-                let end = start + n.len();
-                if !starts_mark(start) && !starts_mark(end) {
-                    from = end;
-                    return Some(self.source_range(start..end));
+            while let Some(&b) = hay.get(at) {
+                while k > 0 && n[k] != b {
+                    k = links[k - 1];
+                }
+                if n[k] == b {
+                    k += 1;
+                }
+                at += 1;
+                if k == n.len() {
+                    let (start, end) = (at - k, at);
+                    if !starts_mark(start) && !starts_mark(end) {
+                        k = 0;
+                        return Some(self.source_range(start..end));
+                    }
+                    k = links[k - 1];
                 }
-                from = start + self.text[start..].chars().next().map_or(1, char::len_utf8);
             }
             done = true;
             None
@@ -351,6 +381,23 @@ mod tests {
 
         let all: Vec<_> = f.fold("Ab ab AB").matches(&f.fold("ab")).collect();
         assert_eq!(all, [0..2, 3..5, 6..8]);
+        let all: Vec<_> = f.fold("aaaaa").matches(&f.fold("aa")).collect();
+        assert_eq!(all, [0..2, 2..4]);
+        assert_eq!(find("abababc", "ababc"), Some("ababc"));
+    }
+
+    #[test]
+    fn occurrences_turned_down_at_a_mark_are_not_searched_again() {
+        // Every occurrence of the needle ends just before an accent, except the one at the end:
+        // each is turned down, and the next one overlaps it.
+        let f = Folder::new("en").keep_accents();
+        let haystack = format!("{}a", "a\u{301}".repeat(100_000));
+        let needle = format!("{}a", "a\u{301}".repeat(50_000));
+        let folded = f.fold(&haystack);
+        let start = std::time::Instant::now();
+        let found = folded.find(&f.fold(&needle));
+        assert_eq!(found, Some(haystack.len() - needle.len()..haystack.len()));
+        assert!(start.elapsed() < std::time::Duration::from_secs(5), "{:?}", start.elapsed());
     }
 
     #[test]

crates/i18n/src/lib.rs

@@ -471,18 +471,24 @@ fn wrap(initiator: char, value: &str) -> String {
     out.push(initiator);
     // One entry per open initiator: true for an isolate, false for an embedding or override.
     let mut open: Vec<bool> = Vec::new();
+    // How many of `open` are isolates, so a PDI needn't search the stack for one.
+    let mut isolates = 0usize;
     for c in value.chars() {
         match c {
-            LRI | RLI | FSI => open.push(true),
+            LRI | RLI | FSI => {
+                open.push(true);
+                isolates += 1;
+            }
             LRE | RLE | LRO | RLO => open.push(false),
             PDF if open.last() == Some(&false) => {
                 open.pop();
             }
             PDF => continue,
-            PDI if open.contains(&true) => {
+            PDI if isolates > 0 => {
                 while open.pop() == Some(false) {
                     out.push(PDF);
                 }
+                isolates -= 1;
             }
             PDI => continue,
             _ => {}
@@ -580,8 +586,17 @@ fn fill(template: &str, args: &[(&str, &str)], isolated: bool) -> String {
 /// script, so Latin Serbian never picks Cyrillic `sr`, while `sr-Cyrl-RS` does. A region that
 /// implies a script (CLDR likely subtags) counts as that script after the tag as written:
 /// `pa-PK` → `pa-pk`, `pa-arab-pk`, `pa-arab`.
+/// A locale tag is no longer than this: a language, script, region and a few variants or a
+/// glibc encoding and modifier fit with room to spare. A longer one is no locale, and isn't
+/// split into a fallback for every subtag.
+const MAX_TAG_LEN: usize = 128;
+
 fn candidates(tag: &str) -> Vec<String> {
-    let tag = tag.trim().to_ascii_lowercase();
+    let tag = tag.trim();
+    if tag.len() > MAX_TAG_LEN {
+        return Vec::new();
+    }
+    let tag = tag.to_ascii_lowercase();
     let base = tag.split(['.', '@']).next().unwrap_or("").replace('_', "-");
     let mut parts: Vec<&str> = base.split('-').filter(|p| !p.is_empty()).collect();
     let Some(&primary) = parts.first() else {
@@ -843,6 +858,12 @@ mod tests {
         assert_eq!(fmt("ש {n}", &[("n", &one)]), format!("ש {one}"));
         // Two isolates side by side are not one, so they get wrapped.
         assert_eq!(fmt("ש{v}", &[("v", "\u{2066}a\u{2069}\u{2066}b\u{2069}")]), "ש\u{2068}\u{2066}a\u{2069}\u{2066}b\u{2069}\u{2069}");
+        // A PDI closes the embeddings opened after its isolate, and one with no isolate open is
+        // dropped, however many embeddings are open.
+        assert_eq!(isolate("\u{2066}\u{202A}a\u{2069}b"), "\u{2068}\u{2066}\u{202A}a\u{202C}\u{2069}b\u{2069}");
+        let embeddings = "\u{202A}".repeat(50_000);
+        let wrapped = isolate(&format!("{embeddings}{}", "\u{2069}".repeat(50_000)));
+        assert_eq!(wrapped, format!("\u{2068}{embeddings}{}\u{2069}", "\u{202C}".repeat(50_000)));
     }
 
     /// `text` in display order, as one line in a left-to-right paragraph, without bidi controls.
@@ -1033,6 +1054,9 @@ mod tests {
         assert_eq!(candidates("kk-CN"), ["kk-cn", "kk-arab-cn", "kk-arab"]);
         assert_eq!(candidates("kk-KZ"), ["kk-kz", "kk"]);
         assert_eq!(candidates("ug-KZ"), ["ug-kz", "ug-cyrl-kz", "ug-cyrl"]);
+        // A tag too long to be a locale has no fallbacks, rather than one per subtag.
+        assert!(candidates(&"x-".repeat(500_000)).is_empty());
+        assert_eq!(candidates(&format!("en-{}", "x-".repeat(60))).len(), 61);
     }
 
     #[test]