Expand description
The single intentional deviation from OpenJDK: UTF-16 code-unit offsets
vs Unicode-codepoint offsets in Matcher.start() / end().
§Differences from OpenJDK’s java.util.regex
This implementation faithfully reproduces OpenJDK’s java.util.regex.Pattern behavior — including OpenJDK’s well-known quirks (those are documented in QUIRKS.md, not here). The only remaining behavioral difference is a structural one rooted in the host language’s string representation.
§1. Position offsets for supplementary code points
java.util.regex.Matcher.start() / Matcher.end() return UTF-16 code-unit offsets, because Java’s String is internally UTF-16. This implementation returns char (Unicode scalar value) offsets, because Rust’s String is UTF-8 and char is a code point.
The two agree exactly on inputs that contain no supplementary characters (every code point ≤ U+FFFF — virtually all text, including all Latin/Greek/Cyrillic/CJK BMP scripts, most emoji-free content, etc.).
For inputs containing supplementary characters (most emoji, several CJK extension blocks, math/historic scripts) a single character is one position in Rust but two positions in Java. Matched text is always identical; only the integer positions differ.
Pattern: \Q\E (zero-width match at every position)
Input: "😀a"
Rust: matches at positions 0, 1, 2 (3 chars)
Java: matches at positions 0, 1, 2, 3 (4 UTF-16 code units)This is a deliberate, fundamental difference rooted in the host language. Code that needs Java-compatible offsets on supplementary input can convert: s.encode_utf16().take(n_chars).count() gives the Java offset for the n-th char.
§Summary
| Behavior | OpenJDK 25 | This implementation |
|---|---|---|
start()/end() indexing | UTF-16 code units | Unicode code points |
That’s it — the only difference. Every other previously-documented divergence has been brought into alignment with OpenJDK:
- Group capture leaks across
find()start positions (was: clean reset, now: leaks like Java does) — see QUIRKS.md §8. - Lookbehind with unbounded multi-char body (was: accepted, now: rejected at compile time like Java).
- Negative lookbehind capture leak (was: clean reset, now: leaks like Java).
For all tested patterns the matched text, match counts, capture group values, split boundaries, and replacement results are identical to OpenJDK 25 across 200,000+ random differential-fuzzer tests. The only mismatches surface on inputs containing supplementary characters, and only in the position-offset integers — never in matched text.