More NotWordBoundaryAscii woes

[SeanRBurton]

I'm sorry to say that another version of issue #241 has reared its ugly head.

Running Group { e: NotWordBoundaryAscii, i: Some(1), name: None } against "\u{28f3e}" in the same way as in that issue gives the match None with the DFA, but Some((0, 0)) with the NFA.

[BurntSushi]

A smallerish repro:

extern crate regex;

fn main() {
    let re0 = regex::Regex::new(r"(?-u)\B").unwrap();
    let re1 = regex::Regex::new(r"(?-u)(\B)").unwrap();
    let s = "\u{28f3e}";
    assert!(re0.is_match(s));
    assert!(re1.is_match(s));
    assert!(re0.find(s).is_some());
    assert!(re1.find(s).is_some());
    assert!(re0.captures(s).is_some());
    assert!(re1.captures(s).is_some());
}

[BurntSushi]

The above actually fails on the last assert. My guess is that something about the DFA -> NFA interaction is bunk, because the presence of the capture group and the call to captures is what will force the exec engine to run the NFA after the DFA.

[BurntSushi]

Indeed. The code was trying to be cute about setting the end of the haystack. In this case, it was adding 1 unconditionally to the end of the match location when in fact it needs to add a number of bytes equal to the next encoded codepoint.