Adjust the end of the haystack after a DFA match.

If the caller asks for captures, and the DFA runs, and there's a match,
and there are actually captures in the regex, then the haystack sent to
the NFA is shortened to correspond to only the match plus some room at the
end for matching zero-width assertions. This "room at the end" needs to be
big enough to at least fit an UTF-8 encoded Unicode codepoint.

Fixes #264.

src/exec.rs

@@ -840,11 +840,7 @@ impl<'c> ExecNoSync<'c> {
     ) -> Option<(usize, usize)> {
         // We can't use match_end directly, because we may need to examine
         // one "character" after the end of a match for lookahead operators.
-        let e = if self.ro.nfa.uses_bytes() {
-            cmp::min(match_end + 1, text.len())
-        } else {
-            cmp::min(next_utf8(text, match_end), text.len())
-        };
+        let e = cmp::min(next_utf8(text, match_end), text.len());
         self.captures_nfa(slots, &text[..e], match_start)
     }
 

tests/regression.rs

@@ -60,3 +60,7 @@ matiter!(word_boundary_dfa, r"\b", "a b c",
 
 // See: https://github.com/rust-lang-nursery/regex/issues/268
 matiter!(partial_anchor, u!(r"^a|b"), "ba", (0, 1));
+
+// See: https://github.com/rust-lang-nursery/regex/issues/264
+mat!(ascii_boundary_no_capture, u!(r"(?-u)\B"), "\u{28f3e}", Some((0, 0)));
+mat!(ascii_boundary_capture, u!(r"(?-u)(\B)"), "\u{28f3e}", Some((0, 0)));