The patch looks fine, except of testing for true value of utf16 in each iteration. it might be the cause of the performance degradation, if you compiled without proper optimizations.

@yoram: the previous implementation have the same problem.

    for (;;) {
    ...
            if (w) {
    ...

For now, there is no performance degradation, but a small improvement (according to my quick bench, ~5%)

Of course we can optimize this (and probably also the "bits" handling in json_escape_string, avoid the REVERSE code and make it more readable)

   smart_str_appendl(buf, "\\u", 2);
   smart_str_appendc(buf, digits[(us & 0xf000) >> 12]);
   smart_str_appendc(buf, digits[(us & 0xf00)  >> 8]);
   smart_str_appendc(buf, digits[(us & 0xf0)   >> 4]);
   smart_str_appendc(buf, digits[(us & 0xf)]);