Interesting result.
It seems https://3v4l.org/d8IAo is a worst case and mbstring's case insensitive code has performance issue in general. 

Tried case sensitive https://3v4l.org/ZF3TK version. The result varies, but it seems mb_* functions have improvement margin. I get slower results consistently with current master on my real PC. (I suppose 3v4l is VM or container) 

If I use 
 $raw = '0123456789abcdefghijklmnopqrstuvwxyzあいうえおかきくけこさしすせそなにぬねのはひふへほまみむめもやいゆえよわをん';
for https://3v4l.org/ZF3TK, case sensitive version mb_strstr() is a little faster. BTW, pcre became a lot faster than 5.6. This is the reason why mb_strstr() is a little faster than pcre version.

If anyone could make more generic benchmark script, i.e. benchmark mb_*() functions, it would be helpful.

Apparently, this is not a general mb_*() issue, but rather related
to case-insensitive operations, as Yasuo has pointed out. It seems
that this is been caused due to the lack of native mbfl_*() case
insensitive operations, so PHP converts $needle and $haystack to
upper-case intermittently, what has some overhead.

Anyhow, this is not a bug, but rather a feature request to improve
the performance of case-insensitive mb_*() operations.

The good news is that the first test case (using mb_stristr) is now more than 6 times faster on master. The bad news is that this is still 3.5 times slower than the PCRE version.