The generalized jbyte version has to perform a real cmpxchg on a full word (jint) so it has to isolate the correct byte within the word. We have a loop:

 while (cur_as_bytes[offset] == compare_value) {
    jint res = cmpxchg(new_val, dest_int, cur);
    if (res == cur) break;
    cur = res;
    new_val = cur;
    new_val_as_bytes[offset] = exchange_value;
  }

but if we fail the initial comparison there is no fence, as required by the semantics for cmpxchg as defined in atomic.hpp.

This was discovered during the review of JDK-8155949.