If you are looking for a performance, I'd suggest really profile the code. Cipher itself has a lot of place to improve

BTW, NID allocation may require some locking.

I re-wrote it with inline assembly so that's 2% of optimized score.
Still, that seems alot.

Performance drop is caused by call of create_NIDs function.
It is true even if I define it like this
(which is similar in context of missing_NIDs having 0 elements):

__attribute__((noinline))
static int create_NIDs() {
    int i;
    asm("");
    return 1;
}

Not sure this will be investigated any further so I close this ticket.