But this way you throw away a lot of functionality, make the existence
of two pointers pointless, cause pthread_self() to change across fork
and force NPTL to copy thread state.

How about instead doing a verify_area in copy_process, putting the
child_settid address and the tid in two child registers and assigning it
in assembly in ret_from_fork?

Alternatively you could also manually call the copy-on-write handler
functions but this adds complexity for little gain.