As long as you define the variables used with millis() as
unsigned long
and do all your math as subtractions (millis_now - millis_before)
the math will work out.
For example, say the time just rolled over and you captured it at 0x0000 0030
and time before was 0xFFFF FF10
then 0000 0030 - FFFF FF10 will yield a correct result of 0x0000 0120