I am afraid there is no such document (at least not for public).
“Divide instructions have a latency of 5–34 cycles depending on the operand data. While the divide is running, the rest of the pipeline is unavailable for additional instructions (blocking divide).”
Let’s imagine you manually divide numbers on the paper. It takes variable number of lines according to operands until you reach indivisible remainder. MCU’s divider works the similar way.
If you do some performance analysis, you should count with worst-case, thus 34 cycles.