Reading a float as uint32_t

This code reads the bits of a float. Why can it give wrong results at -O2, and what is the safe way?

float f = read_sensor();
uint32_t bits = *(uint32_t *)&f;