The direction of truncation when an integral number is converted to a floating-point number that cannot exactly represent the original value (§3.2.1.3)
When an integral number is cast to a floating-point value that cannot exactly represent the value, the value is rounded (up or down) to the nearest suitable value.
For example, casting an unsigned long (with 32 bits of precision) to a float (whose mantissa has 23 bits of precision) rounds the number to the nearest multiple of 256. The long values 4,294,966,913–4,294,967,167 are all rounded to the float value 4,294,967,040.