(6.2.1.3) The direction of truncation when an integral number is converted to a floating-point number that cannot exactly represent the original value:

Numbers are rounded to the nearest value that can be represented.