If you view the raw file and look for the definition of

Item_func_ieee754_to_string::val_str(String *str)

the comment specifies that 17 digits of precision are needed for double. It is also true that 9 digits are needed for single precision. I didn't believe it, but I tested it once with every possible value of a single precision float and found that there were cases where 8 digits was not enough to support a lossless conversion from float to string and back.