Ok so I just implemented this myself and noticed that counting nulls as different values is not ANSI compliant.
The ANSI compliant way of encoding would be:

count(distinct coalesce(nullif(coalesce(col1 || '', '\0'), ''), '\01') || '\0' || coalesce(nullif(coalesce(col2 || '', '\0'), ''), '\02'))

Note that the columns are concatenated with an empty string to get implicit conversion. NULL is encoded as '\0' and empty strings are encoded as '\0' + element number.
This should produce collision free strings as long as the column values don't contain the '\0' character.