This means floatingpoint calculation in a firm graph must be performed
in the mode specified (and may not be performed in a wider mode as
before).