reimplement truncate() to make it optimal.

It is optimal at least up to 7 bits (I've tested all such cases)
This change to truncate() allows a little simplification to the multiplication code,
and it also makes multiplication optimal :)

llvm-svn: 160512