Code for the read-before-write removal optimisation when using 
unique indexes did not work correctly when the base table
had no primary key (i.e. hidden primary key).