<<It would also be rather dubious on the SH4, where single precision is considerably faster than double precision. >> Not if you already have computed the double precision result. In other words CSE could regard a*b as equivalent to double(a)*double(b). I agree this is an unusual case, but certainly might occur.