The f16tof32 opcode was indexing a vector with a float 0, rather
than an int 0.  It may have made no functional difference due to the
identical bit pattern, but code looking at the type could be
confused.