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.