_REG_O3
r[_REG_O3] = tf->tf_out[3];
tf->tf_out[3] = r[_REG_O3];
gr[_REG_O3] = tf->tf_out[3];
tf->tf_out[3] = (uint64_t)gr[_REG_O3];
gr[_REG_O3] = tf->tf_out[3];
tf->tf_out[3] = (uint64_t)gr[_REG_O3];