170 	t[ 0] = MUL31(a[0], b[0]);  in mul9()
171 	t[ 1] = MUL31(a[0], b[1])  in mul9()
172 		+ MUL31(a[1], b[0]);  in mul9()
173 	t[ 2] = MUL31(a[0], b[2])  in mul9()
174 		+ MUL31(a[1], b[1])  in mul9()
175 		+ MUL31(a[2], b[0]);  in mul9()
176 	t[ 3] = MUL31(a[0], b[3])  in mul9()
177 		+ MUL31(a[1], b[2])  in mul9()
178 		+ MUL31(a[2], b[1])  in mul9()
179 		+ MUL31(a[3], b[0]);  in mul9()
180 	t[ 4] = MUL31(a[0], b[4])  in mul9()
181 		+ MUL31(a[1], b[3])  in mul9()
182 		+ MUL31(a[2], b[2])  in mul9()
183 		+ MUL31(a[3], b[1])  in mul9()
184 		+ MUL31(a[4], b[0]);  in mul9()
185 	t[ 5] = MUL31(a[0], b[5])  in mul9()
186 		+ MUL31(a[1], b[4])  in mul9()
187 		+ MUL31(a[2], b[3])  in mul9()
188 		+ MUL31(a[3], b[2])  in mul9()
189 		+ MUL31(a[4], b[1])  in mul9()
190 		+ MUL31(a[5], b[0]);  in mul9()
191 	t[ 6] = MUL31(a[0], b[6])  in mul9()
192 		+ MUL31(a[1], b[5])  in mul9()
193 		+ MUL31(a[2], b[4])  in mul9()
194 		+ MUL31(a[3], b[3])  in mul9()
195 		+ MUL31(a[4], b[2])  in mul9()
196 		+ MUL31(a[5], b[1])  in mul9()
197 		+ MUL31(a[6], b[0]);  in mul9()
198 	t[ 7] = MUL31(a[0], b[7])  in mul9()
199 		+ MUL31(a[1], b[6])  in mul9()
200 		+ MUL31(a[2], b[5])  in mul9()
201 		+ MUL31(a[3], b[4])  in mul9()
202 		+ MUL31(a[4], b[3])  in mul9()
203 		+ MUL31(a[5], b[2])  in mul9()
204 		+ MUL31(a[6], b[1])  in mul9()
205 		+ MUL31(a[7], b[0]);  in mul9()
206 	t[ 8] = MUL31(a[0], b[8])  in mul9()
207 		+ MUL31(a[1], b[7])  in mul9()
208 		+ MUL31(a[2], b[6])  in mul9()
209 		+ MUL31(a[3], b[5])  in mul9()
210 		+ MUL31(a[4], b[4])  in mul9()
211 		+ MUL31(a[5], b[3])  in mul9()
212 		+ MUL31(a[6], b[2])  in mul9()
213 		+ MUL31(a[7], b[1])  in mul9()
214 		+ MUL31(a[8], b[0]);  in mul9()
215 	t[ 9] = MUL31(a[1], b[8])  in mul9()
216 		+ MUL31(a[2], b[7])  in mul9()
217 		+ MUL31(a[3], b[6])  in mul9()
218 		+ MUL31(a[4], b[5])  in mul9()
219 		+ MUL31(a[5], b[4])  in mul9()
220 		+ MUL31(a[6], b[3])  in mul9()
221 		+ MUL31(a[7], b[2])  in mul9()
222 		+ MUL31(a[8], b[1]);  in mul9()
223 	t[10] = MUL31(a[2], b[8])  in mul9()
224 		+ MUL31(a[3], b[7])  in mul9()
225 		+ MUL31(a[4], b[6])  in mul9()
226 		+ MUL31(a[5], b[5])  in mul9()
227 		+ MUL31(a[6], b[4])  in mul9()
228 		+ MUL31(a[7], b[3])  in mul9()
229 		+ MUL31(a[8], b[2]);  in mul9()
230 	t[11] = MUL31(a[3], b[8])  in mul9()
231 		+ MUL31(a[4], b[7])  in mul9()
232 		+ MUL31(a[5], b[6])  in mul9()
233 		+ MUL31(a[6], b[5])  in mul9()
234 		+ MUL31(a[7], b[4])  in mul9()
235 		+ MUL31(a[8], b[3]);  in mul9()
236 	t[12] = MUL31(a[4], b[8])  in mul9()
237 		+ MUL31(a[5], b[7])  in mul9()
238 		+ MUL31(a[6], b[6])  in mul9()
239 		+ MUL31(a[7], b[5])  in mul9()
240 		+ MUL31(a[8], b[4]);  in mul9()
241 	t[13] = MUL31(a[5], b[8])  in mul9()
242 		+ MUL31(a[6], b[7])  in mul9()
243 		+ MUL31(a[7], b[6])  in mul9()
244 		+ MUL31(a[8], b[5]);  in mul9()
245 	t[14] = MUL31(a[6], b[8])  in mul9()
246 		+ MUL31(a[7], b[7])  in mul9()
247 		+ MUL31(a[8], b[6]);  in mul9()
248 	t[15] = MUL31(a[7], b[8])  in mul9()
249 		+ MUL31(a[8], b[7]);  in mul9()
250 	t[16] = MUL31(a[8], b[8]);  in mul9()
277 	t[ 0] = MUL31(a[0], a[0]);  in square9()
278 	t[ 1] = ((MUL31(a[0], a[1])) << 1);  in square9()
279 	t[ 2] = MUL31(a[1], a[1])  in square9()
280 		+ ((MUL31(a[0], a[2])) << 1);  in square9()
281 	t[ 3] = ((MUL31(a[0], a[3])  in square9()
282 		+ MUL31(a[1], a[2])) << 1);  in square9()
283 	t[ 4] = MUL31(a[2], a[2])  in square9()
284 		+ ((MUL31(a[0], a[4])  in square9()
285 		+ MUL31(a[1], a[3])) << 1);  in square9()
286 	t[ 5] = ((MUL31(a[0], a[5])  in square9()
287 		+ MUL31(a[1], a[4])  in square9()
288 		+ MUL31(a[2], a[3])) << 1);  in square9()
289 	t[ 6] = MUL31(a[3], a[3])  in square9()
290 		+ ((MUL31(a[0], a[6])  in square9()
291 		+ MUL31(a[1], a[5])  in square9()
292 		+ MUL31(a[2], a[4])) << 1);  in square9()
293 	t[ 7] = ((MUL31(a[0], a[7])  in square9()
294 		+ MUL31(a[1], a[6])  in square9()
295 		+ MUL31(a[2], a[5])  in square9()
296 		+ MUL31(a[3], a[4])) << 1);  in square9()
297 	t[ 8] = MUL31(a[4], a[4])  in square9()
298 		+ ((MUL31(a[0], a[8])  in square9()
299 		+ MUL31(a[1], a[7])  in square9()
300 		+ MUL31(a[2], a[6])  in square9()
301 		+ MUL31(a[3], a[5])) << 1);  in square9()
302 	t[ 9] = ((MUL31(a[1], a[8])  in square9()
303 		+ MUL31(a[2], a[7])  in square9()
304 		+ MUL31(a[3], a[6])  in square9()
305 		+ MUL31(a[4], a[5])) << 1);  in square9()
306 	t[10] = MUL31(a[5], a[5])  in square9()
307 		+ ((MUL31(a[2], a[8])  in square9()
308 		+ MUL31(a[3], a[7])  in square9()
309 		+ MUL31(a[4], a[6])) << 1);  in square9()
310 	t[11] = ((MUL31(a[3], a[8])  in square9()
311 		+ MUL31(a[4], a[7])  in square9()
312 		+ MUL31(a[5], a[6])) << 1);  in square9()
313 	t[12] = MUL31(a[6], a[6])  in square9()
314 		+ ((MUL31(a[4], a[8])  in square9()
315 		+ MUL31(a[5], a[7])) << 1);  in square9()
316 	t[13] = ((MUL31(a[5], a[8])  in square9()
317 		+ MUL31(a[6], a[7])) << 1);  in square9()
318 	t[14] = MUL31(a[7], a[7])  in square9()
319 		+ ((MUL31(a[6], a[8])) << 1);  in square9()
320 	t[15] = ((MUL31(a[7], a[8])) << 1);  in square9()
321 	t[16] = MUL31(a[8], a[8]);  in square9()
400 		w = (uint64_t)t[i] + (uint64_t)cc + MUL31(t[i + 9], 622592);  in f255_mul()
457 		w = (uint64_t)t[i] + (uint64_t)cc + MUL31(t[i + 9], 622592);  in f255_square()
549 		w = MUL31(a[i], 121665) + (uint64_t)cc;  in f255_mul_a24()
553 	w = MUL31(a[8], 121665) + (uint64_t)cc;  in f255_mul_a24()