163 	__asm__ volatile ("vmovaps  %%ymm0, %0" :"=m" (vec256array[i]));  in store_ymm()
164 	i++; __asm__ volatile ("vmovaps  %%ymm1, %0" :"=m" (vec256array[i]));  in store_ymm()
165 	i++; __asm__ volatile ("vmovaps  %%ymm2, %0" :"=m" (vec256array[i]));  in store_ymm()
166 	i++; __asm__ volatile ("vmovaps  %%ymm3, %0" :"=m" (vec256array[i]));  in store_ymm()
167 	i++; __asm__ volatile ("vmovaps  %%ymm4, %0" :"=m" (vec256array[i]));  in store_ymm()
168 	i++; __asm__ volatile ("vmovaps  %%ymm5, %0" :"=m" (vec256array[i]));  in store_ymm()
169 	i++; __asm__ volatile ("vmovaps  %%ymm6, %0" :"=m" (vec256array[i]));  in store_ymm()
170 	i++; __asm__ volatile ("vmovaps  %%ymm7, %0" :"=m" (vec256array[i]));  in store_ymm()
172 	i++; __asm__ volatile ("vmovaps  %%ymm8, %0" :"=m" (vec256array[i]));  in store_ymm()
173 	i++; __asm__ volatile ("vmovaps  %%ymm9, %0" :"=m" (vec256array[i]));  in store_ymm()
174 	i++; __asm__ volatile ("vmovaps  %%ymm10, %0" :"=m" (vec256array[i]));  in store_ymm()
175 	i++; __asm__ volatile ("vmovaps  %%ymm11, %0" :"=m" (vec256array[i]));  in store_ymm()
176 	i++; __asm__ volatile ("vmovaps  %%ymm12, %0" :"=m" (vec256array[i]));  in store_ymm()
177 	i++; __asm__ volatile ("vmovaps  %%ymm13, %0" :"=m" (vec256array[i]));  in store_ymm()
178 	i++; __asm__ volatile ("vmovaps  %%ymm14, %0" :"=m" (vec256array[i]));  in store_ymm()
179 	i++; __asm__ volatile ("vmovaps  %%ymm15, %0" :"=m" (vec256array[i]));  in store_ymm()
188 	__asm__ volatile ("vmovaps  %0, %%ymm0" :: "m" (*(__m256i*)p) : "ymm0"); p++;  in restore_ymm()
189 	__asm__ volatile ("vmovaps  %0, %%ymm1" :: "m" (*(__m256i*)p) : "ymm1"); p++;  in restore_ymm()
190 	__asm__ volatile ("vmovaps  %0, %%ymm2" :: "m" (*(__m256i*)p) : "ymm2"); p++;  in restore_ymm()
191 	__asm__ volatile ("vmovaps  %0, %%ymm3" :: "m" (*(__m256i*)p) : "ymm3"); p++;  in restore_ymm()
192 	__asm__ volatile ("vmovaps  %0, %%ymm4" :: "m" (*(__m256i*)p) : "ymm4"); p++;  in restore_ymm()
193 	__asm__ volatile ("vmovaps  %0, %%ymm5" :: "m" (*(__m256i*)p) : "ymm5"); p++;  in restore_ymm()
194 	__asm__ volatile ("vmovaps  %0, %%ymm6" :: "m" (*(__m256i*)p) : "ymm6"); p++;  in restore_ymm()
195 	__asm__ volatile ("vmovaps  %0, %%ymm7" :: "m" (*(__m256i*)p) : "ymm7");  in restore_ymm()
198 	++p; __asm__ volatile ("vmovaps  %0, %%ymm8" :: "m" (*(__m256i*)p) : "ymm8"); p++;  in restore_ymm()
199 	__asm__ volatile ("vmovaps  %0, %%ymm9" :: "m" (*(__m256i*)p) : "ymm9"); p++;  in restore_ymm()
200 	__asm__ volatile ("vmovaps  %0, %%ymm10" :: "m" (*(__m256i*)p) : "ymm10"); p++;  in restore_ymm()
201 	__asm__ volatile ("vmovaps  %0, %%ymm11" :: "m" (*(__m256i*)p) : "ymm11"); p++;  in restore_ymm()
202 	__asm__ volatile ("vmovaps  %0, %%ymm12" :: "m" (*(__m256i*)p) : "ymm12"); p++;  in restore_ymm()
203 	__asm__ volatile ("vmovaps  %0, %%ymm13" :: "m" (*(__m256i*)p) : "ymm13"); p++;  in restore_ymm()
204 	__asm__ volatile ("vmovaps  %0, %%ymm14" :: "m" (*(__m256i*)p) : "ymm14"); p++;  in restore_ymm()
205 	__asm__ volatile ("vmovaps  %0, %%ymm15" :: "m" (*(__m256i*)p) : "ymm15");  in restore_ymm()
221 	__asm__ volatile ("vmovaps  %0, %%ymm0" :: "m" (*(__m256i*)p) : "ymm0");  in populate_ymm()
222 	__asm__ volatile ("vmovaps  %0, %%ymm1" :: "m" (*(__m256i*)p) : "ymm1");  in populate_ymm()
223 	__asm__ volatile ("vmovaps  %0, %%ymm2" :: "m" (*(__m256i*)p) : "ymm2");  in populate_ymm()
224 	__asm__ volatile ("vmovaps  %0, %%ymm3" :: "m" (*(__m256i*)p) : "ymm3");  in populate_ymm()
228 	__asm__ volatile ("vmovaps  %0, %%ymm4" :: "m" (*(__m256i*)p) : "ymm4");  in populate_ymm()
229 	__asm__ volatile ("vmovaps  %0, %%ymm5" :: "m" (*(__m256i*)p) : "ymm5");  in populate_ymm()
230 	__asm__ volatile ("vmovaps  %0, %%ymm6" :: "m" (*(__m256i*)p) : "ymm6");  in populate_ymm()
231 	__asm__ volatile ("vmovaps  %0, %%ymm7" :: "m" (*(__m256i*)p) : "ymm7");  in populate_ymm()
236 	__asm__ volatile ("vmovaps  %0, %%ymm8" :: "m" (*(__m256i*)p) : "ymm8");  in populate_ymm()
237 	__asm__ volatile ("vmovaps  %0, %%ymm9" :: "m" (*(__m256i*)p) : "ymm9");  in populate_ymm()
238 	__asm__ volatile ("vmovaps  %0, %%ymm10" :: "m" (*(__m256i*)p) : "ymm10");  in populate_ymm()
239 	__asm__ volatile ("vmovaps  %0, %%ymm11" :: "m" (*(__m256i*)p) : "ymm11");  in populate_ymm()
243 	__asm__ volatile ("vmovaps  %0, %%ymm12" :: "m" (*(__m256i*)p) : "ymm12");  in populate_ymm()
244 	__asm__ volatile ("vmovaps  %0, %%ymm13" :: "m" (*(__m256i*)p) : "ymm13");  in populate_ymm()
245 	__asm__ volatile ("vmovaps  %0, %%ymm14" :: "m" (*(__m256i*)p) : "ymm14");  in populate_ymm()
246 	__asm__ volatile ("vmovaps  %0, %%ymm15" :: "m" (*(__m256i*)p) : "ymm15");  in populate_ymm()
434 	__asm__ volatile ("kmovq %%k0, %0" :"=m" (k[0]));  in store_opmask()
435 	__asm__ volatile ("kmovq %%k1, %0" :"=m" (k[1]));  in store_opmask()
436 	__asm__ volatile ("kmovq %%k2, %0" :"=m" (k[2]));  in store_opmask()
437 	__asm__ volatile ("kmovq %%k3, %0" :"=m" (k[3]));  in store_opmask()
438 	__asm__ volatile ("kmovq %%k4, %0" :"=m" (k[4]));  in store_opmask()
439 	__asm__ volatile ("kmovq %%k5, %0" :"=m" (k[5]));  in store_opmask()
440 	__asm__ volatile ("kmovq %%k6, %0" :"=m" (k[6]));  in store_opmask()
441 	__asm__ volatile ("kmovq %%k7, %0" :"=m" (k[7]));  in store_opmask()
448 	__asm__ volatile ("vmovaps  %%zmm0, %0" :"=m" (vecarray[i]));  in store_zmm()
449 	i++; __asm__ volatile ("vmovaps  %%zmm1, %0" :"=m" (vecarray[i]));  in store_zmm()
450 	i++; __asm__ volatile ("vmovaps  %%zmm2, %0" :"=m" (vecarray[i]));  in store_zmm()
451 	i++; __asm__ volatile ("vmovaps  %%zmm3, %0" :"=m" (vecarray[i]));  in store_zmm()
452 	i++; __asm__ volatile ("vmovaps  %%zmm4, %0" :"=m" (vecarray[i]));  in store_zmm()
453 	i++; __asm__ volatile ("vmovaps  %%zmm5, %0" :"=m" (vecarray[i]));  in store_zmm()
454 	i++; __asm__ volatile ("vmovaps  %%zmm6, %0" :"=m" (vecarray[i]));  in store_zmm()
455 	i++; __asm__ volatile ("vmovaps  %%zmm7, %0" :"=m" (vecarray[i]));  in store_zmm()
457 	i++; __asm__ volatile ("vmovaps  %%zmm8, %0" :"=m" (vecarray[i]));  in store_zmm()
458 	i++; __asm__ volatile ("vmovaps  %%zmm9, %0" :"=m" (vecarray[i]));  in store_zmm()
459 	i++; __asm__ volatile ("vmovaps  %%zmm10, %0" :"=m" (vecarray[i]));  in store_zmm()
460 	i++; __asm__ volatile ("vmovaps  %%zmm11, %0" :"=m" (vecarray[i]));  in store_zmm()
461 	i++; __asm__ volatile ("vmovaps  %%zmm12, %0" :"=m" (vecarray[i]));  in store_zmm()
462 	i++; __asm__ volatile ("vmovaps  %%zmm13, %0" :"=m" (vecarray[i]));  in store_zmm()
463 	i++; __asm__ volatile ("vmovaps  %%zmm14, %0" :"=m" (vecarray[i]));  in store_zmm()
464 	i++; __asm__ volatile ("vmovaps  %%zmm15, %0" :"=m" (vecarray[i]));  in store_zmm()
465 	i++; __asm__ volatile ("vmovaps  %%zmm16, %0" :"=m" (vecarray[i]));  in store_zmm()
466 	i++; __asm__ volatile ("vmovaps  %%zmm17, %0" :"=m" (vecarray[i]));  in store_zmm()
467 	i++; __asm__ volatile ("vmovaps  %%zmm18, %0" :"=m" (vecarray[i]));  in store_zmm()
468 	i++; __asm__ volatile ("vmovaps  %%zmm19, %0" :"=m" (vecarray[i]));  in store_zmm()
469 	i++; __asm__ volatile ("vmovaps  %%zmm20, %0" :"=m" (vecarray[i]));  in store_zmm()
470 	i++; __asm__ volatile ("vmovaps  %%zmm21, %0" :"=m" (vecarray[i]));  in store_zmm()
471 	i++; __asm__ volatile ("vmovaps  %%zmm22, %0" :"=m" (vecarray[i]));  in store_zmm()
472 	i++; __asm__ volatile ("vmovaps  %%zmm23, %0" :"=m" (vecarray[i]));  in store_zmm()
473 	i++; __asm__ volatile ("vmovaps  %%zmm24, %0" :"=m" (vecarray[i]));  in store_zmm()
474 	i++; __asm__ volatile ("vmovaps  %%zmm25, %0" :"=m" (vecarray[i]));  in store_zmm()
475 	i++; __asm__ volatile ("vmovaps  %%zmm26, %0" :"=m" (vecarray[i]));  in store_zmm()
476 	i++; __asm__ volatile ("vmovaps  %%zmm27, %0" :"=m" (vecarray[i]));  in store_zmm()
477 	i++; __asm__ volatile ("vmovaps  %%zmm28, %0" :"=m" (vecarray[i]));  in store_zmm()
478 	i++; __asm__ volatile ("vmovaps  %%zmm29, %0" :"=m" (vecarray[i]));  in store_zmm()
479 	i++; __asm__ volatile ("vmovaps  %%zmm30, %0" :"=m" (vecarray[i]));  in store_zmm()
480 	i++; __asm__ volatile ("vmovaps  %%zmm31, %0" :"=m" (vecarray[i]));  in store_zmm()
489 	__asm__ volatile ("vmovaps  %0, %%zmm0" :: "m" (*(__m512i*)p) : "zmm0"); p++;  in restore_zmm()
490 	__asm__ volatile ("vmovaps  %0, %%zmm1" :: "m" (*(__m512i*)p) : "zmm1"); p++;  in restore_zmm()
491 	__asm__ volatile ("vmovaps  %0, %%zmm2" :: "m" (*(__m512i*)p) : "zmm2"); p++;  in restore_zmm()
492 	__asm__ volatile ("vmovaps  %0, %%zmm3" :: "m" (*(__m512i*)p) : "zmm3"); p++;  in restore_zmm()
493 	__asm__ volatile ("vmovaps  %0, %%zmm4" :: "m" (*(__m512i*)p) : "zmm4"); p++;  in restore_zmm()
494 	__asm__ volatile ("vmovaps  %0, %%zmm5" :: "m" (*(__m512i*)p) : "zmm5"); p++;  in restore_zmm()
495 	__asm__ volatile ("vmovaps  %0, %%zmm6" :: "m" (*(__m512i*)p) : "zmm6"); p++;  in restore_zmm()
496 	__asm__ volatile ("vmovaps  %0, %%zmm7" :: "m" (*(__m512i*)p) : "zmm7");  in restore_zmm()
499 	++p; __asm__ volatile ("vmovaps  %0, %%zmm8" :: "m" (*(__m512i*)p) : "zmm8"); p++;  in restore_zmm()
500 	__asm__ volatile ("vmovaps  %0, %%zmm9" :: "m" (*(__m512i*)p) : "zmm9"); p++;  in restore_zmm()
501 	__asm__ volatile ("vmovaps  %0, %%zmm10" :: "m" (*(__m512i*)p) : "zmm10"); p++;  in restore_zmm()
502 	__asm__ volatile ("vmovaps  %0, %%zmm11" :: "m" (*(__m512i*)p) : "zmm11"); p++;  in restore_zmm()
503 	__asm__ volatile ("vmovaps  %0, %%zmm12" :: "m" (*(__m512i*)p) : "zmm12"); p++;  in restore_zmm()
504 	__asm__ volatile ("vmovaps  %0, %%zmm13" :: "m" (*(__m512i*)p) : "zmm13"); p++;  in restore_zmm()
505 	__asm__ volatile ("vmovaps  %0, %%zmm14" :: "m" (*(__m512i*)p) : "zmm14"); p++;  in restore_zmm()
506 	__asm__ volatile ("vmovaps  %0, %%zmm15" :: "m" (*(__m512i*)p) : "zmm15"); p++;  in restore_zmm()
507 	__asm__ volatile ("vmovaps  %0, %%zmm16" :: "m" (*(__m512i*)p) : "zmm16"); p++;  in restore_zmm()
508 	__asm__ volatile ("vmovaps  %0, %%zmm17" :: "m" (*(__m512i*)p) : "zmm17"); p++;  in restore_zmm()
509 	__asm__ volatile ("vmovaps  %0, %%zmm18" :: "m" (*(__m512i*)p) : "zmm18"); p++;  in restore_zmm()
510 	__asm__ volatile ("vmovaps  %0, %%zmm19" :: "m" (*(__m512i*)p) : "zmm19"); p++;  in restore_zmm()
511 	__asm__ volatile ("vmovaps  %0, %%zmm20" :: "m" (*(__m512i*)p) : "zmm20"); p++;  in restore_zmm()
512 	__asm__ volatile ("vmovaps  %0, %%zmm21" :: "m" (*(__m512i*)p) : "zmm21"); p++;  in restore_zmm()
513 	__asm__ volatile ("vmovaps  %0, %%zmm22" :: "m" (*(__m512i*)p) : "zmm22"); p++;  in restore_zmm()
514 	__asm__ volatile ("vmovaps  %0, %%zmm23" :: "m" (*(__m512i*)p) : "zmm23"); p++;  in restore_zmm()
515 	__asm__ volatile ("vmovaps  %0, %%zmm24" :: "m" (*(__m512i*)p) : "zmm24"); p++;  in restore_zmm()
516 	__asm__ volatile ("vmovaps  %0, %%zmm25" :: "m" (*(__m512i*)p) : "zmm25"); p++;  in restore_zmm()
517 	__asm__ volatile ("vmovaps  %0, %%zmm26" :: "m" (*(__m512i*)p) : "zmm26"); p++;  in restore_zmm()
518 	__asm__ volatile ("vmovaps  %0, %%zmm27" :: "m" (*(__m512i*)p) : "zmm27"); p++;  in restore_zmm()
519 	__asm__ volatile ("vmovaps  %0, %%zmm28" :: "m" (*(__m512i*)p) : "zmm28"); p++;  in restore_zmm()
520 	__asm__ volatile ("vmovaps  %0, %%zmm29" :: "m" (*(__m512i*)p) : "zmm29"); p++;  in restore_zmm()
521 	__asm__ volatile ("vmovaps  %0, %%zmm30" :: "m" (*(__m512i*)p) : "zmm30"); p++;  in restore_zmm()
522 	__asm__ volatile ("vmovaps  %0, %%zmm31" :: "m" (*(__m512i*)p) : "zmm31");  in restore_zmm()
531 	__asm__ volatile ("kmovq %0, %%k0" : :"m" (zero) : "k0");  in zero_opmask()
532 	__asm__ volatile ("kmovq %0, %%k1" : :"m" (zero) : "k1");  in zero_opmask()
533 	__asm__ volatile ("kmovq %0, %%k2" : :"m" (zero) : "k2");  in zero_opmask()
534 	__asm__ volatile ("kmovq %0, %%k3" : :"m" (zero) : "k3");  in zero_opmask()
535 	__asm__ volatile ("kmovq %0, %%k4" : :"m" (zero) : "k4");  in zero_opmask()
536 	__asm__ volatile ("kmovq %0, %%k5" : :"m" (zero) : "k5");  in zero_opmask()
537 	__asm__ volatile ("kmovq %0, %%k6" : :"m" (zero) : "k6");  in zero_opmask()
538 	__asm__ volatile ("kmovq %0, %%k7" : :"m" (zero) : "k7");  in zero_opmask()
551 	__asm__ volatile ("kmovq %0, %%k0" : :"m" (k[0]) : "k0");  in populate_opmask()
552 	__asm__ volatile ("kmovq %0, %%k1" : :"m" (k[1]) : "k1");  in populate_opmask()
553 	__asm__ volatile ("kmovq %0, %%k2" : :"m" (k[2]) : "k2");  in populate_opmask()
554 	__asm__ volatile ("kmovq %0, %%k3" : :"m" (k[3]) : "k3");  in populate_opmask()
555 	__asm__ volatile ("kmovq %0, %%k4" : :"m" (k[4]) : "k4");  in populate_opmask()
556 	__asm__ volatile ("kmovq %0, %%k5" : :"m" (k[5]) : "k5");  in populate_opmask()
557 	__asm__ volatile ("kmovq %0, %%k6" : :"m" (k[6]) : "k6");  in populate_opmask()
558 	__asm__ volatile ("kmovq %0, %%k7" : :"m" (k[7]) : "k7");  in populate_opmask()
597 	__asm__ volatile ("vmovaps  %0, %%zmm0" :: "m" (zero) : "zmm0");  in zero_zmm()
598 	__asm__ volatile ("vmovaps  %0, %%zmm1" :: "m" (zero) : "zmm1");  in zero_zmm()
599 	__asm__ volatile ("vmovaps  %0, %%zmm2" :: "m" (zero) : "zmm2");  in zero_zmm()
600 	__asm__ volatile ("vmovaps  %0, %%zmm3" :: "m" (zero) : "zmm3");  in zero_zmm()
601 	__asm__ volatile ("vmovaps  %0, %%zmm4" :: "m" (zero) : "zmm4");  in zero_zmm()
602 	__asm__ volatile ("vmovaps  %0, %%zmm5" :: "m" (zero) : "zmm5");  in zero_zmm()
603 	__asm__ volatile ("vmovaps  %0, %%zmm6" :: "m" (zero) : "zmm6");  in zero_zmm()
604 	__asm__ volatile ("vmovaps  %0, %%zmm7" :: "m" (zero) : "zmm7");  in zero_zmm()
607 	__asm__ volatile ("vmovaps  %0, %%zmm8" :: "m" (zero) : "zmm8");  in zero_zmm()
608 	__asm__ volatile ("vmovaps  %0, %%zmm9" :: "m" (zero) : "zmm9");  in zero_zmm()
609 	__asm__ volatile ("vmovaps  %0, %%zmm10" :: "m" (zero) : "zmm10");  in zero_zmm()
610 	__asm__ volatile ("vmovaps  %0, %%zmm11" :: "m" (zero) : "zmm11");  in zero_zmm()
611 	__asm__ volatile ("vmovaps  %0, %%zmm12" :: "m" (zero) : "zmm12");  in zero_zmm()
612 	__asm__ volatile ("vmovaps  %0, %%zmm13" :: "m" (zero) : "zmm13");  in zero_zmm()
613 	__asm__ volatile ("vmovaps  %0, %%zmm14" :: "m" (zero) : "zmm14");  in zero_zmm()
614 	__asm__ volatile ("vmovaps  %0, %%zmm15" :: "m" (zero) : "zmm15");  in zero_zmm()
615 	__asm__ volatile ("vmovaps  %0, %%zmm16" :: "m" (zero) : "zmm16");  in zero_zmm()
616 	__asm__ volatile ("vmovaps  %0, %%zmm17" :: "m" (zero) : "zmm17");  in zero_zmm()
617 	__asm__ volatile ("vmovaps  %0, %%zmm18" :: "m" (zero) : "zmm18");  in zero_zmm()
618 	__asm__ volatile ("vmovaps  %0, %%zmm19" :: "m" (zero) : "zmm19");  in zero_zmm()
619 	__asm__ volatile ("vmovaps  %0, %%zmm20" :: "m" (zero) : "zmm20");  in zero_zmm()
620 	__asm__ volatile ("vmovaps  %0, %%zmm21" :: "m" (zero) : "zmm21");  in zero_zmm()
621 	__asm__ volatile ("vmovaps  %0, %%zmm22" :: "m" (zero) : "zmm22");  in zero_zmm()
622 	__asm__ volatile ("vmovaps  %0, %%zmm23" :: "m" (zero) : "zmm23");  in zero_zmm()
623 	__asm__ volatile ("vmovaps  %0, %%zmm24" :: "m" (zero) : "zmm24");  in zero_zmm()
624 	__asm__ volatile ("vmovaps  %0, %%zmm25" :: "m" (zero) : "zmm25");  in zero_zmm()
625 	__asm__ volatile ("vmovaps  %0, %%zmm26" :: "m" (zero) : "zmm26");  in zero_zmm()
626 	__asm__ volatile ("vmovaps  %0, %%zmm27" :: "m" (zero) : "zmm27");  in zero_zmm()
627 	__asm__ volatile ("vmovaps  %0, %%zmm28" :: "m" (zero) : "zmm28");  in zero_zmm()
628 	__asm__ volatile ("vmovaps  %0, %%zmm29" :: "m" (zero) : "zmm29");  in zero_zmm()
629 	__asm__ volatile ("vmovaps  %0, %%zmm30" :: "m" (zero) : "zmm30");  in zero_zmm()
630 	__asm__ volatile ("vmovaps  %0, %%zmm31" :: "m" (zero) : "zmm31");  in zero_zmm()
650 	__asm__ volatile ("vmovaps  %0, %%zmm0" :: "m" (*(__m512i*)p) : "zmm0");  in populate_zmm()
651 	__asm__ volatile ("vmovaps  %0, %%zmm1" :: "m" (*(__m512i*)p) : "zmm1");  in populate_zmm()
652 	__asm__ volatile ("vmovaps  %0, %%zmm2" :: "m" (*(__m512i*)p) : "zmm2");  in populate_zmm()
653 	__asm__ volatile ("vmovaps  %0, %%zmm3" :: "m" (*(__m512i*)p) : "zmm3");  in populate_zmm()
654 	__asm__ volatile ("vmovaps  %0, %%zmm4" :: "m" (*(__m512i*)p) : "zmm4");  in populate_zmm()
655 	__asm__ volatile ("vmovaps  %0, %%zmm5" :: "m" (*(__m512i*)p) : "zmm5");  in populate_zmm()
656 	__asm__ volatile ("vmovaps  %0, %%zmm6" :: "m" (*(__m512i*)p) : "zmm6");  in populate_zmm()
657 	__asm__ volatile ("vmovaps  %0, %%zmm7" :: "m" (*(__m512i*)p) : "zmm7");  in populate_zmm()
664 	__asm__ volatile ("vmovaps  %0, %%zmm8" :: "m" (*(__m512i*)p) : "zmm8");  in populate_zmm()
665 	__asm__ volatile ("vmovaps  %0, %%zmm9" :: "m" (*(__m512i*)p) : "zmm9");  in populate_zmm()
666 	__asm__ volatile ("vmovaps  %0, %%zmm10" :: "m" (*(__m512i*)p) : "zmm10");  in populate_zmm()
667 	__asm__ volatile ("vmovaps  %0, %%zmm11" :: "m" (*(__m512i*)p) : "zmm11");  in populate_zmm()
668 	__asm__ volatile ("vmovaps  %0, %%zmm12" :: "m" (*(__m512i*)p) : "zmm12");  in populate_zmm()
669 	__asm__ volatile ("vmovaps  %0, %%zmm13" :: "m" (*(__m512i*)p) : "zmm13");  in populate_zmm()
670 	__asm__ volatile ("vmovaps  %0, %%zmm14" :: "m" (*(__m512i*)p) : "zmm14");  in populate_zmm()
671 	__asm__ volatile ("vmovaps  %0, %%zmm15" :: "m" (*(__m512i*)p) : "zmm15");  in populate_zmm()
677 	__asm__ volatile ("vmovaps  %0, %%zmm16" :: "m" (*(__m512i*)p) : "zmm16");  in populate_zmm()
678 	__asm__ volatile ("vmovaps  %0, %%zmm17" :: "m" (*(__m512i*)p) : "zmm17");  in populate_zmm()
679 	__asm__ volatile ("vmovaps  %0, %%zmm18" :: "m" (*(__m512i*)p) : "zmm18");  in populate_zmm()
680 	__asm__ volatile ("vmovaps  %0, %%zmm19" :: "m" (*(__m512i*)p) : "zmm19");  in populate_zmm()
681 	__asm__ volatile ("vmovaps  %0, %%zmm20" :: "m" (*(__m512i*)p) : "zmm20");  in populate_zmm()
682 	__asm__ volatile ("vmovaps  %0, %%zmm21" :: "m" (*(__m512i*)p) : "zmm21");  in populate_zmm()
683 	__asm__ volatile ("vmovaps  %0, %%zmm22" :: "m" (*(__m512i*)p) : "zmm22");  in populate_zmm()
684 	__asm__ volatile ("vmovaps  %0, %%zmm23" :: "m" (*(__m512i*)p) : "zmm23");  in populate_zmm()
690 	__asm__ volatile ("vmovaps  %0, %%zmm24" :: "m" (*(__m512i*)p) : "zmm24");  in populate_zmm()
691 	__asm__ volatile ("vmovaps  %0, %%zmm25" :: "m" (*(__m512i*)p) : "zmm25");  in populate_zmm()
692 	__asm__ volatile ("vmovaps  %0, %%zmm26" :: "m" (*(__m512i*)p) : "zmm26");  in populate_zmm()
693 	__asm__ volatile ("vmovaps  %0, %%zmm27" :: "m" (*(__m512i*)p) : "zmm27");  in populate_zmm()
694 	__asm__ volatile ("vmovaps  %0, %%zmm28" :: "m" (*(__m512i*)p) : "zmm28");  in populate_zmm()
695 	__asm__ volatile ("vmovaps  %0, %%zmm29" :: "m" (*(__m512i*)p) : "zmm29");  in populate_zmm()
696 	__asm__ volatile ("vmovaps  %0, %%zmm30" :: "m" (*(__m512i*)p) : "zmm30");  in populate_zmm()
697 	__asm__ volatile ("vmovaps  %0, %%zmm31" :: "m" (*(__m512i*)p) : "zmm31");  in populate_zmm()