Lines Matching full:i
20 int i; \
27 for (i = 0; i < oprsz / (BIT / 8); i++) { \
28 Vd->E1(i) = DO_OP((TD)Vj->E2(2 * i + 1), (TD)Vk->E2(2 * i)); \
38 int i; in HELPER() local
44 for (i = 0; i < oprsz / 16 ; i++) { in HELPER()
45 Vd->Q(i) = int128_add(int128_makes64(Vj->D(2 * i + 1)), in HELPER()
46 int128_makes64(Vk->D(2 * i))); in HELPER()
56 int i; in HELPER() local
62 for (i = 0; i < oprsz / 16; i++) { in HELPER()
63 Vd->Q(i) = int128_sub(int128_makes64(Vj->D(2 * i + 1)), in HELPER()
64 int128_makes64(Vk->D(2 * i))); in HELPER()
74 int i; in HELPER() local
80 for (i = 0; i < oprsz / 16; i ++) { in HELPER()
81 Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i + 1)), in HELPER()
82 int128_make64(Vk->UD(2 * i))); in HELPER()
92 int i; in HELPER() local
98 for (i = 0; i < oprsz / 16; i++) { in HELPER()
99 Vd->Q(i) = int128_sub(int128_make64(Vj->UD(2 * i + 1)), in HELPER()
100 int128_make64(Vk->UD(2 * i))); in HELPER()
107 int i; \
114 for (i = 0; i < oprsz / (BIT / 8); i++) { \
115 Vd->E1(i) = DO_OP((TD)Vj->E2(2 * i) ,(TD)Vk->E2(2 * i)); \
122 int i; \
129 for (i = 0; i < oprsz / (BIT / 8); i++) { \
130 Vd->E1(i) = DO_OP((TD)Vj->E2(2 * i + 1), (TD)Vk->E2(2 * i + 1)); \
136 int i; in HELPER() local
142 for (i = 0; i < oprsz / 16; i++) { in HELPER()
143 Vd->Q(i) = int128_add(int128_makes64(Vj->D(2 * i)), in HELPER()
144 int128_makes64(Vk->D(2 * i))); in HELPER()
154 int i; in HELPER() local
160 for (i = 0; i < oprsz / 16; i++) { in HELPER()
161 Vd->Q(i) = int128_add(int128_makes64(Vj->D(2 * i +1)), in HELPER()
162 int128_makes64(Vk->D(2 * i +1))); in HELPER()
172 int i; in HELPER() local
178 for (i = 0; i < oprsz / 16; i++) { in HELPER()
179 Vd->Q(i) = int128_sub(int128_makes64(Vj->D(2 * i)), in HELPER()
180 int128_makes64(Vk->D(2 * i))); in HELPER()
190 int i; in HELPER() local
196 for (i = 0; i < oprsz / 16; i++) { in HELPER()
197 Vd->Q(i) = int128_sub(int128_makes64(Vj->D(2 * i + 1)), in HELPER()
198 int128_makes64(Vk->D(2 * i + 1))); in HELPER()
208 int i; in HELPER() local
214 for (i = 0; i < oprsz / 16; i++) { in HELPER()
215 Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i)), in HELPER()
216 int128_make64(Vk->UD(2 * i))); in HELPER()
226 int i; in HELPER() local
232 for (i = 0; i < oprsz / 16; i++) { in HELPER()
233 Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i + 1)), in HELPER()
234 int128_make64(Vk->UD(2 * i + 1))); in HELPER()
244 int i; in HELPER() local
250 for (i = 0; i < oprsz / 16; i++) { in HELPER()
251 Vd->Q(i) = int128_sub(int128_make64(Vj->UD(2 * i)), in HELPER()
252 int128_make64(Vk->UD(2 * i))); in HELPER()
262 int i; in HELPER() local
268 for (i = 0; i < oprsz / 16; i++) { in HELPER()
269 Vd->Q(i) = int128_sub(int128_make64(Vj->UD(2 * i + 1)), in HELPER()
270 int128_make64(Vk->UD(2 * i + 1))); in HELPER()
281 int i; \
289 for (i = 0; i < oprsz / (BIT / 8); i++) { \
290 Vd->ES1(i) = DO_OP((TDU)Vj->EU2(2 * i) ,(TDS)Vk->ES2(2 * i)); \
297 int i; \
305 for (i = 0; i < oprsz / (BIT / 8); i++) { \
306 Vd->ES1(i) = DO_OP((TDU)Vj->EU2(2 * i + 1), (TDS)Vk->ES2(2 * i + 1)); \
312 int i; in HELPER() local
318 for (i = 0; i < oprsz / 16; i++) { in HELPER()
319 Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i)), in HELPER()
320 int128_makes64(Vk->D(2 * i))); in HELPER()
330 int i; in HELPER() local
336 for (i = 0; i < oprsz / 16; i++) { in HELPER()
337 Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i + 1)), in HELPER()
338 int128_makes64(Vk->D(2 * i + 1))); in HELPER()
349 int i; \
355 for (i = 0; i < oprsz / (BIT / 8); i++) { \
356 Vd->E(i) = DO_OP(Vj->E(i), Vk->E(i)); \
389 int i; \
395 for (i = 0; i < oprsz / (BIT / 8); i++) { \
396 Vd->E(i) = DO_VABS(Vj->E(i)) + DO_VABS(Vk->E(i)); \
408 int i; \
414 for (i = 0; i < oprsz / (BIT / 8); i++) { \
415 Vd->E(i) = DO_OP(Vj->E(i), (TD)imm); \
439 int i; \
446 for (i = 0; i < oprsz / (BIT / 8); i++) { \
447 Vd->E2(i) = ((T)Vj->E2(i)) * ((T)Vk->E2(i)) >> BIT; \
453 int i; in HELPER() local
460 for (i = 0; i < oprsz / 8; i++) { in HELPER()
461 muls64(&l, &h, Vj->D(i), Vk->D(i)); in HELPER()
462 Vd->D(i) = h; in HELPER()
472 int i; in HELPER() local
479 for (i = 0; i < oprsz / 8; i++) { in HELPER()
480 mulu64(&l, &h, Vj->D(i), Vk->D(i)); in HELPER()
481 Vd->D(i) = h; in HELPER()
516 int i; \
522 for (i = 0; i < oprsz / (BIT / 8); i++) { \
523 Vd->E(i) = DO_OP(Vd->E(i), Vj->E(i) ,Vk->E(i)); \
539 int i; \
546 for (i = 0; i < oprsz / (BIT / 8); i++) { \
547 Vd->E1(i) += DO_OP((TD)Vj->E2(2 * i), (TD)Vk->E2(2 * i)); \
561 int i; \
568 for (i = 0; i < oprsz / (BIT / 8); i++) { \
569 Vd->E1(i) += DO_OP((TD)Vj->E2(2 * i + 1), \
570 (TD)Vk->E2(2 * i + 1)); \
584 int i; \
592 for (i = 0; i < oprsz / (BIT / 8); i++) { \
593 Vd->ES1(i) += DO_OP((TU1)Vj->EU2(2 * i), \
594 (TS1)Vk->ES2(2 * i)); \
605 int i; \
613 for (i = 0; i < oprsz / (BIT / 8); i++) { \
614 Vd->ES1(i) += DO_OP((TU1)Vj->EU2(2 * i + 1), \
615 (TS1)Vk->ES2(2 * i + 1)); \
626 int i; \
632 for (i = 0; i < oprsz / (BIT / 8); i++) { \
633 Vd->E(i) = DO_OP(Vj->E(i), Vk->E(i)); \
657 int i; \
663 for (i = 0; i < oprsz / (BIT / 8); i++) { \
664 Vd->E(i) = Vj->E(i) > (TD)max ? (TD)max : \
665 Vj->E(i) < (TD)~max ? (TD)~max: Vj->E(i); \
677 int i; \
683 for (i = 0; i < oprsz / (BIT / 8); i++) { \
684 Vd->E(i) = Vj->E(i) > (TD)max ? (TD)max : Vj->E(i); \
696 int i, j, ofs; \
702 for (i = 0; i < oprsz / 16; i++) { \
704 Vd->E1(j + i * ofs) = Vj->E2(j + ofs + ofs * 2 * i); \
711 int i; in HELPER() local
716 for (i = 0; i < oprsz / 16; i++) { in HELPER()
717 Vd->Q(i) = int128_makes64(Vj->D(2 * i + 1)); in HELPER()
723 int i; in HELPER() local
728 for (i = 0; i < oprsz / 16; i++) { in HELPER()
729 Vd->Q(i) = int128_make64(Vj->UD(2 * i + 1)); in HELPER()
743 int i; \
749 for (i = 0; i < oprsz / (BIT / 8); i++) { \
750 temp.E1(i) = Vj->E2(i); \
785 int i; in HELPER() local
791 for (i = 0; i < oprsz / 16; i++) { in HELPER()
793 temp = do_vmskltz_b(Vj->D(2 * i)); in HELPER()
794 temp |= (do_vmskltz_b(Vj->D(2 * i + 1)) << 8); in HELPER()
795 Vd->D(2 * i) = temp; in HELPER()
796 Vd->D(2 * i + 1) = 0; in HELPER()
811 int i; in HELPER() local
817 for (i = 0; i < oprsz / 16; i++) { in HELPER()
819 temp = do_vmskltz_h(Vj->D(2 * i)); in HELPER()
820 temp |= (do_vmskltz_h(Vj->D(2 * i + 1)) << 4); in HELPER()
821 Vd->D(2 * i) = temp; in HELPER()
822 Vd->D(2 * i + 1) = 0; in HELPER()
836 int i; in HELPER() local
842 for (i = 0; i < oprsz / 16; i++) { in HELPER()
844 temp = do_vmskltz_w(Vj->D(2 * i)); in HELPER()
845 temp |= (do_vmskltz_w(Vj->D(2 * i + 1)) << 2); in HELPER()
846 Vd->D(2 * i) = temp; in HELPER()
847 Vd->D(2 * i + 1) = 0; in HELPER()
857 int i; in HELPER() local
863 for (i = 0; i < oprsz / 16; i++) { in HELPER()
865 temp = do_vmskltz_d(Vj->D(2 * i)); in HELPER()
866 temp |= (do_vmskltz_d(Vj->D(2 * i + 1)) << 1); in HELPER()
867 Vd->D(2 * i) = temp; in HELPER()
868 Vd->D(2 * i + 1) = 0; in HELPER()
874 int i; in HELPER() local
880 for (i = 0; i < oprsz / 16; i++) { in HELPER()
882 temp = do_vmskltz_b(Vj->D(2 * i)); in HELPER()
883 temp |= (do_vmskltz_b(Vj->D(2 * i + 1)) << 8); in HELPER()
884 Vd->D(2 * i) = (uint16_t)(~temp); in HELPER()
885 Vd->D(2 * i + 1) = 0; in HELPER()
901 int i; in HELPER() local
907 for (i = 0; i < oprsz / 16; i++) { in HELPER()
909 temp = do_vmskez_b(Vj->D(2 * i)); in HELPER()
910 temp |= (do_vmskez_b(Vj->D(2 * i + 1)) << 8); in HELPER()
911 Vd->D(2 * i) = (uint16_t)(~temp); in HELPER()
912 Vd->D(2 * i + 1) = 0; in HELPER()
918 int i; in HELPER() local
922 for (i = 0; i < simd_oprsz(desc); i++) { in HELPER()
923 Vd->B(i) = ~(Vj->B(i) | (uint8_t)imm); in HELPER()
930 int i, j, ofs; \
938 for (i = 0; i < oprsz / 16; i++) { \
940 temp.E1(j + ofs * i) = (TD)Vj->E2(j + ofs * 2 * i) << (imm % BIT); \
949 int i; in HELPER() local
954 for (i = 0; i < oprsz / 16; i++) { in HELPER()
955 Vd->Q(i) = int128_makes64(Vj->D(2 * i)); in HELPER()
961 int i; in HELPER() local
966 for (i = 0; i < oprsz / 16; i++) { in HELPER()
967 Vd->Q(i) = int128_make64(Vj->UD(2 * i)); in HELPER()
996 int i; \ in do_vsrlr()
1002 for (i = 0; i < oprsz / (BIT / 8); i++) { \ in do_vsrlr()
1003 Vd->E(i) = do_vsrlr_ ## E(Vj->E(i), ((T)Vk->E(i))%BIT); \ in do_vsrlr()
1015 int i; \
1020 for (i = 0; i < oprsz / (BIT / 8); i++) { \
1021 Vd->E(i) = do_vsrlr_ ## E(Vj->E(i), imm); \
1048 int i; \
1054 for (i = 0; i < oprsz / (BIT / 8); i++) { \
1055 Vd->E(i) = do_vsrar_ ## E(Vj->E(i), ((T)Vk->E(i))%BIT); \
1067 int i; \
1072 for (i = 0; i < oprsz / (BIT / 8); i++) { \
1073 Vd->E(i) = do_vsrar_ ## E(Vj->E(i), imm); \
1085 int i, j, ofs; \
1092 for (i = 0; i < oprsz / 16; i++) { \
1094 Vd->E1(j + ofs * 2 * i) = R_SHIFT(Vj->E2(j + ofs * i), \
1095 Vk->E2(j + ofs * i) % BIT); \
1097 Vd->D(2 * i + 1) = 0; \
1108 int i, j, ofs; \
1115 for (i = 0; i < oprsz / 16; i++) { \
1117 Vd->E1(j + ofs * 2 * i) = R_SHIFT(Vj->E2(j + ofs * i), \
1118 Vk->E3(j + ofs * i) % BIT); \
1120 Vd->D(2 * i + 1) = 0; \
1131 int i, j, ofs; \
1138 for (i = 0; i < oprsz / 16; i++) { \
1140 temp.E1(j + ofs * 2 * i) = R_SHIFT(Vj->E2(j + ofs * i), imm); \
1141 temp.E1(j + ofs * (2 * i + 1)) = R_SHIFT(Vd->E2(j + ofs * i), \
1150 int i; local
1155 for (i = 0; i < 2; i++) {
1156 temp.D(2 * i) = int128_getlo(int128_urshift(Vj->Q(i), imm % 128));
1157 temp.D(2 * i +1) = int128_getlo(int128_urshift(Vd->Q(i), imm % 128));
1169 int i, j, ofs; \
1176 for (i = 0; i < oprsz / 16; i++) { \
1178 temp.E1(j + ofs * 2 * i) = R_SHIFT(Vj->E2(j + ofs * i), imm); \
1179 temp.E1(j + ofs * (2 * i + 1)) = R_SHIFT(Vd->E2(j + ofs * i), \
1188 int i; in HELPER() local
1193 for (i = 0; i < 2; i++) { in HELPER()
1194 temp.D(2 * i) = int128_getlo(int128_rshift(Vj->Q(i), imm % 128)); in HELPER()
1195 temp.D(2 * i + 1) = int128_getlo(int128_rshift(Vd->Q(i), imm % 128)); in HELPER()
1207 int i, j, ofs; \
1214 for (i = 0; i < oprsz / 16; i++) { \
1216 Vd->E1(j + ofs * 2 * i) = do_vsrlr_ ##E2(Vj->E2(j + ofs * i), \
1217 Vk->E3(j + ofs * i) % BIT); \
1219 Vd->D(2 * i + 1) = 0; \
1230 int i, j, ofs; \
1237 for (i = 0; i < oprsz / 16; i++) { \
1239 Vd->E1(j + ofs * 2 * i) = do_vsrar_ ## E2(Vj->E2(j + ofs * i), \
1240 Vk->E3(j + ofs * i) % BIT); \
1242 Vd->D(2 * i + 1) = 0; \
1253 int i, j, ofs; \
1260 for (i = 0; i < oprsz / 16; i++) { \
1262 temp.E1(j + ofs * 2 * i) = do_vsrlr_ ## E2(Vj->E2(j + ofs * i), imm); \
1263 temp.E1(j + ofs * (2 * i + 1)) = do_vsrlr_ ## E2(Vd->E2(j + ofs * i), \
1272 int i; in HELPER() local
1279 for (i = 0; i < oprsz / 16; i++) { in HELPER()
1281 temp.D(2 * i) = int128_getlo(Vj->Q(i)); in HELPER()
1282 temp.D(2 * i + 1) = int128_getlo(Vd->Q(i)); in HELPER()
1284 r[2 * i] = int128_and(int128_urshift(Vj->Q(i), (imm - 1)), in HELPER()
1286 r[2 * i + 1] = int128_and(int128_urshift(Vd->Q(i), (imm - 1)), in HELPER()
1288 temp.D(2 * i) = int128_getlo(int128_add(int128_urshift(Vj->Q(i), in HELPER()
1289 imm), r[2 * i])); in HELPER()
1290 temp.D(2 * i + 1) = int128_getlo(int128_add(int128_urshift(Vd->Q(i), in HELPER()
1291 imm), r[ 2 * i + 1])); in HELPER()
1304 int i, j, ofs; \
1311 for (i = 0; i < oprsz / 16; i++) { \
1313 temp.E1(j + ofs * 2 * i) = do_vsrar_ ## E2(Vj->E2(j + ofs * i), imm); \
1314 temp.E1(j + ofs * (2 * i + 1)) = do_vsrar_ ## E2(Vd->E2(j + ofs * i), \
1323 int i; in HELPER() local
1330 for (i = 0; i < oprsz / 16; i++) { in HELPER()
1332 temp.D(2 * i) = int128_getlo(Vj->Q(i)); in HELPER()
1333 temp.D(2 * i + 1) = int128_getlo(Vd->Q(i)); in HELPER()
1335 r[2 * i] = int128_and(int128_rshift(Vj->Q(i), (imm - 1)), in HELPER()
1337 r[2 * i + 1] = int128_and(int128_rshift(Vd->Q(i), (imm - 1)), in HELPER()
1339 temp.D(2 * i) = int128_getlo(int128_add(int128_rshift(Vj->Q(i), in HELPER()
1340 imm), r[2 * i])); in HELPER()
1341 temp.D(2 * i + 1) = int128_getlo(int128_add(int128_rshift(Vd->Q(i), in HELPER()
1342 imm), r[2 * i + 1])); in HELPER()
1377 int i, j, ofs; \ in SSRLNS()
1384 for (i = 0; i < oprsz / 16; i++) { \ in SSRLNS()
1386 Vd->E1(j + ofs * 2 * i) = do_ssrlns_ ## E1(Vj->E2(j + ofs * i), \ in SSRLNS()
1387 Vk->E3(j + ofs * i) % BIT, \ in SSRLNS()
1390 Vd->D(2 * i + 1) = 0; \ in SSRLNS()
1425 int i, j, ofs; \
1432 for (i = 0; i < oprsz / 16; i++) { \
1434 Vd->E1(j + ofs * 2 * i) = do_ssrans_ ## E1(Vj->E2(j + ofs * i), \
1435 Vk->E3(j + ofs * i) % BIT, \
1438 Vd->D(2 * i + 1) = 0; \
1471 int i, j, ofs; \
1478 for (i = 0; i < oprsz / 16; i++) { \
1480 Vd->E1(j + ofs * 2 * i) = do_ssrlnu_ ## E1(Vj->E2(j + ofs * i), \
1481 Vk->E3(j + ofs * i) % BIT, \
1484 Vd->D(2 * i + 1) = 0; \
1520 int i, j, ofs; \
1527 for (i = 0; i < oprsz / 16; i++) { \
1529 Vd->E1(j + ofs * 2 * i) = do_ssranu_ ## E1(Vj->E2(j + ofs * i), \
1530 Vk->E3(j + ofs * i) % BIT, \
1533 Vd->D(2 * i + 1) = 0; \
1544 int i, j, ofs; \
1551 for (i = 0; i < oprsz / 16; i++) { \
1553 temp.E1(j + ofs * 2 * i) = do_ssrlns_ ## E1(Vj->E2(j + ofs * i), \
1555 temp.E1(j + ofs * (2 * i + 1)) = do_ssrlns_ ## E1(Vd->E2(j + ofs * i), \
1590 int i; in HELPER() local
1598 for (i = 0; i < oprsz / 16; i++) { in HELPER()
1599 do_vssrlni_q(Vd, Vj, imm, i, mask); in HELPER()
1610 int i, j, ofs; \
1617 for (i = 0; i < oprsz / 16; i++) { \
1619 temp.E1(j + ofs * 2 * i) = do_ssrans_ ## E1(Vj->E2(j + ofs * i), \
1621 temp.E1(j + ofs * (2 * i + 1)) = do_ssrans_ ## E1(Vd->E2(j + ofs * i), \
1660 int i; in HELPER() local
1669 for (i = 0; i < oprsz / 16; i++) { in HELPER()
1670 do_vssrani_d_q(Vd, Vj, imm, i, mask, min); in HELPER()
1682 int i, j, ofs; \
1689 for (i = 0; i < oprsz / 16; i++) { \
1691 temp.E1(j + ofs * 2 * i) = do_ssrlnu_ ## E1(Vj->E2(j + ofs * i), \
1693 temp.E1(j + ofs * (2 * i + 1)) = do_ssrlnu_ ## E1(Vd->E2(j + ofs * i), \
1702 int i; in HELPER() local
1710 for (i = 0; i < oprsz / 16; i++) { in HELPER()
1711 do_vssrlni_q(Vd, Vj, imm, i, mask); in HELPER()
1722 int i, j, ofs; \
1729 for (i = 0; i < oprsz / 16; i++) { \
1731 temp.E1(j + ofs * 2 * i) = do_ssranu_ ## E1(Vj->E2(j + ofs * i), \
1733 temp.E1(j + ofs * (2 * i + 1)) = do_ssranu_ ## E1(Vd->E2(j + ofs * i), \
1776 int i; in HELPER() local
1784 for (i = 0; i < oprsz / 16; i++) { in HELPER()
1785 do_vssrani_du_q(Vd, Vj, imm, i, mask); in HELPER()
1815 int i, j, ofs; \ in SSRLRNS()
1822 for (i = 0; i < oprsz / 16; i++) { \ in SSRLRNS()
1824 Vd->E1(j + ofs * 2 * i) = do_ssrlrns_ ## E1(Vj->E2(j + ofs * i), \ in SSRLRNS()
1825 Vk->E3(j + ofs * i) % BIT, \ in SSRLRNS()
1828 Vd->D(2 * i + 1) = 0; \ in SSRLRNS()
1860 int i, j, ofs; \
1867 for (i = 0; i < oprsz / 16; i++) { \
1869 Vd->E1(j + ofs * 2 * i) = do_ssrarns_ ## E1(Vj->E2(j + ofs * i), \
1870 Vk->E3(j + ofs * i) % BIT, \
1873 Vd->D(2 * i + 1) = 0; \
1904 int i, j, ofs; \
1911 for (i = 0; i < oprsz / 16; i++) { \
1913 Vd->E1(j + ofs * 2 * i) = do_ssrlrnu_ ## E1(Vj->E2(j + ofs * i), \
1914 Vk->E3(j + ofs * i) % BIT, \
1917 Vd->D(2 * i + 1) = 0; \
1951 int i, j, ofs; \
1958 for (i = 0; i < oprsz / 16; i++) { \
1960 Vd->E1(j + ofs * 2 * i) = do_ssrarnu_ ## E1(Vj->E2(j + ofs * i), \
1961 Vk->E3(j + ofs * i) % BIT, \
1964 Vd->D(2 * i + 1) = 0; \
1975 int i, j, ofs; \
1982 for (i = 0; i < oprsz / 16; i++) { \
1984 temp.E1(j + ofs * 2 * i) = do_ssrlrns_ ## E1(Vj->E2(j + ofs * i), \
1986 temp.E1(j + ofs * (2 * i + 1)) = do_ssrlrns_ ## E1(Vd->E2(j + ofs * i), \
2022 int i; in HELPER() local
2030 for (i = 0; i < oprsz / 16; i++) { in HELPER()
2031 do_vssrlrni_q(Vd, Vj, imm, i, mask); in HELPER()
2042 int i, j, ofs; \
2049 for (i = 0; i < oprsz / 16; i++) { \
2051 temp.E1(j + ofs * 2 * i) = do_ssrarns_ ## E1(Vj->E2(j + ofs * i), \
2053 temp.E1(j + ofs * (2 * i + 1)) = do_ssrarns_ ## E1(Vd->E2(j + ofs * i), \
2093 int i; in HELPER() local
2102 for (i = 0; i < oprsz / 16; i++) { in HELPER()
2103 do_vssrarni_d_q(Vd, Vj, imm, i, mask1, mask2); in HELPER()
2114 int i, j, ofs; \
2121 for (i = 0; i < oprsz / 16; i++) { \
2123 temp.E1(j + ofs * 2 * i) = do_ssrlrnu_ ## E1(Vj->E2(j + ofs * i), \
2125 temp.E1(j + ofs * (2 * i + 1)) = do_ssrlrnu_ ## E1(Vd->E2(j + ofs * i), \
2134 int i; in HELPER() local
2142 for (i = 0; i < oprsz / 16; i++) { in HELPER()
2143 do_vssrlrni_q(Vd, Vj, imm, i, mask); in HELPER()
2154 int i, j, ofs; \
2161 for (i = 0; i < oprsz / 16; i++) { \
2163 temp.E1(j + ofs * 2 * i) = do_ssrarnu_ ## E1(Vj->E2(j + ofs * i), \
2165 temp.E1(j + ofs * (2 * i + 1)) = do_ssrarnu_ ## E1(Vd->E2(j + ofs * i), \
2213 int i; in HELPER() local
2222 for (i = 0; i < oprsz / 16; i++) { in HELPER()
2223 do_vssrarni_du_q(Vd, Vj, imm, i, mask1, mask2); in HELPER()
2234 int i; \
2239 for (i = 0; i < oprsz / (BIT / 8); i++) \
2241 Vd->E(i) = DO_OP(Vj->E(i)); \
2257 int i; \
2262 for (i = 0; i < oprsz / (BIT / 8); i++) \
2264 Vd->E(i) = FN(Vj->E(i)); \
2276 int i; \
2282 for (i = 0; i < oprsz / (BIT / 8); i++) { \
2283 Vd->E(i) = DO_OP(Vj->E(i), Vk->E(i)%BIT); \
2303 int i; \
2308 for (i = 0; i < oprsz / (BIT / 8); i++) { \
2309 Vd->E(i) = DO_OP(Vj->E(i), imm); \
2329 int i, j, m, ofs; \
2336 for (i = 0; i < oprsz / 16; i++) { \
2337 m = Vk->E(i * ofs) & MASK; \
2339 if (Vj->E(j + ofs * i) < 0) { \
2343 Vd->E(m + i * ofs) = j; \
2353 int i, j, m, ofs; \
2360 for (i = 0; i < oprsz / 16; i++) { \
2362 if (Vj->E(j + ofs * i) < 0) { \
2366 Vd->E(m + i * ofs) = j; \
2408 int i; \
2415 for (i = 0; i < oprsz / (BIT / 8); i++) { \
2416 Vd->E(i) = FN(Vj->E(i), Vk->E(i), &env->fp_status); \
2442 int i; \
2450 for (i = 0; i < oprsz / (BIT / 8); i++) { \
2451 Vd->E(i) = FN(Vj->E(i), Vk->E(i), Va->E(i), flags, &env->fp_status); \
2471 int i; \
2477 for (i = 0; i < oprsz / (BIT / 8); i++) { \
2478 Vd->E(i) = FN(env, Vj->E(i)); \
2504 int i; \
2509 for (i = 0; i < oprsz / (BIT / 8); i++) { \
2510 Vd->E(i) = FN(env, Vj->E(i)); \
2584 int i, j, ofs; in HELPER() local
2592 for (i = 0; i < oprsz / 16; i++) { in HELPER()
2594 temp.UW(j + ofs * i) =float16_cvt_float32(Vj->UH(j + ofs * 2 * i), in HELPER()
2605 int i, j, ofs; in HELPER() local
2613 for (i = 0; i < oprsz / 16; i++) { in HELPER()
2615 temp.UD(j + ofs * i) = float32_cvt_float64(Vj->UW(j + ofs * 2 * i), in HELPER()
2626 int i, j, ofs; in HELPER() local
2634 for (i = 0; i < oprsz / 16; i++) { in HELPER()
2636 temp.UW(j + ofs * i) = float16_cvt_float32(Vj->UH(j + ofs * (2 * i + 1)), in HELPER()
2647 int i, j, ofs; in HELPER() local
2655 for (i = 0; i < oprsz / 16; i++) { in HELPER()
2657 temp.UD(j + ofs * i) = float32_cvt_float64(Vj->UW(j + ofs * (2 * i + 1)), in HELPER()
2668 int i, j, ofs; in HELPER() local
2677 for(i = 0; i < oprsz / 16; i++) { in HELPER()
2679 temp.UH(j + ofs * (2 * i + 1)) = float32_cvt_float16(Vj->UW(j + ofs * i), in HELPER()
2681 temp.UH(j + ofs * 2 * i) = float32_cvt_float16(Vk->UW(j + ofs * i), in HELPER()
2692 int i, j, ofs; in HELPER() local
2701 for(i = 0; i < oprsz / 16; i++) { in HELPER()
2703 temp.UW(j + ofs * (2 * i + 1)) = float64_cvt_float32(Vj->UD(j + ofs * i), in HELPER()
2705 temp.UW(j + ofs * 2 * i) = float64_cvt_float32(Vk->UD(j + ofs * i), in HELPER()
2716 int i; in HELPER() local
2722 for (i = 0; i < oprsz / 4; i++) { in HELPER()
2723 Vd->W(i) = float32_round_to_int(Vj->UW(i), &env->fp_status); in HELPER()
2731 int i; in HELPER() local
2737 for (i = 0; i < oprsz / 8; i++) { in HELPER()
2738 Vd->D(i) = float64_round_to_int(Vj->UD(i), &env->fp_status); in HELPER()
2747 int i; \
2753 for (i = 0; i < oprsz / (BIT / 8); i++) { \
2756 Vd->E(i) = float## BIT ## _round_to_int(Vj->E(i), &env->fp_status); \
2842 int i, j, ofs; \ in DO_FTINT()
2851 for (i = 0; i < oprsz / 16; i++) { \ in DO_FTINT()
2853 temp.W(j + ofs * (2 * i + 1)) = FN(env, Vj->UD(j + ofs * i)); \ in DO_FTINT()
2854 temp.W(j + ofs * 2 * i) = FN(env, Vk->UD(j + ofs * i)); \ in DO_FTINT()
2879 int i, j, ofs; \
2887 for (i = 0; i < oprsz / 16; i++) { \
2889 temp.D(j + ofs * i) = FN(env, Vj->UW(j + ofs * 2 * i)); \
2905 int i, j, ofs; \
2913 for (i = 0; i < oprsz / 16; i++) { \
2915 temp.D(j + ofs * i) = FN(env, Vj->UW(j + ofs * (2 * i + 1))); \
2950 int i, j, ofs; local
2958 for (i = 0; i < oprsz / 16; i++) {
2960 temp.D(j + ofs * i) = int32_to_float64(Vj->W(j + ofs * 2 * i),
2971 int i, j, ofs; in HELPER() local
2979 for (i = 0; i < oprsz /16; i++) { in HELPER()
2981 temp.D(j + ofs * i) = int32_to_float64(Vj->W(j + ofs * (2 * i + 1)), in HELPER()
2992 int i, j, ofs; in HELPER() local
3001 for (i = 0; i < oprsz / 16; i++) { in HELPER()
3003 temp.W(j + ofs * (2 * i + 1)) = int64_to_float32(Vj->D(j + ofs * i), in HELPER()
3005 temp.W(j + ofs * 2 * i) = int64_to_float32(Vk->D(j + ofs * i), in HELPER()
3016 int i; \
3022 for (i = 0; i < oprsz / (BIT / 8); i++) { \
3023 Vd->E(i) = DO_OP(Vj->E(i), (TD)imm); \
3081 int i; \
3088 for (i = 0; i < oprsz / (BIT / 8); i++) { \
3090 cmp = FN(Vj->E(i), Vk->E(i), &env->fp_status); \
3091 t.E(i) = vfcmp_common(env, cmp, flags); \
3104 int i; in HELPER() local
3108 for (i = 0; i < simd_oprsz(desc); i++) { in HELPER()
3109 Vd->B(i) = (~Vd->B(i) & Vj->B(i)) | (Vd->B(i) & imm); in HELPER()
3179 int i; \
3185 for (i = 1; i < oprsz / (BIT / 8); i++) { \
3186 Vd->E(i) = 0; \
3196 int i; \
3203 for (i = 0; i < oprsz / (BIT / 8); i++) { \
3204 temp.E(2 * i + 1) = Vj->E(2 * i); \
3205 temp.E(2 *i) = Vk->E(2 * i); \
3218 int i; \
3225 for (i = 0; i < oprsz / (BIT / 8); i++) { \
3226 temp.E(2 * i + 1) = Vj->E(2 * i + 1); \
3227 temp.E(2 * i) = Vk->E(2 * i + 1); \
3240 int i, j, ofs; \
3248 for (i = 0; i < oprsz / 16; i++) { \
3250 temp.E(j + ofs * (2 * i + 1)) = Vj->E(2 * (j + ofs * i)); \
3251 temp.E(j + ofs * 2 * i) = Vk->E(2 * (j + ofs * i)); \
3265 int i, j, ofs; \
3273 for (i = 0; i < oprsz / 16; i++) { \
3275 temp.E(j + ofs * (2 * i + 1)) = Vj->E(2 * (j + ofs * i) + 1); \
3276 temp.E(j + ofs * 2 * i) = Vk->E(2 * (j + ofs * i) + 1); \
3290 int i, j, ofs; \
3298 for (i = 0; i < oprsz / 16; i++) { \
3300 temp.E(2 * (j + ofs * i) + 1) = Vj->E(j + ofs * 2 * i); \
3301 temp.E(2 * (j + ofs * i)) = Vk->E(j + ofs * 2 * i); \
3315 int i, j, ofs; \
3323 for (i = 0; i < oprsz / 16; i++) { \
3325 temp.E(2 * (j + ofs * i) + 1) = Vj->E(j + ofs * (2 * i + 1)); \
3326 temp.E(2 * (j + ofs * i)) = Vk->E(j + ofs * (2 * i + 1)); \
3339 int i, j, m; local
3348 for (i = 0; i < (oprsz / 16) * m; i++) {
3349 j = i < m ? 0 : 1;
3350 uint64_t k = (uint8_t)Va->B(i) % (2 * m);
3351 temp.B(i) = k < m ? Vk->B(k + j * m): Vj->B(k + (j - 1) * m);
3359 int i, j, m; \
3367 for (i = 0; i < (oprsz / 16) * m; i++) { \
3368 j = i < m ? 0 : 1; \
3369 uint64_t k = ((uint8_t)Vd->E(i)) % (2 * m); \
3370 temp.E(i) = k < m ? Vk->E(k + j * m) : Vj->E(k + (j - 1) * m); \
3382 int i, j, max; \
3389 for (i = 0; i < oprsz / (BIT / 8); i++) { \
3390 j = i < max ? 1 : 2; \
3391 temp.E(i) = Vj->E(SHF_POS(i - ((j -1)* max), imm) + (j - 1) * max); \
3402 int i; in HELPER() local
3408 for (i = 0; i < oprsz / 16; i++) { in HELPER()
3409 temp.D(2 * i) = (imm & 2 ? Vj : Vd)->D((imm & 1) + 2 * i); in HELPER()
3410 temp.D(2 * i + 1) = (imm & 8 ? Vj : Vd)->D(((imm >> 2) & 1) + 2 * i); in HELPER()
3417 int i, m; in HELPER() local
3424 for (i = 0; i < m ; i++) { in HELPER()
3425 uint64_t k = (uint8_t)Vk->W(i) % 8; in HELPER()
3426 temp.W(i) = Vj->W(k); in HELPER()
3433 int i; in HELPER() local
3439 for (i = 0; i < oprsz / 16; i++) { in HELPER()
3440 temp.W(4 * i) = Vj->W((imm & 0x3) + 4 * i); in HELPER()
3441 temp.W(4 * i + 1) = Vj->W(((imm >> 2) & 0x3) + 4 * i); in HELPER()
3442 temp.W(4 * i + 2) = Vd->W(((imm >> 4) & 0x3) + 4 * i); in HELPER()
3443 temp.W(4 * i + 3) = Vd->W(((imm >> 6) & 0x3) + 4 * i); in HELPER()
3463 int i; in HELPER() local
3468 for (i = 0; i < 2; i++, imm >>= 4) { in HELPER()
3469 temp.Q(i) = (imm & 2 ? Vd: Vj)->Q(imm & 1); in HELPER()
3477 int i, ins, extr, max; \
3485 for (i = 0; i < oprsz / 16; i++) { \
3486 Vd->E(ins + i * max) = Vj->E(extr + i * max); \