diff options
author | beck <> | 2001-08-01 19:51:17 +0000 |
---|---|---|
committer | beck <> | 2001-08-01 19:51:17 +0000 |
commit | 4b790f68539c49ef91f5e82506c2624900c92106 (patch) | |
tree | d14ff9f79630cb04ad006cd2730646f44f62d786 /src/lib/libcrypto/bn | |
parent | edbfd6c7e91e15e92ef0df548474ac76b6dddca0 (diff) | |
download | openbsd-4b790f68539c49ef91f5e82506c2624900c92106.tar.gz openbsd-4b790f68539c49ef91f5e82506c2624900c92106.tar.bz2 openbsd-4b790f68539c49ef91f5e82506c2624900c92106.zip |
merge openssl 0.9.6b-engine
Note that this is a maintenence release, API's appear *not* to have changed.
As such, I have only increased the minor number on these libraries
Diffstat (limited to 'src/lib/libcrypto/bn')
-rw-r--r-- | src/lib/libcrypto/bn/asm/mips3.s | 430 |
1 files changed, 245 insertions, 185 deletions
diff --git a/src/lib/libcrypto/bn/asm/mips3.s b/src/lib/libcrypto/bn/asm/mips3.s index 45786c00a5..dca4105c7d 100644 --- a/src/lib/libcrypto/bn/asm/mips3.s +++ b/src/lib/libcrypto/bn/asm/mips3.s | |||
@@ -1,5 +1,5 @@ | |||
1 | .rdata | 1 | .rdata |
2 | .asciiz "mips3.s, Version 1.0" | 2 | .asciiz "mips3.s, Version 1.1" |
3 | .asciiz "MIPS III/IV ISA artwork by Andy Polyakov <appro@fy.chalmers.se>" | 3 | .asciiz "MIPS III/IV ISA artwork by Andy Polyakov <appro@fy.chalmers.se>" |
4 | 4 | ||
5 | /* | 5 | /* |
@@ -849,6 +849,7 @@ LEAF(bn_mul_comba8) | |||
849 | sltu AT,c_1,t_1 | 849 | sltu AT,c_1,t_1 |
850 | daddu t_2,AT | 850 | daddu t_2,AT |
851 | daddu c_2,t_2 | 851 | daddu c_2,t_2 |
852 | sltu c_3,c_2,t_2 | ||
852 | dmultu a_1,b_2 /* mul_add_c(a[1],b[2],c1,c2,c3); */ | 853 | dmultu a_1,b_2 /* mul_add_c(a[1],b[2],c1,c2,c3); */ |
853 | mflo t_1 | 854 | mflo t_1 |
854 | mfhi t_2 | 855 | mfhi t_2 |
@@ -856,7 +857,8 @@ LEAF(bn_mul_comba8) | |||
856 | sltu AT,c_1,t_1 | 857 | sltu AT,c_1,t_1 |
857 | daddu t_2,AT | 858 | daddu t_2,AT |
858 | daddu c_2,t_2 | 859 | daddu c_2,t_2 |
859 | sltu c_3,c_2,t_2 | 860 | sltu AT,c_2,t_2 |
861 | daddu c_3,AT | ||
860 | dmultu a_2,b_1 /* mul_add_c(a[2],b[1],c1,c2,c3); */ | 862 | dmultu a_2,b_1 /* mul_add_c(a[2],b[1],c1,c2,c3); */ |
861 | mflo t_1 | 863 | mflo t_1 |
862 | mfhi t_2 | 864 | mfhi t_2 |
@@ -884,6 +886,7 @@ LEAF(bn_mul_comba8) | |||
884 | sltu AT,c_2,t_1 | 886 | sltu AT,c_2,t_1 |
885 | daddu t_2,AT | 887 | daddu t_2,AT |
886 | daddu c_3,t_2 | 888 | daddu c_3,t_2 |
889 | sltu c_1,c_3,t_2 | ||
887 | dmultu a_3,b_1 /* mul_add_c(a[3],b[1],c2,c3,c1); */ | 890 | dmultu a_3,b_1 /* mul_add_c(a[3],b[1],c2,c3,c1); */ |
888 | mflo t_1 | 891 | mflo t_1 |
889 | mfhi t_2 | 892 | mfhi t_2 |
@@ -891,7 +894,8 @@ LEAF(bn_mul_comba8) | |||
891 | sltu AT,c_2,t_1 | 894 | sltu AT,c_2,t_1 |
892 | daddu t_2,AT | 895 | daddu t_2,AT |
893 | daddu c_3,t_2 | 896 | daddu c_3,t_2 |
894 | sltu c_1,c_3,t_2 | 897 | sltu AT,c_3,t_2 |
898 | daddu c_1,AT | ||
895 | dmultu a_2,b_2 /* mul_add_c(a[2],b[2],c2,c3,c1); */ | 899 | dmultu a_2,b_2 /* mul_add_c(a[2],b[2],c2,c3,c1); */ |
896 | mflo t_1 | 900 | mflo t_1 |
897 | mfhi t_2 | 901 | mfhi t_2 |
@@ -928,6 +932,7 @@ LEAF(bn_mul_comba8) | |||
928 | sltu AT,c_3,t_1 | 932 | sltu AT,c_3,t_1 |
929 | daddu t_2,AT | 933 | daddu t_2,AT |
930 | daddu c_1,t_2 | 934 | daddu c_1,t_2 |
935 | sltu c_2,c_1,t_2 | ||
931 | dmultu a_1,b_4 /* mul_add_c(a[1],b[4],c3,c1,c2); */ | 936 | dmultu a_1,b_4 /* mul_add_c(a[1],b[4],c3,c1,c2); */ |
932 | mflo t_1 | 937 | mflo t_1 |
933 | mfhi t_2 | 938 | mfhi t_2 |
@@ -935,7 +940,8 @@ LEAF(bn_mul_comba8) | |||
935 | sltu AT,c_3,t_1 | 940 | sltu AT,c_3,t_1 |
936 | daddu t_2,AT | 941 | daddu t_2,AT |
937 | daddu c_1,t_2 | 942 | daddu c_1,t_2 |
938 | sltu c_2,c_1,t_2 | 943 | sltu AT,c_1,t_2 |
944 | daddu c_2,AT | ||
939 | dmultu a_2,b_3 /* mul_add_c(a[2],b[3],c3,c1,c2); */ | 945 | dmultu a_2,b_3 /* mul_add_c(a[2],b[3],c3,c1,c2); */ |
940 | mflo t_1 | 946 | mflo t_1 |
941 | mfhi t_2 | 947 | mfhi t_2 |
@@ -981,6 +987,7 @@ LEAF(bn_mul_comba8) | |||
981 | sltu AT,c_1,t_1 | 987 | sltu AT,c_1,t_1 |
982 | daddu t_2,AT | 988 | daddu t_2,AT |
983 | daddu c_2,t_2 | 989 | daddu c_2,t_2 |
990 | sltu c_3,c_2,t_2 | ||
984 | dmultu a_5,b_1 /* mul_add_c(a[5],b[1],c1,c2,c3); */ | 991 | dmultu a_5,b_1 /* mul_add_c(a[5],b[1],c1,c2,c3); */ |
985 | mflo t_1 | 992 | mflo t_1 |
986 | mfhi t_2 | 993 | mfhi t_2 |
@@ -988,7 +995,8 @@ LEAF(bn_mul_comba8) | |||
988 | sltu AT,c_1,t_1 | 995 | sltu AT,c_1,t_1 |
989 | daddu t_2,AT | 996 | daddu t_2,AT |
990 | daddu c_2,t_2 | 997 | daddu c_2,t_2 |
991 | sltu c_3,c_2,t_2 | 998 | sltu AT,c_2,t_2 |
999 | daddu c_3,AT | ||
992 | dmultu a_4,b_2 /* mul_add_c(a[4],b[2],c1,c2,c3); */ | 1000 | dmultu a_4,b_2 /* mul_add_c(a[4],b[2],c1,c2,c3); */ |
993 | mflo t_1 | 1001 | mflo t_1 |
994 | mfhi t_2 | 1002 | mfhi t_2 |
@@ -1043,6 +1051,7 @@ LEAF(bn_mul_comba8) | |||
1043 | sltu AT,c_2,t_1 | 1051 | sltu AT,c_2,t_1 |
1044 | daddu t_2,AT | 1052 | daddu t_2,AT |
1045 | daddu c_3,t_2 | 1053 | daddu c_3,t_2 |
1054 | sltu c_1,c_3,t_2 | ||
1046 | dmultu a_1,b_6 /* mul_add_c(a[1],b[6],c2,c3,c1); */ | 1055 | dmultu a_1,b_6 /* mul_add_c(a[1],b[6],c2,c3,c1); */ |
1047 | mflo t_1 | 1056 | mflo t_1 |
1048 | mfhi t_2 | 1057 | mfhi t_2 |
@@ -1050,7 +1059,8 @@ LEAF(bn_mul_comba8) | |||
1050 | sltu AT,c_2,t_1 | 1059 | sltu AT,c_2,t_1 |
1051 | daddu t_2,AT | 1060 | daddu t_2,AT |
1052 | daddu c_3,t_2 | 1061 | daddu c_3,t_2 |
1053 | sltu c_1,c_3,t_2 | 1062 | sltu AT,c_3,t_2 |
1063 | daddu c_1,AT | ||
1054 | dmultu a_2,b_5 /* mul_add_c(a[2],b[5],c2,c3,c1); */ | 1064 | dmultu a_2,b_5 /* mul_add_c(a[2],b[5],c2,c3,c1); */ |
1055 | mflo t_1 | 1065 | mflo t_1 |
1056 | mfhi t_2 | 1066 | mfhi t_2 |
@@ -1114,6 +1124,7 @@ LEAF(bn_mul_comba8) | |||
1114 | sltu AT,c_3,t_1 | 1124 | sltu AT,c_3,t_1 |
1115 | daddu t_2,AT | 1125 | daddu t_2,AT |
1116 | daddu c_1,t_2 | 1126 | daddu c_1,t_2 |
1127 | sltu c_2,c_1,t_2 | ||
1117 | dmultu a_6,b_2 /* mul_add_c(a[6],b[2],c3,c1,c2); */ | 1128 | dmultu a_6,b_2 /* mul_add_c(a[6],b[2],c3,c1,c2); */ |
1118 | mflo t_1 | 1129 | mflo t_1 |
1119 | mfhi t_2 | 1130 | mfhi t_2 |
@@ -1121,7 +1132,8 @@ LEAF(bn_mul_comba8) | |||
1121 | sltu AT,c_3,t_1 | 1132 | sltu AT,c_3,t_1 |
1122 | daddu t_2,AT | 1133 | daddu t_2,AT |
1123 | daddu c_1,t_2 | 1134 | daddu c_1,t_2 |
1124 | sltu c_2,c_1,t_2 | 1135 | sltu AT,c_1,t_2 |
1136 | daddu c_2,AT | ||
1125 | dmultu a_5,b_3 /* mul_add_c(a[5],b[3],c3,c1,c2); */ | 1137 | dmultu a_5,b_3 /* mul_add_c(a[5],b[3],c3,c1,c2); */ |
1126 | mflo t_1 | 1138 | mflo t_1 |
1127 | mfhi t_2 | 1139 | mfhi t_2 |
@@ -1176,6 +1188,7 @@ LEAF(bn_mul_comba8) | |||
1176 | sltu AT,c_1,t_1 | 1188 | sltu AT,c_1,t_1 |
1177 | daddu t_2,AT | 1189 | daddu t_2,AT |
1178 | daddu c_2,t_2 | 1190 | daddu c_2,t_2 |
1191 | sltu c_3,c_2,t_2 | ||
1179 | dmultu a_3,b_6 /* mul_add_c(a[3],b[6],c1,c2,c3); */ | 1192 | dmultu a_3,b_6 /* mul_add_c(a[3],b[6],c1,c2,c3); */ |
1180 | mflo t_1 | 1193 | mflo t_1 |
1181 | mfhi t_2 | 1194 | mfhi t_2 |
@@ -1183,7 +1196,8 @@ LEAF(bn_mul_comba8) | |||
1183 | sltu AT,c_1,t_1 | 1196 | sltu AT,c_1,t_1 |
1184 | daddu t_2,AT | 1197 | daddu t_2,AT |
1185 | daddu c_2,t_2 | 1198 | daddu c_2,t_2 |
1186 | sltu c_3,c_2,t_2 | 1199 | sltu AT,c_2,t_2 |
1200 | daddu c_3,AT | ||
1187 | dmultu a_4,b_5 /* mul_add_c(a[4],b[5],c1,c2,c3); */ | 1201 | dmultu a_4,b_5 /* mul_add_c(a[4],b[5],c1,c2,c3); */ |
1188 | mflo t_1 | 1202 | mflo t_1 |
1189 | mfhi t_2 | 1203 | mfhi t_2 |
@@ -1229,6 +1243,7 @@ LEAF(bn_mul_comba8) | |||
1229 | sltu AT,c_2,t_1 | 1243 | sltu AT,c_2,t_1 |
1230 | daddu t_2,AT | 1244 | daddu t_2,AT |
1231 | daddu c_3,t_2 | 1245 | daddu c_3,t_2 |
1246 | sltu c_1,c_3,t_2 | ||
1232 | dmultu a_6,b_4 /* mul_add_c(a[6],b[4],c2,c3,c1); */ | 1247 | dmultu a_6,b_4 /* mul_add_c(a[6],b[4],c2,c3,c1); */ |
1233 | mflo t_1 | 1248 | mflo t_1 |
1234 | mfhi t_2 | 1249 | mfhi t_2 |
@@ -1236,7 +1251,8 @@ LEAF(bn_mul_comba8) | |||
1236 | sltu AT,c_2,t_1 | 1251 | sltu AT,c_2,t_1 |
1237 | daddu t_2,AT | 1252 | daddu t_2,AT |
1238 | daddu c_3,t_2 | 1253 | daddu c_3,t_2 |
1239 | sltu c_1,c_3,t_2 | 1254 | sltu AT,c_3,t_2 |
1255 | daddu c_1,AT | ||
1240 | dmultu a_5,b_5 /* mul_add_c(a[5],b[5],c2,c3,c1); */ | 1256 | dmultu a_5,b_5 /* mul_add_c(a[5],b[5],c2,c3,c1); */ |
1241 | mflo t_1 | 1257 | mflo t_1 |
1242 | mfhi t_2 | 1258 | mfhi t_2 |
@@ -1273,6 +1289,7 @@ LEAF(bn_mul_comba8) | |||
1273 | sltu AT,c_3,t_1 | 1289 | sltu AT,c_3,t_1 |
1274 | daddu t_2,AT | 1290 | daddu t_2,AT |
1275 | daddu c_1,t_2 | 1291 | daddu c_1,t_2 |
1292 | sltu c_2,c_1,t_2 | ||
1276 | dmultu a_5,b_6 /* mul_add_c(a[5],b[6],c3,c1,c2); */ | 1293 | dmultu a_5,b_6 /* mul_add_c(a[5],b[6],c3,c1,c2); */ |
1277 | mflo t_1 | 1294 | mflo t_1 |
1278 | mfhi t_2 | 1295 | mfhi t_2 |
@@ -1280,7 +1297,8 @@ LEAF(bn_mul_comba8) | |||
1280 | sltu AT,c_3,t_1 | 1297 | sltu AT,c_3,t_1 |
1281 | daddu t_2,AT | 1298 | daddu t_2,AT |
1282 | daddu c_1,t_2 | 1299 | daddu c_1,t_2 |
1283 | sltu c_2,c_1,t_2 | 1300 | sltu AT,c_1,t_2 |
1301 | daddu c_2,AT | ||
1284 | dmultu a_6,b_5 /* mul_add_c(a[6],b[5],c3,c1,c2); */ | 1302 | dmultu a_6,b_5 /* mul_add_c(a[6],b[5],c3,c1,c2); */ |
1285 | mflo t_1 | 1303 | mflo t_1 |
1286 | mfhi t_2 | 1304 | mfhi t_2 |
@@ -1308,6 +1326,7 @@ LEAF(bn_mul_comba8) | |||
1308 | sltu AT,c_1,t_1 | 1326 | sltu AT,c_1,t_1 |
1309 | daddu t_2,AT | 1327 | daddu t_2,AT |
1310 | daddu c_2,t_2 | 1328 | daddu c_2,t_2 |
1329 | sltu c_3,c_2,t_2 | ||
1311 | dmultu a_6,b_6 /* mul_add_c(a[6],b[6],c1,c2,c3); */ | 1330 | dmultu a_6,b_6 /* mul_add_c(a[6],b[6],c1,c2,c3); */ |
1312 | mflo t_1 | 1331 | mflo t_1 |
1313 | mfhi t_2 | 1332 | mfhi t_2 |
@@ -1315,7 +1334,8 @@ LEAF(bn_mul_comba8) | |||
1315 | sltu AT,c_1,t_1 | 1334 | sltu AT,c_1,t_1 |
1316 | daddu t_2,AT | 1335 | daddu t_2,AT |
1317 | daddu c_2,t_2 | 1336 | daddu c_2,t_2 |
1318 | sltu c_3,c_2,t_2 | 1337 | sltu AT,c_2,t_2 |
1338 | daddu c_3,AT | ||
1319 | dmultu a_5,b_7 /* mul_add_c(a[5],b[7],c1,c2,c3); */ | 1339 | dmultu a_5,b_7 /* mul_add_c(a[5],b[7],c1,c2,c3); */ |
1320 | mflo t_1 | 1340 | mflo t_1 |
1321 | mfhi t_2 | 1341 | mfhi t_2 |
@@ -1334,6 +1354,7 @@ LEAF(bn_mul_comba8) | |||
1334 | sltu AT,c_2,t_1 | 1354 | sltu AT,c_2,t_1 |
1335 | daddu t_2,AT | 1355 | daddu t_2,AT |
1336 | daddu c_3,t_2 | 1356 | daddu c_3,t_2 |
1357 | sltu c_1,c_3,t_2 | ||
1337 | dmultu a_7,b_6 /* mul_add_c(a[7],b[6],c2,c3,c1); */ | 1358 | dmultu a_7,b_6 /* mul_add_c(a[7],b[6],c2,c3,c1); */ |
1338 | mflo t_1 | 1359 | mflo t_1 |
1339 | mfhi t_2 | 1360 | mfhi t_2 |
@@ -1341,7 +1362,8 @@ LEAF(bn_mul_comba8) | |||
1341 | sltu AT,c_2,t_1 | 1362 | sltu AT,c_2,t_1 |
1342 | daddu t_2,AT | 1363 | daddu t_2,AT |
1343 | daddu c_3,t_2 | 1364 | daddu c_3,t_2 |
1344 | sltu c_1,c_3,t_2 | 1365 | sltu AT,c_3,t_2 |
1366 | daddu c_1,AT | ||
1345 | sd c_2,104(a0) /* r[13]=c2; */ | 1367 | sd c_2,104(a0) /* r[13]=c2; */ |
1346 | 1368 | ||
1347 | dmultu a_7,b_7 /* mul_add_c(a[7],b[7],c3,c1,c2); */ | 1369 | dmultu a_7,b_7 /* mul_add_c(a[7],b[7],c3,c1,c2); */ |
@@ -1430,6 +1452,7 @@ LEAF(bn_mul_comba4) | |||
1430 | sltu AT,c_1,t_1 | 1452 | sltu AT,c_1,t_1 |
1431 | daddu t_2,AT | 1453 | daddu t_2,AT |
1432 | daddu c_2,t_2 | 1454 | daddu c_2,t_2 |
1455 | sltu c_3,c_2,t_2 | ||
1433 | dmultu a_1,b_2 /* mul_add_c(a[1],b[2],c1,c2,c3); */ | 1456 | dmultu a_1,b_2 /* mul_add_c(a[1],b[2],c1,c2,c3); */ |
1434 | mflo t_1 | 1457 | mflo t_1 |
1435 | mfhi t_2 | 1458 | mfhi t_2 |
@@ -1437,7 +1460,8 @@ LEAF(bn_mul_comba4) | |||
1437 | sltu AT,c_1,t_1 | 1460 | sltu AT,c_1,t_1 |
1438 | daddu t_2,AT | 1461 | daddu t_2,AT |
1439 | daddu c_2,t_2 | 1462 | daddu c_2,t_2 |
1440 | sltu c_3,c_2,t_2 | 1463 | sltu AT,c_2,t_2 |
1464 | daddu c_3,AT | ||
1441 | dmultu a_2,b_1 /* mul_add_c(a[2],b[1],c1,c2,c3); */ | 1465 | dmultu a_2,b_1 /* mul_add_c(a[2],b[1],c1,c2,c3); */ |
1442 | mflo t_1 | 1466 | mflo t_1 |
1443 | mfhi t_2 | 1467 | mfhi t_2 |
@@ -1465,6 +1489,7 @@ LEAF(bn_mul_comba4) | |||
1465 | sltu AT,c_2,t_1 | 1489 | sltu AT,c_2,t_1 |
1466 | daddu t_2,AT | 1490 | daddu t_2,AT |
1467 | daddu c_3,t_2 | 1491 | daddu c_3,t_2 |
1492 | sltu c_1,c_3,t_2 | ||
1468 | dmultu a_2,b_2 /* mul_add_c(a[2],b[2],c2,c3,c1); */ | 1493 | dmultu a_2,b_2 /* mul_add_c(a[2],b[2],c2,c3,c1); */ |
1469 | mflo t_1 | 1494 | mflo t_1 |
1470 | mfhi t_2 | 1495 | mfhi t_2 |
@@ -1472,7 +1497,8 @@ LEAF(bn_mul_comba4) | |||
1472 | sltu AT,c_2,t_1 | 1497 | sltu AT,c_2,t_1 |
1473 | daddu t_2,AT | 1498 | daddu t_2,AT |
1474 | daddu c_3,t_2 | 1499 | daddu c_3,t_2 |
1475 | sltu c_1,c_3,t_2 | 1500 | sltu AT,c_3,t_2 |
1501 | daddu c_1,AT | ||
1476 | dmultu a_1,b_3 /* mul_add_c(a[1],b[3],c2,c3,c1); */ | 1502 | dmultu a_1,b_3 /* mul_add_c(a[1],b[3],c2,c3,c1); */ |
1477 | mflo t_1 | 1503 | mflo t_1 |
1478 | mfhi t_2 | 1504 | mfhi t_2 |
@@ -1491,6 +1517,7 @@ LEAF(bn_mul_comba4) | |||
1491 | sltu AT,c_3,t_1 | 1517 | sltu AT,c_3,t_1 |
1492 | daddu t_2,AT | 1518 | daddu t_2,AT |
1493 | daddu c_1,t_2 | 1519 | daddu c_1,t_2 |
1520 | sltu c_2,c_1,t_2 | ||
1494 | dmultu a_3,b_2 /* mul_add_c(a[3],b[2],c3,c1,c2); */ | 1521 | dmultu a_3,b_2 /* mul_add_c(a[3],b[2],c3,c1,c2); */ |
1495 | mflo t_1 | 1522 | mflo t_1 |
1496 | mfhi t_2 | 1523 | mfhi t_2 |
@@ -1498,7 +1525,8 @@ LEAF(bn_mul_comba4) | |||
1498 | sltu AT,c_3,t_1 | 1525 | sltu AT,c_3,t_1 |
1499 | daddu t_2,AT | 1526 | daddu t_2,AT |
1500 | daddu c_1,t_2 | 1527 | daddu c_1,t_2 |
1501 | sltu c_2,c_1,t_2 | 1528 | sltu AT,c_1,t_2 |
1529 | daddu c_2,AT | ||
1502 | sd c_3,40(a0) | 1530 | sd c_3,40(a0) |
1503 | 1531 | ||
1504 | dmultu a_3,b_3 /* mul_add_c(a[3],b[3],c1,c2,c3); */ | 1532 | dmultu a_3,b_3 /* mul_add_c(a[3],b[3],c1,c2,c3); */ |
@@ -1543,28 +1571,30 @@ LEAF(bn_sqr_comba8) | |||
1543 | dmultu a_0,a_1 /* mul_add_c2(a[0],b[1],c2,c3,c1); */ | 1571 | dmultu a_0,a_1 /* mul_add_c2(a[0],b[1],c2,c3,c1); */ |
1544 | mflo t_1 | 1572 | mflo t_1 |
1545 | mfhi t_2 | 1573 | mfhi t_2 |
1574 | slt c_1,t_2,zero | ||
1575 | dsll t_2,1 | ||
1576 | slt a2,t_1,zero | ||
1577 | daddu t_2,a2 | ||
1578 | dsll t_1,1 | ||
1546 | daddu c_2,t_1 | 1579 | daddu c_2,t_1 |
1547 | sltu AT,c_2,t_1 | 1580 | sltu AT,c_2,t_1 |
1548 | daddu c_3,t_2,AT | 1581 | daddu c_3,t_2,AT |
1549 | daddu c_2,t_1 | ||
1550 | sltu AT,c_2,t_1 | ||
1551 | daddu t_2,AT | ||
1552 | daddu c_3,t_2 | ||
1553 | sltu c_1,c_3,t_2 | ||
1554 | sd c_2,8(a0) | 1582 | sd c_2,8(a0) |
1555 | 1583 | ||
1556 | dmultu a_2,a_0 /* mul_add_c2(a[2],b[0],c3,c1,c2); */ | 1584 | dmultu a_2,a_0 /* mul_add_c2(a[2],b[0],c3,c1,c2); */ |
1557 | mflo t_1 | 1585 | mflo t_1 |
1558 | mfhi t_2 | 1586 | mfhi t_2 |
1559 | daddu c_3,t_1 | 1587 | slt c_2,t_2,zero |
1560 | sltu AT,c_3,t_1 | 1588 | dsll t_2,1 |
1561 | daddu a2,t_2,AT | 1589 | slt a2,t_1,zero |
1562 | daddu c_1,a2 | 1590 | daddu t_2,a2 |
1591 | dsll t_1,1 | ||
1563 | daddu c_3,t_1 | 1592 | daddu c_3,t_1 |
1564 | sltu AT,c_3,t_1 | 1593 | sltu AT,c_3,t_1 |
1565 | daddu t_2,AT | 1594 | daddu t_2,AT |
1566 | daddu c_1,t_2 | 1595 | daddu c_1,t_2 |
1567 | sltu c_2,c_1,t_2 | 1596 | sltu AT,c_1,t_2 |
1597 | daddu c_2,AT | ||
1568 | dmultu a_1,a_1 /* mul_add_c(a[1],b[1],c3,c1,c2); */ | 1598 | dmultu a_1,a_1 /* mul_add_c(a[1],b[1],c3,c1,c2); */ |
1569 | mflo t_1 | 1599 | mflo t_1 |
1570 | mfhi t_2 | 1600 | mfhi t_2 |
@@ -1579,24 +1609,26 @@ LEAF(bn_sqr_comba8) | |||
1579 | dmultu a_0,a_3 /* mul_add_c2(a[0],b[3],c1,c2,c3); */ | 1609 | dmultu a_0,a_3 /* mul_add_c2(a[0],b[3],c1,c2,c3); */ |
1580 | mflo t_1 | 1610 | mflo t_1 |
1581 | mfhi t_2 | 1611 | mfhi t_2 |
1582 | daddu c_1,t_1 | 1612 | slt c_3,t_2,zero |
1583 | sltu AT,c_1,t_1 | 1613 | dsll t_2,1 |
1584 | daddu a2,t_2,AT | 1614 | slt a2,t_1,zero |
1585 | daddu c_2,a2 | 1615 | daddu t_2,a2 |
1616 | dsll t_1,1 | ||
1586 | daddu c_1,t_1 | 1617 | daddu c_1,t_1 |
1587 | sltu AT,c_1,t_1 | 1618 | sltu AT,c_1,t_1 |
1588 | daddu t_2,AT | 1619 | daddu t_2,AT |
1589 | daddu c_2,t_2 | 1620 | daddu c_2,t_2 |
1590 | sltu c_3,c_2,t_2 | 1621 | sltu AT,c_2,t_2 |
1622 | daddu c_3,AT | ||
1591 | dmultu a_1,a_2 /* mul_add_c2(a[1],b[2],c1,c2,c3); */ | 1623 | dmultu a_1,a_2 /* mul_add_c2(a[1],b[2],c1,c2,c3); */ |
1592 | mflo t_1 | 1624 | mflo t_1 |
1593 | mfhi t_2 | 1625 | mfhi t_2 |
1594 | daddu c_1,t_1 | 1626 | slt AT,t_2,zero |
1595 | sltu AT,c_1,t_1 | ||
1596 | daddu a2,t_2,AT | ||
1597 | daddu c_2,a2 | ||
1598 | sltu AT,c_2,a2 | ||
1599 | daddu c_3,AT | 1627 | daddu c_3,AT |
1628 | dsll t_2,1 | ||
1629 | slt a2,t_1,zero | ||
1630 | daddu t_2,a2 | ||
1631 | dsll t_1,1 | ||
1600 | daddu c_1,t_1 | 1632 | daddu c_1,t_1 |
1601 | sltu AT,c_1,t_1 | 1633 | sltu AT,c_1,t_1 |
1602 | daddu t_2,AT | 1634 | daddu t_2,AT |
@@ -1608,24 +1640,26 @@ LEAF(bn_sqr_comba8) | |||
1608 | dmultu a_4,a_0 /* mul_add_c2(a[4],b[0],c2,c3,c1); */ | 1640 | dmultu a_4,a_0 /* mul_add_c2(a[4],b[0],c2,c3,c1); */ |
1609 | mflo t_1 | 1641 | mflo t_1 |
1610 | mfhi t_2 | 1642 | mfhi t_2 |
1611 | daddu c_2,t_1 | 1643 | slt c_1,t_2,zero |
1612 | sltu AT,c_2,t_1 | 1644 | dsll t_2,1 |
1613 | daddu a2,t_2,AT | 1645 | slt a2,t_1,zero |
1614 | daddu c_3,a2 | 1646 | daddu t_2,a2 |
1647 | dsll t_1,1 | ||
1615 | daddu c_2,t_1 | 1648 | daddu c_2,t_1 |
1616 | sltu AT,c_2,t_1 | 1649 | sltu AT,c_2,t_1 |
1617 | daddu t_2,AT | 1650 | daddu t_2,AT |
1618 | daddu c_3,t_2 | 1651 | daddu c_3,t_2 |
1619 | sltu c_1,c_3,t_2 | 1652 | sltu AT,c_3,t_2 |
1653 | daddu c_1,AT | ||
1620 | dmultu a_3,a_1 /* mul_add_c2(a[3],b[1],c2,c3,c1); */ | 1654 | dmultu a_3,a_1 /* mul_add_c2(a[3],b[1],c2,c3,c1); */ |
1621 | mflo t_1 | 1655 | mflo t_1 |
1622 | mfhi t_2 | 1656 | mfhi t_2 |
1623 | daddu c_2,t_1 | 1657 | slt AT,t_2,zero |
1624 | sltu AT,c_2,t_1 | ||
1625 | daddu a2,t_2,AT | ||
1626 | daddu c_3,a2 | ||
1627 | sltu AT,c_3,a2 | ||
1628 | daddu c_1,AT | 1658 | daddu c_1,AT |
1659 | dsll t_2,1 | ||
1660 | slt a2,t_1,zero | ||
1661 | daddu t_2,a2 | ||
1662 | dsll t_1,1 | ||
1629 | daddu c_2,t_1 | 1663 | daddu c_2,t_1 |
1630 | sltu AT,c_2,t_1 | 1664 | sltu AT,c_2,t_1 |
1631 | daddu t_2,AT | 1665 | daddu t_2,AT |
@@ -1646,24 +1680,26 @@ LEAF(bn_sqr_comba8) | |||
1646 | dmultu a_0,a_5 /* mul_add_c2(a[0],b[5],c3,c1,c2); */ | 1680 | dmultu a_0,a_5 /* mul_add_c2(a[0],b[5],c3,c1,c2); */ |
1647 | mflo t_1 | 1681 | mflo t_1 |
1648 | mfhi t_2 | 1682 | mfhi t_2 |
1649 | daddu c_3,t_1 | 1683 | slt c_2,t_2,zero |
1650 | sltu AT,c_3,t_1 | 1684 | dsll t_2,1 |
1651 | daddu a2,t_2,AT | 1685 | slt a2,t_1,zero |
1652 | daddu c_1,a2 | 1686 | daddu t_2,a2 |
1687 | dsll t_1,1 | ||
1653 | daddu c_3,t_1 | 1688 | daddu c_3,t_1 |
1654 | sltu AT,c_3,t_1 | 1689 | sltu AT,c_3,t_1 |
1655 | daddu t_2,AT | 1690 | daddu t_2,AT |
1656 | daddu c_1,t_2 | 1691 | daddu c_1,t_2 |
1657 | sltu c_2,c_1,t_2 | 1692 | sltu AT,c_1,t_2 |
1693 | daddu c_2,AT | ||
1658 | dmultu a_1,a_4 /* mul_add_c2(a[1],b[4],c3,c1,c2); */ | 1694 | dmultu a_1,a_4 /* mul_add_c2(a[1],b[4],c3,c1,c2); */ |
1659 | mflo t_1 | 1695 | mflo t_1 |
1660 | mfhi t_2 | 1696 | mfhi t_2 |
1661 | daddu c_3,t_1 | 1697 | slt AT,t_2,zero |
1662 | sltu AT,c_3,t_1 | ||
1663 | daddu a2,t_2,AT | ||
1664 | daddu c_1,a2 | ||
1665 | sltu AT,c_1,a2 | ||
1666 | daddu c_2,AT | 1698 | daddu c_2,AT |
1699 | dsll t_2,1 | ||
1700 | slt a2,t_1,zero | ||
1701 | daddu t_2,a2 | ||
1702 | dsll t_1,1 | ||
1667 | daddu c_3,t_1 | 1703 | daddu c_3,t_1 |
1668 | sltu AT,c_3,t_1 | 1704 | sltu AT,c_3,t_1 |
1669 | daddu t_2,AT | 1705 | daddu t_2,AT |
@@ -1673,12 +1709,12 @@ LEAF(bn_sqr_comba8) | |||
1673 | dmultu a_2,a_3 /* mul_add_c2(a[2],b[3],c3,c1,c2); */ | 1709 | dmultu a_2,a_3 /* mul_add_c2(a[2],b[3],c3,c1,c2); */ |
1674 | mflo t_1 | 1710 | mflo t_1 |
1675 | mfhi t_2 | 1711 | mfhi t_2 |
1676 | daddu c_3,t_1 | 1712 | slt AT,t_2,zero |
1677 | sltu AT,c_3,t_1 | ||
1678 | daddu a2,t_2,AT | ||
1679 | daddu c_1,a2 | ||
1680 | sltu AT,c_1,a2 | ||
1681 | daddu c_2,AT | 1713 | daddu c_2,AT |
1714 | dsll t_2,1 | ||
1715 | slt a2,t_1,zero | ||
1716 | daddu t_2,a2 | ||
1717 | dsll t_1,1 | ||
1682 | daddu c_3,t_1 | 1718 | daddu c_3,t_1 |
1683 | sltu AT,c_3,t_1 | 1719 | sltu AT,c_3,t_1 |
1684 | daddu t_2,AT | 1720 | daddu t_2,AT |
@@ -1690,24 +1726,26 @@ LEAF(bn_sqr_comba8) | |||
1690 | dmultu a_6,a_0 /* mul_add_c2(a[6],b[0],c1,c2,c3); */ | 1726 | dmultu a_6,a_0 /* mul_add_c2(a[6],b[0],c1,c2,c3); */ |
1691 | mflo t_1 | 1727 | mflo t_1 |
1692 | mfhi t_2 | 1728 | mfhi t_2 |
1693 | daddu c_1,t_1 | 1729 | slt c_3,t_2,zero |
1694 | sltu AT,c_1,t_1 | 1730 | dsll t_2,1 |
1695 | daddu a2,t_2,AT | 1731 | slt a2,t_1,zero |
1696 | daddu c_2,a2 | 1732 | daddu t_2,a2 |
1733 | dsll t_1,1 | ||
1697 | daddu c_1,t_1 | 1734 | daddu c_1,t_1 |
1698 | sltu AT,c_1,t_1 | 1735 | sltu AT,c_1,t_1 |
1699 | daddu t_2,AT | 1736 | daddu t_2,AT |
1700 | daddu c_2,t_2 | 1737 | daddu c_2,t_2 |
1701 | sltu c_3,c_2,t_2 | 1738 | sltu AT,c_2,t_2 |
1739 | daddu c_3,AT | ||
1702 | dmultu a_5,a_1 /* mul_add_c2(a[5],b[1],c1,c2,c3); */ | 1740 | dmultu a_5,a_1 /* mul_add_c2(a[5],b[1],c1,c2,c3); */ |
1703 | mflo t_1 | 1741 | mflo t_1 |
1704 | mfhi t_2 | 1742 | mfhi t_2 |
1705 | daddu c_1,t_1 | 1743 | slt AT,t_2,zero |
1706 | sltu AT,c_1,t_1 | ||
1707 | daddu a2,t_2,AT | ||
1708 | daddu c_2,a2 | ||
1709 | sltu AT,c_2,a2 | ||
1710 | daddu c_3,AT | 1744 | daddu c_3,AT |
1745 | dsll t_2,1 | ||
1746 | slt a2,t_1,zero | ||
1747 | daddu t_2,a2 | ||
1748 | dsll t_1,1 | ||
1711 | daddu c_1,t_1 | 1749 | daddu c_1,t_1 |
1712 | sltu AT,c_1,t_1 | 1750 | sltu AT,c_1,t_1 |
1713 | daddu t_2,AT | 1751 | daddu t_2,AT |
@@ -1717,12 +1755,12 @@ LEAF(bn_sqr_comba8) | |||
1717 | dmultu a_4,a_2 /* mul_add_c2(a[4],b[2],c1,c2,c3); */ | 1755 | dmultu a_4,a_2 /* mul_add_c2(a[4],b[2],c1,c2,c3); */ |
1718 | mflo t_1 | 1756 | mflo t_1 |
1719 | mfhi t_2 | 1757 | mfhi t_2 |
1720 | daddu c_1,t_1 | 1758 | slt AT,t_2,zero |
1721 | sltu AT,c_1,t_1 | ||
1722 | daddu a2,t_2,AT | ||
1723 | daddu c_2,a2 | ||
1724 | sltu AT,c_2,a2 | ||
1725 | daddu c_3,AT | 1759 | daddu c_3,AT |
1760 | dsll t_2,1 | ||
1761 | slt a2,t_1,zero | ||
1762 | daddu t_2,a2 | ||
1763 | dsll t_1,1 | ||
1726 | daddu c_1,t_1 | 1764 | daddu c_1,t_1 |
1727 | sltu AT,c_1,t_1 | 1765 | sltu AT,c_1,t_1 |
1728 | daddu t_2,AT | 1766 | daddu t_2,AT |
@@ -1743,24 +1781,26 @@ LEAF(bn_sqr_comba8) | |||
1743 | dmultu a_0,a_7 /* mul_add_c2(a[0],b[7],c2,c3,c1); */ | 1781 | dmultu a_0,a_7 /* mul_add_c2(a[0],b[7],c2,c3,c1); */ |
1744 | mflo t_1 | 1782 | mflo t_1 |
1745 | mfhi t_2 | 1783 | mfhi t_2 |
1746 | daddu c_2,t_1 | 1784 | slt c_1,t_2,zero |
1747 | sltu AT,c_2,t_1 | 1785 | dsll t_2,1 |
1748 | daddu a2,t_2,AT | 1786 | slt a2,t_1,zero |
1749 | daddu c_3,a2 | 1787 | daddu t_2,a2 |
1788 | dsll t_1,1 | ||
1750 | daddu c_2,t_1 | 1789 | daddu c_2,t_1 |
1751 | sltu AT,c_2,t_1 | 1790 | sltu AT,c_2,t_1 |
1752 | daddu t_2,AT | 1791 | daddu t_2,AT |
1753 | daddu c_3,t_2 | 1792 | daddu c_3,t_2 |
1754 | sltu c_1,c_3,t_2 | 1793 | sltu AT,c_3,t_2 |
1794 | daddu c_1,AT | ||
1755 | dmultu a_1,a_6 /* mul_add_c2(a[1],b[6],c2,c3,c1); */ | 1795 | dmultu a_1,a_6 /* mul_add_c2(a[1],b[6],c2,c3,c1); */ |
1756 | mflo t_1 | 1796 | mflo t_1 |
1757 | mfhi t_2 | 1797 | mfhi t_2 |
1758 | daddu c_2,t_1 | 1798 | slt AT,t_2,zero |
1759 | sltu AT,c_2,t_1 | ||
1760 | daddu a2,t_2,AT | ||
1761 | daddu c_3,a2 | ||
1762 | sltu AT,c_3,a2 | ||
1763 | daddu c_1,AT | 1799 | daddu c_1,AT |
1800 | dsll t_2,1 | ||
1801 | slt a2,t_1,zero | ||
1802 | daddu t_2,a2 | ||
1803 | dsll t_1,1 | ||
1764 | daddu c_2,t_1 | 1804 | daddu c_2,t_1 |
1765 | sltu AT,c_2,t_1 | 1805 | sltu AT,c_2,t_1 |
1766 | daddu t_2,AT | 1806 | daddu t_2,AT |
@@ -1770,12 +1810,12 @@ LEAF(bn_sqr_comba8) | |||
1770 | dmultu a_2,a_5 /* mul_add_c2(a[2],b[5],c2,c3,c1); */ | 1810 | dmultu a_2,a_5 /* mul_add_c2(a[2],b[5],c2,c3,c1); */ |
1771 | mflo t_1 | 1811 | mflo t_1 |
1772 | mfhi t_2 | 1812 | mfhi t_2 |
1773 | daddu c_2,t_1 | 1813 | slt AT,t_2,zero |
1774 | sltu AT,c_2,t_1 | ||
1775 | daddu a2,t_2,AT | ||
1776 | daddu c_3,a2 | ||
1777 | sltu AT,c_3,a2 | ||
1778 | daddu c_1,AT | 1814 | daddu c_1,AT |
1815 | dsll t_2,1 | ||
1816 | slt a2,t_1,zero | ||
1817 | daddu t_2,a2 | ||
1818 | dsll t_1,1 | ||
1779 | daddu c_2,t_1 | 1819 | daddu c_2,t_1 |
1780 | sltu AT,c_2,t_1 | 1820 | sltu AT,c_2,t_1 |
1781 | daddu t_2,AT | 1821 | daddu t_2,AT |
@@ -1785,12 +1825,12 @@ LEAF(bn_sqr_comba8) | |||
1785 | dmultu a_3,a_4 /* mul_add_c2(a[3],b[4],c2,c3,c1); */ | 1825 | dmultu a_3,a_4 /* mul_add_c2(a[3],b[4],c2,c3,c1); */ |
1786 | mflo t_1 | 1826 | mflo t_1 |
1787 | mfhi t_2 | 1827 | mfhi t_2 |
1788 | daddu c_2,t_1 | 1828 | slt AT,t_2,zero |
1789 | sltu AT,c_2,t_1 | ||
1790 | daddu a2,t_2,AT | ||
1791 | daddu c_3,a2 | ||
1792 | sltu AT,c_3,a2 | ||
1793 | daddu c_1,AT | 1829 | daddu c_1,AT |
1830 | dsll t_2,1 | ||
1831 | slt a2,t_1,zero | ||
1832 | daddu t_2,a2 | ||
1833 | dsll t_1,1 | ||
1794 | daddu c_2,t_1 | 1834 | daddu c_2,t_1 |
1795 | sltu AT,c_2,t_1 | 1835 | sltu AT,c_2,t_1 |
1796 | daddu t_2,AT | 1836 | daddu t_2,AT |
@@ -1802,24 +1842,26 @@ LEAF(bn_sqr_comba8) | |||
1802 | dmultu a_7,a_1 /* mul_add_c2(a[7],b[1],c3,c1,c2); */ | 1842 | dmultu a_7,a_1 /* mul_add_c2(a[7],b[1],c3,c1,c2); */ |
1803 | mflo t_1 | 1843 | mflo t_1 |
1804 | mfhi t_2 | 1844 | mfhi t_2 |
1805 | daddu c_3,t_1 | 1845 | slt c_2,t_2,zero |
1806 | sltu AT,c_3,t_1 | 1846 | dsll t_2,1 |
1807 | daddu a2,t_2,AT | 1847 | slt a2,t_1,zero |
1808 | daddu c_1,a2 | 1848 | daddu t_2,a2 |
1849 | dsll t_1,1 | ||
1809 | daddu c_3,t_1 | 1850 | daddu c_3,t_1 |
1810 | sltu AT,c_3,t_1 | 1851 | sltu AT,c_3,t_1 |
1811 | daddu t_2,AT | 1852 | daddu t_2,AT |
1812 | daddu c_1,t_2 | 1853 | daddu c_1,t_2 |
1813 | sltu c_2,c_1,t_2 | 1854 | sltu AT,c_1,t_2 |
1855 | daddu c_2,AT | ||
1814 | dmultu a_6,a_2 /* mul_add_c2(a[6],b[2],c3,c1,c2); */ | 1856 | dmultu a_6,a_2 /* mul_add_c2(a[6],b[2],c3,c1,c2); */ |
1815 | mflo t_1 | 1857 | mflo t_1 |
1816 | mfhi t_2 | 1858 | mfhi t_2 |
1817 | daddu c_3,t_1 | 1859 | slt AT,t_2,zero |
1818 | sltu AT,c_3,t_1 | ||
1819 | daddu a2,t_2,AT | ||
1820 | daddu c_1,a2 | ||
1821 | sltu AT,c_1,a2 | ||
1822 | daddu c_2,AT | 1860 | daddu c_2,AT |
1861 | dsll t_2,1 | ||
1862 | slt a2,t_1,zero | ||
1863 | daddu t_2,a2 | ||
1864 | dsll t_1,1 | ||
1823 | daddu c_3,t_1 | 1865 | daddu c_3,t_1 |
1824 | sltu AT,c_3,t_1 | 1866 | sltu AT,c_3,t_1 |
1825 | daddu t_2,AT | 1867 | daddu t_2,AT |
@@ -1829,12 +1871,12 @@ LEAF(bn_sqr_comba8) | |||
1829 | dmultu a_5,a_3 /* mul_add_c2(a[5],b[3],c3,c1,c2); */ | 1871 | dmultu a_5,a_3 /* mul_add_c2(a[5],b[3],c3,c1,c2); */ |
1830 | mflo t_1 | 1872 | mflo t_1 |
1831 | mfhi t_2 | 1873 | mfhi t_2 |
1832 | daddu c_3,t_1 | 1874 | slt AT,t_2,zero |
1833 | sltu AT,c_3,t_1 | ||
1834 | daddu a2,t_2,AT | ||
1835 | daddu c_1,a2 | ||
1836 | sltu AT,c_1,a2 | ||
1837 | daddu c_2,AT | 1875 | daddu c_2,AT |
1876 | dsll t_2,1 | ||
1877 | slt a2,t_1,zero | ||
1878 | daddu t_2,a2 | ||
1879 | dsll t_1,1 | ||
1838 | daddu c_3,t_1 | 1880 | daddu c_3,t_1 |
1839 | sltu AT,c_3,t_1 | 1881 | sltu AT,c_3,t_1 |
1840 | daddu t_2,AT | 1882 | daddu t_2,AT |
@@ -1855,24 +1897,26 @@ LEAF(bn_sqr_comba8) | |||
1855 | dmultu a_2,a_7 /* mul_add_c2(a[2],b[7],c1,c2,c3); */ | 1897 | dmultu a_2,a_7 /* mul_add_c2(a[2],b[7],c1,c2,c3); */ |
1856 | mflo t_1 | 1898 | mflo t_1 |
1857 | mfhi t_2 | 1899 | mfhi t_2 |
1858 | daddu c_1,t_1 | 1900 | slt c_3,t_2,zero |
1859 | sltu AT,c_1,t_1 | 1901 | dsll t_2,1 |
1860 | daddu a2,t_2,AT | 1902 | slt a2,t_1,zero |
1861 | daddu c_2,a2 | 1903 | daddu t_2,a2 |
1904 | dsll t_1,1 | ||
1862 | daddu c_1,t_1 | 1905 | daddu c_1,t_1 |
1863 | sltu AT,c_1,t_1 | 1906 | sltu AT,c_1,t_1 |
1864 | daddu t_2,AT | 1907 | daddu t_2,AT |
1865 | daddu c_2,t_2 | 1908 | daddu c_2,t_2 |
1866 | sltu c_3,c_2,t_2 | 1909 | sltu AT,c_2,t_2 |
1910 | daddu c_3,AT | ||
1867 | dmultu a_3,a_6 /* mul_add_c2(a[3],b[6],c1,c2,c3); */ | 1911 | dmultu a_3,a_6 /* mul_add_c2(a[3],b[6],c1,c2,c3); */ |
1868 | mflo t_1 | 1912 | mflo t_1 |
1869 | mfhi t_2 | 1913 | mfhi t_2 |
1870 | daddu c_1,t_1 | 1914 | slt AT,t_2,zero |
1871 | sltu AT,c_1,t_1 | ||
1872 | daddu a2,t_2,AT | ||
1873 | daddu c_2,a2 | ||
1874 | sltu AT,c_2,a2 | ||
1875 | daddu c_3,AT | 1915 | daddu c_3,AT |
1916 | dsll t_2,1 | ||
1917 | slt a2,t_1,zero | ||
1918 | daddu t_2,a2 | ||
1919 | dsll t_1,1 | ||
1876 | daddu c_1,t_1 | 1920 | daddu c_1,t_1 |
1877 | sltu AT,c_1,t_1 | 1921 | sltu AT,c_1,t_1 |
1878 | daddu t_2,AT | 1922 | daddu t_2,AT |
@@ -1882,12 +1926,12 @@ LEAF(bn_sqr_comba8) | |||
1882 | dmultu a_4,a_5 /* mul_add_c2(a[4],b[5],c1,c2,c3); */ | 1926 | dmultu a_4,a_5 /* mul_add_c2(a[4],b[5],c1,c2,c3); */ |
1883 | mflo t_1 | 1927 | mflo t_1 |
1884 | mfhi t_2 | 1928 | mfhi t_2 |
1885 | daddu c_1,t_1 | 1929 | slt AT,t_2,zero |
1886 | sltu AT,c_1,t_1 | ||
1887 | daddu a2,t_2,AT | ||
1888 | daddu c_2,a2 | ||
1889 | sltu AT,c_2,a2 | ||
1890 | daddu c_3,AT | 1930 | daddu c_3,AT |
1931 | dsll t_2,1 | ||
1932 | slt a2,t_1,zero | ||
1933 | daddu t_2,a2 | ||
1934 | dsll t_1,1 | ||
1891 | daddu c_1,t_1 | 1935 | daddu c_1,t_1 |
1892 | sltu AT,c_1,t_1 | 1936 | sltu AT,c_1,t_1 |
1893 | daddu t_2,AT | 1937 | daddu t_2,AT |
@@ -1899,24 +1943,26 @@ LEAF(bn_sqr_comba8) | |||
1899 | dmultu a_7,a_3 /* mul_add_c2(a[7],b[3],c2,c3,c1); */ | 1943 | dmultu a_7,a_3 /* mul_add_c2(a[7],b[3],c2,c3,c1); */ |
1900 | mflo t_1 | 1944 | mflo t_1 |
1901 | mfhi t_2 | 1945 | mfhi t_2 |
1902 | daddu c_2,t_1 | 1946 | slt c_1,t_2,zero |
1903 | sltu AT,c_2,t_1 | 1947 | dsll t_2,1 |
1904 | daddu a2,t_2,AT | 1948 | slt a2,t_1,zero |
1905 | daddu c_3,a2 | 1949 | daddu t_2,a2 |
1950 | dsll t_1,1 | ||
1906 | daddu c_2,t_1 | 1951 | daddu c_2,t_1 |
1907 | sltu AT,c_2,t_1 | 1952 | sltu AT,c_2,t_1 |
1908 | daddu t_2,AT | 1953 | daddu t_2,AT |
1909 | daddu c_3,t_2 | 1954 | daddu c_3,t_2 |
1910 | sltu c_1,c_3,t_2 | 1955 | sltu AT,c_3,t_2 |
1956 | daddu c_1,AT | ||
1911 | dmultu a_6,a_4 /* mul_add_c2(a[6],b[4],c2,c3,c1); */ | 1957 | dmultu a_6,a_4 /* mul_add_c2(a[6],b[4],c2,c3,c1); */ |
1912 | mflo t_1 | 1958 | mflo t_1 |
1913 | mfhi t_2 | 1959 | mfhi t_2 |
1914 | daddu c_2,t_1 | 1960 | slt AT,t_2,zero |
1915 | sltu AT,c_2,t_1 | ||
1916 | daddu a2,t_2,AT | ||
1917 | daddu c_3,a2 | ||
1918 | sltu AT,c_3,a2 | ||
1919 | daddu c_1,AT | 1961 | daddu c_1,AT |
1962 | dsll t_2,1 | ||
1963 | slt a2,t_1,zero | ||
1964 | daddu t_2,a2 | ||
1965 | dsll t_1,1 | ||
1920 | daddu c_2,t_1 | 1966 | daddu c_2,t_1 |
1921 | sltu AT,c_2,t_1 | 1967 | sltu AT,c_2,t_1 |
1922 | daddu t_2,AT | 1968 | daddu t_2,AT |
@@ -1937,24 +1983,26 @@ LEAF(bn_sqr_comba8) | |||
1937 | dmultu a_4,a_7 /* mul_add_c2(a[4],b[7],c3,c1,c2); */ | 1983 | dmultu a_4,a_7 /* mul_add_c2(a[4],b[7],c3,c1,c2); */ |
1938 | mflo t_1 | 1984 | mflo t_1 |
1939 | mfhi t_2 | 1985 | mfhi t_2 |
1940 | daddu c_3,t_1 | 1986 | slt c_2,t_2,zero |
1941 | sltu AT,c_3,t_1 | 1987 | dsll t_2,1 |
1942 | daddu a2,t_2,AT | 1988 | slt a2,t_1,zero |
1943 | daddu c_1,a2 | 1989 | daddu t_2,a2 |
1990 | dsll t_1,1 | ||
1944 | daddu c_3,t_1 | 1991 | daddu c_3,t_1 |
1945 | sltu AT,c_3,t_1 | 1992 | sltu AT,c_3,t_1 |
1946 | daddu t_2,AT | 1993 | daddu t_2,AT |
1947 | daddu c_1,t_2 | 1994 | daddu c_1,t_2 |
1948 | sltu c_2,c_1,t_2 | 1995 | sltu AT,c_1,t_2 |
1996 | daddu c_2,AT | ||
1949 | dmultu a_5,a_6 /* mul_add_c2(a[5],b[6],c3,c1,c2); */ | 1997 | dmultu a_5,a_6 /* mul_add_c2(a[5],b[6],c3,c1,c2); */ |
1950 | mflo t_1 | 1998 | mflo t_1 |
1951 | mfhi t_2 | 1999 | mfhi t_2 |
1952 | daddu c_3,t_1 | 2000 | slt AT,t_2,zero |
1953 | sltu AT,c_3,t_1 | ||
1954 | daddu a2,t_2,AT | ||
1955 | daddu c_1,a2 | ||
1956 | sltu AT,c_1,a2 | ||
1957 | daddu c_2,AT | 2001 | daddu c_2,AT |
2002 | dsll t_2,1 | ||
2003 | slt a2,t_1,zero | ||
2004 | daddu t_2,a2 | ||
2005 | dsll t_1,1 | ||
1958 | daddu c_3,t_1 | 2006 | daddu c_3,t_1 |
1959 | sltu AT,c_3,t_1 | 2007 | sltu AT,c_3,t_1 |
1960 | daddu t_2,AT | 2008 | daddu t_2,AT |
@@ -1966,15 +2014,17 @@ LEAF(bn_sqr_comba8) | |||
1966 | dmultu a_7,a_5 /* mul_add_c2(a[7],b[5],c1,c2,c3); */ | 2014 | dmultu a_7,a_5 /* mul_add_c2(a[7],b[5],c1,c2,c3); */ |
1967 | mflo t_1 | 2015 | mflo t_1 |
1968 | mfhi t_2 | 2016 | mfhi t_2 |
1969 | daddu c_1,t_1 | 2017 | slt c_3,t_2,zero |
1970 | sltu AT,c_1,t_1 | 2018 | dsll t_2,1 |
1971 | daddu a2,t_2,AT | 2019 | slt a2,t_1,zero |
1972 | daddu c_2,a2 | 2020 | daddu t_2,a2 |
2021 | dsll t_1,1 | ||
1973 | daddu c_1,t_1 | 2022 | daddu c_1,t_1 |
1974 | sltu AT,c_1,t_1 | 2023 | sltu AT,c_1,t_1 |
1975 | daddu t_2,AT | 2024 | daddu t_2,AT |
1976 | daddu c_2,t_2 | 2025 | daddu c_2,t_2 |
1977 | sltu c_3,c_2,t_2 | 2026 | sltu AT,c_2,t_2 |
2027 | daddu c_3,AT | ||
1978 | dmultu a_6,a_6 /* mul_add_c(a[6],b[6],c1,c2,c3); */ | 2028 | dmultu a_6,a_6 /* mul_add_c(a[6],b[6],c1,c2,c3); */ |
1979 | mflo t_1 | 2029 | mflo t_1 |
1980 | mfhi t_2 | 2030 | mfhi t_2 |
@@ -1989,15 +2039,17 @@ LEAF(bn_sqr_comba8) | |||
1989 | dmultu a_6,a_7 /* mul_add_c2(a[6],b[7],c2,c3,c1); */ | 2039 | dmultu a_6,a_7 /* mul_add_c2(a[6],b[7],c2,c3,c1); */ |
1990 | mflo t_1 | 2040 | mflo t_1 |
1991 | mfhi t_2 | 2041 | mfhi t_2 |
1992 | daddu c_2,t_1 | 2042 | slt c_1,t_2,zero |
1993 | sltu AT,c_2,t_1 | 2043 | dsll t_2,1 |
1994 | daddu a2,t_2,AT | 2044 | slt a2,t_1,zero |
1995 | daddu c_3,a2 | 2045 | daddu t_2,a2 |
2046 | dsll t_1,1 | ||
1996 | daddu c_2,t_1 | 2047 | daddu c_2,t_1 |
1997 | sltu AT,c_2,t_1 | 2048 | sltu AT,c_2,t_1 |
1998 | daddu t_2,AT | 2049 | daddu t_2,AT |
1999 | daddu c_3,t_2 | 2050 | daddu c_3,t_2 |
2000 | sltu c_1,c_3,t_2 | 2051 | sltu AT,c_3,t_2 |
2052 | daddu c_1,AT | ||
2001 | sd c_2,104(a0) | 2053 | sd c_2,104(a0) |
2002 | 2054 | ||
2003 | dmultu a_7,a_7 /* mul_add_c(a[7],b[7],c3,c1,c2); */ | 2055 | dmultu a_7,a_7 /* mul_add_c(a[7],b[7],c3,c1,c2); */ |
@@ -2028,28 +2080,30 @@ LEAF(bn_sqr_comba4) | |||
2028 | dmultu a_0,a_1 /* mul_add_c2(a[0],b[1],c2,c3,c1); */ | 2080 | dmultu a_0,a_1 /* mul_add_c2(a[0],b[1],c2,c3,c1); */ |
2029 | mflo t_1 | 2081 | mflo t_1 |
2030 | mfhi t_2 | 2082 | mfhi t_2 |
2083 | slt c_1,t_2,zero | ||
2084 | dsll t_2,1 | ||
2085 | slt a2,t_1,zero | ||
2086 | daddu t_2,a2 | ||
2087 | dsll t_1,1 | ||
2031 | daddu c_2,t_1 | 2088 | daddu c_2,t_1 |
2032 | sltu AT,c_2,t_1 | 2089 | sltu AT,c_2,t_1 |
2033 | daddu c_3,t_2,AT | 2090 | daddu c_3,t_2,AT |
2034 | daddu c_2,t_1 | ||
2035 | sltu AT,c_2,t_1 | ||
2036 | daddu t_2,AT | ||
2037 | daddu c_3,t_2 | ||
2038 | sltu c_1,c_3,t_2 | ||
2039 | sd c_2,8(a0) | 2091 | sd c_2,8(a0) |
2040 | 2092 | ||
2041 | dmultu a_2,a_0 /* mul_add_c2(a[2],b[0],c3,c1,c2); */ | 2093 | dmultu a_2,a_0 /* mul_add_c2(a[2],b[0],c3,c1,c2); */ |
2042 | mflo t_1 | 2094 | mflo t_1 |
2043 | mfhi t_2 | 2095 | mfhi t_2 |
2044 | daddu c_3,t_1 | 2096 | slt c_2,t_2,zero |
2045 | sltu AT,c_3,t_1 | 2097 | dsll t_2,1 |
2046 | daddu a2,t_2,AT | 2098 | slt a2,t_1,zero |
2047 | daddu c_1,a2 | 2099 | daddu t_2,a2 |
2100 | dsll t_1,1 | ||
2048 | daddu c_3,t_1 | 2101 | daddu c_3,t_1 |
2049 | sltu AT,c_3,t_1 | 2102 | sltu AT,c_3,t_1 |
2050 | daddu t_2,AT | 2103 | daddu t_2,AT |
2051 | daddu c_1,t_2 | 2104 | daddu c_1,t_2 |
2052 | sltu c_2,c_1,t_2 | 2105 | sltu AT,c_1,t_2 |
2106 | daddu c_2,AT | ||
2053 | dmultu a_1,a_1 /* mul_add_c(a[1],b[1],c3,c1,c2); */ | 2107 | dmultu a_1,a_1 /* mul_add_c(a[1],b[1],c3,c1,c2); */ |
2054 | mflo t_1 | 2108 | mflo t_1 |
2055 | mfhi t_2 | 2109 | mfhi t_2 |
@@ -2064,24 +2118,26 @@ LEAF(bn_sqr_comba4) | |||
2064 | dmultu a_0,a_3 /* mul_add_c2(a[0],b[3],c1,c2,c3); */ | 2118 | dmultu a_0,a_3 /* mul_add_c2(a[0],b[3],c1,c2,c3); */ |
2065 | mflo t_1 | 2119 | mflo t_1 |
2066 | mfhi t_2 | 2120 | mfhi t_2 |
2067 | daddu c_1,t_1 | 2121 | slt c_3,t_2,zero |
2068 | sltu AT,c_1,t_1 | 2122 | dsll t_2,1 |
2069 | daddu a2,t_2,AT | 2123 | slt a2,t_1,zero |
2070 | daddu c_2,a2 | 2124 | daddu t_2,a2 |
2125 | dsll t_1,1 | ||
2071 | daddu c_1,t_1 | 2126 | daddu c_1,t_1 |
2072 | sltu AT,c_1,t_1 | 2127 | sltu AT,c_1,t_1 |
2073 | daddu t_2,AT | 2128 | daddu t_2,AT |
2074 | daddu c_2,t_2 | 2129 | daddu c_2,t_2 |
2075 | sltu c_3,c_2,t_2 | 2130 | sltu AT,c_2,t_2 |
2131 | daddu c_3,AT | ||
2076 | dmultu a_1,a_2 /* mul_add_c(a2[1],b[2],c1,c2,c3); */ | 2132 | dmultu a_1,a_2 /* mul_add_c(a2[1],b[2],c1,c2,c3); */ |
2077 | mflo t_1 | 2133 | mflo t_1 |
2078 | mfhi t_2 | 2134 | mfhi t_2 |
2079 | daddu c_1,t_1 | 2135 | slt AT,t_2,zero |
2080 | sltu AT,c_1,t_1 | ||
2081 | daddu a2,t_2,AT | ||
2082 | daddu c_2,a2 | ||
2083 | sltu AT,c_2,a2 | ||
2084 | daddu c_3,AT | 2136 | daddu c_3,AT |
2137 | dsll t_2,1 | ||
2138 | slt a2,t_1,zero | ||
2139 | daddu t_2,a2 | ||
2140 | dsll t_1,1 | ||
2085 | daddu c_1,t_1 | 2141 | daddu c_1,t_1 |
2086 | sltu AT,c_1,t_1 | 2142 | sltu AT,c_1,t_1 |
2087 | daddu t_2,AT | 2143 | daddu t_2,AT |
@@ -2093,15 +2149,17 @@ LEAF(bn_sqr_comba4) | |||
2093 | dmultu a_3,a_1 /* mul_add_c2(a[3],b[1],c2,c3,c1); */ | 2149 | dmultu a_3,a_1 /* mul_add_c2(a[3],b[1],c2,c3,c1); */ |
2094 | mflo t_1 | 2150 | mflo t_1 |
2095 | mfhi t_2 | 2151 | mfhi t_2 |
2096 | daddu c_2,t_1 | 2152 | slt c_1,t_2,zero |
2097 | sltu AT,c_2,t_1 | 2153 | dsll t_2,1 |
2098 | daddu a2,t_2,AT | 2154 | slt a2,t_1,zero |
2099 | daddu c_3,a2 | 2155 | daddu t_2,a2 |
2156 | dsll t_1,1 | ||
2100 | daddu c_2,t_1 | 2157 | daddu c_2,t_1 |
2101 | sltu AT,c_2,t_1 | 2158 | sltu AT,c_2,t_1 |
2102 | daddu t_2,AT | 2159 | daddu t_2,AT |
2103 | daddu c_3,t_2 | 2160 | daddu c_3,t_2 |
2104 | sltu c_1,c_3,t_2 | 2161 | sltu AT,c_3,t_2 |
2162 | daddu c_1,AT | ||
2105 | dmultu a_2,a_2 /* mul_add_c(a[2],b[2],c2,c3,c1); */ | 2163 | dmultu a_2,a_2 /* mul_add_c(a[2],b[2],c2,c3,c1); */ |
2106 | mflo t_1 | 2164 | mflo t_1 |
2107 | mfhi t_2 | 2165 | mfhi t_2 |
@@ -2116,15 +2174,17 @@ LEAF(bn_sqr_comba4) | |||
2116 | dmultu a_2,a_3 /* mul_add_c2(a[2],b[3],c3,c1,c2); */ | 2174 | dmultu a_2,a_3 /* mul_add_c2(a[2],b[3],c3,c1,c2); */ |
2117 | mflo t_1 | 2175 | mflo t_1 |
2118 | mfhi t_2 | 2176 | mfhi t_2 |
2119 | daddu c_3,t_1 | 2177 | slt c_2,t_2,zero |
2120 | sltu AT,c_3,t_1 | 2178 | dsll t_2,1 |
2121 | daddu a2,t_2,AT | 2179 | slt a2,t_1,zero |
2122 | daddu c_1,a2 | 2180 | daddu t_2,a2 |
2181 | dsll t_1,1 | ||
2123 | daddu c_3,t_1 | 2182 | daddu c_3,t_1 |
2124 | sltu AT,c_3,t_1 | 2183 | sltu AT,c_3,t_1 |
2125 | daddu t_2,AT | 2184 | daddu t_2,AT |
2126 | daddu c_1,t_2 | 2185 | daddu c_1,t_2 |
2127 | sltu c_2,c_1,t_2 | 2186 | sltu AT,c_1,t_2 |
2187 | daddu c_2,AT | ||
2128 | sd c_3,40(a0) | 2188 | sd c_3,40(a0) |
2129 | 2189 | ||
2130 | dmultu a_3,a_3 /* mul_add_c(a[3],b[3],c1,c2,c3); */ | 2190 | dmultu a_3,a_3 /* mul_add_c(a[3],b[3],c1,c2,c3); */ |