| ... | ... | @@ -633,13 +633,13 @@ static inline uint64_t zig_addo_u64(uint64_t lhs, uint64_t rhs, uint64_t *res, u |
| 633 | 633 | } |
| 634 | 634 | |
| 635 | 635 | static inline uint128_t zig_addo_u128(uint128_t lhs, uint128_t rhs, uint128_t *res, uint128_t max) { |
| 636 | | bool overflow; |
| 636 | int overflow; |
| 637 | 637 | *res = __uaddoti4(lhs, rhs, &overflow); |
| 638 | | if (*res > max && !overflow) { |
| 638 | if (*res > max && overflow == 0) { |
| 639 | 639 | *res -= max - 1; |
| 640 | 640 | return true; |
| 641 | 641 | } |
| 642 | | return overflow; |
| 642 | return overflow != 0; |
| 643 | 643 | } |
| 644 | 644 | |
| 645 | 645 | static inline bool zig_subo_i8(int8_t lhs, int8_t rhs, int8_t *res, int8_t min, int8_t max) { |
| ... | ... | @@ -1095,128 +1095,6 @@ static inline uint128_t zig_mulo_u128(uint128_t lhs, uint128_t rhs, uint128_t *r |
| 1095 | 1095 | return overflow != 0; |
| 1096 | 1096 | } |
| 1097 | 1097 | |
| 1098 | | static inline bool zig_shlo_i8(int8_t lhs, int8_t rhs, int8_t *res, int8_t min, int8_t max) { |
| 1099 | | int16_t big_result = (int16_t)lhs << (int16_t)rhs; |
| 1100 | | if (big_result > max) { |
| 1101 | | *res = big_result - ((int16_t)max - (int16_t)min); |
| 1102 | | return true; |
| 1103 | | } |
| 1104 | | if (big_result < min) { |
| 1105 | | *res = big_result + ((int16_t)max - (int16_t)min); |
| 1106 | | return true; |
| 1107 | | } |
| 1108 | | *res = big_result; |
| 1109 | | return false; |
| 1110 | | } |
| 1111 | | |
| 1112 | | static inline bool zig_shlo_i16(int16_t lhs, int16_t rhs, int16_t *res, int16_t min, int16_t max) { |
| 1113 | | int32_t big_result = (int32_t)lhs << (int32_t)rhs; |
| 1114 | | if (big_result > max) { |
| 1115 | | *res = big_result - ((int32_t)max - (int32_t)min); |
| 1116 | | return true; |
| 1117 | | } |
| 1118 | | if (big_result < min) { |
| 1119 | | *res = big_result + ((int32_t)max - (int32_t)min); |
| 1120 | | return true; |
| 1121 | | } |
| 1122 | | *res = big_result; |
| 1123 | | return false; |
| 1124 | | } |
| 1125 | | |
| 1126 | | static inline bool zig_shlo_i32(int32_t lhs, int32_t rhs, int32_t *res, int32_t min, int32_t max) { |
| 1127 | | int64_t big_result = (int64_t)lhs << (int64_t)rhs; |
| 1128 | | if (big_result > max) { |
| 1129 | | *res = big_result - ((int64_t)max - (int64_t)min); |
| 1130 | | return true; |
| 1131 | | } |
| 1132 | | if (big_result < min) { |
| 1133 | | *res = big_result + ((int64_t)max - (int64_t)min); |
| 1134 | | return true; |
| 1135 | | } |
| 1136 | | *res = big_result; |
| 1137 | | return false; |
| 1138 | | } |
| 1139 | | |
| 1140 | | static inline bool zig_shlo_i64(int64_t lhs, int64_t rhs, int64_t *res, int64_t min, int64_t max) { |
| 1141 | | int overflow; |
| 1142 | | *res = __shlodi4(lhs, rhs, &overflow); |
| 1143 | | if (overflow == 0) { |
| 1144 | | if (*res > max) { |
| 1145 | | // TODO adjust the result to be the truncated bits |
| 1146 | | return true; |
| 1147 | | } else if (*res < min) { |
| 1148 | | // TODO adjust the result to be the truncated bits |
| 1149 | | return true; |
| 1150 | | } |
| 1151 | | } |
| 1152 | | return overflow != 0; |
| 1153 | | } |
| 1154 | | |
| 1155 | | static inline bool zig_shlo_i128(int128_t lhs, int128_t rhs, int128_t *res, int128_t min, int128_t max) { |
| 1156 | | int overflow; |
| 1157 | | *res = __shloti4(lhs, rhs, &overflow); |
| 1158 | | if (overflow == 0) { |
| 1159 | | if (*res > max) { |
| 1160 | | // TODO adjust the result to be the truncated bits |
| 1161 | | return true; |
| 1162 | | } else if (*res < min) { |
| 1163 | | // TODO adjust the result to be the truncated bits |
| 1164 | | return true; |
| 1165 | | } |
| 1166 | | } |
| 1167 | | return overflow != 0; |
| 1168 | | } |
| 1169 | | |
| 1170 | | static inline bool zig_shlo_u8(uint8_t lhs, uint8_t rhs, uint8_t *res, uint8_t max) { |
| 1171 | | uint16_t big_result = (uint16_t)lhs << (uint16_t)rhs; |
| 1172 | | if (big_result > max) { |
| 1173 | | *res = big_result - max - 1; |
| 1174 | | return true; |
| 1175 | | } |
| 1176 | | *res = big_result; |
| 1177 | | return false; |
| 1178 | | } |
| 1179 | | |
| 1180 | | static inline uint16_t zig_shlo_u16(uint16_t lhs, uint16_t rhs, uint16_t *res, uint16_t max) { |
| 1181 | | uint32_t big_result = (uint32_t)lhs << (uint32_t)rhs; |
| 1182 | | if (big_result > max) { |
| 1183 | | *res = big_result - max - 1; |
| 1184 | | return true; |
| 1185 | | } |
| 1186 | | *res = big_result; |
| 1187 | | return false; |
| 1188 | | } |
| 1189 | | |
| 1190 | | static inline uint32_t zig_shlo_u32(uint32_t lhs, uint32_t rhs, uint32_t *res, uint32_t max) { |
| 1191 | | uint64_t big_result = (uint64_t)lhs << (uint64_t)rhs; |
| 1192 | | if (big_result > max) { |
| 1193 | | *res = big_result - max - 1; |
| 1194 | | return true; |
| 1195 | | } |
| 1196 | | *res = big_result; |
| 1197 | | return false; |
| 1198 | | } |
| 1199 | | |
| 1200 | | static inline uint64_t zig_shlo_u64(uint64_t lhs, uint64_t rhs, uint64_t *res, uint64_t max) { |
| 1201 | | int overflow; |
| 1202 | | *res = __ushlodi4(lhs, rhs, &overflow); |
| 1203 | | if (*res > max && overflow == 0) { |
| 1204 | | *res -= max - 1; |
| 1205 | | return true; |
| 1206 | | } |
| 1207 | | return overflow != 0; |
| 1208 | | } |
| 1209 | | |
| 1210 | | static inline uint128_t zig_shlo_u128(uint128_t lhs, uint128_t rhs, uint128_t *res, uint128_t max) { |
| 1211 | | int overflow; |
| 1212 | | *res = __ushloti4(lhs, rhs, &overflow); |
| 1213 | | if (*res > max && overflow == 0) { |
| 1214 | | *res -= max - 1; |
| 1215 | | return true; |
| 1216 | | } |
| 1217 | | return overflow != 0; |
| 1218 | | } |
| 1219 | | |
| 1220 | 1098 | static inline float zig_bitcast_f32_u32(uint32_t arg) { |
| 1221 | 1099 | float dest; |
| 1222 | 1100 | memcpy(&dest, &arg, sizeof dest); |
| ... | ... | @@ -1429,6 +1307,76 @@ static inline int zig_popcount_u128(uint128_t value, uint8_t zig_type_bit_width) |
| 1429 | 1307 | |
| 1430 | 1308 | #define zig_popcount_i128 zig_popcount_u128 |
| 1431 | 1309 | |
| 1310 | static inline bool zig_shlo_i8(int8_t lhs, int8_t rhs, int8_t *res, uint8_t bits) { |
| 1311 | *res = lhs << rhs; |
| 1312 | if (zig_clz_i8(lhs, bits) >= rhs) return false; |
| 1313 | *res &= UINT8_MAX >> (8 - bits); |
| 1314 | return true; |
| 1315 | } |
| 1316 | |
| 1317 | static inline bool zig_shlo_i16(int16_t lhs, int16_t rhs, int16_t *res, uint8_t bits) { |
| 1318 | *res = lhs << rhs; |
| 1319 | if (zig_clz_i16(lhs, bits) >= rhs) return false; |
| 1320 | *res &= UINT16_MAX >> (16 - bits); |
| 1321 | return true; |
| 1322 | } |
| 1323 | |
| 1324 | static inline bool zig_shlo_i32(int32_t lhs, int32_t rhs, int32_t *res, uint8_t bits) { |
| 1325 | *res = lhs << rhs; |
| 1326 | if (zig_clz_i32(lhs, bits) >= rhs) return false; |
| 1327 | *res &= UINT32_MAX >> (32 - bits); |
| 1328 | return true; |
| 1329 | } |
| 1330 | |
| 1331 | static inline bool zig_shlo_i64(int64_t lhs, int64_t rhs, int64_t *res, uint8_t bits) { |
| 1332 | *res = lhs << rhs; |
| 1333 | if (zig_clz_i64(lhs, bits) >= rhs) return false; |
| 1334 | *res &= UINT64_MAX >> (64 - bits); |
| 1335 | return true; |
| 1336 | } |
| 1337 | |
| 1338 | static inline bool zig_shlo_i128(int128_t lhs, int128_t rhs, int128_t *res, uint8_t bits) { |
| 1339 | *res = lhs << rhs; |
| 1340 | if (zig_clz_i128(lhs, bits) >= rhs) return false; |
| 1341 | *res &= UINT128_MAX >> (128 - bits); |
| 1342 | return true; |
| 1343 | } |
| 1344 | |
| 1345 | static inline bool zig_shlo_u8(uint8_t lhs, uint8_t rhs, uint8_t *res, uint8_t bits) { |
| 1346 | *res = lhs << rhs; |
| 1347 | if (zig_clz_u8(lhs, bits) >= rhs) return false; |
| 1348 | *res &= UINT8_MAX >> (8 - bits); |
| 1349 | return true; |
| 1350 | } |
| 1351 | |
| 1352 | static inline uint16_t zig_shlo_u16(uint16_t lhs, uint16_t rhs, uint16_t *res, uint8_t bits) { |
| 1353 | *res = lhs << rhs; |
| 1354 | if (zig_clz_u16(lhs, bits) >= rhs) return false; |
| 1355 | *res &= UINT16_MAX >> (16 - bits); |
| 1356 | return true; |
| 1357 | } |
| 1358 | |
| 1359 | static inline uint32_t zig_shlo_u32(uint32_t lhs, uint32_t rhs, uint32_t *res, uint8_t bits) { |
| 1360 | *res = lhs << rhs; |
| 1361 | if (zig_clz_u32(lhs, bits) >= rhs) return false; |
| 1362 | *res &= UINT32_MAX >> (32 - bits); |
| 1363 | return true; |
| 1364 | } |
| 1365 | |
| 1366 | static inline uint64_t zig_shlo_u64(uint64_t lhs, uint64_t rhs, uint64_t *res, uint8_t bits) { |
| 1367 | *res = lhs << rhs; |
| 1368 | if (zig_clz_u64(lhs, bits) >= rhs) return false; |
| 1369 | *res &= UINT64_MAX >> (64 - bits); |
| 1370 | return true; |
| 1371 | } |
| 1372 | |
| 1373 | static inline uint128_t zig_shlo_u128(uint128_t lhs, uint128_t rhs, uint128_t *res, uint8_t bits) { |
| 1374 | *res = lhs << rhs; |
| 1375 | if (zig_clz_u128(lhs, bits) >= rhs) return false; |
| 1376 | *res &= UINT128_MAX >> (128 - bits); |
| 1377 | return true; |
| 1378 | } |
| 1379 | |
| 1432 | 1380 | #define zig_sign_extend(T) \ |
| 1433 | 1381 | static inline T zig_sign_extend_##T(T value, uint8_t zig_type_bit_width) { \ |
| 1434 | 1382 | const T m = (T)1 << (T)(zig_type_bit_width - 1); \ |