1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866 4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930 4931 4932 4933 4934 4935 4936 4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976 4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137 5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297 5298 5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309 5310 5311 5312 5313 5314 5315 5316 5317 5318 5319 5320 5321 5322 5323 5324 5325 5326 5327 5328 5329 5330 5331 5332 5333 5334 5335 5336 5337 5338 5339 5340 5341 5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355 5356 5357 5358 5359 5360 5361 5362 5363 5364 5365 5366 5367 5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404 5405 5406 5407 5408 5409 5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436 5437 5438 5439 5440 5441 5442 5443 5444 5445 5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498 5499 5500 5501 5502 5503 5504 5505 5506 5507 5508 5509 5510 5511 5512 5513 5514 5515 5516 5517 5518 5519 5520 5521 5522 5523 5524 5525 5526 5527 5528 5529 5530 5531 5532 5533 5534 5535 5536 5537 5538 5539 5540 5541 5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553 5554 5555 5556 5557 5558 5559 5560 5561 5562 5563 5564 5565 5566 5567 5568 5569 5570 5571 5572 5573 5574 5575 5576 5577 5578 5579 5580 5581 5582 5583 5584 5585 5586 5587 5588 5589 5590 5591 5592 5593 5594 5595 5596 5597 5598 5599 5600 5601 5602 5603 5604 5605 5606 5607 5608 5609 5610 5611 5612 5613 5614 5615 5616 5617 5618 5619 5620 5621 5622 5623 5624 5625 5626 5627 5628 5629 5630 5631 5632 5633 5634 5635 5636 5637 5638 5639 5640 5641 5642 5643 5644 5645 5646 5647 5648 5649 5650 5651 5652 5653 5654 5655 5656 5657 5658 5659 5660 5661 5662 5663 5664 5665 5666 5667 5668 5669 5670 5671 5672 5673 5674 5675 5676 5677 5678 5679 5680 5681 5682 5683 5684 5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707 5708 5709 5710 5711 5712 5713 5714 5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733 5734 5735 5736 5737 5738 5739 5740 5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751 5752 5753 5754 5755 5756 5757 5758 5759 5760 5761 5762 5763 5764 5765 5766 5767 5768 5769 5770 5771 5772 5773 5774 5775 5776 5777 5778 5779 5780 5781 5782 5783 5784 5785 5786 5787 5788 5789 5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803 5804 5805 5806 5807 5808 5809 5810 5811 5812 5813 5814 5815 5816 5817 5818 5819 5820 5821 5822 5823 5824 5825 5826 5827 5828 5829 5830 5831 5832 5833 5834 5835 5836 5837 5838 5839 5840 5841 5842 5843 5844 5845 5846 5847 5848 5849 5850 5851 5852 5853 5854 5855 5856 5857 5858 5859 5860 5861 5862 5863 5864 5865 5866 5867 5868 5869 5870 5871 5872 5873 5874 5875 5876 5877 5878 5879 5880 5881 5882 5883 5884 5885 5886 5887 5888 5889 5890 5891 5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902 5903 5904 5905 5906 5907 5908 5909 5910 5911 5912 5913 5914 5915 5916 5917 5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929 5930 5931 5932 5933 5934 5935 5936 5937 5938 5939 5940 5941 5942 5943 5944 5945 5946 5947 5948 5949 5950 5951 5952 5953 5954 5955 5956 5957 5958 5959 5960 5961 5962 5963 5964 5965 5966 5967 5968 5969 5970 5971 5972 5973 5974 5975 5976 5977 5978 5979 5980 5981 5982 5983 5984 5985 5986 5987 5988 5989 5990 5991 5992 5993 5994 5995 5996 5997 5998 5999 6000 6001 6002 6003 6004 6005 6006 6007 6008 6009 6010 6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030 6031 6032 6033 6034 6035 6036 6037 6038 6039 6040 6041 6042 6043 6044 6045 6046 6047 6048 6049 6050 6051 6052 6053 6054 6055 6056 6057 6058 6059 6060 6061 6062 6063 6064 6065 6066 6067 6068 6069 6070 6071 6072 6073 6074 6075 6076 6077 6078 6079 6080 6081 6082 6083 6084 6085 6086 6087 6088 6089 6090 6091 6092 6093 6094 6095 6096 6097 6098 6099 6100 6101 6102 6103 6104 6105 6106 6107 6108 6109 6110 6111 6112 6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126 6127 6128 6129 6130 6131 6132 6133 6134 6135 6136 6137 6138 6139 6140 6141 6142 6143 6144 6145 6146 6147 6148 6149 6150 6151 6152 6153 6154 6155 6156 6157 6158 6159 6160 6161 6162 6163 6164 6165 6166 6167 6168 6169 6170 6171 6172 6173 6174 6175 6176 6177 6178 6179 6180 6181 6182 6183 6184 6185 6186 6187 6188 6189 6190 6191 6192 6193 6194 6195 6196 6197 6198 6199 6200 6201 6202 6203 6204 6205 6206 6207 6208 6209 6210 6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221 6222 6223 6224 6225 6226 6227 6228 6229 6230 6231 6232 6233 6234 6235 6236 6237 6238 6239 6240 6241 6242 6243 6244 6245 6246 6247 6248 6249 6250 6251 6252 6253 6254 6255 6256 6257 6258 6259 6260 6261 6262 6263 6264 6265 6266 6267 6268 6269 6270 6271 6272 6273 6274 6275 6276 6277 6278 6279 6280 6281 6282 6283 6284 6285 6286 6287 6288 6289 6290 6291 6292 6293 6294 6295 6296 6297 6298 6299 6300 6301 6302 6303 6304 6305 6306 6307 6308 6309 6310 6311 6312 6313 6314 6315 6316 6317 6318 6319 6320 6321 6322 6323 6324 6325 6326 6327 6328 6329 6330 6331 6332 6333 6334 6335 6336 6337 6338 6339 6340 6341 6342 6343 6344 6345 6346 6347 6348 6349 6350 6351 6352 6353 6354 6355 6356 6357 6358 6359 6360 6361 6362 6363 6364 6365 6366 6367 6368 6369 6370 6371 6372 6373 6374 6375 6376 6377 6378 6379 6380 6381 6382 6383 6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395 6396 6397 6398 6399 6400 6401 6402 6403 6404 6405 6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424 6425 6426 6427 6428 6429 6430 6431 6432 6433 6434 6435 6436 6437 6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461 6462 6463 6464 6465 6466 6467 6468 6469 6470 6471 6472 6473 6474 6475 6476 6477 6478 6479 6480 6481 6482 6483 6484 6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500 6501 6502 6503 6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522 6523 6524 6525 6526 6527 6528 6529 6530 6531 6532 6533 6534 6535 6536 6537 6538 6539 6540 6541 6542 6543 6544 6545 6546 6547 6548 6549 6550 6551 6552 6553 6554 6555 6556 6557 6558 6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571 6572 6573 6574 6575 6576 6577 6578 6579 6580 6581 6582 6583 6584 6585 6586 6587 6588 6589 6590 6591 6592 6593 6594 6595 6596 6597 6598 6599 6600 6601 6602 6603 6604 6605 6606 6607 6608 6609 6610 6611 6612 6613 6614 6615 6616 6617 6618 6619 6620 6621 6622 6623 6624 6625 6626 6627 6628 6629 6630 6631 6632 6633 6634 6635 6636 6637 6638 6639 6640 6641 6642 6643 6644 6645 6646 6647 6648 6649 6650 6651 6652 6653 6654 6655 6656 6657 6658 6659 6660 6661 6662 6663 6664 6665 6666 6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681 6682 6683 6684 6685 6686 6687 6688 6689 6690 6691 6692 6693 6694 6695 6696 6697 6698 6699 6700 6701 6702 6703 6704 6705 6706 6707 6708 6709 6710 6711 6712 6713 6714 6715 6716 6717 6718 6719 6720 6721 6722 6723 6724 6725 6726 6727 6728 6729 6730 6731 6732 6733 6734 6735 6736 6737 6738 6739 6740 6741 6742 6743 6744 6745 6746 6747 6748 6749 6750 6751 6752 6753 6754 6755 6756 6757 6758 6759 6760 6761 6762 6763 6764 6765 6766 6767 6768 6769 6770 6771 6772 6773 6774 6775 6776 6777 6778 6779 6780 6781 6782 6783 6784 6785 6786 6787 6788 6789 6790 6791 6792 6793 6794 6795 6796 6797 6798 6799 6800 6801 6802 6803 6804 6805 6806 6807 6808 6809 6810 6811 6812 6813 6814 6815 6816 6817 6818 6819 6820 6821 6822 6823 6824 6825 6826 6827 6828 6829 6830 6831 6832 6833 6834 6835 6836 6837 6838 6839 6840 6841 6842 6843 6844 6845 6846 6847 6848 6849 6850 6851 6852 6853 6854 6855 6856 6857 6858 6859 6860 6861 6862 6863 6864 6865 6866 6867 6868 6869 6870 6871 6872 6873 6874 6875 6876 6877 6878 6879 6880 6881 6882 6883 6884 6885 6886 6887 6888 6889 6890 6891 6892 6893 6894 6895 6896 6897 6898 6899 6900 6901 6902 6903 6904 6905 6906 6907 6908 6909 6910 6911 6912 6913 6914 6915 6916 6917 6918 6919 6920 6921 6922 6923 6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939 6940 6941 6942 6943 6944 6945 6946 6947 6948 6949 6950 6951 6952 6953 6954 6955 6956 6957 6958 6959 6960 6961 6962 6963 6964 6965 6966 6967 6968 6969 6970 6971 6972 6973 6974 6975 6976 6977 6978 6979 6980 6981 6982 6983 6984 6985 6986 6987 6988 6989 6990 6991 6992 6993 6994 6995 6996 6997 6998 6999 7000 7001 7002 7003 7004 7005 7006 7007 7008 7009 7010 7011 7012 7013 7014 7015 7016 7017 7018 7019 7020 7021 7022 7023 7024 7025 7026 7027 7028 7029 7030 7031 7032 7033 7034 7035 7036 7037 7038 7039 7040 7041 7042 7043 7044 7045 7046 7047 7048 7049 7050 7051 7052 7053 7054 7055 7056 7057 7058 7059 7060 7061 7062 7063 7064 7065 7066 7067 7068 7069 7070 7071 7072 7073 7074 7075 7076 7077 7078 7079 7080 7081 7082 7083 7084 7085 7086 7087 7088 7089 7090 7091 7092 7093 7094 7095 7096 7097 7098 7099 7100 7101 7102 7103 7104 7105 7106 7107 7108 7109 7110 7111 7112 7113 7114 7115 7116 7117 7118 7119 7120 7121 7122 7123 7124 7125 7126 7127 7128 7129 7130 7131 7132 7133 7134 7135 7136 7137 7138 7139 7140 7141 7142 7143 7144 7145 7146 7147 7148 7149 7150 7151 7152 7153 7154 7155 7156 7157 7158 7159 7160 7161 7162 7163 7164 7165 7166 7167 7168 7169 7170 7171 7172 7173 7174 7175 7176 7177 7178 7179 7180 7181 7182 7183 7184 7185 7186 7187 7188 7189 7190 7191 7192 7193 7194 7195 7196 7197 7198 7199 7200 7201 7202 7203 7204 7205 7206 7207 7208 7209 7210 7211 7212 7213 7214 7215 7216 7217 7218 7219 7220 7221 7222 7223 7224 7225 7226 7227 7228 7229 7230 7231 7232 7233 7234 7235 7236 7237 7238 7239 7240 7241 7242 7243 7244 7245 7246 7247 7248 7249 7250 7251 7252 7253 7254 7255 7256 7257 7258 7259 7260 7261 7262 7263 7264 7265 7266 7267 7268 7269 7270 7271 7272 7273 7274 7275 7276 7277 7278 7279 7280 7281 7282 7283 7284 7285 7286 7287 7288 7289 7290 7291 7292 7293 7294 7295 7296 7297 7298 7299 7300 7301 7302 7303 7304 7305 7306 7307 7308 7309 7310 7311 7312 7313 7314 7315 7316 7317 7318 7319 7320 7321 7322 7323 7324 7325 7326 7327 7328 7329 7330 7331 7332 7333 7334 7335 7336 7337 7338 7339 7340 7341 7342 7343 7344 7345 7346 7347 7348 7349 7350 7351 7352 7353 7354 7355 7356 7357 7358 7359 7360 7361 7362 7363 7364 7365 7366 7367 7368 7369 7370 7371 7372 7373 7374 7375 7376 7377 7378 7379 7380 7381 7382 7383 7384 7385 7386 7387 7388 7389 7390 7391 7392 7393 7394 7395 7396 7397 7398 7399 7400 7401 7402 7403 7404 7405 7406 7407 7408 7409 7410 7411 7412 7413 7414 7415 7416 7417 7418 7419 7420 7421 7422 7423 7424 7425 7426 7427 7428 7429 7430 7431 7432 7433 7434 7435 7436 7437 7438 7439 7440 7441 7442 7443 7444 7445 7446 7447 7448 7449 7450 7451 7452 7453 7454 7455 7456 7457 7458 7459 7460 7461 7462 7463 7464 7465 7466 7467 7468 7469 7470 7471 7472 7473 7474 7475 7476 7477 7478 7479 7480 7481 7482 7483 7484 7485 7486 7487 7488 7489 7490 7491 7492 7493 7494 7495 7496 7497 7498 7499 7500 7501 7502 7503 7504 7505 7506 7507 7508 7509 7510 7511 7512 7513 7514 7515 7516 7517 7518 7519 7520 7521 7522 7523 7524 7525 7526 7527 7528 7529 7530 7531 7532 7533 7534 7535 7536 7537 7538 7539 7540 7541 7542 7543 7544 7545 7546 7547 7548 7549 7550 7551 7552 7553 7554 7555 7556 7557 7558 7559 7560 7561 7562 7563 7564 7565 7566 7567 7568 7569 7570 7571 7572 7573 7574 7575 7576 7577 7578 7579 7580 7581 7582 7583 7584 7585 7586 7587 7588 7589 7590 7591 7592 7593 7594 7595 7596 7597 7598 7599 7600 7601 7602 7603 7604 7605 7606 7607 7608 7609 7610 7611 7612 7613 7614 7615 7616 7617 7618 7619 7620 7621 7622 7623 7624 7625 7626 7627 7628 7629 7630 7631 7632 7633 7634 7635 7636 7637 7638 7639 7640 7641 7642 7643 7644 7645 7646 7647 7648 7649 7650 7651 7652 7653 7654 7655 7656 7657 7658 7659 7660 7661 7662 7663 7664 7665 7666 7667 7668 7669 7670 7671 7672 7673 7674 7675 7676 7677 7678 7679 7680 7681 7682 7683 7684 7685 7686 7687 7688 7689 7690 7691 7692 7693 7694 7695 7696 7697 7698 7699 7700 7701 7702 7703 7704 7705 7706 7707 7708 7709 7710 7711 7712 7713 7714 7715 7716 7717 7718 7719 7720 7721 7722 7723 7724 7725 7726 7727 7728 7729 7730 7731 7732 7733 7734 7735 7736 7737 7738 7739 7740 7741 7742 7743 7744 7745 7746 7747 7748 7749 7750 7751 7752 7753 7754 7755 7756 7757 7758 7759 7760 7761 7762 7763 7764 7765 7766 7767 7768 7769 7770 7771 7772 7773 7774 7775 7776 7777 7778 7779 7780 7781 7782 7783 7784 7785 7786 7787 7788 7789 7790 7791 7792 7793 7794 7795 7796 7797 7798 7799 7800 7801 7802 7803 7804 7805 7806 7807 7808 7809 7810 7811 7812 7813 7814 7815 7816 7817 7818 7819 7820 7821 7822 7823 7824 7825 7826 7827 7828 7829 7830 7831 7832 7833 7834 7835 7836 7837 7838 7839 7840 7841 7842 7843 7844 7845 7846 7847 7848 7849 7850 7851 7852 7853 7854 7855 7856 7857 7858 7859 7860 7861 7862 7863 7864 7865 7866 7867 7868 7869 7870 7871 7872 7873 7874 7875 7876 7877 7878 7879 7880 7881 7882 7883 7884 7885 7886 7887 7888 7889 7890 7891 7892 7893 7894 7895 7896 7897 7898 7899 7900 7901 7902 7903 7904 7905 7906 7907 7908 7909 7910 7911 7912 7913 7914 7915 7916 7917 7918 7919 7920 7921 7922 7923 7924 7925 7926 7927 7928 7929 7930 7931 7932 7933 7934 7935 7936 7937 7938 7939 7940 7941 7942 7943 7944 7945 7946 7947 7948 7949 7950 7951 7952 7953 7954 7955 7956 7957 7958 7959 7960 7961 7962 7963 7964 7965 7966 7967 7968 7969 7970 7971 7972 7973 7974 7975 7976 7977 7978 7979 7980 7981 7982 7983 7984 7985 7986 7987 7988 7989 7990 7991 7992 7993 7994 7995 7996 7997 7998 7999 8000 8001 8002 8003 8004 8005 8006 8007 8008 8009 8010 8011 8012 8013 8014 8015 8016 8017 8018 8019 8020 8021 8022 8023 8024 8025 8026 8027 8028 8029 8030 8031 8032 8033 8034 8035 8036 8037 8038 8039 8040 8041 8042 8043 8044 8045 8046 8047 8048 8049 8050 8051 8052 8053 8054 8055 8056 8057 8058 8059 8060 8061 8062 8063 8064 8065 8066 8067 8068 8069 8070 8071 8072 8073 8074 8075 8076 8077 8078 8079 8080 8081 8082 8083 8084 8085 8086 8087 8088 8089 8090 8091 8092 8093 8094 8095 8096 8097 8098 8099 8100 8101 8102 8103 8104 8105 8106 8107 8108 8109 8110 8111 8112 8113 8114 8115 8116 8117 8118 8119 8120 8121 8122 8123 8124 8125 8126 8127 8128 8129 8130 8131 8132 8133 8134 8135 8136 8137 8138 8139 8140 8141 8142 8143 8144 8145 8146 8147 8148 8149 8150 8151 8152 8153 8154 8155 8156 8157 8158 8159 8160 8161 8162 8163 8164 8165 8166 8167 8168 8169 8170 8171 8172 8173 8174 8175 8176 8177 8178 8179 8180 8181 8182 8183 8184 8185 8186 8187 8188 8189 8190 8191 8192 8193 8194 8195 8196 8197 8198 8199 8200 8201 8202 8203 8204 8205 8206 8207 8208 8209 8210 8211 8212 8213 8214 8215 8216 8217 8218 8219 8220 8221 8222 8223 8224 8225 8226 8227 8228 8229 8230 8231 8232 8233 8234 8235 8236 8237 8238 8239 8240 8241 8242 8243 8244 8245 8246 8247 8248 8249 8250 8251 8252 8253 8254 8255 8256 8257 8258 8259 8260 8261 8262 8263 8264 8265 8266 8267 8268 8269 8270 8271 8272 8273 8274 8275 8276 8277 8278 8279 8280 8281 8282 8283 8284 8285 8286 8287 8288 8289 8290 8291 8292 8293 8294 8295 8296 8297 8298 8299 8300 8301 8302 8303 8304 8305 8306 8307 8308 8309 8310 8311 8312 8313 8314 8315 8316 8317 8318 8319 8320 8321 8322 8323 8324 8325 8326 8327 8328 8329 8330 8331 8332 8333 8334 8335 8336 8337 8338 8339 8340 8341 8342 8343 8344 8345 8346 8347 8348 8349 8350 8351 8352 8353 8354 8355 8356 8357 8358 8359 8360 8361 8362 8363 8364 8365 8366 8367 8368 8369 8370 8371 8372 8373 8374 8375 8376 8377 8378 8379 8380 8381 8382 8383 8384 8385 8386 8387 8388 8389 8390 8391 8392 8393 8394 8395 8396 8397 8398 8399 8400 8401 8402 8403 8404 8405 8406 8407 8408 8409 8410 8411 8412 8413 8414 8415 8416 8417 8418 8419 8420 8421 8422 8423 8424 8425 8426 8427 8428 8429 8430 8431 8432 8433 8434 8435 8436 8437 8438 8439 8440 8441 8442 8443 8444 8445 8446 8447 8448 8449 8450 8451 8452 8453 8454 8455 8456 8457 8458 8459 8460 8461 8462 8463 8464 8465 8466 8467 8468 8469 8470 8471 8472 8473 8474 8475 8476 8477 8478 8479 8480 8481 8482 8483 8484 8485 8486 8487 8488 8489 8490 8491 8492 8493 8494 8495 8496 8497 8498 8499 8500 8501 8502 8503 8504 8505 8506 8507 8508 8509 8510 8511 8512 8513 8514 8515 8516 8517 8518 8519 8520 8521 8522 8523 8524 8525 8526 8527 8528 8529 8530 8531 8532 8533 8534 8535 8536 8537 8538 8539 8540 8541 8542 8543 8544 8545 8546 8547 8548 8549 8550 8551 8552 8553 8554 8555 8556 8557 8558 8559 8560 8561 8562 8563 8564 8565 8566 8567 8568 8569 8570 8571 8572 8573 8574 8575 8576 8577 8578 8579 8580 8581 8582 8583 8584 8585 8586 8587 8588 8589 8590 8591 8592 8593 8594 8595 8596 8597 8598 8599 8600 8601 8602 8603 8604 8605 8606 8607 8608 8609 8610 8611 8612 8613 8614 8615 8616 8617 8618 8619 8620 8621 8622 8623 8624 8625 8626 8627 8628 8629 8630 8631 8632 8633 8634 8635 8636 8637 8638 8639 8640 8641 8642 8643 8644 8645 8646 8647 8648 8649 8650 8651 8652 8653 8654 8655 8656 8657 8658 8659 8660 8661 8662 8663 8664 8665 8666 8667 8668 8669 8670 8671 8672 8673 8674 8675 8676 8677 8678 8679 8680 8681 8682 8683 8684 8685 8686 8687 8688 8689 8690 8691 8692 8693 8694 8695 8696 8697 8698 8699 8700 8701 8702 8703 8704 8705 8706 8707 8708 8709 8710 8711 8712 8713 8714 8715 8716 8717 8718 8719 8720 8721 8722 8723 8724 8725 8726 8727 8728 8729 8730 8731 8732 8733 8734 8735 8736 8737 8738 8739 8740 8741 8742 8743 8744 8745 8746 8747 8748 8749 8750 8751 8752 8753 8754 8755 8756 8757 8758 8759 8760 8761 8762 8763 8764 8765 8766 8767 8768 8769 8770 8771 8772 8773 8774 8775 8776 8777 8778 8779 8780 8781 8782 8783 8784 8785 8786 8787 8788 8789 8790 8791 8792 8793 8794 8795 8796 8797 8798 8799 8800 8801 8802 8803 8804 8805 8806 8807 8808 8809 8810 8811 8812 8813 8814 8815 8816 8817 8818 8819 8820 8821 8822 8823 8824 8825 8826 8827 8828 8829 8830 8831 8832 8833 8834 8835 8836 8837 8838 8839 8840 8841 8842 8843 8844 8845 8846 8847 8848 8849 8850 8851 8852 8853 8854 8855 8856 8857 8858 8859 8860 8861 8862 8863 8864 8865 8866 8867 8868 8869 8870 8871 8872 8873 8874 8875 8876 8877 8878 8879 8880 8881 8882 8883 8884 8885 8886 8887 8888 8889 8890 8891 8892 8893 8894 8895 8896 8897 8898 8899 8900 8901 8902 8903 8904 8905 8906 8907 8908 8909 8910 8911 8912 8913 8914 8915 8916 8917 8918 8919 8920 8921 8922 8923 8924 8925 8926 8927 8928 8929 8930 8931 8932 8933 8934 8935 8936 8937 8938 8939 8940 8941 8942 8943 8944 8945 8946 8947 8948 8949 8950 8951 8952 8953 8954 8955 8956 8957 8958 8959 8960 8961 8962 8963 8964 8965 8966 8967 8968 8969 8970 8971 8972 8973 8974 8975 8976 8977 8978 8979 8980 8981 8982 8983 8984 8985 8986 8987 8988 8989 8990 8991 8992 8993 8994 8995 8996 8997 8998 8999 9000 9001 9002 9003 9004 9005 9006 9007 9008 9009 9010 9011 9012 9013 9014 9015 9016 9017 9018 9019 9020 9021 9022 9023 9024 9025 9026 9027 9028 9029 9030 9031 9032 9033 9034 9035 9036 9037 9038 9039 9040 9041 9042 9043 9044 9045 9046 9047 9048 9049 9050 9051 9052 9053 9054 9055 9056 9057 9058 9059 9060 9061 9062 9063 9064 9065 9066 9067 9068 9069 9070 9071 9072 9073 9074 9075 9076 9077 9078 9079 9080 9081 9082 9083 9084 9085 9086 9087 9088 9089 9090 9091 9092 9093 9094 9095 9096 9097 9098 9099 9100 9101 9102 9103 9104 9105 9106 9107 9108 9109 9110 9111 9112 9113 9114 9115 9116 9117 9118 9119 9120 9121 9122 9123 9124 9125 9126 9127 9128 9129 9130 9131 9132 9133 9134 9135 9136 9137 9138 9139 9140 9141 9142 9143 9144 9145 9146 9147 9148 9149 9150 9151 9152 9153 9154 9155 9156 9157 9158 9159 9160 9161 9162 9163 9164 9165 9166 9167 9168 9169 9170 9171 9172 9173 9174 9175 9176 9177 9178 9179 9180 9181 9182 9183 9184 9185 9186 9187 9188 9189 9190 9191 9192 9193 9194 9195 9196 9197 9198 9199 9200 9201 9202 9203 9204 9205 9206 9207 9208 9209 9210 9211 9212 9213 9214 9215 9216 9217 9218 9219 9220 9221 9222 9223 9224 9225 9226 9227 9228 9229 9230 9231 9232 9233 9234 9235 9236 9237 9238 9239 9240 9241 9242 9243 9244 9245 9246 9247 9248 9249 9250 9251 9252 9253 9254 9255 9256 9257 9258 9259 9260 9261 9262 9263 9264 9265 9266 9267 9268 9269 9270 9271 9272 9273 9274 9275 9276 9277 9278 9279 9280 9281 9282 9283 9284 9285 9286 9287 9288 9289 9290 9291 9292 9293 9294 9295 9296 9297 9298 9299 9300 9301 9302 9303 9304 9305 9306 9307 9308 9309 9310 9311 9312 9313 9314 9315 9316 9317 9318 9319 9320 9321 9322 9323 9324 9325 9326 9327 9328 9329 9330 9331 9332 9333 9334 9335 9336 9337 9338 9339 9340 9341 9342 9343 9344 9345 9346 9347 9348 9349 9350 9351 9352 9353 9354 9355 9356 9357 9358 9359 9360 9361 9362 9363 9364 9365 9366 9367 9368 9369 9370 9371 9372 9373 9374 9375 9376 9377 9378 9379 9380 9381 9382 9383 9384 9385 9386 9387 9388 9389 9390 9391 9392 9393 9394 9395 9396 9397 9398 9399 9400 9401 9402 9403 9404 9405 9406 9407 9408 9409 9410 9411 9412 9413 9414 9415 9416 9417 9418 9419 9420 9421 9422 9423 9424 9425 9426 9427 9428 9429 9430 9431 9432 9433 9434 9435 9436 9437 9438 9439 9440 9441 9442 9443 9444 9445 9446 9447 9448 9449 9450 9451 9452 9453 9454 9455 9456 9457 9458 9459 9460 9461 9462 9463 9464 9465 9466 9467 9468 9469 9470 9471 9472 9473 9474 9475 9476 9477 9478 9479 9480 9481 9482 9483 9484 9485 9486 9487 9488 9489 9490 9491 9492 9493 9494 9495 9496 9497 9498 9499 9500 9501 9502 9503 9504 9505 9506 9507 9508 9509 9510 9511 9512 9513 9514 9515 9516 9517 9518 9519 9520 9521 9522 9523 9524 9525 9526 9527 9528 9529 9530 9531 9532 9533 9534 9535 9536 9537 9538 9539 9540 9541 9542 9543 9544 9545 9546 9547 9548 9549 9550 9551 9552 9553 9554 9555 9556 9557 9558 9559 9560 9561 9562 9563 9564 9565 9566 9567 9568 9569 9570 9571 9572 9573 9574 9575 9576 9577 9578 9579 9580 9581 9582 9583 9584 9585 9586 9587 9588 9589 9590 9591 9592 9593 9594 9595 9596 9597 9598 9599 9600 9601 9602 9603 9604 9605 9606 9607 9608 9609 9610 9611 9612 9613 9614 9615 9616 9617 9618 9619 9620 9621 9622 9623 9624 9625 9626 9627 9628 9629 9630 9631 9632 9633 9634 9635 9636 9637 9638 9639 9640 9641 9642 9643 9644 9645 9646 9647 9648 9649 9650 9651 9652 9653 9654 9655 9656 9657 9658 9659 9660 9661 9662 9663 9664 9665 9666 9667 9668 9669 9670 9671 9672 9673 9674 9675 9676 9677 9678 9679 9680 9681 9682 9683 9684 9685 9686 9687 9688 9689 9690 9691 9692 9693 9694 9695 9696 9697 9698 9699 9700 9701 9702 9703 9704 9705 9706 9707 9708 9709 9710 9711 9712 9713 9714 9715 9716 9717 9718 9719 9720 9721 9722 9723 9724 9725 9726 9727 9728 9729 9730 9731 9732 9733 9734 9735 9736 9737 9738 9739 9740 9741 9742 9743 9744 9745 9746 9747 9748 9749 9750 9751 9752 9753 9754 9755 9756 9757 9758 9759 9760 9761 9762 9763 9764 9765 9766 9767 9768 9769 9770 9771 9772 9773 9774 9775 9776 9777 9778 9779 9780 9781 9782 9783 9784 9785 9786 9787 9788 9789 9790 9791 9792 9793 9794 9795 9796 9797 9798 9799 9800 9801 9802 9803 9804 9805 9806 9807 9808 9809 9810 9811 9812 9813 9814 9815 9816 9817 9818 9819 9820 9821 9822 9823 9824 9825 9826 9827 9828 9829 9830 9831 9832 9833 9834 9835 9836 9837 9838 9839 9840 9841 9842 9843 9844 9845 9846 9847 9848 9849 9850 9851 9852 9853 9854 9855 9856 9857 9858 9859 9860 9861 9862 9863 9864 9865 9866 9867 9868 9869 9870 9871 9872 9873 9874 9875 9876 9877 9878 9879 9880 9881 9882 9883 9884 9885 9886 9887 9888 9889 9890 9891 9892 9893 9894 9895 9896 9897 9898 9899 9900 9901 9902 9903 9904 9905 9906 9907 9908 9909 9910 9911 9912 9913 9914 9915 9916 9917 9918 9919 9920 9921 9922 9923 9924 9925 9926 9927 9928 9929 9930 9931 9932 9933 9934 9935 9936 9937 9938 9939 9940 9941 9942 9943 9944 9945 9946 9947 9948 9949 9950 9951 9952 9953 9954 9955 9956 9957 9958 9959 9960 9961 9962 9963 9964 9965 9966 9967 9968 9969 9970 9971 9972 9973 9974 9975 9976 9977 9978 9979 9980 9981 9982 9983 9984 9985 9986 9987 9988 9989 9990 9991 9992 9993 9994 9995 9996 9997 9998 9999 10000 10001 10002 10003 10004 10005 10006 10007 10008 10009 10010 10011 10012 10013 10014 10015 10016 10017 10018 10019 10020 10021 10022 10023 10024 10025 10026 10027 10028 10029 10030 10031 10032 10033 10034 10035 10036 10037 10038 10039 10040 10041 10042 10043 10044 10045 10046 10047 10048 10049 10050 10051 10052 10053 10054 10055 10056 10057 10058 10059 10060 10061 10062 10063 10064 10065 10066 10067 10068 10069 10070 10071 10072 10073 10074 10075 10076 10077 10078 10079 10080 10081 10082 10083 10084 10085 10086 10087 10088 10089 10090 10091 10092 10093 10094 10095 10096 10097 10098 10099 10100 10101 10102 10103 10104 10105 10106 10107 10108 10109 10110 10111 10112 10113 10114 10115 10116 10117 10118 10119 10120 10121 10122 10123 10124 10125 10126 10127 10128 10129 10130 10131 10132 10133 10134 10135 10136 10137 10138 10139 10140 10141 10142 10143 10144 10145 10146 10147 10148 10149 10150 10151 10152 10153 10154 10155 10156 10157 10158 10159 10160 10161 10162 10163 10164 10165 10166 10167 10168 10169 10170 10171 10172 10173 10174 10175 10176 10177 10178 10179 10180 10181 10182 10183 10184 10185 10186 10187 10188 10189 10190 10191 10192 10193 10194 10195 10196 10197 10198 10199 10200 10201 10202 10203 10204 10205 10206 10207 10208 10209 10210 10211 10212 10213 10214 10215 10216 10217 10218 10219 10220 10221 10222 10223 10224 10225 10226 10227 10228 10229 10230 10231 10232 10233 10234 10235 10236 10237 10238 10239 10240 10241 10242 10243 10244 10245 10246 10247 10248 10249 10250 10251 10252 10253 10254 10255 10256 10257 10258 10259 10260 10261 10262 10263 10264 10265 10266 10267 10268 10269 10270 10271 10272 10273 10274 10275 10276 10277 10278 10279 10280 10281 10282 10283 10284 10285 10286 10287 10288 10289 10290 10291 10292 10293 10294 10295 10296 10297 10298 10299 10300 10301 10302 10303 10304 10305 10306 10307 10308 10309 10310 10311 10312 10313 10314 10315 10316 10317 10318 10319 10320 10321 10322 10323 10324 10325 10326 10327 10328 10329 10330 10331 10332 10333 10334 10335 10336 10337 10338 10339 10340 10341 10342 10343 10344 10345 10346 10347 10348 10349 10350 10351 10352 10353 10354 10355 10356 10357 10358 10359 10360 10361 10362 10363 10364 10365 10366 10367 10368 10369 10370 10371 10372 10373 10374 10375 10376 10377 10378 10379 10380 10381 10382 10383 10384 10385 10386 10387 10388 10389 10390 10391 10392 10393 10394 10395 10396 10397 10398 10399 10400 10401 10402 10403 10404 10405 10406 10407 10408 10409 10410 10411 10412 10413 10414 10415 10416 10417 10418 10419 10420 10421 10422 10423 10424 10425 10426 10427 10428 10429 10430 10431 10432 10433 10434 10435 10436 10437 10438 10439 10440 10441 10442 10443 10444 10445 10446 10447 10448 10449 10450 10451 10452 10453 10454 10455 10456 10457 10458 10459 10460 10461 10462 10463 10464 10465 10466 10467 10468 10469 10470 10471 10472 10473 10474 10475 10476 10477 10478 10479 10480 10481 10482 10483 10484 10485 10486 10487 10488 10489 10490 10491 10492 10493 10494 10495 10496 10497 10498 10499 10500 10501 10502 10503 10504 10505 10506 10507 10508 10509 10510 10511 10512 10513 10514 10515 10516 10517 10518 10519 10520 10521 10522 10523 10524 10525 10526 10527 10528 10529 10530 10531 10532 10533 10534 10535 10536 10537 10538 10539 10540 10541 10542 10543 10544 10545 10546 10547 10548 10549 10550 10551 10552 10553 10554 10555 10556 10557 10558 10559 10560 10561 10562 10563 10564 10565 10566 10567 10568 10569 10570 10571 10572 10573 10574 10575 10576 10577 10578 10579 10580 10581 10582 10583 10584 10585 10586 10587 10588 10589 10590 10591 10592 10593 10594 10595 10596 10597 10598 10599 10600 10601 10602 10603 10604 10605 10606 10607 10608 10609 10610 10611 10612 10613 10614 10615 10616 10617 10618 10619 10620 10621 10622 10623 10624 10625 10626 10627 10628 10629 10630 10631 10632 10633 10634 10635 10636 10637 10638 10639 10640 10641 10642 10643 10644 10645 10646 10647 10648 10649 10650 10651 10652 10653 10654 10655 10656 10657 10658 10659 10660 10661 10662 10663 10664 10665 10666 10667 10668 10669 10670 10671 10672 10673 10674 10675 10676 10677 10678 10679 10680 10681 10682 10683 10684 10685 10686 10687 10688 10689 10690 10691 10692 10693 10694 10695 10696 10697 10698 10699 10700 10701 10702 10703 10704 10705 10706 10707 10708 10709 10710 10711 10712 10713 10714 10715 10716 10717 10718 10719 10720 10721 10722 10723 10724 10725 10726 10727 10728 10729 10730 10731 10732 10733 10734 10735 10736 10737 10738 10739 10740 10741 10742 10743 10744 10745 10746 10747 10748 10749 10750 10751 10752 10753 10754 10755 10756 10757 10758 10759 10760 10761 10762 10763 10764 10765 10766 10767 10768 10769 10770 10771 10772 10773 10774 10775 10776 10777 10778 10779 10780 10781 10782 10783 10784 10785 10786 10787 10788 10789 10790 10791 10792 10793 10794 10795 10796 10797 10798 10799 10800 10801 10802 10803 10804 10805 10806 10807 10808 10809 10810 10811 10812 10813 10814 10815 10816 10817 10818 10819 10820 10821 10822 10823 10824 10825 10826 10827 10828 10829 10830 10831 10832 10833 10834 10835 10836 10837 10838 10839 10840 10841 10842 10843 10844 10845 10846 10847 10848 10849 10850 10851 10852 10853 10854 10855 10856 10857 10858 10859 10860 10861 10862 10863 10864 10865 10866 10867 10868 10869 10870 10871 10872 10873 10874 10875 10876 10877 10878 10879 10880 10881 10882 10883 10884 10885 10886 10887 10888 10889 10890 10891 10892 10893 10894 10895 10896 10897 10898 10899 10900 10901 10902 10903 10904 10905 10906 10907 10908 10909 10910 10911 10912 10913 10914 10915 10916 10917 10918 10919 10920 10921 10922 10923 10924 10925 10926 10927 10928 10929 10930 10931 10932 10933 10934 10935 10936 10937 10938 10939 10940 10941 10942 10943 10944 10945 10946 10947 10948 10949 10950 10951 10952 10953 10954 10955 10956 10957 10958 10959 10960 10961 10962 10963 10964 10965 10966 10967 10968 10969 10970 10971 10972 10973 10974 10975 10976 10977 10978 10979 10980 10981 10982 10983 10984 10985 10986 10987 10988 10989 10990 10991 10992 10993 10994 10995 10996 10997 10998 10999 11000 11001 11002 11003 11004 11005 11006 11007 11008 11009 11010 11011 11012 11013 11014 11015 11016 11017 11018 11019 11020 11021 11022 11023 11024 11025 11026 11027 11028 11029 11030 11031 11032 11033 11034 11035 11036 11037 11038 11039 11040 11041 11042 11043 11044 11045 11046 11047 11048 11049 11050 11051 11052 11053 11054 11055 11056 11057 11058 11059 11060 11061 11062 11063 11064 11065 11066 11067 11068 11069 11070 11071 11072 11073 11074 11075 11076 11077 11078 11079 11080 11081 11082 11083 11084 11085 11086 11087 11088 11089 11090 11091 11092 11093 11094 11095 11096 11097 11098 11099 11100 11101 11102 11103 11104 11105 11106 11107 11108 11109 11110 11111 11112 11113 11114 11115 11116 11117 11118 11119 11120 11121 11122 11123 11124 11125 11126 11127 11128 11129 11130 11131 11132 11133 11134 11135 11136 11137 11138 11139 11140 11141 11142 11143 11144 11145 11146 11147 11148 11149 11150 11151 11152 11153 11154 11155 11156 11157 11158 11159 11160 11161 11162 11163 11164 11165 11166 11167 11168 11169 11170 11171 11172 11173 11174 11175 11176 11177 11178 11179 11180 11181 11182 11183 11184 11185 11186 11187 11188 11189 11190 11191 11192 11193 11194 11195 11196 11197 11198 11199 11200 11201 11202 11203 11204 11205 11206 11207 11208 11209 11210 11211 11212 11213 11214 11215 11216 11217 11218 11219 11220 11221 11222 11223 11224 11225 11226 11227 11228 11229 11230 11231 11232 11233 11234 11235 11236 11237 11238 11239 11240 11241 11242 11243 11244 11245 11246 11247 11248 11249 11250 11251 11252 11253 11254 11255 11256 11257 11258 11259 11260 11261 11262 11263 11264 11265 11266 11267 11268 11269 11270 11271 11272 11273 11274 11275 11276 11277 11278 11279 11280 11281 11282 11283 11284 11285 11286 11287 11288 11289 11290 11291 11292 11293 11294 11295 11296 11297 11298 11299 11300 11301 11302 11303 11304 11305 11306 11307 11308 11309 11310 11311 11312 11313 11314 11315 11316 11317 11318 11319 11320 11321 11322 11323 11324 11325 11326 11327 11328 11329 11330 11331 11332 11333 11334 11335 11336 11337 11338 11339 11340 11341 11342 11343 11344 11345 11346 11347 11348 11349 11350 11351 11352 11353 11354 11355 11356 11357 11358 11359 11360 11361 11362 11363 11364 11365 11366 11367 11368 11369 11370 11371 11372 11373 11374 11375 11376 11377 11378 11379 11380 11381 11382 11383 11384 11385 11386 11387 11388 11389 11390 11391 11392 11393 11394 11395 11396 11397 11398 11399 11400 11401 11402 11403 11404 11405 11406 11407 11408 11409 11410 11411 11412 11413 11414 11415 11416 11417 11418 11419 11420 11421 11422 11423 11424 11425 11426 11427 11428 11429 11430 11431 11432 11433 11434 11435 11436 11437 11438 11439 11440 11441 11442 11443 11444 11445 11446 11447 11448 11449 11450 11451 11452 11453 11454 11455 11456 11457 11458 11459 11460 11461 11462 11463 11464 11465 11466 11467 11468 11469 11470 11471 11472 11473 11474 11475 11476 11477 11478 11479 11480 11481 11482 11483 11484 11485 11486 11487 11488 11489 11490 11491 11492 11493 11494 11495 11496 11497 11498 11499 11500 11501 11502 11503 11504 11505 11506 11507 11508 11509 11510 11511 11512 11513 11514 11515 11516 11517 11518 11519 11520 11521 11522 11523 11524 11525 11526 11527 11528 11529 11530 11531 11532 11533 11534 11535 11536 11537 11538 11539 11540 11541 11542 11543 11544 11545 11546 11547 11548 11549 11550 11551 11552 11553 11554 11555 11556 11557 11558 11559 11560 11561 11562 11563 11564 11565 11566 11567 11568 11569 11570 11571 11572 11573 11574 11575 11576 11577 11578 11579 11580 11581 11582 11583 11584 11585 11586 11587 11588 11589 11590 11591 11592 11593 11594 11595 11596 11597 11598 11599 11600 11601 11602 11603 11604 11605 11606 11607 11608 11609 11610 11611 11612 11613 11614 11615 11616 11617 11618 11619 11620 11621 11622 11623 11624 11625 11626 11627 11628 11629 11630 11631 11632 11633 11634 11635 11636 11637 11638 11639 11640 11641 11642 11643 11644 11645 11646 11647 11648 11649 11650 11651 11652 11653 11654 11655 11656 11657 11658 11659 11660 11661 11662 11663 11664 11665 11666 11667 11668 11669 11670 11671 11672 11673 11674 11675 11676 11677 11678 11679 11680 11681 11682 11683 11684 11685 11686 11687
|
//===-- TargetLowering.cpp - Implement the TargetLowering class -----------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This implements the TargetLowering class.
//
//===----------------------------------------------------------------------===//
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/CodeGenCommonISel.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineModuleInfoImpls.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCExpr.h"
#include "llvm/Support/DivisionByConstantInfo.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetMachine.h"
#include <cctype>
using namespace llvm;
/// NOTE: The TargetMachine owns TLOF.
TargetLowering::TargetLowering(const TargetMachine &tm)
: TargetLoweringBase(tm) {}
const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
return nullptr;
}
bool TargetLowering::isPositionIndependent() const {
return getTargetMachine().isPositionIndependent();
}
/// Check whether a given call node is in tail position within its function. If
/// so, it sets Chain to the input chain of the tail call.
bool TargetLowering::isInTailCallPosition(SelectionDAG &DAG, SDNode *Node,
SDValue &Chain) const {
const Function &F = DAG.getMachineFunction().getFunction();
// First, check if tail calls have been disabled in this function.
if (F.getFnAttribute("disable-tail-calls").getValueAsBool())
return false;
// Conservatively require the attributes of the call to match those of
// the return. Ignore following attributes because they don't affect the
// call sequence.
AttrBuilder CallerAttrs(F.getContext(), F.getAttributes().getRetAttrs());
for (const auto &Attr :
{Attribute::Alignment, Attribute::Dereferenceable,
Attribute::DereferenceableOrNull, Attribute::NoAlias,
Attribute::NonNull, Attribute::NoUndef, Attribute::Range})
CallerAttrs.removeAttribute(Attr);
if (CallerAttrs.hasAttributes())
return false;
// It's not safe to eliminate the sign / zero extension of the return value.
if (CallerAttrs.contains(Attribute::ZExt) ||
CallerAttrs.contains(Attribute::SExt))
return false;
// Check if the only use is a function return node.
return isUsedByReturnOnly(Node, Chain);
}
bool TargetLowering::parametersInCSRMatch(const MachineRegisterInfo &MRI,
const uint32_t *CallerPreservedMask,
const SmallVectorImpl<CCValAssign> &ArgLocs,
const SmallVectorImpl<SDValue> &OutVals) const {
for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
const CCValAssign &ArgLoc = ArgLocs[I];
if (!ArgLoc.isRegLoc())
continue;
MCRegister Reg = ArgLoc.getLocReg();
// Only look at callee saved registers.
if (MachineOperand::clobbersPhysReg(CallerPreservedMask, Reg))
continue;
// Check that we pass the value used for the caller.
// (We look for a CopyFromReg reading a virtual register that is used
// for the function live-in value of register Reg)
SDValue Value = OutVals[I];
if (Value->getOpcode() == ISD::AssertZext)
Value = Value.getOperand(0);
if (Value->getOpcode() != ISD::CopyFromReg)
return false;
Register ArgReg = cast<RegisterSDNode>(Value->getOperand(1))->getReg();
if (MRI.getLiveInPhysReg(ArgReg) != Reg)
return false;
}
return true;
}
/// Set CallLoweringInfo attribute flags based on a call instruction
/// and called function attributes.
void TargetLoweringBase::ArgListEntry::setAttributes(const CallBase *Call,
unsigned ArgIdx) {
IsSExt = Call->paramHasAttr(ArgIdx, Attribute::SExt);
IsZExt = Call->paramHasAttr(ArgIdx, Attribute::ZExt);
IsInReg = Call->paramHasAttr(ArgIdx, Attribute::InReg);
IsSRet = Call->paramHasAttr(ArgIdx, Attribute::StructRet);
IsNest = Call->paramHasAttr(ArgIdx, Attribute::Nest);
IsByVal = Call->paramHasAttr(ArgIdx, Attribute::ByVal);
IsPreallocated = Call->paramHasAttr(ArgIdx, Attribute::Preallocated);
IsInAlloca = Call->paramHasAttr(ArgIdx, Attribute::InAlloca);
IsReturned = Call->paramHasAttr(ArgIdx, Attribute::Returned);
IsSwiftSelf = Call->paramHasAttr(ArgIdx, Attribute::SwiftSelf);
IsSwiftAsync = Call->paramHasAttr(ArgIdx, Attribute::SwiftAsync);
IsSwiftError = Call->paramHasAttr(ArgIdx, Attribute::SwiftError);
Alignment = Call->getParamStackAlign(ArgIdx);
IndirectType = nullptr;
assert(IsByVal + IsPreallocated + IsInAlloca + IsSRet <= 1 &&
"multiple ABI attributes?");
if (IsByVal) {
IndirectType = Call->getParamByValType(ArgIdx);
if (!Alignment)
Alignment = Call->getParamAlign(ArgIdx);
}
if (IsPreallocated)
IndirectType = Call->getParamPreallocatedType(ArgIdx);
if (IsInAlloca)
IndirectType = Call->getParamInAllocaType(ArgIdx);
if (IsSRet)
IndirectType = Call->getParamStructRetType(ArgIdx);
}
/// Generate a libcall taking the given operands as arguments and returning a
/// result of type RetVT.
std::pair<SDValue, SDValue>
TargetLowering::makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC, EVT RetVT,
ArrayRef<SDValue> Ops,
MakeLibCallOptions CallOptions,
const SDLoc &dl,
SDValue InChain) const {
if (!InChain)
InChain = DAG.getEntryNode();
TargetLowering::ArgListTy Args;
Args.reserve(Ops.size());
TargetLowering::ArgListEntry Entry;
for (unsigned i = 0; i < Ops.size(); ++i) {
SDValue NewOp = Ops[i];
Entry.Node = NewOp;
Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
Entry.IsSExt = shouldSignExtendTypeInLibCall(NewOp.getValueType(),
CallOptions.IsSExt);
Entry.IsZExt = !Entry.IsSExt;
if (CallOptions.IsSoften &&
!shouldExtendTypeInLibCall(CallOptions.OpsVTBeforeSoften[i])) {
Entry.IsSExt = Entry.IsZExt = false;
}
Args.push_back(Entry);
}
if (LC == RTLIB::UNKNOWN_LIBCALL)
report_fatal_error("Unsupported library call operation!");
SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
getPointerTy(DAG.getDataLayout()));
Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
TargetLowering::CallLoweringInfo CLI(DAG);
bool signExtend = shouldSignExtendTypeInLibCall(RetVT, CallOptions.IsSExt);
bool zeroExtend = !signExtend;
if (CallOptions.IsSoften &&
!shouldExtendTypeInLibCall(CallOptions.RetVTBeforeSoften)) {
signExtend = zeroExtend = false;
}
CLI.setDebugLoc(dl)
.setChain(InChain)
.setLibCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args))
.setNoReturn(CallOptions.DoesNotReturn)
.setDiscardResult(!CallOptions.IsReturnValueUsed)
.setIsPostTypeLegalization(CallOptions.IsPostTypeLegalization)
.setSExtResult(signExtend)
.setZExtResult(zeroExtend);
return LowerCallTo(CLI);
}
bool TargetLowering::findOptimalMemOpLowering(
std::vector<EVT> &MemOps, unsigned Limit, const MemOp &Op, unsigned DstAS,
unsigned SrcAS, const AttributeList &FuncAttributes) const {
if (Limit != ~unsigned(0) && Op.isMemcpyWithFixedDstAlign() &&
Op.getSrcAlign() < Op.getDstAlign())
return false;
EVT VT = getOptimalMemOpType(Op, FuncAttributes);
if (VT == MVT::Other) {
// Use the largest integer type whose alignment constraints are satisfied.
// We only need to check DstAlign here as SrcAlign is always greater or
// equal to DstAlign (or zero).
VT = MVT::LAST_INTEGER_VALUETYPE;
if (Op.isFixedDstAlign())
while (Op.getDstAlign() < (VT.getSizeInBits() / 8) &&
!allowsMisalignedMemoryAccesses(VT, DstAS, Op.getDstAlign()))
VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
assert(VT.isInteger());
// Find the largest legal integer type.
MVT LVT = MVT::LAST_INTEGER_VALUETYPE;
while (!isTypeLegal(LVT))
LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1);
assert(LVT.isInteger());
// If the type we've chosen is larger than the largest legal integer type
// then use that instead.
if (VT.bitsGT(LVT))
VT = LVT;
}
unsigned NumMemOps = 0;
uint64_t Size = Op.size();
while (Size) {
unsigned VTSize = VT.getSizeInBits() / 8;
while (VTSize > Size) {
// For now, only use non-vector load / store's for the left-over pieces.
EVT NewVT = VT;
unsigned NewVTSize;
bool Found = false;
if (VT.isVector() || VT.isFloatingPoint()) {
NewVT = (VT.getSizeInBits() > 64) ? MVT::i64 : MVT::i32;
if (isOperationLegalOrCustom(ISD::STORE, NewVT) &&
isSafeMemOpType(NewVT.getSimpleVT()))
Found = true;
else if (NewVT == MVT::i64 &&
isOperationLegalOrCustom(ISD::STORE, MVT::f64) &&
isSafeMemOpType(MVT::f64)) {
// i64 is usually not legal on 32-bit targets, but f64 may be.
NewVT = MVT::f64;
Found = true;
}
}
if (!Found) {
do {
NewVT = (MVT::SimpleValueType)(NewVT.getSimpleVT().SimpleTy - 1);
if (NewVT == MVT::i8)
break;
} while (!isSafeMemOpType(NewVT.getSimpleVT()));
}
NewVTSize = NewVT.getSizeInBits() / 8;
// If the new VT cannot cover all of the remaining bits, then consider
// issuing a (or a pair of) unaligned and overlapping load / store.
unsigned Fast;
if (NumMemOps && Op.allowOverlap() && NewVTSize < Size &&
allowsMisalignedMemoryAccesses(
VT, DstAS, Op.isFixedDstAlign() ? Op.getDstAlign() : Align(1),
MachineMemOperand::MONone, &Fast) &&
Fast)
VTSize = Size;
else {
VT = NewVT;
VTSize = NewVTSize;
}
}
if (++NumMemOps > Limit)
return false;
MemOps.push_back(VT);
Size -= VTSize;
}
return true;
}
/// Soften the operands of a comparison. This code is shared among BR_CC,
/// SELECT_CC, and SETCC handlers.
void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT,
SDValue &NewLHS, SDValue &NewRHS,
ISD::CondCode &CCCode,
const SDLoc &dl, const SDValue OldLHS,
const SDValue OldRHS) const {
SDValue Chain;
return softenSetCCOperands(DAG, VT, NewLHS, NewRHS, CCCode, dl, OldLHS,
OldRHS, Chain);
}
void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT,
SDValue &NewLHS, SDValue &NewRHS,
ISD::CondCode &CCCode,
const SDLoc &dl, const SDValue OldLHS,
const SDValue OldRHS,
SDValue &Chain,
bool IsSignaling) const {
// FIXME: Currently we cannot really respect all IEEE predicates due to libgcc
// not supporting it. We can update this code when libgcc provides such
// functions.
assert((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128 || VT == MVT::ppcf128)
&& "Unsupported setcc type!");
// Expand into one or more soft-fp libcall(s).
RTLIB::Libcall LC1 = RTLIB::UNKNOWN_LIBCALL, LC2 = RTLIB::UNKNOWN_LIBCALL;
bool ShouldInvertCC = false;
switch (CCCode) {
case ISD::SETEQ:
case ISD::SETOEQ:
LC1 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
(VT == MVT::f64) ? RTLIB::OEQ_F64 :
(VT == MVT::f128) ? RTLIB::OEQ_F128 : RTLIB::OEQ_PPCF128;
break;
case ISD::SETNE:
case ISD::SETUNE:
LC1 = (VT == MVT::f32) ? RTLIB::UNE_F32 :
(VT == MVT::f64) ? RTLIB::UNE_F64 :
(VT == MVT::f128) ? RTLIB::UNE_F128 : RTLIB::UNE_PPCF128;
break;
case ISD::SETGE:
case ISD::SETOGE:
LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
(VT == MVT::f64) ? RTLIB::OGE_F64 :
(VT == MVT::f128) ? RTLIB::OGE_F128 : RTLIB::OGE_PPCF128;
break;
case ISD::SETLT:
case ISD::SETOLT:
LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
(VT == MVT::f64) ? RTLIB::OLT_F64 :
(VT == MVT::f128) ? RTLIB::OLT_F128 : RTLIB::OLT_PPCF128;
break;
case ISD::SETLE:
case ISD::SETOLE:
LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
(VT == MVT::f64) ? RTLIB::OLE_F64 :
(VT == MVT::f128) ? RTLIB::OLE_F128 : RTLIB::OLE_PPCF128;
break;
case ISD::SETGT:
case ISD::SETOGT:
LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
(VT == MVT::f64) ? RTLIB::OGT_F64 :
(VT == MVT::f128) ? RTLIB::OGT_F128 : RTLIB::OGT_PPCF128;
break;
case ISD::SETO:
ShouldInvertCC = true;
[[fallthrough]];
case ISD::SETUO:
LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
(VT == MVT::f64) ? RTLIB::UO_F64 :
(VT == MVT::f128) ? RTLIB::UO_F128 : RTLIB::UO_PPCF128;
break;
case ISD::SETONE:
// SETONE = O && UNE
ShouldInvertCC = true;
[[fallthrough]];
case ISD::SETUEQ:
LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
(VT == MVT::f64) ? RTLIB::UO_F64 :
(VT == MVT::f128) ? RTLIB::UO_F128 : RTLIB::UO_PPCF128;
LC2 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
(VT == MVT::f64) ? RTLIB::OEQ_F64 :
(VT == MVT::f128) ? RTLIB::OEQ_F128 : RTLIB::OEQ_PPCF128;
break;
default:
// Invert CC for unordered comparisons
ShouldInvertCC = true;
switch (CCCode) {
case ISD::SETULT:
LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
(VT == MVT::f64) ? RTLIB::OGE_F64 :
(VT == MVT::f128) ? RTLIB::OGE_F128 : RTLIB::OGE_PPCF128;
break;
case ISD::SETULE:
LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
(VT == MVT::f64) ? RTLIB::OGT_F64 :
(VT == MVT::f128) ? RTLIB::OGT_F128 : RTLIB::OGT_PPCF128;
break;
case ISD::SETUGT:
LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
(VT == MVT::f64) ? RTLIB::OLE_F64 :
(VT == MVT::f128) ? RTLIB::OLE_F128 : RTLIB::OLE_PPCF128;
break;
case ISD::SETUGE:
LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
(VT == MVT::f64) ? RTLIB::OLT_F64 :
(VT == MVT::f128) ? RTLIB::OLT_F128 : RTLIB::OLT_PPCF128;
break;
default: llvm_unreachable("Do not know how to soften this setcc!");
}
}
// Use the target specific return value for comparison lib calls.
EVT RetVT = getCmpLibcallReturnType();
SDValue Ops[2] = {NewLHS, NewRHS};
TargetLowering::MakeLibCallOptions CallOptions;
EVT OpsVT[2] = { OldLHS.getValueType(),
OldRHS.getValueType() };
CallOptions.setTypeListBeforeSoften(OpsVT, RetVT, true);
auto Call = makeLibCall(DAG, LC1, RetVT, Ops, CallOptions, dl, Chain);
NewLHS = Call.first;
NewRHS = DAG.getConstant(0, dl, RetVT);
CCCode = getCmpLibcallCC(LC1);
if (ShouldInvertCC) {
assert(RetVT.isInteger());
CCCode = getSetCCInverse(CCCode, RetVT);
}
if (LC2 == RTLIB::UNKNOWN_LIBCALL) {
// Update Chain.
Chain = Call.second;
} else {
EVT SetCCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), RetVT);
SDValue Tmp = DAG.getSetCC(dl, SetCCVT, NewLHS, NewRHS, CCCode);
auto Call2 = makeLibCall(DAG, LC2, RetVT, Ops, CallOptions, dl, Chain);
CCCode = getCmpLibcallCC(LC2);
if (ShouldInvertCC)
CCCode = getSetCCInverse(CCCode, RetVT);
NewLHS = DAG.getSetCC(dl, SetCCVT, Call2.first, NewRHS, CCCode);
if (Chain)
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Call.second,
Call2.second);
NewLHS = DAG.getNode(ShouldInvertCC ? ISD::AND : ISD::OR, dl,
Tmp.getValueType(), Tmp, NewLHS);
NewRHS = SDValue();
}
}
/// Return the entry encoding for a jump table in the current function. The
/// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum.
unsigned TargetLowering::getJumpTableEncoding() const {
// In non-pic modes, just use the address of a block.
if (!isPositionIndependent())
return MachineJumpTableInfo::EK_BlockAddress;
// In PIC mode, if the target supports a GPRel32 directive, use it.
if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != nullptr)
return MachineJumpTableInfo::EK_GPRel32BlockAddress;
// Otherwise, use a label difference.
return MachineJumpTableInfo::EK_LabelDifference32;
}
SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
SelectionDAG &DAG) const {
// If our PIC model is GP relative, use the global offset table as the base.
unsigned JTEncoding = getJumpTableEncoding();
if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) ||
(JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress))
return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(DAG.getDataLayout()));
return Table;
}
/// This returns the relocation base for the given PIC jumptable, the same as
/// getPICJumpTableRelocBase, but as an MCExpr.
const MCExpr *
TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
unsigned JTI,MCContext &Ctx) const{
// The normal PIC reloc base is the label at the start of the jump table.
return MCSymbolRefExpr::create(MF->getJTISymbol(JTI, Ctx), Ctx);
}
SDValue TargetLowering::expandIndirectJTBranch(const SDLoc &dl, SDValue Value,
SDValue Addr, int JTI,
SelectionDAG &DAG) const {
SDValue Chain = Value;
// Jump table debug info is only needed if CodeView is enabled.
if (DAG.getTarget().getTargetTriple().isOSBinFormatCOFF()) {
Chain = DAG.getJumpTableDebugInfo(JTI, Chain, dl);
}
return DAG.getNode(ISD::BRIND, dl, MVT::Other, Chain, Addr);
}
bool
TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
const TargetMachine &TM = getTargetMachine();
const GlobalValue *GV = GA->getGlobal();
// If the address is not even local to this DSO we will have to load it from
// a got and then add the offset.
if (!TM.shouldAssumeDSOLocal(GV))
return false;
// If the code is position independent we will have to add a base register.
if (isPositionIndependent())
return false;
// Otherwise we can do it.
return true;
}
//===----------------------------------------------------------------------===//
// Optimization Methods
//===----------------------------------------------------------------------===//
/// If the specified instruction has a constant integer operand and there are
/// bits set in that constant that are not demanded, then clear those bits and
/// return true.
bool TargetLowering::ShrinkDemandedConstant(SDValue Op,
const APInt &DemandedBits,
const APInt &DemandedElts,
TargetLoweringOpt &TLO) const {
SDLoc DL(Op);
unsigned Opcode = Op.getOpcode();
// Early-out if we've ended up calling an undemanded node, leave this to
// constant folding.
if (DemandedBits.isZero() || DemandedElts.isZero())
return false;
// Do target-specific constant optimization.
if (targetShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO))
return TLO.New.getNode();
// FIXME: ISD::SELECT, ISD::SELECT_CC
switch (Opcode) {
default:
break;
case ISD::XOR:
case ISD::AND:
case ISD::OR: {
auto *Op1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
if (!Op1C || Op1C->isOpaque())
return false;
// If this is a 'not' op, don't touch it because that's a canonical form.
const APInt &C = Op1C->getAPIntValue();
if (Opcode == ISD::XOR && DemandedBits.isSubsetOf(C))
return false;
if (!C.isSubsetOf(DemandedBits)) {
EVT VT = Op.getValueType();
SDValue NewC = TLO.DAG.getConstant(DemandedBits & C, DL, VT);
SDValue NewOp = TLO.DAG.getNode(Opcode, DL, VT, Op.getOperand(0), NewC,
Op->getFlags());
return TLO.CombineTo(Op, NewOp);
}
break;
}
}
return false;
}
bool TargetLowering::ShrinkDemandedConstant(SDValue Op,
const APInt &DemandedBits,
TargetLoweringOpt &TLO) const {
EVT VT = Op.getValueType();
APInt DemandedElts = VT.isVector()
? APInt::getAllOnes(VT.getVectorNumElements())
: APInt(1, 1);
return ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO);
}
/// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free.
/// This uses isTruncateFree/isZExtFree and ANY_EXTEND for the widening cast,
/// but it could be generalized for targets with other types of implicit
/// widening casts.
bool TargetLowering::ShrinkDemandedOp(SDValue Op, unsigned BitWidth,
const APInt &DemandedBits,
TargetLoweringOpt &TLO) const {
assert(Op.getNumOperands() == 2 &&
"ShrinkDemandedOp only supports binary operators!");
assert(Op.getNode()->getNumValues() == 1 &&
"ShrinkDemandedOp only supports nodes with one result!");
EVT VT = Op.getValueType();
SelectionDAG &DAG = TLO.DAG;
SDLoc dl(Op);
// Early return, as this function cannot handle vector types.
if (VT.isVector())
return false;
assert(Op.getOperand(0).getValueType().getScalarSizeInBits() == BitWidth &&
Op.getOperand(1).getValueType().getScalarSizeInBits() == BitWidth &&
"ShrinkDemandedOp only supports operands that have the same size!");
// Don't do this if the node has another user, which may require the
// full value.
if (!Op.getNode()->hasOneUse())
return false;
// Search for the smallest integer type with free casts to and from
// Op's type. For expedience, just check power-of-2 integer types.
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
unsigned DemandedSize = DemandedBits.getActiveBits();
for (unsigned SmallVTBits = llvm::bit_ceil(DemandedSize);
SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
if (TLI.isTruncateFree(VT, SmallVT) && TLI.isZExtFree(SmallVT, VT)) {
// We found a type with free casts.
SDValue X = DAG.getNode(
Op.getOpcode(), dl, SmallVT,
DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(0)),
DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(1)));
assert(DemandedSize <= SmallVTBits && "Narrowed below demanded bits?");
SDValue Z = DAG.getNode(ISD::ANY_EXTEND, dl, VT, X);
return TLO.CombineTo(Op, Z);
}
}
return false;
}
bool TargetLowering::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
!DCI.isBeforeLegalizeOps());
KnownBits Known;
bool Simplified = SimplifyDemandedBits(Op, DemandedBits, Known, TLO);
if (Simplified) {
DCI.AddToWorklist(Op.getNode());
DCI.CommitTargetLoweringOpt(TLO);
}
return Simplified;
}
bool TargetLowering::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits,
const APInt &DemandedElts,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
!DCI.isBeforeLegalizeOps());
KnownBits Known;
bool Simplified =
SimplifyDemandedBits(Op, DemandedBits, DemandedElts, Known, TLO);
if (Simplified) {
DCI.AddToWorklist(Op.getNode());
DCI.CommitTargetLoweringOpt(TLO);
}
return Simplified;
}
bool TargetLowering::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits,
KnownBits &Known,
TargetLoweringOpt &TLO,
unsigned Depth,
bool AssumeSingleUse) const {
EVT VT = Op.getValueType();
// Since the number of lanes in a scalable vector is unknown at compile time,
// we track one bit which is implicitly broadcast to all lanes. This means
// that all lanes in a scalable vector are considered demanded.
APInt DemandedElts = VT.isFixedLengthVector()
? APInt::getAllOnes(VT.getVectorNumElements())
: APInt(1, 1);
return SimplifyDemandedBits(Op, DemandedBits, DemandedElts, Known, TLO, Depth,
AssumeSingleUse);
}
// TODO: Under what circumstances can we create nodes? Constant folding?
SDValue TargetLowering::SimplifyMultipleUseDemandedBits(
SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
SelectionDAG &DAG, unsigned Depth) const {
EVT VT = Op.getValueType();
// Limit search depth.
if (Depth >= SelectionDAG::MaxRecursionDepth)
return SDValue();
// Ignore UNDEFs.
if (Op.isUndef())
return SDValue();
// Not demanding any bits/elts from Op.
if (DemandedBits == 0 || DemandedElts == 0)
return DAG.getUNDEF(VT);
bool IsLE = DAG.getDataLayout().isLittleEndian();
unsigned NumElts = DemandedElts.getBitWidth();
unsigned BitWidth = DemandedBits.getBitWidth();
KnownBits LHSKnown, RHSKnown;
switch (Op.getOpcode()) {
case ISD::BITCAST: {
if (VT.isScalableVector())
return SDValue();
SDValue Src = peekThroughBitcasts(Op.getOperand(0));
EVT SrcVT = Src.getValueType();
EVT DstVT = Op.getValueType();
if (SrcVT == DstVT)
return Src;
unsigned NumSrcEltBits = SrcVT.getScalarSizeInBits();
unsigned NumDstEltBits = DstVT.getScalarSizeInBits();
if (NumSrcEltBits == NumDstEltBits)
if (SDValue V = SimplifyMultipleUseDemandedBits(
Src, DemandedBits, DemandedElts, DAG, Depth + 1))
return DAG.getBitcast(DstVT, V);
if (SrcVT.isVector() && (NumDstEltBits % NumSrcEltBits) == 0) {
unsigned Scale = NumDstEltBits / NumSrcEltBits;
unsigned NumSrcElts = SrcVT.getVectorNumElements();
APInt DemandedSrcBits = APInt::getZero(NumSrcEltBits);
APInt DemandedSrcElts = APInt::getZero(NumSrcElts);
for (unsigned i = 0; i != Scale; ++i) {
unsigned EltOffset = IsLE ? i : (Scale - 1 - i);
unsigned BitOffset = EltOffset * NumSrcEltBits;
APInt Sub = DemandedBits.extractBits(NumSrcEltBits, BitOffset);
if (!Sub.isZero()) {
DemandedSrcBits |= Sub;
for (unsigned j = 0; j != NumElts; ++j)
if (DemandedElts[j])
DemandedSrcElts.setBit((j * Scale) + i);
}
}
if (SDValue V = SimplifyMultipleUseDemandedBits(
Src, DemandedSrcBits, DemandedSrcElts, DAG, Depth + 1))
return DAG.getBitcast(DstVT, V);
}
// TODO - bigendian once we have test coverage.
if (IsLE && (NumSrcEltBits % NumDstEltBits) == 0) {
unsigned Scale = NumSrcEltBits / NumDstEltBits;
unsigned NumSrcElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1;
APInt DemandedSrcBits = APInt::getZero(NumSrcEltBits);
APInt DemandedSrcElts = APInt::getZero(NumSrcElts);
for (unsigned i = 0; i != NumElts; ++i)
if (DemandedElts[i]) {
unsigned Offset = (i % Scale) * NumDstEltBits;
DemandedSrcBits.insertBits(DemandedBits, Offset);
DemandedSrcElts.setBit(i / Scale);
}
if (SDValue V = SimplifyMultipleUseDemandedBits(
Src, DemandedSrcBits, DemandedSrcElts, DAG, Depth + 1))
return DAG.getBitcast(DstVT, V);
}
break;
}
case ISD::FREEZE: {
SDValue N0 = Op.getOperand(0);
if (DAG.isGuaranteedNotToBeUndefOrPoison(N0, DemandedElts,
/*PoisonOnly=*/false))
return N0;
break;
}
case ISD::AND: {
LHSKnown = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
RHSKnown = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
// If all of the demanded bits are known 1 on one side, return the other.
// These bits cannot contribute to the result of the 'and' in this
// context.
if (DemandedBits.isSubsetOf(LHSKnown.Zero | RHSKnown.One))
return Op.getOperand(0);
if (DemandedBits.isSubsetOf(RHSKnown.Zero | LHSKnown.One))
return Op.getOperand(1);
break;
}
case ISD::OR: {
LHSKnown = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
RHSKnown = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
// If all of the demanded bits are known zero on one side, return the
// other. These bits cannot contribute to the result of the 'or' in this
// context.
if (DemandedBits.isSubsetOf(LHSKnown.One | RHSKnown.Zero))
return Op.getOperand(0);
if (DemandedBits.isSubsetOf(RHSKnown.One | LHSKnown.Zero))
return Op.getOperand(1);
break;
}
case ISD::XOR: {
LHSKnown = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
RHSKnown = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
// If all of the demanded bits are known zero on one side, return the
// other.
if (DemandedBits.isSubsetOf(RHSKnown.Zero))
return Op.getOperand(0);
if (DemandedBits.isSubsetOf(LHSKnown.Zero))
return Op.getOperand(1);
break;
}
case ISD::SHL: {
// If we are only demanding sign bits then we can use the shift source
// directly.
if (std::optional<uint64_t> MaxSA =
DAG.getValidMaximumShiftAmount(Op, DemandedElts, Depth + 1)) {
SDValue Op0 = Op.getOperand(0);
unsigned ShAmt = *MaxSA;
unsigned NumSignBits =
DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1);
unsigned UpperDemandedBits = BitWidth - DemandedBits.countr_zero();
if (NumSignBits > ShAmt && (NumSignBits - ShAmt) >= (UpperDemandedBits))
return Op0;
}
break;
}
case ISD::SETCC: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
// If (1) we only need the sign-bit, (2) the setcc operands are the same
// width as the setcc result, and (3) the result of a setcc conforms to 0 or
// -1, we may be able to bypass the setcc.
if (DemandedBits.isSignMask() &&
Op0.getScalarValueSizeInBits() == BitWidth &&
getBooleanContents(Op0.getValueType()) ==
BooleanContent::ZeroOrNegativeOneBooleanContent) {
// If we're testing X < 0, then this compare isn't needed - just use X!
// FIXME: We're limiting to integer types here, but this should also work
// if we don't care about FP signed-zero. The use of SETLT with FP means
// that we don't care about NaNs.
if (CC == ISD::SETLT && Op1.getValueType().isInteger() &&
(isNullConstant(Op1) || ISD::isBuildVectorAllZeros(Op1.getNode())))
return Op0;
}
break;
}
case ISD::SIGN_EXTEND_INREG: {
// If none of the extended bits are demanded, eliminate the sextinreg.
SDValue Op0 = Op.getOperand(0);
EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
unsigned ExBits = ExVT.getScalarSizeInBits();
if (DemandedBits.getActiveBits() <= ExBits &&
shouldRemoveRedundantExtend(Op))
return Op0;
// If the input is already sign extended, just drop the extension.
unsigned NumSignBits = DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1);
if (NumSignBits >= (BitWidth - ExBits + 1))
return Op0;
break;
}
case ISD::ANY_EXTEND_VECTOR_INREG:
case ISD::SIGN_EXTEND_VECTOR_INREG:
case ISD::ZERO_EXTEND_VECTOR_INREG: {
if (VT.isScalableVector())
return SDValue();
// If we only want the lowest element and none of extended bits, then we can
// return the bitcasted source vector.
SDValue Src = Op.getOperand(0);
EVT SrcVT = Src.getValueType();
EVT DstVT = Op.getValueType();
if (IsLE && DemandedElts == 1 &&
DstVT.getSizeInBits() == SrcVT.getSizeInBits() &&
DemandedBits.getActiveBits() <= SrcVT.getScalarSizeInBits()) {
return DAG.getBitcast(DstVT, Src);
}
break;
}
case ISD::INSERT_VECTOR_ELT: {
if (VT.isScalableVector())
return SDValue();
// If we don't demand the inserted element, return the base vector.
SDValue Vec = Op.getOperand(0);
auto *CIdx = dyn_cast<ConstantSDNode>(Op.getOperand(2));
EVT VecVT = Vec.getValueType();
if (CIdx && CIdx->getAPIntValue().ult(VecVT.getVectorNumElements()) &&
!DemandedElts[CIdx->getZExtValue()])
return Vec;
break;
}
case ISD::INSERT_SUBVECTOR: {
if (VT.isScalableVector())
return SDValue();
SDValue Vec = Op.getOperand(0);
SDValue Sub = Op.getOperand(1);
uint64_t Idx = Op.getConstantOperandVal(2);
unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx);
// If we don't demand the inserted subvector, return the base vector.
if (DemandedSubElts == 0)
return Vec;
break;
}
case ISD::VECTOR_SHUFFLE: {
assert(!VT.isScalableVector());
ArrayRef<int> ShuffleMask = cast<ShuffleVectorSDNode>(Op)->getMask();
// If all the demanded elts are from one operand and are inline,
// then we can use the operand directly.
bool AllUndef = true, IdentityLHS = true, IdentityRHS = true;
for (unsigned i = 0; i != NumElts; ++i) {
int M = ShuffleMask[i];
if (M < 0 || !DemandedElts[i])
continue;
AllUndef = false;
IdentityLHS &= (M == (int)i);
IdentityRHS &= ((M - NumElts) == i);
}
if (AllUndef)
return DAG.getUNDEF(Op.getValueType());
if (IdentityLHS)
return Op.getOperand(0);
if (IdentityRHS)
return Op.getOperand(1);
break;
}
default:
// TODO: Probably okay to remove after audit; here to reduce change size
// in initial enablement patch for scalable vectors
if (VT.isScalableVector())
return SDValue();
if (Op.getOpcode() >= ISD::BUILTIN_OP_END)
if (SDValue V = SimplifyMultipleUseDemandedBitsForTargetNode(
Op, DemandedBits, DemandedElts, DAG, Depth))
return V;
break;
}
return SDValue();
}
SDValue TargetLowering::SimplifyMultipleUseDemandedBits(
SDValue Op, const APInt &DemandedBits, SelectionDAG &DAG,
unsigned Depth) const {
EVT VT = Op.getValueType();
// Since the number of lanes in a scalable vector is unknown at compile time,
// we track one bit which is implicitly broadcast to all lanes. This means
// that all lanes in a scalable vector are considered demanded.
APInt DemandedElts = VT.isFixedLengthVector()
? APInt::getAllOnes(VT.getVectorNumElements())
: APInt(1, 1);
return SimplifyMultipleUseDemandedBits(Op, DemandedBits, DemandedElts, DAG,
Depth);
}
SDValue TargetLowering::SimplifyMultipleUseDemandedVectorElts(
SDValue Op, const APInt &DemandedElts, SelectionDAG &DAG,
unsigned Depth) const {
APInt DemandedBits = APInt::getAllOnes(Op.getScalarValueSizeInBits());
return SimplifyMultipleUseDemandedBits(Op, DemandedBits, DemandedElts, DAG,
Depth);
}
// Attempt to form ext(avgfloor(A, B)) from shr(add(ext(A), ext(B)), 1).
// or to form ext(avgceil(A, B)) from shr(add(ext(A), ext(B), 1), 1).
static SDValue combineShiftToAVG(SDValue Op,
TargetLowering::TargetLoweringOpt &TLO,
const TargetLowering &TLI,
const APInt &DemandedBits,
const APInt &DemandedElts, unsigned Depth) {
assert((Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) &&
"SRL or SRA node is required here!");
// Is the right shift using an immediate value of 1?
ConstantSDNode *N1C = isConstOrConstSplat(Op.getOperand(1), DemandedElts);
if (!N1C || !N1C->isOne())
return SDValue();
// We are looking for an avgfloor
// add(ext, ext)
// or one of these as a avgceil
// add(add(ext, ext), 1)
// add(add(ext, 1), ext)
// add(ext, add(ext, 1))
SDValue Add = Op.getOperand(0);
if (Add.getOpcode() != ISD::ADD)
return SDValue();
SDValue ExtOpA = Add.getOperand(0);
SDValue ExtOpB = Add.getOperand(1);
SDValue Add2;
auto MatchOperands = [&](SDValue Op1, SDValue Op2, SDValue Op3, SDValue A) {
ConstantSDNode *ConstOp;
if ((ConstOp = isConstOrConstSplat(Op2, DemandedElts)) &&
ConstOp->isOne()) {
ExtOpA = Op1;
ExtOpB = Op3;
Add2 = A;
return true;
}
if ((ConstOp = isConstOrConstSplat(Op3, DemandedElts)) &&
ConstOp->isOne()) {
ExtOpA = Op1;
ExtOpB = Op2;
Add2 = A;
return true;
}
return false;
};
bool IsCeil =
(ExtOpA.getOpcode() == ISD::ADD &&
MatchOperands(ExtOpA.getOperand(0), ExtOpA.getOperand(1), ExtOpB, ExtOpA)) ||
(ExtOpB.getOpcode() == ISD::ADD &&
MatchOperands(ExtOpB.getOperand(0), ExtOpB.getOperand(1), ExtOpA, ExtOpB));
// If the shift is signed (sra):
// - Needs >= 2 sign bit for both operands.
// - Needs >= 2 zero bits.
// If the shift is unsigned (srl):
// - Needs >= 1 zero bit for both operands.
// - Needs 1 demanded bit zero and >= 2 sign bits.
SelectionDAG &DAG = TLO.DAG;
unsigned ShiftOpc = Op.getOpcode();
bool IsSigned = false;
unsigned KnownBits;
unsigned NumSignedA = DAG.ComputeNumSignBits(ExtOpA, DemandedElts, Depth);
unsigned NumSignedB = DAG.ComputeNumSignBits(ExtOpB, DemandedElts, Depth);
unsigned NumSigned = std::min(NumSignedA, NumSignedB) - 1;
unsigned NumZeroA =
DAG.computeKnownBits(ExtOpA, DemandedElts, Depth).countMinLeadingZeros();
unsigned NumZeroB =
DAG.computeKnownBits(ExtOpB, DemandedElts, Depth).countMinLeadingZeros();
unsigned NumZero = std::min(NumZeroA, NumZeroB);
switch (ShiftOpc) {
default:
llvm_unreachable("Unexpected ShiftOpc in combineShiftToAVG");
case ISD::SRA: {
if (NumZero >= 2 && NumSigned < NumZero) {
IsSigned = false;
KnownBits = NumZero;
break;
}
if (NumSigned >= 1) {
IsSigned = true;
KnownBits = NumSigned;
break;
}
return SDValue();
}
case ISD::SRL: {
if (NumZero >= 1 && NumSigned < NumZero) {
IsSigned = false;
KnownBits = NumZero;
break;
}
if (NumSigned >= 1 && DemandedBits.isSignBitClear()) {
IsSigned = true;
KnownBits = NumSigned;
break;
}
return SDValue();
}
}
unsigned AVGOpc = IsCeil ? (IsSigned ? ISD::AVGCEILS : ISD::AVGCEILU)
: (IsSigned ? ISD::AVGFLOORS : ISD::AVGFLOORU);
// Find the smallest power-2 type that is legal for this vector size and
// operation, given the original type size and the number of known sign/zero
// bits.
EVT VT = Op.getValueType();
unsigned MinWidth =
std::max<unsigned>(VT.getScalarSizeInBits() - KnownBits, 8);
EVT NVT = EVT::getIntegerVT(*DAG.getContext(), llvm::bit_ceil(MinWidth));
if (NVT.getScalarSizeInBits() > VT.getScalarSizeInBits())
return SDValue();
if (VT.isVector())
NVT = EVT::getVectorVT(*DAG.getContext(), NVT, VT.getVectorElementCount());
if (TLO.LegalTypes() && !TLI.isOperationLegal(AVGOpc, NVT)) {
// If we could not transform, and (both) adds are nuw/nsw, we can use the
// larger type size to do the transform.
if (TLO.LegalOperations() && !TLI.isOperationLegal(AVGOpc, VT))
return SDValue();
if (DAG.willNotOverflowAdd(IsSigned, Add.getOperand(0),
Add.getOperand(1)) &&
(!Add2 || DAG.willNotOverflowAdd(IsSigned, Add2.getOperand(0),
Add2.getOperand(1))))
NVT = VT;
else
return SDValue();
}
// Don't create a AVGFLOOR node with a scalar constant unless its legal as
// this is likely to stop other folds (reassociation, value tracking etc.)
if (!IsCeil && !TLI.isOperationLegal(AVGOpc, NVT) &&
(isa<ConstantSDNode>(ExtOpA) || isa<ConstantSDNode>(ExtOpB)))
return SDValue();
SDLoc DL(Op);
SDValue ResultAVG =
DAG.getNode(AVGOpc, DL, NVT, DAG.getExtOrTrunc(IsSigned, ExtOpA, DL, NVT),
DAG.getExtOrTrunc(IsSigned, ExtOpB, DL, NVT));
return DAG.getExtOrTrunc(IsSigned, ResultAVG, DL, VT);
}
/// Look at Op. At this point, we know that only the OriginalDemandedBits of the
/// result of Op are ever used downstream. If we can use this information to
/// simplify Op, create a new simplified DAG node and return true, returning the
/// original and new nodes in Old and New. Otherwise, analyze the expression and
/// return a mask of Known bits for the expression (used to simplify the
/// caller). The Known bits may only be accurate for those bits in the
/// OriginalDemandedBits and OriginalDemandedElts.
bool TargetLowering::SimplifyDemandedBits(
SDValue Op, const APInt &OriginalDemandedBits,
const APInt &OriginalDemandedElts, KnownBits &Known, TargetLoweringOpt &TLO,
unsigned Depth, bool AssumeSingleUse) const {
unsigned BitWidth = OriginalDemandedBits.getBitWidth();
assert(Op.getScalarValueSizeInBits() == BitWidth &&
"Mask size mismatches value type size!");
// Don't know anything.
Known = KnownBits(BitWidth);
EVT VT = Op.getValueType();
bool IsLE = TLO.DAG.getDataLayout().isLittleEndian();
unsigned NumElts = OriginalDemandedElts.getBitWidth();
assert((!VT.isFixedLengthVector() || NumElts == VT.getVectorNumElements()) &&
"Unexpected vector size");
APInt DemandedBits = OriginalDemandedBits;
APInt DemandedElts = OriginalDemandedElts;
SDLoc dl(Op);
// Undef operand.
if (Op.isUndef())
return false;
// We can't simplify target constants.
if (Op.getOpcode() == ISD::TargetConstant)
return false;
if (Op.getOpcode() == ISD::Constant) {
// We know all of the bits for a constant!
Known = KnownBits::makeConstant(Op->getAsAPIntVal());
return false;
}
if (Op.getOpcode() == ISD::ConstantFP) {
// We know all of the bits for a floating point constant!
Known = KnownBits::makeConstant(
cast<ConstantFPSDNode>(Op)->getValueAPF().bitcastToAPInt());
return false;
}
// Other users may use these bits.
bool HasMultiUse = false;
if (!AssumeSingleUse && !Op.getNode()->hasOneUse()) {
if (Depth >= SelectionDAG::MaxRecursionDepth) {
// Limit search depth.
return false;
}
// Allow multiple uses, just set the DemandedBits/Elts to all bits.
DemandedBits = APInt::getAllOnes(BitWidth);
DemandedElts = APInt::getAllOnes(NumElts);
HasMultiUse = true;
} else if (OriginalDemandedBits == 0 || OriginalDemandedElts == 0) {
// Not demanding any bits/elts from Op.
return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT));
} else if (Depth >= SelectionDAG::MaxRecursionDepth) {
// Limit search depth.
return false;
}
KnownBits Known2;
switch (Op.getOpcode()) {
case ISD::SCALAR_TO_VECTOR: {
if (VT.isScalableVector())
return false;
if (!DemandedElts[0])
return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT));
KnownBits SrcKnown;
SDValue Src = Op.getOperand(0);
unsigned SrcBitWidth = Src.getScalarValueSizeInBits();
APInt SrcDemandedBits = DemandedBits.zext(SrcBitWidth);
if (SimplifyDemandedBits(Src, SrcDemandedBits, SrcKnown, TLO, Depth + 1))
return true;
// Upper elements are undef, so only get the knownbits if we just demand
// the bottom element.
if (DemandedElts == 1)
Known = SrcKnown.anyextOrTrunc(BitWidth);
break;
}
case ISD::BUILD_VECTOR:
// Collect the known bits that are shared by every demanded element.
// TODO: Call SimplifyDemandedBits for non-constant demanded elements.
Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
return false; // Don't fall through, will infinitely loop.
case ISD::SPLAT_VECTOR: {
SDValue Scl = Op.getOperand(0);
APInt DemandedSclBits = DemandedBits.zextOrTrunc(Scl.getValueSizeInBits());
KnownBits KnownScl;
if (SimplifyDemandedBits(Scl, DemandedSclBits, KnownScl, TLO, Depth + 1))
return true;
// Implicitly truncate the bits to match the official semantics of
// SPLAT_VECTOR.
Known = KnownScl.trunc(BitWidth);
break;
}
case ISD::LOAD: {
auto *LD = cast<LoadSDNode>(Op);
if (getTargetConstantFromLoad(LD)) {
Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
return false; // Don't fall through, will infinitely loop.
}
if (ISD::isZEXTLoad(Op.getNode()) && Op.getResNo() == 0) {
// If this is a ZEXTLoad and we are looking at the loaded value.
EVT MemVT = LD->getMemoryVT();
unsigned MemBits = MemVT.getScalarSizeInBits();
Known.Zero.setBitsFrom(MemBits);
return false; // Don't fall through, will infinitely loop.
}
break;
}
case ISD::INSERT_VECTOR_ELT: {
if (VT.isScalableVector())
return false;
SDValue Vec = Op.getOperand(0);
SDValue Scl = Op.getOperand(1);
auto *CIdx = dyn_cast<ConstantSDNode>(Op.getOperand(2));
EVT VecVT = Vec.getValueType();
// If index isn't constant, assume we need all vector elements AND the
// inserted element.
APInt DemandedVecElts(DemandedElts);
if (CIdx && CIdx->getAPIntValue().ult(VecVT.getVectorNumElements())) {
unsigned Idx = CIdx->getZExtValue();
DemandedVecElts.clearBit(Idx);
// Inserted element is not required.
if (!DemandedElts[Idx])
return TLO.CombineTo(Op, Vec);
}
KnownBits KnownScl;
unsigned NumSclBits = Scl.getScalarValueSizeInBits();
APInt DemandedSclBits = DemandedBits.zextOrTrunc(NumSclBits);
if (SimplifyDemandedBits(Scl, DemandedSclBits, KnownScl, TLO, Depth + 1))
return true;
Known = KnownScl.anyextOrTrunc(BitWidth);
KnownBits KnownVec;
if (SimplifyDemandedBits(Vec, DemandedBits, DemandedVecElts, KnownVec, TLO,
Depth + 1))
return true;
if (!!DemandedVecElts)
Known = Known.intersectWith(KnownVec);
return false;
}
case ISD::INSERT_SUBVECTOR: {
if (VT.isScalableVector())
return false;
// Demand any elements from the subvector and the remainder from the src its
// inserted into.
SDValue Src = Op.getOperand(0);
SDValue Sub = Op.getOperand(1);
uint64_t Idx = Op.getConstantOperandVal(2);
unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx);
APInt DemandedSrcElts = DemandedElts;
DemandedSrcElts.insertBits(APInt::getZero(NumSubElts), Idx);
KnownBits KnownSub, KnownSrc;
if (SimplifyDemandedBits(Sub, DemandedBits, DemandedSubElts, KnownSub, TLO,
Depth + 1))
return true;
if (SimplifyDemandedBits(Src, DemandedBits, DemandedSrcElts, KnownSrc, TLO,
Depth + 1))
return true;
Known.Zero.setAllBits();
Known.One.setAllBits();
if (!!DemandedSubElts)
Known = Known.intersectWith(KnownSub);
if (!!DemandedSrcElts)
Known = Known.intersectWith(KnownSrc);
// Attempt to avoid multi-use src if we don't need anything from it.
if (!DemandedBits.isAllOnes() || !DemandedSubElts.isAllOnes() ||
!DemandedSrcElts.isAllOnes()) {
SDValue NewSub = SimplifyMultipleUseDemandedBits(
Sub, DemandedBits, DemandedSubElts, TLO.DAG, Depth + 1);
SDValue NewSrc = SimplifyMultipleUseDemandedBits(
Src, DemandedBits, DemandedSrcElts, TLO.DAG, Depth + 1);
if (NewSub || NewSrc) {
NewSub = NewSub ? NewSub : Sub;
NewSrc = NewSrc ? NewSrc : Src;
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, NewSrc, NewSub,
Op.getOperand(2));
return TLO.CombineTo(Op, NewOp);
}
}
break;
}
case ISD::EXTRACT_SUBVECTOR: {
if (VT.isScalableVector())
return false;
// Offset the demanded elts by the subvector index.
SDValue Src = Op.getOperand(0);
if (Src.getValueType().isScalableVector())
break;
uint64_t Idx = Op.getConstantOperandVal(1);
unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
if (SimplifyDemandedBits(Src, DemandedBits, DemandedSrcElts, Known, TLO,
Depth + 1))
return true;
// Attempt to avoid multi-use src if we don't need anything from it.
if (!DemandedBits.isAllOnes() || !DemandedSrcElts.isAllOnes()) {
SDValue DemandedSrc = SimplifyMultipleUseDemandedBits(
Src, DemandedBits, DemandedSrcElts, TLO.DAG, Depth + 1);
if (DemandedSrc) {
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, DemandedSrc,
Op.getOperand(1));
return TLO.CombineTo(Op, NewOp);
}
}
break;
}
case ISD::CONCAT_VECTORS: {
if (VT.isScalableVector())
return false;
Known.Zero.setAllBits();
Known.One.setAllBits();
EVT SubVT = Op.getOperand(0).getValueType();
unsigned NumSubVecs = Op.getNumOperands();
unsigned NumSubElts = SubVT.getVectorNumElements();
for (unsigned i = 0; i != NumSubVecs; ++i) {
APInt DemandedSubElts =
DemandedElts.extractBits(NumSubElts, i * NumSubElts);
if (SimplifyDemandedBits(Op.getOperand(i), DemandedBits, DemandedSubElts,
Known2, TLO, Depth + 1))
return true;
// Known bits are shared by every demanded subvector element.
if (!!DemandedSubElts)
Known = Known.intersectWith(Known2);
}
break;
}
case ISD::VECTOR_SHUFFLE: {
assert(!VT.isScalableVector());
ArrayRef<int> ShuffleMask = cast<ShuffleVectorSDNode>(Op)->getMask();
// Collect demanded elements from shuffle operands..
APInt DemandedLHS, DemandedRHS;
if (!getShuffleDemandedElts(NumElts, ShuffleMask, DemandedElts, DemandedLHS,
DemandedRHS))
break;
if (!!DemandedLHS || !!DemandedRHS) {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
Known.Zero.setAllBits();
Known.One.setAllBits();
if (!!DemandedLHS) {
if (SimplifyDemandedBits(Op0, DemandedBits, DemandedLHS, Known2, TLO,
Depth + 1))
return true;
Known = Known.intersectWith(Known2);
}
if (!!DemandedRHS) {
if (SimplifyDemandedBits(Op1, DemandedBits, DemandedRHS, Known2, TLO,
Depth + 1))
return true;
Known = Known.intersectWith(Known2);
}
// Attempt to avoid multi-use ops if we don't need anything from them.
SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits(
Op0, DemandedBits, DemandedLHS, TLO.DAG, Depth + 1);
SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits(
Op1, DemandedBits, DemandedRHS, TLO.DAG, Depth + 1);
if (DemandedOp0 || DemandedOp1) {
Op0 = DemandedOp0 ? DemandedOp0 : Op0;
Op1 = DemandedOp1 ? DemandedOp1 : Op1;
SDValue NewOp = TLO.DAG.getVectorShuffle(VT, dl, Op0, Op1, ShuffleMask);
return TLO.CombineTo(Op, NewOp);
}
}
break;
}
case ISD::AND: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
// If the RHS is a constant, check to see if the LHS would be zero without
// using the bits from the RHS. Below, we use knowledge about the RHS to
// simplify the LHS, here we're using information from the LHS to simplify
// the RHS.
if (ConstantSDNode *RHSC = isConstOrConstSplat(Op1, DemandedElts)) {
// Do not increment Depth here; that can cause an infinite loop.
KnownBits LHSKnown = TLO.DAG.computeKnownBits(Op0, DemandedElts, Depth);
// If the LHS already has zeros where RHSC does, this 'and' is dead.
if ((LHSKnown.Zero & DemandedBits) ==
(~RHSC->getAPIntValue() & DemandedBits))
return TLO.CombineTo(Op, Op0);
// If any of the set bits in the RHS are known zero on the LHS, shrink
// the constant.
if (ShrinkDemandedConstant(Op, ~LHSKnown.Zero & DemandedBits,
DemandedElts, TLO))
return true;
// Bitwise-not (xor X, -1) is a special case: we don't usually shrink its
// constant, but if this 'and' is only clearing bits that were just set by
// the xor, then this 'and' can be eliminated by shrinking the mask of
// the xor. For example, for a 32-bit X:
// and (xor (srl X, 31), -1), 1 --> xor (srl X, 31), 1
if (isBitwiseNot(Op0) && Op0.hasOneUse() &&
LHSKnown.One == ~RHSC->getAPIntValue()) {
SDValue Xor = TLO.DAG.getNode(ISD::XOR, dl, VT, Op0.getOperand(0), Op1);
return TLO.CombineTo(Op, Xor);
}
}
// AND(INSERT_SUBVECTOR(C,X,I),M) -> INSERT_SUBVECTOR(AND(C,M),X,I)
// iff 'C' is Undef/Constant and AND(X,M) == X (for DemandedBits).
if (Op0.getOpcode() == ISD::INSERT_SUBVECTOR && !VT.isScalableVector() &&
(Op0.getOperand(0).isUndef() ||
ISD::isBuildVectorOfConstantSDNodes(Op0.getOperand(0).getNode())) &&
Op0->hasOneUse()) {
unsigned NumSubElts =
Op0.getOperand(1).getValueType().getVectorNumElements();
unsigned SubIdx = Op0.getConstantOperandVal(2);
APInt DemandedSub =
APInt::getBitsSet(NumElts, SubIdx, SubIdx + NumSubElts);
KnownBits KnownSubMask =
TLO.DAG.computeKnownBits(Op1, DemandedSub & DemandedElts, Depth + 1);
if (DemandedBits.isSubsetOf(KnownSubMask.One)) {
SDValue NewAnd =
TLO.DAG.getNode(ISD::AND, dl, VT, Op0.getOperand(0), Op1);
SDValue NewInsert =
TLO.DAG.getNode(ISD::INSERT_SUBVECTOR, dl, VT, NewAnd,
Op0.getOperand(1), Op0.getOperand(2));
return TLO.CombineTo(Op, NewInsert);
}
}
if (SimplifyDemandedBits(Op1, DemandedBits, DemandedElts, Known, TLO,
Depth + 1))
return true;
if (SimplifyDemandedBits(Op0, ~Known.Zero & DemandedBits, DemandedElts,
Known2, TLO, Depth + 1))
return true;
// If all of the demanded bits are known one on one side, return the other.
// These bits cannot contribute to the result of the 'and'.
if (DemandedBits.isSubsetOf(Known2.Zero | Known.One))
return TLO.CombineTo(Op, Op0);
if (DemandedBits.isSubsetOf(Known.Zero | Known2.One))
return TLO.CombineTo(Op, Op1);
// If all of the demanded bits in the inputs are known zeros, return zero.
if (DemandedBits.isSubsetOf(Known.Zero | Known2.Zero))
return TLO.CombineTo(Op, TLO.DAG.getConstant(0, dl, VT));
// If the RHS is a constant, see if we can simplify it.
if (ShrinkDemandedConstant(Op, ~Known2.Zero & DemandedBits, DemandedElts,
TLO))
return true;
// If the operation can be done in a smaller type, do so.
if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO))
return true;
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!DemandedBits.isAllOnes() || !DemandedElts.isAllOnes()) {
SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits(
Op0, DemandedBits, DemandedElts, TLO.DAG, Depth + 1);
SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits(
Op1, DemandedBits, DemandedElts, TLO.DAG, Depth + 1);
if (DemandedOp0 || DemandedOp1) {
Op0 = DemandedOp0 ? DemandedOp0 : Op0;
Op1 = DemandedOp1 ? DemandedOp1 : Op1;
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1);
return TLO.CombineTo(Op, NewOp);
}
}
Known &= Known2;
break;
}
case ISD::OR: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
SDNodeFlags Flags = Op.getNode()->getFlags();
if (SimplifyDemandedBits(Op1, DemandedBits, DemandedElts, Known, TLO,
Depth + 1)) {
if (Flags.hasDisjoint()) {
Flags.setDisjoint(false);
Op->setFlags(Flags);
}
return true;
}
if (SimplifyDemandedBits(Op0, ~Known.One & DemandedBits, DemandedElts,
Known2, TLO, Depth + 1)) {
if (Flags.hasDisjoint()) {
Flags.setDisjoint(false);
Op->setFlags(Flags);
}
return true;
}
// If all of the demanded bits are known zero on one side, return the other.
// These bits cannot contribute to the result of the 'or'.
if (DemandedBits.isSubsetOf(Known2.One | Known.Zero))
return TLO.CombineTo(Op, Op0);
if (DemandedBits.isSubsetOf(Known.One | Known2.Zero))
return TLO.CombineTo(Op, Op1);
// If the RHS is a constant, see if we can simplify it.
if (ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO))
return true;
// If the operation can be done in a smaller type, do so.
if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO))
return true;
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!DemandedBits.isAllOnes() || !DemandedElts.isAllOnes()) {
SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits(
Op0, DemandedBits, DemandedElts, TLO.DAG, Depth + 1);
SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits(
Op1, DemandedBits, DemandedElts, TLO.DAG, Depth + 1);
if (DemandedOp0 || DemandedOp1) {
Op0 = DemandedOp0 ? DemandedOp0 : Op0;
Op1 = DemandedOp1 ? DemandedOp1 : Op1;
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1);
return TLO.CombineTo(Op, NewOp);
}
}
// (or (and X, C1), (and (or X, Y), C2)) -> (or (and X, C1|C2), (and Y, C2))
// TODO: Use SimplifyMultipleUseDemandedBits to peek through masks.
if (Op0.getOpcode() == ISD::AND && Op1.getOpcode() == ISD::AND &&
Op0->hasOneUse() && Op1->hasOneUse()) {
// Attempt to match all commutations - m_c_Or would've been useful!
for (int I = 0; I != 2; ++I) {
SDValue X = Op.getOperand(I).getOperand(0);
SDValue C1 = Op.getOperand(I).getOperand(1);
SDValue Alt = Op.getOperand(1 - I).getOperand(0);
SDValue C2 = Op.getOperand(1 - I).getOperand(1);
if (Alt.getOpcode() == ISD::OR) {
for (int J = 0; J != 2; ++J) {
if (X == Alt.getOperand(J)) {
SDValue Y = Alt.getOperand(1 - J);
if (SDValue C12 = TLO.DAG.FoldConstantArithmetic(ISD::OR, dl, VT,
{C1, C2})) {
SDValue MaskX = TLO.DAG.getNode(ISD::AND, dl, VT, X, C12);
SDValue MaskY = TLO.DAG.getNode(ISD::AND, dl, VT, Y, C2);
return TLO.CombineTo(
Op, TLO.DAG.getNode(ISD::OR, dl, VT, MaskX, MaskY));
}
}
}
}
}
}
Known |= Known2;
break;
}
case ISD::XOR: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
if (SimplifyDemandedBits(Op1, DemandedBits, DemandedElts, Known, TLO,
Depth + 1))
return true;
if (SimplifyDemandedBits(Op0, DemandedBits, DemandedElts, Known2, TLO,
Depth + 1))
return true;
// If all of the demanded bits are known zero on one side, return the other.
// These bits cannot contribute to the result of the 'xor'.
if (DemandedBits.isSubsetOf(Known.Zero))
return TLO.CombineTo(Op, Op0);
if (DemandedBits.isSubsetOf(Known2.Zero))
return TLO.CombineTo(Op, Op1);
// If the operation can be done in a smaller type, do so.
if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO))
return true;
// If all of the unknown bits are known to be zero on one side or the other
// turn this into an *inclusive* or.
// e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
if (DemandedBits.isSubsetOf(Known.Zero | Known2.Zero))
return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, VT, Op0, Op1));
ConstantSDNode *C = isConstOrConstSplat(Op1, DemandedElts);
if (C) {
// If one side is a constant, and all of the set bits in the constant are
// also known set on the other side, turn this into an AND, as we know
// the bits will be cleared.
// e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
// NB: it is okay if more bits are known than are requested
if (C->getAPIntValue() == Known2.One) {
SDValue ANDC =
TLO.DAG.getConstant(~C->getAPIntValue() & DemandedBits, dl, VT);
return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, Op0, ANDC));
}
// If the RHS is a constant, see if we can change it. Don't alter a -1
// constant because that's a 'not' op, and that is better for combining
// and codegen.
if (!C->isAllOnes() && DemandedBits.isSubsetOf(C->getAPIntValue())) {
// We're flipping all demanded bits. Flip the undemanded bits too.
SDValue New = TLO.DAG.getNOT(dl, Op0, VT);
return TLO.CombineTo(Op, New);
}
unsigned Op0Opcode = Op0.getOpcode();
if ((Op0Opcode == ISD::SRL || Op0Opcode == ISD::SHL) && Op0.hasOneUse()) {
if (ConstantSDNode *ShiftC =
isConstOrConstSplat(Op0.getOperand(1), DemandedElts)) {
// Don't crash on an oversized shift. We can not guarantee that a
// bogus shift has been simplified to undef.
if (ShiftC->getAPIntValue().ult(BitWidth)) {
uint64_t ShiftAmt = ShiftC->getZExtValue();
APInt Ones = APInt::getAllOnes(BitWidth);
Ones = Op0Opcode == ISD::SHL ? Ones.shl(ShiftAmt)
: Ones.lshr(ShiftAmt);
const TargetLowering &TLI = TLO.DAG.getTargetLoweringInfo();
if ((DemandedBits & C->getAPIntValue()) == (DemandedBits & Ones) &&
TLI.isDesirableToCommuteXorWithShift(Op.getNode())) {
// If the xor constant is a demanded mask, do a 'not' before the
// shift:
// xor (X << ShiftC), XorC --> (not X) << ShiftC
// xor (X >> ShiftC), XorC --> (not X) >> ShiftC
SDValue Not = TLO.DAG.getNOT(dl, Op0.getOperand(0), VT);
return TLO.CombineTo(Op, TLO.DAG.getNode(Op0Opcode, dl, VT, Not,
Op0.getOperand(1)));
}
}
}
}
}
// If we can't turn this into a 'not', try to shrink the constant.
if (!C || !C->isAllOnes())
if (ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO))
return true;
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!DemandedBits.isAllOnes() || !DemandedElts.isAllOnes()) {
SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits(
Op0, DemandedBits, DemandedElts, TLO.DAG, Depth + 1);
SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits(
Op1, DemandedBits, DemandedElts, TLO.DAG, Depth + 1);
if (DemandedOp0 || DemandedOp1) {
Op0 = DemandedOp0 ? DemandedOp0 : Op0;
Op1 = DemandedOp1 ? DemandedOp1 : Op1;
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1);
return TLO.CombineTo(Op, NewOp);
}
}
Known ^= Known2;
break;
}
case ISD::SELECT:
if (SimplifyDemandedBits(Op.getOperand(2), DemandedBits, DemandedElts,
Known, TLO, Depth + 1))
return true;
if (SimplifyDemandedBits(Op.getOperand(1), DemandedBits, DemandedElts,
Known2, TLO, Depth + 1))
return true;
// If the operands are constants, see if we can simplify them.
if (ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO))
return true;
// Only known if known in both the LHS and RHS.
Known = Known.intersectWith(Known2);
break;
case ISD::VSELECT:
if (SimplifyDemandedBits(Op.getOperand(2), DemandedBits, DemandedElts,
Known, TLO, Depth + 1))
return true;
if (SimplifyDemandedBits(Op.getOperand(1), DemandedBits, DemandedElts,
Known2, TLO, Depth + 1))
return true;
// Only known if known in both the LHS and RHS.
Known = Known.intersectWith(Known2);
break;
case ISD::SELECT_CC:
if (SimplifyDemandedBits(Op.getOperand(3), DemandedBits, DemandedElts,
Known, TLO, Depth + 1))
return true;
if (SimplifyDemandedBits(Op.getOperand(2), DemandedBits, DemandedElts,
Known2, TLO, Depth + 1))
return true;
// If the operands are constants, see if we can simplify them.
if (ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO))
return true;
// Only known if known in both the LHS and RHS.
Known = Known.intersectWith(Known2);
break;
case ISD::SETCC: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
// If (1) we only need the sign-bit, (2) the setcc operands are the same
// width as the setcc result, and (3) the result of a setcc conforms to 0 or
// -1, we may be able to bypass the setcc.
if (DemandedBits.isSignMask() &&
Op0.getScalarValueSizeInBits() == BitWidth &&
getBooleanContents(Op0.getValueType()) ==
BooleanContent::ZeroOrNegativeOneBooleanContent) {
// If we're testing X < 0, then this compare isn't needed - just use X!
// FIXME: We're limiting to integer types here, but this should also work
// if we don't care about FP signed-zero. The use of SETLT with FP means
// that we don't care about NaNs.
if (CC == ISD::SETLT && Op1.getValueType().isInteger() &&
(isNullConstant(Op1) || ISD::isBuildVectorAllZeros(Op1.getNode())))
return TLO.CombineTo(Op, Op0);
// TODO: Should we check for other forms of sign-bit comparisons?
// Examples: X <= -1, X >= 0
}
if (getBooleanContents(Op0.getValueType()) ==
TargetLowering::ZeroOrOneBooleanContent &&
BitWidth > 1)
Known.Zero.setBitsFrom(1);
break;
}
case ISD::SHL: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
EVT ShiftVT = Op1.getValueType();
if (std::optional<uint64_t> KnownSA =
TLO.DAG.getValidShiftAmount(Op, DemandedElts, Depth + 1)) {
unsigned ShAmt = *KnownSA;
if (ShAmt == 0)
return TLO.CombineTo(Op, Op0);
// If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
// single shift. We can do this if the bottom bits (which are shifted
// out) are never demanded.
// TODO - support non-uniform vector amounts.
if (Op0.getOpcode() == ISD::SRL) {
if (!DemandedBits.intersects(APInt::getLowBitsSet(BitWidth, ShAmt))) {
if (std::optional<uint64_t> InnerSA =
TLO.DAG.getValidShiftAmount(Op0, DemandedElts, Depth + 2)) {
unsigned C1 = *InnerSA;
unsigned Opc = ISD::SHL;
int Diff = ShAmt - C1;
if (Diff < 0) {
Diff = -Diff;
Opc = ISD::SRL;
}
SDValue NewSA = TLO.DAG.getConstant(Diff, dl, ShiftVT);
return TLO.CombineTo(
Op, TLO.DAG.getNode(Opc, dl, VT, Op0.getOperand(0), NewSA));
}
}
}
// Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
// are not demanded. This will likely allow the anyext to be folded away.
// TODO - support non-uniform vector amounts.
if (Op0.getOpcode() == ISD::ANY_EXTEND) {
SDValue InnerOp = Op0.getOperand(0);
EVT InnerVT = InnerOp.getValueType();
unsigned InnerBits = InnerVT.getScalarSizeInBits();
if (ShAmt < InnerBits && DemandedBits.getActiveBits() <= InnerBits &&
isTypeDesirableForOp(ISD::SHL, InnerVT)) {
SDValue NarrowShl = TLO.DAG.getNode(
ISD::SHL, dl, InnerVT, InnerOp,
TLO.DAG.getShiftAmountConstant(ShAmt, InnerVT, dl));
return TLO.CombineTo(
Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT, NarrowShl));
}
// Repeat the SHL optimization above in cases where an extension
// intervenes: (shl (anyext (shr x, c1)), c2) to
// (shl (anyext x), c2-c1). This requires that the bottom c1 bits
// aren't demanded (as above) and that the shifted upper c1 bits of
// x aren't demanded.
// TODO - support non-uniform vector amounts.
if (InnerOp.getOpcode() == ISD::SRL && Op0.hasOneUse() &&
InnerOp.hasOneUse()) {
if (std::optional<uint64_t> SA2 = TLO.DAG.getValidShiftAmount(
InnerOp, DemandedElts, Depth + 2)) {
unsigned InnerShAmt = *SA2;
if (InnerShAmt < ShAmt && InnerShAmt < InnerBits &&
DemandedBits.getActiveBits() <=
(InnerBits - InnerShAmt + ShAmt) &&
DemandedBits.countr_zero() >= ShAmt) {
SDValue NewSA =
TLO.DAG.getConstant(ShAmt - InnerShAmt, dl, ShiftVT);
SDValue NewExt = TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT,
InnerOp.getOperand(0));
return TLO.CombineTo(
Op, TLO.DAG.getNode(ISD::SHL, dl, VT, NewExt, NewSA));
}
}
}
}
APInt InDemandedMask = DemandedBits.lshr(ShAmt);
if (SimplifyDemandedBits(Op0, InDemandedMask, DemandedElts, Known, TLO,
Depth + 1)) {
SDNodeFlags Flags = Op.getNode()->getFlags();
if (Flags.hasNoSignedWrap() || Flags.hasNoUnsignedWrap()) {
// Disable the nsw and nuw flags. We can no longer guarantee that we
// won't wrap after simplification.
Flags.setNoSignedWrap(false);
Flags.setNoUnsignedWrap(false);
Op->setFlags(Flags);
}
return true;
}
Known.Zero <<= ShAmt;
Known.One <<= ShAmt;
// low bits known zero.
Known.Zero.setLowBits(ShAmt);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!InDemandedMask.isAllOnes() || !DemandedElts.isAllOnes()) {
SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits(
Op0, InDemandedMask, DemandedElts, TLO.DAG, Depth + 1);
if (DemandedOp0) {
SDValue NewOp = TLO.DAG.getNode(ISD::SHL, dl, VT, DemandedOp0, Op1);
return TLO.CombineTo(Op, NewOp);
}
}
// TODO: Can we merge this fold with the one below?
// Try shrinking the operation as long as the shift amount will still be
// in range.
if (ShAmt < DemandedBits.getActiveBits() && !VT.isVector() &&
Op.getNode()->hasOneUse()) {
// Search for the smallest integer type with free casts to and from
// Op's type. For expedience, just check power-of-2 integer types.
unsigned DemandedSize = DemandedBits.getActiveBits();
for (unsigned SmallVTBits = llvm::bit_ceil(DemandedSize);
SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
EVT SmallVT = EVT::getIntegerVT(*TLO.DAG.getContext(), SmallVTBits);
if (isNarrowingProfitable(VT, SmallVT) &&
isTypeDesirableForOp(ISD::SHL, SmallVT) &&
isTruncateFree(VT, SmallVT) && isZExtFree(SmallVT, VT) &&
(!TLO.LegalOperations() || isOperationLegal(ISD::SHL, SmallVT))) {
assert(DemandedSize <= SmallVTBits &&
"Narrowed below demanded bits?");
// We found a type with free casts.
SDValue NarrowShl = TLO.DAG.getNode(
ISD::SHL, dl, SmallVT,
TLO.DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(0)),
TLO.DAG.getShiftAmountConstant(ShAmt, SmallVT, dl));
return TLO.CombineTo(
Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT, NarrowShl));
}
}
}
// Narrow shift to lower half - similar to ShrinkDemandedOp.
// (shl i64:x, K) -> (i64 zero_extend (shl (i32 (trunc i64:x)), K))
// Only do this if we demand the upper half so the knownbits are correct.
unsigned HalfWidth = BitWidth / 2;
if ((BitWidth % 2) == 0 && !VT.isVector() && ShAmt < HalfWidth &&
DemandedBits.countLeadingOnes() >= HalfWidth) {
EVT HalfVT = EVT::getIntegerVT(*TLO.DAG.getContext(), HalfWidth);
if (isNarrowingProfitable(VT, HalfVT) &&
isTypeDesirableForOp(ISD::SHL, HalfVT) &&
isTruncateFree(VT, HalfVT) && isZExtFree(HalfVT, VT) &&
(!TLO.LegalOperations() || isOperationLegal(ISD::SHL, HalfVT))) {
// If we're demanding the upper bits at all, we must ensure
// that the upper bits of the shift result are known to be zero,
// which is equivalent to the narrow shift being NUW.
if (bool IsNUW = (Known.countMinLeadingZeros() >= HalfWidth)) {
bool IsNSW = Known.countMinSignBits() > HalfWidth;
SDNodeFlags Flags;
Flags.setNoSignedWrap(IsNSW);
Flags.setNoUnsignedWrap(IsNUW);
SDValue NewOp = TLO.DAG.getNode(ISD::TRUNCATE, dl, HalfVT, Op0);
SDValue NewShiftAmt =
TLO.DAG.getShiftAmountConstant(ShAmt, HalfVT, dl);
SDValue NewShift = TLO.DAG.getNode(ISD::SHL, dl, HalfVT, NewOp,
NewShiftAmt, Flags);
SDValue NewExt =
TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, VT, NewShift);
return TLO.CombineTo(Op, NewExt);
}
}
}
} else {
// This is a variable shift, so we can't shift the demand mask by a known
// amount. But if we are not demanding high bits, then we are not
// demanding those bits from the pre-shifted operand either.
if (unsigned CTLZ = DemandedBits.countl_zero()) {
APInt DemandedFromOp(APInt::getLowBitsSet(BitWidth, BitWidth - CTLZ));
if (SimplifyDemandedBits(Op0, DemandedFromOp, DemandedElts, Known, TLO,
Depth + 1)) {
SDNodeFlags Flags = Op.getNode()->getFlags();
if (Flags.hasNoSignedWrap() || Flags.hasNoUnsignedWrap()) {
// Disable the nsw and nuw flags. We can no longer guarantee that we
// won't wrap after simplification.
Flags.setNoSignedWrap(false);
Flags.setNoUnsignedWrap(false);
Op->setFlags(Flags);
}
return true;
}
Known.resetAll();
}
}
// If we are only demanding sign bits then we can use the shift source
// directly.
if (std::optional<uint64_t> MaxSA =
TLO.DAG.getValidMaximumShiftAmount(Op, DemandedElts, Depth + 1)) {
unsigned ShAmt = *MaxSA;
unsigned NumSignBits =
TLO.DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1);
unsigned UpperDemandedBits = BitWidth - DemandedBits.countr_zero();
if (NumSignBits > ShAmt && (NumSignBits - ShAmt) >= (UpperDemandedBits))
return TLO.CombineTo(Op, Op0);
}
break;
}
case ISD::SRL: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
EVT ShiftVT = Op1.getValueType();
if (std::optional<uint64_t> KnownSA =
TLO.DAG.getValidShiftAmount(Op, DemandedElts, Depth + 1)) {
unsigned ShAmt = *KnownSA;
if (ShAmt == 0)
return TLO.CombineTo(Op, Op0);
// If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
// single shift. We can do this if the top bits (which are shifted out)
// are never demanded.
// TODO - support non-uniform vector amounts.
if (Op0.getOpcode() == ISD::SHL) {
if (!DemandedBits.intersects(APInt::getHighBitsSet(BitWidth, ShAmt))) {
if (std::optional<uint64_t> InnerSA =
TLO.DAG.getValidShiftAmount(Op0, DemandedElts, Depth + 2)) {
unsigned C1 = *InnerSA;
unsigned Opc = ISD::SRL;
int Diff = ShAmt - C1;
if (Diff < 0) {
Diff = -Diff;
Opc = ISD::SHL;
}
SDValue NewSA = TLO.DAG.getConstant(Diff, dl, ShiftVT);
return TLO.CombineTo(
Op, TLO.DAG.getNode(Opc, dl, VT, Op0.getOperand(0), NewSA));
}
}
}
APInt InDemandedMask = (DemandedBits << ShAmt);
// If the shift is exact, then it does demand the low bits (and knows that
// they are zero).
if (Op->getFlags().hasExact())
InDemandedMask.setLowBits(ShAmt);
// Narrow shift to lower half - similar to ShrinkDemandedOp.
// (srl i64:x, K) -> (i64 zero_extend (srl (i32 (trunc i64:x)), K))
if ((BitWidth % 2) == 0 && !VT.isVector()) {
APInt HiBits = APInt::getHighBitsSet(BitWidth, BitWidth / 2);
EVT HalfVT = EVT::getIntegerVT(*TLO.DAG.getContext(), BitWidth / 2);
if (isNarrowingProfitable(VT, HalfVT) &&
isTypeDesirableForOp(ISD::SRL, HalfVT) &&
isTruncateFree(VT, HalfVT) && isZExtFree(HalfVT, VT) &&
(!TLO.LegalOperations() || isOperationLegal(ISD::SRL, HalfVT)) &&
((InDemandedMask.countLeadingZeros() >= (BitWidth / 2)) ||
TLO.DAG.MaskedValueIsZero(Op0, HiBits))) {
SDValue NewOp = TLO.DAG.getNode(ISD::TRUNCATE, dl, HalfVT, Op0);
SDValue NewShiftAmt =
TLO.DAG.getShiftAmountConstant(ShAmt, HalfVT, dl);
SDValue NewShift =
TLO.DAG.getNode(ISD::SRL, dl, HalfVT, NewOp, NewShiftAmt);
return TLO.CombineTo(
Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, VT, NewShift));
}
}
// Compute the new bits that are at the top now.
if (SimplifyDemandedBits(Op0, InDemandedMask, DemandedElts, Known, TLO,
Depth + 1))
return true;
Known.Zero.lshrInPlace(ShAmt);
Known.One.lshrInPlace(ShAmt);
// High bits known zero.
Known.Zero.setHighBits(ShAmt);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!InDemandedMask.isAllOnes() || !DemandedElts.isAllOnes()) {
SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits(
Op0, InDemandedMask, DemandedElts, TLO.DAG, Depth + 1);
if (DemandedOp0) {
SDValue NewOp = TLO.DAG.getNode(ISD::SRL, dl, VT, DemandedOp0, Op1);
return TLO.CombineTo(Op, NewOp);
}
}
} else {
// Use generic knownbits computation as it has support for non-uniform
// shift amounts.
Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
}
// Try to match AVG patterns (after shift simplification).
if (SDValue AVG = combineShiftToAVG(Op, TLO, *this, DemandedBits,
DemandedElts, Depth + 1))
return TLO.CombineTo(Op, AVG);
break;
}
case ISD::SRA: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
EVT ShiftVT = Op1.getValueType();
// If we only want bits that already match the signbit then we don't need
// to shift.
unsigned NumHiDemandedBits = BitWidth - DemandedBits.countr_zero();
if (TLO.DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1) >=
NumHiDemandedBits)
return TLO.CombineTo(Op, Op0);
// If this is an arithmetic shift right and only the low-bit is set, we can
// always convert this into a logical shr, even if the shift amount is
// variable. The low bit of the shift cannot be an input sign bit unless
// the shift amount is >= the size of the datatype, which is undefined.
if (DemandedBits.isOne())
return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, Op1));
if (std::optional<uint64_t> KnownSA =
TLO.DAG.getValidShiftAmount(Op, DemandedElts, Depth + 1)) {
unsigned ShAmt = *KnownSA;
if (ShAmt == 0)
return TLO.CombineTo(Op, Op0);
// fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target
// supports sext_inreg.
if (Op0.getOpcode() == ISD::SHL) {
if (std::optional<uint64_t> InnerSA =
TLO.DAG.getValidShiftAmount(Op0, DemandedElts, Depth + 2)) {
unsigned LowBits = BitWidth - ShAmt;
EVT ExtVT = EVT::getIntegerVT(*TLO.DAG.getContext(), LowBits);
if (VT.isVector())
ExtVT = EVT::getVectorVT(*TLO.DAG.getContext(), ExtVT,
VT.getVectorElementCount());
if (*InnerSA == ShAmt) {
if (!TLO.LegalOperations() ||
getOperationAction(ISD::SIGN_EXTEND_INREG, ExtVT) == Legal)
return TLO.CombineTo(
Op, TLO.DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, VT,
Op0.getOperand(0),
TLO.DAG.getValueType(ExtVT)));
// Even if we can't convert to sext_inreg, we might be able to
// remove this shift pair if the input is already sign extended.
unsigned NumSignBits =
TLO.DAG.ComputeNumSignBits(Op0.getOperand(0), DemandedElts);
if (NumSignBits > ShAmt)
return TLO.CombineTo(Op, Op0.getOperand(0));
}
}
}
APInt InDemandedMask = (DemandedBits << ShAmt);
// If the shift is exact, then it does demand the low bits (and knows that
// they are zero).
if (Op->getFlags().hasExact())
InDemandedMask.setLowBits(ShAmt);
// If any of the demanded bits are produced by the sign extension, we also
// demand the input sign bit.
if (DemandedBits.countl_zero() < ShAmt)
InDemandedMask.setSignBit();
if (SimplifyDemandedBits(Op0, InDemandedMask, DemandedElts, Known, TLO,
Depth + 1))
return true;
Known.Zero.lshrInPlace(ShAmt);
Known.One.lshrInPlace(ShAmt);
// If the input sign bit is known to be zero, or if none of the top bits
// are demanded, turn this into an unsigned shift right.
if (Known.Zero[BitWidth - ShAmt - 1] ||
DemandedBits.countl_zero() >= ShAmt) {
SDNodeFlags Flags;
Flags.setExact(Op->getFlags().hasExact());
return TLO.CombineTo(
Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, Op1, Flags));
}
int Log2 = DemandedBits.exactLogBase2();
if (Log2 >= 0) {
// The bit must come from the sign.
SDValue NewSA = TLO.DAG.getConstant(BitWidth - 1 - Log2, dl, ShiftVT);
return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, NewSA));
}
if (Known.One[BitWidth - ShAmt - 1])
// New bits are known one.
Known.One.setHighBits(ShAmt);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!InDemandedMask.isAllOnes() || !DemandedElts.isAllOnes()) {
SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits(
Op0, InDemandedMask, DemandedElts, TLO.DAG, Depth + 1);
if (DemandedOp0) {
SDValue NewOp = TLO.DAG.getNode(ISD::SRA, dl, VT, DemandedOp0, Op1);
return TLO.CombineTo(Op, NewOp);
}
}
}
// Try to match AVG patterns (after shift simplification).
if (SDValue AVG = combineShiftToAVG(Op, TLO, *this, DemandedBits,
DemandedElts, Depth + 1))
return TLO.CombineTo(Op, AVG);
break;
}
case ISD::FSHL:
case ISD::FSHR: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
SDValue Op2 = Op.getOperand(2);
bool IsFSHL = (Op.getOpcode() == ISD::FSHL);
if (ConstantSDNode *SA = isConstOrConstSplat(Op2, DemandedElts)) {
unsigned Amt = SA->getAPIntValue().urem(BitWidth);
// For fshl, 0-shift returns the 1st arg.
// For fshr, 0-shift returns the 2nd arg.
if (Amt == 0) {
if (SimplifyDemandedBits(IsFSHL ? Op0 : Op1, DemandedBits, DemandedElts,
Known, TLO, Depth + 1))
return true;
break;
}
// fshl: (Op0 << Amt) | (Op1 >> (BW - Amt))
// fshr: (Op0 << (BW - Amt)) | (Op1 >> Amt)
APInt Demanded0 = DemandedBits.lshr(IsFSHL ? Amt : (BitWidth - Amt));
APInt Demanded1 = DemandedBits << (IsFSHL ? (BitWidth - Amt) : Amt);
if (SimplifyDemandedBits(Op0, Demanded0, DemandedElts, Known2, TLO,
Depth + 1))
return true;
if (SimplifyDemandedBits(Op1, Demanded1, DemandedElts, Known, TLO,
Depth + 1))
return true;
Known2.One <<= (IsFSHL ? Amt : (BitWidth - Amt));
Known2.Zero <<= (IsFSHL ? Amt : (BitWidth - Amt));
Known.One.lshrInPlace(IsFSHL ? (BitWidth - Amt) : Amt);
Known.Zero.lshrInPlace(IsFSHL ? (BitWidth - Amt) : Amt);
Known = Known.unionWith(Known2);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!Demanded0.isAllOnes() || !Demanded1.isAllOnes() ||
!DemandedElts.isAllOnes()) {
SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits(
Op0, Demanded0, DemandedElts, TLO.DAG, Depth + 1);
SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits(
Op1, Demanded1, DemandedElts, TLO.DAG, Depth + 1);
if (DemandedOp0 || DemandedOp1) {
DemandedOp0 = DemandedOp0 ? DemandedOp0 : Op0;
DemandedOp1 = DemandedOp1 ? DemandedOp1 : Op1;
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, DemandedOp0,
DemandedOp1, Op2);
return TLO.CombineTo(Op, NewOp);
}
}
}
// For pow-2 bitwidths we only demand the bottom modulo amt bits.
if (isPowerOf2_32(BitWidth)) {
APInt DemandedAmtBits(Op2.getScalarValueSizeInBits(), BitWidth - 1);
if (SimplifyDemandedBits(Op2, DemandedAmtBits, DemandedElts,
Known2, TLO, Depth + 1))
return true;
}
break;
}
case ISD::ROTL:
case ISD::ROTR: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
bool IsROTL = (Op.getOpcode() == ISD::ROTL);
// If we're rotating an 0/-1 value, then it stays an 0/-1 value.
if (BitWidth == TLO.DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1))
return TLO.CombineTo(Op, Op0);
if (ConstantSDNode *SA = isConstOrConstSplat(Op1, DemandedElts)) {
unsigned Amt = SA->getAPIntValue().urem(BitWidth);
unsigned RevAmt = BitWidth - Amt;
// rotl: (Op0 << Amt) | (Op0 >> (BW - Amt))
// rotr: (Op0 << (BW - Amt)) | (Op0 >> Amt)
APInt Demanded0 = DemandedBits.rotr(IsROTL ? Amt : RevAmt);
if (SimplifyDemandedBits(Op0, Demanded0, DemandedElts, Known2, TLO,
Depth + 1))
return true;
// rot*(x, 0) --> x
if (Amt == 0)
return TLO.CombineTo(Op, Op0);
// See if we don't demand either half of the rotated bits.
if ((!TLO.LegalOperations() || isOperationLegal(ISD::SHL, VT)) &&
DemandedBits.countr_zero() >= (IsROTL ? Amt : RevAmt)) {
Op1 = TLO.DAG.getConstant(IsROTL ? Amt : RevAmt, dl, Op1.getValueType());
return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT, Op0, Op1));
}
if ((!TLO.LegalOperations() || isOperationLegal(ISD::SRL, VT)) &&
DemandedBits.countl_zero() >= (IsROTL ? RevAmt : Amt)) {
Op1 = TLO.DAG.getConstant(IsROTL ? RevAmt : Amt, dl, Op1.getValueType());
return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, Op1));
}
}
// For pow-2 bitwidths we only demand the bottom modulo amt bits.
if (isPowerOf2_32(BitWidth)) {
APInt DemandedAmtBits(Op1.getScalarValueSizeInBits(), BitWidth - 1);
if (SimplifyDemandedBits(Op1, DemandedAmtBits, DemandedElts, Known2, TLO,
Depth + 1))
return true;
}
break;
}
case ISD::SMIN:
case ISD::SMAX:
case ISD::UMIN:
case ISD::UMAX: {
unsigned Opc = Op.getOpcode();
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
// If we're only demanding signbits, then we can simplify to OR/AND node.
unsigned BitOp =
(Opc == ISD::SMIN || Opc == ISD::UMAX) ? ISD::OR : ISD::AND;
unsigned NumSignBits =
std::min(TLO.DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1),
TLO.DAG.ComputeNumSignBits(Op1, DemandedElts, Depth + 1));
unsigned NumDemandedUpperBits = BitWidth - DemandedBits.countr_zero();
if (NumSignBits >= NumDemandedUpperBits)
return TLO.CombineTo(Op, TLO.DAG.getNode(BitOp, SDLoc(Op), VT, Op0, Op1));
// Check if one arg is always less/greater than (or equal) to the other arg.
KnownBits Known0 = TLO.DAG.computeKnownBits(Op0, DemandedElts, Depth + 1);
KnownBits Known1 = TLO.DAG.computeKnownBits(Op1, DemandedElts, Depth + 1);
switch (Opc) {
case ISD::SMIN:
if (std::optional<bool> IsSLE = KnownBits::sle(Known0, Known1))
return TLO.CombineTo(Op, *IsSLE ? Op0 : Op1);
if (std::optional<bool> IsSLT = KnownBits::slt(Known0, Known1))
return TLO.CombineTo(Op, *IsSLT ? Op0 : Op1);
Known = KnownBits::smin(Known0, Known1);
break;
case ISD::SMAX:
if (std::optional<bool> IsSGE = KnownBits::sge(Known0, Known1))
return TLO.CombineTo(Op, *IsSGE ? Op0 : Op1);
if (std::optional<bool> IsSGT = KnownBits::sgt(Known0, Known1))
return TLO.CombineTo(Op, *IsSGT ? Op0 : Op1);
Known = KnownBits::smax(Known0, Known1);
break;
case ISD::UMIN:
if (std::optional<bool> IsULE = KnownBits::ule(Known0, Known1))
return TLO.CombineTo(Op, *IsULE ? Op0 : Op1);
if (std::optional<bool> IsULT = KnownBits::ult(Known0, Known1))
return TLO.CombineTo(Op, *IsULT ? Op0 : Op1);
Known = KnownBits::umin(Known0, Known1);
break;
case ISD::UMAX:
if (std::optional<bool> IsUGE = KnownBits::uge(Known0, Known1))
return TLO.CombineTo(Op, *IsUGE ? Op0 : Op1);
if (std::optional<bool> IsUGT = KnownBits::ugt(Known0, Known1))
return TLO.CombineTo(Op, *IsUGT ? Op0 : Op1);
Known = KnownBits::umax(Known0, Known1);
break;
}
break;
}
case ISD::BITREVERSE: {
SDValue Src = Op.getOperand(0);
APInt DemandedSrcBits = DemandedBits.reverseBits();
if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedElts, Known2, TLO,
Depth + 1))
return true;
Known.One = Known2.One.reverseBits();
Known.Zero = Known2.Zero.reverseBits();
break;
}
case ISD::BSWAP: {
SDValue Src = Op.getOperand(0);
// If the only bits demanded come from one byte of the bswap result,
// just shift the input byte into position to eliminate the bswap.
unsigned NLZ = DemandedBits.countl_zero();
unsigned NTZ = DemandedBits.countr_zero();
// Round NTZ down to the next byte. If we have 11 trailing zeros, then
// we need all the bits down to bit 8. Likewise, round NLZ. If we
// have 14 leading zeros, round to 8.
NLZ = alignDown(NLZ, 8);
NTZ = alignDown(NTZ, 8);
// If we need exactly one byte, we can do this transformation.
if (BitWidth - NLZ - NTZ == 8) {
// Replace this with either a left or right shift to get the byte into
// the right place.
unsigned ShiftOpcode = NLZ > NTZ ? ISD::SRL : ISD::SHL;
if (!TLO.LegalOperations() || isOperationLegal(ShiftOpcode, VT)) {
unsigned ShiftAmount = NLZ > NTZ ? NLZ - NTZ : NTZ - NLZ;
SDValue ShAmt = TLO.DAG.getShiftAmountConstant(ShiftAmount, VT, dl);
SDValue NewOp = TLO.DAG.getNode(ShiftOpcode, dl, VT, Src, ShAmt);
return TLO.CombineTo(Op, NewOp);
}
}
APInt DemandedSrcBits = DemandedBits.byteSwap();
if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedElts, Known2, TLO,
Depth + 1))
return true;
Known.One = Known2.One.byteSwap();
Known.Zero = Known2.Zero.byteSwap();
break;
}
case ISD::CTPOP: {
// If only 1 bit is demanded, replace with PARITY as long as we're before
// op legalization.
// FIXME: Limit to scalars for now.
if (DemandedBits.isOne() && !TLO.LegalOps && !VT.isVector())
return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::PARITY, dl, VT,
Op.getOperand(0)));
Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
break;
}
case ISD::SIGN_EXTEND_INREG: {
SDValue Op0 = Op.getOperand(0);
EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
unsigned ExVTBits = ExVT.getScalarSizeInBits();
// If we only care about the highest bit, don't bother shifting right.
if (DemandedBits.isSignMask()) {
unsigned MinSignedBits =
TLO.DAG.ComputeMaxSignificantBits(Op0, DemandedElts, Depth + 1);
bool AlreadySignExtended = ExVTBits >= MinSignedBits;
// However if the input is already sign extended we expect the sign
// extension to be dropped altogether later and do not simplify.
if (!AlreadySignExtended) {
// Compute the correct shift amount type, which must be getShiftAmountTy
// for scalar types after legalization.
SDValue ShiftAmt =
TLO.DAG.getShiftAmountConstant(BitWidth - ExVTBits, VT, dl);
return TLO.CombineTo(Op,
TLO.DAG.getNode(ISD::SHL, dl, VT, Op0, ShiftAmt));
}
}
// If none of the extended bits are demanded, eliminate the sextinreg.
if (DemandedBits.getActiveBits() <= ExVTBits)
return TLO.CombineTo(Op, Op0);
APInt InputDemandedBits = DemandedBits.getLoBits(ExVTBits);
// Since the sign extended bits are demanded, we know that the sign
// bit is demanded.
InputDemandedBits.setBit(ExVTBits - 1);
if (SimplifyDemandedBits(Op0, InputDemandedBits, DemandedElts, Known, TLO,
Depth + 1))
return true;
// If the sign bit of the input is known set or clear, then we know the
// top bits of the result.
// If the input sign bit is known zero, convert this into a zero extension.
if (Known.Zero[ExVTBits - 1])
return TLO.CombineTo(Op, TLO.DAG.getZeroExtendInReg(Op0, dl, ExVT));
APInt Mask = APInt::getLowBitsSet(BitWidth, ExVTBits);
if (Known.One[ExVTBits - 1]) { // Input sign bit known set
Known.One.setBitsFrom(ExVTBits);
Known.Zero &= Mask;
} else { // Input sign bit unknown
Known.Zero &= Mask;
Known.One &= Mask;
}
break;
}
case ISD::BUILD_PAIR: {
EVT HalfVT = Op.getOperand(0).getValueType();
unsigned HalfBitWidth = HalfVT.getScalarSizeInBits();
APInt MaskLo = DemandedBits.getLoBits(HalfBitWidth).trunc(HalfBitWidth);
APInt MaskHi = DemandedBits.getHiBits(HalfBitWidth).trunc(HalfBitWidth);
KnownBits KnownLo, KnownHi;
if (SimplifyDemandedBits(Op.getOperand(0), MaskLo, KnownLo, TLO, Depth + 1))
return true;
if (SimplifyDemandedBits(Op.getOperand(1), MaskHi, KnownHi, TLO, Depth + 1))
return true;
Known = KnownHi.concat(KnownLo);
break;
}
case ISD::ZERO_EXTEND_VECTOR_INREG:
if (VT.isScalableVector())
return false;
[[fallthrough]];
case ISD::ZERO_EXTEND: {
SDValue Src = Op.getOperand(0);
EVT SrcVT = Src.getValueType();
unsigned InBits = SrcVT.getScalarSizeInBits();
unsigned InElts = SrcVT.isFixedLengthVector() ? SrcVT.getVectorNumElements() : 1;
bool IsVecInReg = Op.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG;
// If none of the top bits are demanded, convert this into an any_extend.
if (DemandedBits.getActiveBits() <= InBits) {
// If we only need the non-extended bits of the bottom element
// then we can just bitcast to the result.
if (IsLE && IsVecInReg && DemandedElts == 1 &&
VT.getSizeInBits() == SrcVT.getSizeInBits())
return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src));
unsigned Opc =
IsVecInReg ? ISD::ANY_EXTEND_VECTOR_INREG : ISD::ANY_EXTEND;
if (!TLO.LegalOperations() || isOperationLegal(Opc, VT))
return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src));
}
SDNodeFlags Flags = Op->getFlags();
APInt InDemandedBits = DemandedBits.trunc(InBits);
APInt InDemandedElts = DemandedElts.zext(InElts);
if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO,
Depth + 1)) {
if (Flags.hasNonNeg()) {
Flags.setNonNeg(false);
Op->setFlags(Flags);
}
return true;
}
assert(Known.getBitWidth() == InBits && "Src width has changed?");
Known = Known.zext(BitWidth);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (SDValue NewSrc = SimplifyMultipleUseDemandedBits(
Src, InDemandedBits, InDemandedElts, TLO.DAG, Depth + 1))
return TLO.CombineTo(Op, TLO.DAG.getNode(Op.getOpcode(), dl, VT, NewSrc));
break;
}
case ISD::SIGN_EXTEND_VECTOR_INREG:
if (VT.isScalableVector())
return false;
[[fallthrough]];
case ISD::SIGN_EXTEND: {
SDValue Src = Op.getOperand(0);
EVT SrcVT = Src.getValueType();
unsigned InBits = SrcVT.getScalarSizeInBits();
unsigned InElts = SrcVT.isFixedLengthVector() ? SrcVT.getVectorNumElements() : 1;
bool IsVecInReg = Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG;
APInt InDemandedElts = DemandedElts.zext(InElts);
APInt InDemandedBits = DemandedBits.trunc(InBits);
// Since some of the sign extended bits are demanded, we know that the sign
// bit is demanded.
InDemandedBits.setBit(InBits - 1);
// If none of the top bits are demanded, convert this into an any_extend.
if (DemandedBits.getActiveBits() <= InBits) {
// If we only need the non-extended bits of the bottom element
// then we can just bitcast to the result.
if (IsLE && IsVecInReg && DemandedElts == 1 &&
VT.getSizeInBits() == SrcVT.getSizeInBits())
return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src));
// Don't lose an all signbits 0/-1 splat on targets with 0/-1 booleans.
if (getBooleanContents(VT) != ZeroOrNegativeOneBooleanContent ||
TLO.DAG.ComputeNumSignBits(Src, InDemandedElts, Depth + 1) !=
InBits) {
unsigned Opc =
IsVecInReg ? ISD::ANY_EXTEND_VECTOR_INREG : ISD::ANY_EXTEND;
if (!TLO.LegalOperations() || isOperationLegal(Opc, VT))
return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src));
}
}
if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO,
Depth + 1))
return true;
assert(Known.getBitWidth() == InBits && "Src width has changed?");
// If the sign bit is known one, the top bits match.
Known = Known.sext(BitWidth);
// If the sign bit is known zero, convert this to a zero extend.
if (Known.isNonNegative()) {
unsigned Opc =
IsVecInReg ? ISD::ZERO_EXTEND_VECTOR_INREG : ISD::ZERO_EXTEND;
if (!TLO.LegalOperations() || isOperationLegal(Opc, VT)) {
SDNodeFlags Flags;
if (!IsVecInReg)
Flags.setNonNeg(true);
return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src, Flags));
}
}
// Attempt to avoid multi-use ops if we don't need anything from them.
if (SDValue NewSrc = SimplifyMultipleUseDemandedBits(
Src, InDemandedBits, InDemandedElts, TLO.DAG, Depth + 1))
return TLO.CombineTo(Op, TLO.DAG.getNode(Op.getOpcode(), dl, VT, NewSrc));
break;
}
case ISD::ANY_EXTEND_VECTOR_INREG:
if (VT.isScalableVector())
return false;
[[fallthrough]];
case ISD::ANY_EXTEND: {
SDValue Src = Op.getOperand(0);
EVT SrcVT = Src.getValueType();
unsigned InBits = SrcVT.getScalarSizeInBits();
unsigned InElts = SrcVT.isFixedLengthVector() ? SrcVT.getVectorNumElements() : 1;
bool IsVecInReg = Op.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG;
// If we only need the bottom element then we can just bitcast.
// TODO: Handle ANY_EXTEND?
if (IsLE && IsVecInReg && DemandedElts == 1 &&
VT.getSizeInBits() == SrcVT.getSizeInBits())
return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src));
APInt InDemandedBits = DemandedBits.trunc(InBits);
APInt InDemandedElts = DemandedElts.zext(InElts);
if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO,
Depth + 1))
return true;
assert(Known.getBitWidth() == InBits && "Src width has changed?");
Known = Known.anyext(BitWidth);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (SDValue NewSrc = SimplifyMultipleUseDemandedBits(
Src, InDemandedBits, InDemandedElts, TLO.DAG, Depth + 1))
return TLO.CombineTo(Op, TLO.DAG.getNode(Op.getOpcode(), dl, VT, NewSrc));
break;
}
case ISD::TRUNCATE: {
SDValue Src = Op.getOperand(0);
// Simplify the input, using demanded bit information, and compute the known
// zero/one bits live out.
unsigned OperandBitWidth = Src.getScalarValueSizeInBits();
APInt TruncMask = DemandedBits.zext(OperandBitWidth);
if (SimplifyDemandedBits(Src, TruncMask, DemandedElts, Known, TLO,
Depth + 1))
return true;
Known = Known.trunc(BitWidth);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (SDValue NewSrc = SimplifyMultipleUseDemandedBits(
Src, TruncMask, DemandedElts, TLO.DAG, Depth + 1))
return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::TRUNCATE, dl, VT, NewSrc));
// If the input is only used by this truncate, see if we can shrink it based
// on the known demanded bits.
switch (Src.getOpcode()) {
default:
break;
case ISD::SRL:
// Shrink SRL by a constant if none of the high bits shifted in are
// demanded.
if (TLO.LegalTypes() && !isTypeDesirableForOp(ISD::SRL, VT))
// Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
// undesirable.
break;
if (Src.getNode()->hasOneUse()) {
if (isTruncateFree(Src, VT) &&
!isTruncateFree(Src.getValueType(), VT)) {
// If truncate is only free at trunc(srl), do not turn it into
// srl(trunc). The check is done by first check the truncate is free
// at Src's opcode(srl), then check the truncate is not done by
// referencing sub-register. In test, if both trunc(srl) and
// srl(trunc)'s trunc are free, srl(trunc) performs better. If only
// trunc(srl)'s trunc is free, trunc(srl) is better.
break;
}
std::optional<uint64_t> ShAmtC =
TLO.DAG.getValidShiftAmount(Src, DemandedElts, Depth + 2);
if (!ShAmtC || *ShAmtC >= BitWidth)
break;
uint64_t ShVal = *ShAmtC;
APInt HighBits =
APInt::getHighBitsSet(OperandBitWidth, OperandBitWidth - BitWidth);
HighBits.lshrInPlace(ShVal);
HighBits = HighBits.trunc(BitWidth);
if (!(HighBits & DemandedBits)) {
// None of the shifted in bits are needed. Add a truncate of the
// shift input, then shift it.
SDValue NewShAmt = TLO.DAG.getShiftAmountConstant(ShVal, VT, dl);
SDValue NewTrunc =
TLO.DAG.getNode(ISD::TRUNCATE, dl, VT, Src.getOperand(0));
return TLO.CombineTo(
Op, TLO.DAG.getNode(ISD::SRL, dl, VT, NewTrunc, NewShAmt));
}
}
break;
}
break;
}
case ISD::AssertZext: {
// AssertZext demands all of the high bits, plus any of the low bits
// demanded by its users.
EVT ZVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
APInt InMask = APInt::getLowBitsSet(BitWidth, ZVT.getSizeInBits());
if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | DemandedBits, Known,
TLO, Depth + 1))
return true;
Known.Zero |= ~InMask;
Known.One &= (~Known.Zero);
break;
}
case ISD::EXTRACT_VECTOR_ELT: {
SDValue Src = Op.getOperand(0);
SDValue Idx = Op.getOperand(1);
ElementCount SrcEltCnt = Src.getValueType().getVectorElementCount();
unsigned EltBitWidth = Src.getScalarValueSizeInBits();
if (SrcEltCnt.isScalable())
return false;
// Demand the bits from every vector element without a constant index.
unsigned NumSrcElts = SrcEltCnt.getFixedValue();
APInt DemandedSrcElts = APInt::getAllOnes(NumSrcElts);
if (auto *CIdx = dyn_cast<ConstantSDNode>(Idx))
if (CIdx->getAPIntValue().ult(NumSrcElts))
DemandedSrcElts = APInt::getOneBitSet(NumSrcElts, CIdx->getZExtValue());
// If BitWidth > EltBitWidth the value is anyext:ed. So we do not know
// anything about the extended bits.
APInt DemandedSrcBits = DemandedBits;
if (BitWidth > EltBitWidth)
DemandedSrcBits = DemandedSrcBits.trunc(EltBitWidth);
if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts, Known2, TLO,
Depth + 1))
return true;
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!DemandedSrcBits.isAllOnes() || !DemandedSrcElts.isAllOnes()) {
if (SDValue DemandedSrc = SimplifyMultipleUseDemandedBits(
Src, DemandedSrcBits, DemandedSrcElts, TLO.DAG, Depth + 1)) {
SDValue NewOp =
TLO.DAG.getNode(Op.getOpcode(), dl, VT, DemandedSrc, Idx);
return TLO.CombineTo(Op, NewOp);
}
}
Known = Known2;
if (BitWidth > EltBitWidth)
Known = Known.anyext(BitWidth);
break;
}
case ISD::BITCAST: {
if (VT.isScalableVector())
return false;
SDValue Src = Op.getOperand(0);
EVT SrcVT = Src.getValueType();
unsigned NumSrcEltBits = SrcVT.getScalarSizeInBits();
// If this is an FP->Int bitcast and if the sign bit is the only
// thing demanded, turn this into a FGETSIGN.
if (!TLO.LegalOperations() && !VT.isVector() && !SrcVT.isVector() &&
DemandedBits == APInt::getSignMask(Op.getValueSizeInBits()) &&
SrcVT.isFloatingPoint()) {
bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, VT);
bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32);
if ((OpVTLegal || i32Legal) && VT.isSimple() && SrcVT != MVT::f16 &&
SrcVT != MVT::f128) {
// Cannot eliminate/lower SHL for f128 yet.
EVT Ty = OpVTLegal ? VT : MVT::i32;
// Make a FGETSIGN + SHL to move the sign bit into the appropriate
// place. We expect the SHL to be eliminated by other optimizations.
SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Src);
unsigned OpVTSizeInBits = Op.getValueSizeInBits();
if (!OpVTLegal && OpVTSizeInBits > 32)
Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Sign);
unsigned ShVal = Op.getValueSizeInBits() - 1;
SDValue ShAmt = TLO.DAG.getConstant(ShVal, dl, VT);
return TLO.CombineTo(Op,
TLO.DAG.getNode(ISD::SHL, dl, VT, Sign, ShAmt));
}
}
// Bitcast from a vector using SimplifyDemanded Bits/VectorElts.
// Demand the elt/bit if any of the original elts/bits are demanded.
if (SrcVT.isVector() && (BitWidth % NumSrcEltBits) == 0) {
unsigned Scale = BitWidth / NumSrcEltBits;
unsigned NumSrcElts = SrcVT.getVectorNumElements();
APInt DemandedSrcBits = APInt::getZero(NumSrcEltBits);
APInt DemandedSrcElts = APInt::getZero(NumSrcElts);
for (unsigned i = 0; i != Scale; ++i) {
unsigned EltOffset = IsLE ? i : (Scale - 1 - i);
unsigned BitOffset = EltOffset * NumSrcEltBits;
APInt Sub = DemandedBits.extractBits(NumSrcEltBits, BitOffset);
if (!Sub.isZero()) {
DemandedSrcBits |= Sub;
for (unsigned j = 0; j != NumElts; ++j)
if (DemandedElts[j])
DemandedSrcElts.setBit((j * Scale) + i);
}
}
APInt KnownSrcUndef, KnownSrcZero;
if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownSrcUndef,
KnownSrcZero, TLO, Depth + 1))
return true;
KnownBits KnownSrcBits;
if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts,
KnownSrcBits, TLO, Depth + 1))
return true;
} else if (IsLE && (NumSrcEltBits % BitWidth) == 0) {
// TODO - bigendian once we have test coverage.
unsigned Scale = NumSrcEltBits / BitWidth;
unsigned NumSrcElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1;
APInt DemandedSrcBits = APInt::getZero(NumSrcEltBits);
APInt DemandedSrcElts = APInt::getZero(NumSrcElts);
for (unsigned i = 0; i != NumElts; ++i)
if (DemandedElts[i]) {
unsigned Offset = (i % Scale) * BitWidth;
DemandedSrcBits.insertBits(DemandedBits, Offset);
DemandedSrcElts.setBit(i / Scale);
}
if (SrcVT.isVector()) {
APInt KnownSrcUndef, KnownSrcZero;
if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownSrcUndef,
KnownSrcZero, TLO, Depth + 1))
return true;
}
KnownBits KnownSrcBits;
if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts,
KnownSrcBits, TLO, Depth + 1))
return true;
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!DemandedSrcBits.isAllOnes() || !DemandedSrcElts.isAllOnes()) {
if (SDValue DemandedSrc = SimplifyMultipleUseDemandedBits(
Src, DemandedSrcBits, DemandedSrcElts, TLO.DAG, Depth + 1)) {
SDValue NewOp = TLO.DAG.getBitcast(VT, DemandedSrc);
return TLO.CombineTo(Op, NewOp);
}
}
}
// If this is a bitcast, let computeKnownBits handle it. Only do this on a
// recursive call where Known may be useful to the caller.
if (Depth > 0) {
Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
return false;
}
break;
}
case ISD::MUL:
if (DemandedBits.isPowerOf2()) {
// The LSB of X*Y is set only if (X & 1) == 1 and (Y & 1) == 1.
// If we demand exactly one bit N and we have "X * (C' << N)" where C' is
// odd (has LSB set), then the left-shifted low bit of X is the answer.
unsigned CTZ = DemandedBits.countr_zero();
ConstantSDNode *C = isConstOrConstSplat(Op.getOperand(1), DemandedElts);
if (C && C->getAPIntValue().countr_zero() == CTZ) {
SDValue AmtC = TLO.DAG.getShiftAmountConstant(CTZ, VT, dl);
SDValue Shl = TLO.DAG.getNode(ISD::SHL, dl, VT, Op.getOperand(0), AmtC);
return TLO.CombineTo(Op, Shl);
}
}
// For a squared value "X * X", the bottom 2 bits are 0 and X[0] because:
// X * X is odd iff X is odd.
// 'Quadratic Reciprocity': X * X -> 0 for bit[1]
if (Op.getOperand(0) == Op.getOperand(1) && DemandedBits.ult(4)) {
SDValue One = TLO.DAG.getConstant(1, dl, VT);
SDValue And1 = TLO.DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), One);
return TLO.CombineTo(Op, And1);
}
[[fallthrough]];
case ISD::ADD:
case ISD::SUB: {
// Add, Sub, and Mul don't demand any bits in positions beyond that
// of the highest bit demanded of them.
SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
SDNodeFlags Flags = Op.getNode()->getFlags();
unsigned DemandedBitsLZ = DemandedBits.countl_zero();
APInt LoMask = APInt::getLowBitsSet(BitWidth, BitWidth - DemandedBitsLZ);
KnownBits KnownOp0, KnownOp1;
auto GetDemandedBitsLHSMask = [&](APInt Demanded,
const KnownBits &KnownRHS) {
if (Op.getOpcode() == ISD::MUL)
Demanded.clearHighBits(KnownRHS.countMinTrailingZeros());
return Demanded;
};
if (SimplifyDemandedBits(Op1, LoMask, DemandedElts, KnownOp1, TLO,
Depth + 1) ||
SimplifyDemandedBits(Op0, GetDemandedBitsLHSMask(LoMask, KnownOp1),
DemandedElts, KnownOp0, TLO, Depth + 1) ||
// See if the operation should be performed at a smaller bit width.
ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) {
if (Flags.hasNoSignedWrap() || Flags.hasNoUnsignedWrap()) {
// Disable the nsw and nuw flags. We can no longer guarantee that we
// won't wrap after simplification.
Flags.setNoSignedWrap(false);
Flags.setNoUnsignedWrap(false);
Op->setFlags(Flags);
}
return true;
}
// neg x with only low bit demanded is simply x.
if (Op.getOpcode() == ISD::SUB && DemandedBits.isOne() &&
isNullConstant(Op0))
return TLO.CombineTo(Op, Op1);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!LoMask.isAllOnes() || !DemandedElts.isAllOnes()) {
SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits(
Op0, LoMask, DemandedElts, TLO.DAG, Depth + 1);
SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits(
Op1, LoMask, DemandedElts, TLO.DAG, Depth + 1);
if (DemandedOp0 || DemandedOp1) {
Flags.setNoSignedWrap(false);
Flags.setNoUnsignedWrap(false);
Op0 = DemandedOp0 ? DemandedOp0 : Op0;
Op1 = DemandedOp1 ? DemandedOp1 : Op1;
SDValue NewOp =
TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1, Flags);
return TLO.CombineTo(Op, NewOp);
}
}
// If we have a constant operand, we may be able to turn it into -1 if we
// do not demand the high bits. This can make the constant smaller to
// encode, allow more general folding, or match specialized instruction
// patterns (eg, 'blsr' on x86). Don't bother changing 1 to -1 because that
// is probably not useful (and could be detrimental).
ConstantSDNode *C = isConstOrConstSplat(Op1);
APInt HighMask = APInt::getHighBitsSet(BitWidth, DemandedBitsLZ);
if (C && !C->isAllOnes() && !C->isOne() &&
(C->getAPIntValue() | HighMask).isAllOnes()) {
SDValue Neg1 = TLO.DAG.getAllOnesConstant(dl, VT);
// Disable the nsw and nuw flags. We can no longer guarantee that we
// won't wrap after simplification.
Flags.setNoSignedWrap(false);
Flags.setNoUnsignedWrap(false);
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Neg1, Flags);
return TLO.CombineTo(Op, NewOp);
}
// Match a multiply with a disguised negated-power-of-2 and convert to a
// an equivalent shift-left amount.
// Example: (X * MulC) + Op1 --> Op1 - (X << log2(-MulC))
auto getShiftLeftAmt = [&HighMask](SDValue Mul) -> unsigned {
if (Mul.getOpcode() != ISD::MUL || !Mul.hasOneUse())
return 0;
// Don't touch opaque constants. Also, ignore zero and power-of-2
// multiplies. Those will get folded later.
ConstantSDNode *MulC = isConstOrConstSplat(Mul.getOperand(1));
if (MulC && !MulC->isOpaque() && !MulC->isZero() &&
!MulC->getAPIntValue().isPowerOf2()) {
APInt UnmaskedC = MulC->getAPIntValue() | HighMask;
if (UnmaskedC.isNegatedPowerOf2())
return (-UnmaskedC).logBase2();
}
return 0;
};
auto foldMul = [&](ISD::NodeType NT, SDValue X, SDValue Y,
unsigned ShlAmt) {
SDValue ShlAmtC = TLO.DAG.getShiftAmountConstant(ShlAmt, VT, dl);
SDValue Shl = TLO.DAG.getNode(ISD::SHL, dl, VT, X, ShlAmtC);
SDValue Res = TLO.DAG.getNode(NT, dl, VT, Y, Shl);
return TLO.CombineTo(Op, Res);
};
if (isOperationLegalOrCustom(ISD::SHL, VT)) {
if (Op.getOpcode() == ISD::ADD) {
// (X * MulC) + Op1 --> Op1 - (X << log2(-MulC))
if (unsigned ShAmt = getShiftLeftAmt(Op0))
return foldMul(ISD::SUB, Op0.getOperand(0), Op1, ShAmt);
// Op0 + (X * MulC) --> Op0 - (X << log2(-MulC))
if (unsigned ShAmt = getShiftLeftAmt(Op1))
return foldMul(ISD::SUB, Op1.getOperand(0), Op0, ShAmt);
}
if (Op.getOpcode() == ISD::SUB) {
// Op0 - (X * MulC) --> Op0 + (X << log2(-MulC))
if (unsigned ShAmt = getShiftLeftAmt(Op1))
return foldMul(ISD::ADD, Op1.getOperand(0), Op0, ShAmt);
}
}
if (Op.getOpcode() == ISD::MUL) {
Known = KnownBits::mul(KnownOp0, KnownOp1);
} else { // Op.getOpcode() is either ISD::ADD or ISD::SUB.
Known = KnownBits::computeForAddSub(
Op.getOpcode() == ISD::ADD, Flags.hasNoSignedWrap(),
Flags.hasNoUnsignedWrap(), KnownOp0, KnownOp1);
}
break;
}
default:
// We also ask the target about intrinsics (which could be specific to it).
if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
// TODO: Probably okay to remove after audit; here to reduce change size
// in initial enablement patch for scalable vectors
if (Op.getValueType().isScalableVector())
break;
if (SimplifyDemandedBitsForTargetNode(Op, DemandedBits, DemandedElts,
Known, TLO, Depth))
return true;
break;
}
// Just use computeKnownBits to compute output bits.
Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
break;
}
// If we know the value of all of the demanded bits, return this as a
// constant.
if (!isTargetCanonicalConstantNode(Op) &&
DemandedBits.isSubsetOf(Known.Zero | Known.One)) {
// Avoid folding to a constant if any OpaqueConstant is involved.
const SDNode *N = Op.getNode();
for (SDNode *Op :
llvm::make_range(SDNodeIterator::begin(N), SDNodeIterator::end(N))) {
if (auto *C = dyn_cast<ConstantSDNode>(Op))
if (C->isOpaque())
return false;
}
if (VT.isInteger())
return TLO.CombineTo(Op, TLO.DAG.getConstant(Known.One, dl, VT));
if (VT.isFloatingPoint())
return TLO.CombineTo(
Op,
TLO.DAG.getConstantFP(
APFloat(TLO.DAG.EVTToAPFloatSemantics(VT), Known.One), dl, VT));
}
// A multi use 'all demanded elts' simplify failed to find any knownbits.
// Try again just for the original demanded elts.
// Ensure we do this AFTER constant folding above.
if (HasMultiUse && Known.isUnknown() && !OriginalDemandedElts.isAllOnes())
Known = TLO.DAG.computeKnownBits(Op, OriginalDemandedElts, Depth);
return false;
}
bool TargetLowering::SimplifyDemandedVectorElts(SDValue Op,
const APInt &DemandedElts,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
!DCI.isBeforeLegalizeOps());
APInt KnownUndef, KnownZero;
bool Simplified =
SimplifyDemandedVectorElts(Op, DemandedElts, KnownUndef, KnownZero, TLO);
if (Simplified) {
DCI.AddToWorklist(Op.getNode());
DCI.CommitTargetLoweringOpt(TLO);
}
return Simplified;
}
/// Given a vector binary operation and known undefined elements for each input
/// operand, compute whether each element of the output is undefined.
static APInt getKnownUndefForVectorBinop(SDValue BO, SelectionDAG &DAG,
const APInt &UndefOp0,
const APInt &UndefOp1) {
EVT VT = BO.getValueType();
assert(DAG.getTargetLoweringInfo().isBinOp(BO.getOpcode()) && VT.isVector() &&
"Vector binop only");
EVT EltVT = VT.getVectorElementType();
unsigned NumElts = VT.isFixedLengthVector() ? VT.getVectorNumElements() : 1;
assert(UndefOp0.getBitWidth() == NumElts &&
UndefOp1.getBitWidth() == NumElts && "Bad type for undef analysis");
auto getUndefOrConstantElt = [&](SDValue V, unsigned Index,
const APInt &UndefVals) {
if (UndefVals[Index])
return DAG.getUNDEF(EltVT);
if (auto *BV = dyn_cast<BuildVectorSDNode>(V)) {
// Try hard to make sure that the getNode() call is not creating temporary
// nodes. Ignore opaque integers because they do not constant fold.
SDValue Elt = BV->getOperand(Index);
auto *C = dyn_cast<ConstantSDNode>(Elt);
if (isa<ConstantFPSDNode>(Elt) || Elt.isUndef() || (C && !C->isOpaque()))
return Elt;
}
return SDValue();
};
APInt KnownUndef = APInt::getZero(NumElts);
for (unsigned i = 0; i != NumElts; ++i) {
// If both inputs for this element are either constant or undef and match
// the element type, compute the constant/undef result for this element of
// the vector.
// TODO: Ideally we would use FoldConstantArithmetic() here, but that does
// not handle FP constants. The code within getNode() should be refactored
// to avoid the danger of creating a bogus temporary node here.
SDValue C0 = getUndefOrConstantElt(BO.getOperand(0), i, UndefOp0);
SDValue C1 = getUndefOrConstantElt(BO.getOperand(1), i, UndefOp1);
if (C0 && C1 && C0.getValueType() == EltVT && C1.getValueType() == EltVT)
if (DAG.getNode(BO.getOpcode(), SDLoc(BO), EltVT, C0, C1).isUndef())
KnownUndef.setBit(i);
}
return KnownUndef;
}
bool TargetLowering::SimplifyDemandedVectorElts(
SDValue Op, const APInt &OriginalDemandedElts, APInt &KnownUndef,
APInt &KnownZero, TargetLoweringOpt &TLO, unsigned Depth,
bool AssumeSingleUse) const {
EVT VT = Op.getValueType();
unsigned Opcode = Op.getOpcode();
APInt DemandedElts = OriginalDemandedElts;
unsigned NumElts = DemandedElts.getBitWidth();
assert(VT.isVector() && "Expected vector op");
KnownUndef = KnownZero = APInt::getZero(NumElts);
const TargetLowering &TLI = TLO.DAG.getTargetLoweringInfo();
if (!TLI.shouldSimplifyDemandedVectorElts(Op, TLO))
return false;
// TODO: For now we assume we know nothing about scalable vectors.
if (VT.isScalableVector())
return false;
assert(VT.getVectorNumElements() == NumElts &&
"Mask size mismatches value type element count!");
// Undef operand.
if (Op.isUndef()) {
KnownUndef.setAllBits();
return false;
}
// If Op has other users, assume that all elements are needed.
if (!AssumeSingleUse && !Op.getNode()->hasOneUse())
DemandedElts.setAllBits();
// Not demanding any elements from Op.
if (DemandedElts == 0) {
KnownUndef.setAllBits();
return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT));
}
// Limit search depth.
if (Depth >= SelectionDAG::MaxRecursionDepth)
return false;
SDLoc DL(Op);
unsigned EltSizeInBits = VT.getScalarSizeInBits();
bool IsLE = TLO.DAG.getDataLayout().isLittleEndian();
// Helper for demanding the specified elements and all the bits of both binary
// operands.
auto SimplifyDemandedVectorEltsBinOp = [&](SDValue Op0, SDValue Op1) {
SDValue NewOp0 = SimplifyMultipleUseDemandedVectorElts(Op0, DemandedElts,
TLO.DAG, Depth + 1);
SDValue NewOp1 = SimplifyMultipleUseDemandedVectorElts(Op1, DemandedElts,
TLO.DAG, Depth + 1);
if (NewOp0 || NewOp1) {
SDValue NewOp =
TLO.DAG.getNode(Opcode, SDLoc(Op), VT, NewOp0 ? NewOp0 : Op0,
NewOp1 ? NewOp1 : Op1, Op->getFlags());
return TLO.CombineTo(Op, NewOp);
}
return false;
};
switch (Opcode) {
case ISD::SCALAR_TO_VECTOR: {
if (!DemandedElts[0]) {
KnownUndef.setAllBits();
return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT));
}
SDValue ScalarSrc = Op.getOperand(0);
if (ScalarSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
SDValue Src = ScalarSrc.getOperand(0);
SDValue Idx = ScalarSrc.getOperand(1);
EVT SrcVT = Src.getValueType();
ElementCount SrcEltCnt = SrcVT.getVectorElementCount();
if (SrcEltCnt.isScalable())
return false;
unsigned NumSrcElts = SrcEltCnt.getFixedValue();
if (isNullConstant(Idx)) {
APInt SrcDemandedElts = APInt::getOneBitSet(NumSrcElts, 0);
APInt SrcUndef = KnownUndef.zextOrTrunc(NumSrcElts);
APInt SrcZero = KnownZero.zextOrTrunc(NumSrcElts);
if (SimplifyDemandedVectorElts(Src, SrcDemandedElts, SrcUndef, SrcZero,
TLO, Depth + 1))
return true;
}
}
KnownUndef.setHighBits(NumElts - 1);
break;
}
case ISD::BITCAST: {
SDValue Src = Op.getOperand(0);
EVT SrcVT = Src.getValueType();
// We only handle vectors here.
// TODO - investigate calling SimplifyDemandedBits/ComputeKnownBits?
if (!SrcVT.isVector())
break;
// Fast handling of 'identity' bitcasts.
unsigned NumSrcElts = SrcVT.getVectorNumElements();
if (NumSrcElts == NumElts)
return SimplifyDemandedVectorElts(Src, DemandedElts, KnownUndef,
KnownZero, TLO, Depth + 1);
APInt SrcDemandedElts, SrcZero, SrcUndef;
// Bitcast from 'large element' src vector to 'small element' vector, we
// must demand a source element if any DemandedElt maps to it.
if ((NumElts % NumSrcElts) == 0) {
unsigned Scale = NumElts / NumSrcElts;
SrcDemandedElts = APIntOps::ScaleBitMask(DemandedElts, NumSrcElts);
if (SimplifyDemandedVectorElts(Src, SrcDemandedElts, SrcUndef, SrcZero,
TLO, Depth + 1))
return true;
// Try calling SimplifyDemandedBits, converting demanded elts to the bits
// of the large element.
// TODO - bigendian once we have test coverage.
if (IsLE) {
unsigned SrcEltSizeInBits = SrcVT.getScalarSizeInBits();
APInt SrcDemandedBits = APInt::getZero(SrcEltSizeInBits);
for (unsigned i = 0; i != NumElts; ++i)
if (DemandedElts[i]) {
unsigned Ofs = (i % Scale) * EltSizeInBits;
SrcDemandedBits.setBits(Ofs, Ofs + EltSizeInBits);
}
KnownBits Known;
if (SimplifyDemandedBits(Src, SrcDemandedBits, SrcDemandedElts, Known,
TLO, Depth + 1))
return true;
// The bitcast has split each wide element into a number of
// narrow subelements. We have just computed the Known bits
// for wide elements. See if element splitting results in
// some subelements being zero. Only for demanded elements!
for (unsigned SubElt = 0; SubElt != Scale; ++SubElt) {
if (!Known.Zero.extractBits(EltSizeInBits, SubElt * EltSizeInBits)
.isAllOnes())
continue;
for (unsigned SrcElt = 0; SrcElt != NumSrcElts; ++SrcElt) {
unsigned Elt = Scale * SrcElt + SubElt;
if (DemandedElts[Elt])
KnownZero.setBit(Elt);
}
}
}
// If the src element is zero/undef then all the output elements will be -
// only demanded elements are guaranteed to be correct.
for (unsigned i = 0; i != NumSrcElts; ++i) {
if (SrcDemandedElts[i]) {
if (SrcZero[i])
KnownZero.setBits(i * Scale, (i + 1) * Scale);
if (SrcUndef[i])
KnownUndef.setBits(i * Scale, (i + 1) * Scale);
}
}
}
// Bitcast from 'small element' src vector to 'large element' vector, we
// demand all smaller source elements covered by the larger demanded element
// of this vector.
if ((NumSrcElts % NumElts) == 0) {
unsigned Scale = NumSrcElts / NumElts;
SrcDemandedElts = APIntOps::ScaleBitMask(DemandedElts, NumSrcElts);
if (SimplifyDemandedVectorElts(Src, SrcDemandedElts, SrcUndef, SrcZero,
TLO, Depth + 1))
return true;
// If all the src elements covering an output element are zero/undef, then
// the output element will be as well, assuming it was demanded.
for (unsigned i = 0; i != NumElts; ++i) {
if (DemandedElts[i]) {
if (SrcZero.extractBits(Scale, i * Scale).isAllOnes())
KnownZero.setBit(i);
if (SrcUndef.extractBits(Scale, i * Scale).isAllOnes())
KnownUndef.setBit(i);
}
}
}
break;
}
case ISD::FREEZE: {
SDValue N0 = Op.getOperand(0);
if (TLO.DAG.isGuaranteedNotToBeUndefOrPoison(N0, DemandedElts,
/*PoisonOnly=*/false))
return TLO.CombineTo(Op, N0);
// TODO: Replace this with the general fold from DAGCombiner::visitFREEZE
// freeze(op(x, ...)) -> op(freeze(x), ...).
if (N0.getOpcode() == ISD::SCALAR_TO_VECTOR && DemandedElts == 1)
return TLO.CombineTo(
Op, TLO.DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT,
TLO.DAG.getFreeze(N0.getOperand(0))));
break;
}
case ISD::BUILD_VECTOR: {
// Check all elements and simplify any unused elements with UNDEF.
if (!DemandedElts.isAllOnes()) {
// Don't simplify BROADCASTS.
if (llvm::any_of(Op->op_values(),
[&](SDValue Elt) { return Op.getOperand(0) != Elt; })) {
SmallVector<SDValue, 32> Ops(Op->op_begin(), Op->op_end());
bool Updated = false;
for (unsigned i = 0; i != NumElts; ++i) {
if (!DemandedElts[i] && !Ops[i].isUndef()) {
Ops[i] = TLO.DAG.getUNDEF(Ops[0].getValueType());
KnownUndef.setBit(i);
Updated = true;
}
}
if (Updated)
return TLO.CombineTo(Op, TLO.DAG.getBuildVector(VT, DL, Ops));
}
}
for (unsigned i = 0; i != NumElts; ++i) {
SDValue SrcOp = Op.getOperand(i);
if (SrcOp.isUndef()) {
KnownUndef.setBit(i);
} else if (EltSizeInBits == SrcOp.getScalarValueSizeInBits() &&
(isNullConstant(SrcOp) || isNullFPConstant(SrcOp))) {
KnownZero.setBit(i);
}
}
break;
}
case ISD::CONCAT_VECTORS: {
EVT SubVT = Op.getOperand(0).getValueType();
unsigned NumSubVecs = Op.getNumOperands();
unsigned NumSubElts = SubVT.getVectorNumElements();
for (unsigned i = 0; i != NumSubVecs; ++i) {
SDValue SubOp = Op.getOperand(i);
APInt SubElts = DemandedElts.extractBits(NumSubElts, i * NumSubElts);
APInt SubUndef, SubZero;
if (SimplifyDemandedVectorElts(SubOp, SubElts, SubUndef, SubZero, TLO,
Depth + 1))
return true;
KnownUndef.insertBits(SubUndef, i * NumSubElts);
KnownZero.insertBits(SubZero, i * NumSubElts);
}
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!DemandedElts.isAllOnes()) {
bool FoundNewSub = false;
SmallVector<SDValue, 2> DemandedSubOps;
for (unsigned i = 0; i != NumSubVecs; ++i) {
SDValue SubOp = Op.getOperand(i);
APInt SubElts = DemandedElts.extractBits(NumSubElts, i * NumSubElts);
SDValue NewSubOp = SimplifyMultipleUseDemandedVectorElts(
SubOp, SubElts, TLO.DAG, Depth + 1);
DemandedSubOps.push_back(NewSubOp ? NewSubOp : SubOp);
FoundNewSub = NewSubOp ? true : FoundNewSub;
}
if (FoundNewSub) {
SDValue NewOp =
TLO.DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, DemandedSubOps);
return TLO.CombineTo(Op, NewOp);
}
}
break;
}
case ISD::INSERT_SUBVECTOR: {
// Demand any elements from the subvector and the remainder from the src its
// inserted into.
SDValue Src = Op.getOperand(0);
SDValue Sub = Op.getOperand(1);
uint64_t Idx = Op.getConstantOperandVal(2);
unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx);
APInt DemandedSrcElts = DemandedElts;
DemandedSrcElts.insertBits(APInt::getZero(NumSubElts), Idx);
APInt SubUndef, SubZero;
if (SimplifyDemandedVectorElts(Sub, DemandedSubElts, SubUndef, SubZero, TLO,
Depth + 1))
return true;
// If none of the src operand elements are demanded, replace it with undef.
if (!DemandedSrcElts && !Src.isUndef())
return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
TLO.DAG.getUNDEF(VT), Sub,
Op.getOperand(2)));
if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownUndef, KnownZero,
TLO, Depth + 1))
return true;
KnownUndef.insertBits(SubUndef, Idx);
KnownZero.insertBits(SubZero, Idx);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!DemandedSrcElts.isAllOnes() || !DemandedSubElts.isAllOnes()) {
SDValue NewSrc = SimplifyMultipleUseDemandedVectorElts(
Src, DemandedSrcElts, TLO.DAG, Depth + 1);
SDValue NewSub = SimplifyMultipleUseDemandedVectorElts(
Sub, DemandedSubElts, TLO.DAG, Depth + 1);
if (NewSrc || NewSub) {
NewSrc = NewSrc ? NewSrc : Src;
NewSub = NewSub ? NewSub : Sub;
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, NewSrc,
NewSub, Op.getOperand(2));
return TLO.CombineTo(Op, NewOp);
}
}
break;
}
case ISD::EXTRACT_SUBVECTOR: {
// Offset the demanded elts by the subvector index.
SDValue Src = Op.getOperand(0);
if (Src.getValueType().isScalableVector())
break;
uint64_t Idx = Op.getConstantOperandVal(1);
unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
APInt SrcUndef, SrcZero;
if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, SrcUndef, SrcZero, TLO,
Depth + 1))
return true;
KnownUndef = SrcUndef.extractBits(NumElts, Idx);
KnownZero = SrcZero.extractBits(NumElts, Idx);
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!DemandedElts.isAllOnes()) {
SDValue NewSrc = SimplifyMultipleUseDemandedVectorElts(
Src, DemandedSrcElts, TLO.DAG, Depth + 1);
if (NewSrc) {
SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, NewSrc,
Op.getOperand(1));
return TLO.CombineTo(Op, NewOp);
}
}
break;
}
case ISD::INSERT_VECTOR_ELT: {
SDValue Vec = Op.getOperand(0);
SDValue Scl = Op.getOperand(1);
auto *CIdx = dyn_cast<ConstantSDNode>(Op.getOperand(2));
// For a legal, constant insertion index, if we don't need this insertion
// then strip it, else remove it from the demanded elts.
if (CIdx && CIdx->getAPIntValue().ult(NumElts)) {
unsigned Idx = CIdx->getZExtValue();
if (!DemandedElts[Idx])
return TLO.CombineTo(Op, Vec);
APInt DemandedVecElts(DemandedElts);
DemandedVecElts.clearBit(Idx);
if (SimplifyDemandedVectorElts(Vec, DemandedVecElts, KnownUndef,
KnownZero, TLO, Depth + 1))
return true;
KnownUndef.setBitVal(Idx, Scl.isUndef());
KnownZero.setBitVal(Idx, isNullConstant(Scl) || isNullFPConstant(Scl));
break;
}
APInt VecUndef, VecZero;
if (SimplifyDemandedVectorElts(Vec, DemandedElts, VecUndef, VecZero, TLO,
Depth + 1))
return true;
// Without knowing the insertion index we can't set KnownUndef/KnownZero.
break;
}
case ISD::VSELECT: {
SDValue Sel = Op.getOperand(0);
SDValue LHS = Op.getOperand(1);
SDValue RHS = Op.getOperand(2);
// Try to transform the select condition based on the current demanded
// elements.
APInt UndefSel, ZeroSel;
if (SimplifyDemandedVectorElts(Sel, DemandedElts, UndefSel, ZeroSel, TLO,
Depth + 1))
return true;
// See if we can simplify either vselect operand.
APInt DemandedLHS(DemandedElts);
APInt DemandedRHS(DemandedElts);
APInt UndefLHS, ZeroLHS;
APInt UndefRHS, ZeroRHS;
if (SimplifyDemandedVectorElts(LHS, DemandedLHS, UndefLHS, ZeroLHS, TLO,
Depth + 1))
return true;
if (SimplifyDemandedVectorElts(RHS, DemandedRHS, UndefRHS, ZeroRHS, TLO,
Depth + 1))
return true;
KnownUndef = UndefLHS & UndefRHS;
KnownZero = ZeroLHS & ZeroRHS;
// If we know that the selected element is always zero, we don't need the
// select value element.
APInt DemandedSel = DemandedElts & ~KnownZero;
if (DemandedSel != DemandedElts)
if (SimplifyDemandedVectorElts(Sel, DemandedSel, UndefSel, ZeroSel, TLO,
Depth + 1))
return true;
break;
}
case ISD::VECTOR_SHUFFLE: {
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
ArrayRef<int> ShuffleMask = cast<ShuffleVectorSDNode>(Op)->getMask();
// Collect demanded elements from shuffle operands..
APInt DemandedLHS(NumElts, 0);
APInt DemandedRHS(NumElts, 0);
for (unsigned i = 0; i != NumElts; ++i) {
int M = ShuffleMask[i];
if (M < 0 || !DemandedElts[i])
continue;
assert(0 <= M && M < (int)(2 * NumElts) && "Shuffle index out of range");
if (M < (int)NumElts)
DemandedLHS.setBit(M);
else
DemandedRHS.setBit(M - NumElts);
}
// See if we can simplify either shuffle operand.
APInt UndefLHS, ZeroLHS;
APInt UndefRHS, ZeroRHS;
if (SimplifyDemandedVectorElts(LHS, DemandedLHS, UndefLHS, ZeroLHS, TLO,
Depth + 1))
return true;
if (SimplifyDemandedVectorElts(RHS, DemandedRHS, UndefRHS, ZeroRHS, TLO,
Depth + 1))
return true;
// Simplify mask using undef elements from LHS/RHS.
bool Updated = false;
bool IdentityLHS = true, IdentityRHS = true;
SmallVector<int, 32> NewMask(ShuffleMask);
for (unsigned i = 0; i != NumElts; ++i) {
int &M = NewMask[i];
if (M < 0)
continue;
if (!DemandedElts[i] || (M < (int)NumElts && UndefLHS[M]) ||
(M >= (int)NumElts && UndefRHS[M - NumElts])) {
Updated = true;
M = -1;
}
IdentityLHS &= (M < 0) || (M == (int)i);
IdentityRHS &= (M < 0) || ((M - NumElts) == i);
}
// Update legal shuffle masks based on demanded elements if it won't reduce
// to Identity which can cause premature removal of the shuffle mask.
if (Updated && !IdentityLHS && !IdentityRHS && !TLO.LegalOps) {
SDValue LegalShuffle =
buildLegalVectorShuffle(VT, DL, LHS, RHS, NewMask, TLO.DAG);
if (LegalShuffle)
return TLO.CombineTo(Op, LegalShuffle);
}
// Propagate undef/zero elements from LHS/RHS.
for (unsigned i = 0; i != NumElts; ++i) {
int M = ShuffleMask[i];
if (M < 0) {
KnownUndef.setBit(i);
} else if (M < (int)NumElts) {
if (UndefLHS[M])
KnownUndef.setBit(i);
if (ZeroLHS[M])
KnownZero.setBit(i);
} else {
if (UndefRHS[M - NumElts])
KnownUndef.setBit(i);
if (ZeroRHS[M - NumElts])
KnownZero.setBit(i);
}
}
break;
}
case ISD::ANY_EXTEND_VECTOR_INREG:
case ISD::SIGN_EXTEND_VECTOR_INREG:
case ISD::ZERO_EXTEND_VECTOR_INREG: {
APInt SrcUndef, SrcZero;
SDValue Src = Op.getOperand(0);
unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts);
if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, SrcUndef, SrcZero, TLO,
Depth + 1))
return true;
KnownZero = SrcZero.zextOrTrunc(NumElts);
KnownUndef = SrcUndef.zextOrTrunc(NumElts);
if (IsLE && Op.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG &&
Op.getValueSizeInBits() == Src.getValueSizeInBits() &&
DemandedSrcElts == 1) {
// aext - if we just need the bottom element then we can bitcast.
return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src));
}
if (Op.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG) {
// zext(undef) upper bits are guaranteed to be zero.
if (DemandedElts.isSubsetOf(KnownUndef))
return TLO.CombineTo(Op, TLO.DAG.getConstant(0, SDLoc(Op), VT));
KnownUndef.clearAllBits();
// zext - if we just need the bottom element then we can mask:
// zext(and(x,c)) -> and(x,c') iff the zext is the only user of the and.
if (IsLE && DemandedSrcElts == 1 && Src.getOpcode() == ISD::AND &&
Op->isOnlyUserOf(Src.getNode()) &&
Op.getValueSizeInBits() == Src.getValueSizeInBits()) {
SDLoc DL(Op);
EVT SrcVT = Src.getValueType();
EVT SrcSVT = SrcVT.getScalarType();
SmallVector<SDValue> MaskElts;
MaskElts.push_back(TLO.DAG.getAllOnesConstant(DL, SrcSVT));
MaskElts.append(NumSrcElts - 1, TLO.DAG.getConstant(0, DL, SrcSVT));
SDValue Mask = TLO.DAG.getBuildVector(SrcVT, DL, MaskElts);
if (SDValue Fold = TLO.DAG.FoldConstantArithmetic(
ISD::AND, DL, SrcVT, {Src.getOperand(1), Mask})) {
Fold = TLO.DAG.getNode(ISD::AND, DL, SrcVT, Src.getOperand(0), Fold);
return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Fold));
}
}
}
break;
}
// TODO: There are more binop opcodes that could be handled here - MIN,
// MAX, saturated math, etc.
case ISD::ADD: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
if (Op0 == Op1 && Op->isOnlyUserOf(Op0.getNode())) {
APInt UndefLHS, ZeroLHS;
if (SimplifyDemandedVectorElts(Op0, DemandedElts, UndefLHS, ZeroLHS, TLO,
Depth + 1, /*AssumeSingleUse*/ true))
return true;
}
[[fallthrough]];
}
case ISD::AVGCEILS:
case ISD::AVGCEILU:
case ISD::AVGFLOORS:
case ISD::AVGFLOORU:
case ISD::OR:
case ISD::XOR:
case ISD::SUB:
case ISD::FADD:
case ISD::FSUB:
case ISD::FMUL:
case ISD::FDIV:
case ISD::FREM: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
APInt UndefRHS, ZeroRHS;
if (SimplifyDemandedVectorElts(Op1, DemandedElts, UndefRHS, ZeroRHS, TLO,
Depth + 1))
return true;
APInt UndefLHS, ZeroLHS;
if (SimplifyDemandedVectorElts(Op0, DemandedElts, UndefLHS, ZeroLHS, TLO,
Depth + 1))
return true;
KnownZero = ZeroLHS & ZeroRHS;
KnownUndef = getKnownUndefForVectorBinop(Op, TLO.DAG, UndefLHS, UndefRHS);
// Attempt to avoid multi-use ops if we don't need anything from them.
// TODO - use KnownUndef to relax the demandedelts?
if (!DemandedElts.isAllOnes())
if (SimplifyDemandedVectorEltsBinOp(Op0, Op1))
return true;
break;
}
case ISD::SHL:
case ISD::SRL:
case ISD::SRA:
case ISD::ROTL:
case ISD::ROTR: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
APInt UndefRHS, ZeroRHS;
if (SimplifyDemandedVectorElts(Op1, DemandedElts, UndefRHS, ZeroRHS, TLO,
Depth + 1))
return true;
APInt UndefLHS, ZeroLHS;
if (SimplifyDemandedVectorElts(Op0, DemandedElts, UndefLHS, ZeroLHS, TLO,
Depth + 1))
return true;
KnownZero = ZeroLHS;
KnownUndef = UndefLHS & UndefRHS; // TODO: use getKnownUndefForVectorBinop?
// Attempt to avoid multi-use ops if we don't need anything from them.
// TODO - use KnownUndef to relax the demandedelts?
if (!DemandedElts.isAllOnes())
if (SimplifyDemandedVectorEltsBinOp(Op0, Op1))
return true;
break;
}
case ISD::MUL:
case ISD::MULHU:
case ISD::MULHS:
case ISD::AND: {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
APInt SrcUndef, SrcZero;
if (SimplifyDemandedVectorElts(Op1, DemandedElts, SrcUndef, SrcZero, TLO,
Depth + 1))
return true;
// If we know that a demanded element was zero in Op1 we don't need to
// demand it in Op0 - its guaranteed to be zero.
APInt DemandedElts0 = DemandedElts & ~SrcZero;
if (SimplifyDemandedVectorElts(Op0, DemandedElts0, KnownUndef, KnownZero,
TLO, Depth + 1))
return true;
KnownUndef &= DemandedElts0;
KnownZero &= DemandedElts0;
// If every element pair has a zero/undef then just fold to zero.
// fold (and x, undef) -> 0 / (and x, 0) -> 0
// fold (mul x, undef) -> 0 / (mul x, 0) -> 0
if (DemandedElts.isSubsetOf(SrcZero | KnownZero | SrcUndef | KnownUndef))
return TLO.CombineTo(Op, TLO.DAG.getConstant(0, SDLoc(Op), VT));
// If either side has a zero element, then the result element is zero, even
// if the other is an UNDEF.
// TODO: Extend getKnownUndefForVectorBinop to also deal with known zeros
// and then handle 'and' nodes with the rest of the binop opcodes.
KnownZero |= SrcZero;
KnownUndef &= SrcUndef;
KnownUndef &= ~KnownZero;
// Attempt to avoid multi-use ops if we don't need anything from them.
if (!DemandedElts.isAllOnes())
if (SimplifyDemandedVectorEltsBinOp(Op0, Op1))
return true;
break;
}
case ISD::TRUNCATE:
case ISD::SIGN_EXTEND:
case ISD::ZERO_EXTEND:
if (SimplifyDemandedVectorElts(Op.getOperand(0), DemandedElts, KnownUndef,
KnownZero, TLO, Depth + 1))
return true;
if (Op.getOpcode() == ISD::ZERO_EXTEND) {
// zext(undef) upper bits are guaranteed to be zero.
if (DemandedElts.isSubsetOf(KnownUndef))
return TLO.CombineTo(Op, TLO.DAG.getConstant(0, SDLoc(Op), VT));
KnownUndef.clearAllBits();
}
break;
default: {
if (Op.getOpcode() >= ISD::BUILTIN_OP_END) {
if (SimplifyDemandedVectorEltsForTargetNode(Op, DemandedElts, KnownUndef,
KnownZero, TLO, Depth))
return true;
} else {
KnownBits Known;
APInt DemandedBits = APInt::getAllOnes(EltSizeInBits);
if (SimplifyDemandedBits(Op, DemandedBits, OriginalDemandedElts, Known,
TLO, Depth, AssumeSingleUse))
return true;
}
break;
}
}
assert((KnownUndef & KnownZero) == 0 && "Elements flagged as undef AND zero");
// Constant fold all undef cases.
// TODO: Handle zero cases as well.
if (DemandedElts.isSubsetOf(KnownUndef))
return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT));
return false;
}
/// Determine which of the bits specified in Mask are known to be either zero or
/// one and return them in the Known.
void TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
KnownBits &Known,
const APInt &DemandedElts,
const SelectionDAG &DAG,
unsigned Depth) const {
assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use MaskedValueIsZero if you don't know whether Op"
" is a target node!");
Known.resetAll();
}
void TargetLowering::computeKnownBitsForTargetInstr(
GISelKnownBits &Analysis, Register R, KnownBits &Known,
const APInt &DemandedElts, const MachineRegisterInfo &MRI,
unsigned Depth) const {
Known.resetAll();
}
void TargetLowering::computeKnownBitsForFrameIndex(
const int FrameIdx, KnownBits &Known, const MachineFunction &MF) const {
// The low bits are known zero if the pointer is aligned.
Known.Zero.setLowBits(Log2(MF.getFrameInfo().getObjectAlign(FrameIdx)));
}
Align TargetLowering::computeKnownAlignForTargetInstr(
GISelKnownBits &Analysis, Register R, const MachineRegisterInfo &MRI,
unsigned Depth) const {
return Align(1);
}
/// This method can be implemented by targets that want to expose additional
/// information about sign bits to the DAG Combiner.
unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
const APInt &,
const SelectionDAG &,
unsigned Depth) const {
assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use ComputeNumSignBits if you don't know whether Op"
" is a target node!");
return 1;
}
unsigned TargetLowering::computeNumSignBitsForTargetInstr(
GISelKnownBits &Analysis, Register R, const APInt &DemandedElts,
const MachineRegisterInfo &MRI, unsigned Depth) const {
return 1;
}
bool TargetLowering::SimplifyDemandedVectorEltsForTargetNode(
SDValue Op, const APInt &DemandedElts, APInt &KnownUndef, APInt &KnownZero,
TargetLoweringOpt &TLO, unsigned Depth) const {
assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use SimplifyDemandedVectorElts if you don't know whether Op"
" is a target node!");
return false;
}
bool TargetLowering::SimplifyDemandedBitsForTargetNode(
SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
KnownBits &Known, TargetLoweringOpt &TLO, unsigned Depth) const {
assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use SimplifyDemandedBits if you don't know whether Op"
" is a target node!");
computeKnownBitsForTargetNode(Op, Known, DemandedElts, TLO.DAG, Depth);
return false;
}
SDValue TargetLowering::SimplifyMultipleUseDemandedBitsForTargetNode(
SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
SelectionDAG &DAG, unsigned Depth) const {
assert(
(Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use SimplifyMultipleUseDemandedBits if you don't know whether Op"
" is a target node!");
return SDValue();
}
SDValue
TargetLowering::buildLegalVectorShuffle(EVT VT, const SDLoc &DL, SDValue N0,
SDValue N1, MutableArrayRef<int> Mask,
SelectionDAG &DAG) const {
bool LegalMask = isShuffleMaskLegal(Mask, VT);
if (!LegalMask) {
std::swap(N0, N1);
ShuffleVectorSDNode::commuteMask(Mask);
LegalMask = isShuffleMaskLegal(Mask, VT);
}
if (!LegalMask)
return SDValue();
return DAG.getVectorShuffle(VT, DL, N0, N1, Mask);
}
const Constant *TargetLowering::getTargetConstantFromLoad(LoadSDNode*) const {
return nullptr;
}
bool TargetLowering::isGuaranteedNotToBeUndefOrPoisonForTargetNode(
SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
bool PoisonOnly, unsigned Depth) const {
assert(
(Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use isGuaranteedNotToBeUndefOrPoison if you don't know whether Op"
" is a target node!");
// If Op can't create undef/poison and none of its operands are undef/poison
// then Op is never undef/poison.
return !canCreateUndefOrPoisonForTargetNode(Op, DemandedElts, DAG, PoisonOnly,
/*ConsiderFlags*/ true, Depth) &&
all_of(Op->ops(), [&](SDValue V) {
return DAG.isGuaranteedNotToBeUndefOrPoison(V, PoisonOnly,
Depth + 1);
});
}
bool TargetLowering::canCreateUndefOrPoisonForTargetNode(
SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
bool PoisonOnly, bool ConsiderFlags, unsigned Depth) const {
assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use canCreateUndefOrPoison if you don't know whether Op"
" is a target node!");
// Be conservative and return true.
return true;
}
bool TargetLowering::isKnownNeverNaNForTargetNode(SDValue Op,
const SelectionDAG &DAG,
bool SNaN,
unsigned Depth) const {
assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use isKnownNeverNaN if you don't know whether Op"
" is a target node!");
return false;
}
bool TargetLowering::isSplatValueForTargetNode(SDValue Op,
const APInt &DemandedElts,
APInt &UndefElts,
const SelectionDAG &DAG,
unsigned Depth) const {
assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use isSplatValue if you don't know whether Op"
" is a target node!");
return false;
}
// FIXME: Ideally, this would use ISD::isConstantSplatVector(), but that must
// work with truncating build vectors and vectors with elements of less than
// 8 bits.
bool TargetLowering::isConstTrueVal(SDValue N) const {
if (!N)
return false;
unsigned EltWidth;
APInt CVal;
if (ConstantSDNode *CN = isConstOrConstSplat(N, /*AllowUndefs=*/false,
/*AllowTruncation=*/true)) {
CVal = CN->getAPIntValue();
EltWidth = N.getValueType().getScalarSizeInBits();
} else
return false;
// If this is a truncating splat, truncate the splat value.
// Otherwise, we may fail to match the expected values below.
if (EltWidth < CVal.getBitWidth())
CVal = CVal.trunc(EltWidth);
switch (getBooleanContents(N.getValueType())) {
case UndefinedBooleanContent:
return CVal[0];
case ZeroOrOneBooleanContent:
return CVal.isOne();
case ZeroOrNegativeOneBooleanContent:
return CVal.isAllOnes();
}
llvm_unreachable("Invalid boolean contents");
}
bool TargetLowering::isConstFalseVal(SDValue N) const {
if (!N)
return false;
const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
if (!CN) {
const BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
if (!BV)
return false;
// Only interested in constant splats, we don't care about undef
// elements in identifying boolean constants and getConstantSplatNode
// returns NULL if all ops are undef;
CN = BV->getConstantSplatNode();
if (!CN)
return false;
}
if (getBooleanContents(N->getValueType(0)) == UndefinedBooleanContent)
return !CN->getAPIntValue()[0];
return CN->isZero();
}
bool TargetLowering::isExtendedTrueVal(const ConstantSDNode *N, EVT VT,
bool SExt) const {
if (VT == MVT::i1)
return N->isOne();
TargetLowering::BooleanContent Cnt = getBooleanContents(VT);
switch (Cnt) {
case TargetLowering::ZeroOrOneBooleanContent:
// An extended value of 1 is always true, unless its original type is i1,
// in which case it will be sign extended to -1.
return (N->isOne() && !SExt) || (SExt && (N->getValueType(0) != MVT::i1));
case TargetLowering::UndefinedBooleanContent:
case TargetLowering::ZeroOrNegativeOneBooleanContent:
return N->isAllOnes() && SExt;
}
llvm_unreachable("Unexpected enumeration.");
}
/// This helper function of SimplifySetCC tries to optimize the comparison when
/// either operand of the SetCC node is a bitwise-and instruction.
SDValue TargetLowering::foldSetCCWithAnd(EVT VT, SDValue N0, SDValue N1,
ISD::CondCode Cond, const SDLoc &DL,
DAGCombinerInfo &DCI) const {
if (N1.getOpcode() == ISD::AND && N0.getOpcode() != ISD::AND)
std::swap(N0, N1);
SelectionDAG &DAG = DCI.DAG;
EVT OpVT = N0.getValueType();
if (N0.getOpcode() != ISD::AND || !OpVT.isInteger() ||
(Cond != ISD::SETEQ && Cond != ISD::SETNE))
return SDValue();
// (X & Y) != 0 --> zextOrTrunc(X & Y)
// iff everything but LSB is known zero:
if (Cond == ISD::SETNE && isNullConstant(N1) &&
(getBooleanContents(OpVT) == TargetLowering::UndefinedBooleanContent ||
getBooleanContents(OpVT) == TargetLowering::ZeroOrOneBooleanContent)) {
unsigned NumEltBits = OpVT.getScalarSizeInBits();
APInt UpperBits = APInt::getHighBitsSet(NumEltBits, NumEltBits - 1);
if (DAG.MaskedValueIsZero(N0, UpperBits))
return DAG.getBoolExtOrTrunc(N0, DL, VT, OpVT);
}
// Try to eliminate a power-of-2 mask constant by converting to a signbit
// test in a narrow type that we can truncate to with no cost. Examples:
// (i32 X & 32768) == 0 --> (trunc X to i16) >= 0
// (i32 X & 32768) != 0 --> (trunc X to i16) < 0
// TODO: This conservatively checks for type legality on the source and
// destination types. That may inhibit optimizations, but it also
// allows setcc->shift transforms that may be more beneficial.
auto *AndC = dyn_cast<ConstantSDNode>(N0.getOperand(1));
if (AndC && isNullConstant(N1) && AndC->getAPIntValue().isPowerOf2() &&
isTypeLegal(OpVT) && N0.hasOneUse()) {
EVT NarrowVT = EVT::getIntegerVT(*DAG.getContext(),
AndC->getAPIntValue().getActiveBits());
if (isTruncateFree(OpVT, NarrowVT) && isTypeLegal(NarrowVT)) {
SDValue Trunc = DAG.getZExtOrTrunc(N0.getOperand(0), DL, NarrowVT);
SDValue Zero = DAG.getConstant(0, DL, NarrowVT);
return DAG.getSetCC(DL, VT, Trunc, Zero,
Cond == ISD::SETEQ ? ISD::SETGE : ISD::SETLT);
}
}
// Match these patterns in any of their permutations:
// (X & Y) == Y
// (X & Y) != Y
SDValue X, Y;
if (N0.getOperand(0) == N1) {
X = N0.getOperand(1);
Y = N0.getOperand(0);
} else if (N0.getOperand(1) == N1) {
X = N0.getOperand(0);
Y = N0.getOperand(1);
} else {
return SDValue();
}
// TODO: We should invert (X & Y) eq/ne 0 -> (X & Y) ne/eq Y if
// `isXAndYEqZeroPreferableToXAndYEqY` is false. This is a bit difficult as
// its liable to create and infinite loop.
SDValue Zero = DAG.getConstant(0, DL, OpVT);
if (isXAndYEqZeroPreferableToXAndYEqY(Cond, OpVT) &&
DAG.isKnownToBeAPowerOfTwo(Y)) {
// Simplify X & Y == Y to X & Y != 0 if Y has exactly one bit set.
// Note that where Y is variable and is known to have at most one bit set
// (for example, if it is Z & 1) we cannot do this; the expressions are not
// equivalent when Y == 0.
assert(OpVT.isInteger());
Cond = ISD::getSetCCInverse(Cond, OpVT);
if (DCI.isBeforeLegalizeOps() ||
isCondCodeLegal(Cond, N0.getSimpleValueType()))
return DAG.getSetCC(DL, VT, N0, Zero, Cond);
} else if (N0.hasOneUse() && hasAndNotCompare(Y)) {
// If the target supports an 'and-not' or 'and-complement' logic operation,
// try to use that to make a comparison operation more efficient.
// But don't do this transform if the mask is a single bit because there are
// more efficient ways to deal with that case (for example, 'bt' on x86 or
// 'rlwinm' on PPC).
// Bail out if the compare operand that we want to turn into a zero is
// already a zero (otherwise, infinite loop).
if (isNullConstant(Y))
return SDValue();
// Transform this into: ~X & Y == 0.
SDValue NotX = DAG.getNOT(SDLoc(X), X, OpVT);
SDValue NewAnd = DAG.getNode(ISD::AND, SDLoc(N0), OpVT, NotX, Y);
return DAG.getSetCC(DL, VT, NewAnd, Zero, Cond);
}
return SDValue();
}
/// There are multiple IR patterns that could be checking whether certain
/// truncation of a signed number would be lossy or not. The pattern which is
/// best at IR level, may not lower optimally. Thus, we want to unfold it.
/// We are looking for the following pattern: (KeptBits is a constant)
/// (add %x, (1 << (KeptBits-1))) srccond (1 << KeptBits)
/// KeptBits won't be bitwidth(x), that will be constant-folded to true/false.
/// KeptBits also can't be 1, that would have been folded to %x dstcond 0
/// We will unfold it into the natural trunc+sext pattern:
/// ((%x << C) a>> C) dstcond %x
/// Where C = bitwidth(x) - KeptBits and C u< bitwidth(x)
SDValue TargetLowering::optimizeSetCCOfSignedTruncationCheck(
EVT SCCVT, SDValue N0, SDValue N1, ISD::CondCode Cond, DAGCombinerInfo &DCI,
const SDLoc &DL) const {
// We must be comparing with a constant.
ConstantSDNode *C1;
if (!(C1 = dyn_cast<ConstantSDNode>(N1)))
return SDValue();
// N0 should be: add %x, (1 << (KeptBits-1))
if (N0->getOpcode() != ISD::ADD)
return SDValue();
// And we must be 'add'ing a constant.
ConstantSDNode *C01;
if (!(C01 = dyn_cast<ConstantSDNode>(N0->getOperand(1))))
return SDValue();
SDValue X = N0->getOperand(0);
EVT XVT = X.getValueType();
// Validate constants ...
APInt I1 = C1->getAPIntValue();
ISD::CondCode NewCond;
if (Cond == ISD::CondCode::SETULT) {
NewCond = ISD::CondCode::SETEQ;
} else if (Cond == ISD::CondCode::SETULE) {
NewCond = ISD::CondCode::SETEQ;
// But need to 'canonicalize' the constant.
I1 += 1;
} else if (Cond == ISD::CondCode::SETUGT) {
NewCond = ISD::CondCode::SETNE;
// But need to 'canonicalize' the constant.
I1 += 1;
} else if (Cond == ISD::CondCode::SETUGE) {
NewCond = ISD::CondCode::SETNE;
} else
return SDValue();
APInt I01 = C01->getAPIntValue();
auto checkConstants = [&I1, &I01]() -> bool {
// Both of them must be power-of-two, and the constant from setcc is bigger.
return I1.ugt(I01) && I1.isPowerOf2() && I01.isPowerOf2();
};
if (checkConstants()) {
// Great, e.g. got icmp ult i16 (add i16 %x, 128), 256
} else {
// What if we invert constants? (and the target predicate)
I1.negate();
I01.negate();
assert(XVT.isInteger());
NewCond = getSetCCInverse(NewCond, XVT);
if (!checkConstants())
return SDValue();
// Great, e.g. got icmp uge i16 (add i16 %x, -128), -256
}
// They are power-of-two, so which bit is set?
const unsigned KeptBits = I1.logBase2();
const unsigned KeptBitsMinusOne = I01.logBase2();
// Magic!
if (KeptBits != (KeptBitsMinusOne + 1))
return SDValue();
assert(KeptBits > 0 && KeptBits < XVT.getSizeInBits() && "unreachable");
// We don't want to do this in every single case.
SelectionDAG &DAG = DCI.DAG;
if (!DAG.getTargetLoweringInfo().shouldTransformSignedTruncationCheck(
XVT, KeptBits))
return SDValue();
// Unfold into: sext_inreg(%x) cond %x
// Where 'cond' will be either 'eq' or 'ne'.
SDValue SExtInReg = DAG.getNode(
ISD::SIGN_EXTEND_INREG, DL, XVT, X,
DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), KeptBits)));
return DAG.getSetCC(DL, SCCVT, SExtInReg, X, NewCond);
}
// (X & (C l>>/<< Y)) ==/!= 0 --> ((X <</l>> Y) & C) ==/!= 0
SDValue TargetLowering::optimizeSetCCByHoistingAndByConstFromLogicalShift(
EVT SCCVT, SDValue N0, SDValue N1C, ISD::CondCode Cond,
DAGCombinerInfo &DCI, const SDLoc &DL) const {
assert(isConstOrConstSplat(N1C) && isConstOrConstSplat(N1C)->isZero() &&
"Should be a comparison with 0.");
assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
"Valid only for [in]equality comparisons.");
unsigned NewShiftOpcode;
SDValue X, C, Y;
SelectionDAG &DAG = DCI.DAG;
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
// Look for '(C l>>/<< Y)'.
auto Match = [&NewShiftOpcode, &X, &C, &Y, &TLI, &DAG](SDValue V) {
// The shift should be one-use.
if (!V.hasOneUse())
return false;
unsigned OldShiftOpcode = V.getOpcode();
switch (OldShiftOpcode) {
case ISD::SHL:
NewShiftOpcode = ISD::SRL;
break;
case ISD::SRL:
NewShiftOpcode = ISD::SHL;
break;
default:
return false; // must be a logical shift.
}
// We should be shifting a constant.
// FIXME: best to use isConstantOrConstantVector().
C = V.getOperand(0);
ConstantSDNode *CC =
isConstOrConstSplat(C, /*AllowUndefs=*/true, /*AllowTruncation=*/true);
if (!CC)
return false;
Y = V.getOperand(1);
ConstantSDNode *XC =
isConstOrConstSplat(X, /*AllowUndefs=*/true, /*AllowTruncation=*/true);
return TLI.shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
X, XC, CC, Y, OldShiftOpcode, NewShiftOpcode, DAG);
};
// LHS of comparison should be an one-use 'and'.
if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
return SDValue();
X = N0.getOperand(0);
SDValue Mask = N0.getOperand(1);
// 'and' is commutative!
if (!Match(Mask)) {
std::swap(X, Mask);
if (!Match(Mask))
return SDValue();
}
EVT VT = X.getValueType();
// Produce:
// ((X 'OppositeShiftOpcode' Y) & C) Cond 0
SDValue T0 = DAG.getNode(NewShiftOpcode, DL, VT, X, Y);
SDValue T1 = DAG.getNode(ISD::AND, DL, VT, T0, C);
SDValue T2 = DAG.getSetCC(DL, SCCVT, T1, N1C, Cond);
return T2;
}
/// Try to fold an equality comparison with a {add/sub/xor} binary operation as
/// the 1st operand (N0). Callers are expected to swap the N0/N1 parameters to
/// handle the commuted versions of these patterns.
SDValue TargetLowering::foldSetCCWithBinOp(EVT VT, SDValue N0, SDValue N1,
ISD::CondCode Cond, const SDLoc &DL,
DAGCombinerInfo &DCI) const {
unsigned BOpcode = N0.getOpcode();
assert((BOpcode == ISD::ADD || BOpcode == ISD::SUB || BOpcode == ISD::XOR) &&
"Unexpected binop");
assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Unexpected condcode");
// (X + Y) == X --> Y == 0
// (X - Y) == X --> Y == 0
// (X ^ Y) == X --> Y == 0
SelectionDAG &DAG = DCI.DAG;
EVT OpVT = N0.getValueType();
SDValue X = N0.getOperand(0);
SDValue Y = N0.getOperand(1);
if (X == N1)
return DAG.getSetCC(DL, VT, Y, DAG.getConstant(0, DL, OpVT), Cond);
if (Y != N1)
return SDValue();
// (X + Y) == Y --> X == 0
// (X ^ Y) == Y --> X == 0
if (BOpcode == ISD::ADD || BOpcode == ISD::XOR)
return DAG.getSetCC(DL, VT, X, DAG.getConstant(0, DL, OpVT), Cond);
// The shift would not be valid if the operands are boolean (i1).
if (!N0.hasOneUse() || OpVT.getScalarSizeInBits() == 1)
return SDValue();
// (X - Y) == Y --> X == Y << 1
SDValue One = DAG.getShiftAmountConstant(1, OpVT, DL);
SDValue YShl1 = DAG.getNode(ISD::SHL, DL, N1.getValueType(), Y, One);
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(YShl1.getNode());
return DAG.getSetCC(DL, VT, X, YShl1, Cond);
}
static SDValue simplifySetCCWithCTPOP(const TargetLowering &TLI, EVT VT,
SDValue N0, const APInt &C1,
ISD::CondCode Cond, const SDLoc &dl,
SelectionDAG &DAG) {
// Look through truncs that don't change the value of a ctpop.
// FIXME: Add vector support? Need to be careful with setcc result type below.
SDValue CTPOP = N0;
if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() && !VT.isVector() &&
N0.getScalarValueSizeInBits() > Log2_32(N0.getOperand(0).getScalarValueSizeInBits()))
CTPOP = N0.getOperand(0);
if (CTPOP.getOpcode() != ISD::CTPOP || !CTPOP.hasOneUse())
return SDValue();
EVT CTVT = CTPOP.getValueType();
SDValue CTOp = CTPOP.getOperand(0);
// Expand a power-of-2-or-zero comparison based on ctpop:
// (ctpop x) u< 2 -> (x & x-1) == 0
// (ctpop x) u> 1 -> (x & x-1) != 0
if (Cond == ISD::SETULT || Cond == ISD::SETUGT) {
// Keep the CTPOP if it is a cheap vector op.
if (CTVT.isVector() && TLI.isCtpopFast(CTVT))
return SDValue();
unsigned CostLimit = TLI.getCustomCtpopCost(CTVT, Cond);
if (C1.ugt(CostLimit + (Cond == ISD::SETULT)))
return SDValue();
if (C1 == 0 && (Cond == ISD::SETULT))
return SDValue(); // This is handled elsewhere.
unsigned Passes = C1.getLimitedValue() - (Cond == ISD::SETULT);
SDValue NegOne = DAG.getAllOnesConstant(dl, CTVT);
SDValue Result = CTOp;
for (unsigned i = 0; i < Passes; i++) {
SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, Result, NegOne);
Result = DAG.getNode(ISD::AND, dl, CTVT, Result, Add);
}
ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
return DAG.getSetCC(dl, VT, Result, DAG.getConstant(0, dl, CTVT), CC);
}
// Expand a power-of-2 comparison based on ctpop
if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && C1 == 1) {
// Keep the CTPOP if it is cheap.
if (TLI.isCtpopFast(CTVT))
return SDValue();
SDValue Zero = DAG.getConstant(0, dl, CTVT);
SDValue NegOne = DAG.getAllOnesConstant(dl, CTVT);
assert(CTVT.isInteger());
SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, CTOp, NegOne);
// Its not uncommon for known-never-zero X to exist in (ctpop X) eq/ne 1, so
// check before emitting a potentially unnecessary op.
if (DAG.isKnownNeverZero(CTOp)) {
// (ctpop x) == 1 --> (x & x-1) == 0
// (ctpop x) != 1 --> (x & x-1) != 0
SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Add);
SDValue RHS = DAG.getSetCC(dl, VT, And, Zero, Cond);
return RHS;
}
// (ctpop x) == 1 --> (x ^ x-1) > x-1
// (ctpop x) != 1 --> (x ^ x-1) <= x-1
SDValue Xor = DAG.getNode(ISD::XOR, dl, CTVT, CTOp, Add);
ISD::CondCode CmpCond = Cond == ISD::SETEQ ? ISD::SETUGT : ISD::SETULE;
return DAG.getSetCC(dl, VT, Xor, Add, CmpCond);
}
return SDValue();
}
static SDValue foldSetCCWithRotate(EVT VT, SDValue N0, SDValue N1,
ISD::CondCode Cond, const SDLoc &dl,
SelectionDAG &DAG) {
if (Cond != ISD::SETEQ && Cond != ISD::SETNE)
return SDValue();
auto *C1 = isConstOrConstSplat(N1, /* AllowUndefs */ true);
if (!C1 || !(C1->isZero() || C1->isAllOnes()))
return SDValue();
auto getRotateSource = [](SDValue X) {
if (X.getOpcode() == ISD::ROTL || X.getOpcode() == ISD::ROTR)
return X.getOperand(0);
return SDValue();
};
// Peek through a rotated value compared against 0 or -1:
// (rot X, Y) == 0/-1 --> X == 0/-1
// (rot X, Y) != 0/-1 --> X != 0/-1
if (SDValue R = getRotateSource(N0))
return DAG.getSetCC(dl, VT, R, N1, Cond);
// Peek through an 'or' of a rotated value compared against 0:
// or (rot X, Y), Z ==/!= 0 --> (or X, Z) ==/!= 0
// or Z, (rot X, Y) ==/!= 0 --> (or X, Z) ==/!= 0
//
// TODO: Add the 'and' with -1 sibling.
// TODO: Recurse through a series of 'or' ops to find the rotate.
EVT OpVT = N0.getValueType();
if (N0.hasOneUse() && N0.getOpcode() == ISD::OR && C1->isZero()) {
if (SDValue R = getRotateSource(N0.getOperand(0))) {
SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, R, N0.getOperand(1));
return DAG.getSetCC(dl, VT, NewOr, N1, Cond);
}
if (SDValue R = getRotateSource(N0.getOperand(1))) {
SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, R, N0.getOperand(0));
return DAG.getSetCC(dl, VT, NewOr, N1, Cond);
}
}
return SDValue();
}
static SDValue foldSetCCWithFunnelShift(EVT VT, SDValue N0, SDValue N1,
ISD::CondCode Cond, const SDLoc &dl,
SelectionDAG &DAG) {
// If we are testing for all-bits-clear, we might be able to do that with
// less shifting since bit-order does not matter.
if (Cond != ISD::SETEQ && Cond != ISD::SETNE)
return SDValue();
auto *C1 = isConstOrConstSplat(N1, /* AllowUndefs */ true);
if (!C1 || !C1->isZero())
return SDValue();
if (!N0.hasOneUse() ||
(N0.getOpcode() != ISD::FSHL && N0.getOpcode() != ISD::FSHR))
return SDValue();
unsigned BitWidth = N0.getScalarValueSizeInBits();
auto *ShAmtC = isConstOrConstSplat(N0.getOperand(2));
if (!ShAmtC || ShAmtC->getAPIntValue().uge(BitWidth))
return SDValue();
// Canonicalize fshr as fshl to reduce pattern-matching.
unsigned ShAmt = ShAmtC->getZExtValue();
if (N0.getOpcode() == ISD::FSHR)
ShAmt = BitWidth - ShAmt;
// Match an 'or' with a specific operand 'Other' in either commuted variant.
SDValue X, Y;
auto matchOr = [&X, &Y](SDValue Or, SDValue Other) {
if (Or.getOpcode() != ISD::OR || !Or.hasOneUse())
return false;
if (Or.getOperand(0) == Other) {
X = Or.getOperand(0);
Y = Or.getOperand(1);
return true;
}
if (Or.getOperand(1) == Other) {
X = Or.getOperand(1);
Y = Or.getOperand(0);
return true;
}
return false;
};
EVT OpVT = N0.getValueType();
EVT ShAmtVT = N0.getOperand(2).getValueType();
SDValue F0 = N0.getOperand(0);
SDValue F1 = N0.getOperand(1);
if (matchOr(F0, F1)) {
// fshl (or X, Y), X, C ==/!= 0 --> or (shl Y, C), X ==/!= 0
SDValue NewShAmt = DAG.getConstant(ShAmt, dl, ShAmtVT);
SDValue Shift = DAG.getNode(ISD::SHL, dl, OpVT, Y, NewShAmt);
SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, Shift, X);
return DAG.getSetCC(dl, VT, NewOr, N1, Cond);
}
if (matchOr(F1, F0)) {
// fshl X, (or X, Y), C ==/!= 0 --> or (srl Y, BW-C), X ==/!= 0
SDValue NewShAmt = DAG.getConstant(BitWidth - ShAmt, dl, ShAmtVT);
SDValue Shift = DAG.getNode(ISD::SRL, dl, OpVT, Y, NewShAmt);
SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, Shift, X);
return DAG.getSetCC(dl, VT, NewOr, N1, Cond);
}
return SDValue();
}
/// Try to simplify a setcc built with the specified operands and cc. If it is
/// unable to simplify it, return a null SDValue.
SDValue TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
ISD::CondCode Cond, bool foldBooleans,
DAGCombinerInfo &DCI,
const SDLoc &dl) const {
SelectionDAG &DAG = DCI.DAG;
const DataLayout &Layout = DAG.getDataLayout();
EVT OpVT = N0.getValueType();
AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
// Constant fold or commute setcc.
if (SDValue Fold = DAG.FoldSetCC(VT, N0, N1, Cond, dl))
return Fold;
bool N0ConstOrSplat =
isConstOrConstSplat(N0, /*AllowUndefs*/ false, /*AllowTruncate*/ true);
bool N1ConstOrSplat =
isConstOrConstSplat(N1, /*AllowUndefs*/ false, /*AllowTruncate*/ true);
// Canonicalize toward having the constant on the RHS.
// TODO: Handle non-splat vector constants. All undef causes trouble.
// FIXME: We can't yet fold constant scalable vector splats, so avoid an
// infinite loop here when we encounter one.
ISD::CondCode SwappedCC = ISD::getSetCCSwappedOperands(Cond);
if (N0ConstOrSplat && !N1ConstOrSplat &&
(DCI.isBeforeLegalizeOps() ||
isCondCodeLegal(SwappedCC, N0.getSimpleValueType())))
return DAG.getSetCC(dl, VT, N1, N0, SwappedCC);
// If we have a subtract with the same 2 non-constant operands as this setcc
// -- but in reverse order -- then try to commute the operands of this setcc
// to match. A matching pair of setcc (cmp) and sub may be combined into 1
// instruction on some targets.
if (!N0ConstOrSplat && !N1ConstOrSplat &&
(DCI.isBeforeLegalizeOps() ||
isCondCodeLegal(SwappedCC, N0.getSimpleValueType())) &&
DAG.doesNodeExist(ISD::SUB, DAG.getVTList(OpVT), {N1, N0}) &&
!DAG.doesNodeExist(ISD::SUB, DAG.getVTList(OpVT), {N0, N1}))
return DAG.getSetCC(dl, VT, N1, N0, SwappedCC);
if (SDValue V = foldSetCCWithRotate(VT, N0, N1, Cond, dl, DAG))
return V;
if (SDValue V = foldSetCCWithFunnelShift(VT, N0, N1, Cond, dl, DAG))
return V;
if (auto *N1C = isConstOrConstSplat(N1)) {
const APInt &C1 = N1C->getAPIntValue();
// Optimize some CTPOP cases.
if (SDValue V = simplifySetCCWithCTPOP(*this, VT, N0, C1, Cond, dl, DAG))
return V;
// For equality to 0 of a no-wrap multiply, decompose and test each op:
// X * Y == 0 --> (X == 0) || (Y == 0)
// X * Y != 0 --> (X != 0) && (Y != 0)
// TODO: This bails out if minsize is set, but if the target doesn't have a
// single instruction multiply for this type, it would likely be
// smaller to decompose.
if (C1.isZero() && (Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
N0.getOpcode() == ISD::MUL && N0.hasOneUse() &&
(N0->getFlags().hasNoUnsignedWrap() ||
N0->getFlags().hasNoSignedWrap()) &&
!Attr.hasFnAttr(Attribute::MinSize)) {
SDValue IsXZero = DAG.getSetCC(dl, VT, N0.getOperand(0), N1, Cond);
SDValue IsYZero = DAG.getSetCC(dl, VT, N0.getOperand(1), N1, Cond);
unsigned LogicOp = Cond == ISD::SETEQ ? ISD::OR : ISD::AND;
return DAG.getNode(LogicOp, dl, VT, IsXZero, IsYZero);
}
// If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
// equality comparison, then we're just comparing whether X itself is
// zero.
if (N0.getOpcode() == ISD::SRL && (C1.isZero() || C1.isOne()) &&
N0.getOperand(0).getOpcode() == ISD::CTLZ &&
llvm::has_single_bit<uint32_t>(N0.getScalarValueSizeInBits())) {
if (ConstantSDNode *ShAmt = isConstOrConstSplat(N0.getOperand(1))) {
if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
ShAmt->getAPIntValue() == Log2_32(N0.getScalarValueSizeInBits())) {
if ((C1 == 0) == (Cond == ISD::SETEQ)) {
// (srl (ctlz x), 5) == 0 -> X != 0
// (srl (ctlz x), 5) != 1 -> X != 0
Cond = ISD::SETNE;
} else {
// (srl (ctlz x), 5) != 0 -> X == 0
// (srl (ctlz x), 5) == 1 -> X == 0
Cond = ISD::SETEQ;
}
SDValue Zero = DAG.getConstant(0, dl, N0.getValueType());
return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), Zero,
Cond);
}
}
}
}
// FIXME: Support vectors.
if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
const APInt &C1 = N1C->getAPIntValue();
// (zext x) == C --> x == (trunc C)
// (sext x) == C --> x == (trunc C)
if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
DCI.isBeforeLegalize() && N0->hasOneUse()) {
unsigned MinBits = N0.getValueSizeInBits();
SDValue PreExt;
bool Signed = false;
if (N0->getOpcode() == ISD::ZERO_EXTEND) {
// ZExt
MinBits = N0->getOperand(0).getValueSizeInBits();
PreExt = N0->getOperand(0);
} else if (N0->getOpcode() == ISD::AND) {
// DAGCombine turns costly ZExts into ANDs
if (auto *C = dyn_cast<ConstantSDNode>(N0->getOperand(1)))
if ((C->getAPIntValue()+1).isPowerOf2()) {
MinBits = C->getAPIntValue().countr_one();
PreExt = N0->getOperand(0);
}
} else if (N0->getOpcode() == ISD::SIGN_EXTEND) {
// SExt
MinBits = N0->getOperand(0).getValueSizeInBits();
PreExt = N0->getOperand(0);
Signed = true;
} else if (auto *LN0 = dyn_cast<LoadSDNode>(N0)) {
// ZEXTLOAD / SEXTLOAD
if (LN0->getExtensionType() == ISD::ZEXTLOAD) {
MinBits = LN0->getMemoryVT().getSizeInBits();
PreExt = N0;
} else if (LN0->getExtensionType() == ISD::SEXTLOAD) {
Signed = true;
MinBits = LN0->getMemoryVT().getSizeInBits();
PreExt = N0;
}
}
// Figure out how many bits we need to preserve this constant.
unsigned ReqdBits = Signed ? C1.getSignificantBits() : C1.getActiveBits();
// Make sure we're not losing bits from the constant.
if (MinBits > 0 &&
MinBits < C1.getBitWidth() &&
MinBits >= ReqdBits) {
EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits);
if (isTypeDesirableForOp(ISD::SETCC, MinVT)) {
// Will get folded away.
SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreExt);
if (MinBits == 1 && C1 == 1)
// Invert the condition.
return DAG.getSetCC(dl, VT, Trunc, DAG.getConstant(0, dl, MVT::i1),
Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
SDValue C = DAG.getConstant(C1.trunc(MinBits), dl, MinVT);
return DAG.getSetCC(dl, VT, Trunc, C, Cond);
}
// If truncating the setcc operands is not desirable, we can still
// simplify the expression in some cases:
// setcc ([sz]ext (setcc x, y, cc)), 0, setne) -> setcc (x, y, cc)
// setcc ([sz]ext (setcc x, y, cc)), 0, seteq) -> setcc (x, y, inv(cc))
// setcc (zext (setcc x, y, cc)), 1, setne) -> setcc (x, y, inv(cc))
// setcc (zext (setcc x, y, cc)), 1, seteq) -> setcc (x, y, cc)
// setcc (sext (setcc x, y, cc)), -1, setne) -> setcc (x, y, inv(cc))
// setcc (sext (setcc x, y, cc)), -1, seteq) -> setcc (x, y, cc)
SDValue TopSetCC = N0->getOperand(0);
unsigned N0Opc = N0->getOpcode();
bool SExt = (N0Opc == ISD::SIGN_EXTEND);
if (TopSetCC.getValueType() == MVT::i1 && VT == MVT::i1 &&
TopSetCC.getOpcode() == ISD::SETCC &&
(N0Opc == ISD::ZERO_EXTEND || N0Opc == ISD::SIGN_EXTEND) &&
(isConstFalseVal(N1) ||
isExtendedTrueVal(N1C, N0->getValueType(0), SExt))) {
bool Inverse = (N1C->isZero() && Cond == ISD::SETEQ) ||
(!N1C->isZero() && Cond == ISD::SETNE);
if (!Inverse)
return TopSetCC;
ISD::CondCode InvCond = ISD::getSetCCInverse(
cast<CondCodeSDNode>(TopSetCC.getOperand(2))->get(),
TopSetCC.getOperand(0).getValueType());
return DAG.getSetCC(dl, VT, TopSetCC.getOperand(0),
TopSetCC.getOperand(1),
InvCond);
}
}
}
// If the LHS is '(and load, const)', the RHS is 0, the test is for
// equality or unsigned, and all 1 bits of the const are in the same
// partial word, see if we can shorten the load.
if (DCI.isBeforeLegalize() &&
!ISD::isSignedIntSetCC(Cond) &&
N0.getOpcode() == ISD::AND && C1 == 0 &&
N0.getNode()->hasOneUse() &&
isa<LoadSDNode>(N0.getOperand(0)) &&
N0.getOperand(0).getNode()->hasOneUse() &&
isa<ConstantSDNode>(N0.getOperand(1))) {
auto *Lod = cast<LoadSDNode>(N0.getOperand(0));
APInt bestMask;
unsigned bestWidth = 0, bestOffset = 0;
if (Lod->isSimple() && Lod->isUnindexed() &&
(Lod->getMemoryVT().isByteSized() ||
isPaddedAtMostSignificantBitsWhenStored(Lod->getMemoryVT()))) {
unsigned memWidth = Lod->getMemoryVT().getStoreSizeInBits();
unsigned origWidth = N0.getValueSizeInBits();
unsigned maskWidth = origWidth;
// We can narrow (e.g.) 16-bit extending loads on 32-bit target to
// 8 bits, but have to be careful...
if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
origWidth = Lod->getMemoryVT().getSizeInBits();
const APInt &Mask = N0.getConstantOperandAPInt(1);
// Only consider power-of-2 widths (and at least one byte) as candiates
// for the narrowed load.
for (unsigned width = 8; width < origWidth; width *= 2) {
EVT newVT = EVT::getIntegerVT(*DAG.getContext(), width);
if (!shouldReduceLoadWidth(Lod, ISD::NON_EXTLOAD, newVT))
continue;
APInt newMask = APInt::getLowBitsSet(maskWidth, width);
// Avoid accessing any padding here for now (we could use memWidth
// instead of origWidth here otherwise).
unsigned maxOffset = origWidth - width;
for (unsigned offset = 0; offset <= maxOffset; offset += 8) {
if (Mask.isSubsetOf(newMask)) {
unsigned ptrOffset =
Layout.isLittleEndian() ? offset : memWidth - width - offset;
unsigned IsFast = 0;
Align NewAlign = commonAlignment(Lod->getAlign(), ptrOffset / 8);
if (allowsMemoryAccess(
*DAG.getContext(), Layout, newVT, Lod->getAddressSpace(),
NewAlign, Lod->getMemOperand()->getFlags(), &IsFast) &&
IsFast) {
bestOffset = ptrOffset / 8;
bestMask = Mask.lshr(offset);
bestWidth = width;
break;
}
}
newMask <<= 8;
}
if (bestWidth)
break;
}
}
if (bestWidth) {
EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth);
SDValue Ptr = Lod->getBasePtr();
if (bestOffset != 0)
Ptr = DAG.getObjectPtrOffset(dl, Ptr, TypeSize::getFixed(bestOffset));
SDValue NewLoad =
DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
Lod->getPointerInfo().getWithOffset(bestOffset),
Lod->getOriginalAlign());
SDValue And =
DAG.getNode(ISD::AND, dl, newVT, NewLoad,
DAG.getConstant(bestMask.trunc(bestWidth), dl, newVT));
return DAG.getSetCC(dl, VT, And, DAG.getConstant(0LL, dl, newVT), Cond);
}
}
// If the LHS is a ZERO_EXTEND, perform the comparison on the input.
if (N0.getOpcode() == ISD::ZERO_EXTEND) {
unsigned InSize = N0.getOperand(0).getValueSizeInBits();
// If the comparison constant has bits in the upper part, the
// zero-extended value could never match.
if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
C1.getBitWidth() - InSize))) {
switch (Cond) {
case ISD::SETUGT:
case ISD::SETUGE:
case ISD::SETEQ:
return DAG.getConstant(0, dl, VT);
case ISD::SETULT:
case ISD::SETULE:
case ISD::SETNE:
return DAG.getConstant(1, dl, VT);
case ISD::SETGT:
case ISD::SETGE:
// True if the sign bit of C1 is set.
return DAG.getConstant(C1.isNegative(), dl, VT);
case ISD::SETLT:
case ISD::SETLE:
// True if the sign bit of C1 isn't set.
return DAG.getConstant(C1.isNonNegative(), dl, VT);
default:
break;
}
}
// Otherwise, we can perform the comparison with the low bits.
switch (Cond) {
case ISD::SETEQ:
case ISD::SETNE:
case ISD::SETUGT:
case ISD::SETUGE:
case ISD::SETULT:
case ISD::SETULE: {
EVT newVT = N0.getOperand(0).getValueType();
if (DCI.isBeforeLegalizeOps() ||
(isOperationLegal(ISD::SETCC, newVT) &&
isCondCodeLegal(Cond, newVT.getSimpleVT()))) {
EVT NewSetCCVT = getSetCCResultType(Layout, *DAG.getContext(), newVT);
SDValue NewConst = DAG.getConstant(C1.trunc(InSize), dl, newVT);
SDValue NewSetCC = DAG.getSetCC(dl, NewSetCCVT, N0.getOperand(0),
NewConst, Cond);
return DAG.getBoolExtOrTrunc(NewSetCC, dl, VT, N0.getValueType());
}
break;
}
default:
break; // todo, be more careful with signed comparisons
}
} else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
(Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
!isSExtCheaperThanZExt(cast<VTSDNode>(N0.getOperand(1))->getVT(),
OpVT)) {
EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
EVT ExtDstTy = N0.getValueType();
unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
// If the constant doesn't fit into the number of bits for the source of
// the sign extension, it is impossible for both sides to be equal.
if (C1.getSignificantBits() > ExtSrcTyBits)
return DAG.getBoolConstant(Cond == ISD::SETNE, dl, VT, OpVT);
assert(ExtDstTy == N0.getOperand(0).getValueType() &&
ExtDstTy != ExtSrcTy && "Unexpected types!");
APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
SDValue ZextOp = DAG.getNode(ISD::AND, dl, ExtDstTy, N0.getOperand(0),
DAG.getConstant(Imm, dl, ExtDstTy));
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(ZextOp.getNode());
// Otherwise, make this a use of a zext.
return DAG.getSetCC(dl, VT, ZextOp,
DAG.getConstant(C1 & Imm, dl, ExtDstTy), Cond);
} else if ((N1C->isZero() || N1C->isOne()) &&
(Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
// SETCC (X), [0|1], [EQ|NE] -> X if X is known 0/1. i1 types are
// excluded as they are handled below whilst checking for foldBooleans.
if ((N0.getOpcode() == ISD::SETCC || VT.getScalarType() != MVT::i1) &&
isTypeLegal(VT) && VT.bitsLE(N0.getValueType()) &&
(N0.getValueType() == MVT::i1 ||
getBooleanContents(N0.getValueType()) == ZeroOrOneBooleanContent) &&
DAG.MaskedValueIsZero(
N0, APInt::getBitsSetFrom(N0.getValueSizeInBits(), 1))) {
bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (!N1C->isOne());
if (TrueWhenTrue)
return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
// Invert the condition.
if (N0.getOpcode() == ISD::SETCC) {
ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
CC = ISD::getSetCCInverse(CC, N0.getOperand(0).getValueType());
if (DCI.isBeforeLegalizeOps() ||
isCondCodeLegal(CC, N0.getOperand(0).getSimpleValueType()))
return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
}
}
if ((N0.getOpcode() == ISD::XOR ||
(N0.getOpcode() == ISD::AND &&
N0.getOperand(0).getOpcode() == ISD::XOR &&
N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
isOneConstant(N0.getOperand(1))) {
// If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
// can only do this if the top bits are known zero.
unsigned BitWidth = N0.getValueSizeInBits();
if (DAG.MaskedValueIsZero(N0,
APInt::getHighBitsSet(BitWidth,
BitWidth-1))) {
// Okay, get the un-inverted input value.
SDValue Val;
if (N0.getOpcode() == ISD::XOR) {
Val = N0.getOperand(0);
} else {
assert(N0.getOpcode() == ISD::AND &&
N0.getOperand(0).getOpcode() == ISD::XOR);
// ((X^1)&1)^1 -> X & 1
Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
N0.getOperand(0).getOperand(0),
N0.getOperand(1));
}
return DAG.getSetCC(dl, VT, Val, N1,
Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
}
} else if (N1C->isOne()) {
SDValue Op0 = N0;
if (Op0.getOpcode() == ISD::TRUNCATE)
Op0 = Op0.getOperand(0);
if ((Op0.getOpcode() == ISD::XOR) &&
Op0.getOperand(0).getOpcode() == ISD::SETCC &&
Op0.getOperand(1).getOpcode() == ISD::SETCC) {
SDValue XorLHS = Op0.getOperand(0);
SDValue XorRHS = Op0.getOperand(1);
// Ensure that the input setccs return an i1 type or 0/1 value.
if (Op0.getValueType() == MVT::i1 ||
(getBooleanContents(XorLHS.getOperand(0).getValueType()) ==
ZeroOrOneBooleanContent &&
getBooleanContents(XorRHS.getOperand(0).getValueType()) ==
ZeroOrOneBooleanContent)) {
// (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
return DAG.getSetCC(dl, VT, XorLHS, XorRHS, Cond);
}
}
if (Op0.getOpcode() == ISD::AND && isOneConstant(Op0.getOperand(1))) {
// If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
if (Op0.getValueType().bitsGT(VT))
Op0 = DAG.getNode(ISD::AND, dl, VT,
DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
DAG.getConstant(1, dl, VT));
else if (Op0.getValueType().bitsLT(VT))
Op0 = DAG.getNode(ISD::AND, dl, VT,
DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
DAG.getConstant(1, dl, VT));
return DAG.getSetCC(dl, VT, Op0,
DAG.getConstant(0, dl, Op0.getValueType()),
Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
}
if (Op0.getOpcode() == ISD::AssertZext &&
cast<VTSDNode>(Op0.getOperand(1))->getVT() == MVT::i1)
return DAG.getSetCC(dl, VT, Op0,
DAG.getConstant(0, dl, Op0.getValueType()),
Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
}
}
// Given:
// icmp eq/ne (urem %x, %y), 0
// Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
// icmp eq/ne %x, 0
if (N0.getOpcode() == ISD::UREM && N1C->isZero() &&
(Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
KnownBits XKnown = DAG.computeKnownBits(N0.getOperand(0));
KnownBits YKnown = DAG.computeKnownBits(N0.getOperand(1));
if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
return DAG.getSetCC(dl, VT, N0.getOperand(0), N1, Cond);
}
// Fold set_cc seteq (ashr X, BW-1), -1 -> set_cc setlt X, 0
// and set_cc setne (ashr X, BW-1), -1 -> set_cc setge X, 0
if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
N0.getOpcode() == ISD::SRA && isa<ConstantSDNode>(N0.getOperand(1)) &&
N0.getConstantOperandAPInt(1) == OpVT.getScalarSizeInBits() - 1 &&
N1C && N1C->isAllOnes()) {
return DAG.getSetCC(dl, VT, N0.getOperand(0),
DAG.getConstant(0, dl, OpVT),
Cond == ISD::SETEQ ? ISD::SETLT : ISD::SETGE);
}
if (SDValue V =
optimizeSetCCOfSignedTruncationCheck(VT, N0, N1, Cond, DCI, dl))
return V;
}
// These simplifications apply to splat vectors as well.
// TODO: Handle more splat vector cases.
if (auto *N1C = isConstOrConstSplat(N1)) {
const APInt &C1 = N1C->getAPIntValue();
APInt MinVal, MaxVal;
unsigned OperandBitSize = N1C->getValueType(0).getScalarSizeInBits();
if (ISD::isSignedIntSetCC(Cond)) {
MinVal = APInt::getSignedMinValue(OperandBitSize);
MaxVal = APInt::getSignedMaxValue(OperandBitSize);
} else {
MinVal = APInt::getMinValue(OperandBitSize);
MaxVal = APInt::getMaxValue(OperandBitSize);
}
// Canonicalize GE/LE comparisons to use GT/LT comparisons.
if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
// X >= MIN --> true
if (C1 == MinVal)
return DAG.getBoolConstant(true, dl, VT, OpVT);
if (!VT.isVector()) { // TODO: Support this for vectors.
// X >= C0 --> X > (C0 - 1)
APInt C = C1 - 1;
ISD::CondCode NewCC = (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT;
if ((DCI.isBeforeLegalizeOps() ||
isCondCodeLegal(NewCC, VT.getSimpleVT())) &&
(!N1C->isOpaque() || (C.getBitWidth() <= 64 &&
isLegalICmpImmediate(C.getSExtValue())))) {
return DAG.getSetCC(dl, VT, N0,
DAG.getConstant(C, dl, N1.getValueType()),
NewCC);
}
}
}
if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
// X <= MAX --> true
if (C1 == MaxVal)
return DAG.getBoolConstant(true, dl, VT, OpVT);
// X <= C0 --> X < (C0 + 1)
if (!VT.isVector()) { // TODO: Support this for vectors.
APInt C = C1 + 1;
ISD::CondCode NewCC = (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT;
if ((DCI.isBeforeLegalizeOps() ||
isCondCodeLegal(NewCC, VT.getSimpleVT())) &&
(!N1C->isOpaque() || (C.getBitWidth() <= 64 &&
isLegalICmpImmediate(C.getSExtValue())))) {
return DAG.getSetCC(dl, VT, N0,
DAG.getConstant(C, dl, N1.getValueType()),
NewCC);
}
}
}
if (Cond == ISD::SETLT || Cond == ISD::SETULT) {
if (C1 == MinVal)
return DAG.getBoolConstant(false, dl, VT, OpVT); // X < MIN --> false
// TODO: Support this for vectors after legalize ops.
if (!VT.isVector() || DCI.isBeforeLegalizeOps()) {
// Canonicalize setlt X, Max --> setne X, Max
if (C1 == MaxVal)
return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
// If we have setult X, 1, turn it into seteq X, 0
if (C1 == MinVal+1)
return DAG.getSetCC(dl, VT, N0,
DAG.getConstant(MinVal, dl, N0.getValueType()),
ISD::SETEQ);
}
}
if (Cond == ISD::SETGT || Cond == ISD::SETUGT) {
if (C1 == MaxVal)
return DAG.getBoolConstant(false, dl, VT, OpVT); // X > MAX --> false
// TODO: Support this for vectors after legalize ops.
if (!VT.isVector() || DCI.isBeforeLegalizeOps()) {
// Canonicalize setgt X, Min --> setne X, Min
if (C1 == MinVal)
return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
// If we have setugt X, Max-1, turn it into seteq X, Max
if (C1 == MaxVal-1)
return DAG.getSetCC(dl, VT, N0,
DAG.getConstant(MaxVal, dl, N0.getValueType()),
ISD::SETEQ);
}
}
if (Cond == ISD::SETEQ || Cond == ISD::SETNE) {
// (X & (C l>>/<< Y)) ==/!= 0 --> ((X <</l>> Y) & C) ==/!= 0
if (C1.isZero())
if (SDValue CC = optimizeSetCCByHoistingAndByConstFromLogicalShift(
VT, N0, N1, Cond, DCI, dl))
return CC;
// For all/any comparisons, replace or(x,shl(y,bw/2)) with and/or(x,y).
// For example, when high 32-bits of i64 X are known clear:
// all bits clear: (X | (Y<<32)) == 0 --> (X | Y) == 0
// all bits set: (X | (Y<<32)) == -1 --> (X & Y) == -1
bool CmpZero = N1C->isZero();
bool CmpNegOne = N1C->isAllOnes();
if ((CmpZero || CmpNegOne) && N0.hasOneUse()) {
// Match or(lo,shl(hi,bw/2)) pattern.
auto IsConcat = [&](SDValue V, SDValue &Lo, SDValue &Hi) {
unsigned EltBits = V.getScalarValueSizeInBits();
if (V.getOpcode() != ISD::OR || (EltBits % 2) != 0)
return false;
SDValue LHS = V.getOperand(0);
SDValue RHS = V.getOperand(1);
APInt HiBits = APInt::getHighBitsSet(EltBits, EltBits / 2);
// Unshifted element must have zero upperbits.
if (RHS.getOpcode() == ISD::SHL &&
isa<ConstantSDNode>(RHS.getOperand(1)) &&
RHS.getConstantOperandAPInt(1) == (EltBits / 2) &&
DAG.MaskedValueIsZero(LHS, HiBits)) {
Lo = LHS;
Hi = RHS.getOperand(0);
return true;
}
if (LHS.getOpcode() == ISD::SHL &&
isa<ConstantSDNode>(LHS.getOperand(1)) &&
LHS.getConstantOperandAPInt(1) == (EltBits / 2) &&
DAG.MaskedValueIsZero(RHS, HiBits)) {
Lo = RHS;
Hi = LHS.getOperand(0);
return true;
}
return false;
};
auto MergeConcat = [&](SDValue Lo, SDValue Hi) {
unsigned EltBits = N0.getScalarValueSizeInBits();
unsigned HalfBits = EltBits / 2;
APInt HiBits = APInt::getHighBitsSet(EltBits, HalfBits);
SDValue LoBits = DAG.getConstant(~HiBits, dl, OpVT);
SDValue HiMask = DAG.getNode(ISD::AND, dl, OpVT, Hi, LoBits);
SDValue NewN0 =
DAG.getNode(CmpZero ? ISD::OR : ISD::AND, dl, OpVT, Lo, HiMask);
SDValue NewN1 = CmpZero ? DAG.getConstant(0, dl, OpVT) : LoBits;
return DAG.getSetCC(dl, VT, NewN0, NewN1, Cond);
};
SDValue Lo, Hi;
if (IsConcat(N0, Lo, Hi))
return MergeConcat(Lo, Hi);
if (N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR) {
SDValue Lo0, Lo1, Hi0, Hi1;
if (IsConcat(N0.getOperand(0), Lo0, Hi0) &&
IsConcat(N0.getOperand(1), Lo1, Hi1)) {
return MergeConcat(DAG.getNode(N0.getOpcode(), dl, OpVT, Lo0, Lo1),
DAG.getNode(N0.getOpcode(), dl, OpVT, Hi0, Hi1));
}
}
}
}
// If we have "setcc X, C0", check to see if we can shrink the immediate
// by changing cc.
// TODO: Support this for vectors after legalize ops.
if (!VT.isVector() || DCI.isBeforeLegalizeOps()) {
// SETUGT X, SINTMAX -> SETLT X, 0
// SETUGE X, SINTMIN -> SETLT X, 0
if ((Cond == ISD::SETUGT && C1.isMaxSignedValue()) ||
(Cond == ISD::SETUGE && C1.isMinSignedValue()))
return DAG.getSetCC(dl, VT, N0,
DAG.getConstant(0, dl, N1.getValueType()),
ISD::SETLT);
// SETULT X, SINTMIN -> SETGT X, -1
// SETULE X, SINTMAX -> SETGT X, -1
if ((Cond == ISD::SETULT && C1.isMinSignedValue()) ||
(Cond == ISD::SETULE && C1.isMaxSignedValue()))
return DAG.getSetCC(dl, VT, N0,
DAG.getAllOnesConstant(dl, N1.getValueType()),
ISD::SETGT);
}
}
// Back to non-vector simplifications.
// TODO: Can we do these for vector splats?
if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
const APInt &C1 = N1C->getAPIntValue();
EVT ShValTy = N0.getValueType();
// Fold bit comparisons when we can. This will result in an
// incorrect value when boolean false is negative one, unless
// the bitsize is 1 in which case the false value is the same
// in practice regardless of the representation.
if ((VT.getSizeInBits() == 1 ||
getBooleanContents(N0.getValueType()) == ZeroOrOneBooleanContent) &&
(Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
(VT == ShValTy || (isTypeLegal(VT) && VT.bitsLE(ShValTy))) &&
N0.getOpcode() == ISD::AND) {
if (auto *AndRHS = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
// Perform the xform if the AND RHS is a single bit.
unsigned ShCt = AndRHS->getAPIntValue().logBase2();
if (AndRHS->getAPIntValue().isPowerOf2() &&
!TLI.shouldAvoidTransformToShift(ShValTy, ShCt)) {
return DAG.getNode(
ISD::TRUNCATE, dl, VT,
DAG.getNode(ISD::SRL, dl, ShValTy, N0,
DAG.getShiftAmountConstant(ShCt, ShValTy, dl)));
}
} else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
// (X & 8) == 8 --> (X & 8) >> 3
// Perform the xform if C1 is a single bit.
unsigned ShCt = C1.logBase2();
if (C1.isPowerOf2() &&
!TLI.shouldAvoidTransformToShift(ShValTy, ShCt)) {
return DAG.getNode(
ISD::TRUNCATE, dl, VT,
DAG.getNode(ISD::SRL, dl, ShValTy, N0,
DAG.getShiftAmountConstant(ShCt, ShValTy, dl)));
}
}
}
}
if (C1.getSignificantBits() <= 64 &&
!isLegalICmpImmediate(C1.getSExtValue())) {
// (X & -256) == 256 -> (X >> 8) == 1
if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
N0.getOpcode() == ISD::AND && N0.hasOneUse()) {
if (auto *AndRHS = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
const APInt &AndRHSC = AndRHS->getAPIntValue();
if (AndRHSC.isNegatedPowerOf2() && (AndRHSC & C1) == C1) {
unsigned ShiftBits = AndRHSC.countr_zero();
if (!TLI.shouldAvoidTransformToShift(ShValTy, ShiftBits)) {
SDValue Shift = DAG.getNode(
ISD::SRL, dl, ShValTy, N0.getOperand(0),
DAG.getShiftAmountConstant(ShiftBits, ShValTy, dl));
SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), dl, ShValTy);
return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond);
}
}
}
} else if (Cond == ISD::SETULT || Cond == ISD::SETUGE ||
Cond == ISD::SETULE || Cond == ISD::SETUGT) {
bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT);
// X < 0x100000000 -> (X >> 32) < 1
// X >= 0x100000000 -> (X >> 32) >= 1
// X <= 0x0ffffffff -> (X >> 32) < 1
// X > 0x0ffffffff -> (X >> 32) >= 1
unsigned ShiftBits;
APInt NewC = C1;
ISD::CondCode NewCond = Cond;
if (AdjOne) {
ShiftBits = C1.countr_one();
NewC = NewC + 1;
NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
} else {
ShiftBits = C1.countr_zero();
}
NewC.lshrInPlace(ShiftBits);
if (ShiftBits && NewC.getSignificantBits() <= 64 &&
isLegalICmpImmediate(NewC.getSExtValue()) &&
!TLI.shouldAvoidTransformToShift(ShValTy, ShiftBits)) {
SDValue Shift =
DAG.getNode(ISD::SRL, dl, ShValTy, N0,
DAG.getShiftAmountConstant(ShiftBits, ShValTy, dl));
SDValue CmpRHS = DAG.getConstant(NewC, dl, ShValTy);
return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond);
}
}
}
}
if (!isa<ConstantFPSDNode>(N0) && isa<ConstantFPSDNode>(N1)) {
auto *CFP = cast<ConstantFPSDNode>(N1);
assert(!CFP->getValueAPF().isNaN() && "Unexpected NaN value");
// Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
// constant if knowing that the operand is non-nan is enough. We prefer to
// have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
// materialize 0.0.
if (Cond == ISD::SETO || Cond == ISD::SETUO)
return DAG.getSetCC(dl, VT, N0, N0, Cond);
// setcc (fneg x), C -> setcc swap(pred) x, -C
if (N0.getOpcode() == ISD::FNEG) {
ISD::CondCode SwapCond = ISD::getSetCCSwappedOperands(Cond);
if (DCI.isBeforeLegalizeOps() ||
isCondCodeLegal(SwapCond, N0.getSimpleValueType())) {
SDValue NegN1 = DAG.getNode(ISD::FNEG, dl, N0.getValueType(), N1);
return DAG.getSetCC(dl, VT, N0.getOperand(0), NegN1, SwapCond);
}
}
// setueq/setoeq X, (fabs Inf) -> is_fpclass X, fcInf
if (isOperationLegalOrCustom(ISD::IS_FPCLASS, N0.getValueType()) &&
!isFPImmLegal(CFP->getValueAPF(), CFP->getValueType(0))) {
bool IsFabs = N0.getOpcode() == ISD::FABS;
SDValue Op = IsFabs ? N0.getOperand(0) : N0;
if ((Cond == ISD::SETOEQ || Cond == ISD::SETUEQ) && CFP->isInfinity()) {
FPClassTest Flag = CFP->isNegative() ? (IsFabs ? fcNone : fcNegInf)
: (IsFabs ? fcInf : fcPosInf);
if (Cond == ISD::SETUEQ)
Flag |= fcNan;
return DAG.getNode(ISD::IS_FPCLASS, dl, VT, Op,
DAG.getTargetConstant(Flag, dl, MVT::i32));
}
}
// If the condition is not legal, see if we can find an equivalent one
// which is legal.
if (!isCondCodeLegal(Cond, N0.getSimpleValueType())) {
// If the comparison was an awkward floating-point == or != and one of
// the comparison operands is infinity or negative infinity, convert the
// condition to a less-awkward <= or >=.
if (CFP->getValueAPF().isInfinity()) {
bool IsNegInf = CFP->getValueAPF().isNegative();
ISD::CondCode NewCond = ISD::SETCC_INVALID;
switch (Cond) {
case ISD::SETOEQ: NewCond = IsNegInf ? ISD::SETOLE : ISD::SETOGE; break;
case ISD::SETUEQ: NewCond = IsNegInf ? ISD::SETULE : ISD::SETUGE; break;
case ISD::SETUNE: NewCond = IsNegInf ? ISD::SETUGT : ISD::SETULT; break;
case ISD::SETONE: NewCond = IsNegInf ? ISD::SETOGT : ISD::SETOLT; break;
default: break;
}
if (NewCond != ISD::SETCC_INVALID &&
isCondCodeLegal(NewCond, N0.getSimpleValueType()))
return DAG.getSetCC(dl, VT, N0, N1, NewCond);
}
}
}
if (N0 == N1) {
// The sext(setcc()) => setcc() optimization relies on the appropriate
// constant being emitted.
assert(!N0.getValueType().isInteger() &&
"Integer types should be handled by FoldSetCC");
bool EqTrue = ISD::isTrueWhenEqual(Cond);
unsigned UOF = ISD::getUnorderedFlavor(Cond);
if (UOF == 2) // FP operators that are undefined on NaNs.
return DAG.getBoolConstant(EqTrue, dl, VT, OpVT);
if (UOF == unsigned(EqTrue))
return DAG.getBoolConstant(EqTrue, dl, VT, OpVT);
// Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
// if it is not already.
ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
if (NewCond != Cond &&
(DCI.isBeforeLegalizeOps() ||
isCondCodeLegal(NewCond, N0.getSimpleValueType())))
return DAG.getSetCC(dl, VT, N0, N1, NewCond);
}
// ~X > ~Y --> Y > X
// ~X < ~Y --> Y < X
// ~X < C --> X > ~C
// ~X > C --> X < ~C
if ((isSignedIntSetCC(Cond) || isUnsignedIntSetCC(Cond)) &&
N0.getValueType().isInteger()) {
if (isBitwiseNot(N0)) {
if (isBitwiseNot(N1))
return DAG.getSetCC(dl, VT, N1.getOperand(0), N0.getOperand(0), Cond);
if (DAG.isConstantIntBuildVectorOrConstantInt(N1) &&
!DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(0))) {
SDValue Not = DAG.getNOT(dl, N1, OpVT);
return DAG.getSetCC(dl, VT, Not, N0.getOperand(0), Cond);
}
}
}
if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
N0.getValueType().isInteger()) {
if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
N0.getOpcode() == ISD::XOR) {
// Simplify (X+Y) == (X+Z) --> Y == Z
if (N0.getOpcode() == N1.getOpcode()) {
if (N0.getOperand(0) == N1.getOperand(0))
return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
if (N0.getOperand(1) == N1.getOperand(1))
return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
if (isCommutativeBinOp(N0.getOpcode())) {
// If X op Y == Y op X, try other combinations.
if (N0.getOperand(0) == N1.getOperand(1))
return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
Cond);
if (N0.getOperand(1) == N1.getOperand(0))
return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
Cond);
}
}
// If RHS is a legal immediate value for a compare instruction, we need
// to be careful about increasing register pressure needlessly.
bool LegalRHSImm = false;
if (auto *RHSC = dyn_cast<ConstantSDNode>(N1)) {
if (auto *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
// Turn (X+C1) == C2 --> X == C2-C1
if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse())
return DAG.getSetCC(
dl, VT, N0.getOperand(0),
DAG.getConstant(RHSC->getAPIntValue() - LHSR->getAPIntValue(),
dl, N0.getValueType()),
Cond);
// Turn (X^C1) == C2 --> X == C1^C2
if (N0.getOpcode() == ISD::XOR && N0.getNode()->hasOneUse())
return DAG.getSetCC(
dl, VT, N0.getOperand(0),
DAG.getConstant(LHSR->getAPIntValue() ^ RHSC->getAPIntValue(),
dl, N0.getValueType()),
Cond);
}
// Turn (C1-X) == C2 --> X == C1-C2
if (auto *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse())
return DAG.getSetCC(
dl, VT, N0.getOperand(1),
DAG.getConstant(SUBC->getAPIntValue() - RHSC->getAPIntValue(),
dl, N0.getValueType()),
Cond);
// Could RHSC fold directly into a compare?
if (RHSC->getValueType(0).getSizeInBits() <= 64)
LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue());
}
// (X+Y) == X --> Y == 0 and similar folds.
// Don't do this if X is an immediate that can fold into a cmp
// instruction and X+Y has other uses. It could be an induction variable
// chain, and the transform would increase register pressure.
if (!LegalRHSImm || N0.hasOneUse())
if (SDValue V = foldSetCCWithBinOp(VT, N0, N1, Cond, dl, DCI))
return V;
}
if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
N1.getOpcode() == ISD::XOR)
if (SDValue V = foldSetCCWithBinOp(VT, N1, N0, Cond, dl, DCI))
return V;
if (SDValue V = foldSetCCWithAnd(VT, N0, N1, Cond, dl, DCI))
return V;
}
// Fold remainder of division by a constant.
if ((N0.getOpcode() == ISD::UREM || N0.getOpcode() == ISD::SREM) &&
N0.hasOneUse() && (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
// When division is cheap or optimizing for minimum size,
// fall through to DIVREM creation by skipping this fold.
if (!isIntDivCheap(VT, Attr) && !Attr.hasFnAttr(Attribute::MinSize)) {
if (N0.getOpcode() == ISD::UREM) {
if (SDValue Folded = buildUREMEqFold(VT, N0, N1, Cond, DCI, dl))
return Folded;
} else if (N0.getOpcode() == ISD::SREM) {
if (SDValue Folded = buildSREMEqFold(VT, N0, N1, Cond, DCI, dl))
return Folded;
}
}
}
// Fold away ALL boolean setcc's.
if (N0.getValueType().getScalarType() == MVT::i1 && foldBooleans) {
SDValue Temp;
switch (Cond) {
default: llvm_unreachable("Unknown integer setcc!");
case ISD::SETEQ: // X == Y -> ~(X^Y)
Temp = DAG.getNode(ISD::XOR, dl, OpVT, N0, N1);
N0 = DAG.getNOT(dl, Temp, OpVT);
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(Temp.getNode());
break;
case ISD::SETNE: // X != Y --> (X^Y)
N0 = DAG.getNode(ISD::XOR, dl, OpVT, N0, N1);
break;
case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
Temp = DAG.getNOT(dl, N0, OpVT);
N0 = DAG.getNode(ISD::AND, dl, OpVT, N1, Temp);
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(Temp.getNode());
break;
case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
Temp = DAG.getNOT(dl, N1, OpVT);
N0 = DAG.getNode(ISD::AND, dl, OpVT, N0, Temp);
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(Temp.getNode());
break;
case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
Temp = DAG.getNOT(dl, N0, OpVT);
N0 = DAG.getNode(ISD::OR, dl, OpVT, N1, Temp);
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(Temp.getNode());
break;
case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
Temp = DAG.getNOT(dl, N1, OpVT);
N0 = DAG.getNode(ISD::OR, dl, OpVT, N0, Temp);
break;
}
if (VT.getScalarType() != MVT::i1) {
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(N0.getNode());
// FIXME: If running after legalize, we probably can't do this.
ISD::NodeType ExtendCode = getExtendForContent(getBooleanContents(OpVT));
N0 = DAG.getNode(ExtendCode, dl, VT, N0);
}
return N0;
}
// Could not fold it.
return SDValue();
}
/// Returns true (and the GlobalValue and the offset) if the node is a
/// GlobalAddress + offset.
bool TargetLowering::isGAPlusOffset(SDNode *WN, const GlobalValue *&GA,
int64_t &Offset) const {
SDNode *N = unwrapAddress(SDValue(WN, 0)).getNode();
if (auto *GASD = dyn_cast<GlobalAddressSDNode>(N)) {
GA = GASD->getGlobal();
Offset += GASD->getOffset();
return true;
}
if (N->getOpcode() == ISD::ADD) {
SDValue N1 = N->getOperand(0);
SDValue N2 = N->getOperand(1);
if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
if (auto *V = dyn_cast<ConstantSDNode>(N2)) {
Offset += V->getSExtValue();
return true;
}
} else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
if (auto *V = dyn_cast<ConstantSDNode>(N1)) {
Offset += V->getSExtValue();
return true;
}
}
}
return false;
}
SDValue TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
// Default implementation: no optimization.
return SDValue();
}
//===----------------------------------------------------------------------===//
// Inline Assembler Implementation Methods
//===----------------------------------------------------------------------===//
TargetLowering::ConstraintType
TargetLowering::getConstraintType(StringRef Constraint) const {
unsigned S = Constraint.size();
if (S == 1) {
switch (Constraint[0]) {
default: break;
case 'r':
return C_RegisterClass;
case 'm': // memory
case 'o': // offsetable
case 'V': // not offsetable
return C_Memory;
case 'p': // Address.
return C_Address;
case 'n': // Simple Integer
case 'E': // Floating Point Constant
case 'F': // Floating Point Constant
return C_Immediate;
case 'i': // Simple Integer or Relocatable Constant
case 's': // Relocatable Constant
case 'X': // Allow ANY value.
case 'I': // Target registers.
case 'J':
case 'K':
case 'L':
case 'M':
case 'N':
case 'O':
case 'P':
case '<':
case '>':
return C_Other;
}
}
if (S > 1 && Constraint[0] == '{' && Constraint[S - 1] == '}') {
if (S == 8 && Constraint.substr(1, 6) == "memory") // "{memory}"
return C_Memory;
return C_Register;
}
return C_Unknown;
}
/// Try to replace an X constraint, which matches anything, with another that
/// has more specific requirements based on the type of the corresponding
/// operand.
const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const {
if (ConstraintVT.isInteger())
return "r";
if (ConstraintVT.isFloatingPoint())
return "f"; // works for many targets
return nullptr;
}
SDValue TargetLowering::LowerAsmOutputForConstraint(
SDValue &Chain, SDValue &Glue, const SDLoc &DL,
const AsmOperandInfo &OpInfo, SelectionDAG &DAG) const {
return SDValue();
}
/// Lower the specified operand into the Ops vector.
/// If it is invalid, don't add anything to Ops.
void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
StringRef Constraint,
std::vector<SDValue> &Ops,
SelectionDAG &DAG) const {
if (Constraint.size() > 1)
return;
char ConstraintLetter = Constraint[0];
switch (ConstraintLetter) {
default: break;
case 'X': // Allows any operand
case 'i': // Simple Integer or Relocatable Constant
case 'n': // Simple Integer
case 's': { // Relocatable Constant
ConstantSDNode *C;
uint64_t Offset = 0;
// Match (GA) or (C) or (GA+C) or (GA-C) or ((GA+C)+C) or (((GA+C)+C)+C),
// etc., since getelementpointer is variadic. We can't use
// SelectionDAG::FoldSymbolOffset because it expects the GA to be accessible
// while in this case the GA may be furthest from the root node which is
// likely an ISD::ADD.
while (true) {
if ((C = dyn_cast<ConstantSDNode>(Op)) && ConstraintLetter != 's') {
// gcc prints these as sign extended. Sign extend value to 64 bits
// now; without this it would get ZExt'd later in
// ScheduleDAGSDNodes::EmitNode, which is very generic.
bool IsBool = C->getConstantIntValue()->getBitWidth() == 1;
BooleanContent BCont = getBooleanContents(MVT::i64);
ISD::NodeType ExtOpc =
IsBool ? getExtendForContent(BCont) : ISD::SIGN_EXTEND;
int64_t ExtVal =
ExtOpc == ISD::ZERO_EXTEND ? C->getZExtValue() : C->getSExtValue();
Ops.push_back(
DAG.getTargetConstant(Offset + ExtVal, SDLoc(C), MVT::i64));
return;
}
if (ConstraintLetter != 'n') {
if (const auto *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), SDLoc(Op),
GA->getValueType(0),
Offset + GA->getOffset()));
return;
}
if (const auto *BA = dyn_cast<BlockAddressSDNode>(Op)) {
Ops.push_back(DAG.getTargetBlockAddress(
BA->getBlockAddress(), BA->getValueType(0),
Offset + BA->getOffset(), BA->getTargetFlags()));
return;
}
if (isa<BasicBlockSDNode>(Op)) {
Ops.push_back(Op);
return;
}
}
const unsigned OpCode = Op.getOpcode();
if (OpCode == ISD::ADD || OpCode == ISD::SUB) {
if ((C = dyn_cast<ConstantSDNode>(Op.getOperand(0))))
Op = Op.getOperand(1);
// Subtraction is not commutative.
else if (OpCode == ISD::ADD &&
(C = dyn_cast<ConstantSDNode>(Op.getOperand(1))))
Op = Op.getOperand(0);
else
return;
Offset += (OpCode == ISD::ADD ? 1 : -1) * C->getSExtValue();
continue;
}
return;
}
break;
}
}
}
void TargetLowering::CollectTargetIntrinsicOperands(
const CallInst &I, SmallVectorImpl<SDValue> &Ops, SelectionDAG &DAG) const {
}
std::pair<unsigned, const TargetRegisterClass *>
TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *RI,
StringRef Constraint,
MVT VT) const {
if (!Constraint.starts_with("{"))
return std::make_pair(0u, static_cast<TargetRegisterClass *>(nullptr));
assert(*(Constraint.end() - 1) == '}' && "Not a brace enclosed constraint?");
// Remove the braces from around the name.
StringRef RegName(Constraint.data() + 1, Constraint.size() - 2);
std::pair<unsigned, const TargetRegisterClass *> R =
std::make_pair(0u, static_cast<const TargetRegisterClass *>(nullptr));
// Figure out which register class contains this reg.
for (const TargetRegisterClass *RC : RI->regclasses()) {
// If none of the value types for this register class are valid, we
// can't use it. For example, 64-bit reg classes on 32-bit targets.
if (!isLegalRC(*RI, *RC))
continue;
for (const MCPhysReg &PR : *RC) {
if (RegName.equals_insensitive(RI->getRegAsmName(PR))) {
std::pair<unsigned, const TargetRegisterClass *> S =
std::make_pair(PR, RC);
// If this register class has the requested value type, return it,
// otherwise keep searching and return the first class found
// if no other is found which explicitly has the requested type.
if (RI->isTypeLegalForClass(*RC, VT))
return S;
if (!R.second)
R = S;
}
}
}
return R;
}
//===----------------------------------------------------------------------===//
// Constraint Selection.
/// Return true of this is an input operand that is a matching constraint like
/// "4".
bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
assert(!ConstraintCode.empty() && "No known constraint!");
return isdigit(static_cast<unsigned char>(ConstraintCode[0]));
}
/// If this is an input matching constraint, this method returns the output
/// operand it matches.
unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
assert(!ConstraintCode.empty() && "No known constraint!");
return atoi(ConstraintCode.c_str());
}
/// Split up the constraint string from the inline assembly value into the
/// specific constraints and their prefixes, and also tie in the associated
/// operand values.
/// If this returns an empty vector, and if the constraint string itself
/// isn't empty, there was an error parsing.
TargetLowering::AsmOperandInfoVector
TargetLowering::ParseConstraints(const DataLayout &DL,
const TargetRegisterInfo *TRI,
const CallBase &Call) const {
/// Information about all of the constraints.
AsmOperandInfoVector ConstraintOperands;
const InlineAsm *IA = cast<InlineAsm>(Call.getCalledOperand());
unsigned maCount = 0; // Largest number of multiple alternative constraints.
// Do a prepass over the constraints, canonicalizing them, and building up the
// ConstraintOperands list.
unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
unsigned ResNo = 0; // ResNo - The result number of the next output.
unsigned LabelNo = 0; // LabelNo - CallBr indirect dest number.
for (InlineAsm::ConstraintInfo &CI : IA->ParseConstraints()) {
ConstraintOperands.emplace_back(std::move(CI));
AsmOperandInfo &OpInfo = ConstraintOperands.back();
// Update multiple alternative constraint count.
if (OpInfo.multipleAlternatives.size() > maCount)
maCount = OpInfo.multipleAlternatives.size();
OpInfo.ConstraintVT = MVT::Other;
// Compute the value type for each operand.
switch (OpInfo.Type) {
case InlineAsm::isOutput:
// Indirect outputs just consume an argument.
if (OpInfo.isIndirect) {
OpInfo.CallOperandVal = Call.getArgOperand(ArgNo);
break;
}
// The return value of the call is this value. As such, there is no
// corresponding argument.
assert(!Call.getType()->isVoidTy() && "Bad inline asm!");
if (auto *STy = dyn_cast<StructType>(Call.getType())) {
OpInfo.ConstraintVT =
getSimpleValueType(DL, STy->getElementType(ResNo));
} else {
assert(ResNo == 0 && "Asm only has one result!");
OpInfo.ConstraintVT =
getAsmOperandValueType(DL, Call.getType()).getSimpleVT();
}
++ResNo;
break;
case InlineAsm::isInput:
OpInfo.CallOperandVal = Call.getArgOperand(ArgNo);
break;
case InlineAsm::isLabel:
OpInfo.CallOperandVal = cast<CallBrInst>(&Call)->getIndirectDest(LabelNo);
++LabelNo;
continue;
case InlineAsm::isClobber:
// Nothing to do.
break;
}
if (OpInfo.CallOperandVal) {
llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
if (OpInfo.isIndirect) {
OpTy = Call.getParamElementType(ArgNo);
assert(OpTy && "Indirect operand must have elementtype attribute");
}
// Look for vector wrapped in a struct. e.g. { <16 x i8> }.
if (StructType *STy = dyn_cast<StructType>(OpTy))
if (STy->getNumElements() == 1)
OpTy = STy->getElementType(0);
// If OpTy is not a single value, it may be a struct/union that we
// can tile with integers.
if (!OpTy->isSingleValueType() && OpTy->isSized()) {
unsigned BitSize = DL.getTypeSizeInBits(OpTy);
switch (BitSize) {
default: break;
case 1:
case 8:
case 16:
case 32:
case 64:
case 128:
OpTy = IntegerType::get(OpTy->getContext(), BitSize);
break;
}
}
EVT VT = getAsmOperandValueType(DL, OpTy, true);
OpInfo.ConstraintVT = VT.isSimple() ? VT.getSimpleVT() : MVT::Other;
ArgNo++;
}
}
// If we have multiple alternative constraints, select the best alternative.
if (!ConstraintOperands.empty()) {
if (maCount) {
unsigned bestMAIndex = 0;
int bestWeight = -1;
// weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
int weight = -1;
unsigned maIndex;
// Compute the sums of the weights for each alternative, keeping track
// of the best (highest weight) one so far.
for (maIndex = 0; maIndex < maCount; ++maIndex) {
int weightSum = 0;
for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
cIndex != eIndex; ++cIndex) {
AsmOperandInfo &OpInfo = ConstraintOperands[cIndex];
if (OpInfo.Type == InlineAsm::isClobber)
continue;
// If this is an output operand with a matching input operand,
// look up the matching input. If their types mismatch, e.g. one
// is an integer, the other is floating point, or their sizes are
// different, flag it as an maCantMatch.
if (OpInfo.hasMatchingInput()) {
AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
if (OpInfo.ConstraintVT != Input.ConstraintVT) {
if ((OpInfo.ConstraintVT.isInteger() !=
Input.ConstraintVT.isInteger()) ||
(OpInfo.ConstraintVT.getSizeInBits() !=
Input.ConstraintVT.getSizeInBits())) {
weightSum = -1; // Can't match.
break;
}
}
}
weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
if (weight == -1) {
weightSum = -1;
break;
}
weightSum += weight;
}
// Update best.
if (weightSum > bestWeight) {
bestWeight = weightSum;
bestMAIndex = maIndex;
}
}
// Now select chosen alternative in each constraint.
for (AsmOperandInfo &cInfo : ConstraintOperands)
if (cInfo.Type != InlineAsm::isClobber)
cInfo.selectAlternative(bestMAIndex);
}
}
// Check and hook up tied operands, choose constraint code to use.
for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
cIndex != eIndex; ++cIndex) {
AsmOperandInfo &OpInfo = ConstraintOperands[cIndex];
// If this is an output operand with a matching input operand, look up the
// matching input. If their types mismatch, e.g. one is an integer, the
// other is floating point, or their sizes are different, flag it as an
// error.
if (OpInfo.hasMatchingInput()) {
AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
if (OpInfo.ConstraintVT != Input.ConstraintVT) {
std::pair<unsigned, const TargetRegisterClass *> MatchRC =
getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode,
OpInfo.ConstraintVT);
std::pair<unsigned, const TargetRegisterClass *> InputRC =
getRegForInlineAsmConstraint(TRI, Input.ConstraintCode,
Input.ConstraintVT);
if ((OpInfo.ConstraintVT.isInteger() !=
Input.ConstraintVT.isInteger()) ||
(MatchRC.second != InputRC.second)) {
report_fatal_error("Unsupported asm: input constraint"
" with a matching output constraint of"
" incompatible type!");
}
}
}
}
return ConstraintOperands;
}
/// Return a number indicating our preference for chosing a type of constraint
/// over another, for the purpose of sorting them. Immediates are almost always
/// preferrable (when they can be emitted). A higher return value means a
/// stronger preference for one constraint type relative to another.
/// FIXME: We should prefer registers over memory but doing so may lead to
/// unrecoverable register exhaustion later.
/// https://github.com/llvm/llvm-project/issues/20571
static unsigned getConstraintPiority(TargetLowering::ConstraintType CT) {
switch (CT) {
case TargetLowering::C_Immediate:
case TargetLowering::C_Other:
return 4;
case TargetLowering::C_Memory:
case TargetLowering::C_Address:
return 3;
case TargetLowering::C_RegisterClass:
return 2;
case TargetLowering::C_Register:
return 1;
case TargetLowering::C_Unknown:
return 0;
}
llvm_unreachable("Invalid constraint type");
}
/// Examine constraint type and operand type and determine a weight value.
/// This object must already have been set up with the operand type
/// and the current alternative constraint selected.
TargetLowering::ConstraintWeight
TargetLowering::getMultipleConstraintMatchWeight(
AsmOperandInfo &info, int maIndex) const {
InlineAsm::ConstraintCodeVector *rCodes;
if (maIndex >= (int)info.multipleAlternatives.size())
rCodes = &info.Codes;
else
rCodes = &info.multipleAlternatives[maIndex].Codes;
ConstraintWeight BestWeight = CW_Invalid;
// Loop over the options, keeping track of the most general one.
for (const std::string &rCode : *rCodes) {
ConstraintWeight weight =
getSingleConstraintMatchWeight(info, rCode.c_str());
if (weight > BestWeight)
BestWeight = weight;
}
return BestWeight;
}
/// Examine constraint type and operand type and determine a weight value.
/// This object must already have been set up with the operand type
/// and the current alternative constraint selected.
TargetLowering::ConstraintWeight
TargetLowering::getSingleConstraintMatchWeight(
AsmOperandInfo &info, const char *constraint) const {
ConstraintWeight weight = CW_Invalid;
Value *CallOperandVal = info.CallOperandVal;
// If we don't have a value, we can't do a match,
// but allow it at the lowest weight.
if (!CallOperandVal)
return CW_Default;
// Look at the constraint type.
switch (*constraint) {
case 'i': // immediate integer.
case 'n': // immediate integer with a known value.
if (isa<ConstantInt>(CallOperandVal))
weight = CW_Constant;
break;
case 's': // non-explicit intregal immediate.
if (isa<GlobalValue>(CallOperandVal))
weight = CW_Constant;
break;
case 'E': // immediate float if host format.
case 'F': // immediate float.
if (isa<ConstantFP>(CallOperandVal))
weight = CW_Constant;
break;
case '<': // memory operand with autodecrement.
case '>': // memory operand with autoincrement.
case 'm': // memory operand.
case 'o': // offsettable memory operand
case 'V': // non-offsettable memory operand
weight = CW_Memory;
break;
case 'r': // general register.
case 'g': // general register, memory operand or immediate integer.
// note: Clang converts "g" to "imr".
if (CallOperandVal->getType()->isIntegerTy())
weight = CW_Register;
break;
case 'X': // any operand.
default:
weight = CW_Default;
break;
}
return weight;
}
/// If there are multiple different constraints that we could pick for this
/// operand (e.g. "imr") try to pick the 'best' one.
/// This is somewhat tricky: constraints (TargetLowering::ConstraintType) fall
/// into seven classes:
/// Register -> one specific register
/// RegisterClass -> a group of regs
/// Memory -> memory
/// Address -> a symbolic memory reference
/// Immediate -> immediate values
/// Other -> magic values (such as "Flag Output Operands")
/// Unknown -> something we don't recognize yet and can't handle
/// Ideally, we would pick the most specific constraint possible: if we have
/// something that fits into a register, we would pick it. The problem here
/// is that if we have something that could either be in a register or in
/// memory that use of the register could cause selection of *other*
/// operands to fail: they might only succeed if we pick memory. Because of
/// this the heuristic we use is:
///
/// 1) If there is an 'other' constraint, and if the operand is valid for
/// that constraint, use it. This makes us take advantage of 'i'
/// constraints when available.
/// 2) Otherwise, pick the most general constraint present. This prefers
/// 'm' over 'r', for example.
///
TargetLowering::ConstraintGroup TargetLowering::getConstraintPreferences(
TargetLowering::AsmOperandInfo &OpInfo) const {
ConstraintGroup Ret;
Ret.reserve(OpInfo.Codes.size());
for (StringRef Code : OpInfo.Codes) {
TargetLowering::ConstraintType CType = getConstraintType(Code);
// Indirect 'other' or 'immediate' constraints are not allowed.
if (OpInfo.isIndirect && !(CType == TargetLowering::C_Memory ||
CType == TargetLowering::C_Register ||
CType == TargetLowering::C_RegisterClass))
continue;
// Things with matching constraints can only be registers, per gcc
// documentation. This mainly affects "g" constraints.
if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
continue;
Ret.emplace_back(Code, CType);
}
std::stable_sort(
Ret.begin(), Ret.end(), [](ConstraintPair a, ConstraintPair b) {
return getConstraintPiority(a.second) > getConstraintPiority(b.second);
});
return Ret;
}
/// If we have an immediate, see if we can lower it. Return true if we can,
/// false otherwise.
static bool lowerImmediateIfPossible(TargetLowering::ConstraintPair &P,
SDValue Op, SelectionDAG *DAG,
const TargetLowering &TLI) {
assert((P.second == TargetLowering::C_Other ||
P.second == TargetLowering::C_Immediate) &&
"need immediate or other");
if (!Op.getNode())
return false;
std::vector<SDValue> ResultOps;
TLI.LowerAsmOperandForConstraint(Op, P.first, ResultOps, *DAG);
return !ResultOps.empty();
}
/// Determines the constraint code and constraint type to use for the specific
/// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType.
void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
SDValue Op,
SelectionDAG *DAG) const {
assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
// Single-letter constraints ('r') are very common.
if (OpInfo.Codes.size() == 1) {
OpInfo.ConstraintCode = OpInfo.Codes[0];
OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
} else {
ConstraintGroup G = getConstraintPreferences(OpInfo);
if (G.empty())
return;
unsigned BestIdx = 0;
for (const unsigned E = G.size();
BestIdx < E && (G[BestIdx].second == TargetLowering::C_Other ||
G[BestIdx].second == TargetLowering::C_Immediate);
++BestIdx) {
if (lowerImmediateIfPossible(G[BestIdx], Op, DAG, *this))
break;
// If we're out of constraints, just pick the first one.
if (BestIdx + 1 == E) {
BestIdx = 0;
break;
}
}
OpInfo.ConstraintCode = G[BestIdx].first;
OpInfo.ConstraintType = G[BestIdx].second;
}
// 'X' matches anything.
if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
// Constants are handled elsewhere. For Functions, the type here is the
// type of the result, which is not what we want to look at; leave them
// alone.
Value *v = OpInfo.CallOperandVal;
if (isa<ConstantInt>(v) || isa<Function>(v)) {
return;
}
if (isa<BasicBlock>(v) || isa<BlockAddress>(v)) {
OpInfo.ConstraintCode = "i";
return;
}
// Otherwise, try to resolve it to something we know about by looking at
// the actual operand type.
if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
OpInfo.ConstraintCode = Repl;
OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
}
}
}
/// Given an exact SDIV by a constant, create a multiplication
/// with the multiplicative inverse of the constant.
/// Ref: "Hacker's Delight" by Henry Warren, 2nd Edition, p. 242
static SDValue BuildExactSDIV(const TargetLowering &TLI, SDNode *N,
const SDLoc &dl, SelectionDAG &DAG,
SmallVectorImpl<SDNode *> &Created) {
SDValue Op0 = N->getOperand(0);
SDValue Op1 = N->getOperand(1);
EVT VT = N->getValueType(0);
EVT SVT = VT.getScalarType();
EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
EVT ShSVT = ShVT.getScalarType();
bool UseSRA = false;
SmallVector<SDValue, 16> Shifts, Factors;
auto BuildSDIVPattern = [&](ConstantSDNode *C) {
if (C->isZero())
return false;
APInt Divisor = C->getAPIntValue();
unsigned Shift = Divisor.countr_zero();
if (Shift) {
Divisor.ashrInPlace(Shift);
UseSRA = true;
}
APInt Factor = Divisor.multiplicativeInverse();
Shifts.push_back(DAG.getConstant(Shift, dl, ShSVT));
Factors.push_back(DAG.getConstant(Factor, dl, SVT));
return true;
};
// Collect all magic values from the build vector.
if (!ISD::matchUnaryPredicate(Op1, BuildSDIVPattern))
return SDValue();
SDValue Shift, Factor;
if (Op1.getOpcode() == ISD::BUILD_VECTOR) {
Shift = DAG.getBuildVector(ShVT, dl, Shifts);
Factor = DAG.getBuildVector(VT, dl, Factors);
} else if (Op1.getOpcode() == ISD::SPLAT_VECTOR) {
assert(Shifts.size() == 1 && Factors.size() == 1 &&
"Expected matchUnaryPredicate to return one element for scalable "
"vectors");
Shift = DAG.getSplatVector(ShVT, dl, Shifts[0]);
Factor = DAG.getSplatVector(VT, dl, Factors[0]);
} else {
assert(isa<ConstantSDNode>(Op1) && "Expected a constant");
Shift = Shifts[0];
Factor = Factors[0];
}
SDValue Res = Op0;
if (UseSRA) {
SDNodeFlags Flags;
Flags.setExact(true);
Res = DAG.getNode(ISD::SRA, dl, VT, Res, Shift, Flags);
Created.push_back(Res.getNode());
}
return DAG.getNode(ISD::MUL, dl, VT, Res, Factor);
}
/// Given an exact UDIV by a constant, create a multiplication
/// with the multiplicative inverse of the constant.
/// Ref: "Hacker's Delight" by Henry Warren, 2nd Edition, p. 242
static SDValue BuildExactUDIV(const TargetLowering &TLI, SDNode *N,
const SDLoc &dl, SelectionDAG &DAG,
SmallVectorImpl<SDNode *> &Created) {
EVT VT = N->getValueType(0);
EVT SVT = VT.getScalarType();
EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
EVT ShSVT = ShVT.getScalarType();
bool UseSRL = false;
SmallVector<SDValue, 16> Shifts, Factors;
auto BuildUDIVPattern = [&](ConstantSDNode *C) {
if (C->isZero())
return false;
APInt Divisor = C->getAPIntValue();
unsigned Shift = Divisor.countr_zero();
if (Shift) {
Divisor.lshrInPlace(Shift);
UseSRL = true;
}
// Calculate the multiplicative inverse modulo BW.
APInt Factor = Divisor.multiplicativeInverse();
Shifts.push_back(DAG.getConstant(Shift, dl, ShSVT));
Factors.push_back(DAG.getConstant(Factor, dl, SVT));
return true;
};
SDValue Op1 = N->getOperand(1);
// Collect all magic values from the build vector.
if (!ISD::matchUnaryPredicate(Op1, BuildUDIVPattern))
return SDValue();
SDValue Shift, Factor;
if (Op1.getOpcode() == ISD::BUILD_VECTOR) {
Shift = DAG.getBuildVector(ShVT, dl, Shifts);
Factor = DAG.getBuildVector(VT, dl, Factors);
} else if (Op1.getOpcode() == ISD::SPLAT_VECTOR) {
assert(Shifts.size() == 1 && Factors.size() == 1 &&
"Expected matchUnaryPredicate to return one element for scalable "
"vectors");
Shift = DAG.getSplatVector(ShVT, dl, Shifts[0]);
Factor = DAG.getSplatVector(VT, dl, Factors[0]);
} else {
assert(isa<ConstantSDNode>(Op1) && "Expected a constant");
Shift = Shifts[0];
Factor = Factors[0];
}
SDValue Res = N->getOperand(0);
if (UseSRL) {
SDNodeFlags Flags;
Flags.setExact(true);
Res = DAG.getNode(ISD::SRL, dl, VT, Res, Shift, Flags);
Created.push_back(Res.getNode());
}
return DAG.getNode(ISD::MUL, dl, VT, Res, Factor);
}
SDValue TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
SelectionDAG &DAG,
SmallVectorImpl<SDNode *> &Created) const {
AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
if (TLI.isIntDivCheap(N->getValueType(0), Attr))
return SDValue(N, 0); // Lower SDIV as SDIV
return SDValue();
}
SDValue
TargetLowering::BuildSREMPow2(SDNode *N, const APInt &Divisor,
SelectionDAG &DAG,
SmallVectorImpl<SDNode *> &Created) const {
AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
if (TLI.isIntDivCheap(N->getValueType(0), Attr))
return SDValue(N, 0); // Lower SREM as SREM
return SDValue();
}
/// Build sdiv by power-of-2 with conditional move instructions
/// Ref: "Hacker's Delight" by Henry Warren 10-1
/// If conditional move/branch is preferred, we lower sdiv x, +/-2**k into:
/// bgez x, label
/// add x, x, 2**k-1
/// label:
/// sra res, x, k
/// neg res, res (when the divisor is negative)
SDValue TargetLowering::buildSDIVPow2WithCMov(
SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
SmallVectorImpl<SDNode *> &Created) const {
unsigned Lg2 = Divisor.countr_zero();
EVT VT = N->getValueType(0);
SDLoc DL(N);
SDValue N0 = N->getOperand(0);
SDValue Zero = DAG.getConstant(0, DL, VT);
APInt Lg2Mask = APInt::getLowBitsSet(VT.getSizeInBits(), Lg2);
SDValue Pow2MinusOne = DAG.getConstant(Lg2Mask, DL, VT);
// If N0 is negative, we need to add (Pow2 - 1) to it before shifting right.
EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue Cmp = DAG.getSetCC(DL, CCVT, N0, Zero, ISD::SETLT);
SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
SDValue CMov = DAG.getNode(ISD::SELECT, DL, VT, Cmp, Add, N0);
Created.push_back(Cmp.getNode());
Created.push_back(Add.getNode());
Created.push_back(CMov.getNode());
// Divide by pow2.
SDValue SRA =
DAG.getNode(ISD::SRA, DL, VT, CMov, DAG.getConstant(Lg2, DL, VT));
// If we're dividing by a positive value, we're done. Otherwise, we must
// negate the result.
if (Divisor.isNonNegative())
return SRA;
Created.push_back(SRA.getNode());
return DAG.getNode(ISD::SUB, DL, VT, Zero, SRA);
}
/// Given an ISD::SDIV node expressing a divide by constant,
/// return a DAG expression to select that will generate the same value by
/// multiplying by a magic number.
/// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
bool IsAfterLegalization,
SmallVectorImpl<SDNode *> &Created) const {
SDLoc dl(N);
EVT VT = N->getValueType(0);
EVT SVT = VT.getScalarType();
EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout());
EVT ShSVT = ShVT.getScalarType();
unsigned EltBits = VT.getScalarSizeInBits();
EVT MulVT;
// Check to see if we can do this.
// FIXME: We should be more aggressive here.
if (!isTypeLegal(VT)) {
// Limit this to simple scalars for now.
if (VT.isVector() || !VT.isSimple())
return SDValue();
// If this type will be promoted to a large enough type with a legal
// multiply operation, we can go ahead and do this transform.
if (getTypeAction(VT.getSimpleVT()) != TypePromoteInteger)
return SDValue();
MulVT = getTypeToTransformTo(*DAG.getContext(), VT);
if (MulVT.getSizeInBits() < (2 * EltBits) ||
!isOperationLegal(ISD::MUL, MulVT))
return SDValue();
}
// If the sdiv has an 'exact' bit we can use a simpler lowering.
if (N->getFlags().hasExact())
return BuildExactSDIV(*this, N, dl, DAG, Created);
SmallVector<SDValue, 16> MagicFactors, Factors, Shifts, ShiftMasks;
auto BuildSDIVPattern = [&](ConstantSDNode *C) {
if (C->isZero())
return false;
const APInt &Divisor = C->getAPIntValue();
SignedDivisionByConstantInfo magics = SignedDivisionByConstantInfo::get(Divisor);
int NumeratorFactor = 0;
int ShiftMask = -1;
if (Divisor.isOne() || Divisor.isAllOnes()) {
// If d is +1/-1, we just multiply the numerator by +1/-1.
NumeratorFactor = Divisor.getSExtValue();
magics.Magic = 0;
magics.ShiftAmount = 0;
ShiftMask = 0;
} else if (Divisor.isStrictlyPositive() && magics.Magic.isNegative()) {
// If d > 0 and m < 0, add the numerator.
NumeratorFactor = 1;
} else if (Divisor.isNegative() && magics.Magic.isStrictlyPositive()) {
// If d < 0 and m > 0, subtract the numerator.
NumeratorFactor = -1;
}
MagicFactors.push_back(DAG.getConstant(magics.Magic, dl, SVT));
Factors.push_back(DAG.getConstant(NumeratorFactor, dl, SVT));
Shifts.push_back(DAG.getConstant(magics.ShiftAmount, dl, ShSVT));
ShiftMasks.push_back(DAG.getConstant(ShiftMask, dl, SVT));
return true;
};
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
// Collect the shifts / magic values from each element.
if (!ISD::matchUnaryPredicate(N1, BuildSDIVPattern))
return SDValue();
SDValue MagicFactor, Factor, Shift, ShiftMask;
if (N1.getOpcode() == ISD::BUILD_VECTOR) {
MagicFactor = DAG.getBuildVector(VT, dl, MagicFactors);
Factor = DAG.getBuildVector(VT, dl, Factors);
Shift = DAG.getBuildVector(ShVT, dl, Shifts);
ShiftMask = DAG.getBuildVector(VT, dl, ShiftMasks);
} else if (N1.getOpcode() == ISD::SPLAT_VECTOR) {
assert(MagicFactors.size() == 1 && Factors.size() == 1 &&
Shifts.size() == 1 && ShiftMasks.size() == 1 &&
"Expected matchUnaryPredicate to return one element for scalable "
"vectors");
MagicFactor = DAG.getSplatVector(VT, dl, MagicFactors[0]);
Factor = DAG.getSplatVector(VT, dl, Factors[0]);
Shift = DAG.getSplatVector(ShVT, dl, Shifts[0]);
ShiftMask = DAG.getSplatVector(VT, dl, ShiftMasks[0]);
} else {
assert(isa<ConstantSDNode>(N1) && "Expected a constant");
MagicFactor = MagicFactors[0];
Factor = Factors[0];
Shift = Shifts[0];
ShiftMask = ShiftMasks[0];
}
// Multiply the numerator (operand 0) by the magic value.
// FIXME: We should support doing a MUL in a wider type.
auto GetMULHS = [&](SDValue X, SDValue Y) {
// If the type isn't legal, use a wider mul of the type calculated
// earlier.
if (!isTypeLegal(VT)) {
X = DAG.getNode(ISD::SIGN_EXTEND, dl, MulVT, X);
Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MulVT, Y);
Y = DAG.getNode(ISD::MUL, dl, MulVT, X, Y);
Y = DAG.getNode(ISD::SRL, dl, MulVT, Y,
DAG.getShiftAmountConstant(EltBits, MulVT, dl));
return DAG.getNode(ISD::TRUNCATE, dl, VT, Y);
}
if (isOperationLegalOrCustom(ISD::MULHS, VT, IsAfterLegalization))
return DAG.getNode(ISD::MULHS, dl, VT, X, Y);
if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT, IsAfterLegalization)) {
SDValue LoHi =
DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), X, Y);
return SDValue(LoHi.getNode(), 1);
}
// If type twice as wide legal, widen and use a mul plus a shift.
unsigned Size = VT.getScalarSizeInBits();
EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), Size * 2);
if (VT.isVector())
WideVT = EVT::getVectorVT(*DAG.getContext(), WideVT,
VT.getVectorElementCount());
if (isOperationLegalOrCustom(ISD::MUL, WideVT)) {
X = DAG.getNode(ISD::SIGN_EXTEND, dl, WideVT, X);
Y = DAG.getNode(ISD::SIGN_EXTEND, dl, WideVT, Y);
Y = DAG.getNode(ISD::MUL, dl, WideVT, X, Y);
Y = DAG.getNode(ISD::SRL, dl, WideVT, Y,
DAG.getShiftAmountConstant(EltBits, WideVT, dl));
return DAG.getNode(ISD::TRUNCATE, dl, VT, Y);
}
return SDValue();
};
SDValue Q = GetMULHS(N0, MagicFactor);
if (!Q)
return SDValue();
Created.push_back(Q.getNode());
// (Optionally) Add/subtract the numerator using Factor.
Factor = DAG.getNode(ISD::MUL, dl, VT, N0, Factor);
Created.push_back(Factor.getNode());
Q = DAG.getNode(ISD::ADD, dl, VT, Q, Factor);
Created.push_back(Q.getNode());
// Shift right algebraic by shift value.
Q = DAG.getNode(ISD::SRA, dl, VT, Q, Shift);
Created.push_back(Q.getNode());
// Extract the sign bit, mask it and add it to the quotient.
SDValue SignShift = DAG.getConstant(EltBits - 1, dl, ShVT);
SDValue T = DAG.getNode(ISD::SRL, dl, VT, Q, SignShift);
Created.push_back(T.getNode());
T = DAG.getNode(ISD::AND, dl, VT, T, ShiftMask);
Created.push_back(T.getNode());
return DAG.getNode(ISD::ADD, dl, VT, Q, T);
}
/// Given an ISD::UDIV node expressing a divide by constant,
/// return a DAG expression to select that will generate the same value by
/// multiplying by a magic number.
/// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
bool IsAfterLegalization,
SmallVectorImpl<SDNode *> &Created) const {
SDLoc dl(N);
EVT VT = N->getValueType(0);
EVT SVT = VT.getScalarType();
EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout());
EVT ShSVT = ShVT.getScalarType();
unsigned EltBits = VT.getScalarSizeInBits();
EVT MulVT;
// Check to see if we can do this.
// FIXME: We should be more aggressive here.
if (!isTypeLegal(VT)) {
// Limit this to simple scalars for now.
if (VT.isVector() || !VT.isSimple())
return SDValue();
// If this type will be promoted to a large enough type with a legal
// multiply operation, we can go ahead and do this transform.
if (getTypeAction(VT.getSimpleVT()) != TypePromoteInteger)
return SDValue();
MulVT = getTypeToTransformTo(*DAG.getContext(), VT);
if (MulVT.getSizeInBits() < (2 * EltBits) ||
!isOperationLegal(ISD::MUL, MulVT))
return SDValue();
}
// If the udiv has an 'exact' bit we can use a simpler lowering.
if (N->getFlags().hasExact())
return BuildExactUDIV(*this, N, dl, DAG, Created);
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
// Try to use leading zeros of the dividend to reduce the multiplier and
// avoid expensive fixups.
unsigned KnownLeadingZeros = DAG.computeKnownBits(N0).countMinLeadingZeros();
bool UseNPQ = false, UsePreShift = false, UsePostShift = false;
SmallVector<SDValue, 16> PreShifts, PostShifts, MagicFactors, NPQFactors;
auto BuildUDIVPattern = [&](ConstantSDNode *C) {
if (C->isZero())
return false;
const APInt& Divisor = C->getAPIntValue();
SDValue PreShift, MagicFactor, NPQFactor, PostShift;
// Magic algorithm doesn't work for division by 1. We need to emit a select
// at the end.
if (Divisor.isOne()) {
PreShift = PostShift = DAG.getUNDEF(ShSVT);
MagicFactor = NPQFactor = DAG.getUNDEF(SVT);
} else {
UnsignedDivisionByConstantInfo magics =
UnsignedDivisionByConstantInfo::get(
Divisor, std::min(KnownLeadingZeros, Divisor.countl_zero()));
MagicFactor = DAG.getConstant(magics.Magic, dl, SVT);
assert(magics.PreShift < Divisor.getBitWidth() &&
"We shouldn't generate an undefined shift!");
assert(magics.PostShift < Divisor.getBitWidth() &&
"We shouldn't generate an undefined shift!");
assert((!magics.IsAdd || magics.PreShift == 0) &&
"Unexpected pre-shift");
PreShift = DAG.getConstant(magics.PreShift, dl, ShSVT);
PostShift = DAG.getConstant(magics.PostShift, dl, ShSVT);
NPQFactor = DAG.getConstant(
magics.IsAdd ? APInt::getOneBitSet(EltBits, EltBits - 1)
: APInt::getZero(EltBits),
dl, SVT);
UseNPQ |= magics.IsAdd;
UsePreShift |= magics.PreShift != 0;
UsePostShift |= magics.PostShift != 0;
}
PreShifts.push_back(PreShift);
MagicFactors.push_back(MagicFactor);
NPQFactors.push_back(NPQFactor);
PostShifts.push_back(PostShift);
return true;
};
// Collect the shifts/magic values from each element.
if (!ISD::matchUnaryPredicate(N1, BuildUDIVPattern))
return SDValue();
SDValue PreShift, PostShift, MagicFactor, NPQFactor;
if (N1.getOpcode() == ISD::BUILD_VECTOR) {
PreShift = DAG.getBuildVector(ShVT, dl, PreShifts);
MagicFactor = DAG.getBuildVector(VT, dl, MagicFactors);
NPQFactor = DAG.getBuildVector(VT, dl, NPQFactors);
PostShift = DAG.getBuildVector(ShVT, dl, PostShifts);
} else if (N1.getOpcode() == ISD::SPLAT_VECTOR) {
assert(PreShifts.size() == 1 && MagicFactors.size() == 1 &&
NPQFactors.size() == 1 && PostShifts.size() == 1 &&
"Expected matchUnaryPredicate to return one for scalable vectors");
PreShift = DAG.getSplatVector(ShVT, dl, PreShifts[0]);
MagicFactor = DAG.getSplatVector(VT, dl, MagicFactors[0]);
NPQFactor = DAG.getSplatVector(VT, dl, NPQFactors[0]);
PostShift = DAG.getSplatVector(ShVT, dl, PostShifts[0]);
} else {
assert(isa<ConstantSDNode>(N1) && "Expected a constant");
PreShift = PreShifts[0];
MagicFactor = MagicFactors[0];
PostShift = PostShifts[0];
}
SDValue Q = N0;
if (UsePreShift) {
Q = DAG.getNode(ISD::SRL, dl, VT, Q, PreShift);
Created.push_back(Q.getNode());
}
// FIXME: We should support doing a MUL in a wider type.
auto GetMULHU = [&](SDValue X, SDValue Y) {
// If the type isn't legal, use a wider mul of the type calculated
// earlier.
if (!isTypeLegal(VT)) {
X = DAG.getNode(ISD::ZERO_EXTEND, dl, MulVT, X);
Y = DAG.getNode(ISD::ZERO_EXTEND, dl, MulVT, Y);
Y = DAG.getNode(ISD::MUL, dl, MulVT, X, Y);
Y = DAG.getNode(ISD::SRL, dl, MulVT, Y,
DAG.getShiftAmountConstant(EltBits, MulVT, dl));
return DAG.getNode(ISD::TRUNCATE, dl, VT, Y);
}
if (isOperationLegalOrCustom(ISD::MULHU, VT, IsAfterLegalization))
return DAG.getNode(ISD::MULHU, dl, VT, X, Y);
if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT, IsAfterLegalization)) {
SDValue LoHi =
DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), X, Y);
return SDValue(LoHi.getNode(), 1);
}
// If type twice as wide legal, widen and use a mul plus a shift.
unsigned Size = VT.getScalarSizeInBits();
EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), Size * 2);
if (VT.isVector())
WideVT = EVT::getVectorVT(*DAG.getContext(), WideVT,
VT.getVectorElementCount());
if (isOperationLegalOrCustom(ISD::MUL, WideVT)) {
X = DAG.getNode(ISD::ZERO_EXTEND, dl, WideVT, X);
Y = DAG.getNode(ISD::ZERO_EXTEND, dl, WideVT, Y);
Y = DAG.getNode(ISD::MUL, dl, WideVT, X, Y);
Y = DAG.getNode(ISD::SRL, dl, WideVT, Y,
DAG.getShiftAmountConstant(EltBits, WideVT, dl));
return DAG.getNode(ISD::TRUNCATE, dl, VT, Y);
}
return SDValue(); // No mulhu or equivalent
};
// Multiply the numerator (operand 0) by the magic value.
Q = GetMULHU(Q, MagicFactor);
if (!Q)
return SDValue();
Created.push_back(Q.getNode());
if (UseNPQ) {
SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N0, Q);
Created.push_back(NPQ.getNode());
// For vectors we might have a mix of non-NPQ/NPQ paths, so use
// MULHU to act as a SRL-by-1 for NPQ, else multiply by zero.
if (VT.isVector())
NPQ = GetMULHU(NPQ, NPQFactor);
else
NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, DAG.getConstant(1, dl, ShVT));
Created.push_back(NPQ.getNode());
Q = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
Created.push_back(Q.getNode());
}
if (UsePostShift) {
Q = DAG.getNode(ISD::SRL, dl, VT, Q, PostShift);
Created.push_back(Q.getNode());
}
EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue One = DAG.getConstant(1, dl, VT);
SDValue IsOne = DAG.getSetCC(dl, SetCCVT, N1, One, ISD::SETEQ);
return DAG.getSelect(dl, VT, IsOne, N0, Q);
}
/// If all values in Values that *don't* match the predicate are same 'splat'
/// value, then replace all values with that splat value.
/// Else, if AlternativeReplacement was provided, then replace all values that
/// do match predicate with AlternativeReplacement value.
static void
turnVectorIntoSplatVector(MutableArrayRef<SDValue> Values,
std::function<bool(SDValue)> Predicate,
SDValue AlternativeReplacement = SDValue()) {
SDValue Replacement;
// Is there a value for which the Predicate does *NOT* match? What is it?
auto SplatValue = llvm::find_if_not(Values, Predicate);
if (SplatValue != Values.end()) {
// Does Values consist only of SplatValue's and values matching Predicate?
if (llvm::all_of(Values, [Predicate, SplatValue](SDValue Value) {
return Value == *SplatValue || Predicate(Value);
})) // Then we shall replace values matching predicate with SplatValue.
Replacement = *SplatValue;
}
if (!Replacement) {
// Oops, we did not find the "baseline" splat value.
if (!AlternativeReplacement)
return; // Nothing to do.
// Let's replace with provided value then.
Replacement = AlternativeReplacement;
}
std::replace_if(Values.begin(), Values.end(), Predicate, Replacement);
}
/// Given an ISD::UREM used only by an ISD::SETEQ or ISD::SETNE
/// where the divisor is constant and the comparison target is zero,
/// return a DAG expression that will generate the same comparison result
/// using only multiplications, additions and shifts/rotations.
/// Ref: "Hacker's Delight" 10-17.
SDValue TargetLowering::buildUREMEqFold(EVT SETCCVT, SDValue REMNode,
SDValue CompTargetNode,
ISD::CondCode Cond,
DAGCombinerInfo &DCI,
const SDLoc &DL) const {
SmallVector<SDNode *, 5> Built;
if (SDValue Folded = prepareUREMEqFold(SETCCVT, REMNode, CompTargetNode, Cond,
DCI, DL, Built)) {
for (SDNode *N : Built)
DCI.AddToWorklist(N);
return Folded;
}
return SDValue();
}
SDValue
TargetLowering::prepareUREMEqFold(EVT SETCCVT, SDValue REMNode,
SDValue CompTargetNode, ISD::CondCode Cond,
DAGCombinerInfo &DCI, const SDLoc &DL,
SmallVectorImpl<SDNode *> &Created) const {
// fold (seteq/ne (urem N, D), 0) -> (setule/ugt (rotr (mul N, P), K), Q)
// - D must be constant, with D = D0 * 2^K where D0 is odd
// - P is the multiplicative inverse of D0 modulo 2^W
// - Q = floor(((2^W) - 1) / D)
// where W is the width of the common type of N and D.
assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
"Only applicable for (in)equality comparisons.");
SelectionDAG &DAG = DCI.DAG;
EVT VT = REMNode.getValueType();
EVT SVT = VT.getScalarType();
EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout());
EVT ShSVT = ShVT.getScalarType();
// If MUL is unavailable, we cannot proceed in any case.
if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::MUL, VT))
return SDValue();
bool ComparingWithAllZeros = true;
bool AllComparisonsWithNonZerosAreTautological = true;
bool HadTautologicalLanes = false;
bool AllLanesAreTautological = true;
bool HadEvenDivisor = false;
bool AllDivisorsArePowerOfTwo = true;
bool HadTautologicalInvertedLanes = false;
SmallVector<SDValue, 16> PAmts, KAmts, QAmts, IAmts;
auto BuildUREMPattern = [&](ConstantSDNode *CDiv, ConstantSDNode *CCmp) {
// Division by 0 is UB. Leave it to be constant-folded elsewhere.
if (CDiv->isZero())
return false;
const APInt &D = CDiv->getAPIntValue();
const APInt &Cmp = CCmp->getAPIntValue();
ComparingWithAllZeros &= Cmp.isZero();
// x u% C1` is *always* less than C1. So given `x u% C1 == C2`,
// if C2 is not less than C1, the comparison is always false.
// But we will only be able to produce the comparison that will give the
// opposive tautological answer. So this lane would need to be fixed up.
bool TautologicalInvertedLane = D.ule(Cmp);
HadTautologicalInvertedLanes |= TautologicalInvertedLane;
// If all lanes are tautological (either all divisors are ones, or divisor
// is not greater than the constant we are comparing with),
// we will prefer to avoid the fold.
bool TautologicalLane = D.isOne() || TautologicalInvertedLane;
HadTautologicalLanes |= TautologicalLane;
AllLanesAreTautological &= TautologicalLane;
// If we are comparing with non-zero, we need'll need to subtract said
// comparison value from the LHS. But there is no point in doing that if
// every lane where we are comparing with non-zero is tautological..
if (!Cmp.isZero())
AllComparisonsWithNonZerosAreTautological &= TautologicalLane;
// Decompose D into D0 * 2^K
unsigned K = D.countr_zero();
assert((!D.isOne() || (K == 0)) && "For divisor '1' we won't rotate.");
APInt D0 = D.lshr(K);
// D is even if it has trailing zeros.
HadEvenDivisor |= (K != 0);
// D is a power-of-two if D0 is one.
// If all divisors are power-of-two, we will prefer to avoid the fold.
AllDivisorsArePowerOfTwo &= D0.isOne();
// P = inv(D0, 2^W)
// 2^W requires W + 1 bits, so we have to extend and then truncate.
unsigned W = D.getBitWidth();
APInt P = D0.multiplicativeInverse();
assert((D0 * P).isOne() && "Multiplicative inverse basic check failed.");
// Q = floor((2^W - 1) u/ D)
// R = ((2^W - 1) u% D)
APInt Q, R;
APInt::udivrem(APInt::getAllOnes(W), D, Q, R);
// If we are comparing with zero, then that comparison constant is okay,
// else it may need to be one less than that.
if (Cmp.ugt(R))
Q -= 1;
assert(APInt::getAllOnes(ShSVT.getSizeInBits()).ugt(K) &&
"We are expecting that K is always less than all-ones for ShSVT");
// If the lane is tautological the result can be constant-folded.
if (TautologicalLane) {
// Set P and K amount to a bogus values so we can try to splat them.
P = 0;
K = -1;
// And ensure that comparison constant is tautological,
// it will always compare true/false.
Q = -1;
}
PAmts.push_back(DAG.getConstant(P, DL, SVT));
KAmts.push_back(
DAG.getConstant(APInt(ShSVT.getSizeInBits(), K), DL, ShSVT));
QAmts.push_back(DAG.getConstant(Q, DL, SVT));
return true;
};
SDValue N = REMNode.getOperand(0);
SDValue D = REMNode.getOperand(1);
// Collect the values from each element.
if (!ISD::matchBinaryPredicate(D, CompTargetNode, BuildUREMPattern))
return SDValue();
// If all lanes are tautological, the result can be constant-folded.
if (AllLanesAreTautological)
return SDValue();
// If this is a urem by a powers-of-two, avoid the fold since it can be
// best implemented as a bit test.
if (AllDivisorsArePowerOfTwo)
return SDValue();
SDValue PVal, KVal, QVal;
if (D.getOpcode() == ISD::BUILD_VECTOR) {
if (HadTautologicalLanes) {
// Try to turn PAmts into a splat, since we don't care about the values
// that are currently '0'. If we can't, just keep '0'`s.
turnVectorIntoSplatVector(PAmts, isNullConstant);
// Try to turn KAmts into a splat, since we don't care about the values
// that are currently '-1'. If we can't, change them to '0'`s.
turnVectorIntoSplatVector(KAmts, isAllOnesConstant,
DAG.getConstant(0, DL, ShSVT));
}
PVal = DAG.getBuildVector(VT, DL, PAmts);
KVal = DAG.getBuildVector(ShVT, DL, KAmts);
QVal = DAG.getBuildVector(VT, DL, QAmts);
} else if (D.getOpcode() == ISD::SPLAT_VECTOR) {
assert(PAmts.size() == 1 && KAmts.size() == 1 && QAmts.size() == 1 &&
"Expected matchBinaryPredicate to return one element for "
"SPLAT_VECTORs");
PVal = DAG.getSplatVector(VT, DL, PAmts[0]);
KVal = DAG.getSplatVector(ShVT, DL, KAmts[0]);
QVal = DAG.getSplatVector(VT, DL, QAmts[0]);
} else {
PVal = PAmts[0];
KVal = KAmts[0];
QVal = QAmts[0];
}
if (!ComparingWithAllZeros && !AllComparisonsWithNonZerosAreTautological) {
if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::SUB, VT))
return SDValue(); // FIXME: Could/should use `ISD::ADD`?
assert(CompTargetNode.getValueType() == N.getValueType() &&
"Expecting that the types on LHS and RHS of comparisons match.");
N = DAG.getNode(ISD::SUB, DL, VT, N, CompTargetNode);
}
// (mul N, P)
SDValue Op0 = DAG.getNode(ISD::MUL, DL, VT, N, PVal);
Created.push_back(Op0.getNode());
// Rotate right only if any divisor was even. We avoid rotates for all-odd
// divisors as a performance improvement, since rotating by 0 is a no-op.
if (HadEvenDivisor) {
// We need ROTR to do this.
if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::ROTR, VT))
return SDValue();
// UREM: (rotr (mul N, P), K)
Op0 = DAG.getNode(ISD::ROTR, DL, VT, Op0, KVal);
Created.push_back(Op0.getNode());
}
// UREM: (setule/setugt (rotr (mul N, P), K), Q)
SDValue NewCC =
DAG.getSetCC(DL, SETCCVT, Op0, QVal,
((Cond == ISD::SETEQ) ? ISD::SETULE : ISD::SETUGT));
if (!HadTautologicalInvertedLanes)
return NewCC;
// If any lanes previously compared always-false, the NewCC will give
// always-true result for them, so we need to fixup those lanes.
// Or the other way around for inequality predicate.
assert(VT.isVector() && "Can/should only get here for vectors.");
Created.push_back(NewCC.getNode());
// x u% C1` is *always* less than C1. So given `x u% C1 == C2`,
// if C2 is not less than C1, the comparison is always false.
// But we have produced the comparison that will give the
// opposive tautological answer. So these lanes would need to be fixed up.
SDValue TautologicalInvertedChannels =
DAG.getSetCC(DL, SETCCVT, D, CompTargetNode, ISD::SETULE);
Created.push_back(TautologicalInvertedChannels.getNode());
// NOTE: we avoid letting illegal types through even if we're before legalize
// ops – legalization has a hard time producing good code for this.
if (isOperationLegalOrCustom(ISD::VSELECT, SETCCVT)) {
// If we have a vector select, let's replace the comparison results in the
// affected lanes with the correct tautological result.
SDValue Replacement = DAG.getBoolConstant(Cond == ISD::SETEQ ? false : true,
DL, SETCCVT, SETCCVT);
return DAG.getNode(ISD::VSELECT, DL, SETCCVT, TautologicalInvertedChannels,
Replacement, NewCC);
}
// Else, we can just invert the comparison result in the appropriate lanes.
//
// NOTE: see the note above VSELECT above.
if (isOperationLegalOrCustom(ISD::XOR, SETCCVT))
return DAG.getNode(ISD::XOR, DL, SETCCVT, NewCC,
TautologicalInvertedChannels);
return SDValue(); // Don't know how to lower.
}
/// Given an ISD::SREM used only by an ISD::SETEQ or ISD::SETNE
/// where the divisor is constant and the comparison target is zero,
/// return a DAG expression that will generate the same comparison result
/// using only multiplications, additions and shifts/rotations.
/// Ref: "Hacker's Delight" 10-17.
SDValue TargetLowering::buildSREMEqFold(EVT SETCCVT, SDValue REMNode,
SDValue CompTargetNode,
ISD::CondCode Cond,
DAGCombinerInfo &DCI,
const SDLoc &DL) const {
SmallVector<SDNode *, 7> Built;
if (SDValue Folded = prepareSREMEqFold(SETCCVT, REMNode, CompTargetNode, Cond,
DCI, DL, Built)) {
assert(Built.size() <= 7 && "Max size prediction failed.");
for (SDNode *N : Built)
DCI.AddToWorklist(N);
return Folded;
}
return SDValue();
}
SDValue
TargetLowering::prepareSREMEqFold(EVT SETCCVT, SDValue REMNode,
SDValue CompTargetNode, ISD::CondCode Cond,
DAGCombinerInfo &DCI, const SDLoc &DL,
SmallVectorImpl<SDNode *> &Created) const {
// Derived from Hacker's Delight, 2nd Edition, by Hank Warren. Section 10-17.
// Fold:
// (seteq/ne (srem N, D), 0)
// To:
// (setule/ugt (rotr (add (mul N, P), A), K), Q)
//
// - D must be constant, with D = D0 * 2^K where D0 is odd
// - P is the multiplicative inverse of D0 modulo 2^W
// - A = bitwiseand(floor((2^(W - 1) - 1) / D0), (-(2^k)))
// - Q = floor((2 * A) / (2^K))
// where W is the width of the common type of N and D.
//
// When D is a power of two (and thus D0 is 1), the normal
// formula for A and Q don't apply, because the derivation
// depends on D not dividing 2^(W-1), and thus theorem ZRS
// does not apply. This specifically fails when N = INT_MIN.
//
// Instead, for power-of-two D, we use:
// - A = 2^(W-1)
// |-> Order-preserving map from [-2^(W-1), 2^(W-1) - 1] to [0,2^W - 1])
// - Q = 2^(W-K) - 1
// |-> Test that the top K bits are zero after rotation
assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
"Only applicable for (in)equality comparisons.");
SelectionDAG &DAG = DCI.DAG;
EVT VT = REMNode.getValueType();
EVT SVT = VT.getScalarType();
EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout());
EVT ShSVT = ShVT.getScalarType();
// If we are after ops legalization, and MUL is unavailable, we can not
// proceed.
if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::MUL, VT))
return SDValue();
// TODO: Could support comparing with non-zero too.
ConstantSDNode *CompTarget = isConstOrConstSplat(CompTargetNode);
if (!CompTarget || !CompTarget->isZero())
return SDValue();
bool HadIntMinDivisor = false;
bool HadOneDivisor = false;
bool AllDivisorsAreOnes = true;
bool HadEvenDivisor = false;
bool NeedToApplyOffset = false;
bool AllDivisorsArePowerOfTwo = true;
SmallVector<SDValue, 16> PAmts, AAmts, KAmts, QAmts;
auto BuildSREMPattern = [&](ConstantSDNode *C) {
// Division by 0 is UB. Leave it to be constant-folded elsewhere.
if (C->isZero())
return false;
// FIXME: we don't fold `rem %X, -C` to `rem %X, C` in DAGCombine.
// WARNING: this fold is only valid for positive divisors!
APInt D = C->getAPIntValue();
if (D.isNegative())
D.negate(); // `rem %X, -C` is equivalent to `rem %X, C`
HadIntMinDivisor |= D.isMinSignedValue();
// If all divisors are ones, we will prefer to avoid the fold.
HadOneDivisor |= D.isOne();
AllDivisorsAreOnes &= D.isOne();
// Decompose D into D0 * 2^K
unsigned K = D.countr_zero();
assert((!D.isOne() || (K == 0)) && "For divisor '1' we won't rotate.");
APInt D0 = D.lshr(K);
if (!D.isMinSignedValue()) {
// D is even if it has trailing zeros; unless it's INT_MIN, in which case
// we don't care about this lane in this fold, we'll special-handle it.
HadEvenDivisor |= (K != 0);
}
// D is a power-of-two if D0 is one. This includes INT_MIN.
// If all divisors are power-of-two, we will prefer to avoid the fold.
AllDivisorsArePowerOfTwo &= D0.isOne();
// P = inv(D0, 2^W)
// 2^W requires W + 1 bits, so we have to extend and then truncate.
unsigned W = D.getBitWidth();
APInt P = D0.multiplicativeInverse();
assert((D0 * P).isOne() && "Multiplicative inverse basic check failed.");
// A = floor((2^(W - 1) - 1) / D0) & -2^K
APInt A = APInt::getSignedMaxValue(W).udiv(D0);
A.clearLowBits(K);
if (!D.isMinSignedValue()) {
// If divisor INT_MIN, then we don't care about this lane in this fold,
// we'll special-handle it.
NeedToApplyOffset |= A != 0;
}
// Q = floor((2 * A) / (2^K))
APInt Q = (2 * A).udiv(APInt::getOneBitSet(W, K));
assert(APInt::getAllOnes(SVT.getSizeInBits()).ugt(A) &&
"We are expecting that A is always less than all-ones for SVT");
assert(APInt::getAllOnes(ShSVT.getSizeInBits()).ugt(K) &&
"We are expecting that K is always less than all-ones for ShSVT");
// If D was a power of two, apply the alternate constant derivation.
if (D0.isOne()) {
// A = 2^(W-1)
A = APInt::getSignedMinValue(W);
// - Q = 2^(W-K) - 1
Q = APInt::getAllOnes(W - K).zext(W);
}
// If the divisor is 1 the result can be constant-folded. Likewise, we
// don't care about INT_MIN lanes, those can be set to undef if appropriate.
if (D.isOne()) {
// Set P, A and K to a bogus values so we can try to splat them.
P = 0;
A = -1;
K = -1;
// x ?% 1 == 0 <--> true <--> x u<= -1
Q = -1;
}
PAmts.push_back(DAG.getConstant(P, DL, SVT));
AAmts.push_back(DAG.getConstant(A, DL, SVT));
KAmts.push_back(
DAG.getConstant(APInt(ShSVT.getSizeInBits(), K), DL, ShSVT));
QAmts.push_back(DAG.getConstant(Q, DL, SVT));
return true;
};
SDValue N = REMNode.getOperand(0);
SDValue D = REMNode.getOperand(1);
// Collect the values from each element.
if (!ISD::matchUnaryPredicate(D, BuildSREMPattern))
return SDValue();
// If this is a srem by a one, avoid the fold since it can be constant-folded.
if (AllDivisorsAreOnes)
return SDValue();
// If this is a srem by a powers-of-two (including INT_MIN), avoid the fold
// since it can be best implemented as a bit test.
if (AllDivisorsArePowerOfTwo)
return SDValue();
SDValue PVal, AVal, KVal, QVal;
if (D.getOpcode() == ISD::BUILD_VECTOR) {
if (HadOneDivisor) {
// Try to turn PAmts into a splat, since we don't care about the values
// that are currently '0'. If we can't, just keep '0'`s.
turnVectorIntoSplatVector(PAmts, isNullConstant);
// Try to turn AAmts into a splat, since we don't care about the
// values that are currently '-1'. If we can't, change them to '0'`s.
turnVectorIntoSplatVector(AAmts, isAllOnesConstant,
DAG.getConstant(0, DL, SVT));
// Try to turn KAmts into a splat, since we don't care about the values
// that are currently '-1'. If we can't, change them to '0'`s.
turnVectorIntoSplatVector(KAmts, isAllOnesConstant,
DAG.getConstant(0, DL, ShSVT));
}
PVal = DAG.getBuildVector(VT, DL, PAmts);
AVal = DAG.getBuildVector(VT, DL, AAmts);
KVal = DAG.getBuildVector(ShVT, DL, KAmts);
QVal = DAG.getBuildVector(VT, DL, QAmts);
} else if (D.getOpcode() == ISD::SPLAT_VECTOR) {
assert(PAmts.size() == 1 && AAmts.size() == 1 && KAmts.size() == 1 &&
QAmts.size() == 1 &&
"Expected matchUnaryPredicate to return one element for scalable "
"vectors");
PVal = DAG.getSplatVector(VT, DL, PAmts[0]);
AVal = DAG.getSplatVector(VT, DL, AAmts[0]);
KVal = DAG.getSplatVector(ShVT, DL, KAmts[0]);
QVal = DAG.getSplatVector(VT, DL, QAmts[0]);
} else {
assert(isa<ConstantSDNode>(D) && "Expected a constant");
PVal = PAmts[0];
AVal = AAmts[0];
KVal = KAmts[0];
QVal = QAmts[0];
}
// (mul N, P)
SDValue Op0 = DAG.getNode(ISD::MUL, DL, VT, N, PVal);
Created.push_back(Op0.getNode());
if (NeedToApplyOffset) {
// We need ADD to do this.
if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::ADD, VT))
return SDValue();
// (add (mul N, P), A)
Op0 = DAG.getNode(ISD::ADD, DL, VT, Op0, AVal);
Created.push_back(Op0.getNode());
}
// Rotate right only if any divisor was even. We avoid rotates for all-odd
// divisors as a performance improvement, since rotating by 0 is a no-op.
if (HadEvenDivisor) {
// We need ROTR to do this.
if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::ROTR, VT))
return SDValue();
// SREM: (rotr (add (mul N, P), A), K)
Op0 = DAG.getNode(ISD::ROTR, DL, VT, Op0, KVal);
Created.push_back(Op0.getNode());
}
// SREM: (setule/setugt (rotr (add (mul N, P), A), K), Q)
SDValue Fold =
DAG.getSetCC(DL, SETCCVT, Op0, QVal,
((Cond == ISD::SETEQ) ? ISD::SETULE : ISD::SETUGT));
// If we didn't have lanes with INT_MIN divisor, then we're done.
if (!HadIntMinDivisor)
return Fold;
// That fold is only valid for positive divisors. Which effectively means,
// it is invalid for INT_MIN divisors. So if we have such a lane,
// we must fix-up results for said lanes.
assert(VT.isVector() && "Can/should only get here for vectors.");
// NOTE: we avoid letting illegal types through even if we're before legalize
// ops – legalization has a hard time producing good code for the code that
// follows.
if (!isOperationLegalOrCustom(ISD::SETCC, SETCCVT) ||
!isOperationLegalOrCustom(ISD::AND, VT) ||
!isCondCodeLegalOrCustom(Cond, VT.getSimpleVT()) ||
!isOperationLegalOrCustom(ISD::VSELECT, SETCCVT))
return SDValue();
Created.push_back(Fold.getNode());
SDValue IntMin = DAG.getConstant(
APInt::getSignedMinValue(SVT.getScalarSizeInBits()), DL, VT);
SDValue IntMax = DAG.getConstant(
APInt::getSignedMaxValue(SVT.getScalarSizeInBits()), DL, VT);
SDValue Zero =
DAG.getConstant(APInt::getZero(SVT.getScalarSizeInBits()), DL, VT);
// Which lanes had INT_MIN divisors? Divisor is constant, so const-folded.
SDValue DivisorIsIntMin = DAG.getSetCC(DL, SETCCVT, D, IntMin, ISD::SETEQ);
Created.push_back(DivisorIsIntMin.getNode());
// (N s% INT_MIN) ==/!= 0 <--> (N & INT_MAX) ==/!= 0
SDValue Masked = DAG.getNode(ISD::AND, DL, VT, N, IntMax);
Created.push_back(Masked.getNode());
SDValue MaskedIsZero = DAG.getSetCC(DL, SETCCVT, Masked, Zero, Cond);
Created.push_back(MaskedIsZero.getNode());
// To produce final result we need to blend 2 vectors: 'SetCC' and
// 'MaskedIsZero'. If the divisor for channel was *NOT* INT_MIN, we pick
// from 'Fold', else pick from 'MaskedIsZero'. Since 'DivisorIsIntMin' is
// constant-folded, select can get lowered to a shuffle with constant mask.
SDValue Blended = DAG.getNode(ISD::VSELECT, DL, SETCCVT, DivisorIsIntMin,
MaskedIsZero, Fold);
return Blended;
}
bool TargetLowering::
verifyReturnAddressArgumentIsConstant(SDValue Op, SelectionDAG &DAG) const {
if (!isa<ConstantSDNode>(Op.getOperand(0))) {
DAG.getContext()->emitError("argument to '__builtin_return_address' must "
"be a constant integer");
return true;
}
return false;
}
SDValue TargetLowering::getSqrtInputTest(SDValue Op, SelectionDAG &DAG,
const DenormalMode &Mode) const {
SDLoc DL(Op);
EVT VT = Op.getValueType();
EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue FPZero = DAG.getConstantFP(0.0, DL, VT);
// This is specifically a check for the handling of denormal inputs, not the
// result.
if (Mode.Input == DenormalMode::PreserveSign ||
Mode.Input == DenormalMode::PositiveZero) {
// Test = X == 0.0
return DAG.getSetCC(DL, CCVT, Op, FPZero, ISD::SETEQ);
}
// Testing it with denormal inputs to avoid wrong estimate.
//
// Test = fabs(X) < SmallestNormal
const fltSemantics &FltSem = DAG.EVTToAPFloatSemantics(VT);
APFloat SmallestNorm = APFloat::getSmallestNormalized(FltSem);
SDValue NormC = DAG.getConstantFP(SmallestNorm, DL, VT);
SDValue Fabs = DAG.getNode(ISD::FABS, DL, VT, Op);
return DAG.getSetCC(DL, CCVT, Fabs, NormC, ISD::SETLT);
}
SDValue TargetLowering::getNegatedExpression(SDValue Op, SelectionDAG &DAG,
bool LegalOps, bool OptForSize,
NegatibleCost &Cost,
unsigned Depth) const {
// fneg is removable even if it has multiple uses.
if (Op.getOpcode() == ISD::FNEG || Op.getOpcode() == ISD::VP_FNEG) {
Cost = NegatibleCost::Cheaper;
return Op.getOperand(0);
}
// Don't recurse exponentially.
if (Depth > SelectionDAG::MaxRecursionDepth)
return SDValue();
// Pre-increment recursion depth for use in recursive calls.
++Depth;
const SDNodeFlags Flags = Op->getFlags();
const TargetOptions &Options = DAG.getTarget().Options;
EVT VT = Op.getValueType();
unsigned Opcode = Op.getOpcode();
// Don't allow anything with multiple uses unless we know it is free.
if (!Op.hasOneUse() && Opcode != ISD::ConstantFP) {
bool IsFreeExtend = Opcode == ISD::FP_EXTEND &&
isFPExtFree(VT, Op.getOperand(0).getValueType());
if (!IsFreeExtend)
return SDValue();
}
auto RemoveDeadNode = [&](SDValue N) {
if (N && N.getNode()->use_empty())
DAG.RemoveDeadNode(N.getNode());
};
SDLoc DL(Op);
// Because getNegatedExpression can delete nodes we need a handle to keep
// temporary nodes alive in case the recursion manages to create an identical
// node.
std::list<HandleSDNode> Handles;
switch (Opcode) {
case ISD::ConstantFP: {
// Don't invert constant FP values after legalization unless the target says
// the negated constant is legal.
bool IsOpLegal =
isOperationLegal(ISD::ConstantFP, VT) ||
isFPImmLegal(neg(cast<ConstantFPSDNode>(Op)->getValueAPF()), VT,
OptForSize);
if (LegalOps && !IsOpLegal)
break;
APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
V.changeSign();
SDValue CFP = DAG.getConstantFP(V, DL, VT);
// If we already have the use of the negated floating constant, it is free
// to negate it even it has multiple uses.
if (!Op.hasOneUse() && CFP.use_empty())
break;
Cost = NegatibleCost::Neutral;
return CFP;
}
case ISD::BUILD_VECTOR: {
// Only permit BUILD_VECTOR of constants.
if (llvm::any_of(Op->op_values(), [&](SDValue N) {
return !N.isUndef() && !isa<ConstantFPSDNode>(N);
}))
break;
bool IsOpLegal =
(isOperationLegal(ISD::ConstantFP, VT) &&
isOperationLegal(ISD::BUILD_VECTOR, VT)) ||
llvm::all_of(Op->op_values(), [&](SDValue N) {
return N.isUndef() ||
isFPImmLegal(neg(cast<ConstantFPSDNode>(N)->getValueAPF()), VT,
OptForSize);
});
if (LegalOps && !IsOpLegal)
break;
SmallVector<SDValue, 4> Ops;
for (SDValue C : Op->op_values()) {
if (C.isUndef()) {
Ops.push_back(C);
continue;
}
APFloat V = cast<ConstantFPSDNode>(C)->getValueAPF();
V.changeSign();
Ops.push_back(DAG.getConstantFP(V, DL, C.getValueType()));
}
Cost = NegatibleCost::Neutral;
return DAG.getBuildVector(VT, DL, Ops);
}
case ISD::FADD: {
if (!Options.NoSignedZerosFPMath && !Flags.hasNoSignedZeros())
break;
// After operation legalization, it might not be legal to create new FSUBs.
if (LegalOps && !isOperationLegalOrCustom(ISD::FSUB, VT))
break;
SDValue X = Op.getOperand(0), Y = Op.getOperand(1);
// fold (fneg (fadd X, Y)) -> (fsub (fneg X), Y)
NegatibleCost CostX = NegatibleCost::Expensive;
SDValue NegX =
getNegatedExpression(X, DAG, LegalOps, OptForSize, CostX, Depth);
// Prevent this node from being deleted by the next call.
if (NegX)
Handles.emplace_back(NegX);
// fold (fneg (fadd X, Y)) -> (fsub (fneg Y), X)
NegatibleCost CostY = NegatibleCost::Expensive;
SDValue NegY =
getNegatedExpression(Y, DAG, LegalOps, OptForSize, CostY, Depth);
// We're done with the handles.
Handles.clear();
// Negate the X if its cost is less or equal than Y.
if (NegX && (CostX <= CostY)) {
Cost = CostX;
SDValue N = DAG.getNode(ISD::FSUB, DL, VT, NegX, Y, Flags);
if (NegY != N)
RemoveDeadNode(NegY);
return N;
}
// Negate the Y if it is not expensive.
if (NegY) {
Cost = CostY;
SDValue N = DAG.getNode(ISD::FSUB, DL, VT, NegY, X, Flags);
if (NegX != N)
RemoveDeadNode(NegX);
return N;
}
break;
}
case ISD::FSUB: {
// We can't turn -(A-B) into B-A when we honor signed zeros.
if (!Options.NoSignedZerosFPMath && !Flags.hasNoSignedZeros())
break;
SDValue X = Op.getOperand(0), Y = Op.getOperand(1);
// fold (fneg (fsub 0, Y)) -> Y
if (ConstantFPSDNode *C = isConstOrConstSplatFP(X, /*AllowUndefs*/ true))
if (C->isZero()) {
Cost = NegatibleCost::Cheaper;
return Y;
}
// fold (fneg (fsub X, Y)) -> (fsub Y, X)
Cost = NegatibleCost::Neutral;
return DAG.getNode(ISD::FSUB, DL, VT, Y, X, Flags);
}
case ISD::FMUL:
case ISD::FDIV: {
SDValue X = Op.getOperand(0), Y = Op.getOperand(1);
// fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
NegatibleCost CostX = NegatibleCost::Expensive;
SDValue NegX =
getNegatedExpression(X, DAG, LegalOps, OptForSize, CostX, Depth);
// Prevent this node from being deleted by the next call.
if (NegX)
Handles.emplace_back(NegX);
// fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
NegatibleCost CostY = NegatibleCost::Expensive;
SDValue NegY =
getNegatedExpression(Y, DAG, LegalOps, OptForSize, CostY, Depth);
// We're done with the handles.
Handles.clear();
// Negate the X if its cost is less or equal than Y.
if (NegX && (CostX <= CostY)) {
Cost = CostX;
SDValue N = DAG.getNode(Opcode, DL, VT, NegX, Y, Flags);
if (NegY != N)
RemoveDeadNode(NegY);
return N;
}
// Ignore X * 2.0 because that is expected to be canonicalized to X + X.
if (auto *C = isConstOrConstSplatFP(Op.getOperand(1)))
if (C->isExactlyValue(2.0) && Op.getOpcode() == ISD::FMUL)
break;
// Negate the Y if it is not expensive.
if (NegY) {
Cost = CostY;
SDValue N = DAG.getNode(Opcode, DL, VT, X, NegY, Flags);
if (NegX != N)
RemoveDeadNode(NegX);
return N;
}
break;
}
case ISD::FMA:
case ISD::FMAD: {
if (!Options.NoSignedZerosFPMath && !Flags.hasNoSignedZeros())
break;
SDValue X = Op.getOperand(0), Y = Op.getOperand(1), Z = Op.getOperand(2);
NegatibleCost CostZ = NegatibleCost::Expensive;
SDValue NegZ =
getNegatedExpression(Z, DAG, LegalOps, OptForSize, CostZ, Depth);
// Give up if fail to negate the Z.
if (!NegZ)
break;
// Prevent this node from being deleted by the next two calls.
Handles.emplace_back(NegZ);
// fold (fneg (fma X, Y, Z)) -> (fma (fneg X), Y, (fneg Z))
NegatibleCost CostX = NegatibleCost::Expensive;
SDValue NegX =
getNegatedExpression(X, DAG, LegalOps, OptForSize, CostX, Depth);
// Prevent this node from being deleted by the next call.
if (NegX)
Handles.emplace_back(NegX);
// fold (fneg (fma X, Y, Z)) -> (fma X, (fneg Y), (fneg Z))
NegatibleCost CostY = NegatibleCost::Expensive;
SDValue NegY =
getNegatedExpression(Y, DAG, LegalOps, OptForSize, CostY, Depth);
// We're done with the handles.
Handles.clear();
// Negate the X if its cost is less or equal than Y.
if (NegX && (CostX <= CostY)) {
Cost = std::min(CostX, CostZ);
SDValue N = DAG.getNode(Opcode, DL, VT, NegX, Y, NegZ, Flags);
if (NegY != N)
RemoveDeadNode(NegY);
return N;
}
// Negate the Y if it is not expensive.
if (NegY) {
Cost = std::min(CostY, CostZ);
SDValue N = DAG.getNode(Opcode, DL, VT, X, NegY, NegZ, Flags);
if (NegX != N)
RemoveDeadNode(NegX);
return N;
}
break;
}
case ISD::FP_EXTEND:
case ISD::FSIN:
if (SDValue NegV = getNegatedExpression(Op.getOperand(0), DAG, LegalOps,
OptForSize, Cost, Depth))
return DAG.getNode(Opcode, DL, VT, NegV);
break;
case ISD::FP_ROUND:
if (SDValue NegV = getNegatedExpression(Op.getOperand(0), DAG, LegalOps,
OptForSize, Cost, Depth))
return DAG.getNode(ISD::FP_ROUND, DL, VT, NegV, Op.getOperand(1));
break;
case ISD::SELECT:
case ISD::VSELECT: {
// fold (fneg (select C, LHS, RHS)) -> (select C, (fneg LHS), (fneg RHS))
// iff at least one cost is cheaper and the other is neutral/cheaper
SDValue LHS = Op.getOperand(1);
NegatibleCost CostLHS = NegatibleCost::Expensive;
SDValue NegLHS =
getNegatedExpression(LHS, DAG, LegalOps, OptForSize, CostLHS, Depth);
if (!NegLHS || CostLHS > NegatibleCost::Neutral) {
RemoveDeadNode(NegLHS);
break;
}
// Prevent this node from being deleted by the next call.
Handles.emplace_back(NegLHS);
SDValue RHS = Op.getOperand(2);
NegatibleCost CostRHS = NegatibleCost::Expensive;
SDValue NegRHS =
getNegatedExpression(RHS, DAG, LegalOps, OptForSize, CostRHS, Depth);
// We're done with the handles.
Handles.clear();
if (!NegRHS || CostRHS > NegatibleCost::Neutral ||
(CostLHS != NegatibleCost::Cheaper &&
CostRHS != NegatibleCost::Cheaper)) {
RemoveDeadNode(NegLHS);
RemoveDeadNode(NegRHS);
break;
}
Cost = std::min(CostLHS, CostRHS);
return DAG.getSelect(DL, VT, Op.getOperand(0), NegLHS, NegRHS);
}
}
return SDValue();
}
//===----------------------------------------------------------------------===//
// Legalization Utilities
//===----------------------------------------------------------------------===//
bool TargetLowering::expandMUL_LOHI(unsigned Opcode, EVT VT, const SDLoc &dl,
SDValue LHS, SDValue RHS,
SmallVectorImpl<SDValue> &Result,
EVT HiLoVT, SelectionDAG &DAG,
MulExpansionKind Kind, SDValue LL,
SDValue LH, SDValue RL, SDValue RH) const {
assert(Opcode == ISD::MUL || Opcode == ISD::UMUL_LOHI ||
Opcode == ISD::SMUL_LOHI);
bool HasMULHS = (Kind == MulExpansionKind::Always) ||
isOperationLegalOrCustom(ISD::MULHS, HiLoVT);
bool HasMULHU = (Kind == MulExpansionKind::Always) ||
isOperationLegalOrCustom(ISD::MULHU, HiLoVT);
bool HasSMUL_LOHI = (Kind == MulExpansionKind::Always) ||
isOperationLegalOrCustom(ISD::SMUL_LOHI, HiLoVT);
bool HasUMUL_LOHI = (Kind == MulExpansionKind::Always) ||
isOperationLegalOrCustom(ISD::UMUL_LOHI, HiLoVT);
if (!HasMULHU && !HasMULHS && !HasUMUL_LOHI && !HasSMUL_LOHI)
return false;
unsigned OuterBitSize = VT.getScalarSizeInBits();
unsigned InnerBitSize = HiLoVT.getScalarSizeInBits();
// LL, LH, RL, and RH must be either all NULL or all set to a value.
assert((LL.getNode() && LH.getNode() && RL.getNode() && RH.getNode()) ||
(!LL.getNode() && !LH.getNode() && !RL.getNode() && !RH.getNode()));
SDVTList VTs = DAG.getVTList(HiLoVT, HiLoVT);
auto MakeMUL_LOHI = [&](SDValue L, SDValue R, SDValue &Lo, SDValue &Hi,
bool Signed) -> bool {
if ((Signed && HasSMUL_LOHI) || (!Signed && HasUMUL_LOHI)) {
Lo = DAG.getNode(Signed ? ISD::SMUL_LOHI : ISD::UMUL_LOHI, dl, VTs, L, R);
Hi = SDValue(Lo.getNode(), 1);
return true;
}
if ((Signed && HasMULHS) || (!Signed && HasMULHU)) {
Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, L, R);
Hi = DAG.getNode(Signed ? ISD::MULHS : ISD::MULHU, dl, HiLoVT, L, R);
return true;
}
return false;
};
SDValue Lo, Hi;
if (!LL.getNode() && !RL.getNode() &&
isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) {
LL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, LHS);
RL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, RHS);
}
if (!LL.getNode())
return false;
APInt HighMask = APInt::getHighBitsSet(OuterBitSize, InnerBitSize);
if (DAG.MaskedValueIsZero(LHS, HighMask) &&
DAG.MaskedValueIsZero(RHS, HighMask)) {
// The inputs are both zero-extended.
if (MakeMUL_LOHI(LL, RL, Lo, Hi, false)) {
Result.push_back(Lo);
Result.push_back(Hi);
if (Opcode != ISD::MUL) {
SDValue Zero = DAG.getConstant(0, dl, HiLoVT);
Result.push_back(Zero);
Result.push_back(Zero);
}
return true;
}
}
if (!VT.isVector() && Opcode == ISD::MUL &&
DAG.ComputeMaxSignificantBits(LHS) <= InnerBitSize &&
DAG.ComputeMaxSignificantBits(RHS) <= InnerBitSize) {
// The input values are both sign-extended.
// TODO non-MUL case?
if (MakeMUL_LOHI(LL, RL, Lo, Hi, true)) {
Result.push_back(Lo);
Result.push_back(Hi);
return true;
}
}
unsigned ShiftAmount = OuterBitSize - InnerBitSize;
SDValue Shift = DAG.getShiftAmountConstant(ShiftAmount, VT, dl);
if (!LH.getNode() && !RH.getNode() &&
isOperationLegalOrCustom(ISD::SRL, VT) &&
isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) {
LH = DAG.getNode(ISD::SRL, dl, VT, LHS, Shift);
LH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, LH);
RH = DAG.getNode(ISD::SRL, dl, VT, RHS, Shift);
RH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, RH);
}
if (!LH.getNode())
return false;
if (!MakeMUL_LOHI(LL, RL, Lo, Hi, false))
return false;
Result.push_back(Lo);
if (Opcode == ISD::MUL) {
RH = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RH);
LH = DAG.getNode(ISD::MUL, dl, HiLoVT, LH, RL);
Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, RH);
Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, LH);
Result.push_back(Hi);
return true;
}
// Compute the full width result.
auto Merge = [&](SDValue Lo, SDValue Hi) -> SDValue {
Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
Hi = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Hi);
Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
return DAG.getNode(ISD::OR, dl, VT, Lo, Hi);
};
SDValue Next = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Hi);
if (!MakeMUL_LOHI(LL, RH, Lo, Hi, false))
return false;
// This is effectively the add part of a multiply-add of half-sized operands,
// so it cannot overflow.
Next = DAG.getNode(ISD::ADD, dl, VT, Next, Merge(Lo, Hi));
if (!MakeMUL_LOHI(LH, RL, Lo, Hi, false))
return false;
SDValue Zero = DAG.getConstant(0, dl, HiLoVT);
EVT BoolType = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
bool UseGlue = (isOperationLegalOrCustom(ISD::ADDC, VT) &&
isOperationLegalOrCustom(ISD::ADDE, VT));
if (UseGlue)
Next = DAG.getNode(ISD::ADDC, dl, DAG.getVTList(VT, MVT::Glue), Next,
Merge(Lo, Hi));
else
Next = DAG.getNode(ISD::UADDO_CARRY, dl, DAG.getVTList(VT, BoolType), Next,
Merge(Lo, Hi), DAG.getConstant(0, dl, BoolType));
SDValue Carry = Next.getValue(1);
Result.push_back(DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, Next));
Next = DAG.getNode(ISD::SRL, dl, VT, Next, Shift);
if (!MakeMUL_LOHI(LH, RH, Lo, Hi, Opcode == ISD::SMUL_LOHI))
return false;
if (UseGlue)
Hi = DAG.getNode(ISD::ADDE, dl, DAG.getVTList(HiLoVT, MVT::Glue), Hi, Zero,
Carry);
else
Hi = DAG.getNode(ISD::UADDO_CARRY, dl, DAG.getVTList(HiLoVT, BoolType), Hi,
Zero, Carry);
Next = DAG.getNode(ISD::ADD, dl, VT, Next, Merge(Lo, Hi));
if (Opcode == ISD::SMUL_LOHI) {
SDValue NextSub = DAG.getNode(ISD::SUB, dl, VT, Next,
DAG.getNode(ISD::ZERO_EXTEND, dl, VT, RL));
Next = DAG.getSelectCC(dl, LH, Zero, NextSub, Next, ISD::SETLT);
NextSub = DAG.getNode(ISD::SUB, dl, VT, Next,
DAG.getNode(ISD::ZERO_EXTEND, dl, VT, LL));
Next = DAG.getSelectCC(dl, RH, Zero, NextSub, Next, ISD::SETLT);
}
Result.push_back(DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, Next));
Next = DAG.getNode(ISD::SRL, dl, VT, Next, Shift);
Result.push_back(DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, Next));
return true;
}
bool TargetLowering::expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT,
SelectionDAG &DAG, MulExpansionKind Kind,
SDValue LL, SDValue LH, SDValue RL,
SDValue RH) const {
SmallVector<SDValue, 2> Result;
bool Ok = expandMUL_LOHI(N->getOpcode(), N->getValueType(0), SDLoc(N),
N->getOperand(0), N->getOperand(1), Result, HiLoVT,
DAG, Kind, LL, LH, RL, RH);
if (Ok) {
assert(Result.size() == 2);
Lo = Result[0];
Hi = Result[1];
}
return Ok;
}
// Optimize unsigned division or remainder by constants for types twice as large
// as a legal VT.
//
// If (1 << (BitWidth / 2)) % Constant == 1, then the remainder
// can be computed
// as:
// Sum += __builtin_uadd_overflow(Lo, High, &Sum);
// Remainder = Sum % Constant
// This is based on "Remainder by Summing Digits" from Hacker's Delight.
//
// For division, we can compute the remainder using the algorithm described
// above, subtract it from the dividend to get an exact multiple of Constant.
// Then multiply that exact multiply by the multiplicative inverse modulo
// (1 << (BitWidth / 2)) to get the quotient.
// If Constant is even, we can shift right the dividend and the divisor by the
// number of trailing zeros in Constant before applying the remainder algorithm.
// If we're after the quotient, we can subtract this value from the shifted
// dividend and multiply by the multiplicative inverse of the shifted divisor.
// If we want the remainder, we shift the value left by the number of trailing
// zeros and add the bits that were shifted out of the dividend.
bool TargetLowering::expandDIVREMByConstant(SDNode *N,
SmallVectorImpl<SDValue> &Result,
EVT HiLoVT, SelectionDAG &DAG,
SDValue LL, SDValue LH) const {
unsigned Opcode = N->getOpcode();
EVT VT = N->getValueType(0);
// TODO: Support signed division/remainder.
if (Opcode == ISD::SREM || Opcode == ISD::SDIV || Opcode == ISD::SDIVREM)
return false;
assert(
(Opcode == ISD::UREM || Opcode == ISD::UDIV || Opcode == ISD::UDIVREM) &&
"Unexpected opcode");
auto *CN = dyn_cast<ConstantSDNode>(N->getOperand(1));
if (!CN)
return false;
APInt Divisor = CN->getAPIntValue();
unsigned BitWidth = Divisor.getBitWidth();
unsigned HBitWidth = BitWidth / 2;
assert(VT.getScalarSizeInBits() == BitWidth &&
HiLoVT.getScalarSizeInBits() == HBitWidth && "Unexpected VTs");
// Divisor needs to less than (1 << HBitWidth).
APInt HalfMaxPlus1 = APInt::getOneBitSet(BitWidth, HBitWidth);
if (Divisor.uge(HalfMaxPlus1))
return false;
// We depend on the UREM by constant optimization in DAGCombiner that requires
// high multiply.
if (!isOperationLegalOrCustom(ISD::MULHU, HiLoVT) &&
!isOperationLegalOrCustom(ISD::UMUL_LOHI, HiLoVT))
return false;
// Don't expand if optimizing for size.
if (DAG.shouldOptForSize())
return false;
// Early out for 0 or 1 divisors.
if (Divisor.ule(1))
return false;
// If the divisor is even, shift it until it becomes odd.
unsigned TrailingZeros = 0;
if (!Divisor[0]) {
TrailingZeros = Divisor.countr_zero();
Divisor.lshrInPlace(TrailingZeros);
}
SDLoc dl(N);
SDValue Sum;
SDValue PartialRem;
// If (1 << HBitWidth) % divisor == 1, we can add the two halves together and
// then add in the carry.
// TODO: If we can't split it in half, we might be able to split into 3 or
// more pieces using a smaller bit width.
if (HalfMaxPlus1.urem(Divisor).isOne()) {
assert(!LL == !LH && "Expected both input halves or no input halves!");
if (!LL)
std::tie(LL, LH) = DAG.SplitScalar(N->getOperand(0), dl, HiLoVT, HiLoVT);
// Shift the input by the number of TrailingZeros in the divisor. The
// shifted out bits will be added to the remainder later.
if (TrailingZeros) {
// Save the shifted off bits if we need the remainder.
if (Opcode != ISD::UDIV) {
APInt Mask = APInt::getLowBitsSet(HBitWidth, TrailingZeros);
PartialRem = DAG.getNode(ISD::AND, dl, HiLoVT, LL,
DAG.getConstant(Mask, dl, HiLoVT));
}
LL = DAG.getNode(
ISD::OR, dl, HiLoVT,
DAG.getNode(ISD::SRL, dl, HiLoVT, LL,
DAG.getShiftAmountConstant(TrailingZeros, HiLoVT, dl)),
DAG.getNode(ISD::SHL, dl, HiLoVT, LH,
DAG.getShiftAmountConstant(HBitWidth - TrailingZeros,
HiLoVT, dl)));
LH = DAG.getNode(ISD::SRL, dl, HiLoVT, LH,
DAG.getShiftAmountConstant(TrailingZeros, HiLoVT, dl));
}
// Use uaddo_carry if we can, otherwise use a compare to detect overflow.
EVT SetCCType =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), HiLoVT);
if (isOperationLegalOrCustom(ISD::UADDO_CARRY, HiLoVT)) {
SDVTList VTList = DAG.getVTList(HiLoVT, SetCCType);
Sum = DAG.getNode(ISD::UADDO, dl, VTList, LL, LH);
Sum = DAG.getNode(ISD::UADDO_CARRY, dl, VTList, Sum,
DAG.getConstant(0, dl, HiLoVT), Sum.getValue(1));
} else {
Sum = DAG.getNode(ISD::ADD, dl, HiLoVT, LL, LH);
SDValue Carry = DAG.getSetCC(dl, SetCCType, Sum, LL, ISD::SETULT);
// If the boolean for the target is 0 or 1, we can add the setcc result
// directly.
if (getBooleanContents(HiLoVT) ==
TargetLoweringBase::ZeroOrOneBooleanContent)
Carry = DAG.getZExtOrTrunc(Carry, dl, HiLoVT);
else
Carry = DAG.getSelect(dl, HiLoVT, Carry, DAG.getConstant(1, dl, HiLoVT),
DAG.getConstant(0, dl, HiLoVT));
Sum = DAG.getNode(ISD::ADD, dl, HiLoVT, Sum, Carry);
}
}
// If we didn't find a sum, we can't do the expansion.
if (!Sum)
return false;
// Perform a HiLoVT urem on the Sum using truncated divisor.
SDValue RemL =
DAG.getNode(ISD::UREM, dl, HiLoVT, Sum,
DAG.getConstant(Divisor.trunc(HBitWidth), dl, HiLoVT));
SDValue RemH = DAG.getConstant(0, dl, HiLoVT);
if (Opcode != ISD::UREM) {
// Subtract the remainder from the shifted dividend.
SDValue Dividend = DAG.getNode(ISD::BUILD_PAIR, dl, VT, LL, LH);
SDValue Rem = DAG.getNode(ISD::BUILD_PAIR, dl, VT, RemL, RemH);
Dividend = DAG.getNode(ISD::SUB, dl, VT, Dividend, Rem);
// Multiply by the multiplicative inverse of the divisor modulo
// (1 << BitWidth).
APInt MulFactor = Divisor.multiplicativeInverse();
SDValue Quotient = DAG.getNode(ISD::MUL, dl, VT, Dividend,
DAG.getConstant(MulFactor, dl, VT));
// Split the quotient into low and high parts.
SDValue QuotL, QuotH;
std::tie(QuotL, QuotH) = DAG.SplitScalar(Quotient, dl, HiLoVT, HiLoVT);
Result.push_back(QuotL);
Result.push_back(QuotH);
}
if (Opcode != ISD::UDIV) {
// If we shifted the input, shift the remainder left and add the bits we
// shifted off the input.
if (TrailingZeros) {
APInt Mask = APInt::getLowBitsSet(HBitWidth, TrailingZeros);
RemL = DAG.getNode(ISD::SHL, dl, HiLoVT, RemL,
DAG.getShiftAmountConstant(TrailingZeros, HiLoVT, dl));
RemL = DAG.getNode(ISD::ADD, dl, HiLoVT, RemL, PartialRem);
}
Result.push_back(RemL);
Result.push_back(DAG.getConstant(0, dl, HiLoVT));
}
return true;
}
// Check that (every element of) Z is undef or not an exact multiple of BW.
static bool isNonZeroModBitWidthOrUndef(SDValue Z, unsigned BW) {
return ISD::matchUnaryPredicate(
Z,
[=](ConstantSDNode *C) { return !C || C->getAPIntValue().urem(BW) != 0; },
true);
}
static SDValue expandVPFunnelShift(SDNode *Node, SelectionDAG &DAG) {
EVT VT = Node->getValueType(0);
SDValue ShX, ShY;
SDValue ShAmt, InvShAmt;
SDValue X = Node->getOperand(0);
SDValue Y = Node->getOperand(1);
SDValue Z = Node->getOperand(2);
SDValue Mask = Node->getOperand(3);
SDValue VL = Node->getOperand(4);
unsigned BW = VT.getScalarSizeInBits();
bool IsFSHL = Node->getOpcode() == ISD::VP_FSHL;
SDLoc DL(SDValue(Node, 0));
EVT ShVT = Z.getValueType();
if (isNonZeroModBitWidthOrUndef(Z, BW)) {
// fshl: X << C | Y >> (BW - C)
// fshr: X << (BW - C) | Y >> C
// where C = Z % BW is not zero
SDValue BitWidthC = DAG.getConstant(BW, DL, ShVT);
ShAmt = DAG.getNode(ISD::VP_UREM, DL, ShVT, Z, BitWidthC, Mask, VL);
InvShAmt = DAG.getNode(ISD::VP_SUB, DL, ShVT, BitWidthC, ShAmt, Mask, VL);
ShX = DAG.getNode(ISD::VP_SHL, DL, VT, X, IsFSHL ? ShAmt : InvShAmt, Mask,
VL);
ShY = DAG.getNode(ISD::VP_SRL, DL, VT, Y, IsFSHL ? InvShAmt : ShAmt, Mask,
VL);
} else {
// fshl: X << (Z % BW) | Y >> 1 >> (BW - 1 - (Z % BW))
// fshr: X << 1 << (BW - 1 - (Z % BW)) | Y >> (Z % BW)
SDValue BitMask = DAG.getConstant(BW - 1, DL, ShVT);
if (isPowerOf2_32(BW)) {
// Z % BW -> Z & (BW - 1)
ShAmt = DAG.getNode(ISD::VP_AND, DL, ShVT, Z, BitMask, Mask, VL);
// (BW - 1) - (Z % BW) -> ~Z & (BW - 1)
SDValue NotZ = DAG.getNode(ISD::VP_XOR, DL, ShVT, Z,
DAG.getAllOnesConstant(DL, ShVT), Mask, VL);
InvShAmt = DAG.getNode(ISD::VP_AND, DL, ShVT, NotZ, BitMask, Mask, VL);
} else {
SDValue BitWidthC = DAG.getConstant(BW, DL, ShVT);
ShAmt = DAG.getNode(ISD::VP_UREM, DL, ShVT, Z, BitWidthC, Mask, VL);
InvShAmt = DAG.getNode(ISD::VP_SUB, DL, ShVT, BitMask, ShAmt, Mask, VL);
}
SDValue One = DAG.getConstant(1, DL, ShVT);
if (IsFSHL) {
ShX = DAG.getNode(ISD::VP_SHL, DL, VT, X, ShAmt, Mask, VL);
SDValue ShY1 = DAG.getNode(ISD::VP_SRL, DL, VT, Y, One, Mask, VL);
ShY = DAG.getNode(ISD::VP_SRL, DL, VT, ShY1, InvShAmt, Mask, VL);
} else {
SDValue ShX1 = DAG.getNode(ISD::VP_SHL, DL, VT, X, One, Mask, VL);
ShX = DAG.getNode(ISD::VP_SHL, DL, VT, ShX1, InvShAmt, Mask, VL);
ShY = DAG.getNode(ISD::VP_SRL, DL, VT, Y, ShAmt, Mask, VL);
}
}
return DAG.getNode(ISD::VP_OR, DL, VT, ShX, ShY, Mask, VL);
}
SDValue TargetLowering::expandFunnelShift(SDNode *Node,
SelectionDAG &DAG) const {
if (Node->isVPOpcode())
return expandVPFunnelShift(Node, DAG);
EVT VT = Node->getValueType(0);
if (VT.isVector() && (!isOperationLegalOrCustom(ISD::SHL, VT) ||
!isOperationLegalOrCustom(ISD::SRL, VT) ||
!isOperationLegalOrCustom(ISD::SUB, VT) ||
!isOperationLegalOrCustomOrPromote(ISD::OR, VT)))
return SDValue();
SDValue X = Node->getOperand(0);
SDValue Y = Node->getOperand(1);
SDValue Z = Node->getOperand(2);
unsigned BW = VT.getScalarSizeInBits();
bool IsFSHL = Node->getOpcode() == ISD::FSHL;
SDLoc DL(SDValue(Node, 0));
EVT ShVT = Z.getValueType();
// If a funnel shift in the other direction is more supported, use it.
unsigned RevOpcode = IsFSHL ? ISD::FSHR : ISD::FSHL;
if (!isOperationLegalOrCustom(Node->getOpcode(), VT) &&
isOperationLegalOrCustom(RevOpcode, VT) && isPowerOf2_32(BW)) {
if (isNonZeroModBitWidthOrUndef(Z, BW)) {
// fshl X, Y, Z -> fshr X, Y, -Z
// fshr X, Y, Z -> fshl X, Y, -Z
SDValue Zero = DAG.getConstant(0, DL, ShVT);
Z = DAG.getNode(ISD::SUB, DL, VT, Zero, Z);
} else {
// fshl X, Y, Z -> fshr (srl X, 1), (fshr X, Y, 1), ~Z
// fshr X, Y, Z -> fshl (fshl X, Y, 1), (shl Y, 1), ~Z
SDValue One = DAG.getConstant(1, DL, ShVT);
if (IsFSHL) {
Y = DAG.getNode(RevOpcode, DL, VT, X, Y, One);
X = DAG.getNode(ISD::SRL, DL, VT, X, One);
} else {
X = DAG.getNode(RevOpcode, DL, VT, X, Y, One);
Y = DAG.getNode(ISD::SHL, DL, VT, Y, One);
}
Z = DAG.getNOT(DL, Z, ShVT);
}
return DAG.getNode(RevOpcode, DL, VT, X, Y, Z);
}
SDValue ShX, ShY;
SDValue ShAmt, InvShAmt;
if (isNonZeroModBitWidthOrUndef(Z, BW)) {
// fshl: X << C | Y >> (BW - C)
// fshr: X << (BW - C) | Y >> C
// where C = Z % BW is not zero
SDValue BitWidthC = DAG.getConstant(BW, DL, ShVT);
ShAmt = DAG.getNode(ISD::UREM, DL, ShVT, Z, BitWidthC);
InvShAmt = DAG.getNode(ISD::SUB, DL, ShVT, BitWidthC, ShAmt);
ShX = DAG.getNode(ISD::SHL, DL, VT, X, IsFSHL ? ShAmt : InvShAmt);
ShY = DAG.getNode(ISD::SRL, DL, VT, Y, IsFSHL ? InvShAmt : ShAmt);
} else {
// fshl: X << (Z % BW) | Y >> 1 >> (BW - 1 - (Z % BW))
// fshr: X << 1 << (BW - 1 - (Z % BW)) | Y >> (Z % BW)
SDValue Mask = DAG.getConstant(BW - 1, DL, ShVT);
if (isPowerOf2_32(BW)) {
// Z % BW -> Z & (BW - 1)
ShAmt = DAG.getNode(ISD::AND, DL, ShVT, Z, Mask);
// (BW - 1) - (Z % BW) -> ~Z & (BW - 1)
InvShAmt = DAG.getNode(ISD::AND, DL, ShVT, DAG.getNOT(DL, Z, ShVT), Mask);
} else {
SDValue BitWidthC = DAG.getConstant(BW, DL, ShVT);
ShAmt = DAG.getNode(ISD::UREM, DL, ShVT, Z, BitWidthC);
InvShAmt = DAG.getNode(ISD::SUB, DL, ShVT, Mask, ShAmt);
}
SDValue One = DAG.getConstant(1, DL, ShVT);
if (IsFSHL) {
ShX = DAG.getNode(ISD::SHL, DL, VT, X, ShAmt);
SDValue ShY1 = DAG.getNode(ISD::SRL, DL, VT, Y, One);
ShY = DAG.getNode(ISD::SRL, DL, VT, ShY1, InvShAmt);
} else {
SDValue ShX1 = DAG.getNode(ISD::SHL, DL, VT, X, One);
ShX = DAG.getNode(ISD::SHL, DL, VT, ShX1, InvShAmt);
ShY = DAG.getNode(ISD::SRL, DL, VT, Y, ShAmt);
}
}
return DAG.getNode(ISD::OR, DL, VT, ShX, ShY);
}
// TODO: Merge with expandFunnelShift.
SDValue TargetLowering::expandROT(SDNode *Node, bool AllowVectorOps,
SelectionDAG &DAG) const {
EVT VT = Node->getValueType(0);
unsigned EltSizeInBits = VT.getScalarSizeInBits();
bool IsLeft = Node->getOpcode() == ISD::ROTL;
SDValue Op0 = Node->getOperand(0);
SDValue Op1 = Node->getOperand(1);
SDLoc DL(SDValue(Node, 0));
EVT ShVT = Op1.getValueType();
SDValue Zero = DAG.getConstant(0, DL, ShVT);
// If a rotate in the other direction is more supported, use it.
unsigned RevRot = IsLeft ? ISD::ROTR : ISD::ROTL;
if (!isOperationLegalOrCustom(Node->getOpcode(), VT) &&
isOperationLegalOrCustom(RevRot, VT) && isPowerOf2_32(EltSizeInBits)) {
SDValue Sub = DAG.getNode(ISD::SUB, DL, ShVT, Zero, Op1);
return DAG.getNode(RevRot, DL, VT, Op0, Sub);
}
if (!AllowVectorOps && VT.isVector() &&
(!isOperationLegalOrCustom(ISD::SHL, VT) ||
!isOperationLegalOrCustom(ISD::SRL, VT) ||
!isOperationLegalOrCustom(ISD::SUB, VT) ||
!isOperationLegalOrCustomOrPromote(ISD::OR, VT) ||
!isOperationLegalOrCustomOrPromote(ISD::AND, VT)))
return SDValue();
unsigned ShOpc = IsLeft ? ISD::SHL : ISD::SRL;
unsigned HsOpc = IsLeft ? ISD::SRL : ISD::SHL;
SDValue BitWidthMinusOneC = DAG.getConstant(EltSizeInBits - 1, DL, ShVT);
SDValue ShVal;
SDValue HsVal;
if (isPowerOf2_32(EltSizeInBits)) {
// (rotl x, c) -> x << (c & (w - 1)) | x >> (-c & (w - 1))
// (rotr x, c) -> x >> (c & (w - 1)) | x << (-c & (w - 1))
SDValue NegOp1 = DAG.getNode(ISD::SUB, DL, ShVT, Zero, Op1);
SDValue ShAmt = DAG.getNode(ISD::AND, DL, ShVT, Op1, BitWidthMinusOneC);
ShVal = DAG.getNode(ShOpc, DL, VT, Op0, ShAmt);
SDValue HsAmt = DAG.getNode(ISD::AND, DL, ShVT, NegOp1, BitWidthMinusOneC);
HsVal = DAG.getNode(HsOpc, DL, VT, Op0, HsAmt);
} else {
// (rotl x, c) -> x << (c % w) | x >> 1 >> (w - 1 - (c % w))
// (rotr x, c) -> x >> (c % w) | x << 1 << (w - 1 - (c % w))
SDValue BitWidthC = DAG.getConstant(EltSizeInBits, DL, ShVT);
SDValue ShAmt = DAG.getNode(ISD::UREM, DL, ShVT, Op1, BitWidthC);
ShVal = DAG.getNode(ShOpc, DL, VT, Op0, ShAmt);
SDValue HsAmt = DAG.getNode(ISD::SUB, DL, ShVT, BitWidthMinusOneC, ShAmt);
SDValue One = DAG.getConstant(1, DL, ShVT);
HsVal =
DAG.getNode(HsOpc, DL, VT, DAG.getNode(HsOpc, DL, VT, Op0, One), HsAmt);
}
return DAG.getNode(ISD::OR, DL, VT, ShVal, HsVal);
}
void TargetLowering::expandShiftParts(SDNode *Node, SDValue &Lo, SDValue &Hi,
SelectionDAG &DAG) const {
assert(Node->getNumOperands() == 3 && "Not a double-shift!");
EVT VT = Node->getValueType(0);
unsigned VTBits = VT.getScalarSizeInBits();
assert(isPowerOf2_32(VTBits) && "Power-of-two integer type expected");
bool IsSHL = Node->getOpcode() == ISD::SHL_PARTS;
bool IsSRA = Node->getOpcode() == ISD::SRA_PARTS;
SDValue ShOpLo = Node->getOperand(0);
SDValue ShOpHi = Node->getOperand(1);
SDValue ShAmt = Node->getOperand(2);
EVT ShAmtVT = ShAmt.getValueType();
EVT ShAmtCCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), ShAmtVT);
SDLoc dl(Node);
// ISD::FSHL and ISD::FSHR have defined overflow behavior but ISD::SHL and
// ISD::SRA/L nodes haven't. Insert an AND to be safe, it's usually optimized
// away during isel.
SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, ShAmtVT, ShAmt,
DAG.getConstant(VTBits - 1, dl, ShAmtVT));
SDValue Tmp1 = IsSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
DAG.getConstant(VTBits - 1, dl, ShAmtVT))
: DAG.getConstant(0, dl, VT);
SDValue Tmp2, Tmp3;
if (IsSHL) {
Tmp2 = DAG.getNode(ISD::FSHL, dl, VT, ShOpHi, ShOpLo, ShAmt);
Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
} else {
Tmp2 = DAG.getNode(ISD::FSHR, dl, VT, ShOpHi, ShOpLo, ShAmt);
Tmp3 = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
}
// If the shift amount is larger or equal than the width of a part we don't
// use the result from the FSHL/FSHR. Insert a test and select the appropriate
// values for large shift amounts.
SDValue AndNode = DAG.getNode(ISD::AND, dl, ShAmtVT, ShAmt,
DAG.getConstant(VTBits, dl, ShAmtVT));
SDValue Cond = DAG.getSetCC(dl, ShAmtCCVT, AndNode,
DAG.getConstant(0, dl, ShAmtVT), ISD::SETNE);
if (IsSHL) {
Hi = DAG.getNode(ISD::SELECT, dl, VT, Cond, Tmp3, Tmp2);
Lo = DAG.getNode(ISD::SELECT, dl, VT, Cond, Tmp1, Tmp3);
} else {
Lo = DAG.getNode(ISD::SELECT, dl, VT, Cond, Tmp3, Tmp2);
Hi = DAG.getNode(ISD::SELECT, dl, VT, Cond, Tmp1, Tmp3);
}
}
bool TargetLowering::expandFP_TO_SINT(SDNode *Node, SDValue &Result,
SelectionDAG &DAG) const {
unsigned OpNo = Node->isStrictFPOpcode() ? 1 : 0;
SDValue Src = Node->getOperand(OpNo);
EVT SrcVT = Src.getValueType();
EVT DstVT = Node->getValueType(0);
SDLoc dl(SDValue(Node, 0));
// FIXME: Only f32 to i64 conversions are supported.
if (SrcVT != MVT::f32 || DstVT != MVT::i64)
return false;
if (Node->isStrictFPOpcode())
// When a NaN is converted to an integer a trap is allowed. We can't
// use this expansion here because it would eliminate that trap. Other
// traps are also allowed and cannot be eliminated. See
// IEEE 754-2008 sec 5.8.
return false;
// Expand f32 -> i64 conversion
// This algorithm comes from compiler-rt's implementation of fixsfdi:
// https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/builtins/fixsfdi.c
unsigned SrcEltBits = SrcVT.getScalarSizeInBits();
EVT IntVT = SrcVT.changeTypeToInteger();
EVT IntShVT = getShiftAmountTy(IntVT, DAG.getDataLayout());
SDValue ExponentMask = DAG.getConstant(0x7F800000, dl, IntVT);
SDValue ExponentLoBit = DAG.getConstant(23, dl, IntVT);
SDValue Bias = DAG.getConstant(127, dl, IntVT);
SDValue SignMask = DAG.getConstant(APInt::getSignMask(SrcEltBits), dl, IntVT);
SDValue SignLowBit = DAG.getConstant(SrcEltBits - 1, dl, IntVT);
SDValue MantissaMask = DAG.getConstant(0x007FFFFF, dl, IntVT);
SDValue Bits = DAG.getNode(ISD::BITCAST, dl, IntVT, Src);
SDValue ExponentBits = DAG.getNode(
ISD::SRL, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, ExponentMask),
DAG.getZExtOrTrunc(ExponentLoBit, dl, IntShVT));
SDValue Exponent = DAG.getNode(ISD::SUB, dl, IntVT, ExponentBits, Bias);
SDValue Sign = DAG.getNode(ISD::SRA, dl, IntVT,
DAG.getNode(ISD::AND, dl, IntVT, Bits, SignMask),
DAG.getZExtOrTrunc(SignLowBit, dl, IntShVT));
Sign = DAG.getSExtOrTrunc(Sign, dl, DstVT);
SDValue R = DAG.getNode(ISD::OR, dl, IntVT,
DAG.getNode(ISD::AND, dl, IntVT, Bits, MantissaMask),
DAG.getConstant(0x00800000, dl, IntVT));
R = DAG.getZExtOrTrunc(R, dl, DstVT);
R = DAG.getSelectCC(
dl, Exponent, ExponentLoBit,
DAG.getNode(ISD::SHL, dl, DstVT, R,
DAG.getZExtOrTrunc(
DAG.getNode(ISD::SUB, dl, IntVT, Exponent, ExponentLoBit),
dl, IntShVT)),
DAG.getNode(ISD::SRL, dl, DstVT, R,
DAG.getZExtOrTrunc(
DAG.getNode(ISD::SUB, dl, IntVT, ExponentLoBit, Exponent),
dl, IntShVT)),
ISD::SETGT);
SDValue Ret = DAG.getNode(ISD::SUB, dl, DstVT,
DAG.getNode(ISD::XOR, dl, DstVT, R, Sign), Sign);
Result = DAG.getSelectCC(dl, Exponent, DAG.getConstant(0, dl, IntVT),
DAG.getConstant(0, dl, DstVT), Ret, ISD::SETLT);
return true;
}
bool TargetLowering::expandFP_TO_UINT(SDNode *Node, SDValue &Result,
SDValue &Chain,
SelectionDAG &DAG) const {
SDLoc dl(SDValue(Node, 0));
unsigned OpNo = Node->isStrictFPOpcode() ? 1 : 0;
SDValue Src = Node->getOperand(OpNo);
EVT SrcVT = Src.getValueType();
EVT DstVT = Node->getValueType(0);
EVT SetCCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT);
EVT DstSetCCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), DstVT);
// Only expand vector types if we have the appropriate vector bit operations.
unsigned SIntOpcode = Node->isStrictFPOpcode() ? ISD::STRICT_FP_TO_SINT :
ISD::FP_TO_SINT;
if (DstVT.isVector() && (!isOperationLegalOrCustom(SIntOpcode, DstVT) ||
!isOperationLegalOrCustomOrPromote(ISD::XOR, SrcVT)))
return false;
// If the maximum float value is smaller then the signed integer range,
// the destination signmask can't be represented by the float, so we can
// just use FP_TO_SINT directly.
const fltSemantics &APFSem = DAG.EVTToAPFloatSemantics(SrcVT);
APFloat APF(APFSem, APInt::getZero(SrcVT.getScalarSizeInBits()));
APInt SignMask = APInt::getSignMask(DstVT.getScalarSizeInBits());
if (APFloat::opOverflow &
APF.convertFromAPInt(SignMask, false, APFloat::rmNearestTiesToEven)) {
if (Node->isStrictFPOpcode()) {
Result = DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, { DstVT, MVT::Other },
{ Node->getOperand(0), Src });
Chain = Result.getValue(1);
} else
Result = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, Src);
return true;
}
// Don't expand it if there isn't cheap fsub instruction.
if (!isOperationLegalOrCustom(
Node->isStrictFPOpcode() ? ISD::STRICT_FSUB : ISD::FSUB, SrcVT))
return false;
SDValue Cst = DAG.getConstantFP(APF, dl, SrcVT);
SDValue Sel;
if (Node->isStrictFPOpcode()) {
Sel = DAG.getSetCC(dl, SetCCVT, Src, Cst, ISD::SETLT,
Node->getOperand(0), /*IsSignaling*/ true);
Chain = Sel.getValue(1);
} else {
Sel = DAG.getSetCC(dl, SetCCVT, Src, Cst, ISD::SETLT);
}
bool Strict = Node->isStrictFPOpcode() ||
shouldUseStrictFP_TO_INT(SrcVT, DstVT, /*IsSigned*/ false);
if (Strict) {
// Expand based on maximum range of FP_TO_SINT, if the value exceeds the
// signmask then offset (the result of which should be fully representable).
// Sel = Src < 0x8000000000000000
// FltOfs = select Sel, 0, 0x8000000000000000
// IntOfs = select Sel, 0, 0x8000000000000000
// Result = fp_to_sint(Src - FltOfs) ^ IntOfs
// TODO: Should any fast-math-flags be set for the FSUB?
SDValue FltOfs = DAG.getSelect(dl, SrcVT, Sel,
DAG.getConstantFP(0.0, dl, SrcVT), Cst);
Sel = DAG.getBoolExtOrTrunc(Sel, dl, DstSetCCVT, DstVT);
SDValue IntOfs = DAG.getSelect(dl, DstVT, Sel,
DAG.getConstant(0, dl, DstVT),
DAG.getConstant(SignMask, dl, DstVT));
SDValue SInt;
if (Node->isStrictFPOpcode()) {
SDValue Val = DAG.getNode(ISD::STRICT_FSUB, dl, { SrcVT, MVT::Other },
{ Chain, Src, FltOfs });
SInt = DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, { DstVT, MVT::Other },
{ Val.getValue(1), Val });
Chain = SInt.getValue(1);
} else {
SDValue Val = DAG.getNode(ISD::FSUB, dl, SrcVT, Src, FltOfs);
SInt = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, Val);
}
Result = DAG.getNode(ISD::XOR, dl, DstVT, SInt, IntOfs);
} else {
// Expand based on maximum range of FP_TO_SINT:
// True = fp_to_sint(Src)
// False = 0x8000000000000000 + fp_to_sint(Src - 0x8000000000000000)
// Result = select (Src < 0x8000000000000000), True, False
SDValue True = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, Src);
// TODO: Should any fast-math-flags be set for the FSUB?
SDValue False = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT,
DAG.getNode(ISD::FSUB, dl, SrcVT, Src, Cst));
False = DAG.getNode(ISD::XOR, dl, DstVT, False,
DAG.getConstant(SignMask, dl, DstVT));
Sel = DAG.getBoolExtOrTrunc(Sel, dl, DstSetCCVT, DstVT);
Result = DAG.getSelect(dl, DstVT, Sel, True, False);
}
return true;
}
bool TargetLowering::expandUINT_TO_FP(SDNode *Node, SDValue &Result,
SDValue &Chain,
SelectionDAG &DAG) const {
// This transform is not correct for converting 0 when rounding mode is set
// to round toward negative infinity which will produce -0.0. So disable under
// strictfp.
if (Node->isStrictFPOpcode())
return false;
SDValue Src = Node->getOperand(0);
EVT SrcVT = Src.getValueType();
EVT DstVT = Node->getValueType(0);
if (SrcVT.getScalarType() != MVT::i64 || DstVT.getScalarType() != MVT::f64)
return false;
// Only expand vector types if we have the appropriate vector bit operations.
if (SrcVT.isVector() && (!isOperationLegalOrCustom(ISD::SRL, SrcVT) ||
!isOperationLegalOrCustom(ISD::FADD, DstVT) ||
!isOperationLegalOrCustom(ISD::FSUB, DstVT) ||
!isOperationLegalOrCustomOrPromote(ISD::OR, SrcVT) ||
!isOperationLegalOrCustomOrPromote(ISD::AND, SrcVT)))
return false;
SDLoc dl(SDValue(Node, 0));
EVT ShiftVT = getShiftAmountTy(SrcVT, DAG.getDataLayout());
// Implementation of unsigned i64 to f64 following the algorithm in
// __floatundidf in compiler_rt. This implementation performs rounding
// correctly in all rounding modes with the exception of converting 0
// when rounding toward negative infinity. In that case the fsub will produce
// -0.0. This will be added to +0.0 and produce -0.0 which is incorrect.
SDValue TwoP52 = DAG.getConstant(UINT64_C(0x4330000000000000), dl, SrcVT);
SDValue TwoP84PlusTwoP52 = DAG.getConstantFP(
llvm::bit_cast<double>(UINT64_C(0x4530000000100000)), dl, DstVT);
SDValue TwoP84 = DAG.getConstant(UINT64_C(0x4530000000000000), dl, SrcVT);
SDValue LoMask = DAG.getConstant(UINT64_C(0x00000000FFFFFFFF), dl, SrcVT);
SDValue HiShift = DAG.getConstant(32, dl, ShiftVT);
SDValue Lo = DAG.getNode(ISD::AND, dl, SrcVT, Src, LoMask);
SDValue Hi = DAG.getNode(ISD::SRL, dl, SrcVT, Src, HiShift);
SDValue LoOr = DAG.getNode(ISD::OR, dl, SrcVT, Lo, TwoP52);
SDValue HiOr = DAG.getNode(ISD::OR, dl, SrcVT, Hi, TwoP84);
SDValue LoFlt = DAG.getBitcast(DstVT, LoOr);
SDValue HiFlt = DAG.getBitcast(DstVT, HiOr);
SDValue HiSub =
DAG.getNode(ISD::FSUB, dl, DstVT, HiFlt, TwoP84PlusTwoP52);
Result = DAG.getNode(ISD::FADD, dl, DstVT, LoFlt, HiSub);
return true;
}
SDValue
TargetLowering::createSelectForFMINNUM_FMAXNUM(SDNode *Node,
SelectionDAG &DAG) const {
unsigned Opcode = Node->getOpcode();
assert((Opcode == ISD::FMINNUM || Opcode == ISD::FMAXNUM ||
Opcode == ISD::STRICT_FMINNUM || Opcode == ISD::STRICT_FMAXNUM) &&
"Wrong opcode");
if (Node->getFlags().hasNoNaNs()) {
ISD::CondCode Pred = Opcode == ISD::FMINNUM ? ISD::SETLT : ISD::SETGT;
SDValue Op1 = Node->getOperand(0);
SDValue Op2 = Node->getOperand(1);
SDValue SelCC = DAG.getSelectCC(SDLoc(Node), Op1, Op2, Op1, Op2, Pred);
// Copy FMF flags, but always set the no-signed-zeros flag
// as this is implied by the FMINNUM/FMAXNUM semantics.
SDNodeFlags Flags = Node->getFlags();
Flags.setNoSignedZeros(true);
SelCC->setFlags(Flags);
return SelCC;
}
return SDValue();
}
SDValue TargetLowering::expandFMINNUM_FMAXNUM(SDNode *Node,
SelectionDAG &DAG) const {
SDLoc dl(Node);
unsigned NewOp = Node->getOpcode() == ISD::FMINNUM ?
ISD::FMINNUM_IEEE : ISD::FMAXNUM_IEEE;
EVT VT = Node->getValueType(0);
if (VT.isScalableVector())
report_fatal_error(
"Expanding fminnum/fmaxnum for scalable vectors is undefined.");
if (isOperationLegalOrCustom(NewOp, VT)) {
SDValue Quiet0 = Node->getOperand(0);
SDValue Quiet1 = Node->getOperand(1);
if (!Node->getFlags().hasNoNaNs()) {
// Insert canonicalizes if it's possible we need to quiet to get correct
// sNaN behavior.
if (!DAG.isKnownNeverSNaN(Quiet0)) {
Quiet0 = DAG.getNode(ISD::FCANONICALIZE, dl, VT, Quiet0,
Node->getFlags());
}
if (!DAG.isKnownNeverSNaN(Quiet1)) {
Quiet1 = DAG.getNode(ISD::FCANONICALIZE, dl, VT, Quiet1,
Node->getFlags());
}
}
return DAG.getNode(NewOp, dl, VT, Quiet0, Quiet1, Node->getFlags());
}
// If the target has FMINIMUM/FMAXIMUM but not FMINNUM/FMAXNUM use that
// instead if there are no NaNs and there can't be an incompatible zero
// compare: at least one operand isn't +/-0, or there are no signed-zeros.
if ((Node->getFlags().hasNoNaNs() ||
(DAG.isKnownNeverNaN(Node->getOperand(0)) &&
DAG.isKnownNeverNaN(Node->getOperand(1)))) &&
(Node->getFlags().hasNoSignedZeros() ||
DAG.isKnownNeverZeroFloat(Node->getOperand(0)) ||
DAG.isKnownNeverZeroFloat(Node->getOperand(1)))) {
unsigned IEEE2018Op =
Node->getOpcode() == ISD::FMINNUM ? ISD::FMINIMUM : ISD::FMAXIMUM;
if (isOperationLegalOrCustom(IEEE2018Op, VT))
return DAG.getNode(IEEE2018Op, dl, VT, Node->getOperand(0),
Node->getOperand(1), Node->getFlags());
}
if (SDValue SelCC = createSelectForFMINNUM_FMAXNUM(Node, DAG))
return SelCC;
return SDValue();
}
SDValue TargetLowering::expandFMINIMUM_FMAXIMUM(SDNode *N,
SelectionDAG &DAG) const {
SDLoc DL(N);
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
unsigned Opc = N->getOpcode();
EVT VT = N->getValueType(0);
EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
bool IsMax = Opc == ISD::FMAXIMUM;
SDNodeFlags Flags = N->getFlags();
// First, implement comparison not propagating NaN. If no native fmin or fmax
// available, use plain select with setcc instead.
SDValue MinMax;
unsigned CompOpcIeee = IsMax ? ISD::FMAXNUM_IEEE : ISD::FMINNUM_IEEE;
unsigned CompOpc = IsMax ? ISD::FMAXNUM : ISD::FMINNUM;
// FIXME: We should probably define fminnum/fmaxnum variants with correct
// signed zero behavior.
bool MinMaxMustRespectOrderedZero = false;
if (isOperationLegalOrCustom(CompOpcIeee, VT)) {
MinMax = DAG.getNode(CompOpcIeee, DL, VT, LHS, RHS, Flags);
MinMaxMustRespectOrderedZero = true;
} else if (isOperationLegalOrCustom(CompOpc, VT)) {
MinMax = DAG.getNode(CompOpc, DL, VT, LHS, RHS, Flags);
} else {
if (VT.isVector() && !isOperationLegalOrCustom(ISD::VSELECT, VT))
return DAG.UnrollVectorOp(N);
// NaN (if exists) will be propagated later, so orderness doesn't matter.
SDValue Compare =
DAG.getSetCC(DL, CCVT, LHS, RHS, IsMax ? ISD::SETGT : ISD::SETLT);
MinMax = DAG.getSelect(DL, VT, Compare, LHS, RHS, Flags);
}
// Propagate any NaN of both operands
if (!N->getFlags().hasNoNaNs() &&
(!DAG.isKnownNeverNaN(RHS) || !DAG.isKnownNeverNaN(LHS))) {
ConstantFP *FPNaN = ConstantFP::get(
*DAG.getContext(), APFloat::getNaN(DAG.EVTToAPFloatSemantics(VT)));
MinMax = DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, LHS, RHS, ISD::SETUO),
DAG.getConstantFP(*FPNaN, DL, VT), MinMax, Flags);
}
// fminimum/fmaximum requires -0.0 less than +0.0
if (!MinMaxMustRespectOrderedZero && !N->getFlags().hasNoSignedZeros() &&
!DAG.isKnownNeverZeroFloat(RHS) && !DAG.isKnownNeverZeroFloat(LHS)) {
SDValue IsZero = DAG.getSetCC(DL, CCVT, MinMax,
DAG.getConstantFP(0.0, DL, VT), ISD::SETEQ);
SDValue TestZero =
DAG.getTargetConstant(IsMax ? fcPosZero : fcNegZero, DL, MVT::i32);
SDValue LCmp = DAG.getSelect(
DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, LHS, TestZero), LHS,
MinMax, Flags);
SDValue RCmp = DAG.getSelect(
DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, RHS, TestZero), RHS,
LCmp, Flags);
MinMax = DAG.getSelect(DL, VT, IsZero, RCmp, MinMax, Flags);
}
return MinMax;
}
/// Returns a true value if if this FPClassTest can be performed with an ordered
/// fcmp to 0, and a false value if it's an unordered fcmp to 0. Returns
/// std::nullopt if it cannot be performed as a compare with 0.
static std::optional<bool> isFCmpEqualZero(FPClassTest Test,
const fltSemantics &Semantics,
const MachineFunction &MF) {
FPClassTest OrderedMask = Test & ~fcNan;
FPClassTest NanTest = Test & fcNan;
bool IsOrdered = NanTest == fcNone;
bool IsUnordered = NanTest == fcNan;
// Skip cases that are testing for only a qnan or snan.
if (!IsOrdered && !IsUnordered)
return std::nullopt;
if (OrderedMask == fcZero &&
MF.getDenormalMode(Semantics).Input == DenormalMode::IEEE)
return IsOrdered;
if (OrderedMask == (fcZero | fcSubnormal) &&
MF.getDenormalMode(Semantics).inputsAreZero())
return IsOrdered;
return std::nullopt;
}
SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,
FPClassTest Test, SDNodeFlags Flags,
const SDLoc &DL,
SelectionDAG &DAG) const {
EVT OperandVT = Op.getValueType();
assert(OperandVT.isFloatingPoint());
// Degenerated cases.
if (Test == fcNone)
return DAG.getBoolConstant(false, DL, ResultVT, OperandVT);
if ((Test & fcAllFlags) == fcAllFlags)
return DAG.getBoolConstant(true, DL, ResultVT, OperandVT);
// PPC double double is a pair of doubles, of which the higher part determines
// the value class.
if (OperandVT == MVT::ppcf128) {
Op = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::f64, Op,
DAG.getConstant(1, DL, MVT::i32));
OperandVT = MVT::f64;
}
// Some checks may be represented as inversion of simpler check, for example
// "inf|normal|subnormal|zero" => !"nan".
bool IsInverted = false;
if (FPClassTest InvertedCheck = invertFPClassTestIfSimpler(Test)) {
IsInverted = true;
Test = InvertedCheck;
}
// Floating-point type properties.
EVT ScalarFloatVT = OperandVT.getScalarType();
const Type *FloatTy = ScalarFloatVT.getTypeForEVT(*DAG.getContext());
const llvm::fltSemantics &Semantics = FloatTy->getFltSemantics();
bool IsF80 = (ScalarFloatVT == MVT::f80);
// Some checks can be implemented using float comparisons, if floating point
// exceptions are ignored.
if (Flags.hasNoFPExcept() &&
isOperationLegalOrCustom(ISD::SETCC, OperandVT.getScalarType())) {
ISD::CondCode OrderedCmpOpcode = IsInverted ? ISD::SETUNE : ISD::SETOEQ;
ISD::CondCode UnorderedCmpOpcode = IsInverted ? ISD::SETONE : ISD::SETUEQ;
if (std::optional<bool> IsCmp0 =
isFCmpEqualZero(Test, Semantics, DAG.getMachineFunction());
IsCmp0 && (isCondCodeLegalOrCustom(
*IsCmp0 ? OrderedCmpOpcode : UnorderedCmpOpcode,
OperandVT.getScalarType().getSimpleVT()))) {
// If denormals could be implicitly treated as 0, this is not equivalent
// to a compare with 0 since it will also be true for denormals.
return DAG.getSetCC(DL, ResultVT, Op,
DAG.getConstantFP(0.0, DL, OperandVT),
*IsCmp0 ? OrderedCmpOpcode : UnorderedCmpOpcode);
}
if (Test == fcNan &&
isCondCodeLegalOrCustom(IsInverted ? ISD::SETO : ISD::SETUO,
OperandVT.getScalarType().getSimpleVT())) {
return DAG.getSetCC(DL, ResultVT, Op, Op,
IsInverted ? ISD::SETO : ISD::SETUO);
}
if (Test == fcInf &&
isCondCodeLegalOrCustom(IsInverted ? ISD::SETUNE : ISD::SETOEQ,
OperandVT.getScalarType().getSimpleVT()) &&
isOperationLegalOrCustom(ISD::FABS, OperandVT.getScalarType())) {
// isinf(x) --> fabs(x) == inf
SDValue Abs = DAG.getNode(ISD::FABS, DL, OperandVT, Op);
SDValue Inf =
DAG.getConstantFP(APFloat::getInf(Semantics), DL, OperandVT);
return DAG.getSetCC(DL, ResultVT, Abs, Inf,
IsInverted ? ISD::SETUNE : ISD::SETOEQ);
}
}
// In the general case use integer operations.
unsigned BitSize = OperandVT.getScalarSizeInBits();
EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), BitSize);
if (OperandVT.isVector())
IntVT = EVT::getVectorVT(*DAG.getContext(), IntVT,
OperandVT.getVectorElementCount());
SDValue OpAsInt = DAG.getBitcast(IntVT, Op);
// Various masks.
APInt SignBit = APInt::getSignMask(BitSize);
APInt ValueMask = APInt::getSignedMaxValue(BitSize); // All bits but sign.
APInt Inf = APFloat::getInf(Semantics).bitcastToAPInt(); // Exp and int bit.
const unsigned ExplicitIntBitInF80 = 63;
APInt ExpMask = Inf;
if (IsF80)
ExpMask.clearBit(ExplicitIntBitInF80);
APInt AllOneMantissa = APFloat::getLargest(Semantics).bitcastToAPInt() & ~Inf;
APInt QNaNBitMask =
APInt::getOneBitSet(BitSize, AllOneMantissa.getActiveBits() - 1);
APInt InvertionMask = APInt::getAllOnes(ResultVT.getScalarSizeInBits());
SDValue ValueMaskV = DAG.getConstant(ValueMask, DL, IntVT);
SDValue SignBitV = DAG.getConstant(SignBit, DL, IntVT);
SDValue ExpMaskV = DAG.getConstant(ExpMask, DL, IntVT);
SDValue ZeroV = DAG.getConstant(0, DL, IntVT);
SDValue InfV = DAG.getConstant(Inf, DL, IntVT);
SDValue ResultInvertionMask = DAG.getConstant(InvertionMask, DL, ResultVT);
SDValue Res;
const auto appendResult = [&](SDValue PartialRes) {
if (PartialRes) {
if (Res)
Res = DAG.getNode(ISD::OR, DL, ResultVT, Res, PartialRes);
else
Res = PartialRes;
}
};
SDValue IntBitIsSetV; // Explicit integer bit in f80 mantissa is set.
const auto getIntBitIsSet = [&]() -> SDValue {
if (!IntBitIsSetV) {
APInt IntBitMask(BitSize, 0);
IntBitMask.setBit(ExplicitIntBitInF80);
SDValue IntBitMaskV = DAG.getConstant(IntBitMask, DL, IntVT);
SDValue IntBitV = DAG.getNode(ISD::AND, DL, IntVT, OpAsInt, IntBitMaskV);
IntBitIsSetV = DAG.getSetCC(DL, ResultVT, IntBitV, ZeroV, ISD::SETNE);
}
return IntBitIsSetV;
};
// Split the value into sign bit and absolute value.
SDValue AbsV = DAG.getNode(ISD::AND, DL, IntVT, OpAsInt, ValueMaskV);
SDValue SignV = DAG.getSetCC(DL, ResultVT, OpAsInt,
DAG.getConstant(0.0, DL, IntVT), ISD::SETLT);
// Tests that involve more than one class should be processed first.
SDValue PartialRes;
if (IsF80)
; // Detect finite numbers of f80 by checking individual classes because
// they have different settings of the explicit integer bit.
else if ((Test & fcFinite) == fcFinite) {
// finite(V) ==> abs(V) < exp_mask
PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, ExpMaskV, ISD::SETLT);
Test &= ~fcFinite;
} else if ((Test & fcFinite) == fcPosFinite) {
// finite(V) && V > 0 ==> V < exp_mask
PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, ExpMaskV, ISD::SETULT);
Test &= ~fcPosFinite;
} else if ((Test & fcFinite) == fcNegFinite) {
// finite(V) && V < 0 ==> abs(V) < exp_mask && signbit == 1
PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, ExpMaskV, ISD::SETLT);
PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, SignV);
Test &= ~fcNegFinite;
}
appendResult(PartialRes);
if (FPClassTest PartialCheck = Test & (fcZero | fcSubnormal)) {
// fcZero | fcSubnormal => test all exponent bits are 0
// TODO: Handle sign bit specific cases
if (PartialCheck == (fcZero | fcSubnormal)) {
SDValue ExpBits = DAG.getNode(ISD::AND, DL, IntVT, OpAsInt, ExpMaskV);
SDValue ExpIsZero =
DAG.getSetCC(DL, ResultVT, ExpBits, ZeroV, ISD::SETEQ);
appendResult(ExpIsZero);
Test &= ~PartialCheck & fcAllFlags;
}
}
// Check for individual classes.
if (unsigned PartialCheck = Test & fcZero) {
if (PartialCheck == fcPosZero)
PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, ZeroV, ISD::SETEQ);
else if (PartialCheck == fcZero)
PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, ZeroV, ISD::SETEQ);
else // ISD::fcNegZero
PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, SignBitV, ISD::SETEQ);
appendResult(PartialRes);
}
if (unsigned PartialCheck = Test & fcSubnormal) {
// issubnormal(V) ==> unsigned(abs(V) - 1) < (all mantissa bits set)
// issubnormal(V) && V>0 ==> unsigned(V - 1) < (all mantissa bits set)
SDValue V = (PartialCheck == fcPosSubnormal) ? OpAsInt : AbsV;
SDValue MantissaV = DAG.getConstant(AllOneMantissa, DL, IntVT);
SDValue VMinusOneV =
DAG.getNode(ISD::SUB, DL, IntVT, V, DAG.getConstant(1, DL, IntVT));
PartialRes = DAG.getSetCC(DL, ResultVT, VMinusOneV, MantissaV, ISD::SETULT);
if (PartialCheck == fcNegSubnormal)
PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, SignV);
appendResult(PartialRes);
}
if (unsigned PartialCheck = Test & fcInf) {
if (PartialCheck == fcPosInf)
PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, InfV, ISD::SETEQ);
else if (PartialCheck == fcInf)
PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, InfV, ISD::SETEQ);
else { // ISD::fcNegInf
APInt NegInf = APFloat::getInf(Semantics, true).bitcastToAPInt();
SDValue NegInfV = DAG.getConstant(NegInf, DL, IntVT);
PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, NegInfV, ISD::SETEQ);
}
appendResult(PartialRes);
}
if (unsigned PartialCheck = Test & fcNan) {
APInt InfWithQnanBit = Inf | QNaNBitMask;
SDValue InfWithQnanBitV = DAG.getConstant(InfWithQnanBit, DL, IntVT);
if (PartialCheck == fcNan) {
// isnan(V) ==> abs(V) > int(inf)
PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, InfV, ISD::SETGT);
if (IsF80) {
// Recognize unsupported values as NaNs for compatibility with glibc.
// In them (exp(V)==0) == int_bit.
SDValue ExpBits = DAG.getNode(ISD::AND, DL, IntVT, AbsV, ExpMaskV);
SDValue ExpIsZero =
DAG.getSetCC(DL, ResultVT, ExpBits, ZeroV, ISD::SETEQ);
SDValue IsPseudo =
DAG.getSetCC(DL, ResultVT, getIntBitIsSet(), ExpIsZero, ISD::SETEQ);
PartialRes = DAG.getNode(ISD::OR, DL, ResultVT, PartialRes, IsPseudo);
}
} else if (PartialCheck == fcQNan) {
// isquiet(V) ==> abs(V) >= (unsigned(Inf) | quiet_bit)
PartialRes =
DAG.getSetCC(DL, ResultVT, AbsV, InfWithQnanBitV, ISD::SETGE);
} else { // ISD::fcSNan
// issignaling(V) ==> abs(V) > unsigned(Inf) &&
// abs(V) < (unsigned(Inf) | quiet_bit)
SDValue IsNan = DAG.getSetCC(DL, ResultVT, AbsV, InfV, ISD::SETGT);
SDValue IsNotQnan =
DAG.getSetCC(DL, ResultVT, AbsV, InfWithQnanBitV, ISD::SETLT);
PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, IsNan, IsNotQnan);
}
appendResult(PartialRes);
}
if (unsigned PartialCheck = Test & fcNormal) {
// isnormal(V) ==> (0 < exp < max_exp) ==> (unsigned(exp-1) < (max_exp-1))
APInt ExpLSB = ExpMask & ~(ExpMask.shl(1));
SDValue ExpLSBV = DAG.getConstant(ExpLSB, DL, IntVT);
SDValue ExpMinus1 = DAG.getNode(ISD::SUB, DL, IntVT, AbsV, ExpLSBV);
APInt ExpLimit = ExpMask - ExpLSB;
SDValue ExpLimitV = DAG.getConstant(ExpLimit, DL, IntVT);
PartialRes = DAG.getSetCC(DL, ResultVT, ExpMinus1, ExpLimitV, ISD::SETULT);
if (PartialCheck == fcNegNormal)
PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, SignV);
else if (PartialCheck == fcPosNormal) {
SDValue PosSignV =
DAG.getNode(ISD::XOR, DL, ResultVT, SignV, ResultInvertionMask);
PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, PosSignV);
}
if (IsF80)
PartialRes =
DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, getIntBitIsSet());
appendResult(PartialRes);
}
if (!Res)
return DAG.getConstant(IsInverted, DL, ResultVT);
if (IsInverted)
Res = DAG.getNode(ISD::XOR, DL, ResultVT, Res, ResultInvertionMask);
return Res;
}
// Only expand vector types if we have the appropriate vector bit operations.
static bool canExpandVectorCTPOP(const TargetLowering &TLI, EVT VT) {
assert(VT.isVector() && "Expected vector type");
unsigned Len = VT.getScalarSizeInBits();
return TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
TLI.isOperationLegalOrCustom(ISD::SUB, VT) &&
TLI.isOperationLegalOrCustom(ISD::SRL, VT) &&
(Len == 8 || TLI.isOperationLegalOrCustom(ISD::MUL, VT)) &&
TLI.isOperationLegalOrCustomOrPromote(ISD::AND, VT);
}
SDValue TargetLowering::expandCTPOP(SDNode *Node, SelectionDAG &DAG) const {
SDLoc dl(Node);
EVT VT = Node->getValueType(0);
EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout());
SDValue Op = Node->getOperand(0);
unsigned Len = VT.getScalarSizeInBits();
assert(VT.isInteger() && "CTPOP not implemented for this type.");
// TODO: Add support for irregular type lengths.
if (!(Len <= 128 && Len % 8 == 0))
return SDValue();
// Only expand vector types if we have the appropriate vector bit operations.
if (VT.isVector() && !canExpandVectorCTPOP(*this, VT))
return SDValue();
// This is the "best" algorithm from
// http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
SDValue Mask55 =
DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), dl, VT);
SDValue Mask33 =
DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), dl, VT);
SDValue Mask0F =
DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), dl, VT);
// v = v - ((v >> 1) & 0x55555555...)
Op = DAG.getNode(ISD::SUB, dl, VT, Op,
DAG.getNode(ISD::AND, dl, VT,
DAG.getNode(ISD::SRL, dl, VT, Op,
DAG.getConstant(1, dl, ShVT)),
Mask55));
// v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
Op = DAG.getNode(ISD::ADD, dl, VT, DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
DAG.getNode(ISD::AND, dl, VT,
DAG.getNode(ISD::SRL, dl, VT, Op,
DAG.getConstant(2, dl, ShVT)),
Mask33));
// v = (v + (v >> 4)) & 0x0F0F0F0F...
Op = DAG.getNode(ISD::AND, dl, VT,
DAG.getNode(ISD::ADD, dl, VT, Op,
DAG.getNode(ISD::SRL, dl, VT, Op,
DAG.getConstant(4, dl, ShVT))),
Mask0F);
if (Len <= 8)
return Op;
// Avoid the multiply if we only have 2 bytes to add.
// TODO: Only doing this for scalars because vectors weren't as obviously
// improved.
if (Len == 16 && !VT.isVector()) {
// v = (v + (v >> 8)) & 0x00FF;
return DAG.getNode(ISD::AND, dl, VT,
DAG.getNode(ISD::ADD, dl, VT, Op,
DAG.getNode(ISD::SRL, dl, VT, Op,
DAG.getConstant(8, dl, ShVT))),
DAG.getConstant(0xFF, dl, VT));
}
// v = (v * 0x01010101...) >> (Len - 8)
SDValue V;
if (isOperationLegalOrCustomOrPromote(
ISD::MUL, getTypeToTransformTo(*DAG.getContext(), VT))) {
SDValue Mask01 =
DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), dl, VT);
V = DAG.getNode(ISD::MUL, dl, VT, Op, Mask01);
} else {
V = Op;
for (unsigned Shift = 8; Shift < Len; Shift *= 2) {
SDValue ShiftC = DAG.getShiftAmountConstant(Shift, VT, dl);
V = DAG.getNode(ISD::ADD, dl, VT, V,
DAG.getNode(ISD::SHL, dl, VT, V, ShiftC));
}
}
return DAG.getNode(ISD::SRL, dl, VT, V, DAG.getConstant(Len - 8, dl, ShVT));
}
SDValue TargetLowering::expandVPCTPOP(SDNode *Node, SelectionDAG &DAG) const {
SDLoc dl(Node);
EVT VT = Node->getValueType(0);
EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout());
SDValue Op = Node->getOperand(0);
SDValue Mask = Node->getOperand(1);
SDValue VL = Node->getOperand(2);
unsigned Len = VT.getScalarSizeInBits();
assert(VT.isInteger() && "VP_CTPOP not implemented for this type.");
// TODO: Add support for irregular type lengths.
if (!(Len <= 128 && Len % 8 == 0))
return SDValue();
// This is same algorithm of expandCTPOP from
// http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
SDValue Mask55 =
DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), dl, VT);
SDValue Mask33 =
DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), dl, VT);
SDValue Mask0F =
DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), dl, VT);
SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5;
// v = v - ((v >> 1) & 0x55555555...)
Tmp1 = DAG.getNode(ISD::VP_AND, dl, VT,
DAG.getNode(ISD::VP_SRL, dl, VT, Op,
DAG.getConstant(1, dl, ShVT), Mask, VL),
Mask55, Mask, VL);
Op = DAG.getNode(ISD::VP_SUB, dl, VT, Op, Tmp1, Mask, VL);
// v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Op, Mask33, Mask, VL);
Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT,
DAG.getNode(ISD::VP_SRL, dl, VT, Op,
DAG.getConstant(2, dl, ShVT), Mask, VL),
Mask33, Mask, VL);
Op = DAG.getNode(ISD::VP_ADD, dl, VT, Tmp2, Tmp3, Mask, VL);
// v = (v + (v >> 4)) & 0x0F0F0F0F...
Tmp4 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(4, dl, ShVT),
Mask, VL),
Tmp5 = DAG.getNode(ISD::VP_ADD, dl, VT, Op, Tmp4, Mask, VL);
Op = DAG.getNode(ISD::VP_AND, dl, VT, Tmp5, Mask0F, Mask, VL);
if (Len <= 8)
return Op;
// v = (v * 0x01010101...) >> (Len - 8)
SDValue V;
if (isOperationLegalOrCustomOrPromote(
ISD::VP_MUL, getTypeToTransformTo(*DAG.getContext(), VT))) {
SDValue Mask01 =
DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), dl, VT);
V = DAG.getNode(ISD::VP_MUL, dl, VT, Op, Mask01, Mask, VL);
} else {
V = Op;
for (unsigned Shift = 8; Shift < Len; Shift *= 2) {
SDValue ShiftC = DAG.getShiftAmountConstant(Shift, VT, dl);
V = DAG.getNode(ISD::VP_ADD, dl, VT, V,
DAG.getNode(ISD::VP_SHL, dl, VT, V, ShiftC, Mask, VL),
Mask, VL);
}
}
return DAG.getNode(ISD::VP_SRL, dl, VT, V, DAG.getConstant(Len - 8, dl, ShVT),
Mask, VL);
}
SDValue TargetLowering::expandCTLZ(SDNode *Node, SelectionDAG &DAG) const {
SDLoc dl(Node);
EVT VT = Node->getValueType(0);
EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout());
SDValue Op = Node->getOperand(0);
unsigned NumBitsPerElt = VT.getScalarSizeInBits();
// If the non-ZERO_UNDEF version is supported we can use that instead.
if (Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF &&
isOperationLegalOrCustom(ISD::CTLZ, VT))
return DAG.getNode(ISD::CTLZ, dl, VT, Op);
// If the ZERO_UNDEF version is supported use that and handle the zero case.
if (isOperationLegalOrCustom(ISD::CTLZ_ZERO_UNDEF, VT)) {
EVT SetCCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue CTLZ = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, dl, VT, Op);
SDValue Zero = DAG.getConstant(0, dl, VT);
SDValue SrcIsZero = DAG.getSetCC(dl, SetCCVT, Op, Zero, ISD::SETEQ);
return DAG.getSelect(dl, VT, SrcIsZero,
DAG.getConstant(NumBitsPerElt, dl, VT), CTLZ);
}
// Only expand vector types if we have the appropriate vector bit operations.
// This includes the operations needed to expand CTPOP if it isn't supported.
if (VT.isVector() && (!isPowerOf2_32(NumBitsPerElt) ||
(!isOperationLegalOrCustom(ISD::CTPOP, VT) &&
!canExpandVectorCTPOP(*this, VT)) ||
!isOperationLegalOrCustom(ISD::SRL, VT) ||
!isOperationLegalOrCustomOrPromote(ISD::OR, VT)))
return SDValue();
// for now, we do this:
// x = x | (x >> 1);
// x = x | (x >> 2);
// ...
// x = x | (x >>16);
// x = x | (x >>32); // for 64-bit input
// return popcount(~x);
//
// Ref: "Hacker's Delight" by Henry Warren
for (unsigned i = 0; (1U << i) < NumBitsPerElt; ++i) {
SDValue Tmp = DAG.getConstant(1ULL << i, dl, ShVT);
Op = DAG.getNode(ISD::OR, dl, VT, Op,
DAG.getNode(ISD::SRL, dl, VT, Op, Tmp));
}
Op = DAG.getNOT(dl, Op, VT);
return DAG.getNode(ISD::CTPOP, dl, VT, Op);
}
SDValue TargetLowering::expandVPCTLZ(SDNode *Node, SelectionDAG &DAG) const {
SDLoc dl(Node);
EVT VT = Node->getValueType(0);
EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout());
SDValue Op = Node->getOperand(0);
SDValue Mask = Node->getOperand(1);
SDValue VL = Node->getOperand(2);
unsigned NumBitsPerElt = VT.getScalarSizeInBits();
// do this:
// x = x | (x >> 1);
// x = x | (x >> 2);
// ...
// x = x | (x >>16);
// x = x | (x >>32); // for 64-bit input
// return popcount(~x);
for (unsigned i = 0; (1U << i) < NumBitsPerElt; ++i) {
SDValue Tmp = DAG.getConstant(1ULL << i, dl, ShVT);
Op = DAG.getNode(ISD::VP_OR, dl, VT, Op,
DAG.getNode(ISD::VP_SRL, dl, VT, Op, Tmp, Mask, VL), Mask,
VL);
}
Op = DAG.getNode(ISD::VP_XOR, dl, VT, Op, DAG.getConstant(-1, dl, VT), Mask,
VL);
return DAG.getNode(ISD::VP_CTPOP, dl, VT, Op, Mask, VL);
}
SDValue TargetLowering::CTTZTableLookup(SDNode *Node, SelectionDAG &DAG,
const SDLoc &DL, EVT VT, SDValue Op,
unsigned BitWidth) const {
if (BitWidth != 32 && BitWidth != 64)
return SDValue();
APInt DeBruijn = BitWidth == 32 ? APInt(32, 0x077CB531U)
: APInt(64, 0x0218A392CD3D5DBFULL);
const DataLayout &TD = DAG.getDataLayout();
MachinePointerInfo PtrInfo =
MachinePointerInfo::getConstantPool(DAG.getMachineFunction());
unsigned ShiftAmt = BitWidth - Log2_32(BitWidth);
SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Op);
SDValue Lookup = DAG.getNode(
ISD::SRL, DL, VT,
DAG.getNode(ISD::MUL, DL, VT, DAG.getNode(ISD::AND, DL, VT, Op, Neg),
DAG.getConstant(DeBruijn, DL, VT)),
DAG.getConstant(ShiftAmt, DL, VT));
Lookup = DAG.getSExtOrTrunc(Lookup, DL, getPointerTy(TD));
SmallVector<uint8_t> Table(BitWidth, 0);
for (unsigned i = 0; i < BitWidth; i++) {
APInt Shl = DeBruijn.shl(i);
APInt Lshr = Shl.lshr(ShiftAmt);
Table[Lshr.getZExtValue()] = i;
}
// Create a ConstantArray in Constant Pool
auto *CA = ConstantDataArray::get(*DAG.getContext(), Table);
SDValue CPIdx = DAG.getConstantPool(CA, getPointerTy(TD),
TD.getPrefTypeAlign(CA->getType()));
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, VT, DAG.getEntryNode(),
DAG.getMemBasePlusOffset(CPIdx, Lookup, DL),
PtrInfo, MVT::i8);
if (Node->getOpcode() == ISD::CTTZ_ZERO_UNDEF)
return ExtLoad;
EVT SetCCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue Zero = DAG.getConstant(0, DL, VT);
SDValue SrcIsZero = DAG.getSetCC(DL, SetCCVT, Op, Zero, ISD::SETEQ);
return DAG.getSelect(DL, VT, SrcIsZero,
DAG.getConstant(BitWidth, DL, VT), ExtLoad);
}
SDValue TargetLowering::expandCTTZ(SDNode *Node, SelectionDAG &DAG) const {
SDLoc dl(Node);
EVT VT = Node->getValueType(0);
SDValue Op = Node->getOperand(0);
unsigned NumBitsPerElt = VT.getScalarSizeInBits();
// If the non-ZERO_UNDEF version is supported we can use that instead.
if (Node->getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
isOperationLegalOrCustom(ISD::CTTZ, VT))
return DAG.getNode(ISD::CTTZ, dl, VT, Op);
// If the ZERO_UNDEF version is supported use that and handle the zero case.
if (isOperationLegalOrCustom(ISD::CTTZ_ZERO_UNDEF, VT)) {
EVT SetCCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue CTTZ = DAG.getNode(ISD::CTTZ_ZERO_UNDEF, dl, VT, Op);
SDValue Zero = DAG.getConstant(0, dl, VT);
SDValue SrcIsZero = DAG.getSetCC(dl, SetCCVT, Op, Zero, ISD::SETEQ);
return DAG.getSelect(dl, VT, SrcIsZero,
DAG.getConstant(NumBitsPerElt, dl, VT), CTTZ);
}
// Only expand vector types if we have the appropriate vector bit operations.
// This includes the operations needed to expand CTPOP if it isn't supported.
if (VT.isVector() && (!isPowerOf2_32(NumBitsPerElt) ||
(!isOperationLegalOrCustom(ISD::CTPOP, VT) &&
!isOperationLegalOrCustom(ISD::CTLZ, VT) &&
!canExpandVectorCTPOP(*this, VT)) ||
!isOperationLegalOrCustom(ISD::SUB, VT) ||
!isOperationLegalOrCustomOrPromote(ISD::AND, VT) ||
!isOperationLegalOrCustomOrPromote(ISD::XOR, VT)))
return SDValue();
// Emit Table Lookup if ISD::CTLZ and ISD::CTPOP are not legal.
if (!VT.isVector() && isOperationExpand(ISD::CTPOP, VT) &&
!isOperationLegal(ISD::CTLZ, VT))
if (SDValue V = CTTZTableLookup(Node, DAG, dl, VT, Op, NumBitsPerElt))
return V;
// for now, we use: { return popcount(~x & (x - 1)); }
// unless the target has ctlz but not ctpop, in which case we use:
// { return 32 - nlz(~x & (x-1)); }
// Ref: "Hacker's Delight" by Henry Warren
SDValue Tmp = DAG.getNode(
ISD::AND, dl, VT, DAG.getNOT(dl, Op, VT),
DAG.getNode(ISD::SUB, dl, VT, Op, DAG.getConstant(1, dl, VT)));
// If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
if (isOperationLegal(ISD::CTLZ, VT) && !isOperationLegal(ISD::CTPOP, VT)) {
return DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(NumBitsPerElt, dl, VT),
DAG.getNode(ISD::CTLZ, dl, VT, Tmp));
}
return DAG.getNode(ISD::CTPOP, dl, VT, Tmp);
}
SDValue TargetLowering::expandVPCTTZ(SDNode *Node, SelectionDAG &DAG) const {
SDValue Op = Node->getOperand(0);
SDValue Mask = Node->getOperand(1);
SDValue VL = Node->getOperand(2);
SDLoc dl(Node);
EVT VT = Node->getValueType(0);
// Same as the vector part of expandCTTZ, use: popcount(~x & (x - 1))
SDValue Not = DAG.getNode(ISD::VP_XOR, dl, VT, Op,
DAG.getConstant(-1, dl, VT), Mask, VL);
SDValue MinusOne = DAG.getNode(ISD::VP_SUB, dl, VT, Op,
DAG.getConstant(1, dl, VT), Mask, VL);
SDValue Tmp = DAG.getNode(ISD::VP_AND, dl, VT, Not, MinusOne, Mask, VL);
return DAG.getNode(ISD::VP_CTPOP, dl, VT, Tmp, Mask, VL);
}
SDValue TargetLowering::expandVPCTTZElements(SDNode *N,
SelectionDAG &DAG) const {
// %cond = to_bool_vec %source
// %splat = splat /*val=*/VL
// %tz = step_vector
// %v = vp.select %cond, /*true=*/tz, /*false=*/%splat
// %r = vp.reduce.umin %v
SDLoc DL(N);
SDValue Source = N->getOperand(0);
SDValue Mask = N->getOperand(1);
SDValue EVL = N->getOperand(2);
EVT SrcVT = Source.getValueType();
EVT ResVT = N->getValueType(0);
EVT ResVecVT =
EVT::getVectorVT(*DAG.getContext(), ResVT, SrcVT.getVectorElementCount());
// Convert to boolean vector.
if (SrcVT.getScalarType() != MVT::i1) {
SDValue AllZero = DAG.getConstant(0, DL, SrcVT);
SrcVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
SrcVT.getVectorElementCount());
Source = DAG.getNode(ISD::VP_SETCC, DL, SrcVT, Source, AllZero,
DAG.getCondCode(ISD::SETNE), Mask, EVL);
}
SDValue ExtEVL = DAG.getZExtOrTrunc(EVL, DL, ResVT);
SDValue Splat = DAG.getSplat(ResVecVT, DL, ExtEVL);
SDValue StepVec = DAG.getStepVector(DL, ResVecVT);
SDValue Select =
DAG.getNode(ISD::VP_SELECT, DL, ResVecVT, Source, StepVec, Splat, EVL);
return DAG.getNode(ISD::VP_REDUCE_UMIN, DL, ResVT, ExtEVL, Select, Mask, EVL);
}
SDValue TargetLowering::expandABS(SDNode *N, SelectionDAG &DAG,
bool IsNegative) const {
SDLoc dl(N);
EVT VT = N->getValueType(0);
SDValue Op = N->getOperand(0);
// abs(x) -> smax(x,sub(0,x))
if (!IsNegative && isOperationLegal(ISD::SUB, VT) &&
isOperationLegal(ISD::SMAX, VT)) {
SDValue Zero = DAG.getConstant(0, dl, VT);
Op = DAG.getFreeze(Op);
return DAG.getNode(ISD::SMAX, dl, VT, Op,
DAG.getNode(ISD::SUB, dl, VT, Zero, Op));
}
// abs(x) -> umin(x,sub(0,x))
if (!IsNegative && isOperationLegal(ISD::SUB, VT) &&
isOperationLegal(ISD::UMIN, VT)) {
SDValue Zero = DAG.getConstant(0, dl, VT);
Op = DAG.getFreeze(Op);
return DAG.getNode(ISD::UMIN, dl, VT, Op,
DAG.getNode(ISD::SUB, dl, VT, Zero, Op));
}
// 0 - abs(x) -> smin(x, sub(0,x))
if (IsNegative && isOperationLegal(ISD::SUB, VT) &&
isOperationLegal(ISD::SMIN, VT)) {
SDValue Zero = DAG.getConstant(0, dl, VT);
Op = DAG.getFreeze(Op);
return DAG.getNode(ISD::SMIN, dl, VT, Op,
DAG.getNode(ISD::SUB, dl, VT, Zero, Op));
}
// Only expand vector types if we have the appropriate vector operations.
if (VT.isVector() &&
(!isOperationLegalOrCustom(ISD::SRA, VT) ||
(!IsNegative && !isOperationLegalOrCustom(ISD::ADD, VT)) ||
(IsNegative && !isOperationLegalOrCustom(ISD::SUB, VT)) ||
!isOperationLegalOrCustomOrPromote(ISD::XOR, VT)))
return SDValue();
Op = DAG.getFreeze(Op);
SDValue Shift = DAG.getNode(
ISD::SRA, dl, VT, Op,
DAG.getShiftAmountConstant(VT.getScalarSizeInBits() - 1, VT, dl));
SDValue Xor = DAG.getNode(ISD::XOR, dl, VT, Op, Shift);
// abs(x) -> Y = sra (X, size(X)-1); sub (xor (X, Y), Y)
if (!IsNegative)
return DAG.getNode(ISD::SUB, dl, VT, Xor, Shift);
// 0 - abs(x) -> Y = sra (X, size(X)-1); sub (Y, xor (X, Y))
return DAG.getNode(ISD::SUB, dl, VT, Shift, Xor);
}
SDValue TargetLowering::expandABD(SDNode *N, SelectionDAG &DAG) const {
SDLoc dl(N);
EVT VT = N->getValueType(0);
SDValue LHS = DAG.getFreeze(N->getOperand(0));
SDValue RHS = DAG.getFreeze(N->getOperand(1));
bool IsSigned = N->getOpcode() == ISD::ABDS;
// abds(lhs, rhs) -> sub(smax(lhs,rhs), smin(lhs,rhs))
// abdu(lhs, rhs) -> sub(umax(lhs,rhs), umin(lhs,rhs))
unsigned MaxOpc = IsSigned ? ISD::SMAX : ISD::UMAX;
unsigned MinOpc = IsSigned ? ISD::SMIN : ISD::UMIN;
if (isOperationLegal(MaxOpc, VT) && isOperationLegal(MinOpc, VT)) {
SDValue Max = DAG.getNode(MaxOpc, dl, VT, LHS, RHS);
SDValue Min = DAG.getNode(MinOpc, dl, VT, LHS, RHS);
return DAG.getNode(ISD::SUB, dl, VT, Max, Min);
}
// abdu(lhs, rhs) -> or(usubsat(lhs,rhs), usubsat(rhs,lhs))
if (!IsSigned && isOperationLegal(ISD::USUBSAT, VT))
return DAG.getNode(ISD::OR, dl, VT,
DAG.getNode(ISD::USUBSAT, dl, VT, LHS, RHS),
DAG.getNode(ISD::USUBSAT, dl, VT, RHS, LHS));
EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
ISD::CondCode CC = IsSigned ? ISD::CondCode::SETGT : ISD::CondCode::SETUGT;
SDValue Cmp = DAG.getSetCC(dl, CCVT, LHS, RHS, CC);
// Branchless expansion iff cmp result is allbits:
// abds(lhs, rhs) -> sub(sgt(lhs, rhs), xor(sgt(lhs, rhs), sub(lhs, rhs)))
// abdu(lhs, rhs) -> sub(ugt(lhs, rhs), xor(ugt(lhs, rhs), sub(lhs, rhs)))
if (CCVT == VT && getBooleanContents(VT) == ZeroOrNegativeOneBooleanContent) {
SDValue Diff = DAG.getNode(ISD::SUB, dl, VT, LHS, RHS);
SDValue Xor = DAG.getNode(ISD::XOR, dl, VT, Diff, Cmp);
return DAG.getNode(ISD::SUB, dl, VT, Cmp, Xor);
}
// abds(lhs, rhs) -> select(sgt(lhs,rhs), sub(lhs,rhs), sub(rhs,lhs))
// abdu(lhs, rhs) -> select(ugt(lhs,rhs), sub(lhs,rhs), sub(rhs,lhs))
return DAG.getSelect(dl, VT, Cmp, DAG.getNode(ISD::SUB, dl, VT, LHS, RHS),
DAG.getNode(ISD::SUB, dl, VT, RHS, LHS));
}
SDValue TargetLowering::expandAVG(SDNode *N, SelectionDAG &DAG) const {
SDLoc dl(N);
EVT VT = N->getValueType(0);
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
unsigned Opc = N->getOpcode();
bool IsFloor = Opc == ISD::AVGFLOORS || Opc == ISD::AVGFLOORU;
bool IsSigned = Opc == ISD::AVGCEILS || Opc == ISD::AVGFLOORS;
unsigned SumOpc = IsFloor ? ISD::ADD : ISD::SUB;
unsigned SignOpc = IsFloor ? ISD::AND : ISD::OR;
unsigned ShiftOpc = IsSigned ? ISD::SRA : ISD::SRL;
unsigned ExtOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
assert((Opc == ISD::AVGFLOORS || Opc == ISD::AVGCEILS ||
Opc == ISD::AVGFLOORU || Opc == ISD::AVGCEILU) &&
"Unknown AVG node");
// If the operands are already extended, we can add+shift.
bool IsExt =
(IsSigned && DAG.ComputeNumSignBits(LHS) >= 2 &&
DAG.ComputeNumSignBits(RHS) >= 2) ||
(!IsSigned && DAG.computeKnownBits(LHS).countMinLeadingZeros() >= 1 &&
DAG.computeKnownBits(RHS).countMinLeadingZeros() >= 1);
if (IsExt) {
SDValue Sum = DAG.getNode(ISD::ADD, dl, VT, LHS, RHS);
if (!IsFloor)
Sum = DAG.getNode(ISD::ADD, dl, VT, Sum, DAG.getConstant(1, dl, VT));
return DAG.getNode(ShiftOpc, dl, VT, Sum,
DAG.getShiftAmountConstant(1, VT, dl));
}
// For scalars, see if we can efficiently extend/truncate to use add+shift.
if (VT.isScalarInteger()) {
unsigned BW = VT.getScalarSizeInBits();
EVT ExtVT = VT.getIntegerVT(*DAG.getContext(), 2 * BW);
if (isTypeLegal(ExtVT) && isTruncateFree(ExtVT, VT)) {
LHS = DAG.getNode(ExtOpc, dl, ExtVT, LHS);
RHS = DAG.getNode(ExtOpc, dl, ExtVT, RHS);
SDValue Avg = DAG.getNode(ISD::ADD, dl, ExtVT, LHS, RHS);
if (!IsFloor)
Avg = DAG.getNode(ISD::ADD, dl, ExtVT, Avg,
DAG.getConstant(1, dl, ExtVT));
// Just use SRL as we will be truncating away the extended sign bits.
Avg = DAG.getNode(ISD::SRL, dl, ExtVT, Avg,
DAG.getShiftAmountConstant(1, ExtVT, dl));
return DAG.getNode(ISD::TRUNCATE, dl, VT, Avg);
}
}
// avgceils(lhs, rhs) -> sub(or(lhs,rhs),ashr(xor(lhs,rhs),1))
// avgceilu(lhs, rhs) -> sub(or(lhs,rhs),lshr(xor(lhs,rhs),1))
// avgfloors(lhs, rhs) -> add(and(lhs,rhs),ashr(xor(lhs,rhs),1))
// avgflooru(lhs, rhs) -> add(and(lhs,rhs),lshr(xor(lhs,rhs),1))
LHS = DAG.getFreeze(LHS);
RHS = DAG.getFreeze(RHS);
SDValue Sign = DAG.getNode(SignOpc, dl, VT, LHS, RHS);
SDValue Xor = DAG.getNode(ISD::XOR, dl, VT, LHS, RHS);
SDValue Shift =
DAG.getNode(ShiftOpc, dl, VT, Xor, DAG.getShiftAmountConstant(1, VT, dl));
return DAG.getNode(SumOpc, dl, VT, Sign, Shift);
}
SDValue TargetLowering::expandBSWAP(SDNode *N, SelectionDAG &DAG) const {
SDLoc dl(N);
EVT VT = N->getValueType(0);
SDValue Op = N->getOperand(0);
if (!VT.isSimple())
return SDValue();
EVT SHVT = getShiftAmountTy(VT, DAG.getDataLayout());
SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
switch (VT.getSimpleVT().getScalarType().SimpleTy) {
default:
return SDValue();
case MVT::i16:
// Use a rotate by 8. This can be further expanded if necessary.
return DAG.getNode(ISD::ROTL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
case MVT::i32:
Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
Tmp3 = DAG.getNode(ISD::AND, dl, VT, Op,
DAG.getConstant(0xFF00, dl, VT));
Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(8, dl, SHVT));
Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, dl, VT));
Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
case MVT::i64:
Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
Tmp7 = DAG.getNode(ISD::AND, dl, VT, Op,
DAG.getConstant(255ULL<<8, dl, VT));
Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Tmp7, DAG.getConstant(40, dl, SHVT));
Tmp6 = DAG.getNode(ISD::AND, dl, VT, Op,
DAG.getConstant(255ULL<<16, dl, VT));
Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Tmp6, DAG.getConstant(24, dl, SHVT));
Tmp5 = DAG.getNode(ISD::AND, dl, VT, Op,
DAG.getConstant(255ULL<<24, dl, VT));
Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Tmp5, DAG.getConstant(8, dl, SHVT));
Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4,
DAG.getConstant(255ULL<<24, dl, VT));
Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3,
DAG.getConstant(255ULL<<16, dl, VT));
Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, dl, SHVT));
Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2,
DAG.getConstant(255ULL<<8, dl, VT));
Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
}
}
SDValue TargetLowering::expandVPBSWAP(SDNode *N, SelectionDAG &DAG) const {
SDLoc dl(N);
EVT VT = N->getValueType(0);
SDValue Op = N->getOperand(0);
SDValue Mask = N->getOperand(1);
SDValue EVL = N->getOperand(2);
if (!VT.isSimple())
return SDValue();
EVT SHVT = getShiftAmountTy(VT, DAG.getDataLayout());
SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
switch (VT.getSimpleVT().getScalarType().SimpleTy) {
default:
return SDValue();
case MVT::i16:
Tmp1 = DAG.getNode(ISD::VP_SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT),
Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT),
Mask, EVL);
return DAG.getNode(ISD::VP_OR, dl, VT, Tmp1, Tmp2, Mask, EVL);
case MVT::i32:
Tmp4 = DAG.getNode(ISD::VP_SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT),
Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Op, DAG.getConstant(0xFF00, dl, VT),
Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp3, DAG.getConstant(8, dl, SHVT),
Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT),
Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2,
DAG.getConstant(0xFF00, dl, VT), Mask, EVL);
Tmp1 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT),
Mask, EVL);
Tmp4 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp4, Tmp3, Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp1, Mask, EVL);
return DAG.getNode(ISD::VP_OR, dl, VT, Tmp4, Tmp2, Mask, EVL);
case MVT::i64:
Tmp8 = DAG.getNode(ISD::VP_SHL, dl, VT, Op, DAG.getConstant(56, dl, SHVT),
Mask, EVL);
Tmp7 = DAG.getNode(ISD::VP_AND, dl, VT, Op,
DAG.getConstant(255ULL << 8, dl, VT), Mask, EVL);
Tmp7 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp7, DAG.getConstant(40, dl, SHVT),
Mask, EVL);
Tmp6 = DAG.getNode(ISD::VP_AND, dl, VT, Op,
DAG.getConstant(255ULL << 16, dl, VT), Mask, EVL);
Tmp6 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp6, DAG.getConstant(24, dl, SHVT),
Mask, EVL);
Tmp5 = DAG.getNode(ISD::VP_AND, dl, VT, Op,
DAG.getConstant(255ULL << 24, dl, VT), Mask, EVL);
Tmp5 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp5, DAG.getConstant(8, dl, SHVT),
Mask, EVL);
Tmp4 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT),
Mask, EVL);
Tmp4 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp4,
DAG.getConstant(255ULL << 24, dl, VT), Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT),
Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp3,
DAG.getConstant(255ULL << 16, dl, VT), Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(40, dl, SHVT),
Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2,
DAG.getConstant(255ULL << 8, dl, VT), Mask, EVL);
Tmp1 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(56, dl, SHVT),
Mask, EVL);
Tmp8 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp8, Tmp7, Mask, EVL);
Tmp6 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp6, Tmp5, Mask, EVL);
Tmp4 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp4, Tmp3, Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp1, Mask, EVL);
Tmp8 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp8, Tmp6, Mask, EVL);
Tmp4 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp4, Tmp2, Mask, EVL);
return DAG.getNode(ISD::VP_OR, dl, VT, Tmp8, Tmp4, Mask, EVL);
}
}
SDValue TargetLowering::expandBITREVERSE(SDNode *N, SelectionDAG &DAG) const {
SDLoc dl(N);
EVT VT = N->getValueType(0);
SDValue Op = N->getOperand(0);
EVT SHVT = getShiftAmountTy(VT, DAG.getDataLayout());
unsigned Sz = VT.getScalarSizeInBits();
SDValue Tmp, Tmp2, Tmp3;
// If we can, perform BSWAP first and then the mask+swap the i4, then i2
// and finally the i1 pairs.
// TODO: We can easily support i4/i2 legal types if any target ever does.
if (Sz >= 8 && isPowerOf2_32(Sz)) {
// Create the masks - repeating the pattern every byte.
APInt Mask4 = APInt::getSplat(Sz, APInt(8, 0x0F));
APInt Mask2 = APInt::getSplat(Sz, APInt(8, 0x33));
APInt Mask1 = APInt::getSplat(Sz, APInt(8, 0x55));
// BSWAP if the type is wider than a single byte.
Tmp = (Sz > 8 ? DAG.getNode(ISD::BSWAP, dl, VT, Op) : Op);
// swap i4: ((V >> 4) & 0x0F) | ((V & 0x0F) << 4)
Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp, DAG.getConstant(4, dl, SHVT));
Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Mask4, dl, VT));
Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(Mask4, dl, VT));
Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(4, dl, SHVT));
Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
// swap i2: ((V >> 2) & 0x33) | ((V & 0x33) << 2)
Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp, DAG.getConstant(2, dl, SHVT));
Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Mask2, dl, VT));
Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(Mask2, dl, VT));
Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(2, dl, SHVT));
Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
// swap i1: ((V >> 1) & 0x55) | ((V & 0x55) << 1)
Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp, DAG.getConstant(1, dl, SHVT));
Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Mask1, dl, VT));
Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(Mask1, dl, VT));
Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(1, dl, SHVT));
Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
return Tmp;
}
Tmp = DAG.getConstant(0, dl, VT);
for (unsigned I = 0, J = Sz-1; I < Sz; ++I, --J) {
if (I < J)
Tmp2 =
DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(J - I, dl, SHVT));
else
Tmp2 =
DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(I - J, dl, SHVT));
APInt Shift = APInt::getOneBitSet(Sz, J);
Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Shift, dl, VT));
Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp, Tmp2);
}
return Tmp;
}
SDValue TargetLowering::expandVPBITREVERSE(SDNode *N, SelectionDAG &DAG) const {
assert(N->getOpcode() == ISD::VP_BITREVERSE);
SDLoc dl(N);
EVT VT = N->getValueType(0);
SDValue Op = N->getOperand(0);
SDValue Mask = N->getOperand(1);
SDValue EVL = N->getOperand(2);
EVT SHVT = getShiftAmountTy(VT, DAG.getDataLayout());
unsigned Sz = VT.getScalarSizeInBits();
SDValue Tmp, Tmp2, Tmp3;
// If we can, perform BSWAP first and then the mask+swap the i4, then i2
// and finally the i1 pairs.
// TODO: We can easily support i4/i2 legal types if any target ever does.
if (Sz >= 8 && isPowerOf2_32(Sz)) {
// Create the masks - repeating the pattern every byte.
APInt Mask4 = APInt::getSplat(Sz, APInt(8, 0x0F));
APInt Mask2 = APInt::getSplat(Sz, APInt(8, 0x33));
APInt Mask1 = APInt::getSplat(Sz, APInt(8, 0x55));
// BSWAP if the type is wider than a single byte.
Tmp = (Sz > 8 ? DAG.getNode(ISD::VP_BSWAP, dl, VT, Op, Mask, EVL) : Op);
// swap i4: ((V >> 4) & 0x0F) | ((V & 0x0F) << 4)
Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Tmp, DAG.getConstant(4, dl, SHVT),
Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2,
DAG.getConstant(Mask4, dl, VT), Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp, DAG.getConstant(Mask4, dl, VT),
Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp3, DAG.getConstant(4, dl, SHVT),
Mask, EVL);
Tmp = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp3, Mask, EVL);
// swap i2: ((V >> 2) & 0x33) | ((V & 0x33) << 2)
Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Tmp, DAG.getConstant(2, dl, SHVT),
Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2,
DAG.getConstant(Mask2, dl, VT), Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp, DAG.getConstant(Mask2, dl, VT),
Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp3, DAG.getConstant(2, dl, SHVT),
Mask, EVL);
Tmp = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp3, Mask, EVL);
// swap i1: ((V >> 1) & 0x55) | ((V & 0x55) << 1)
Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Tmp, DAG.getConstant(1, dl, SHVT),
Mask, EVL);
Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2,
DAG.getConstant(Mask1, dl, VT), Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp, DAG.getConstant(Mask1, dl, VT),
Mask, EVL);
Tmp3 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp3, DAG.getConstant(1, dl, SHVT),
Mask, EVL);
Tmp = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp3, Mask, EVL);
return Tmp;
}
return SDValue();
}
std::pair<SDValue, SDValue>
TargetLowering::scalarizeVectorLoad(LoadSDNode *LD,
SelectionDAG &DAG) const {
SDLoc SL(LD);
SDValue Chain = LD->getChain();
SDValue BasePTR = LD->getBasePtr();
EVT SrcVT = LD->getMemoryVT();
EVT DstVT = LD->getValueType(0);
ISD::LoadExtType ExtType = LD->getExtensionType();
if (SrcVT.isScalableVector())
report_fatal_error("Cannot scalarize scalable vector loads");
unsigned NumElem = SrcVT.getVectorNumElements();
EVT SrcEltVT = SrcVT.getScalarType();
EVT DstEltVT = DstVT.getScalarType();
// A vector must always be stored in memory as-is, i.e. without any padding
// between the elements, since various code depend on it, e.g. in the
// handling of a bitcast of a vector type to int, which may be done with a
// vector store followed by an integer load. A vector that does not have
// elements that are byte-sized must therefore be stored as an integer
// built out of the extracted vector elements.
if (!SrcEltVT.isByteSized()) {
unsigned NumLoadBits = SrcVT.getStoreSizeInBits();
EVT LoadVT = EVT::getIntegerVT(*DAG.getContext(), NumLoadBits);
unsigned NumSrcBits = SrcVT.getSizeInBits();
EVT SrcIntVT = EVT::getIntegerVT(*DAG.getContext(), NumSrcBits);
unsigned SrcEltBits = SrcEltVT.getSizeInBits();
SDValue SrcEltBitMask = DAG.getConstant(
APInt::getLowBitsSet(NumLoadBits, SrcEltBits), SL, LoadVT);
// Load the whole vector and avoid masking off the top bits as it makes
// the codegen worse.
SDValue Load =
DAG.getExtLoad(ISD::EXTLOAD, SL, LoadVT, Chain, BasePTR,
LD->getPointerInfo(), SrcIntVT, LD->getOriginalAlign(),
LD->getMemOperand()->getFlags(), LD->getAAInfo());
SmallVector<SDValue, 8> Vals;
for (unsigned Idx = 0; Idx < NumElem; ++Idx) {
unsigned ShiftIntoIdx =
(DAG.getDataLayout().isBigEndian() ? (NumElem - 1) - Idx : Idx);
SDValue ShiftAmount = DAG.getShiftAmountConstant(
ShiftIntoIdx * SrcEltVT.getSizeInBits(), LoadVT, SL);
SDValue ShiftedElt = DAG.getNode(ISD::SRL, SL, LoadVT, Load, ShiftAmount);
SDValue Elt =
DAG.getNode(ISD::AND, SL, LoadVT, ShiftedElt, SrcEltBitMask);
SDValue Scalar = DAG.getNode(ISD::TRUNCATE, SL, SrcEltVT, Elt);
if (ExtType != ISD::NON_EXTLOAD) {
unsigned ExtendOp = ISD::getExtForLoadExtType(false, ExtType);
Scalar = DAG.getNode(ExtendOp, SL, DstEltVT, Scalar);
}
Vals.push_back(Scalar);
}
SDValue Value = DAG.getBuildVector(DstVT, SL, Vals);
return std::make_pair(Value, Load.getValue(1));
}
unsigned Stride = SrcEltVT.getSizeInBits() / 8;
assert(SrcEltVT.isByteSized());
SmallVector<SDValue, 8> Vals;
SmallVector<SDValue, 8> LoadChains;
for (unsigned Idx = 0; Idx < NumElem; ++Idx) {
SDValue ScalarLoad =
DAG.getExtLoad(ExtType, SL, DstEltVT, Chain, BasePTR,
LD->getPointerInfo().getWithOffset(Idx * Stride),
SrcEltVT, LD->getOriginalAlign(),
LD->getMemOperand()->getFlags(), LD->getAAInfo());
BasePTR = DAG.getObjectPtrOffset(SL, BasePTR, TypeSize::getFixed(Stride));
Vals.push_back(ScalarLoad.getValue(0));
LoadChains.push_back(ScalarLoad.getValue(1));
}
SDValue NewChain = DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoadChains);
SDValue Value = DAG.getBuildVector(DstVT, SL, Vals);
return std::make_pair(Value, NewChain);
}
SDValue TargetLowering::scalarizeVectorStore(StoreSDNode *ST,
SelectionDAG &DAG) const {
SDLoc SL(ST);
SDValue Chain = ST->getChain();
SDValue BasePtr = ST->getBasePtr();
SDValue Value = ST->getValue();
EVT StVT = ST->getMemoryVT();
if (StVT.isScalableVector())
report_fatal_error("Cannot scalarize scalable vector stores");
// The type of the data we want to save
EVT RegVT = Value.getValueType();
EVT RegSclVT = RegVT.getScalarType();
// The type of data as saved in memory.
EVT MemSclVT = StVT.getScalarType();
unsigned NumElem = StVT.getVectorNumElements();
// A vector must always be stored in memory as-is, i.e. without any padding
// between the elements, since various code depend on it, e.g. in the
// handling of a bitcast of a vector type to int, which may be done with a
// vector store followed by an integer load. A vector that does not have
// elements that are byte-sized must therefore be stored as an integer
// built out of the extracted vector elements.
if (!MemSclVT.isByteSized()) {
unsigned NumBits = StVT.getSizeInBits();
EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), NumBits);
SDValue CurrVal = DAG.getConstant(0, SL, IntVT);
for (unsigned Idx = 0; Idx < NumElem; ++Idx) {
SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, RegSclVT, Value,
DAG.getVectorIdxConstant(Idx, SL));
SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, MemSclVT, Elt);
SDValue ExtElt = DAG.getNode(ISD::ZERO_EXTEND, SL, IntVT, Trunc);
unsigned ShiftIntoIdx =
(DAG.getDataLayout().isBigEndian() ? (NumElem - 1) - Idx : Idx);
SDValue ShiftAmount =
DAG.getConstant(ShiftIntoIdx * MemSclVT.getSizeInBits(), SL, IntVT);
SDValue ShiftedElt =
DAG.getNode(ISD::SHL, SL, IntVT, ExtElt, ShiftAmount);
CurrVal = DAG.getNode(ISD::OR, SL, IntVT, CurrVal, ShiftedElt);
}
return DAG.getStore(Chain, SL, CurrVal, BasePtr, ST->getPointerInfo(),
ST->getOriginalAlign(), ST->getMemOperand()->getFlags(),
ST->getAAInfo());
}
// Store Stride in bytes
unsigned Stride = MemSclVT.getSizeInBits() / 8;
assert(Stride && "Zero stride!");
// Extract each of the elements from the original vector and save them into
// memory individually.
SmallVector<SDValue, 8> Stores;
for (unsigned Idx = 0; Idx < NumElem; ++Idx) {
SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, RegSclVT, Value,
DAG.getVectorIdxConstant(Idx, SL));
SDValue Ptr =
DAG.getObjectPtrOffset(SL, BasePtr, TypeSize::getFixed(Idx * Stride));
// This scalar TruncStore may be illegal, but we legalize it later.
SDValue Store = DAG.getTruncStore(
Chain, SL, Elt, Ptr, ST->getPointerInfo().getWithOffset(Idx * Stride),
MemSclVT, ST->getOriginalAlign(), ST->getMemOperand()->getFlags(),
ST->getAAInfo());
Stores.push_back(Store);
}
return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Stores);
}
std::pair<SDValue, SDValue>
TargetLowering::expandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG) const {
assert(LD->getAddressingMode() == ISD::UNINDEXED &&
"unaligned indexed loads not implemented!");
SDValue Chain = LD->getChain();
SDValue Ptr = LD->getBasePtr();
EVT VT = LD->getValueType(0);
EVT LoadedVT = LD->getMemoryVT();
SDLoc dl(LD);
auto &MF = DAG.getMachineFunction();
if (VT.isFloatingPoint() || VT.isVector()) {
EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
if (isTypeLegal(intVT) && isTypeLegal(LoadedVT)) {
if (!isOperationLegalOrCustom(ISD::LOAD, intVT) &&
LoadedVT.isVector()) {
// Scalarize the load and let the individual components be handled.
return scalarizeVectorLoad(LD, DAG);
}
// Expand to a (misaligned) integer load of the same size,
// then bitconvert to floating point or vector.
SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr,
LD->getMemOperand());
SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
if (LoadedVT != VT)
Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
ISD::ANY_EXTEND, dl, VT, Result);
return std::make_pair(Result, newLoad.getValue(1));
}
// Copy the value to a (aligned) stack slot using (unaligned) integer
// loads and stores, then do a (aligned) load from the stack slot.
MVT RegVT = getRegisterType(*DAG.getContext(), intVT);
unsigned LoadedBytes = LoadedVT.getStoreSize();
unsigned RegBytes = RegVT.getSizeInBits() / 8;
unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
// Make sure the stack slot is also aligned for the register type.
SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
auto FrameIndex = cast<FrameIndexSDNode>(StackBase.getNode())->getIndex();
SmallVector<SDValue, 8> Stores;
SDValue StackPtr = StackBase;
unsigned Offset = 0;
EVT PtrVT = Ptr.getValueType();
EVT StackPtrVT = StackPtr.getValueType();
SDValue PtrIncrement = DAG.getConstant(RegBytes, dl, PtrVT);
SDValue StackPtrIncrement = DAG.getConstant(RegBytes, dl, StackPtrVT);
// Do all but one copies using the full register width.
for (unsigned i = 1; i < NumRegs; i++) {
// Load one integer register's worth from the original location.
SDValue Load = DAG.getLoad(
RegVT, dl, Chain, Ptr, LD->getPointerInfo().getWithOffset(Offset),
LD->getOriginalAlign(), LD->getMemOperand()->getFlags(),
LD->getAAInfo());
// Follow the load with a store to the stack slot. Remember the store.
Stores.push_back(DAG.getStore(
Load.getValue(1), dl, Load, StackPtr,
MachinePointerInfo::getFixedStack(MF, FrameIndex, Offset)));
// Increment the pointers.
Offset += RegBytes;
Ptr = DAG.getObjectPtrOffset(dl, Ptr, PtrIncrement);
StackPtr = DAG.getObjectPtrOffset(dl, StackPtr, StackPtrIncrement);
}
// The last copy may be partial. Do an extending load.
EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
8 * (LoadedBytes - Offset));
SDValue Load =
DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(Offset), MemVT,
LD->getOriginalAlign(), LD->getMemOperand()->getFlags(),
LD->getAAInfo());
// Follow the load with a store to the stack slot. Remember the store.
// On big-endian machines this requires a truncating store to ensure
// that the bits end up in the right place.
Stores.push_back(DAG.getTruncStore(
Load.getValue(1), dl, Load, StackPtr,
MachinePointerInfo::getFixedStack(MF, FrameIndex, Offset), MemVT));
// The order of the stores doesn't matter - say it with a TokenFactor.
SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
// Finally, perform the original load only redirected to the stack slot.
Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
MachinePointerInfo::getFixedStack(MF, FrameIndex, 0),
LoadedVT);
// Callers expect a MERGE_VALUES node.
return std::make_pair(Load, TF);
}
assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
"Unaligned load of unsupported type.");
// Compute the new VT that is half the size of the old one. This is an
// integer MVT.
unsigned NumBits = LoadedVT.getSizeInBits();
EVT NewLoadedVT;
NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
NumBits >>= 1;
Align Alignment = LD->getOriginalAlign();
unsigned IncrementSize = NumBits / 8;
ISD::LoadExtType HiExtType = LD->getExtensionType();
// If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
if (HiExtType == ISD::NON_EXTLOAD)
HiExtType = ISD::ZEXTLOAD;
// Load the value in two parts
SDValue Lo, Hi;
if (DAG.getDataLayout().isLittleEndian()) {
Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(),
LD->getAAInfo());
Ptr = DAG.getObjectPtrOffset(dl, Ptr, TypeSize::getFixed(IncrementSize));
Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(IncrementSize),
NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(),
LD->getAAInfo());
} else {
Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(),
LD->getAAInfo());
Ptr = DAG.getObjectPtrOffset(dl, Ptr, TypeSize::getFixed(IncrementSize));
Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(IncrementSize),
NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(),
LD->getAAInfo());
}
// aggregate the two parts
SDValue ShiftAmount = DAG.getShiftAmountConstant(NumBits, VT, dl);
SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
Hi.getValue(1));
return std::make_pair(Result, TF);
}
SDValue TargetLowering::expandUnalignedStore(StoreSDNode *ST,
SelectionDAG &DAG) const {
assert(ST->getAddressingMode() == ISD::UNINDEXED &&
"unaligned indexed stores not implemented!");
SDValue Chain = ST->getChain();
SDValue Ptr = ST->getBasePtr();
SDValue Val = ST->getValue();
EVT VT = Val.getValueType();
Align Alignment = ST->getOriginalAlign();
auto &MF = DAG.getMachineFunction();
EVT StoreMemVT = ST->getMemoryVT();
SDLoc dl(ST);
if (StoreMemVT.isFloatingPoint() || StoreMemVT.isVector()) {
EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
if (isTypeLegal(intVT)) {
if (!isOperationLegalOrCustom(ISD::STORE, intVT) &&
StoreMemVT.isVector()) {
// Scalarize the store and let the individual components be handled.
SDValue Result = scalarizeVectorStore(ST, DAG);
return Result;
}
// Expand to a bitconvert of the value to the integer type of the
// same size, then a (misaligned) int store.
// FIXME: Does not handle truncating floating point stores!
SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
Alignment, ST->getMemOperand()->getFlags());
return Result;
}
// Do a (aligned) store to a stack slot, then copy from the stack slot
// to the final destination using (unaligned) integer loads and stores.
MVT RegVT = getRegisterType(
*DAG.getContext(),
EVT::getIntegerVT(*DAG.getContext(), StoreMemVT.getSizeInBits()));
EVT PtrVT = Ptr.getValueType();
unsigned StoredBytes = StoreMemVT.getStoreSize();
unsigned RegBytes = RegVT.getSizeInBits() / 8;
unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
// Make sure the stack slot is also aligned for the register type.
SDValue StackPtr = DAG.CreateStackTemporary(StoreMemVT, RegVT);
auto FrameIndex = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
// Perform the original store, only redirected to the stack slot.
SDValue Store = DAG.getTruncStore(
Chain, dl, Val, StackPtr,
MachinePointerInfo::getFixedStack(MF, FrameIndex, 0), StoreMemVT);
EVT StackPtrVT = StackPtr.getValueType();
SDValue PtrIncrement = DAG.getConstant(RegBytes, dl, PtrVT);
SDValue StackPtrIncrement = DAG.getConstant(RegBytes, dl, StackPtrVT);
SmallVector<SDValue, 8> Stores;
unsigned Offset = 0;
// Do all but one copies using the full register width.
for (unsigned i = 1; i < NumRegs; i++) {
// Load one integer register's worth from the stack slot.
SDValue Load = DAG.getLoad(
RegVT, dl, Store, StackPtr,
MachinePointerInfo::getFixedStack(MF, FrameIndex, Offset));
// Store it to the final location. Remember the store.
Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
ST->getPointerInfo().getWithOffset(Offset),
ST->getOriginalAlign(),
ST->getMemOperand()->getFlags()));
// Increment the pointers.
Offset += RegBytes;
StackPtr = DAG.getObjectPtrOffset(dl, StackPtr, StackPtrIncrement);
Ptr = DAG.getObjectPtrOffset(dl, Ptr, PtrIncrement);
}
// The last store may be partial. Do a truncating store. On big-endian
// machines this requires an extending load from the stack slot to ensure
// that the bits are in the right place.
EVT LoadMemVT =
EVT::getIntegerVT(*DAG.getContext(), 8 * (StoredBytes - Offset));
// Load from the stack slot.
SDValue Load = DAG.getExtLoad(
ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
MachinePointerInfo::getFixedStack(MF, FrameIndex, Offset), LoadMemVT);
Stores.push_back(
DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
ST->getPointerInfo().getWithOffset(Offset), LoadMemVT,
ST->getOriginalAlign(),
ST->getMemOperand()->getFlags(), ST->getAAInfo()));
// The order of the stores doesn't matter - say it with a TokenFactor.
SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
return Result;
}
assert(StoreMemVT.isInteger() && !StoreMemVT.isVector() &&
"Unaligned store of unknown type.");
// Get the half-size VT
EVT NewStoredVT = StoreMemVT.getHalfSizedIntegerVT(*DAG.getContext());
unsigned NumBits = NewStoredVT.getFixedSizeInBits();
unsigned IncrementSize = NumBits / 8;
// Divide the stored value in two parts.
SDValue ShiftAmount =
DAG.getShiftAmountConstant(NumBits, Val.getValueType(), dl);
SDValue Lo = Val;
// If Val is a constant, replace the upper bits with 0. The SRL will constant
// fold and not use the upper bits. A smaller constant may be easier to
// materialize.
if (auto *C = dyn_cast<ConstantSDNode>(Lo); C && !C->isOpaque())
Lo = DAG.getNode(
ISD::AND, dl, VT, Lo,
DAG.getConstant(APInt::getLowBitsSet(VT.getSizeInBits(), NumBits), dl,
VT));
SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
// Store the two parts
SDValue Store1, Store2;
Store1 = DAG.getTruncStore(Chain, dl,
DAG.getDataLayout().isLittleEndian() ? Lo : Hi,
Ptr, ST->getPointerInfo(), NewStoredVT, Alignment,
ST->getMemOperand()->getFlags());
Ptr = DAG.getObjectPtrOffset(dl, Ptr, TypeSize::getFixed(IncrementSize));
Store2 = DAG.getTruncStore(
Chain, dl, DAG.getDataLayout().isLittleEndian() ? Hi : Lo, Ptr,
ST->getPointerInfo().getWithOffset(IncrementSize), NewStoredVT, Alignment,
ST->getMemOperand()->getFlags(), ST->getAAInfo());
SDValue Result =
DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
return Result;
}
SDValue
TargetLowering::IncrementMemoryAddress(SDValue Addr, SDValue Mask,
const SDLoc &DL, EVT DataVT,
SelectionDAG &DAG,
bool IsCompressedMemory) const {
SDValue Increment;
EVT AddrVT = Addr.getValueType();
EVT MaskVT = Mask.getValueType();
assert(DataVT.getVectorElementCount() == MaskVT.getVectorElementCount() &&
"Incompatible types of Data and Mask");
if (IsCompressedMemory) {
if (DataVT.isScalableVector())
report_fatal_error(
"Cannot currently handle compressed memory with scalable vectors");
// Incrementing the pointer according to number of '1's in the mask.
EVT MaskIntVT = EVT::getIntegerVT(*DAG.getContext(), MaskVT.getSizeInBits());
SDValue MaskInIntReg = DAG.getBitcast(MaskIntVT, Mask);
if (MaskIntVT.getSizeInBits() < 32) {
MaskInIntReg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, MaskInIntReg);
MaskIntVT = MVT::i32;
}
// Count '1's with POPCNT.
Increment = DAG.getNode(ISD::CTPOP, DL, MaskIntVT, MaskInIntReg);
Increment = DAG.getZExtOrTrunc(Increment, DL, AddrVT);
// Scale is an element size in bytes.
SDValue Scale = DAG.getConstant(DataVT.getScalarSizeInBits() / 8, DL,
AddrVT);
Increment = DAG.getNode(ISD::MUL, DL, AddrVT, Increment, Scale);
} else if (DataVT.isScalableVector()) {
Increment = DAG.getVScale(DL, AddrVT,
APInt(AddrVT.getFixedSizeInBits(),
DataVT.getStoreSize().getKnownMinValue()));
} else
Increment = DAG.getConstant(DataVT.getStoreSize(), DL, AddrVT);
return DAG.getNode(ISD::ADD, DL, AddrVT, Addr, Increment);
}
static SDValue clampDynamicVectorIndex(SelectionDAG &DAG, SDValue Idx,
EVT VecVT, const SDLoc &dl,
ElementCount SubEC) {
assert(!(SubEC.isScalable() && VecVT.isFixedLengthVector()) &&
"Cannot index a scalable vector within a fixed-width vector");
unsigned NElts = VecVT.getVectorMinNumElements();
unsigned NumSubElts = SubEC.getKnownMinValue();
EVT IdxVT = Idx.getValueType();
if (VecVT.isScalableVector() && !SubEC.isScalable()) {
// If this is a constant index and we know the value plus the number of the
// elements in the subvector minus one is less than the minimum number of
// elements then it's safe to return Idx.
if (auto *IdxCst = dyn_cast<ConstantSDNode>(Idx))
if (IdxCst->getZExtValue() + (NumSubElts - 1) < NElts)
return Idx;
SDValue VS =
DAG.getVScale(dl, IdxVT, APInt(IdxVT.getFixedSizeInBits(), NElts));
unsigned SubOpcode = NumSubElts <= NElts ? ISD::SUB : ISD::USUBSAT;
SDValue Sub = DAG.getNode(SubOpcode, dl, IdxVT, VS,
DAG.getConstant(NumSubElts, dl, IdxVT));
return DAG.getNode(ISD::UMIN, dl, IdxVT, Idx, Sub);
}
if (isPowerOf2_32(NElts) && NumSubElts == 1) {
APInt Imm = APInt::getLowBitsSet(IdxVT.getSizeInBits(), Log2_32(NElts));
return DAG.getNode(ISD::AND, dl, IdxVT, Idx,
DAG.getConstant(Imm, dl, IdxVT));
}
unsigned MaxIndex = NumSubElts < NElts ? NElts - NumSubElts : 0;
return DAG.getNode(ISD::UMIN, dl, IdxVT, Idx,
DAG.getConstant(MaxIndex, dl, IdxVT));
}
SDValue TargetLowering::getVectorElementPointer(SelectionDAG &DAG,
SDValue VecPtr, EVT VecVT,
SDValue Index) const {
return getVectorSubVecPointer(
DAG, VecPtr, VecVT,
EVT::getVectorVT(*DAG.getContext(), VecVT.getVectorElementType(), 1),
Index);
}
SDValue TargetLowering::getVectorSubVecPointer(SelectionDAG &DAG,
SDValue VecPtr, EVT VecVT,
EVT SubVecVT,
SDValue Index) const {
SDLoc dl(Index);
// Make sure the index type is big enough to compute in.
Index = DAG.getZExtOrTrunc(Index, dl, VecPtr.getValueType());
EVT EltVT = VecVT.getVectorElementType();
// Calculate the element offset and add it to the pointer.
unsigned EltSize = EltVT.getFixedSizeInBits() / 8; // FIXME: should be ABI size.
assert(EltSize * 8 == EltVT.getFixedSizeInBits() &&
"Converting bits to bytes lost precision");
assert(SubVecVT.getVectorElementType() == EltVT &&
"Sub-vector must be a vector with matching element type");
Index = clampDynamicVectorIndex(DAG, Index, VecVT, dl,
SubVecVT.getVectorElementCount());
EVT IdxVT = Index.getValueType();
if (SubVecVT.isScalableVector())
Index =
DAG.getNode(ISD::MUL, dl, IdxVT, Index,
DAG.getVScale(dl, IdxVT, APInt(IdxVT.getSizeInBits(), 1)));
Index = DAG.getNode(ISD::MUL, dl, IdxVT, Index,
DAG.getConstant(EltSize, dl, IdxVT));
return DAG.getMemBasePlusOffset(VecPtr, Index, dl);
}
//===----------------------------------------------------------------------===//
// Implementation of Emulated TLS Model
//===----------------------------------------------------------------------===//
SDValue TargetLowering::LowerToTLSEmulatedModel(const GlobalAddressSDNode *GA,
SelectionDAG &DAG) const {
// Access to address of TLS varialbe xyz is lowered to a function call:
// __emutls_get_address( address of global variable named "__emutls_v.xyz" )
EVT PtrVT = getPointerTy(DAG.getDataLayout());
PointerType *VoidPtrType = PointerType::get(*DAG.getContext(), 0);
SDLoc dl(GA);
ArgListTy Args;
ArgListEntry Entry;
std::string NameString = ("__emutls_v." + GA->getGlobal()->getName()).str();
Module *VariableModule = const_cast<Module*>(GA->getGlobal()->getParent());
StringRef EmuTlsVarName(NameString);
GlobalVariable *EmuTlsVar = VariableModule->getNamedGlobal(EmuTlsVarName);
assert(EmuTlsVar && "Cannot find EmuTlsVar ");
Entry.Node = DAG.getGlobalAddress(EmuTlsVar, dl, PtrVT);
Entry.Ty = VoidPtrType;
Args.push_back(Entry);
SDValue EmuTlsGetAddr = DAG.getExternalSymbol("__emutls_get_address", PtrVT);
TargetLowering::CallLoweringInfo CLI(DAG);
CLI.setDebugLoc(dl).setChain(DAG.getEntryNode());
CLI.setLibCallee(CallingConv::C, VoidPtrType, EmuTlsGetAddr, std::move(Args));
std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
// TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
// At last for X86 targets, maybe good for other targets too?
MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
MFI.setAdjustsStack(true); // Is this only for X86 target?
MFI.setHasCalls(true);
assert((GA->getOffset() == 0) &&
"Emulated TLS must have zero offset in GlobalAddressSDNode");
return CallResult.first;
}
SDValue TargetLowering::lowerCmpEqZeroToCtlzSrl(SDValue Op,
SelectionDAG &DAG) const {
assert((Op->getOpcode() == ISD::SETCC) && "Input has to be a SETCC node.");
if (!isCtlzFast())
return SDValue();
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
SDLoc dl(Op);
if (isNullConstant(Op.getOperand(1)) && CC == ISD::SETEQ) {
EVT VT = Op.getOperand(0).getValueType();
SDValue Zext = Op.getOperand(0);
if (VT.bitsLT(MVT::i32)) {
VT = MVT::i32;
Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
}
unsigned Log2b = Log2_32(VT.getSizeInBits());
SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
DAG.getConstant(Log2b, dl, MVT::i32));
return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
}
return SDValue();
}
SDValue TargetLowering::expandIntMINMAX(SDNode *Node, SelectionDAG &DAG) const {
SDValue Op0 = Node->getOperand(0);
SDValue Op1 = Node->getOperand(1);
EVT VT = Op0.getValueType();
EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
unsigned Opcode = Node->getOpcode();
SDLoc DL(Node);
// umax(x,1) --> sub(x,cmpeq(x,0)) iff cmp result is allbits
if (Opcode == ISD::UMAX && llvm::isOneOrOneSplat(Op1, true) && BoolVT == VT &&
getBooleanContents(VT) == ZeroOrNegativeOneBooleanContent) {
Op0 = DAG.getFreeze(Op0);
SDValue Zero = DAG.getConstant(0, DL, VT);
return DAG.getNode(ISD::SUB, DL, VT, Op0,
DAG.getSetCC(DL, VT, Op0, Zero, ISD::SETEQ));
}
// umin(x,y) -> sub(x,usubsat(x,y))
// TODO: Missing freeze(Op0)?
if (Opcode == ISD::UMIN && isOperationLegal(ISD::SUB, VT) &&
isOperationLegal(ISD::USUBSAT, VT)) {
return DAG.getNode(ISD::SUB, DL, VT, Op0,
DAG.getNode(ISD::USUBSAT, DL, VT, Op0, Op1));
}
// umax(x,y) -> add(x,usubsat(y,x))
// TODO: Missing freeze(Op0)?
if (Opcode == ISD::UMAX && isOperationLegal(ISD::ADD, VT) &&
isOperationLegal(ISD::USUBSAT, VT)) {
return DAG.getNode(ISD::ADD, DL, VT, Op0,
DAG.getNode(ISD::USUBSAT, DL, VT, Op1, Op0));
}
// FIXME: Should really try to split the vector in case it's legal on a
// subvector.
if (VT.isVector() && !isOperationLegalOrCustom(ISD::VSELECT, VT))
return DAG.UnrollVectorOp(Node);
// Attempt to find an existing SETCC node that we can reuse.
// TODO: Do we need a generic doesSETCCNodeExist?
// TODO: Missing freeze(Op0)/freeze(Op1)?
auto buildMinMax = [&](ISD::CondCode PrefCC, ISD::CondCode AltCC,
ISD::CondCode PrefCommuteCC,
ISD::CondCode AltCommuteCC) {
SDVTList BoolVTList = DAG.getVTList(BoolVT);
for (ISD::CondCode CC : {PrefCC, AltCC}) {
if (DAG.doesNodeExist(ISD::SETCC, BoolVTList,
{Op0, Op1, DAG.getCondCode(CC)})) {
SDValue Cond = DAG.getSetCC(DL, BoolVT, Op0, Op1, CC);
return DAG.getSelect(DL, VT, Cond, Op0, Op1);
}
}
for (ISD::CondCode CC : {PrefCommuteCC, AltCommuteCC}) {
if (DAG.doesNodeExist(ISD::SETCC, BoolVTList,
{Op0, Op1, DAG.getCondCode(CC)})) {
SDValue Cond = DAG.getSetCC(DL, BoolVT, Op0, Op1, CC);
return DAG.getSelect(DL, VT, Cond, Op1, Op0);
}
}
SDValue Cond = DAG.getSetCC(DL, BoolVT, Op0, Op1, PrefCC);
return DAG.getSelect(DL, VT, Cond, Op0, Op1);
};
// Expand Y = MAX(A, B) -> Y = (A > B) ? A : B
// -> Y = (A < B) ? B : A
// -> Y = (A >= B) ? A : B
// -> Y = (A <= B) ? B : A
switch (Opcode) {
case ISD::SMAX:
return buildMinMax(ISD::SETGT, ISD::SETGE, ISD::SETLT, ISD::SETLE);
case ISD::SMIN:
return buildMinMax(ISD::SETLT, ISD::SETLE, ISD::SETGT, ISD::SETGE);
case ISD::UMAX:
return buildMinMax(ISD::SETUGT, ISD::SETUGE, ISD::SETULT, ISD::SETULE);
case ISD::UMIN:
return buildMinMax(ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE);
}
llvm_unreachable("How did we get here?");
}
SDValue TargetLowering::expandAddSubSat(SDNode *Node, SelectionDAG &DAG) const {
unsigned Opcode = Node->getOpcode();
SDValue LHS = Node->getOperand(0);
SDValue RHS = Node->getOperand(1);
EVT VT = LHS.getValueType();
SDLoc dl(Node);
assert(VT == RHS.getValueType() && "Expected operands to be the same type");
assert(VT.isInteger() && "Expected operands to be integers");
// usub.sat(a, b) -> umax(a, b) - b
if (Opcode == ISD::USUBSAT && isOperationLegal(ISD::UMAX, VT)) {
SDValue Max = DAG.getNode(ISD::UMAX, dl, VT, LHS, RHS);
return DAG.getNode(ISD::SUB, dl, VT, Max, RHS);
}
// uadd.sat(a, b) -> umin(a, ~b) + b
if (Opcode == ISD::UADDSAT && isOperationLegal(ISD::UMIN, VT)) {
SDValue InvRHS = DAG.getNOT(dl, RHS, VT);
SDValue Min = DAG.getNode(ISD::UMIN, dl, VT, LHS, InvRHS);
return DAG.getNode(ISD::ADD, dl, VT, Min, RHS);
}
unsigned OverflowOp;
switch (Opcode) {
case ISD::SADDSAT:
OverflowOp = ISD::SADDO;
break;
case ISD::UADDSAT:
OverflowOp = ISD::UADDO;
break;
case ISD::SSUBSAT:
OverflowOp = ISD::SSUBO;
break;
case ISD::USUBSAT:
OverflowOp = ISD::USUBO;
break;
default:
llvm_unreachable("Expected method to receive signed or unsigned saturation "
"addition or subtraction node.");
}
// FIXME: Should really try to split the vector in case it's legal on a
// subvector.
if (VT.isVector() && !isOperationLegalOrCustom(ISD::VSELECT, VT))
return DAG.UnrollVectorOp(Node);
unsigned BitWidth = LHS.getScalarValueSizeInBits();
EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue Result = DAG.getNode(OverflowOp, dl, DAG.getVTList(VT, BoolVT), LHS, RHS);
SDValue SumDiff = Result.getValue(0);
SDValue Overflow = Result.getValue(1);
SDValue Zero = DAG.getConstant(0, dl, VT);
SDValue AllOnes = DAG.getAllOnesConstant(dl, VT);
if (Opcode == ISD::UADDSAT) {
if (getBooleanContents(VT) == ZeroOrNegativeOneBooleanContent) {
// (LHS + RHS) | OverflowMask
SDValue OverflowMask = DAG.getSExtOrTrunc(Overflow, dl, VT);
return DAG.getNode(ISD::OR, dl, VT, SumDiff, OverflowMask);
}
// Overflow ? 0xffff.... : (LHS + RHS)
return DAG.getSelect(dl, VT, Overflow, AllOnes, SumDiff);
}
if (Opcode == ISD::USUBSAT) {
if (getBooleanContents(VT) == ZeroOrNegativeOneBooleanContent) {
// (LHS - RHS) & ~OverflowMask
SDValue OverflowMask = DAG.getSExtOrTrunc(Overflow, dl, VT);
SDValue Not = DAG.getNOT(dl, OverflowMask, VT);
return DAG.getNode(ISD::AND, dl, VT, SumDiff, Not);
}
// Overflow ? 0 : (LHS - RHS)
return DAG.getSelect(dl, VT, Overflow, Zero, SumDiff);
}
if (Opcode == ISD::SADDSAT || Opcode == ISD::SSUBSAT) {
APInt MinVal = APInt::getSignedMinValue(BitWidth);
APInt MaxVal = APInt::getSignedMaxValue(BitWidth);
KnownBits KnownLHS = DAG.computeKnownBits(LHS);
KnownBits KnownRHS = DAG.computeKnownBits(RHS);
// If either of the operand signs are known, then they are guaranteed to
// only saturate in one direction. If non-negative they will saturate
// towards SIGNED_MAX, if negative they will saturate towards SIGNED_MIN.
//
// In the case of ISD::SSUBSAT, 'x - y' is equivalent to 'x + (-y)', so the
// sign of 'y' has to be flipped.
bool LHSIsNonNegative = KnownLHS.isNonNegative();
bool RHSIsNonNegative = Opcode == ISD::SADDSAT ? KnownRHS.isNonNegative()
: KnownRHS.isNegative();
if (LHSIsNonNegative || RHSIsNonNegative) {
SDValue SatMax = DAG.getConstant(MaxVal, dl, VT);
return DAG.getSelect(dl, VT, Overflow, SatMax, SumDiff);
}
bool LHSIsNegative = KnownLHS.isNegative();
bool RHSIsNegative = Opcode == ISD::SADDSAT ? KnownRHS.isNegative()
: KnownRHS.isNonNegative();
if (LHSIsNegative || RHSIsNegative) {
SDValue SatMin = DAG.getConstant(MinVal, dl, VT);
return DAG.getSelect(dl, VT, Overflow, SatMin, SumDiff);
}
}
// Overflow ? (SumDiff >> BW) ^ MinVal : SumDiff
APInt MinVal = APInt::getSignedMinValue(BitWidth);
SDValue SatMin = DAG.getConstant(MinVal, dl, VT);
SDValue Shift = DAG.getNode(ISD::SRA, dl, VT, SumDiff,
DAG.getConstant(BitWidth - 1, dl, VT));
Result = DAG.getNode(ISD::XOR, dl, VT, Shift, SatMin);
return DAG.getSelect(dl, VT, Overflow, Result, SumDiff);
}
SDValue TargetLowering::expandCMP(SDNode *Node, SelectionDAG &DAG) const {
unsigned Opcode = Node->getOpcode();
SDValue LHS = Node->getOperand(0);
SDValue RHS = Node->getOperand(1);
EVT VT = LHS.getValueType();
EVT ResVT = Node->getValueType(0);
EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDLoc dl(Node);
auto LTPredicate = (Opcode == ISD::UCMP ? ISD::SETULT : ISD::SETLT);
auto GTPredicate = (Opcode == ISD::UCMP ? ISD::SETUGT : ISD::SETGT);
SDValue IsLT = DAG.getSetCC(dl, BoolVT, LHS, RHS, LTPredicate);
SDValue IsGT = DAG.getSetCC(dl, BoolVT, LHS, RHS, GTPredicate);
// We can't perform arithmetic on i1 values. Extending them would
// probably result in worse codegen, so let's just use two selects instead.
// Some targets are also just better off using selects rather than subtraction
// because one of the conditions can be merged with one of the selects.
// And finally, if we don't know the contents of high bits of a boolean value
// we can't perform any arithmetic either.
if (shouldExpandCmpUsingSelects() || BoolVT.getScalarSizeInBits() == 1 ||
getBooleanContents(BoolVT) == UndefinedBooleanContent) {
SDValue SelectZeroOrOne =
DAG.getSelect(dl, ResVT, IsGT, DAG.getConstant(1, dl, ResVT),
DAG.getConstant(0, dl, ResVT));
return DAG.getSelect(dl, ResVT, IsLT, DAG.getConstant(-1, dl, ResVT),
SelectZeroOrOne);
}
if (getBooleanContents(BoolVT) == ZeroOrNegativeOneBooleanContent)
std::swap(IsGT, IsLT);
return DAG.getSExtOrTrunc(DAG.getNode(ISD::SUB, dl, BoolVT, IsGT, IsLT), dl,
ResVT);
}
SDValue TargetLowering::expandShlSat(SDNode *Node, SelectionDAG &DAG) const {
unsigned Opcode = Node->getOpcode();
bool IsSigned = Opcode == ISD::SSHLSAT;
SDValue LHS = Node->getOperand(0);
SDValue RHS = Node->getOperand(1);
EVT VT = LHS.getValueType();
SDLoc dl(Node);
assert((Node->getOpcode() == ISD::SSHLSAT ||
Node->getOpcode() == ISD::USHLSAT) &&
"Expected a SHLSAT opcode");
assert(VT == RHS.getValueType() && "Expected operands to be the same type");
assert(VT.isInteger() && "Expected operands to be integers");
if (VT.isVector() && !isOperationLegalOrCustom(ISD::VSELECT, VT))
return DAG.UnrollVectorOp(Node);
// If LHS != (LHS << RHS) >> RHS, we have overflow and must saturate.
unsigned BW = VT.getScalarSizeInBits();
EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue Result = DAG.getNode(ISD::SHL, dl, VT, LHS, RHS);
SDValue Orig =
DAG.getNode(IsSigned ? ISD::SRA : ISD::SRL, dl, VT, Result, RHS);
SDValue SatVal;
if (IsSigned) {
SDValue SatMin = DAG.getConstant(APInt::getSignedMinValue(BW), dl, VT);
SDValue SatMax = DAG.getConstant(APInt::getSignedMaxValue(BW), dl, VT);
SDValue Cond =
DAG.getSetCC(dl, BoolVT, LHS, DAG.getConstant(0, dl, VT), ISD::SETLT);
SatVal = DAG.getSelect(dl, VT, Cond, SatMin, SatMax);
} else {
SatVal = DAG.getConstant(APInt::getMaxValue(BW), dl, VT);
}
SDValue Cond = DAG.getSetCC(dl, BoolVT, LHS, Orig, ISD::SETNE);
return DAG.getSelect(dl, VT, Cond, SatVal, Result);
}
void TargetLowering::forceExpandWideMUL(SelectionDAG &DAG, const SDLoc &dl,
bool Signed, EVT WideVT,
const SDValue LL, const SDValue LH,
const SDValue RL, const SDValue RH,
SDValue &Lo, SDValue &Hi) const {
// We can fall back to a libcall with an illegal type for the MUL if we
// have a libcall big enough.
// Also, we can fall back to a division in some cases, but that's a big
// performance hit in the general case.
RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
if (WideVT == MVT::i16)
LC = RTLIB::MUL_I16;
else if (WideVT == MVT::i32)
LC = RTLIB::MUL_I32;
else if (WideVT == MVT::i64)
LC = RTLIB::MUL_I64;
else if (WideVT == MVT::i128)
LC = RTLIB::MUL_I128;
if (LC == RTLIB::UNKNOWN_LIBCALL || !getLibcallName(LC)) {
// We'll expand the multiplication by brute force because we have no other
// options. This is a trivially-generalized version of the code from
// Hacker's Delight (itself derived from Knuth's Algorithm M from section
// 4.3.1).
EVT VT = LL.getValueType();
unsigned Bits = VT.getSizeInBits();
unsigned HalfBits = Bits >> 1;
SDValue Mask =
DAG.getConstant(APInt::getLowBitsSet(Bits, HalfBits), dl, VT);
SDValue LLL = DAG.getNode(ISD::AND, dl, VT, LL, Mask);
SDValue RLL = DAG.getNode(ISD::AND, dl, VT, RL, Mask);
SDValue T = DAG.getNode(ISD::MUL, dl, VT, LLL, RLL);
SDValue TL = DAG.getNode(ISD::AND, dl, VT, T, Mask);
SDValue Shift = DAG.getShiftAmountConstant(HalfBits, VT, dl);
SDValue TH = DAG.getNode(ISD::SRL, dl, VT, T, Shift);
SDValue LLH = DAG.getNode(ISD::SRL, dl, VT, LL, Shift);
SDValue RLH = DAG.getNode(ISD::SRL, dl, VT, RL, Shift);
SDValue U = DAG.getNode(ISD::ADD, dl, VT,
DAG.getNode(ISD::MUL, dl, VT, LLH, RLL), TH);
SDValue UL = DAG.getNode(ISD::AND, dl, VT, U, Mask);
SDValue UH = DAG.getNode(ISD::SRL, dl, VT, U, Shift);
SDValue V = DAG.getNode(ISD::ADD, dl, VT,
DAG.getNode(ISD::MUL, dl, VT, LLL, RLH), UL);
SDValue VH = DAG.getNode(ISD::SRL, dl, VT, V, Shift);
SDValue W =
DAG.getNode(ISD::ADD, dl, VT, DAG.getNode(ISD::MUL, dl, VT, LLH, RLH),
DAG.getNode(ISD::ADD, dl, VT, UH, VH));
Lo = DAG.getNode(ISD::ADD, dl, VT, TL,
DAG.getNode(ISD::SHL, dl, VT, V, Shift));
Hi = DAG.getNode(ISD::ADD, dl, VT, W,
DAG.getNode(ISD::ADD, dl, VT,
DAG.getNode(ISD::MUL, dl, VT, RH, LL),
DAG.getNode(ISD::MUL, dl, VT, RL, LH)));
} else {
// Attempt a libcall.
SDValue Ret;
TargetLowering::MakeLibCallOptions CallOptions;
CallOptions.setSExt(Signed);
CallOptions.setIsPostTypeLegalization(true);
if (shouldSplitFunctionArgumentsAsLittleEndian(DAG.getDataLayout())) {
// Halves of WideVT are packed into registers in different order
// depending on platform endianness. This is usually handled by
// the C calling convention, but we can't defer to it in
// the legalizer.
SDValue Args[] = {LL, LH, RL, RH};
Ret = makeLibCall(DAG, LC, WideVT, Args, CallOptions, dl).first;
} else {
SDValue Args[] = {LH, LL, RH, RL};
Ret = makeLibCall(DAG, LC, WideVT, Args, CallOptions, dl).first;
}
assert(Ret.getOpcode() == ISD::MERGE_VALUES &&
"Ret value is a collection of constituent nodes holding result.");
if (DAG.getDataLayout().isLittleEndian()) {
// Same as above.
Lo = Ret.getOperand(0);
Hi = Ret.getOperand(1);
} else {
Lo = Ret.getOperand(1);
Hi = Ret.getOperand(0);
}
}
}
void TargetLowering::forceExpandWideMUL(SelectionDAG &DAG, const SDLoc &dl,
bool Signed, const SDValue LHS,
const SDValue RHS, SDValue &Lo,
SDValue &Hi) const {
EVT VT = LHS.getValueType();
assert(RHS.getValueType() == VT && "Mismatching operand types");
SDValue HiLHS;
SDValue HiRHS;
if (Signed) {
// The high part is obtained by SRA'ing all but one of the bits of low
// part.
unsigned LoSize = VT.getFixedSizeInBits();
HiLHS = DAG.getNode(
ISD::SRA, dl, VT, LHS,
DAG.getConstant(LoSize - 1, dl, getPointerTy(DAG.getDataLayout())));
HiRHS = DAG.getNode(
ISD::SRA, dl, VT, RHS,
DAG.getConstant(LoSize - 1, dl, getPointerTy(DAG.getDataLayout())));
} else {
HiLHS = DAG.getConstant(0, dl, VT);
HiRHS = DAG.getConstant(0, dl, VT);
}
EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
forceExpandWideMUL(DAG, dl, Signed, WideVT, LHS, HiLHS, RHS, HiRHS, Lo, Hi);
}
SDValue
TargetLowering::expandFixedPointMul(SDNode *Node, SelectionDAG &DAG) const {
assert((Node->getOpcode() == ISD::SMULFIX ||
Node->getOpcode() == ISD::UMULFIX ||
Node->getOpcode() == ISD::SMULFIXSAT ||
Node->getOpcode() == ISD::UMULFIXSAT) &&
"Expected a fixed point multiplication opcode");
SDLoc dl(Node);
SDValue LHS = Node->getOperand(0);
SDValue RHS = Node->getOperand(1);
EVT VT = LHS.getValueType();
unsigned Scale = Node->getConstantOperandVal(2);
bool Saturating = (Node->getOpcode() == ISD::SMULFIXSAT ||
Node->getOpcode() == ISD::UMULFIXSAT);
bool Signed = (Node->getOpcode() == ISD::SMULFIX ||
Node->getOpcode() == ISD::SMULFIXSAT);
EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
unsigned VTSize = VT.getScalarSizeInBits();
if (!Scale) {
// [us]mul.fix(a, b, 0) -> mul(a, b)
if (!Saturating) {
if (isOperationLegalOrCustom(ISD::MUL, VT))
return DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
} else if (Signed && isOperationLegalOrCustom(ISD::SMULO, VT)) {
SDValue Result =
DAG.getNode(ISD::SMULO, dl, DAG.getVTList(VT, BoolVT), LHS, RHS);
SDValue Product = Result.getValue(0);
SDValue Overflow = Result.getValue(1);
SDValue Zero = DAG.getConstant(0, dl, VT);
APInt MinVal = APInt::getSignedMinValue(VTSize);
APInt MaxVal = APInt::getSignedMaxValue(VTSize);
SDValue SatMin = DAG.getConstant(MinVal, dl, VT);
SDValue SatMax = DAG.getConstant(MaxVal, dl, VT);
// Xor the inputs, if resulting sign bit is 0 the product will be
// positive, else negative.
SDValue Xor = DAG.getNode(ISD::XOR, dl, VT, LHS, RHS);
SDValue ProdNeg = DAG.getSetCC(dl, BoolVT, Xor, Zero, ISD::SETLT);
Result = DAG.getSelect(dl, VT, ProdNeg, SatMin, SatMax);
return DAG.getSelect(dl, VT, Overflow, Result, Product);
} else if (!Signed && isOperationLegalOrCustom(ISD::UMULO, VT)) {
SDValue Result =
DAG.getNode(ISD::UMULO, dl, DAG.getVTList(VT, BoolVT), LHS, RHS);
SDValue Product = Result.getValue(0);
SDValue Overflow = Result.getValue(1);
APInt MaxVal = APInt::getMaxValue(VTSize);
SDValue SatMax = DAG.getConstant(MaxVal, dl, VT);
return DAG.getSelect(dl, VT, Overflow, SatMax, Product);
}
}
assert(((Signed && Scale < VTSize) || (!Signed && Scale <= VTSize)) &&
"Expected scale to be less than the number of bits if signed or at "
"most the number of bits if unsigned.");
assert(LHS.getValueType() == RHS.getValueType() &&
"Expected both operands to be the same type");
// Get the upper and lower bits of the result.
SDValue Lo, Hi;
unsigned LoHiOp = Signed ? ISD::SMUL_LOHI : ISD::UMUL_LOHI;
unsigned HiOp = Signed ? ISD::MULHS : ISD::MULHU;
EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VTSize * 2);
if (isOperationLegalOrCustom(LoHiOp, VT)) {
SDValue Result = DAG.getNode(LoHiOp, dl, DAG.getVTList(VT, VT), LHS, RHS);
Lo = Result.getValue(0);
Hi = Result.getValue(1);
} else if (isOperationLegalOrCustom(HiOp, VT)) {
Lo = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
Hi = DAG.getNode(HiOp, dl, VT, LHS, RHS);
} else if (isOperationLegalOrCustom(ISD::MUL, WideVT)) {
// Try for a multiplication using a wider type.
unsigned Ext = Signed ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
SDValue LHSExt = DAG.getNode(Ext, dl, WideVT, LHS);
SDValue RHSExt = DAG.getNode(Ext, dl, WideVT, RHS);
SDValue Res = DAG.getNode(ISD::MUL, dl, WideVT, LHSExt, RHSExt);
Lo = DAG.getNode(ISD::TRUNCATE, dl, VT, Res);
SDValue Shifted =
DAG.getNode(ISD::SRA, dl, WideVT, Res,
DAG.getShiftAmountConstant(VTSize, WideVT, dl));
Hi = DAG.getNode(ISD::TRUNCATE, dl, VT, Shifted);
} else if (VT.isVector()) {
return SDValue();
} else {
forceExpandWideMUL(DAG, dl, Signed, LHS, RHS, Lo, Hi);
}
if (Scale == VTSize)
// Result is just the top half since we'd be shifting by the width of the
// operand. Overflow impossible so this works for both UMULFIX and
// UMULFIXSAT.
return Hi;
// The result will need to be shifted right by the scale since both operands
// are scaled. The result is given to us in 2 halves, so we only want part of
// both in the result.
SDValue Result = DAG.getNode(ISD::FSHR, dl, VT, Hi, Lo,
DAG.getShiftAmountConstant(Scale, VT, dl));
if (!Saturating)
return Result;
if (!Signed) {
// Unsigned overflow happened if the upper (VTSize - Scale) bits (of the
// widened multiplication) aren't all zeroes.
// Saturate to max if ((Hi >> Scale) != 0),
// which is the same as if (Hi > ((1 << Scale) - 1))
APInt MaxVal = APInt::getMaxValue(VTSize);
SDValue LowMask = DAG.getConstant(APInt::getLowBitsSet(VTSize, Scale),
dl, VT);
Result = DAG.getSelectCC(dl, Hi, LowMask,
DAG.getConstant(MaxVal, dl, VT), Result,
ISD::SETUGT);
return Result;
}
// Signed overflow happened if the upper (VTSize - Scale + 1) bits (of the
// widened multiplication) aren't all ones or all zeroes.
SDValue SatMin = DAG.getConstant(APInt::getSignedMinValue(VTSize), dl, VT);
SDValue SatMax = DAG.getConstant(APInt::getSignedMaxValue(VTSize), dl, VT);
if (Scale == 0) {
SDValue Sign = DAG.getNode(ISD::SRA, dl, VT, Lo,
DAG.getShiftAmountConstant(VTSize - 1, VT, dl));
SDValue Overflow = DAG.getSetCC(dl, BoolVT, Hi, Sign, ISD::SETNE);
// Saturated to SatMin if wide product is negative, and SatMax if wide
// product is positive ...
SDValue Zero = DAG.getConstant(0, dl, VT);
SDValue ResultIfOverflow = DAG.getSelectCC(dl, Hi, Zero, SatMin, SatMax,
ISD::SETLT);
// ... but only if we overflowed.
return DAG.getSelect(dl, VT, Overflow, ResultIfOverflow, Result);
}
// We handled Scale==0 above so all the bits to examine is in Hi.
// Saturate to max if ((Hi >> (Scale - 1)) > 0),
// which is the same as if (Hi > (1 << (Scale - 1)) - 1)
SDValue LowMask = DAG.getConstant(APInt::getLowBitsSet(VTSize, Scale - 1),
dl, VT);
Result = DAG.getSelectCC(dl, Hi, LowMask, SatMax, Result, ISD::SETGT);
// Saturate to min if (Hi >> (Scale - 1)) < -1),
// which is the same as if (HI < (-1 << (Scale - 1))
SDValue HighMask =
DAG.getConstant(APInt::getHighBitsSet(VTSize, VTSize - Scale + 1),
dl, VT);
Result = DAG.getSelectCC(dl, Hi, HighMask, SatMin, Result, ISD::SETLT);
return Result;
}
SDValue
TargetLowering::expandFixedPointDiv(unsigned Opcode, const SDLoc &dl,
SDValue LHS, SDValue RHS,
unsigned Scale, SelectionDAG &DAG) const {
assert((Opcode == ISD::SDIVFIX || Opcode == ISD::SDIVFIXSAT ||
Opcode == ISD::UDIVFIX || Opcode == ISD::UDIVFIXSAT) &&
"Expected a fixed point division opcode");
EVT VT = LHS.getValueType();
bool Signed = Opcode == ISD::SDIVFIX || Opcode == ISD::SDIVFIXSAT;
bool Saturating = Opcode == ISD::SDIVFIXSAT || Opcode == ISD::UDIVFIXSAT;
EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
// If there is enough room in the type to upscale the LHS or downscale the
// RHS before the division, we can perform it in this type without having to
// resize. For signed operations, the LHS headroom is the number of
// redundant sign bits, and for unsigned ones it is the number of zeroes.
// The headroom for the RHS is the number of trailing zeroes.
unsigned LHSLead = Signed ? DAG.ComputeNumSignBits(LHS) - 1
: DAG.computeKnownBits(LHS).countMinLeadingZeros();
unsigned RHSTrail = DAG.computeKnownBits(RHS).countMinTrailingZeros();
// For signed saturating operations, we need to be able to detect true integer
// division overflow; that is, when you have MIN / -EPS. However, this
// is undefined behavior and if we emit divisions that could take such
// values it may cause undesired behavior (arithmetic exceptions on x86, for
// example).
// Avoid this by requiring an extra bit so that we never get this case.
// FIXME: This is a bit unfortunate as it means that for an 8-bit 7-scale
// signed saturating division, we need to emit a whopping 32-bit division.
if (LHSLead + RHSTrail < Scale + (unsigned)(Saturating && Signed))
return SDValue();
unsigned LHSShift = std::min(LHSLead, Scale);
unsigned RHSShift = Scale - LHSShift;
// At this point, we know that if we shift the LHS up by LHSShift and the
// RHS down by RHSShift, we can emit a regular division with a final scaling
// factor of Scale.
if (LHSShift)
LHS = DAG.getNode(ISD::SHL, dl, VT, LHS,
DAG.getShiftAmountConstant(LHSShift, VT, dl));
if (RHSShift)
RHS = DAG.getNode(Signed ? ISD::SRA : ISD::SRL, dl, VT, RHS,
DAG.getShiftAmountConstant(RHSShift, VT, dl));
SDValue Quot;
if (Signed) {
// For signed operations, if the resulting quotient is negative and the
// remainder is nonzero, subtract 1 from the quotient to round towards
// negative infinity.
SDValue Rem;
// FIXME: Ideally we would always produce an SDIVREM here, but if the
// type isn't legal, SDIVREM cannot be expanded. There is no reason why
// we couldn't just form a libcall, but the type legalizer doesn't do it.
if (isTypeLegal(VT) &&
isOperationLegalOrCustom(ISD::SDIVREM, VT)) {
Quot = DAG.getNode(ISD::SDIVREM, dl,
DAG.getVTList(VT, VT),
LHS, RHS);
Rem = Quot.getValue(1);
Quot = Quot.getValue(0);
} else {
Quot = DAG.getNode(ISD::SDIV, dl, VT,
LHS, RHS);
Rem = DAG.getNode(ISD::SREM, dl, VT,
LHS, RHS);
}
SDValue Zero = DAG.getConstant(0, dl, VT);
SDValue RemNonZero = DAG.getSetCC(dl, BoolVT, Rem, Zero, ISD::SETNE);
SDValue LHSNeg = DAG.getSetCC(dl, BoolVT, LHS, Zero, ISD::SETLT);
SDValue RHSNeg = DAG.getSetCC(dl, BoolVT, RHS, Zero, ISD::SETLT);
SDValue QuotNeg = DAG.getNode(ISD::XOR, dl, BoolVT, LHSNeg, RHSNeg);
SDValue Sub1 = DAG.getNode(ISD::SUB, dl, VT, Quot,
DAG.getConstant(1, dl, VT));
Quot = DAG.getSelect(dl, VT,
DAG.getNode(ISD::AND, dl, BoolVT, RemNonZero, QuotNeg),
Sub1, Quot);
} else
Quot = DAG.getNode(ISD::UDIV, dl, VT,
LHS, RHS);
return Quot;
}
void TargetLowering::expandUADDSUBO(
SDNode *Node, SDValue &Result, SDValue &Overflow, SelectionDAG &DAG) const {
SDLoc dl(Node);
SDValue LHS = Node->getOperand(0);
SDValue RHS = Node->getOperand(1);
bool IsAdd = Node->getOpcode() == ISD::UADDO;
// If UADDO_CARRY/SUBO_CARRY is legal, use that instead.
unsigned OpcCarry = IsAdd ? ISD::UADDO_CARRY : ISD::USUBO_CARRY;
if (isOperationLegalOrCustom(OpcCarry, Node->getValueType(0))) {
SDValue CarryIn = DAG.getConstant(0, dl, Node->getValueType(1));
SDValue NodeCarry = DAG.getNode(OpcCarry, dl, Node->getVTList(),
{ LHS, RHS, CarryIn });
Result = SDValue(NodeCarry.getNode(), 0);
Overflow = SDValue(NodeCarry.getNode(), 1);
return;
}
Result = DAG.getNode(IsAdd ? ISD::ADD : ISD::SUB, dl,
LHS.getValueType(), LHS, RHS);
EVT ResultType = Node->getValueType(1);
EVT SetCCType = getSetCCResultType(
DAG.getDataLayout(), *DAG.getContext(), Node->getValueType(0));
SDValue SetCC;
if (IsAdd && isOneConstant(RHS)) {
// Special case: uaddo X, 1 overflowed if X+1 is 0. This potential reduces
// the live range of X. We assume comparing with 0 is cheap.
// The general case (X + C) < C is not necessarily beneficial. Although we
// reduce the live range of X, we may introduce the materialization of
// constant C.
SetCC =
DAG.getSetCC(dl, SetCCType, Result,
DAG.getConstant(0, dl, Node->getValueType(0)), ISD::SETEQ);
} else if (IsAdd && isAllOnesConstant(RHS)) {
// Special case: uaddo X, -1 overflows if X != 0.
SetCC =
DAG.getSetCC(dl, SetCCType, LHS,
DAG.getConstant(0, dl, Node->getValueType(0)), ISD::SETNE);
} else {
ISD::CondCode CC = IsAdd ? ISD::SETULT : ISD::SETUGT;
SetCC = DAG.getSetCC(dl, SetCCType, Result, LHS, CC);
}
Overflow = DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType);
}
void TargetLowering::expandSADDSUBO(
SDNode *Node, SDValue &Result, SDValue &Overflow, SelectionDAG &DAG) const {
SDLoc dl(Node);
SDValue LHS = Node->getOperand(0);
SDValue RHS = Node->getOperand(1);
bool IsAdd = Node->getOpcode() == ISD::SADDO;
Result = DAG.getNode(IsAdd ? ISD::ADD : ISD::SUB, dl,
LHS.getValueType(), LHS, RHS);
EVT ResultType = Node->getValueType(1);
EVT OType = getSetCCResultType(
DAG.getDataLayout(), *DAG.getContext(), Node->getValueType(0));
// If SADDSAT/SSUBSAT is legal, compare results to detect overflow.
unsigned OpcSat = IsAdd ? ISD::SADDSAT : ISD::SSUBSAT;
if (isOperationLegal(OpcSat, LHS.getValueType())) {
SDValue Sat = DAG.getNode(OpcSat, dl, LHS.getValueType(), LHS, RHS);
SDValue SetCC = DAG.getSetCC(dl, OType, Result, Sat, ISD::SETNE);
Overflow = DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType);
return;
}
SDValue Zero = DAG.getConstant(0, dl, LHS.getValueType());
// For an addition, the result should be less than one of the operands (LHS)
// if and only if the other operand (RHS) is negative, otherwise there will
// be overflow.
// For a subtraction, the result should be less than one of the operands
// (LHS) if and only if the other operand (RHS) is (non-zero) positive,
// otherwise there will be overflow.
SDValue ResultLowerThanLHS = DAG.getSetCC(dl, OType, Result, LHS, ISD::SETLT);
SDValue ConditionRHS =
DAG.getSetCC(dl, OType, RHS, Zero, IsAdd ? ISD::SETLT : ISD::SETGT);
Overflow = DAG.getBoolExtOrTrunc(
DAG.getNode(ISD::XOR, dl, OType, ConditionRHS, ResultLowerThanLHS), dl,
ResultType, ResultType);
}
bool TargetLowering::expandMULO(SDNode *Node, SDValue &Result,
SDValue &Overflow, SelectionDAG &DAG) const {
SDLoc dl(Node);
EVT VT = Node->getValueType(0);
EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue LHS = Node->getOperand(0);
SDValue RHS = Node->getOperand(1);
bool isSigned = Node->getOpcode() == ISD::SMULO;
// For power-of-two multiplications we can use a simpler shift expansion.
if (ConstantSDNode *RHSC = isConstOrConstSplat(RHS)) {
const APInt &C = RHSC->getAPIntValue();
// mulo(X, 1 << S) -> { X << S, (X << S) >> S != X }
if (C.isPowerOf2()) {
// smulo(x, signed_min) is same as umulo(x, signed_min).
bool UseArithShift = isSigned && !C.isMinSignedValue();
SDValue ShiftAmt = DAG.getShiftAmountConstant(C.logBase2(), VT, dl);
Result = DAG.getNode(ISD::SHL, dl, VT, LHS, ShiftAmt);
Overflow = DAG.getSetCC(dl, SetCCVT,
DAG.getNode(UseArithShift ? ISD::SRA : ISD::SRL,
dl, VT, Result, ShiftAmt),
LHS, ISD::SETNE);
return true;
}
}
EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getScalarSizeInBits() * 2);
if (VT.isVector())
WideVT =
EVT::getVectorVT(*DAG.getContext(), WideVT, VT.getVectorElementCount());
SDValue BottomHalf;
SDValue TopHalf;
static const unsigned Ops[2][3] =
{ { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
{ ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
if (isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
} else if (isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
RHS);
TopHalf = BottomHalf.getValue(1);
} else if (isTypeLegal(WideVT)) {
LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
SDValue Mul = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
BottomHalf = DAG.getNode(ISD::TRUNCATE, dl, VT, Mul);
SDValue ShiftAmt =
DAG.getShiftAmountConstant(VT.getScalarSizeInBits(), WideVT, dl);
TopHalf = DAG.getNode(ISD::TRUNCATE, dl, VT,
DAG.getNode(ISD::SRL, dl, WideVT, Mul, ShiftAmt));
} else {
if (VT.isVector())
return false;
forceExpandWideMUL(DAG, dl, isSigned, LHS, RHS, BottomHalf, TopHalf);
}
Result = BottomHalf;
if (isSigned) {
SDValue ShiftAmt = DAG.getShiftAmountConstant(
VT.getScalarSizeInBits() - 1, BottomHalf.getValueType(), dl);
SDValue Sign = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, ShiftAmt);
Overflow = DAG.getSetCC(dl, SetCCVT, TopHalf, Sign, ISD::SETNE);
} else {
Overflow = DAG.getSetCC(dl, SetCCVT, TopHalf,
DAG.getConstant(0, dl, VT), ISD::SETNE);
}
// Truncate the result if SetCC returns a larger type than needed.
EVT RType = Node->getValueType(1);
if (RType.bitsLT(Overflow.getValueType()))
Overflow = DAG.getNode(ISD::TRUNCATE, dl, RType, Overflow);
assert(RType.getSizeInBits() == Overflow.getValueSizeInBits() &&
"Unexpected result type for S/UMULO legalization");
return true;
}
SDValue TargetLowering::expandVecReduce(SDNode *Node, SelectionDAG &DAG) const {
SDLoc dl(Node);
unsigned BaseOpcode = ISD::getVecReduceBaseOpcode(Node->getOpcode());
SDValue Op = Node->getOperand(0);
EVT VT = Op.getValueType();
if (VT.isScalableVector())
report_fatal_error(
"Expanding reductions for scalable vectors is undefined.");
// Try to use a shuffle reduction for power of two vectors.
if (VT.isPow2VectorType()) {
while (VT.getVectorNumElements() > 1) {
EVT HalfVT = VT.getHalfNumVectorElementsVT(*DAG.getContext());
if (!isOperationLegalOrCustom(BaseOpcode, HalfVT))
break;
SDValue Lo, Hi;
std::tie(Lo, Hi) = DAG.SplitVector(Op, dl);
Op = DAG.getNode(BaseOpcode, dl, HalfVT, Lo, Hi, Node->getFlags());
VT = HalfVT;
}
}
EVT EltVT = VT.getVectorElementType();
unsigned NumElts = VT.getVectorNumElements();
SmallVector<SDValue, 8> Ops;
DAG.ExtractVectorElements(Op, Ops, 0, NumElts);
SDValue Res = Ops[0];
for (unsigned i = 1; i < NumElts; i++)
Res = DAG.getNode(BaseOpcode, dl, EltVT, Res, Ops[i], Node->getFlags());
// Result type may be wider than element type.
if (EltVT != Node->getValueType(0))
Res = DAG.getNode(ISD::ANY_EXTEND, dl, Node->getValueType(0), Res);
return Res;
}
SDValue TargetLowering::expandVecReduceSeq(SDNode *Node, SelectionDAG &DAG) const {
SDLoc dl(Node);
SDValue AccOp = Node->getOperand(0);
SDValue VecOp = Node->getOperand(1);
SDNodeFlags Flags = Node->getFlags();
EVT VT = VecOp.getValueType();
EVT EltVT = VT.getVectorElementType();
if (VT.isScalableVector())
report_fatal_error(
"Expanding reductions for scalable vectors is undefined.");
unsigned NumElts = VT.getVectorNumElements();
SmallVector<SDValue, 8> Ops;
DAG.ExtractVectorElements(VecOp, Ops, 0, NumElts);
unsigned BaseOpcode = ISD::getVecReduceBaseOpcode(Node->getOpcode());
SDValue Res = AccOp;
for (unsigned i = 0; i < NumElts; i++)
Res = DAG.getNode(BaseOpcode, dl, EltVT, Res, Ops[i], Flags);
return Res;
}
bool TargetLowering::expandREM(SDNode *Node, SDValue &Result,
SelectionDAG &DAG) const {
EVT VT = Node->getValueType(0);
SDLoc dl(Node);
bool isSigned = Node->getOpcode() == ISD::SREM;
unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
SDValue Dividend = Node->getOperand(0);
SDValue Divisor = Node->getOperand(1);
if (isOperationLegalOrCustom(DivRemOpc, VT)) {
SDVTList VTs = DAG.getVTList(VT, VT);
Result = DAG.getNode(DivRemOpc, dl, VTs, Dividend, Divisor).getValue(1);
return true;
}
if (isOperationLegalOrCustom(DivOpc, VT)) {
// X % Y -> X-X/Y*Y
SDValue Divide = DAG.getNode(DivOpc, dl, VT, Dividend, Divisor);
SDValue Mul = DAG.getNode(ISD::MUL, dl, VT, Divide, Divisor);
Result = DAG.getNode(ISD::SUB, dl, VT, Dividend, Mul);
return true;
}
return false;
}
SDValue TargetLowering::expandFP_TO_INT_SAT(SDNode *Node,
SelectionDAG &DAG) const {
bool IsSigned = Node->getOpcode() == ISD::FP_TO_SINT_SAT;
SDLoc dl(SDValue(Node, 0));
SDValue Src = Node->getOperand(0);
// DstVT is the result type, while SatVT is the size to which we saturate
EVT SrcVT = Src.getValueType();
EVT DstVT = Node->getValueType(0);
EVT SatVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
unsigned SatWidth = SatVT.getScalarSizeInBits();
unsigned DstWidth = DstVT.getScalarSizeInBits();
assert(SatWidth <= DstWidth &&
"Expected saturation width smaller than result width");
// Determine minimum and maximum integer values and their corresponding
// floating-point values.
APInt MinInt, MaxInt;
if (IsSigned) {
MinInt = APInt::getSignedMinValue(SatWidth).sext(DstWidth);
MaxInt = APInt::getSignedMaxValue(SatWidth).sext(DstWidth);
} else {
MinInt = APInt::getMinValue(SatWidth).zext(DstWidth);
MaxInt = APInt::getMaxValue(SatWidth).zext(DstWidth);
}
// We cannot risk emitting FP_TO_XINT nodes with a source VT of [b]f16, as
// libcall emission cannot handle this. Large result types will fail.
if (SrcVT == MVT::f16 || SrcVT == MVT::bf16) {
Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Src);
SrcVT = Src.getValueType();
}
APFloat MinFloat(DAG.EVTToAPFloatSemantics(SrcVT));
APFloat MaxFloat(DAG.EVTToAPFloatSemantics(SrcVT));
APFloat::opStatus MinStatus =
MinFloat.convertFromAPInt(MinInt, IsSigned, APFloat::rmTowardZero);
APFloat::opStatus MaxStatus =
MaxFloat.convertFromAPInt(MaxInt, IsSigned, APFloat::rmTowardZero);
bool AreExactFloatBounds = !(MinStatus & APFloat::opStatus::opInexact) &&
!(MaxStatus & APFloat::opStatus::opInexact);
SDValue MinFloatNode = DAG.getConstantFP(MinFloat, dl, SrcVT);
SDValue MaxFloatNode = DAG.getConstantFP(MaxFloat, dl, SrcVT);
// If the integer bounds are exactly representable as floats and min/max are
// legal, emit a min+max+fptoi sequence. Otherwise we have to use a sequence
// of comparisons and selects.
bool MinMaxLegal = isOperationLegal(ISD::FMINNUM, SrcVT) &&
isOperationLegal(ISD::FMAXNUM, SrcVT);
if (AreExactFloatBounds && MinMaxLegal) {
SDValue Clamped = Src;
// Clamp Src by MinFloat from below. If Src is NaN the result is MinFloat.
Clamped = DAG.getNode(ISD::FMAXNUM, dl, SrcVT, Clamped, MinFloatNode);
// Clamp by MaxFloat from above. NaN cannot occur.
Clamped = DAG.getNode(ISD::FMINNUM, dl, SrcVT, Clamped, MaxFloatNode);
// Convert clamped value to integer.
SDValue FpToInt = DAG.getNode(IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT,
dl, DstVT, Clamped);
// In the unsigned case we're done, because we mapped NaN to MinFloat,
// which will cast to zero.
if (!IsSigned)
return FpToInt;
// Otherwise, select 0 if Src is NaN.
SDValue ZeroInt = DAG.getConstant(0, dl, DstVT);
EVT SetCCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT);
SDValue IsNan = DAG.getSetCC(dl, SetCCVT, Src, Src, ISD::CondCode::SETUO);
return DAG.getSelect(dl, DstVT, IsNan, ZeroInt, FpToInt);
}
SDValue MinIntNode = DAG.getConstant(MinInt, dl, DstVT);
SDValue MaxIntNode = DAG.getConstant(MaxInt, dl, DstVT);
// Result of direct conversion. The assumption here is that the operation is
// non-trapping and it's fine to apply it to an out-of-range value if we
// select it away later.
SDValue FpToInt =
DAG.getNode(IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, dl, DstVT, Src);
SDValue Select = FpToInt;
EVT SetCCVT =
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT);
// If Src ULT MinFloat, select MinInt. In particular, this also selects
// MinInt if Src is NaN.
SDValue ULT = DAG.getSetCC(dl, SetCCVT, Src, MinFloatNode, ISD::SETULT);
Select = DAG.getSelect(dl, DstVT, ULT, MinIntNode, Select);
// If Src OGT MaxFloat, select MaxInt.
SDValue OGT = DAG.getSetCC(dl, SetCCVT, Src, MaxFloatNode, ISD::SETOGT);
Select = DAG.getSelect(dl, DstVT, OGT, MaxIntNode, Select);
// In the unsigned case we are done, because we mapped NaN to MinInt, which
// is already zero.
if (!IsSigned)
return Select;
// Otherwise, select 0 if Src is NaN.
SDValue ZeroInt = DAG.getConstant(0, dl, DstVT);
SDValue IsNan = DAG.getSetCC(dl, SetCCVT, Src, Src, ISD::CondCode::SETUO);
return DAG.getSelect(dl, DstVT, IsNan, ZeroInt, Select);
}
SDValue TargetLowering::expandRoundInexactToOdd(EVT ResultVT, SDValue Op,
const SDLoc &dl,
SelectionDAG &DAG) const {
EVT OperandVT = Op.getValueType();
if (OperandVT.getScalarType() == ResultVT.getScalarType())
return Op;
EVT ResultIntVT = ResultVT.changeTypeToInteger();
// We are rounding binary64/binary128 -> binary32 -> bfloat16. This
// can induce double-rounding which may alter the results. We can
// correct for this using a trick explained in: Boldo, Sylvie, and
// Guillaume Melquiond. "When double rounding is odd." 17th IMACS
// World Congress. 2005.
unsigned BitSize = OperandVT.getScalarSizeInBits();
EVT WideIntVT = OperandVT.changeTypeToInteger();
SDValue OpAsInt = DAG.getBitcast(WideIntVT, Op);
SDValue SignBit =
DAG.getNode(ISD::AND, dl, WideIntVT, OpAsInt,
DAG.getConstant(APInt::getSignMask(BitSize), dl, WideIntVT));
SDValue AbsWide;
if (isOperationLegalOrCustom(ISD::FABS, OperandVT)) {
AbsWide = DAG.getNode(ISD::FABS, dl, OperandVT, Op);
} else {
SDValue ClearedSign = DAG.getNode(
ISD::AND, dl, WideIntVT, OpAsInt,
DAG.getConstant(APInt::getSignedMaxValue(BitSize), dl, WideIntVT));
AbsWide = DAG.getBitcast(OperandVT, ClearedSign);
}
SDValue AbsNarrow = DAG.getFPExtendOrRound(AbsWide, dl, ResultVT);
SDValue AbsNarrowAsWide = DAG.getFPExtendOrRound(AbsNarrow, dl, OperandVT);
// We can keep the narrow value as-is if narrowing was exact (no
// rounding error), the wide value was NaN (the narrow value is also
// NaN and should be preserved) or if we rounded to the odd value.
SDValue NarrowBits = DAG.getNode(ISD::BITCAST, dl, ResultIntVT, AbsNarrow);
SDValue One = DAG.getConstant(1, dl, ResultIntVT);
SDValue NegativeOne = DAG.getAllOnesConstant(dl, ResultIntVT);
SDValue And = DAG.getNode(ISD::AND, dl, ResultIntVT, NarrowBits, One);
EVT ResultIntVTCCVT = getSetCCResultType(
DAG.getDataLayout(), *DAG.getContext(), And.getValueType());
SDValue Zero = DAG.getConstant(0, dl, ResultIntVT);
// The result is already odd so we don't need to do anything.
SDValue AlreadyOdd = DAG.getSetCC(dl, ResultIntVTCCVT, And, Zero, ISD::SETNE);
EVT WideSetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
AbsWide.getValueType());
// We keep results which are exact, odd or NaN.
SDValue KeepNarrow =
DAG.getSetCC(dl, WideSetCCVT, AbsWide, AbsNarrowAsWide, ISD::SETUEQ);
KeepNarrow = DAG.getNode(ISD::OR, dl, WideSetCCVT, KeepNarrow, AlreadyOdd);
// We morally performed a round-down if AbsNarrow is smaller than
// AbsWide.
SDValue NarrowIsRd =
DAG.getSetCC(dl, WideSetCCVT, AbsWide, AbsNarrowAsWide, ISD::SETOGT);
// If the narrow value is odd or exact, pick it.
// Otherwise, narrow is even and corresponds to either the rounded-up
// or rounded-down value. If narrow is the rounded-down value, we want
// the rounded-up value as it will be odd.
SDValue Adjust = DAG.getSelect(dl, ResultIntVT, NarrowIsRd, One, NegativeOne);
SDValue Adjusted = DAG.getNode(ISD::ADD, dl, ResultIntVT, NarrowBits, Adjust);
Op = DAG.getSelect(dl, ResultIntVT, KeepNarrow, NarrowBits, Adjusted);
int ShiftAmount = BitSize - ResultVT.getScalarSizeInBits();
SDValue ShiftCnst = DAG.getShiftAmountConstant(ShiftAmount, WideIntVT, dl);
SignBit = DAG.getNode(ISD::SRL, dl, WideIntVT, SignBit, ShiftCnst);
SignBit = DAG.getNode(ISD::TRUNCATE, dl, ResultIntVT, SignBit);
Op = DAG.getNode(ISD::OR, dl, ResultIntVT, Op, SignBit);
return DAG.getNode(ISD::BITCAST, dl, ResultVT, Op);
}
SDValue TargetLowering::expandFP_ROUND(SDNode *Node, SelectionDAG &DAG) const {
assert(Node->getOpcode() == ISD::FP_ROUND && "Unexpected opcode!");
SDValue Op = Node->getOperand(0);
EVT VT = Node->getValueType(0);
SDLoc dl(Node);
if (VT.getScalarType() == MVT::bf16) {
if (Node->getConstantOperandVal(1) == 1) {
return DAG.getNode(ISD::FP_TO_BF16, dl, VT, Node->getOperand(0));
}
EVT OperandVT = Op.getValueType();
SDValue IsNaN = DAG.getSetCC(
dl,
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), OperandVT),
Op, Op, ISD::SETUO);
// We are rounding binary64/binary128 -> binary32 -> bfloat16. This
// can induce double-rounding which may alter the results. We can
// correct for this using a trick explained in: Boldo, Sylvie, and
// Guillaume Melquiond. "When double rounding is odd." 17th IMACS
// World Congress. 2005.
EVT F32 = VT.isVector() ? VT.changeVectorElementType(MVT::f32) : MVT::f32;
EVT I32 = F32.changeTypeToInteger();
Op = expandRoundInexactToOdd(F32, Op, dl, DAG);
Op = DAG.getNode(ISD::BITCAST, dl, I32, Op);
// Conversions should set NaN's quiet bit. This also prevents NaNs from
// turning into infinities.
SDValue NaN =
DAG.getNode(ISD::OR, dl, I32, Op, DAG.getConstant(0x400000, dl, I32));
// Factor in the contribution of the low 16 bits.
SDValue One = DAG.getConstant(1, dl, I32);
SDValue Lsb = DAG.getNode(ISD::SRL, dl, I32, Op,
DAG.getShiftAmountConstant(16, I32, dl));
Lsb = DAG.getNode(ISD::AND, dl, I32, Lsb, One);
SDValue RoundingBias =
DAG.getNode(ISD::ADD, dl, I32, DAG.getConstant(0x7fff, dl, I32), Lsb);
SDValue Add = DAG.getNode(ISD::ADD, dl, I32, Op, RoundingBias);
// Don't round if we had a NaN, we don't want to turn 0x7fffffff into
// 0x80000000.
Op = DAG.getSelect(dl, I32, IsNaN, NaN, Add);
// Now that we have rounded, shift the bits into position.
Op = DAG.getNode(ISD::SRL, dl, I32, Op,
DAG.getShiftAmountConstant(16, I32, dl));
Op = DAG.getNode(ISD::BITCAST, dl, I32, Op);
EVT I16 = I32.isVector() ? I32.changeVectorElementType(MVT::i16) : MVT::i16;
Op = DAG.getNode(ISD::TRUNCATE, dl, I16, Op);
return DAG.getNode(ISD::BITCAST, dl, VT, Op);
}
return SDValue();
}
SDValue TargetLowering::expandVectorSplice(SDNode *Node,
SelectionDAG &DAG) const {
assert(Node->getOpcode() == ISD::VECTOR_SPLICE && "Unexpected opcode!");
assert(Node->getValueType(0).isScalableVector() &&
"Fixed length vector types expected to use SHUFFLE_VECTOR!");
EVT VT = Node->getValueType(0);
SDValue V1 = Node->getOperand(0);
SDValue V2 = Node->getOperand(1);
int64_t Imm = cast<ConstantSDNode>(Node->getOperand(2))->getSExtValue();
SDLoc DL(Node);
// Expand through memory thusly:
// Alloca CONCAT_VECTORS_TYPES(V1, V2) Ptr
// Store V1, Ptr
// Store V2, Ptr + sizeof(V1)
// If (Imm < 0)
// TrailingElts = -Imm
// Ptr = Ptr + sizeof(V1) - (TrailingElts * sizeof(VT.Elt))
// else
// Ptr = Ptr + (Imm * sizeof(VT.Elt))
// Res = Load Ptr
Align Alignment = DAG.getReducedAlign(VT, /*UseABI=*/false);
EVT MemVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
VT.getVectorElementCount() * 2);
SDValue StackPtr = DAG.CreateStackTemporary(MemVT.getStoreSize(), Alignment);
EVT PtrVT = StackPtr.getValueType();
auto &MF = DAG.getMachineFunction();
auto FrameIndex = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FrameIndex);
// Store the lo part of CONCAT_VECTORS(V1, V2)
SDValue StoreV1 = DAG.getStore(DAG.getEntryNode(), DL, V1, StackPtr, PtrInfo);
// Store the hi part of CONCAT_VECTORS(V1, V2)
SDValue OffsetToV2 = DAG.getVScale(
DL, PtrVT,
APInt(PtrVT.getFixedSizeInBits(), VT.getStoreSize().getKnownMinValue()));
SDValue StackPtr2 = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, OffsetToV2);
SDValue StoreV2 = DAG.getStore(StoreV1, DL, V2, StackPtr2, PtrInfo);
if (Imm >= 0) {
// Load back the required element. getVectorElementPointer takes care of
// clamping the index if it's out-of-bounds.
StackPtr = getVectorElementPointer(DAG, StackPtr, VT, Node->getOperand(2));
// Load the spliced result
return DAG.getLoad(VT, DL, StoreV2, StackPtr,
MachinePointerInfo::getUnknownStack(MF));
}
uint64_t TrailingElts = -Imm;
// NOTE: TrailingElts must be clamped so as not to read outside of V1:V2.
TypeSize EltByteSize = VT.getVectorElementType().getStoreSize();
SDValue TrailingBytes =
DAG.getConstant(TrailingElts * EltByteSize, DL, PtrVT);
if (TrailingElts > VT.getVectorMinNumElements()) {
SDValue VLBytes =
DAG.getVScale(DL, PtrVT,
APInt(PtrVT.getFixedSizeInBits(),
VT.getStoreSize().getKnownMinValue()));
TrailingBytes = DAG.getNode(ISD::UMIN, DL, PtrVT, TrailingBytes, VLBytes);
}
// Calculate the start address of the spliced result.
StackPtr2 = DAG.getNode(ISD::SUB, DL, PtrVT, StackPtr2, TrailingBytes);
// Load the spliced result
return DAG.getLoad(VT, DL, StoreV2, StackPtr2,
MachinePointerInfo::getUnknownStack(MF));
}
SDValue TargetLowering::expandVECTOR_COMPRESS(SDNode *Node,
SelectionDAG &DAG) const {
SDLoc DL(Node);
SDValue Vec = Node->getOperand(0);
SDValue Mask = Node->getOperand(1);
SDValue Passthru = Node->getOperand(2);
EVT VecVT = Vec.getValueType();
EVT ScalarVT = VecVT.getScalarType();
EVT MaskVT = Mask.getValueType();
EVT MaskScalarVT = MaskVT.getScalarType();
// Needs to be handled by targets that have scalable vector types.
if (VecVT.isScalableVector())
report_fatal_error("Cannot expand masked_compress for scalable vectors.");
SDValue StackPtr = DAG.CreateStackTemporary(
VecVT.getStoreSize(), DAG.getReducedAlign(VecVT, /*UseABI=*/false));
int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
MachinePointerInfo PtrInfo =
MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
MVT PositionVT = getVectorIdxTy(DAG.getDataLayout());
SDValue Chain = DAG.getEntryNode();
SDValue OutPos = DAG.getConstant(0, DL, PositionVT);
bool HasPassthru = !Passthru.isUndef();
// If we have a passthru vector, store it on the stack, overwrite the matching
// positions and then re-write the last element that was potentially
// overwritten even though mask[i] = false.
if (HasPassthru)
Chain = DAG.getStore(Chain, DL, Passthru, StackPtr, PtrInfo);
SDValue LastWriteVal;
APInt PassthruSplatVal;
bool IsSplatPassthru =
ISD::isConstantSplatVector(Passthru.getNode(), PassthruSplatVal);
if (IsSplatPassthru) {
// As we do not know which position we wrote to last, we cannot simply
// access that index from the passthru vector. So we first check if passthru
// is a splat vector, to use any element ...
LastWriteVal = DAG.getConstant(PassthruSplatVal, DL, ScalarVT);
} else if (HasPassthru) {
// ... if it is not a splat vector, we need to get the passthru value at
// position = popcount(mask) and re-load it from the stack before it is
// overwritten in the loop below.
SDValue Popcount = DAG.getNode(
ISD::TRUNCATE, DL, MaskVT.changeVectorElementType(MVT::i1), Mask);
Popcount = DAG.getNode(ISD::ZERO_EXTEND, DL,
MaskVT.changeVectorElementType(ScalarVT), Popcount);
Popcount = DAG.getNode(ISD::VECREDUCE_ADD, DL, ScalarVT, Popcount);
SDValue LastElmtPtr =
getVectorElementPointer(DAG, StackPtr, VecVT, Popcount);
LastWriteVal = DAG.getLoad(
ScalarVT, DL, Chain, LastElmtPtr,
MachinePointerInfo::getUnknownStack(DAG.getMachineFunction()));
Chain = LastWriteVal.getValue(1);
}
unsigned NumElms = VecVT.getVectorNumElements();
for (unsigned I = 0; I < NumElms; I++) {
SDValue Idx = DAG.getVectorIdxConstant(I, DL);
SDValue ValI = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ScalarVT, Vec, Idx);
SDValue OutPtr = getVectorElementPointer(DAG, StackPtr, VecVT, OutPos);
Chain = DAG.getStore(
Chain, DL, ValI, OutPtr,
MachinePointerInfo::getUnknownStack(DAG.getMachineFunction()));
// Get the mask value and add it to the current output position. This
// either increments by 1 if MaskI is true or adds 0 otherwise.
// Freeze in case we have poison/undef mask entries.
SDValue MaskI = DAG.getFreeze(
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MaskScalarVT, Mask, Idx));
MaskI = DAG.getFreeze(MaskI);
MaskI = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, MaskI);
MaskI = DAG.getNode(ISD::ZERO_EXTEND, DL, PositionVT, MaskI);
OutPos = DAG.getNode(ISD::ADD, DL, PositionVT, OutPos, MaskI);
if (HasPassthru && I == NumElms - 1) {
SDValue EndOfVector =
DAG.getConstant(VecVT.getVectorNumElements() - 1, DL, PositionVT);
SDValue AllLanesSelected =
DAG.getSetCC(DL, MVT::i1, OutPos, EndOfVector, ISD::CondCode::SETUGT);
OutPos = DAG.getNode(ISD::UMIN, DL, PositionVT, OutPos, EndOfVector);
OutPtr = getVectorElementPointer(DAG, StackPtr, VecVT, OutPos);
// Re-write the last ValI if all lanes were selected. Otherwise,
// overwrite the last write it with the passthru value.
LastWriteVal =
DAG.getSelect(DL, ScalarVT, AllLanesSelected, ValI, LastWriteVal);
Chain = DAG.getStore(
Chain, DL, LastWriteVal, OutPtr,
MachinePointerInfo::getUnknownStack(DAG.getMachineFunction()));
}
}
return DAG.getLoad(VecVT, DL, Chain, StackPtr, PtrInfo);
}
bool TargetLowering::LegalizeSetCCCondCode(SelectionDAG &DAG, EVT VT,
SDValue &LHS, SDValue &RHS,
SDValue &CC, SDValue Mask,
SDValue EVL, bool &NeedInvert,
const SDLoc &dl, SDValue &Chain,
bool IsSignaling) const {
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
MVT OpVT = LHS.getSimpleValueType();
ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
NeedInvert = false;
assert(!EVL == !Mask && "VP Mask and EVL must either both be set or unset");
bool IsNonVP = !EVL;
switch (TLI.getCondCodeAction(CCCode, OpVT)) {
default:
llvm_unreachable("Unknown condition code action!");
case TargetLowering::Legal:
// Nothing to do.
break;
case TargetLowering::Expand: {
ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
if (TLI.isCondCodeLegalOrCustom(InvCC, OpVT)) {
std::swap(LHS, RHS);
CC = DAG.getCondCode(InvCC);
return true;
}
// Swapping operands didn't work. Try inverting the condition.
bool NeedSwap = false;
InvCC = getSetCCInverse(CCCode, OpVT);
if (!TLI.isCondCodeLegalOrCustom(InvCC, OpVT)) {
// If inverting the condition is not enough, try swapping operands
// on top of it.
InvCC = ISD::getSetCCSwappedOperands(InvCC);
NeedSwap = true;
}
if (TLI.isCondCodeLegalOrCustom(InvCC, OpVT)) {
CC = DAG.getCondCode(InvCC);
NeedInvert = true;
if (NeedSwap)
std::swap(LHS, RHS);
return true;
}
ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
unsigned Opc = 0;
switch (CCCode) {
default:
llvm_unreachable("Don't know how to expand this condition!");
case ISD::SETUO:
if (TLI.isCondCodeLegal(ISD::SETUNE, OpVT)) {
CC1 = ISD::SETUNE;
CC2 = ISD::SETUNE;
Opc = ISD::OR;
break;
}
assert(TLI.isCondCodeLegal(ISD::SETOEQ, OpVT) &&
"If SETUE is expanded, SETOEQ or SETUNE must be legal!");
NeedInvert = true;
[[fallthrough]];
case ISD::SETO:
assert(TLI.isCondCodeLegal(ISD::SETOEQ, OpVT) &&
"If SETO is expanded, SETOEQ must be legal!");
CC1 = ISD::SETOEQ;
CC2 = ISD::SETOEQ;
Opc = ISD::AND;
break;
case ISD::SETONE:
case ISD::SETUEQ:
// If the SETUO or SETO CC isn't legal, we might be able to use
// SETOGT || SETOLT, inverting the result for SETUEQ. We only need one
// of SETOGT/SETOLT to be legal, the other can be emulated by swapping
// the operands.
CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
if (!TLI.isCondCodeLegal(CC2, OpVT) &&
(TLI.isCondCodeLegal(ISD::SETOGT, OpVT) ||
TLI.isCondCodeLegal(ISD::SETOLT, OpVT))) {
CC1 = ISD::SETOGT;
CC2 = ISD::SETOLT;
Opc = ISD::OR;
NeedInvert = ((unsigned)CCCode & 0x8U);
break;
}
[[fallthrough]];
case ISD::SETOEQ:
case ISD::SETOGT:
case ISD::SETOGE:
case ISD::SETOLT:
case ISD::SETOLE:
case ISD::SETUNE:
case ISD::SETUGT:
case ISD::SETUGE:
case ISD::SETULT:
case ISD::SETULE:
// If we are floating point, assign and break, otherwise fall through.
if (!OpVT.isInteger()) {
// We can use the 4th bit to tell if we are the unordered
// or ordered version of the opcode.
CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
break;
}
// Fallthrough if we are unsigned integer.
[[fallthrough]];
case ISD::SETLE:
case ISD::SETGT:
case ISD::SETGE:
case ISD::SETLT:
case ISD::SETNE:
case ISD::SETEQ:
// If all combinations of inverting the condition and swapping operands
// didn't work then we have no means to expand the condition.
llvm_unreachable("Don't know how to expand this condition!");
}
SDValue SetCC1, SetCC2;
if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
// If we aren't the ordered or unorder operation,
// then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
if (IsNonVP) {
SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1, Chain, IsSignaling);
SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2, Chain, IsSignaling);
} else {
SetCC1 = DAG.getSetCCVP(dl, VT, LHS, RHS, CC1, Mask, EVL);
SetCC2 = DAG.getSetCCVP(dl, VT, LHS, RHS, CC2, Mask, EVL);
}
} else {
// Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
if (IsNonVP) {
SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1, Chain, IsSignaling);
SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2, Chain, IsSignaling);
} else {
SetCC1 = DAG.getSetCCVP(dl, VT, LHS, LHS, CC1, Mask, EVL);
SetCC2 = DAG.getSetCCVP(dl, VT, RHS, RHS, CC2, Mask, EVL);
}
}
if (Chain)
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, SetCC1.getValue(1),
SetCC2.getValue(1));
if (IsNonVP)
LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
else {
// Transform the binary opcode to the VP equivalent.
assert((Opc == ISD::OR || Opc == ISD::AND) && "Unexpected opcode");
Opc = Opc == ISD::OR ? ISD::VP_OR : ISD::VP_AND;
LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2, Mask, EVL);
}
RHS = SDValue();
CC = SDValue();
return true;
}
}
return false;
}
|