TaSTT.git - Free self-hosted STT for VRChat.

diff options

Diffstat (limited to 'BrowserSource')

0 files changed, 0 insertions, 0 deletions

ref='#n50'>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

// slang-ir-specialize.cpp
#include "slang-ir-specialize.h"

#include "slang-ir.h"
#include "slang-ir-clone.h"
#include "slang-ir-insts.h"

namespace Slang
{

// This file implements the primary specialization pass, that takes
// generic/polymorphic Slang code and specializes/monomorphises it.
//
// At present this primarily means generating specialized copies
// of generic functions/types based on the concrete types used
// at specialization sites, and also specializing instances
// of witness-table lookup to directly refer to the concrete
// values for witnesses when witness tables are known.
//
// This pass also performs some amount of simplification and
// specialization for code using existential (interface) types
// for local variables and function parameters/results.
//
// Eventually, this pass will also need to perform specialization
// of functions to argument values for parameters that must
// be compile-time constants,
//
// All of these passes are inter-related in that applying
// simplifications/specializations of one category can open
// up opportunities for transformations in the other categories.

struct SpecializationContext;

IRInst* specializeGenericImpl(
    IRGeneric*              genericVal,
    IRSpecialize*           specializeInst,
    IRModule*               module,
    SpecializationContext*  context);

struct SpecializationContext
{
    // For convenience, we will keep a pointer to the module
    // we are specializing.
    IRModule* module;

    // We know that we can only perform generic specialization when all
    // of the arguments to a generic are also fully specialized.
    // The "is fully specialized" condition is something we
    // need to solve for over the program, because the fully-
    // specialized-ness of an instruction depends on the
    // fully-specialized-ness of its operands.
    //
    // We will build an explicit hash set to encode those
    // instructions that are fully specialized.
    //
    HashSet<IRInst*> fullySpecializedInsts;

    // An instruction is then fully specialized if and only
    // if it is in our set.
    //
    bool isInstFullySpecialized(
        IRInst*                     inst)
    {
        // A small wrinkle is that a null instruction pointer
        // sometimes appears a a type, and so should be treated
        // as fully specialized too.
        //
        // TODO: It would be nice to remove this wrinkle.
        //
        if(!inst) return true;

        // An interface requirement entry should always be considered
        // to be fully specialized, even if it hasn't been visited.
        //
        // Note: This logic is here to stop a circularity, where we
        // can't mark an interface as used until its requirements are
        // used, etc.
        //
        if(inst->getOp() == kIROp_InterfaceRequirementEntry)
            return true;

        return fullySpecializedInsts.Contains(inst);
    }

    // When an instruction isn't fully specialized, but its operands *are*
    // then it is a candidate for specialization itself, so we will have
    // a query to check for the "all operands fully specialized" case.
    //
    bool areAllOperandsFullySpecialized(
        IRInst*                     inst)
    {
        if(!isInstFullySpecialized(inst->getFullType()))
            return false;

        UInt operandCount = inst->getOperandCount();
        for(UInt ii = 0; ii < operandCount; ++ii)
        {
            IRInst* operand = inst->getOperand(ii);
            if(!isInstFullySpecialized(operand))
                return false;
        }

        return true;
    }

    // We will use a single work list of instructions that need
    // to be considered for specialization or simplification,
    // whether generic, existential, etc.
    //
    OrderedHashSet<IRInst*> workList;
    HashSet<IRInst*> cleanInsts;

    void addToWorkList(
        IRInst* inst)
    {
        // We will ignore any code that is nested under a generic,
        // because it doesn't make sense to perform specialization
        // on such code.
        //
        for( auto ii = inst->getParent(); ii; ii = ii->getParent() )
        {
            if(as<IRGeneric>(ii))
                return;
        }

        if (workList.Add(inst))
        {
            cleanInsts.Remove(inst);

            addUsersToWorkList(inst);
        }
    }

    // When a transformation makes a change to an instruction,
    // we may need to re-consider transformations for instructions
    // that use its value. In those cases we will call `addUsersToWorkList`
    // on the instruction that is being modified or replaced.
    //
    void addUsersToWorkList(
        IRInst* inst)
    {
        for( auto use = inst->firstUse; use; use = use->nextUse )
        {
            auto user = use->getUser();
            
            addToWorkList(user);
        }
    }

    // One of the main transformations we will apply is to
    // consider an instruction as being fully specialized.
    //
    void markInstAsFullySpecialized(
        IRInst* inst)
    {
        if(fullySpecializedInsts.Contains(inst))
            return;
        fullySpecializedInsts.Add(inst);

        // If we know that an instruction is fully specialized,
        // then we should start to consider its uses and children
        // as candidates for being fully specialized too...
        //
        addUsersToWorkList(inst);
    }


    // Of course, somewhere along the way we expect
    // to run into uses of `specialize(...)` instructions
    // to bind a generic to arguments that we want to
    // specialize into concrete code.
    //
    // We also know that if we encouter `specialize(g, a, b, c)`
    // and then later `specialize(g, a, b, c)` again, we
    // only want to generate the specialized code for `g<a,b,c>`
    // *once*, and re-use it for both versions.
    //
    // We will cache existing specializations of generic function/types
    // using the simple key type defined as part of the IR cloning infrastructure.
    //
    typedef IRSimpleSpecializationKey Key;
    Dictionary<Key, IRInst*> genericSpecializations;

    // We will also use some shared IR building state across
    // all of our specialization/cloning steps.
    //
    SharedIRBuilder sharedBuilderStorage;

    // Now let's look at the task of finding or generation a
    // specialization of some generic `g`, given a specialization
    // instruction like `specialize(g, a, b, c)`.
    //
    // The `specializeGeneric` function will return a value
    // suitable for use as a replacement for the `specialize(...)`
    // instruction.
    //
    IRInst* specializeGeneric(
        IRGeneric*      genericVal,
        IRSpecialize*   specializeInst)
    {
        // First, we want to see if an existing specialization
        // has already been made. To do that we will construct a key
        // for lookup in the generic specialization context.
        //
        // Our key will consist of the identity of the generic
        // being specialized, and each of the argument values
        // being pased to it. In our hypothetical example of
        // `specialize(g, a, b, c)` the key will then be
        // the array `[g, a, b, c]`.
        //
        Key key;
        key.vals.add(specializeInst->getBase());
        UInt argCount = specializeInst->getArgCount();
        for( UInt ii = 0; ii < argCount; ++ii )
        {
            key.vals.add(specializeInst->getArg(ii));
        }

        {
            // We use our generated key to look for an
            // existing specialization that has been registered.
            // If one is found, our work is done.
            //
            IRInst* specializedVal = nullptr;
            if(genericSpecializations.TryGetValue(key, specializedVal))
                return specializedVal;
        }

        // If no existing specialization is found, we need
        // to create the specialization instead.
        // This mostly amounts to evaluating the generic as
        // if it were a function being called.
        //
        // We will use a free function to do the actual work
        // of evaluating the generic, so that the logic
        // can be re-used in other cases that need to
        // do one-off specialization.
        //
        IRInst* specializedVal = specializeGenericImpl(genericVal, specializeInst, module, this);


        // The value that was returned from evaluating
        // the generic is the specialized value, and we
        // need to remember it in our dictionary of
        // specializations so that we don't instantiate
        // this generic again for the same arguments.
        //
        genericSpecializations.Add(key, specializedVal);

        return specializedVal;
    }

    // The logic for generating a specialization of an IR generic
    // relies on the ability to "evaluate" the code in the body of
    // the generic, but that obviously doesn't work if we don't
    // actually have the full definition for the body.
    //
    // This can arise in particular for builtin operations/types.
    //
    // Before calling `specializeGeneric()` we need to make sure
    // that the generic is actually amenable to specialization,
    // by looking at whether it is a definition or a declaration.
    //
    bool canSpecializeGeneric(
        IRGeneric*  generic)
    {
        // It is possible to have multiple "layers" of generics
        // (e.g., when a generic function is nested in a generic
        // type). Therefore we need to drill down through all
        // of the layers present to see if at the leaf we have
        // something that looks like a definition.
        //
        IRGeneric* g = generic;
        for(;;)
        {
            // We can't specialize a generic if it is marked as
            // being imported from an external module (in which
            // case its definition is not available to us).
            //
            if(!isDefinition(g))
                return false;

            // Given the generic `g`, we will find the value
            // it appears to return in its body.
            //
            auto val = findGenericReturnVal(g);
            if(!val)
                return false;

            // If `g` returns an inner generic, then we need
            // to drill down further.
            //
            if (auto nestedGeneric = as<IRGeneric>(val))
            {
                g = nestedGeneric;
                continue;
            }

            // We should never specialize intrinsic types.
            //
            // TODO: This logic assumes that having *any* target
            // intrinsic decoration makes a type skip specialization,
            // even if the decoration isn't applicable to the
            // current target. This should be made true in practice
            // by having the linking step strip/skip decorations
            // that aren't applicable to the chosen target at link time.
            //
            if(as<IRStructType>(val) && val->findDecoration<IRTargetIntrinsicDecoration>())
                return false;

            // Once we've found the leaf value that will be produced
            // after all specialization is complete, we can check
            // whether it looks like a definition or not.
            //
            return isDefinition(val);
        }
    }

    // Now that we know when we can specialize a generic, and how
    // to do it, we can write a subroutine that takes a
    // `specialize(g, a, b, c, ...)` instruction and performs
    // specialization if it is possible.
    //
    void maybeSpecializeGeneric(
        IRSpecialize* specInst)
    {
        // We will only attempt to specialize when all of the
        // operands to the `speicalize(...)` instruction are
        // themselves fully specialized.
        //
        if(!areAllOperandsFullySpecialized(specInst))
            return;

        // The invariant that the arguments are fully specialized
        // should mean that `a, b, c, ...` are in a form that
        // we can work with, but it does *not* guarantee
        // that the `g` operand is something we can work with.
        //
        // We can only perform specialization in the case where
        // the base `g` is a known `generic` instruction.
        //
        auto baseVal = specInst->getBase();
        auto genericVal = as<IRGeneric>(baseVal);
        if(!genericVal)
            return;

        // We can also only specialize a generic if it
        // represents a definition rather than a declaration.
        //
        if(!canSpecializeGeneric(genericVal))
            return;

        // Once we know that specialization is possible,
        // the actual work is fairly simple.
        //
        // First, we find or generate a specialized
        // version of the result of the generic (a specialized
        // type, function, or whatever).
        //
        auto specializedVal = specializeGeneric(genericVal, specInst);

        // Any uses of this `specialize(...)` instruction will
        // become uses of `specializeVal`, so we want to re-consider
        // them for subsequent transformations.
        //
        addUsersToWorkList(specInst);

        // Then we simply replace any uses of the `specialize(...)`
        // instruction with the specialized value and delete
        // the `specialize(...)` instruction from existence.
        //
        specInst->replaceUsesWith(specializedVal);
        specInst->removeAndDeallocate();
    }

    // Generic specialization depends on identifying when
    // instructions are fully specialized.
    //
    void maybeMarkAsFullySpecialized(
        IRInst* inst)
    {
        // TODO: The logic here is completely bogus and
        // we need to revisit the notion of fully-specialized-ness
        // to only involve things that are semantically *values*
        // rather than computations/expressions.
        //
        // The rules should be something like:
        //
        // * Literals are values
        // * Composite type constructors where all the operands are value are values
        // * References to nominal types are values
        // * Built-in types where all the operands are values are values
        //
        // The system for defining value-ness probably needs
        // to combine with the system for deduplicating instructions,
        // since values are an important class of instruction we want
        // to deduplicate.

        switch(inst->getOp())
        {
        default:
            // The default case is that an instruction can
            // be considered as fully specialized as soon
            // as all of its operands are.
            //
            // TODO: We realistically need a more refined
            // check here that uses an allow-list of instructions
            // that can represent values suitable for use
            // as generic arguments.
            //
            if(areAllOperandsFullySpecialized(inst))
            {
                markInstAsFullySpecialized(inst);
            }
            break;

            // Certain instructions cannot ever be considered
            // fully specialized because they should never
            // be substituted into a generic as its arguments.
        case kIROp_lookup_interface_method:
        case kIROp_ExtractExistentialType:
        case kIROp_BindExistentialsType:
            break;

            // An interface type is always fully specialized.
        case kIROp_InterfaceType:
            markInstAsFullySpecialized(inst);
            break;

        case kIROp_Specialize:
            // The `specialize` instruction is a bit sepcial,
            // because it is possible to have a `specialize`
            // of a built-in type so that it never gets
            // substituted for another type. (e.g., the specific
            // case where this code path first showed up
            // as necessary was `RayQuery<>`)
            //
            {
                auto specialize = cast<IRSpecialize>(inst);
                auto base = specialize->getBase();
                if( auto generic = as<IRGeneric>(base) )
                {
                    // If the thing being specialized can be resolved,
                    // *and* it is a target intrinsic, ...
                    //
                    if( auto result = findGenericReturnVal(generic) )
                    {
                        if( result->findDecoration<IRTargetIntrinsicDecoration>() )
                        {
                            // ... then we should consider the instruction as
                            // "fully specialized" in the same cases as for
                            // any ordinary instruciton.
                            //

                            if( areAllOperandsFullySpecialized(inst) )
                            {
                                markInstAsFullySpecialized(inst);
                            }
                            return;
                        }
                    }
                }

                // Otherwise, a `specialize` instruction falls into
                // the case of instructions that should never be
                // considered to be fully specialized.
            }
            break;
        }
    }

    // The core of this pass is to look at one instruction
    // at a time, and try to perform whatever specialization
    // is appropriate based on its opcode.
    //
    void maybeSpecializeInst(
        IRInst*                     inst)
    {
        switch(inst->getOp())
        {
        default:
            // By default we assume that specialization is
            // not possible for a given opcode.
            //
            break;

        case kIROp_Specialize:
            // The logic for specializing a `specialize(...)`
            // instruction has already been elaborated above.
            //
            maybeSpecializeGeneric(cast<IRSpecialize>(inst));
            break;

        case kIROp_lookup_interface_method:
            // The remaining case we need to consider here for generics
            // is when we have a `lookup_witness_method` instruction
            // that is being applied to a concrete witness table,
            // because we can specialize it to just be a direct
            // reference to the actual witness value from the table.
            //
            maybeSpecializeWitnessLookup(cast<IRLookupWitnessMethod>(inst));
            break;

        case kIROp_Call:
            // When writing functions with existential-type parameters,
            // we need additional support to specialize a callee
            // function based on the concrete type encapsulated in
            // an argument of existential type.
            //
            maybeSpecializeExistentialsForCall(cast<IRCall>(inst));
            break;

            // The specialization of functions with existential-type
            // parameters can create further opportunities for specialization,
            // but in order to realize these we often need to propagate
            // through local simplification on values of existential type.
            //
        case kIROp_ExtractExistentialType:
            maybeSpecializeExtractExistentialType(inst);
            break;
        case kIROp_ExtractExistentialValue:
            maybeSpecializeExtractExistentialValue(inst);
            break;
        case kIROp_ExtractExistentialWitnessTable:
            maybeSpecializeExtractExistentialWitnessTable(inst);
            break;

        case kIROp_Load:
            maybeSpecializeLoad(as<IRLoad>(inst));
            break;

        case kIROp_FieldExtract:
            maybeSpecializeFieldExtract(as<IRFieldExtract>(inst));
            break;
        case kIROp_FieldAddress:
            maybeSpecializeFieldAddress(as<IRFieldAddress>(inst));
            break;

        case kIROp_getElement:
            maybeSpecializeGetElement(as<IRGetElement>(inst));
            break;
        case kIROp_getElementPtr:
            maybeSpecializeGetElementAddress(as<IRGetElementPtr>(inst));
            break;

        case kIROp_BindExistentialsType:
            maybeSpecializeBindExistentialsType(as<IRBindExistentialsType>(inst));
            break;
        }
    }

    // Specializing lookup on witness tables is a general
    // transformation that helps with both generic and
    // existential-based code. 
    //
    void maybeSpecializeWitnessLookup(
        IRLookupWitnessMethod* lookupInst)
    {
        // Note: While we currently have named the instruction
        // `lookup_witness_method`, the `method` part is a misnomer
        // and the same instruction can look up *any* interface
        // requirement based on the witness table that provides
        // a conformance, and the "key" that indicates the interface
        // requirement.

        // We can only specialize in the case where the lookup
        // is being done on a concrete witness table, and not
        // the result of a `specialize` instruction or other
        // operation that will yield such a table.
        //
        auto witnessTable = as<IRWitnessTable>(lookupInst->getWitnessTable());
        if(!witnessTable)
            return;

        // Because we have a concrete witness table, we can
        // use it to look up the IR value that satisfies
        // the given interface requirement.
        //
        auto requirementKey = lookupInst->getRequirementKey();
        auto satisfyingVal = findWitnessVal(witnessTable, requirementKey);

        // We expect to always find a satisfying value, but
        // we will go ahead and code defensively so that
        // we leave "correct" but unspecialized code if
        // we cannot find a concrete value to use.
        //
        if(!satisfyingVal)
            return;

        // At this point, we know that `satisfyingVal` is what
        // would result from executing this `lookup_witness_method`
        // instruction dynamically, so we can go ahead and
        // replace the original instruction with that value.
        //
        // We also make sure to add any uses of the lookup
        // instruction to our work list, because subsequent
        // simplifications might be possible now.
        //
        addUsersToWorkList(lookupInst);
        lookupInst->replaceUsesWith(satisfyingVal);
        lookupInst->removeAndDeallocate();
    }

    // The above subroutine needed a way to look up
    // the satisfying value for a given requirement
    // key in a concrete witness table, so let's
    // define that now.
    //
    IRInst* findWitnessVal(
        IRWitnessTable* witnessTable,
        IRInst*         requirementKey)
    {
        // A witness table is basically just a container
        // for key-value pairs, and so the best we can
        // do for now is a naive linear search.
        //
        for( auto entry : witnessTable->getEntries() )
        {
            if (requirementKey == entry->getRequirementKey())
            {
                return entry->getSatisfyingVal();
            }
        }
        return nullptr;
    }

    // All of the machinery for generic specialization
    // has been defined above, so we will now walk
    // through the flow of the overall specialization pass.
    //
    void processModule()
    {
        // We start by initializing our shared IR building state,
        // since we will re-use that state for any code we
        // generate along the way.
        //
        SharedIRBuilder* sharedBuilder = &sharedBuilderStorage;
        sharedBuilder->module = module;
        sharedBuilder->session = module->session;

        // The unspecialized IR we receive as input will have
        // `IRBindGlobalGenericParam` instructions that associate
        // each global-scope generic parameter (a type, witness
        // table, or what-have-you) with the value that it should
        // be bound to for the purposes of this code-generation
        // pass.
        //
        // Before doing any other specialization work, we will
        // iterate over these instructions (which may only
        // appear at the global scope) and use them to drive
        // replacement of the given generic type parameter with
        // the desired concrete value.
        //
        // TODO: When we start to support global shader parameters
        // that include existential/interface types, we will need
        // to support a similar specialization step for them.
        //
        specializeGlobalGenericParameters();

        // Now that we've eliminated all cases of global generic parameters,
        // we should now have the properties that:
        //
        // 1. Execution starts in non-generic code, with no unbound
        //    generic parameters in scope.
        //
        // 2. Any case where non-generic code makes use of a generic
        //    type/function, there will be a `specialize` instruction
        //    that specifies both the generic and the (concrete) type
        //    arguments that should be provided to it.
        //
        // The basic approach now is to look for opportunities to apply
        // our specialization rules (e.g., a `specialize` instruction
        // where all the type arguments are concrete types) and then
        // processing any additional opportunities created along the way.
        //
        // We start out simple by putting the root instruction for the
        // module onto our work list.
        //
        addToWorkList(module->getModuleInst());

        while(workList.Count() != 0)
        {

        // We will then iterate until our work list goes dry.
        //
        while(workList.Count() != 0)
        {
            IRInst* inst = workList.getLast();

            workList.removeLast();

            cleanInsts.Add(inst);

            // For each instruction we process, we want to perform
            // a few steps.
            //
            // First we will do any checking required to tag an
            // instruction as being fully specialized.
            //
            maybeMarkAsFullySpecialized(inst);

            // Next we will look for all the general-purpose
            // specialization opportunities (generic specialization,
            // existential specialization, simplifications, etc.)
            //
            maybeSpecializeInst(inst);

            // Finally, we need to make our logic recurse through
            // the whole IR module, so we want to add the children
            // of any parent instructions to our work list so that
            // we process them too.
            //
            // Note that we are adding the children of an instruction
            // in reverse order. This is because the way we are
            // using the work list treats it like a stack (LIFO) and
            // we know that fully-specialized-ness will tend to flow
            // top-down through the program, so that we want to process
            // the children of an instruction in their original order.
            //
            for(auto child = inst->getLastChild(); child; child = child->getPrevInst())
            {
                // Also note that `addToWorkList` has been written
                // to avoid adding any instruction that is a descendent
                // of an IR generic, because we don't actually want
                // to perform specialization inside of generics.
                //
                addToWorkList(child);
            }
        }

        addDirtyInstsToWorkListRec(module->getModuleInst());

        }

        // Once the work list has gone dry, we should have the invariant
        // that there are no `specialize` instructions inside of non-generic
        // functions that in turn reference a generic type/function, *except*
        // in the case where that generic is for a builtin type/function, in
        // which case we wouldn't want to specialize it anyway.
    }

    void addDirtyInstsToWorkListRec(IRInst* inst)
    {
        if( !cleanInsts.Contains(inst) )
        {
            addToWorkList(inst);
        }

        for(auto child = inst->getLastChild(); child; child = child->getPrevInst())
        {
            addDirtyInstsToWorkListRec(child);
        }
    }

    // Returns true if the call inst represents a call to
    // StructuredBuffer::operator[]/Load/Consume methods.
    bool isBufferLoadCall(IRCall* inst)
    {
        // TODO: We should have something like a `[knownSemantics(...)]` decoration
        // that can identify that an `IRFunc` has semantics that are consistent
        // with a list of compiler-known behaviors. The operand to `knownSemantics`
        // could come from a `KnownSemantics` enumeration or something similar,
        // so that we don't have to make string-based checks here.
        //
        // Note that `[knownSemantics(...)]` would be independent of any targets,
        // and could even apply to functions that are implemented entirely in Slang.

        if (auto targetIntrinsic = findAnyTargetIntrinsicDecoration(inst->getCallee()))
        {
            auto name = targetIntrinsic->getDefinition();
            if (name == ".operator[]" || name == ".Load" || name == ".Consume")
            {
                return true;
            }
        }
        return false;
    }

    /// Used by `maybeSpecailizeBufferLoadCall`, this function returns a new specialized callee that
    /// replaces a `specialize(.operator[], oldType)` to `specialize(.operator[], newElementType)`.
    IRInst* getNewSpecializedBufferLoadCallee(
        IRInst* oldSpecializedCallee,
        IRType* newContainerType,
        IRType* newElementType)
    {
        auto oldSpecialize = cast<IRSpecialize>(oldSpecializedCallee);
        SLANG_ASSERT(oldSpecialize->getArgCount() == 1);
        IRBuilder builder;
        builder.sharedBuilder = &sharedBuilderStorage;
        builder.setInsertBefore(oldSpecializedCallee);
        auto calleeType = builder.getFuncType(1, &newContainerType, newElementType);
        auto newSpecialize = builder.emitSpecializeInst(
            calleeType, oldSpecialize->getBase(), 1, (IRInst**)&newElementType);
        return newSpecialize;
    }

    /// Transform a buffer load intrinsic call.
    /// `bufferLoad(wrapExistential(bufferObj, wrapArgs), loadArgs)` should be transformed into
    /// `wrapExistential(bufferLoad(bufferObj, loadArgs), wragArgs)`.
    /// Returns true if `inst` matches the pattern and the load is transformed, otherwise,
    /// returns false.
    bool maybeSpecializeBufferLoadCall(IRCall* inst)
    {
        if (isBufferLoadCall(inst))
        {
            SLANG_ASSERT(inst->getArgCount() > 0);
            if (auto wrapExistential = as<IRWrapExistential>(inst->getArg(0)))
            {
                if (auto sbType = as<IRHLSLStructuredBufferTypeBase>(
                        wrapExistential->getWrappedValue()->getDataType()))
                {
                    // We are seeing the instruction sequence in the form of
                    // .operator[](wrapExistential(structuredBuffer), idx).
                    // Similar to handling load(wrapExistential(..)) insts,
                    // we need to replace it into wrapExistential(.operator[](sb, idx))
                    auto resultType = inst->getFullType();
                    auto elementType = sbType->getElementType();

                    IRBuilder builder;
                    builder.sharedBuilder = &sharedBuilderStorage;
                    builder.setInsertBefore(inst);

                    List<IRInst*> args;
                    args.add(wrapExistential->getWrappedValue());
                    for (UInt i = 1; i < inst->getArgCount(); i++)
                        args.add(inst->getArg(i));
                    List<IRInst*> slotOperands;
                    UInt slotOperandCount = wrapExistential->getSlotOperandCount();
                    for (UInt ii = 0; ii < slotOperandCount; ++ii)
                    {
                        slotOperands.add(wrapExistential->getSlotOperand(ii));
                    }
                    // The old callee should be in the form of `specialize(.operator[], IInterfaceType)`,
                    // we should update it to be `specialize(.operator[], elementType)`, so the return type
                    // of the load call is `elementType`.
                    auto oldCallee = inst->getCallee();

                    // A subscript operation on mutable buffers returns a ptr type instead of a value type.
                    // We need to make sure the pointer-ness is preserved correctly.
                    auto innerResultType = elementType;
                    if (auto ptrResultType = as<IRPtrType>(inst->getDataType()))
                    {
                        innerResultType = builder.getPtrType(elementType);
                    }
                    auto newCallee = getNewSpecializedBufferLoadCallee(inst->getCallee(), sbType, innerResultType);
                    auto newCall = builder.emitCallInst(innerResultType, newCallee, args);
                    auto newWrapExistential = builder.emitWrapExistential(
                        resultType, newCall, slotOperandCount, slotOperands.getBuffer());
                    inst->replaceUsesWith(newWrapExistential);
                    workList.Remove(inst);
                    inst->removeAndDeallocate();
                    SLANG_ASSERT(!oldCallee->hasUses());
                    workList.Remove(oldCallee);
                    oldCallee->removeAndDeallocate();
                    addUsersToWorkList(newWrapExistential);

                    workList.Remove(wrapExistential);
                    SLANG_ASSERT(!wrapExistential->hasUses());
                    wrapExistential->removeAndDeallocate();
                    return true;
                }
            }
        }
        return false;
    }

    // Given a `call` instruction in the IR, we need to detect the case
    // where the callee has some interface-type parameter(s) and at the
    // call site it is statically clear what concrete type(s) the arguments
    // will have.
    //
    void maybeSpecializeExistentialsForCall(IRCall* inst)
    {
        // Handle a special case of `StructuredBuffer.operator[]/Load/Consume`
        // calls first. These calls on builtin generic types should be handled
        // the same way as a `load` inst.
        if (maybeSpecializeBufferLoadCall(inst))
            return;

        // We can only specialize a call when the callee function is known.
        //
        auto calleeFunc = as<IRFunc>(inst->getCallee());
        if(!calleeFunc)
            return;

        // Update result type since the callee may have been changed.
        if (inst->getDataType() != calleeFunc->getResultType())
        {
            inst->setFullType(calleeFunc->getResultType());
        }

        // We can only specialize if we have access to a body for the callee.
        //
        if(!calleeFunc->isDefinition())
            return;

        // We shouldn't bother specializing unless the callee has at least
        // one parameter that has an existential/interface type.
        //
        bool shouldSpecialize = false;
        UInt argCounter = 0;
        for( auto param : calleeFunc->getParams() )
        {
            auto arg = inst->getArg(argCounter++);
            if( !isExistentialType(param->getDataType()) )
                continue;

            shouldSpecialize = true;

            // We *cannot* specialize unless the argument value corresponding
            // to such a parameter is one we can specialize.
            //
            if( !canSpecializeExistentialArg(arg))
                return;

        }
        // If we never found a parameter worth specializing, we should bail out.
        //
        if(!shouldSpecialize)
            return;

        // At this point, we believe we *should* and *can* specialize.
        //
        // We need a specialized variant of the callee (with the concrete
        // types substituted in for existential-type parameters), and then
        // we can replace the call site to call the new function instead.
        //
        // Any two call sites where the argument types are the same can
        // re-use the same callee, so we will  cache and re-use the
        // specialized functions that we generate (similar to how generic
        // specialization works). Therefore we will construct a key
        // for use when caching the specialized functions.
        // 
        IRSimpleSpecializationKey key;

        // The specialized callee will always depend on the unspecialized
        // function from which it is generated, so we add that to our key.
        //
        key.vals.add(calleeFunc);

        // Also, for any parameter that has an existential type, the
        // specialized function will depend on the concrete type of the
        // argument.
        //
        argCounter = 0;
        for( auto param : calleeFunc->getParams() )
        {
            auto arg = inst->getArg(argCounter++);
            if( !isExistentialType(param->getDataType()) )
                continue;

            if( auto makeExistential = as<IRMakeExistential>(arg) )
            {
                // Note that we use the *type* stored in the
                // existential-type argument, but not anything to
                // do with the particular value (otherwise we'd only
                // be able to re-use the specialized callee for
                // call sites that pass in the exact same argument).
                //
                auto val = makeExistential->getWrappedValue();
                auto valType = val->getFullType();
                key.vals.add(valType);

                // We are also including the witness table in the key.
                // This isn't required with our current language model,
                // since a given type can only conform to a given interface
                // in one way (so there can be only one witness table).
                // That means that the `valType` and the existential
                // type of `param` above should uniquely determine
                // the witness table we see.
                //
                // There are forward-looking cases where supporting
                // "overlapping conformances" could be required, and
                // there is low incremental cost to future-proofing
                // this code, so we go ahead and add the witness
                // table even if it is redundant.
                //
                auto witnessTable = makeExistential->getWitnessTable();
                key.vals.add(witnessTable);
            }
            else if( auto wrapExistential = as<IRWrapExistential>(arg) )
            {
                auto val = wrapExistential->getWrappedValue();
                auto valType = val->getFullType();
                key.vals.add(valType);

                UInt slotOperandCount = wrapExistential->getSlotOperandCount();
                for( UInt ii = 0; ii < slotOperandCount; ++ii )
                {
                    auto slotOperand = wrapExistential->getSlotOperand(ii);
                    key.vals.add(slotOperand);
                }
            }
            else
            {
                SLANG_UNEXPECTED("missing case for existential argument");
            }
        }

        // Once we've constructed our key, we can try to look for an
        // existing specialization of the callee that we can use.
        //
        IRFunc* specializedCallee = nullptr;
        if( !existentialSpecializedFuncs.TryGetValue(key, specializedCallee) )
        {
            // If we didn't find a specialized callee already made, then we
            // will go ahead and create one, and then register it in our cache.
            //
            specializedCallee = createExistentialSpecializedFunc(inst, calleeFunc);
            existentialSpecializedFuncs.Add(key, specializedCallee);
        }

        // At this point we have found or generated a specialized version
        // of the callee, and we need to emit a call to it.
        //
        // We will start by constructing the argument list for the new call.
        //
        argCounter = 0;
        List<IRInst*> newArgs;
        for( auto param : calleeFunc->getParams() )
        {
            auto arg = inst->getArg(argCounter++);

            // How we handle each argument depends on whether the corresponding
            // parameter has an existential type or not.
            //
            if( !isExistentialType(param->getDataType()) )
            {
                // If the parameter doesn't have an existential type, then we
                // don't want to change up the argument we pass at all.
                //
                newArgs.add(arg);
            }
            else
            {
                // Any place where the original function had a parameter of
                // existential type, we will now be passing in the concrete
                // argument value instead of an existential wrapper.
                //
                if( auto makeExistential = as<IRMakeExistential>(arg) )
                {
                    auto val = makeExistential->getWrappedValue();
                    newArgs.add(val);
                }
                else if( auto wrapExistential = as<IRWrapExistential>(arg) )
                {
                    auto val = wrapExistential->getWrappedValue();
                    newArgs.add(val);
                }
                else
                {
                    SLANG_UNEXPECTED("missing case for existential argument");
                }
            }
        }

        // Now that we've built up our argument list, it is simple enough
        // to construct a new `call` instruction.
        //
        IRBuilder builderStorage;
        auto builder = &builderStorage;
        builder->sharedBuilder = &sharedBuilderStorage;

        builder->setInsertBefore(inst);
        auto newCall = builder->emitCallInst(
            inst->getFullType(), specializedCallee, newArgs);

        // We will completely replace the old `call` instruction with the
        // new one, and will go so far as to transfer any decorations
        // that were attached to the old call over to the new one.
        //
        inst->transferDecorationsTo(newCall);
        inst->replaceUsesWith(newCall);
        inst->removeAndDeallocate();

        // Just in case, we will add any instructions that used the
        // result of this call to our work list for re-consideration.
        // At this moment this shouldn't open up new opportunities
        // for specialization, but we can always play it safe.
        //
        addUsersToWorkList(newCall);
    }

    // The above `maybeSpecializeExistentialsForCall` routine needed
    // a few utilities, which we will now define.

    // First, we want to be able to test whether a type (used by
    // a parameter) is an existential type so that we should specialize it.
    //
    bool isExistentialType(IRType* type)
    {
        // An IR-level interface type is always an existential.
        //
        if(as<IRInterfaceType>(type))
            return true;
        if (calcExistentialTypeParamSlotCount(type) != 0)
            return true;
        // Eventually we will also want to handle arrays over
        // existential types, but that will require careful
        // handling in many places.

        return false;
    }

    // Similarly, we want to be able to test whether an instruction
    // used as an argument for an existential-type parameter is
    // suitable for use in specialization.
    //
    bool canSpecializeExistentialArg(IRInst* inst)
    {
        // A `makeExistential(v, w)` instruction can be used
        // for specialization, since we have the concrete value `v`
        // (which implicitly determines the concrete type), and
        // the witness table `w.
        //
        if( auto makeExistential = as<IRMakeExistential>(inst) )
        {
            // We need to be careful about the type that we'd be specializing
            // to, since it needs to be visible to both the caller and calee.
            //
            // In particular, if the type is the result of a function-local
            // operation like `extractExistentialType`, then we can't possibly
            // specialize the callee, since it wouldn't be able to access
            // the same type (since the type is the result of an instruction in
            // the body of the caller)
            //
            auto concreteVal = makeExistential->getWrappedValue();
            auto concreteType = concreteVal->getDataType();

            // TODO: We probably need/want a more robust test here.
            // For now we are just listing the single opcode that is
            // causing problems.
            //
            // TODO: eventually this check would become unnecessary because
            // we can simply check if the `concreteType` is a compile-time
            // constant value.
            //
            if(concreteType->getOp() == kIROp_ExtractExistentialType)
                return false;

            return true;
        }

        // A `wrapExistential(v, T0,w0, T1, w1, ...)` instruction
        // is just a generalization of `makeExistential`, so it
        // can apply in the same cases.
        //
        if(as<IRWrapExistential>(inst))
            return true;

        // If we start to specialize functions that take arrays
        // of existentials as input, we will need a strategy to
        // determine arguments suitable for use in specializing
        // them (these would need to be arrays that nominally
        // have an existential element type, but somehow have
        // annotations to indicate that the concrete type
        // underlying the elements in homogeneous).

        return false;
    }

    // In order to cache and re-use functions that have had existential-type
    // parameters specialized, we need storage for the cache.
    //
    Dictionary<IRSimpleSpecializationKey, IRFunc*> existentialSpecializedFuncs;

    // The logic for creating a specialized callee function by plugging
    // in concrete types for existentials is similar to other cases of
    // specialization in the compiler.
    //
    IRFunc* createExistentialSpecializedFunc(
        IRCall* oldCall,
        IRFunc* oldFunc)
    {
        // We will make use of the infrastructure for cloning
        // IR code, that is defined in `ir-clone.{h,cpp}`.
        //
        // In order to do the cloning work we need an
        // "environment" that will map old values to
        // their replacements.
        //
        IRCloneEnv cloneEnv;

        // We also need some IR building state, for any
        // new instructions we will emit.
        //
        IRBuilder builderStorage;
        auto builder = &builderStorage;
        builder->sharedBuilder = &sharedBuilderStorage;

        // We will start out by determining what the parameters
        // of the specialized function should be, based on
        // the parameters of the original, and the concrete
        // type of selected arguments at the call site.
        //
        // Along the way we will build up explicit lists of
        // the parameters, as well as any new instructions
        // that need to be added to the body of the function
        // we generate (as a kind of "prologue"). We build
        // the lists here because we don't yet have a basic
        // block, or even a function, to insert them into.
        //
        List<IRParam*> newParams;
        List<IRInst*> newBodyInsts;
        UInt argCounter = 0;
        for( auto oldParam : oldFunc->getParams() )
        {
            auto arg = oldCall->getArg(argCounter++);

            // Given an old parameter, and the argument value at
            // the (old) call site, we need to determine what
            // value should stand in for that parameter in
            // the specialized callee.
            //
            IRInst* replacementVal = nullptr;

            // The trickier case is when we have an existential-type
            // parameter, because we need to extract out the concrete
            // type that is coming from the call site.
            //
            if( auto oldMakeExistential = as<IRMakeExistential>(arg) )
            {
                // In this case, the `arg` is `makeExistential(val, witnessTable)`
                // and we know that the specialized call site will just be
                // passing in `val`.
                //
                auto val = oldMakeExistential->getWrappedValue();
                auto witnessTable = oldMakeExistential->getWitnessTable();

                // Our specialized function needs to take a parameter with the
                // same type as `val`, to match the call site(s) that will be
                // created.
                //
                auto valType = val->getFullType();
                auto newParam = builder->createParam(valType);
                newParams.add(newParam);

                // Within the body of the function we cannot just use `val`
                // directly, because the existing code expects an existential
                // value, including its witness table.
                //
                // Therefore we will create a `makeExistential(newParam, witnessTable)`
                // in the body of the new function and use *that* as the replacement
                // value for the original parameter (since it will have the
                // correct existential type, and stores the right witness table).
                //
                auto newMakeExistential = builder->emitMakeExistential(oldParam->getFullType(), newParam, witnessTable);
                newBodyInsts.add(newMakeExistential);
                replacementVal = newMakeExistential;
            }
            else if( auto oldWrapExistential = as<IRWrapExistential>(arg) )
            {
                auto val = oldWrapExistential->getWrappedValue();
                auto valType = val->getFullType();

                auto newParam = builder->createParam(valType);
                newParams.add(newParam);

                // Within the body of the function we cannot just use `val`
                // directly, because the existing code expects an existential
                // value, including its witness table.
                //
                // Therefore we will create a `makeExistential(newParam, witnessTable)`
                // in the body of the new function and use *that* as the replacement
                // value for the original parameter (since it will have the
                // correct existential type, and stores the right witness table).
                //
                auto newWrapExistential = builder->emitWrapExistential(
                    oldParam->getFullType(),
                    newParam,
                    oldWrapExistential->getSlotOperandCount(),
                    oldWrapExistential->getSlotOperands());
                newBodyInsts.add(newWrapExistential);
                replacementVal = newWrapExistential;
            }
            else
            {
                // For parameters that don't have an existential type,
                // there is nothing interesting to do. The new function
                // will also have a parameter of the exact same type,
                // and we'll use that instead of the original parameter.
                //
                auto newParam = builder->createParam(oldParam->getFullType());
                newParams.add(newParam);
                replacementVal = newParam;
            }

            // Whatever replacement value was constructed, we need to
            // register it as the replacement for the original parameter.
            //
            cloneEnv.mapOldValToNew.Add(oldParam, replacementVal);
        }

        // Next we will create the skeleton of the new
        // specialized function, including its type.
        //
        // In order to construct the type of the new function, we
        // need to extract the types of all its parameters.
        //
        List<IRType*> newParamTypes;
        for( auto newParam : newParams )
        {
            newParamTypes.add(newParam->getFullType());
        }
        IRType* newFuncType = builder->getFuncType(
            newParamTypes.getCount(),
            newParamTypes.getBuffer(),
            oldFunc->getResultType());
        IRFunc* newFunc = builder->createFunc();
        newFunc->setFullType(newFuncType);

        // By construction, our new function type will be
        // "fully specialized" by the rules used for doing
        // generic specialization elsewhere in this pass.
        //
        fullySpecializedInsts.Add(newFuncType);

        // The above steps have accomplished the "first phase"
        // of cloning the function (since `IRFunc`s have no
        // operands).
        //
        // We can now use the shared IR cloning infrastructure
        // to perform the second phase of cloning, which will recursively
        // clone any nested decorations, blocks, and instructions.
        //
        cloneInstDecorationsAndChildren(
            &cloneEnv,
            builder->sharedBuilder,
            oldFunc,
            newFunc);

        // Now that the main body of existing isntructions have
        // been cloned into the new function, we can go ahead
        // and insert all the parameters and body instructions
        // we built up into the function at the right place.
        //
        // We expect the function to always have at least one
        // block (this was an invariant established before
        // we decided to specialize).
        //
        auto newEntryBlock = newFunc->getFirstBlock();
        SLANG_ASSERT(newEntryBlock);

        // We expect every valid block to have at least one
        // "ordinary" instruction (it will at least have
        // a terminator like a `return`).
        //
        auto newFirstOrdinary = newEntryBlock->getFirstOrdinaryInst();
        SLANG_ASSERT(newFirstOrdinary);

        // All of our parameters will get inserted before
        // the first ordinary instruction (since the function parameters
        // should come at the start of the first block).
        //
        for( auto newParam : newParams )
        {
            newParam->insertBefore(newFirstOrdinary);
        }

        // All of our new body instructions will *also* be inserted
        // before the first ordinary instruction (but will come
        // *after* the parameters by the order of these two loops).
        //
        for( auto newBodyInst : newBodyInsts )
        {
            newBodyInst->insertBefore(newFirstOrdinary);
        }

        // After all this work we have a valid `newFunc` that has been
        // specialized to match the types at the call site.
        //
        // There might be further opportunities for simplification and
        // specialization in the function body now that we've plugged
        // in some more concrete type information, so we will
        // add the whole function to our work list for subsequent
        // consideration.
        //
        addToWorkList(newFunc);

        return newFunc;
    }

    // When we've specialized a function with an interface-type parameter
    // we will still end up with a `makeExistential` operation in its
    // body, which could impede subequent specializations.
    //
    // For example, if we have the following after specialization:
    //
    //      e = makeExistential(v, w1);
    //      w2 = extractExistentialWitnessTable(e);
    //      f = lookup_witness_method(w2, k);
    //      call(f, ...);
    //
    // We cannot then specialize the lookup for `f` in this code as written,
    // but it seems obvious that we could replace `w2` with `w1` and maybe
    // get further along.
    //
    // In order to set up further specialization opportunities we need
    // to implement a few simplification rules around operations that
    // extract from an existential, when their operand is a `makeExistential`.
    //
    // Let's start with the routine for the case above of extracting
    // a witness table.
    //
    void maybeSpecializeExtractExistentialWitnessTable(IRInst* inst)
    {
        // We know `inst` is `extractExistentialWitnessTable(existentialArg)`.
        //
        auto existentialArg = inst->getOperand(0);

        if( auto makeExistential = as<IRMakeExistential>(existentialArg) )
        {
            // In this case we know `inst` is:
            //
            //      extractExistentialWitnessTable(makeExistential(..., witnessTable))
            //
            // and we can just simplify that to `witnessTable`.
            //
            auto witnessTable = makeExistential->getWitnessTable();

            // Anything that used this instruction is now a candidate for
            // further simplification or specialization (e.g., one of
            // the users of this instruction could be a `lookup_witness_method`
            // that we can now specialize).
            //
            addUsersToWorkList(inst);

            inst->replaceUsesWith(witnessTable);
            inst->removeAndDeallocate();
        }
    }

    // The cases for simplifying `extractExistentialValue` is more or less the same
    // as for witness tables.
    //
    void maybeSpecializeExtractExistentialValue(IRInst* inst)
    {
        // We know `inst` is `extractExistentialValue(existentialArg)`.
        //
        auto existentialArg = inst->getOperand(0);
        if( auto makeExistential = as<IRMakeExistential>(existentialArg) )
        {
            // Now we know `inst` is:
            //
            //      extractExistentialValue(makeExistential(val, ...))
            //
            // and we can just simplify that to `val`.
            //
            auto val = makeExistential->getWrappedValue();

            addUsersToWorkList(inst);

            inst->replaceUsesWith(val);
            inst->removeAndDeallocate();
        }
    }

    // The cases for simplifying `extractExistentialType` is more or less the same
    // as for witness tables.
    //
    void maybeSpecializeExtractExistentialType(IRInst* inst)
    {
        // We know `inst` is `extractExistentialValue(existentialArg)`.
        //
        auto existentialArg = inst->getOperand(0);
        if( auto makeExistential = as<IRMakeExistential>(existentialArg) )
        {
            // Now we know `inst` is:
            //
            //      extractExistentialType(makeExistential(val, ...))
            //
            // and we can just simplify that to type type of `val`.
            //
            auto val = makeExistential->getWrappedValue();
            auto valType = val->getFullType();

            addUsersToWorkList(inst);

            inst->replaceUsesWith(valType);
            inst->removeAndDeallocate();
        }
    }

    void maybeSpecializeLoad(IRLoad* inst)
    {
        auto ptrArg = inst->ptr.get();

        if( auto wrapInst = as<IRWrapExistential>(ptrArg) )
        {
            // We have an instruction of the form `load(wrapExistential(val, ...))`
            //
            auto val = wrapInst->getWrappedValue();

            // We know what type we are expected to
            // produce (which should be the pointed-to
            // type for whatever the type of the
            // `wrapExistential` is).
            //
            auto resultType = inst->getFullType();

            IRBuilder builder;
            builder.sharedBuilder = &sharedBuilderStorage;
            builder.setInsertBefore(inst);

            // We'd *like* to replace this instruction with
            // `wrapExistential(load(val))` instead, since that
            // will enable subsequent specializations.
            //
            // To do that, we need to be able to determine
            // the type that `load(val)` should return.
            //
            auto elementType = tryGetPointedToType(&builder, val->getDataType());
            if(!elementType)
                return;


            List<IRInst*> slotOperands;
            UInt slotOperandCount = wrapInst->getSlotOperandCount();
            for( UInt ii = 0; ii < slotOperandCount; ++ii )
            {
                slotOperands.add(wrapInst->getSlotOperand(ii));
            }

            auto newLoadInst = builder.emitLoad(elementType, val);
            auto newWrapExistentialInst = builder.emitWrapExistential(
                resultType,
                newLoadInst,
                slotOperandCount,
                slotOperands.getBuffer());

            addUsersToWorkList(inst);

            inst->replaceUsesWith(newWrapExistentialInst);
            inst->removeAndDeallocate();
        }
    }

    UInt calcExistentialBoxSlotCount(IRType* type)
    {
    top:
        if( as<IRBoundInterfaceType>(type) )
        {
            return 2;
        }
        else if( as<IRInterfaceType>(type) )
        {
            return 2;
        }
        else if( auto ptrType = as<IRPtrTypeBase>(type) )
        {
            type = ptrType->getValueType();
            goto top;
        }
        else if( auto ptrLikeType = as<IRPointerLikeType>(type) )
        {
            type = ptrLikeType->getElementType();
            goto top;
        }
        else if (auto arrayType = as<IRArrayTypeBase>(type))
        {
            type = arrayType->getElementType();
            goto top;
        }
        else if (auto sbType = as<IRHLSLStructuredBufferTypeBase>(type))
        {
            type = sbType->getElementType();
            goto top;
        }
        else if( auto structType = as<IRStructType>(type) )
        {
            UInt count = 0;
            for( auto field : structType->getFields() )
            {
                count += calcExistentialBoxSlotCount(field->getFieldType());
            }
            return count;
        }
        else
        {
            return 0;
        }
    }

    void maybeSpecializeFieldExtract(IRFieldExtract* inst)
    {
        auto baseArg = inst->getBase();
        auto fieldKey = inst->getField();

        if( auto wrapInst = as<IRWrapExistential>(baseArg) )
        {
            // We have `getField(wrapExistential(val, ...), fieldKey)`
            //
            auto val = wrapInst->getWrappedValue();

            // We know what type we are expected to produce.
            //
            auto resultType = inst->getFullType();

            IRBuilder builder;
            builder.sharedBuilder = &sharedBuilderStorage;
            builder.setInsertBefore(inst);

            // We'd *like* to replace this instruction with
            // `wrapExistential(getField(val, fieldKey), ...)` instead, since that
            // will enable subsequent specializations.
            //
            // To do that, we need to figure out:
            //
            // 1. What type that inner `getField` would return (what
            // is the type of the `fieldKey` field in `val`?)
            //
            // 2. Which of the existential slot operands in `...` there
            // actually apply to the given field.
            //

            // To determine these things, we need the type of
            // `val` to be a structure type so that we can look
            // up the field corresponding to `fieldKey`.
            //
            auto valType = val->getDataType();
            auto valStructType = as<IRStructType>(valType);
            if(!valStructType)
                return;

            UInt slotOperandOffset = 0;

            IRStructField* foundField = nullptr;
            for( auto valField : valStructType->getFields() )
            {
                if( valField->getKey() == fieldKey )
                {
                    foundField = valField;
                    break;
                }

                slotOperandOffset += calcExistentialBoxSlotCount(valField->getFieldType());
            }

            if(!foundField)
                return;

            auto foundFieldType = foundField->getFieldType();

            List<IRInst*> slotOperands;
            UInt slotOperandCount = calcExistentialBoxSlotCount(foundFieldType);

            for( UInt ii = 0; ii < slotOperandCount; ++ii )
            {
                slotOperands.add(wrapInst->getSlotOperand(slotOperandOffset + ii));
            }

            auto newGetField = builder.emitFieldExtract(
                foundFieldType,
                val,
                fieldKey);

            auto newWrapExistentialInst = builder.emitWrapExistential(
                resultType,
                newGetField,
                slotOperandCount,
                slotOperands.getBuffer());

            addUsersToWorkList(inst);
            inst->replaceUsesWith(newWrapExistentialInst);
            inst->removeAndDeallocate();
        }
    }


    void maybeSpecializeFieldAddress(IRFieldAddress* inst)
    {
        auto baseArg = inst->getBase();
        auto fieldKey = inst->getField();

        if( auto wrapInst = as<IRWrapExistential>(baseArg) )
        {
            // We have `getFieldAddr(wrapExistential(val, ...), fieldKey)`
            //
            auto val = wrapInst->getWrappedValue();

            // We know what type we are expected to produce.
            //
            auto resultType = inst->getFullType();

            IRBuilder builder;
            builder.sharedBuilder = &sharedBuilderStorage;
            builder.setInsertBefore(inst);

            // We'd *like* to replace this instruction with
            // `wrapExistential(getFieldAddr(val, fieldKey), ...)` instead, since that
            // will enable subsequent specializations.
            //
            // To do that, we need to figure out:
            //
            // 1. What type that inner `getFieldAddr` would return (what
            // is the type of the `fieldKey` field in `val`?)
            //
            // 2. Which of the existential slot operands in `...` there
            // actually apply to the given field.
            //

            // To determine these things, we need the type of
            // `val` to be a (pointer to a) structure type so that we can look
            // up the field corresponding to `fieldKey`.
            //
            auto valType = tryGetPointedToType(&builder, val->getDataType());
            if(!valType)
                return;

            auto valStructType = as<IRStructType>(valType);
            if(!valStructType)
                return;

            UInt slotOperandOffset = 0;

            IRStructField* foundField = nullptr;
            for( auto valField : valStructType->getFields() )
            {
                if( valField->getKey() == fieldKey )
                {
                    foundField = valField;
                    break;
                }

                slotOperandOffset += calcExistentialBoxSlotCount(valField->getFieldType());
            }

            if(!foundField)
                return;

            auto foundFieldType = foundField->getFieldType();

            List<IRInst*> slotOperands;
            UInt slotOperandCount = calcExistentialBoxSlotCount(foundFieldType);

            for( UInt ii = 0; ii < slotOperandCount; ++ii )
            {
                slotOperands.add(wrapInst->getSlotOperand(slotOperandOffset + ii));
            }

            auto newGetFieldAddr = builder.emitFieldAddress(
                builder.getPtrType(foundFieldType),
                val,
                fieldKey);

            auto newWrapExistentialInst = builder.emitWrapExistential(
                resultType,
                newGetFieldAddr,
                slotOperandCount,
                slotOperands.getBuffer());

            addUsersToWorkList(inst);
            inst->replaceUsesWith(newWrapExistentialInst);
            inst->removeAndDeallocate();
        }
    }

    void maybeSpecializeGetElement(IRGetElement* inst)
    {
        auto baseArg = inst->getBase();
        if (auto wrapInst = as<IRWrapExistential>(baseArg))
        {
            // We have `getElement(wrapExistential(val, ...), index)`
            // We need to replace this instruction with
            // `wrapExistential(getElement(val, index), ...)`
            auto index = inst->getIndex();

            auto val = wrapInst->getWrappedValue();
            auto resultType = inst->getFullType();

            IRBuilder builder;
            builder.sharedBuilder = &sharedBuilderStorage;
            builder.setInsertBefore(inst);

            auto elementType = cast<IRArrayTypeBase>(val->getDataType())->getElementType();

            List<IRInst*> slotOperands;
            UInt slotOperandCount = wrapInst->getSlotOperandCount();

            for (UInt ii = 0; ii < slotOperandCount; ++ii)
            {
                slotOperands.add(wrapInst->getSlotOperand(ii));
            }

            auto newGetElement = builder.emitElementExtract(elementType, val, index);

            auto newWrapExistentialInst = builder.emitWrapExistential(
                resultType, newGetElement, slotOperandCount, slotOperands.getBuffer());

            addUsersToWorkList(inst);
            inst->replaceUsesWith(newWrapExistentialInst);
            inst->removeAndDeallocate();
        }
    }

    void maybeSpecializeGetElementAddress(IRGetElementPtr* inst)
    {
        auto baseArg = inst->getBase();
        if (auto wrapInst = as<IRWrapExistential>(baseArg))
        {
            // We have `getElementPtr(wrapExistential(val, ...), index)`
            // We need to replace this instruction with
            // `wrapExistential(getElementPtr(val, index), ...)`
            auto index = inst->getIndex();

            auto val = wrapInst->getWrappedValue();
            auto ptrType = cast<IRPtrTypeBase>(val->getDataType());
            auto arrayType = cast<IRArrayTypeBase>(ptrType->getValueType());
            auto elementType = arrayType->getElementType();

            auto resultType = inst->getFullType();

            IRBuilder builder;
            builder.sharedBuilder = &sharedBuilderStorage;
            builder.setInsertBefore(inst);

            List<IRInst*> slotOperands;
            UInt slotOperandCount = wrapInst->getSlotOperandCount();

            for (UInt ii = 0; ii < slotOperandCount; ++ii)
            {
                slotOperands.add(wrapInst->getSlotOperand(ii));
            }

            auto elementPtrType = builder.getPtrType(ptrType->getOp(), elementType);
            auto newElementAddr = builder.emitElementAddress(elementPtrType, val, index);

            auto newWrapExistentialInst = builder.emitWrapExistential(
                resultType, newElementAddr, slotOperandCount, slotOperands.getBuffer());

            addUsersToWorkList(inst);
            inst->replaceUsesWith(newWrapExistentialInst);
            inst->removeAndDeallocate();
        }
    }

    UInt calcExistentialTypeParamSlotCount(IRType* type)
    {
    top:
        if( as<IRInterfaceType>(type) )
        {
            return 2;
        }
        else if( auto ptrType = as<IRPtrTypeBase>(type) )
        {
            type = ptrType->getValueType();
            goto top;
        }
        else if( auto ptrLikeType = as<IRPointerLikeType>(type) )
        {
            type = ptrLikeType->getElementType();
            goto top;
        }
        else if (auto sbType = as<IRHLSLStructuredBufferTypeBase>(type))
        {
            type = sbType->getElementType();
            goto top;
        }
        else if( auto structType = as<IRStructType>(type) )
        {
            UInt count = 0;
            for( auto field : structType->getFields() )
            {
                count += calcExistentialTypeParamSlotCount(field->getFieldType());
            }
            return count;
        }
        else
        {
            return 0;
        }
    }

    Dictionary<IRSimpleSpecializationKey, IRStructType*> existentialSpecializedStructs;

    void maybeSpecializeBindExistentialsType(IRBindExistentialsType* type)
    {
        auto baseType = type->getBaseType();
        UInt slotOperandCount = type->getExistentialArgCount();

        IRBuilder builder;
        builder.sharedBuilder = &sharedBuilderStorage;
        builder.setInsertBefore(type);

        if( auto baseInterfaceType = as<IRInterfaceType>(baseType) )
        {
            // We always expect two slot operands, one for the concrete type
            // and one for the witness table.
            //
            SLANG_ASSERT(slotOperandCount == 2);
            if(slotOperandCount < 2) return;

            auto concreteType = (IRType*) type->getExistentialArg(0);
            auto witnessTable = type->getExistentialArg(1);
            auto newVal = builder.getBoundInterfaceType(baseInterfaceType, concreteType, witnessTable);

            addUsersToWorkList(type);
            type->replaceUsesWith(newVal);
            type->removeAndDeallocate();
            return;
        }
        else if( as<IRPointerLikeType>(baseType) ||
                 as<IRHLSLStructuredBufferTypeBase>(baseType) ||
                 as<IRArrayTypeBase>(baseType))
        {
            // A `BindExistentials<P<T>, ...>` can be simplified to
            // `P<BindExistentials<T, ...>>` when `P` is a pointer-like
            // type constructor.
            //
            IRType* baseElementType = nullptr;
            if (auto basePtrLikeType = as<IRPointerLikeType>(baseType))
                baseElementType = basePtrLikeType->getElementType();
            else if (auto arrayType = as<IRArrayTypeBase>(baseType))
                baseElementType = arrayType->getElementType();
            else if (auto baseSBType = as<IRHLSLStructuredBufferTypeBase>(baseType))
                baseElementType = baseSBType->getElementType();

            IRInst* wrappedElementType = builder.getBindExistentialsType(
                baseElementType,
                slotOperandCount,
                type->getExistentialArgs());

            auto newPtrLikeType = builder.getType(
                baseType->getOp(),
                1,
                &wrappedElementType);
            addUsersToWorkList(type);
            addToWorkList(newPtrLikeType);
            addToWorkList(wrappedElementType);

            type->replaceUsesWith(newPtrLikeType);
            type->removeAndDeallocate();
            return;
        }
        else if( auto baseStructType = as<IRStructType>(baseType) )
        {
            // In order to bind a `struct` type we will generate
            // a new specialized `struct` type on demand and then
            // cache and re-use it.
            //
            // We don't want to start specializing here unless
            // all the operand types (and witness tables) we
            // will be specializing to are themselves fully
            // specialized, so that we can be sure that we
            // have a unique type.
            //
            if( !areAllOperandsFullySpecialized(type) )
                return;


context:
space:
mode: