rustc_codegen_ssa/back/write.rs
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
use std::any::Any;
use std::assert_matches::assert_matches;
use std::marker::PhantomData;
use std::path::{Path, PathBuf};
use std::sync::Arc;
use std::sync::mpsc::{Receiver, Sender, channel};
use std::{fs, io, mem, str, thread};
use jobserver::{Acquired, Client};
use rustc_ast::attr;
use rustc_data_structures::fx::{FxHashMap, FxIndexMap};
use rustc_data_structures::memmap::Mmap;
use rustc_data_structures::profiling::{SelfProfilerRef, VerboseTimingGuard};
use rustc_data_structures::sync::Lrc;
use rustc_errors::emitter::Emitter;
use rustc_errors::translation::Translate;
use rustc_errors::{
Diag, DiagArgMap, DiagCtxt, DiagMessage, ErrCode, FatalError, FluentBundle, Level, MultiSpan,
Style, Suggestions,
};
use rustc_fs_util::link_or_copy;
use rustc_hir::def_id::{CrateNum, LOCAL_CRATE};
use rustc_incremental::{
copy_cgu_workproduct_to_incr_comp_cache_dir, in_incr_comp_dir, in_incr_comp_dir_sess,
};
use rustc_metadata::EncodedMetadata;
use rustc_metadata::fs::copy_to_stdout;
use rustc_middle::bug;
use rustc_middle::dep_graph::{WorkProduct, WorkProductId};
use rustc_middle::middle::exported_symbols::SymbolExportInfo;
use rustc_middle::ty::TyCtxt;
use rustc_session::Session;
use rustc_session::config::{
self, CrateType, Lto, OutFileName, OutputFilenames, OutputType, Passes, SwitchWithOptPath,
};
use rustc_span::source_map::SourceMap;
use rustc_span::symbol::sym;
use rustc_span::{BytePos, FileName, InnerSpan, Pos, Span};
use rustc_target::spec::{MergeFunctions, SanitizerSet};
use tracing::debug;
use super::link::{self, ensure_removed};
use super::lto::{self, SerializedModule};
use super::symbol_export::symbol_name_for_instance_in_crate;
use crate::errors::ErrorCreatingRemarkDir;
use crate::traits::*;
use crate::{
CachedModuleCodegen, CodegenResults, CompiledModule, CrateInfo, ModuleCodegen, ModuleKind,
errors,
};
const PRE_LTO_BC_EXT: &str = "pre-lto.bc";
/// What kind of object file to emit.
#[derive(Clone, Copy, PartialEq)]
pub enum EmitObj {
// No object file.
None,
// Just uncompressed llvm bitcode. Provides easy compatibility with
// emscripten's ecc compiler, when used as the linker.
Bitcode,
// Object code, possibly augmented with a bitcode section.
ObjectCode(BitcodeSection),
}
/// What kind of llvm bitcode section to embed in an object file.
#[derive(Clone, Copy, PartialEq)]
pub enum BitcodeSection {
// No bitcode section.
None,
// A full, uncompressed bitcode section.
Full,
}
/// Module-specific configuration for `optimize_and_codegen`.
pub struct ModuleConfig {
/// Names of additional optimization passes to run.
pub passes: Vec<String>,
/// Some(level) to optimize at a certain level, or None to run
/// absolutely no optimizations (used for the metadata module).
pub opt_level: Option<config::OptLevel>,
/// Some(level) to optimize binary size, or None to not affect program size.
pub opt_size: Option<config::OptLevel>,
pub pgo_gen: SwitchWithOptPath,
pub pgo_use: Option<PathBuf>,
pub pgo_sample_use: Option<PathBuf>,
pub debug_info_for_profiling: bool,
pub instrument_coverage: bool,
pub instrument_gcov: bool,
pub sanitizer: SanitizerSet,
pub sanitizer_recover: SanitizerSet,
pub sanitizer_dataflow_abilist: Vec<String>,
pub sanitizer_memory_track_origins: usize,
// Flags indicating which outputs to produce.
pub emit_pre_lto_bc: bool,
pub emit_no_opt_bc: bool,
pub emit_bc: bool,
pub emit_ir: bool,
pub emit_asm: bool,
pub emit_obj: EmitObj,
pub emit_thin_lto: bool,
pub emit_thin_lto_summary: bool,
pub bc_cmdline: String,
// Miscellaneous flags. These are mostly copied from command-line
// options.
pub verify_llvm_ir: bool,
pub lint_llvm_ir: bool,
pub no_prepopulate_passes: bool,
pub no_builtins: bool,
pub time_module: bool,
pub vectorize_loop: bool,
pub vectorize_slp: bool,
pub merge_functions: bool,
pub emit_lifetime_markers: bool,
pub llvm_plugins: Vec<String>,
}
impl ModuleConfig {
fn new(
kind: ModuleKind,
tcx: TyCtxt<'_>,
no_builtins: bool,
is_compiler_builtins: bool,
) -> ModuleConfig {
// If it's a regular module, use `$regular`, otherwise use `$other`.
// `$regular` and `$other` are evaluated lazily.
macro_rules! if_regular {
($regular: expr, $other: expr) => {
if let ModuleKind::Regular = kind { $regular } else { $other }
};
}
let sess = tcx.sess;
let opt_level_and_size = if_regular!(Some(sess.opts.optimize), None);
let save_temps = sess.opts.cg.save_temps;
let should_emit_obj = sess.opts.output_types.contains_key(&OutputType::Exe)
|| match kind {
ModuleKind::Regular => sess.opts.output_types.contains_key(&OutputType::Object),
ModuleKind::Allocator => false,
ModuleKind::Metadata => sess.opts.output_types.contains_key(&OutputType::Metadata),
};
let emit_obj = if !should_emit_obj {
EmitObj::None
} else if sess.target.obj_is_bitcode
|| (sess.opts.cg.linker_plugin_lto.enabled() && !no_builtins)
{
// This case is selected if the target uses objects as bitcode, or
// if linker plugin LTO is enabled. In the linker plugin LTO case
// the assumption is that the final link-step will read the bitcode
// and convert it to object code. This may be done by either the
// native linker or rustc itself.
//
// Note, however, that the linker-plugin-lto requested here is
// explicitly ignored for `#![no_builtins]` crates. These crates are
// specifically ignored by rustc's LTO passes and wouldn't work if
// loaded into the linker. These crates define symbols that LLVM
// lowers intrinsics to, and these symbol dependencies aren't known
// until after codegen. As a result any crate marked
// `#![no_builtins]` is assumed to not participate in LTO and
// instead goes on to generate object code.
EmitObj::Bitcode
} else if need_bitcode_in_object(tcx) {
EmitObj::ObjectCode(BitcodeSection::Full)
} else {
EmitObj::ObjectCode(BitcodeSection::None)
};
ModuleConfig {
passes: if_regular!(sess.opts.cg.passes.clone(), vec![]),
opt_level: opt_level_and_size,
opt_size: opt_level_and_size,
pgo_gen: if_regular!(
sess.opts.cg.profile_generate.clone(),
SwitchWithOptPath::Disabled
),
pgo_use: if_regular!(sess.opts.cg.profile_use.clone(), None),
pgo_sample_use: if_regular!(sess.opts.unstable_opts.profile_sample_use.clone(), None),
debug_info_for_profiling: sess.opts.unstable_opts.debug_info_for_profiling,
instrument_coverage: if_regular!(sess.instrument_coverage(), false),
instrument_gcov: if_regular!(
// compiler_builtins overrides the codegen-units settings,
// which is incompatible with -Zprofile which requires that
// only a single codegen unit is used per crate.
sess.opts.unstable_opts.profile && !is_compiler_builtins,
false
),
sanitizer: if_regular!(sess.opts.unstable_opts.sanitizer, SanitizerSet::empty()),
sanitizer_dataflow_abilist: if_regular!(
sess.opts.unstable_opts.sanitizer_dataflow_abilist.clone(),
Vec::new()
),
sanitizer_recover: if_regular!(
sess.opts.unstable_opts.sanitizer_recover,
SanitizerSet::empty()
),
sanitizer_memory_track_origins: if_regular!(
sess.opts.unstable_opts.sanitizer_memory_track_origins,
0
),
emit_pre_lto_bc: if_regular!(
save_temps || need_pre_lto_bitcode_for_incr_comp(sess),
false
),
emit_no_opt_bc: if_regular!(save_temps, false),
emit_bc: if_regular!(
save_temps || sess.opts.output_types.contains_key(&OutputType::Bitcode),
save_temps
),
emit_ir: if_regular!(
sess.opts.output_types.contains_key(&OutputType::LlvmAssembly),
false
),
emit_asm: if_regular!(
sess.opts.output_types.contains_key(&OutputType::Assembly),
false
),
emit_obj,
emit_thin_lto: sess.opts.unstable_opts.emit_thin_lto,
emit_thin_lto_summary: if_regular!(
sess.opts.output_types.contains_key(&OutputType::ThinLinkBitcode),
false
),
bc_cmdline: sess.target.bitcode_llvm_cmdline.to_string(),
verify_llvm_ir: sess.verify_llvm_ir(),
lint_llvm_ir: sess.opts.unstable_opts.lint_llvm_ir,
no_prepopulate_passes: sess.opts.cg.no_prepopulate_passes,
no_builtins: no_builtins || sess.target.no_builtins,
// Exclude metadata and allocator modules from time_passes output,
// since they throw off the "LLVM passes" measurement.
time_module: if_regular!(true, false),
// Copy what clang does by turning on loop vectorization at O2 and
// slp vectorization at O3.
vectorize_loop: !sess.opts.cg.no_vectorize_loops
&& (sess.opts.optimize == config::OptLevel::Default
|| sess.opts.optimize == config::OptLevel::Aggressive),
vectorize_slp: !sess.opts.cg.no_vectorize_slp
&& sess.opts.optimize == config::OptLevel::Aggressive,
// Some targets (namely, NVPTX) interact badly with the
// MergeFunctions pass. This is because MergeFunctions can generate
// new function calls which may interfere with the target calling
// convention; e.g. for the NVPTX target, PTX kernels should not
// call other PTX kernels. MergeFunctions can also be configured to
// generate aliases instead, but aliases are not supported by some
// backends (again, NVPTX). Therefore, allow targets to opt out of
// the MergeFunctions pass, but otherwise keep the pass enabled (at
// O2 and O3) since it can be useful for reducing code size.
merge_functions: match sess
.opts
.unstable_opts
.merge_functions
.unwrap_or(sess.target.merge_functions)
{
MergeFunctions::Disabled => false,
MergeFunctions::Trampolines | MergeFunctions::Aliases => {
use config::OptLevel::*;
match sess.opts.optimize {
Aggressive | Default | SizeMin | Size => true,
Less | No => false,
}
}
},
emit_lifetime_markers: sess.emit_lifetime_markers(),
llvm_plugins: if_regular!(sess.opts.unstable_opts.llvm_plugins.clone(), vec![]),
}
}
pub fn bitcode_needed(&self) -> bool {
self.emit_bc
|| self.emit_thin_lto_summary
|| self.emit_obj == EmitObj::Bitcode
|| self.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full)
}
}
/// Configuration passed to the function returned by the `target_machine_factory`.
pub struct TargetMachineFactoryConfig {
/// Split DWARF is enabled in LLVM by checking that `TM.MCOptions.SplitDwarfFile` isn't empty,
/// so the path to the dwarf object has to be provided when we create the target machine.
/// This can be ignored by backends which do not need it for their Split DWARF support.
pub split_dwarf_file: Option<PathBuf>,
/// The name of the output object file. Used for setting OutputFilenames in target options
/// so that LLVM can emit the CodeView S_OBJNAME record in pdb files
pub output_obj_file: Option<PathBuf>,
}
impl TargetMachineFactoryConfig {
pub fn new(
cgcx: &CodegenContext<impl WriteBackendMethods>,
module_name: &str,
) -> TargetMachineFactoryConfig {
let split_dwarf_file = if cgcx.target_can_use_split_dwarf {
cgcx.output_filenames.split_dwarf_path(
cgcx.split_debuginfo,
cgcx.split_dwarf_kind,
Some(module_name),
)
} else {
None
};
let output_obj_file =
Some(cgcx.output_filenames.temp_path(OutputType::Object, Some(module_name)));
TargetMachineFactoryConfig { split_dwarf_file, output_obj_file }
}
}
pub type TargetMachineFactoryFn<B> = Arc<
dyn Fn(
TargetMachineFactoryConfig,
) -> Result<
<B as WriteBackendMethods>::TargetMachine,
<B as WriteBackendMethods>::TargetMachineError,
> + Send
+ Sync,
>;
type ExportedSymbols = FxHashMap<CrateNum, Arc<Vec<(String, SymbolExportInfo)>>>;
/// Additional resources used by optimize_and_codegen (not module specific)
#[derive(Clone)]
pub struct CodegenContext<B: WriteBackendMethods> {
// Resources needed when running LTO
pub prof: SelfProfilerRef,
pub lto: Lto,
pub save_temps: bool,
pub fewer_names: bool,
pub time_trace: bool,
pub exported_symbols: Option<Arc<ExportedSymbols>>,
pub opts: Arc<config::Options>,
pub crate_types: Vec<CrateType>,
pub each_linked_rlib_for_lto: Vec<(CrateNum, PathBuf)>,
pub output_filenames: Arc<OutputFilenames>,
pub regular_module_config: Arc<ModuleConfig>,
pub metadata_module_config: Arc<ModuleConfig>,
pub allocator_module_config: Arc<ModuleConfig>,
pub tm_factory: TargetMachineFactoryFn<B>,
pub msvc_imps_needed: bool,
pub is_pe_coff: bool,
pub target_can_use_split_dwarf: bool,
pub target_arch: String,
pub split_debuginfo: rustc_target::spec::SplitDebuginfo,
pub split_dwarf_kind: rustc_session::config::SplitDwarfKind,
/// All commandline args used to invoke the compiler, with @file args fully expanded.
/// This will only be used within debug info, e.g. in the pdb file on windows
/// This is mainly useful for other tools that reads that debuginfo to figure out
/// how to call the compiler with the same arguments.
pub expanded_args: Vec<String>,
/// Emitter to use for diagnostics produced during codegen.
pub diag_emitter: SharedEmitter,
/// LLVM optimizations for which we want to print remarks.
pub remark: Passes,
/// Directory into which should the LLVM optimization remarks be written.
/// If `None`, they will be written to stderr.
pub remark_dir: Option<PathBuf>,
/// The incremental compilation session directory, or None if we are not
/// compiling incrementally
pub incr_comp_session_dir: Option<PathBuf>,
/// Channel back to the main control thread to send messages to
pub coordinator_send: Sender<Box<dyn Any + Send>>,
/// `true` if the codegen should be run in parallel.
///
/// Depends on [`CodegenBackend::supports_parallel()`] and `-Zno_parallel_backend`.
pub parallel: bool,
}
impl<B: WriteBackendMethods> CodegenContext<B> {
pub fn create_dcx(&self) -> DiagCtxt {
DiagCtxt::new(Box::new(self.diag_emitter.clone()))
}
pub fn config(&self, kind: ModuleKind) -> &ModuleConfig {
match kind {
ModuleKind::Regular => &self.regular_module_config,
ModuleKind::Metadata => &self.metadata_module_config,
ModuleKind::Allocator => &self.allocator_module_config,
}
}
}
fn generate_lto_work<B: ExtraBackendMethods>(
cgcx: &CodegenContext<B>,
needs_fat_lto: Vec<FatLtoInput<B>>,
needs_thin_lto: Vec<(String, B::ThinBuffer)>,
import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>,
) -> Vec<(WorkItem<B>, u64)> {
let _prof_timer = cgcx.prof.generic_activity("codegen_generate_lto_work");
if !needs_fat_lto.is_empty() {
assert!(needs_thin_lto.is_empty());
let module =
B::run_fat_lto(cgcx, needs_fat_lto, import_only_modules).unwrap_or_else(|e| e.raise());
// We are adding a single work item, so the cost doesn't matter.
vec![(WorkItem::LTO(module), 0)]
} else {
assert!(needs_fat_lto.is_empty());
let (lto_modules, copy_jobs) = B::run_thin_lto(cgcx, needs_thin_lto, import_only_modules)
.unwrap_or_else(|e| e.raise());
lto_modules
.into_iter()
.map(|module| {
let cost = module.cost();
(WorkItem::LTO(module), cost)
})
.chain(copy_jobs.into_iter().map(|wp| {
(
WorkItem::CopyPostLtoArtifacts(CachedModuleCodegen {
name: wp.cgu_name.clone(),
source: wp,
}),
0, // copying is very cheap
)
}))
.collect()
}
}
struct CompiledModules {
modules: Vec<CompiledModule>,
allocator_module: Option<CompiledModule>,
}
fn need_bitcode_in_object(tcx: TyCtxt<'_>) -> bool {
let sess = tcx.sess;
let requested_for_rlib = sess.opts.cg.embed_bitcode
&& tcx.crate_types().contains(&CrateType::Rlib)
&& sess.opts.output_types.contains_key(&OutputType::Exe);
let forced_by_target = sess.target.forces_embed_bitcode;
requested_for_rlib || forced_by_target
}
fn need_pre_lto_bitcode_for_incr_comp(sess: &Session) -> bool {
if sess.opts.incremental.is_none() {
return false;
}
match sess.lto() {
Lto::No => false,
Lto::Fat | Lto::Thin | Lto::ThinLocal => true,
}
}
pub(crate) fn start_async_codegen<B: ExtraBackendMethods>(
backend: B,
tcx: TyCtxt<'_>,
target_cpu: String,
metadata: EncodedMetadata,
metadata_module: Option<CompiledModule>,
) -> OngoingCodegen<B> {
let (coordinator_send, coordinator_receive) = channel();
let sess = tcx.sess;
let crate_attrs = tcx.hir().attrs(rustc_hir::CRATE_HIR_ID);
let no_builtins = attr::contains_name(crate_attrs, sym::no_builtins);
let is_compiler_builtins = attr::contains_name(crate_attrs, sym::compiler_builtins);
let crate_info = CrateInfo::new(tcx, target_cpu);
let regular_config =
ModuleConfig::new(ModuleKind::Regular, tcx, no_builtins, is_compiler_builtins);
let metadata_config =
ModuleConfig::new(ModuleKind::Metadata, tcx, no_builtins, is_compiler_builtins);
let allocator_config =
ModuleConfig::new(ModuleKind::Allocator, tcx, no_builtins, is_compiler_builtins);
let (shared_emitter, shared_emitter_main) = SharedEmitter::new();
let (codegen_worker_send, codegen_worker_receive) = channel();
let coordinator_thread = start_executing_work(
backend.clone(),
tcx,
&crate_info,
shared_emitter,
codegen_worker_send,
coordinator_receive,
sess.jobserver.clone(),
Arc::new(regular_config),
Arc::new(metadata_config),
Arc::new(allocator_config),
coordinator_send.clone(),
);
OngoingCodegen {
backend,
metadata,
metadata_module,
crate_info,
codegen_worker_receive,
shared_emitter_main,
coordinator: Coordinator {
sender: coordinator_send,
future: Some(coordinator_thread),
phantom: PhantomData,
},
output_filenames: tcx.output_filenames(()).clone(),
}
}
fn copy_all_cgu_workproducts_to_incr_comp_cache_dir(
sess: &Session,
compiled_modules: &CompiledModules,
) -> FxIndexMap<WorkProductId, WorkProduct> {
let mut work_products = FxIndexMap::default();
if sess.opts.incremental.is_none() {
return work_products;
}
let _timer = sess.timer("copy_all_cgu_workproducts_to_incr_comp_cache_dir");
for module in compiled_modules.modules.iter().filter(|m| m.kind == ModuleKind::Regular) {
let mut files = Vec::new();
if let Some(object_file_path) = &module.object {
files.push((OutputType::Object.extension(), object_file_path.as_path()));
}
if let Some(dwarf_object_file_path) = &module.dwarf_object {
files.push(("dwo", dwarf_object_file_path.as_path()));
}
if let Some(path) = &module.assembly {
files.push((OutputType::Assembly.extension(), path.as_path()));
}
if let Some(path) = &module.llvm_ir {
files.push((OutputType::LlvmAssembly.extension(), path.as_path()));
}
if let Some(path) = &module.bytecode {
files.push((OutputType::Bitcode.extension(), path.as_path()));
}
if let Some((id, product)) =
copy_cgu_workproduct_to_incr_comp_cache_dir(sess, &module.name, files.as_slice())
{
work_products.insert(id, product);
}
}
work_products
}
fn produce_final_output_artifacts(
sess: &Session,
compiled_modules: &CompiledModules,
crate_output: &OutputFilenames,
) {
let mut user_wants_bitcode = false;
let mut user_wants_objects = false;
// Produce final compile outputs.
let copy_gracefully = |from: &Path, to: &OutFileName| match to {
OutFileName::Stdout => {
if let Err(e) = copy_to_stdout(from) {
sess.dcx().emit_err(errors::CopyPath::new(from, to.as_path(), e));
}
}
OutFileName::Real(path) => {
if let Err(e) = fs::copy(from, path) {
sess.dcx().emit_err(errors::CopyPath::new(from, path, e));
}
}
};
let copy_if_one_unit = |output_type: OutputType, keep_numbered: bool| {
if compiled_modules.modules.len() == 1 {
// 1) Only one codegen unit. In this case it's no difficulty
// to copy `foo.0.x` to `foo.x`.
let module_name = Some(&compiled_modules.modules[0].name[..]);
let path = crate_output.temp_path(output_type, module_name);
let output = crate_output.path(output_type);
if !output_type.is_text_output() && output.is_tty() {
sess.dcx()
.emit_err(errors::BinaryOutputToTty { shorthand: output_type.shorthand() });
} else {
copy_gracefully(&path, &output);
}
if !sess.opts.cg.save_temps && !keep_numbered {
// The user just wants `foo.x`, not `foo.#module-name#.x`.
ensure_removed(sess.dcx(), &path);
}
} else {
let extension = crate_output
.temp_path(output_type, None)
.extension()
.unwrap()
.to_str()
.unwrap()
.to_owned();
if crate_output.outputs.contains_explicit_name(&output_type) {
// 2) Multiple codegen units, with `--emit foo=some_name`. We have
// no good solution for this case, so warn the user.
sess.dcx().emit_warn(errors::IgnoringEmitPath { extension });
} else if crate_output.single_output_file.is_some() {
// 3) Multiple codegen units, with `-o some_name`. We have
// no good solution for this case, so warn the user.
sess.dcx().emit_warn(errors::IgnoringOutput { extension });
} else {
// 4) Multiple codegen units, but no explicit name. We
// just leave the `foo.0.x` files in place.
// (We don't have to do any work in this case.)
}
}
};
// Flag to indicate whether the user explicitly requested bitcode.
// Otherwise, we produced it only as a temporary output, and will need
// to get rid of it.
for output_type in crate_output.outputs.keys() {
match *output_type {
OutputType::Bitcode => {
user_wants_bitcode = true;
// Copy to .bc, but always keep the .0.bc. There is a later
// check to figure out if we should delete .0.bc files, or keep
// them for making an rlib.
copy_if_one_unit(OutputType::Bitcode, true);
}
OutputType::ThinLinkBitcode => {
copy_if_one_unit(OutputType::ThinLinkBitcode, false);
}
OutputType::LlvmAssembly => {
copy_if_one_unit(OutputType::LlvmAssembly, false);
}
OutputType::Assembly => {
copy_if_one_unit(OutputType::Assembly, false);
}
OutputType::Object => {
user_wants_objects = true;
copy_if_one_unit(OutputType::Object, true);
}
OutputType::Mir | OutputType::Metadata | OutputType::Exe | OutputType::DepInfo => {}
}
}
// Clean up unwanted temporary files.
// We create the following files by default:
// - #crate#.#module-name#.bc
// - #crate#.#module-name#.o
// - #crate#.crate.metadata.bc
// - #crate#.crate.metadata.o
// - #crate#.o (linked from crate.##.o)
// - #crate#.bc (copied from crate.##.bc)
// We may create additional files if requested by the user (through
// `-C save-temps` or `--emit=` flags).
if !sess.opts.cg.save_temps {
// Remove the temporary .#module-name#.o objects. If the user didn't
// explicitly request bitcode (with --emit=bc), and the bitcode is not
// needed for building an rlib, then we must remove .#module-name#.bc as
// well.
// Specific rules for keeping .#module-name#.bc:
// - If the user requested bitcode (`user_wants_bitcode`), and
// codegen_units > 1, then keep it.
// - If the user requested bitcode but codegen_units == 1, then we
// can toss .#module-name#.bc because we copied it to .bc earlier.
// - If we're not building an rlib and the user didn't request
// bitcode, then delete .#module-name#.bc.
// If you change how this works, also update back::link::link_rlib,
// where .#module-name#.bc files are (maybe) deleted after making an
// rlib.
let needs_crate_object = crate_output.outputs.contains_key(&OutputType::Exe);
let keep_numbered_bitcode = user_wants_bitcode && sess.codegen_units().as_usize() > 1;
let keep_numbered_objects =
needs_crate_object || (user_wants_objects && sess.codegen_units().as_usize() > 1);
for module in compiled_modules.modules.iter() {
if let Some(ref path) = module.object {
if !keep_numbered_objects {
ensure_removed(sess.dcx(), path);
}
}
if let Some(ref path) = module.dwarf_object {
if !keep_numbered_objects {
ensure_removed(sess.dcx(), path);
}
}
if let Some(ref path) = module.bytecode {
if !keep_numbered_bitcode {
ensure_removed(sess.dcx(), path);
}
}
}
if !user_wants_bitcode {
if let Some(ref allocator_module) = compiled_modules.allocator_module {
if let Some(ref path) = allocator_module.bytecode {
ensure_removed(sess.dcx(), path);
}
}
}
}
if sess.opts.json_artifact_notifications {
if compiled_modules.modules.len() == 1 {
compiled_modules.modules[0].for_each_output(|_path, ty| {
if sess.opts.output_types.contains_key(&ty) {
let descr = ty.shorthand();
// for single cgu file is renamed to drop cgu specific suffix
// so we regenerate it the same way
let path = crate_output.path(ty);
sess.dcx().emit_artifact_notification(path.as_path(), descr);
}
});
} else {
for module in &compiled_modules.modules {
module.for_each_output(|path, ty| {
if sess.opts.output_types.contains_key(&ty) {
let descr = ty.shorthand();
sess.dcx().emit_artifact_notification(&path, descr);
}
});
}
}
}
// We leave the following files around by default:
// - #crate#.o
// - #crate#.crate.metadata.o
// - #crate#.bc
// These are used in linking steps and will be cleaned up afterward.
}
pub(crate) enum WorkItem<B: WriteBackendMethods> {
/// Optimize a newly codegened, totally unoptimized module.
Optimize(ModuleCodegen<B::Module>),
/// Copy the post-LTO artifacts from the incremental cache to the output
/// directory.
CopyPostLtoArtifacts(CachedModuleCodegen),
/// Performs (Thin)LTO on the given module.
LTO(lto::LtoModuleCodegen<B>),
}
impl<B: WriteBackendMethods> WorkItem<B> {
fn module_kind(&self) -> ModuleKind {
match *self {
WorkItem::Optimize(ref m) => m.kind,
WorkItem::CopyPostLtoArtifacts(_) | WorkItem::LTO(_) => ModuleKind::Regular,
}
}
/// Generate a short description of this work item suitable for use as a thread name.
fn short_description(&self) -> String {
// `pthread_setname()` on *nix ignores anything beyond the first 15
// bytes. Use short descriptions to maximize the space available for
// the module name.
#[cfg(not(windows))]
fn desc(short: &str, _long: &str, name: &str) -> String {
// The short label is three bytes, and is followed by a space. That
// leaves 11 bytes for the CGU name. How we obtain those 11 bytes
// depends on the CGU name form.
//
// - Non-incremental, e.g. `regex.f10ba03eb5ec7975-cgu.0`: the part
// before the `-cgu.0` is the same for every CGU, so use the
// `cgu.0` part. The number suffix will be different for each
// CGU.
//
// - Incremental (normal), e.g. `2i52vvl2hco29us0`: use the whole
// name because each CGU will have a unique ASCII hash, and the
// first 11 bytes will be enough to identify it.
//
// - Incremental (with `-Zhuman-readable-cgu-names`), e.g.
// `regex.f10ba03eb5ec7975-re_builder.volatile`: use the whole
// name. The first 11 bytes won't be enough to uniquely identify
// it, but no obvious substring will, and this is a rarely used
// option so it doesn't matter much.
//
assert_eq!(short.len(), 3);
let name = if let Some(index) = name.find("-cgu.") {
&name[index + 1..] // +1 skips the leading '-'.
} else {
name
};
format!("{short} {name}")
}
// Windows has no thread name length limit, so use more descriptive names.
#[cfg(windows)]
fn desc(_short: &str, long: &str, name: &str) -> String {
format!("{long} {name}")
}
match self {
WorkItem::Optimize(m) => desc("opt", "optimize module", &m.name),
WorkItem::CopyPostLtoArtifacts(m) => desc("cpy", "copy LTO artifacts for", &m.name),
WorkItem::LTO(m) => desc("lto", "LTO module", m.name()),
}
}
}
/// A result produced by the backend.
pub(crate) enum WorkItemResult<B: WriteBackendMethods> {
/// The backend has finished compiling a CGU, nothing more required.
Finished(CompiledModule),
/// The backend has finished compiling a CGU, which now needs linking
/// because `-Zcombine-cgu` was specified.
NeedsLink(ModuleCodegen<B::Module>),
/// The backend has finished compiling a CGU, which now needs to go through
/// fat LTO.
NeedsFatLto(FatLtoInput<B>),
/// The backend has finished compiling a CGU, which now needs to go through
/// thin LTO.
NeedsThinLto(String, B::ThinBuffer),
}
pub enum FatLtoInput<B: WriteBackendMethods> {
Serialized { name: String, buffer: B::ModuleBuffer },
InMemory(ModuleCodegen<B::Module>),
}
/// Actual LTO type we end up choosing based on multiple factors.
pub(crate) enum ComputedLtoType {
No,
Thin,
Fat,
}
pub(crate) fn compute_per_cgu_lto_type(
sess_lto: &Lto,
opts: &config::Options,
sess_crate_types: &[CrateType],
module_kind: ModuleKind,
) -> ComputedLtoType {
// Metadata modules never participate in LTO regardless of the lto
// settings.
if module_kind == ModuleKind::Metadata {
return ComputedLtoType::No;
}
// If the linker does LTO, we don't have to do it. Note that we
// keep doing full LTO, if it is requested, as not to break the
// assumption that the output will be a single module.
let linker_does_lto = opts.cg.linker_plugin_lto.enabled();
// When we're automatically doing ThinLTO for multi-codegen-unit
// builds we don't actually want to LTO the allocator modules if
// it shows up. This is due to various linker shenanigans that
// we'll encounter later.
let is_allocator = module_kind == ModuleKind::Allocator;
// We ignore a request for full crate graph LTO if the crate type
// is only an rlib, as there is no full crate graph to process,
// that'll happen later.
//
// This use case currently comes up primarily for targets that
// require LTO so the request for LTO is always unconditionally
// passed down to the backend, but we don't actually want to do
// anything about it yet until we've got a final product.
let is_rlib = sess_crate_types.len() == 1 && sess_crate_types[0] == CrateType::Rlib;
match sess_lto {
Lto::ThinLocal if !linker_does_lto && !is_allocator => ComputedLtoType::Thin,
Lto::Thin if !linker_does_lto && !is_rlib => ComputedLtoType::Thin,
Lto::Fat if !is_rlib => ComputedLtoType::Fat,
_ => ComputedLtoType::No,
}
}
fn execute_optimize_work_item<B: ExtraBackendMethods>(
cgcx: &CodegenContext<B>,
module: ModuleCodegen<B::Module>,
module_config: &ModuleConfig,
) -> Result<WorkItemResult<B>, FatalError> {
let dcx = cgcx.create_dcx();
let dcx = dcx.handle();
unsafe {
B::optimize(cgcx, dcx, &module, module_config)?;
}
// After we've done the initial round of optimizations we need to
// decide whether to synchronously codegen this module or ship it
// back to the coordinator thread for further LTO processing (which
// has to wait for all the initial modules to be optimized).
let lto_type = compute_per_cgu_lto_type(&cgcx.lto, &cgcx.opts, &cgcx.crate_types, module.kind);
// If we're doing some form of incremental LTO then we need to be sure to
// save our module to disk first.
let bitcode = if cgcx.config(module.kind).emit_pre_lto_bc {
let filename = pre_lto_bitcode_filename(&module.name);
cgcx.incr_comp_session_dir.as_ref().map(|path| path.join(&filename))
} else {
None
};
match lto_type {
ComputedLtoType::No => finish_intra_module_work(cgcx, module, module_config),
ComputedLtoType::Thin => {
let (name, thin_buffer) = B::prepare_thin(module, false);
if let Some(path) = bitcode {
fs::write(&path, thin_buffer.data()).unwrap_or_else(|e| {
panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e);
});
}
Ok(WorkItemResult::NeedsThinLto(name, thin_buffer))
}
ComputedLtoType::Fat => match bitcode {
Some(path) => {
let (name, buffer) = B::serialize_module(module);
fs::write(&path, buffer.data()).unwrap_or_else(|e| {
panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e);
});
Ok(WorkItemResult::NeedsFatLto(FatLtoInput::Serialized { name, buffer }))
}
None => Ok(WorkItemResult::NeedsFatLto(FatLtoInput::InMemory(module))),
},
}
}
fn execute_copy_from_cache_work_item<B: ExtraBackendMethods>(
cgcx: &CodegenContext<B>,
module: CachedModuleCodegen,
module_config: &ModuleConfig,
) -> WorkItemResult<B> {
let incr_comp_session_dir = cgcx.incr_comp_session_dir.as_ref().unwrap();
let load_from_incr_comp_dir = |output_path: PathBuf, saved_path: &str| {
let source_file = in_incr_comp_dir(incr_comp_session_dir, saved_path);
debug!(
"copying preexisting module `{}` from {:?} to {}",
module.name,
source_file,
output_path.display()
);
match link_or_copy(&source_file, &output_path) {
Ok(_) => Some(output_path),
Err(error) => {
cgcx.create_dcx().handle().emit_err(errors::CopyPathBuf {
source_file,
output_path,
error,
});
None
}
}
};
let dwarf_object =
module.source.saved_files.get("dwo").as_ref().and_then(|saved_dwarf_object_file| {
let dwarf_obj_out = cgcx
.output_filenames
.split_dwarf_path(cgcx.split_debuginfo, cgcx.split_dwarf_kind, Some(&module.name))
.expect(
"saved dwarf object in work product but `split_dwarf_path` returned `None`",
);
load_from_incr_comp_dir(dwarf_obj_out, saved_dwarf_object_file)
});
let load_from_incr_cache = |perform, output_type: OutputType| {
if perform {
let saved_file = module.source.saved_files.get(output_type.extension())?;
let output_path = cgcx.output_filenames.temp_path(output_type, Some(&module.name));
load_from_incr_comp_dir(output_path, &saved_file)
} else {
None
}
};
let should_emit_obj = module_config.emit_obj != EmitObj::None;
let assembly = load_from_incr_cache(module_config.emit_asm, OutputType::Assembly);
let llvm_ir = load_from_incr_cache(module_config.emit_ir, OutputType::LlvmAssembly);
let bytecode = load_from_incr_cache(module_config.emit_bc, OutputType::Bitcode);
let object = load_from_incr_cache(should_emit_obj, OutputType::Object);
if should_emit_obj && object.is_none() {
cgcx.create_dcx().handle().emit_fatal(errors::NoSavedObjectFile { cgu_name: &module.name })
}
WorkItemResult::Finished(CompiledModule {
name: module.name,
kind: ModuleKind::Regular,
object,
dwarf_object,
bytecode,
assembly,
llvm_ir,
})
}
fn execute_lto_work_item<B: ExtraBackendMethods>(
cgcx: &CodegenContext<B>,
module: lto::LtoModuleCodegen<B>,
module_config: &ModuleConfig,
) -> Result<WorkItemResult<B>, FatalError> {
let module = unsafe { module.optimize(cgcx)? };
finish_intra_module_work(cgcx, module, module_config)
}
fn finish_intra_module_work<B: ExtraBackendMethods>(
cgcx: &CodegenContext<B>,
module: ModuleCodegen<B::Module>,
module_config: &ModuleConfig,
) -> Result<WorkItemResult<B>, FatalError> {
let dcx = cgcx.create_dcx();
let dcx = dcx.handle();
if !cgcx.opts.unstable_opts.combine_cgu
|| module.kind == ModuleKind::Metadata
|| module.kind == ModuleKind::Allocator
{
let module = unsafe { B::codegen(cgcx, dcx, module, module_config)? };
Ok(WorkItemResult::Finished(module))
} else {
Ok(WorkItemResult::NeedsLink(module))
}
}
/// Messages sent to the coordinator.
pub(crate) enum Message<B: WriteBackendMethods> {
/// A jobserver token has become available. Sent from the jobserver helper
/// thread.
Token(io::Result<Acquired>),
/// The backend has finished processing a work item for a codegen unit.
/// Sent from a backend worker thread.
WorkItem { result: Result<WorkItemResult<B>, Option<WorkerFatalError>>, worker_id: usize },
/// The frontend has finished generating something (backend IR or a
/// post-LTO artifact) for a codegen unit, and it should be passed to the
/// backend. Sent from the main thread.
CodegenDone { llvm_work_item: WorkItem<B>, cost: u64 },
/// Similar to `CodegenDone`, but for reusing a pre-LTO artifact
/// Sent from the main thread.
AddImportOnlyModule {
module_data: SerializedModule<B::ModuleBuffer>,
work_product: WorkProduct,
},
/// The frontend has finished generating everything for all codegen units.
/// Sent from the main thread.
CodegenComplete,
/// Some normal-ish compiler error occurred, and codegen should be wound
/// down. Sent from the main thread.
CodegenAborted,
}
/// A message sent from the coordinator thread to the main thread telling it to
/// process another codegen unit.
pub struct CguMessage;
// A cut-down version of `rustc_errors::DiagInner` that impls `Send`, which
// can be used to send diagnostics from codegen threads to the main thread.
// It's missing the following fields from `rustc_errors::DiagInner`.
// - `span`: it doesn't impl `Send`.
// - `suggestions`: it doesn't impl `Send`, and isn't used for codegen
// diagnostics.
// - `sort_span`: it doesn't impl `Send`.
// - `is_lint`: lints aren't relevant during codegen.
// - `emitted_at`: not used for codegen diagnostics.
struct Diagnostic {
level: Level,
messages: Vec<(DiagMessage, Style)>,
code: Option<ErrCode>,
children: Vec<Subdiagnostic>,
args: DiagArgMap,
}
// A cut-down version of `rustc_errors::Subdiag` that impls `Send`. It's
// missing the following fields from `rustc_errors::Subdiag`.
// - `span`: it doesn't impl `Send`.
pub(crate) struct Subdiagnostic {
level: Level,
messages: Vec<(DiagMessage, Style)>,
}
#[derive(PartialEq, Clone, Copy, Debug)]
enum MainThreadState {
/// Doing nothing.
Idle,
/// Doing codegen, i.e. MIR-to-LLVM-IR conversion.
Codegenning,
/// Idle, but lending the compiler process's Token to an LLVM thread so it can do useful work.
Lending,
}
fn start_executing_work<B: ExtraBackendMethods>(
backend: B,
tcx: TyCtxt<'_>,
crate_info: &CrateInfo,
shared_emitter: SharedEmitter,
codegen_worker_send: Sender<CguMessage>,
coordinator_receive: Receiver<Box<dyn Any + Send>>,
jobserver: Client,
regular_config: Arc<ModuleConfig>,
metadata_config: Arc<ModuleConfig>,
allocator_config: Arc<ModuleConfig>,
tx_to_llvm_workers: Sender<Box<dyn Any + Send>>,
) -> thread::JoinHandle<Result<CompiledModules, ()>> {
let coordinator_send = tx_to_llvm_workers;
let sess = tcx.sess;
let mut each_linked_rlib_for_lto = Vec::new();
drop(link::each_linked_rlib(crate_info, None, &mut |cnum, path| {
if link::ignored_for_lto(sess, crate_info, cnum) {
return;
}
each_linked_rlib_for_lto.push((cnum, path.to_path_buf()));
}));
// Compute the set of symbols we need to retain when doing LTO (if we need to)
let exported_symbols = {
let mut exported_symbols = FxHashMap::default();
let copy_symbols = |cnum| {
let symbols = tcx
.exported_symbols(cnum)
.iter()
.map(|&(s, lvl)| (symbol_name_for_instance_in_crate(tcx, s, cnum), lvl))
.collect();
Arc::new(symbols)
};
match sess.lto() {
Lto::No => None,
Lto::ThinLocal => {
exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
Some(Arc::new(exported_symbols))
}
Lto::Fat | Lto::Thin => {
exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
for &(cnum, ref _path) in &each_linked_rlib_for_lto {
exported_symbols.insert(cnum, copy_symbols(cnum));
}
Some(Arc::new(exported_symbols))
}
}
};
// First up, convert our jobserver into a helper thread so we can use normal
// mpsc channels to manage our messages and such.
// After we've requested tokens then we'll, when we can,
// get tokens on `coordinator_receive` which will
// get managed in the main loop below.
let coordinator_send2 = coordinator_send.clone();
let helper = jobserver
.into_helper_thread(move |token| {
drop(coordinator_send2.send(Box::new(Message::Token::<B>(token))));
})
.expect("failed to spawn helper thread");
let ol =
if tcx.sess.opts.unstable_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() {
// If we know that we won’t be doing codegen, create target machines without optimisation.
config::OptLevel::No
} else {
tcx.backend_optimization_level(())
};
let backend_features = tcx.global_backend_features(());
let remark_dir = if let Some(ref dir) = sess.opts.unstable_opts.remark_dir {
let result = fs::create_dir_all(dir).and_then(|_| dir.canonicalize());
match result {
Ok(dir) => Some(dir),
Err(error) => sess.dcx().emit_fatal(ErrorCreatingRemarkDir { error }),
}
} else {
None
};
let cgcx = CodegenContext::<B> {
crate_types: tcx.crate_types().to_vec(),
each_linked_rlib_for_lto,
lto: sess.lto(),
fewer_names: sess.fewer_names(),
save_temps: sess.opts.cg.save_temps,
time_trace: sess.opts.unstable_opts.llvm_time_trace,
opts: Arc::new(sess.opts.clone()),
prof: sess.prof.clone(),
exported_symbols,
remark: sess.opts.cg.remark.clone(),
remark_dir,
incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone()),
coordinator_send,
expanded_args: tcx.sess.expanded_args.clone(),
diag_emitter: shared_emitter.clone(),
output_filenames: tcx.output_filenames(()).clone(),
regular_module_config: regular_config,
metadata_module_config: metadata_config,
allocator_module_config: allocator_config,
tm_factory: backend.target_machine_factory(tcx.sess, ol, backend_features),
msvc_imps_needed: msvc_imps_needed(tcx),
is_pe_coff: tcx.sess.target.is_like_windows,
target_can_use_split_dwarf: tcx.sess.target_can_use_split_dwarf(),
target_arch: tcx.sess.target.arch.to_string(),
split_debuginfo: tcx.sess.split_debuginfo(),
split_dwarf_kind: tcx.sess.opts.unstable_opts.split_dwarf_kind,
parallel: backend.supports_parallel() && !sess.opts.unstable_opts.no_parallel_backend,
};
// This is the "main loop" of parallel work happening for parallel codegen.
// It's here that we manage parallelism, schedule work, and work with
// messages coming from clients.
//
// There are a few environmental pre-conditions that shape how the system
// is set up:
//
// - Error reporting can only happen on the main thread because that's the
// only place where we have access to the compiler `Session`.
// - LLVM work can be done on any thread.
// - Codegen can only happen on the main thread.
// - Each thread doing substantial work must be in possession of a `Token`
// from the `Jobserver`.
// - The compiler process always holds one `Token`. Any additional `Tokens`
// have to be requested from the `Jobserver`.
//
// Error Reporting
// ===============
// The error reporting restriction is handled separately from the rest: We
// set up a `SharedEmitter` that holds an open channel to the main thread.
// When an error occurs on any thread, the shared emitter will send the
// error message to the receiver main thread (`SharedEmitterMain`). The
// main thread will periodically query this error message queue and emit
// any error messages it has received. It might even abort compilation if
// it has received a fatal error. In this case we rely on all other threads
// being torn down automatically with the main thread.
// Since the main thread will often be busy doing codegen work, error
// reporting will be somewhat delayed, since the message queue can only be
// checked in between two work packages.
//
// Work Processing Infrastructure
// ==============================
// The work processing infrastructure knows three major actors:
//
// - the coordinator thread,
// - the main thread, and
// - LLVM worker threads
//
// The coordinator thread is running a message loop. It instructs the main
// thread about what work to do when, and it will spawn off LLVM worker
// threads as open LLVM WorkItems become available.
//
// The job of the main thread is to codegen CGUs into LLVM work packages
// (since the main thread is the only thread that can do this). The main
// thread will block until it receives a message from the coordinator, upon
// which it will codegen one CGU, send it to the coordinator and block
// again. This way the coordinator can control what the main thread is
// doing.
//
// The coordinator keeps a queue of LLVM WorkItems, and when a `Token` is
// available, it will spawn off a new LLVM worker thread and let it process
// a WorkItem. When a LLVM worker thread is done with its WorkItem,
// it will just shut down, which also frees all resources associated with
// the given LLVM module, and sends a message to the coordinator that the
// WorkItem has been completed.
//
// Work Scheduling
// ===============
// The scheduler's goal is to minimize the time it takes to complete all
// work there is, however, we also want to keep memory consumption low
// if possible. These two goals are at odds with each other: If memory
// consumption were not an issue, we could just let the main thread produce
// LLVM WorkItems at full speed, assuring maximal utilization of
// Tokens/LLVM worker threads. However, since codegen is usually faster
// than LLVM processing, the queue of LLVM WorkItems would fill up and each
// WorkItem potentially holds on to a substantial amount of memory.
//
// So the actual goal is to always produce just enough LLVM WorkItems as
// not to starve our LLVM worker threads. That means, once we have enough
// WorkItems in our queue, we can block the main thread, so it does not
// produce more until we need them.
//
// Doing LLVM Work on the Main Thread
// ----------------------------------
// Since the main thread owns the compiler process's implicit `Token`, it is
// wasteful to keep it blocked without doing any work. Therefore, what we do
// in this case is: We spawn off an additional LLVM worker thread that helps
// reduce the queue. The work it is doing corresponds to the implicit
// `Token`. The coordinator will mark the main thread as being busy with
// LLVM work. (The actual work happens on another OS thread but we just care
// about `Tokens`, not actual threads).
//
// When any LLVM worker thread finishes while the main thread is marked as
// "busy with LLVM work", we can do a little switcheroo: We give the Token
// of the just finished thread to the LLVM worker thread that is working on
// behalf of the main thread's implicit Token, thus freeing up the main
// thread again. The coordinator can then again decide what the main thread
// should do. This allows the coordinator to make decisions at more points
// in time.
//
// Striking a Balance between Throughput and Memory Consumption
// ------------------------------------------------------------
// Since our two goals, (1) use as many Tokens as possible and (2) keep
// memory consumption as low as possible, are in conflict with each other,
// we have to find a trade off between them. Right now, the goal is to keep
// all workers busy, which means that no worker should find the queue empty
// when it is ready to start.
// How do we do achieve this? Good question :) We actually never know how
// many `Tokens` are potentially available so it's hard to say how much to
// fill up the queue before switching the main thread to LLVM work. Also we
// currently don't have a means to estimate how long a running LLVM worker
// will still be busy with it's current WorkItem. However, we know the
// maximal count of available Tokens that makes sense (=the number of CPU
// cores), so we can take a conservative guess. The heuristic we use here
// is implemented in the `queue_full_enough()` function.
//
// Some Background on Jobservers
// -----------------------------
// It's worth also touching on the management of parallelism here. We don't
// want to just spawn a thread per work item because while that's optimal
// parallelism it may overload a system with too many threads or violate our
// configuration for the maximum amount of cpu to use for this process. To
// manage this we use the `jobserver` crate.
//
// Job servers are an artifact of GNU make and are used to manage
// parallelism between processes. A jobserver is a glorified IPC semaphore
// basically. Whenever we want to run some work we acquire the semaphore,
// and whenever we're done with that work we release the semaphore. In this
// manner we can ensure that the maximum number of parallel workers is
// capped at any one point in time.
//
// LTO and the coordinator thread
// ------------------------------
//
// The final job the coordinator thread is responsible for is managing LTO
// and how that works. When LTO is requested what we'll do is collect all
// optimized LLVM modules into a local vector on the coordinator. Once all
// modules have been codegened and optimized we hand this to the `lto`
// module for further optimization. The `lto` module will return back a list
// of more modules to work on, which the coordinator will continue to spawn
// work for.
//
// Each LLVM module is automatically sent back to the coordinator for LTO if
// necessary. There's already optimizations in place to avoid sending work
// back to the coordinator if LTO isn't requested.
return B::spawn_named_thread(cgcx.time_trace, "coordinator".to_string(), move || {
let mut worker_id_counter = 0;
let mut free_worker_ids = Vec::new();
let mut get_worker_id = |free_worker_ids: &mut Vec<usize>| {
if let Some(id) = free_worker_ids.pop() {
id
} else {
let id = worker_id_counter;
worker_id_counter += 1;
id
}
};
// This is where we collect codegen units that have gone all the way
// through codegen and LLVM.
let mut compiled_modules = vec![];
let mut compiled_allocator_module = None;
let mut needs_link = Vec::new();
let mut needs_fat_lto = Vec::new();
let mut needs_thin_lto = Vec::new();
let mut lto_import_only_modules = Vec::new();
let mut started_lto = false;
/// Possible state transitions:
/// - Ongoing -> Completed
/// - Ongoing -> Aborted
/// - Completed -> Aborted
#[derive(Debug, PartialEq)]
enum CodegenState {
Ongoing,
Completed,
Aborted,
}
use CodegenState::*;
let mut codegen_state = Ongoing;
// This is the queue of LLVM work items that still need processing.
let mut work_items = Vec::<(WorkItem<B>, u64)>::new();
// This are the Jobserver Tokens we currently hold. Does not include
// the implicit Token the compiler process owns no matter what.
let mut tokens = Vec::new();
let mut main_thread_state = MainThreadState::Idle;
// How many LLVM worker threads are running while holding a Token. This
// *excludes* any that the main thread is lending a Token to.
let mut running_with_own_token = 0;
// How many LLVM worker threads are running in total. This *includes*
// any that the main thread is lending a Token to.
let running_with_any_token = |main_thread_state, running_with_own_token| {
running_with_own_token
+ if main_thread_state == MainThreadState::Lending { 1 } else { 0 }
};
let mut llvm_start_time: Option<VerboseTimingGuard<'_>> = None;
// Run the message loop while there's still anything that needs message
// processing. Note that as soon as codegen is aborted we simply want to
// wait for all existing work to finish, so many of the conditions here
// only apply if codegen hasn't been aborted as they represent pending
// work to be done.
loop {
// While there are still CGUs to be codegened, the coordinator has
// to decide how to utilize the compiler processes implicit Token:
// For codegenning more CGU or for running them through LLVM.
if codegen_state == Ongoing {
if main_thread_state == MainThreadState::Idle {
// Compute the number of workers that will be running once we've taken as many
// items from the work queue as we can, plus one for the main thread. It's not
// critically important that we use this instead of just
// `running_with_own_token`, but it prevents the `queue_full_enough` heuristic
// from fluctuating just because a worker finished up and we decreased the
// `running_with_own_token` count, even though we're just going to increase it
// right after this when we put a new worker to work.
let extra_tokens = tokens.len().checked_sub(running_with_own_token).unwrap();
let additional_running = std::cmp::min(extra_tokens, work_items.len());
let anticipated_running = running_with_own_token + additional_running + 1;
if !queue_full_enough(work_items.len(), anticipated_running) {
// The queue is not full enough, process more codegen units:
if codegen_worker_send.send(CguMessage).is_err() {
panic!("Could not send CguMessage to main thread")
}
main_thread_state = MainThreadState::Codegenning;
} else {
// The queue is full enough to not let the worker
// threads starve. Use the implicit Token to do some
// LLVM work too.
let (item, _) =
work_items.pop().expect("queue empty - queue_full_enough() broken?");
main_thread_state = MainThreadState::Lending;
spawn_work(
&cgcx,
&mut llvm_start_time,
get_worker_id(&mut free_worker_ids),
item,
);
}
}
} else if codegen_state == Completed {
if running_with_any_token(main_thread_state, running_with_own_token) == 0
&& work_items.is_empty()
{
// All codegen work is done. Do we have LTO work to do?
if needs_fat_lto.is_empty()
&& needs_thin_lto.is_empty()
&& lto_import_only_modules.is_empty()
{
// Nothing more to do!
break;
}
// We have LTO work to do. Perform the serial work here of
// figuring out what we're going to LTO and then push a
// bunch of work items onto our queue to do LTO. This all
// happens on the coordinator thread but it's very quick so
// we don't worry about tokens.
assert!(!started_lto);
started_lto = true;
let needs_fat_lto = mem::take(&mut needs_fat_lto);
let needs_thin_lto = mem::take(&mut needs_thin_lto);
let import_only_modules = mem::take(&mut lto_import_only_modules);
for (work, cost) in
generate_lto_work(&cgcx, needs_fat_lto, needs_thin_lto, import_only_modules)
{
let insertion_index = work_items
.binary_search_by_key(&cost, |&(_, cost)| cost)
.unwrap_or_else(|e| e);
work_items.insert(insertion_index, (work, cost));
if cgcx.parallel {
helper.request_token();
}
}
}
// In this branch, we know that everything has been codegened,
// so it's just a matter of determining whether the implicit
// Token is free to use for LLVM work.
match main_thread_state {
MainThreadState::Idle => {
if let Some((item, _)) = work_items.pop() {
main_thread_state = MainThreadState::Lending;
spawn_work(
&cgcx,
&mut llvm_start_time,
get_worker_id(&mut free_worker_ids),
item,
);
} else {
// There is no unstarted work, so let the main thread
// take over for a running worker. Otherwise the
// implicit token would just go to waste.
// We reduce the `running` counter by one. The
// `tokens.truncate()` below will take care of
// giving the Token back.
assert!(running_with_own_token > 0);
running_with_own_token -= 1;
main_thread_state = MainThreadState::Lending;
}
}
MainThreadState::Codegenning => bug!(
"codegen worker should not be codegenning after \
codegen was already completed"
),
MainThreadState::Lending => {
// Already making good use of that token
}
}
} else {
// Don't queue up any more work if codegen was aborted, we're
// just waiting for our existing children to finish.
assert!(codegen_state == Aborted);
if running_with_any_token(main_thread_state, running_with_own_token) == 0 {
break;
}
}
// Spin up what work we can, only doing this while we've got available
// parallelism slots and work left to spawn.
if codegen_state != Aborted {
while !work_items.is_empty() && running_with_own_token < tokens.len() {
let (item, _) = work_items.pop().unwrap();
spawn_work(
&cgcx,
&mut llvm_start_time,
get_worker_id(&mut free_worker_ids),
item,
);
running_with_own_token += 1;
}
}
// Relinquish accidentally acquired extra tokens.
tokens.truncate(running_with_own_token);
// If a thread exits successfully then we drop a token associated
// with that worker and update our `running_with_own_token` count.
// We may later re-acquire a token to continue running more work.
// We may also not actually drop a token here if the worker was
// running with an "ephemeral token".
let mut free_worker = |worker_id| {
if main_thread_state == MainThreadState::Lending {
main_thread_state = MainThreadState::Idle;
} else {
running_with_own_token -= 1;
}
free_worker_ids.push(worker_id);
};
let msg = coordinator_receive.recv().unwrap();
match *msg.downcast::<Message<B>>().ok().unwrap() {
// Save the token locally and the next turn of the loop will use
// this to spawn a new unit of work, or it may get dropped
// immediately if we have no more work to spawn.
Message::Token(token) => {
match token {
Ok(token) => {
tokens.push(token);
if main_thread_state == MainThreadState::Lending {
// If the main thread token is used for LLVM work
// at the moment, we turn that thread into a regular
// LLVM worker thread, so the main thread is free
// to react to codegen demand.
main_thread_state = MainThreadState::Idle;
running_with_own_token += 1;
}
}
Err(e) => {
let msg = &format!("failed to acquire jobserver token: {e}");
shared_emitter.fatal(msg);
codegen_state = Aborted;
}
}
}
Message::CodegenDone { llvm_work_item, cost } => {
// We keep the queue sorted by estimated processing cost,
// so that more expensive items are processed earlier. This
// is good for throughput as it gives the main thread more
// time to fill up the queue and it avoids scheduling
// expensive items to the end.
// Note, however, that this is not ideal for memory
// consumption, as LLVM module sizes are not evenly
// distributed.
let insertion_index = work_items.binary_search_by_key(&cost, |&(_, cost)| cost);
let insertion_index = match insertion_index {
Ok(idx) | Err(idx) => idx,
};
work_items.insert(insertion_index, (llvm_work_item, cost));
if cgcx.parallel {
helper.request_token();
}
assert_eq!(main_thread_state, MainThreadState::Codegenning);
main_thread_state = MainThreadState::Idle;
}
Message::CodegenComplete => {
if codegen_state != Aborted {
codegen_state = Completed;
}
assert_eq!(main_thread_state, MainThreadState::Codegenning);
main_thread_state = MainThreadState::Idle;
}
// If codegen is aborted that means translation was aborted due
// to some normal-ish compiler error. In this situation we want
// to exit as soon as possible, but we want to make sure all
// existing work has finished. Flag codegen as being done, and
// then conditions above will ensure no more work is spawned but
// we'll keep executing this loop until `running_with_own_token`
// hits 0.
Message::CodegenAborted => {
codegen_state = Aborted;
}
Message::WorkItem { result, worker_id } => {
free_worker(worker_id);
match result {
Ok(WorkItemResult::Finished(compiled_module)) => {
match compiled_module.kind {
ModuleKind::Regular => {
assert!(needs_link.is_empty());
compiled_modules.push(compiled_module);
}
ModuleKind::Allocator => {
assert!(compiled_allocator_module.is_none());
compiled_allocator_module = Some(compiled_module);
}
ModuleKind::Metadata => bug!("Should be handled separately"),
}
}
Ok(WorkItemResult::NeedsLink(module)) => {
assert!(compiled_modules.is_empty());
needs_link.push(module);
}
Ok(WorkItemResult::NeedsFatLto(fat_lto_input)) => {
assert!(!started_lto);
assert!(needs_thin_lto.is_empty());
needs_fat_lto.push(fat_lto_input);
}
Ok(WorkItemResult::NeedsThinLto(name, thin_buffer)) => {
assert!(!started_lto);
assert!(needs_fat_lto.is_empty());
needs_thin_lto.push((name, thin_buffer));
}
Err(Some(WorkerFatalError)) => {
// Like `CodegenAborted`, wait for remaining work to finish.
codegen_state = Aborted;
}
Err(None) => {
// If the thread failed that means it panicked, so
// we abort immediately.
bug!("worker thread panicked");
}
}
}
Message::AddImportOnlyModule { module_data, work_product } => {
assert!(!started_lto);
assert_eq!(codegen_state, Ongoing);
assert_eq!(main_thread_state, MainThreadState::Codegenning);
lto_import_only_modules.push((module_data, work_product));
main_thread_state = MainThreadState::Idle;
}
}
}
if codegen_state == Aborted {
return Err(());
}
let needs_link = mem::take(&mut needs_link);
if !needs_link.is_empty() {
assert!(compiled_modules.is_empty());
let dcx = cgcx.create_dcx();
let dcx = dcx.handle();
let module = B::run_link(&cgcx, dcx, needs_link).map_err(|_| ())?;
let module = unsafe {
B::codegen(&cgcx, dcx, module, cgcx.config(ModuleKind::Regular)).map_err(|_| ())?
};
compiled_modules.push(module);
}
// Drop to print timings
drop(llvm_start_time);
// Regardless of what order these modules completed in, report them to
// the backend in the same order every time to ensure that we're handing
// out deterministic results.
compiled_modules.sort_by(|a, b| a.name.cmp(&b.name));
Ok(CompiledModules {
modules: compiled_modules,
allocator_module: compiled_allocator_module,
})
})
.expect("failed to spawn coordinator thread");
// A heuristic that determines if we have enough LLVM WorkItems in the
// queue so that the main thread can do LLVM work instead of codegen
fn queue_full_enough(items_in_queue: usize, workers_running: usize) -> bool {
// This heuristic scales ahead-of-time codegen according to available
// concurrency, as measured by `workers_running`. The idea is that the
// more concurrency we have available, the more demand there will be for
// work items, and the fuller the queue should be kept to meet demand.
// An important property of this approach is that we codegen ahead of
// time only as much as necessary, so as to keep fewer LLVM modules in
// memory at once, thereby reducing memory consumption.
//
// When the number of workers running is less than the max concurrency
// available to us, this heuristic can cause us to instruct the main
// thread to work on an LLVM item (that is, tell it to "LLVM") instead
// of codegen, even though it seems like it *should* be codegenning so
// that we can create more work items and spawn more LLVM workers.
//
// But this is not a problem. When the main thread is told to LLVM,
// according to this heuristic and how work is scheduled, there is
// always at least one item in the queue, and therefore at least one
// pending jobserver token request. If there *is* more concurrency
// available, we will immediately receive a token, which will upgrade
// the main thread's LLVM worker to a real one (conceptually), and free
// up the main thread to codegen if necessary. On the other hand, if
// there isn't more concurrency, then the main thread working on an LLVM
// item is appropriate, as long as the queue is full enough for demand.
//
// Speaking of which, how full should we keep the queue? Probably less
// full than you'd think. A lot has to go wrong for the queue not to be
// full enough and for that to have a negative effect on compile times.
//
// Workers are unlikely to finish at exactly the same time, so when one
// finishes and takes another work item off the queue, we often have
// ample time to codegen at that point before the next worker finishes.
// But suppose that codegen takes so long that the workers exhaust the
// queue, and we have one or more workers that have nothing to work on.
// Well, it might not be so bad. Of all the LLVM modules we create and
// optimize, one has to finish last. It's not necessarily the case that
// by losing some concurrency for a moment, we delay the point at which
// that last LLVM module is finished and the rest of compilation can
// proceed. Also, when we can't take advantage of some concurrency, we
// give tokens back to the job server. That enables some other rustc to
// potentially make use of the available concurrency. That could even
// *decrease* overall compile time if we're lucky. But yes, if no other
// rustc can make use of the concurrency, then we've squandered it.
//
// However, keeping the queue full is also beneficial when we have a
// surge in available concurrency. Then items can be taken from the
// queue immediately, without having to wait for codegen.
//
// So, the heuristic below tries to keep one item in the queue for every
// four running workers. Based on limited benchmarking, this appears to
// be more than sufficient to avoid increasing compilation times.
let quarter_of_workers = workers_running - 3 * workers_running / 4;
items_in_queue > 0 && items_in_queue >= quarter_of_workers
}
}
/// `FatalError` is explicitly not `Send`.
#[must_use]
pub(crate) struct WorkerFatalError;
fn spawn_work<'a, B: ExtraBackendMethods>(
cgcx: &'a CodegenContext<B>,
llvm_start_time: &mut Option<VerboseTimingGuard<'a>>,
worker_id: usize,
work: WorkItem<B>,
) {
if cgcx.config(work.module_kind()).time_module && llvm_start_time.is_none() {
*llvm_start_time = Some(cgcx.prof.verbose_generic_activity("LLVM_passes"));
}
let cgcx = cgcx.clone();
B::spawn_named_thread(cgcx.time_trace, work.short_description(), move || {
// Set up a destructor which will fire off a message that we're done as
// we exit.
struct Bomb<B: ExtraBackendMethods> {
coordinator_send: Sender<Box<dyn Any + Send>>,
result: Option<Result<WorkItemResult<B>, FatalError>>,
worker_id: usize,
}
impl<B: ExtraBackendMethods> Drop for Bomb<B> {
fn drop(&mut self) {
let worker_id = self.worker_id;
let msg = match self.result.take() {
Some(Ok(result)) => Message::WorkItem::<B> { result: Ok(result), worker_id },
Some(Err(FatalError)) => {
Message::WorkItem::<B> { result: Err(Some(WorkerFatalError)), worker_id }
}
None => Message::WorkItem::<B> { result: Err(None), worker_id },
};
drop(self.coordinator_send.send(Box::new(msg)));
}
}
let mut bomb =
Bomb::<B> { coordinator_send: cgcx.coordinator_send.clone(), result: None, worker_id };
// Execute the work itself, and if it finishes successfully then flag
// ourselves as a success as well.
//
// Note that we ignore any `FatalError` coming out of `execute_work_item`,
// as a diagnostic was already sent off to the main thread - just
// surface that there was an error in this worker.
bomb.result = {
let module_config = cgcx.config(work.module_kind());
Some(match work {
WorkItem::Optimize(m) => {
let _timer =
cgcx.prof.generic_activity_with_arg("codegen_module_optimize", &*m.name);
execute_optimize_work_item(&cgcx, m, module_config)
}
WorkItem::CopyPostLtoArtifacts(m) => {
let _timer = cgcx.prof.generic_activity_with_arg(
"codegen_copy_artifacts_from_incr_cache",
&*m.name,
);
Ok(execute_copy_from_cache_work_item(&cgcx, m, module_config))
}
WorkItem::LTO(m) => {
let _timer =
cgcx.prof.generic_activity_with_arg("codegen_module_perform_lto", m.name());
execute_lto_work_item(&cgcx, m, module_config)
}
})
};
})
.expect("failed to spawn work thread");
}
enum SharedEmitterMessage {
Diagnostic(Diagnostic),
InlineAsmError(u32, String, Level, Option<(String, Vec<InnerSpan>)>),
Fatal(String),
}
#[derive(Clone)]
pub struct SharedEmitter {
sender: Sender<SharedEmitterMessage>,
}
pub struct SharedEmitterMain {
receiver: Receiver<SharedEmitterMessage>,
}
impl SharedEmitter {
fn new() -> (SharedEmitter, SharedEmitterMain) {
let (sender, receiver) = channel();
(SharedEmitter { sender }, SharedEmitterMain { receiver })
}
pub fn inline_asm_error(
&self,
cookie: u32,
msg: String,
level: Level,
source: Option<(String, Vec<InnerSpan>)>,
) {
drop(self.sender.send(SharedEmitterMessage::InlineAsmError(cookie, msg, level, source)));
}
fn fatal(&self, msg: &str) {
drop(self.sender.send(SharedEmitterMessage::Fatal(msg.to_string())));
}
}
impl Translate for SharedEmitter {
fn fluent_bundle(&self) -> Option<&Lrc<FluentBundle>> {
None
}
fn fallback_fluent_bundle(&self) -> &FluentBundle {
panic!("shared emitter attempted to translate a diagnostic");
}
}
impl Emitter for SharedEmitter {
fn emit_diagnostic(&mut self, mut diag: rustc_errors::DiagInner) {
// Check that we aren't missing anything interesting when converting to
// the cut-down local `DiagInner`.
assert_eq!(diag.span, MultiSpan::new());
assert_eq!(diag.suggestions, Suggestions::Enabled(vec![]));
assert_eq!(diag.sort_span, rustc_span::DUMMY_SP);
assert_eq!(diag.is_lint, None);
// No sensible check for `diag.emitted_at`.
let args = mem::replace(&mut diag.args, DiagArgMap::default());
drop(
self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
level: diag.level(),
messages: diag.messages,
code: diag.code,
children: diag
.children
.into_iter()
.map(|child| Subdiagnostic { level: child.level, messages: child.messages })
.collect(),
args,
})),
);
}
fn source_map(&self) -> Option<&Lrc<SourceMap>> {
None
}
}
impl SharedEmitterMain {
fn check(&self, sess: &Session, blocking: bool) {
loop {
let message = if blocking {
match self.receiver.recv() {
Ok(message) => Ok(message),
Err(_) => Err(()),
}
} else {
match self.receiver.try_recv() {
Ok(message) => Ok(message),
Err(_) => Err(()),
}
};
match message {
Ok(SharedEmitterMessage::Diagnostic(diag)) => {
// The diagnostic has been received on the main thread.
// Convert it back to a full `Diagnostic` and emit.
let dcx = sess.dcx();
let mut d =
rustc_errors::DiagInner::new_with_messages(diag.level, diag.messages);
d.code = diag.code; // may be `None`, that's ok
d.children = diag
.children
.into_iter()
.map(|sub| rustc_errors::Subdiag {
level: sub.level,
messages: sub.messages,
span: MultiSpan::new(),
})
.collect();
d.args = diag.args;
dcx.emit_diagnostic(d);
sess.dcx().abort_if_errors();
}
Ok(SharedEmitterMessage::InlineAsmError(cookie, msg, level, source)) => {
assert_matches!(level, Level::Error | Level::Warning | Level::Note);
let msg = msg.strip_prefix("error: ").unwrap_or(&msg).to_string();
let mut err = Diag::<()>::new(sess.dcx(), level, msg);
// If the cookie is 0 then we don't have span information.
if cookie != 0 {
let pos = BytePos::from_u32(cookie);
let span = Span::with_root_ctxt(pos, pos);
err.span(span);
};
// Point to the generated assembly if it is available.
if let Some((buffer, spans)) = source {
let source = sess
.source_map()
.new_source_file(FileName::inline_asm_source_code(&buffer), buffer);
let spans: Vec<_> = spans
.iter()
.map(|sp| {
Span::with_root_ctxt(
source.normalized_byte_pos(sp.start as u32),
source.normalized_byte_pos(sp.end as u32),
)
})
.collect();
err.span_note(spans, "instantiated into assembly here");
}
err.emit();
}
Ok(SharedEmitterMessage::Fatal(msg)) => {
sess.dcx().fatal(msg);
}
Err(_) => {
break;
}
}
}
}
}
pub struct Coordinator<B: ExtraBackendMethods> {
pub sender: Sender<Box<dyn Any + Send>>,
future: Option<thread::JoinHandle<Result<CompiledModules, ()>>>,
// Only used for the Message type.
phantom: PhantomData<B>,
}
impl<B: ExtraBackendMethods> Coordinator<B> {
fn join(mut self) -> std::thread::Result<Result<CompiledModules, ()>> {
self.future.take().unwrap().join()
}
}
impl<B: ExtraBackendMethods> Drop for Coordinator<B> {
fn drop(&mut self) {
if let Some(future) = self.future.take() {
// If we haven't joined yet, signal to the coordinator that it should spawn no more
// work, and wait for worker threads to finish.
drop(self.sender.send(Box::new(Message::CodegenAborted::<B>)));
drop(future.join());
}
}
}
pub struct OngoingCodegen<B: ExtraBackendMethods> {
pub backend: B,
pub metadata: EncodedMetadata,
pub metadata_module: Option<CompiledModule>,
pub crate_info: CrateInfo,
pub codegen_worker_receive: Receiver<CguMessage>,
pub shared_emitter_main: SharedEmitterMain,
pub output_filenames: Arc<OutputFilenames>,
pub coordinator: Coordinator<B>,
}
impl<B: ExtraBackendMethods> OngoingCodegen<B> {
pub fn join(self, sess: &Session) -> (CodegenResults, FxIndexMap<WorkProductId, WorkProduct>) {
let _timer = sess.timer("finish_ongoing_codegen");
self.shared_emitter_main.check(sess, true);
let compiled_modules = sess.time("join_worker_thread", || match self.coordinator.join() {
Ok(Ok(compiled_modules)) => compiled_modules,
Ok(Err(())) => {
sess.dcx().abort_if_errors();
panic!("expected abort due to worker thread errors")
}
Err(_) => {
bug!("panic during codegen/LLVM phase");
}
});
sess.dcx().abort_if_errors();
let work_products =
copy_all_cgu_workproducts_to_incr_comp_cache_dir(sess, &compiled_modules);
produce_final_output_artifacts(sess, &compiled_modules, &self.output_filenames);
// FIXME: time_llvm_passes support - does this use a global context or
// something?
if sess.codegen_units().as_usize() == 1 && sess.opts.unstable_opts.time_llvm_passes {
self.backend.print_pass_timings()
}
if sess.print_llvm_stats() {
self.backend.print_statistics()
}
(
CodegenResults {
metadata: self.metadata,
crate_info: self.crate_info,
modules: compiled_modules.modules,
allocator_module: compiled_modules.allocator_module,
metadata_module: self.metadata_module,
},
work_products,
)
}
pub(crate) fn codegen_finished(&self, tcx: TyCtxt<'_>) {
self.wait_for_signal_to_codegen_item();
self.check_for_errors(tcx.sess);
drop(self.coordinator.sender.send(Box::new(Message::CodegenComplete::<B>)));
}
pub(crate) fn check_for_errors(&self, sess: &Session) {
self.shared_emitter_main.check(sess, false);
}
pub(crate) fn wait_for_signal_to_codegen_item(&self) {
match self.codegen_worker_receive.recv() {
Ok(CguMessage) => {
// Ok to proceed.
}
Err(_) => {
// One of the LLVM threads must have panicked, fall through so
// error handling can be reached.
}
}
}
}
pub(crate) fn submit_codegened_module_to_llvm<B: ExtraBackendMethods>(
_backend: &B,
tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>,
module: ModuleCodegen<B::Module>,
cost: u64,
) {
let llvm_work_item = WorkItem::Optimize(module);
drop(tx_to_llvm_workers.send(Box::new(Message::CodegenDone::<B> { llvm_work_item, cost })));
}
pub(crate) fn submit_post_lto_module_to_llvm<B: ExtraBackendMethods>(
_backend: &B,
tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>,
module: CachedModuleCodegen,
) {
let llvm_work_item = WorkItem::CopyPostLtoArtifacts(module);
drop(tx_to_llvm_workers.send(Box::new(Message::CodegenDone::<B> { llvm_work_item, cost: 0 })));
}
pub(crate) fn submit_pre_lto_module_to_llvm<B: ExtraBackendMethods>(
_backend: &B,
tcx: TyCtxt<'_>,
tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>,
module: CachedModuleCodegen,
) {
let filename = pre_lto_bitcode_filename(&module.name);
let bc_path = in_incr_comp_dir_sess(tcx.sess, &filename);
let file = fs::File::open(&bc_path)
.unwrap_or_else(|e| panic!("failed to open bitcode file `{}`: {}", bc_path.display(), e));
let mmap = unsafe {
Mmap::map(file).unwrap_or_else(|e| {
panic!("failed to mmap bitcode file `{}`: {}", bc_path.display(), e)
})
};
// Schedule the module to be loaded
drop(tx_to_llvm_workers.send(Box::new(Message::AddImportOnlyModule::<B> {
module_data: SerializedModule::FromUncompressedFile(mmap),
work_product: module.source,
})));
}
fn pre_lto_bitcode_filename(module_name: &str) -> String {
format!("{module_name}.{PRE_LTO_BC_EXT}")
}
fn msvc_imps_needed(tcx: TyCtxt<'_>) -> bool {
// This should never be true (because it's not supported). If it is true,
// something is wrong with commandline arg validation.
assert!(
!(tcx.sess.opts.cg.linker_plugin_lto.enabled()
&& tcx.sess.target.is_like_windows
&& tcx.sess.opts.cg.prefer_dynamic)
);
tcx.sess.target.is_like_windows &&
tcx.crate_types().iter().any(|ct| *ct == CrateType::Rlib) &&
// ThinLTO can't handle this workaround in all cases, so we don't
// emit the `__imp_` symbols. Instead we make them unnecessary by disallowing
// dynamic linking when linker plugin LTO is enabled.
!tcx.sess.opts.cg.linker_plugin_lto.enabled()
}