1 //===- Writer.cpp ---------------------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "Writer.h" 10 #include "CallGraphSort.h" 11 #include "Config.h" 12 #include "DLL.h" 13 #include "InputFiles.h" 14 #include "LLDMapFile.h" 15 #include "MapFile.h" 16 #include "PDB.h" 17 #include "SymbolTable.h" 18 #include "Symbols.h" 19 #include "lld/Common/ErrorHandler.h" 20 #include "lld/Common/Memory.h" 21 #include "lld/Common/Timer.h" 22 #include "llvm/ADT/DenseMap.h" 23 #include "llvm/ADT/STLExtras.h" 24 #include "llvm/ADT/StringSet.h" 25 #include "llvm/ADT/StringSwitch.h" 26 #include "llvm/Support/BinaryStreamReader.h" 27 #include "llvm/Support/Debug.h" 28 #include "llvm/Support/Endian.h" 29 #include "llvm/Support/FileOutputBuffer.h" 30 #include "llvm/Support/Parallel.h" 31 #include "llvm/Support/Path.h" 32 #include "llvm/Support/RandomNumberGenerator.h" 33 #include "llvm/Support/xxhash.h" 34 #include <algorithm> 35 #include <cstdio> 36 #include <map> 37 #include <memory> 38 #include <utility> 39 40 using namespace llvm; 41 using namespace llvm::COFF; 42 using namespace llvm::object; 43 using namespace llvm::support; 44 using namespace llvm::support::endian; 45 using namespace lld; 46 using namespace lld::coff; 47 48 /* To re-generate DOSProgram: 49 $ cat > /tmp/DOSProgram.asm 50 org 0 51 ; Copy cs to ds. 52 push cs 53 pop ds 54 ; Point ds:dx at the $-terminated string. 55 mov dx, str 56 ; Int 21/AH=09h: Write string to standard output. 57 mov ah, 0x9 58 int 0x21 59 ; Int 21/AH=4Ch: Exit with return code (in AL). 60 mov ax, 0x4C01 61 int 0x21 62 str: 63 db 'This program cannot be run in DOS mode.$' 64 align 8, db 0 65 $ nasm -fbin /tmp/DOSProgram.asm -o /tmp/DOSProgram.bin 66 $ xxd -i /tmp/DOSProgram.bin 67 */ 68 static unsigned char dosProgram[] = { 69 0x0e, 0x1f, 0xba, 0x0e, 0x00, 0xb4, 0x09, 0xcd, 0x21, 0xb8, 0x01, 0x4c, 70 0xcd, 0x21, 0x54, 0x68, 0x69, 0x73, 0x20, 0x70, 0x72, 0x6f, 0x67, 0x72, 71 0x61, 0x6d, 0x20, 0x63, 0x61, 0x6e, 0x6e, 0x6f, 0x74, 0x20, 0x62, 0x65, 72 0x20, 0x72, 0x75, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x44, 0x4f, 0x53, 0x20, 73 0x6d, 0x6f, 0x64, 0x65, 0x2e, 0x24, 0x00, 0x00 74 }; 75 static_assert(sizeof(dosProgram) % 8 == 0, 76 "DOSProgram size must be multiple of 8"); 77 78 static const int dosStubSize = sizeof(dos_header) + sizeof(dosProgram); 79 static_assert(dosStubSize % 8 == 0, "DOSStub size must be multiple of 8"); 80 81 static const int numberOfDataDirectory = 16; 82 83 // Global vector of all output sections. After output sections are finalized, 84 // this can be indexed by Chunk::getOutputSection. 85 static std::vector<OutputSection *> outputSections; 86 87 OutputSection *Chunk::getOutputSection() const { 88 return osidx == 0 ? nullptr : outputSections[osidx - 1]; 89 } 90 91 void OutputSection::clear() { outputSections.clear(); } 92 93 namespace { 94 95 class DebugDirectoryChunk : public NonSectionChunk { 96 public: 97 DebugDirectoryChunk(const std::vector<std::pair<COFF::DebugType, Chunk *>> &r, 98 bool writeRepro) 99 : records(r), writeRepro(writeRepro) {} 100 101 size_t getSize() const override { 102 return (records.size() + int(writeRepro)) * sizeof(debug_directory); 103 } 104 105 void writeTo(uint8_t *b) const override { 106 auto *d = reinterpret_cast<debug_directory *>(b); 107 108 for (const std::pair<COFF::DebugType, Chunk *>& record : records) { 109 Chunk *c = record.second; 110 OutputSection *os = c->getOutputSection(); 111 uint64_t offs = os->getFileOff() + (c->getRVA() - os->getRVA()); 112 fillEntry(d, record.first, c->getSize(), c->getRVA(), offs); 113 ++d; 114 } 115 116 if (writeRepro) { 117 // FIXME: The COFF spec allows either a 0-sized entry to just say 118 // "the timestamp field is really a hash", or a 4-byte size field 119 // followed by that many bytes containing a longer hash (with the 120 // lowest 4 bytes usually being the timestamp in little-endian order). 121 // Consider storing the full 8 bytes computed by xxHash64 here. 122 fillEntry(d, COFF::IMAGE_DEBUG_TYPE_REPRO, 0, 0, 0); 123 } 124 } 125 126 void setTimeDateStamp(uint32_t timeDateStamp) { 127 for (support::ulittle32_t *tds : timeDateStamps) 128 *tds = timeDateStamp; 129 } 130 131 private: 132 void fillEntry(debug_directory *d, COFF::DebugType debugType, size_t size, 133 uint64_t rva, uint64_t offs) const { 134 d->Characteristics = 0; 135 d->TimeDateStamp = 0; 136 d->MajorVersion = 0; 137 d->MinorVersion = 0; 138 d->Type = debugType; 139 d->SizeOfData = size; 140 d->AddressOfRawData = rva; 141 d->PointerToRawData = offs; 142 143 timeDateStamps.push_back(&d->TimeDateStamp); 144 } 145 146 mutable std::vector<support::ulittle32_t *> timeDateStamps; 147 const std::vector<std::pair<COFF::DebugType, Chunk *>> &records; 148 bool writeRepro; 149 }; 150 151 class CVDebugRecordChunk : public NonSectionChunk { 152 public: 153 size_t getSize() const override { 154 return sizeof(codeview::DebugInfo) + config->pdbAltPath.size() + 1; 155 } 156 157 void writeTo(uint8_t *b) const override { 158 // Save off the DebugInfo entry to backfill the file signature (build id) 159 // in Writer::writeBuildId 160 buildId = reinterpret_cast<codeview::DebugInfo *>(b); 161 162 // variable sized field (PDB Path) 163 char *p = reinterpret_cast<char *>(b + sizeof(*buildId)); 164 if (!config->pdbAltPath.empty()) 165 memcpy(p, config->pdbAltPath.data(), config->pdbAltPath.size()); 166 p[config->pdbAltPath.size()] = '\0'; 167 } 168 169 mutable codeview::DebugInfo *buildId = nullptr; 170 }; 171 172 class ExtendedDllCharacteristicsChunk : public NonSectionChunk { 173 public: 174 ExtendedDllCharacteristicsChunk(uint32_t c) : characteristics(c) {} 175 176 size_t getSize() const override { return 4; } 177 178 void writeTo(uint8_t *buf) const override { write32le(buf, characteristics); } 179 180 uint32_t characteristics = 0; 181 }; 182 183 // PartialSection represents a group of chunks that contribute to an 184 // OutputSection. Collating a collection of PartialSections of same name and 185 // characteristics constitutes the OutputSection. 186 class PartialSectionKey { 187 public: 188 StringRef name; 189 unsigned characteristics; 190 191 bool operator<(const PartialSectionKey &other) const { 192 int c = name.compare(other.name); 193 if (c == 1) 194 return false; 195 if (c == 0) 196 return characteristics < other.characteristics; 197 return true; 198 } 199 }; 200 201 // The writer writes a SymbolTable result to a file. 202 class Writer { 203 public: 204 Writer() : buffer(errorHandler().outputBuffer) {} 205 void run(); 206 207 private: 208 void createSections(); 209 void createMiscChunks(); 210 void createImportTables(); 211 void appendImportThunks(); 212 void locateImportTables(); 213 void createExportTable(); 214 void mergeSections(); 215 void removeUnusedSections(); 216 void assignAddresses(); 217 void finalizeAddresses(); 218 void removeEmptySections(); 219 void assignOutputSectionIndices(); 220 void createSymbolAndStringTable(); 221 void openFile(StringRef outputPath); 222 template <typename PEHeaderTy> void writeHeader(); 223 void createSEHTable(); 224 void createRuntimePseudoRelocs(); 225 void insertCtorDtorSymbols(); 226 void createGuardCFTables(); 227 void markSymbolsForRVATable(ObjFile *file, 228 ArrayRef<SectionChunk *> symIdxChunks, 229 SymbolRVASet &tableSymbols); 230 void getSymbolsFromSections(ObjFile *file, 231 ArrayRef<SectionChunk *> symIdxChunks, 232 std::vector<Symbol *> &symbols); 233 void maybeAddRVATable(SymbolRVASet tableSymbols, StringRef tableSym, 234 StringRef countSym); 235 void setSectionPermissions(); 236 void writeSections(); 237 void writeBuildId(); 238 void sortSections(); 239 void sortExceptionTable(); 240 void sortCRTSectionChunks(std::vector<Chunk *> &chunks); 241 void addSyntheticIdata(); 242 void fixPartialSectionChars(StringRef name, uint32_t chars); 243 bool fixGnuImportChunks(); 244 void fixTlsAlignment(); 245 PartialSection *createPartialSection(StringRef name, uint32_t outChars); 246 PartialSection *findPartialSection(StringRef name, uint32_t outChars); 247 248 llvm::Optional<coff_symbol16> createSymbol(Defined *d); 249 size_t addEntryToStringTable(StringRef str); 250 251 OutputSection *findSection(StringRef name); 252 void addBaserels(); 253 void addBaserelBlocks(std::vector<Baserel> &v); 254 255 uint32_t getSizeOfInitializedData(); 256 257 std::unique_ptr<FileOutputBuffer> &buffer; 258 std::map<PartialSectionKey, PartialSection *> partialSections; 259 std::vector<char> strtab; 260 std::vector<llvm::object::coff_symbol16> outputSymtab; 261 IdataContents idata; 262 Chunk *importTableStart = nullptr; 263 uint64_t importTableSize = 0; 264 Chunk *edataStart = nullptr; 265 Chunk *edataEnd = nullptr; 266 Chunk *iatStart = nullptr; 267 uint64_t iatSize = 0; 268 DelayLoadContents delayIdata; 269 EdataContents edata; 270 bool setNoSEHCharacteristic = false; 271 uint32_t tlsAlignment = 0; 272 273 DebugDirectoryChunk *debugDirectory = nullptr; 274 std::vector<std::pair<COFF::DebugType, Chunk *>> debugRecords; 275 CVDebugRecordChunk *buildId = nullptr; 276 ArrayRef<uint8_t> sectionTable; 277 278 uint64_t fileSize; 279 uint32_t pointerToSymbolTable = 0; 280 uint64_t sizeOfImage; 281 uint64_t sizeOfHeaders; 282 283 OutputSection *textSec; 284 OutputSection *rdataSec; 285 OutputSection *buildidSec; 286 OutputSection *dataSec; 287 OutputSection *pdataSec; 288 OutputSection *idataSec; 289 OutputSection *edataSec; 290 OutputSection *didatSec; 291 OutputSection *rsrcSec; 292 OutputSection *relocSec; 293 OutputSection *ctorsSec; 294 OutputSection *dtorsSec; 295 296 // The first and last .pdata sections in the output file. 297 // 298 // We need to keep track of the location of .pdata in whichever section it 299 // gets merged into so that we can sort its contents and emit a correct data 300 // directory entry for the exception table. This is also the case for some 301 // other sections (such as .edata) but because the contents of those sections 302 // are entirely linker-generated we can keep track of their locations using 303 // the chunks that the linker creates. All .pdata chunks come from input 304 // files, so we need to keep track of them separately. 305 Chunk *firstPdata = nullptr; 306 Chunk *lastPdata; 307 }; 308 } // anonymous namespace 309 310 static Timer codeLayoutTimer("Code Layout", Timer::root()); 311 static Timer diskCommitTimer("Commit Output File", Timer::root()); 312 313 void lld::coff::writeResult() { Writer().run(); } 314 315 void OutputSection::addChunk(Chunk *c) { 316 chunks.push_back(c); 317 } 318 319 void OutputSection::insertChunkAtStart(Chunk *c) { 320 chunks.insert(chunks.begin(), c); 321 } 322 323 void OutputSection::setPermissions(uint32_t c) { 324 header.Characteristics &= ~permMask; 325 header.Characteristics |= c; 326 } 327 328 void OutputSection::merge(OutputSection *other) { 329 chunks.insert(chunks.end(), other->chunks.begin(), other->chunks.end()); 330 other->chunks.clear(); 331 contribSections.insert(contribSections.end(), other->contribSections.begin(), 332 other->contribSections.end()); 333 other->contribSections.clear(); 334 } 335 336 // Write the section header to a given buffer. 337 void OutputSection::writeHeaderTo(uint8_t *buf) { 338 auto *hdr = reinterpret_cast<coff_section *>(buf); 339 *hdr = header; 340 if (stringTableOff) { 341 // If name is too long, write offset into the string table as a name. 342 sprintf(hdr->Name, "/%d", stringTableOff); 343 } else { 344 assert(!config->debug || name.size() <= COFF::NameSize || 345 (hdr->Characteristics & IMAGE_SCN_MEM_DISCARDABLE) == 0); 346 strncpy(hdr->Name, name.data(), 347 std::min(name.size(), (size_t)COFF::NameSize)); 348 } 349 } 350 351 void OutputSection::addContributingPartialSection(PartialSection *sec) { 352 contribSections.push_back(sec); 353 } 354 355 // Check whether the target address S is in range from a relocation 356 // of type relType at address P. 357 static bool isInRange(uint16_t relType, uint64_t s, uint64_t p, int margin) { 358 if (config->machine == ARMNT) { 359 int64_t diff = AbsoluteDifference(s, p + 4) + margin; 360 switch (relType) { 361 case IMAGE_REL_ARM_BRANCH20T: 362 return isInt<21>(diff); 363 case IMAGE_REL_ARM_BRANCH24T: 364 case IMAGE_REL_ARM_BLX23T: 365 return isInt<25>(diff); 366 default: 367 return true; 368 } 369 } else if (config->machine == ARM64) { 370 int64_t diff = AbsoluteDifference(s, p) + margin; 371 switch (relType) { 372 case IMAGE_REL_ARM64_BRANCH26: 373 return isInt<28>(diff); 374 case IMAGE_REL_ARM64_BRANCH19: 375 return isInt<21>(diff); 376 case IMAGE_REL_ARM64_BRANCH14: 377 return isInt<16>(diff); 378 default: 379 return true; 380 } 381 } else { 382 llvm_unreachable("Unexpected architecture"); 383 } 384 } 385 386 // Return the last thunk for the given target if it is in range, 387 // or create a new one. 388 static std::pair<Defined *, bool> 389 getThunk(DenseMap<uint64_t, Defined *> &lastThunks, Defined *target, uint64_t p, 390 uint16_t type, int margin) { 391 Defined *&lastThunk = lastThunks[target->getRVA()]; 392 if (lastThunk && isInRange(type, lastThunk->getRVA(), p, margin)) 393 return {lastThunk, false}; 394 Chunk *c; 395 switch (config->machine) { 396 case ARMNT: 397 c = make<RangeExtensionThunkARM>(target); 398 break; 399 case ARM64: 400 c = make<RangeExtensionThunkARM64>(target); 401 break; 402 default: 403 llvm_unreachable("Unexpected architecture"); 404 } 405 Defined *d = make<DefinedSynthetic>("", c); 406 lastThunk = d; 407 return {d, true}; 408 } 409 410 // This checks all relocations, and for any relocation which isn't in range 411 // it adds a thunk after the section chunk that contains the relocation. 412 // If the latest thunk for the specific target is in range, that is used 413 // instead of creating a new thunk. All range checks are done with the 414 // specified margin, to make sure that relocations that originally are in 415 // range, but only barely, also get thunks - in case other added thunks makes 416 // the target go out of range. 417 // 418 // After adding thunks, we verify that all relocations are in range (with 419 // no extra margin requirements). If this failed, we restart (throwing away 420 // the previously created thunks) and retry with a wider margin. 421 static bool createThunks(OutputSection *os, int margin) { 422 bool addressesChanged = false; 423 DenseMap<uint64_t, Defined *> lastThunks; 424 DenseMap<std::pair<ObjFile *, Defined *>, uint32_t> thunkSymtabIndices; 425 size_t thunksSize = 0; 426 // Recheck Chunks.size() each iteration, since we can insert more 427 // elements into it. 428 for (size_t i = 0; i != os->chunks.size(); ++i) { 429 SectionChunk *sc = dyn_cast_or_null<SectionChunk>(os->chunks[i]); 430 if (!sc) 431 continue; 432 size_t thunkInsertionSpot = i + 1; 433 434 // Try to get a good enough estimate of where new thunks will be placed. 435 // Offset this by the size of the new thunks added so far, to make the 436 // estimate slightly better. 437 size_t thunkInsertionRVA = sc->getRVA() + sc->getSize() + thunksSize; 438 ObjFile *file = sc->file; 439 std::vector<std::pair<uint32_t, uint32_t>> relocReplacements; 440 ArrayRef<coff_relocation> originalRelocs = 441 file->getCOFFObj()->getRelocations(sc->header); 442 for (size_t j = 0, e = originalRelocs.size(); j < e; ++j) { 443 const coff_relocation &rel = originalRelocs[j]; 444 Symbol *relocTarget = file->getSymbol(rel.SymbolTableIndex); 445 446 // The estimate of the source address P should be pretty accurate, 447 // but we don't know whether the target Symbol address should be 448 // offset by thunksSize or not (or by some of thunksSize but not all of 449 // it), giving us some uncertainty once we have added one thunk. 450 uint64_t p = sc->getRVA() + rel.VirtualAddress + thunksSize; 451 452 Defined *sym = dyn_cast_or_null<Defined>(relocTarget); 453 if (!sym) 454 continue; 455 456 uint64_t s = sym->getRVA(); 457 458 if (isInRange(rel.Type, s, p, margin)) 459 continue; 460 461 // If the target isn't in range, hook it up to an existing or new 462 // thunk. 463 Defined *thunk; 464 bool wasNew; 465 std::tie(thunk, wasNew) = getThunk(lastThunks, sym, p, rel.Type, margin); 466 if (wasNew) { 467 Chunk *thunkChunk = thunk->getChunk(); 468 thunkChunk->setRVA( 469 thunkInsertionRVA); // Estimate of where it will be located. 470 os->chunks.insert(os->chunks.begin() + thunkInsertionSpot, thunkChunk); 471 thunkInsertionSpot++; 472 thunksSize += thunkChunk->getSize(); 473 thunkInsertionRVA += thunkChunk->getSize(); 474 addressesChanged = true; 475 } 476 477 // To redirect the relocation, add a symbol to the parent object file's 478 // symbol table, and replace the relocation symbol table index with the 479 // new index. 480 auto insertion = thunkSymtabIndices.insert({{file, thunk}, ~0U}); 481 uint32_t &thunkSymbolIndex = insertion.first->second; 482 if (insertion.second) 483 thunkSymbolIndex = file->addRangeThunkSymbol(thunk); 484 relocReplacements.push_back({j, thunkSymbolIndex}); 485 } 486 487 // Get a writable copy of this section's relocations so they can be 488 // modified. If the relocations point into the object file, allocate new 489 // memory. Otherwise, this must be previously allocated memory that can be 490 // modified in place. 491 ArrayRef<coff_relocation> curRelocs = sc->getRelocs(); 492 MutableArrayRef<coff_relocation> newRelocs; 493 if (originalRelocs.data() == curRelocs.data()) { 494 newRelocs = makeMutableArrayRef( 495 bAlloc.Allocate<coff_relocation>(originalRelocs.size()), 496 originalRelocs.size()); 497 } else { 498 newRelocs = makeMutableArrayRef( 499 const_cast<coff_relocation *>(curRelocs.data()), curRelocs.size()); 500 } 501 502 // Copy each relocation, but replace the symbol table indices which need 503 // thunks. 504 auto nextReplacement = relocReplacements.begin(); 505 auto endReplacement = relocReplacements.end(); 506 for (size_t i = 0, e = originalRelocs.size(); i != e; ++i) { 507 newRelocs[i] = originalRelocs[i]; 508 if (nextReplacement != endReplacement && nextReplacement->first == i) { 509 newRelocs[i].SymbolTableIndex = nextReplacement->second; 510 ++nextReplacement; 511 } 512 } 513 514 sc->setRelocs(newRelocs); 515 } 516 return addressesChanged; 517 } 518 519 // Verify that all relocations are in range, with no extra margin requirements. 520 static bool verifyRanges(const std::vector<Chunk *> chunks) { 521 for (Chunk *c : chunks) { 522 SectionChunk *sc = dyn_cast_or_null<SectionChunk>(c); 523 if (!sc) 524 continue; 525 526 ArrayRef<coff_relocation> relocs = sc->getRelocs(); 527 for (size_t j = 0, e = relocs.size(); j < e; ++j) { 528 const coff_relocation &rel = relocs[j]; 529 Symbol *relocTarget = sc->file->getSymbol(rel.SymbolTableIndex); 530 531 Defined *sym = dyn_cast_or_null<Defined>(relocTarget); 532 if (!sym) 533 continue; 534 535 uint64_t p = sc->getRVA() + rel.VirtualAddress; 536 uint64_t s = sym->getRVA(); 537 538 if (!isInRange(rel.Type, s, p, 0)) 539 return false; 540 } 541 } 542 return true; 543 } 544 545 // Assign addresses and add thunks if necessary. 546 void Writer::finalizeAddresses() { 547 assignAddresses(); 548 if (config->machine != ARMNT && config->machine != ARM64) 549 return; 550 551 size_t origNumChunks = 0; 552 for (OutputSection *sec : outputSections) { 553 sec->origChunks = sec->chunks; 554 origNumChunks += sec->chunks.size(); 555 } 556 557 int pass = 0; 558 int margin = 1024 * 100; 559 while (true) { 560 // First check whether we need thunks at all, or if the previous pass of 561 // adding them turned out ok. 562 bool rangesOk = true; 563 size_t numChunks = 0; 564 for (OutputSection *sec : outputSections) { 565 if (!verifyRanges(sec->chunks)) { 566 rangesOk = false; 567 break; 568 } 569 numChunks += sec->chunks.size(); 570 } 571 if (rangesOk) { 572 if (pass > 0) 573 log("Added " + Twine(numChunks - origNumChunks) + " thunks with " + 574 "margin " + Twine(margin) + " in " + Twine(pass) + " passes"); 575 return; 576 } 577 578 if (pass >= 10) 579 fatal("adding thunks hasn't converged after " + Twine(pass) + " passes"); 580 581 if (pass > 0) { 582 // If the previous pass didn't work out, reset everything back to the 583 // original conditions before retrying with a wider margin. This should 584 // ideally never happen under real circumstances. 585 for (OutputSection *sec : outputSections) 586 sec->chunks = sec->origChunks; 587 margin *= 2; 588 } 589 590 // Try adding thunks everywhere where it is needed, with a margin 591 // to avoid things going out of range due to the added thunks. 592 bool addressesChanged = false; 593 for (OutputSection *sec : outputSections) 594 addressesChanged |= createThunks(sec, margin); 595 // If the verification above thought we needed thunks, we should have 596 // added some. 597 assert(addressesChanged); 598 599 // Recalculate the layout for the whole image (and verify the ranges at 600 // the start of the next round). 601 assignAddresses(); 602 603 pass++; 604 } 605 } 606 607 // The main function of the writer. 608 void Writer::run() { 609 ScopedTimer t1(codeLayoutTimer); 610 611 // First, clear the output sections from previous runs 612 outputSections.clear(); 613 614 createImportTables(); 615 createSections(); 616 appendImportThunks(); 617 // Import thunks must be added before the Control Flow Guard tables are added. 618 createMiscChunks(); 619 createExportTable(); 620 mergeSections(); 621 removeUnusedSections(); 622 finalizeAddresses(); 623 removeEmptySections(); 624 assignOutputSectionIndices(); 625 setSectionPermissions(); 626 createSymbolAndStringTable(); 627 628 if (fileSize > UINT32_MAX) 629 fatal("image size (" + Twine(fileSize) + ") " + 630 "exceeds maximum allowable size (" + Twine(UINT32_MAX) + ")"); 631 632 openFile(config->outputFile); 633 if (config->is64()) { 634 writeHeader<pe32plus_header>(); 635 } else { 636 writeHeader<pe32_header>(); 637 } 638 writeSections(); 639 sortExceptionTable(); 640 641 // Fix up the alignment in the TLS Directory's characteristic field, 642 // if a specific alignment value is needed 643 if (tlsAlignment) 644 fixTlsAlignment(); 645 646 t1.stop(); 647 648 if (!config->pdbPath.empty() && config->debug) { 649 assert(buildId); 650 createPDB(symtab, outputSections, sectionTable, buildId->buildId); 651 } 652 writeBuildId(); 653 654 writeLLDMapFile(outputSections); 655 writeMapFile(outputSections); 656 657 if (errorCount()) 658 return; 659 660 ScopedTimer t2(diskCommitTimer); 661 if (auto e = buffer->commit()) 662 fatal("failed to write the output file: " + toString(std::move(e))); 663 } 664 665 static StringRef getOutputSectionName(StringRef name) { 666 StringRef s = name.split('$').first; 667 668 // Treat a later period as a separator for MinGW, for sections like 669 // ".ctors.01234". 670 return s.substr(0, s.find('.', 1)); 671 } 672 673 // For /order. 674 static void sortBySectionOrder(std::vector<Chunk *> &chunks) { 675 auto getPriority = [](const Chunk *c) { 676 if (auto *sec = dyn_cast<SectionChunk>(c)) 677 if (sec->sym) 678 return config->order.lookup(sec->sym->getName()); 679 return 0; 680 }; 681 682 llvm::stable_sort(chunks, [=](const Chunk *a, const Chunk *b) { 683 return getPriority(a) < getPriority(b); 684 }); 685 } 686 687 // Change the characteristics of existing PartialSections that belong to the 688 // section Name to Chars. 689 void Writer::fixPartialSectionChars(StringRef name, uint32_t chars) { 690 for (auto it : partialSections) { 691 PartialSection *pSec = it.second; 692 StringRef curName = pSec->name; 693 if (!curName.consume_front(name) || 694 (!curName.empty() && !curName.startswith("$"))) 695 continue; 696 if (pSec->characteristics == chars) 697 continue; 698 PartialSection *destSec = createPartialSection(pSec->name, chars); 699 destSec->chunks.insert(destSec->chunks.end(), pSec->chunks.begin(), 700 pSec->chunks.end()); 701 pSec->chunks.clear(); 702 } 703 } 704 705 // Sort concrete section chunks from GNU import libraries. 706 // 707 // GNU binutils doesn't use short import files, but instead produces import 708 // libraries that consist of object files, with section chunks for the .idata$* 709 // sections. These are linked just as regular static libraries. Each import 710 // library consists of one header object, one object file for every imported 711 // symbol, and one trailer object. In order for the .idata tables/lists to 712 // be formed correctly, the section chunks within each .idata$* section need 713 // to be grouped by library, and sorted alphabetically within each library 714 // (which makes sure the header comes first and the trailer last). 715 bool Writer::fixGnuImportChunks() { 716 uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ; 717 718 // Make sure all .idata$* section chunks are mapped as RDATA in order to 719 // be sorted into the same sections as our own synthesized .idata chunks. 720 fixPartialSectionChars(".idata", rdata); 721 722 bool hasIdata = false; 723 // Sort all .idata$* chunks, grouping chunks from the same library, 724 // with alphabetical ordering of the object fils within a library. 725 for (auto it : partialSections) { 726 PartialSection *pSec = it.second; 727 if (!pSec->name.startswith(".idata")) 728 continue; 729 730 if (!pSec->chunks.empty()) 731 hasIdata = true; 732 llvm::stable_sort(pSec->chunks, [&](Chunk *s, Chunk *t) { 733 SectionChunk *sc1 = dyn_cast_or_null<SectionChunk>(s); 734 SectionChunk *sc2 = dyn_cast_or_null<SectionChunk>(t); 735 if (!sc1 || !sc2) { 736 // if SC1, order them ascending. If SC2 or both null, 737 // S is not less than T. 738 return sc1 != nullptr; 739 } 740 // Make a string with "libraryname/objectfile" for sorting, achieving 741 // both grouping by library and sorting of objects within a library, 742 // at once. 743 std::string key1 = 744 (sc1->file->parentName + "/" + sc1->file->getName()).str(); 745 std::string key2 = 746 (sc2->file->parentName + "/" + sc2->file->getName()).str(); 747 return key1 < key2; 748 }); 749 } 750 return hasIdata; 751 } 752 753 // Add generated idata chunks, for imported symbols and DLLs, and a 754 // terminator in .idata$2. 755 void Writer::addSyntheticIdata() { 756 uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ; 757 idata.create(); 758 759 // Add the .idata content in the right section groups, to allow 760 // chunks from other linked in object files to be grouped together. 761 // See Microsoft PE/COFF spec 5.4 for details. 762 auto add = [&](StringRef n, std::vector<Chunk *> &v) { 763 PartialSection *pSec = createPartialSection(n, rdata); 764 pSec->chunks.insert(pSec->chunks.end(), v.begin(), v.end()); 765 }; 766 767 // The loader assumes a specific order of data. 768 // Add each type in the correct order. 769 add(".idata$2", idata.dirs); 770 add(".idata$4", idata.lookups); 771 add(".idata$5", idata.addresses); 772 if (!idata.hints.empty()) 773 add(".idata$6", idata.hints); 774 add(".idata$7", idata.dllNames); 775 } 776 777 // Locate the first Chunk and size of the import directory list and the 778 // IAT. 779 void Writer::locateImportTables() { 780 uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ; 781 782 if (PartialSection *importDirs = findPartialSection(".idata$2", rdata)) { 783 if (!importDirs->chunks.empty()) 784 importTableStart = importDirs->chunks.front(); 785 for (Chunk *c : importDirs->chunks) 786 importTableSize += c->getSize(); 787 } 788 789 if (PartialSection *importAddresses = findPartialSection(".idata$5", rdata)) { 790 if (!importAddresses->chunks.empty()) 791 iatStart = importAddresses->chunks.front(); 792 for (Chunk *c : importAddresses->chunks) 793 iatSize += c->getSize(); 794 } 795 } 796 797 // Return whether a SectionChunk's suffix (the dollar and any trailing 798 // suffix) should be removed and sorted into the main suffixless 799 // PartialSection. 800 static bool shouldStripSectionSuffix(SectionChunk *sc, StringRef name) { 801 // On MinGW, comdat groups are formed by putting the comdat group name 802 // after the '$' in the section name. For .eh_frame$<symbol>, that must 803 // still be sorted before the .eh_frame trailer from crtend.o, thus just 804 // strip the section name trailer. For other sections, such as 805 // .tls$$<symbol> (where non-comdat .tls symbols are otherwise stored in 806 // ".tls$"), they must be strictly sorted after .tls. And for the 807 // hypothetical case of comdat .CRT$XCU, we definitely need to keep the 808 // suffix for sorting. Thus, to play it safe, only strip the suffix for 809 // the standard sections. 810 if (!config->mingw) 811 return false; 812 if (!sc || !sc->isCOMDAT()) 813 return false; 814 return name.startswith(".text$") || name.startswith(".data$") || 815 name.startswith(".rdata$") || name.startswith(".pdata$") || 816 name.startswith(".xdata$") || name.startswith(".eh_frame$"); 817 } 818 819 void Writer::sortSections() { 820 if (!config->callGraphProfile.empty()) { 821 DenseMap<const SectionChunk *, int> order = computeCallGraphProfileOrder(); 822 for (auto it : order) { 823 if (DefinedRegular *sym = it.first->sym) 824 config->order[sym->getName()] = it.second; 825 } 826 } 827 if (!config->order.empty()) 828 for (auto it : partialSections) 829 sortBySectionOrder(it.second->chunks); 830 } 831 832 // Create output section objects and add them to OutputSections. 833 void Writer::createSections() { 834 // First, create the builtin sections. 835 const uint32_t data = IMAGE_SCN_CNT_INITIALIZED_DATA; 836 const uint32_t bss = IMAGE_SCN_CNT_UNINITIALIZED_DATA; 837 const uint32_t code = IMAGE_SCN_CNT_CODE; 838 const uint32_t discardable = IMAGE_SCN_MEM_DISCARDABLE; 839 const uint32_t r = IMAGE_SCN_MEM_READ; 840 const uint32_t w = IMAGE_SCN_MEM_WRITE; 841 const uint32_t x = IMAGE_SCN_MEM_EXECUTE; 842 843 SmallDenseMap<std::pair<StringRef, uint32_t>, OutputSection *> sections; 844 auto createSection = [&](StringRef name, uint32_t outChars) { 845 OutputSection *&sec = sections[{name, outChars}]; 846 if (!sec) { 847 sec = make<OutputSection>(name, outChars); 848 outputSections.push_back(sec); 849 } 850 return sec; 851 }; 852 853 // Try to match the section order used by link.exe. 854 textSec = createSection(".text", code | r | x); 855 createSection(".bss", bss | r | w); 856 rdataSec = createSection(".rdata", data | r); 857 buildidSec = createSection(".buildid", data | r); 858 dataSec = createSection(".data", data | r | w); 859 pdataSec = createSection(".pdata", data | r); 860 idataSec = createSection(".idata", data | r); 861 edataSec = createSection(".edata", data | r); 862 didatSec = createSection(".didat", data | r); 863 rsrcSec = createSection(".rsrc", data | r); 864 relocSec = createSection(".reloc", data | discardable | r); 865 ctorsSec = createSection(".ctors", data | r | w); 866 dtorsSec = createSection(".dtors", data | r | w); 867 868 // Then bin chunks by name and output characteristics. 869 for (Chunk *c : symtab->getChunks()) { 870 auto *sc = dyn_cast<SectionChunk>(c); 871 if (sc && !sc->live) { 872 if (config->verbose) 873 sc->printDiscardedMessage(); 874 continue; 875 } 876 StringRef name = c->getSectionName(); 877 if (shouldStripSectionSuffix(sc, name)) 878 name = name.split('$').first; 879 880 if (name.startswith(".tls")) 881 tlsAlignment = std::max(tlsAlignment, c->getAlignment()); 882 883 PartialSection *pSec = createPartialSection(name, 884 c->getOutputCharacteristics()); 885 pSec->chunks.push_back(c); 886 } 887 888 fixPartialSectionChars(".rsrc", data | r); 889 fixPartialSectionChars(".edata", data | r); 890 // Even in non MinGW cases, we might need to link against GNU import 891 // libraries. 892 bool hasIdata = fixGnuImportChunks(); 893 if (!idata.empty()) 894 hasIdata = true; 895 896 if (hasIdata) 897 addSyntheticIdata(); 898 899 sortSections(); 900 901 if (hasIdata) 902 locateImportTables(); 903 904 // Then create an OutputSection for each section. 905 // '$' and all following characters in input section names are 906 // discarded when determining output section. So, .text$foo 907 // contributes to .text, for example. See PE/COFF spec 3.2. 908 for (auto it : partialSections) { 909 PartialSection *pSec = it.second; 910 StringRef name = getOutputSectionName(pSec->name); 911 uint32_t outChars = pSec->characteristics; 912 913 if (name == ".CRT") { 914 // In link.exe, there is a special case for the I386 target where .CRT 915 // sections are treated as if they have output characteristics DATA | R if 916 // their characteristics are DATA | R | W. This implements the same 917 // special case for all architectures. 918 outChars = data | r; 919 920 log("Processing section " + pSec->name + " -> " + name); 921 922 sortCRTSectionChunks(pSec->chunks); 923 } 924 925 OutputSection *sec = createSection(name, outChars); 926 for (Chunk *c : pSec->chunks) 927 sec->addChunk(c); 928 929 sec->addContributingPartialSection(pSec); 930 } 931 932 // Finally, move some output sections to the end. 933 auto sectionOrder = [&](const OutputSection *s) { 934 // Move DISCARDABLE (or non-memory-mapped) sections to the end of file 935 // because the loader cannot handle holes. Stripping can remove other 936 // discardable ones than .reloc, which is first of them (created early). 937 if (s->header.Characteristics & IMAGE_SCN_MEM_DISCARDABLE) 938 return 2; 939 // .rsrc should come at the end of the non-discardable sections because its 940 // size may change by the Win32 UpdateResources() function, causing 941 // subsequent sections to move (see https://crbug.com/827082). 942 if (s == rsrcSec) 943 return 1; 944 return 0; 945 }; 946 llvm::stable_sort(outputSections, 947 [&](const OutputSection *s, const OutputSection *t) { 948 return sectionOrder(s) < sectionOrder(t); 949 }); 950 } 951 952 void Writer::createMiscChunks() { 953 for (MergeChunk *p : MergeChunk::instances) { 954 if (p) { 955 p->finalizeContents(); 956 rdataSec->addChunk(p); 957 } 958 } 959 960 // Create thunks for locally-dllimported symbols. 961 if (!symtab->localImportChunks.empty()) { 962 for (Chunk *c : symtab->localImportChunks) 963 rdataSec->addChunk(c); 964 } 965 966 // Create Debug Information Chunks 967 OutputSection *debugInfoSec = config->mingw ? buildidSec : rdataSec; 968 if (config->debug || config->repro || config->cetCompat) { 969 debugDirectory = make<DebugDirectoryChunk>(debugRecords, config->repro); 970 debugDirectory->setAlignment(4); 971 debugInfoSec->addChunk(debugDirectory); 972 } 973 974 if (config->debug) { 975 // Make a CVDebugRecordChunk even when /DEBUG:CV is not specified. We 976 // output a PDB no matter what, and this chunk provides the only means of 977 // allowing a debugger to match a PDB and an executable. So we need it even 978 // if we're ultimately not going to write CodeView data to the PDB. 979 buildId = make<CVDebugRecordChunk>(); 980 debugRecords.push_back({COFF::IMAGE_DEBUG_TYPE_CODEVIEW, buildId}); 981 } 982 983 if (config->cetCompat) { 984 debugRecords.push_back({COFF::IMAGE_DEBUG_TYPE_EX_DLLCHARACTERISTICS, 985 make<ExtendedDllCharacteristicsChunk>( 986 IMAGE_DLL_CHARACTERISTICS_EX_CET_COMPAT)}); 987 } 988 989 // Align and add each chunk referenced by the debug data directory. 990 for (std::pair<COFF::DebugType, Chunk *> r : debugRecords) { 991 r.second->setAlignment(4); 992 debugInfoSec->addChunk(r.second); 993 } 994 995 // Create SEH table. x86-only. 996 if (config->safeSEH) 997 createSEHTable(); 998 999 // Create /guard:cf tables if requested. 1000 if (config->guardCF != GuardCFLevel::Off) 1001 createGuardCFTables(); 1002 1003 if (config->autoImport) 1004 createRuntimePseudoRelocs(); 1005 1006 if (config->mingw) 1007 insertCtorDtorSymbols(); 1008 } 1009 1010 // Create .idata section for the DLL-imported symbol table. 1011 // The format of this section is inherently Windows-specific. 1012 // IdataContents class abstracted away the details for us, 1013 // so we just let it create chunks and add them to the section. 1014 void Writer::createImportTables() { 1015 // Initialize DLLOrder so that import entries are ordered in 1016 // the same order as in the command line. (That affects DLL 1017 // initialization order, and this ordering is MSVC-compatible.) 1018 for (ImportFile *file : ImportFile::instances) { 1019 if (!file->live) 1020 continue; 1021 1022 std::string dll = StringRef(file->dllName).lower(); 1023 if (config->dllOrder.count(dll) == 0) 1024 config->dllOrder[dll] = config->dllOrder.size(); 1025 1026 if (file->impSym && !isa<DefinedImportData>(file->impSym)) 1027 fatal(toString(*file->impSym) + " was replaced"); 1028 DefinedImportData *impSym = cast_or_null<DefinedImportData>(file->impSym); 1029 if (config->delayLoads.count(StringRef(file->dllName).lower())) { 1030 if (!file->thunkSym) 1031 fatal("cannot delay-load " + toString(file) + 1032 " due to import of data: " + toString(*impSym)); 1033 delayIdata.add(impSym); 1034 } else { 1035 idata.add(impSym); 1036 } 1037 } 1038 } 1039 1040 void Writer::appendImportThunks() { 1041 if (ImportFile::instances.empty()) 1042 return; 1043 1044 for (ImportFile *file : ImportFile::instances) { 1045 if (!file->live) 1046 continue; 1047 1048 if (!file->thunkSym) 1049 continue; 1050 1051 if (!isa<DefinedImportThunk>(file->thunkSym)) 1052 fatal(toString(*file->thunkSym) + " was replaced"); 1053 DefinedImportThunk *thunk = cast<DefinedImportThunk>(file->thunkSym); 1054 if (file->thunkLive) 1055 textSec->addChunk(thunk->getChunk()); 1056 } 1057 1058 if (!delayIdata.empty()) { 1059 Defined *helper = cast<Defined>(config->delayLoadHelper); 1060 delayIdata.create(helper); 1061 for (Chunk *c : delayIdata.getChunks()) 1062 didatSec->addChunk(c); 1063 for (Chunk *c : delayIdata.getDataChunks()) 1064 dataSec->addChunk(c); 1065 for (Chunk *c : delayIdata.getCodeChunks()) 1066 textSec->addChunk(c); 1067 } 1068 } 1069 1070 void Writer::createExportTable() { 1071 if (!edataSec->chunks.empty()) { 1072 // Allow using a custom built export table from input object files, instead 1073 // of having the linker synthesize the tables. 1074 if (config->hadExplicitExports) 1075 warn("literal .edata sections override exports"); 1076 } else if (!config->exports.empty()) { 1077 for (Chunk *c : edata.chunks) 1078 edataSec->addChunk(c); 1079 } 1080 if (!edataSec->chunks.empty()) { 1081 edataStart = edataSec->chunks.front(); 1082 edataEnd = edataSec->chunks.back(); 1083 } 1084 } 1085 1086 void Writer::removeUnusedSections() { 1087 // Remove sections that we can be sure won't get content, to avoid 1088 // allocating space for their section headers. 1089 auto isUnused = [this](OutputSection *s) { 1090 if (s == relocSec) 1091 return false; // This section is populated later. 1092 // MergeChunks have zero size at this point, as their size is finalized 1093 // later. Only remove sections that have no Chunks at all. 1094 return s->chunks.empty(); 1095 }; 1096 outputSections.erase( 1097 std::remove_if(outputSections.begin(), outputSections.end(), isUnused), 1098 outputSections.end()); 1099 } 1100 1101 // The Windows loader doesn't seem to like empty sections, 1102 // so we remove them if any. 1103 void Writer::removeEmptySections() { 1104 auto isEmpty = [](OutputSection *s) { return s->getVirtualSize() == 0; }; 1105 outputSections.erase( 1106 std::remove_if(outputSections.begin(), outputSections.end(), isEmpty), 1107 outputSections.end()); 1108 } 1109 1110 void Writer::assignOutputSectionIndices() { 1111 // Assign final output section indices, and assign each chunk to its output 1112 // section. 1113 uint32_t idx = 1; 1114 for (OutputSection *os : outputSections) { 1115 os->sectionIndex = idx; 1116 for (Chunk *c : os->chunks) 1117 c->setOutputSectionIdx(idx); 1118 ++idx; 1119 } 1120 1121 // Merge chunks are containers of chunks, so assign those an output section 1122 // too. 1123 for (MergeChunk *mc : MergeChunk::instances) 1124 if (mc) 1125 for (SectionChunk *sc : mc->sections) 1126 if (sc && sc->live) 1127 sc->setOutputSectionIdx(mc->getOutputSectionIdx()); 1128 } 1129 1130 size_t Writer::addEntryToStringTable(StringRef str) { 1131 assert(str.size() > COFF::NameSize); 1132 size_t offsetOfEntry = strtab.size() + 4; // +4 for the size field 1133 strtab.insert(strtab.end(), str.begin(), str.end()); 1134 strtab.push_back('\0'); 1135 return offsetOfEntry; 1136 } 1137 1138 Optional<coff_symbol16> Writer::createSymbol(Defined *def) { 1139 coff_symbol16 sym; 1140 switch (def->kind()) { 1141 case Symbol::DefinedAbsoluteKind: 1142 sym.Value = def->getRVA(); 1143 sym.SectionNumber = IMAGE_SYM_ABSOLUTE; 1144 break; 1145 case Symbol::DefinedSyntheticKind: 1146 // Relative symbols are unrepresentable in a COFF symbol table. 1147 return None; 1148 default: { 1149 // Don't write symbols that won't be written to the output to the symbol 1150 // table. 1151 Chunk *c = def->getChunk(); 1152 if (!c) 1153 return None; 1154 OutputSection *os = c->getOutputSection(); 1155 if (!os) 1156 return None; 1157 1158 sym.Value = def->getRVA() - os->getRVA(); 1159 sym.SectionNumber = os->sectionIndex; 1160 break; 1161 } 1162 } 1163 1164 // Symbols that are runtime pseudo relocations don't point to the actual 1165 // symbol data itself (as they are imported), but points to the IAT entry 1166 // instead. Avoid emitting them to the symbol table, as they can confuse 1167 // debuggers. 1168 if (def->isRuntimePseudoReloc) 1169 return None; 1170 1171 StringRef name = def->getName(); 1172 if (name.size() > COFF::NameSize) { 1173 sym.Name.Offset.Zeroes = 0; 1174 sym.Name.Offset.Offset = addEntryToStringTable(name); 1175 } else { 1176 memset(sym.Name.ShortName, 0, COFF::NameSize); 1177 memcpy(sym.Name.ShortName, name.data(), name.size()); 1178 } 1179 1180 if (auto *d = dyn_cast<DefinedCOFF>(def)) { 1181 COFFSymbolRef ref = d->getCOFFSymbol(); 1182 sym.Type = ref.getType(); 1183 sym.StorageClass = ref.getStorageClass(); 1184 } else { 1185 sym.Type = IMAGE_SYM_TYPE_NULL; 1186 sym.StorageClass = IMAGE_SYM_CLASS_EXTERNAL; 1187 } 1188 sym.NumberOfAuxSymbols = 0; 1189 return sym; 1190 } 1191 1192 void Writer::createSymbolAndStringTable() { 1193 // PE/COFF images are limited to 8 byte section names. Longer names can be 1194 // supported by writing a non-standard string table, but this string table is 1195 // not mapped at runtime and the long names will therefore be inaccessible. 1196 // link.exe always truncates section names to 8 bytes, whereas binutils always 1197 // preserves long section names via the string table. LLD adopts a hybrid 1198 // solution where discardable sections have long names preserved and 1199 // non-discardable sections have their names truncated, to ensure that any 1200 // section which is mapped at runtime also has its name mapped at runtime. 1201 for (OutputSection *sec : outputSections) { 1202 if (sec->name.size() <= COFF::NameSize) 1203 continue; 1204 if ((sec->header.Characteristics & IMAGE_SCN_MEM_DISCARDABLE) == 0) 1205 continue; 1206 if (config->warnLongSectionNames) { 1207 warn("section name " + sec->name + 1208 " is longer than 8 characters and will use a non-standard string " 1209 "table"); 1210 } 1211 sec->setStringTableOff(addEntryToStringTable(sec->name)); 1212 } 1213 1214 if (config->debugDwarf || config->debugSymtab) { 1215 for (ObjFile *file : ObjFile::instances) { 1216 for (Symbol *b : file->getSymbols()) { 1217 auto *d = dyn_cast_or_null<Defined>(b); 1218 if (!d || d->writtenToSymtab) 1219 continue; 1220 d->writtenToSymtab = true; 1221 1222 if (Optional<coff_symbol16> sym = createSymbol(d)) 1223 outputSymtab.push_back(*sym); 1224 } 1225 } 1226 } 1227 1228 if (outputSymtab.empty() && strtab.empty()) 1229 return; 1230 1231 // We position the symbol table to be adjacent to the end of the last section. 1232 uint64_t fileOff = fileSize; 1233 pointerToSymbolTable = fileOff; 1234 fileOff += outputSymtab.size() * sizeof(coff_symbol16); 1235 fileOff += 4 + strtab.size(); 1236 fileSize = alignTo(fileOff, config->fileAlign); 1237 } 1238 1239 void Writer::mergeSections() { 1240 if (!pdataSec->chunks.empty()) { 1241 firstPdata = pdataSec->chunks.front(); 1242 lastPdata = pdataSec->chunks.back(); 1243 } 1244 1245 for (auto &p : config->merge) { 1246 StringRef toName = p.second; 1247 if (p.first == toName) 1248 continue; 1249 StringSet<> names; 1250 while (1) { 1251 if (!names.insert(toName).second) 1252 fatal("/merge: cycle found for section '" + p.first + "'"); 1253 auto i = config->merge.find(toName); 1254 if (i == config->merge.end()) 1255 break; 1256 toName = i->second; 1257 } 1258 OutputSection *from = findSection(p.first); 1259 OutputSection *to = findSection(toName); 1260 if (!from) 1261 continue; 1262 if (!to) { 1263 from->name = toName; 1264 continue; 1265 } 1266 to->merge(from); 1267 } 1268 } 1269 1270 // Visits all sections to assign incremental, non-overlapping RVAs and 1271 // file offsets. 1272 void Writer::assignAddresses() { 1273 sizeOfHeaders = dosStubSize + sizeof(PEMagic) + sizeof(coff_file_header) + 1274 sizeof(data_directory) * numberOfDataDirectory + 1275 sizeof(coff_section) * outputSections.size(); 1276 sizeOfHeaders += 1277 config->is64() ? sizeof(pe32plus_header) : sizeof(pe32_header); 1278 sizeOfHeaders = alignTo(sizeOfHeaders, config->fileAlign); 1279 fileSize = sizeOfHeaders; 1280 1281 // The first page is kept unmapped. 1282 uint64_t rva = alignTo(sizeOfHeaders, config->align); 1283 1284 for (OutputSection *sec : outputSections) { 1285 if (sec == relocSec) 1286 addBaserels(); 1287 uint64_t rawSize = 0, virtualSize = 0; 1288 sec->header.VirtualAddress = rva; 1289 1290 // If /FUNCTIONPADMIN is used, functions are padded in order to create a 1291 // hotpatchable image. 1292 const bool isCodeSection = 1293 (sec->header.Characteristics & IMAGE_SCN_CNT_CODE) && 1294 (sec->header.Characteristics & IMAGE_SCN_MEM_READ) && 1295 (sec->header.Characteristics & IMAGE_SCN_MEM_EXECUTE); 1296 uint32_t padding = isCodeSection ? config->functionPadMin : 0; 1297 1298 for (Chunk *c : sec->chunks) { 1299 if (padding && c->isHotPatchable()) 1300 virtualSize += padding; 1301 virtualSize = alignTo(virtualSize, c->getAlignment()); 1302 c->setRVA(rva + virtualSize); 1303 virtualSize += c->getSize(); 1304 if (c->hasData) 1305 rawSize = alignTo(virtualSize, config->fileAlign); 1306 } 1307 if (virtualSize > UINT32_MAX) 1308 error("section larger than 4 GiB: " + sec->name); 1309 sec->header.VirtualSize = virtualSize; 1310 sec->header.SizeOfRawData = rawSize; 1311 if (rawSize != 0) 1312 sec->header.PointerToRawData = fileSize; 1313 rva += alignTo(virtualSize, config->align); 1314 fileSize += alignTo(rawSize, config->fileAlign); 1315 } 1316 sizeOfImage = alignTo(rva, config->align); 1317 1318 // Assign addresses to sections in MergeChunks. 1319 for (MergeChunk *mc : MergeChunk::instances) 1320 if (mc) 1321 mc->assignSubsectionRVAs(); 1322 } 1323 1324 template <typename PEHeaderTy> void Writer::writeHeader() { 1325 // Write DOS header. For backwards compatibility, the first part of a PE/COFF 1326 // executable consists of an MS-DOS MZ executable. If the executable is run 1327 // under DOS, that program gets run (usually to just print an error message). 1328 // When run under Windows, the loader looks at AddressOfNewExeHeader and uses 1329 // the PE header instead. 1330 uint8_t *buf = buffer->getBufferStart(); 1331 auto *dos = reinterpret_cast<dos_header *>(buf); 1332 buf += sizeof(dos_header); 1333 dos->Magic[0] = 'M'; 1334 dos->Magic[1] = 'Z'; 1335 dos->UsedBytesInTheLastPage = dosStubSize % 512; 1336 dos->FileSizeInPages = divideCeil(dosStubSize, 512); 1337 dos->HeaderSizeInParagraphs = sizeof(dos_header) / 16; 1338 1339 dos->AddressOfRelocationTable = sizeof(dos_header); 1340 dos->AddressOfNewExeHeader = dosStubSize; 1341 1342 // Write DOS program. 1343 memcpy(buf, dosProgram, sizeof(dosProgram)); 1344 buf += sizeof(dosProgram); 1345 1346 // Write PE magic 1347 memcpy(buf, PEMagic, sizeof(PEMagic)); 1348 buf += sizeof(PEMagic); 1349 1350 // Write COFF header 1351 auto *coff = reinterpret_cast<coff_file_header *>(buf); 1352 buf += sizeof(*coff); 1353 coff->Machine = config->machine; 1354 coff->NumberOfSections = outputSections.size(); 1355 coff->Characteristics = IMAGE_FILE_EXECUTABLE_IMAGE; 1356 if (config->largeAddressAware) 1357 coff->Characteristics |= IMAGE_FILE_LARGE_ADDRESS_AWARE; 1358 if (!config->is64()) 1359 coff->Characteristics |= IMAGE_FILE_32BIT_MACHINE; 1360 if (config->dll) 1361 coff->Characteristics |= IMAGE_FILE_DLL; 1362 if (config->driverUponly) 1363 coff->Characteristics |= IMAGE_FILE_UP_SYSTEM_ONLY; 1364 if (!config->relocatable) 1365 coff->Characteristics |= IMAGE_FILE_RELOCS_STRIPPED; 1366 if (config->swaprunCD) 1367 coff->Characteristics |= IMAGE_FILE_REMOVABLE_RUN_FROM_SWAP; 1368 if (config->swaprunNet) 1369 coff->Characteristics |= IMAGE_FILE_NET_RUN_FROM_SWAP; 1370 coff->SizeOfOptionalHeader = 1371 sizeof(PEHeaderTy) + sizeof(data_directory) * numberOfDataDirectory; 1372 1373 // Write PE header 1374 auto *pe = reinterpret_cast<PEHeaderTy *>(buf); 1375 buf += sizeof(*pe); 1376 pe->Magic = config->is64() ? PE32Header::PE32_PLUS : PE32Header::PE32; 1377 1378 // If {Major,Minor}LinkerVersion is left at 0.0, then for some 1379 // reason signing the resulting PE file with Authenticode produces a 1380 // signature that fails to validate on Windows 7 (but is OK on 10). 1381 // Set it to 14.0, which is what VS2015 outputs, and which avoids 1382 // that problem. 1383 pe->MajorLinkerVersion = 14; 1384 pe->MinorLinkerVersion = 0; 1385 1386 pe->ImageBase = config->imageBase; 1387 pe->SectionAlignment = config->align; 1388 pe->FileAlignment = config->fileAlign; 1389 pe->MajorImageVersion = config->majorImageVersion; 1390 pe->MinorImageVersion = config->minorImageVersion; 1391 pe->MajorOperatingSystemVersion = config->majorOSVersion; 1392 pe->MinorOperatingSystemVersion = config->minorOSVersion; 1393 pe->MajorSubsystemVersion = config->majorSubsystemVersion; 1394 pe->MinorSubsystemVersion = config->minorSubsystemVersion; 1395 pe->Subsystem = config->subsystem; 1396 pe->SizeOfImage = sizeOfImage; 1397 pe->SizeOfHeaders = sizeOfHeaders; 1398 if (!config->noEntry) { 1399 Defined *entry = cast<Defined>(config->entry); 1400 pe->AddressOfEntryPoint = entry->getRVA(); 1401 // Pointer to thumb code must have the LSB set, so adjust it. 1402 if (config->machine == ARMNT) 1403 pe->AddressOfEntryPoint |= 1; 1404 } 1405 pe->SizeOfStackReserve = config->stackReserve; 1406 pe->SizeOfStackCommit = config->stackCommit; 1407 pe->SizeOfHeapReserve = config->heapReserve; 1408 pe->SizeOfHeapCommit = config->heapCommit; 1409 if (config->appContainer) 1410 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_APPCONTAINER; 1411 if (config->driverWdm) 1412 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_WDM_DRIVER; 1413 if (config->dynamicBase) 1414 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_DYNAMIC_BASE; 1415 if (config->highEntropyVA) 1416 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_HIGH_ENTROPY_VA; 1417 if (!config->allowBind) 1418 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NO_BIND; 1419 if (config->nxCompat) 1420 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NX_COMPAT; 1421 if (!config->allowIsolation) 1422 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NO_ISOLATION; 1423 if (config->guardCF != GuardCFLevel::Off) 1424 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_GUARD_CF; 1425 if (config->integrityCheck) 1426 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_FORCE_INTEGRITY; 1427 if (setNoSEHCharacteristic || config->noSEH) 1428 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NO_SEH; 1429 if (config->terminalServerAware) 1430 pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_TERMINAL_SERVER_AWARE; 1431 pe->NumberOfRvaAndSize = numberOfDataDirectory; 1432 if (textSec->getVirtualSize()) { 1433 pe->BaseOfCode = textSec->getRVA(); 1434 pe->SizeOfCode = textSec->getRawSize(); 1435 } 1436 pe->SizeOfInitializedData = getSizeOfInitializedData(); 1437 1438 // Write data directory 1439 auto *dir = reinterpret_cast<data_directory *>(buf); 1440 buf += sizeof(*dir) * numberOfDataDirectory; 1441 if (edataStart) { 1442 dir[EXPORT_TABLE].RelativeVirtualAddress = edataStart->getRVA(); 1443 dir[EXPORT_TABLE].Size = 1444 edataEnd->getRVA() + edataEnd->getSize() - edataStart->getRVA(); 1445 } 1446 if (importTableStart) { 1447 dir[IMPORT_TABLE].RelativeVirtualAddress = importTableStart->getRVA(); 1448 dir[IMPORT_TABLE].Size = importTableSize; 1449 } 1450 if (iatStart) { 1451 dir[IAT].RelativeVirtualAddress = iatStart->getRVA(); 1452 dir[IAT].Size = iatSize; 1453 } 1454 if (rsrcSec->getVirtualSize()) { 1455 dir[RESOURCE_TABLE].RelativeVirtualAddress = rsrcSec->getRVA(); 1456 dir[RESOURCE_TABLE].Size = rsrcSec->getVirtualSize(); 1457 } 1458 if (firstPdata) { 1459 dir[EXCEPTION_TABLE].RelativeVirtualAddress = firstPdata->getRVA(); 1460 dir[EXCEPTION_TABLE].Size = 1461 lastPdata->getRVA() + lastPdata->getSize() - firstPdata->getRVA(); 1462 } 1463 if (relocSec->getVirtualSize()) { 1464 dir[BASE_RELOCATION_TABLE].RelativeVirtualAddress = relocSec->getRVA(); 1465 dir[BASE_RELOCATION_TABLE].Size = relocSec->getVirtualSize(); 1466 } 1467 if (Symbol *sym = symtab->findUnderscore("_tls_used")) { 1468 if (Defined *b = dyn_cast<Defined>(sym)) { 1469 dir[TLS_TABLE].RelativeVirtualAddress = b->getRVA(); 1470 dir[TLS_TABLE].Size = config->is64() 1471 ? sizeof(object::coff_tls_directory64) 1472 : sizeof(object::coff_tls_directory32); 1473 } 1474 } 1475 if (debugDirectory) { 1476 dir[DEBUG_DIRECTORY].RelativeVirtualAddress = debugDirectory->getRVA(); 1477 dir[DEBUG_DIRECTORY].Size = debugDirectory->getSize(); 1478 } 1479 if (Symbol *sym = symtab->findUnderscore("_load_config_used")) { 1480 if (auto *b = dyn_cast<DefinedRegular>(sym)) { 1481 SectionChunk *sc = b->getChunk(); 1482 assert(b->getRVA() >= sc->getRVA()); 1483 uint64_t offsetInChunk = b->getRVA() - sc->getRVA(); 1484 if (!sc->hasData || offsetInChunk + 4 > sc->getSize()) 1485 fatal("_load_config_used is malformed"); 1486 1487 ArrayRef<uint8_t> secContents = sc->getContents(); 1488 uint32_t loadConfigSize = 1489 *reinterpret_cast<const ulittle32_t *>(&secContents[offsetInChunk]); 1490 if (offsetInChunk + loadConfigSize > sc->getSize()) 1491 fatal("_load_config_used is too large"); 1492 dir[LOAD_CONFIG_TABLE].RelativeVirtualAddress = b->getRVA(); 1493 dir[LOAD_CONFIG_TABLE].Size = loadConfigSize; 1494 } 1495 } 1496 if (!delayIdata.empty()) { 1497 dir[DELAY_IMPORT_DESCRIPTOR].RelativeVirtualAddress = 1498 delayIdata.getDirRVA(); 1499 dir[DELAY_IMPORT_DESCRIPTOR].Size = delayIdata.getDirSize(); 1500 } 1501 1502 // Write section table 1503 for (OutputSection *sec : outputSections) { 1504 sec->writeHeaderTo(buf); 1505 buf += sizeof(coff_section); 1506 } 1507 sectionTable = ArrayRef<uint8_t>( 1508 buf - outputSections.size() * sizeof(coff_section), buf); 1509 1510 if (outputSymtab.empty() && strtab.empty()) 1511 return; 1512 1513 coff->PointerToSymbolTable = pointerToSymbolTable; 1514 uint32_t numberOfSymbols = outputSymtab.size(); 1515 coff->NumberOfSymbols = numberOfSymbols; 1516 auto *symbolTable = reinterpret_cast<coff_symbol16 *>( 1517 buffer->getBufferStart() + coff->PointerToSymbolTable); 1518 for (size_t i = 0; i != numberOfSymbols; ++i) 1519 symbolTable[i] = outputSymtab[i]; 1520 // Create the string table, it follows immediately after the symbol table. 1521 // The first 4 bytes is length including itself. 1522 buf = reinterpret_cast<uint8_t *>(&symbolTable[numberOfSymbols]); 1523 write32le(buf, strtab.size() + 4); 1524 if (!strtab.empty()) 1525 memcpy(buf + 4, strtab.data(), strtab.size()); 1526 } 1527 1528 void Writer::openFile(StringRef path) { 1529 buffer = CHECK( 1530 FileOutputBuffer::create(path, fileSize, FileOutputBuffer::F_executable), 1531 "failed to open " + path); 1532 } 1533 1534 void Writer::createSEHTable() { 1535 SymbolRVASet handlers; 1536 for (ObjFile *file : ObjFile::instances) { 1537 if (!file->hasSafeSEH()) 1538 error("/safeseh: " + file->getName() + " is not compatible with SEH"); 1539 markSymbolsForRVATable(file, file->getSXDataChunks(), handlers); 1540 } 1541 1542 // Set the "no SEH" characteristic if there really were no handlers, or if 1543 // there is no load config object to point to the table of handlers. 1544 setNoSEHCharacteristic = 1545 handlers.empty() || !symtab->findUnderscore("_load_config_used"); 1546 1547 maybeAddRVATable(std::move(handlers), "__safe_se_handler_table", 1548 "__safe_se_handler_count"); 1549 } 1550 1551 // Add a symbol to an RVA set. Two symbols may have the same RVA, but an RVA set 1552 // cannot contain duplicates. Therefore, the set is uniqued by Chunk and the 1553 // symbol's offset into that Chunk. 1554 static void addSymbolToRVASet(SymbolRVASet &rvaSet, Defined *s) { 1555 Chunk *c = s->getChunk(); 1556 if (auto *sc = dyn_cast<SectionChunk>(c)) 1557 c = sc->repl; // Look through ICF replacement. 1558 uint32_t off = s->getRVA() - (c ? c->getRVA() : 0); 1559 rvaSet.insert({c, off}); 1560 } 1561 1562 // Given a symbol, add it to the GFIDs table if it is a live, defined, function 1563 // symbol in an executable section. 1564 static void maybeAddAddressTakenFunction(SymbolRVASet &addressTakenSyms, 1565 Symbol *s) { 1566 if (!s) 1567 return; 1568 1569 switch (s->kind()) { 1570 case Symbol::DefinedLocalImportKind: 1571 case Symbol::DefinedImportDataKind: 1572 // Defines an __imp_ pointer, so it is data, so it is ignored. 1573 break; 1574 case Symbol::DefinedCommonKind: 1575 // Common is always data, so it is ignored. 1576 break; 1577 case Symbol::DefinedAbsoluteKind: 1578 case Symbol::DefinedSyntheticKind: 1579 // Absolute is never code, synthetic generally isn't and usually isn't 1580 // determinable. 1581 break; 1582 case Symbol::LazyArchiveKind: 1583 case Symbol::LazyObjectKind: 1584 case Symbol::UndefinedKind: 1585 // Undefined symbols resolve to zero, so they don't have an RVA. Lazy 1586 // symbols shouldn't have relocations. 1587 break; 1588 1589 case Symbol::DefinedImportThunkKind: 1590 // Thunks are always code, include them. 1591 addSymbolToRVASet(addressTakenSyms, cast<Defined>(s)); 1592 break; 1593 1594 case Symbol::DefinedRegularKind: { 1595 // This is a regular, defined, symbol from a COFF file. Mark the symbol as 1596 // address taken if the symbol type is function and it's in an executable 1597 // section. 1598 auto *d = cast<DefinedRegular>(s); 1599 if (d->getCOFFSymbol().getComplexType() == COFF::IMAGE_SYM_DTYPE_FUNCTION) { 1600 SectionChunk *sc = dyn_cast<SectionChunk>(d->getChunk()); 1601 if (sc && sc->live && 1602 sc->getOutputCharacteristics() & IMAGE_SCN_MEM_EXECUTE) 1603 addSymbolToRVASet(addressTakenSyms, d); 1604 } 1605 break; 1606 } 1607 } 1608 } 1609 1610 // Visit all relocations from all section contributions of this object file and 1611 // mark the relocation target as address-taken. 1612 static void markSymbolsWithRelocations(ObjFile *file, 1613 SymbolRVASet &usedSymbols) { 1614 for (Chunk *c : file->getChunks()) { 1615 // We only care about live section chunks. Common chunks and other chunks 1616 // don't generally contain relocations. 1617 SectionChunk *sc = dyn_cast<SectionChunk>(c); 1618 if (!sc || !sc->live) 1619 continue; 1620 1621 for (const coff_relocation &reloc : sc->getRelocs()) { 1622 if (config->machine == I386 && reloc.Type == COFF::IMAGE_REL_I386_REL32) 1623 // Ignore relative relocations on x86. On x86_64 they can't be ignored 1624 // since they're also used to compute absolute addresses. 1625 continue; 1626 1627 Symbol *ref = sc->file->getSymbol(reloc.SymbolTableIndex); 1628 maybeAddAddressTakenFunction(usedSymbols, ref); 1629 } 1630 } 1631 } 1632 1633 // Create the guard function id table. This is a table of RVAs of all 1634 // address-taken functions. It is sorted and uniqued, just like the safe SEH 1635 // table. 1636 void Writer::createGuardCFTables() { 1637 SymbolRVASet addressTakenSyms; 1638 SymbolRVASet giatsRVASet; 1639 std::vector<Symbol *> giatsSymbols; 1640 SymbolRVASet longJmpTargets; 1641 for (ObjFile *file : ObjFile::instances) { 1642 // If the object was compiled with /guard:cf, the address taken symbols 1643 // are in .gfids$y sections, and the longjmp targets are in .gljmp$y 1644 // sections. If the object was not compiled with /guard:cf, we assume there 1645 // were no setjmp targets, and that all code symbols with relocations are 1646 // possibly address-taken. 1647 if (file->hasGuardCF()) { 1648 markSymbolsForRVATable(file, file->getGuardFidChunks(), addressTakenSyms); 1649 markSymbolsForRVATable(file, file->getGuardIATChunks(), giatsRVASet); 1650 getSymbolsFromSections(file, file->getGuardIATChunks(), giatsSymbols); 1651 markSymbolsForRVATable(file, file->getGuardLJmpChunks(), longJmpTargets); 1652 } else { 1653 markSymbolsWithRelocations(file, addressTakenSyms); 1654 } 1655 } 1656 1657 // Mark the image entry as address-taken. 1658 if (config->entry) 1659 maybeAddAddressTakenFunction(addressTakenSyms, config->entry); 1660 1661 // Mark exported symbols in executable sections as address-taken. 1662 for (Export &e : config->exports) 1663 maybeAddAddressTakenFunction(addressTakenSyms, e.sym); 1664 1665 // For each entry in the .giats table, check if it has a corresponding load 1666 // thunk (e.g. because the DLL that defines it will be delay-loaded) and, if 1667 // so, add the load thunk to the address taken (.gfids) table. 1668 for (Symbol *s : giatsSymbols) { 1669 if (auto *di = dyn_cast<DefinedImportData>(s)) { 1670 if (di->loadThunkSym) 1671 addSymbolToRVASet(addressTakenSyms, di->loadThunkSym); 1672 } 1673 } 1674 1675 // Ensure sections referenced in the gfid table are 16-byte aligned. 1676 for (const ChunkAndOffset &c : addressTakenSyms) 1677 if (c.inputChunk->getAlignment() < 16) 1678 c.inputChunk->setAlignment(16); 1679 1680 maybeAddRVATable(std::move(addressTakenSyms), "__guard_fids_table", 1681 "__guard_fids_count"); 1682 1683 // Add the Guard Address Taken IAT Entry Table (.giats). 1684 maybeAddRVATable(std::move(giatsRVASet), "__guard_iat_table", 1685 "__guard_iat_count"); 1686 1687 // Add the longjmp target table unless the user told us not to. 1688 if (config->guardCF == GuardCFLevel::Full) 1689 maybeAddRVATable(std::move(longJmpTargets), "__guard_longjmp_table", 1690 "__guard_longjmp_count"); 1691 1692 // Set __guard_flags, which will be used in the load config to indicate that 1693 // /guard:cf was enabled. 1694 uint32_t guardFlags = uint32_t(coff_guard_flags::CFInstrumented) | 1695 uint32_t(coff_guard_flags::HasFidTable); 1696 if (config->guardCF == GuardCFLevel::Full) 1697 guardFlags |= uint32_t(coff_guard_flags::HasLongJmpTable); 1698 Symbol *flagSym = symtab->findUnderscore("__guard_flags"); 1699 cast<DefinedAbsolute>(flagSym)->setVA(guardFlags); 1700 } 1701 1702 // Take a list of input sections containing symbol table indices and add those 1703 // symbols to a vector. The challenge is that symbol RVAs are not known and 1704 // depend on the table size, so we can't directly build a set of integers. 1705 void Writer::getSymbolsFromSections(ObjFile *file, 1706 ArrayRef<SectionChunk *> symIdxChunks, 1707 std::vector<Symbol *> &symbols) { 1708 for (SectionChunk *c : symIdxChunks) { 1709 // Skip sections discarded by linker GC. This comes up when a .gfids section 1710 // is associated with something like a vtable and the vtable is discarded. 1711 // In this case, the associated gfids section is discarded, and we don't 1712 // mark the virtual member functions as address-taken by the vtable. 1713 if (!c->live) 1714 continue; 1715 1716 // Validate that the contents look like symbol table indices. 1717 ArrayRef<uint8_t> data = c->getContents(); 1718 if (data.size() % 4 != 0) { 1719 warn("ignoring " + c->getSectionName() + 1720 " symbol table index section in object " + toString(file)); 1721 continue; 1722 } 1723 1724 // Read each symbol table index and check if that symbol was included in the 1725 // final link. If so, add it to the vector of symbols. 1726 ArrayRef<ulittle32_t> symIndices( 1727 reinterpret_cast<const ulittle32_t *>(data.data()), data.size() / 4); 1728 ArrayRef<Symbol *> objSymbols = file->getSymbols(); 1729 for (uint32_t symIndex : symIndices) { 1730 if (symIndex >= objSymbols.size()) { 1731 warn("ignoring invalid symbol table index in section " + 1732 c->getSectionName() + " in object " + toString(file)); 1733 continue; 1734 } 1735 if (Symbol *s = objSymbols[symIndex]) { 1736 if (s->isLive()) 1737 symbols.push_back(cast<Symbol>(s)); 1738 } 1739 } 1740 } 1741 } 1742 1743 // Take a list of input sections containing symbol table indices and add those 1744 // symbols to an RVA table. 1745 void Writer::markSymbolsForRVATable(ObjFile *file, 1746 ArrayRef<SectionChunk *> symIdxChunks, 1747 SymbolRVASet &tableSymbols) { 1748 std::vector<Symbol *> syms; 1749 getSymbolsFromSections(file, symIdxChunks, syms); 1750 1751 for (Symbol *s : syms) 1752 addSymbolToRVASet(tableSymbols, cast<Defined>(s)); 1753 } 1754 1755 // Replace the absolute table symbol with a synthetic symbol pointing to 1756 // tableChunk so that we can emit base relocations for it and resolve section 1757 // relative relocations. 1758 void Writer::maybeAddRVATable(SymbolRVASet tableSymbols, StringRef tableSym, 1759 StringRef countSym) { 1760 if (tableSymbols.empty()) 1761 return; 1762 1763 RVATableChunk *tableChunk = make<RVATableChunk>(std::move(tableSymbols)); 1764 rdataSec->addChunk(tableChunk); 1765 1766 Symbol *t = symtab->findUnderscore(tableSym); 1767 Symbol *c = symtab->findUnderscore(countSym); 1768 replaceSymbol<DefinedSynthetic>(t, t->getName(), tableChunk); 1769 cast<DefinedAbsolute>(c)->setVA(tableChunk->getSize() / 4); 1770 } 1771 1772 // MinGW specific. Gather all relocations that are imported from a DLL even 1773 // though the code didn't expect it to, produce the table that the runtime 1774 // uses for fixing them up, and provide the synthetic symbols that the 1775 // runtime uses for finding the table. 1776 void Writer::createRuntimePseudoRelocs() { 1777 std::vector<RuntimePseudoReloc> rels; 1778 1779 for (Chunk *c : symtab->getChunks()) { 1780 auto *sc = dyn_cast<SectionChunk>(c); 1781 if (!sc || !sc->live) 1782 continue; 1783 sc->getRuntimePseudoRelocs(rels); 1784 } 1785 1786 if (!config->pseudoRelocs) { 1787 // Not writing any pseudo relocs; if some were needed, error out and 1788 // indicate what required them. 1789 for (const RuntimePseudoReloc &rpr : rels) 1790 error("automatic dllimport of " + rpr.sym->getName() + " in " + 1791 toString(rpr.target->file) + " requires pseudo relocations"); 1792 return; 1793 } 1794 1795 if (!rels.empty()) 1796 log("Writing " + Twine(rels.size()) + " runtime pseudo relocations"); 1797 PseudoRelocTableChunk *table = make<PseudoRelocTableChunk>(rels); 1798 rdataSec->addChunk(table); 1799 EmptyChunk *endOfList = make<EmptyChunk>(); 1800 rdataSec->addChunk(endOfList); 1801 1802 Symbol *headSym = symtab->findUnderscore("__RUNTIME_PSEUDO_RELOC_LIST__"); 1803 Symbol *endSym = symtab->findUnderscore("__RUNTIME_PSEUDO_RELOC_LIST_END__"); 1804 replaceSymbol<DefinedSynthetic>(headSym, headSym->getName(), table); 1805 replaceSymbol<DefinedSynthetic>(endSym, endSym->getName(), endOfList); 1806 } 1807 1808 // MinGW specific. 1809 // The MinGW .ctors and .dtors lists have sentinels at each end; 1810 // a (uintptr_t)-1 at the start and a (uintptr_t)0 at the end. 1811 // There's a symbol pointing to the start sentinel pointer, __CTOR_LIST__ 1812 // and __DTOR_LIST__ respectively. 1813 void Writer::insertCtorDtorSymbols() { 1814 AbsolutePointerChunk *ctorListHead = make<AbsolutePointerChunk>(-1); 1815 AbsolutePointerChunk *ctorListEnd = make<AbsolutePointerChunk>(0); 1816 AbsolutePointerChunk *dtorListHead = make<AbsolutePointerChunk>(-1); 1817 AbsolutePointerChunk *dtorListEnd = make<AbsolutePointerChunk>(0); 1818 ctorsSec->insertChunkAtStart(ctorListHead); 1819 ctorsSec->addChunk(ctorListEnd); 1820 dtorsSec->insertChunkAtStart(dtorListHead); 1821 dtorsSec->addChunk(dtorListEnd); 1822 1823 Symbol *ctorListSym = symtab->findUnderscore("__CTOR_LIST__"); 1824 Symbol *dtorListSym = symtab->findUnderscore("__DTOR_LIST__"); 1825 replaceSymbol<DefinedSynthetic>(ctorListSym, ctorListSym->getName(), 1826 ctorListHead); 1827 replaceSymbol<DefinedSynthetic>(dtorListSym, dtorListSym->getName(), 1828 dtorListHead); 1829 } 1830 1831 // Handles /section options to allow users to overwrite 1832 // section attributes. 1833 void Writer::setSectionPermissions() { 1834 for (auto &p : config->section) { 1835 StringRef name = p.first; 1836 uint32_t perm = p.second; 1837 for (OutputSection *sec : outputSections) 1838 if (sec->name == name) 1839 sec->setPermissions(perm); 1840 } 1841 } 1842 1843 // Write section contents to a mmap'ed file. 1844 void Writer::writeSections() { 1845 // Record the number of sections to apply section index relocations 1846 // against absolute symbols. See applySecIdx in Chunks.cpp.. 1847 DefinedAbsolute::numOutputSections = outputSections.size(); 1848 1849 uint8_t *buf = buffer->getBufferStart(); 1850 for (OutputSection *sec : outputSections) { 1851 uint8_t *secBuf = buf + sec->getFileOff(); 1852 // Fill gaps between functions in .text with INT3 instructions 1853 // instead of leaving as NUL bytes (which can be interpreted as 1854 // ADD instructions). 1855 if (sec->header.Characteristics & IMAGE_SCN_CNT_CODE) 1856 memset(secBuf, 0xCC, sec->getRawSize()); 1857 parallelForEach(sec->chunks, [&](Chunk *c) { 1858 c->writeTo(secBuf + c->getRVA() - sec->getRVA()); 1859 }); 1860 } 1861 } 1862 1863 void Writer::writeBuildId() { 1864 // There are two important parts to the build ID. 1865 // 1) If building with debug info, the COFF debug directory contains a 1866 // timestamp as well as a Guid and Age of the PDB. 1867 // 2) In all cases, the PE COFF file header also contains a timestamp. 1868 // For reproducibility, instead of a timestamp we want to use a hash of the 1869 // PE contents. 1870 if (config->debug) { 1871 assert(buildId && "BuildId is not set!"); 1872 // BuildId->BuildId was filled in when the PDB was written. 1873 } 1874 1875 // At this point the only fields in the COFF file which remain unset are the 1876 // "timestamp" in the COFF file header, and the ones in the coff debug 1877 // directory. Now we can hash the file and write that hash to the various 1878 // timestamp fields in the file. 1879 StringRef outputFileData( 1880 reinterpret_cast<const char *>(buffer->getBufferStart()), 1881 buffer->getBufferSize()); 1882 1883 uint32_t timestamp = config->timestamp; 1884 uint64_t hash = 0; 1885 bool generateSyntheticBuildId = 1886 config->mingw && config->debug && config->pdbPath.empty(); 1887 1888 if (config->repro || generateSyntheticBuildId) 1889 hash = xxHash64(outputFileData); 1890 1891 if (config->repro) 1892 timestamp = static_cast<uint32_t>(hash); 1893 1894 if (generateSyntheticBuildId) { 1895 // For MinGW builds without a PDB file, we still generate a build id 1896 // to allow associating a crash dump to the executable. 1897 buildId->buildId->PDB70.CVSignature = OMF::Signature::PDB70; 1898 buildId->buildId->PDB70.Age = 1; 1899 memcpy(buildId->buildId->PDB70.Signature, &hash, 8); 1900 // xxhash only gives us 8 bytes, so put some fixed data in the other half. 1901 memcpy(&buildId->buildId->PDB70.Signature[8], "LLD PDB.", 8); 1902 } 1903 1904 if (debugDirectory) 1905 debugDirectory->setTimeDateStamp(timestamp); 1906 1907 uint8_t *buf = buffer->getBufferStart(); 1908 buf += dosStubSize + sizeof(PEMagic); 1909 object::coff_file_header *coffHeader = 1910 reinterpret_cast<coff_file_header *>(buf); 1911 coffHeader->TimeDateStamp = timestamp; 1912 } 1913 1914 // Sort .pdata section contents according to PE/COFF spec 5.5. 1915 void Writer::sortExceptionTable() { 1916 if (!firstPdata) 1917 return; 1918 // We assume .pdata contains function table entries only. 1919 auto bufAddr = [&](Chunk *c) { 1920 OutputSection *os = c->getOutputSection(); 1921 return buffer->getBufferStart() + os->getFileOff() + c->getRVA() - 1922 os->getRVA(); 1923 }; 1924 uint8_t *begin = bufAddr(firstPdata); 1925 uint8_t *end = bufAddr(lastPdata) + lastPdata->getSize(); 1926 if (config->machine == AMD64) { 1927 struct Entry { ulittle32_t begin, end, unwind; }; 1928 if ((end - begin) % sizeof(Entry) != 0) { 1929 fatal("unexpected .pdata size: " + Twine(end - begin) + 1930 " is not a multiple of " + Twine(sizeof(Entry))); 1931 } 1932 parallelSort( 1933 MutableArrayRef<Entry>((Entry *)begin, (Entry *)end), 1934 [](const Entry &a, const Entry &b) { return a.begin < b.begin; }); 1935 return; 1936 } 1937 if (config->machine == ARMNT || config->machine == ARM64) { 1938 struct Entry { ulittle32_t begin, unwind; }; 1939 if ((end - begin) % sizeof(Entry) != 0) { 1940 fatal("unexpected .pdata size: " + Twine(end - begin) + 1941 " is not a multiple of " + Twine(sizeof(Entry))); 1942 } 1943 parallelSort( 1944 MutableArrayRef<Entry>((Entry *)begin, (Entry *)end), 1945 [](const Entry &a, const Entry &b) { return a.begin < b.begin; }); 1946 return; 1947 } 1948 lld::errs() << "warning: don't know how to handle .pdata.\n"; 1949 } 1950 1951 // The CRT section contains, among other things, the array of function 1952 // pointers that initialize every global variable that is not trivially 1953 // constructed. The CRT calls them one after the other prior to invoking 1954 // main(). 1955 // 1956 // As per C++ spec, 3.6.2/2.3, 1957 // "Variables with ordered initialization defined within a single 1958 // translation unit shall be initialized in the order of their definitions 1959 // in the translation unit" 1960 // 1961 // It is therefore critical to sort the chunks containing the function 1962 // pointers in the order that they are listed in the object file (top to 1963 // bottom), otherwise global objects might not be initialized in the 1964 // correct order. 1965 void Writer::sortCRTSectionChunks(std::vector<Chunk *> &chunks) { 1966 auto sectionChunkOrder = [](const Chunk *a, const Chunk *b) { 1967 auto sa = dyn_cast<SectionChunk>(a); 1968 auto sb = dyn_cast<SectionChunk>(b); 1969 assert(sa && sb && "Non-section chunks in CRT section!"); 1970 1971 StringRef sAObj = sa->file->mb.getBufferIdentifier(); 1972 StringRef sBObj = sb->file->mb.getBufferIdentifier(); 1973 1974 return sAObj == sBObj && sa->getSectionNumber() < sb->getSectionNumber(); 1975 }; 1976 llvm::stable_sort(chunks, sectionChunkOrder); 1977 1978 if (config->verbose) { 1979 for (auto &c : chunks) { 1980 auto sc = dyn_cast<SectionChunk>(c); 1981 log(" " + sc->file->mb.getBufferIdentifier().str() + 1982 ", SectionID: " + Twine(sc->getSectionNumber())); 1983 } 1984 } 1985 } 1986 1987 OutputSection *Writer::findSection(StringRef name) { 1988 for (OutputSection *sec : outputSections) 1989 if (sec->name == name) 1990 return sec; 1991 return nullptr; 1992 } 1993 1994 uint32_t Writer::getSizeOfInitializedData() { 1995 uint32_t res = 0; 1996 for (OutputSection *s : outputSections) 1997 if (s->header.Characteristics & IMAGE_SCN_CNT_INITIALIZED_DATA) 1998 res += s->getRawSize(); 1999 return res; 2000 } 2001 2002 // Add base relocations to .reloc section. 2003 void Writer::addBaserels() { 2004 if (!config->relocatable) 2005 return; 2006 relocSec->chunks.clear(); 2007 std::vector<Baserel> v; 2008 for (OutputSection *sec : outputSections) { 2009 if (sec->header.Characteristics & IMAGE_SCN_MEM_DISCARDABLE) 2010 continue; 2011 // Collect all locations for base relocations. 2012 for (Chunk *c : sec->chunks) 2013 c->getBaserels(&v); 2014 // Add the addresses to .reloc section. 2015 if (!v.empty()) 2016 addBaserelBlocks(v); 2017 v.clear(); 2018 } 2019 } 2020 2021 // Add addresses to .reloc section. Note that addresses are grouped by page. 2022 void Writer::addBaserelBlocks(std::vector<Baserel> &v) { 2023 const uint32_t mask = ~uint32_t(pageSize - 1); 2024 uint32_t page = v[0].rva & mask; 2025 size_t i = 0, j = 1; 2026 for (size_t e = v.size(); j < e; ++j) { 2027 uint32_t p = v[j].rva & mask; 2028 if (p == page) 2029 continue; 2030 relocSec->addChunk(make<BaserelChunk>(page, &v[i], &v[0] + j)); 2031 i = j; 2032 page = p; 2033 } 2034 if (i == j) 2035 return; 2036 relocSec->addChunk(make<BaserelChunk>(page, &v[i], &v[0] + j)); 2037 } 2038 2039 PartialSection *Writer::createPartialSection(StringRef name, 2040 uint32_t outChars) { 2041 PartialSection *&pSec = partialSections[{name, outChars}]; 2042 if (pSec) 2043 return pSec; 2044 pSec = make<PartialSection>(name, outChars); 2045 return pSec; 2046 } 2047 2048 PartialSection *Writer::findPartialSection(StringRef name, uint32_t outChars) { 2049 auto it = partialSections.find({name, outChars}); 2050 if (it != partialSections.end()) 2051 return it->second; 2052 return nullptr; 2053 } 2054 2055 void Writer::fixTlsAlignment() { 2056 Defined *tlsSym = 2057 dyn_cast_or_null<Defined>(symtab->findUnderscore("_tls_used")); 2058 if (!tlsSym) 2059 return; 2060 2061 OutputSection *sec = tlsSym->getChunk()->getOutputSection(); 2062 assert(sec && tlsSym->getRVA() >= sec->getRVA() && 2063 "no output section for _tls_used"); 2064 2065 uint8_t *secBuf = buffer->getBufferStart() + sec->getFileOff(); 2066 uint64_t tlsOffset = tlsSym->getRVA() - sec->getRVA(); 2067 uint64_t directorySize = config->is64() 2068 ? sizeof(object::coff_tls_directory64) 2069 : sizeof(object::coff_tls_directory32); 2070 2071 if (tlsOffset + directorySize > sec->getRawSize()) 2072 fatal("_tls_used sym is malformed"); 2073 2074 if (config->is64()) { 2075 object::coff_tls_directory64 *tlsDir = 2076 reinterpret_cast<object::coff_tls_directory64 *>(&secBuf[tlsOffset]); 2077 tlsDir->setAlignment(tlsAlignment); 2078 } else { 2079 object::coff_tls_directory32 *tlsDir = 2080 reinterpret_cast<object::coff_tls_directory32 *>(&secBuf[tlsOffset]); 2081 tlsDir->setAlignment(tlsAlignment); 2082 } 2083 } 2084