| 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 "COFFLinkerContext.h" |
| 11 | #include "CallGraphSort.h" |
| 12 | #include "Config.h" |
| 13 | #include "DLL.h" |
| 14 | #include "InputFiles.h" |
| 15 | #include "LLDMapFile.h" |
| 16 | #include "MapFile.h" |
| 17 | #include "PDB.h" |
| 18 | #include "SymbolTable.h" |
| 19 | #include "Symbols.h" |
| 20 | #include "lld/Common/ErrorHandler.h" |
| 21 | #include "lld/Common/Memory.h" |
| 22 | #include "lld/Common/Timer.h" |
| 23 | #include "llvm/ADT/DenseMap.h" |
| 24 | #include "llvm/ADT/STLExtras.h" |
| 25 | #include "llvm/ADT/StringSet.h" |
| 26 | #include "llvm/BinaryFormat/COFF.h" |
| 27 | #include "llvm/Support/Endian.h" |
| 28 | #include "llvm/Support/FileOutputBuffer.h" |
| 29 | #include "llvm/Support/Parallel.h" |
| 30 | #include "llvm/Support/RandomNumberGenerator.h" |
| 31 | #include "llvm/Support/TimeProfiler.h" |
| 32 | #include "llvm/Support/xxhash.h" |
| 33 | #include <algorithm> |
| 34 | #include <cstdio> |
| 35 | #include <map> |
| 36 | #include <memory> |
| 37 | #include <utility> |
| 38 | |
| 39 | using namespace llvm; |
| 40 | using namespace llvm::COFF; |
| 41 | using namespace llvm::object; |
| 42 | using namespace llvm::support; |
| 43 | using namespace llvm::support::endian; |
| 44 | using namespace lld; |
| 45 | using namespace lld::coff; |
| 46 | |
| 47 | /* To re-generate DOSProgram: |
| 48 | $ cat > /tmp/DOSProgram.asm |
| 49 | org 0 |
| 50 | ; Copy cs to ds. |
| 51 | push cs |
| 52 | pop ds |
| 53 | ; Point ds:dx at the $-terminated string. |
| 54 | mov dx, str |
| 55 | ; Int 21/AH=09h: Write string to standard output. |
| 56 | mov ah, 0x9 |
| 57 | int 0x21 |
| 58 | ; Int 21/AH=4Ch: Exit with return code (in AL). |
| 59 | mov ax, 0x4C01 |
| 60 | int 0x21 |
| 61 | str: |
| 62 | db 'This program cannot be run in DOS mode.$' |
| 63 | align 8, db 0 |
| 64 | $ nasm -fbin /tmp/DOSProgram.asm -o /tmp/DOSProgram.bin |
| 65 | $ xxd -i /tmp/DOSProgram.bin |
| 66 | */ |
| 67 | static unsigned char dosProgram[] = { |
| 68 | 0x0e, 0x1f, 0xba, 0x0e, 0x00, 0xb4, 0x09, 0xcd, 0x21, 0xb8, 0x01, 0x4c, |
| 69 | 0xcd, 0x21, 0x54, 0x68, 0x69, 0x73, 0x20, 0x70, 0x72, 0x6f, 0x67, 0x72, |
| 70 | 0x61, 0x6d, 0x20, 0x63, 0x61, 0x6e, 0x6e, 0x6f, 0x74, 0x20, 0x62, 0x65, |
| 71 | 0x20, 0x72, 0x75, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x44, 0x4f, 0x53, 0x20, |
| 72 | 0x6d, 0x6f, 0x64, 0x65, 0x2e, 0x24, 0x00, 0x00 |
| 73 | }; |
| 74 | static_assert(sizeof(dosProgram) % 8 == 0, |
| 75 | "DOSProgram size must be multiple of 8" ); |
| 76 | static_assert((sizeof(dos_header) + sizeof(dosProgram)) % 8 == 0, |
| 77 | "DOSStub size must be multiple of 8" ); |
| 78 | |
| 79 | static const int numberOfDataDirectory = 16; |
| 80 | |
| 81 | namespace { |
| 82 | |
| 83 | class DebugDirectoryChunk : public NonSectionChunk { |
| 84 | public: |
| 85 | DebugDirectoryChunk(const COFFLinkerContext &c, |
| 86 | const std::vector<std::pair<COFF::DebugType, Chunk *>> &r, |
| 87 | bool writeRepro) |
| 88 | : records(r), writeRepro(writeRepro), ctx(c) {} |
| 89 | |
| 90 | size_t getSize() const override { |
| 91 | return (records.size() + int(writeRepro)) * sizeof(debug_directory); |
| 92 | } |
| 93 | |
| 94 | void writeTo(uint8_t *b) const override { |
| 95 | auto *d = reinterpret_cast<debug_directory *>(b); |
| 96 | |
| 97 | for (const std::pair<COFF::DebugType, Chunk *>& record : records) { |
| 98 | Chunk *c = record.second; |
| 99 | const OutputSection *os = ctx.getOutputSection(c); |
| 100 | uint64_t offs = os->getFileOff() + (c->getRVA() - os->getRVA()); |
| 101 | fillEntry(d, debugType: record.first, size: c->getSize(), rva: c->getRVA(), offs); |
| 102 | ++d; |
| 103 | } |
| 104 | |
| 105 | if (writeRepro) { |
| 106 | // FIXME: The COFF spec allows either a 0-sized entry to just say |
| 107 | // "the timestamp field is really a hash", or a 4-byte size field |
| 108 | // followed by that many bytes containing a longer hash (with the |
| 109 | // lowest 4 bytes usually being the timestamp in little-endian order). |
| 110 | // Consider storing the full 8 bytes computed by xxh3_64bits here. |
| 111 | fillEntry(d, debugType: COFF::IMAGE_DEBUG_TYPE_REPRO, size: 0, rva: 0, offs: 0); |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | void setTimeDateStamp(uint32_t timeDateStamp) { |
| 116 | for (support::ulittle32_t *tds : timeDateStamps) |
| 117 | *tds = timeDateStamp; |
| 118 | } |
| 119 | |
| 120 | private: |
| 121 | void fillEntry(debug_directory *d, COFF::DebugType debugType, size_t size, |
| 122 | uint64_t rva, uint64_t offs) const { |
| 123 | d->Characteristics = 0; |
| 124 | d->TimeDateStamp = 0; |
| 125 | d->MajorVersion = 0; |
| 126 | d->MinorVersion = 0; |
| 127 | d->Type = debugType; |
| 128 | d->SizeOfData = size; |
| 129 | d->AddressOfRawData = rva; |
| 130 | d->PointerToRawData = offs; |
| 131 | |
| 132 | timeDateStamps.push_back(x: &d->TimeDateStamp); |
| 133 | } |
| 134 | |
| 135 | mutable std::vector<support::ulittle32_t *> timeDateStamps; |
| 136 | const std::vector<std::pair<COFF::DebugType, Chunk *>> &records; |
| 137 | bool writeRepro; |
| 138 | const COFFLinkerContext &ctx; |
| 139 | }; |
| 140 | |
| 141 | class CVDebugRecordChunk : public NonSectionChunk { |
| 142 | public: |
| 143 | CVDebugRecordChunk(const COFFLinkerContext &c) : ctx(c) {} |
| 144 | |
| 145 | size_t getSize() const override { |
| 146 | return sizeof(codeview::DebugInfo) + ctx.config.pdbAltPath.size() + 1; |
| 147 | } |
| 148 | |
| 149 | void writeTo(uint8_t *b) const override { |
| 150 | // Save off the DebugInfo entry to backfill the file signature (build id) |
| 151 | // in Writer::writeBuildId |
| 152 | buildId = reinterpret_cast<codeview::DebugInfo *>(b); |
| 153 | |
| 154 | // variable sized field (PDB Path) |
| 155 | char *p = reinterpret_cast<char *>(b + sizeof(*buildId)); |
| 156 | if (!ctx.config.pdbAltPath.empty()) |
| 157 | memcpy(dest: p, src: ctx.config.pdbAltPath.data(), n: ctx.config.pdbAltPath.size()); |
| 158 | p[ctx.config.pdbAltPath.size()] = '\0'; |
| 159 | } |
| 160 | |
| 161 | mutable codeview::DebugInfo *buildId = nullptr; |
| 162 | |
| 163 | private: |
| 164 | const COFFLinkerContext &ctx; |
| 165 | }; |
| 166 | |
| 167 | class ExtendedDllCharacteristicsChunk : public NonSectionChunk { |
| 168 | public: |
| 169 | ExtendedDllCharacteristicsChunk(uint32_t c) : characteristics(c) {} |
| 170 | |
| 171 | size_t getSize() const override { return 4; } |
| 172 | |
| 173 | void writeTo(uint8_t *buf) const override { write32le(P: buf, V: characteristics); } |
| 174 | |
| 175 | uint32_t characteristics = 0; |
| 176 | }; |
| 177 | |
| 178 | // PartialSection represents a group of chunks that contribute to an |
| 179 | // OutputSection. Collating a collection of PartialSections of same name and |
| 180 | // characteristics constitutes the OutputSection. |
| 181 | class PartialSectionKey { |
| 182 | public: |
| 183 | StringRef name; |
| 184 | unsigned characteristics; |
| 185 | |
| 186 | bool operator<(const PartialSectionKey &other) const { |
| 187 | int c = name.compare(RHS: other.name); |
| 188 | if (c > 0) |
| 189 | return false; |
| 190 | if (c == 0) |
| 191 | return characteristics < other.characteristics; |
| 192 | return true; |
| 193 | } |
| 194 | }; |
| 195 | |
| 196 | struct ChunkRange { |
| 197 | Chunk *first = nullptr, *last; |
| 198 | }; |
| 199 | |
| 200 | // The writer writes a SymbolTable result to a file. |
| 201 | class Writer { |
| 202 | public: |
| 203 | Writer(COFFLinkerContext &c) |
| 204 | : buffer(c.e.outputBuffer), delayIdata(c), ctx(c) {} |
| 205 | void run(); |
| 206 | |
| 207 | private: |
| 208 | void calculateStubDependentSizes(); |
| 209 | void createSections(); |
| 210 | void createMiscChunks(); |
| 211 | void createImportTables(); |
| 212 | void appendImportThunks(); |
| 213 | void locateImportTables(); |
| 214 | void createExportTable(); |
| 215 | void mergeSection(const std::map<StringRef, StringRef>::value_type &p); |
| 216 | void mergeSections(); |
| 217 | void sortECChunks(); |
| 218 | void appendECImportTables(); |
| 219 | void removeUnusedSections(); |
| 220 | void assignAddresses(); |
| 221 | bool isInRange(uint16_t relType, uint64_t s, uint64_t p, int margin, |
| 222 | MachineTypes machine); |
| 223 | std::pair<Defined *, bool> getThunk(DenseMap<uint64_t, Defined *> &lastThunks, |
| 224 | Defined *target, uint64_t p, |
| 225 | uint16_t type, int margin, |
| 226 | MachineTypes machine); |
| 227 | bool createThunks(OutputSection *os, int margin); |
| 228 | bool verifyRanges(const std::vector<Chunk *> chunks); |
| 229 | void createECCodeMap(); |
| 230 | void finalizeAddresses(); |
| 231 | void removeEmptySections(); |
| 232 | void assignOutputSectionIndices(); |
| 233 | void createSymbolAndStringTable(); |
| 234 | void openFile(StringRef outputPath); |
| 235 | template <typename PEHeaderTy> void writeHeader(); |
| 236 | void createSEHTable(); |
| 237 | void createRuntimePseudoRelocs(); |
| 238 | void createECChunks(); |
| 239 | void insertCtorDtorSymbols(); |
| 240 | void insertBssDataStartEndSymbols(); |
| 241 | void markSymbolsWithRelocations(ObjFile *file, SymbolRVASet &usedSymbols); |
| 242 | void createGuardCFTables(); |
| 243 | void markSymbolsForRVATable(ObjFile *file, |
| 244 | ArrayRef<SectionChunk *> symIdxChunks, |
| 245 | SymbolRVASet &tableSymbols); |
| 246 | void getSymbolsFromSections(ObjFile *file, |
| 247 | ArrayRef<SectionChunk *> symIdxChunks, |
| 248 | std::vector<Symbol *> &symbols); |
| 249 | void maybeAddRVATable(SymbolRVASet tableSymbols, StringRef tableSym, |
| 250 | StringRef countSym, bool hasFlag=false); |
| 251 | void setSectionPermissions(); |
| 252 | void setECSymbols(); |
| 253 | void writeSections(); |
| 254 | void writeBuildId(); |
| 255 | void writePEChecksum(); |
| 256 | void sortSections(); |
| 257 | template <typename T> void sortExceptionTable(ChunkRange &exceptionTable); |
| 258 | void sortExceptionTables(); |
| 259 | void sortCRTSectionChunks(std::vector<Chunk *> &chunks); |
| 260 | void addSyntheticIdata(); |
| 261 | void sortBySectionOrder(std::vector<Chunk *> &chunks); |
| 262 | void fixPartialSectionChars(StringRef name, uint32_t chars); |
| 263 | bool fixGnuImportChunks(); |
| 264 | void fixTlsAlignment(); |
| 265 | PartialSection *createPartialSection(StringRef name, uint32_t outChars); |
| 266 | PartialSection *findPartialSection(StringRef name, uint32_t outChars); |
| 267 | |
| 268 | std::optional<coff_symbol16> createSymbol(Defined *d); |
| 269 | size_t addEntryToStringTable(StringRef str); |
| 270 | |
| 271 | OutputSection *findSection(StringRef name); |
| 272 | void addBaserels(); |
| 273 | void addBaserelBlocks(std::vector<Baserel> &v); |
| 274 | void createDynamicRelocs(); |
| 275 | |
| 276 | uint32_t getSizeOfInitializedData(); |
| 277 | |
| 278 | void prepareLoadConfig(); |
| 279 | template <typename T> |
| 280 | void prepareLoadConfig(SymbolTable &symtab, T *loadConfig); |
| 281 | |
| 282 | std::unique_ptr<FileOutputBuffer> &buffer; |
| 283 | std::map<PartialSectionKey, PartialSection *> partialSections; |
| 284 | std::vector<char> strtab; |
| 285 | std::vector<llvm::object::coff_symbol16> outputSymtab; |
| 286 | std::vector<ECCodeMapEntry> codeMap; |
| 287 | IdataContents idata; |
| 288 | Chunk *importTableStart = nullptr; |
| 289 | uint64_t importTableSize = 0; |
| 290 | Chunk *iatStart = nullptr; |
| 291 | uint64_t iatSize = 0; |
| 292 | DelayLoadContents delayIdata; |
| 293 | bool setNoSEHCharacteristic = false; |
| 294 | uint32_t tlsAlignment = 0; |
| 295 | |
| 296 | DebugDirectoryChunk *debugDirectory = nullptr; |
| 297 | std::vector<std::pair<COFF::DebugType, Chunk *>> debugRecords; |
| 298 | CVDebugRecordChunk *buildId = nullptr; |
| 299 | ArrayRef<uint8_t> sectionTable; |
| 300 | |
| 301 | // List of Arm64EC export thunks. |
| 302 | std::vector<std::pair<Chunk *, Defined *>> exportThunks; |
| 303 | |
| 304 | uint64_t fileSize; |
| 305 | uint32_t pointerToSymbolTable = 0; |
| 306 | uint64_t sizeOfImage; |
| 307 | uint64_t ; |
| 308 | |
| 309 | uint32_t dosStubSize; |
| 310 | uint32_t ; |
| 311 | uint32_t ; |
| 312 | uint32_t dataDirOffset64; |
| 313 | |
| 314 | OutputSection *textSec; |
| 315 | OutputSection *hexpthkSec; |
| 316 | OutputSection *bssSec; |
| 317 | OutputSection *rdataSec; |
| 318 | OutputSection *buildidSec; |
| 319 | OutputSection *dataSec; |
| 320 | OutputSection *pdataSec; |
| 321 | OutputSection *idataSec; |
| 322 | OutputSection *edataSec; |
| 323 | OutputSection *didatSec; |
| 324 | OutputSection *a64xrmSec; |
| 325 | OutputSection *rsrcSec; |
| 326 | OutputSection *relocSec; |
| 327 | OutputSection *; |
| 328 | OutputSection *; |
| 329 | // Either .rdata section or .buildid section. |
| 330 | OutputSection *debugInfoSec; |
| 331 | |
| 332 | // The range of .pdata sections in the output file. |
| 333 | // |
| 334 | // We need to keep track of the location of .pdata in whichever section it |
| 335 | // gets merged into so that we can sort its contents and emit a correct data |
| 336 | // directory entry for the exception table. This is also the case for some |
| 337 | // other sections (such as .edata) but because the contents of those sections |
| 338 | // are entirely linker-generated we can keep track of their locations using |
| 339 | // the chunks that the linker creates. All .pdata chunks come from input |
| 340 | // files, so we need to keep track of them separately. |
| 341 | ChunkRange pdata; |
| 342 | |
| 343 | // x86_64 .pdata sections on ARM64EC/ARM64X targets. |
| 344 | ChunkRange hybridPdata; |
| 345 | |
| 346 | // CHPE metadata symbol on ARM64C target. |
| 347 | DefinedRegular *chpeSym = nullptr; |
| 348 | |
| 349 | COFFLinkerContext &ctx; |
| 350 | }; |
| 351 | } // anonymous namespace |
| 352 | |
| 353 | void lld::coff::writeResult(COFFLinkerContext &ctx) { |
| 354 | llvm::TimeTraceScope timeScope("Write output(s)" ); |
| 355 | Writer(ctx).run(); |
| 356 | } |
| 357 | |
| 358 | void OutputSection::addChunk(Chunk *c) { |
| 359 | chunks.push_back(x: c); |
| 360 | } |
| 361 | |
| 362 | void OutputSection::insertChunkAtStart(Chunk *c) { |
| 363 | chunks.insert(position: chunks.begin(), x: c); |
| 364 | } |
| 365 | |
| 366 | void OutputSection::setPermissions(uint32_t c) { |
| 367 | header.Characteristics &= ~permMask; |
| 368 | header.Characteristics |= c; |
| 369 | } |
| 370 | |
| 371 | void OutputSection::merge(OutputSection *other) { |
| 372 | chunks.insert(position: chunks.end(), first: other->chunks.begin(), last: other->chunks.end()); |
| 373 | other->chunks.clear(); |
| 374 | contribSections.insert(position: contribSections.end(), first: other->contribSections.begin(), |
| 375 | last: other->contribSections.end()); |
| 376 | other->contribSections.clear(); |
| 377 | |
| 378 | // MS link.exe compatibility: when merging a code section into a data section, |
| 379 | // mark the target section as a code section. |
| 380 | if (other->header.Characteristics & IMAGE_SCN_CNT_CODE) { |
| 381 | header.Characteristics |= IMAGE_SCN_CNT_CODE; |
| 382 | header.Characteristics &= |
| 383 | ~(IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_CNT_UNINITIALIZED_DATA); |
| 384 | } |
| 385 | } |
| 386 | |
| 387 | // Write the section header to a given buffer. |
| 388 | void OutputSection::(uint8_t *buf, bool isDebug) { |
| 389 | auto *hdr = reinterpret_cast<coff_section *>(buf); |
| 390 | *hdr = header; |
| 391 | if (stringTableOff) { |
| 392 | // If name is too long, write offset into the string table as a name. |
| 393 | encodeSectionName(Out: hdr->Name, Offset: stringTableOff); |
| 394 | } else { |
| 395 | assert(!isDebug || name.size() <= COFF::NameSize || |
| 396 | (hdr->Characteristics & IMAGE_SCN_MEM_DISCARDABLE) == 0); |
| 397 | strncpy(dest: hdr->Name, src: name.data(), |
| 398 | n: std::min(a: name.size(), b: (size_t)COFF::NameSize)); |
| 399 | } |
| 400 | } |
| 401 | |
| 402 | void OutputSection::addContributingPartialSection(PartialSection *sec) { |
| 403 | contribSections.push_back(x: sec); |
| 404 | } |
| 405 | |
| 406 | void OutputSection::splitECChunks() { |
| 407 | llvm::stable_sort(Range&: chunks, C: [=](const Chunk *a, const Chunk *b) { |
| 408 | return (a->getMachine() != ARM64) < (b->getMachine() != ARM64); |
| 409 | }); |
| 410 | } |
| 411 | |
| 412 | // Check whether the target address S is in range from a relocation |
| 413 | // of type relType at address P. |
| 414 | bool Writer::isInRange(uint16_t relType, uint64_t s, uint64_t p, int margin, |
| 415 | MachineTypes machine) { |
| 416 | if (machine == ARMNT) { |
| 417 | int64_t diff = AbsoluteDifference(X: s, Y: p + 4) + margin; |
| 418 | switch (relType) { |
| 419 | case IMAGE_REL_ARM_BRANCH20T: |
| 420 | return isInt<21>(x: diff); |
| 421 | case IMAGE_REL_ARM_BRANCH24T: |
| 422 | case IMAGE_REL_ARM_BLX23T: |
| 423 | return isInt<25>(x: diff); |
| 424 | default: |
| 425 | return true; |
| 426 | } |
| 427 | } else if (isAnyArm64(Machine: machine)) { |
| 428 | int64_t diff = AbsoluteDifference(X: s, Y: p) + margin; |
| 429 | switch (relType) { |
| 430 | case IMAGE_REL_ARM64_BRANCH26: |
| 431 | return isInt<28>(x: diff); |
| 432 | case IMAGE_REL_ARM64_BRANCH19: |
| 433 | return isInt<21>(x: diff); |
| 434 | case IMAGE_REL_ARM64_BRANCH14: |
| 435 | return isInt<16>(x: diff); |
| 436 | default: |
| 437 | return true; |
| 438 | } |
| 439 | } else { |
| 440 | return true; |
| 441 | } |
| 442 | } |
| 443 | |
| 444 | // Return the last thunk for the given target if it is in range, |
| 445 | // or create a new one. |
| 446 | std::pair<Defined *, bool> |
| 447 | Writer::getThunk(DenseMap<uint64_t, Defined *> &lastThunks, Defined *target, |
| 448 | uint64_t p, uint16_t type, int margin, MachineTypes machine) { |
| 449 | Defined *&lastThunk = lastThunks[target->getRVA()]; |
| 450 | if (lastThunk && isInRange(relType: type, s: lastThunk->getRVA(), p, margin, machine)) |
| 451 | return {lastThunk, false}; |
| 452 | Chunk *c; |
| 453 | switch (getMachineArchType(machine)) { |
| 454 | case Triple::thumb: |
| 455 | c = make<RangeExtensionThunkARM>(args&: ctx, args&: target); |
| 456 | break; |
| 457 | case Triple::aarch64: |
| 458 | c = make<RangeExtensionThunkARM64>(args&: machine, args&: target); |
| 459 | break; |
| 460 | default: |
| 461 | llvm_unreachable("Unexpected architecture" ); |
| 462 | } |
| 463 | Defined *d = make<DefinedSynthetic>(args: "range_extension_thunk" , args&: c); |
| 464 | lastThunk = d; |
| 465 | return {d, true}; |
| 466 | } |
| 467 | |
| 468 | // This checks all relocations, and for any relocation which isn't in range |
| 469 | // it adds a thunk after the section chunk that contains the relocation. |
| 470 | // If the latest thunk for the specific target is in range, that is used |
| 471 | // instead of creating a new thunk. All range checks are done with the |
| 472 | // specified margin, to make sure that relocations that originally are in |
| 473 | // range, but only barely, also get thunks - in case other added thunks makes |
| 474 | // the target go out of range. |
| 475 | // |
| 476 | // After adding thunks, we verify that all relocations are in range (with |
| 477 | // no extra margin requirements). If this failed, we restart (throwing away |
| 478 | // the previously created thunks) and retry with a wider margin. |
| 479 | bool Writer::createThunks(OutputSection *os, int margin) { |
| 480 | bool addressesChanged = false; |
| 481 | DenseMap<uint64_t, Defined *> lastThunks; |
| 482 | DenseMap<std::pair<ObjFile *, Defined *>, uint32_t> thunkSymtabIndices; |
| 483 | size_t thunksSize = 0; |
| 484 | // Recheck Chunks.size() each iteration, since we can insert more |
| 485 | // elements into it. |
| 486 | for (size_t i = 0; i != os->chunks.size(); ++i) { |
| 487 | SectionChunk *sc = dyn_cast<SectionChunk>(Val: os->chunks[i]); |
| 488 | if (!sc) { |
| 489 | auto chunk = cast<NonSectionChunk>(Val: os->chunks[i]); |
| 490 | if (uint32_t size = chunk->extendRanges()) { |
| 491 | thunksSize += size; |
| 492 | addressesChanged = true; |
| 493 | } |
| 494 | continue; |
| 495 | } |
| 496 | MachineTypes machine = sc->getMachine(); |
| 497 | size_t thunkInsertionSpot = i + 1; |
| 498 | |
| 499 | // Try to get a good enough estimate of where new thunks will be placed. |
| 500 | // Offset this by the size of the new thunks added so far, to make the |
| 501 | // estimate slightly better. |
| 502 | size_t thunkInsertionRVA = sc->getRVA() + sc->getSize() + thunksSize; |
| 503 | ObjFile *file = sc->file; |
| 504 | std::vector<std::pair<uint32_t, uint32_t>> relocReplacements; |
| 505 | ArrayRef<coff_relocation> originalRelocs = |
| 506 | file->getCOFFObj()->getRelocations(Sec: sc->header); |
| 507 | for (size_t j = 0, e = originalRelocs.size(); j < e; ++j) { |
| 508 | const coff_relocation &rel = originalRelocs[j]; |
| 509 | Symbol *relocTarget = file->getSymbol(symbolIndex: rel.SymbolTableIndex); |
| 510 | |
| 511 | // The estimate of the source address P should be pretty accurate, |
| 512 | // but we don't know whether the target Symbol address should be |
| 513 | // offset by thunksSize or not (or by some of thunksSize but not all of |
| 514 | // it), giving us some uncertainty once we have added one thunk. |
| 515 | uint64_t p = sc->getRVA() + rel.VirtualAddress + thunksSize; |
| 516 | |
| 517 | Defined *sym = dyn_cast_or_null<Defined>(Val: relocTarget); |
| 518 | if (!sym) |
| 519 | continue; |
| 520 | |
| 521 | uint64_t s = sym->getRVA(); |
| 522 | |
| 523 | if (isInRange(relType: rel.Type, s, p, margin, machine)) |
| 524 | continue; |
| 525 | |
| 526 | // If the target isn't in range, hook it up to an existing or new thunk. |
| 527 | auto [thunk, wasNew] = |
| 528 | getThunk(lastThunks, target: sym, p, type: rel.Type, margin, machine); |
| 529 | if (wasNew) { |
| 530 | Chunk *thunkChunk = thunk->getChunk(); |
| 531 | thunkChunk->setRVA( |
| 532 | thunkInsertionRVA); // Estimate of where it will be located. |
| 533 | os->chunks.insert(position: os->chunks.begin() + thunkInsertionSpot, x: thunkChunk); |
| 534 | thunkInsertionSpot++; |
| 535 | thunksSize += thunkChunk->getSize(); |
| 536 | thunkInsertionRVA += thunkChunk->getSize(); |
| 537 | addressesChanged = true; |
| 538 | } |
| 539 | |
| 540 | // To redirect the relocation, add a symbol to the parent object file's |
| 541 | // symbol table, and replace the relocation symbol table index with the |
| 542 | // new index. |
| 543 | auto insertion = thunkSymtabIndices.insert(KV: {{file, thunk}, ~0U}); |
| 544 | uint32_t &thunkSymbolIndex = insertion.first->second; |
| 545 | if (insertion.second) |
| 546 | thunkSymbolIndex = file->addRangeThunkSymbol(thunk); |
| 547 | relocReplacements.emplace_back(args&: j, args&: thunkSymbolIndex); |
| 548 | } |
| 549 | |
| 550 | // Get a writable copy of this section's relocations so they can be |
| 551 | // modified. If the relocations point into the object file, allocate new |
| 552 | // memory. Otherwise, this must be previously allocated memory that can be |
| 553 | // modified in place. |
| 554 | ArrayRef<coff_relocation> curRelocs = sc->getRelocs(); |
| 555 | MutableArrayRef<coff_relocation> newRelocs; |
| 556 | if (originalRelocs.data() == curRelocs.data()) { |
| 557 | newRelocs = MutableArrayRef( |
| 558 | bAlloc().Allocate<coff_relocation>(Num: originalRelocs.size()), |
| 559 | originalRelocs.size()); |
| 560 | } else { |
| 561 | newRelocs = MutableArrayRef( |
| 562 | const_cast<coff_relocation *>(curRelocs.data()), curRelocs.size()); |
| 563 | } |
| 564 | |
| 565 | // Copy each relocation, but replace the symbol table indices which need |
| 566 | // thunks. |
| 567 | auto nextReplacement = relocReplacements.begin(); |
| 568 | auto endReplacement = relocReplacements.end(); |
| 569 | for (size_t i = 0, e = originalRelocs.size(); i != e; ++i) { |
| 570 | newRelocs[i] = originalRelocs[i]; |
| 571 | if (nextReplacement != endReplacement && nextReplacement->first == i) { |
| 572 | newRelocs[i].SymbolTableIndex = nextReplacement->second; |
| 573 | ++nextReplacement; |
| 574 | } |
| 575 | } |
| 576 | |
| 577 | sc->setRelocs(newRelocs); |
| 578 | } |
| 579 | return addressesChanged; |
| 580 | } |
| 581 | |
| 582 | // Create a code map for CHPE metadata. |
| 583 | void Writer::createECCodeMap() { |
| 584 | if (!ctx.symtab.isEC()) |
| 585 | return; |
| 586 | |
| 587 | // Clear the map in case we were're recomputing the map after adding |
| 588 | // a range extension thunk. |
| 589 | codeMap.clear(); |
| 590 | |
| 591 | std::optional<chpe_range_type> lastType; |
| 592 | Chunk *first, *last; |
| 593 | |
| 594 | auto closeRange = [&]() { |
| 595 | if (lastType) { |
| 596 | codeMap.push_back(x: {first, last, *lastType}); |
| 597 | lastType.reset(); |
| 598 | } |
| 599 | }; |
| 600 | |
| 601 | for (OutputSection *sec : ctx.outputSections) { |
| 602 | for (Chunk *c : sec->chunks) { |
| 603 | // Skip empty section chunks. MS link.exe does not seem to do that and |
| 604 | // generates empty code ranges in some cases. |
| 605 | if (isa<SectionChunk>(Val: c) && !c->getSize()) |
| 606 | continue; |
| 607 | |
| 608 | std::optional<chpe_range_type> chunkType = c->getArm64ECRangeType(); |
| 609 | if (chunkType != lastType) { |
| 610 | closeRange(); |
| 611 | first = c; |
| 612 | lastType = chunkType; |
| 613 | } |
| 614 | last = c; |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | closeRange(); |
| 619 | |
| 620 | Symbol *tableCountSym = ctx.symtab.findUnderscore(name: "__hybrid_code_map_count" ); |
| 621 | cast<DefinedAbsolute>(Val: tableCountSym)->setVA(codeMap.size()); |
| 622 | } |
| 623 | |
| 624 | // Verify that all relocations are in range, with no extra margin requirements. |
| 625 | bool Writer::verifyRanges(const std::vector<Chunk *> chunks) { |
| 626 | for (Chunk *c : chunks) { |
| 627 | SectionChunk *sc = dyn_cast<SectionChunk>(Val: c); |
| 628 | if (!sc) { |
| 629 | if (!cast<NonSectionChunk>(Val: c)->verifyRanges()) |
| 630 | return false; |
| 631 | continue; |
| 632 | } |
| 633 | MachineTypes machine = sc->getMachine(); |
| 634 | |
| 635 | ArrayRef<coff_relocation> relocs = sc->getRelocs(); |
| 636 | for (const coff_relocation &rel : relocs) { |
| 637 | Symbol *relocTarget = sc->file->getSymbol(symbolIndex: rel.SymbolTableIndex); |
| 638 | |
| 639 | Defined *sym = dyn_cast_or_null<Defined>(Val: relocTarget); |
| 640 | if (!sym) |
| 641 | continue; |
| 642 | |
| 643 | uint64_t p = sc->getRVA() + rel.VirtualAddress; |
| 644 | uint64_t s = sym->getRVA(); |
| 645 | |
| 646 | if (!isInRange(relType: rel.Type, s, p, margin: 0, machine)) |
| 647 | return false; |
| 648 | } |
| 649 | } |
| 650 | return true; |
| 651 | } |
| 652 | |
| 653 | // Assign addresses and add thunks if necessary. |
| 654 | void Writer::finalizeAddresses() { |
| 655 | assignAddresses(); |
| 656 | if (ctx.config.machine != ARMNT && !isAnyArm64(Machine: ctx.config.machine)) |
| 657 | return; |
| 658 | |
| 659 | size_t origNumChunks = 0; |
| 660 | for (OutputSection *sec : ctx.outputSections) { |
| 661 | sec->origChunks = sec->chunks; |
| 662 | origNumChunks += sec->chunks.size(); |
| 663 | } |
| 664 | |
| 665 | int pass = 0; |
| 666 | int margin = 1024 * 100; |
| 667 | while (true) { |
| 668 | llvm::TimeTraceScope timeScope2("Add thunks pass" ); |
| 669 | |
| 670 | // First check whether we need thunks at all, or if the previous pass of |
| 671 | // adding them turned out ok. |
| 672 | bool rangesOk = true; |
| 673 | size_t numChunks = 0; |
| 674 | { |
| 675 | llvm::TimeTraceScope timeScope3("Verify ranges" ); |
| 676 | for (OutputSection *sec : ctx.outputSections) { |
| 677 | if (!verifyRanges(chunks: sec->chunks)) { |
| 678 | rangesOk = false; |
| 679 | break; |
| 680 | } |
| 681 | numChunks += sec->chunks.size(); |
| 682 | } |
| 683 | } |
| 684 | if (rangesOk) { |
| 685 | if (pass > 0) |
| 686 | Log(ctx) << "Added " << (numChunks - origNumChunks) << " thunks with " |
| 687 | << "margin " << margin << " in " << pass << " passes" ; |
| 688 | return; |
| 689 | } |
| 690 | |
| 691 | if (pass >= 10) |
| 692 | Fatal(ctx) << "adding thunks hasn't converged after " << pass |
| 693 | << " passes" ; |
| 694 | |
| 695 | if (pass > 0) { |
| 696 | // If the previous pass didn't work out, reset everything back to the |
| 697 | // original conditions before retrying with a wider margin. This should |
| 698 | // ideally never happen under real circumstances. |
| 699 | for (OutputSection *sec : ctx.outputSections) |
| 700 | sec->chunks = sec->origChunks; |
| 701 | margin *= 2; |
| 702 | } |
| 703 | |
| 704 | // Try adding thunks everywhere where it is needed, with a margin |
| 705 | // to avoid things going out of range due to the added thunks. |
| 706 | bool addressesChanged = false; |
| 707 | { |
| 708 | llvm::TimeTraceScope timeScope3("Create thunks" ); |
| 709 | for (OutputSection *sec : ctx.outputSections) |
| 710 | addressesChanged |= createThunks(os: sec, margin); |
| 711 | } |
| 712 | // If the verification above thought we needed thunks, we should have |
| 713 | // added some. |
| 714 | assert(addressesChanged); |
| 715 | (void)addressesChanged; |
| 716 | |
| 717 | // Recalculate the layout for the whole image (and verify the ranges at |
| 718 | // the start of the next round). |
| 719 | assignAddresses(); |
| 720 | |
| 721 | pass++; |
| 722 | } |
| 723 | } |
| 724 | |
| 725 | void Writer::writePEChecksum() { |
| 726 | if (!ctx.config.writeCheckSum) { |
| 727 | return; |
| 728 | } |
| 729 | |
| 730 | llvm::TimeTraceScope timeScope("PE checksum" ); |
| 731 | |
| 732 | // https://docs.microsoft.com/en-us/windows/win32/debug/pe-format#checksum |
| 733 | uint32_t *buf = (uint32_t *)buffer->getBufferStart(); |
| 734 | uint32_t size = (uint32_t)(buffer->getBufferSize()); |
| 735 | |
| 736 | pe32_header * = (pe32_header *)((uint8_t *)buf + coffHeaderOffset + |
| 737 | sizeof(coff_file_header)); |
| 738 | |
| 739 | uint64_t sum = 0; |
| 740 | uint32_t count = size; |
| 741 | ulittle16_t *addr = (ulittle16_t *)buf; |
| 742 | |
| 743 | // The PE checksum algorithm, implemented as suggested in RFC1071 |
| 744 | while (count > 1) { |
| 745 | sum += *addr++; |
| 746 | count -= 2; |
| 747 | } |
| 748 | |
| 749 | // Add left-over byte, if any |
| 750 | if (count > 0) |
| 751 | sum += *(unsigned char *)addr; |
| 752 | |
| 753 | // Fold 32-bit sum to 16 bits |
| 754 | while (sum >> 16) { |
| 755 | sum = (sum & 0xffff) + (sum >> 16); |
| 756 | } |
| 757 | |
| 758 | sum += size; |
| 759 | peHeader->CheckSum = sum; |
| 760 | } |
| 761 | |
| 762 | // The main function of the writer. |
| 763 | void Writer::run() { |
| 764 | { |
| 765 | llvm::TimeTraceScope timeScope("Write PE" ); |
| 766 | ScopedTimer t1(ctx.codeLayoutTimer); |
| 767 | |
| 768 | calculateStubDependentSizes(); |
| 769 | if (ctx.config.machine == ARM64X) |
| 770 | ctx.dynamicRelocs = make<DynamicRelocsChunk>(); |
| 771 | createImportTables(); |
| 772 | createSections(); |
| 773 | appendImportThunks(); |
| 774 | // Import thunks must be added before the Control Flow Guard tables are |
| 775 | // added. |
| 776 | createMiscChunks(); |
| 777 | createExportTable(); |
| 778 | mergeSections(); |
| 779 | sortECChunks(); |
| 780 | appendECImportTables(); |
| 781 | createDynamicRelocs(); |
| 782 | removeUnusedSections(); |
| 783 | finalizeAddresses(); |
| 784 | removeEmptySections(); |
| 785 | assignOutputSectionIndices(); |
| 786 | setSectionPermissions(); |
| 787 | setECSymbols(); |
| 788 | createSymbolAndStringTable(); |
| 789 | |
| 790 | if (fileSize > UINT32_MAX) |
| 791 | Fatal(ctx) << "image size (" << fileSize << ") " |
| 792 | << "exceeds maximum allowable size (" << UINT32_MAX << ")" ; |
| 793 | |
| 794 | openFile(outputPath: ctx.config.outputFile); |
| 795 | if (ctx.config.is64()) { |
| 796 | writeHeader<pe32plus_header>(); |
| 797 | } else { |
| 798 | writeHeader<pe32_header>(); |
| 799 | } |
| 800 | writeSections(); |
| 801 | prepareLoadConfig(); |
| 802 | sortExceptionTables(); |
| 803 | |
| 804 | // Fix up the alignment in the TLS Directory's characteristic field, |
| 805 | // if a specific alignment value is needed |
| 806 | if (tlsAlignment) |
| 807 | fixTlsAlignment(); |
| 808 | } |
| 809 | |
| 810 | if (!ctx.config.pdbPath.empty() && ctx.config.debug) { |
| 811 | assert(buildId); |
| 812 | createPDB(ctx, sectionTable, buildId: buildId->buildId); |
| 813 | } |
| 814 | writeBuildId(); |
| 815 | |
| 816 | writeLLDMapFile(ctx); |
| 817 | writeMapFile(ctx); |
| 818 | |
| 819 | writePEChecksum(); |
| 820 | |
| 821 | if (errorCount()) |
| 822 | return; |
| 823 | |
| 824 | llvm::TimeTraceScope timeScope("Commit PE to disk" ); |
| 825 | ScopedTimer t2(ctx.outputCommitTimer); |
| 826 | if (auto e = buffer->commit()) |
| 827 | Fatal(ctx) << "failed to write output '" << buffer->getPath() |
| 828 | << "': " << toString(E: std::move(e)); |
| 829 | } |
| 830 | |
| 831 | static StringRef getOutputSectionName(StringRef name) { |
| 832 | StringRef s = name.split(Separator: '$').first; |
| 833 | |
| 834 | // Treat a later period as a separator for MinGW, for sections like |
| 835 | // ".ctors.01234". |
| 836 | return s.substr(Start: 0, N: s.find(C: '.', From: 1)); |
| 837 | } |
| 838 | |
| 839 | // For /order. |
| 840 | void Writer::sortBySectionOrder(std::vector<Chunk *> &chunks) { |
| 841 | auto getPriority = [&ctx = ctx](const Chunk *c) { |
| 842 | if (auto *sec = dyn_cast<SectionChunk>(Val: c)) |
| 843 | if (sec->sym) |
| 844 | return ctx.config.order.lookup(Key: sec->sym->getName()); |
| 845 | return 0; |
| 846 | }; |
| 847 | |
| 848 | llvm::stable_sort(Range&: chunks, C: [=](const Chunk *a, const Chunk *b) { |
| 849 | return getPriority(a) < getPriority(b); |
| 850 | }); |
| 851 | } |
| 852 | |
| 853 | // Change the characteristics of existing PartialSections that belong to the |
| 854 | // section Name to Chars. |
| 855 | void Writer::fixPartialSectionChars(StringRef name, uint32_t chars) { |
| 856 | for (auto it : partialSections) { |
| 857 | PartialSection *pSec = it.second; |
| 858 | StringRef curName = pSec->name; |
| 859 | if (!curName.consume_front(Prefix: name) || |
| 860 | (!curName.empty() && !curName.starts_with(Prefix: "$" ))) |
| 861 | continue; |
| 862 | if (pSec->characteristics == chars) |
| 863 | continue; |
| 864 | PartialSection *destSec = createPartialSection(name: pSec->name, outChars: chars); |
| 865 | destSec->chunks.insert(position: destSec->chunks.end(), first: pSec->chunks.begin(), |
| 866 | last: pSec->chunks.end()); |
| 867 | pSec->chunks.clear(); |
| 868 | } |
| 869 | } |
| 870 | |
| 871 | // Sort concrete section chunks from GNU import libraries. |
| 872 | // |
| 873 | // GNU binutils doesn't use short import files, but instead produces import |
| 874 | // libraries that consist of object files, with section chunks for the .idata$* |
| 875 | // sections. These are linked just as regular static libraries. Each import |
| 876 | // library consists of one header object, one object file for every imported |
| 877 | // symbol, and one trailer object. In order for the .idata tables/lists to |
| 878 | // be formed correctly, the section chunks within each .idata$* section need |
| 879 | // to be grouped by library, and sorted alphabetically within each library |
| 880 | // (which makes sure the header comes first and the trailer last). |
| 881 | bool Writer::fixGnuImportChunks() { |
| 882 | uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ; |
| 883 | |
| 884 | // Make sure all .idata$* section chunks are mapped as RDATA in order to |
| 885 | // be sorted into the same sections as our own synthesized .idata chunks. |
| 886 | fixPartialSectionChars(name: ".idata" , chars: rdata); |
| 887 | |
| 888 | bool hasIdata = false; |
| 889 | // Sort all .idata$* chunks, grouping chunks from the same library, |
| 890 | // with alphabetical ordering of the object files within a library. |
| 891 | for (auto it : partialSections) { |
| 892 | PartialSection *pSec = it.second; |
| 893 | if (!pSec->name.starts_with(Prefix: ".idata" )) |
| 894 | continue; |
| 895 | |
| 896 | if (!pSec->chunks.empty()) |
| 897 | hasIdata = true; |
| 898 | llvm::stable_sort(Range&: pSec->chunks, C: [&](Chunk *s, Chunk *t) { |
| 899 | SectionChunk *sc1 = dyn_cast<SectionChunk>(Val: s); |
| 900 | SectionChunk *sc2 = dyn_cast<SectionChunk>(Val: t); |
| 901 | if (!sc1 || !sc2) { |
| 902 | // if SC1, order them ascending. If SC2 or both null, |
| 903 | // S is not less than T. |
| 904 | return sc1 != nullptr; |
| 905 | } |
| 906 | // Make a string with "libraryname/objectfile" for sorting, achieving |
| 907 | // both grouping by library and sorting of objects within a library, |
| 908 | // at once. |
| 909 | std::string key1 = |
| 910 | (sc1->file->parentName + "/" + sc1->file->getName()).str(); |
| 911 | std::string key2 = |
| 912 | (sc2->file->parentName + "/" + sc2->file->getName()).str(); |
| 913 | return key1 < key2; |
| 914 | }); |
| 915 | } |
| 916 | return hasIdata; |
| 917 | } |
| 918 | |
| 919 | // Add generated idata chunks, for imported symbols and DLLs, and a |
| 920 | // terminator in .idata$2. |
| 921 | void Writer::addSyntheticIdata() { |
| 922 | uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ; |
| 923 | idata.create(ctx); |
| 924 | |
| 925 | // Add the .idata content in the right section groups, to allow |
| 926 | // chunks from other linked in object files to be grouped together. |
| 927 | // See Microsoft PE/COFF spec 5.4 for details. |
| 928 | auto add = [&](StringRef n, std::vector<Chunk *> &v) { |
| 929 | PartialSection *pSec = createPartialSection(name: n, outChars: rdata); |
| 930 | pSec->chunks.insert(position: pSec->chunks.end(), first: v.begin(), last: v.end()); |
| 931 | }; |
| 932 | |
| 933 | // The loader assumes a specific order of data. |
| 934 | // Add each type in the correct order. |
| 935 | add(".idata$2" , idata.dirs); |
| 936 | add(".idata$4" , idata.lookups); |
| 937 | add(".idata$5" , idata.addresses); |
| 938 | if (!idata.hints.empty()) |
| 939 | add(".idata$6" , idata.hints); |
| 940 | add(".idata$7" , idata.dllNames); |
| 941 | if (!idata.auxIat.empty()) |
| 942 | add(".idata$9" , idata.auxIat); |
| 943 | if (!idata.auxIatCopy.empty()) |
| 944 | add(".idata$a" , idata.auxIatCopy); |
| 945 | } |
| 946 | |
| 947 | void Writer::appendECImportTables() { |
| 948 | if (!isArm64EC(Machine: ctx.config.machine)) |
| 949 | return; |
| 950 | |
| 951 | const uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ; |
| 952 | |
| 953 | // IAT is always placed at the beginning of .rdata section and its size |
| 954 | // is aligned to 4KB. Insert it here, after all merges all done. |
| 955 | if (PartialSection *importAddresses = findPartialSection(name: ".idata$5" , outChars: rdata)) { |
| 956 | if (!rdataSec->chunks.empty()) |
| 957 | rdataSec->chunks.front()->setAlignment( |
| 958 | std::max(a: 0x1000u, b: rdataSec->chunks.front()->getAlignment())); |
| 959 | iatSize = alignTo(Value: iatSize, Align: 0x1000); |
| 960 | |
| 961 | rdataSec->chunks.insert(position: rdataSec->chunks.begin(), |
| 962 | first: importAddresses->chunks.begin(), |
| 963 | last: importAddresses->chunks.end()); |
| 964 | rdataSec->contribSections.insert(position: rdataSec->contribSections.begin(), |
| 965 | x: importAddresses); |
| 966 | } |
| 967 | |
| 968 | // The auxiliary IAT is always placed at the end of the .rdata section |
| 969 | // and is aligned to 4KB. |
| 970 | if (PartialSection *auxIat = findPartialSection(name: ".idata$9" , outChars: rdata)) { |
| 971 | auxIat->chunks.front()->setAlignment(0x1000); |
| 972 | rdataSec->chunks.insert(position: rdataSec->chunks.end(), first: auxIat->chunks.begin(), |
| 973 | last: auxIat->chunks.end()); |
| 974 | rdataSec->addContributingPartialSection(sec: auxIat); |
| 975 | } |
| 976 | |
| 977 | if (!delayIdata.getAuxIat().empty()) { |
| 978 | delayIdata.getAuxIat().front()->setAlignment(0x1000); |
| 979 | rdataSec->chunks.insert(position: rdataSec->chunks.end(), |
| 980 | first: delayIdata.getAuxIat().begin(), |
| 981 | last: delayIdata.getAuxIat().end()); |
| 982 | } |
| 983 | } |
| 984 | |
| 985 | // Locate the first Chunk and size of the import directory list and the |
| 986 | // IAT. |
| 987 | void Writer::locateImportTables() { |
| 988 | uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ; |
| 989 | |
| 990 | if (PartialSection *importDirs = findPartialSection(name: ".idata$2" , outChars: rdata)) { |
| 991 | if (!importDirs->chunks.empty()) |
| 992 | importTableStart = importDirs->chunks.front(); |
| 993 | for (Chunk *c : importDirs->chunks) |
| 994 | importTableSize += c->getSize(); |
| 995 | } |
| 996 | |
| 997 | if (PartialSection *importAddresses = findPartialSection(name: ".idata$5" , outChars: rdata)) { |
| 998 | if (!importAddresses->chunks.empty()) |
| 999 | iatStart = importAddresses->chunks.front(); |
| 1000 | for (Chunk *c : importAddresses->chunks) |
| 1001 | iatSize += c->getSize(); |
| 1002 | } |
| 1003 | } |
| 1004 | |
| 1005 | // Return whether a SectionChunk's suffix (the dollar and any trailing |
| 1006 | // suffix) should be removed and sorted into the main suffixless |
| 1007 | // PartialSection. |
| 1008 | static bool shouldStripSectionSuffix(SectionChunk *sc, StringRef name, |
| 1009 | bool isMinGW) { |
| 1010 | // On MinGW, comdat groups are formed by putting the comdat group name |
| 1011 | // after the '$' in the section name. For .eh_frame$<symbol>, that must |
| 1012 | // still be sorted before the .eh_frame trailer from crtend.o, thus just |
| 1013 | // strip the section name trailer. For other sections, such as |
| 1014 | // .tls$$<symbol> (where non-comdat .tls symbols are otherwise stored in |
| 1015 | // ".tls$"), they must be strictly sorted after .tls. And for the |
| 1016 | // hypothetical case of comdat .CRT$XCU, we definitely need to keep the |
| 1017 | // suffix for sorting. Thus, to play it safe, only strip the suffix for |
| 1018 | // the standard sections. |
| 1019 | if (!isMinGW) |
| 1020 | return false; |
| 1021 | if (!sc || !sc->isCOMDAT()) |
| 1022 | return false; |
| 1023 | return name.starts_with(Prefix: ".text$" ) || name.starts_with(Prefix: ".data$" ) || |
| 1024 | name.starts_with(Prefix: ".rdata$" ) || name.starts_with(Prefix: ".pdata$" ) || |
| 1025 | name.starts_with(Prefix: ".xdata$" ) || name.starts_with(Prefix: ".eh_frame$" ); |
| 1026 | } |
| 1027 | |
| 1028 | void Writer::sortSections() { |
| 1029 | if (!ctx.config.callGraphProfile.empty()) { |
| 1030 | DenseMap<const SectionChunk *, int> order = |
| 1031 | computeCallGraphProfileOrder(ctx); |
| 1032 | for (auto it : order) { |
| 1033 | if (DefinedRegular *sym = it.first->sym) |
| 1034 | ctx.config.order[sym->getName()] = it.second; |
| 1035 | } |
| 1036 | } |
| 1037 | if (!ctx.config.order.empty()) |
| 1038 | for (auto it : partialSections) |
| 1039 | sortBySectionOrder(chunks&: it.second->chunks); |
| 1040 | } |
| 1041 | |
| 1042 | void Writer::calculateStubDependentSizes() { |
| 1043 | if (ctx.config.dosStub) |
| 1044 | dosStubSize = alignTo(Value: ctx.config.dosStub->getBufferSize(), Align: 8); |
| 1045 | else |
| 1046 | dosStubSize = sizeof(dos_header) + sizeof(dosProgram); |
| 1047 | |
| 1048 | coffHeaderOffset = dosStubSize + sizeof(PEMagic); |
| 1049 | peHeaderOffset = coffHeaderOffset + sizeof(coff_file_header); |
| 1050 | dataDirOffset64 = peHeaderOffset + sizeof(pe32plus_header); |
| 1051 | } |
| 1052 | |
| 1053 | // Create output section objects and add them to OutputSections. |
| 1054 | void Writer::createSections() { |
| 1055 | llvm::TimeTraceScope timeScope("Output sections" ); |
| 1056 | // First, create the builtin sections. |
| 1057 | const uint32_t data = IMAGE_SCN_CNT_INITIALIZED_DATA; |
| 1058 | const uint32_t bss = IMAGE_SCN_CNT_UNINITIALIZED_DATA; |
| 1059 | const uint32_t code = IMAGE_SCN_CNT_CODE; |
| 1060 | const uint32_t discardable = IMAGE_SCN_MEM_DISCARDABLE; |
| 1061 | const uint32_t r = IMAGE_SCN_MEM_READ; |
| 1062 | const uint32_t w = IMAGE_SCN_MEM_WRITE; |
| 1063 | const uint32_t x = IMAGE_SCN_MEM_EXECUTE; |
| 1064 | |
| 1065 | SmallDenseMap<std::pair<StringRef, uint32_t>, OutputSection *> sections; |
| 1066 | auto createSection = [&](StringRef name, uint32_t outChars) { |
| 1067 | OutputSection *&sec = sections[{name, outChars}]; |
| 1068 | if (!sec) { |
| 1069 | sec = make<OutputSection>(args&: name, args&: outChars); |
| 1070 | ctx.outputSections.push_back(x: sec); |
| 1071 | } |
| 1072 | return sec; |
| 1073 | }; |
| 1074 | |
| 1075 | // Try to match the section order used by link.exe. |
| 1076 | textSec = createSection(".text" , code | r | x); |
| 1077 | if (isArm64EC(Machine: ctx.config.machine)) |
| 1078 | hexpthkSec = createSection(".hexpthk" , code | r | x); |
| 1079 | bssSec = createSection(".bss" , bss | r | w); |
| 1080 | rdataSec = createSection(".rdata" , data | r); |
| 1081 | buildidSec = createSection(".buildid" , data | r); |
| 1082 | dataSec = createSection(".data" , data | r | w); |
| 1083 | pdataSec = createSection(".pdata" , data | r); |
| 1084 | idataSec = createSection(".idata" , data | r); |
| 1085 | edataSec = createSection(".edata" , data | r); |
| 1086 | didatSec = createSection(".didat" , data | r); |
| 1087 | if (isArm64EC(Machine: ctx.config.machine)) |
| 1088 | a64xrmSec = createSection(".a64xrm" , data | r); |
| 1089 | rsrcSec = createSection(".rsrc" , data | r); |
| 1090 | relocSec = createSection(".reloc" , data | discardable | r); |
| 1091 | ctorsSec = createSection(".ctors" , data | r | w); |
| 1092 | dtorsSec = createSection(".dtors" , data | r | w); |
| 1093 | |
| 1094 | // Then bin chunks by name and output characteristics. |
| 1095 | for (Chunk *c : ctx.driver.getChunks()) { |
| 1096 | auto *sc = dyn_cast<SectionChunk>(Val: c); |
| 1097 | if (sc && !sc->live) { |
| 1098 | if (ctx.config.verbose) |
| 1099 | sc->printDiscardedMessage(); |
| 1100 | continue; |
| 1101 | } |
| 1102 | StringRef name = c->getSectionName(); |
| 1103 | if (shouldStripSectionSuffix(sc, name, isMinGW: ctx.config.mingw)) |
| 1104 | name = name.split(Separator: '$').first; |
| 1105 | |
| 1106 | if (name.starts_with(Prefix: ".tls" )) |
| 1107 | tlsAlignment = std::max(a: tlsAlignment, b: c->getAlignment()); |
| 1108 | |
| 1109 | PartialSection *pSec = createPartialSection(name, |
| 1110 | outChars: c->getOutputCharacteristics()); |
| 1111 | pSec->chunks.push_back(x: c); |
| 1112 | } |
| 1113 | |
| 1114 | fixPartialSectionChars(name: ".rsrc" , chars: data | r); |
| 1115 | fixPartialSectionChars(name: ".edata" , chars: data | r); |
| 1116 | // Even in non MinGW cases, we might need to link against GNU import |
| 1117 | // libraries. |
| 1118 | bool hasIdata = fixGnuImportChunks(); |
| 1119 | if (!idata.empty()) |
| 1120 | hasIdata = true; |
| 1121 | |
| 1122 | if (hasIdata) |
| 1123 | addSyntheticIdata(); |
| 1124 | |
| 1125 | sortSections(); |
| 1126 | |
| 1127 | if (hasIdata) |
| 1128 | locateImportTables(); |
| 1129 | |
| 1130 | // Then create an OutputSection for each section. |
| 1131 | // '$' and all following characters in input section names are |
| 1132 | // discarded when determining output section. So, .text$foo |
| 1133 | // contributes to .text, for example. See PE/COFF spec 3.2. |
| 1134 | for (auto it : partialSections) { |
| 1135 | PartialSection *pSec = it.second; |
| 1136 | StringRef name = getOutputSectionName(name: pSec->name); |
| 1137 | uint32_t outChars = pSec->characteristics; |
| 1138 | |
| 1139 | if (name == ".CRT" ) { |
| 1140 | // In link.exe, there is a special case for the I386 target where .CRT |
| 1141 | // sections are treated as if they have output characteristics DATA | R if |
| 1142 | // their characteristics are DATA | R | W. This implements the same |
| 1143 | // special case for all architectures. |
| 1144 | outChars = data | r; |
| 1145 | |
| 1146 | Log(ctx) << "Processing section " << pSec->name << " -> " << name; |
| 1147 | |
| 1148 | sortCRTSectionChunks(chunks&: pSec->chunks); |
| 1149 | } |
| 1150 | |
| 1151 | // ARM64EC has specific placement and alignment requirements for the IAT. |
| 1152 | // Delay adding its chunks until appendECImportTables. |
| 1153 | if (isArm64EC(Machine: ctx.config.machine) && |
| 1154 | (pSec->name == ".idata$5" || pSec->name == ".idata$9" )) |
| 1155 | continue; |
| 1156 | |
| 1157 | OutputSection *sec = createSection(name, outChars); |
| 1158 | for (Chunk *c : pSec->chunks) |
| 1159 | sec->addChunk(c); |
| 1160 | |
| 1161 | sec->addContributingPartialSection(sec: pSec); |
| 1162 | } |
| 1163 | |
| 1164 | if (ctx.hybridSymtab) { |
| 1165 | if (OutputSection *sec = findSection(name: ".CRT" )) |
| 1166 | sec->splitECChunks(); |
| 1167 | } |
| 1168 | |
| 1169 | // Finally, move some output sections to the end. |
| 1170 | auto sectionOrder = [&](const OutputSection *s) { |
| 1171 | // Move DISCARDABLE (or non-memory-mapped) sections to the end of file |
| 1172 | // because the loader cannot handle holes. Stripping can remove other |
| 1173 | // discardable ones than .reloc, which is first of them (created early). |
| 1174 | if (s->header.Characteristics & IMAGE_SCN_MEM_DISCARDABLE) { |
| 1175 | // Move discardable sections named .debug_ to the end, after other |
| 1176 | // discardable sections. Stripping only removes the sections named |
| 1177 | // .debug_* - thus try to avoid leaving holes after stripping. |
| 1178 | if (s->name.starts_with(Prefix: ".debug_" )) |
| 1179 | return 3; |
| 1180 | return 2; |
| 1181 | } |
| 1182 | // .rsrc should come at the end of the non-discardable sections because its |
| 1183 | // size may change by the Win32 UpdateResources() function, causing |
| 1184 | // subsequent sections to move (see https://crbug.com/827082). |
| 1185 | if (s == rsrcSec) |
| 1186 | return 1; |
| 1187 | return 0; |
| 1188 | }; |
| 1189 | llvm::stable_sort(Range&: ctx.outputSections, |
| 1190 | C: [&](const OutputSection *s, const OutputSection *t) { |
| 1191 | return sectionOrder(s) < sectionOrder(t); |
| 1192 | }); |
| 1193 | } |
| 1194 | |
| 1195 | void Writer::createMiscChunks() { |
| 1196 | llvm::TimeTraceScope timeScope("Misc chunks" ); |
| 1197 | Configuration *config = &ctx.config; |
| 1198 | |
| 1199 | for (MergeChunk *p : ctx.mergeChunkInstances) { |
| 1200 | if (p) { |
| 1201 | p->finalizeContents(); |
| 1202 | rdataSec->addChunk(c: p); |
| 1203 | } |
| 1204 | } |
| 1205 | |
| 1206 | // Create thunks for locally-dllimported symbols. |
| 1207 | ctx.forEachSymtab(f: [&](SymbolTable &symtab) { |
| 1208 | if (!symtab.localImportChunks.empty()) { |
| 1209 | for (Chunk *c : symtab.localImportChunks) |
| 1210 | rdataSec->addChunk(c); |
| 1211 | } |
| 1212 | }); |
| 1213 | |
| 1214 | // Create Debug Information Chunks |
| 1215 | debugInfoSec = config->mingw ? buildidSec : rdataSec; |
| 1216 | if (config->buildIDHash != BuildIDHash::None || config->debug || |
| 1217 | config->repro || config->cetCompat) { |
| 1218 | debugDirectory = |
| 1219 | make<DebugDirectoryChunk>(args&: ctx, args&: debugRecords, args&: config->repro); |
| 1220 | debugDirectory->setAlignment(4); |
| 1221 | debugInfoSec->addChunk(c: debugDirectory); |
| 1222 | } |
| 1223 | |
| 1224 | if (config->debug || config->buildIDHash != BuildIDHash::None) { |
| 1225 | // Make a CVDebugRecordChunk even when /DEBUG:CV is not specified. We |
| 1226 | // output a PDB no matter what, and this chunk provides the only means of |
| 1227 | // allowing a debugger to match a PDB and an executable. So we need it even |
| 1228 | // if we're ultimately not going to write CodeView data to the PDB. |
| 1229 | buildId = make<CVDebugRecordChunk>(args&: ctx); |
| 1230 | debugRecords.emplace_back(args: COFF::IMAGE_DEBUG_TYPE_CODEVIEW, args&: buildId); |
| 1231 | ctx.forEachSymtab(f: [&](SymbolTable &symtab) { |
| 1232 | if (Symbol *buildidSym = symtab.findUnderscore(name: "__buildid" )) |
| 1233 | replaceSymbol<DefinedSynthetic>(s: buildidSym, arg: buildidSym->getName(), |
| 1234 | arg&: buildId, arg: 4); |
| 1235 | }); |
| 1236 | } |
| 1237 | |
| 1238 | if (config->cetCompat) { |
| 1239 | debugRecords.emplace_back(args: COFF::IMAGE_DEBUG_TYPE_EX_DLLCHARACTERISTICS, |
| 1240 | args: make<ExtendedDllCharacteristicsChunk>( |
| 1241 | args: IMAGE_DLL_CHARACTERISTICS_EX_CET_COMPAT)); |
| 1242 | } |
| 1243 | |
| 1244 | // Align and add each chunk referenced by the debug data directory. |
| 1245 | for (std::pair<COFF::DebugType, Chunk *> r : debugRecords) { |
| 1246 | r.second->setAlignment(4); |
| 1247 | debugInfoSec->addChunk(c: r.second); |
| 1248 | } |
| 1249 | |
| 1250 | // Create SEH table. x86-only. |
| 1251 | if (config->safeSEH) |
| 1252 | createSEHTable(); |
| 1253 | |
| 1254 | // Create /guard:cf tables if requested. |
| 1255 | createGuardCFTables(); |
| 1256 | |
| 1257 | createECChunks(); |
| 1258 | |
| 1259 | if (config->autoImport) |
| 1260 | createRuntimePseudoRelocs(); |
| 1261 | |
| 1262 | if (config->mingw) { |
| 1263 | insertCtorDtorSymbols(); |
| 1264 | insertBssDataStartEndSymbols(); |
| 1265 | } |
| 1266 | } |
| 1267 | |
| 1268 | // Create .idata section for the DLL-imported symbol table. |
| 1269 | // The format of this section is inherently Windows-specific. |
| 1270 | // IdataContents class abstracted away the details for us, |
| 1271 | // so we just let it create chunks and add them to the section. |
| 1272 | void Writer::createImportTables() { |
| 1273 | llvm::TimeTraceScope timeScope("Import tables" ); |
| 1274 | // Initialize DLLOrder so that import entries are ordered in |
| 1275 | // the same order as in the command line. (That affects DLL |
| 1276 | // initialization order, and this ordering is MSVC-compatible.) |
| 1277 | for (ImportFile *file : ctx.importFileInstances) { |
| 1278 | if (!file->live) |
| 1279 | continue; |
| 1280 | |
| 1281 | std::string dll = StringRef(file->dllName).lower(); |
| 1282 | ctx.config.dllOrder.try_emplace(k: dll, args: ctx.config.dllOrder.size()); |
| 1283 | |
| 1284 | if (file->impSym && !isa<DefinedImportData>(Val: file->impSym)) |
| 1285 | Fatal(ctx) << file->symtab.printSymbol(sym: file->impSym) << " was replaced" ; |
| 1286 | DefinedImportData *impSym = cast_or_null<DefinedImportData>(Val: file->impSym); |
| 1287 | if (ctx.config.delayLoads.count(x: StringRef(file->dllName).lower())) { |
| 1288 | if (!file->thunkSym) |
| 1289 | Fatal(ctx) << "cannot delay-load " << toString(file) |
| 1290 | << " due to import of data: " |
| 1291 | << file->symtab.printSymbol(sym: impSym); |
| 1292 | delayIdata.add(sym: impSym); |
| 1293 | } else { |
| 1294 | idata.add(sym: impSym); |
| 1295 | } |
| 1296 | } |
| 1297 | } |
| 1298 | |
| 1299 | void Writer::appendImportThunks() { |
| 1300 | if (ctx.importFileInstances.empty()) |
| 1301 | return; |
| 1302 | |
| 1303 | llvm::TimeTraceScope timeScope("Import thunks" ); |
| 1304 | for (ImportFile *file : ctx.importFileInstances) { |
| 1305 | if (!file->live) |
| 1306 | continue; |
| 1307 | |
| 1308 | if (file->thunkSym) { |
| 1309 | if (!isa<DefinedImportThunk>(Val: file->thunkSym)) |
| 1310 | Fatal(ctx) << file->symtab.printSymbol(sym: file->thunkSym) |
| 1311 | << " was replaced" ; |
| 1312 | auto *chunk = cast<DefinedImportThunk>(Val: file->thunkSym)->getChunk(); |
| 1313 | if (chunk->live) |
| 1314 | textSec->addChunk(c: chunk); |
| 1315 | } |
| 1316 | |
| 1317 | if (file->auxThunkSym) { |
| 1318 | if (!isa<DefinedImportThunk>(Val: file->auxThunkSym)) |
| 1319 | Fatal(ctx) << file->symtab.printSymbol(sym: file->auxThunkSym) |
| 1320 | << " was replaced" ; |
| 1321 | auto *chunk = cast<DefinedImportThunk>(Val: file->auxThunkSym)->getChunk(); |
| 1322 | if (chunk->live) |
| 1323 | textSec->addChunk(c: chunk); |
| 1324 | } |
| 1325 | |
| 1326 | if (file->impchkThunk) |
| 1327 | textSec->addChunk(c: file->impchkThunk); |
| 1328 | } |
| 1329 | |
| 1330 | if (!delayIdata.empty()) { |
| 1331 | delayIdata.create(); |
| 1332 | for (Chunk *c : delayIdata.getChunks()) |
| 1333 | didatSec->addChunk(c); |
| 1334 | for (Chunk *c : delayIdata.getDataChunks()) |
| 1335 | dataSec->addChunk(c); |
| 1336 | for (Chunk *c : delayIdata.getCodeChunks()) |
| 1337 | textSec->addChunk(c); |
| 1338 | for (Chunk *c : delayIdata.getCodePData()) |
| 1339 | pdataSec->addChunk(c); |
| 1340 | for (Chunk *c : delayIdata.getAuxIatCopy()) |
| 1341 | rdataSec->addChunk(c); |
| 1342 | for (Chunk *c : delayIdata.getCodeUnwindInfo()) |
| 1343 | rdataSec->addChunk(c); |
| 1344 | } |
| 1345 | } |
| 1346 | |
| 1347 | void Writer::createExportTable() { |
| 1348 | llvm::TimeTraceScope timeScope("Export table" ); |
| 1349 | if (!edataSec->chunks.empty()) { |
| 1350 | // Allow using a custom built export table from input object files, instead |
| 1351 | // of having the linker synthesize the tables. |
| 1352 | if (!ctx.hybridSymtab) { |
| 1353 | ctx.symtab.edataStart = edataSec->chunks.front(); |
| 1354 | ctx.symtab.edataEnd = edataSec->chunks.back(); |
| 1355 | } else { |
| 1356 | // On hybrid target, split EC and native chunks. |
| 1357 | llvm::stable_sort(Range&: edataSec->chunks, C: [=](const Chunk *a, const Chunk *b) { |
| 1358 | return (a->getMachine() != ARM64) < (b->getMachine() != ARM64); |
| 1359 | }); |
| 1360 | |
| 1361 | for (auto chunk : edataSec->chunks) { |
| 1362 | if (chunk->getMachine() != ARM64) { |
| 1363 | ctx.symtab.edataStart = chunk; |
| 1364 | ctx.symtab.edataEnd = edataSec->chunks.back(); |
| 1365 | break; |
| 1366 | } |
| 1367 | |
| 1368 | if (!ctx.hybridSymtab->edataStart) |
| 1369 | ctx.hybridSymtab->edataStart = chunk; |
| 1370 | ctx.hybridSymtab->edataEnd = chunk; |
| 1371 | } |
| 1372 | } |
| 1373 | } |
| 1374 | ctx.forEachActiveSymtab(f: [&](SymbolTable &symtab) { |
| 1375 | if (symtab.edataStart) { |
| 1376 | if (symtab.hadExplicitExports) |
| 1377 | Warn(ctx) << "literal .edata sections override exports" ; |
| 1378 | } else if (!symtab.exports.empty()) { |
| 1379 | std::vector<Chunk *> edataChunks; |
| 1380 | createEdataChunks(symtab, chunks&: edataChunks); |
| 1381 | for (Chunk *c : edataChunks) |
| 1382 | edataSec->addChunk(c); |
| 1383 | symtab.edataStart = edataChunks.front(); |
| 1384 | symtab.edataEnd = edataChunks.back(); |
| 1385 | } |
| 1386 | |
| 1387 | // Warn on exported deleting destructor. |
| 1388 | for (auto e : symtab.exports) |
| 1389 | if (e.sym && e.sym->getName().starts_with(Prefix: "??_G" )) |
| 1390 | Warn(ctx) << "export of deleting dtor: " << toString(ctx, b&: *e.sym); |
| 1391 | }); |
| 1392 | } |
| 1393 | |
| 1394 | void Writer::removeUnusedSections() { |
| 1395 | llvm::TimeTraceScope timeScope("Remove unused sections" ); |
| 1396 | // Remove sections that we can be sure won't get content, to avoid |
| 1397 | // allocating space for their section headers. |
| 1398 | auto isUnused = [this](OutputSection *s) { |
| 1399 | if (s == relocSec) |
| 1400 | return false; // This section is populated later. |
| 1401 | // MergeChunks have zero size at this point, as their size is finalized |
| 1402 | // later. Only remove sections that have no Chunks at all. |
| 1403 | return s->chunks.empty(); |
| 1404 | }; |
| 1405 | llvm::erase_if(C&: ctx.outputSections, P: isUnused); |
| 1406 | } |
| 1407 | |
| 1408 | // The Windows loader doesn't seem to like empty sections, |
| 1409 | // so we remove them if any. |
| 1410 | void Writer::removeEmptySections() { |
| 1411 | llvm::TimeTraceScope timeScope("Remove empty sections" ); |
| 1412 | auto isEmpty = [](OutputSection *s) { return s->getVirtualSize() == 0; }; |
| 1413 | llvm::erase_if(C&: ctx.outputSections, P: isEmpty); |
| 1414 | } |
| 1415 | |
| 1416 | void Writer::assignOutputSectionIndices() { |
| 1417 | llvm::TimeTraceScope timeScope("Output sections indices" ); |
| 1418 | // Assign final output section indices, and assign each chunk to its output |
| 1419 | // section. |
| 1420 | uint32_t idx = 1; |
| 1421 | for (OutputSection *os : ctx.outputSections) { |
| 1422 | os->sectionIndex = idx; |
| 1423 | for (Chunk *c : os->chunks) |
| 1424 | c->setOutputSectionIdx(idx); |
| 1425 | ++idx; |
| 1426 | } |
| 1427 | |
| 1428 | // Merge chunks are containers of chunks, so assign those an output section |
| 1429 | // too. |
| 1430 | for (MergeChunk *mc : ctx.mergeChunkInstances) |
| 1431 | if (mc) |
| 1432 | for (SectionChunk *sc : mc->sections) |
| 1433 | if (sc && sc->live) |
| 1434 | sc->setOutputSectionIdx(mc->getOutputSectionIdx()); |
| 1435 | } |
| 1436 | |
| 1437 | size_t Writer::addEntryToStringTable(StringRef str) { |
| 1438 | assert(str.size() > COFF::NameSize); |
| 1439 | size_t offsetOfEntry = strtab.size() + 4; // +4 for the size field |
| 1440 | strtab.insert(position: strtab.end(), first: str.begin(), last: str.end()); |
| 1441 | strtab.push_back(x: '\0'); |
| 1442 | return offsetOfEntry; |
| 1443 | } |
| 1444 | |
| 1445 | std::optional<coff_symbol16> Writer::createSymbol(Defined *def) { |
| 1446 | coff_symbol16 sym; |
| 1447 | switch (def->kind()) { |
| 1448 | case Symbol::DefinedAbsoluteKind: { |
| 1449 | auto *da = dyn_cast<DefinedAbsolute>(Val: def); |
| 1450 | // Note: COFF symbol can only store 32-bit values, so 64-bit absolute |
| 1451 | // values will be truncated. |
| 1452 | sym.Value = da->getVA(); |
| 1453 | sym.SectionNumber = IMAGE_SYM_ABSOLUTE; |
| 1454 | break; |
| 1455 | } |
| 1456 | default: { |
| 1457 | // Don't write symbols that won't be written to the output to the symbol |
| 1458 | // table. |
| 1459 | // We also try to write DefinedSynthetic as a normal symbol. Some of these |
| 1460 | // symbols do point to an actual chunk, like __safe_se_handler_table. Others |
| 1461 | // like __ImageBase are outside of sections and thus cannot be represented. |
| 1462 | Chunk *c = def->getChunk(); |
| 1463 | if (!c) |
| 1464 | return std::nullopt; |
| 1465 | OutputSection *os = ctx.getOutputSection(c); |
| 1466 | if (!os) |
| 1467 | return std::nullopt; |
| 1468 | |
| 1469 | sym.Value = def->getRVA() - os->getRVA(); |
| 1470 | sym.SectionNumber = os->sectionIndex; |
| 1471 | break; |
| 1472 | } |
| 1473 | } |
| 1474 | |
| 1475 | // Symbols that are runtime pseudo relocations don't point to the actual |
| 1476 | // symbol data itself (as they are imported), but points to the IAT entry |
| 1477 | // instead. Avoid emitting them to the symbol table, as they can confuse |
| 1478 | // debuggers. |
| 1479 | if (def->isRuntimePseudoReloc) |
| 1480 | return std::nullopt; |
| 1481 | |
| 1482 | StringRef name = def->getName(); |
| 1483 | if (name.size() > COFF::NameSize) { |
| 1484 | sym.Name.Offset.Zeroes = 0; |
| 1485 | sym.Name.Offset.Offset = addEntryToStringTable(str: name); |
| 1486 | } else { |
| 1487 | memset(s: sym.Name.ShortName, c: 0, n: COFF::NameSize); |
| 1488 | memcpy(dest: sym.Name.ShortName, src: name.data(), n: name.size()); |
| 1489 | } |
| 1490 | |
| 1491 | if (auto *d = dyn_cast<DefinedCOFF>(Val: def)) { |
| 1492 | COFFSymbolRef ref = d->getCOFFSymbol(); |
| 1493 | sym.Type = ref.getType(); |
| 1494 | sym.StorageClass = ref.getStorageClass(); |
| 1495 | } else if (def->kind() == Symbol::DefinedImportThunkKind) { |
| 1496 | sym.Type = (IMAGE_SYM_DTYPE_FUNCTION << SCT_COMPLEX_TYPE_SHIFT) | |
| 1497 | IMAGE_SYM_TYPE_NULL; |
| 1498 | sym.StorageClass = IMAGE_SYM_CLASS_EXTERNAL; |
| 1499 | } else { |
| 1500 | sym.Type = IMAGE_SYM_TYPE_NULL; |
| 1501 | sym.StorageClass = IMAGE_SYM_CLASS_EXTERNAL; |
| 1502 | } |
| 1503 | sym.NumberOfAuxSymbols = 0; |
| 1504 | return sym; |
| 1505 | } |
| 1506 | |
| 1507 | void Writer::createSymbolAndStringTable() { |
| 1508 | llvm::TimeTraceScope timeScope("Symbol and string table" ); |
| 1509 | // PE/COFF images are limited to 8 byte section names. Longer names can be |
| 1510 | // supported by writing a non-standard string table, but this string table is |
| 1511 | // not mapped at runtime and the long names will therefore be inaccessible. |
| 1512 | // link.exe always truncates section names to 8 bytes, whereas binutils always |
| 1513 | // preserves long section names via the string table. LLD adopts a hybrid |
| 1514 | // solution where discardable sections have long names preserved and |
| 1515 | // non-discardable sections have their names truncated, to ensure that any |
| 1516 | // section which is mapped at runtime also has its name mapped at runtime. |
| 1517 | for (OutputSection *sec : ctx.outputSections) { |
| 1518 | if (sec->name.size() <= COFF::NameSize) |
| 1519 | continue; |
| 1520 | if ((sec->header.Characteristics & IMAGE_SCN_MEM_DISCARDABLE) == 0) |
| 1521 | continue; |
| 1522 | if (ctx.config.warnLongSectionNames) { |
| 1523 | Warn(ctx) |
| 1524 | << "section name " << sec->name |
| 1525 | << " is longer than 8 characters and will use a non-standard string " |
| 1526 | "table" ; |
| 1527 | } |
| 1528 | sec->setStringTableOff(addEntryToStringTable(str: sec->name)); |
| 1529 | } |
| 1530 | |
| 1531 | if (ctx.config.writeSymtab) { |
| 1532 | for (ObjFile *file : ctx.objFileInstances) { |
| 1533 | for (Symbol *b : file->getSymbols()) { |
| 1534 | auto *d = dyn_cast_or_null<Defined>(Val: b); |
| 1535 | if (!d || d->writtenToSymtab) |
| 1536 | continue; |
| 1537 | d->writtenToSymtab = true; |
| 1538 | if (auto *dc = dyn_cast_or_null<DefinedCOFF>(Val: d)) { |
| 1539 | COFFSymbolRef symRef = dc->getCOFFSymbol(); |
| 1540 | if (symRef.isSectionDefinition() || |
| 1541 | symRef.getStorageClass() == COFF::IMAGE_SYM_CLASS_LABEL) |
| 1542 | continue; |
| 1543 | } |
| 1544 | |
| 1545 | if (std::optional<coff_symbol16> sym = createSymbol(def: d)) |
| 1546 | outputSymtab.push_back(x: *sym); |
| 1547 | |
| 1548 | if (auto *dthunk = dyn_cast<DefinedImportThunk>(Val: d)) { |
| 1549 | if (!dthunk->wrappedSym->writtenToSymtab) { |
| 1550 | dthunk->wrappedSym->writtenToSymtab = true; |
| 1551 | if (std::optional<coff_symbol16> sym = |
| 1552 | createSymbol(def: dthunk->wrappedSym)) |
| 1553 | outputSymtab.push_back(x: *sym); |
| 1554 | } |
| 1555 | } |
| 1556 | } |
| 1557 | } |
| 1558 | } |
| 1559 | |
| 1560 | if (outputSymtab.empty() && strtab.empty()) |
| 1561 | return; |
| 1562 | |
| 1563 | // We position the symbol table to be adjacent to the end of the last section. |
| 1564 | uint64_t fileOff = fileSize; |
| 1565 | pointerToSymbolTable = fileOff; |
| 1566 | fileOff += outputSymtab.size() * sizeof(coff_symbol16); |
| 1567 | fileOff += 4 + strtab.size(); |
| 1568 | fileSize = alignTo(Value: fileOff, Align: ctx.config.fileAlign); |
| 1569 | } |
| 1570 | |
| 1571 | void Writer::mergeSection(const std::map<StringRef, StringRef>::value_type &p) { |
| 1572 | StringRef toName = p.second; |
| 1573 | if (p.first == toName) |
| 1574 | return; |
| 1575 | StringSet<> names; |
| 1576 | while (true) { |
| 1577 | if (!names.insert(key: toName).second) |
| 1578 | Fatal(ctx) << "/merge: cycle found for section '" << p.first << "'" ; |
| 1579 | auto i = ctx.config.merge.find(x: toName); |
| 1580 | if (i == ctx.config.merge.end()) |
| 1581 | break; |
| 1582 | toName = i->second; |
| 1583 | } |
| 1584 | OutputSection *from = findSection(name: p.first); |
| 1585 | OutputSection *to = findSection(name: toName); |
| 1586 | if (!from) |
| 1587 | return; |
| 1588 | if (!to) { |
| 1589 | from->name = toName; |
| 1590 | return; |
| 1591 | } |
| 1592 | to->merge(other: from); |
| 1593 | } |
| 1594 | |
| 1595 | void Writer::mergeSections() { |
| 1596 | llvm::TimeTraceScope timeScope("Merge sections" ); |
| 1597 | if (!pdataSec->chunks.empty()) { |
| 1598 | if (isArm64EC(Machine: ctx.config.machine)) { |
| 1599 | // On ARM64EC .pdata may contain both ARM64 and X64 data. Split them by |
| 1600 | // sorting and store their regions separately. |
| 1601 | llvm::stable_sort(Range&: pdataSec->chunks, C: [=](const Chunk *a, const Chunk *b) { |
| 1602 | return (a->getMachine() == AMD64) < (b->getMachine() == AMD64); |
| 1603 | }); |
| 1604 | |
| 1605 | for (auto chunk : pdataSec->chunks) { |
| 1606 | if (chunk->getMachine() == AMD64) { |
| 1607 | hybridPdata.first = chunk; |
| 1608 | hybridPdata.last = pdataSec->chunks.back(); |
| 1609 | break; |
| 1610 | } |
| 1611 | |
| 1612 | if (!pdata.first) |
| 1613 | pdata.first = chunk; |
| 1614 | pdata.last = chunk; |
| 1615 | } |
| 1616 | } else { |
| 1617 | pdata.first = pdataSec->chunks.front(); |
| 1618 | pdata.last = pdataSec->chunks.back(); |
| 1619 | } |
| 1620 | } |
| 1621 | |
| 1622 | for (auto &p : ctx.config.merge) { |
| 1623 | if (p.first != ".bss" ) |
| 1624 | mergeSection(p); |
| 1625 | } |
| 1626 | |
| 1627 | // Because .bss contains all zeros, it should be merged at the end of |
| 1628 | // whatever section it is being merged into (usually .data) so that the image |
| 1629 | // need not actually contain all of the zeros. |
| 1630 | auto it = ctx.config.merge.find(x: ".bss" ); |
| 1631 | if (it != ctx.config.merge.end()) |
| 1632 | mergeSection(p: *it); |
| 1633 | } |
| 1634 | |
| 1635 | // EC targets may have chunks of various architectures mixed together at this |
| 1636 | // point. Group code chunks of the same architecture together by sorting chunks |
| 1637 | // by their EC range type. |
| 1638 | void Writer::sortECChunks() { |
| 1639 | if (!isArm64EC(Machine: ctx.config.machine)) |
| 1640 | return; |
| 1641 | |
| 1642 | for (OutputSection *sec : ctx.outputSections) { |
| 1643 | if (sec->isCodeSection()) |
| 1644 | llvm::stable_sort(Range&: sec->chunks, C: [=](const Chunk *a, const Chunk *b) { |
| 1645 | std::optional<chpe_range_type> aType = a->getArm64ECRangeType(), |
| 1646 | bType = b->getArm64ECRangeType(); |
| 1647 | return bType && (!aType || *aType < *bType); |
| 1648 | }); |
| 1649 | } |
| 1650 | } |
| 1651 | |
| 1652 | // Visits all sections to assign incremental, non-overlapping RVAs and |
| 1653 | // file offsets. |
| 1654 | void Writer::assignAddresses() { |
| 1655 | llvm::TimeTraceScope timeScope("Assign addresses" ); |
| 1656 | Configuration *config = &ctx.config; |
| 1657 | |
| 1658 | // We need to create EC code map so that ECCodeMapChunk knows its size. |
| 1659 | // We do it here to make sure that we account for range extension chunks. |
| 1660 | createECCodeMap(); |
| 1661 | |
| 1662 | sizeOfHeaders = dosStubSize + sizeof(PEMagic) + sizeof(coff_file_header) + |
| 1663 | sizeof(data_directory) * numberOfDataDirectory + |
| 1664 | sizeof(coff_section) * ctx.outputSections.size(); |
| 1665 | sizeOfHeaders += |
| 1666 | config->is64() ? sizeof(pe32plus_header) : sizeof(pe32_header); |
| 1667 | sizeOfHeaders = alignTo(Value: sizeOfHeaders, Align: config->fileAlign); |
| 1668 | fileSize = sizeOfHeaders; |
| 1669 | |
| 1670 | // The first page is kept unmapped. |
| 1671 | uint64_t rva = alignTo(Value: sizeOfHeaders, Align: config->align); |
| 1672 | |
| 1673 | for (OutputSection *sec : ctx.outputSections) { |
| 1674 | llvm::TimeTraceScope timeScope("Section: " , sec->name); |
| 1675 | if (sec == relocSec) { |
| 1676 | sec->chunks.clear(); |
| 1677 | addBaserels(); |
| 1678 | if (ctx.dynamicRelocs) { |
| 1679 | ctx.dynamicRelocs->finalize(); |
| 1680 | relocSec->addChunk(c: ctx.dynamicRelocs); |
| 1681 | } |
| 1682 | } |
| 1683 | uint64_t rawSize = 0, virtualSize = 0; |
| 1684 | sec->header.VirtualAddress = rva; |
| 1685 | |
| 1686 | // If /FUNCTIONPADMIN is used, functions are padded in order to create a |
| 1687 | // hotpatchable image. |
| 1688 | uint32_t padding = sec->isCodeSection() ? config->functionPadMin : 0; |
| 1689 | std::optional<chpe_range_type> prevECRange; |
| 1690 | |
| 1691 | for (Chunk *c : sec->chunks) { |
| 1692 | // Alignment EC code range baudaries. |
| 1693 | if (isArm64EC(Machine: ctx.config.machine) && sec->isCodeSection()) { |
| 1694 | std::optional<chpe_range_type> rangeType = c->getArm64ECRangeType(); |
| 1695 | if (rangeType != prevECRange) { |
| 1696 | virtualSize = alignTo(Value: virtualSize, Align: 4096); |
| 1697 | prevECRange = rangeType; |
| 1698 | } |
| 1699 | } |
| 1700 | if (padding && c->isHotPatchable()) |
| 1701 | virtualSize += padding; |
| 1702 | // If chunk has EC entry thunk, reserve a space for an offset to the |
| 1703 | // thunk. |
| 1704 | if (c->getEntryThunk()) |
| 1705 | virtualSize += sizeof(uint32_t); |
| 1706 | virtualSize = alignTo(Value: virtualSize, Align: c->getAlignment()); |
| 1707 | c->setRVA(rva + virtualSize); |
| 1708 | virtualSize += c->getSize(); |
| 1709 | if (c->hasData) |
| 1710 | rawSize = alignTo(Value: virtualSize, Align: config->fileAlign); |
| 1711 | } |
| 1712 | if (virtualSize > UINT32_MAX) |
| 1713 | Err(ctx) << "section larger than 4 GiB: " << sec->name; |
| 1714 | sec->header.VirtualSize = virtualSize; |
| 1715 | sec->header.SizeOfRawData = rawSize; |
| 1716 | if (rawSize != 0) |
| 1717 | sec->header.PointerToRawData = fileSize; |
| 1718 | rva += alignTo(Value: virtualSize, Align: config->align); |
| 1719 | fileSize += alignTo(Value: rawSize, Align: config->fileAlign); |
| 1720 | } |
| 1721 | sizeOfImage = alignTo(Value: rva, Align: config->align); |
| 1722 | |
| 1723 | // Assign addresses to sections in MergeChunks. |
| 1724 | for (MergeChunk *mc : ctx.mergeChunkInstances) |
| 1725 | if (mc) |
| 1726 | mc->assignSubsectionRVAs(); |
| 1727 | } |
| 1728 | |
| 1729 | template <typename PEHeaderTy> void Writer::() { |
| 1730 | // Write DOS header. For backwards compatibility, the first part of a PE/COFF |
| 1731 | // executable consists of an MS-DOS MZ executable. If the executable is run |
| 1732 | // under DOS, that program gets run (usually to just print an error message). |
| 1733 | // When run under Windows, the loader looks at AddressOfNewExeHeader and uses |
| 1734 | // the PE header instead. |
| 1735 | Configuration *config = &ctx.config; |
| 1736 | |
| 1737 | uint8_t *buf = buffer->getBufferStart(); |
| 1738 | auto *dos = reinterpret_cast<dos_header *>(buf); |
| 1739 | |
| 1740 | // Write DOS program. |
| 1741 | if (config->dosStub) { |
| 1742 | memcpy(dest: buf, src: config->dosStub->getBufferStart(), |
| 1743 | n: config->dosStub->getBufferSize()); |
| 1744 | // MS link.exe accepts an invalid `e_lfanew` (AddressOfNewExeHeader) and |
| 1745 | // updates it automatically. Replicate the same behaviour. |
| 1746 | dos->AddressOfNewExeHeader = alignTo(Value: config->dosStub->getBufferSize(), Align: 8); |
| 1747 | // Unlike MS link.exe, LLD accepts non-8-byte-aligned stubs. |
| 1748 | // In that case, we add zero paddings ourselves. |
| 1749 | buf += alignTo(Value: config->dosStub->getBufferSize(), Align: 8); |
| 1750 | } else { |
| 1751 | buf += sizeof(dos_header); |
| 1752 | dos->Magic[0] = 'M'; |
| 1753 | dos->Magic[1] = 'Z'; |
| 1754 | dos->UsedBytesInTheLastPage = dosStubSize % 512; |
| 1755 | dos->FileSizeInPages = divideCeil(Numerator: dosStubSize, Denominator: 512); |
| 1756 | dos->HeaderSizeInParagraphs = sizeof(dos_header) / 16; |
| 1757 | |
| 1758 | dos->AddressOfRelocationTable = sizeof(dos_header); |
| 1759 | dos->AddressOfNewExeHeader = dosStubSize; |
| 1760 | |
| 1761 | memcpy(dest: buf, src: dosProgram, n: sizeof(dosProgram)); |
| 1762 | buf += sizeof(dosProgram); |
| 1763 | } |
| 1764 | |
| 1765 | // Make sure DOS stub is aligned to 8 bytes at this point |
| 1766 | assert((buf - buffer->getBufferStart()) % 8 == 0); |
| 1767 | |
| 1768 | // Write PE magic |
| 1769 | memcpy(dest: buf, src: PEMagic, n: sizeof(PEMagic)); |
| 1770 | buf += sizeof(PEMagic); |
| 1771 | |
| 1772 | // Write COFF header |
| 1773 | assert(coffHeaderOffset == buf - buffer->getBufferStart()); |
| 1774 | auto *coff = reinterpret_cast<coff_file_header *>(buf); |
| 1775 | buf += sizeof(*coff); |
| 1776 | SymbolTable &symtab = |
| 1777 | ctx.config.machine == ARM64X ? *ctx.hybridSymtab : ctx.symtab; |
| 1778 | coff->Machine = symtab.isEC() ? AMD64 : symtab.machine; |
| 1779 | coff->NumberOfSections = ctx.outputSections.size(); |
| 1780 | coff->Characteristics = IMAGE_FILE_EXECUTABLE_IMAGE; |
| 1781 | if (config->largeAddressAware) |
| 1782 | coff->Characteristics |= IMAGE_FILE_LARGE_ADDRESS_AWARE; |
| 1783 | if (!config->is64()) |
| 1784 | coff->Characteristics |= IMAGE_FILE_32BIT_MACHINE; |
| 1785 | if (config->dll) |
| 1786 | coff->Characteristics |= IMAGE_FILE_DLL; |
| 1787 | if (config->driverUponly) |
| 1788 | coff->Characteristics |= IMAGE_FILE_UP_SYSTEM_ONLY; |
| 1789 | if (!config->relocatable) |
| 1790 | coff->Characteristics |= IMAGE_FILE_RELOCS_STRIPPED; |
| 1791 | if (config->swaprunCD) |
| 1792 | coff->Characteristics |= IMAGE_FILE_REMOVABLE_RUN_FROM_SWAP; |
| 1793 | if (config->swaprunNet) |
| 1794 | coff->Characteristics |= IMAGE_FILE_NET_RUN_FROM_SWAP; |
| 1795 | coff->SizeOfOptionalHeader = |
| 1796 | sizeof(PEHeaderTy) + sizeof(data_directory) * numberOfDataDirectory; |
| 1797 | |
| 1798 | // Write PE header |
| 1799 | assert(peHeaderOffset == buf - buffer->getBufferStart()); |
| 1800 | auto *pe = reinterpret_cast<PEHeaderTy *>(buf); |
| 1801 | buf += sizeof(*pe); |
| 1802 | pe->Magic = config->is64() ? PE32Header::PE32_PLUS : PE32Header::PE32; |
| 1803 | |
| 1804 | // If {Major,Minor}LinkerVersion is left at 0.0, then for some |
| 1805 | // reason signing the resulting PE file with Authenticode produces a |
| 1806 | // signature that fails to validate on Windows 7 (but is OK on 10). |
| 1807 | // Set it to 14.0, which is what VS2015 outputs, and which avoids |
| 1808 | // that problem. |
| 1809 | pe->MajorLinkerVersion = 14; |
| 1810 | pe->MinorLinkerVersion = 0; |
| 1811 | |
| 1812 | pe->ImageBase = config->imageBase; |
| 1813 | pe->SectionAlignment = config->align; |
| 1814 | pe->FileAlignment = config->fileAlign; |
| 1815 | pe->MajorImageVersion = config->majorImageVersion; |
| 1816 | pe->MinorImageVersion = config->minorImageVersion; |
| 1817 | pe->MajorOperatingSystemVersion = config->majorOSVersion; |
| 1818 | pe->MinorOperatingSystemVersion = config->minorOSVersion; |
| 1819 | pe->MajorSubsystemVersion = config->majorSubsystemVersion; |
| 1820 | pe->MinorSubsystemVersion = config->minorSubsystemVersion; |
| 1821 | pe->Subsystem = config->subsystem; |
| 1822 | pe->SizeOfImage = sizeOfImage; |
| 1823 | pe->SizeOfHeaders = sizeOfHeaders; |
| 1824 | if (!config->noEntry) { |
| 1825 | Defined *entry = cast<Defined>(Val: symtab.entry); |
| 1826 | pe->AddressOfEntryPoint = entry->getRVA(); |
| 1827 | // Pointer to thumb code must have the LSB set, so adjust it. |
| 1828 | if (config->machine == ARMNT) |
| 1829 | pe->AddressOfEntryPoint |= 1; |
| 1830 | } |
| 1831 | pe->SizeOfStackReserve = config->stackReserve; |
| 1832 | pe->SizeOfStackCommit = config->stackCommit; |
| 1833 | pe->SizeOfHeapReserve = config->heapReserve; |
| 1834 | pe->SizeOfHeapCommit = config->heapCommit; |
| 1835 | if (config->appContainer) |
| 1836 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_APPCONTAINER; |
| 1837 | if (config->driverWdm) |
| 1838 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_WDM_DRIVER; |
| 1839 | if (config->dynamicBase) |
| 1840 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_DYNAMIC_BASE; |
| 1841 | if (config->highEntropyVA) |
| 1842 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_HIGH_ENTROPY_VA; |
| 1843 | if (!config->allowBind) |
| 1844 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NO_BIND; |
| 1845 | if (config->nxCompat) |
| 1846 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NX_COMPAT; |
| 1847 | if (!config->allowIsolation) |
| 1848 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NO_ISOLATION; |
| 1849 | if (config->guardCF != GuardCFLevel::Off) |
| 1850 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_GUARD_CF; |
| 1851 | if (config->integrityCheck) |
| 1852 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_FORCE_INTEGRITY; |
| 1853 | if (setNoSEHCharacteristic || config->noSEH) |
| 1854 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NO_SEH; |
| 1855 | if (config->terminalServerAware) |
| 1856 | pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_TERMINAL_SERVER_AWARE; |
| 1857 | pe->NumberOfRvaAndSize = numberOfDataDirectory; |
| 1858 | if (textSec->getVirtualSize()) { |
| 1859 | pe->BaseOfCode = textSec->getRVA(); |
| 1860 | pe->SizeOfCode = textSec->getRawSize(); |
| 1861 | } |
| 1862 | pe->SizeOfInitializedData = getSizeOfInitializedData(); |
| 1863 | |
| 1864 | // Write data directory |
| 1865 | assert(!ctx.config.is64() || |
| 1866 | dataDirOffset64 == buf - buffer->getBufferStart()); |
| 1867 | auto *dir = reinterpret_cast<data_directory *>(buf); |
| 1868 | buf += sizeof(*dir) * numberOfDataDirectory; |
| 1869 | if (symtab.edataStart) { |
| 1870 | dir[EXPORT_TABLE].RelativeVirtualAddress = symtab.edataStart->getRVA(); |
| 1871 | dir[EXPORT_TABLE].Size = symtab.edataEnd->getRVA() + |
| 1872 | symtab.edataEnd->getSize() - |
| 1873 | symtab.edataStart->getRVA(); |
| 1874 | } |
| 1875 | if (importTableStart) { |
| 1876 | dir[IMPORT_TABLE].RelativeVirtualAddress = importTableStart->getRVA(); |
| 1877 | dir[IMPORT_TABLE].Size = importTableSize; |
| 1878 | } |
| 1879 | if (iatStart) { |
| 1880 | dir[IAT].RelativeVirtualAddress = iatStart->getRVA(); |
| 1881 | dir[IAT].Size = iatSize; |
| 1882 | } |
| 1883 | if (rsrcSec->getVirtualSize()) { |
| 1884 | dir[RESOURCE_TABLE].RelativeVirtualAddress = rsrcSec->getRVA(); |
| 1885 | dir[RESOURCE_TABLE].Size = rsrcSec->getVirtualSize(); |
| 1886 | } |
| 1887 | // ARM64EC (but not ARM64X) contains x86_64 exception table in data directory. |
| 1888 | ChunkRange &exceptionTable = |
| 1889 | ctx.config.machine == ARM64EC ? hybridPdata : pdata; |
| 1890 | if (exceptionTable.first) { |
| 1891 | dir[EXCEPTION_TABLE].RelativeVirtualAddress = |
| 1892 | exceptionTable.first->getRVA(); |
| 1893 | dir[EXCEPTION_TABLE].Size = exceptionTable.last->getRVA() + |
| 1894 | exceptionTable.last->getSize() - |
| 1895 | exceptionTable.first->getRVA(); |
| 1896 | } |
| 1897 | size_t relocSize = relocSec->getVirtualSize(); |
| 1898 | if (ctx.dynamicRelocs) |
| 1899 | relocSize -= ctx.dynamicRelocs->getSize(); |
| 1900 | if (relocSize) { |
| 1901 | dir[BASE_RELOCATION_TABLE].RelativeVirtualAddress = relocSec->getRVA(); |
| 1902 | dir[BASE_RELOCATION_TABLE].Size = relocSize; |
| 1903 | } |
| 1904 | if (Symbol *sym = symtab.findUnderscore(name: "_tls_used" )) { |
| 1905 | if (Defined *b = dyn_cast<Defined>(Val: sym)) { |
| 1906 | dir[TLS_TABLE].RelativeVirtualAddress = b->getRVA(); |
| 1907 | dir[TLS_TABLE].Size = config->is64() |
| 1908 | ? sizeof(object::coff_tls_directory64) |
| 1909 | : sizeof(object::coff_tls_directory32); |
| 1910 | } |
| 1911 | } |
| 1912 | if (debugDirectory) { |
| 1913 | dir[DEBUG_DIRECTORY].RelativeVirtualAddress = debugDirectory->getRVA(); |
| 1914 | dir[DEBUG_DIRECTORY].Size = debugDirectory->getSize(); |
| 1915 | } |
| 1916 | if (symtab.loadConfigSym) { |
| 1917 | dir[LOAD_CONFIG_TABLE].RelativeVirtualAddress = |
| 1918 | symtab.loadConfigSym->getRVA(); |
| 1919 | dir[LOAD_CONFIG_TABLE].Size = symtab.loadConfigSize; |
| 1920 | } |
| 1921 | if (!delayIdata.empty()) { |
| 1922 | dir[DELAY_IMPORT_DESCRIPTOR].RelativeVirtualAddress = |
| 1923 | delayIdata.getDirRVA(); |
| 1924 | dir[DELAY_IMPORT_DESCRIPTOR].Size = delayIdata.getDirSize(); |
| 1925 | } |
| 1926 | |
| 1927 | // Write section table |
| 1928 | for (OutputSection *sec : ctx.outputSections) { |
| 1929 | sec->writeHeaderTo(buf, isDebug: config->debug); |
| 1930 | buf += sizeof(coff_section); |
| 1931 | } |
| 1932 | sectionTable = ArrayRef<uint8_t>( |
| 1933 | buf - ctx.outputSections.size() * sizeof(coff_section), buf); |
| 1934 | |
| 1935 | if (outputSymtab.empty() && strtab.empty()) |
| 1936 | return; |
| 1937 | |
| 1938 | coff->PointerToSymbolTable = pointerToSymbolTable; |
| 1939 | uint32_t numberOfSymbols = outputSymtab.size(); |
| 1940 | coff->NumberOfSymbols = numberOfSymbols; |
| 1941 | auto *symbolTable = reinterpret_cast<coff_symbol16 *>( |
| 1942 | buffer->getBufferStart() + coff->PointerToSymbolTable); |
| 1943 | for (size_t i = 0; i != numberOfSymbols; ++i) |
| 1944 | symbolTable[i] = outputSymtab[i]; |
| 1945 | // Create the string table, it follows immediately after the symbol table. |
| 1946 | // The first 4 bytes is length including itself. |
| 1947 | buf = reinterpret_cast<uint8_t *>(&symbolTable[numberOfSymbols]); |
| 1948 | write32le(P: buf, V: strtab.size() + 4); |
| 1949 | if (!strtab.empty()) |
| 1950 | memcpy(dest: buf + 4, src: strtab.data(), n: strtab.size()); |
| 1951 | } |
| 1952 | |
| 1953 | void Writer::openFile(StringRef path) { |
| 1954 | buffer = CHECK( |
| 1955 | FileOutputBuffer::create(path, fileSize, FileOutputBuffer::F_executable), |
| 1956 | "failed to open " + path); |
| 1957 | } |
| 1958 | |
| 1959 | void Writer::createSEHTable() { |
| 1960 | SymbolRVASet handlers; |
| 1961 | for (ObjFile *file : ctx.objFileInstances) { |
| 1962 | if (!file->hasSafeSEH()) |
| 1963 | Err(ctx) << "/safeseh: " << file->getName() |
| 1964 | << " is not compatible with SEH" ; |
| 1965 | markSymbolsForRVATable(file, symIdxChunks: file->getSXDataChunks(), tableSymbols&: handlers); |
| 1966 | } |
| 1967 | |
| 1968 | // Set the "no SEH" characteristic if there really were no handlers, or if |
| 1969 | // there is no load config object to point to the table of handlers. |
| 1970 | setNoSEHCharacteristic = |
| 1971 | handlers.empty() || !ctx.symtab.findUnderscore(name: "_load_config_used" ); |
| 1972 | |
| 1973 | maybeAddRVATable(tableSymbols: std::move(handlers), tableSym: "__safe_se_handler_table" , |
| 1974 | countSym: "__safe_se_handler_count" ); |
| 1975 | } |
| 1976 | |
| 1977 | // Add a symbol to an RVA set. Two symbols may have the same RVA, but an RVA set |
| 1978 | // cannot contain duplicates. Therefore, the set is uniqued by Chunk and the |
| 1979 | // symbol's offset into that Chunk. |
| 1980 | static void addSymbolToRVASet(SymbolRVASet &rvaSet, Defined *s) { |
| 1981 | Chunk *c = s->getChunk(); |
| 1982 | if (!c) |
| 1983 | return; |
| 1984 | if (auto *sc = dyn_cast<SectionChunk>(Val: c)) |
| 1985 | c = sc->repl; // Look through ICF replacement. |
| 1986 | uint32_t off = s->getRVA() - (c ? c->getRVA() : 0); |
| 1987 | rvaSet.insert(V: {.inputChunk: c, .offset: off}); |
| 1988 | } |
| 1989 | |
| 1990 | // Given a symbol, add it to the GFIDs table if it is a live, defined, function |
| 1991 | // symbol in an executable section. |
| 1992 | static void maybeAddAddressTakenFunction(SymbolRVASet &addressTakenSyms, |
| 1993 | Symbol *s) { |
| 1994 | if (!s) |
| 1995 | return; |
| 1996 | |
| 1997 | switch (s->kind()) { |
| 1998 | case Symbol::DefinedLocalImportKind: |
| 1999 | case Symbol::DefinedImportDataKind: |
| 2000 | // Defines an __imp_ pointer, so it is data, so it is ignored. |
| 2001 | break; |
| 2002 | case Symbol::DefinedCommonKind: |
| 2003 | // Common is always data, so it is ignored. |
| 2004 | break; |
| 2005 | case Symbol::DefinedAbsoluteKind: |
| 2006 | // Absolute is never code, synthetic generally isn't and usually isn't |
| 2007 | // determinable. |
| 2008 | break; |
| 2009 | case Symbol::DefinedSyntheticKind: |
| 2010 | // For EC export thunks, mark both the thunk itself and its target. |
| 2011 | if (auto expChunk = dyn_cast_or_null<ECExportThunkChunk>( |
| 2012 | Val: cast<Defined>(Val: s)->getChunk())) { |
| 2013 | addSymbolToRVASet(rvaSet&: addressTakenSyms, s: cast<Defined>(Val: s)); |
| 2014 | addSymbolToRVASet(rvaSet&: addressTakenSyms, s: expChunk->target); |
| 2015 | } |
| 2016 | break; |
| 2017 | case Symbol::LazyArchiveKind: |
| 2018 | case Symbol::LazyObjectKind: |
| 2019 | case Symbol::LazyDLLSymbolKind: |
| 2020 | case Symbol::UndefinedKind: |
| 2021 | // Undefined symbols resolve to zero, so they don't have an RVA. Lazy |
| 2022 | // symbols shouldn't have relocations. |
| 2023 | break; |
| 2024 | |
| 2025 | case Symbol::DefinedImportThunkKind: |
| 2026 | // Thunks are always code, include them. |
| 2027 | addSymbolToRVASet(rvaSet&: addressTakenSyms, s: cast<Defined>(Val: s)); |
| 2028 | break; |
| 2029 | |
| 2030 | case Symbol::DefinedRegularKind: { |
| 2031 | // This is a regular, defined, symbol from a COFF file. Mark the symbol as |
| 2032 | // address taken if the symbol type is function and it's in an executable |
| 2033 | // section. |
| 2034 | auto *d = cast<DefinedRegular>(Val: s); |
| 2035 | if (d->getCOFFSymbol().getComplexType() == COFF::IMAGE_SYM_DTYPE_FUNCTION) { |
| 2036 | SectionChunk *sc = dyn_cast<SectionChunk>(Val: d->getChunk()); |
| 2037 | if (sc && sc->live && |
| 2038 | sc->getOutputCharacteristics() & IMAGE_SCN_MEM_EXECUTE) |
| 2039 | addSymbolToRVASet(rvaSet&: addressTakenSyms, s: d); |
| 2040 | } |
| 2041 | break; |
| 2042 | } |
| 2043 | } |
| 2044 | } |
| 2045 | |
| 2046 | // Visit all relocations from all section contributions of this object file and |
| 2047 | // mark the relocation target as address-taken. |
| 2048 | void Writer::markSymbolsWithRelocations(ObjFile *file, |
| 2049 | SymbolRVASet &usedSymbols) { |
| 2050 | for (Chunk *c : file->getChunks()) { |
| 2051 | // We only care about live section chunks. Common chunks and other chunks |
| 2052 | // don't generally contain relocations. |
| 2053 | SectionChunk *sc = dyn_cast<SectionChunk>(Val: c); |
| 2054 | if (!sc || !sc->live) |
| 2055 | continue; |
| 2056 | |
| 2057 | for (const coff_relocation &reloc : sc->getRelocs()) { |
| 2058 | if (ctx.config.machine == I386 && |
| 2059 | reloc.Type == COFF::IMAGE_REL_I386_REL32) |
| 2060 | // Ignore relative relocations on x86. On x86_64 they can't be ignored |
| 2061 | // since they're also used to compute absolute addresses. |
| 2062 | continue; |
| 2063 | |
| 2064 | Symbol *ref = sc->file->getSymbol(symbolIndex: reloc.SymbolTableIndex); |
| 2065 | maybeAddAddressTakenFunction(addressTakenSyms&: usedSymbols, s: ref); |
| 2066 | } |
| 2067 | } |
| 2068 | } |
| 2069 | |
| 2070 | // Create the guard function id table. This is a table of RVAs of all |
| 2071 | // address-taken functions. It is sorted and uniqued, just like the safe SEH |
| 2072 | // table. |
| 2073 | void Writer::createGuardCFTables() { |
| 2074 | Configuration *config = &ctx.config; |
| 2075 | |
| 2076 | if (config->guardCF == GuardCFLevel::Off) { |
| 2077 | // MSVC marks the entire image as instrumented if any input object was built |
| 2078 | // with /guard:cf. |
| 2079 | for (ObjFile *file : ctx.objFileInstances) { |
| 2080 | if (file->hasGuardCF()) { |
| 2081 | ctx.forEachSymtab(f: [&](SymbolTable &symtab) { |
| 2082 | Symbol *flagSym = symtab.findUnderscore(name: "__guard_flags" ); |
| 2083 | cast<DefinedAbsolute>(Val: flagSym)->setVA( |
| 2084 | uint32_t(GuardFlags::CF_INSTRUMENTED)); |
| 2085 | }); |
| 2086 | break; |
| 2087 | } |
| 2088 | } |
| 2089 | return; |
| 2090 | } |
| 2091 | |
| 2092 | SymbolRVASet addressTakenSyms; |
| 2093 | SymbolRVASet giatsRVASet; |
| 2094 | std::vector<Symbol *> giatsSymbols; |
| 2095 | SymbolRVASet longJmpTargets; |
| 2096 | SymbolRVASet ehContTargets; |
| 2097 | for (ObjFile *file : ctx.objFileInstances) { |
| 2098 | // If the object was compiled with /guard:cf, the address taken symbols |
| 2099 | // are in .gfids$y sections, and the longjmp targets are in .gljmp$y |
| 2100 | // sections. If the object was not compiled with /guard:cf, we assume there |
| 2101 | // were no setjmp targets, and that all code symbols with relocations are |
| 2102 | // possibly address-taken. |
| 2103 | if (file->hasGuardCF()) { |
| 2104 | markSymbolsForRVATable(file, symIdxChunks: file->getGuardFidChunks(), tableSymbols&: addressTakenSyms); |
| 2105 | markSymbolsForRVATable(file, symIdxChunks: file->getGuardIATChunks(), tableSymbols&: giatsRVASet); |
| 2106 | getSymbolsFromSections(file, symIdxChunks: file->getGuardIATChunks(), symbols&: giatsSymbols); |
| 2107 | markSymbolsForRVATable(file, symIdxChunks: file->getGuardLJmpChunks(), tableSymbols&: longJmpTargets); |
| 2108 | } else { |
| 2109 | markSymbolsWithRelocations(file, usedSymbols&: addressTakenSyms); |
| 2110 | } |
| 2111 | // If the object was compiled with /guard:ehcont, the ehcont targets are in |
| 2112 | // .gehcont$y sections. |
| 2113 | if (file->hasGuardEHCont()) |
| 2114 | markSymbolsForRVATable(file, symIdxChunks: file->getGuardEHContChunks(), tableSymbols&: ehContTargets); |
| 2115 | } |
| 2116 | |
| 2117 | // Mark the image entry as address-taken. |
| 2118 | ctx.forEachSymtab(f: [&](SymbolTable &symtab) { |
| 2119 | if (symtab.entry) |
| 2120 | maybeAddAddressTakenFunction(addressTakenSyms, s: symtab.entry); |
| 2121 | |
| 2122 | // Mark exported symbols in executable sections as address-taken. |
| 2123 | for (Export &e : symtab.exports) |
| 2124 | maybeAddAddressTakenFunction(addressTakenSyms, s: e.sym); |
| 2125 | }); |
| 2126 | |
| 2127 | // For each entry in the .giats table, check if it has a corresponding load |
| 2128 | // thunk (e.g. because the DLL that defines it will be delay-loaded) and, if |
| 2129 | // so, add the load thunk to the address taken (.gfids) table. |
| 2130 | for (Symbol *s : giatsSymbols) { |
| 2131 | if (auto *di = dyn_cast<DefinedImportData>(Val: s)) { |
| 2132 | if (di->loadThunkSym) |
| 2133 | addSymbolToRVASet(rvaSet&: addressTakenSyms, s: di->loadThunkSym); |
| 2134 | } |
| 2135 | } |
| 2136 | |
| 2137 | // Ensure sections referenced in the gfid table are 16-byte aligned. |
| 2138 | for (const ChunkAndOffset &c : addressTakenSyms) |
| 2139 | if (c.inputChunk->getAlignment() < 16) |
| 2140 | c.inputChunk->setAlignment(16); |
| 2141 | |
| 2142 | maybeAddRVATable(tableSymbols: std::move(addressTakenSyms), tableSym: "__guard_fids_table" , |
| 2143 | countSym: "__guard_fids_count" ); |
| 2144 | |
| 2145 | // Add the Guard Address Taken IAT Entry Table (.giats). |
| 2146 | maybeAddRVATable(tableSymbols: std::move(giatsRVASet), tableSym: "__guard_iat_table" , |
| 2147 | countSym: "__guard_iat_count" ); |
| 2148 | |
| 2149 | // Add the longjmp target table unless the user told us not to. |
| 2150 | if (config->guardCF & GuardCFLevel::LongJmp) |
| 2151 | maybeAddRVATable(tableSymbols: std::move(longJmpTargets), tableSym: "__guard_longjmp_table" , |
| 2152 | countSym: "__guard_longjmp_count" ); |
| 2153 | |
| 2154 | // Add the ehcont target table unless the user told us not to. |
| 2155 | if (config->guardCF & GuardCFLevel::EHCont) |
| 2156 | maybeAddRVATable(tableSymbols: std::move(ehContTargets), tableSym: "__guard_eh_cont_table" , |
| 2157 | countSym: "__guard_eh_cont_count" ); |
| 2158 | |
| 2159 | // Set __guard_flags, which will be used in the load config to indicate that |
| 2160 | // /guard:cf was enabled. |
| 2161 | uint32_t guardFlags = uint32_t(GuardFlags::CF_INSTRUMENTED) | |
| 2162 | uint32_t(GuardFlags::CF_FUNCTION_TABLE_PRESENT); |
| 2163 | if (config->guardCF & GuardCFLevel::LongJmp) |
| 2164 | guardFlags |= uint32_t(GuardFlags::CF_LONGJUMP_TABLE_PRESENT); |
| 2165 | if (config->guardCF & GuardCFLevel::EHCont) |
| 2166 | guardFlags |= uint32_t(GuardFlags::EH_CONTINUATION_TABLE_PRESENT); |
| 2167 | ctx.forEachSymtab(f: [guardFlags](SymbolTable &symtab) { |
| 2168 | Symbol *flagSym = symtab.findUnderscore(name: "__guard_flags" ); |
| 2169 | cast<DefinedAbsolute>(Val: flagSym)->setVA(guardFlags); |
| 2170 | }); |
| 2171 | } |
| 2172 | |
| 2173 | // Take a list of input sections containing symbol table indices and add those |
| 2174 | // symbols to a vector. The challenge is that symbol RVAs are not known and |
| 2175 | // depend on the table size, so we can't directly build a set of integers. |
| 2176 | void Writer::getSymbolsFromSections(ObjFile *file, |
| 2177 | ArrayRef<SectionChunk *> symIdxChunks, |
| 2178 | std::vector<Symbol *> &symbols) { |
| 2179 | for (SectionChunk *c : symIdxChunks) { |
| 2180 | // Skip sections discarded by linker GC. This comes up when a .gfids section |
| 2181 | // is associated with something like a vtable and the vtable is discarded. |
| 2182 | // In this case, the associated gfids section is discarded, and we don't |
| 2183 | // mark the virtual member functions as address-taken by the vtable. |
| 2184 | if (!c->live) |
| 2185 | continue; |
| 2186 | |
| 2187 | // Validate that the contents look like symbol table indices. |
| 2188 | ArrayRef<uint8_t> data = c->getContents(); |
| 2189 | if (data.size() % 4 != 0) { |
| 2190 | Warn(ctx) << "ignoring " << c->getSectionName() |
| 2191 | << " symbol table index section in object " << file; |
| 2192 | continue; |
| 2193 | } |
| 2194 | |
| 2195 | // Read each symbol table index and check if that symbol was included in the |
| 2196 | // final link. If so, add it to the vector of symbols. |
| 2197 | ArrayRef<ulittle32_t> symIndices( |
| 2198 | reinterpret_cast<const ulittle32_t *>(data.data()), data.size() / 4); |
| 2199 | ArrayRef<Symbol *> objSymbols = file->getSymbols(); |
| 2200 | for (uint32_t symIndex : symIndices) { |
| 2201 | if (symIndex >= objSymbols.size()) { |
| 2202 | Warn(ctx) << "ignoring invalid symbol table index in section " |
| 2203 | << c->getSectionName() << " in object " << file; |
| 2204 | continue; |
| 2205 | } |
| 2206 | if (Symbol *s = objSymbols[symIndex]) { |
| 2207 | if (s->isLive()) |
| 2208 | symbols.push_back(x: cast<Symbol>(Val: s)); |
| 2209 | } |
| 2210 | } |
| 2211 | } |
| 2212 | } |
| 2213 | |
| 2214 | // Take a list of input sections containing symbol table indices and add those |
| 2215 | // symbols to an RVA table. |
| 2216 | void Writer::markSymbolsForRVATable(ObjFile *file, |
| 2217 | ArrayRef<SectionChunk *> symIdxChunks, |
| 2218 | SymbolRVASet &tableSymbols) { |
| 2219 | std::vector<Symbol *> syms; |
| 2220 | getSymbolsFromSections(file, symIdxChunks, symbols&: syms); |
| 2221 | |
| 2222 | for (Symbol *s : syms) |
| 2223 | addSymbolToRVASet(rvaSet&: tableSymbols, s: cast<Defined>(Val: s)); |
| 2224 | } |
| 2225 | |
| 2226 | // Replace the absolute table symbol with a synthetic symbol pointing to |
| 2227 | // tableChunk so that we can emit base relocations for it and resolve section |
| 2228 | // relative relocations. |
| 2229 | void Writer::maybeAddRVATable(SymbolRVASet tableSymbols, StringRef tableSym, |
| 2230 | StringRef countSym, bool hasFlag) { |
| 2231 | if (tableSymbols.empty()) |
| 2232 | return; |
| 2233 | |
| 2234 | NonSectionChunk *tableChunk; |
| 2235 | if (hasFlag) |
| 2236 | tableChunk = make<RVAFlagTableChunk>(args: std::move(tableSymbols)); |
| 2237 | else |
| 2238 | tableChunk = make<RVATableChunk>(args: std::move(tableSymbols)); |
| 2239 | rdataSec->addChunk(c: tableChunk); |
| 2240 | |
| 2241 | ctx.forEachSymtab(f: [&](SymbolTable &symtab) { |
| 2242 | Symbol *t = symtab.findUnderscore(name: tableSym); |
| 2243 | Symbol *c = symtab.findUnderscore(name: countSym); |
| 2244 | replaceSymbol<DefinedSynthetic>(s: t, arg: t->getName(), arg&: tableChunk); |
| 2245 | cast<DefinedAbsolute>(Val: c)->setVA(tableChunk->getSize() / (hasFlag ? 5 : 4)); |
| 2246 | }); |
| 2247 | } |
| 2248 | |
| 2249 | // Create CHPE metadata chunks. |
| 2250 | void Writer::createECChunks() { |
| 2251 | if (!ctx.symtab.isEC()) |
| 2252 | return; |
| 2253 | |
| 2254 | for (Symbol *s : ctx.symtab.expSymbols) { |
| 2255 | auto sym = dyn_cast<Defined>(Val: s); |
| 2256 | if (!sym || !sym->getChunk()) |
| 2257 | continue; |
| 2258 | if (auto thunk = dyn_cast<ECExportThunkChunk>(Val: sym->getChunk())) { |
| 2259 | hexpthkSec->addChunk(c: thunk); |
| 2260 | exportThunks.push_back(x: {thunk, thunk->target}); |
| 2261 | } else if (auto def = dyn_cast<DefinedRegular>(Val: sym)) { |
| 2262 | // Allow section chunk to be treated as an export thunk if it looks like |
| 2263 | // one. |
| 2264 | SectionChunk *chunk = def->getChunk(); |
| 2265 | if (!chunk->live || chunk->getMachine() != AMD64) |
| 2266 | continue; |
| 2267 | assert(sym->getName().starts_with("EXP+" )); |
| 2268 | StringRef targetName = sym->getName().substr(Start: strlen(s: "EXP+" )); |
| 2269 | // If EXP+#foo is an export thunk of a hybrid patchable function, |
| 2270 | // we should use the #foo$hp_target symbol as the redirection target. |
| 2271 | // First, try to look up the $hp_target symbol. If it can't be found, |
| 2272 | // assume it's a regular function and look for #foo instead. |
| 2273 | Symbol *targetSym = ctx.symtab.find(name: (targetName + "$hp_target" ).str()); |
| 2274 | if (!targetSym) |
| 2275 | targetSym = ctx.symtab.find(name: targetName); |
| 2276 | Defined *t = dyn_cast_or_null<Defined>(Val: targetSym); |
| 2277 | if (t && isArm64EC(Machine: t->getChunk()->getMachine())) |
| 2278 | exportThunks.push_back(x: {chunk, t}); |
| 2279 | } |
| 2280 | } |
| 2281 | |
| 2282 | auto codeMapChunk = make<ECCodeMapChunk>(args&: codeMap); |
| 2283 | rdataSec->addChunk(c: codeMapChunk); |
| 2284 | Symbol *codeMapSym = ctx.symtab.findUnderscore(name: "__hybrid_code_map" ); |
| 2285 | replaceSymbol<DefinedSynthetic>(s: codeMapSym, arg: codeMapSym->getName(), |
| 2286 | arg&: codeMapChunk); |
| 2287 | |
| 2288 | CHPECodeRangesChunk *ranges = make<CHPECodeRangesChunk>(args&: exportThunks); |
| 2289 | rdataSec->addChunk(c: ranges); |
| 2290 | Symbol *rangesSym = |
| 2291 | ctx.symtab.findUnderscore(name: "__x64_code_ranges_to_entry_points" ); |
| 2292 | replaceSymbol<DefinedSynthetic>(s: rangesSym, arg: rangesSym->getName(), arg&: ranges); |
| 2293 | |
| 2294 | CHPERedirectionChunk *entryPoints = make<CHPERedirectionChunk>(args&: exportThunks); |
| 2295 | a64xrmSec->addChunk(c: entryPoints); |
| 2296 | Symbol *entryPointsSym = |
| 2297 | ctx.symtab.findUnderscore(name: "__arm64x_redirection_metadata" ); |
| 2298 | replaceSymbol<DefinedSynthetic>(s: entryPointsSym, arg: entryPointsSym->getName(), |
| 2299 | arg&: entryPoints); |
| 2300 | } |
| 2301 | |
| 2302 | // MinGW specific. Gather all relocations that are imported from a DLL even |
| 2303 | // though the code didn't expect it to, produce the table that the runtime |
| 2304 | // uses for fixing them up, and provide the synthetic symbols that the |
| 2305 | // runtime uses for finding the table. |
| 2306 | void Writer::createRuntimePseudoRelocs() { |
| 2307 | ctx.forEachSymtab(f: [&](SymbolTable &symtab) { |
| 2308 | std::vector<RuntimePseudoReloc> rels; |
| 2309 | |
| 2310 | for (Chunk *c : ctx.driver.getChunks()) { |
| 2311 | auto *sc = dyn_cast<SectionChunk>(Val: c); |
| 2312 | if (!sc || !sc->live || &sc->file->symtab != &symtab) |
| 2313 | continue; |
| 2314 | // Don't create pseudo relocations for sections that won't be |
| 2315 | // mapped at runtime. |
| 2316 | if (sc->header->Characteristics & IMAGE_SCN_MEM_DISCARDABLE) |
| 2317 | continue; |
| 2318 | sc->getRuntimePseudoRelocs(res&: rels); |
| 2319 | } |
| 2320 | |
| 2321 | if (!ctx.config.pseudoRelocs) { |
| 2322 | // Not writing any pseudo relocs; if some were needed, error out and |
| 2323 | // indicate what required them. |
| 2324 | for (const RuntimePseudoReloc &rpr : rels) |
| 2325 | Err(ctx) << "automatic dllimport of " << rpr.sym->getName() << " in " |
| 2326 | << toString(file: rpr.target->file) |
| 2327 | << " requires pseudo relocations" ; |
| 2328 | return; |
| 2329 | } |
| 2330 | |
| 2331 | if (!rels.empty()) { |
| 2332 | Log(ctx) << "Writing " << Twine(rels.size()) |
| 2333 | << " runtime pseudo relocations" ; |
| 2334 | const char *symbolName = "_pei386_runtime_relocator" ; |
| 2335 | Symbol *relocator = symtab.findUnderscore(name: symbolName); |
| 2336 | if (!relocator) |
| 2337 | Err(ctx) |
| 2338 | << "output image has runtime pseudo relocations, but the function " |
| 2339 | << symbolName |
| 2340 | << " is missing; it is needed for fixing the relocations at " |
| 2341 | "runtime" ; |
| 2342 | } |
| 2343 | |
| 2344 | PseudoRelocTableChunk *table = make<PseudoRelocTableChunk>(args&: rels); |
| 2345 | rdataSec->addChunk(c: table); |
| 2346 | EmptyChunk *endOfList = make<EmptyChunk>(); |
| 2347 | rdataSec->addChunk(c: endOfList); |
| 2348 | |
| 2349 | Symbol *headSym = symtab.findUnderscore(name: "__RUNTIME_PSEUDO_RELOC_LIST__" ); |
| 2350 | Symbol *endSym = symtab.findUnderscore(name: "__RUNTIME_PSEUDO_RELOC_LIST_END__" ); |
| 2351 | replaceSymbol<DefinedSynthetic>(s: headSym, arg: headSym->getName(), arg&: table); |
| 2352 | replaceSymbol<DefinedSynthetic>(s: endSym, arg: endSym->getName(), arg&: endOfList); |
| 2353 | }); |
| 2354 | } |
| 2355 | |
| 2356 | // MinGW specific. |
| 2357 | // The MinGW .ctors and .dtors lists have sentinels at each end; |
| 2358 | // a (uintptr_t)-1 at the start and a (uintptr_t)0 at the end. |
| 2359 | // There's a symbol pointing to the start sentinel pointer, __CTOR_LIST__ |
| 2360 | // and __DTOR_LIST__ respectively. |
| 2361 | void Writer::insertCtorDtorSymbols() { |
| 2362 | ctx.forEachSymtab(f: [&](SymbolTable &symtab) { |
| 2363 | AbsolutePointerChunk *ctorListHead = make<AbsolutePointerChunk>(args&: symtab, args: -1); |
| 2364 | AbsolutePointerChunk *ctorListEnd = make<AbsolutePointerChunk>(args&: symtab, args: 0); |
| 2365 | AbsolutePointerChunk *dtorListHead = make<AbsolutePointerChunk>(args&: symtab, args: -1); |
| 2366 | AbsolutePointerChunk *dtorListEnd = make<AbsolutePointerChunk>(args&: symtab, args: 0); |
| 2367 | ctorsSec->insertChunkAtStart(c: ctorListHead); |
| 2368 | ctorsSec->addChunk(c: ctorListEnd); |
| 2369 | dtorsSec->insertChunkAtStart(c: dtorListHead); |
| 2370 | dtorsSec->addChunk(c: dtorListEnd); |
| 2371 | |
| 2372 | Symbol *ctorListSym = symtab.findUnderscore(name: "__CTOR_LIST__" ); |
| 2373 | Symbol *dtorListSym = symtab.findUnderscore(name: "__DTOR_LIST__" ); |
| 2374 | replaceSymbol<DefinedSynthetic>(s: ctorListSym, arg: ctorListSym->getName(), |
| 2375 | arg&: ctorListHead); |
| 2376 | replaceSymbol<DefinedSynthetic>(s: dtorListSym, arg: dtorListSym->getName(), |
| 2377 | arg&: dtorListHead); |
| 2378 | }); |
| 2379 | |
| 2380 | if (ctx.hybridSymtab) { |
| 2381 | ctorsSec->splitECChunks(); |
| 2382 | dtorsSec->splitECChunks(); |
| 2383 | } |
| 2384 | } |
| 2385 | |
| 2386 | // MinGW (really, Cygwin) specific. |
| 2387 | // The Cygwin startup code uses __data_start__ __data_end__ __bss_start__ |
| 2388 | // and __bss_end__ to know what to copy during fork emulation. |
| 2389 | void Writer::insertBssDataStartEndSymbols() { |
| 2390 | if (!dataSec->chunks.empty()) { |
| 2391 | Symbol *dataStartSym = ctx.symtab.find(name: "__data_start__" ); |
| 2392 | Symbol *dataEndSym = ctx.symtab.find(name: "__data_end__" ); |
| 2393 | Chunk *endChunk = dataSec->chunks.back(); |
| 2394 | replaceSymbol<DefinedSynthetic>(s: dataStartSym, arg: dataStartSym->getName(), |
| 2395 | arg&: dataSec->chunks.front()); |
| 2396 | replaceSymbol<DefinedSynthetic>(s: dataEndSym, arg: dataEndSym->getName(), arg&: endChunk, |
| 2397 | arg: endChunk->getSize()); |
| 2398 | } |
| 2399 | |
| 2400 | if (!bssSec->chunks.empty()) { |
| 2401 | Symbol *bssStartSym = ctx.symtab.find(name: "__bss_start__" ); |
| 2402 | Symbol *bssEndSym = ctx.symtab.find(name: "__bss_end__" ); |
| 2403 | Chunk *endChunk = bssSec->chunks.back(); |
| 2404 | replaceSymbol<DefinedSynthetic>(s: bssStartSym, arg: bssStartSym->getName(), |
| 2405 | arg&: bssSec->chunks.front()); |
| 2406 | replaceSymbol<DefinedSynthetic>(s: bssEndSym, arg: bssEndSym->getName(), arg&: endChunk, |
| 2407 | arg: endChunk->getSize()); |
| 2408 | } |
| 2409 | } |
| 2410 | |
| 2411 | // Handles /section options to allow users to overwrite |
| 2412 | // section attributes. |
| 2413 | void Writer::setSectionPermissions() { |
| 2414 | llvm::TimeTraceScope timeScope("Sections permissions" ); |
| 2415 | for (auto &p : ctx.config.section) { |
| 2416 | StringRef name = p.first; |
| 2417 | uint32_t perm = p.second; |
| 2418 | for (OutputSection *sec : ctx.outputSections) |
| 2419 | if (sec->name == name) |
| 2420 | sec->setPermissions(perm); |
| 2421 | } |
| 2422 | } |
| 2423 | |
| 2424 | // Set symbols used by ARM64EC metadata. |
| 2425 | void Writer::setECSymbols() { |
| 2426 | if (!ctx.symtab.isEC()) |
| 2427 | return; |
| 2428 | |
| 2429 | llvm::stable_sort(Range&: exportThunks, C: [](const std::pair<Chunk *, Defined *> &a, |
| 2430 | const std::pair<Chunk *, Defined *> &b) { |
| 2431 | return a.first->getRVA() < b.first->getRVA(); |
| 2432 | }); |
| 2433 | |
| 2434 | ChunkRange &chpePdata = ctx.config.machine == ARM64X ? hybridPdata : pdata; |
| 2435 | Symbol *rfeTableSym = ctx.symtab.findUnderscore(name: "__arm64x_extra_rfe_table" ); |
| 2436 | replaceSymbol<DefinedSynthetic>(s: rfeTableSym, arg: "__arm64x_extra_rfe_table" , |
| 2437 | arg&: chpePdata.first); |
| 2438 | |
| 2439 | if (chpePdata.first) { |
| 2440 | Symbol *rfeSizeSym = |
| 2441 | ctx.symtab.findUnderscore(name: "__arm64x_extra_rfe_table_size" ); |
| 2442 | cast<DefinedAbsolute>(Val: rfeSizeSym) |
| 2443 | ->setVA(chpePdata.last->getRVA() + chpePdata.last->getSize() - |
| 2444 | chpePdata.first->getRVA()); |
| 2445 | } |
| 2446 | |
| 2447 | Symbol *rangesCountSym = |
| 2448 | ctx.symtab.findUnderscore(name: "__x64_code_ranges_to_entry_points_count" ); |
| 2449 | cast<DefinedAbsolute>(Val: rangesCountSym)->setVA(exportThunks.size()); |
| 2450 | |
| 2451 | Symbol *entryPointCountSym = |
| 2452 | ctx.symtab.findUnderscore(name: "__arm64x_redirection_metadata_count" ); |
| 2453 | cast<DefinedAbsolute>(Val: entryPointCountSym)->setVA(exportThunks.size()); |
| 2454 | |
| 2455 | Symbol *iatSym = ctx.symtab.findUnderscore(name: "__hybrid_auxiliary_iat" ); |
| 2456 | replaceSymbol<DefinedSynthetic>(s: iatSym, arg: "__hybrid_auxiliary_iat" , |
| 2457 | arg: idata.auxIat.empty() ? nullptr |
| 2458 | : idata.auxIat.front()); |
| 2459 | |
| 2460 | Symbol *iatCopySym = ctx.symtab.findUnderscore(name: "__hybrid_auxiliary_iat_copy" ); |
| 2461 | replaceSymbol<DefinedSynthetic>( |
| 2462 | s: iatCopySym, arg: "__hybrid_auxiliary_iat_copy" , |
| 2463 | arg: idata.auxIatCopy.empty() ? nullptr : idata.auxIatCopy.front()); |
| 2464 | |
| 2465 | Symbol *delayIatSym = |
| 2466 | ctx.symtab.findUnderscore(name: "__hybrid_auxiliary_delayload_iat" ); |
| 2467 | replaceSymbol<DefinedSynthetic>( |
| 2468 | s: delayIatSym, arg: "__hybrid_auxiliary_delayload_iat" , |
| 2469 | arg: delayIdata.getAuxIat().empty() ? nullptr |
| 2470 | : delayIdata.getAuxIat().front()); |
| 2471 | |
| 2472 | Symbol *delayIatCopySym = |
| 2473 | ctx.symtab.findUnderscore(name: "__hybrid_auxiliary_delayload_iat_copy" ); |
| 2474 | replaceSymbol<DefinedSynthetic>( |
| 2475 | s: delayIatCopySym, arg: "__hybrid_auxiliary_delayload_iat_copy" , |
| 2476 | arg: delayIdata.getAuxIatCopy().empty() ? nullptr |
| 2477 | : delayIdata.getAuxIatCopy().front()); |
| 2478 | |
| 2479 | if (ctx.config.machine == ARM64X) { |
| 2480 | // For the hybrid image, set the alternate entry point to the EC entry |
| 2481 | // point. In the hybrid view, it is swapped to the native entry point |
| 2482 | // using ARM64X relocations. |
| 2483 | if (auto altEntrySym = cast_or_null<Defined>(Val: ctx.symtab.entry)) { |
| 2484 | // If the entry is an EC export thunk, use its target instead. |
| 2485 | if (auto thunkChunk = |
| 2486 | dyn_cast<ECExportThunkChunk>(Val: altEntrySym->getChunk())) |
| 2487 | altEntrySym = thunkChunk->target; |
| 2488 | ctx.symtab.findUnderscore(name: "__arm64x_native_entrypoint" ) |
| 2489 | ->replaceKeepingName(other: altEntrySym, size: sizeof(SymbolUnion)); |
| 2490 | } |
| 2491 | |
| 2492 | if (ctx.symtab.edataStart) |
| 2493 | ctx.dynamicRelocs->set( |
| 2494 | rva: dataDirOffset64 + EXPORT_TABLE * sizeof(data_directory) + |
| 2495 | offsetof(data_directory, Size), |
| 2496 | value: ctx.symtab.edataEnd->getRVA() - ctx.symtab.edataStart->getRVA() + |
| 2497 | ctx.symtab.edataEnd->getSize()); |
| 2498 | if (hybridPdata.first) { |
| 2499 | ctx.dynamicRelocs->set( |
| 2500 | rva: dataDirOffset64 + EXCEPTION_TABLE * sizeof(data_directory) + |
| 2501 | offsetof(data_directory, Size), |
| 2502 | value: hybridPdata.last->getRVA() - hybridPdata.first->getRVA() + |
| 2503 | hybridPdata.last->getSize()); |
| 2504 | if (chpeSym) { |
| 2505 | size_t size = 0; |
| 2506 | if (pdata.first) |
| 2507 | size = pdata.last->getRVA() + pdata.last->getSize() - |
| 2508 | pdata.first->getRVA(); |
| 2509 | ctx.dynamicRelocs->set(rva: chpeSym->getRVA() + |
| 2510 | offsetof(chpe_metadata, ExtraRFETableSize), |
| 2511 | value: size); |
| 2512 | } |
| 2513 | } |
| 2514 | } |
| 2515 | } |
| 2516 | |
| 2517 | // Write section contents to a mmap'ed file. |
| 2518 | void Writer::writeSections() { |
| 2519 | llvm::TimeTraceScope timeScope("Write sections" ); |
| 2520 | uint8_t *buf = buffer->getBufferStart(); |
| 2521 | for (OutputSection *sec : ctx.outputSections) { |
| 2522 | uint8_t *secBuf = buf + sec->getFileOff(); |
| 2523 | // Fill gaps between functions in .text with INT3 instructions |
| 2524 | // instead of leaving as NUL bytes (which can be interpreted as |
| 2525 | // ADD instructions). Only fill the gaps between chunks. Most |
| 2526 | // chunks overwrite it anyway, but uninitialized data chunks |
| 2527 | // merged into a code section don't. |
| 2528 | if ((sec->header.Characteristics & IMAGE_SCN_CNT_CODE) && |
| 2529 | (ctx.config.machine == AMD64 || ctx.config.machine == I386)) { |
| 2530 | uint32_t prevEnd = 0; |
| 2531 | for (Chunk *c : sec->chunks) { |
| 2532 | uint32_t off = c->getRVA() - sec->getRVA(); |
| 2533 | memset(s: secBuf + prevEnd, c: 0xCC, n: off - prevEnd); |
| 2534 | prevEnd = off + c->getSize(); |
| 2535 | } |
| 2536 | memset(s: secBuf + prevEnd, c: 0xCC, n: sec->getRawSize() - prevEnd); |
| 2537 | } |
| 2538 | |
| 2539 | parallelForEach(R&: sec->chunks, Fn: [&](Chunk *c) { |
| 2540 | c->writeTo(buf: secBuf + c->getRVA() - sec->getRVA()); |
| 2541 | }); |
| 2542 | } |
| 2543 | } |
| 2544 | |
| 2545 | void Writer::writeBuildId() { |
| 2546 | llvm::TimeTraceScope timeScope("Write build ID" ); |
| 2547 | |
| 2548 | // There are two important parts to the build ID. |
| 2549 | // 1) If building with debug info, the COFF debug directory contains a |
| 2550 | // timestamp as well as a Guid and Age of the PDB. |
| 2551 | // 2) In all cases, the PE COFF file header also contains a timestamp. |
| 2552 | // For reproducibility, instead of a timestamp we want to use a hash of the |
| 2553 | // PE contents. |
| 2554 | Configuration *config = &ctx.config; |
| 2555 | bool generateSyntheticBuildId = config->buildIDHash == BuildIDHash::Binary; |
| 2556 | if (generateSyntheticBuildId) { |
| 2557 | assert(buildId && "BuildId is not set!" ); |
| 2558 | // BuildId->BuildId was filled in when the PDB was written. |
| 2559 | } |
| 2560 | |
| 2561 | // At this point the only fields in the COFF file which remain unset are the |
| 2562 | // "timestamp" in the COFF file header, and the ones in the coff debug |
| 2563 | // directory. Now we can hash the file and write that hash to the various |
| 2564 | // timestamp fields in the file. |
| 2565 | StringRef outputFileData( |
| 2566 | reinterpret_cast<const char *>(buffer->getBufferStart()), |
| 2567 | buffer->getBufferSize()); |
| 2568 | |
| 2569 | uint32_t timestamp = config->timestamp; |
| 2570 | uint64_t hash = 0; |
| 2571 | |
| 2572 | if (config->repro || generateSyntheticBuildId) |
| 2573 | hash = xxh3_64bits(data: outputFileData); |
| 2574 | |
| 2575 | if (config->repro) |
| 2576 | timestamp = static_cast<uint32_t>(hash); |
| 2577 | |
| 2578 | if (generateSyntheticBuildId) { |
| 2579 | buildId->buildId->PDB70.CVSignature = OMF::Signature::PDB70; |
| 2580 | buildId->buildId->PDB70.Age = 1; |
| 2581 | memcpy(dest: buildId->buildId->PDB70.Signature, src: &hash, n: 8); |
| 2582 | // xxhash only gives us 8 bytes, so put some fixed data in the other half. |
| 2583 | memcpy(dest: &buildId->buildId->PDB70.Signature[8], src: "LLD PDB." , n: 8); |
| 2584 | } |
| 2585 | |
| 2586 | if (debugDirectory) |
| 2587 | debugDirectory->setTimeDateStamp(timestamp); |
| 2588 | |
| 2589 | uint8_t *buf = buffer->getBufferStart(); |
| 2590 | buf += dosStubSize + sizeof(PEMagic); |
| 2591 | object::coff_file_header * = |
| 2592 | reinterpret_cast<coff_file_header *>(buf); |
| 2593 | coffHeader->TimeDateStamp = timestamp; |
| 2594 | } |
| 2595 | |
| 2596 | // Sort .pdata section contents according to PE/COFF spec 5.5. |
| 2597 | template <typename T> |
| 2598 | void Writer::sortExceptionTable(ChunkRange &exceptionTable) { |
| 2599 | if (!exceptionTable.first) |
| 2600 | return; |
| 2601 | |
| 2602 | // We assume .pdata contains function table entries only. |
| 2603 | auto bufAddr = [&](Chunk *c) { |
| 2604 | OutputSection *os = ctx.getOutputSection(c); |
| 2605 | return buffer->getBufferStart() + os->getFileOff() + c->getRVA() - |
| 2606 | os->getRVA(); |
| 2607 | }; |
| 2608 | uint8_t *begin = bufAddr(exceptionTable.first); |
| 2609 | uint8_t *end = bufAddr(exceptionTable.last) + exceptionTable.last->getSize(); |
| 2610 | if ((end - begin) % sizeof(T) != 0) { |
| 2611 | Fatal(ctx) << "unexpected .pdata size: " << (end - begin) |
| 2612 | << " is not a multiple of " << sizeof(T); |
| 2613 | } |
| 2614 | |
| 2615 | parallelSort(MutableArrayRef<T>(reinterpret_cast<T *>(begin), |
| 2616 | reinterpret_cast<T *>(end)), |
| 2617 | [](const T &a, const T &b) { return a.begin < b.begin; }); |
| 2618 | } |
| 2619 | |
| 2620 | // Sort .pdata section contents according to PE/COFF spec 5.5. |
| 2621 | void Writer::sortExceptionTables() { |
| 2622 | llvm::TimeTraceScope timeScope("Sort exception table" ); |
| 2623 | |
| 2624 | struct EntryX64 { |
| 2625 | ulittle32_t begin, end, unwind; |
| 2626 | }; |
| 2627 | struct EntryArm { |
| 2628 | ulittle32_t begin, unwind; |
| 2629 | }; |
| 2630 | |
| 2631 | switch (ctx.config.machine) { |
| 2632 | case AMD64: |
| 2633 | sortExceptionTable<EntryX64>(exceptionTable&: pdata); |
| 2634 | break; |
| 2635 | case ARM64EC: |
| 2636 | case ARM64X: |
| 2637 | sortExceptionTable<EntryX64>(exceptionTable&: hybridPdata); |
| 2638 | [[fallthrough]]; |
| 2639 | case ARMNT: |
| 2640 | case ARM64: |
| 2641 | sortExceptionTable<EntryArm>(exceptionTable&: pdata); |
| 2642 | break; |
| 2643 | default: |
| 2644 | if (pdata.first) |
| 2645 | ctx.e.errs() << "warning: don't know how to handle .pdata\n" ; |
| 2646 | break; |
| 2647 | } |
| 2648 | } |
| 2649 | |
| 2650 | // The CRT section contains, among other things, the array of function |
| 2651 | // pointers that initialize every global variable that is not trivially |
| 2652 | // constructed. The CRT calls them one after the other prior to invoking |
| 2653 | // main(). |
| 2654 | // |
| 2655 | // As per C++ spec, 3.6.2/2.3, |
| 2656 | // "Variables with ordered initialization defined within a single |
| 2657 | // translation unit shall be initialized in the order of their definitions |
| 2658 | // in the translation unit" |
| 2659 | // |
| 2660 | // It is therefore critical to sort the chunks containing the function |
| 2661 | // pointers in the order that they are listed in the object file (top to |
| 2662 | // bottom), otherwise global objects might not be initialized in the |
| 2663 | // correct order. |
| 2664 | void Writer::sortCRTSectionChunks(std::vector<Chunk *> &chunks) { |
| 2665 | auto sectionChunkOrder = [](const Chunk *a, const Chunk *b) { |
| 2666 | auto sa = dyn_cast<SectionChunk>(Val: a); |
| 2667 | auto sb = dyn_cast<SectionChunk>(Val: b); |
| 2668 | assert(sa && sb && "Non-section chunks in CRT section!" ); |
| 2669 | |
| 2670 | StringRef sAObj = sa->file->mb.getBufferIdentifier(); |
| 2671 | StringRef sBObj = sb->file->mb.getBufferIdentifier(); |
| 2672 | |
| 2673 | return sAObj == sBObj && sa->getSectionNumber() < sb->getSectionNumber(); |
| 2674 | }; |
| 2675 | llvm::stable_sort(Range&: chunks, C: sectionChunkOrder); |
| 2676 | |
| 2677 | if (ctx.config.verbose) { |
| 2678 | for (auto &c : chunks) { |
| 2679 | auto sc = dyn_cast<SectionChunk>(Val: c); |
| 2680 | Log(ctx) << " " << sc->file->mb.getBufferIdentifier().str() |
| 2681 | << ", SectionID: " << sc->getSectionNumber(); |
| 2682 | } |
| 2683 | } |
| 2684 | } |
| 2685 | |
| 2686 | OutputSection *Writer::findSection(StringRef name) { |
| 2687 | for (OutputSection *sec : ctx.outputSections) |
| 2688 | if (sec->name == name) |
| 2689 | return sec; |
| 2690 | return nullptr; |
| 2691 | } |
| 2692 | |
| 2693 | uint32_t Writer::getSizeOfInitializedData() { |
| 2694 | uint32_t res = 0; |
| 2695 | for (OutputSection *s : ctx.outputSections) |
| 2696 | if (s->header.Characteristics & IMAGE_SCN_CNT_INITIALIZED_DATA) |
| 2697 | res += s->getRawSize(); |
| 2698 | return res; |
| 2699 | } |
| 2700 | |
| 2701 | // Add base relocations to .reloc section. |
| 2702 | void Writer::addBaserels() { |
| 2703 | if (!ctx.config.relocatable) |
| 2704 | return; |
| 2705 | std::vector<Baserel> v; |
| 2706 | for (OutputSection *sec : ctx.outputSections) { |
| 2707 | if (sec->header.Characteristics & IMAGE_SCN_MEM_DISCARDABLE) |
| 2708 | continue; |
| 2709 | llvm::TimeTraceScope timeScope("Base relocations: " , sec->name); |
| 2710 | // Collect all locations for base relocations. |
| 2711 | for (Chunk *c : sec->chunks) |
| 2712 | c->getBaserels(res: &v); |
| 2713 | // Add the addresses to .reloc section. |
| 2714 | if (!v.empty()) |
| 2715 | addBaserelBlocks(v); |
| 2716 | v.clear(); |
| 2717 | } |
| 2718 | } |
| 2719 | |
| 2720 | // Add addresses to .reloc section. Note that addresses are grouped by page. |
| 2721 | void Writer::addBaserelBlocks(std::vector<Baserel> &v) { |
| 2722 | const uint32_t mask = ~uint32_t(pageSize - 1); |
| 2723 | uint32_t page = v[0].rva & mask; |
| 2724 | size_t i = 0, j = 1; |
| 2725 | llvm::sort(C&: v, |
| 2726 | Comp: [](const Baserel &x, const Baserel &y) { return x.rva < y.rva; }); |
| 2727 | for (size_t e = v.size(); j < e; ++j) { |
| 2728 | uint32_t p = v[j].rva & mask; |
| 2729 | if (p == page) |
| 2730 | continue; |
| 2731 | relocSec->addChunk(c: make<BaserelChunk>(args&: page, args: &v[i], args: &v[0] + j)); |
| 2732 | i = j; |
| 2733 | page = p; |
| 2734 | } |
| 2735 | if (i == j) |
| 2736 | return; |
| 2737 | relocSec->addChunk(c: make<BaserelChunk>(args&: page, args: &v[i], args: &v[0] + j)); |
| 2738 | } |
| 2739 | |
| 2740 | void Writer::createDynamicRelocs() { |
| 2741 | if (!ctx.dynamicRelocs) |
| 2742 | return; |
| 2743 | |
| 2744 | // Adjust the Machine field in the COFF header to AMD64. |
| 2745 | ctx.dynamicRelocs->add(type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint16_t), |
| 2746 | offset: coffHeaderOffset + offsetof(coff_file_header, Machine), |
| 2747 | value: AMD64); |
| 2748 | |
| 2749 | if (ctx.symtab.entry != ctx.hybridSymtab->entry || |
| 2750 | pdata.first != hybridPdata.first) { |
| 2751 | chpeSym = cast_or_null<DefinedRegular>( |
| 2752 | Val: ctx.symtab.findUnderscore(name: "__chpe_metadata" )); |
| 2753 | if (!chpeSym) |
| 2754 | Warn(ctx) << "'__chpe_metadata' is missing for ARM64X target" ; |
| 2755 | } |
| 2756 | |
| 2757 | if (ctx.symtab.entry != ctx.hybridSymtab->entry) { |
| 2758 | ctx.dynamicRelocs->add(type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2759 | offset: peHeaderOffset + |
| 2760 | offsetof(pe32plus_header, AddressOfEntryPoint), |
| 2761 | value: cast_or_null<Defined>(Val: ctx.symtab.entry)); |
| 2762 | |
| 2763 | // Swap the alternate entry point in the CHPE metadata. |
| 2764 | if (chpeSym) |
| 2765 | ctx.dynamicRelocs->add( |
| 2766 | type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2767 | offset: Arm64XRelocVal(chpeSym, offsetof(chpe_metadata, AlternateEntryPoint)), |
| 2768 | value: cast_or_null<Defined>(Val: ctx.hybridSymtab->entry)); |
| 2769 | } |
| 2770 | |
| 2771 | if (ctx.symtab.edataStart != ctx.hybridSymtab->edataStart) { |
| 2772 | ctx.dynamicRelocs->add(type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2773 | offset: dataDirOffset64 + |
| 2774 | EXPORT_TABLE * sizeof(data_directory) + |
| 2775 | offsetof(data_directory, RelativeVirtualAddress), |
| 2776 | value: ctx.symtab.edataStart); |
| 2777 | // The Size value is assigned after addresses are finalized. |
| 2778 | ctx.dynamicRelocs->add(type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2779 | offset: dataDirOffset64 + |
| 2780 | EXPORT_TABLE * sizeof(data_directory) + |
| 2781 | offsetof(data_directory, Size)); |
| 2782 | } |
| 2783 | |
| 2784 | if (pdata.first != hybridPdata.first) { |
| 2785 | ctx.dynamicRelocs->add(type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2786 | offset: dataDirOffset64 + |
| 2787 | EXCEPTION_TABLE * sizeof(data_directory) + |
| 2788 | offsetof(data_directory, RelativeVirtualAddress), |
| 2789 | value: hybridPdata.first); |
| 2790 | // The Size value is assigned after addresses are finalized. |
| 2791 | ctx.dynamicRelocs->add(type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2792 | offset: dataDirOffset64 + |
| 2793 | EXCEPTION_TABLE * sizeof(data_directory) + |
| 2794 | offsetof(data_directory, Size)); |
| 2795 | |
| 2796 | // Swap ExtraRFETable in the CHPE metadata. |
| 2797 | if (chpeSym) { |
| 2798 | ctx.dynamicRelocs->add( |
| 2799 | type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2800 | offset: Arm64XRelocVal(chpeSym, offsetof(chpe_metadata, ExtraRFETable)), |
| 2801 | value: pdata.first); |
| 2802 | // The Size value is assigned after addresses are finalized. |
| 2803 | ctx.dynamicRelocs->add( |
| 2804 | type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2805 | offset: Arm64XRelocVal(chpeSym, offsetof(chpe_metadata, ExtraRFETableSize))); |
| 2806 | } |
| 2807 | } |
| 2808 | |
| 2809 | // Set the hybrid load config to the EC load config. |
| 2810 | ctx.dynamicRelocs->add(type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2811 | offset: dataDirOffset64 + |
| 2812 | LOAD_CONFIG_TABLE * sizeof(data_directory) + |
| 2813 | offsetof(data_directory, RelativeVirtualAddress), |
| 2814 | value: ctx.symtab.loadConfigSym); |
| 2815 | ctx.dynamicRelocs->add(type: IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE, size: sizeof(uint32_t), |
| 2816 | offset: dataDirOffset64 + |
| 2817 | LOAD_CONFIG_TABLE * sizeof(data_directory) + |
| 2818 | offsetof(data_directory, Size), |
| 2819 | value: ctx.symtab.loadConfigSize); |
| 2820 | } |
| 2821 | |
| 2822 | PartialSection *Writer::createPartialSection(StringRef name, |
| 2823 | uint32_t outChars) { |
| 2824 | PartialSection *&pSec = partialSections[{.name: name, .characteristics: outChars}]; |
| 2825 | if (pSec) |
| 2826 | return pSec; |
| 2827 | pSec = make<PartialSection>(args&: name, args&: outChars); |
| 2828 | return pSec; |
| 2829 | } |
| 2830 | |
| 2831 | PartialSection *Writer::findPartialSection(StringRef name, uint32_t outChars) { |
| 2832 | auto it = partialSections.find(x: {.name: name, .characteristics: outChars}); |
| 2833 | if (it != partialSections.end()) |
| 2834 | return it->second; |
| 2835 | return nullptr; |
| 2836 | } |
| 2837 | |
| 2838 | void Writer::fixTlsAlignment() { |
| 2839 | Defined *tlsSym = |
| 2840 | dyn_cast_or_null<Defined>(Val: ctx.symtab.findUnderscore(name: "_tls_used" )); |
| 2841 | if (!tlsSym) |
| 2842 | return; |
| 2843 | |
| 2844 | OutputSection *sec = ctx.getOutputSection(c: tlsSym->getChunk()); |
| 2845 | assert(sec && tlsSym->getRVA() >= sec->getRVA() && |
| 2846 | "no output section for _tls_used" ); |
| 2847 | |
| 2848 | uint8_t *secBuf = buffer->getBufferStart() + sec->getFileOff(); |
| 2849 | uint64_t tlsOffset = tlsSym->getRVA() - sec->getRVA(); |
| 2850 | uint64_t directorySize = ctx.config.is64() |
| 2851 | ? sizeof(object::coff_tls_directory64) |
| 2852 | : sizeof(object::coff_tls_directory32); |
| 2853 | |
| 2854 | if (tlsOffset + directorySize > sec->getRawSize()) |
| 2855 | Fatal(ctx) << "_tls_used sym is malformed" ; |
| 2856 | |
| 2857 | if (ctx.config.is64()) { |
| 2858 | object::coff_tls_directory64 *tlsDir = |
| 2859 | reinterpret_cast<object::coff_tls_directory64 *>(&secBuf[tlsOffset]); |
| 2860 | tlsDir->setAlignment(tlsAlignment); |
| 2861 | } else { |
| 2862 | object::coff_tls_directory32 *tlsDir = |
| 2863 | reinterpret_cast<object::coff_tls_directory32 *>(&secBuf[tlsOffset]); |
| 2864 | tlsDir->setAlignment(tlsAlignment); |
| 2865 | } |
| 2866 | } |
| 2867 | |
| 2868 | void Writer::prepareLoadConfig() { |
| 2869 | ctx.forEachActiveSymtab(f: [&](SymbolTable &symtab) { |
| 2870 | if (!symtab.loadConfigSym) |
| 2871 | return; |
| 2872 | |
| 2873 | OutputSection *sec = ctx.getOutputSection(c: symtab.loadConfigSym->getChunk()); |
| 2874 | uint8_t *secBuf = buffer->getBufferStart() + sec->getFileOff(); |
| 2875 | uint8_t *symBuf = secBuf + (symtab.loadConfigSym->getRVA() - sec->getRVA()); |
| 2876 | |
| 2877 | if (ctx.config.is64()) |
| 2878 | prepareLoadConfig(symtab, |
| 2879 | loadConfig: reinterpret_cast<coff_load_configuration64 *>(symBuf)); |
| 2880 | else |
| 2881 | prepareLoadConfig(symtab, |
| 2882 | loadConfig: reinterpret_cast<coff_load_configuration32 *>(symBuf)); |
| 2883 | }); |
| 2884 | } |
| 2885 | |
| 2886 | template <typename T> |
| 2887 | void Writer::prepareLoadConfig(SymbolTable &symtab, T *loadConfig) { |
| 2888 | size_t loadConfigSize = loadConfig->Size; |
| 2889 | |
| 2890 | #define RETURN_IF_NOT_CONTAINS(field) \ |
| 2891 | if (loadConfigSize < offsetof(T, field) + sizeof(T::field)) { \ |
| 2892 | Warn(ctx) << "'_load_config_used' structure too small to include " #field; \ |
| 2893 | return; \ |
| 2894 | } |
| 2895 | |
| 2896 | #define IF_CONTAINS(field) \ |
| 2897 | if (loadConfigSize >= offsetof(T, field) + sizeof(T::field)) |
| 2898 | |
| 2899 | #define CHECK_VA(field, sym) \ |
| 2900 | if (auto *s = dyn_cast<DefinedSynthetic>(symtab.findUnderscore(sym))) \ |
| 2901 | if (loadConfig->field != ctx.config.imageBase + s->getRVA()) \ |
| 2902 | Warn(ctx) << #field " not set correctly in '_load_config_used'"; |
| 2903 | |
| 2904 | #define CHECK_ABSOLUTE(field, sym) \ |
| 2905 | if (auto *s = dyn_cast<DefinedAbsolute>(symtab.findUnderscore(sym))) \ |
| 2906 | if (loadConfig->field != s->getVA()) \ |
| 2907 | Warn(ctx) << #field " not set correctly in '_load_config_used'"; |
| 2908 | |
| 2909 | if (ctx.config.dependentLoadFlags) { |
| 2910 | RETURN_IF_NOT_CONTAINS(DependentLoadFlags) |
| 2911 | loadConfig->DependentLoadFlags = ctx.config.dependentLoadFlags; |
| 2912 | } |
| 2913 | |
| 2914 | if (ctx.dynamicRelocs) { |
| 2915 | IF_CONTAINS(DynamicValueRelocTableSection) { |
| 2916 | loadConfig->DynamicValueRelocTableSection = relocSec->sectionIndex; |
| 2917 | loadConfig->DynamicValueRelocTableOffset = |
| 2918 | ctx.dynamicRelocs->getRVA() - relocSec->getRVA(); |
| 2919 | } |
| 2920 | else { |
| 2921 | Warn(ctx) << "'_load_config_used' structure too small to include dynamic " |
| 2922 | "relocations" ; |
| 2923 | } |
| 2924 | } |
| 2925 | |
| 2926 | IF_CONTAINS(CHPEMetadataPointer) { |
| 2927 | // On ARM64X, only the EC version of the load config contains |
| 2928 | // CHPEMetadataPointer. Copy its value to the native load config. |
| 2929 | if (ctx.config.machine == ARM64X && !symtab.isEC() && |
| 2930 | ctx.symtab.loadConfigSize >= |
| 2931 | offsetof(T, CHPEMetadataPointer) + sizeof(T::CHPEMetadataPointer)) { |
| 2932 | OutputSection *sec = |
| 2933 | ctx.getOutputSection(c: ctx.symtab.loadConfigSym->getChunk()); |
| 2934 | uint8_t *secBuf = buffer->getBufferStart() + sec->getFileOff(); |
| 2935 | auto hybridLoadConfig = |
| 2936 | reinterpret_cast<const coff_load_configuration64 *>( |
| 2937 | secBuf + (ctx.symtab.loadConfigSym->getRVA() - sec->getRVA())); |
| 2938 | loadConfig->CHPEMetadataPointer = hybridLoadConfig->CHPEMetadataPointer; |
| 2939 | } |
| 2940 | } |
| 2941 | |
| 2942 | if (ctx.config.guardCF == GuardCFLevel::Off) |
| 2943 | return; |
| 2944 | RETURN_IF_NOT_CONTAINS(GuardFlags) |
| 2945 | CHECK_VA(GuardCFFunctionTable, "__guard_fids_table" ) |
| 2946 | CHECK_ABSOLUTE(GuardCFFunctionCount, "__guard_fids_count" ) |
| 2947 | CHECK_ABSOLUTE(GuardFlags, "__guard_flags" ) |
| 2948 | IF_CONTAINS(GuardAddressTakenIatEntryCount) { |
| 2949 | CHECK_VA(GuardAddressTakenIatEntryTable, "__guard_iat_table" ) |
| 2950 | CHECK_ABSOLUTE(GuardAddressTakenIatEntryCount, "__guard_iat_count" ) |
| 2951 | } |
| 2952 | |
| 2953 | if (!(ctx.config.guardCF & GuardCFLevel::LongJmp)) |
| 2954 | return; |
| 2955 | RETURN_IF_NOT_CONTAINS(GuardLongJumpTargetCount) |
| 2956 | CHECK_VA(GuardLongJumpTargetTable, "__guard_longjmp_table" ) |
| 2957 | CHECK_ABSOLUTE(GuardLongJumpTargetCount, "__guard_longjmp_count" ) |
| 2958 | |
| 2959 | if (!(ctx.config.guardCF & GuardCFLevel::EHCont)) |
| 2960 | return; |
| 2961 | RETURN_IF_NOT_CONTAINS(GuardEHContinuationCount) |
| 2962 | CHECK_VA(GuardEHContinuationTable, "__guard_eh_cont_table" ) |
| 2963 | CHECK_ABSOLUTE(GuardEHContinuationCount, "__guard_eh_cont_count" ) |
| 2964 | |
| 2965 | #undef RETURN_IF_NOT_CONTAINS |
| 2966 | #undef IF_CONTAINS |
| 2967 | #undef CHECK_VA |
| 2968 | #undef CHECK_ABSOLUTE |
| 2969 | } |
| 2970 | |