1 //===- SyntheticSections.cpp ----------------------------------------------===// 2 // 3 // The LLVM Linker 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains linker-synthesized sections. Currently, 11 // synthetic sections are created either output sections or input sections, 12 // but we are rewriting code so that all synthetic sections are created as 13 // input sections. 14 // 15 //===----------------------------------------------------------------------===// 16 17 #include "SyntheticSections.h" 18 #include "Config.h" 19 #include "Error.h" 20 #include "InputFiles.h" 21 #include "LinkerScript.h" 22 #include "Memory.h" 23 #include "OutputSections.h" 24 #include "Strings.h" 25 #include "SymbolTable.h" 26 #include "Target.h" 27 #include "Threads.h" 28 #include "Writer.h" 29 #include "lld/Common/Version.h" 30 #include "llvm/BinaryFormat/Dwarf.h" 31 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h" 32 #include "llvm/Object/Decompressor.h" 33 #include "llvm/Object/ELFObjectFile.h" 34 #include "llvm/Support/Endian.h" 35 #include "llvm/Support/MD5.h" 36 #include "llvm/Support/RandomNumberGenerator.h" 37 #include "llvm/Support/SHA1.h" 38 #include "llvm/Support/xxhash.h" 39 #include <cstdlib> 40 #include <thread> 41 42 using namespace llvm; 43 using namespace llvm::dwarf; 44 using namespace llvm::ELF; 45 using namespace llvm::object; 46 using namespace llvm::support; 47 using namespace llvm::support::endian; 48 49 using namespace lld; 50 using namespace lld::elf; 51 52 constexpr size_t MergeNoTailSection::NumShards; 53 54 uint64_t SyntheticSection::getVA() const { 55 if (OutputSection *Sec = getParent()) 56 return Sec->Addr + OutSecOff; 57 return 0; 58 } 59 60 // Create a .bss section for each common section and replace the common symbol 61 // with a DefinedRegular symbol. 62 template <class ELFT> void elf::createCommonSections() { 63 for (Symbol *S : Symtab->getSymbols()) { 64 auto *Sym = dyn_cast<DefinedCommon>(S->body()); 65 66 if (!Sym) 67 continue; 68 69 // Create a synthetic section for the common data. 70 auto *Section = make<BssSection>("COMMON", Sym->Size, Sym->Alignment); 71 Section->File = Sym->getFile(); 72 Section->Live = !Config->GcSections; 73 InputSections.push_back(Section); 74 75 // Replace all DefinedCommon symbols with DefinedRegular symbols so that we 76 // don't have to care about DefinedCommon symbols beyond this point. 77 replaceBody<DefinedRegular>(S, Sym->getFile(), Sym->getName(), 78 static_cast<bool>(Sym->IsLocal), Sym->StOther, 79 Sym->Type, 0, Sym->getSize<ELFT>(), Section); 80 } 81 } 82 83 // Returns an LLD version string. 84 static ArrayRef<uint8_t> getVersion() { 85 // Check LLD_VERSION first for ease of testing. 86 // You can get consitent output by using the environment variable. 87 // This is only for testing. 88 StringRef S = getenv("LLD_VERSION"); 89 if (S.empty()) 90 S = Saver.save(Twine("Linker: ") + getLLDVersion()); 91 92 // +1 to include the terminating '\0'. 93 return {(const uint8_t *)S.data(), S.size() + 1}; 94 } 95 96 // Creates a .comment section containing LLD version info. 97 // With this feature, you can identify LLD-generated binaries easily 98 // by "readelf --string-dump .comment <file>". 99 // The returned object is a mergeable string section. 100 template <class ELFT> MergeInputSection *elf::createCommentSection() { 101 typename ELFT::Shdr Hdr = {}; 102 Hdr.sh_flags = SHF_MERGE | SHF_STRINGS; 103 Hdr.sh_type = SHT_PROGBITS; 104 Hdr.sh_entsize = 1; 105 Hdr.sh_addralign = 1; 106 107 auto *Ret = 108 make<MergeInputSection>((ObjFile<ELFT> *)nullptr, &Hdr, ".comment"); 109 Ret->Data = getVersion(); 110 return Ret; 111 } 112 113 // .MIPS.abiflags section. 114 template <class ELFT> 115 MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags Flags) 116 : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags"), 117 Flags(Flags) { 118 this->Entsize = sizeof(Elf_Mips_ABIFlags); 119 } 120 121 template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *Buf) { 122 memcpy(Buf, &Flags, sizeof(Flags)); 123 } 124 125 template <class ELFT> 126 MipsAbiFlagsSection<ELFT> *MipsAbiFlagsSection<ELFT>::create() { 127 Elf_Mips_ABIFlags Flags = {}; 128 bool Create = false; 129 130 for (InputSectionBase *Sec : InputSections) { 131 if (Sec->Type != SHT_MIPS_ABIFLAGS) 132 continue; 133 Sec->Live = false; 134 Create = true; 135 136 std::string Filename = toString(Sec->getFile<ELFT>()); 137 const size_t Size = Sec->Data.size(); 138 // Older version of BFD (such as the default FreeBSD linker) concatenate 139 // .MIPS.abiflags instead of merging. To allow for this case (or potential 140 // zero padding) we ignore everything after the first Elf_Mips_ABIFlags 141 if (Size < sizeof(Elf_Mips_ABIFlags)) { 142 error(Filename + ": invalid size of .MIPS.abiflags section: got " + 143 Twine(Size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags))); 144 return nullptr; 145 } 146 auto *S = reinterpret_cast<const Elf_Mips_ABIFlags *>(Sec->Data.data()); 147 if (S->version != 0) { 148 error(Filename + ": unexpected .MIPS.abiflags version " + 149 Twine(S->version)); 150 return nullptr; 151 } 152 153 // LLD checks ISA compatibility in calcMipsEFlags(). Here we just 154 // select the highest number of ISA/Rev/Ext. 155 Flags.isa_level = std::max(Flags.isa_level, S->isa_level); 156 Flags.isa_rev = std::max(Flags.isa_rev, S->isa_rev); 157 Flags.isa_ext = std::max(Flags.isa_ext, S->isa_ext); 158 Flags.gpr_size = std::max(Flags.gpr_size, S->gpr_size); 159 Flags.cpr1_size = std::max(Flags.cpr1_size, S->cpr1_size); 160 Flags.cpr2_size = std::max(Flags.cpr2_size, S->cpr2_size); 161 Flags.ases |= S->ases; 162 Flags.flags1 |= S->flags1; 163 Flags.flags2 |= S->flags2; 164 Flags.fp_abi = elf::getMipsFpAbiFlag(Flags.fp_abi, S->fp_abi, Filename); 165 }; 166 167 if (Create) 168 return make<MipsAbiFlagsSection<ELFT>>(Flags); 169 return nullptr; 170 } 171 172 // .MIPS.options section. 173 template <class ELFT> 174 MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo Reginfo) 175 : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"), 176 Reginfo(Reginfo) { 177 this->Entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo); 178 } 179 180 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *Buf) { 181 auto *Options = reinterpret_cast<Elf_Mips_Options *>(Buf); 182 Options->kind = ODK_REGINFO; 183 Options->size = getSize(); 184 185 if (!Config->Relocatable) 186 Reginfo.ri_gp_value = InX::MipsGot->getGp(); 187 memcpy(Buf + sizeof(Elf_Mips_Options), &Reginfo, sizeof(Reginfo)); 188 } 189 190 template <class ELFT> 191 MipsOptionsSection<ELFT> *MipsOptionsSection<ELFT>::create() { 192 // N64 ABI only. 193 if (!ELFT::Is64Bits) 194 return nullptr; 195 196 Elf_Mips_RegInfo Reginfo = {}; 197 bool Create = false; 198 199 for (InputSectionBase *Sec : InputSections) { 200 if (Sec->Type != SHT_MIPS_OPTIONS) 201 continue; 202 Sec->Live = false; 203 Create = true; 204 205 std::string Filename = toString(Sec->getFile<ELFT>()); 206 ArrayRef<uint8_t> D = Sec->Data; 207 208 while (!D.empty()) { 209 if (D.size() < sizeof(Elf_Mips_Options)) { 210 error(Filename + ": invalid size of .MIPS.options section"); 211 break; 212 } 213 214 auto *Opt = reinterpret_cast<const Elf_Mips_Options *>(D.data()); 215 if (Opt->kind == ODK_REGINFO) { 216 if (Config->Relocatable && Opt->getRegInfo().ri_gp_value) 217 error(Filename + ": unsupported non-zero ri_gp_value"); 218 Reginfo.ri_gprmask |= Opt->getRegInfo().ri_gprmask; 219 Sec->getFile<ELFT>()->MipsGp0 = Opt->getRegInfo().ri_gp_value; 220 break; 221 } 222 223 if (!Opt->size) 224 fatal(Filename + ": zero option descriptor size"); 225 D = D.slice(Opt->size); 226 } 227 }; 228 229 if (Create) 230 return make<MipsOptionsSection<ELFT>>(Reginfo); 231 return nullptr; 232 } 233 234 // MIPS .reginfo section. 235 template <class ELFT> 236 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo Reginfo) 237 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"), 238 Reginfo(Reginfo) { 239 this->Entsize = sizeof(Elf_Mips_RegInfo); 240 } 241 242 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *Buf) { 243 if (!Config->Relocatable) 244 Reginfo.ri_gp_value = InX::MipsGot->getGp(); 245 memcpy(Buf, &Reginfo, sizeof(Reginfo)); 246 } 247 248 template <class ELFT> 249 MipsReginfoSection<ELFT> *MipsReginfoSection<ELFT>::create() { 250 // Section should be alive for O32 and N32 ABIs only. 251 if (ELFT::Is64Bits) 252 return nullptr; 253 254 Elf_Mips_RegInfo Reginfo = {}; 255 bool Create = false; 256 257 for (InputSectionBase *Sec : InputSections) { 258 if (Sec->Type != SHT_MIPS_REGINFO) 259 continue; 260 Sec->Live = false; 261 Create = true; 262 263 if (Sec->Data.size() != sizeof(Elf_Mips_RegInfo)) { 264 error(toString(Sec->getFile<ELFT>()) + 265 ": invalid size of .reginfo section"); 266 return nullptr; 267 } 268 auto *R = reinterpret_cast<const Elf_Mips_RegInfo *>(Sec->Data.data()); 269 if (Config->Relocatable && R->ri_gp_value) 270 error(toString(Sec->getFile<ELFT>()) + 271 ": unsupported non-zero ri_gp_value"); 272 273 Reginfo.ri_gprmask |= R->ri_gprmask; 274 Sec->getFile<ELFT>()->MipsGp0 = R->ri_gp_value; 275 }; 276 277 if (Create) 278 return make<MipsReginfoSection<ELFT>>(Reginfo); 279 return nullptr; 280 } 281 282 InputSection *elf::createInterpSection() { 283 // StringSaver guarantees that the returned string ends with '\0'. 284 StringRef S = Saver.save(Config->DynamicLinker); 285 ArrayRef<uint8_t> Contents = {(const uint8_t *)S.data(), S.size() + 1}; 286 287 auto *Sec = 288 make<InputSection>(SHF_ALLOC, SHT_PROGBITS, 1, Contents, ".interp"); 289 Sec->Live = true; 290 return Sec; 291 } 292 293 SymbolBody *elf::addSyntheticLocal(StringRef Name, uint8_t Type, uint64_t Value, 294 uint64_t Size, InputSectionBase *Section) { 295 auto *S = make<DefinedRegular>(Name, /*IsLocal*/ true, STV_DEFAULT, Type, 296 Value, Size, Section); 297 if (InX::SymTab) 298 InX::SymTab->addSymbol(S); 299 return S; 300 } 301 302 static size_t getHashSize() { 303 switch (Config->BuildId) { 304 case BuildIdKind::Fast: 305 return 8; 306 case BuildIdKind::Md5: 307 case BuildIdKind::Uuid: 308 return 16; 309 case BuildIdKind::Sha1: 310 return 20; 311 case BuildIdKind::Hexstring: 312 return Config->BuildIdVector.size(); 313 default: 314 llvm_unreachable("unknown BuildIdKind"); 315 } 316 } 317 318 BuildIdSection::BuildIdSection() 319 : SyntheticSection(SHF_ALLOC, SHT_NOTE, 1, ".note.gnu.build-id"), 320 HashSize(getHashSize()) {} 321 322 void BuildIdSection::writeTo(uint8_t *Buf) { 323 endianness E = Config->Endianness; 324 write32(Buf, 4, E); // Name size 325 write32(Buf + 4, HashSize, E); // Content size 326 write32(Buf + 8, NT_GNU_BUILD_ID, E); // Type 327 memcpy(Buf + 12, "GNU", 4); // Name string 328 HashBuf = Buf + 16; 329 } 330 331 // Split one uint8 array into small pieces of uint8 arrays. 332 static std::vector<ArrayRef<uint8_t>> split(ArrayRef<uint8_t> Arr, 333 size_t ChunkSize) { 334 std::vector<ArrayRef<uint8_t>> Ret; 335 while (Arr.size() > ChunkSize) { 336 Ret.push_back(Arr.take_front(ChunkSize)); 337 Arr = Arr.drop_front(ChunkSize); 338 } 339 if (!Arr.empty()) 340 Ret.push_back(Arr); 341 return Ret; 342 } 343 344 // Computes a hash value of Data using a given hash function. 345 // In order to utilize multiple cores, we first split data into 1MB 346 // chunks, compute a hash for each chunk, and then compute a hash value 347 // of the hash values. 348 void BuildIdSection::computeHash( 349 llvm::ArrayRef<uint8_t> Data, 350 std::function<void(uint8_t *Dest, ArrayRef<uint8_t> Arr)> HashFn) { 351 std::vector<ArrayRef<uint8_t>> Chunks = split(Data, 1024 * 1024); 352 std::vector<uint8_t> Hashes(Chunks.size() * HashSize); 353 354 // Compute hash values. 355 parallelForEachN(0, Chunks.size(), [&](size_t I) { 356 HashFn(Hashes.data() + I * HashSize, Chunks[I]); 357 }); 358 359 // Write to the final output buffer. 360 HashFn(HashBuf, Hashes); 361 } 362 363 BssSection::BssSection(StringRef Name, uint64_t Size, uint32_t Alignment) 364 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, Alignment, Name) { 365 if (OutputSection *Sec = getParent()) 366 Sec->updateAlignment(Alignment); 367 this->Size = Size; 368 } 369 370 void BuildIdSection::writeBuildId(ArrayRef<uint8_t> Buf) { 371 switch (Config->BuildId) { 372 case BuildIdKind::Fast: 373 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) { 374 write64le(Dest, xxHash64(toStringRef(Arr))); 375 }); 376 break; 377 case BuildIdKind::Md5: 378 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) { 379 memcpy(Dest, MD5::hash(Arr).data(), 16); 380 }); 381 break; 382 case BuildIdKind::Sha1: 383 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) { 384 memcpy(Dest, SHA1::hash(Arr).data(), 20); 385 }); 386 break; 387 case BuildIdKind::Uuid: 388 if (getRandomBytes(HashBuf, HashSize)) 389 error("entropy source failure"); 390 break; 391 case BuildIdKind::Hexstring: 392 memcpy(HashBuf, Config->BuildIdVector.data(), Config->BuildIdVector.size()); 393 break; 394 default: 395 llvm_unreachable("unknown BuildIdKind"); 396 } 397 } 398 399 template <class ELFT> 400 EhFrameSection<ELFT>::EhFrameSection() 401 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame") {} 402 403 // Search for an existing CIE record or create a new one. 404 // CIE records from input object files are uniquified by their contents 405 // and where their relocations point to. 406 template <class ELFT> 407 template <class RelTy> 408 CieRecord *EhFrameSection<ELFT>::addCie(EhSectionPiece &Cie, 409 ArrayRef<RelTy> Rels) { 410 auto *Sec = cast<EhInputSection>(Cie.Sec); 411 const endianness E = ELFT::TargetEndianness; 412 if (read32<E>(Cie.data().data() + 4) != 0) 413 fatal(toString(Sec) + ": CIE expected at beginning of .eh_frame"); 414 415 SymbolBody *Personality = nullptr; 416 unsigned FirstRelI = Cie.FirstRelocation; 417 if (FirstRelI != (unsigned)-1) 418 Personality = 419 &Sec->template getFile<ELFT>()->getRelocTargetSym(Rels[FirstRelI]); 420 421 // Search for an existing CIE by CIE contents/relocation target pair. 422 CieRecord *&Rec = CieMap[{Cie.data(), Personality}]; 423 424 // If not found, create a new one. 425 if (!Rec) { 426 Rec = make<CieRecord>(); 427 Rec->Cie = &Cie; 428 CieRecords.push_back(Rec); 429 } 430 return Rec; 431 } 432 433 // There is one FDE per function. Returns true if a given FDE 434 // points to a live function. 435 template <class ELFT> 436 template <class RelTy> 437 bool EhFrameSection<ELFT>::isFdeLive(EhSectionPiece &Fde, 438 ArrayRef<RelTy> Rels) { 439 auto *Sec = cast<EhInputSection>(Fde.Sec); 440 unsigned FirstRelI = Fde.FirstRelocation; 441 442 // An FDE should point to some function because FDEs are to describe 443 // functions. That's however not always the case due to an issue of 444 // ld.gold with -r. ld.gold may discard only functions and leave their 445 // corresponding FDEs, which results in creating bad .eh_frame sections. 446 // To deal with that, we ignore such FDEs. 447 if (FirstRelI == (unsigned)-1) 448 return false; 449 450 const RelTy &Rel = Rels[FirstRelI]; 451 SymbolBody &B = Sec->template getFile<ELFT>()->getRelocTargetSym(Rel); 452 if (auto *D = dyn_cast<DefinedRegular>(&B)) 453 if (D->Section) 454 return cast<InputSectionBase>(D->Section)->Repl->Live; 455 return false; 456 } 457 458 // .eh_frame is a sequence of CIE or FDE records. In general, there 459 // is one CIE record per input object file which is followed by 460 // a list of FDEs. This function searches an existing CIE or create a new 461 // one and associates FDEs to the CIE. 462 template <class ELFT> 463 template <class RelTy> 464 void EhFrameSection<ELFT>::addSectionAux(EhInputSection *Sec, 465 ArrayRef<RelTy> Rels) { 466 const endianness E = ELFT::TargetEndianness; 467 468 DenseMap<size_t, CieRecord *> OffsetToCie; 469 for (EhSectionPiece &Piece : Sec->Pieces) { 470 // The empty record is the end marker. 471 if (Piece.Size == 4) 472 return; 473 474 size_t Offset = Piece.InputOff; 475 uint32_t ID = read32<E>(Piece.data().data() + 4); 476 if (ID == 0) { 477 OffsetToCie[Offset] = addCie(Piece, Rels); 478 continue; 479 } 480 481 uint32_t CieOffset = Offset + 4 - ID; 482 CieRecord *Rec = OffsetToCie[CieOffset]; 483 if (!Rec) 484 fatal(toString(Sec) + ": invalid CIE reference"); 485 486 if (!isFdeLive(Piece, Rels)) 487 continue; 488 Rec->Fdes.push_back(&Piece); 489 NumFdes++; 490 } 491 } 492 493 template <class ELFT> 494 void EhFrameSection<ELFT>::addSection(InputSectionBase *C) { 495 auto *Sec = cast<EhInputSection>(C); 496 Sec->Parent = this; 497 updateAlignment(Sec->Alignment); 498 Sections.push_back(Sec); 499 for (auto *DS : Sec->DependentSections) 500 DependentSections.push_back(DS); 501 502 // .eh_frame is a sequence of CIE or FDE records. This function 503 // splits it into pieces so that we can call 504 // SplitInputSection::getSectionPiece on the section. 505 Sec->split<ELFT>(); 506 if (Sec->Pieces.empty()) 507 return; 508 509 if (Sec->NumRelocations == 0) 510 addSectionAux(Sec, makeArrayRef<Elf_Rela>(nullptr, nullptr)); 511 else if (Sec->AreRelocsRela) 512 addSectionAux(Sec, Sec->template relas<ELFT>()); 513 else 514 addSectionAux(Sec, Sec->template rels<ELFT>()); 515 } 516 517 template <class ELFT> 518 static void writeCieFde(uint8_t *Buf, ArrayRef<uint8_t> D) { 519 memcpy(Buf, D.data(), D.size()); 520 521 size_t Aligned = alignTo(D.size(), sizeof(typename ELFT::uint)); 522 523 // Zero-clear trailing padding if it exists. 524 memset(Buf + D.size(), 0, Aligned - D.size()); 525 526 // Fix the size field. -4 since size does not include the size field itself. 527 const endianness E = ELFT::TargetEndianness; 528 write32<E>(Buf, Aligned - 4); 529 } 530 531 template <class ELFT> void EhFrameSection<ELFT>::finalizeContents() { 532 if (this->Size) 533 return; // Already finalized. 534 535 size_t Off = 0; 536 for (CieRecord *Rec : CieRecords) { 537 Rec->Cie->OutputOff = Off; 538 Off += alignTo(Rec->Cie->Size, Config->Wordsize); 539 540 for (EhSectionPiece *Fde : Rec->Fdes) { 541 Fde->OutputOff = Off; 542 Off += alignTo(Fde->Size, Config->Wordsize); 543 } 544 } 545 546 // The LSB standard does not allow a .eh_frame section with zero 547 // Call Frame Information records. Therefore add a CIE record length 548 // 0 as a terminator if this .eh_frame section is empty. 549 if (Off == 0) 550 Off = 4; 551 552 this->Size = Off; 553 } 554 555 template <class ELFT> static uint64_t readFdeAddr(uint8_t *Buf, int Size) { 556 const endianness E = ELFT::TargetEndianness; 557 switch (Size) { 558 case DW_EH_PE_udata2: 559 return read16<E>(Buf); 560 case DW_EH_PE_udata4: 561 return read32<E>(Buf); 562 case DW_EH_PE_udata8: 563 return read64<E>(Buf); 564 case DW_EH_PE_absptr: 565 if (ELFT::Is64Bits) 566 return read64<E>(Buf); 567 return read32<E>(Buf); 568 } 569 fatal("unknown FDE size encoding"); 570 } 571 572 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to. 573 // We need it to create .eh_frame_hdr section. 574 template <class ELFT> 575 uint64_t EhFrameSection<ELFT>::getFdePc(uint8_t *Buf, size_t FdeOff, 576 uint8_t Enc) { 577 // The starting address to which this FDE applies is 578 // stored at FDE + 8 byte. 579 size_t Off = FdeOff + 8; 580 uint64_t Addr = readFdeAddr<ELFT>(Buf + Off, Enc & 0x7); 581 if ((Enc & 0x70) == DW_EH_PE_absptr) 582 return Addr; 583 if ((Enc & 0x70) == DW_EH_PE_pcrel) 584 return Addr + getParent()->Addr + Off; 585 fatal("unknown FDE size relative encoding"); 586 } 587 588 template <class ELFT> void EhFrameSection<ELFT>::writeTo(uint8_t *Buf) { 589 const endianness E = ELFT::TargetEndianness; 590 for (CieRecord *Rec : CieRecords) { 591 size_t CieOffset = Rec->Cie->OutputOff; 592 writeCieFde<ELFT>(Buf + CieOffset, Rec->Cie->data()); 593 594 for (EhSectionPiece *Fde : Rec->Fdes) { 595 size_t Off = Fde->OutputOff; 596 writeCieFde<ELFT>(Buf + Off, Fde->data()); 597 598 // FDE's second word should have the offset to an associated CIE. 599 // Write it. 600 write32<E>(Buf + Off + 4, Off + 4 - CieOffset); 601 } 602 } 603 604 for (EhInputSection *S : Sections) 605 S->relocateAlloc(Buf, nullptr); 606 607 // Construct .eh_frame_hdr. .eh_frame_hdr is a binary search table 608 // to get a FDE from an address to which FDE is applied. So here 609 // we obtain two addresses and pass them to EhFrameHdr object. 610 if (In<ELFT>::EhFrameHdr) { 611 for (CieRecord *Rec : CieRecords) { 612 uint8_t Enc = getFdeEncoding<ELFT>(Rec->Cie); 613 for (EhSectionPiece *Fde : Rec->Fdes) { 614 uint64_t Pc = getFdePc(Buf, Fde->OutputOff, Enc); 615 uint64_t FdeVA = getParent()->Addr + Fde->OutputOff; 616 In<ELFT>::EhFrameHdr->addFde(Pc, FdeVA); 617 } 618 } 619 } 620 } 621 622 GotSection::GotSection() 623 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 624 Target->GotEntrySize, ".got") {} 625 626 void GotSection::addEntry(SymbolBody &Sym) { 627 Sym.GotIndex = NumEntries; 628 ++NumEntries; 629 } 630 631 bool GotSection::addDynTlsEntry(SymbolBody &Sym) { 632 if (Sym.GlobalDynIndex != -1U) 633 return false; 634 Sym.GlobalDynIndex = NumEntries; 635 // Global Dynamic TLS entries take two GOT slots. 636 NumEntries += 2; 637 return true; 638 } 639 640 // Reserves TLS entries for a TLS module ID and a TLS block offset. 641 // In total it takes two GOT slots. 642 bool GotSection::addTlsIndex() { 643 if (TlsIndexOff != uint32_t(-1)) 644 return false; 645 TlsIndexOff = NumEntries * Config->Wordsize; 646 NumEntries += 2; 647 return true; 648 } 649 650 uint64_t GotSection::getGlobalDynAddr(const SymbolBody &B) const { 651 return this->getVA() + B.GlobalDynIndex * Config->Wordsize; 652 } 653 654 uint64_t GotSection::getGlobalDynOffset(const SymbolBody &B) const { 655 return B.GlobalDynIndex * Config->Wordsize; 656 } 657 658 void GotSection::finalizeContents() { Size = NumEntries * Config->Wordsize; } 659 660 bool GotSection::empty() const { 661 // We need to emit a GOT even if it's empty if there's a relocation that is 662 // relative to GOT(such as GOTOFFREL) or there's a symbol that points to a GOT 663 // (i.e. _GLOBAL_OFFSET_TABLE_). 664 return NumEntries == 0 && !HasGotOffRel && !ElfSym::GlobalOffsetTable; 665 } 666 667 void GotSection::writeTo(uint8_t *Buf) { 668 // Buf points to the start of this section's buffer, 669 // whereas InputSectionBase::relocateAlloc() expects its argument 670 // to point to the start of the output section. 671 relocateAlloc(Buf - OutSecOff, Buf - OutSecOff + Size); 672 } 673 674 MipsGotSection::MipsGotSection() 675 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16, 676 ".got") {} 677 678 void MipsGotSection::addEntry(SymbolBody &Sym, int64_t Addend, RelExpr Expr) { 679 // For "true" local symbols which can be referenced from the same module 680 // only compiler creates two instructions for address loading: 681 // 682 // lw $8, 0($gp) # R_MIPS_GOT16 683 // addi $8, $8, 0 # R_MIPS_LO16 684 // 685 // The first instruction loads high 16 bits of the symbol address while 686 // the second adds an offset. That allows to reduce number of required 687 // GOT entries because only one global offset table entry is necessary 688 // for every 64 KBytes of local data. So for local symbols we need to 689 // allocate number of GOT entries to hold all required "page" addresses. 690 // 691 // All global symbols (hidden and regular) considered by compiler uniformly. 692 // It always generates a single `lw` instruction and R_MIPS_GOT16 relocation 693 // to load address of the symbol. So for each such symbol we need to 694 // allocate dedicated GOT entry to store its address. 695 // 696 // If a symbol is preemptible we need help of dynamic linker to get its 697 // final address. The corresponding GOT entries are allocated in the 698 // "global" part of GOT. Entries for non preemptible global symbol allocated 699 // in the "local" part of GOT. 700 // 701 // See "Global Offset Table" in Chapter 5: 702 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 703 if (Expr == R_MIPS_GOT_LOCAL_PAGE) { 704 // At this point we do not know final symbol value so to reduce number 705 // of allocated GOT entries do the following trick. Save all output 706 // sections referenced by GOT relocations. Then later in the `finalize` 707 // method calculate number of "pages" required to cover all saved output 708 // section and allocate appropriate number of GOT entries. 709 PageIndexMap.insert({Sym.getOutputSection(), 0}); 710 return; 711 } 712 if (Sym.isTls()) { 713 // GOT entries created for MIPS TLS relocations behave like 714 // almost GOT entries from other ABIs. They go to the end 715 // of the global offset table. 716 Sym.GotIndex = TlsEntries.size(); 717 TlsEntries.push_back(&Sym); 718 return; 719 } 720 auto AddEntry = [&](SymbolBody &S, uint64_t A, GotEntries &Items) { 721 if (S.isInGot() && !A) 722 return; 723 size_t NewIndex = Items.size(); 724 if (!EntryIndexMap.insert({{&S, A}, NewIndex}).second) 725 return; 726 Items.emplace_back(&S, A); 727 if (!A) 728 S.GotIndex = NewIndex; 729 }; 730 if (Sym.isPreemptible()) { 731 // Ignore addends for preemptible symbols. They got single GOT entry anyway. 732 AddEntry(Sym, 0, GlobalEntries); 733 Sym.IsInGlobalMipsGot = true; 734 } else if (Expr == R_MIPS_GOT_OFF32) { 735 AddEntry(Sym, Addend, LocalEntries32); 736 Sym.Is32BitMipsGot = true; 737 } else { 738 // Hold local GOT entries accessed via a 16-bit index separately. 739 // That allows to write them in the beginning of the GOT and keep 740 // their indexes as less as possible to escape relocation's overflow. 741 AddEntry(Sym, Addend, LocalEntries); 742 } 743 } 744 745 bool MipsGotSection::addDynTlsEntry(SymbolBody &Sym) { 746 if (Sym.GlobalDynIndex != -1U) 747 return false; 748 Sym.GlobalDynIndex = TlsEntries.size(); 749 // Global Dynamic TLS entries take two GOT slots. 750 TlsEntries.push_back(nullptr); 751 TlsEntries.push_back(&Sym); 752 return true; 753 } 754 755 // Reserves TLS entries for a TLS module ID and a TLS block offset. 756 // In total it takes two GOT slots. 757 bool MipsGotSection::addTlsIndex() { 758 if (TlsIndexOff != uint32_t(-1)) 759 return false; 760 TlsIndexOff = TlsEntries.size() * Config->Wordsize; 761 TlsEntries.push_back(nullptr); 762 TlsEntries.push_back(nullptr); 763 return true; 764 } 765 766 static uint64_t getMipsPageAddr(uint64_t Addr) { 767 return (Addr + 0x8000) & ~0xffff; 768 } 769 770 static uint64_t getMipsPageCount(uint64_t Size) { 771 return (Size + 0xfffe) / 0xffff + 1; 772 } 773 774 uint64_t MipsGotSection::getPageEntryOffset(const SymbolBody &B, 775 int64_t Addend) const { 776 const OutputSection *OutSec = B.getOutputSection(); 777 uint64_t SecAddr = getMipsPageAddr(OutSec->Addr); 778 uint64_t SymAddr = getMipsPageAddr(B.getVA(Addend)); 779 uint64_t Index = PageIndexMap.lookup(OutSec) + (SymAddr - SecAddr) / 0xffff; 780 assert(Index < PageEntriesNum); 781 return (HeaderEntriesNum + Index) * Config->Wordsize; 782 } 783 784 uint64_t MipsGotSection::getBodyEntryOffset(const SymbolBody &B, 785 int64_t Addend) const { 786 // Calculate offset of the GOT entries block: TLS, global, local. 787 uint64_t Index = HeaderEntriesNum + PageEntriesNum; 788 if (B.isTls()) 789 Index += LocalEntries.size() + LocalEntries32.size() + GlobalEntries.size(); 790 else if (B.IsInGlobalMipsGot) 791 Index += LocalEntries.size() + LocalEntries32.size(); 792 else if (B.Is32BitMipsGot) 793 Index += LocalEntries.size(); 794 // Calculate offset of the GOT entry in the block. 795 if (B.isInGot()) 796 Index += B.GotIndex; 797 else { 798 auto It = EntryIndexMap.find({&B, Addend}); 799 assert(It != EntryIndexMap.end()); 800 Index += It->second; 801 } 802 return Index * Config->Wordsize; 803 } 804 805 uint64_t MipsGotSection::getTlsOffset() const { 806 return (getLocalEntriesNum() + GlobalEntries.size()) * Config->Wordsize; 807 } 808 809 uint64_t MipsGotSection::getGlobalDynOffset(const SymbolBody &B) const { 810 return B.GlobalDynIndex * Config->Wordsize; 811 } 812 813 const SymbolBody *MipsGotSection::getFirstGlobalEntry() const { 814 return GlobalEntries.empty() ? nullptr : GlobalEntries.front().first; 815 } 816 817 unsigned MipsGotSection::getLocalEntriesNum() const { 818 return HeaderEntriesNum + PageEntriesNum + LocalEntries.size() + 819 LocalEntries32.size(); 820 } 821 822 void MipsGotSection::finalizeContents() { updateAllocSize(); } 823 824 void MipsGotSection::updateAllocSize() { 825 PageEntriesNum = 0; 826 for (std::pair<const OutputSection *, size_t> &P : PageIndexMap) { 827 // For each output section referenced by GOT page relocations calculate 828 // and save into PageIndexMap an upper bound of MIPS GOT entries required 829 // to store page addresses of local symbols. We assume the worst case - 830 // each 64kb page of the output section has at least one GOT relocation 831 // against it. And take in account the case when the section intersects 832 // page boundaries. 833 P.second = PageEntriesNum; 834 PageEntriesNum += getMipsPageCount(P.first->Size); 835 } 836 Size = (getLocalEntriesNum() + GlobalEntries.size() + TlsEntries.size()) * 837 Config->Wordsize; 838 } 839 840 bool MipsGotSection::empty() const { 841 // We add the .got section to the result for dynamic MIPS target because 842 // its address and properties are mentioned in the .dynamic section. 843 return Config->Relocatable; 844 } 845 846 uint64_t MipsGotSection::getGp() const { return ElfSym::MipsGp->getVA(0); } 847 848 static uint64_t readUint(uint8_t *Buf) { 849 if (Config->Is64) 850 return read64(Buf, Config->Endianness); 851 return read32(Buf, Config->Endianness); 852 } 853 854 static void writeUint(uint8_t *Buf, uint64_t Val) { 855 if (Config->Is64) 856 write64(Buf, Val, Config->Endianness); 857 else 858 write32(Buf, Val, Config->Endianness); 859 } 860 861 void MipsGotSection::writeTo(uint8_t *Buf) { 862 // Set the MSB of the second GOT slot. This is not required by any 863 // MIPS ABI documentation, though. 864 // 865 // There is a comment in glibc saying that "The MSB of got[1] of a 866 // gnu object is set to identify gnu objects," and in GNU gold it 867 // says "the second entry will be used by some runtime loaders". 868 // But how this field is being used is unclear. 869 // 870 // We are not really willing to mimic other linkers behaviors 871 // without understanding why they do that, but because all files 872 // generated by GNU tools have this special GOT value, and because 873 // we've been doing this for years, it is probably a safe bet to 874 // keep doing this for now. We really need to revisit this to see 875 // if we had to do this. 876 writeUint(Buf + Config->Wordsize, (uint64_t)1 << (Config->Wordsize * 8 - 1)); 877 Buf += HeaderEntriesNum * Config->Wordsize; 878 // Write 'page address' entries to the local part of the GOT. 879 for (std::pair<const OutputSection *, size_t> &L : PageIndexMap) { 880 size_t PageCount = getMipsPageCount(L.first->Size); 881 uint64_t FirstPageAddr = getMipsPageAddr(L.first->Addr); 882 for (size_t PI = 0; PI < PageCount; ++PI) { 883 uint8_t *Entry = Buf + (L.second + PI) * Config->Wordsize; 884 writeUint(Entry, FirstPageAddr + PI * 0x10000); 885 } 886 } 887 Buf += PageEntriesNum * Config->Wordsize; 888 auto AddEntry = [&](const GotEntry &SA) { 889 uint8_t *Entry = Buf; 890 Buf += Config->Wordsize; 891 const SymbolBody *Body = SA.first; 892 uint64_t VA = Body->getVA(SA.second); 893 writeUint(Entry, VA); 894 }; 895 std::for_each(std::begin(LocalEntries), std::end(LocalEntries), AddEntry); 896 std::for_each(std::begin(LocalEntries32), std::end(LocalEntries32), AddEntry); 897 std::for_each(std::begin(GlobalEntries), std::end(GlobalEntries), AddEntry); 898 // Initialize TLS-related GOT entries. If the entry has a corresponding 899 // dynamic relocations, leave it initialized by zero. Write down adjusted 900 // TLS symbol's values otherwise. To calculate the adjustments use offsets 901 // for thread-local storage. 902 // https://www.linux-mips.org/wiki/NPTL 903 if (TlsIndexOff != -1U && !Config->Pic) 904 writeUint(Buf + TlsIndexOff, 1); 905 for (const SymbolBody *B : TlsEntries) { 906 if (!B || B->isPreemptible()) 907 continue; 908 uint64_t VA = B->getVA(); 909 if (B->GotIndex != -1U) { 910 uint8_t *Entry = Buf + B->GotIndex * Config->Wordsize; 911 writeUint(Entry, VA - 0x7000); 912 } 913 if (B->GlobalDynIndex != -1U) { 914 uint8_t *Entry = Buf + B->GlobalDynIndex * Config->Wordsize; 915 writeUint(Entry, 1); 916 Entry += Config->Wordsize; 917 writeUint(Entry, VA - 0x8000); 918 } 919 } 920 } 921 922 GotPltSection::GotPltSection() 923 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 924 Target->GotPltEntrySize, ".got.plt") {} 925 926 void GotPltSection::addEntry(SymbolBody &Sym) { 927 Sym.GotPltIndex = Target->GotPltHeaderEntriesNum + Entries.size(); 928 Entries.push_back(&Sym); 929 } 930 931 size_t GotPltSection::getSize() const { 932 return (Target->GotPltHeaderEntriesNum + Entries.size()) * 933 Target->GotPltEntrySize; 934 } 935 936 void GotPltSection::writeTo(uint8_t *Buf) { 937 Target->writeGotPltHeader(Buf); 938 Buf += Target->GotPltHeaderEntriesNum * Target->GotPltEntrySize; 939 for (const SymbolBody *B : Entries) { 940 Target->writeGotPlt(Buf, *B); 941 Buf += Config->Wordsize; 942 } 943 } 944 945 // On ARM the IgotPltSection is part of the GotSection, on other Targets it is 946 // part of the .got.plt 947 IgotPltSection::IgotPltSection() 948 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 949 Target->GotPltEntrySize, 950 Config->EMachine == EM_ARM ? ".got" : ".got.plt") {} 951 952 void IgotPltSection::addEntry(SymbolBody &Sym) { 953 Sym.IsInIgot = true; 954 Sym.GotPltIndex = Entries.size(); 955 Entries.push_back(&Sym); 956 } 957 958 size_t IgotPltSection::getSize() const { 959 return Entries.size() * Target->GotPltEntrySize; 960 } 961 962 void IgotPltSection::writeTo(uint8_t *Buf) { 963 for (const SymbolBody *B : Entries) { 964 Target->writeIgotPlt(Buf, *B); 965 Buf += Config->Wordsize; 966 } 967 } 968 969 StringTableSection::StringTableSection(StringRef Name, bool Dynamic) 970 : SyntheticSection(Dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, Name), 971 Dynamic(Dynamic) { 972 // ELF string tables start with a NUL byte. 973 addString(""); 974 } 975 976 // Adds a string to the string table. If HashIt is true we hash and check for 977 // duplicates. It is optional because the name of global symbols are already 978 // uniqued and hashing them again has a big cost for a small value: uniquing 979 // them with some other string that happens to be the same. 980 unsigned StringTableSection::addString(StringRef S, bool HashIt) { 981 if (HashIt) { 982 auto R = StringMap.insert(std::make_pair(S, this->Size)); 983 if (!R.second) 984 return R.first->second; 985 } 986 unsigned Ret = this->Size; 987 this->Size = this->Size + S.size() + 1; 988 Strings.push_back(S); 989 return Ret; 990 } 991 992 void StringTableSection::writeTo(uint8_t *Buf) { 993 for (StringRef S : Strings) { 994 memcpy(Buf, S.data(), S.size()); 995 Buf[S.size()] = '\0'; 996 Buf += S.size() + 1; 997 } 998 } 999 1000 // Returns the number of version definition entries. Because the first entry 1001 // is for the version definition itself, it is the number of versioned symbols 1002 // plus one. Note that we don't support multiple versions yet. 1003 static unsigned getVerDefNum() { return Config->VersionDefinitions.size() + 1; } 1004 1005 template <class ELFT> 1006 DynamicSection<ELFT>::DynamicSection() 1007 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, Config->Wordsize, 1008 ".dynamic") { 1009 this->Entsize = ELFT::Is64Bits ? 16 : 8; 1010 1011 // .dynamic section is not writable on MIPS and on Fuchsia OS 1012 // which passes -z rodynamic. 1013 // See "Special Section" in Chapter 4 in the following document: 1014 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 1015 if (Config->EMachine == EM_MIPS || Config->ZRodynamic) 1016 this->Flags = SHF_ALLOC; 1017 1018 addEntries(); 1019 } 1020 1021 // There are some dynamic entries that don't depend on other sections. 1022 // Such entries can be set early. 1023 template <class ELFT> void DynamicSection<ELFT>::addEntries() { 1024 // Add strings to .dynstr early so that .dynstr's size will be 1025 // fixed early. 1026 for (StringRef S : Config->FilterList) 1027 add({DT_FILTER, InX::DynStrTab->addString(S)}); 1028 for (StringRef S : Config->AuxiliaryList) 1029 add({DT_AUXILIARY, InX::DynStrTab->addString(S)}); 1030 if (!Config->Rpath.empty()) 1031 add({Config->EnableNewDtags ? DT_RUNPATH : DT_RPATH, 1032 InX::DynStrTab->addString(Config->Rpath)}); 1033 for (InputFile *File : SharedFiles) { 1034 SharedFile<ELFT> *F = cast<SharedFile<ELFT>>(File); 1035 if (F->isNeeded()) 1036 add({DT_NEEDED, InX::DynStrTab->addString(F->SoName)}); 1037 } 1038 if (!Config->SoName.empty()) 1039 add({DT_SONAME, InX::DynStrTab->addString(Config->SoName)}); 1040 1041 // Set DT_FLAGS and DT_FLAGS_1. 1042 uint32_t DtFlags = 0; 1043 uint32_t DtFlags1 = 0; 1044 if (Config->Bsymbolic) 1045 DtFlags |= DF_SYMBOLIC; 1046 if (Config->ZNodelete) 1047 DtFlags1 |= DF_1_NODELETE; 1048 if (Config->ZNodlopen) 1049 DtFlags1 |= DF_1_NOOPEN; 1050 if (Config->ZNow) { 1051 DtFlags |= DF_BIND_NOW; 1052 DtFlags1 |= DF_1_NOW; 1053 } 1054 if (Config->ZOrigin) { 1055 DtFlags |= DF_ORIGIN; 1056 DtFlags1 |= DF_1_ORIGIN; 1057 } 1058 1059 if (DtFlags) 1060 add({DT_FLAGS, DtFlags}); 1061 if (DtFlags1) 1062 add({DT_FLAGS_1, DtFlags1}); 1063 1064 // DT_DEBUG is a pointer to debug informaion used by debuggers at runtime. We 1065 // need it for each process, so we don't write it for DSOs. The loader writes 1066 // the pointer into this entry. 1067 // 1068 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some 1069 // systems (currently only Fuchsia OS) provide other means to give the 1070 // debugger this information. Such systems may choose make .dynamic read-only. 1071 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG. 1072 if (!Config->Shared && !Config->Relocatable && !Config->ZRodynamic) 1073 add({DT_DEBUG, (uint64_t)0}); 1074 } 1075 1076 // Add remaining entries to complete .dynamic contents. 1077 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() { 1078 if (this->Size) 1079 return; // Already finalized. 1080 1081 this->Link = InX::DynStrTab->getParent()->SectionIndex; 1082 if (In<ELFT>::RelaDyn->getParent() && !In<ELFT>::RelaDyn->empty()) { 1083 bool IsRela = Config->IsRela; 1084 add({IsRela ? DT_RELA : DT_REL, In<ELFT>::RelaDyn}); 1085 add({IsRela ? DT_RELASZ : DT_RELSZ, In<ELFT>::RelaDyn->getParent(), 1086 Entry::SecSize}); 1087 add({IsRela ? DT_RELAENT : DT_RELENT, 1088 uint64_t(IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel))}); 1089 1090 // MIPS dynamic loader does not support RELCOUNT tag. 1091 // The problem is in the tight relation between dynamic 1092 // relocations and GOT. So do not emit this tag on MIPS. 1093 if (Config->EMachine != EM_MIPS) { 1094 size_t NumRelativeRels = In<ELFT>::RelaDyn->getRelativeRelocCount(); 1095 if (Config->ZCombreloc && NumRelativeRels) 1096 add({IsRela ? DT_RELACOUNT : DT_RELCOUNT, NumRelativeRels}); 1097 } 1098 } 1099 if (In<ELFT>::RelaPlt->getParent() && !In<ELFT>::RelaPlt->empty()) { 1100 add({DT_JMPREL, In<ELFT>::RelaPlt}); 1101 add({DT_PLTRELSZ, In<ELFT>::RelaPlt->getParent(), Entry::SecSize}); 1102 switch (Config->EMachine) { 1103 case EM_MIPS: 1104 add({DT_MIPS_PLTGOT, In<ELFT>::GotPlt}); 1105 break; 1106 case EM_SPARCV9: 1107 add({DT_PLTGOT, In<ELFT>::Plt}); 1108 break; 1109 default: 1110 add({DT_PLTGOT, In<ELFT>::GotPlt}); 1111 break; 1112 } 1113 add({DT_PLTREL, uint64_t(Config->IsRela ? DT_RELA : DT_REL)}); 1114 } 1115 1116 add({DT_SYMTAB, InX::DynSymTab}); 1117 add({DT_SYMENT, sizeof(Elf_Sym)}); 1118 add({DT_STRTAB, InX::DynStrTab}); 1119 add({DT_STRSZ, InX::DynStrTab->getSize()}); 1120 if (!Config->ZText) 1121 add({DT_TEXTREL, (uint64_t)0}); 1122 if (InX::GnuHashTab) 1123 add({DT_GNU_HASH, InX::GnuHashTab}); 1124 if (InX::HashTab) 1125 add({DT_HASH, InX::HashTab}); 1126 1127 if (Out::PreinitArray) { 1128 add({DT_PREINIT_ARRAY, Out::PreinitArray}); 1129 add({DT_PREINIT_ARRAYSZ, Out::PreinitArray, Entry::SecSize}); 1130 } 1131 if (Out::InitArray) { 1132 add({DT_INIT_ARRAY, Out::InitArray}); 1133 add({DT_INIT_ARRAYSZ, Out::InitArray, Entry::SecSize}); 1134 } 1135 if (Out::FiniArray) { 1136 add({DT_FINI_ARRAY, Out::FiniArray}); 1137 add({DT_FINI_ARRAYSZ, Out::FiniArray, Entry::SecSize}); 1138 } 1139 1140 if (SymbolBody *B = Symtab->find(Config->Init)) 1141 if (B->isInCurrentDSO()) 1142 add({DT_INIT, B}); 1143 if (SymbolBody *B = Symtab->find(Config->Fini)) 1144 if (B->isInCurrentDSO()) 1145 add({DT_FINI, B}); 1146 1147 bool HasVerNeed = In<ELFT>::VerNeed->getNeedNum() != 0; 1148 if (HasVerNeed || In<ELFT>::VerDef) 1149 add({DT_VERSYM, In<ELFT>::VerSym}); 1150 if (In<ELFT>::VerDef) { 1151 add({DT_VERDEF, In<ELFT>::VerDef}); 1152 add({DT_VERDEFNUM, getVerDefNum()}); 1153 } 1154 if (HasVerNeed) { 1155 add({DT_VERNEED, In<ELFT>::VerNeed}); 1156 add({DT_VERNEEDNUM, In<ELFT>::VerNeed->getNeedNum()}); 1157 } 1158 1159 if (Config->EMachine == EM_MIPS) { 1160 add({DT_MIPS_RLD_VERSION, 1}); 1161 add({DT_MIPS_FLAGS, RHF_NOTPOT}); 1162 add({DT_MIPS_BASE_ADDRESS, Config->ImageBase}); 1163 add({DT_MIPS_SYMTABNO, InX::DynSymTab->getNumSymbols()}); 1164 add({DT_MIPS_LOCAL_GOTNO, InX::MipsGot->getLocalEntriesNum()}); 1165 if (const SymbolBody *B = InX::MipsGot->getFirstGlobalEntry()) 1166 add({DT_MIPS_GOTSYM, B->DynsymIndex}); 1167 else 1168 add({DT_MIPS_GOTSYM, InX::DynSymTab->getNumSymbols()}); 1169 add({DT_PLTGOT, InX::MipsGot}); 1170 if (InX::MipsRldMap) 1171 add({DT_MIPS_RLD_MAP, InX::MipsRldMap}); 1172 } 1173 1174 getParent()->Link = this->Link; 1175 1176 // +1 for DT_NULL 1177 this->Size = (Entries.size() + 1) * this->Entsize; 1178 } 1179 1180 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *Buf) { 1181 auto *P = reinterpret_cast<Elf_Dyn *>(Buf); 1182 1183 for (const Entry &E : Entries) { 1184 P->d_tag = E.Tag; 1185 switch (E.Kind) { 1186 case Entry::SecAddr: 1187 P->d_un.d_ptr = E.OutSec->Addr; 1188 break; 1189 case Entry::InSecAddr: 1190 P->d_un.d_ptr = E.InSec->getParent()->Addr + E.InSec->OutSecOff; 1191 break; 1192 case Entry::SecSize: 1193 P->d_un.d_val = E.OutSec->Size; 1194 break; 1195 case Entry::SymAddr: 1196 P->d_un.d_ptr = E.Sym->getVA(); 1197 break; 1198 case Entry::PlainInt: 1199 P->d_un.d_val = E.Val; 1200 break; 1201 } 1202 ++P; 1203 } 1204 } 1205 1206 uint64_t DynamicReloc::getOffset() const { 1207 return InputSec->getOutputSection()->Addr + InputSec->getOffset(OffsetInSec); 1208 } 1209 1210 int64_t DynamicReloc::getAddend() const { 1211 if (UseSymVA) 1212 return Sym->getVA(Addend); 1213 return Addend; 1214 } 1215 1216 uint32_t DynamicReloc::getSymIndex() const { 1217 if (Sym && !UseSymVA) 1218 return Sym->DynsymIndex; 1219 return 0; 1220 } 1221 1222 template <class ELFT> 1223 RelocationSection<ELFT>::RelocationSection(StringRef Name, bool Sort) 1224 : SyntheticSection(SHF_ALLOC, Config->IsRela ? SHT_RELA : SHT_REL, 1225 Config->Wordsize, Name), 1226 Sort(Sort) { 1227 this->Entsize = Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1228 } 1229 1230 template <class ELFT> 1231 void RelocationSection<ELFT>::addReloc(const DynamicReloc &Reloc) { 1232 if (Reloc.Type == Target->RelativeRel) 1233 ++NumRelativeRelocs; 1234 Relocs.push_back(Reloc); 1235 } 1236 1237 template <class ELFT, class RelTy> 1238 static bool compRelocations(const RelTy &A, const RelTy &B) { 1239 bool AIsRel = A.getType(Config->IsMips64EL) == Target->RelativeRel; 1240 bool BIsRel = B.getType(Config->IsMips64EL) == Target->RelativeRel; 1241 if (AIsRel != BIsRel) 1242 return AIsRel; 1243 1244 return A.getSymbol(Config->IsMips64EL) < B.getSymbol(Config->IsMips64EL); 1245 } 1246 1247 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *Buf) { 1248 uint8_t *BufBegin = Buf; 1249 for (const DynamicReloc &Rel : Relocs) { 1250 auto *P = reinterpret_cast<Elf_Rela *>(Buf); 1251 Buf += Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1252 1253 if (Config->IsRela) 1254 P->r_addend = Rel.getAddend(); 1255 P->r_offset = Rel.getOffset(); 1256 if (Config->EMachine == EM_MIPS && Rel.getInputSec() == InX::MipsGot) 1257 // Dynamic relocation against MIPS GOT section make deal TLS entries 1258 // allocated in the end of the GOT. We need to adjust the offset to take 1259 // in account 'local' and 'global' GOT entries. 1260 P->r_offset += InX::MipsGot->getTlsOffset(); 1261 P->setSymbolAndType(Rel.getSymIndex(), Rel.Type, Config->IsMips64EL); 1262 } 1263 1264 if (Sort) { 1265 if (Config->IsRela) 1266 std::stable_sort((Elf_Rela *)BufBegin, 1267 (Elf_Rela *)BufBegin + Relocs.size(), 1268 compRelocations<ELFT, Elf_Rela>); 1269 else 1270 std::stable_sort((Elf_Rel *)BufBegin, (Elf_Rel *)BufBegin + Relocs.size(), 1271 compRelocations<ELFT, Elf_Rel>); 1272 } 1273 } 1274 1275 template <class ELFT> unsigned RelocationSection<ELFT>::getRelocOffset() { 1276 return this->Entsize * Relocs.size(); 1277 } 1278 1279 template <class ELFT> void RelocationSection<ELFT>::finalizeContents() { 1280 this->Link = InX::DynSymTab ? InX::DynSymTab->getParent()->SectionIndex 1281 : InX::SymTab->getParent()->SectionIndex; 1282 1283 // Set required output section properties. 1284 getParent()->Link = this->Link; 1285 } 1286 1287 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &StrTabSec) 1288 : SyntheticSection(StrTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0, 1289 StrTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB, 1290 Config->Wordsize, 1291 StrTabSec.isDynamic() ? ".dynsym" : ".symtab"), 1292 StrTabSec(StrTabSec) {} 1293 1294 // Orders symbols according to their positions in the GOT, 1295 // in compliance with MIPS ABI rules. 1296 // See "Global Offset Table" in Chapter 5 in the following document 1297 // for detailed description: 1298 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 1299 static bool sortMipsSymbols(const SymbolTableEntry &L, 1300 const SymbolTableEntry &R) { 1301 // Sort entries related to non-local preemptible symbols by GOT indexes. 1302 // All other entries go to the first part of GOT in arbitrary order. 1303 bool LIsInLocalGot = !L.Symbol->IsInGlobalMipsGot; 1304 bool RIsInLocalGot = !R.Symbol->IsInGlobalMipsGot; 1305 if (LIsInLocalGot || RIsInLocalGot) 1306 return !RIsInLocalGot; 1307 return L.Symbol->GotIndex < R.Symbol->GotIndex; 1308 } 1309 1310 // Finalize a symbol table. The ELF spec requires that all local 1311 // symbols precede global symbols, so we sort symbol entries in this 1312 // function. (For .dynsym, we don't do that because symbols for 1313 // dynamic linking are inherently all globals.) 1314 void SymbolTableBaseSection::finalizeContents() { 1315 getParent()->Link = StrTabSec.getParent()->SectionIndex; 1316 1317 // If it is a .dynsym, there should be no local symbols, but we need 1318 // to do a few things for the dynamic linker. 1319 if (this->Type == SHT_DYNSYM) { 1320 // Section's Info field has the index of the first non-local symbol. 1321 // Because the first symbol entry is a null entry, 1 is the first. 1322 getParent()->Info = 1; 1323 1324 if (InX::GnuHashTab) { 1325 // NB: It also sorts Symbols to meet the GNU hash table requirements. 1326 InX::GnuHashTab->addSymbols(Symbols); 1327 } else if (Config->EMachine == EM_MIPS) { 1328 std::stable_sort(Symbols.begin(), Symbols.end(), sortMipsSymbols); 1329 } 1330 1331 size_t I = 0; 1332 for (const SymbolTableEntry &S : Symbols) 1333 S.Symbol->DynsymIndex = ++I; 1334 return; 1335 } 1336 } 1337 1338 void SymbolTableBaseSection::postThunkContents() { 1339 if (this->Type == SHT_DYNSYM) 1340 return; 1341 // move all local symbols before global symbols. 1342 auto It = std::stable_partition( 1343 Symbols.begin(), Symbols.end(), [](const SymbolTableEntry &S) { 1344 return S.Symbol->isLocal() || 1345 S.Symbol->symbol()->computeBinding() == STB_LOCAL; 1346 }); 1347 size_t NumLocals = It - Symbols.begin(); 1348 getParent()->Info = NumLocals + 1; 1349 } 1350 1351 void SymbolTableBaseSection::addSymbol(SymbolBody *B) { 1352 // Adding a local symbol to a .dynsym is a bug. 1353 assert(this->Type != SHT_DYNSYM || !B->isLocal()); 1354 1355 bool HashIt = B->isLocal(); 1356 Symbols.push_back({B, StrTabSec.addString(B->getName(), HashIt)}); 1357 } 1358 1359 size_t SymbolTableBaseSection::getSymbolIndex(SymbolBody *Body) { 1360 // Initializes symbol lookup tables lazily. This is used only 1361 // for -r or -emit-relocs. 1362 llvm::call_once(OnceFlag, [&] { 1363 SymbolIndexMap.reserve(Symbols.size()); 1364 size_t I = 0; 1365 for (const SymbolTableEntry &E : Symbols) { 1366 if (E.Symbol->Type == STT_SECTION) 1367 SectionIndexMap[E.Symbol->getOutputSection()] = ++I; 1368 else 1369 SymbolIndexMap[E.Symbol] = ++I; 1370 } 1371 }); 1372 1373 // Section symbols are mapped based on their output sections 1374 // to maintain their semantics. 1375 if (Body->Type == STT_SECTION) 1376 return SectionIndexMap.lookup(Body->getOutputSection()); 1377 return SymbolIndexMap.lookup(Body); 1378 } 1379 1380 template <class ELFT> 1381 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &StrTabSec) 1382 : SymbolTableBaseSection(StrTabSec) { 1383 this->Entsize = sizeof(Elf_Sym); 1384 } 1385 1386 // Write the internal symbol table contents to the output symbol table. 1387 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *Buf) { 1388 // The first entry is a null entry as per the ELF spec. 1389 Buf += sizeof(Elf_Sym); 1390 1391 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf); 1392 1393 for (SymbolTableEntry &Ent : Symbols) { 1394 SymbolBody *Body = Ent.Symbol; 1395 1396 // Set st_info and st_other. 1397 if (Body->isLocal()) { 1398 ESym->setBindingAndType(STB_LOCAL, Body->Type); 1399 } else { 1400 ESym->setBindingAndType(Body->symbol()->computeBinding(), Body->Type); 1401 ESym->setVisibility(Body->symbol()->Visibility); 1402 } 1403 1404 ESym->st_name = Ent.StrTabOffset; 1405 1406 // Set a section index. 1407 if (const OutputSection *OutSec = Body->getOutputSection()) 1408 ESym->st_shndx = OutSec->SectionIndex; 1409 else if (isa<DefinedRegular>(Body)) 1410 ESym->st_shndx = SHN_ABS; 1411 else if (isa<DefinedCommon>(Body)) 1412 ESym->st_shndx = SHN_COMMON; 1413 1414 // Copy symbol size if it is a defined symbol. st_size is not significant 1415 // for undefined symbols, so whether copying it or not is up to us if that's 1416 // the case. We'll leave it as zero because by not setting a value, we can 1417 // get the exact same outputs for two sets of input files that differ only 1418 // in undefined symbol size in DSOs. 1419 if (ESym->st_shndx != SHN_UNDEF) 1420 ESym->st_size = Body->getSize<ELFT>(); 1421 1422 // st_value is usually an address of a symbol, but that has a 1423 // special meaining for uninstantiated common symbols (this can 1424 // occur if -r is given). 1425 if (!Config->DefineCommon && isa<DefinedCommon>(Body)) 1426 ESym->st_value = cast<DefinedCommon>(Body)->Alignment; 1427 else 1428 ESym->st_value = Body->getVA(); 1429 1430 ++ESym; 1431 } 1432 1433 // On MIPS we need to mark symbol which has a PLT entry and requires 1434 // pointer equality by STO_MIPS_PLT flag. That is necessary to help 1435 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs. 1436 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt 1437 if (Config->EMachine == EM_MIPS) { 1438 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf); 1439 1440 for (SymbolTableEntry &Ent : Symbols) { 1441 SymbolBody *Body = Ent.Symbol; 1442 if (Body->isInPlt() && Body->NeedsPltAddr) 1443 ESym->st_other |= STO_MIPS_PLT; 1444 1445 if (Config->Relocatable) 1446 if (auto *D = dyn_cast<DefinedRegular>(Body)) 1447 if (D->isMipsPIC<ELFT>()) 1448 ESym->st_other |= STO_MIPS_PIC; 1449 ++ESym; 1450 } 1451 } 1452 } 1453 1454 // .hash and .gnu.hash sections contain on-disk hash tables that map 1455 // symbol names to their dynamic symbol table indices. Their purpose 1456 // is to help the dynamic linker resolve symbols quickly. If ELF files 1457 // don't have them, the dynamic linker has to do linear search on all 1458 // dynamic symbols, which makes programs slower. Therefore, a .hash 1459 // section is added to a DSO by default. A .gnu.hash is added if you 1460 // give the -hash-style=gnu or -hash-style=both option. 1461 // 1462 // The Unix semantics of resolving dynamic symbols is somewhat expensive. 1463 // Each ELF file has a list of DSOs that the ELF file depends on and a 1464 // list of dynamic symbols that need to be resolved from any of the 1465 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n) 1466 // where m is the number of DSOs and n is the number of dynamic 1467 // symbols. For modern large programs, both m and n are large. So 1468 // making each step faster by using hash tables substiantially 1469 // improves time to load programs. 1470 // 1471 // (Note that this is not the only way to design the shared library. 1472 // For instance, the Windows DLL takes a different approach. On 1473 // Windows, each dynamic symbol has a name of DLL from which the symbol 1474 // has to be resolved. That makes the cost of symbol resolution O(n). 1475 // This disables some hacky techniques you can use on Unix such as 1476 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.) 1477 // 1478 // Due to historical reasons, we have two different hash tables, .hash 1479 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new 1480 // and better version of .hash. .hash is just an on-disk hash table, but 1481 // .gnu.hash has a bloom filter in addition to a hash table to skip 1482 // DSOs very quickly. If you are sure that your dynamic linker knows 1483 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a 1484 // safe bet is to specify -hash-style=both for backward compatibilty. 1485 GnuHashTableSection::GnuHashTableSection() 1486 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, Config->Wordsize, ".gnu.hash") { 1487 } 1488 1489 void GnuHashTableSection::finalizeContents() { 1490 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex; 1491 1492 // Computes bloom filter size in word size. We want to allocate 8 1493 // bits for each symbol. It must be a power of two. 1494 if (Symbols.empty()) 1495 MaskWords = 1; 1496 else 1497 MaskWords = NextPowerOf2((Symbols.size() - 1) / Config->Wordsize); 1498 1499 Size = 16; // Header 1500 Size += Config->Wordsize * MaskWords; // Bloom filter 1501 Size += NBuckets * 4; // Hash buckets 1502 Size += Symbols.size() * 4; // Hash values 1503 } 1504 1505 void GnuHashTableSection::writeTo(uint8_t *Buf) { 1506 // Write a header. 1507 write32(Buf, NBuckets, Config->Endianness); 1508 write32(Buf + 4, InX::DynSymTab->getNumSymbols() - Symbols.size(), 1509 Config->Endianness); 1510 write32(Buf + 8, MaskWords, Config->Endianness); 1511 write32(Buf + 12, getShift2(), Config->Endianness); 1512 Buf += 16; 1513 1514 // Write a bloom filter and a hash table. 1515 writeBloomFilter(Buf); 1516 Buf += Config->Wordsize * MaskWords; 1517 writeHashTable(Buf); 1518 } 1519 1520 // This function writes a 2-bit bloom filter. This bloom filter alone 1521 // usually filters out 80% or more of all symbol lookups [1]. 1522 // The dynamic linker uses the hash table only when a symbol is not 1523 // filtered out by a bloom filter. 1524 // 1525 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2), 1526 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf 1527 void GnuHashTableSection::writeBloomFilter(uint8_t *Buf) { 1528 const unsigned C = Config->Wordsize * 8; 1529 for (const Entry &Sym : Symbols) { 1530 size_t I = (Sym.Hash / C) & (MaskWords - 1); 1531 uint64_t Val = readUint(Buf + I * Config->Wordsize); 1532 Val |= uint64_t(1) << (Sym.Hash % C); 1533 Val |= uint64_t(1) << ((Sym.Hash >> getShift2()) % C); 1534 writeUint(Buf + I * Config->Wordsize, Val); 1535 } 1536 } 1537 1538 void GnuHashTableSection::writeHashTable(uint8_t *Buf) { 1539 // Group symbols by hash value. 1540 std::vector<std::vector<Entry>> Syms(NBuckets); 1541 for (const Entry &Ent : Symbols) 1542 Syms[Ent.Hash % NBuckets].push_back(Ent); 1543 1544 // Write hash buckets. Hash buckets contain indices in the following 1545 // hash value table. 1546 uint32_t *Buckets = reinterpret_cast<uint32_t *>(Buf); 1547 for (size_t I = 0; I < NBuckets; ++I) 1548 if (!Syms[I].empty()) 1549 write32(Buckets + I, Syms[I][0].Body->DynsymIndex, Config->Endianness); 1550 1551 // Write a hash value table. It represents a sequence of chains that 1552 // share the same hash modulo value. The last element of each chain 1553 // is terminated by LSB 1. 1554 uint32_t *Values = Buckets + NBuckets; 1555 size_t I = 0; 1556 for (std::vector<Entry> &Vec : Syms) { 1557 if (Vec.empty()) 1558 continue; 1559 for (const Entry &Ent : makeArrayRef(Vec).drop_back()) 1560 write32(Values + I++, Ent.Hash & ~1, Config->Endianness); 1561 write32(Values + I++, Vec.back().Hash | 1, Config->Endianness); 1562 } 1563 } 1564 1565 static uint32_t hashGnu(StringRef Name) { 1566 uint32_t H = 5381; 1567 for (uint8_t C : Name) 1568 H = (H << 5) + H + C; 1569 return H; 1570 } 1571 1572 // Returns a number of hash buckets to accomodate given number of elements. 1573 // We want to choose a moderate number that is not too small (which 1574 // causes too many hash collisions) and not too large (which wastes 1575 // disk space.) 1576 // 1577 // We return a prime number because it (is believed to) achieve good 1578 // hash distribution. 1579 static size_t getBucketSize(size_t NumSymbols) { 1580 // List of largest prime numbers that are not greater than 2^n + 1. 1581 for (size_t N : {131071, 65521, 32749, 16381, 8191, 4093, 2039, 1021, 509, 1582 251, 127, 61, 31, 13, 7, 3, 1}) 1583 if (N <= NumSymbols) 1584 return N; 1585 return 0; 1586 } 1587 1588 // Add symbols to this symbol hash table. Note that this function 1589 // destructively sort a given vector -- which is needed because 1590 // GNU-style hash table places some sorting requirements. 1591 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &V) { 1592 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce 1593 // its type correctly. 1594 std::vector<SymbolTableEntry>::iterator Mid = 1595 std::stable_partition(V.begin(), V.end(), [](const SymbolTableEntry &S) { 1596 return S.Symbol->isUndefined(); 1597 }); 1598 if (Mid == V.end()) 1599 return; 1600 1601 for (SymbolTableEntry &Ent : llvm::make_range(Mid, V.end())) { 1602 SymbolBody *B = Ent.Symbol; 1603 Symbols.push_back({B, Ent.StrTabOffset, hashGnu(B->getName())}); 1604 } 1605 1606 NBuckets = getBucketSize(Symbols.size()); 1607 std::stable_sort(Symbols.begin(), Symbols.end(), 1608 [&](const Entry &L, const Entry &R) { 1609 return L.Hash % NBuckets < R.Hash % NBuckets; 1610 }); 1611 1612 V.erase(Mid, V.end()); 1613 for (const Entry &Ent : Symbols) 1614 V.push_back({Ent.Body, Ent.StrTabOffset}); 1615 } 1616 1617 HashTableSection::HashTableSection() 1618 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") { 1619 this->Entsize = 4; 1620 } 1621 1622 void HashTableSection::finalizeContents() { 1623 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex; 1624 1625 unsigned NumEntries = 2; // nbucket and nchain. 1626 NumEntries += InX::DynSymTab->getNumSymbols(); // The chain entries. 1627 1628 // Create as many buckets as there are symbols. 1629 // FIXME: This is simplistic. We can try to optimize it, but implementing 1630 // support for SHT_GNU_HASH is probably even more profitable. 1631 NumEntries += InX::DynSymTab->getNumSymbols(); 1632 this->Size = NumEntries * 4; 1633 } 1634 1635 void HashTableSection::writeTo(uint8_t *Buf) { 1636 unsigned NumSymbols = InX::DynSymTab->getNumSymbols(); 1637 1638 uint32_t *P = reinterpret_cast<uint32_t *>(Buf); 1639 write32(P++, NumSymbols, Config->Endianness); // nbucket 1640 write32(P++, NumSymbols, Config->Endianness); // nchain 1641 1642 uint32_t *Buckets = P; 1643 uint32_t *Chains = P + NumSymbols; 1644 1645 for (const SymbolTableEntry &S : InX::DynSymTab->getSymbols()) { 1646 SymbolBody *Body = S.Symbol; 1647 StringRef Name = Body->getName(); 1648 unsigned I = Body->DynsymIndex; 1649 uint32_t Hash = hashSysV(Name) % NumSymbols; 1650 Chains[I] = Buckets[Hash]; 1651 write32(Buckets + Hash, I, Config->Endianness); 1652 } 1653 } 1654 1655 PltSection::PltSection(size_t S) 1656 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt"), 1657 HeaderSize(S) { 1658 // The PLT needs to be writable on SPARC as the dynamic linker will 1659 // modify the instructions in the PLT entries. 1660 if (Config->EMachine == EM_SPARCV9) 1661 this->Flags |= SHF_WRITE; 1662 } 1663 1664 void PltSection::writeTo(uint8_t *Buf) { 1665 // At beginning of PLT but not the IPLT, we have code to call the dynamic 1666 // linker to resolve dynsyms at runtime. Write such code. 1667 if (HeaderSize != 0) 1668 Target->writePltHeader(Buf); 1669 size_t Off = HeaderSize; 1670 // The IPlt is immediately after the Plt, account for this in RelOff 1671 unsigned PltOff = getPltRelocOff(); 1672 1673 for (auto &I : Entries) { 1674 const SymbolBody *B = I.first; 1675 unsigned RelOff = I.second + PltOff; 1676 uint64_t Got = B->getGotPltVA(); 1677 uint64_t Plt = this->getVA() + Off; 1678 Target->writePlt(Buf + Off, Got, Plt, B->PltIndex, RelOff); 1679 Off += Target->PltEntrySize; 1680 } 1681 } 1682 1683 template <class ELFT> void PltSection::addEntry(SymbolBody &Sym) { 1684 Sym.PltIndex = Entries.size(); 1685 RelocationSection<ELFT> *PltRelocSection = In<ELFT>::RelaPlt; 1686 if (HeaderSize == 0) { 1687 PltRelocSection = In<ELFT>::RelaIplt; 1688 Sym.IsInIplt = true; 1689 } 1690 unsigned RelOff = PltRelocSection->getRelocOffset(); 1691 Entries.push_back(std::make_pair(&Sym, RelOff)); 1692 } 1693 1694 size_t PltSection::getSize() const { 1695 return HeaderSize + Entries.size() * Target->PltEntrySize; 1696 } 1697 1698 // Some architectures such as additional symbols in the PLT section. For 1699 // example ARM uses mapping symbols to aid disassembly 1700 void PltSection::addSymbols() { 1701 // The PLT may have symbols defined for the Header, the IPLT has no header 1702 if (HeaderSize != 0) 1703 Target->addPltHeaderSymbols(this); 1704 size_t Off = HeaderSize; 1705 for (size_t I = 0; I < Entries.size(); ++I) { 1706 Target->addPltSymbols(this, Off); 1707 Off += Target->PltEntrySize; 1708 } 1709 } 1710 1711 unsigned PltSection::getPltRelocOff() const { 1712 return (HeaderSize == 0) ? InX::Plt->getSize() : 0; 1713 } 1714 1715 // The string hash function for .gdb_index. 1716 static uint32_t computeGdbHash(StringRef S) { 1717 uint32_t H = 0; 1718 for (uint8_t C : S) 1719 H = H * 67 + tolower(C) - 113; 1720 return H; 1721 } 1722 1723 static std::vector<GdbIndexChunk::CuEntry> readCuList(DWARFContext &Dwarf) { 1724 std::vector<GdbIndexChunk::CuEntry> Ret; 1725 for (std::unique_ptr<DWARFCompileUnit> &Cu : Dwarf.compile_units()) 1726 Ret.push_back({Cu->getOffset(), Cu->getLength() + 4}); 1727 return Ret; 1728 } 1729 1730 static std::vector<GdbIndexChunk::AddressEntry> 1731 readAddressAreas(DWARFContext &Dwarf, InputSection *Sec) { 1732 std::vector<GdbIndexChunk::AddressEntry> Ret; 1733 1734 uint32_t CuIdx = 0; 1735 for (std::unique_ptr<DWARFCompileUnit> &Cu : Dwarf.compile_units()) { 1736 DWARFAddressRangesVector Ranges; 1737 Cu->collectAddressRanges(Ranges); 1738 1739 ArrayRef<InputSectionBase *> Sections = Sec->File->getSections(); 1740 for (DWARFAddressRange &R : Ranges) { 1741 InputSectionBase *S = Sections[R.SectionIndex]; 1742 if (!S || S == &InputSection::Discarded || !S->Live) 1743 continue; 1744 // Range list with zero size has no effect. 1745 if (R.LowPC == R.HighPC) 1746 continue; 1747 auto *IS = cast<InputSection>(S); 1748 uint64_t Offset = IS->getOffsetInFile(); 1749 Ret.push_back({IS, R.LowPC - Offset, R.HighPC - Offset, CuIdx}); 1750 } 1751 ++CuIdx; 1752 } 1753 return Ret; 1754 } 1755 1756 static std::vector<GdbIndexChunk::NameTypeEntry> 1757 readPubNamesAndTypes(DWARFContext &Dwarf) { 1758 StringRef Sec1 = Dwarf.getDWARFObj().getGnuPubNamesSection(); 1759 StringRef Sec2 = Dwarf.getDWARFObj().getGnuPubTypesSection(); 1760 1761 std::vector<GdbIndexChunk::NameTypeEntry> Ret; 1762 for (StringRef Sec : {Sec1, Sec2}) { 1763 DWARFDebugPubTable Table(Sec, Config->IsLE, true); 1764 for (const DWARFDebugPubTable::Set &Set : Table.getData()) { 1765 for (const DWARFDebugPubTable::Entry &Ent : Set.Entries) { 1766 CachedHashStringRef S(Ent.Name, computeGdbHash(Ent.Name)); 1767 Ret.push_back({S, Ent.Descriptor.toBits()}); 1768 } 1769 } 1770 } 1771 return Ret; 1772 } 1773 1774 static std::vector<InputSection *> getDebugInfoSections() { 1775 std::vector<InputSection *> Ret; 1776 for (InputSectionBase *S : InputSections) 1777 if (InputSection *IS = dyn_cast<InputSection>(S)) 1778 if (IS->Name == ".debug_info") 1779 Ret.push_back(IS); 1780 return Ret; 1781 } 1782 1783 void GdbIndexSection::fixCuIndex() { 1784 uint32_t Idx = 0; 1785 for (GdbIndexChunk &Chunk : Chunks) { 1786 for (GdbIndexChunk::AddressEntry &Ent : Chunk.AddressAreas) 1787 Ent.CuIndex += Idx; 1788 Idx += Chunk.CompilationUnits.size(); 1789 } 1790 } 1791 1792 std::vector<std::vector<uint32_t>> GdbIndexSection::createCuVectors() { 1793 std::vector<std::vector<uint32_t>> Ret; 1794 uint32_t Idx = 0; 1795 uint32_t Off = 0; 1796 1797 for (GdbIndexChunk &Chunk : Chunks) { 1798 for (GdbIndexChunk::NameTypeEntry &Ent : Chunk.NamesAndTypes) { 1799 GdbSymbol *&Sym = Symbols[Ent.Name]; 1800 if (!Sym) { 1801 Sym = make<GdbSymbol>(GdbSymbol{Ent.Name.hash(), Off, Ret.size()}); 1802 Off += Ent.Name.size() + 1; 1803 Ret.push_back({}); 1804 } 1805 1806 // gcc 5.4.1 produces a buggy .debug_gnu_pubnames that contains 1807 // duplicate entries, so we want to dedup them. 1808 std::vector<uint32_t> &Vec = Ret[Sym->CuVectorIndex]; 1809 uint32_t Val = (Ent.Type << 24) | Idx; 1810 if (Vec.empty() || Vec.back() != Val) 1811 Vec.push_back(Val); 1812 } 1813 Idx += Chunk.CompilationUnits.size(); 1814 } 1815 1816 StringPoolSize = Off; 1817 return Ret; 1818 } 1819 1820 template <class ELFT> GdbIndexSection *elf::createGdbIndex() { 1821 std::vector<InputSection *> Sections = getDebugInfoSections(); 1822 std::vector<GdbIndexChunk> Chunks(Sections.size()); 1823 1824 parallelForEachN(0, Chunks.size(), [&](size_t I) { 1825 ObjFile<ELFT> *File = Sections[I]->getFile<ELFT>(); 1826 DWARFContext Dwarf(make_unique<LLDDwarfObj<ELFT>>(File)); 1827 1828 Chunks[I].DebugInfoSec = Sections[I]; 1829 Chunks[I].CompilationUnits = readCuList(Dwarf); 1830 Chunks[I].AddressAreas = readAddressAreas(Dwarf, Sections[I]); 1831 Chunks[I].NamesAndTypes = readPubNamesAndTypes(Dwarf); 1832 }); 1833 1834 return make<GdbIndexSection>(std::move(Chunks)); 1835 } 1836 1837 static size_t getCuSize(ArrayRef<GdbIndexChunk> Arr) { 1838 size_t Ret = 0; 1839 for (const GdbIndexChunk &D : Arr) 1840 Ret += D.CompilationUnits.size(); 1841 return Ret; 1842 } 1843 1844 static size_t getAddressAreaSize(ArrayRef<GdbIndexChunk> Arr) { 1845 size_t Ret = 0; 1846 for (const GdbIndexChunk &D : Arr) 1847 Ret += D.AddressAreas.size(); 1848 return Ret; 1849 } 1850 1851 std::vector<GdbSymbol *> GdbIndexSection::createGdbSymtab() { 1852 uint32_t Size = NextPowerOf2(Symbols.size() * 4 / 3); 1853 if (Size < 1024) 1854 Size = 1024; 1855 1856 uint32_t Mask = Size - 1; 1857 std::vector<GdbSymbol *> Ret(Size); 1858 1859 for (auto &KV : Symbols) { 1860 GdbSymbol *Sym = KV.second; 1861 uint32_t I = Sym->NameHash & Mask; 1862 uint32_t Step = ((Sym->NameHash * 17) & Mask) | 1; 1863 1864 while (Ret[I]) 1865 I = (I + Step) & Mask; 1866 Ret[I] = Sym; 1867 } 1868 return Ret; 1869 } 1870 1871 GdbIndexSection::GdbIndexSection(std::vector<GdbIndexChunk> &&C) 1872 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index"), Chunks(std::move(C)) { 1873 fixCuIndex(); 1874 CuVectors = createCuVectors(); 1875 GdbSymtab = createGdbSymtab(); 1876 1877 // Compute offsets early to know the section size. 1878 // Each chunk size needs to be in sync with what we write in writeTo. 1879 CuTypesOffset = CuListOffset + getCuSize(Chunks) * 16; 1880 SymtabOffset = CuTypesOffset + getAddressAreaSize(Chunks) * 20; 1881 ConstantPoolOffset = SymtabOffset + GdbSymtab.size() * 8; 1882 1883 size_t Off = 0; 1884 for (ArrayRef<uint32_t> Vec : CuVectors) { 1885 CuVectorOffsets.push_back(Off); 1886 Off += (Vec.size() + 1) * 4; 1887 } 1888 StringPoolOffset = ConstantPoolOffset + Off; 1889 } 1890 1891 size_t GdbIndexSection::getSize() const { 1892 return StringPoolOffset + StringPoolSize; 1893 } 1894 1895 void GdbIndexSection::writeTo(uint8_t *Buf) { 1896 // Write the section header. 1897 write32le(Buf, 7); 1898 write32le(Buf + 4, CuListOffset); 1899 write32le(Buf + 8, CuTypesOffset); 1900 write32le(Buf + 12, CuTypesOffset); 1901 write32le(Buf + 16, SymtabOffset); 1902 write32le(Buf + 20, ConstantPoolOffset); 1903 Buf += 24; 1904 1905 // Write the CU list. 1906 for (GdbIndexChunk &D : Chunks) { 1907 for (GdbIndexChunk::CuEntry &Cu : D.CompilationUnits) { 1908 write64le(Buf, D.DebugInfoSec->OutSecOff + Cu.CuOffset); 1909 write64le(Buf + 8, Cu.CuLength); 1910 Buf += 16; 1911 } 1912 } 1913 1914 // Write the address area. 1915 for (GdbIndexChunk &D : Chunks) { 1916 for (GdbIndexChunk::AddressEntry &E : D.AddressAreas) { 1917 uint64_t BaseAddr = 1918 E.Section->getParent()->Addr + E.Section->getOffset(0); 1919 write64le(Buf, BaseAddr + E.LowAddress); 1920 write64le(Buf + 8, BaseAddr + E.HighAddress); 1921 write32le(Buf + 16, E.CuIndex); 1922 Buf += 20; 1923 } 1924 } 1925 1926 // Write the symbol table. 1927 for (GdbSymbol *Sym : GdbSymtab) { 1928 if (Sym) { 1929 write32le(Buf, Sym->NameOffset + StringPoolOffset - ConstantPoolOffset); 1930 write32le(Buf + 4, CuVectorOffsets[Sym->CuVectorIndex]); 1931 } 1932 Buf += 8; 1933 } 1934 1935 // Write the CU vectors. 1936 for (ArrayRef<uint32_t> Vec : CuVectors) { 1937 write32le(Buf, Vec.size()); 1938 Buf += 4; 1939 for (uint32_t Val : Vec) { 1940 write32le(Buf, Val); 1941 Buf += 4; 1942 } 1943 } 1944 1945 // Write the string pool. 1946 for (auto &KV : Symbols) { 1947 CachedHashStringRef S = KV.first; 1948 GdbSymbol *Sym = KV.second; 1949 size_t Off = Sym->NameOffset; 1950 memcpy(Buf + Off, S.val().data(), S.size()); 1951 Buf[Off + S.size()] = '\0'; 1952 } 1953 } 1954 1955 bool GdbIndexSection::empty() const { return !Out::DebugInfo; } 1956 1957 template <class ELFT> 1958 EhFrameHeader<ELFT>::EhFrameHeader() 1959 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame_hdr") {} 1960 1961 // .eh_frame_hdr contains a binary search table of pointers to FDEs. 1962 // Each entry of the search table consists of two values, 1963 // the starting PC from where FDEs covers, and the FDE's address. 1964 // It is sorted by PC. 1965 template <class ELFT> void EhFrameHeader<ELFT>::writeTo(uint8_t *Buf) { 1966 const endianness E = ELFT::TargetEndianness; 1967 1968 // Sort the FDE list by their PC and uniqueify. Usually there is only 1969 // one FDE for a PC (i.e. function), but if ICF merges two functions 1970 // into one, there can be more than one FDEs pointing to the address. 1971 auto Less = [](const FdeData &A, const FdeData &B) { return A.Pc < B.Pc; }; 1972 std::stable_sort(Fdes.begin(), Fdes.end(), Less); 1973 auto Eq = [](const FdeData &A, const FdeData &B) { return A.Pc == B.Pc; }; 1974 Fdes.erase(std::unique(Fdes.begin(), Fdes.end(), Eq), Fdes.end()); 1975 1976 Buf[0] = 1; 1977 Buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4; 1978 Buf[2] = DW_EH_PE_udata4; 1979 Buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4; 1980 write32<E>(Buf + 4, In<ELFT>::EhFrame->getParent()->Addr - this->getVA() - 4); 1981 write32<E>(Buf + 8, Fdes.size()); 1982 Buf += 12; 1983 1984 uint64_t VA = this->getVA(); 1985 for (FdeData &Fde : Fdes) { 1986 write32<E>(Buf, Fde.Pc - VA); 1987 write32<E>(Buf + 4, Fde.FdeVA - VA); 1988 Buf += 8; 1989 } 1990 } 1991 1992 template <class ELFT> size_t EhFrameHeader<ELFT>::getSize() const { 1993 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs. 1994 return 12 + In<ELFT>::EhFrame->NumFdes * 8; 1995 } 1996 1997 template <class ELFT> 1998 void EhFrameHeader<ELFT>::addFde(uint32_t Pc, uint32_t FdeVA) { 1999 Fdes.push_back({Pc, FdeVA}); 2000 } 2001 2002 template <class ELFT> bool EhFrameHeader<ELFT>::empty() const { 2003 return In<ELFT>::EhFrame->empty(); 2004 } 2005 2006 template <class ELFT> 2007 VersionDefinitionSection<ELFT>::VersionDefinitionSection() 2008 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t), 2009 ".gnu.version_d") {} 2010 2011 static StringRef getFileDefName() { 2012 if (!Config->SoName.empty()) 2013 return Config->SoName; 2014 return Config->OutputFile; 2015 } 2016 2017 template <class ELFT> void VersionDefinitionSection<ELFT>::finalizeContents() { 2018 FileDefNameOff = InX::DynStrTab->addString(getFileDefName()); 2019 for (VersionDefinition &V : Config->VersionDefinitions) 2020 V.NameOff = InX::DynStrTab->addString(V.Name); 2021 2022 getParent()->Link = InX::DynStrTab->getParent()->SectionIndex; 2023 2024 // sh_info should be set to the number of definitions. This fact is missed in 2025 // documentation, but confirmed by binutils community: 2026 // https://sourceware.org/ml/binutils/2014-11/msg00355.html 2027 getParent()->Info = getVerDefNum(); 2028 } 2029 2030 template <class ELFT> 2031 void VersionDefinitionSection<ELFT>::writeOne(uint8_t *Buf, uint32_t Index, 2032 StringRef Name, size_t NameOff) { 2033 auto *Verdef = reinterpret_cast<Elf_Verdef *>(Buf); 2034 Verdef->vd_version = 1; 2035 Verdef->vd_cnt = 1; 2036 Verdef->vd_aux = sizeof(Elf_Verdef); 2037 Verdef->vd_next = sizeof(Elf_Verdef) + sizeof(Elf_Verdaux); 2038 Verdef->vd_flags = (Index == 1 ? VER_FLG_BASE : 0); 2039 Verdef->vd_ndx = Index; 2040 Verdef->vd_hash = hashSysV(Name); 2041 2042 auto *Verdaux = reinterpret_cast<Elf_Verdaux *>(Buf + sizeof(Elf_Verdef)); 2043 Verdaux->vda_name = NameOff; 2044 Verdaux->vda_next = 0; 2045 } 2046 2047 template <class ELFT> 2048 void VersionDefinitionSection<ELFT>::writeTo(uint8_t *Buf) { 2049 writeOne(Buf, 1, getFileDefName(), FileDefNameOff); 2050 2051 for (VersionDefinition &V : Config->VersionDefinitions) { 2052 Buf += sizeof(Elf_Verdef) + sizeof(Elf_Verdaux); 2053 writeOne(Buf, V.Id, V.Name, V.NameOff); 2054 } 2055 2056 // Need to terminate the last version definition. 2057 Elf_Verdef *Verdef = reinterpret_cast<Elf_Verdef *>(Buf); 2058 Verdef->vd_next = 0; 2059 } 2060 2061 template <class ELFT> size_t VersionDefinitionSection<ELFT>::getSize() const { 2062 return (sizeof(Elf_Verdef) + sizeof(Elf_Verdaux)) * getVerDefNum(); 2063 } 2064 2065 template <class ELFT> 2066 VersionTableSection<ELFT>::VersionTableSection() 2067 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t), 2068 ".gnu.version") { 2069 this->Entsize = sizeof(Elf_Versym); 2070 } 2071 2072 template <class ELFT> void VersionTableSection<ELFT>::finalizeContents() { 2073 // At the moment of june 2016 GNU docs does not mention that sh_link field 2074 // should be set, but Sun docs do. Also readelf relies on this field. 2075 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex; 2076 } 2077 2078 template <class ELFT> size_t VersionTableSection<ELFT>::getSize() const { 2079 return sizeof(Elf_Versym) * (InX::DynSymTab->getSymbols().size() + 1); 2080 } 2081 2082 template <class ELFT> void VersionTableSection<ELFT>::writeTo(uint8_t *Buf) { 2083 auto *OutVersym = reinterpret_cast<Elf_Versym *>(Buf) + 1; 2084 for (const SymbolTableEntry &S : InX::DynSymTab->getSymbols()) { 2085 OutVersym->vs_index = S.Symbol->symbol()->VersionId; 2086 ++OutVersym; 2087 } 2088 } 2089 2090 template <class ELFT> bool VersionTableSection<ELFT>::empty() const { 2091 return !In<ELFT>::VerDef && In<ELFT>::VerNeed->empty(); 2092 } 2093 2094 template <class ELFT> 2095 VersionNeedSection<ELFT>::VersionNeedSection() 2096 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t), 2097 ".gnu.version_r") { 2098 // Identifiers in verneed section start at 2 because 0 and 1 are reserved 2099 // for VER_NDX_LOCAL and VER_NDX_GLOBAL. 2100 // First identifiers are reserved by verdef section if it exist. 2101 NextIndex = getVerDefNum() + 1; 2102 } 2103 2104 template <class ELFT> 2105 void VersionNeedSection<ELFT>::addSymbol(SharedSymbol *SS) { 2106 auto *Ver = reinterpret_cast<const typename ELFT::Verdef *>(SS->Verdef); 2107 if (!Ver) { 2108 SS->symbol()->VersionId = VER_NDX_GLOBAL; 2109 return; 2110 } 2111 2112 SharedFile<ELFT> *File = SS->getFile<ELFT>(); 2113 2114 // If we don't already know that we need an Elf_Verneed for this DSO, prepare 2115 // to create one by adding it to our needed list and creating a dynstr entry 2116 // for the soname. 2117 if (File->VerdefMap.empty()) 2118 Needed.push_back({File, InX::DynStrTab->addString(File->SoName)}); 2119 typename SharedFile<ELFT>::NeededVer &NV = File->VerdefMap[Ver]; 2120 // If we don't already know that we need an Elf_Vernaux for this Elf_Verdef, 2121 // prepare to create one by allocating a version identifier and creating a 2122 // dynstr entry for the version name. 2123 if (NV.Index == 0) { 2124 NV.StrTab = InX::DynStrTab->addString(File->getStringTable().data() + 2125 Ver->getAux()->vda_name); 2126 NV.Index = NextIndex++; 2127 } 2128 SS->symbol()->VersionId = NV.Index; 2129 } 2130 2131 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *Buf) { 2132 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs. 2133 auto *Verneed = reinterpret_cast<Elf_Verneed *>(Buf); 2134 auto *Vernaux = reinterpret_cast<Elf_Vernaux *>(Verneed + Needed.size()); 2135 2136 for (std::pair<SharedFile<ELFT> *, size_t> &P : Needed) { 2137 // Create an Elf_Verneed for this DSO. 2138 Verneed->vn_version = 1; 2139 Verneed->vn_cnt = P.first->VerdefMap.size(); 2140 Verneed->vn_file = P.second; 2141 Verneed->vn_aux = 2142 reinterpret_cast<char *>(Vernaux) - reinterpret_cast<char *>(Verneed); 2143 Verneed->vn_next = sizeof(Elf_Verneed); 2144 ++Verneed; 2145 2146 // Create the Elf_Vernauxs for this Elf_Verneed. The loop iterates over 2147 // VerdefMap, which will only contain references to needed version 2148 // definitions. Each Elf_Vernaux is based on the information contained in 2149 // the Elf_Verdef in the source DSO. This loop iterates over a std::map of 2150 // pointers, but is deterministic because the pointers refer to Elf_Verdef 2151 // data structures within a single input file. 2152 for (auto &NV : P.first->VerdefMap) { 2153 Vernaux->vna_hash = NV.first->vd_hash; 2154 Vernaux->vna_flags = 0; 2155 Vernaux->vna_other = NV.second.Index; 2156 Vernaux->vna_name = NV.second.StrTab; 2157 Vernaux->vna_next = sizeof(Elf_Vernaux); 2158 ++Vernaux; 2159 } 2160 2161 Vernaux[-1].vna_next = 0; 2162 } 2163 Verneed[-1].vn_next = 0; 2164 } 2165 2166 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() { 2167 getParent()->Link = InX::DynStrTab->getParent()->SectionIndex; 2168 getParent()->Info = Needed.size(); 2169 } 2170 2171 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const { 2172 unsigned Size = Needed.size() * sizeof(Elf_Verneed); 2173 for (const std::pair<SharedFile<ELFT> *, size_t> &P : Needed) 2174 Size += P.first->VerdefMap.size() * sizeof(Elf_Vernaux); 2175 return Size; 2176 } 2177 2178 template <class ELFT> bool VersionNeedSection<ELFT>::empty() const { 2179 return getNeedNum() == 0; 2180 } 2181 2182 void MergeSyntheticSection::addSection(MergeInputSection *MS) { 2183 MS->Parent = this; 2184 Sections.push_back(MS); 2185 } 2186 2187 MergeTailSection::MergeTailSection(StringRef Name, uint32_t Type, 2188 uint64_t Flags, uint32_t Alignment) 2189 : MergeSyntheticSection(Name, Type, Flags, Alignment), 2190 Builder(StringTableBuilder::RAW, Alignment) {} 2191 2192 size_t MergeTailSection::getSize() const { return Builder.getSize(); } 2193 2194 void MergeTailSection::writeTo(uint8_t *Buf) { Builder.write(Buf); } 2195 2196 void MergeTailSection::finalizeContents() { 2197 // Add all string pieces to the string table builder to create section 2198 // contents. 2199 for (MergeInputSection *Sec : Sections) 2200 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2201 if (Sec->Pieces[I].Live) 2202 Builder.add(Sec->getData(I)); 2203 2204 // Fix the string table content. After this, the contents will never change. 2205 Builder.finalize(); 2206 2207 // finalize() fixed tail-optimized strings, so we can now get 2208 // offsets of strings. Get an offset for each string and save it 2209 // to a corresponding StringPiece for easy access. 2210 for (MergeInputSection *Sec : Sections) 2211 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2212 if (Sec->Pieces[I].Live) 2213 Sec->Pieces[I].OutputOff = Builder.getOffset(Sec->getData(I)); 2214 } 2215 2216 void MergeNoTailSection::writeTo(uint8_t *Buf) { 2217 for (size_t I = 0; I < NumShards; ++I) 2218 Shards[I].write(Buf + ShardOffsets[I]); 2219 } 2220 2221 // This function is very hot (i.e. it can take several seconds to finish) 2222 // because sometimes the number of inputs is in an order of magnitude of 2223 // millions. So, we use multi-threading. 2224 // 2225 // For any strings S and T, we know S is not mergeable with T if S's hash 2226 // value is different from T's. If that's the case, we can safely put S and 2227 // T into different string builders without worrying about merge misses. 2228 // We do it in parallel. 2229 void MergeNoTailSection::finalizeContents() { 2230 // Initializes string table builders. 2231 for (size_t I = 0; I < NumShards; ++I) 2232 Shards.emplace_back(StringTableBuilder::RAW, Alignment); 2233 2234 // Concurrency level. Must be a power of 2 to avoid expensive modulo 2235 // operations in the following tight loop. 2236 size_t Concurrency = 1; 2237 if (Config->Threads) 2238 if (int N = std::thread::hardware_concurrency()) 2239 Concurrency = std::min<size_t>(PowerOf2Floor(N), NumShards); 2240 2241 // Add section pieces to the builders. 2242 parallelForEachN(0, Concurrency, [&](size_t ThreadId) { 2243 for (MergeInputSection *Sec : Sections) { 2244 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) { 2245 if (!Sec->Pieces[I].Live) 2246 continue; 2247 CachedHashStringRef Str = Sec->getData(I); 2248 size_t ShardId = getShardId(Str.hash()); 2249 if ((ShardId & (Concurrency - 1)) == ThreadId) 2250 Sec->Pieces[I].OutputOff = Shards[ShardId].add(Str); 2251 } 2252 } 2253 }); 2254 2255 // Compute an in-section offset for each shard. 2256 size_t Off = 0; 2257 for (size_t I = 0; I < NumShards; ++I) { 2258 Shards[I].finalizeInOrder(); 2259 if (Shards[I].getSize() > 0) 2260 Off = alignTo(Off, Alignment); 2261 ShardOffsets[I] = Off; 2262 Off += Shards[I].getSize(); 2263 } 2264 Size = Off; 2265 2266 // So far, section pieces have offsets from beginning of shards, but 2267 // we want offsets from beginning of the whole section. Fix them. 2268 parallelForEach(Sections, [&](MergeInputSection *Sec) { 2269 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2270 if (Sec->Pieces[I].Live) 2271 Sec->Pieces[I].OutputOff += 2272 ShardOffsets[getShardId(Sec->getData(I).hash())]; 2273 }); 2274 } 2275 2276 static MergeSyntheticSection *createMergeSynthetic(StringRef Name, 2277 uint32_t Type, 2278 uint64_t Flags, 2279 uint32_t Alignment) { 2280 bool ShouldTailMerge = (Flags & SHF_STRINGS) && Config->Optimize >= 2; 2281 if (ShouldTailMerge) 2282 return make<MergeTailSection>(Name, Type, Flags, Alignment); 2283 return make<MergeNoTailSection>(Name, Type, Flags, Alignment); 2284 } 2285 2286 // This function decompresses compressed sections and scans over the input 2287 // sections to create mergeable synthetic sections. It removes 2288 // MergeInputSections from the input section array and adds new synthetic 2289 // sections at the location of the first input section that it replaces. It then 2290 // finalizes each synthetic section in order to compute an output offset for 2291 // each piece of each input section. 2292 void elf::decompressAndMergeSections() { 2293 // splitIntoPieces needs to be called on each MergeInputSection before calling 2294 // finalizeContents(). Do that first. 2295 parallelForEach(InputSections, [](InputSectionBase *S) { 2296 if (!S->Live) 2297 return; 2298 S->maybeUncompress(); 2299 if (auto *MS = dyn_cast<MergeInputSection>(S)) 2300 MS->splitIntoPieces(); 2301 }); 2302 2303 std::vector<MergeSyntheticSection *> MergeSections; 2304 for (InputSectionBase *&S : InputSections) { 2305 MergeInputSection *MS = dyn_cast<MergeInputSection>(S); 2306 if (!MS) 2307 continue; 2308 2309 // We do not want to handle sections that are not alive, so just remove 2310 // them instead of trying to merge. 2311 if (!MS->Live) 2312 continue; 2313 2314 StringRef OutsecName = getOutputSectionName(MS->Name); 2315 uint32_t Alignment = std::max<uint32_t>(MS->Alignment, MS->Entsize); 2316 2317 auto I = llvm::find_if(MergeSections, [=](MergeSyntheticSection *Sec) { 2318 return Sec->Name == OutsecName && Sec->Flags == MS->Flags && 2319 Sec->Alignment == Alignment; 2320 }); 2321 if (I == MergeSections.end()) { 2322 MergeSyntheticSection *Syn = 2323 createMergeSynthetic(OutsecName, MS->Type, MS->Flags, Alignment); 2324 MergeSections.push_back(Syn); 2325 I = std::prev(MergeSections.end()); 2326 S = Syn; 2327 } else { 2328 S = nullptr; 2329 } 2330 (*I)->addSection(MS); 2331 } 2332 for (auto *MS : MergeSections) 2333 MS->finalizeContents(); 2334 2335 std::vector<InputSectionBase *> &V = InputSections; 2336 V.erase(std::remove(V.begin(), V.end(), nullptr), V.end()); 2337 } 2338 2339 MipsRldMapSection::MipsRldMapSection() 2340 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, Config->Wordsize, 2341 ".rld_map") {} 2342 2343 ARMExidxSentinelSection::ARMExidxSentinelSection() 2344 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX, 2345 Config->Wordsize, ".ARM.exidx") {} 2346 2347 // Write a terminating sentinel entry to the end of the .ARM.exidx table. 2348 // This section will have been sorted last in the .ARM.exidx table. 2349 // This table entry will have the form: 2350 // | PREL31 upper bound of code that has exception tables | EXIDX_CANTUNWIND | 2351 // The sentinel must have the PREL31 value of an address higher than any 2352 // address described by any other table entry. 2353 void ARMExidxSentinelSection::writeTo(uint8_t *Buf) { 2354 // The Sections are sorted in order of ascending PREL31 address with the 2355 // sentinel last. We need to find the InputSection that precedes the 2356 // sentinel. By construction the Sentinel is in the last 2357 // InputSectionDescription as the InputSection that precedes it. 2358 OutputSection *C = getParent(); 2359 auto ISD = std::find_if(C->Commands.rbegin(), C->Commands.rend(), 2360 [](const BaseCommand *Base) { 2361 return isa<InputSectionDescription>(Base); 2362 }); 2363 auto L = cast<InputSectionDescription>(*ISD); 2364 InputSection *Highest = L->Sections[L->Sections.size() - 2]; 2365 InputSection *LS = Highest->getLinkOrderDep(); 2366 uint64_t S = LS->getParent()->Addr + LS->getOffset(LS->getSize()); 2367 uint64_t P = getVA(); 2368 Target->relocateOne(Buf, R_ARM_PREL31, S - P); 2369 write32le(Buf + 4, 0x1); 2370 } 2371 2372 ThunkSection::ThunkSection(OutputSection *OS, uint64_t Off) 2373 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 2374 Config->Wordsize, ".text.thunk") { 2375 this->Parent = OS; 2376 this->OutSecOff = Off; 2377 } 2378 2379 void ThunkSection::addThunk(Thunk *T) { 2380 uint64_t Off = alignTo(Size, T->Alignment); 2381 T->Offset = Off; 2382 Thunks.push_back(T); 2383 T->addSymbols(*this); 2384 Size = Off + T->size(); 2385 } 2386 2387 void ThunkSection::writeTo(uint8_t *Buf) { 2388 for (const Thunk *T : Thunks) 2389 T->writeTo(Buf + T->Offset, *this); 2390 } 2391 2392 InputSection *ThunkSection::getTargetInputSection() const { 2393 const Thunk *T = Thunks.front(); 2394 return T->getTargetInputSection(); 2395 } 2396 2397 InputSection *InX::ARMAttributes; 2398 BssSection *InX::Bss; 2399 BssSection *InX::BssRelRo; 2400 BuildIdSection *InX::BuildId; 2401 SyntheticSection *InX::Dynamic; 2402 StringTableSection *InX::DynStrTab; 2403 SymbolTableBaseSection *InX::DynSymTab; 2404 InputSection *InX::Interp; 2405 GdbIndexSection *InX::GdbIndex; 2406 GotSection *InX::Got; 2407 GotPltSection *InX::GotPlt; 2408 GnuHashTableSection *InX::GnuHashTab; 2409 HashTableSection *InX::HashTab; 2410 IgotPltSection *InX::IgotPlt; 2411 MipsGotSection *InX::MipsGot; 2412 MipsRldMapSection *InX::MipsRldMap; 2413 PltSection *InX::Plt; 2414 PltSection *InX::Iplt; 2415 StringTableSection *InX::ShStrTab; 2416 StringTableSection *InX::StrTab; 2417 SymbolTableBaseSection *InX::SymTab; 2418 2419 template GdbIndexSection *elf::createGdbIndex<ELF32LE>(); 2420 template GdbIndexSection *elf::createGdbIndex<ELF32BE>(); 2421 template GdbIndexSection *elf::createGdbIndex<ELF64LE>(); 2422 template GdbIndexSection *elf::createGdbIndex<ELF64BE>(); 2423 2424 template void PltSection::addEntry<ELF32LE>(SymbolBody &Sym); 2425 template void PltSection::addEntry<ELF32BE>(SymbolBody &Sym); 2426 template void PltSection::addEntry<ELF64LE>(SymbolBody &Sym); 2427 template void PltSection::addEntry<ELF64BE>(SymbolBody &Sym); 2428 2429 template void elf::createCommonSections<ELF32LE>(); 2430 template void elf::createCommonSections<ELF32BE>(); 2431 template void elf::createCommonSections<ELF64LE>(); 2432 template void elf::createCommonSections<ELF64BE>(); 2433 2434 template MergeInputSection *elf::createCommentSection<ELF32LE>(); 2435 template MergeInputSection *elf::createCommentSection<ELF32BE>(); 2436 template MergeInputSection *elf::createCommentSection<ELF64LE>(); 2437 template MergeInputSection *elf::createCommentSection<ELF64BE>(); 2438 2439 template class elf::MipsAbiFlagsSection<ELF32LE>; 2440 template class elf::MipsAbiFlagsSection<ELF32BE>; 2441 template class elf::MipsAbiFlagsSection<ELF64LE>; 2442 template class elf::MipsAbiFlagsSection<ELF64BE>; 2443 2444 template class elf::MipsOptionsSection<ELF32LE>; 2445 template class elf::MipsOptionsSection<ELF32BE>; 2446 template class elf::MipsOptionsSection<ELF64LE>; 2447 template class elf::MipsOptionsSection<ELF64BE>; 2448 2449 template class elf::MipsReginfoSection<ELF32LE>; 2450 template class elf::MipsReginfoSection<ELF32BE>; 2451 template class elf::MipsReginfoSection<ELF64LE>; 2452 template class elf::MipsReginfoSection<ELF64BE>; 2453 2454 template class elf::DynamicSection<ELF32LE>; 2455 template class elf::DynamicSection<ELF32BE>; 2456 template class elf::DynamicSection<ELF64LE>; 2457 template class elf::DynamicSection<ELF64BE>; 2458 2459 template class elf::RelocationSection<ELF32LE>; 2460 template class elf::RelocationSection<ELF32BE>; 2461 template class elf::RelocationSection<ELF64LE>; 2462 template class elf::RelocationSection<ELF64BE>; 2463 2464 template class elf::SymbolTableSection<ELF32LE>; 2465 template class elf::SymbolTableSection<ELF32BE>; 2466 template class elf::SymbolTableSection<ELF64LE>; 2467 template class elf::SymbolTableSection<ELF64BE>; 2468 2469 template class elf::EhFrameHeader<ELF32LE>; 2470 template class elf::EhFrameHeader<ELF32BE>; 2471 template class elf::EhFrameHeader<ELF64LE>; 2472 template class elf::EhFrameHeader<ELF64BE>; 2473 2474 template class elf::VersionTableSection<ELF32LE>; 2475 template class elf::VersionTableSection<ELF32BE>; 2476 template class elf::VersionTableSection<ELF64LE>; 2477 template class elf::VersionTableSection<ELF64BE>; 2478 2479 template class elf::VersionNeedSection<ELF32LE>; 2480 template class elf::VersionNeedSection<ELF32BE>; 2481 template class elf::VersionNeedSection<ELF64LE>; 2482 template class elf::VersionNeedSection<ELF64BE>; 2483 2484 template class elf::VersionDefinitionSection<ELF32LE>; 2485 template class elf::VersionDefinitionSection<ELF32BE>; 2486 template class elf::VersionDefinitionSection<ELF64LE>; 2487 template class elf::VersionDefinitionSection<ELF64BE>; 2488 2489 template class elf::EhFrameSection<ELF32LE>; 2490 template class elf::EhFrameSection<ELF32BE>; 2491 template class elf::EhFrameSection<ELF64LE>; 2492 template class elf::EhFrameSection<ELF64BE>; 2493