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 symbol 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->Alignment = std::max(Sec->Alignment, 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; 497 498 Alignment = std::max(Alignment, Sec->Alignment); 499 Sections.push_back(Sec); 500 501 for (auto *DS : Sec->DependentSections) 502 DependentSections.push_back(DS); 503 504 // .eh_frame is a sequence of CIE or FDE records. This function 505 // splits it into pieces so that we can call 506 // SplitInputSection::getSectionPiece on the section. 507 Sec->split<ELFT>(); 508 if (Sec->Pieces.empty()) 509 return; 510 511 if (Sec->NumRelocations == 0) 512 addSectionAux(Sec, makeArrayRef<Elf_Rela>(nullptr, nullptr)); 513 else if (Sec->AreRelocsRela) 514 addSectionAux(Sec, Sec->template relas<ELFT>()); 515 else 516 addSectionAux(Sec, Sec->template rels<ELFT>()); 517 } 518 519 template <class ELFT> 520 static void writeCieFde(uint8_t *Buf, ArrayRef<uint8_t> D) { 521 memcpy(Buf, D.data(), D.size()); 522 523 size_t Aligned = alignTo(D.size(), sizeof(typename ELFT::uint)); 524 525 // Zero-clear trailing padding if it exists. 526 memset(Buf + D.size(), 0, Aligned - D.size()); 527 528 // Fix the size field. -4 since size does not include the size field itself. 529 const endianness E = ELFT::TargetEndianness; 530 write32<E>(Buf, Aligned - 4); 531 } 532 533 template <class ELFT> void EhFrameSection<ELFT>::finalizeContents() { 534 if (this->Size) 535 return; // Already finalized. 536 537 size_t Off = 0; 538 for (CieRecord *Rec : CieRecords) { 539 Rec->Cie->OutputOff = Off; 540 Off += alignTo(Rec->Cie->Size, Config->Wordsize); 541 542 for (EhSectionPiece *Fde : Rec->Fdes) { 543 Fde->OutputOff = Off; 544 Off += alignTo(Fde->Size, Config->Wordsize); 545 } 546 } 547 548 // The LSB standard does not allow a .eh_frame section with zero 549 // Call Frame Information records. Therefore add a CIE record length 550 // 0 as a terminator if this .eh_frame section is empty. 551 if (Off == 0) 552 Off = 4; 553 554 this->Size = Off; 555 } 556 557 template <class ELFT> static uint64_t readFdeAddr(uint8_t *Buf, int Size) { 558 const endianness E = ELFT::TargetEndianness; 559 switch (Size) { 560 case DW_EH_PE_udata2: 561 return read16<E>(Buf); 562 case DW_EH_PE_udata4: 563 return read32<E>(Buf); 564 case DW_EH_PE_udata8: 565 return read64<E>(Buf); 566 case DW_EH_PE_absptr: 567 if (ELFT::Is64Bits) 568 return read64<E>(Buf); 569 return read32<E>(Buf); 570 } 571 fatal("unknown FDE size encoding"); 572 } 573 574 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to. 575 // We need it to create .eh_frame_hdr section. 576 template <class ELFT> 577 uint64_t EhFrameSection<ELFT>::getFdePc(uint8_t *Buf, size_t FdeOff, 578 uint8_t Enc) { 579 // The starting address to which this FDE applies is 580 // stored at FDE + 8 byte. 581 size_t Off = FdeOff + 8; 582 uint64_t Addr = readFdeAddr<ELFT>(Buf + Off, Enc & 0x7); 583 if ((Enc & 0x70) == DW_EH_PE_absptr) 584 return Addr; 585 if ((Enc & 0x70) == DW_EH_PE_pcrel) 586 return Addr + getParent()->Addr + Off; 587 fatal("unknown FDE size relative encoding"); 588 } 589 590 template <class ELFT> void EhFrameSection<ELFT>::writeTo(uint8_t *Buf) { 591 const endianness E = ELFT::TargetEndianness; 592 for (CieRecord *Rec : CieRecords) { 593 size_t CieOffset = Rec->Cie->OutputOff; 594 writeCieFde<ELFT>(Buf + CieOffset, Rec->Cie->data()); 595 596 for (EhSectionPiece *Fde : Rec->Fdes) { 597 size_t Off = Fde->OutputOff; 598 writeCieFde<ELFT>(Buf + Off, Fde->data()); 599 600 // FDE's second word should have the offset to an associated CIE. 601 // Write it. 602 write32<E>(Buf + Off + 4, Off + 4 - CieOffset); 603 } 604 } 605 606 for (EhInputSection *S : Sections) 607 S->relocateAlloc(Buf, nullptr); 608 609 // Construct .eh_frame_hdr. .eh_frame_hdr is a binary search table 610 // to get a FDE from an address to which FDE is applied. So here 611 // we obtain two addresses and pass them to EhFrameHdr object. 612 if (In<ELFT>::EhFrameHdr) { 613 for (CieRecord *Rec : CieRecords) { 614 uint8_t Enc = getFdeEncoding<ELFT>(Rec->Cie); 615 for (EhSectionPiece *Fde : Rec->Fdes) { 616 uint64_t Pc = getFdePc(Buf, Fde->OutputOff, Enc); 617 uint64_t FdeVA = getParent()->Addr + Fde->OutputOff; 618 In<ELFT>::EhFrameHdr->addFde(Pc, FdeVA); 619 } 620 } 621 } 622 } 623 624 GotSection::GotSection() 625 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 626 Target->GotEntrySize, ".got") {} 627 628 void GotSection::addEntry(SymbolBody &Sym) { 629 Sym.GotIndex = NumEntries; 630 ++NumEntries; 631 } 632 633 bool GotSection::addDynTlsEntry(SymbolBody &Sym) { 634 if (Sym.GlobalDynIndex != -1U) 635 return false; 636 Sym.GlobalDynIndex = NumEntries; 637 // Global Dynamic TLS entries take two GOT slots. 638 NumEntries += 2; 639 return true; 640 } 641 642 // Reserves TLS entries for a TLS module ID and a TLS block offset. 643 // In total it takes two GOT slots. 644 bool GotSection::addTlsIndex() { 645 if (TlsIndexOff != uint32_t(-1)) 646 return false; 647 TlsIndexOff = NumEntries * Config->Wordsize; 648 NumEntries += 2; 649 return true; 650 } 651 652 uint64_t GotSection::getGlobalDynAddr(const SymbolBody &B) const { 653 return this->getVA() + B.GlobalDynIndex * Config->Wordsize; 654 } 655 656 uint64_t GotSection::getGlobalDynOffset(const SymbolBody &B) const { 657 return B.GlobalDynIndex * Config->Wordsize; 658 } 659 660 void GotSection::finalizeContents() { Size = NumEntries * Config->Wordsize; } 661 662 bool GotSection::empty() const { 663 // We need to emit a GOT even if it's empty if there's a relocation that is 664 // relative to GOT(such as GOTOFFREL) or there's a symbol that points to a GOT 665 // (i.e. _GLOBAL_OFFSET_TABLE_). 666 return NumEntries == 0 && !HasGotOffRel && !ElfSym::GlobalOffsetTable; 667 } 668 669 void GotSection::writeTo(uint8_t *Buf) { 670 // Buf points to the start of this section's buffer, 671 // whereas InputSectionBase::relocateAlloc() expects its argument 672 // to point to the start of the output section. 673 relocateAlloc(Buf - OutSecOff, Buf - OutSecOff + Size); 674 } 675 676 MipsGotSection::MipsGotSection() 677 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16, 678 ".got") {} 679 680 void MipsGotSection::addEntry(SymbolBody &Sym, int64_t Addend, RelExpr Expr) { 681 // For "true" local symbols which can be referenced from the same module 682 // only compiler creates two instructions for address loading: 683 // 684 // lw $8, 0($gp) # R_MIPS_GOT16 685 // addi $8, $8, 0 # R_MIPS_LO16 686 // 687 // The first instruction loads high 16 bits of the symbol address while 688 // the second adds an offset. That allows to reduce number of required 689 // GOT entries because only one global offset table entry is necessary 690 // for every 64 KBytes of local data. So for local symbols we need to 691 // allocate number of GOT entries to hold all required "page" addresses. 692 // 693 // All global symbols (hidden and regular) considered by compiler uniformly. 694 // It always generates a single `lw` instruction and R_MIPS_GOT16 relocation 695 // to load address of the symbol. So for each such symbol we need to 696 // allocate dedicated GOT entry to store its address. 697 // 698 // If a symbol is preemptible we need help of dynamic linker to get its 699 // final address. The corresponding GOT entries are allocated in the 700 // "global" part of GOT. Entries for non preemptible global symbol allocated 701 // in the "local" part of GOT. 702 // 703 // See "Global Offset Table" in Chapter 5: 704 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 705 if (Expr == R_MIPS_GOT_LOCAL_PAGE) { 706 // At this point we do not know final symbol value so to reduce number 707 // of allocated GOT entries do the following trick. Save all output 708 // sections referenced by GOT relocations. Then later in the `finalize` 709 // method calculate number of "pages" required to cover all saved output 710 // section and allocate appropriate number of GOT entries. 711 PageIndexMap.insert({Sym.getOutputSection(), 0}); 712 return; 713 } 714 if (Sym.isTls()) { 715 // GOT entries created for MIPS TLS relocations behave like 716 // almost GOT entries from other ABIs. They go to the end 717 // of the global offset table. 718 Sym.GotIndex = TlsEntries.size(); 719 TlsEntries.push_back(&Sym); 720 return; 721 } 722 auto AddEntry = [&](SymbolBody &S, uint64_t A, GotEntries &Items) { 723 if (S.isInGot() && !A) 724 return; 725 size_t NewIndex = Items.size(); 726 if (!EntryIndexMap.insert({{&S, A}, NewIndex}).second) 727 return; 728 Items.emplace_back(&S, A); 729 if (!A) 730 S.GotIndex = NewIndex; 731 }; 732 if (Sym.isPreemptible()) { 733 // Ignore addends for preemptible symbols. They got single GOT entry anyway. 734 AddEntry(Sym, 0, GlobalEntries); 735 Sym.IsInGlobalMipsGot = true; 736 } else if (Expr == R_MIPS_GOT_OFF32) { 737 AddEntry(Sym, Addend, LocalEntries32); 738 Sym.Is32BitMipsGot = true; 739 } else { 740 // Hold local GOT entries accessed via a 16-bit index separately. 741 // That allows to write them in the beginning of the GOT and keep 742 // their indexes as less as possible to escape relocation's overflow. 743 AddEntry(Sym, Addend, LocalEntries); 744 } 745 } 746 747 bool MipsGotSection::addDynTlsEntry(SymbolBody &Sym) { 748 if (Sym.GlobalDynIndex != -1U) 749 return false; 750 Sym.GlobalDynIndex = TlsEntries.size(); 751 // Global Dynamic TLS entries take two GOT slots. 752 TlsEntries.push_back(nullptr); 753 TlsEntries.push_back(&Sym); 754 return true; 755 } 756 757 // Reserves TLS entries for a TLS module ID and a TLS block offset. 758 // In total it takes two GOT slots. 759 bool MipsGotSection::addTlsIndex() { 760 if (TlsIndexOff != uint32_t(-1)) 761 return false; 762 TlsIndexOff = TlsEntries.size() * Config->Wordsize; 763 TlsEntries.push_back(nullptr); 764 TlsEntries.push_back(nullptr); 765 return true; 766 } 767 768 static uint64_t getMipsPageAddr(uint64_t Addr) { 769 return (Addr + 0x8000) & ~0xffff; 770 } 771 772 static uint64_t getMipsPageCount(uint64_t Size) { 773 return (Size + 0xfffe) / 0xffff + 1; 774 } 775 776 uint64_t MipsGotSection::getPageEntryOffset(const SymbolBody &B, 777 int64_t Addend) const { 778 const OutputSection *OutSec = B.getOutputSection(); 779 uint64_t SecAddr = getMipsPageAddr(OutSec->Addr); 780 uint64_t SymAddr = getMipsPageAddr(B.getVA(Addend)); 781 uint64_t Index = PageIndexMap.lookup(OutSec) + (SymAddr - SecAddr) / 0xffff; 782 assert(Index < PageEntriesNum); 783 return (HeaderEntriesNum + Index) * Config->Wordsize; 784 } 785 786 uint64_t MipsGotSection::getBodyEntryOffset(const SymbolBody &B, 787 int64_t Addend) const { 788 // Calculate offset of the GOT entries block: TLS, global, local. 789 uint64_t Index = HeaderEntriesNum + PageEntriesNum; 790 if (B.isTls()) 791 Index += LocalEntries.size() + LocalEntries32.size() + GlobalEntries.size(); 792 else if (B.IsInGlobalMipsGot) 793 Index += LocalEntries.size() + LocalEntries32.size(); 794 else if (B.Is32BitMipsGot) 795 Index += LocalEntries.size(); 796 // Calculate offset of the GOT entry in the block. 797 if (B.isInGot()) 798 Index += B.GotIndex; 799 else { 800 auto It = EntryIndexMap.find({&B, Addend}); 801 assert(It != EntryIndexMap.end()); 802 Index += It->second; 803 } 804 return Index * Config->Wordsize; 805 } 806 807 uint64_t MipsGotSection::getTlsOffset() const { 808 return (getLocalEntriesNum() + GlobalEntries.size()) * Config->Wordsize; 809 } 810 811 uint64_t MipsGotSection::getGlobalDynOffset(const SymbolBody &B) const { 812 return B.GlobalDynIndex * Config->Wordsize; 813 } 814 815 const SymbolBody *MipsGotSection::getFirstGlobalEntry() const { 816 return GlobalEntries.empty() ? nullptr : GlobalEntries.front().first; 817 } 818 819 unsigned MipsGotSection::getLocalEntriesNum() const { 820 return HeaderEntriesNum + PageEntriesNum + LocalEntries.size() + 821 LocalEntries32.size(); 822 } 823 824 void MipsGotSection::finalizeContents() { updateAllocSize(); } 825 826 void MipsGotSection::updateAllocSize() { 827 PageEntriesNum = 0; 828 for (std::pair<const OutputSection *, size_t> &P : PageIndexMap) { 829 // For each output section referenced by GOT page relocations calculate 830 // and save into PageIndexMap an upper bound of MIPS GOT entries required 831 // to store page addresses of local symbols. We assume the worst case - 832 // each 64kb page of the output section has at least one GOT relocation 833 // against it. And take in account the case when the section intersects 834 // page boundaries. 835 P.second = PageEntriesNum; 836 PageEntriesNum += getMipsPageCount(P.first->Size); 837 } 838 Size = (getLocalEntriesNum() + GlobalEntries.size() + TlsEntries.size()) * 839 Config->Wordsize; 840 } 841 842 bool MipsGotSection::empty() const { 843 // We add the .got section to the result for dynamic MIPS target because 844 // its address and properties are mentioned in the .dynamic section. 845 return Config->Relocatable; 846 } 847 848 uint64_t MipsGotSection::getGp() const { return ElfSym::MipsGp->getVA(0); } 849 850 static uint64_t readUint(uint8_t *Buf) { 851 if (Config->Is64) 852 return read64(Buf, Config->Endianness); 853 return read32(Buf, Config->Endianness); 854 } 855 856 static void writeUint(uint8_t *Buf, uint64_t Val) { 857 if (Config->Is64) 858 write64(Buf, Val, Config->Endianness); 859 else 860 write32(Buf, Val, Config->Endianness); 861 } 862 863 void MipsGotSection::writeTo(uint8_t *Buf) { 864 // Set the MSB of the second GOT slot. This is not required by any 865 // MIPS ABI documentation, though. 866 // 867 // There is a comment in glibc saying that "The MSB of got[1] of a 868 // gnu object is set to identify gnu objects," and in GNU gold it 869 // says "the second entry will be used by some runtime loaders". 870 // But how this field is being used is unclear. 871 // 872 // We are not really willing to mimic other linkers behaviors 873 // without understanding why they do that, but because all files 874 // generated by GNU tools have this special GOT value, and because 875 // we've been doing this for years, it is probably a safe bet to 876 // keep doing this for now. We really need to revisit this to see 877 // if we had to do this. 878 writeUint(Buf + Config->Wordsize, (uint64_t)1 << (Config->Wordsize * 8 - 1)); 879 Buf += HeaderEntriesNum * Config->Wordsize; 880 // Write 'page address' entries to the local part of the GOT. 881 for (std::pair<const OutputSection *, size_t> &L : PageIndexMap) { 882 size_t PageCount = getMipsPageCount(L.first->Size); 883 uint64_t FirstPageAddr = getMipsPageAddr(L.first->Addr); 884 for (size_t PI = 0; PI < PageCount; ++PI) { 885 uint8_t *Entry = Buf + (L.second + PI) * Config->Wordsize; 886 writeUint(Entry, FirstPageAddr + PI * 0x10000); 887 } 888 } 889 Buf += PageEntriesNum * Config->Wordsize; 890 auto AddEntry = [&](const GotEntry &SA) { 891 uint8_t *Entry = Buf; 892 Buf += Config->Wordsize; 893 const SymbolBody *Body = SA.first; 894 uint64_t VA = Body->getVA(SA.second); 895 writeUint(Entry, VA); 896 }; 897 std::for_each(std::begin(LocalEntries), std::end(LocalEntries), AddEntry); 898 std::for_each(std::begin(LocalEntries32), std::end(LocalEntries32), AddEntry); 899 std::for_each(std::begin(GlobalEntries), std::end(GlobalEntries), AddEntry); 900 // Initialize TLS-related GOT entries. If the entry has a corresponding 901 // dynamic relocations, leave it initialized by zero. Write down adjusted 902 // TLS symbol's values otherwise. To calculate the adjustments use offsets 903 // for thread-local storage. 904 // https://www.linux-mips.org/wiki/NPTL 905 if (TlsIndexOff != -1U && !Config->Pic) 906 writeUint(Buf + TlsIndexOff, 1); 907 for (const SymbolBody *B : TlsEntries) { 908 if (!B || B->isPreemptible()) 909 continue; 910 uint64_t VA = B->getVA(); 911 if (B->GotIndex != -1U) { 912 uint8_t *Entry = Buf + B->GotIndex * Config->Wordsize; 913 writeUint(Entry, VA - 0x7000); 914 } 915 if (B->GlobalDynIndex != -1U) { 916 uint8_t *Entry = Buf + B->GlobalDynIndex * Config->Wordsize; 917 writeUint(Entry, 1); 918 Entry += Config->Wordsize; 919 writeUint(Entry, VA - 0x8000); 920 } 921 } 922 } 923 924 GotPltSection::GotPltSection() 925 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 926 Target->GotPltEntrySize, ".got.plt") {} 927 928 void GotPltSection::addEntry(SymbolBody &Sym) { 929 Sym.GotPltIndex = Target->GotPltHeaderEntriesNum + Entries.size(); 930 Entries.push_back(&Sym); 931 } 932 933 size_t GotPltSection::getSize() const { 934 return (Target->GotPltHeaderEntriesNum + Entries.size()) * 935 Target->GotPltEntrySize; 936 } 937 938 void GotPltSection::writeTo(uint8_t *Buf) { 939 Target->writeGotPltHeader(Buf); 940 Buf += Target->GotPltHeaderEntriesNum * Target->GotPltEntrySize; 941 for (const SymbolBody *B : Entries) { 942 Target->writeGotPlt(Buf, *B); 943 Buf += Config->Wordsize; 944 } 945 } 946 947 // On ARM the IgotPltSection is part of the GotSection, on other Targets it is 948 // part of the .got.plt 949 IgotPltSection::IgotPltSection() 950 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 951 Target->GotPltEntrySize, 952 Config->EMachine == EM_ARM ? ".got" : ".got.plt") {} 953 954 void IgotPltSection::addEntry(SymbolBody &Sym) { 955 Sym.IsInIgot = true; 956 Sym.GotPltIndex = Entries.size(); 957 Entries.push_back(&Sym); 958 } 959 960 size_t IgotPltSection::getSize() const { 961 return Entries.size() * Target->GotPltEntrySize; 962 } 963 964 void IgotPltSection::writeTo(uint8_t *Buf) { 965 for (const SymbolBody *B : Entries) { 966 Target->writeIgotPlt(Buf, *B); 967 Buf += Config->Wordsize; 968 } 969 } 970 971 StringTableSection::StringTableSection(StringRef Name, bool Dynamic) 972 : SyntheticSection(Dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, Name), 973 Dynamic(Dynamic) { 974 // ELF string tables start with a NUL byte. 975 addString(""); 976 } 977 978 // Adds a string to the string table. If HashIt is true we hash and check for 979 // duplicates. It is optional because the name of global symbols are already 980 // uniqued and hashing them again has a big cost for a small value: uniquing 981 // them with some other string that happens to be the same. 982 unsigned StringTableSection::addString(StringRef S, bool HashIt) { 983 if (HashIt) { 984 auto R = StringMap.insert(std::make_pair(S, this->Size)); 985 if (!R.second) 986 return R.first->second; 987 } 988 unsigned Ret = this->Size; 989 this->Size = this->Size + S.size() + 1; 990 Strings.push_back(S); 991 return Ret; 992 } 993 994 void StringTableSection::writeTo(uint8_t *Buf) { 995 for (StringRef S : Strings) { 996 memcpy(Buf, S.data(), S.size()); 997 Buf[S.size()] = '\0'; 998 Buf += S.size() + 1; 999 } 1000 } 1001 1002 // Returns the number of version definition entries. Because the first entry 1003 // is for the version definition itself, it is the number of versioned symbols 1004 // plus one. Note that we don't support multiple versions yet. 1005 static unsigned getVerDefNum() { return Config->VersionDefinitions.size() + 1; } 1006 1007 template <class ELFT> 1008 DynamicSection<ELFT>::DynamicSection() 1009 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, Config->Wordsize, 1010 ".dynamic") { 1011 this->Entsize = ELFT::Is64Bits ? 16 : 8; 1012 1013 // .dynamic section is not writable on MIPS and on Fuchsia OS 1014 // which passes -z rodynamic. 1015 // See "Special Section" in Chapter 4 in the following document: 1016 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 1017 if (Config->EMachine == EM_MIPS || Config->ZRodynamic) 1018 this->Flags = SHF_ALLOC; 1019 1020 addEntries(); 1021 } 1022 1023 // There are some dynamic entries that don't depend on other sections. 1024 // Such entries can be set early. 1025 template <class ELFT> void DynamicSection<ELFT>::addEntries() { 1026 // Add strings to .dynstr early so that .dynstr's size will be 1027 // fixed early. 1028 for (StringRef S : Config->FilterList) 1029 add({DT_FILTER, InX::DynStrTab->addString(S)}); 1030 for (StringRef S : Config->AuxiliaryList) 1031 add({DT_AUXILIARY, InX::DynStrTab->addString(S)}); 1032 if (!Config->Rpath.empty()) 1033 add({Config->EnableNewDtags ? DT_RUNPATH : DT_RPATH, 1034 InX::DynStrTab->addString(Config->Rpath)}); 1035 for (InputFile *File : SharedFiles) { 1036 SharedFile<ELFT> *F = cast<SharedFile<ELFT>>(File); 1037 if (F->isNeeded()) 1038 add({DT_NEEDED, InX::DynStrTab->addString(F->SoName)}); 1039 } 1040 if (!Config->SoName.empty()) 1041 add({DT_SONAME, InX::DynStrTab->addString(Config->SoName)}); 1042 1043 // Set DT_FLAGS and DT_FLAGS_1. 1044 uint32_t DtFlags = 0; 1045 uint32_t DtFlags1 = 0; 1046 if (Config->Bsymbolic) 1047 DtFlags |= DF_SYMBOLIC; 1048 if (Config->ZNodelete) 1049 DtFlags1 |= DF_1_NODELETE; 1050 if (Config->ZNodlopen) 1051 DtFlags1 |= DF_1_NOOPEN; 1052 if (Config->ZNow) { 1053 DtFlags |= DF_BIND_NOW; 1054 DtFlags1 |= DF_1_NOW; 1055 } 1056 if (Config->ZOrigin) { 1057 DtFlags |= DF_ORIGIN; 1058 DtFlags1 |= DF_1_ORIGIN; 1059 } 1060 1061 if (DtFlags) 1062 add({DT_FLAGS, DtFlags}); 1063 if (DtFlags1) 1064 add({DT_FLAGS_1, DtFlags1}); 1065 1066 // DT_DEBUG is a pointer to debug informaion used by debuggers at runtime. We 1067 // need it for each process, so we don't write it for DSOs. The loader writes 1068 // the pointer into this entry. 1069 // 1070 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some 1071 // systems (currently only Fuchsia OS) provide other means to give the 1072 // debugger this information. Such systems may choose make .dynamic read-only. 1073 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG. 1074 if (!Config->Shared && !Config->Relocatable && !Config->ZRodynamic) 1075 add({DT_DEBUG, (uint64_t)0}); 1076 } 1077 1078 // Add remaining entries to complete .dynamic contents. 1079 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() { 1080 if (this->Size) 1081 return; // Already finalized. 1082 1083 this->Link = InX::DynStrTab->getParent()->SectionIndex; 1084 if (In<ELFT>::RelaDyn->getParent() && !In<ELFT>::RelaDyn->empty()) { 1085 bool IsRela = Config->IsRela; 1086 add({IsRela ? DT_RELA : DT_REL, In<ELFT>::RelaDyn}); 1087 add({IsRela ? DT_RELASZ : DT_RELSZ, In<ELFT>::RelaDyn->getParent(), 1088 Entry::SecSize}); 1089 add({IsRela ? DT_RELAENT : DT_RELENT, 1090 uint64_t(IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel))}); 1091 1092 // MIPS dynamic loader does not support RELCOUNT tag. 1093 // The problem is in the tight relation between dynamic 1094 // relocations and GOT. So do not emit this tag on MIPS. 1095 if (Config->EMachine != EM_MIPS) { 1096 size_t NumRelativeRels = In<ELFT>::RelaDyn->getRelativeRelocCount(); 1097 if (Config->ZCombreloc && NumRelativeRels) 1098 add({IsRela ? DT_RELACOUNT : DT_RELCOUNT, NumRelativeRels}); 1099 } 1100 } 1101 if (In<ELFT>::RelaPlt->getParent() && !In<ELFT>::RelaPlt->empty()) { 1102 add({DT_JMPREL, In<ELFT>::RelaPlt}); 1103 add({DT_PLTRELSZ, In<ELFT>::RelaPlt->getParent(), Entry::SecSize}); 1104 switch (Config->EMachine) { 1105 case EM_MIPS: 1106 add({DT_MIPS_PLTGOT, In<ELFT>::GotPlt}); 1107 break; 1108 case EM_SPARCV9: 1109 add({DT_PLTGOT, In<ELFT>::Plt}); 1110 break; 1111 default: 1112 add({DT_PLTGOT, In<ELFT>::GotPlt}); 1113 break; 1114 } 1115 add({DT_PLTREL, uint64_t(Config->IsRela ? DT_RELA : DT_REL)}); 1116 } 1117 1118 add({DT_SYMTAB, InX::DynSymTab}); 1119 add({DT_SYMENT, sizeof(Elf_Sym)}); 1120 add({DT_STRTAB, InX::DynStrTab}); 1121 add({DT_STRSZ, InX::DynStrTab->getSize()}); 1122 if (!Config->ZText) 1123 add({DT_TEXTREL, (uint64_t)0}); 1124 if (InX::GnuHashTab) 1125 add({DT_GNU_HASH, InX::GnuHashTab}); 1126 if (InX::HashTab) 1127 add({DT_HASH, InX::HashTab}); 1128 1129 if (Out::PreinitArray) { 1130 add({DT_PREINIT_ARRAY, Out::PreinitArray}); 1131 add({DT_PREINIT_ARRAYSZ, Out::PreinitArray, Entry::SecSize}); 1132 } 1133 if (Out::InitArray) { 1134 add({DT_INIT_ARRAY, Out::InitArray}); 1135 add({DT_INIT_ARRAYSZ, Out::InitArray, Entry::SecSize}); 1136 } 1137 if (Out::FiniArray) { 1138 add({DT_FINI_ARRAY, Out::FiniArray}); 1139 add({DT_FINI_ARRAYSZ, Out::FiniArray, Entry::SecSize}); 1140 } 1141 1142 if (SymbolBody *B = Symtab->find(Config->Init)) 1143 if (B->isInCurrentDSO()) 1144 add({DT_INIT, B}); 1145 if (SymbolBody *B = Symtab->find(Config->Fini)) 1146 if (B->isInCurrentDSO()) 1147 add({DT_FINI, B}); 1148 1149 bool HasVerNeed = In<ELFT>::VerNeed->getNeedNum() != 0; 1150 if (HasVerNeed || In<ELFT>::VerDef) 1151 add({DT_VERSYM, In<ELFT>::VerSym}); 1152 if (In<ELFT>::VerDef) { 1153 add({DT_VERDEF, In<ELFT>::VerDef}); 1154 add({DT_VERDEFNUM, getVerDefNum()}); 1155 } 1156 if (HasVerNeed) { 1157 add({DT_VERNEED, In<ELFT>::VerNeed}); 1158 add({DT_VERNEEDNUM, In<ELFT>::VerNeed->getNeedNum()}); 1159 } 1160 1161 if (Config->EMachine == EM_MIPS) { 1162 add({DT_MIPS_RLD_VERSION, 1}); 1163 add({DT_MIPS_FLAGS, RHF_NOTPOT}); 1164 add({DT_MIPS_BASE_ADDRESS, Config->ImageBase}); 1165 add({DT_MIPS_SYMTABNO, InX::DynSymTab->getNumSymbols()}); 1166 add({DT_MIPS_LOCAL_GOTNO, InX::MipsGot->getLocalEntriesNum()}); 1167 if (const SymbolBody *B = InX::MipsGot->getFirstGlobalEntry()) 1168 add({DT_MIPS_GOTSYM, B->DynsymIndex}); 1169 else 1170 add({DT_MIPS_GOTSYM, InX::DynSymTab->getNumSymbols()}); 1171 add({DT_PLTGOT, InX::MipsGot}); 1172 if (InX::MipsRldMap) 1173 add({DT_MIPS_RLD_MAP, InX::MipsRldMap}); 1174 } 1175 1176 add({DT_NULL, (uint64_t)0}); 1177 1178 getParent()->Link = this->Link; 1179 this->Size = Entries.size() * this->Entsize; 1180 } 1181 1182 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *Buf) { 1183 auto *P = reinterpret_cast<Elf_Dyn *>(Buf); 1184 1185 for (const Entry &E : Entries) { 1186 P->d_tag = E.Tag; 1187 switch (E.Kind) { 1188 case Entry::SecAddr: 1189 P->d_un.d_ptr = E.OutSec->Addr; 1190 break; 1191 case Entry::InSecAddr: 1192 P->d_un.d_ptr = E.InSec->getParent()->Addr + E.InSec->OutSecOff; 1193 break; 1194 case Entry::SecSize: 1195 P->d_un.d_val = E.OutSec->Size; 1196 break; 1197 case Entry::SymAddr: 1198 P->d_un.d_ptr = E.Sym->getVA(); 1199 break; 1200 case Entry::PlainInt: 1201 P->d_un.d_val = E.Val; 1202 break; 1203 } 1204 ++P; 1205 } 1206 } 1207 1208 uint64_t DynamicReloc::getOffset() const { 1209 return InputSec->getOutputSection()->Addr + InputSec->getOffset(OffsetInSec); 1210 } 1211 1212 int64_t DynamicReloc::getAddend() const { 1213 if (UseSymVA) 1214 return Sym->getVA(Addend); 1215 return Addend; 1216 } 1217 1218 uint32_t DynamicReloc::getSymIndex() const { 1219 if (Sym && !UseSymVA) 1220 return Sym->DynsymIndex; 1221 return 0; 1222 } 1223 1224 template <class ELFT> 1225 RelocationSection<ELFT>::RelocationSection(StringRef Name, bool Sort) 1226 : SyntheticSection(SHF_ALLOC, Config->IsRela ? SHT_RELA : SHT_REL, 1227 Config->Wordsize, Name), 1228 Sort(Sort) { 1229 this->Entsize = Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1230 } 1231 1232 template <class ELFT> 1233 void RelocationSection<ELFT>::addReloc(const DynamicReloc &Reloc) { 1234 if (Reloc.Type == Target->RelativeRel) 1235 ++NumRelativeRelocs; 1236 Relocs.push_back(Reloc); 1237 } 1238 1239 template <class ELFT, class RelTy> 1240 static bool compRelocations(const RelTy &A, const RelTy &B) { 1241 bool AIsRel = A.getType(Config->IsMips64EL) == Target->RelativeRel; 1242 bool BIsRel = B.getType(Config->IsMips64EL) == Target->RelativeRel; 1243 if (AIsRel != BIsRel) 1244 return AIsRel; 1245 1246 return A.getSymbol(Config->IsMips64EL) < B.getSymbol(Config->IsMips64EL); 1247 } 1248 1249 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *Buf) { 1250 uint8_t *BufBegin = Buf; 1251 for (const DynamicReloc &Rel : Relocs) { 1252 auto *P = reinterpret_cast<Elf_Rela *>(Buf); 1253 Buf += Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1254 1255 if (Config->IsRela) 1256 P->r_addend = Rel.getAddend(); 1257 P->r_offset = Rel.getOffset(); 1258 if (Config->EMachine == EM_MIPS && Rel.getInputSec() == InX::MipsGot) 1259 // Dynamic relocation against MIPS GOT section make deal TLS entries 1260 // allocated in the end of the GOT. We need to adjust the offset to take 1261 // in account 'local' and 'global' GOT entries. 1262 P->r_offset += InX::MipsGot->getTlsOffset(); 1263 P->setSymbolAndType(Rel.getSymIndex(), Rel.Type, Config->IsMips64EL); 1264 } 1265 1266 if (Sort) { 1267 if (Config->IsRela) 1268 std::stable_sort((Elf_Rela *)BufBegin, 1269 (Elf_Rela *)BufBegin + Relocs.size(), 1270 compRelocations<ELFT, Elf_Rela>); 1271 else 1272 std::stable_sort((Elf_Rel *)BufBegin, (Elf_Rel *)BufBegin + Relocs.size(), 1273 compRelocations<ELFT, Elf_Rel>); 1274 } 1275 } 1276 1277 template <class ELFT> unsigned RelocationSection<ELFT>::getRelocOffset() { 1278 return this->Entsize * Relocs.size(); 1279 } 1280 1281 template <class ELFT> void RelocationSection<ELFT>::finalizeContents() { 1282 this->Link = InX::DynSymTab ? InX::DynSymTab->getParent()->SectionIndex 1283 : InX::SymTab->getParent()->SectionIndex; 1284 1285 // Set required output section properties. 1286 getParent()->Link = this->Link; 1287 } 1288 1289 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &StrTabSec) 1290 : SyntheticSection(StrTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0, 1291 StrTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB, 1292 Config->Wordsize, 1293 StrTabSec.isDynamic() ? ".dynsym" : ".symtab"), 1294 StrTabSec(StrTabSec) {} 1295 1296 // Orders symbols according to their positions in the GOT, 1297 // in compliance with MIPS ABI rules. 1298 // See "Global Offset Table" in Chapter 5 in the following document 1299 // for detailed description: 1300 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 1301 static bool sortMipsSymbols(const SymbolTableEntry &L, 1302 const SymbolTableEntry &R) { 1303 // Sort entries related to non-local preemptible symbols by GOT indexes. 1304 // All other entries go to the first part of GOT in arbitrary order. 1305 bool LIsInLocalGot = !L.Symbol->IsInGlobalMipsGot; 1306 bool RIsInLocalGot = !R.Symbol->IsInGlobalMipsGot; 1307 if (LIsInLocalGot || RIsInLocalGot) 1308 return !RIsInLocalGot; 1309 return L.Symbol->GotIndex < R.Symbol->GotIndex; 1310 } 1311 1312 // Finalize a symbol table. The ELF spec requires that all local 1313 // symbols precede global symbols, so we sort symbol entries in this 1314 // function. (For .dynsym, we don't do that because symbols for 1315 // dynamic linking are inherently all globals.) 1316 void SymbolTableBaseSection::finalizeContents() { 1317 getParent()->Link = StrTabSec.getParent()->SectionIndex; 1318 1319 // If it is a .dynsym, there should be no local symbols, but we need 1320 // to do a few things for the dynamic linker. 1321 if (this->Type == SHT_DYNSYM) { 1322 // Section's Info field has the index of the first non-local symbol. 1323 // Because the first symbol entry is a null entry, 1 is the first. 1324 getParent()->Info = 1; 1325 1326 if (InX::GnuHashTab) { 1327 // NB: It also sorts Symbols to meet the GNU hash table requirements. 1328 InX::GnuHashTab->addSymbols(Symbols); 1329 } else if (Config->EMachine == EM_MIPS) { 1330 std::stable_sort(Symbols.begin(), Symbols.end(), sortMipsSymbols); 1331 } 1332 1333 size_t I = 0; 1334 for (const SymbolTableEntry &S : Symbols) 1335 S.Symbol->DynsymIndex = ++I; 1336 return; 1337 } 1338 } 1339 1340 void SymbolTableBaseSection::postThunkContents() { 1341 if (this->Type == SHT_DYNSYM) 1342 return; 1343 // move all local symbols before global symbols. 1344 auto It = std::stable_partition( 1345 Symbols.begin(), Symbols.end(), [](const SymbolTableEntry &S) { 1346 return S.Symbol->isLocal() || 1347 S.Symbol->symbol()->computeBinding() == STB_LOCAL; 1348 }); 1349 size_t NumLocals = It - Symbols.begin(); 1350 getParent()->Info = NumLocals + 1; 1351 } 1352 1353 void SymbolTableBaseSection::addSymbol(SymbolBody *B) { 1354 // Adding a local symbol to a .dynsym is a bug. 1355 assert(this->Type != SHT_DYNSYM || !B->isLocal()); 1356 1357 bool HashIt = B->isLocal(); 1358 Symbols.push_back({B, StrTabSec.addString(B->getName(), HashIt)}); 1359 } 1360 1361 size_t SymbolTableBaseSection::getSymbolIndex(SymbolBody *Body) { 1362 // Initializes symbol lookup tables lazily. This is used only 1363 // for -r or -emit-relocs. 1364 llvm::call_once(OnceFlag, [&] { 1365 SymbolIndexMap.reserve(Symbols.size()); 1366 size_t I = 0; 1367 for (const SymbolTableEntry &E : Symbols) { 1368 if (E.Symbol->Type == STT_SECTION) 1369 SectionIndexMap[E.Symbol->getOutputSection()] = ++I; 1370 else 1371 SymbolIndexMap[E.Symbol] = ++I; 1372 } 1373 }); 1374 1375 // Section symbols are mapped based on their output sections 1376 // to maintain their semantics. 1377 if (Body->Type == STT_SECTION) 1378 return SectionIndexMap.lookup(Body->getOutputSection()); 1379 return SymbolIndexMap.lookup(Body); 1380 } 1381 1382 template <class ELFT> 1383 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &StrTabSec) 1384 : SymbolTableBaseSection(StrTabSec) { 1385 this->Entsize = sizeof(Elf_Sym); 1386 } 1387 1388 // Write the internal symbol table contents to the output symbol table. 1389 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *Buf) { 1390 // The first entry is a null entry as per the ELF spec. 1391 memset(Buf, 0, sizeof(Elf_Sym)); 1392 Buf += sizeof(Elf_Sym); 1393 1394 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf); 1395 1396 for (SymbolTableEntry &Ent : Symbols) { 1397 SymbolBody *Body = Ent.Symbol; 1398 1399 // Set st_info and st_other. 1400 ESym->st_other = 0; 1401 if (Body->isLocal()) { 1402 ESym->setBindingAndType(STB_LOCAL, Body->Type); 1403 } else { 1404 ESym->setBindingAndType(Body->symbol()->computeBinding(), Body->Type); 1405 ESym->setVisibility(Body->symbol()->Visibility); 1406 } 1407 1408 ESym->st_name = Ent.StrTabOffset; 1409 1410 // Set a section index. 1411 if (const OutputSection *OutSec = Body->getOutputSection()) 1412 ESym->st_shndx = OutSec->SectionIndex; 1413 else if (isa<DefinedRegular>(Body)) 1414 ESym->st_shndx = SHN_ABS; 1415 else if (isa<DefinedCommon>(Body)) 1416 ESym->st_shndx = SHN_COMMON; 1417 else 1418 ESym->st_shndx = SHN_UNDEF; 1419 1420 // Copy symbol size if it is a defined symbol. st_size is not significant 1421 // for undefined symbols, so whether copying it or not is up to us if that's 1422 // the case. We'll leave it as zero because by not setting a value, we can 1423 // get the exact same outputs for two sets of input files that differ only 1424 // in undefined symbol size in DSOs. 1425 if (ESym->st_shndx == SHN_UNDEF) 1426 ESym->st_size = 0; 1427 else 1428 ESym->st_size = Body->getSize<ELFT>(); 1429 1430 // st_value is usually an address of a symbol, but that has a 1431 // special meaining for uninstantiated common symbols (this can 1432 // occur if -r is given). 1433 if (!Config->DefineCommon && isa<DefinedCommon>(Body)) 1434 ESym->st_value = cast<DefinedCommon>(Body)->Alignment; 1435 else 1436 ESym->st_value = Body->getVA(); 1437 1438 ++ESym; 1439 } 1440 1441 // On MIPS we need to mark symbol which has a PLT entry and requires 1442 // pointer equality by STO_MIPS_PLT flag. That is necessary to help 1443 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs. 1444 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt 1445 if (Config->EMachine == EM_MIPS) { 1446 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf); 1447 1448 for (SymbolTableEntry &Ent : Symbols) { 1449 SymbolBody *Body = Ent.Symbol; 1450 if (Body->isInPlt() && Body->NeedsPltAddr) 1451 ESym->st_other |= STO_MIPS_PLT; 1452 1453 if (Config->Relocatable) 1454 if (auto *D = dyn_cast<DefinedRegular>(Body)) 1455 if (D->isMipsPIC<ELFT>()) 1456 ESym->st_other |= STO_MIPS_PIC; 1457 ++ESym; 1458 } 1459 } 1460 } 1461 1462 // .hash and .gnu.hash sections contain on-disk hash tables that map 1463 // symbol names to their dynamic symbol table indices. Their purpose 1464 // is to help the dynamic linker resolve symbols quickly. If ELF files 1465 // don't have them, the dynamic linker has to do linear search on all 1466 // dynamic symbols, which makes programs slower. Therefore, a .hash 1467 // section is added to a DSO by default. A .gnu.hash is added if you 1468 // give the -hash-style=gnu or -hash-style=both option. 1469 // 1470 // The Unix semantics of resolving dynamic symbols is somewhat expensive. 1471 // Each ELF file has a list of DSOs that the ELF file depends on and a 1472 // list of dynamic symbols that need to be resolved from any of the 1473 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n) 1474 // where m is the number of DSOs and n is the number of dynamic 1475 // symbols. For modern large programs, both m and n are large. So 1476 // making each step faster by using hash tables substiantially 1477 // improves time to load programs. 1478 // 1479 // (Note that this is not the only way to design the shared library. 1480 // For instance, the Windows DLL takes a different approach. On 1481 // Windows, each dynamic symbol has a name of DLL from which the symbol 1482 // has to be resolved. That makes the cost of symbol resolution O(n). 1483 // This disables some hacky techniques you can use on Unix such as 1484 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.) 1485 // 1486 // Due to historical reasons, we have two different hash tables, .hash 1487 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new 1488 // and better version of .hash. .hash is just an on-disk hash table, but 1489 // .gnu.hash has a bloom filter in addition to a hash table to skip 1490 // DSOs very quickly. If you are sure that your dynamic linker knows 1491 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a 1492 // safe bet is to specify -hash-style=both for backward compatibilty. 1493 GnuHashTableSection::GnuHashTableSection() 1494 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, Config->Wordsize, ".gnu.hash") { 1495 } 1496 1497 void GnuHashTableSection::finalizeContents() { 1498 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex; 1499 1500 // Computes bloom filter size in word size. We want to allocate 8 1501 // bits for each symbol. It must be a power of two. 1502 if (Symbols.empty()) 1503 MaskWords = 1; 1504 else 1505 MaskWords = NextPowerOf2((Symbols.size() - 1) / Config->Wordsize); 1506 1507 Size = 16; // Header 1508 Size += Config->Wordsize * MaskWords; // Bloom filter 1509 Size += NBuckets * 4; // Hash buckets 1510 Size += Symbols.size() * 4; // Hash values 1511 } 1512 1513 void GnuHashTableSection::writeTo(uint8_t *Buf) { 1514 // Write a header. 1515 write32(Buf, NBuckets, Config->Endianness); 1516 write32(Buf + 4, InX::DynSymTab->getNumSymbols() - Symbols.size(), 1517 Config->Endianness); 1518 write32(Buf + 8, MaskWords, Config->Endianness); 1519 write32(Buf + 12, getShift2(), Config->Endianness); 1520 Buf += 16; 1521 1522 // Write a bloom filter and a hash table. 1523 writeBloomFilter(Buf); 1524 Buf += Config->Wordsize * MaskWords; 1525 writeHashTable(Buf); 1526 } 1527 1528 // This function writes a 2-bit bloom filter. This bloom filter alone 1529 // usually filters out 80% or more of all symbol lookups [1]. 1530 // The dynamic linker uses the hash table only when a symbol is not 1531 // filtered out by a bloom filter. 1532 // 1533 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2), 1534 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf 1535 void GnuHashTableSection::writeBloomFilter(uint8_t *Buf) { 1536 const unsigned C = Config->Wordsize * 8; 1537 for (const Entry &Sym : Symbols) { 1538 size_t I = (Sym.Hash / C) & (MaskWords - 1); 1539 uint64_t Val = readUint(Buf + I * Config->Wordsize); 1540 Val |= uint64_t(1) << (Sym.Hash % C); 1541 Val |= uint64_t(1) << ((Sym.Hash >> getShift2()) % C); 1542 writeUint(Buf + I * Config->Wordsize, Val); 1543 } 1544 } 1545 1546 void GnuHashTableSection::writeHashTable(uint8_t *Buf) { 1547 // Group symbols by hash value. 1548 std::vector<std::vector<Entry>> Syms(NBuckets); 1549 for (const Entry &Ent : Symbols) 1550 Syms[Ent.Hash % NBuckets].push_back(Ent); 1551 1552 // Write hash buckets. Hash buckets contain indices in the following 1553 // hash value table. 1554 uint32_t *Buckets = reinterpret_cast<uint32_t *>(Buf); 1555 for (size_t I = 0; I < NBuckets; ++I) 1556 if (!Syms[I].empty()) 1557 write32(Buckets + I, Syms[I][0].Body->DynsymIndex, Config->Endianness); 1558 1559 // Write a hash value table. It represents a sequence of chains that 1560 // share the same hash modulo value. The last element of each chain 1561 // is terminated by LSB 1. 1562 uint32_t *Values = Buckets + NBuckets; 1563 size_t I = 0; 1564 for (std::vector<Entry> &Vec : Syms) { 1565 if (Vec.empty()) 1566 continue; 1567 for (const Entry &Ent : makeArrayRef(Vec).drop_back()) 1568 write32(Values + I++, Ent.Hash & ~1, Config->Endianness); 1569 write32(Values + I++, Vec.back().Hash | 1, Config->Endianness); 1570 } 1571 } 1572 1573 static uint32_t hashGnu(StringRef Name) { 1574 uint32_t H = 5381; 1575 for (uint8_t C : Name) 1576 H = (H << 5) + H + C; 1577 return H; 1578 } 1579 1580 // Returns a number of hash buckets to accomodate given number of elements. 1581 // We want to choose a moderate number that is not too small (which 1582 // causes too many hash collisions) and not too large (which wastes 1583 // disk space.) 1584 // 1585 // We return a prime number because it (is believed to) achieve good 1586 // hash distribution. 1587 static size_t getBucketSize(size_t NumSymbols) { 1588 // List of largest prime numbers that are not greater than 2^n + 1. 1589 for (size_t N : {131071, 65521, 32749, 16381, 8191, 4093, 2039, 1021, 509, 1590 251, 127, 61, 31, 13, 7, 3, 1}) 1591 if (N <= NumSymbols) 1592 return N; 1593 return 0; 1594 } 1595 1596 // Add symbols to this symbol hash table. Note that this function 1597 // destructively sort a given vector -- which is needed because 1598 // GNU-style hash table places some sorting requirements. 1599 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &V) { 1600 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce 1601 // its type correctly. 1602 std::vector<SymbolTableEntry>::iterator Mid = 1603 std::stable_partition(V.begin(), V.end(), [](const SymbolTableEntry &S) { 1604 return S.Symbol->isUndefined(); 1605 }); 1606 if (Mid == V.end()) 1607 return; 1608 1609 for (SymbolTableEntry &Ent : llvm::make_range(Mid, V.end())) { 1610 SymbolBody *B = Ent.Symbol; 1611 Symbols.push_back({B, Ent.StrTabOffset, hashGnu(B->getName())}); 1612 } 1613 1614 NBuckets = getBucketSize(Symbols.size()); 1615 std::stable_sort(Symbols.begin(), Symbols.end(), 1616 [&](const Entry &L, const Entry &R) { 1617 return L.Hash % NBuckets < R.Hash % NBuckets; 1618 }); 1619 1620 V.erase(Mid, V.end()); 1621 for (const Entry &Ent : Symbols) 1622 V.push_back({Ent.Body, Ent.StrTabOffset}); 1623 } 1624 1625 HashTableSection::HashTableSection() 1626 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") { 1627 this->Entsize = 4; 1628 } 1629 1630 void HashTableSection::finalizeContents() { 1631 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex; 1632 1633 unsigned NumEntries = 2; // nbucket and nchain. 1634 NumEntries += InX::DynSymTab->getNumSymbols(); // The chain entries. 1635 1636 // Create as many buckets as there are symbols. 1637 // FIXME: This is simplistic. We can try to optimize it, but implementing 1638 // support for SHT_GNU_HASH is probably even more profitable. 1639 NumEntries += InX::DynSymTab->getNumSymbols(); 1640 this->Size = NumEntries * 4; 1641 } 1642 1643 void HashTableSection::writeTo(uint8_t *Buf) { 1644 unsigned NumSymbols = InX::DynSymTab->getNumSymbols(); 1645 1646 uint32_t *P = reinterpret_cast<uint32_t *>(Buf); 1647 write32(P++, NumSymbols, Config->Endianness); // nbucket 1648 write32(P++, NumSymbols, Config->Endianness); // nchain 1649 1650 uint32_t *Buckets = P; 1651 uint32_t *Chains = P + NumSymbols; 1652 1653 for (const SymbolTableEntry &S : InX::DynSymTab->getSymbols()) { 1654 SymbolBody *Body = S.Symbol; 1655 StringRef Name = Body->getName(); 1656 unsigned I = Body->DynsymIndex; 1657 uint32_t Hash = hashSysV(Name) % NumSymbols; 1658 Chains[I] = Buckets[Hash]; 1659 write32(Buckets + Hash, I, Config->Endianness); 1660 } 1661 } 1662 1663 PltSection::PltSection(size_t S) 1664 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt"), 1665 HeaderSize(S) { 1666 // The PLT needs to be writable on SPARC as the dynamic linker will 1667 // modify the instructions in the PLT entries. 1668 if (Config->EMachine == EM_SPARCV9) 1669 this->Flags |= SHF_WRITE; 1670 } 1671 1672 void PltSection::writeTo(uint8_t *Buf) { 1673 // At beginning of PLT but not the IPLT, we have code to call the dynamic 1674 // linker to resolve dynsyms at runtime. Write such code. 1675 if (HeaderSize != 0) 1676 Target->writePltHeader(Buf); 1677 size_t Off = HeaderSize; 1678 // The IPlt is immediately after the Plt, account for this in RelOff 1679 unsigned PltOff = getPltRelocOff(); 1680 1681 for (auto &I : Entries) { 1682 const SymbolBody *B = I.first; 1683 unsigned RelOff = I.second + PltOff; 1684 uint64_t Got = B->getGotPltVA(); 1685 uint64_t Plt = this->getVA() + Off; 1686 Target->writePlt(Buf + Off, Got, Plt, B->PltIndex, RelOff); 1687 Off += Target->PltEntrySize; 1688 } 1689 } 1690 1691 template <class ELFT> void PltSection::addEntry(SymbolBody &Sym) { 1692 Sym.PltIndex = Entries.size(); 1693 RelocationSection<ELFT> *PltRelocSection = In<ELFT>::RelaPlt; 1694 if (HeaderSize == 0) { 1695 PltRelocSection = In<ELFT>::RelaIplt; 1696 Sym.IsInIplt = true; 1697 } 1698 unsigned RelOff = PltRelocSection->getRelocOffset(); 1699 Entries.push_back(std::make_pair(&Sym, RelOff)); 1700 } 1701 1702 size_t PltSection::getSize() const { 1703 return HeaderSize + Entries.size() * Target->PltEntrySize; 1704 } 1705 1706 // Some architectures such as additional symbols in the PLT section. For 1707 // example ARM uses mapping symbols to aid disassembly 1708 void PltSection::addSymbols() { 1709 // The PLT may have symbols defined for the Header, the IPLT has no header 1710 if (HeaderSize != 0) 1711 Target->addPltHeaderSymbols(this); 1712 size_t Off = HeaderSize; 1713 for (size_t I = 0; I < Entries.size(); ++I) { 1714 Target->addPltSymbols(this, Off); 1715 Off += Target->PltEntrySize; 1716 } 1717 } 1718 1719 unsigned PltSection::getPltRelocOff() const { 1720 return (HeaderSize == 0) ? InX::Plt->getSize() : 0; 1721 } 1722 1723 // The string hash function for .gdb_index. 1724 static uint32_t computeGdbHash(StringRef S) { 1725 uint32_t H = 0; 1726 for (uint8_t C : S) 1727 H = H * 67 + tolower(C) - 113; 1728 return H; 1729 } 1730 1731 static std::vector<GdbIndexChunk::CuEntry> readCuList(DWARFContext &Dwarf) { 1732 std::vector<GdbIndexChunk::CuEntry> Ret; 1733 for (std::unique_ptr<DWARFCompileUnit> &Cu : Dwarf.compile_units()) 1734 Ret.push_back({Cu->getOffset(), Cu->getLength() + 4}); 1735 return Ret; 1736 } 1737 1738 static std::vector<GdbIndexChunk::AddressEntry> 1739 readAddressAreas(DWARFContext &Dwarf, InputSection *Sec) { 1740 std::vector<GdbIndexChunk::AddressEntry> Ret; 1741 1742 uint32_t CuIdx = 0; 1743 for (std::unique_ptr<DWARFCompileUnit> &Cu : Dwarf.compile_units()) { 1744 DWARFAddressRangesVector Ranges; 1745 Cu->collectAddressRanges(Ranges); 1746 1747 ArrayRef<InputSectionBase *> Sections = Sec->File->getSections(); 1748 for (DWARFAddressRange &R : Ranges) { 1749 InputSectionBase *S = Sections[R.SectionIndex]; 1750 if (!S || S == &InputSection::Discarded || !S->Live) 1751 continue; 1752 // Range list with zero size has no effect. 1753 if (R.LowPC == R.HighPC) 1754 continue; 1755 auto *IS = cast<InputSection>(S); 1756 uint64_t Offset = IS->getOffsetInFile(); 1757 Ret.push_back({IS, R.LowPC - Offset, R.HighPC - Offset, CuIdx}); 1758 } 1759 ++CuIdx; 1760 } 1761 return Ret; 1762 } 1763 1764 static std::vector<GdbIndexChunk::NameTypeEntry> 1765 readPubNamesAndTypes(DWARFContext &Dwarf) { 1766 StringRef Sec1 = Dwarf.getDWARFObj().getGnuPubNamesSection(); 1767 StringRef Sec2 = Dwarf.getDWARFObj().getGnuPubTypesSection(); 1768 1769 std::vector<GdbIndexChunk::NameTypeEntry> Ret; 1770 for (StringRef Sec : {Sec1, Sec2}) { 1771 DWARFDebugPubTable Table(Sec, Config->IsLE, true); 1772 for (const DWARFDebugPubTable::Set &Set : Table.getData()) { 1773 for (const DWARFDebugPubTable::Entry &Ent : Set.Entries) { 1774 CachedHashStringRef S(Ent.Name, computeGdbHash(Ent.Name)); 1775 Ret.push_back({S, Ent.Descriptor.toBits()}); 1776 } 1777 } 1778 } 1779 return Ret; 1780 } 1781 1782 static std::vector<InputSection *> getDebugInfoSections() { 1783 std::vector<InputSection *> Ret; 1784 for (InputSectionBase *S : InputSections) 1785 if (InputSection *IS = dyn_cast<InputSection>(S)) 1786 if (IS->Name == ".debug_info") 1787 Ret.push_back(IS); 1788 return Ret; 1789 } 1790 1791 void GdbIndexSection::fixCuIndex() { 1792 uint32_t Idx = 0; 1793 for (GdbIndexChunk &Chunk : Chunks) { 1794 for (GdbIndexChunk::AddressEntry &Ent : Chunk.AddressAreas) 1795 Ent.CuIndex += Idx; 1796 Idx += Chunk.CompilationUnits.size(); 1797 } 1798 } 1799 1800 std::vector<std::vector<uint32_t>> GdbIndexSection::createCuVectors() { 1801 std::vector<std::vector<uint32_t>> Ret; 1802 uint32_t Idx = 0; 1803 uint32_t Off = 0; 1804 1805 for (GdbIndexChunk &Chunk : Chunks) { 1806 for (GdbIndexChunk::NameTypeEntry &Ent : Chunk.NamesAndTypes) { 1807 GdbSymbol *&Sym = Symbols[Ent.Name]; 1808 if (!Sym) { 1809 Sym = make<GdbSymbol>(GdbSymbol{Ent.Name.hash(), Off, Ret.size()}); 1810 Off += Ent.Name.size() + 1; 1811 Ret.push_back({}); 1812 } 1813 1814 // gcc 5.4.1 produces a buggy .debug_gnu_pubnames that contains 1815 // duplicate entries, so we want to dedup them. 1816 std::vector<uint32_t> &Vec = Ret[Sym->CuVectorIndex]; 1817 uint32_t Val = (Ent.Type << 24) | Idx; 1818 if (Vec.empty() || Vec.back() != Val) 1819 Vec.push_back(Val); 1820 } 1821 Idx += Chunk.CompilationUnits.size(); 1822 } 1823 1824 StringPoolSize = Off; 1825 return Ret; 1826 } 1827 1828 template <class ELFT> GdbIndexSection *elf::createGdbIndex() { 1829 std::vector<InputSection *> Sections = getDebugInfoSections(); 1830 std::vector<GdbIndexChunk> Chunks(Sections.size()); 1831 1832 parallelForEachN(0, Chunks.size(), [&](size_t I) { 1833 ObjFile<ELFT> *File = Sections[I]->getFile<ELFT>(); 1834 DWARFContext Dwarf(make_unique<LLDDwarfObj<ELFT>>(File)); 1835 1836 Chunks[I].DebugInfoSec = Sections[I]; 1837 Chunks[I].CompilationUnits = readCuList(Dwarf); 1838 Chunks[I].AddressAreas = readAddressAreas(Dwarf, Sections[I]); 1839 Chunks[I].NamesAndTypes = readPubNamesAndTypes(Dwarf); 1840 }); 1841 1842 return make<GdbIndexSection>(std::move(Chunks)); 1843 } 1844 1845 static size_t getCuSize(ArrayRef<GdbIndexChunk> Arr) { 1846 size_t Ret = 0; 1847 for (const GdbIndexChunk &D : Arr) 1848 Ret += D.CompilationUnits.size(); 1849 return Ret; 1850 } 1851 1852 static size_t getAddressAreaSize(ArrayRef<GdbIndexChunk> Arr) { 1853 size_t Ret = 0; 1854 for (const GdbIndexChunk &D : Arr) 1855 Ret += D.AddressAreas.size(); 1856 return Ret; 1857 } 1858 1859 std::vector<GdbSymbol *> GdbIndexSection::createGdbSymtab() { 1860 uint32_t Size = NextPowerOf2(Symbols.size() * 4 / 3); 1861 if (Size < 1024) 1862 Size = 1024; 1863 1864 uint32_t Mask = Size - 1; 1865 std::vector<GdbSymbol *> Ret(Size); 1866 1867 for (auto &KV : Symbols) { 1868 GdbSymbol *Sym = KV.second; 1869 uint32_t I = Sym->NameHash & Mask; 1870 uint32_t Step = ((Sym->NameHash * 17) & Mask) | 1; 1871 1872 while (Ret[I]) 1873 I = (I + Step) & Mask; 1874 Ret[I] = Sym; 1875 } 1876 return Ret; 1877 } 1878 1879 GdbIndexSection::GdbIndexSection(std::vector<GdbIndexChunk> &&C) 1880 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index"), Chunks(std::move(C)) { 1881 fixCuIndex(); 1882 CuVectors = createCuVectors(); 1883 GdbSymtab = createGdbSymtab(); 1884 1885 // Compute offsets early to know the section size. 1886 // Each chunk size needs to be in sync with what we write in writeTo. 1887 CuTypesOffset = CuListOffset + getCuSize(Chunks) * 16; 1888 SymtabOffset = CuTypesOffset + getAddressAreaSize(Chunks) * 20; 1889 ConstantPoolOffset = SymtabOffset + GdbSymtab.size() * 8; 1890 1891 size_t Off = 0; 1892 for (ArrayRef<uint32_t> Vec : CuVectors) { 1893 CuVectorOffsets.push_back(Off); 1894 Off += (Vec.size() + 1) * 4; 1895 } 1896 StringPoolOffset = ConstantPoolOffset + Off; 1897 } 1898 1899 size_t GdbIndexSection::getSize() const { 1900 return StringPoolOffset + StringPoolSize; 1901 } 1902 1903 void GdbIndexSection::writeTo(uint8_t *Buf) { 1904 // Write the section header. 1905 write32le(Buf, 7); 1906 write32le(Buf + 4, CuListOffset); 1907 write32le(Buf + 8, CuTypesOffset); 1908 write32le(Buf + 12, CuTypesOffset); 1909 write32le(Buf + 16, SymtabOffset); 1910 write32le(Buf + 20, ConstantPoolOffset); 1911 Buf += 24; 1912 1913 // Write the CU list. 1914 for (GdbIndexChunk &D : Chunks) { 1915 for (GdbIndexChunk::CuEntry &Cu : D.CompilationUnits) { 1916 write64le(Buf, D.DebugInfoSec->OutSecOff + Cu.CuOffset); 1917 write64le(Buf + 8, Cu.CuLength); 1918 Buf += 16; 1919 } 1920 } 1921 1922 // Write the address area. 1923 for (GdbIndexChunk &D : Chunks) { 1924 for (GdbIndexChunk::AddressEntry &E : D.AddressAreas) { 1925 uint64_t BaseAddr = 1926 E.Section->getParent()->Addr + E.Section->getOffset(0); 1927 write64le(Buf, BaseAddr + E.LowAddress); 1928 write64le(Buf + 8, BaseAddr + E.HighAddress); 1929 write32le(Buf + 16, E.CuIndex); 1930 Buf += 20; 1931 } 1932 } 1933 1934 // Write the symbol table. 1935 for (GdbSymbol *Sym : GdbSymtab) { 1936 if (Sym) { 1937 write32le(Buf, Sym->NameOffset + StringPoolOffset - ConstantPoolOffset); 1938 write32le(Buf + 4, CuVectorOffsets[Sym->CuVectorIndex]); 1939 } 1940 Buf += 8; 1941 } 1942 1943 // Write the CU vectors. 1944 for (ArrayRef<uint32_t> Vec : CuVectors) { 1945 write32le(Buf, Vec.size()); 1946 Buf += 4; 1947 for (uint32_t Val : Vec) { 1948 write32le(Buf, Val); 1949 Buf += 4; 1950 } 1951 } 1952 1953 // Write the string pool. 1954 for (auto &KV : Symbols) { 1955 CachedHashStringRef S = KV.first; 1956 GdbSymbol *Sym = KV.second; 1957 size_t Off = Sym->NameOffset; 1958 memcpy(Buf + Off, S.val().data(), S.size()); 1959 Buf[Off + S.size()] = '\0'; 1960 } 1961 } 1962 1963 bool GdbIndexSection::empty() const { return !Out::DebugInfo; } 1964 1965 template <class ELFT> 1966 EhFrameHeader<ELFT>::EhFrameHeader() 1967 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame_hdr") {} 1968 1969 // .eh_frame_hdr contains a binary search table of pointers to FDEs. 1970 // Each entry of the search table consists of two values, 1971 // the starting PC from where FDEs covers, and the FDE's address. 1972 // It is sorted by PC. 1973 template <class ELFT> void EhFrameHeader<ELFT>::writeTo(uint8_t *Buf) { 1974 const endianness E = ELFT::TargetEndianness; 1975 1976 // Sort the FDE list by their PC and uniqueify. Usually there is only 1977 // one FDE for a PC (i.e. function), but if ICF merges two functions 1978 // into one, there can be more than one FDEs pointing to the address. 1979 auto Less = [](const FdeData &A, const FdeData &B) { return A.Pc < B.Pc; }; 1980 std::stable_sort(Fdes.begin(), Fdes.end(), Less); 1981 auto Eq = [](const FdeData &A, const FdeData &B) { return A.Pc == B.Pc; }; 1982 Fdes.erase(std::unique(Fdes.begin(), Fdes.end(), Eq), Fdes.end()); 1983 1984 Buf[0] = 1; 1985 Buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4; 1986 Buf[2] = DW_EH_PE_udata4; 1987 Buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4; 1988 write32<E>(Buf + 4, In<ELFT>::EhFrame->getParent()->Addr - this->getVA() - 4); 1989 write32<E>(Buf + 8, Fdes.size()); 1990 Buf += 12; 1991 1992 uint64_t VA = this->getVA(); 1993 for (FdeData &Fde : Fdes) { 1994 write32<E>(Buf, Fde.Pc - VA); 1995 write32<E>(Buf + 4, Fde.FdeVA - VA); 1996 Buf += 8; 1997 } 1998 } 1999 2000 template <class ELFT> size_t EhFrameHeader<ELFT>::getSize() const { 2001 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs. 2002 return 12 + In<ELFT>::EhFrame->NumFdes * 8; 2003 } 2004 2005 template <class ELFT> 2006 void EhFrameHeader<ELFT>::addFde(uint32_t Pc, uint32_t FdeVA) { 2007 Fdes.push_back({Pc, FdeVA}); 2008 } 2009 2010 template <class ELFT> bool EhFrameHeader<ELFT>::empty() const { 2011 return In<ELFT>::EhFrame->empty(); 2012 } 2013 2014 template <class ELFT> 2015 VersionDefinitionSection<ELFT>::VersionDefinitionSection() 2016 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t), 2017 ".gnu.version_d") {} 2018 2019 static StringRef getFileDefName() { 2020 if (!Config->SoName.empty()) 2021 return Config->SoName; 2022 return Config->OutputFile; 2023 } 2024 2025 template <class ELFT> void VersionDefinitionSection<ELFT>::finalizeContents() { 2026 FileDefNameOff = InX::DynStrTab->addString(getFileDefName()); 2027 for (VersionDefinition &V : Config->VersionDefinitions) 2028 V.NameOff = InX::DynStrTab->addString(V.Name); 2029 2030 getParent()->Link = InX::DynStrTab->getParent()->SectionIndex; 2031 2032 // sh_info should be set to the number of definitions. This fact is missed in 2033 // documentation, but confirmed by binutils community: 2034 // https://sourceware.org/ml/binutils/2014-11/msg00355.html 2035 getParent()->Info = getVerDefNum(); 2036 } 2037 2038 template <class ELFT> 2039 void VersionDefinitionSection<ELFT>::writeOne(uint8_t *Buf, uint32_t Index, 2040 StringRef Name, size_t NameOff) { 2041 auto *Verdef = reinterpret_cast<Elf_Verdef *>(Buf); 2042 Verdef->vd_version = 1; 2043 Verdef->vd_cnt = 1; 2044 Verdef->vd_aux = sizeof(Elf_Verdef); 2045 Verdef->vd_next = sizeof(Elf_Verdef) + sizeof(Elf_Verdaux); 2046 Verdef->vd_flags = (Index == 1 ? VER_FLG_BASE : 0); 2047 Verdef->vd_ndx = Index; 2048 Verdef->vd_hash = hashSysV(Name); 2049 2050 auto *Verdaux = reinterpret_cast<Elf_Verdaux *>(Buf + sizeof(Elf_Verdef)); 2051 Verdaux->vda_name = NameOff; 2052 Verdaux->vda_next = 0; 2053 } 2054 2055 template <class ELFT> 2056 void VersionDefinitionSection<ELFT>::writeTo(uint8_t *Buf) { 2057 writeOne(Buf, 1, getFileDefName(), FileDefNameOff); 2058 2059 for (VersionDefinition &V : Config->VersionDefinitions) { 2060 Buf += sizeof(Elf_Verdef) + sizeof(Elf_Verdaux); 2061 writeOne(Buf, V.Id, V.Name, V.NameOff); 2062 } 2063 2064 // Need to terminate the last version definition. 2065 Elf_Verdef *Verdef = reinterpret_cast<Elf_Verdef *>(Buf); 2066 Verdef->vd_next = 0; 2067 } 2068 2069 template <class ELFT> size_t VersionDefinitionSection<ELFT>::getSize() const { 2070 return (sizeof(Elf_Verdef) + sizeof(Elf_Verdaux)) * getVerDefNum(); 2071 } 2072 2073 template <class ELFT> 2074 VersionTableSection<ELFT>::VersionTableSection() 2075 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t), 2076 ".gnu.version") { 2077 this->Entsize = sizeof(Elf_Versym); 2078 } 2079 2080 template <class ELFT> void VersionTableSection<ELFT>::finalizeContents() { 2081 // At the moment of june 2016 GNU docs does not mention that sh_link field 2082 // should be set, but Sun docs do. Also readelf relies on this field. 2083 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex; 2084 } 2085 2086 template <class ELFT> size_t VersionTableSection<ELFT>::getSize() const { 2087 return sizeof(Elf_Versym) * (InX::DynSymTab->getSymbols().size() + 1); 2088 } 2089 2090 template <class ELFT> void VersionTableSection<ELFT>::writeTo(uint8_t *Buf) { 2091 auto *OutVersym = reinterpret_cast<Elf_Versym *>(Buf) + 1; 2092 for (const SymbolTableEntry &S : InX::DynSymTab->getSymbols()) { 2093 OutVersym->vs_index = S.Symbol->symbol()->VersionId; 2094 ++OutVersym; 2095 } 2096 } 2097 2098 template <class ELFT> bool VersionTableSection<ELFT>::empty() const { 2099 return !In<ELFT>::VerDef && In<ELFT>::VerNeed->empty(); 2100 } 2101 2102 template <class ELFT> 2103 VersionNeedSection<ELFT>::VersionNeedSection() 2104 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t), 2105 ".gnu.version_r") { 2106 // Identifiers in verneed section start at 2 because 0 and 1 are reserved 2107 // for VER_NDX_LOCAL and VER_NDX_GLOBAL. 2108 // First identifiers are reserved by verdef section if it exist. 2109 NextIndex = getVerDefNum() + 1; 2110 } 2111 2112 template <class ELFT> 2113 void VersionNeedSection<ELFT>::addSymbol(SharedSymbol *SS) { 2114 auto *Ver = reinterpret_cast<const typename ELFT::Verdef *>(SS->Verdef); 2115 if (!Ver) { 2116 SS->symbol()->VersionId = VER_NDX_GLOBAL; 2117 return; 2118 } 2119 2120 SharedFile<ELFT> *File = SS->getFile<ELFT>(); 2121 2122 // If we don't already know that we need an Elf_Verneed for this DSO, prepare 2123 // to create one by adding it to our needed list and creating a dynstr entry 2124 // for the soname. 2125 if (File->VerdefMap.empty()) 2126 Needed.push_back({File, InX::DynStrTab->addString(File->SoName)}); 2127 typename SharedFile<ELFT>::NeededVer &NV = File->VerdefMap[Ver]; 2128 // If we don't already know that we need an Elf_Vernaux for this Elf_Verdef, 2129 // prepare to create one by allocating a version identifier and creating a 2130 // dynstr entry for the version name. 2131 if (NV.Index == 0) { 2132 NV.StrTab = InX::DynStrTab->addString(File->getStringTable().data() + 2133 Ver->getAux()->vda_name); 2134 NV.Index = NextIndex++; 2135 } 2136 SS->symbol()->VersionId = NV.Index; 2137 } 2138 2139 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *Buf) { 2140 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs. 2141 auto *Verneed = reinterpret_cast<Elf_Verneed *>(Buf); 2142 auto *Vernaux = reinterpret_cast<Elf_Vernaux *>(Verneed + Needed.size()); 2143 2144 for (std::pair<SharedFile<ELFT> *, size_t> &P : Needed) { 2145 // Create an Elf_Verneed for this DSO. 2146 Verneed->vn_version = 1; 2147 Verneed->vn_cnt = P.first->VerdefMap.size(); 2148 Verneed->vn_file = P.second; 2149 Verneed->vn_aux = 2150 reinterpret_cast<char *>(Vernaux) - reinterpret_cast<char *>(Verneed); 2151 Verneed->vn_next = sizeof(Elf_Verneed); 2152 ++Verneed; 2153 2154 // Create the Elf_Vernauxs for this Elf_Verneed. The loop iterates over 2155 // VerdefMap, which will only contain references to needed version 2156 // definitions. Each Elf_Vernaux is based on the information contained in 2157 // the Elf_Verdef in the source DSO. This loop iterates over a std::map of 2158 // pointers, but is deterministic because the pointers refer to Elf_Verdef 2159 // data structures within a single input file. 2160 for (auto &NV : P.first->VerdefMap) { 2161 Vernaux->vna_hash = NV.first->vd_hash; 2162 Vernaux->vna_flags = 0; 2163 Vernaux->vna_other = NV.second.Index; 2164 Vernaux->vna_name = NV.second.StrTab; 2165 Vernaux->vna_next = sizeof(Elf_Vernaux); 2166 ++Vernaux; 2167 } 2168 2169 Vernaux[-1].vna_next = 0; 2170 } 2171 Verneed[-1].vn_next = 0; 2172 } 2173 2174 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() { 2175 getParent()->Link = InX::DynStrTab->getParent()->SectionIndex; 2176 getParent()->Info = Needed.size(); 2177 } 2178 2179 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const { 2180 unsigned Size = Needed.size() * sizeof(Elf_Verneed); 2181 for (const std::pair<SharedFile<ELFT> *, size_t> &P : Needed) 2182 Size += P.first->VerdefMap.size() * sizeof(Elf_Vernaux); 2183 return Size; 2184 } 2185 2186 template <class ELFT> bool VersionNeedSection<ELFT>::empty() const { 2187 return getNeedNum() == 0; 2188 } 2189 2190 void MergeSyntheticSection::addSection(MergeInputSection *MS) { 2191 MS->Parent = this; 2192 Sections.push_back(MS); 2193 } 2194 2195 MergeTailSection::MergeTailSection(StringRef Name, uint32_t Type, 2196 uint64_t Flags, uint32_t Alignment) 2197 : MergeSyntheticSection(Name, Type, Flags, Alignment), 2198 Builder(StringTableBuilder::RAW, Alignment) {} 2199 2200 size_t MergeTailSection::getSize() const { return Builder.getSize(); } 2201 2202 void MergeTailSection::writeTo(uint8_t *Buf) { Builder.write(Buf); } 2203 2204 void MergeTailSection::finalizeContents() { 2205 // Add all string pieces to the string table builder to create section 2206 // contents. 2207 for (MergeInputSection *Sec : Sections) 2208 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2209 if (Sec->Pieces[I].Live) 2210 Builder.add(Sec->getData(I)); 2211 2212 // Fix the string table content. After this, the contents will never change. 2213 Builder.finalize(); 2214 2215 // finalize() fixed tail-optimized strings, so we can now get 2216 // offsets of strings. Get an offset for each string and save it 2217 // to a corresponding StringPiece for easy access. 2218 for (MergeInputSection *Sec : Sections) 2219 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2220 if (Sec->Pieces[I].Live) 2221 Sec->Pieces[I].OutputOff = Builder.getOffset(Sec->getData(I)); 2222 } 2223 2224 void MergeNoTailSection::writeTo(uint8_t *Buf) { 2225 for (size_t I = 0; I < NumShards; ++I) 2226 Shards[I].write(Buf + ShardOffsets[I]); 2227 } 2228 2229 // This function is very hot (i.e. it can take several seconds to finish) 2230 // because sometimes the number of inputs is in an order of magnitude of 2231 // millions. So, we use multi-threading. 2232 // 2233 // For any strings S and T, we know S is not mergeable with T if S's hash 2234 // value is different from T's. If that's the case, we can safely put S and 2235 // T into different string builders without worrying about merge misses. 2236 // We do it in parallel. 2237 void MergeNoTailSection::finalizeContents() { 2238 // Initializes string table builders. 2239 for (size_t I = 0; I < NumShards; ++I) 2240 Shards.emplace_back(StringTableBuilder::RAW, Alignment); 2241 2242 // Concurrency level. Must be a power of 2 to avoid expensive modulo 2243 // operations in the following tight loop. 2244 size_t Concurrency = 1; 2245 if (Config->Threads) 2246 Concurrency = 2247 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), NumShards); 2248 2249 // Add section pieces to the builders. 2250 parallelForEachN(0, Concurrency, [&](size_t ThreadId) { 2251 for (MergeInputSection *Sec : Sections) { 2252 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) { 2253 if (!Sec->Pieces[I].Live) 2254 continue; 2255 CachedHashStringRef Str = Sec->getData(I); 2256 size_t ShardId = getShardId(Str.hash()); 2257 if ((ShardId & (Concurrency - 1)) == ThreadId) 2258 Sec->Pieces[I].OutputOff = Shards[ShardId].add(Str); 2259 } 2260 } 2261 }); 2262 2263 // Compute an in-section offset for each shard. 2264 size_t Off = 0; 2265 for (size_t I = 0; I < NumShards; ++I) { 2266 Shards[I].finalizeInOrder(); 2267 if (Shards[I].getSize() > 0) 2268 Off = alignTo(Off, Alignment); 2269 ShardOffsets[I] = Off; 2270 Off += Shards[I].getSize(); 2271 } 2272 Size = Off; 2273 2274 // So far, section pieces have offsets from beginning of shards, but 2275 // we want offsets from beginning of the whole section. Fix them. 2276 parallelForEach(Sections, [&](MergeInputSection *Sec) { 2277 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2278 if (Sec->Pieces[I].Live) 2279 Sec->Pieces[I].OutputOff += 2280 ShardOffsets[getShardId(Sec->getData(I).hash())]; 2281 }); 2282 } 2283 2284 static MergeSyntheticSection *createMergeSynthetic(StringRef Name, 2285 uint32_t Type, 2286 uint64_t Flags, 2287 uint32_t Alignment) { 2288 bool ShouldTailMerge = (Flags & SHF_STRINGS) && Config->Optimize >= 2; 2289 if (ShouldTailMerge) 2290 return make<MergeTailSection>(Name, Type, Flags, Alignment); 2291 return make<MergeNoTailSection>(Name, Type, Flags, Alignment); 2292 } 2293 2294 // This function decompresses compressed sections and scans over the input 2295 // sections to create mergeable synthetic sections. It removes 2296 // MergeInputSections from the input section array and adds new synthetic 2297 // sections at the location of the first input section that it replaces. It then 2298 // finalizes each synthetic section in order to compute an output offset for 2299 // each piece of each input section. 2300 void elf::decompressAndMergeSections() { 2301 // splitIntoPieces needs to be called on each MergeInputSection before calling 2302 // finalizeContents(). Do that first. 2303 parallelForEach(InputSections, [](InputSectionBase *S) { 2304 if (!S->Live) 2305 return; 2306 S->maybeUncompress(); 2307 if (auto *MS = dyn_cast<MergeInputSection>(S)) 2308 MS->splitIntoPieces(); 2309 }); 2310 2311 std::vector<MergeSyntheticSection *> MergeSections; 2312 for (InputSectionBase *&S : InputSections) { 2313 MergeInputSection *MS = dyn_cast<MergeInputSection>(S); 2314 if (!MS) 2315 continue; 2316 2317 // We do not want to handle sections that are not alive, so just remove 2318 // them instead of trying to merge. 2319 if (!MS->Live) 2320 continue; 2321 2322 StringRef OutsecName = getOutputSectionName(MS->Name); 2323 uint32_t Alignment = std::max<uint32_t>(MS->Alignment, MS->Entsize); 2324 2325 auto I = llvm::find_if(MergeSections, [=](MergeSyntheticSection *Sec) { 2326 return Sec->Name == OutsecName && Sec->Flags == MS->Flags && 2327 Sec->Alignment == Alignment; 2328 }); 2329 if (I == MergeSections.end()) { 2330 MergeSyntheticSection *Syn = 2331 createMergeSynthetic(OutsecName, MS->Type, MS->Flags, Alignment); 2332 MergeSections.push_back(Syn); 2333 I = std::prev(MergeSections.end()); 2334 S = Syn; 2335 } else { 2336 S = nullptr; 2337 } 2338 (*I)->addSection(MS); 2339 } 2340 for (auto *MS : MergeSections) 2341 MS->finalizeContents(); 2342 2343 std::vector<InputSectionBase *> &V = InputSections; 2344 V.erase(std::remove(V.begin(), V.end(), nullptr), V.end()); 2345 } 2346 2347 MipsRldMapSection::MipsRldMapSection() 2348 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, Config->Wordsize, 2349 ".rld_map") {} 2350 2351 ARMExidxSentinelSection::ARMExidxSentinelSection() 2352 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX, 2353 Config->Wordsize, ".ARM.exidx") {} 2354 2355 // Write a terminating sentinel entry to the end of the .ARM.exidx table. 2356 // This section will have been sorted last in the .ARM.exidx table. 2357 // This table entry will have the form: 2358 // | PREL31 upper bound of code that has exception tables | EXIDX_CANTUNWIND | 2359 // The sentinel must have the PREL31 value of an address higher than any 2360 // address described by any other table entry. 2361 void ARMExidxSentinelSection::writeTo(uint8_t *Buf) { 2362 // The Sections are sorted in order of ascending PREL31 address with the 2363 // sentinel last. We need to find the InputSection that precedes the 2364 // sentinel. By construction the Sentinel is in the last 2365 // InputSectionDescription as the InputSection that precedes it. 2366 OutputSection *C = getParent(); 2367 auto ISD = std::find_if(C->Commands.rbegin(), C->Commands.rend(), 2368 [](const BaseCommand *Base) { 2369 return isa<InputSectionDescription>(Base); 2370 }); 2371 auto L = cast<InputSectionDescription>(*ISD); 2372 InputSection *Highest = L->Sections[L->Sections.size() - 2]; 2373 InputSection *LS = Highest->getLinkOrderDep(); 2374 uint64_t S = LS->getParent()->Addr + LS->getOffset(LS->getSize()); 2375 uint64_t P = getVA(); 2376 Target->relocateOne(Buf, R_ARM_PREL31, S - P); 2377 write32le(Buf + 4, 0x1); 2378 } 2379 2380 ThunkSection::ThunkSection(OutputSection *OS, uint64_t Off) 2381 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 2382 Config->Wordsize, ".text.thunk") { 2383 this->Parent = OS; 2384 this->OutSecOff = Off; 2385 } 2386 2387 void ThunkSection::addThunk(Thunk *T) { 2388 uint64_t Off = alignTo(Size, T->Alignment); 2389 T->Offset = Off; 2390 Thunks.push_back(T); 2391 T->addSymbols(*this); 2392 Size = Off + T->size(); 2393 } 2394 2395 void ThunkSection::writeTo(uint8_t *Buf) { 2396 for (const Thunk *T : Thunks) 2397 T->writeTo(Buf + T->Offset, *this); 2398 } 2399 2400 InputSection *ThunkSection::getTargetInputSection() const { 2401 const Thunk *T = Thunks.front(); 2402 return T->getTargetInputSection(); 2403 } 2404 2405 InputSection *InX::ARMAttributes; 2406 BssSection *InX::Bss; 2407 BssSection *InX::BssRelRo; 2408 BuildIdSection *InX::BuildId; 2409 SyntheticSection *InX::Dynamic; 2410 StringTableSection *InX::DynStrTab; 2411 SymbolTableBaseSection *InX::DynSymTab; 2412 InputSection *InX::Interp; 2413 GdbIndexSection *InX::GdbIndex; 2414 GotSection *InX::Got; 2415 GotPltSection *InX::GotPlt; 2416 GnuHashTableSection *InX::GnuHashTab; 2417 HashTableSection *InX::HashTab; 2418 IgotPltSection *InX::IgotPlt; 2419 MipsGotSection *InX::MipsGot; 2420 MipsRldMapSection *InX::MipsRldMap; 2421 PltSection *InX::Plt; 2422 PltSection *InX::Iplt; 2423 StringTableSection *InX::ShStrTab; 2424 StringTableSection *InX::StrTab; 2425 SymbolTableBaseSection *InX::SymTab; 2426 2427 template GdbIndexSection *elf::createGdbIndex<ELF32LE>(); 2428 template GdbIndexSection *elf::createGdbIndex<ELF32BE>(); 2429 template GdbIndexSection *elf::createGdbIndex<ELF64LE>(); 2430 template GdbIndexSection *elf::createGdbIndex<ELF64BE>(); 2431 2432 template void PltSection::addEntry<ELF32LE>(SymbolBody &Sym); 2433 template void PltSection::addEntry<ELF32BE>(SymbolBody &Sym); 2434 template void PltSection::addEntry<ELF64LE>(SymbolBody &Sym); 2435 template void PltSection::addEntry<ELF64BE>(SymbolBody &Sym); 2436 2437 template void elf::createCommonSections<ELF32LE>(); 2438 template void elf::createCommonSections<ELF32BE>(); 2439 template void elf::createCommonSections<ELF64LE>(); 2440 template void elf::createCommonSections<ELF64BE>(); 2441 2442 template MergeInputSection *elf::createCommentSection<ELF32LE>(); 2443 template MergeInputSection *elf::createCommentSection<ELF32BE>(); 2444 template MergeInputSection *elf::createCommentSection<ELF64LE>(); 2445 template MergeInputSection *elf::createCommentSection<ELF64BE>(); 2446 2447 template class elf::MipsAbiFlagsSection<ELF32LE>; 2448 template class elf::MipsAbiFlagsSection<ELF32BE>; 2449 template class elf::MipsAbiFlagsSection<ELF64LE>; 2450 template class elf::MipsAbiFlagsSection<ELF64BE>; 2451 2452 template class elf::MipsOptionsSection<ELF32LE>; 2453 template class elf::MipsOptionsSection<ELF32BE>; 2454 template class elf::MipsOptionsSection<ELF64LE>; 2455 template class elf::MipsOptionsSection<ELF64BE>; 2456 2457 template class elf::MipsReginfoSection<ELF32LE>; 2458 template class elf::MipsReginfoSection<ELF32BE>; 2459 template class elf::MipsReginfoSection<ELF64LE>; 2460 template class elf::MipsReginfoSection<ELF64BE>; 2461 2462 template class elf::DynamicSection<ELF32LE>; 2463 template class elf::DynamicSection<ELF32BE>; 2464 template class elf::DynamicSection<ELF64LE>; 2465 template class elf::DynamicSection<ELF64BE>; 2466 2467 template class elf::RelocationSection<ELF32LE>; 2468 template class elf::RelocationSection<ELF32BE>; 2469 template class elf::RelocationSection<ELF64LE>; 2470 template class elf::RelocationSection<ELF64BE>; 2471 2472 template class elf::SymbolTableSection<ELF32LE>; 2473 template class elf::SymbolTableSection<ELF32BE>; 2474 template class elf::SymbolTableSection<ELF64LE>; 2475 template class elf::SymbolTableSection<ELF64BE>; 2476 2477 template class elf::EhFrameHeader<ELF32LE>; 2478 template class elf::EhFrameHeader<ELF32BE>; 2479 template class elf::EhFrameHeader<ELF64LE>; 2480 template class elf::EhFrameHeader<ELF64BE>; 2481 2482 template class elf::VersionTableSection<ELF32LE>; 2483 template class elf::VersionTableSection<ELF32BE>; 2484 template class elf::VersionTableSection<ELF64LE>; 2485 template class elf::VersionTableSection<ELF64BE>; 2486 2487 template class elf::VersionNeedSection<ELF32LE>; 2488 template class elf::VersionNeedSection<ELF32BE>; 2489 template class elf::VersionNeedSection<ELF64LE>; 2490 template class elf::VersionNeedSection<ELF64BE>; 2491 2492 template class elf::VersionDefinitionSection<ELF32LE>; 2493 template class elf::VersionDefinitionSection<ELF32BE>; 2494 template class elf::VersionDefinitionSection<ELF64LE>; 2495 template class elf::VersionDefinitionSection<ELF64BE>; 2496 2497 template class elf::EhFrameSection<ELF32LE>; 2498 template class elf::EhFrameSection<ELF32BE>; 2499 template class elf::EhFrameSection<ELF64LE>; 2500 template class elf::EhFrameSection<ELF64BE>; 2501