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 "Bits.h" 19 #include "Config.h" 20 #include "InputFiles.h" 21 #include "LinkerScript.h" 22 #include "OutputSections.h" 23 #include "SymbolTable.h" 24 #include "Symbols.h" 25 #include "Target.h" 26 #include "Writer.h" 27 #include "lld/Common/ErrorHandler.h" 28 #include "lld/Common/Memory.h" 29 #include "lld/Common/Strings.h" 30 #include "lld/Common/Threads.h" 31 #include "lld/Common/Version.h" 32 #include "llvm/ADT/SetOperations.h" 33 #include "llvm/BinaryFormat/Dwarf.h" 34 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h" 35 #include "llvm/Object/Decompressor.h" 36 #include "llvm/Object/ELFObjectFile.h" 37 #include "llvm/Support/Endian.h" 38 #include "llvm/Support/LEB128.h" 39 #include "llvm/Support/MD5.h" 40 #include "llvm/Support/RandomNumberGenerator.h" 41 #include "llvm/Support/SHA1.h" 42 #include "llvm/Support/xxhash.h" 43 #include <cstdlib> 44 #include <thread> 45 46 using namespace llvm; 47 using namespace llvm::dwarf; 48 using namespace llvm::ELF; 49 using namespace llvm::object; 50 using namespace llvm::support; 51 52 using namespace lld; 53 using namespace lld::elf; 54 55 using llvm::support::endian::read32le; 56 using llvm::support::endian::write32le; 57 using llvm::support::endian::write64le; 58 59 constexpr size_t MergeNoTailSection::NumShards; 60 61 // Returns an LLD version string. 62 static ArrayRef<uint8_t> getVersion() { 63 // Check LLD_VERSION first for ease of testing. 64 // You can get consistent output by using the environment variable. 65 // This is only for testing. 66 StringRef S = getenv("LLD_VERSION"); 67 if (S.empty()) 68 S = Saver.save(Twine("Linker: ") + getLLDVersion()); 69 70 // +1 to include the terminating '\0'. 71 return {(const uint8_t *)S.data(), S.size() + 1}; 72 } 73 74 // Creates a .comment section containing LLD version info. 75 // With this feature, you can identify LLD-generated binaries easily 76 // by "readelf --string-dump .comment <file>". 77 // The returned object is a mergeable string section. 78 MergeInputSection *elf::createCommentSection() { 79 return make<MergeInputSection>(SHF_MERGE | SHF_STRINGS, SHT_PROGBITS, 1, 80 getVersion(), ".comment"); 81 } 82 83 // .MIPS.abiflags section. 84 template <class ELFT> 85 MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags Flags) 86 : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags"), 87 Flags(Flags) { 88 this->Entsize = sizeof(Elf_Mips_ABIFlags); 89 } 90 91 template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *Buf) { 92 memcpy(Buf, &Flags, sizeof(Flags)); 93 } 94 95 template <class ELFT> 96 MipsAbiFlagsSection<ELFT> *MipsAbiFlagsSection<ELFT>::create() { 97 Elf_Mips_ABIFlags Flags = {}; 98 bool Create = false; 99 100 for (InputSectionBase *Sec : InputSections) { 101 if (Sec->Type != SHT_MIPS_ABIFLAGS) 102 continue; 103 Sec->Live = false; 104 Create = true; 105 106 std::string Filename = toString(Sec->File); 107 const size_t Size = Sec->Data.size(); 108 // Older version of BFD (such as the default FreeBSD linker) concatenate 109 // .MIPS.abiflags instead of merging. To allow for this case (or potential 110 // zero padding) we ignore everything after the first Elf_Mips_ABIFlags 111 if (Size < sizeof(Elf_Mips_ABIFlags)) { 112 error(Filename + ": invalid size of .MIPS.abiflags section: got " + 113 Twine(Size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags))); 114 return nullptr; 115 } 116 auto *S = reinterpret_cast<const Elf_Mips_ABIFlags *>(Sec->Data.data()); 117 if (S->version != 0) { 118 error(Filename + ": unexpected .MIPS.abiflags version " + 119 Twine(S->version)); 120 return nullptr; 121 } 122 123 // LLD checks ISA compatibility in calcMipsEFlags(). Here we just 124 // select the highest number of ISA/Rev/Ext. 125 Flags.isa_level = std::max(Flags.isa_level, S->isa_level); 126 Flags.isa_rev = std::max(Flags.isa_rev, S->isa_rev); 127 Flags.isa_ext = std::max(Flags.isa_ext, S->isa_ext); 128 Flags.gpr_size = std::max(Flags.gpr_size, S->gpr_size); 129 Flags.cpr1_size = std::max(Flags.cpr1_size, S->cpr1_size); 130 Flags.cpr2_size = std::max(Flags.cpr2_size, S->cpr2_size); 131 Flags.ases |= S->ases; 132 Flags.flags1 |= S->flags1; 133 Flags.flags2 |= S->flags2; 134 Flags.fp_abi = elf::getMipsFpAbiFlag(Flags.fp_abi, S->fp_abi, Filename); 135 }; 136 137 if (Create) 138 return make<MipsAbiFlagsSection<ELFT>>(Flags); 139 return nullptr; 140 } 141 142 // .MIPS.options section. 143 template <class ELFT> 144 MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo Reginfo) 145 : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"), 146 Reginfo(Reginfo) { 147 this->Entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo); 148 } 149 150 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *Buf) { 151 auto *Options = reinterpret_cast<Elf_Mips_Options *>(Buf); 152 Options->kind = ODK_REGINFO; 153 Options->size = getSize(); 154 155 if (!Config->Relocatable) 156 Reginfo.ri_gp_value = InX::MipsGot->getGp(); 157 memcpy(Buf + sizeof(Elf_Mips_Options), &Reginfo, sizeof(Reginfo)); 158 } 159 160 template <class ELFT> 161 MipsOptionsSection<ELFT> *MipsOptionsSection<ELFT>::create() { 162 // N64 ABI only. 163 if (!ELFT::Is64Bits) 164 return nullptr; 165 166 std::vector<InputSectionBase *> Sections; 167 for (InputSectionBase *Sec : InputSections) 168 if (Sec->Type == SHT_MIPS_OPTIONS) 169 Sections.push_back(Sec); 170 171 if (Sections.empty()) 172 return nullptr; 173 174 Elf_Mips_RegInfo Reginfo = {}; 175 for (InputSectionBase *Sec : Sections) { 176 Sec->Live = false; 177 178 std::string Filename = toString(Sec->File); 179 ArrayRef<uint8_t> D = Sec->Data; 180 181 while (!D.empty()) { 182 if (D.size() < sizeof(Elf_Mips_Options)) { 183 error(Filename + ": invalid size of .MIPS.options section"); 184 break; 185 } 186 187 auto *Opt = reinterpret_cast<const Elf_Mips_Options *>(D.data()); 188 if (Opt->kind == ODK_REGINFO) { 189 Reginfo.ri_gprmask |= Opt->getRegInfo().ri_gprmask; 190 Sec->getFile<ELFT>()->MipsGp0 = Opt->getRegInfo().ri_gp_value; 191 break; 192 } 193 194 if (!Opt->size) 195 fatal(Filename + ": zero option descriptor size"); 196 D = D.slice(Opt->size); 197 } 198 }; 199 200 return make<MipsOptionsSection<ELFT>>(Reginfo); 201 } 202 203 // MIPS .reginfo section. 204 template <class ELFT> 205 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo Reginfo) 206 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"), 207 Reginfo(Reginfo) { 208 this->Entsize = sizeof(Elf_Mips_RegInfo); 209 } 210 211 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *Buf) { 212 if (!Config->Relocatable) 213 Reginfo.ri_gp_value = InX::MipsGot->getGp(); 214 memcpy(Buf, &Reginfo, sizeof(Reginfo)); 215 } 216 217 template <class ELFT> 218 MipsReginfoSection<ELFT> *MipsReginfoSection<ELFT>::create() { 219 // Section should be alive for O32 and N32 ABIs only. 220 if (ELFT::Is64Bits) 221 return nullptr; 222 223 std::vector<InputSectionBase *> Sections; 224 for (InputSectionBase *Sec : InputSections) 225 if (Sec->Type == SHT_MIPS_REGINFO) 226 Sections.push_back(Sec); 227 228 if (Sections.empty()) 229 return nullptr; 230 231 Elf_Mips_RegInfo Reginfo = {}; 232 for (InputSectionBase *Sec : Sections) { 233 Sec->Live = false; 234 235 if (Sec->Data.size() != sizeof(Elf_Mips_RegInfo)) { 236 error(toString(Sec->File) + ": invalid size of .reginfo section"); 237 return nullptr; 238 } 239 240 auto *R = reinterpret_cast<const Elf_Mips_RegInfo *>(Sec->Data.data()); 241 Reginfo.ri_gprmask |= R->ri_gprmask; 242 Sec->getFile<ELFT>()->MipsGp0 = R->ri_gp_value; 243 }; 244 245 return make<MipsReginfoSection<ELFT>>(Reginfo); 246 } 247 248 InputSection *elf::createInterpSection() { 249 // StringSaver guarantees that the returned string ends with '\0'. 250 StringRef S = Saver.save(Config->DynamicLinker); 251 ArrayRef<uint8_t> Contents = {(const uint8_t *)S.data(), S.size() + 1}; 252 253 auto *Sec = make<InputSection>(nullptr, SHF_ALLOC, SHT_PROGBITS, 1, Contents, 254 ".interp"); 255 Sec->Live = true; 256 return Sec; 257 } 258 259 Defined *elf::addSyntheticLocal(StringRef Name, uint8_t Type, uint64_t Value, 260 uint64_t Size, InputSectionBase &Section) { 261 auto *S = make<Defined>(Section.File, Name, STB_LOCAL, STV_DEFAULT, Type, 262 Value, Size, &Section); 263 if (InX::SymTab) 264 InX::SymTab->addSymbol(S); 265 return S; 266 } 267 268 static size_t getHashSize() { 269 switch (Config->BuildId) { 270 case BuildIdKind::Fast: 271 return 8; 272 case BuildIdKind::Md5: 273 case BuildIdKind::Uuid: 274 return 16; 275 case BuildIdKind::Sha1: 276 return 20; 277 case BuildIdKind::Hexstring: 278 return Config->BuildIdVector.size(); 279 default: 280 llvm_unreachable("unknown BuildIdKind"); 281 } 282 } 283 284 BuildIdSection::BuildIdSection() 285 : SyntheticSection(SHF_ALLOC, SHT_NOTE, 4, ".note.gnu.build-id"), 286 HashSize(getHashSize()) {} 287 288 void BuildIdSection::writeTo(uint8_t *Buf) { 289 write32(Buf, 4); // Name size 290 write32(Buf + 4, HashSize); // Content size 291 write32(Buf + 8, NT_GNU_BUILD_ID); // Type 292 memcpy(Buf + 12, "GNU", 4); // Name string 293 HashBuf = Buf + 16; 294 } 295 296 // Split one uint8 array into small pieces of uint8 arrays. 297 static std::vector<ArrayRef<uint8_t>> split(ArrayRef<uint8_t> Arr, 298 size_t ChunkSize) { 299 std::vector<ArrayRef<uint8_t>> Ret; 300 while (Arr.size() > ChunkSize) { 301 Ret.push_back(Arr.take_front(ChunkSize)); 302 Arr = Arr.drop_front(ChunkSize); 303 } 304 if (!Arr.empty()) 305 Ret.push_back(Arr); 306 return Ret; 307 } 308 309 // Computes a hash value of Data using a given hash function. 310 // In order to utilize multiple cores, we first split data into 1MB 311 // chunks, compute a hash for each chunk, and then compute a hash value 312 // of the hash values. 313 void BuildIdSection::computeHash( 314 llvm::ArrayRef<uint8_t> Data, 315 std::function<void(uint8_t *Dest, ArrayRef<uint8_t> Arr)> HashFn) { 316 std::vector<ArrayRef<uint8_t>> Chunks = split(Data, 1024 * 1024); 317 std::vector<uint8_t> Hashes(Chunks.size() * HashSize); 318 319 // Compute hash values. 320 parallelForEachN(0, Chunks.size(), [&](size_t I) { 321 HashFn(Hashes.data() + I * HashSize, Chunks[I]); 322 }); 323 324 // Write to the final output buffer. 325 HashFn(HashBuf, Hashes); 326 } 327 328 BssSection::BssSection(StringRef Name, uint64_t Size, uint32_t Alignment) 329 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, Alignment, Name) { 330 this->Bss = true; 331 this->Size = Size; 332 } 333 334 void BuildIdSection::writeBuildId(ArrayRef<uint8_t> Buf) { 335 switch (Config->BuildId) { 336 case BuildIdKind::Fast: 337 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) { 338 write64le(Dest, xxHash64(toStringRef(Arr))); 339 }); 340 break; 341 case BuildIdKind::Md5: 342 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) { 343 memcpy(Dest, MD5::hash(Arr).data(), 16); 344 }); 345 break; 346 case BuildIdKind::Sha1: 347 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) { 348 memcpy(Dest, SHA1::hash(Arr).data(), 20); 349 }); 350 break; 351 case BuildIdKind::Uuid: 352 if (auto EC = getRandomBytes(HashBuf, HashSize)) 353 error("entropy source failure: " + EC.message()); 354 break; 355 case BuildIdKind::Hexstring: 356 memcpy(HashBuf, Config->BuildIdVector.data(), Config->BuildIdVector.size()); 357 break; 358 default: 359 llvm_unreachable("unknown BuildIdKind"); 360 } 361 } 362 363 EhFrameSection::EhFrameSection() 364 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame") {} 365 366 // Search for an existing CIE record or create a new one. 367 // CIE records from input object files are uniquified by their contents 368 // and where their relocations point to. 369 template <class ELFT, class RelTy> 370 CieRecord *EhFrameSection::addCie(EhSectionPiece &Cie, ArrayRef<RelTy> Rels) { 371 Symbol *Personality = nullptr; 372 unsigned FirstRelI = Cie.FirstRelocation; 373 if (FirstRelI != (unsigned)-1) 374 Personality = 375 &Cie.Sec->template getFile<ELFT>()->getRelocTargetSym(Rels[FirstRelI]); 376 377 // Search for an existing CIE by CIE contents/relocation target pair. 378 CieRecord *&Rec = CieMap[{Cie.data(), Personality}]; 379 380 // If not found, create a new one. 381 if (!Rec) { 382 Rec = make<CieRecord>(); 383 Rec->Cie = &Cie; 384 CieRecords.push_back(Rec); 385 } 386 return Rec; 387 } 388 389 // There is one FDE per function. Returns true if a given FDE 390 // points to a live function. 391 template <class ELFT, class RelTy> 392 bool EhFrameSection::isFdeLive(EhSectionPiece &Fde, ArrayRef<RelTy> Rels) { 393 auto *Sec = cast<EhInputSection>(Fde.Sec); 394 unsigned FirstRelI = Fde.FirstRelocation; 395 396 // An FDE should point to some function because FDEs are to describe 397 // functions. That's however not always the case due to an issue of 398 // ld.gold with -r. ld.gold may discard only functions and leave their 399 // corresponding FDEs, which results in creating bad .eh_frame sections. 400 // To deal with that, we ignore such FDEs. 401 if (FirstRelI == (unsigned)-1) 402 return false; 403 404 const RelTy &Rel = Rels[FirstRelI]; 405 Symbol &B = Sec->template getFile<ELFT>()->getRelocTargetSym(Rel); 406 407 // FDEs for garbage-collected or merged-by-ICF sections are dead. 408 if (auto *D = dyn_cast<Defined>(&B)) 409 if (SectionBase *Sec = D->Section) 410 return Sec->Live; 411 return false; 412 } 413 414 // .eh_frame is a sequence of CIE or FDE records. In general, there 415 // is one CIE record per input object file which is followed by 416 // a list of FDEs. This function searches an existing CIE or create a new 417 // one and associates FDEs to the CIE. 418 template <class ELFT, class RelTy> 419 void EhFrameSection::addSectionAux(EhInputSection *Sec, ArrayRef<RelTy> Rels) { 420 OffsetToCie.clear(); 421 for (EhSectionPiece &Piece : Sec->Pieces) { 422 // The empty record is the end marker. 423 if (Piece.Size == 4) 424 return; 425 426 size_t Offset = Piece.InputOff; 427 uint32_t ID = read32(Piece.data().data() + 4); 428 if (ID == 0) { 429 OffsetToCie[Offset] = addCie<ELFT>(Piece, Rels); 430 continue; 431 } 432 433 uint32_t CieOffset = Offset + 4 - ID; 434 CieRecord *Rec = OffsetToCie[CieOffset]; 435 if (!Rec) 436 fatal(toString(Sec) + ": invalid CIE reference"); 437 438 if (!isFdeLive<ELFT>(Piece, Rels)) 439 continue; 440 Rec->Fdes.push_back(&Piece); 441 NumFdes++; 442 } 443 } 444 445 template <class ELFT> void EhFrameSection::addSection(InputSectionBase *C) { 446 auto *Sec = cast<EhInputSection>(C); 447 Sec->Parent = this; 448 449 Alignment = std::max(Alignment, Sec->Alignment); 450 Sections.push_back(Sec); 451 452 for (auto *DS : Sec->DependentSections) 453 DependentSections.push_back(DS); 454 455 if (Sec->Pieces.empty()) 456 return; 457 458 if (Sec->AreRelocsRela) 459 addSectionAux<ELFT>(Sec, Sec->template relas<ELFT>()); 460 else 461 addSectionAux<ELFT>(Sec, Sec->template rels<ELFT>()); 462 } 463 464 static void writeCieFde(uint8_t *Buf, ArrayRef<uint8_t> D) { 465 memcpy(Buf, D.data(), D.size()); 466 467 size_t Aligned = alignTo(D.size(), Config->Wordsize); 468 469 // Zero-clear trailing padding if it exists. 470 memset(Buf + D.size(), 0, Aligned - D.size()); 471 472 // Fix the size field. -4 since size does not include the size field itself. 473 write32(Buf, Aligned - 4); 474 } 475 476 void EhFrameSection::finalizeContents() { 477 assert(!this->Size); // Not finalized. 478 size_t Off = 0; 479 for (CieRecord *Rec : CieRecords) { 480 Rec->Cie->OutputOff = Off; 481 Off += alignTo(Rec->Cie->Size, Config->Wordsize); 482 483 for (EhSectionPiece *Fde : Rec->Fdes) { 484 Fde->OutputOff = Off; 485 Off += alignTo(Fde->Size, Config->Wordsize); 486 } 487 } 488 489 // The LSB standard does not allow a .eh_frame section with zero 490 // Call Frame Information records. glibc unwind-dw2-fde.c 491 // classify_object_over_fdes expects there is a CIE record length 0 as a 492 // terminator. Thus we add one unconditionally. 493 Off += 4; 494 495 this->Size = Off; 496 } 497 498 // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table 499 // to get an FDE from an address to which FDE is applied. This function 500 // returns a list of such pairs. 501 std::vector<EhFrameSection::FdeData> EhFrameSection::getFdeData() const { 502 uint8_t *Buf = getParent()->Loc + OutSecOff; 503 std::vector<FdeData> Ret; 504 505 for (CieRecord *Rec : CieRecords) { 506 uint8_t Enc = getFdeEncoding(Rec->Cie); 507 for (EhSectionPiece *Fde : Rec->Fdes) { 508 uint64_t Pc = getFdePc(Buf, Fde->OutputOff, Enc); 509 if (Pc > UINT32_MAX) 510 fatal(toString(Fde->Sec) + ": PC address is too large: " + Twine(Pc)); 511 uint32_t FdeVA = getParent()->Addr + Fde->OutputOff; 512 Ret.push_back({(uint32_t)Pc, FdeVA}); 513 } 514 } 515 return Ret; 516 } 517 518 static uint64_t readFdeAddr(uint8_t *Buf, int Size) { 519 switch (Size) { 520 case DW_EH_PE_udata2: 521 return read16(Buf); 522 case DW_EH_PE_udata4: 523 return read32(Buf); 524 case DW_EH_PE_udata8: 525 return read64(Buf); 526 case DW_EH_PE_absptr: 527 return readUint(Buf); 528 } 529 fatal("unknown FDE size encoding"); 530 } 531 532 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to. 533 // We need it to create .eh_frame_hdr section. 534 uint64_t EhFrameSection::getFdePc(uint8_t *Buf, size_t FdeOff, 535 uint8_t Enc) const { 536 // The starting address to which this FDE applies is 537 // stored at FDE + 8 byte. 538 size_t Off = FdeOff + 8; 539 uint64_t Addr = readFdeAddr(Buf + Off, Enc & 0x7); 540 if ((Enc & 0x70) == DW_EH_PE_absptr) 541 return Addr; 542 if ((Enc & 0x70) == DW_EH_PE_pcrel) 543 return Addr + getParent()->Addr + Off; 544 fatal("unknown FDE size relative encoding"); 545 } 546 547 void EhFrameSection::writeTo(uint8_t *Buf) { 548 // Write CIE and FDE records. 549 for (CieRecord *Rec : CieRecords) { 550 size_t CieOffset = Rec->Cie->OutputOff; 551 writeCieFde(Buf + CieOffset, Rec->Cie->data()); 552 553 for (EhSectionPiece *Fde : Rec->Fdes) { 554 size_t Off = Fde->OutputOff; 555 writeCieFde(Buf + Off, Fde->data()); 556 557 // FDE's second word should have the offset to an associated CIE. 558 // Write it. 559 write32(Buf + Off + 4, Off + 4 - CieOffset); 560 } 561 } 562 563 // Apply relocations. .eh_frame section contents are not contiguous 564 // in the output buffer, but relocateAlloc() still works because 565 // getOffset() takes care of discontiguous section pieces. 566 for (EhInputSection *S : Sections) 567 S->relocateAlloc(Buf, nullptr); 568 } 569 570 GotSection::GotSection() 571 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 572 Target->GotEntrySize, ".got") { 573 // PPC64 saves the ElfSym::GlobalOffsetTable .TOC. as the first entry in the 574 // .got. If there are no references to .TOC. in the symbol table, 575 // ElfSym::GlobalOffsetTable will not be defined and we won't need to save 576 // .TOC. in the .got. When it is defined, we increase NumEntries by the number 577 // of entries used to emit ElfSym::GlobalOffsetTable. 578 if (ElfSym::GlobalOffsetTable && !Target->GotBaseSymInGotPlt) 579 NumEntries += Target->GotHeaderEntriesNum; 580 } 581 582 void GotSection::addEntry(Symbol &Sym) { 583 Sym.GotIndex = NumEntries; 584 ++NumEntries; 585 } 586 587 bool GotSection::addDynTlsEntry(Symbol &Sym) { 588 if (Sym.GlobalDynIndex != -1U) 589 return false; 590 Sym.GlobalDynIndex = NumEntries; 591 // Global Dynamic TLS entries take two GOT slots. 592 NumEntries += 2; 593 return true; 594 } 595 596 // Reserves TLS entries for a TLS module ID and a TLS block offset. 597 // In total it takes two GOT slots. 598 bool GotSection::addTlsIndex() { 599 if (TlsIndexOff != uint32_t(-1)) 600 return false; 601 TlsIndexOff = NumEntries * Config->Wordsize; 602 NumEntries += 2; 603 return true; 604 } 605 606 uint64_t GotSection::getGlobalDynAddr(const Symbol &B) const { 607 return this->getVA() + B.GlobalDynIndex * Config->Wordsize; 608 } 609 610 uint64_t GotSection::getGlobalDynOffset(const Symbol &B) const { 611 return B.GlobalDynIndex * Config->Wordsize; 612 } 613 614 void GotSection::finalizeContents() { 615 Size = NumEntries * Config->Wordsize; 616 } 617 618 bool GotSection::empty() const { 619 // We need to emit a GOT even if it's empty if there's a relocation that is 620 // relative to GOT(such as GOTOFFREL) or there's a symbol that points to a GOT 621 // (i.e. _GLOBAL_OFFSET_TABLE_) that the target defines relative to the .got. 622 return NumEntries == 0 && !HasGotOffRel && 623 !(ElfSym::GlobalOffsetTable && !Target->GotBaseSymInGotPlt); 624 } 625 626 void GotSection::writeTo(uint8_t *Buf) { 627 // Buf points to the start of this section's buffer, 628 // whereas InputSectionBase::relocateAlloc() expects its argument 629 // to point to the start of the output section. 630 Target->writeGotHeader(Buf); 631 relocateAlloc(Buf - OutSecOff, Buf - OutSecOff + Size); 632 } 633 634 static uint64_t getMipsPageAddr(uint64_t Addr) { 635 return (Addr + 0x8000) & ~0xffff; 636 } 637 638 static uint64_t getMipsPageCount(uint64_t Size) { 639 return (Size + 0xfffe) / 0xffff + 1; 640 } 641 642 MipsGotSection::MipsGotSection() 643 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16, 644 ".got") {} 645 646 void MipsGotSection::addEntry(InputFile &File, Symbol &Sym, int64_t Addend, 647 RelExpr Expr) { 648 FileGot &G = getGot(File); 649 if (Expr == R_MIPS_GOT_LOCAL_PAGE) { 650 if (const OutputSection *OS = Sym.getOutputSection()) 651 G.PagesMap.insert({OS, {}}); 652 else 653 G.Local16.insert({{nullptr, getMipsPageAddr(Sym.getVA(Addend))}, 0}); 654 } else if (Sym.isTls()) 655 G.Tls.insert({&Sym, 0}); 656 else if (Sym.IsPreemptible && Expr == R_ABS) 657 G.Relocs.insert({&Sym, 0}); 658 else if (Sym.IsPreemptible) 659 G.Global.insert({&Sym, 0}); 660 else if (Expr == R_MIPS_GOT_OFF32) 661 G.Local32.insert({{&Sym, Addend}, 0}); 662 else 663 G.Local16.insert({{&Sym, Addend}, 0}); 664 } 665 666 void MipsGotSection::addDynTlsEntry(InputFile &File, Symbol &Sym) { 667 getGot(File).DynTlsSymbols.insert({&Sym, 0}); 668 } 669 670 void MipsGotSection::addTlsIndex(InputFile &File) { 671 getGot(File).DynTlsSymbols.insert({nullptr, 0}); 672 } 673 674 size_t MipsGotSection::FileGot::getEntriesNum() const { 675 return getPageEntriesNum() + Local16.size() + Global.size() + Relocs.size() + 676 Tls.size() + DynTlsSymbols.size() * 2; 677 } 678 679 size_t MipsGotSection::FileGot::getPageEntriesNum() const { 680 size_t Num = 0; 681 for (const std::pair<const OutputSection *, FileGot::PageBlock> &P : PagesMap) 682 Num += P.second.Count; 683 return Num; 684 } 685 686 size_t MipsGotSection::FileGot::getIndexedEntriesNum() const { 687 size_t Count = getPageEntriesNum() + Local16.size() + Global.size(); 688 // If there are relocation-only entries in the GOT, TLS entries 689 // are allocated after them. TLS entries should be addressable 690 // by 16-bit index so count both reloc-only and TLS entries. 691 if (!Tls.empty() || !DynTlsSymbols.empty()) 692 Count += Relocs.size() + Tls.size() + DynTlsSymbols.size() * 2; 693 return Count; 694 } 695 696 MipsGotSection::FileGot &MipsGotSection::getGot(InputFile &F) { 697 if (!F.MipsGotIndex.hasValue()) { 698 Gots.emplace_back(); 699 Gots.back().File = &F; 700 F.MipsGotIndex = Gots.size() - 1; 701 } 702 return Gots[*F.MipsGotIndex]; 703 } 704 705 uint64_t MipsGotSection::getPageEntryOffset(const InputFile *F, 706 const Symbol &Sym, 707 int64_t Addend) const { 708 const FileGot &G = Gots[*F->MipsGotIndex]; 709 uint64_t Index = 0; 710 if (const OutputSection *OutSec = Sym.getOutputSection()) { 711 uint64_t SecAddr = getMipsPageAddr(OutSec->Addr); 712 uint64_t SymAddr = getMipsPageAddr(Sym.getVA(Addend)); 713 Index = G.PagesMap.lookup(OutSec).FirstIndex + (SymAddr - SecAddr) / 0xffff; 714 } else { 715 Index = G.Local16.lookup({nullptr, getMipsPageAddr(Sym.getVA(Addend))}); 716 } 717 return Index * Config->Wordsize; 718 } 719 720 uint64_t MipsGotSection::getSymEntryOffset(const InputFile *F, const Symbol &S, 721 int64_t Addend) const { 722 const FileGot &G = Gots[*F->MipsGotIndex]; 723 Symbol *Sym = const_cast<Symbol *>(&S); 724 if (Sym->isTls()) 725 return G.Tls.lookup(Sym) * Config->Wordsize; 726 if (Sym->IsPreemptible) 727 return G.Global.lookup(Sym) * Config->Wordsize; 728 return G.Local16.lookup({Sym, Addend}) * Config->Wordsize; 729 } 730 731 uint64_t MipsGotSection::getTlsIndexOffset(const InputFile *F) const { 732 const FileGot &G = Gots[*F->MipsGotIndex]; 733 return G.DynTlsSymbols.lookup(nullptr) * Config->Wordsize; 734 } 735 736 uint64_t MipsGotSection::getGlobalDynOffset(const InputFile *F, 737 const Symbol &S) const { 738 const FileGot &G = Gots[*F->MipsGotIndex]; 739 Symbol *Sym = const_cast<Symbol *>(&S); 740 return G.DynTlsSymbols.lookup(Sym) * Config->Wordsize; 741 } 742 743 const Symbol *MipsGotSection::getFirstGlobalEntry() const { 744 if (Gots.empty()) 745 return nullptr; 746 const FileGot &PrimGot = Gots.front(); 747 if (!PrimGot.Global.empty()) 748 return PrimGot.Global.front().first; 749 if (!PrimGot.Relocs.empty()) 750 return PrimGot.Relocs.front().first; 751 return nullptr; 752 } 753 754 unsigned MipsGotSection::getLocalEntriesNum() const { 755 if (Gots.empty()) 756 return HeaderEntriesNum; 757 return HeaderEntriesNum + Gots.front().getPageEntriesNum() + 758 Gots.front().Local16.size(); 759 } 760 761 bool MipsGotSection::tryMergeGots(FileGot &Dst, FileGot &Src, bool IsPrimary) { 762 FileGot Tmp = Dst; 763 set_union(Tmp.PagesMap, Src.PagesMap); 764 set_union(Tmp.Local16, Src.Local16); 765 set_union(Tmp.Global, Src.Global); 766 set_union(Tmp.Relocs, Src.Relocs); 767 set_union(Tmp.Tls, Src.Tls); 768 set_union(Tmp.DynTlsSymbols, Src.DynTlsSymbols); 769 770 size_t Count = IsPrimary ? HeaderEntriesNum : 0; 771 Count += Tmp.getIndexedEntriesNum(); 772 773 if (Count * Config->Wordsize > Config->MipsGotSize) 774 return false; 775 776 std::swap(Tmp, Dst); 777 return true; 778 } 779 780 void MipsGotSection::finalizeContents() { updateAllocSize(); } 781 782 bool MipsGotSection::updateAllocSize() { 783 Size = HeaderEntriesNum * Config->Wordsize; 784 for (const FileGot &G : Gots) 785 Size += G.getEntriesNum() * Config->Wordsize; 786 return false; 787 } 788 789 template <class ELFT> void MipsGotSection::build() { 790 if (Gots.empty()) 791 return; 792 793 std::vector<FileGot> MergedGots(1); 794 795 // For each GOT move non-preemptible symbols from the `Global` 796 // to `Local16` list. Preemptible symbol might become non-preemptible 797 // one if, for example, it gets a related copy relocation. 798 for (FileGot &Got : Gots) { 799 for (auto &P: Got.Global) 800 if (!P.first->IsPreemptible) 801 Got.Local16.insert({{P.first, 0}, 0}); 802 Got.Global.remove_if([&](const std::pair<Symbol *, size_t> &P) { 803 return !P.first->IsPreemptible; 804 }); 805 } 806 807 // For each GOT remove "reloc-only" entry if there is "global" 808 // entry for the same symbol. And add local entries which indexed 809 // using 32-bit value at the end of 16-bit entries. 810 for (FileGot &Got : Gots) { 811 Got.Relocs.remove_if([&](const std::pair<Symbol *, size_t> &P) { 812 return Got.Global.count(P.first); 813 }); 814 set_union(Got.Local16, Got.Local32); 815 Got.Local32.clear(); 816 } 817 818 // Evaluate number of "reloc-only" entries in the resulting GOT. 819 // To do that put all unique "reloc-only" and "global" entries 820 // from all GOTs to the future primary GOT. 821 FileGot *PrimGot = &MergedGots.front(); 822 for (FileGot &Got : Gots) { 823 set_union(PrimGot->Relocs, Got.Global); 824 set_union(PrimGot->Relocs, Got.Relocs); 825 Got.Relocs.clear(); 826 } 827 828 // Evaluate number of "page" entries in each GOT. 829 for (FileGot &Got : Gots) { 830 for (std::pair<const OutputSection *, FileGot::PageBlock> &P : 831 Got.PagesMap) { 832 const OutputSection *OS = P.first; 833 uint64_t SecSize = 0; 834 for (BaseCommand *Cmd : OS->SectionCommands) { 835 if (auto *ISD = dyn_cast<InputSectionDescription>(Cmd)) 836 for (InputSection *IS : ISD->Sections) { 837 uint64_t Off = alignTo(SecSize, IS->Alignment); 838 SecSize = Off + IS->getSize(); 839 } 840 } 841 P.second.Count = getMipsPageCount(SecSize); 842 } 843 } 844 845 // Merge GOTs. Try to join as much as possible GOTs but do not exceed 846 // maximum GOT size. At first, try to fill the primary GOT because 847 // the primary GOT can be accessed in the most effective way. If it 848 // is not possible, try to fill the last GOT in the list, and finally 849 // create a new GOT if both attempts failed. 850 for (FileGot &SrcGot : Gots) { 851 InputFile *File = SrcGot.File; 852 if (tryMergeGots(MergedGots.front(), SrcGot, true)) { 853 File->MipsGotIndex = 0; 854 } else { 855 if (!tryMergeGots(MergedGots.back(), SrcGot, false)) { 856 MergedGots.emplace_back(); 857 std::swap(MergedGots.back(), SrcGot); 858 } 859 File->MipsGotIndex = MergedGots.size() - 1; 860 } 861 } 862 std::swap(Gots, MergedGots); 863 864 // Reduce number of "reloc-only" entries in the primary GOT 865 // by substracting "global" entries exist in the primary GOT. 866 PrimGot = &Gots.front(); 867 PrimGot->Relocs.remove_if([&](const std::pair<Symbol *, size_t> &P) { 868 return PrimGot->Global.count(P.first); 869 }); 870 871 // Calculate indexes for each GOT entry. 872 size_t Index = HeaderEntriesNum; 873 for (FileGot &Got : Gots) { 874 Got.StartIndex = &Got == PrimGot ? 0 : Index; 875 for (std::pair<const OutputSection *, FileGot::PageBlock> &P : 876 Got.PagesMap) { 877 // For each output section referenced by GOT page relocations calculate 878 // and save into PagesMap an upper bound of MIPS GOT entries required 879 // to store page addresses of local symbols. We assume the worst case - 880 // each 64kb page of the output section has at least one GOT relocation 881 // against it. And take in account the case when the section intersects 882 // page boundaries. 883 P.second.FirstIndex = Index; 884 Index += P.second.Count; 885 } 886 for (auto &P: Got.Local16) 887 P.second = Index++; 888 for (auto &P: Got.Global) 889 P.second = Index++; 890 for (auto &P: Got.Relocs) 891 P.second = Index++; 892 for (auto &P: Got.Tls) 893 P.second = Index++; 894 for (auto &P: Got.DynTlsSymbols) { 895 P.second = Index; 896 Index += 2; 897 } 898 } 899 900 // Update Symbol::GotIndex field to use this 901 // value later in the `sortMipsSymbols` function. 902 for (auto &P : PrimGot->Global) 903 P.first->GotIndex = P.second; 904 for (auto &P : PrimGot->Relocs) 905 P.first->GotIndex = P.second; 906 907 // Create dynamic relocations. 908 for (FileGot &Got : Gots) { 909 // Create dynamic relocations for TLS entries. 910 for (std::pair<Symbol *, size_t> &P : Got.Tls) { 911 Symbol *S = P.first; 912 uint64_t Offset = P.second * Config->Wordsize; 913 if (S->IsPreemptible) 914 InX::RelaDyn->addReloc(Target->TlsGotRel, this, Offset, S); 915 } 916 for (std::pair<Symbol *, size_t> &P : Got.DynTlsSymbols) { 917 Symbol *S = P.first; 918 uint64_t Offset = P.second * Config->Wordsize; 919 if (S == nullptr) { 920 if (!Config->Pic) 921 continue; 922 InX::RelaDyn->addReloc(Target->TlsModuleIndexRel, this, Offset, S); 923 } else { 924 // When building a shared library we still need a dynamic relocation 925 // for the module index. Therefore only checking for 926 // S->IsPreemptible is not sufficient (this happens e.g. for 927 // thread-locals that have been marked as local through a linker script) 928 if (!S->IsPreemptible && !Config->Pic) 929 continue; 930 InX::RelaDyn->addReloc(Target->TlsModuleIndexRel, this, Offset, S); 931 // However, we can skip writing the TLS offset reloc for non-preemptible 932 // symbols since it is known even in shared libraries 933 if (!S->IsPreemptible) 934 continue; 935 Offset += Config->Wordsize; 936 InX::RelaDyn->addReloc(Target->TlsOffsetRel, this, Offset, S); 937 } 938 } 939 940 // Do not create dynamic relocations for non-TLS 941 // entries in the primary GOT. 942 if (&Got == PrimGot) 943 continue; 944 945 // Dynamic relocations for "global" entries. 946 for (const std::pair<Symbol *, size_t> &P : Got.Global) { 947 uint64_t Offset = P.second * Config->Wordsize; 948 InX::RelaDyn->addReloc(Target->RelativeRel, this, Offset, P.first); 949 } 950 if (!Config->Pic) 951 continue; 952 // Dynamic relocations for "local" entries in case of PIC. 953 for (const std::pair<const OutputSection *, FileGot::PageBlock> &L : 954 Got.PagesMap) { 955 size_t PageCount = L.second.Count; 956 for (size_t PI = 0; PI < PageCount; ++PI) { 957 uint64_t Offset = (L.second.FirstIndex + PI) * Config->Wordsize; 958 InX::RelaDyn->addReloc({Target->RelativeRel, this, Offset, L.first, 959 int64_t(PI * 0x10000)}); 960 } 961 } 962 for (const std::pair<GotEntry, size_t> &P : Got.Local16) { 963 uint64_t Offset = P.second * Config->Wordsize; 964 InX::RelaDyn->addReloc({Target->RelativeRel, this, Offset, true, 965 P.first.first, P.first.second}); 966 } 967 } 968 } 969 970 bool MipsGotSection::empty() const { 971 // We add the .got section to the result for dynamic MIPS target because 972 // its address and properties are mentioned in the .dynamic section. 973 return Config->Relocatable; 974 } 975 976 uint64_t MipsGotSection::getGp(const InputFile *F) const { 977 // For files without related GOT or files refer a primary GOT 978 // returns "common" _gp value. For secondary GOTs calculate 979 // individual _gp values. 980 if (!F || !F->MipsGotIndex.hasValue() || *F->MipsGotIndex == 0) 981 return ElfSym::MipsGp->getVA(0); 982 return getVA() + Gots[*F->MipsGotIndex].StartIndex * Config->Wordsize + 983 0x7ff0; 984 } 985 986 void MipsGotSection::writeTo(uint8_t *Buf) { 987 // Set the MSB of the second GOT slot. This is not required by any 988 // MIPS ABI documentation, though. 989 // 990 // There is a comment in glibc saying that "The MSB of got[1] of a 991 // gnu object is set to identify gnu objects," and in GNU gold it 992 // says "the second entry will be used by some runtime loaders". 993 // But how this field is being used is unclear. 994 // 995 // We are not really willing to mimic other linkers behaviors 996 // without understanding why they do that, but because all files 997 // generated by GNU tools have this special GOT value, and because 998 // we've been doing this for years, it is probably a safe bet to 999 // keep doing this for now. We really need to revisit this to see 1000 // if we had to do this. 1001 writeUint(Buf + Config->Wordsize, (uint64_t)1 << (Config->Wordsize * 8 - 1)); 1002 for (const FileGot &G : Gots) { 1003 auto Write = [&](size_t I, const Symbol *S, int64_t A) { 1004 uint64_t VA = A; 1005 if (S) { 1006 VA = S->getVA(A); 1007 if (S->StOther & STO_MIPS_MICROMIPS) 1008 VA |= 1; 1009 } 1010 writeUint(Buf + I * Config->Wordsize, VA); 1011 }; 1012 // Write 'page address' entries to the local part of the GOT. 1013 for (const std::pair<const OutputSection *, FileGot::PageBlock> &L : 1014 G.PagesMap) { 1015 size_t PageCount = L.second.Count; 1016 uint64_t FirstPageAddr = getMipsPageAddr(L.first->Addr); 1017 for (size_t PI = 0; PI < PageCount; ++PI) 1018 Write(L.second.FirstIndex + PI, nullptr, FirstPageAddr + PI * 0x10000); 1019 } 1020 // Local, global, TLS, reloc-only entries. 1021 // If TLS entry has a corresponding dynamic relocations, leave it 1022 // initialized by zero. Write down adjusted TLS symbol's values otherwise. 1023 // To calculate the adjustments use offsets for thread-local storage. 1024 // https://www.linux-mips.org/wiki/NPTL 1025 for (const std::pair<GotEntry, size_t> &P : G.Local16) 1026 Write(P.second, P.first.first, P.first.second); 1027 // Write VA to the primary GOT only. For secondary GOTs that 1028 // will be done by REL32 dynamic relocations. 1029 if (&G == &Gots.front()) 1030 for (const std::pair<const Symbol *, size_t> &P : G.Global) 1031 Write(P.second, P.first, 0); 1032 for (const std::pair<Symbol *, size_t> &P : G.Relocs) 1033 Write(P.second, P.first, 0); 1034 for (const std::pair<Symbol *, size_t> &P : G.Tls) 1035 Write(P.second, P.first, P.first->IsPreemptible ? 0 : -0x7000); 1036 for (const std::pair<Symbol *, size_t> &P : G.DynTlsSymbols) { 1037 if (P.first == nullptr && !Config->Pic) 1038 Write(P.second, nullptr, 1); 1039 else if (P.first && !P.first->IsPreemptible) { 1040 // If we are emitting PIC code with relocations we mustn't write 1041 // anything to the GOT here. When using Elf_Rel relocations the value 1042 // one will be treated as an addend and will cause crashes at runtime 1043 if (!Config->Pic) 1044 Write(P.second, nullptr, 1); 1045 Write(P.second + 1, P.first, -0x8000); 1046 } 1047 } 1048 } 1049 } 1050 1051 // On PowerPC the .plt section is used to hold the table of function addresses 1052 // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss 1053 // section. I don't know why we have a BSS style type for the section but it is 1054 // consitent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI. 1055 GotPltSection::GotPltSection() 1056 : SyntheticSection(SHF_ALLOC | SHF_WRITE, 1057 Config->EMachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS, 1058 Target->GotPltEntrySize, 1059 Config->EMachine == EM_PPC64 ? ".plt" : ".got.plt") {} 1060 1061 void GotPltSection::addEntry(Symbol &Sym) { 1062 assert(Sym.PltIndex == Entries.size()); 1063 Entries.push_back(&Sym); 1064 } 1065 1066 size_t GotPltSection::getSize() const { 1067 return (Target->GotPltHeaderEntriesNum + Entries.size()) * 1068 Target->GotPltEntrySize; 1069 } 1070 1071 void GotPltSection::writeTo(uint8_t *Buf) { 1072 Target->writeGotPltHeader(Buf); 1073 Buf += Target->GotPltHeaderEntriesNum * Target->GotPltEntrySize; 1074 for (const Symbol *B : Entries) { 1075 Target->writeGotPlt(Buf, *B); 1076 Buf += Config->Wordsize; 1077 } 1078 } 1079 1080 bool GotPltSection::empty() const { 1081 // We need to emit a GOT.PLT even if it's empty if there's a symbol that 1082 // references the _GLOBAL_OFFSET_TABLE_ and the Target defines the symbol 1083 // relative to the .got.plt section. 1084 return Entries.empty() && 1085 !(ElfSym::GlobalOffsetTable && Target->GotBaseSymInGotPlt); 1086 } 1087 1088 static StringRef getIgotPltName() { 1089 // On ARM the IgotPltSection is part of the GotSection. 1090 if (Config->EMachine == EM_ARM) 1091 return ".got"; 1092 1093 // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection 1094 // needs to be named the same. 1095 if (Config->EMachine == EM_PPC64) 1096 return ".plt"; 1097 1098 return ".got.plt"; 1099 } 1100 1101 // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit 1102 // with the IgotPltSection. 1103 IgotPltSection::IgotPltSection() 1104 : SyntheticSection(SHF_ALLOC | SHF_WRITE, 1105 Config->EMachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS, 1106 Target->GotPltEntrySize, getIgotPltName()) {} 1107 1108 void IgotPltSection::addEntry(Symbol &Sym) { 1109 Sym.IsInIgot = true; 1110 assert(Sym.PltIndex == Entries.size()); 1111 Entries.push_back(&Sym); 1112 } 1113 1114 size_t IgotPltSection::getSize() const { 1115 return Entries.size() * Target->GotPltEntrySize; 1116 } 1117 1118 void IgotPltSection::writeTo(uint8_t *Buf) { 1119 for (const Symbol *B : Entries) { 1120 Target->writeIgotPlt(Buf, *B); 1121 Buf += Config->Wordsize; 1122 } 1123 } 1124 1125 StringTableSection::StringTableSection(StringRef Name, bool Dynamic) 1126 : SyntheticSection(Dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, Name), 1127 Dynamic(Dynamic) { 1128 // ELF string tables start with a NUL byte. 1129 addString(""); 1130 } 1131 1132 // Adds a string to the string table. If HashIt is true we hash and check for 1133 // duplicates. It is optional because the name of global symbols are already 1134 // uniqued and hashing them again has a big cost for a small value: uniquing 1135 // them with some other string that happens to be the same. 1136 unsigned StringTableSection::addString(StringRef S, bool HashIt) { 1137 if (HashIt) { 1138 auto R = StringMap.insert(std::make_pair(S, this->Size)); 1139 if (!R.second) 1140 return R.first->second; 1141 } 1142 unsigned Ret = this->Size; 1143 this->Size = this->Size + S.size() + 1; 1144 Strings.push_back(S); 1145 return Ret; 1146 } 1147 1148 void StringTableSection::writeTo(uint8_t *Buf) { 1149 for (StringRef S : Strings) { 1150 memcpy(Buf, S.data(), S.size()); 1151 Buf[S.size()] = '\0'; 1152 Buf += S.size() + 1; 1153 } 1154 } 1155 1156 // Returns the number of version definition entries. Because the first entry 1157 // is for the version definition itself, it is the number of versioned symbols 1158 // plus one. Note that we don't support multiple versions yet. 1159 static unsigned getVerDefNum() { return Config->VersionDefinitions.size() + 1; } 1160 1161 template <class ELFT> 1162 DynamicSection<ELFT>::DynamicSection() 1163 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, Config->Wordsize, 1164 ".dynamic") { 1165 this->Entsize = ELFT::Is64Bits ? 16 : 8; 1166 1167 // .dynamic section is not writable on MIPS and on Fuchsia OS 1168 // which passes -z rodynamic. 1169 // See "Special Section" in Chapter 4 in the following document: 1170 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 1171 if (Config->EMachine == EM_MIPS || Config->ZRodynamic) 1172 this->Flags = SHF_ALLOC; 1173 1174 // Add strings to .dynstr early so that .dynstr's size will be 1175 // fixed early. 1176 for (StringRef S : Config->FilterList) 1177 addInt(DT_FILTER, InX::DynStrTab->addString(S)); 1178 for (StringRef S : Config->AuxiliaryList) 1179 addInt(DT_AUXILIARY, InX::DynStrTab->addString(S)); 1180 1181 if (!Config->Rpath.empty()) 1182 addInt(Config->EnableNewDtags ? DT_RUNPATH : DT_RPATH, 1183 InX::DynStrTab->addString(Config->Rpath)); 1184 1185 for (InputFile *File : SharedFiles) { 1186 SharedFile<ELFT> *F = cast<SharedFile<ELFT>>(File); 1187 if (F->IsNeeded) 1188 addInt(DT_NEEDED, InX::DynStrTab->addString(F->SoName)); 1189 } 1190 if (!Config->SoName.empty()) 1191 addInt(DT_SONAME, InX::DynStrTab->addString(Config->SoName)); 1192 } 1193 1194 template <class ELFT> 1195 void DynamicSection<ELFT>::add(int32_t Tag, std::function<uint64_t()> Fn) { 1196 Entries.push_back({Tag, Fn}); 1197 } 1198 1199 template <class ELFT> 1200 void DynamicSection<ELFT>::addInt(int32_t Tag, uint64_t Val) { 1201 Entries.push_back({Tag, [=] { return Val; }}); 1202 } 1203 1204 template <class ELFT> 1205 void DynamicSection<ELFT>::addInSec(int32_t Tag, InputSection *Sec) { 1206 Entries.push_back({Tag, [=] { return Sec->getVA(0); }}); 1207 } 1208 1209 template <class ELFT> 1210 void DynamicSection<ELFT>::addInSecRelative(int32_t Tag, InputSection *Sec) { 1211 size_t TagOffset = Entries.size() * Entsize; 1212 Entries.push_back( 1213 {Tag, [=] { return Sec->getVA(0) - (getVA() + TagOffset); }}); 1214 } 1215 1216 template <class ELFT> 1217 void DynamicSection<ELFT>::addOutSec(int32_t Tag, OutputSection *Sec) { 1218 Entries.push_back({Tag, [=] { return Sec->Addr; }}); 1219 } 1220 1221 template <class ELFT> 1222 void DynamicSection<ELFT>::addSize(int32_t Tag, OutputSection *Sec) { 1223 Entries.push_back({Tag, [=] { return Sec->Size; }}); 1224 } 1225 1226 template <class ELFT> 1227 void DynamicSection<ELFT>::addSym(int32_t Tag, Symbol *Sym) { 1228 Entries.push_back({Tag, [=] { return Sym->getVA(); }}); 1229 } 1230 1231 // Add remaining entries to complete .dynamic contents. 1232 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() { 1233 if (this->Size) 1234 return; // Already finalized. 1235 1236 // Set DT_FLAGS and DT_FLAGS_1. 1237 uint32_t DtFlags = 0; 1238 uint32_t DtFlags1 = 0; 1239 if (Config->Bsymbolic) 1240 DtFlags |= DF_SYMBOLIC; 1241 if (Config->ZInitfirst) 1242 DtFlags1 |= DF_1_INITFIRST; 1243 if (Config->ZNodelete) 1244 DtFlags1 |= DF_1_NODELETE; 1245 if (Config->ZNodlopen) 1246 DtFlags1 |= DF_1_NOOPEN; 1247 if (Config->ZNow) { 1248 DtFlags |= DF_BIND_NOW; 1249 DtFlags1 |= DF_1_NOW; 1250 } 1251 if (Config->ZOrigin) { 1252 DtFlags |= DF_ORIGIN; 1253 DtFlags1 |= DF_1_ORIGIN; 1254 } 1255 if (!Config->ZText) 1256 DtFlags |= DF_TEXTREL; 1257 1258 if (DtFlags) 1259 addInt(DT_FLAGS, DtFlags); 1260 if (DtFlags1) 1261 addInt(DT_FLAGS_1, DtFlags1); 1262 1263 // DT_DEBUG is a pointer to debug informaion used by debuggers at runtime. We 1264 // need it for each process, so we don't write it for DSOs. The loader writes 1265 // the pointer into this entry. 1266 // 1267 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some 1268 // systems (currently only Fuchsia OS) provide other means to give the 1269 // debugger this information. Such systems may choose make .dynamic read-only. 1270 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG. 1271 if (!Config->Shared && !Config->Relocatable && !Config->ZRodynamic) 1272 addInt(DT_DEBUG, 0); 1273 1274 this->Link = InX::DynStrTab->getParent()->SectionIndex; 1275 if (!InX::RelaDyn->empty()) { 1276 addInSec(InX::RelaDyn->DynamicTag, InX::RelaDyn); 1277 addSize(InX::RelaDyn->SizeDynamicTag, InX::RelaDyn->getParent()); 1278 1279 bool IsRela = Config->IsRela; 1280 addInt(IsRela ? DT_RELAENT : DT_RELENT, 1281 IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel)); 1282 1283 // MIPS dynamic loader does not support RELCOUNT tag. 1284 // The problem is in the tight relation between dynamic 1285 // relocations and GOT. So do not emit this tag on MIPS. 1286 if (Config->EMachine != EM_MIPS) { 1287 size_t NumRelativeRels = InX::RelaDyn->getRelativeRelocCount(); 1288 if (Config->ZCombreloc && NumRelativeRels) 1289 addInt(IsRela ? DT_RELACOUNT : DT_RELCOUNT, NumRelativeRels); 1290 } 1291 } 1292 if (InX::RelrDyn && !InX::RelrDyn->Relocs.empty()) { 1293 addInSec(Config->UseAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR, 1294 InX::RelrDyn); 1295 addSize(Config->UseAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ, 1296 InX::RelrDyn->getParent()); 1297 addInt(Config->UseAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT, 1298 sizeof(Elf_Relr)); 1299 } 1300 // .rel[a].plt section usually consists of two parts, containing plt and 1301 // iplt relocations. It is possible to have only iplt relocations in the 1302 // output. In that case RelaPlt is empty and have zero offset, the same offset 1303 // as RelaIplt have. And we still want to emit proper dynamic tags for that 1304 // case, so here we always use RelaPlt as marker for the begining of 1305 // .rel[a].plt section. 1306 if (InX::RelaPlt->getParent()->Live) { 1307 addInSec(DT_JMPREL, InX::RelaPlt); 1308 addSize(DT_PLTRELSZ, InX::RelaPlt->getParent()); 1309 switch (Config->EMachine) { 1310 case EM_MIPS: 1311 addInSec(DT_MIPS_PLTGOT, InX::GotPlt); 1312 break; 1313 case EM_SPARCV9: 1314 addInSec(DT_PLTGOT, InX::Plt); 1315 break; 1316 default: 1317 addInSec(DT_PLTGOT, InX::GotPlt); 1318 break; 1319 } 1320 addInt(DT_PLTREL, Config->IsRela ? DT_RELA : DT_REL); 1321 } 1322 1323 addInSec(DT_SYMTAB, InX::DynSymTab); 1324 addInt(DT_SYMENT, sizeof(Elf_Sym)); 1325 addInSec(DT_STRTAB, InX::DynStrTab); 1326 addInt(DT_STRSZ, InX::DynStrTab->getSize()); 1327 if (!Config->ZText) 1328 addInt(DT_TEXTREL, 0); 1329 if (InX::GnuHashTab) 1330 addInSec(DT_GNU_HASH, InX::GnuHashTab); 1331 if (InX::HashTab) 1332 addInSec(DT_HASH, InX::HashTab); 1333 1334 if (Out::PreinitArray) { 1335 addOutSec(DT_PREINIT_ARRAY, Out::PreinitArray); 1336 addSize(DT_PREINIT_ARRAYSZ, Out::PreinitArray); 1337 } 1338 if (Out::InitArray) { 1339 addOutSec(DT_INIT_ARRAY, Out::InitArray); 1340 addSize(DT_INIT_ARRAYSZ, Out::InitArray); 1341 } 1342 if (Out::FiniArray) { 1343 addOutSec(DT_FINI_ARRAY, Out::FiniArray); 1344 addSize(DT_FINI_ARRAYSZ, Out::FiniArray); 1345 } 1346 1347 if (Symbol *B = Symtab->find(Config->Init)) 1348 if (B->isDefined()) 1349 addSym(DT_INIT, B); 1350 if (Symbol *B = Symtab->find(Config->Fini)) 1351 if (B->isDefined()) 1352 addSym(DT_FINI, B); 1353 1354 bool HasVerNeed = In<ELFT>::VerNeed->getNeedNum() != 0; 1355 if (HasVerNeed || In<ELFT>::VerDef) 1356 addInSec(DT_VERSYM, In<ELFT>::VerSym); 1357 if (In<ELFT>::VerDef) { 1358 addInSec(DT_VERDEF, In<ELFT>::VerDef); 1359 addInt(DT_VERDEFNUM, getVerDefNum()); 1360 } 1361 if (HasVerNeed) { 1362 addInSec(DT_VERNEED, In<ELFT>::VerNeed); 1363 addInt(DT_VERNEEDNUM, In<ELFT>::VerNeed->getNeedNum()); 1364 } 1365 1366 if (Config->EMachine == EM_MIPS) { 1367 addInt(DT_MIPS_RLD_VERSION, 1); 1368 addInt(DT_MIPS_FLAGS, RHF_NOTPOT); 1369 addInt(DT_MIPS_BASE_ADDRESS, Target->getImageBase()); 1370 addInt(DT_MIPS_SYMTABNO, InX::DynSymTab->getNumSymbols()); 1371 1372 add(DT_MIPS_LOCAL_GOTNO, [] { return InX::MipsGot->getLocalEntriesNum(); }); 1373 1374 if (const Symbol *B = InX::MipsGot->getFirstGlobalEntry()) 1375 addInt(DT_MIPS_GOTSYM, B->DynsymIndex); 1376 else 1377 addInt(DT_MIPS_GOTSYM, InX::DynSymTab->getNumSymbols()); 1378 addInSec(DT_PLTGOT, InX::MipsGot); 1379 if (InX::MipsRldMap) { 1380 if (!Config->Pie) 1381 addInSec(DT_MIPS_RLD_MAP, InX::MipsRldMap); 1382 // Store the offset to the .rld_map section 1383 // relative to the address of the tag. 1384 addInSecRelative(DT_MIPS_RLD_MAP_REL, InX::MipsRldMap); 1385 } 1386 } 1387 1388 // Glink dynamic tag is required by the V2 abi if the plt section isn't empty. 1389 if (Config->EMachine == EM_PPC64 && !InX::Plt->empty()) { 1390 // The Glink tag points to 32 bytes before the first lazy symbol resolution 1391 // stub, which starts directly after the header. 1392 Entries.push_back({DT_PPC64_GLINK, [=] { 1393 unsigned Offset = Target->PltHeaderSize - 32; 1394 return InX::Plt->getVA(0) + Offset; 1395 }}); 1396 } 1397 1398 addInt(DT_NULL, 0); 1399 1400 getParent()->Link = this->Link; 1401 this->Size = Entries.size() * this->Entsize; 1402 } 1403 1404 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *Buf) { 1405 auto *P = reinterpret_cast<Elf_Dyn *>(Buf); 1406 1407 for (std::pair<int32_t, std::function<uint64_t()>> &KV : Entries) { 1408 P->d_tag = KV.first; 1409 P->d_un.d_val = KV.second(); 1410 ++P; 1411 } 1412 } 1413 1414 uint64_t DynamicReloc::getOffset() const { 1415 return InputSec->getVA(OffsetInSec); 1416 } 1417 1418 int64_t DynamicReloc::computeAddend() const { 1419 if (UseSymVA) 1420 return Sym->getVA(Addend); 1421 if (!OutputSec) 1422 return Addend; 1423 // See the comment in the DynamicReloc ctor. 1424 return getMipsPageAddr(OutputSec->Addr) + Addend; 1425 } 1426 1427 uint32_t DynamicReloc::getSymIndex() const { 1428 if (Sym && !UseSymVA) 1429 return Sym->DynsymIndex; 1430 return 0; 1431 } 1432 1433 RelocationBaseSection::RelocationBaseSection(StringRef Name, uint32_t Type, 1434 int32_t DynamicTag, 1435 int32_t SizeDynamicTag) 1436 : SyntheticSection(SHF_ALLOC, Type, Config->Wordsize, Name), 1437 DynamicTag(DynamicTag), SizeDynamicTag(SizeDynamicTag) {} 1438 1439 void RelocationBaseSection::addReloc(RelType DynType, InputSectionBase *IS, 1440 uint64_t OffsetInSec, Symbol *Sym) { 1441 addReloc({DynType, IS, OffsetInSec, false, Sym, 0}); 1442 } 1443 1444 void RelocationBaseSection::addReloc(RelType DynType, 1445 InputSectionBase *InputSec, 1446 uint64_t OffsetInSec, Symbol *Sym, 1447 int64_t Addend, RelExpr Expr, 1448 RelType Type) { 1449 // Write the addends to the relocated address if required. We skip 1450 // it if the written value would be zero. 1451 if (Config->WriteAddends && (Expr != R_ADDEND || Addend != 0)) 1452 InputSec->Relocations.push_back({Expr, Type, OffsetInSec, Addend, Sym}); 1453 addReloc({DynType, InputSec, OffsetInSec, Expr != R_ADDEND, Sym, Addend}); 1454 } 1455 1456 void RelocationBaseSection::addReloc(const DynamicReloc &Reloc) { 1457 if (Reloc.Type == Target->RelativeRel) 1458 ++NumRelativeRelocs; 1459 Relocs.push_back(Reloc); 1460 } 1461 1462 void RelocationBaseSection::finalizeContents() { 1463 // If all relocations are R_*_RELATIVE they don't refer to any 1464 // dynamic symbol and we don't need a dynamic symbol table. If that 1465 // is the case, just use 0 as the link. 1466 Link = InX::DynSymTab ? InX::DynSymTab->getParent()->SectionIndex : 0; 1467 1468 // Set required output section properties. 1469 getParent()->Link = Link; 1470 } 1471 1472 RelrBaseSection::RelrBaseSection() 1473 : SyntheticSection(SHF_ALLOC, 1474 Config->UseAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR, 1475 Config->Wordsize, ".relr.dyn") {} 1476 1477 template <class ELFT> 1478 static void encodeDynamicReloc(typename ELFT::Rela *P, 1479 const DynamicReloc &Rel) { 1480 if (Config->IsRela) 1481 P->r_addend = Rel.computeAddend(); 1482 P->r_offset = Rel.getOffset(); 1483 P->setSymbolAndType(Rel.getSymIndex(), Rel.Type, Config->IsMips64EL); 1484 } 1485 1486 template <class ELFT> 1487 RelocationSection<ELFT>::RelocationSection(StringRef Name, bool Sort) 1488 : RelocationBaseSection(Name, Config->IsRela ? SHT_RELA : SHT_REL, 1489 Config->IsRela ? DT_RELA : DT_REL, 1490 Config->IsRela ? DT_RELASZ : DT_RELSZ), 1491 Sort(Sort) { 1492 this->Entsize = Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1493 } 1494 1495 static bool compRelocations(const DynamicReloc &A, const DynamicReloc &B) { 1496 bool AIsRel = A.Type == Target->RelativeRel; 1497 bool BIsRel = B.Type == Target->RelativeRel; 1498 if (AIsRel != BIsRel) 1499 return AIsRel; 1500 return A.getSymIndex() < B.getSymIndex(); 1501 } 1502 1503 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *Buf) { 1504 if (Sort) 1505 std::stable_sort(Relocs.begin(), Relocs.end(), compRelocations); 1506 1507 for (const DynamicReloc &Rel : Relocs) { 1508 encodeDynamicReloc<ELFT>(reinterpret_cast<Elf_Rela *>(Buf), Rel); 1509 Buf += Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1510 } 1511 } 1512 1513 template <class ELFT> unsigned RelocationSection<ELFT>::getRelocOffset() { 1514 return this->Entsize * Relocs.size(); 1515 } 1516 1517 template <class ELFT> 1518 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection( 1519 StringRef Name) 1520 : RelocationBaseSection( 1521 Name, Config->IsRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL, 1522 Config->IsRela ? DT_ANDROID_RELA : DT_ANDROID_REL, 1523 Config->IsRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ) { 1524 this->Entsize = 1; 1525 } 1526 1527 template <class ELFT> 1528 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() { 1529 // This function computes the contents of an Android-format packed relocation 1530 // section. 1531 // 1532 // This format compresses relocations by using relocation groups to factor out 1533 // fields that are common between relocations and storing deltas from previous 1534 // relocations in SLEB128 format (which has a short representation for small 1535 // numbers). A good example of a relocation type with common fields is 1536 // R_*_RELATIVE, which is normally used to represent function pointers in 1537 // vtables. In the REL format, each relative relocation has the same r_info 1538 // field, and is only different from other relative relocations in terms of 1539 // the r_offset field. By sorting relocations by offset, grouping them by 1540 // r_info and representing each relocation with only the delta from the 1541 // previous offset, each 8-byte relocation can be compressed to as little as 1 1542 // byte (or less with run-length encoding). This relocation packer was able to 1543 // reduce the size of the relocation section in an Android Chromium DSO from 1544 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size. 1545 // 1546 // A relocation section consists of a header containing the literal bytes 1547 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two 1548 // elements are the total number of relocations in the section and an initial 1549 // r_offset value. The remaining elements define a sequence of relocation 1550 // groups. Each relocation group starts with a header consisting of the 1551 // following elements: 1552 // 1553 // - the number of relocations in the relocation group 1554 // - flags for the relocation group 1555 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta 1556 // for each relocation in the group. 1557 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info 1558 // field for each relocation in the group. 1559 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and 1560 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for 1561 // each relocation in the group. 1562 // 1563 // Following the relocation group header are descriptions of each of the 1564 // relocations in the group. They consist of the following elements: 1565 // 1566 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset 1567 // delta for this relocation. 1568 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info 1569 // field for this relocation. 1570 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and 1571 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for 1572 // this relocation. 1573 1574 size_t OldSize = RelocData.size(); 1575 1576 RelocData = {'A', 'P', 'S', '2'}; 1577 raw_svector_ostream OS(RelocData); 1578 auto Add = [&](int64_t V) { encodeSLEB128(V, OS); }; 1579 1580 // The format header includes the number of relocations and the initial 1581 // offset (we set this to zero because the first relocation group will 1582 // perform the initial adjustment). 1583 Add(Relocs.size()); 1584 Add(0); 1585 1586 std::vector<Elf_Rela> Relatives, NonRelatives; 1587 1588 for (const DynamicReloc &Rel : Relocs) { 1589 Elf_Rela R; 1590 encodeDynamicReloc<ELFT>(&R, Rel); 1591 1592 if (R.getType(Config->IsMips64EL) == Target->RelativeRel) 1593 Relatives.push_back(R); 1594 else 1595 NonRelatives.push_back(R); 1596 } 1597 1598 llvm::sort(Relatives.begin(), Relatives.end(), 1599 [](const Elf_Rel &A, const Elf_Rel &B) { 1600 return A.r_offset < B.r_offset; 1601 }); 1602 1603 // Try to find groups of relative relocations which are spaced one word 1604 // apart from one another. These generally correspond to vtable entries. The 1605 // format allows these groups to be encoded using a sort of run-length 1606 // encoding, but each group will cost 7 bytes in addition to the offset from 1607 // the previous group, so it is only profitable to do this for groups of 1608 // size 8 or larger. 1609 std::vector<Elf_Rela> UngroupedRelatives; 1610 std::vector<std::vector<Elf_Rela>> RelativeGroups; 1611 for (auto I = Relatives.begin(), E = Relatives.end(); I != E;) { 1612 std::vector<Elf_Rela> Group; 1613 do { 1614 Group.push_back(*I++); 1615 } while (I != E && (I - 1)->r_offset + Config->Wordsize == I->r_offset); 1616 1617 if (Group.size() < 8) 1618 UngroupedRelatives.insert(UngroupedRelatives.end(), Group.begin(), 1619 Group.end()); 1620 else 1621 RelativeGroups.emplace_back(std::move(Group)); 1622 } 1623 1624 unsigned HasAddendIfRela = 1625 Config->IsRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0; 1626 1627 uint64_t Offset = 0; 1628 uint64_t Addend = 0; 1629 1630 // Emit the run-length encoding for the groups of adjacent relative 1631 // relocations. Each group is represented using two groups in the packed 1632 // format. The first is used to set the current offset to the start of the 1633 // group (and also encodes the first relocation), and the second encodes the 1634 // remaining relocations. 1635 for (std::vector<Elf_Rela> &G : RelativeGroups) { 1636 // The first relocation in the group. 1637 Add(1); 1638 Add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG | 1639 RELOCATION_GROUPED_BY_INFO_FLAG | HasAddendIfRela); 1640 Add(G[0].r_offset - Offset); 1641 Add(Target->RelativeRel); 1642 if (Config->IsRela) { 1643 Add(G[0].r_addend - Addend); 1644 Addend = G[0].r_addend; 1645 } 1646 1647 // The remaining relocations. 1648 Add(G.size() - 1); 1649 Add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG | 1650 RELOCATION_GROUPED_BY_INFO_FLAG | HasAddendIfRela); 1651 Add(Config->Wordsize); 1652 Add(Target->RelativeRel); 1653 if (Config->IsRela) { 1654 for (auto I = G.begin() + 1, E = G.end(); I != E; ++I) { 1655 Add(I->r_addend - Addend); 1656 Addend = I->r_addend; 1657 } 1658 } 1659 1660 Offset = G.back().r_offset; 1661 } 1662 1663 // Now the ungrouped relatives. 1664 if (!UngroupedRelatives.empty()) { 1665 Add(UngroupedRelatives.size()); 1666 Add(RELOCATION_GROUPED_BY_INFO_FLAG | HasAddendIfRela); 1667 Add(Target->RelativeRel); 1668 for (Elf_Rela &R : UngroupedRelatives) { 1669 Add(R.r_offset - Offset); 1670 Offset = R.r_offset; 1671 if (Config->IsRela) { 1672 Add(R.r_addend - Addend); 1673 Addend = R.r_addend; 1674 } 1675 } 1676 } 1677 1678 // Finally the non-relative relocations. 1679 llvm::sort(NonRelatives.begin(), NonRelatives.end(), 1680 [](const Elf_Rela &A, const Elf_Rela &B) { 1681 return A.r_offset < B.r_offset; 1682 }); 1683 if (!NonRelatives.empty()) { 1684 Add(NonRelatives.size()); 1685 Add(HasAddendIfRela); 1686 for (Elf_Rela &R : NonRelatives) { 1687 Add(R.r_offset - Offset); 1688 Offset = R.r_offset; 1689 Add(R.r_info); 1690 if (Config->IsRela) { 1691 Add(R.r_addend - Addend); 1692 Addend = R.r_addend; 1693 } 1694 } 1695 } 1696 1697 // Returns whether the section size changed. We need to keep recomputing both 1698 // section layout and the contents of this section until the size converges 1699 // because changing this section's size can affect section layout, which in 1700 // turn can affect the sizes of the LEB-encoded integers stored in this 1701 // section. 1702 return RelocData.size() != OldSize; 1703 } 1704 1705 template <class ELFT> RelrSection<ELFT>::RelrSection() { 1706 this->Entsize = Config->Wordsize; 1707 } 1708 1709 template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() { 1710 // This function computes the contents of an SHT_RELR packed relocation 1711 // section. 1712 // 1713 // Proposal for adding SHT_RELR sections to generic-abi is here: 1714 // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg 1715 // 1716 // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks 1717 // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ] 1718 // 1719 // i.e. start with an address, followed by any number of bitmaps. The address 1720 // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63 1721 // relocations each, at subsequent offsets following the last address entry. 1722 // 1723 // The bitmap entries must have 1 in the least significant bit. The assumption 1724 // here is that an address cannot have 1 in lsb. Odd addresses are not 1725 // supported. 1726 // 1727 // Excluding the least significant bit in the bitmap, each non-zero bit in 1728 // the bitmap represents a relocation to be applied to a corresponding machine 1729 // word that follows the base address word. The second least significant bit 1730 // represents the machine word immediately following the initial address, and 1731 // each bit that follows represents the next word, in linear order. As such, 1732 // a single bitmap can encode up to 31 relocations in a 32-bit object, and 1733 // 63 relocations in a 64-bit object. 1734 // 1735 // This encoding has a couple of interesting properties: 1736 // 1. Looking at any entry, it is clear whether it's an address or a bitmap: 1737 // even means address, odd means bitmap. 1738 // 2. Just a simple list of addresses is a valid encoding. 1739 1740 size_t OldSize = RelrRelocs.size(); 1741 RelrRelocs.clear(); 1742 1743 // Same as Config->Wordsize but faster because this is a compile-time 1744 // constant. 1745 const size_t Wordsize = sizeof(typename ELFT::uint); 1746 1747 // Number of bits to use for the relocation offsets bitmap. 1748 // Must be either 63 or 31. 1749 const size_t NBits = Wordsize * 8 - 1; 1750 1751 // Get offsets for all relative relocations and sort them. 1752 std::vector<uint64_t> Offsets; 1753 for (const RelativeReloc &Rel : Relocs) 1754 Offsets.push_back(Rel.getOffset()); 1755 llvm::sort(Offsets.begin(), Offsets.end()); 1756 1757 // For each leading relocation, find following ones that can be folded 1758 // as a bitmap and fold them. 1759 for (size_t I = 0, E = Offsets.size(); I < E;) { 1760 // Add a leading relocation. 1761 RelrRelocs.push_back(Elf_Relr(Offsets[I])); 1762 uint64_t Base = Offsets[I] + Wordsize; 1763 ++I; 1764 1765 // Find foldable relocations to construct bitmaps. 1766 while (I < E) { 1767 uint64_t Bitmap = 0; 1768 1769 while (I < E) { 1770 uint64_t Delta = Offsets[I] - Base; 1771 1772 // If it is too far, it cannot be folded. 1773 if (Delta >= NBits * Wordsize) 1774 break; 1775 1776 // If it is not a multiple of wordsize away, it cannot be folded. 1777 if (Delta % Wordsize) 1778 break; 1779 1780 // Fold it. 1781 Bitmap |= 1ULL << (Delta / Wordsize); 1782 ++I; 1783 } 1784 1785 if (!Bitmap) 1786 break; 1787 1788 RelrRelocs.push_back(Elf_Relr((Bitmap << 1) | 1)); 1789 Base += NBits * Wordsize; 1790 } 1791 } 1792 1793 return RelrRelocs.size() != OldSize; 1794 } 1795 1796 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &StrTabSec) 1797 : SyntheticSection(StrTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0, 1798 StrTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB, 1799 Config->Wordsize, 1800 StrTabSec.isDynamic() ? ".dynsym" : ".symtab"), 1801 StrTabSec(StrTabSec) {} 1802 1803 // Orders symbols according to their positions in the GOT, 1804 // in compliance with MIPS ABI rules. 1805 // See "Global Offset Table" in Chapter 5 in the following document 1806 // for detailed description: 1807 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 1808 static bool sortMipsSymbols(const SymbolTableEntry &L, 1809 const SymbolTableEntry &R) { 1810 // Sort entries related to non-local preemptible symbols by GOT indexes. 1811 // All other entries go to the beginning of a dynsym in arbitrary order. 1812 if (L.Sym->isInGot() && R.Sym->isInGot()) 1813 return L.Sym->GotIndex < R.Sym->GotIndex; 1814 if (!L.Sym->isInGot() && !R.Sym->isInGot()) 1815 return false; 1816 return !L.Sym->isInGot(); 1817 } 1818 1819 void SymbolTableBaseSection::finalizeContents() { 1820 getParent()->Link = StrTabSec.getParent()->SectionIndex; 1821 1822 if (this->Type != SHT_DYNSYM) 1823 return; 1824 1825 // If it is a .dynsym, there should be no local symbols, but we need 1826 // to do a few things for the dynamic linker. 1827 1828 // Section's Info field has the index of the first non-local symbol. 1829 // Because the first symbol entry is a null entry, 1 is the first. 1830 getParent()->Info = 1; 1831 1832 if (InX::GnuHashTab) { 1833 // NB: It also sorts Symbols to meet the GNU hash table requirements. 1834 InX::GnuHashTab->addSymbols(Symbols); 1835 } else if (Config->EMachine == EM_MIPS) { 1836 std::stable_sort(Symbols.begin(), Symbols.end(), sortMipsSymbols); 1837 } 1838 1839 size_t I = 0; 1840 for (const SymbolTableEntry &S : Symbols) 1841 S.Sym->DynsymIndex = ++I; 1842 } 1843 1844 // The ELF spec requires that all local symbols precede global symbols, so we 1845 // sort symbol entries in this function. (For .dynsym, we don't do that because 1846 // symbols for dynamic linking are inherently all globals.) 1847 // 1848 // Aside from above, we put local symbols in groups starting with the STT_FILE 1849 // symbol. That is convenient for purpose of identifying where are local symbols 1850 // coming from. 1851 void SymbolTableBaseSection::postThunkContents() { 1852 if (this->Type == SHT_DYNSYM) 1853 return; 1854 1855 // Move all local symbols before global symbols. 1856 auto E = std::stable_partition( 1857 Symbols.begin(), Symbols.end(), [](const SymbolTableEntry &S) { 1858 return S.Sym->isLocal() || S.Sym->computeBinding() == STB_LOCAL; 1859 }); 1860 size_t NumLocals = E - Symbols.begin(); 1861 getParent()->Info = NumLocals + 1; 1862 1863 // We want to group the local symbols by file. For that we rebuild the local 1864 // part of the symbols vector. We do not need to care about the STT_FILE 1865 // symbols, they are already naturally placed first in each group. That 1866 // happens because STT_FILE is always the first symbol in the object and hence 1867 // precede all other local symbols we add for a file. 1868 MapVector<InputFile *, std::vector<SymbolTableEntry>> Arr; 1869 for (const SymbolTableEntry &S : llvm::make_range(Symbols.begin(), E)) 1870 Arr[S.Sym->File].push_back(S); 1871 1872 auto I = Symbols.begin(); 1873 for (std::pair<InputFile *, std::vector<SymbolTableEntry>> &P : Arr) 1874 for (SymbolTableEntry &Entry : P.second) 1875 *I++ = Entry; 1876 } 1877 1878 void SymbolTableBaseSection::addSymbol(Symbol *B) { 1879 // Adding a local symbol to a .dynsym is a bug. 1880 assert(this->Type != SHT_DYNSYM || !B->isLocal()); 1881 1882 bool HashIt = B->isLocal(); 1883 Symbols.push_back({B, StrTabSec.addString(B->getName(), HashIt)}); 1884 } 1885 1886 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *Sym) { 1887 // Initializes symbol lookup tables lazily. This is used only 1888 // for -r or -emit-relocs. 1889 llvm::call_once(OnceFlag, [&] { 1890 SymbolIndexMap.reserve(Symbols.size()); 1891 size_t I = 0; 1892 for (const SymbolTableEntry &E : Symbols) { 1893 if (E.Sym->Type == STT_SECTION) 1894 SectionIndexMap[E.Sym->getOutputSection()] = ++I; 1895 else 1896 SymbolIndexMap[E.Sym] = ++I; 1897 } 1898 }); 1899 1900 // Section symbols are mapped based on their output sections 1901 // to maintain their semantics. 1902 if (Sym->Type == STT_SECTION) 1903 return SectionIndexMap.lookup(Sym->getOutputSection()); 1904 return SymbolIndexMap.lookup(Sym); 1905 } 1906 1907 template <class ELFT> 1908 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &StrTabSec) 1909 : SymbolTableBaseSection(StrTabSec) { 1910 this->Entsize = sizeof(Elf_Sym); 1911 } 1912 1913 // Write the internal symbol table contents to the output symbol table. 1914 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *Buf) { 1915 // The first entry is a null entry as per the ELF spec. 1916 memset(Buf, 0, sizeof(Elf_Sym)); 1917 Buf += sizeof(Elf_Sym); 1918 1919 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf); 1920 1921 for (SymbolTableEntry &Ent : Symbols) { 1922 Symbol *Sym = Ent.Sym; 1923 1924 // Set st_info and st_other. 1925 ESym->st_other = 0; 1926 if (Sym->isLocal()) { 1927 ESym->setBindingAndType(STB_LOCAL, Sym->Type); 1928 } else { 1929 ESym->setBindingAndType(Sym->computeBinding(), Sym->Type); 1930 ESym->setVisibility(Sym->Visibility); 1931 } 1932 1933 ESym->st_name = Ent.StrTabOffset; 1934 1935 // Set a section index. 1936 BssSection *CommonSec = nullptr; 1937 if (!Config->DefineCommon) 1938 if (auto *D = dyn_cast<Defined>(Sym)) 1939 CommonSec = dyn_cast_or_null<BssSection>(D->Section); 1940 if (CommonSec) 1941 ESym->st_shndx = SHN_COMMON; 1942 else if (Sym->NeedsPltAddr) 1943 ESym->st_shndx = SHN_UNDEF; 1944 else if (const OutputSection *OutSec = Sym->getOutputSection()) 1945 ESym->st_shndx = OutSec->SectionIndex; 1946 else if (isa<Defined>(Sym)) 1947 ESym->st_shndx = SHN_ABS; 1948 else 1949 ESym->st_shndx = SHN_UNDEF; 1950 1951 // Copy symbol size if it is a defined symbol. st_size is not significant 1952 // for undefined symbols, so whether copying it or not is up to us if that's 1953 // the case. We'll leave it as zero because by not setting a value, we can 1954 // get the exact same outputs for two sets of input files that differ only 1955 // in undefined symbol size in DSOs. 1956 if (ESym->st_shndx == SHN_UNDEF) 1957 ESym->st_size = 0; 1958 else 1959 ESym->st_size = Sym->getSize(); 1960 1961 // st_value is usually an address of a symbol, but that has a 1962 // special meaining for uninstantiated common symbols (this can 1963 // occur if -r is given). 1964 if (CommonSec) 1965 ESym->st_value = CommonSec->Alignment; 1966 else 1967 ESym->st_value = Sym->getVA(); 1968 1969 ++ESym; 1970 } 1971 1972 // On MIPS we need to mark symbol which has a PLT entry and requires 1973 // pointer equality by STO_MIPS_PLT flag. That is necessary to help 1974 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs. 1975 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt 1976 if (Config->EMachine == EM_MIPS) { 1977 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf); 1978 1979 for (SymbolTableEntry &Ent : Symbols) { 1980 Symbol *Sym = Ent.Sym; 1981 if (Sym->isInPlt() && Sym->NeedsPltAddr) 1982 ESym->st_other |= STO_MIPS_PLT; 1983 if (isMicroMips()) { 1984 // Set STO_MIPS_MICROMIPS flag and less-significant bit for 1985 // a defined microMIPS symbol and symbol should point to its 1986 // PLT entry (in case of microMIPS, PLT entries always contain 1987 // microMIPS code). 1988 if (Sym->isDefined() && 1989 ((Sym->StOther & STO_MIPS_MICROMIPS) || Sym->NeedsPltAddr)) { 1990 if (StrTabSec.isDynamic()) 1991 ESym->st_value |= 1; 1992 ESym->st_other |= STO_MIPS_MICROMIPS; 1993 } 1994 } 1995 if (Config->Relocatable) 1996 if (auto *D = dyn_cast<Defined>(Sym)) 1997 if (isMipsPIC<ELFT>(D)) 1998 ESym->st_other |= STO_MIPS_PIC; 1999 ++ESym; 2000 } 2001 } 2002 } 2003 2004 // .hash and .gnu.hash sections contain on-disk hash tables that map 2005 // symbol names to their dynamic symbol table indices. Their purpose 2006 // is to help the dynamic linker resolve symbols quickly. If ELF files 2007 // don't have them, the dynamic linker has to do linear search on all 2008 // dynamic symbols, which makes programs slower. Therefore, a .hash 2009 // section is added to a DSO by default. A .gnu.hash is added if you 2010 // give the -hash-style=gnu or -hash-style=both option. 2011 // 2012 // The Unix semantics of resolving dynamic symbols is somewhat expensive. 2013 // Each ELF file has a list of DSOs that the ELF file depends on and a 2014 // list of dynamic symbols that need to be resolved from any of the 2015 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n) 2016 // where m is the number of DSOs and n is the number of dynamic 2017 // symbols. For modern large programs, both m and n are large. So 2018 // making each step faster by using hash tables substiantially 2019 // improves time to load programs. 2020 // 2021 // (Note that this is not the only way to design the shared library. 2022 // For instance, the Windows DLL takes a different approach. On 2023 // Windows, each dynamic symbol has a name of DLL from which the symbol 2024 // has to be resolved. That makes the cost of symbol resolution O(n). 2025 // This disables some hacky techniques you can use on Unix such as 2026 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.) 2027 // 2028 // Due to historical reasons, we have two different hash tables, .hash 2029 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new 2030 // and better version of .hash. .hash is just an on-disk hash table, but 2031 // .gnu.hash has a bloom filter in addition to a hash table to skip 2032 // DSOs very quickly. If you are sure that your dynamic linker knows 2033 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a 2034 // safe bet is to specify -hash-style=both for backward compatibilty. 2035 GnuHashTableSection::GnuHashTableSection() 2036 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, Config->Wordsize, ".gnu.hash") { 2037 } 2038 2039 void GnuHashTableSection::finalizeContents() { 2040 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex; 2041 2042 // Computes bloom filter size in word size. We want to allocate 12 2043 // bits for each symbol. It must be a power of two. 2044 if (Symbols.empty()) { 2045 MaskWords = 1; 2046 } else { 2047 uint64_t NumBits = Symbols.size() * 12; 2048 MaskWords = NextPowerOf2(NumBits / (Config->Wordsize * 8)); 2049 } 2050 2051 Size = 16; // Header 2052 Size += Config->Wordsize * MaskWords; // Bloom filter 2053 Size += NBuckets * 4; // Hash buckets 2054 Size += Symbols.size() * 4; // Hash values 2055 } 2056 2057 void GnuHashTableSection::writeTo(uint8_t *Buf) { 2058 // The output buffer is not guaranteed to be zero-cleared because we pre- 2059 // fill executable sections with trap instructions. This is a precaution 2060 // for that case, which happens only when -no-rosegment is given. 2061 memset(Buf, 0, Size); 2062 2063 // Write a header. 2064 write32(Buf, NBuckets); 2065 write32(Buf + 4, InX::DynSymTab->getNumSymbols() - Symbols.size()); 2066 write32(Buf + 8, MaskWords); 2067 write32(Buf + 12, Shift2); 2068 Buf += 16; 2069 2070 // Write a bloom filter and a hash table. 2071 writeBloomFilter(Buf); 2072 Buf += Config->Wordsize * MaskWords; 2073 writeHashTable(Buf); 2074 } 2075 2076 // This function writes a 2-bit bloom filter. This bloom filter alone 2077 // usually filters out 80% or more of all symbol lookups [1]. 2078 // The dynamic linker uses the hash table only when a symbol is not 2079 // filtered out by a bloom filter. 2080 // 2081 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2), 2082 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf 2083 void GnuHashTableSection::writeBloomFilter(uint8_t *Buf) { 2084 unsigned C = Config->Is64 ? 64 : 32; 2085 for (const Entry &Sym : Symbols) { 2086 size_t I = (Sym.Hash / C) & (MaskWords - 1); 2087 uint64_t Val = readUint(Buf + I * Config->Wordsize); 2088 Val |= uint64_t(1) << (Sym.Hash % C); 2089 Val |= uint64_t(1) << ((Sym.Hash >> Shift2) % C); 2090 writeUint(Buf + I * Config->Wordsize, Val); 2091 } 2092 } 2093 2094 void GnuHashTableSection::writeHashTable(uint8_t *Buf) { 2095 uint32_t *Buckets = reinterpret_cast<uint32_t *>(Buf); 2096 uint32_t OldBucket = -1; 2097 uint32_t *Values = Buckets + NBuckets; 2098 for (auto I = Symbols.begin(), E = Symbols.end(); I != E; ++I) { 2099 // Write a hash value. It represents a sequence of chains that share the 2100 // same hash modulo value. The last element of each chain is terminated by 2101 // LSB 1. 2102 uint32_t Hash = I->Hash; 2103 bool IsLastInChain = (I + 1) == E || I->BucketIdx != (I + 1)->BucketIdx; 2104 Hash = IsLastInChain ? Hash | 1 : Hash & ~1; 2105 write32(Values++, Hash); 2106 2107 if (I->BucketIdx == OldBucket) 2108 continue; 2109 // Write a hash bucket. Hash buckets contain indices in the following hash 2110 // value table. 2111 write32(Buckets + I->BucketIdx, I->Sym->DynsymIndex); 2112 OldBucket = I->BucketIdx; 2113 } 2114 } 2115 2116 static uint32_t hashGnu(StringRef Name) { 2117 uint32_t H = 5381; 2118 for (uint8_t C : Name) 2119 H = (H << 5) + H + C; 2120 return H; 2121 } 2122 2123 // Add symbols to this symbol hash table. Note that this function 2124 // destructively sort a given vector -- which is needed because 2125 // GNU-style hash table places some sorting requirements. 2126 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &V) { 2127 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce 2128 // its type correctly. 2129 std::vector<SymbolTableEntry>::iterator Mid = 2130 std::stable_partition(V.begin(), V.end(), [](const SymbolTableEntry &S) { 2131 return !S.Sym->isDefined(); 2132 }); 2133 2134 // We chose load factor 4 for the on-disk hash table. For each hash 2135 // collision, the dynamic linker will compare a uint32_t hash value. 2136 // Since the integer comparison is quite fast, we believe we can 2137 // make the load factor even larger. 4 is just a conservative choice. 2138 // 2139 // Note that we don't want to create a zero-sized hash table because 2140 // Android loader as of 2018 doesn't like a .gnu.hash containing such 2141 // table. If that's the case, we create a hash table with one unused 2142 // dummy slot. 2143 NBuckets = std::max<size_t>((V.end() - Mid) / 4, 1); 2144 2145 if (Mid == V.end()) 2146 return; 2147 2148 for (SymbolTableEntry &Ent : llvm::make_range(Mid, V.end())) { 2149 Symbol *B = Ent.Sym; 2150 uint32_t Hash = hashGnu(B->getName()); 2151 uint32_t BucketIdx = Hash % NBuckets; 2152 Symbols.push_back({B, Ent.StrTabOffset, Hash, BucketIdx}); 2153 } 2154 2155 std::stable_sort( 2156 Symbols.begin(), Symbols.end(), 2157 [](const Entry &L, const Entry &R) { return L.BucketIdx < R.BucketIdx; }); 2158 2159 V.erase(Mid, V.end()); 2160 for (const Entry &Ent : Symbols) 2161 V.push_back({Ent.Sym, Ent.StrTabOffset}); 2162 } 2163 2164 HashTableSection::HashTableSection() 2165 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") { 2166 this->Entsize = 4; 2167 } 2168 2169 void HashTableSection::finalizeContents() { 2170 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex; 2171 2172 unsigned NumEntries = 2; // nbucket and nchain. 2173 NumEntries += InX::DynSymTab->getNumSymbols(); // The chain entries. 2174 2175 // Create as many buckets as there are symbols. 2176 NumEntries += InX::DynSymTab->getNumSymbols(); 2177 this->Size = NumEntries * 4; 2178 } 2179 2180 void HashTableSection::writeTo(uint8_t *Buf) { 2181 // See comment in GnuHashTableSection::writeTo. 2182 memset(Buf, 0, Size); 2183 2184 unsigned NumSymbols = InX::DynSymTab->getNumSymbols(); 2185 2186 uint32_t *P = reinterpret_cast<uint32_t *>(Buf); 2187 write32(P++, NumSymbols); // nbucket 2188 write32(P++, NumSymbols); // nchain 2189 2190 uint32_t *Buckets = P; 2191 uint32_t *Chains = P + NumSymbols; 2192 2193 for (const SymbolTableEntry &S : InX::DynSymTab->getSymbols()) { 2194 Symbol *Sym = S.Sym; 2195 StringRef Name = Sym->getName(); 2196 unsigned I = Sym->DynsymIndex; 2197 uint32_t Hash = hashSysV(Name) % NumSymbols; 2198 Chains[I] = Buckets[Hash]; 2199 write32(Buckets + Hash, I); 2200 } 2201 } 2202 2203 // On PowerPC64 the lazy symbol resolvers go into the `global linkage table` 2204 // in the .glink section, rather then the typical .plt section. 2205 PltSection::PltSection(bool IsIplt) 2206 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, 2207 Config->EMachine == EM_PPC64 ? ".glink" : ".plt"), 2208 HeaderSize(IsIplt ? 0 : Target->PltHeaderSize), IsIplt(IsIplt) { 2209 // The PLT needs to be writable on SPARC as the dynamic linker will 2210 // modify the instructions in the PLT entries. 2211 if (Config->EMachine == EM_SPARCV9) 2212 this->Flags |= SHF_WRITE; 2213 } 2214 2215 void PltSection::writeTo(uint8_t *Buf) { 2216 // At beginning of PLT but not the IPLT, we have code to call the dynamic 2217 // linker to resolve dynsyms at runtime. Write such code. 2218 if (!IsIplt) 2219 Target->writePltHeader(Buf); 2220 size_t Off = HeaderSize; 2221 // The IPlt is immediately after the Plt, account for this in RelOff 2222 unsigned PltOff = getPltRelocOff(); 2223 2224 for (auto &I : Entries) { 2225 const Symbol *B = I.first; 2226 unsigned RelOff = I.second + PltOff; 2227 uint64_t Got = B->getGotPltVA(); 2228 uint64_t Plt = this->getVA() + Off; 2229 Target->writePlt(Buf + Off, Got, Plt, B->PltIndex, RelOff); 2230 Off += Target->PltEntrySize; 2231 } 2232 } 2233 2234 template <class ELFT> void PltSection::addEntry(Symbol &Sym) { 2235 Sym.PltIndex = Entries.size(); 2236 RelocationBaseSection *PltRelocSection = InX::RelaPlt; 2237 if (IsIplt) { 2238 PltRelocSection = InX::RelaIplt; 2239 Sym.IsInIplt = true; 2240 } 2241 unsigned RelOff = 2242 static_cast<RelocationSection<ELFT> *>(PltRelocSection)->getRelocOffset(); 2243 Entries.push_back(std::make_pair(&Sym, RelOff)); 2244 } 2245 2246 size_t PltSection::getSize() const { 2247 return HeaderSize + Entries.size() * Target->PltEntrySize; 2248 } 2249 2250 // Some architectures such as additional symbols in the PLT section. For 2251 // example ARM uses mapping symbols to aid disassembly 2252 void PltSection::addSymbols() { 2253 // The PLT may have symbols defined for the Header, the IPLT has no header 2254 if (!IsIplt) 2255 Target->addPltHeaderSymbols(*this); 2256 size_t Off = HeaderSize; 2257 for (size_t I = 0; I < Entries.size(); ++I) { 2258 Target->addPltSymbols(*this, Off); 2259 Off += Target->PltEntrySize; 2260 } 2261 } 2262 2263 unsigned PltSection::getPltRelocOff() const { 2264 return IsIplt ? InX::Plt->getSize() : 0; 2265 } 2266 2267 // The string hash function for .gdb_index. 2268 static uint32_t computeGdbHash(StringRef S) { 2269 uint32_t H = 0; 2270 for (uint8_t C : S) 2271 H = H * 67 + tolower(C) - 113; 2272 return H; 2273 } 2274 2275 GdbIndexSection::GdbIndexSection() 2276 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index") {} 2277 2278 // Returns the desired size of an on-disk hash table for a .gdb_index section. 2279 // There's a tradeoff between size and collision rate. We aim 75% utilization. 2280 size_t GdbIndexSection::computeSymtabSize() const { 2281 return std::max<size_t>(NextPowerOf2(Symbols.size() * 4 / 3), 1024); 2282 } 2283 2284 // Compute the output section size. 2285 void GdbIndexSection::initOutputSize() { 2286 Size = sizeof(GdbIndexHeader) + computeSymtabSize() * 8; 2287 2288 for (GdbChunk &Chunk : Chunks) 2289 Size += Chunk.CompilationUnits.size() * 16 + Chunk.AddressAreas.size() * 20; 2290 2291 // Add the constant pool size if exists. 2292 if (!Symbols.empty()) { 2293 GdbSymbol &Sym = Symbols.back(); 2294 Size += Sym.NameOff + Sym.Name.size() + 1; 2295 } 2296 } 2297 2298 static std::vector<InputSection *> getDebugInfoSections() { 2299 std::vector<InputSection *> Ret; 2300 for (InputSectionBase *S : InputSections) 2301 if (InputSection *IS = dyn_cast<InputSection>(S)) 2302 if (IS->Name == ".debug_info") 2303 Ret.push_back(IS); 2304 return Ret; 2305 } 2306 2307 static std::vector<GdbIndexSection::CuEntry> readCuList(DWARFContext &Dwarf) { 2308 std::vector<GdbIndexSection::CuEntry> Ret; 2309 for (std::unique_ptr<DWARFCompileUnit> &Cu : Dwarf.compile_units()) 2310 Ret.push_back({Cu->getOffset(), Cu->getLength() + 4}); 2311 return Ret; 2312 } 2313 2314 static std::vector<GdbIndexSection::AddressEntry> 2315 readAddressAreas(DWARFContext &Dwarf, InputSection *Sec) { 2316 std::vector<GdbIndexSection::AddressEntry> Ret; 2317 2318 uint32_t CuIdx = 0; 2319 for (std::unique_ptr<DWARFCompileUnit> &Cu : Dwarf.compile_units()) { 2320 DWARFAddressRangesVector Ranges; 2321 Cu->collectAddressRanges(Ranges); 2322 2323 ArrayRef<InputSectionBase *> Sections = Sec->File->getSections(); 2324 for (DWARFAddressRange &R : Ranges) { 2325 InputSectionBase *S = Sections[R.SectionIndex]; 2326 if (!S || S == &InputSection::Discarded || !S->Live) 2327 continue; 2328 // Range list with zero size has no effect. 2329 if (R.LowPC == R.HighPC) 2330 continue; 2331 auto *IS = cast<InputSection>(S); 2332 uint64_t Offset = IS->getOffsetInFile(); 2333 Ret.push_back({IS, R.LowPC - Offset, R.HighPC - Offset, CuIdx}); 2334 } 2335 ++CuIdx; 2336 } 2337 return Ret; 2338 } 2339 2340 static std::vector<GdbIndexSection::NameTypeEntry> 2341 readPubNamesAndTypes(DWARFContext &Dwarf, uint32_t Idx) { 2342 StringRef Sec1 = Dwarf.getDWARFObj().getGnuPubNamesSection(); 2343 StringRef Sec2 = Dwarf.getDWARFObj().getGnuPubTypesSection(); 2344 2345 std::vector<GdbIndexSection::NameTypeEntry> Ret; 2346 for (StringRef Sec : {Sec1, Sec2}) { 2347 DWARFDebugPubTable Table(Sec, Config->IsLE, true); 2348 for (const DWARFDebugPubTable::Set &Set : Table.getData()) 2349 for (const DWARFDebugPubTable::Entry &Ent : Set.Entries) 2350 Ret.push_back({{Ent.Name, computeGdbHash(Ent.Name)}, 2351 (Ent.Descriptor.toBits() << 24) | Idx}); 2352 } 2353 return Ret; 2354 } 2355 2356 // Create a list of symbols from a given list of symbol names and types 2357 // by uniquifying them by name. 2358 static std::vector<GdbIndexSection::GdbSymbol> 2359 createSymbols(ArrayRef<std::vector<GdbIndexSection::NameTypeEntry>> NameTypes) { 2360 typedef GdbIndexSection::GdbSymbol GdbSymbol; 2361 typedef GdbIndexSection::NameTypeEntry NameTypeEntry; 2362 2363 // The number of symbols we will handle in this function is of the order 2364 // of millions for very large executables, so we use multi-threading to 2365 // speed it up. 2366 size_t NumShards = 32; 2367 size_t Concurrency = 1; 2368 if (ThreadsEnabled) 2369 Concurrency = 2370 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), NumShards); 2371 2372 // A sharded map to uniquify symbols by name. 2373 std::vector<DenseMap<CachedHashStringRef, size_t>> Map(NumShards); 2374 size_t Shift = 32 - countTrailingZeros(NumShards); 2375 2376 // Instantiate GdbSymbols while uniqufying them by name. 2377 std::vector<std::vector<GdbSymbol>> Symbols(NumShards); 2378 parallelForEachN(0, Concurrency, [&](size_t ThreadId) { 2379 for (ArrayRef<NameTypeEntry> Entries : NameTypes) { 2380 for (const NameTypeEntry &Ent : Entries) { 2381 size_t ShardId = Ent.Name.hash() >> Shift; 2382 if ((ShardId & (Concurrency - 1)) != ThreadId) 2383 continue; 2384 2385 size_t &Idx = Map[ShardId][Ent.Name]; 2386 if (Idx) { 2387 Symbols[ShardId][Idx - 1].CuVector.push_back(Ent.Type); 2388 continue; 2389 } 2390 2391 Idx = Symbols[ShardId].size() + 1; 2392 Symbols[ShardId].push_back({Ent.Name, {Ent.Type}, 0, 0}); 2393 } 2394 } 2395 }); 2396 2397 size_t NumSymbols = 0; 2398 for (ArrayRef<GdbSymbol> V : Symbols) 2399 NumSymbols += V.size(); 2400 2401 // The return type is a flattened vector, so we'll copy each vector 2402 // contents to Ret. 2403 std::vector<GdbSymbol> Ret; 2404 Ret.reserve(NumSymbols); 2405 for (std::vector<GdbSymbol> &Vec : Symbols) 2406 for (GdbSymbol &Sym : Vec) 2407 Ret.push_back(std::move(Sym)); 2408 2409 // CU vectors and symbol names are adjacent in the output file. 2410 // We can compute their offsets in the output file now. 2411 size_t Off = 0; 2412 for (GdbSymbol &Sym : Ret) { 2413 Sym.CuVectorOff = Off; 2414 Off += (Sym.CuVector.size() + 1) * 4; 2415 } 2416 for (GdbSymbol &Sym : Ret) { 2417 Sym.NameOff = Off; 2418 Off += Sym.Name.size() + 1; 2419 } 2420 2421 return Ret; 2422 } 2423 2424 // Returns a newly-created .gdb_index section. 2425 template <class ELFT> GdbIndexSection *GdbIndexSection::create() { 2426 std::vector<InputSection *> Sections = getDebugInfoSections(); 2427 2428 // .debug_gnu_pub{names,types} are useless in executables. 2429 // They are present in input object files solely for creating 2430 // a .gdb_index. So we can remove them from the output. 2431 for (InputSectionBase *S : InputSections) 2432 if (S->Name == ".debug_gnu_pubnames" || S->Name == ".debug_gnu_pubtypes") 2433 S->Live = false; 2434 2435 std::vector<GdbChunk> Chunks(Sections.size()); 2436 std::vector<std::vector<NameTypeEntry>> NameTypes(Sections.size()); 2437 2438 parallelForEachN(0, Sections.size(), [&](size_t I) { 2439 ObjFile<ELFT> *File = Sections[I]->getFile<ELFT>(); 2440 DWARFContext Dwarf(make_unique<LLDDwarfObj<ELFT>>(File)); 2441 2442 Chunks[I].Sec = Sections[I]; 2443 Chunks[I].CompilationUnits = readCuList(Dwarf); 2444 Chunks[I].AddressAreas = readAddressAreas(Dwarf, Sections[I]); 2445 NameTypes[I] = readPubNamesAndTypes(Dwarf, I); 2446 }); 2447 2448 auto *Ret = make<GdbIndexSection>(); 2449 Ret->Chunks = std::move(Chunks); 2450 Ret->Symbols = createSymbols(NameTypes); 2451 Ret->initOutputSize(); 2452 return Ret; 2453 } 2454 2455 void GdbIndexSection::writeTo(uint8_t *Buf) { 2456 // Write the header. 2457 auto *Hdr = reinterpret_cast<GdbIndexHeader *>(Buf); 2458 uint8_t *Start = Buf; 2459 Hdr->Version = 7; 2460 Buf += sizeof(*Hdr); 2461 2462 // Write the CU list. 2463 Hdr->CuListOff = Buf - Start; 2464 for (GdbChunk &Chunk : Chunks) { 2465 for (CuEntry &Cu : Chunk.CompilationUnits) { 2466 write64le(Buf, Chunk.Sec->OutSecOff + Cu.CuOffset); 2467 write64le(Buf + 8, Cu.CuLength); 2468 Buf += 16; 2469 } 2470 } 2471 2472 // Write the address area. 2473 Hdr->CuTypesOff = Buf - Start; 2474 Hdr->AddressAreaOff = Buf - Start; 2475 uint32_t CuOff = 0; 2476 for (GdbChunk &Chunk : Chunks) { 2477 for (AddressEntry &E : Chunk.AddressAreas) { 2478 uint64_t BaseAddr = E.Section->getVA(0); 2479 write64le(Buf, BaseAddr + E.LowAddress); 2480 write64le(Buf + 8, BaseAddr + E.HighAddress); 2481 write32le(Buf + 16, E.CuIndex + CuOff); 2482 Buf += 20; 2483 } 2484 CuOff += Chunk.CompilationUnits.size(); 2485 } 2486 2487 // Write the on-disk open-addressing hash table containing symbols. 2488 Hdr->SymtabOff = Buf - Start; 2489 size_t SymtabSize = computeSymtabSize(); 2490 uint32_t Mask = SymtabSize - 1; 2491 2492 for (GdbSymbol &Sym : Symbols) { 2493 uint32_t H = Sym.Name.hash(); 2494 uint32_t I = H & Mask; 2495 uint32_t Step = ((H * 17) & Mask) | 1; 2496 2497 while (read32le(Buf + I * 8)) 2498 I = (I + Step) & Mask; 2499 2500 write32le(Buf + I * 8, Sym.NameOff); 2501 write32le(Buf + I * 8 + 4, Sym.CuVectorOff); 2502 } 2503 2504 Buf += SymtabSize * 8; 2505 2506 // Write the string pool. 2507 Hdr->ConstantPoolOff = Buf - Start; 2508 for (GdbSymbol &Sym : Symbols) 2509 memcpy(Buf + Sym.NameOff, Sym.Name.data(), Sym.Name.size()); 2510 2511 // Write the CU vectors. 2512 for (GdbSymbol &Sym : Symbols) { 2513 write32le(Buf, Sym.CuVector.size()); 2514 Buf += 4; 2515 for (uint32_t Val : Sym.CuVector) { 2516 write32le(Buf, Val); 2517 Buf += 4; 2518 } 2519 } 2520 } 2521 2522 bool GdbIndexSection::empty() const { return !Out::DebugInfo; } 2523 2524 EhFrameHeader::EhFrameHeader() 2525 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr") {} 2526 2527 // .eh_frame_hdr contains a binary search table of pointers to FDEs. 2528 // Each entry of the search table consists of two values, 2529 // the starting PC from where FDEs covers, and the FDE's address. 2530 // It is sorted by PC. 2531 void EhFrameHeader::writeTo(uint8_t *Buf) { 2532 typedef EhFrameSection::FdeData FdeData; 2533 2534 std::vector<FdeData> Fdes = InX::EhFrame->getFdeData(); 2535 2536 // Sort the FDE list by their PC and uniqueify. Usually there is only 2537 // one FDE for a PC (i.e. function), but if ICF merges two functions 2538 // into one, there can be more than one FDEs pointing to the address. 2539 auto Less = [](const FdeData &A, const FdeData &B) { return A.Pc < B.Pc; }; 2540 std::stable_sort(Fdes.begin(), Fdes.end(), Less); 2541 auto Eq = [](const FdeData &A, const FdeData &B) { return A.Pc == B.Pc; }; 2542 Fdes.erase(std::unique(Fdes.begin(), Fdes.end(), Eq), Fdes.end()); 2543 2544 Buf[0] = 1; 2545 Buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4; 2546 Buf[2] = DW_EH_PE_udata4; 2547 Buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4; 2548 write32(Buf + 4, InX::EhFrame->getParent()->Addr - this->getVA() - 4); 2549 write32(Buf + 8, Fdes.size()); 2550 Buf += 12; 2551 2552 uint64_t VA = this->getVA(); 2553 for (FdeData &Fde : Fdes) { 2554 write32(Buf, Fde.Pc - VA); 2555 write32(Buf + 4, Fde.FdeVA - VA); 2556 Buf += 8; 2557 } 2558 } 2559 2560 size_t EhFrameHeader::getSize() const { 2561 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs. 2562 return 12 + InX::EhFrame->NumFdes * 8; 2563 } 2564 2565 bool EhFrameHeader::empty() const { return InX::EhFrame->empty(); } 2566 2567 template <class ELFT> 2568 VersionDefinitionSection<ELFT>::VersionDefinitionSection() 2569 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t), 2570 ".gnu.version_d") {} 2571 2572 static StringRef getFileDefName() { 2573 if (!Config->SoName.empty()) 2574 return Config->SoName; 2575 return Config->OutputFile; 2576 } 2577 2578 template <class ELFT> void VersionDefinitionSection<ELFT>::finalizeContents() { 2579 FileDefNameOff = InX::DynStrTab->addString(getFileDefName()); 2580 for (VersionDefinition &V : Config->VersionDefinitions) 2581 V.NameOff = InX::DynStrTab->addString(V.Name); 2582 2583 getParent()->Link = InX::DynStrTab->getParent()->SectionIndex; 2584 2585 // sh_info should be set to the number of definitions. This fact is missed in 2586 // documentation, but confirmed by binutils community: 2587 // https://sourceware.org/ml/binutils/2014-11/msg00355.html 2588 getParent()->Info = getVerDefNum(); 2589 } 2590 2591 template <class ELFT> 2592 void VersionDefinitionSection<ELFT>::writeOne(uint8_t *Buf, uint32_t Index, 2593 StringRef Name, size_t NameOff) { 2594 auto *Verdef = reinterpret_cast<Elf_Verdef *>(Buf); 2595 Verdef->vd_version = 1; 2596 Verdef->vd_cnt = 1; 2597 Verdef->vd_aux = sizeof(Elf_Verdef); 2598 Verdef->vd_next = sizeof(Elf_Verdef) + sizeof(Elf_Verdaux); 2599 Verdef->vd_flags = (Index == 1 ? VER_FLG_BASE : 0); 2600 Verdef->vd_ndx = Index; 2601 Verdef->vd_hash = hashSysV(Name); 2602 2603 auto *Verdaux = reinterpret_cast<Elf_Verdaux *>(Buf + sizeof(Elf_Verdef)); 2604 Verdaux->vda_name = NameOff; 2605 Verdaux->vda_next = 0; 2606 } 2607 2608 template <class ELFT> 2609 void VersionDefinitionSection<ELFT>::writeTo(uint8_t *Buf) { 2610 writeOne(Buf, 1, getFileDefName(), FileDefNameOff); 2611 2612 for (VersionDefinition &V : Config->VersionDefinitions) { 2613 Buf += sizeof(Elf_Verdef) + sizeof(Elf_Verdaux); 2614 writeOne(Buf, V.Id, V.Name, V.NameOff); 2615 } 2616 2617 // Need to terminate the last version definition. 2618 Elf_Verdef *Verdef = reinterpret_cast<Elf_Verdef *>(Buf); 2619 Verdef->vd_next = 0; 2620 } 2621 2622 template <class ELFT> size_t VersionDefinitionSection<ELFT>::getSize() const { 2623 return (sizeof(Elf_Verdef) + sizeof(Elf_Verdaux)) * getVerDefNum(); 2624 } 2625 2626 template <class ELFT> 2627 VersionTableSection<ELFT>::VersionTableSection() 2628 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t), 2629 ".gnu.version") { 2630 this->Entsize = sizeof(Elf_Versym); 2631 } 2632 2633 template <class ELFT> void VersionTableSection<ELFT>::finalizeContents() { 2634 // At the moment of june 2016 GNU docs does not mention that sh_link field 2635 // should be set, but Sun docs do. Also readelf relies on this field. 2636 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex; 2637 } 2638 2639 template <class ELFT> size_t VersionTableSection<ELFT>::getSize() const { 2640 return sizeof(Elf_Versym) * (InX::DynSymTab->getSymbols().size() + 1); 2641 } 2642 2643 template <class ELFT> void VersionTableSection<ELFT>::writeTo(uint8_t *Buf) { 2644 auto *OutVersym = reinterpret_cast<Elf_Versym *>(Buf) + 1; 2645 for (const SymbolTableEntry &S : InX::DynSymTab->getSymbols()) { 2646 OutVersym->vs_index = S.Sym->VersionId; 2647 ++OutVersym; 2648 } 2649 } 2650 2651 template <class ELFT> bool VersionTableSection<ELFT>::empty() const { 2652 return !In<ELFT>::VerDef && In<ELFT>::VerNeed->empty(); 2653 } 2654 2655 template <class ELFT> 2656 VersionNeedSection<ELFT>::VersionNeedSection() 2657 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t), 2658 ".gnu.version_r") { 2659 // Identifiers in verneed section start at 2 because 0 and 1 are reserved 2660 // for VER_NDX_LOCAL and VER_NDX_GLOBAL. 2661 // First identifiers are reserved by verdef section if it exist. 2662 NextIndex = getVerDefNum() + 1; 2663 } 2664 2665 template <class ELFT> void VersionNeedSection<ELFT>::addSymbol(Symbol *SS) { 2666 auto &File = cast<SharedFile<ELFT>>(*SS->File); 2667 if (SS->VerdefIndex == VER_NDX_GLOBAL) { 2668 SS->VersionId = VER_NDX_GLOBAL; 2669 return; 2670 } 2671 2672 // If we don't already know that we need an Elf_Verneed for this DSO, prepare 2673 // to create one by adding it to our needed list and creating a dynstr entry 2674 // for the soname. 2675 if (File.VerdefMap.empty()) 2676 Needed.push_back({&File, InX::DynStrTab->addString(File.SoName)}); 2677 const typename ELFT::Verdef *Ver = File.Verdefs[SS->VerdefIndex]; 2678 typename SharedFile<ELFT>::NeededVer &NV = File.VerdefMap[Ver]; 2679 2680 // If we don't already know that we need an Elf_Vernaux for this Elf_Verdef, 2681 // prepare to create one by allocating a version identifier and creating a 2682 // dynstr entry for the version name. 2683 if (NV.Index == 0) { 2684 NV.StrTab = InX::DynStrTab->addString(File.getStringTable().data() + 2685 Ver->getAux()->vda_name); 2686 NV.Index = NextIndex++; 2687 } 2688 SS->VersionId = NV.Index; 2689 } 2690 2691 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *Buf) { 2692 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs. 2693 auto *Verneed = reinterpret_cast<Elf_Verneed *>(Buf); 2694 auto *Vernaux = reinterpret_cast<Elf_Vernaux *>(Verneed + Needed.size()); 2695 2696 for (std::pair<SharedFile<ELFT> *, size_t> &P : Needed) { 2697 // Create an Elf_Verneed for this DSO. 2698 Verneed->vn_version = 1; 2699 Verneed->vn_cnt = P.first->VerdefMap.size(); 2700 Verneed->vn_file = P.second; 2701 Verneed->vn_aux = 2702 reinterpret_cast<char *>(Vernaux) - reinterpret_cast<char *>(Verneed); 2703 Verneed->vn_next = sizeof(Elf_Verneed); 2704 ++Verneed; 2705 2706 // Create the Elf_Vernauxs for this Elf_Verneed. The loop iterates over 2707 // VerdefMap, which will only contain references to needed version 2708 // definitions. Each Elf_Vernaux is based on the information contained in 2709 // the Elf_Verdef in the source DSO. This loop iterates over a std::map of 2710 // pointers, but is deterministic because the pointers refer to Elf_Verdef 2711 // data structures within a single input file. 2712 for (auto &NV : P.first->VerdefMap) { 2713 Vernaux->vna_hash = NV.first->vd_hash; 2714 Vernaux->vna_flags = 0; 2715 Vernaux->vna_other = NV.second.Index; 2716 Vernaux->vna_name = NV.second.StrTab; 2717 Vernaux->vna_next = sizeof(Elf_Vernaux); 2718 ++Vernaux; 2719 } 2720 2721 Vernaux[-1].vna_next = 0; 2722 } 2723 Verneed[-1].vn_next = 0; 2724 } 2725 2726 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() { 2727 getParent()->Link = InX::DynStrTab->getParent()->SectionIndex; 2728 getParent()->Info = Needed.size(); 2729 } 2730 2731 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const { 2732 unsigned Size = Needed.size() * sizeof(Elf_Verneed); 2733 for (const std::pair<SharedFile<ELFT> *, size_t> &P : Needed) 2734 Size += P.first->VerdefMap.size() * sizeof(Elf_Vernaux); 2735 return Size; 2736 } 2737 2738 template <class ELFT> bool VersionNeedSection<ELFT>::empty() const { 2739 return getNeedNum() == 0; 2740 } 2741 2742 void MergeSyntheticSection::addSection(MergeInputSection *MS) { 2743 MS->Parent = this; 2744 Sections.push_back(MS); 2745 } 2746 2747 MergeTailSection::MergeTailSection(StringRef Name, uint32_t Type, 2748 uint64_t Flags, uint32_t Alignment) 2749 : MergeSyntheticSection(Name, Type, Flags, Alignment), 2750 Builder(StringTableBuilder::RAW, Alignment) {} 2751 2752 size_t MergeTailSection::getSize() const { return Builder.getSize(); } 2753 2754 void MergeTailSection::writeTo(uint8_t *Buf) { Builder.write(Buf); } 2755 2756 void MergeTailSection::finalizeContents() { 2757 // Add all string pieces to the string table builder to create section 2758 // contents. 2759 for (MergeInputSection *Sec : Sections) 2760 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2761 if (Sec->Pieces[I].Live) 2762 Builder.add(Sec->getData(I)); 2763 2764 // Fix the string table content. After this, the contents will never change. 2765 Builder.finalize(); 2766 2767 // finalize() fixed tail-optimized strings, so we can now get 2768 // offsets of strings. Get an offset for each string and save it 2769 // to a corresponding StringPiece for easy access. 2770 for (MergeInputSection *Sec : Sections) 2771 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2772 if (Sec->Pieces[I].Live) 2773 Sec->Pieces[I].OutputOff = Builder.getOffset(Sec->getData(I)); 2774 } 2775 2776 void MergeNoTailSection::writeTo(uint8_t *Buf) { 2777 for (size_t I = 0; I < NumShards; ++I) 2778 Shards[I].write(Buf + ShardOffsets[I]); 2779 } 2780 2781 // This function is very hot (i.e. it can take several seconds to finish) 2782 // because sometimes the number of inputs is in an order of magnitude of 2783 // millions. So, we use multi-threading. 2784 // 2785 // For any strings S and T, we know S is not mergeable with T if S's hash 2786 // value is different from T's. If that's the case, we can safely put S and 2787 // T into different string builders without worrying about merge misses. 2788 // We do it in parallel. 2789 void MergeNoTailSection::finalizeContents() { 2790 // Initializes string table builders. 2791 for (size_t I = 0; I < NumShards; ++I) 2792 Shards.emplace_back(StringTableBuilder::RAW, Alignment); 2793 2794 // Concurrency level. Must be a power of 2 to avoid expensive modulo 2795 // operations in the following tight loop. 2796 size_t Concurrency = 1; 2797 if (ThreadsEnabled) 2798 Concurrency = 2799 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), NumShards); 2800 2801 // Add section pieces to the builders. 2802 parallelForEachN(0, Concurrency, [&](size_t ThreadId) { 2803 for (MergeInputSection *Sec : Sections) { 2804 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) { 2805 size_t ShardId = getShardId(Sec->Pieces[I].Hash); 2806 if ((ShardId & (Concurrency - 1)) == ThreadId && Sec->Pieces[I].Live) 2807 Sec->Pieces[I].OutputOff = Shards[ShardId].add(Sec->getData(I)); 2808 } 2809 } 2810 }); 2811 2812 // Compute an in-section offset for each shard. 2813 size_t Off = 0; 2814 for (size_t I = 0; I < NumShards; ++I) { 2815 Shards[I].finalizeInOrder(); 2816 if (Shards[I].getSize() > 0) 2817 Off = alignTo(Off, Alignment); 2818 ShardOffsets[I] = Off; 2819 Off += Shards[I].getSize(); 2820 } 2821 Size = Off; 2822 2823 // So far, section pieces have offsets from beginning of shards, but 2824 // we want offsets from beginning of the whole section. Fix them. 2825 parallelForEach(Sections, [&](MergeInputSection *Sec) { 2826 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2827 if (Sec->Pieces[I].Live) 2828 Sec->Pieces[I].OutputOff += 2829 ShardOffsets[getShardId(Sec->Pieces[I].Hash)]; 2830 }); 2831 } 2832 2833 static MergeSyntheticSection *createMergeSynthetic(StringRef Name, 2834 uint32_t Type, 2835 uint64_t Flags, 2836 uint32_t Alignment) { 2837 bool ShouldTailMerge = (Flags & SHF_STRINGS) && Config->Optimize >= 2; 2838 if (ShouldTailMerge) 2839 return make<MergeTailSection>(Name, Type, Flags, Alignment); 2840 return make<MergeNoTailSection>(Name, Type, Flags, Alignment); 2841 } 2842 2843 // Debug sections may be compressed by zlib. Decompress if exists. 2844 void elf::decompressSections() { 2845 parallelForEach(InputSections, 2846 [](InputSectionBase *Sec) { Sec->maybeDecompress(); }); 2847 } 2848 2849 template <class ELFT> void elf::splitSections() { 2850 // splitIntoPieces needs to be called on each MergeInputSection 2851 // before calling finalizeContents(). 2852 parallelForEach(InputSections, [](InputSectionBase *Sec) { 2853 if (auto *S = dyn_cast<MergeInputSection>(Sec)) 2854 S->splitIntoPieces(); 2855 else if (auto *Eh = dyn_cast<EhInputSection>(Sec)) 2856 Eh->split<ELFT>(); 2857 }); 2858 } 2859 2860 // This function scans over the inputsections to create mergeable 2861 // synthetic sections. 2862 // 2863 // It removes MergeInputSections from the input section array and adds 2864 // new synthetic sections at the location of the first input section 2865 // that it replaces. It then finalizes each synthetic section in order 2866 // to compute an output offset for each piece of each input section. 2867 void elf::mergeSections() { 2868 std::vector<MergeSyntheticSection *> MergeSections; 2869 for (InputSectionBase *&S : InputSections) { 2870 MergeInputSection *MS = dyn_cast<MergeInputSection>(S); 2871 if (!MS) 2872 continue; 2873 2874 // We do not want to handle sections that are not alive, so just remove 2875 // them instead of trying to merge. 2876 if (!MS->Live) 2877 continue; 2878 2879 StringRef OutsecName = getOutputSectionName(MS); 2880 uint32_t Alignment = std::max<uint32_t>(MS->Alignment, MS->Entsize); 2881 2882 auto I = llvm::find_if(MergeSections, [=](MergeSyntheticSection *Sec) { 2883 // While we could create a single synthetic section for two different 2884 // values of Entsize, it is better to take Entsize into consideration. 2885 // 2886 // With a single synthetic section no two pieces with different Entsize 2887 // could be equal, so we may as well have two sections. 2888 // 2889 // Using Entsize in here also allows us to propagate it to the synthetic 2890 // section. 2891 return Sec->Name == OutsecName && Sec->Flags == MS->Flags && 2892 Sec->Entsize == MS->Entsize && Sec->Alignment == Alignment; 2893 }); 2894 if (I == MergeSections.end()) { 2895 MergeSyntheticSection *Syn = 2896 createMergeSynthetic(OutsecName, MS->Type, MS->Flags, Alignment); 2897 MergeSections.push_back(Syn); 2898 I = std::prev(MergeSections.end()); 2899 S = Syn; 2900 Syn->Entsize = MS->Entsize; 2901 } else { 2902 S = nullptr; 2903 } 2904 (*I)->addSection(MS); 2905 } 2906 for (auto *MS : MergeSections) 2907 MS->finalizeContents(); 2908 2909 std::vector<InputSectionBase *> &V = InputSections; 2910 V.erase(std::remove(V.begin(), V.end(), nullptr), V.end()); 2911 } 2912 2913 MipsRldMapSection::MipsRldMapSection() 2914 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, Config->Wordsize, 2915 ".rld_map") {} 2916 2917 ARMExidxSentinelSection::ARMExidxSentinelSection() 2918 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX, 2919 Config->Wordsize, ".ARM.exidx") {} 2920 2921 // Write a terminating sentinel entry to the end of the .ARM.exidx table. 2922 // This section will have been sorted last in the .ARM.exidx table. 2923 // This table entry will have the form: 2924 // | PREL31 upper bound of code that has exception tables | EXIDX_CANTUNWIND | 2925 // The sentinel must have the PREL31 value of an address higher than any 2926 // address described by any other table entry. 2927 void ARMExidxSentinelSection::writeTo(uint8_t *Buf) { 2928 assert(Highest); 2929 uint64_t S = Highest->getVA(Highest->getSize()); 2930 uint64_t P = getVA(); 2931 Target->relocateOne(Buf, R_ARM_PREL31, S - P); 2932 write32le(Buf + 4, 1); 2933 } 2934 2935 // The sentinel has to be removed if there are no other .ARM.exidx entries. 2936 bool ARMExidxSentinelSection::empty() const { 2937 for (InputSection *IS : getInputSections(getParent())) 2938 if (!isa<ARMExidxSentinelSection>(IS)) 2939 return false; 2940 return true; 2941 } 2942 2943 bool ARMExidxSentinelSection::classof(const SectionBase *D) { 2944 return D->kind() == InputSectionBase::Synthetic && D->Type == SHT_ARM_EXIDX; 2945 } 2946 2947 ThunkSection::ThunkSection(OutputSection *OS, uint64_t Off) 2948 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 2949 Config->Wordsize, ".text.thunk") { 2950 this->Parent = OS; 2951 this->OutSecOff = Off; 2952 } 2953 2954 void ThunkSection::addThunk(Thunk *T) { 2955 Thunks.push_back(T); 2956 T->addSymbols(*this); 2957 } 2958 2959 void ThunkSection::writeTo(uint8_t *Buf) { 2960 for (Thunk *T : Thunks) 2961 T->writeTo(Buf + T->Offset); 2962 } 2963 2964 InputSection *ThunkSection::getTargetInputSection() const { 2965 if (Thunks.empty()) 2966 return nullptr; 2967 const Thunk *T = Thunks.front(); 2968 return T->getTargetInputSection(); 2969 } 2970 2971 bool ThunkSection::assignOffsets() { 2972 uint64_t Off = 0; 2973 for (Thunk *T : Thunks) { 2974 Off = alignTo(Off, T->Alignment); 2975 T->setOffset(Off); 2976 uint32_t Size = T->size(); 2977 T->getThunkTargetSym()->Size = Size; 2978 Off += Size; 2979 } 2980 bool Changed = Off != Size; 2981 Size = Off; 2982 return Changed; 2983 } 2984 2985 InputSection *InX::ARMAttributes; 2986 BssSection *InX::Bss; 2987 BssSection *InX::BssRelRo; 2988 BuildIdSection *InX::BuildId; 2989 EhFrameHeader *InX::EhFrameHdr; 2990 EhFrameSection *InX::EhFrame; 2991 SyntheticSection *InX::Dynamic; 2992 StringTableSection *InX::DynStrTab; 2993 SymbolTableBaseSection *InX::DynSymTab; 2994 InputSection *InX::Interp; 2995 GdbIndexSection *InX::GdbIndex; 2996 GotSection *InX::Got; 2997 GotPltSection *InX::GotPlt; 2998 GnuHashTableSection *InX::GnuHashTab; 2999 HashTableSection *InX::HashTab; 3000 IgotPltSection *InX::IgotPlt; 3001 MipsGotSection *InX::MipsGot; 3002 MipsRldMapSection *InX::MipsRldMap; 3003 PltSection *InX::Plt; 3004 PltSection *InX::Iplt; 3005 RelocationBaseSection *InX::RelaDyn; 3006 RelrBaseSection *InX::RelrDyn; 3007 RelocationBaseSection *InX::RelaPlt; 3008 RelocationBaseSection *InX::RelaIplt; 3009 StringTableSection *InX::ShStrTab; 3010 StringTableSection *InX::StrTab; 3011 SymbolTableBaseSection *InX::SymTab; 3012 3013 template GdbIndexSection *GdbIndexSection::create<ELF32LE>(); 3014 template GdbIndexSection *GdbIndexSection::create<ELF32BE>(); 3015 template GdbIndexSection *GdbIndexSection::create<ELF64LE>(); 3016 template GdbIndexSection *GdbIndexSection::create<ELF64BE>(); 3017 3018 template void elf::splitSections<ELF32LE>(); 3019 template void elf::splitSections<ELF32BE>(); 3020 template void elf::splitSections<ELF64LE>(); 3021 template void elf::splitSections<ELF64BE>(); 3022 3023 template void EhFrameSection::addSection<ELF32LE>(InputSectionBase *); 3024 template void EhFrameSection::addSection<ELF32BE>(InputSectionBase *); 3025 template void EhFrameSection::addSection<ELF64LE>(InputSectionBase *); 3026 template void EhFrameSection::addSection<ELF64BE>(InputSectionBase *); 3027 3028 template void PltSection::addEntry<ELF32LE>(Symbol &Sym); 3029 template void PltSection::addEntry<ELF32BE>(Symbol &Sym); 3030 template void PltSection::addEntry<ELF64LE>(Symbol &Sym); 3031 template void PltSection::addEntry<ELF64BE>(Symbol &Sym); 3032 3033 template void MipsGotSection::build<ELF32LE>(); 3034 template void MipsGotSection::build<ELF32BE>(); 3035 template void MipsGotSection::build<ELF64LE>(); 3036 template void MipsGotSection::build<ELF64BE>(); 3037 3038 template class elf::MipsAbiFlagsSection<ELF32LE>; 3039 template class elf::MipsAbiFlagsSection<ELF32BE>; 3040 template class elf::MipsAbiFlagsSection<ELF64LE>; 3041 template class elf::MipsAbiFlagsSection<ELF64BE>; 3042 3043 template class elf::MipsOptionsSection<ELF32LE>; 3044 template class elf::MipsOptionsSection<ELF32BE>; 3045 template class elf::MipsOptionsSection<ELF64LE>; 3046 template class elf::MipsOptionsSection<ELF64BE>; 3047 3048 template class elf::MipsReginfoSection<ELF32LE>; 3049 template class elf::MipsReginfoSection<ELF32BE>; 3050 template class elf::MipsReginfoSection<ELF64LE>; 3051 template class elf::MipsReginfoSection<ELF64BE>; 3052 3053 template class elf::DynamicSection<ELF32LE>; 3054 template class elf::DynamicSection<ELF32BE>; 3055 template class elf::DynamicSection<ELF64LE>; 3056 template class elf::DynamicSection<ELF64BE>; 3057 3058 template class elf::RelocationSection<ELF32LE>; 3059 template class elf::RelocationSection<ELF32BE>; 3060 template class elf::RelocationSection<ELF64LE>; 3061 template class elf::RelocationSection<ELF64BE>; 3062 3063 template class elf::AndroidPackedRelocationSection<ELF32LE>; 3064 template class elf::AndroidPackedRelocationSection<ELF32BE>; 3065 template class elf::AndroidPackedRelocationSection<ELF64LE>; 3066 template class elf::AndroidPackedRelocationSection<ELF64BE>; 3067 3068 template class elf::RelrSection<ELF32LE>; 3069 template class elf::RelrSection<ELF32BE>; 3070 template class elf::RelrSection<ELF64LE>; 3071 template class elf::RelrSection<ELF64BE>; 3072 3073 template class elf::SymbolTableSection<ELF32LE>; 3074 template class elf::SymbolTableSection<ELF32BE>; 3075 template class elf::SymbolTableSection<ELF64LE>; 3076 template class elf::SymbolTableSection<ELF64BE>; 3077 3078 template class elf::VersionTableSection<ELF32LE>; 3079 template class elf::VersionTableSection<ELF32BE>; 3080 template class elf::VersionTableSection<ELF64LE>; 3081 template class elf::VersionTableSection<ELF64BE>; 3082 3083 template class elf::VersionNeedSection<ELF32LE>; 3084 template class elf::VersionNeedSection<ELF32BE>; 3085 template class elf::VersionNeedSection<ELF64LE>; 3086 template class elf::VersionNeedSection<ELF64BE>; 3087 3088 template class elf::VersionDefinitionSection<ELF32LE>; 3089 template class elf::VersionDefinitionSection<ELF32BE>; 3090 template class elf::VersionDefinitionSection<ELF64LE>; 3091 template class elf::VersionDefinitionSection<ELF64BE>; 3092