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/ADT/StringExtras.h" 34 #include "llvm/BinaryFormat/Dwarf.h" 35 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h" 36 #include "llvm/Object/ELFObjectFile.h" 37 #include "llvm/Support/Compression.h" 38 #include "llvm/Support/Endian.h" 39 #include "llvm/Support/LEB128.h" 40 #include "llvm/Support/MD5.h" 41 #include "llvm/Support/RandomNumberGenerator.h" 42 #include "llvm/Support/SHA1.h" 43 #include "llvm/Support/xxhash.h" 44 #include <cstdlib> 45 #include <thread> 46 47 using namespace llvm; 48 using namespace llvm::dwarf; 49 using namespace llvm::ELF; 50 using namespace llvm::object; 51 using namespace llvm::support; 52 53 using namespace lld; 54 using namespace lld::elf; 55 56 using llvm::support::endian::read32le; 57 using llvm::support::endian::write32le; 58 using llvm::support::endian::write64le; 59 60 constexpr size_t MergeNoTailSection::NumShards; 61 62 // Returns an LLD version string. 63 static ArrayRef<uint8_t> getVersion() { 64 // Check LLD_VERSION first for ease of testing. 65 // You can get consistent output by using the environment variable. 66 // This is only for testing. 67 StringRef S = getenv("LLD_VERSION"); 68 if (S.empty()) 69 S = Saver.save(Twine("Linker: ") + getLLDVersion()); 70 71 // +1 to include the terminating '\0'. 72 return {(const uint8_t *)S.data(), S.size() + 1}; 73 } 74 75 // Creates a .comment section containing LLD version info. 76 // With this feature, you can identify LLD-generated binaries easily 77 // by "readelf --string-dump .comment <file>". 78 // The returned object is a mergeable string section. 79 MergeInputSection *elf::createCommentSection() { 80 return make<MergeInputSection>(SHF_MERGE | SHF_STRINGS, SHT_PROGBITS, 1, 81 getVersion(), ".comment"); 82 } 83 84 // .MIPS.abiflags section. 85 template <class ELFT> 86 MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags Flags) 87 : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags"), 88 Flags(Flags) { 89 this->Entsize = sizeof(Elf_Mips_ABIFlags); 90 } 91 92 template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *Buf) { 93 memcpy(Buf, &Flags, sizeof(Flags)); 94 } 95 96 template <class ELFT> 97 MipsAbiFlagsSection<ELFT> *MipsAbiFlagsSection<ELFT>::create() { 98 Elf_Mips_ABIFlags Flags = {}; 99 bool Create = false; 100 101 for (InputSectionBase *Sec : InputSections) { 102 if (Sec->Type != SHT_MIPS_ABIFLAGS) 103 continue; 104 Sec->Live = false; 105 Create = true; 106 107 std::string Filename = toString(Sec->File); 108 const size_t Size = Sec->data().size(); 109 // Older version of BFD (such as the default FreeBSD linker) concatenate 110 // .MIPS.abiflags instead of merging. To allow for this case (or potential 111 // zero padding) we ignore everything after the first Elf_Mips_ABIFlags 112 if (Size < sizeof(Elf_Mips_ABIFlags)) { 113 error(Filename + ": invalid size of .MIPS.abiflags section: got " + 114 Twine(Size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags))); 115 return nullptr; 116 } 117 auto *S = reinterpret_cast<const Elf_Mips_ABIFlags *>(Sec->data().data()); 118 if (S->version != 0) { 119 error(Filename + ": unexpected .MIPS.abiflags version " + 120 Twine(S->version)); 121 return nullptr; 122 } 123 124 // LLD checks ISA compatibility in calcMipsEFlags(). Here we just 125 // select the highest number of ISA/Rev/Ext. 126 Flags.isa_level = std::max(Flags.isa_level, S->isa_level); 127 Flags.isa_rev = std::max(Flags.isa_rev, S->isa_rev); 128 Flags.isa_ext = std::max(Flags.isa_ext, S->isa_ext); 129 Flags.gpr_size = std::max(Flags.gpr_size, S->gpr_size); 130 Flags.cpr1_size = std::max(Flags.cpr1_size, S->cpr1_size); 131 Flags.cpr2_size = std::max(Flags.cpr2_size, S->cpr2_size); 132 Flags.ases |= S->ases; 133 Flags.flags1 |= S->flags1; 134 Flags.flags2 |= S->flags2; 135 Flags.fp_abi = elf::getMipsFpAbiFlag(Flags.fp_abi, S->fp_abi, Filename); 136 }; 137 138 if (Create) 139 return make<MipsAbiFlagsSection<ELFT>>(Flags); 140 return nullptr; 141 } 142 143 // .MIPS.options section. 144 template <class ELFT> 145 MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo Reginfo) 146 : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"), 147 Reginfo(Reginfo) { 148 this->Entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo); 149 } 150 151 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *Buf) { 152 auto *Options = reinterpret_cast<Elf_Mips_Options *>(Buf); 153 Options->kind = ODK_REGINFO; 154 Options->size = getSize(); 155 156 if (!Config->Relocatable) 157 Reginfo.ri_gp_value = In.MipsGot->getGp(); 158 memcpy(Buf + sizeof(Elf_Mips_Options), &Reginfo, sizeof(Reginfo)); 159 } 160 161 template <class ELFT> 162 MipsOptionsSection<ELFT> *MipsOptionsSection<ELFT>::create() { 163 // N64 ABI only. 164 if (!ELFT::Is64Bits) 165 return nullptr; 166 167 std::vector<InputSectionBase *> Sections; 168 for (InputSectionBase *Sec : InputSections) 169 if (Sec->Type == SHT_MIPS_OPTIONS) 170 Sections.push_back(Sec); 171 172 if (Sections.empty()) 173 return nullptr; 174 175 Elf_Mips_RegInfo Reginfo = {}; 176 for (InputSectionBase *Sec : Sections) { 177 Sec->Live = false; 178 179 std::string Filename = toString(Sec->File); 180 ArrayRef<uint8_t> D = Sec->data(); 181 182 while (!D.empty()) { 183 if (D.size() < sizeof(Elf_Mips_Options)) { 184 error(Filename + ": invalid size of .MIPS.options section"); 185 break; 186 } 187 188 auto *Opt = reinterpret_cast<const Elf_Mips_Options *>(D.data()); 189 if (Opt->kind == ODK_REGINFO) { 190 Reginfo.ri_gprmask |= Opt->getRegInfo().ri_gprmask; 191 Sec->getFile<ELFT>()->MipsGp0 = Opt->getRegInfo().ri_gp_value; 192 break; 193 } 194 195 if (!Opt->size) 196 fatal(Filename + ": zero option descriptor size"); 197 D = D.slice(Opt->size); 198 } 199 }; 200 201 return make<MipsOptionsSection<ELFT>>(Reginfo); 202 } 203 204 // MIPS .reginfo section. 205 template <class ELFT> 206 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo Reginfo) 207 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"), 208 Reginfo(Reginfo) { 209 this->Entsize = sizeof(Elf_Mips_RegInfo); 210 } 211 212 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *Buf) { 213 if (!Config->Relocatable) 214 Reginfo.ri_gp_value = In.MipsGot->getGp(); 215 memcpy(Buf, &Reginfo, sizeof(Reginfo)); 216 } 217 218 template <class ELFT> 219 MipsReginfoSection<ELFT> *MipsReginfoSection<ELFT>::create() { 220 // Section should be alive for O32 and N32 ABIs only. 221 if (ELFT::Is64Bits) 222 return nullptr; 223 224 std::vector<InputSectionBase *> Sections; 225 for (InputSectionBase *Sec : InputSections) 226 if (Sec->Type == SHT_MIPS_REGINFO) 227 Sections.push_back(Sec); 228 229 if (Sections.empty()) 230 return nullptr; 231 232 Elf_Mips_RegInfo Reginfo = {}; 233 for (InputSectionBase *Sec : Sections) { 234 Sec->Live = false; 235 236 if (Sec->data().size() != sizeof(Elf_Mips_RegInfo)) { 237 error(toString(Sec->File) + ": invalid size of .reginfo section"); 238 return nullptr; 239 } 240 241 auto *R = reinterpret_cast<const Elf_Mips_RegInfo *>(Sec->data().data()); 242 Reginfo.ri_gprmask |= R->ri_gprmask; 243 Sec->getFile<ELFT>()->MipsGp0 = R->ri_gp_value; 244 }; 245 246 return make<MipsReginfoSection<ELFT>>(Reginfo); 247 } 248 249 InputSection *elf::createInterpSection() { 250 // StringSaver guarantees that the returned string ends with '\0'. 251 StringRef S = Saver.save(Config->DynamicLinker); 252 ArrayRef<uint8_t> Contents = {(const uint8_t *)S.data(), S.size() + 1}; 253 254 auto *Sec = make<InputSection>(nullptr, SHF_ALLOC, SHT_PROGBITS, 1, Contents, 255 ".interp"); 256 Sec->Live = true; 257 return Sec; 258 } 259 260 Defined *elf::addSyntheticLocal(StringRef Name, uint8_t Type, uint64_t Value, 261 uint64_t Size, InputSectionBase &Section) { 262 auto *S = make<Defined>(Section.File, Name, STB_LOCAL, STV_DEFAULT, Type, 263 Value, Size, &Section); 264 if (In.SymTab) 265 In.SymTab->addSymbol(S); 266 return S; 267 } 268 269 static size_t getHashSize() { 270 switch (Config->BuildId) { 271 case BuildIdKind::Fast: 272 return 8; 273 case BuildIdKind::Md5: 274 case BuildIdKind::Uuid: 275 return 16; 276 case BuildIdKind::Sha1: 277 return 20; 278 case BuildIdKind::Hexstring: 279 return Config->BuildIdVector.size(); 280 default: 281 llvm_unreachable("unknown BuildIdKind"); 282 } 283 } 284 285 BuildIdSection::BuildIdSection() 286 : SyntheticSection(SHF_ALLOC, SHT_NOTE, 4, ".note.gnu.build-id"), 287 HashSize(getHashSize()) {} 288 289 void BuildIdSection::writeTo(uint8_t *Buf) { 290 write32(Buf, 4); // Name size 291 write32(Buf + 4, HashSize); // Content size 292 write32(Buf + 8, NT_GNU_BUILD_ID); // Type 293 memcpy(Buf + 12, "GNU", 4); // Name string 294 HashBuf = Buf + 16; 295 } 296 297 // Split one uint8 array into small pieces of uint8 arrays. 298 static std::vector<ArrayRef<uint8_t>> split(ArrayRef<uint8_t> Arr, 299 size_t ChunkSize) { 300 std::vector<ArrayRef<uint8_t>> Ret; 301 while (Arr.size() > ChunkSize) { 302 Ret.push_back(Arr.take_front(ChunkSize)); 303 Arr = Arr.drop_front(ChunkSize); 304 } 305 if (!Arr.empty()) 306 Ret.push_back(Arr); 307 return Ret; 308 } 309 310 // Computes a hash value of Data using a given hash function. 311 // In order to utilize multiple cores, we first split data into 1MB 312 // chunks, compute a hash for each chunk, and then compute a hash value 313 // of the hash values. 314 void BuildIdSection::computeHash( 315 llvm::ArrayRef<uint8_t> Data, 316 std::function<void(uint8_t *Dest, ArrayRef<uint8_t> Arr)> HashFn) { 317 std::vector<ArrayRef<uint8_t>> Chunks = split(Data, 1024 * 1024); 318 std::vector<uint8_t> Hashes(Chunks.size() * HashSize); 319 320 // Compute hash values. 321 parallelForEachN(0, Chunks.size(), [&](size_t I) { 322 HashFn(Hashes.data() + I * HashSize, Chunks[I]); 323 }); 324 325 // Write to the final output buffer. 326 HashFn(HashBuf, Hashes); 327 } 328 329 BssSection::BssSection(StringRef Name, uint64_t Size, uint32_t Alignment) 330 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, Alignment, Name) { 331 this->Bss = true; 332 this->Size = Size; 333 } 334 335 void BuildIdSection::writeBuildId(ArrayRef<uint8_t> Buf) { 336 switch (Config->BuildId) { 337 case BuildIdKind::Fast: 338 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) { 339 write64le(Dest, xxHash64(Arr)); 340 }); 341 break; 342 case BuildIdKind::Md5: 343 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) { 344 memcpy(Dest, MD5::hash(Arr).data(), 16); 345 }); 346 break; 347 case BuildIdKind::Sha1: 348 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) { 349 memcpy(Dest, SHA1::hash(Arr).data(), 20); 350 }); 351 break; 352 case BuildIdKind::Uuid: 353 if (auto EC = getRandomBytes(HashBuf, HashSize)) 354 error("entropy source failure: " + EC.message()); 355 break; 356 case BuildIdKind::Hexstring: 357 memcpy(HashBuf, Config->BuildIdVector.data(), Config->BuildIdVector.size()); 358 break; 359 default: 360 llvm_unreachable("unknown BuildIdKind"); 361 } 362 } 363 364 EhFrameSection::EhFrameSection() 365 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame") {} 366 367 // Search for an existing CIE record or create a new one. 368 // CIE records from input object files are uniquified by their contents 369 // and where their relocations point to. 370 template <class ELFT, class RelTy> 371 CieRecord *EhFrameSection::addCie(EhSectionPiece &Cie, ArrayRef<RelTy> Rels) { 372 Symbol *Personality = nullptr; 373 unsigned FirstRelI = Cie.FirstRelocation; 374 if (FirstRelI != (unsigned)-1) 375 Personality = 376 &Cie.Sec->template getFile<ELFT>()->getRelocTargetSym(Rels[FirstRelI]); 377 378 // Search for an existing CIE by CIE contents/relocation target pair. 379 CieRecord *&Rec = CieMap[{Cie.data(), Personality}]; 380 381 // If not found, create a new one. 382 if (!Rec) { 383 Rec = make<CieRecord>(); 384 Rec->Cie = &Cie; 385 CieRecords.push_back(Rec); 386 } 387 return Rec; 388 } 389 390 // There is one FDE per function. Returns true if a given FDE 391 // points to a live function. 392 template <class ELFT, class RelTy> 393 bool EhFrameSection::isFdeLive(EhSectionPiece &Fde, ArrayRef<RelTy> Rels) { 394 auto *Sec = cast<EhInputSection>(Fde.Sec); 395 unsigned FirstRelI = Fde.FirstRelocation; 396 397 // An FDE should point to some function because FDEs are to describe 398 // functions. That's however not always the case due to an issue of 399 // ld.gold with -r. ld.gold may discard only functions and leave their 400 // corresponding FDEs, which results in creating bad .eh_frame sections. 401 // To deal with that, we ignore such FDEs. 402 if (FirstRelI == (unsigned)-1) 403 return false; 404 405 const RelTy &Rel = Rels[FirstRelI]; 406 Symbol &B = Sec->template getFile<ELFT>()->getRelocTargetSym(Rel); 407 408 // FDEs for garbage-collected or merged-by-ICF sections are dead. 409 if (auto *D = dyn_cast<Defined>(&B)) 410 if (SectionBase *Sec = D->Section) 411 return Sec->Live; 412 return false; 413 } 414 415 // .eh_frame is a sequence of CIE or FDE records. In general, there 416 // is one CIE record per input object file which is followed by 417 // a list of FDEs. This function searches an existing CIE or create a new 418 // one and associates FDEs to the CIE. 419 template <class ELFT, class RelTy> 420 void EhFrameSection::addSectionAux(EhInputSection *Sec, ArrayRef<RelTy> Rels) { 421 OffsetToCie.clear(); 422 for (EhSectionPiece &Piece : Sec->Pieces) { 423 // The empty record is the end marker. 424 if (Piece.Size == 4) 425 return; 426 427 size_t Offset = Piece.InputOff; 428 uint32_t ID = read32(Piece.data().data() + 4); 429 if (ID == 0) { 430 OffsetToCie[Offset] = addCie<ELFT>(Piece, Rels); 431 continue; 432 } 433 434 uint32_t CieOffset = Offset + 4 - ID; 435 CieRecord *Rec = OffsetToCie[CieOffset]; 436 if (!Rec) 437 fatal(toString(Sec) + ": invalid CIE reference"); 438 439 if (!isFdeLive<ELFT>(Piece, Rels)) 440 continue; 441 Rec->Fdes.push_back(&Piece); 442 NumFdes++; 443 } 444 } 445 446 template <class ELFT> void EhFrameSection::addSection(InputSectionBase *C) { 447 auto *Sec = cast<EhInputSection>(C); 448 Sec->Parent = this; 449 450 Alignment = std::max(Alignment, Sec->Alignment); 451 Sections.push_back(Sec); 452 453 for (auto *DS : Sec->DependentSections) 454 DependentSections.push_back(DS); 455 456 if (Sec->Pieces.empty()) 457 return; 458 459 if (Sec->AreRelocsRela) 460 addSectionAux<ELFT>(Sec, Sec->template relas<ELFT>()); 461 else 462 addSectionAux<ELFT>(Sec, Sec->template rels<ELFT>()); 463 } 464 465 static void writeCieFde(uint8_t *Buf, ArrayRef<uint8_t> D) { 466 memcpy(Buf, D.data(), D.size()); 467 468 size_t Aligned = alignTo(D.size(), Config->Wordsize); 469 470 // Zero-clear trailing padding if it exists. 471 memset(Buf + D.size(), 0, Aligned - D.size()); 472 473 // Fix the size field. -4 since size does not include the size field itself. 474 write32(Buf, Aligned - 4); 475 } 476 477 void EhFrameSection::finalizeContents() { 478 assert(!this->Size); // Not finalized. 479 size_t Off = 0; 480 for (CieRecord *Rec : CieRecords) { 481 Rec->Cie->OutputOff = Off; 482 Off += alignTo(Rec->Cie->Size, Config->Wordsize); 483 484 for (EhSectionPiece *Fde : Rec->Fdes) { 485 Fde->OutputOff = Off; 486 Off += alignTo(Fde->Size, Config->Wordsize); 487 } 488 } 489 490 // The LSB standard does not allow a .eh_frame section with zero 491 // Call Frame Information records. glibc unwind-dw2-fde.c 492 // classify_object_over_fdes expects there is a CIE record length 0 as a 493 // terminator. Thus we add one unconditionally. 494 Off += 4; 495 496 this->Size = Off; 497 } 498 499 // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table 500 // to get an FDE from an address to which FDE is applied. This function 501 // returns a list of such pairs. 502 std::vector<EhFrameSection::FdeData> EhFrameSection::getFdeData() const { 503 uint8_t *Buf = getParent()->Loc + OutSecOff; 504 std::vector<FdeData> Ret; 505 506 uint64_t VA = In.EhFrameHdr->getVA(); 507 for (CieRecord *Rec : CieRecords) { 508 uint8_t Enc = getFdeEncoding(Rec->Cie); 509 for (EhSectionPiece *Fde : Rec->Fdes) { 510 uint64_t Pc = getFdePc(Buf, Fde->OutputOff, Enc); 511 uint64_t FdeVA = getParent()->Addr + Fde->OutputOff; 512 if (!isInt<32>(Pc - VA)) 513 fatal(toString(Fde->Sec) + ": PC offset is too large: 0x" + 514 Twine::utohexstr(Pc - VA)); 515 Ret.push_back({uint32_t(Pc - VA), uint32_t(FdeVA - VA)}); 516 } 517 } 518 519 // Sort the FDE list by their PC and uniqueify. Usually there is only 520 // one FDE for a PC (i.e. function), but if ICF merges two functions 521 // into one, there can be more than one FDEs pointing to the address. 522 auto Less = [](const FdeData &A, const FdeData &B) { 523 return A.PcRel < B.PcRel; 524 }; 525 std::stable_sort(Ret.begin(), Ret.end(), Less); 526 auto Eq = [](const FdeData &A, const FdeData &B) { 527 return A.PcRel == B.PcRel; 528 }; 529 Ret.erase(std::unique(Ret.begin(), Ret.end(), Eq), Ret.end()); 530 531 return Ret; 532 } 533 534 static uint64_t readFdeAddr(uint8_t *Buf, int Size) { 535 switch (Size) { 536 case DW_EH_PE_udata2: 537 return read16(Buf); 538 case DW_EH_PE_sdata2: 539 return (int16_t)read16(Buf); 540 case DW_EH_PE_udata4: 541 return read32(Buf); 542 case DW_EH_PE_sdata4: 543 return (int32_t)read32(Buf); 544 case DW_EH_PE_udata8: 545 case DW_EH_PE_sdata8: 546 return read64(Buf); 547 case DW_EH_PE_absptr: 548 return readUint(Buf); 549 } 550 fatal("unknown FDE size encoding"); 551 } 552 553 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to. 554 // We need it to create .eh_frame_hdr section. 555 uint64_t EhFrameSection::getFdePc(uint8_t *Buf, size_t FdeOff, 556 uint8_t Enc) const { 557 // The starting address to which this FDE applies is 558 // stored at FDE + 8 byte. 559 size_t Off = FdeOff + 8; 560 uint64_t Addr = readFdeAddr(Buf + Off, Enc & 0xf); 561 if ((Enc & 0x70) == DW_EH_PE_absptr) 562 return Addr; 563 if ((Enc & 0x70) == DW_EH_PE_pcrel) 564 return Addr + getParent()->Addr + Off; 565 fatal("unknown FDE size relative encoding"); 566 } 567 568 void EhFrameSection::writeTo(uint8_t *Buf) { 569 // Write CIE and FDE records. 570 for (CieRecord *Rec : CieRecords) { 571 size_t CieOffset = Rec->Cie->OutputOff; 572 writeCieFde(Buf + CieOffset, Rec->Cie->data()); 573 574 for (EhSectionPiece *Fde : Rec->Fdes) { 575 size_t Off = Fde->OutputOff; 576 writeCieFde(Buf + Off, Fde->data()); 577 578 // FDE's second word should have the offset to an associated CIE. 579 // Write it. 580 write32(Buf + Off + 4, Off + 4 - CieOffset); 581 } 582 } 583 584 // Apply relocations. .eh_frame section contents are not contiguous 585 // in the output buffer, but relocateAlloc() still works because 586 // getOffset() takes care of discontiguous section pieces. 587 for (EhInputSection *S : Sections) 588 S->relocateAlloc(Buf, nullptr); 589 } 590 591 GotSection::GotSection() 592 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 593 Target->GotEntrySize, ".got") { 594 // PPC64 saves the ElfSym::GlobalOffsetTable .TOC. as the first entry in the 595 // .got. If there are no references to .TOC. in the symbol table, 596 // ElfSym::GlobalOffsetTable will not be defined and we won't need to save 597 // .TOC. in the .got. When it is defined, we increase NumEntries by the number 598 // of entries used to emit ElfSym::GlobalOffsetTable. 599 if (ElfSym::GlobalOffsetTable && !Target->GotBaseSymInGotPlt) 600 NumEntries += Target->GotHeaderEntriesNum; 601 } 602 603 void GotSection::addEntry(Symbol &Sym) { 604 Sym.GotIndex = NumEntries; 605 ++NumEntries; 606 } 607 608 bool GotSection::addDynTlsEntry(Symbol &Sym) { 609 if (Sym.GlobalDynIndex != -1U) 610 return false; 611 Sym.GlobalDynIndex = NumEntries; 612 // Global Dynamic TLS entries take two GOT slots. 613 NumEntries += 2; 614 return true; 615 } 616 617 // Reserves TLS entries for a TLS module ID and a TLS block offset. 618 // In total it takes two GOT slots. 619 bool GotSection::addTlsIndex() { 620 if (TlsIndexOff != uint32_t(-1)) 621 return false; 622 TlsIndexOff = NumEntries * Config->Wordsize; 623 NumEntries += 2; 624 return true; 625 } 626 627 uint64_t GotSection::getGlobalDynAddr(const Symbol &B) const { 628 return this->getVA() + B.GlobalDynIndex * Config->Wordsize; 629 } 630 631 uint64_t GotSection::getGlobalDynOffset(const Symbol &B) const { 632 return B.GlobalDynIndex * Config->Wordsize; 633 } 634 635 void GotSection::finalizeContents() { 636 Size = NumEntries * Config->Wordsize; 637 } 638 639 bool GotSection::empty() const { 640 // We need to emit a GOT even if it's empty if there's a relocation that is 641 // relative to GOT(such as GOTOFFREL) or there's a symbol that points to a GOT 642 // (i.e. _GLOBAL_OFFSET_TABLE_) that the target defines relative to the .got. 643 return NumEntries == 0 && !HasGotOffRel && 644 !(ElfSym::GlobalOffsetTable && !Target->GotBaseSymInGotPlt); 645 } 646 647 void GotSection::writeTo(uint8_t *Buf) { 648 // Buf points to the start of this section's buffer, 649 // whereas InputSectionBase::relocateAlloc() expects its argument 650 // to point to the start of the output section. 651 Target->writeGotHeader(Buf); 652 relocateAlloc(Buf - OutSecOff, Buf - OutSecOff + Size); 653 } 654 655 static uint64_t getMipsPageAddr(uint64_t Addr) { 656 return (Addr + 0x8000) & ~0xffff; 657 } 658 659 static uint64_t getMipsPageCount(uint64_t Size) { 660 return (Size + 0xfffe) / 0xffff + 1; 661 } 662 663 MipsGotSection::MipsGotSection() 664 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16, 665 ".got") {} 666 667 void MipsGotSection::addEntry(InputFile &File, Symbol &Sym, int64_t Addend, 668 RelExpr Expr) { 669 FileGot &G = getGot(File); 670 if (Expr == R_MIPS_GOT_LOCAL_PAGE) { 671 if (const OutputSection *OS = Sym.getOutputSection()) 672 G.PagesMap.insert({OS, {}}); 673 else 674 G.Local16.insert({{nullptr, getMipsPageAddr(Sym.getVA(Addend))}, 0}); 675 } else if (Sym.isTls()) 676 G.Tls.insert({&Sym, 0}); 677 else if (Sym.IsPreemptible && Expr == R_ABS) 678 G.Relocs.insert({&Sym, 0}); 679 else if (Sym.IsPreemptible) 680 G.Global.insert({&Sym, 0}); 681 else if (Expr == R_MIPS_GOT_OFF32) 682 G.Local32.insert({{&Sym, Addend}, 0}); 683 else 684 G.Local16.insert({{&Sym, Addend}, 0}); 685 } 686 687 void MipsGotSection::addDynTlsEntry(InputFile &File, Symbol &Sym) { 688 getGot(File).DynTlsSymbols.insert({&Sym, 0}); 689 } 690 691 void MipsGotSection::addTlsIndex(InputFile &File) { 692 getGot(File).DynTlsSymbols.insert({nullptr, 0}); 693 } 694 695 size_t MipsGotSection::FileGot::getEntriesNum() const { 696 return getPageEntriesNum() + Local16.size() + Global.size() + Relocs.size() + 697 Tls.size() + DynTlsSymbols.size() * 2; 698 } 699 700 size_t MipsGotSection::FileGot::getPageEntriesNum() const { 701 size_t Num = 0; 702 for (const std::pair<const OutputSection *, FileGot::PageBlock> &P : PagesMap) 703 Num += P.second.Count; 704 return Num; 705 } 706 707 size_t MipsGotSection::FileGot::getIndexedEntriesNum() const { 708 size_t Count = getPageEntriesNum() + Local16.size() + Global.size(); 709 // If there are relocation-only entries in the GOT, TLS entries 710 // are allocated after them. TLS entries should be addressable 711 // by 16-bit index so count both reloc-only and TLS entries. 712 if (!Tls.empty() || !DynTlsSymbols.empty()) 713 Count += Relocs.size() + Tls.size() + DynTlsSymbols.size() * 2; 714 return Count; 715 } 716 717 MipsGotSection::FileGot &MipsGotSection::getGot(InputFile &F) { 718 if (!F.MipsGotIndex.hasValue()) { 719 Gots.emplace_back(); 720 Gots.back().File = &F; 721 F.MipsGotIndex = Gots.size() - 1; 722 } 723 return Gots[*F.MipsGotIndex]; 724 } 725 726 uint64_t MipsGotSection::getPageEntryOffset(const InputFile *F, 727 const Symbol &Sym, 728 int64_t Addend) const { 729 const FileGot &G = Gots[*F->MipsGotIndex]; 730 uint64_t Index = 0; 731 if (const OutputSection *OutSec = Sym.getOutputSection()) { 732 uint64_t SecAddr = getMipsPageAddr(OutSec->Addr); 733 uint64_t SymAddr = getMipsPageAddr(Sym.getVA(Addend)); 734 Index = G.PagesMap.lookup(OutSec).FirstIndex + (SymAddr - SecAddr) / 0xffff; 735 } else { 736 Index = G.Local16.lookup({nullptr, getMipsPageAddr(Sym.getVA(Addend))}); 737 } 738 return Index * Config->Wordsize; 739 } 740 741 uint64_t MipsGotSection::getSymEntryOffset(const InputFile *F, const Symbol &S, 742 int64_t Addend) const { 743 const FileGot &G = Gots[*F->MipsGotIndex]; 744 Symbol *Sym = const_cast<Symbol *>(&S); 745 if (Sym->isTls()) 746 return G.Tls.lookup(Sym) * Config->Wordsize; 747 if (Sym->IsPreemptible) 748 return G.Global.lookup(Sym) * Config->Wordsize; 749 return G.Local16.lookup({Sym, Addend}) * Config->Wordsize; 750 } 751 752 uint64_t MipsGotSection::getTlsIndexOffset(const InputFile *F) const { 753 const FileGot &G = Gots[*F->MipsGotIndex]; 754 return G.DynTlsSymbols.lookup(nullptr) * Config->Wordsize; 755 } 756 757 uint64_t MipsGotSection::getGlobalDynOffset(const InputFile *F, 758 const Symbol &S) const { 759 const FileGot &G = Gots[*F->MipsGotIndex]; 760 Symbol *Sym = const_cast<Symbol *>(&S); 761 return G.DynTlsSymbols.lookup(Sym) * Config->Wordsize; 762 } 763 764 const Symbol *MipsGotSection::getFirstGlobalEntry() const { 765 if (Gots.empty()) 766 return nullptr; 767 const FileGot &PrimGot = Gots.front(); 768 if (!PrimGot.Global.empty()) 769 return PrimGot.Global.front().first; 770 if (!PrimGot.Relocs.empty()) 771 return PrimGot.Relocs.front().first; 772 return nullptr; 773 } 774 775 unsigned MipsGotSection::getLocalEntriesNum() const { 776 if (Gots.empty()) 777 return HeaderEntriesNum; 778 return HeaderEntriesNum + Gots.front().getPageEntriesNum() + 779 Gots.front().Local16.size(); 780 } 781 782 bool MipsGotSection::tryMergeGots(FileGot &Dst, FileGot &Src, bool IsPrimary) { 783 FileGot Tmp = Dst; 784 set_union(Tmp.PagesMap, Src.PagesMap); 785 set_union(Tmp.Local16, Src.Local16); 786 set_union(Tmp.Global, Src.Global); 787 set_union(Tmp.Relocs, Src.Relocs); 788 set_union(Tmp.Tls, Src.Tls); 789 set_union(Tmp.DynTlsSymbols, Src.DynTlsSymbols); 790 791 size_t Count = IsPrimary ? HeaderEntriesNum : 0; 792 Count += Tmp.getIndexedEntriesNum(); 793 794 if (Count * Config->Wordsize > Config->MipsGotSize) 795 return false; 796 797 std::swap(Tmp, Dst); 798 return true; 799 } 800 801 void MipsGotSection::finalizeContents() { updateAllocSize(); } 802 803 bool MipsGotSection::updateAllocSize() { 804 Size = HeaderEntriesNum * Config->Wordsize; 805 for (const FileGot &G : Gots) 806 Size += G.getEntriesNum() * Config->Wordsize; 807 return false; 808 } 809 810 template <class ELFT> void MipsGotSection::build() { 811 if (Gots.empty()) 812 return; 813 814 std::vector<FileGot> MergedGots(1); 815 816 // For each GOT move non-preemptible symbols from the `Global` 817 // to `Local16` list. Preemptible symbol might become non-preemptible 818 // one if, for example, it gets a related copy relocation. 819 for (FileGot &Got : Gots) { 820 for (auto &P: Got.Global) 821 if (!P.first->IsPreemptible) 822 Got.Local16.insert({{P.first, 0}, 0}); 823 Got.Global.remove_if([&](const std::pair<Symbol *, size_t> &P) { 824 return !P.first->IsPreemptible; 825 }); 826 } 827 828 // For each GOT remove "reloc-only" entry if there is "global" 829 // entry for the same symbol. And add local entries which indexed 830 // using 32-bit value at the end of 16-bit entries. 831 for (FileGot &Got : Gots) { 832 Got.Relocs.remove_if([&](const std::pair<Symbol *, size_t> &P) { 833 return Got.Global.count(P.first); 834 }); 835 set_union(Got.Local16, Got.Local32); 836 Got.Local32.clear(); 837 } 838 839 // Evaluate number of "reloc-only" entries in the resulting GOT. 840 // To do that put all unique "reloc-only" and "global" entries 841 // from all GOTs to the future primary GOT. 842 FileGot *PrimGot = &MergedGots.front(); 843 for (FileGot &Got : Gots) { 844 set_union(PrimGot->Relocs, Got.Global); 845 set_union(PrimGot->Relocs, Got.Relocs); 846 Got.Relocs.clear(); 847 } 848 849 // Evaluate number of "page" entries in each GOT. 850 for (FileGot &Got : Gots) { 851 for (std::pair<const OutputSection *, FileGot::PageBlock> &P : 852 Got.PagesMap) { 853 const OutputSection *OS = P.first; 854 uint64_t SecSize = 0; 855 for (BaseCommand *Cmd : OS->SectionCommands) { 856 if (auto *ISD = dyn_cast<InputSectionDescription>(Cmd)) 857 for (InputSection *IS : ISD->Sections) { 858 uint64_t Off = alignTo(SecSize, IS->Alignment); 859 SecSize = Off + IS->getSize(); 860 } 861 } 862 P.second.Count = getMipsPageCount(SecSize); 863 } 864 } 865 866 // Merge GOTs. Try to join as much as possible GOTs but do not exceed 867 // maximum GOT size. At first, try to fill the primary GOT because 868 // the primary GOT can be accessed in the most effective way. If it 869 // is not possible, try to fill the last GOT in the list, and finally 870 // create a new GOT if both attempts failed. 871 for (FileGot &SrcGot : Gots) { 872 InputFile *File = SrcGot.File; 873 if (tryMergeGots(MergedGots.front(), SrcGot, true)) { 874 File->MipsGotIndex = 0; 875 } else { 876 // If this is the first time we failed to merge with the primary GOT, 877 // MergedGots.back() will also be the primary GOT. We must make sure not 878 // to try to merge again with IsPrimary=false, as otherwise, if the 879 // inputs are just right, we could allow the primary GOT to become 1 or 2 880 // words too big due to ignoring the header size. 881 if (MergedGots.size() == 1 || 882 !tryMergeGots(MergedGots.back(), SrcGot, false)) { 883 MergedGots.emplace_back(); 884 std::swap(MergedGots.back(), SrcGot); 885 } 886 File->MipsGotIndex = MergedGots.size() - 1; 887 } 888 } 889 std::swap(Gots, MergedGots); 890 891 // Reduce number of "reloc-only" entries in the primary GOT 892 // by substracting "global" entries exist in the primary GOT. 893 PrimGot = &Gots.front(); 894 PrimGot->Relocs.remove_if([&](const std::pair<Symbol *, size_t> &P) { 895 return PrimGot->Global.count(P.first); 896 }); 897 898 // Calculate indexes for each GOT entry. 899 size_t Index = HeaderEntriesNum; 900 for (FileGot &Got : Gots) { 901 Got.StartIndex = &Got == PrimGot ? 0 : Index; 902 for (std::pair<const OutputSection *, FileGot::PageBlock> &P : 903 Got.PagesMap) { 904 // For each output section referenced by GOT page relocations calculate 905 // and save into PagesMap an upper bound of MIPS GOT entries required 906 // to store page addresses of local symbols. We assume the worst case - 907 // each 64kb page of the output section has at least one GOT relocation 908 // against it. And take in account the case when the section intersects 909 // page boundaries. 910 P.second.FirstIndex = Index; 911 Index += P.second.Count; 912 } 913 for (auto &P: Got.Local16) 914 P.second = Index++; 915 for (auto &P: Got.Global) 916 P.second = Index++; 917 for (auto &P: Got.Relocs) 918 P.second = Index++; 919 for (auto &P: Got.Tls) 920 P.second = Index++; 921 for (auto &P: Got.DynTlsSymbols) { 922 P.second = Index; 923 Index += 2; 924 } 925 } 926 927 // Update Symbol::GotIndex field to use this 928 // value later in the `sortMipsSymbols` function. 929 for (auto &P : PrimGot->Global) 930 P.first->GotIndex = P.second; 931 for (auto &P : PrimGot->Relocs) 932 P.first->GotIndex = P.second; 933 934 // Create dynamic relocations. 935 for (FileGot &Got : Gots) { 936 // Create dynamic relocations for TLS entries. 937 for (std::pair<Symbol *, size_t> &P : Got.Tls) { 938 Symbol *S = P.first; 939 uint64_t Offset = P.second * Config->Wordsize; 940 if (S->IsPreemptible) 941 In.RelaDyn->addReloc(Target->TlsGotRel, this, Offset, S); 942 } 943 for (std::pair<Symbol *, size_t> &P : Got.DynTlsSymbols) { 944 Symbol *S = P.first; 945 uint64_t Offset = P.second * Config->Wordsize; 946 if (S == nullptr) { 947 if (!Config->Pic) 948 continue; 949 In.RelaDyn->addReloc(Target->TlsModuleIndexRel, this, Offset, S); 950 } else { 951 // When building a shared library we still need a dynamic relocation 952 // for the module index. Therefore only checking for 953 // S->IsPreemptible is not sufficient (this happens e.g. for 954 // thread-locals that have been marked as local through a linker script) 955 if (!S->IsPreemptible && !Config->Pic) 956 continue; 957 In.RelaDyn->addReloc(Target->TlsModuleIndexRel, this, Offset, S); 958 // However, we can skip writing the TLS offset reloc for non-preemptible 959 // symbols since it is known even in shared libraries 960 if (!S->IsPreemptible) 961 continue; 962 Offset += Config->Wordsize; 963 In.RelaDyn->addReloc(Target->TlsOffsetRel, this, Offset, S); 964 } 965 } 966 967 // Do not create dynamic relocations for non-TLS 968 // entries in the primary GOT. 969 if (&Got == PrimGot) 970 continue; 971 972 // Dynamic relocations for "global" entries. 973 for (const std::pair<Symbol *, size_t> &P : Got.Global) { 974 uint64_t Offset = P.second * Config->Wordsize; 975 In.RelaDyn->addReloc(Target->RelativeRel, this, Offset, P.first); 976 } 977 if (!Config->Pic) 978 continue; 979 // Dynamic relocations for "local" entries in case of PIC. 980 for (const std::pair<const OutputSection *, FileGot::PageBlock> &L : 981 Got.PagesMap) { 982 size_t PageCount = L.second.Count; 983 for (size_t PI = 0; PI < PageCount; ++PI) { 984 uint64_t Offset = (L.second.FirstIndex + PI) * Config->Wordsize; 985 In.RelaDyn->addReloc({Target->RelativeRel, this, Offset, L.first, 986 int64_t(PI * 0x10000)}); 987 } 988 } 989 for (const std::pair<GotEntry, size_t> &P : Got.Local16) { 990 uint64_t Offset = P.second * Config->Wordsize; 991 In.RelaDyn->addReloc({Target->RelativeRel, this, Offset, true, 992 P.first.first, P.first.second}); 993 } 994 } 995 } 996 997 bool MipsGotSection::empty() const { 998 // We add the .got section to the result for dynamic MIPS target because 999 // its address and properties are mentioned in the .dynamic section. 1000 return Config->Relocatable; 1001 } 1002 1003 uint64_t MipsGotSection::getGp(const InputFile *F) const { 1004 // For files without related GOT or files refer a primary GOT 1005 // returns "common" _gp value. For secondary GOTs calculate 1006 // individual _gp values. 1007 if (!F || !F->MipsGotIndex.hasValue() || *F->MipsGotIndex == 0) 1008 return ElfSym::MipsGp->getVA(0); 1009 return getVA() + Gots[*F->MipsGotIndex].StartIndex * Config->Wordsize + 1010 0x7ff0; 1011 } 1012 1013 void MipsGotSection::writeTo(uint8_t *Buf) { 1014 // Set the MSB of the second GOT slot. This is not required by any 1015 // MIPS ABI documentation, though. 1016 // 1017 // There is a comment in glibc saying that "The MSB of got[1] of a 1018 // gnu object is set to identify gnu objects," and in GNU gold it 1019 // says "the second entry will be used by some runtime loaders". 1020 // But how this field is being used is unclear. 1021 // 1022 // We are not really willing to mimic other linkers behaviors 1023 // without understanding why they do that, but because all files 1024 // generated by GNU tools have this special GOT value, and because 1025 // we've been doing this for years, it is probably a safe bet to 1026 // keep doing this for now. We really need to revisit this to see 1027 // if we had to do this. 1028 writeUint(Buf + Config->Wordsize, (uint64_t)1 << (Config->Wordsize * 8 - 1)); 1029 for (const FileGot &G : Gots) { 1030 auto Write = [&](size_t I, const Symbol *S, int64_t A) { 1031 uint64_t VA = A; 1032 if (S) { 1033 VA = S->getVA(A); 1034 if (S->StOther & STO_MIPS_MICROMIPS) 1035 VA |= 1; 1036 } 1037 writeUint(Buf + I * Config->Wordsize, VA); 1038 }; 1039 // Write 'page address' entries to the local part of the GOT. 1040 for (const std::pair<const OutputSection *, FileGot::PageBlock> &L : 1041 G.PagesMap) { 1042 size_t PageCount = L.second.Count; 1043 uint64_t FirstPageAddr = getMipsPageAddr(L.first->Addr); 1044 for (size_t PI = 0; PI < PageCount; ++PI) 1045 Write(L.second.FirstIndex + PI, nullptr, FirstPageAddr + PI * 0x10000); 1046 } 1047 // Local, global, TLS, reloc-only entries. 1048 // If TLS entry has a corresponding dynamic relocations, leave it 1049 // initialized by zero. Write down adjusted TLS symbol's values otherwise. 1050 // To calculate the adjustments use offsets for thread-local storage. 1051 // https://www.linux-mips.org/wiki/NPTL 1052 for (const std::pair<GotEntry, size_t> &P : G.Local16) 1053 Write(P.second, P.first.first, P.first.second); 1054 // Write VA to the primary GOT only. For secondary GOTs that 1055 // will be done by REL32 dynamic relocations. 1056 if (&G == &Gots.front()) 1057 for (const std::pair<const Symbol *, size_t> &P : G.Global) 1058 Write(P.second, P.first, 0); 1059 for (const std::pair<Symbol *, size_t> &P : G.Relocs) 1060 Write(P.second, P.first, 0); 1061 for (const std::pair<Symbol *, size_t> &P : G.Tls) 1062 Write(P.second, P.first, P.first->IsPreemptible ? 0 : -0x7000); 1063 for (const std::pair<Symbol *, size_t> &P : G.DynTlsSymbols) { 1064 if (P.first == nullptr && !Config->Pic) 1065 Write(P.second, nullptr, 1); 1066 else if (P.first && !P.first->IsPreemptible) { 1067 // If we are emitting PIC code with relocations we mustn't write 1068 // anything to the GOT here. When using Elf_Rel relocations the value 1069 // one will be treated as an addend and will cause crashes at runtime 1070 if (!Config->Pic) 1071 Write(P.second, nullptr, 1); 1072 Write(P.second + 1, P.first, -0x8000); 1073 } 1074 } 1075 } 1076 } 1077 1078 // On PowerPC the .plt section is used to hold the table of function addresses 1079 // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss 1080 // section. I don't know why we have a BSS style type for the section but it is 1081 // consitent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI. 1082 GotPltSection::GotPltSection() 1083 : SyntheticSection(SHF_ALLOC | SHF_WRITE, 1084 Config->EMachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS, 1085 Target->GotPltEntrySize, 1086 Config->EMachine == EM_PPC64 ? ".plt" : ".got.plt") {} 1087 1088 void GotPltSection::addEntry(Symbol &Sym) { 1089 assert(Sym.PltIndex == Entries.size()); 1090 Entries.push_back(&Sym); 1091 } 1092 1093 size_t GotPltSection::getSize() const { 1094 return (Target->GotPltHeaderEntriesNum + Entries.size()) * 1095 Target->GotPltEntrySize; 1096 } 1097 1098 void GotPltSection::writeTo(uint8_t *Buf) { 1099 Target->writeGotPltHeader(Buf); 1100 Buf += Target->GotPltHeaderEntriesNum * Target->GotPltEntrySize; 1101 for (const Symbol *B : Entries) { 1102 Target->writeGotPlt(Buf, *B); 1103 Buf += Config->Wordsize; 1104 } 1105 } 1106 1107 bool GotPltSection::empty() const { 1108 // We need to emit a GOT.PLT even if it's empty if there's a symbol that 1109 // references the _GLOBAL_OFFSET_TABLE_ and the Target defines the symbol 1110 // relative to the .got.plt section. 1111 return Entries.empty() && 1112 !(ElfSym::GlobalOffsetTable && Target->GotBaseSymInGotPlt); 1113 } 1114 1115 static StringRef getIgotPltName() { 1116 // On ARM the IgotPltSection is part of the GotSection. 1117 if (Config->EMachine == EM_ARM) 1118 return ".got"; 1119 1120 // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection 1121 // needs to be named the same. 1122 if (Config->EMachine == EM_PPC64) 1123 return ".plt"; 1124 1125 return ".got.plt"; 1126 } 1127 1128 // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit 1129 // with the IgotPltSection. 1130 IgotPltSection::IgotPltSection() 1131 : SyntheticSection(SHF_ALLOC | SHF_WRITE, 1132 Config->EMachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS, 1133 Target->GotPltEntrySize, getIgotPltName()) {} 1134 1135 void IgotPltSection::addEntry(Symbol &Sym) { 1136 Sym.IsInIgot = true; 1137 assert(Sym.PltIndex == Entries.size()); 1138 Entries.push_back(&Sym); 1139 } 1140 1141 size_t IgotPltSection::getSize() const { 1142 return Entries.size() * Target->GotPltEntrySize; 1143 } 1144 1145 void IgotPltSection::writeTo(uint8_t *Buf) { 1146 for (const Symbol *B : Entries) { 1147 Target->writeIgotPlt(Buf, *B); 1148 Buf += Config->Wordsize; 1149 } 1150 } 1151 1152 StringTableSection::StringTableSection(StringRef Name, bool Dynamic) 1153 : SyntheticSection(Dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, Name), 1154 Dynamic(Dynamic) { 1155 // ELF string tables start with a NUL byte. 1156 addString(""); 1157 } 1158 1159 // Adds a string to the string table. If HashIt is true we hash and check for 1160 // duplicates. It is optional because the name of global symbols are already 1161 // uniqued and hashing them again has a big cost for a small value: uniquing 1162 // them with some other string that happens to be the same. 1163 unsigned StringTableSection::addString(StringRef S, bool HashIt) { 1164 if (HashIt) { 1165 auto R = StringMap.insert(std::make_pair(S, this->Size)); 1166 if (!R.second) 1167 return R.first->second; 1168 } 1169 unsigned Ret = this->Size; 1170 this->Size = this->Size + S.size() + 1; 1171 Strings.push_back(S); 1172 return Ret; 1173 } 1174 1175 void StringTableSection::writeTo(uint8_t *Buf) { 1176 for (StringRef S : Strings) { 1177 memcpy(Buf, S.data(), S.size()); 1178 Buf[S.size()] = '\0'; 1179 Buf += S.size() + 1; 1180 } 1181 } 1182 1183 // Returns the number of version definition entries. Because the first entry 1184 // is for the version definition itself, it is the number of versioned symbols 1185 // plus one. Note that we don't support multiple versions yet. 1186 static unsigned getVerDefNum() { return Config->VersionDefinitions.size() + 1; } 1187 1188 template <class ELFT> 1189 DynamicSection<ELFT>::DynamicSection() 1190 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, Config->Wordsize, 1191 ".dynamic") { 1192 this->Entsize = ELFT::Is64Bits ? 16 : 8; 1193 1194 // .dynamic section is not writable on MIPS and on Fuchsia OS 1195 // which passes -z rodynamic. 1196 // See "Special Section" in Chapter 4 in the following document: 1197 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 1198 if (Config->EMachine == EM_MIPS || Config->ZRodynamic) 1199 this->Flags = SHF_ALLOC; 1200 1201 // Add strings to .dynstr early so that .dynstr's size will be 1202 // fixed early. 1203 for (StringRef S : Config->FilterList) 1204 addInt(DT_FILTER, In.DynStrTab->addString(S)); 1205 for (StringRef S : Config->AuxiliaryList) 1206 addInt(DT_AUXILIARY, In.DynStrTab->addString(S)); 1207 1208 if (!Config->Rpath.empty()) 1209 addInt(Config->EnableNewDtags ? DT_RUNPATH : DT_RPATH, 1210 In.DynStrTab->addString(Config->Rpath)); 1211 1212 for (InputFile *File : SharedFiles) { 1213 SharedFile<ELFT> *F = cast<SharedFile<ELFT>>(File); 1214 if (F->IsNeeded) 1215 addInt(DT_NEEDED, In.DynStrTab->addString(F->SoName)); 1216 } 1217 if (!Config->SoName.empty()) 1218 addInt(DT_SONAME, In.DynStrTab->addString(Config->SoName)); 1219 } 1220 1221 template <class ELFT> 1222 void DynamicSection<ELFT>::add(int32_t Tag, std::function<uint64_t()> Fn) { 1223 Entries.push_back({Tag, Fn}); 1224 } 1225 1226 template <class ELFT> 1227 void DynamicSection<ELFT>::addInt(int32_t Tag, uint64_t Val) { 1228 Entries.push_back({Tag, [=] { return Val; }}); 1229 } 1230 1231 template <class ELFT> 1232 void DynamicSection<ELFT>::addInSec(int32_t Tag, InputSection *Sec) { 1233 Entries.push_back({Tag, [=] { return Sec->getVA(0); }}); 1234 } 1235 1236 template <class ELFT> 1237 void DynamicSection<ELFT>::addInSecRelative(int32_t Tag, InputSection *Sec) { 1238 size_t TagOffset = Entries.size() * Entsize; 1239 Entries.push_back( 1240 {Tag, [=] { return Sec->getVA(0) - (getVA() + TagOffset); }}); 1241 } 1242 1243 template <class ELFT> 1244 void DynamicSection<ELFT>::addOutSec(int32_t Tag, OutputSection *Sec) { 1245 Entries.push_back({Tag, [=] { return Sec->Addr; }}); 1246 } 1247 1248 template <class ELFT> 1249 void DynamicSection<ELFT>::addSize(int32_t Tag, OutputSection *Sec) { 1250 Entries.push_back({Tag, [=] { return Sec->Size; }}); 1251 } 1252 1253 template <class ELFT> 1254 void DynamicSection<ELFT>::addSym(int32_t Tag, Symbol *Sym) { 1255 Entries.push_back({Tag, [=] { return Sym->getVA(); }}); 1256 } 1257 1258 // A Linker script may assign the RELA relocation sections to the same 1259 // output section. When this occurs we cannot just use the OutputSection 1260 // Size. Moreover the [DT_JMPREL, DT_JMPREL + DT_PLTRELSZ) is permitted to 1261 // overlap with the [DT_RELA, DT_RELA + DT_RELASZ). 1262 static uint64_t addPltRelSz() { 1263 size_t Size = In.RelaPlt->getSize(); 1264 if (In.RelaIplt->getParent() == In.RelaPlt->getParent() && 1265 In.RelaIplt->Name == In.RelaPlt->Name) 1266 Size += In.RelaIplt->getSize(); 1267 return Size; 1268 } 1269 1270 // Add remaining entries to complete .dynamic contents. 1271 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() { 1272 // Set DT_FLAGS and DT_FLAGS_1. 1273 uint32_t DtFlags = 0; 1274 uint32_t DtFlags1 = 0; 1275 if (Config->Bsymbolic) 1276 DtFlags |= DF_SYMBOLIC; 1277 if (Config->ZGlobal) 1278 DtFlags1 |= DF_1_GLOBAL; 1279 if (Config->ZInitfirst) 1280 DtFlags1 |= DF_1_INITFIRST; 1281 if (Config->ZInterpose) 1282 DtFlags1 |= DF_1_INTERPOSE; 1283 if (Config->ZNodefaultlib) 1284 DtFlags1 |= DF_1_NODEFLIB; 1285 if (Config->ZNodelete) 1286 DtFlags1 |= DF_1_NODELETE; 1287 if (Config->ZNodlopen) 1288 DtFlags1 |= DF_1_NOOPEN; 1289 if (Config->ZNow) { 1290 DtFlags |= DF_BIND_NOW; 1291 DtFlags1 |= DF_1_NOW; 1292 } 1293 if (Config->ZOrigin) { 1294 DtFlags |= DF_ORIGIN; 1295 DtFlags1 |= DF_1_ORIGIN; 1296 } 1297 if (!Config->ZText) 1298 DtFlags |= DF_TEXTREL; 1299 1300 if (DtFlags) 1301 addInt(DT_FLAGS, DtFlags); 1302 if (DtFlags1) 1303 addInt(DT_FLAGS_1, DtFlags1); 1304 1305 // DT_DEBUG is a pointer to debug informaion used by debuggers at runtime. We 1306 // need it for each process, so we don't write it for DSOs. The loader writes 1307 // the pointer into this entry. 1308 // 1309 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some 1310 // systems (currently only Fuchsia OS) provide other means to give the 1311 // debugger this information. Such systems may choose make .dynamic read-only. 1312 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG. 1313 if (!Config->Shared && !Config->Relocatable && !Config->ZRodynamic) 1314 addInt(DT_DEBUG, 0); 1315 1316 if (OutputSection *Sec = In.DynStrTab->getParent()) 1317 this->Link = Sec->SectionIndex; 1318 1319 if (!In.RelaDyn->empty()) { 1320 addInSec(In.RelaDyn->DynamicTag, In.RelaDyn); 1321 addSize(In.RelaDyn->SizeDynamicTag, In.RelaDyn->getParent()); 1322 1323 bool IsRela = Config->IsRela; 1324 addInt(IsRela ? DT_RELAENT : DT_RELENT, 1325 IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel)); 1326 1327 // MIPS dynamic loader does not support RELCOUNT tag. 1328 // The problem is in the tight relation between dynamic 1329 // relocations and GOT. So do not emit this tag on MIPS. 1330 if (Config->EMachine != EM_MIPS) { 1331 size_t NumRelativeRels = In.RelaDyn->getRelativeRelocCount(); 1332 if (Config->ZCombreloc && NumRelativeRels) 1333 addInt(IsRela ? DT_RELACOUNT : DT_RELCOUNT, NumRelativeRels); 1334 } 1335 } 1336 if (In.RelrDyn && !In.RelrDyn->Relocs.empty()) { 1337 addInSec(Config->UseAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR, 1338 In.RelrDyn); 1339 addSize(Config->UseAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ, 1340 In.RelrDyn->getParent()); 1341 addInt(Config->UseAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT, 1342 sizeof(Elf_Relr)); 1343 } 1344 // .rel[a].plt section usually consists of two parts, containing plt and 1345 // iplt relocations. It is possible to have only iplt relocations in the 1346 // output. In that case RelaPlt is empty and have zero offset, the same offset 1347 // as RelaIplt have. And we still want to emit proper dynamic tags for that 1348 // case, so here we always use RelaPlt as marker for the begining of 1349 // .rel[a].plt section. 1350 if (In.RelaPlt->getParent()->Live) { 1351 addInSec(DT_JMPREL, In.RelaPlt); 1352 Entries.push_back({DT_PLTRELSZ, addPltRelSz}); 1353 switch (Config->EMachine) { 1354 case EM_MIPS: 1355 addInSec(DT_MIPS_PLTGOT, In.GotPlt); 1356 break; 1357 case EM_SPARCV9: 1358 addInSec(DT_PLTGOT, In.Plt); 1359 break; 1360 default: 1361 addInSec(DT_PLTGOT, In.GotPlt); 1362 break; 1363 } 1364 addInt(DT_PLTREL, Config->IsRela ? DT_RELA : DT_REL); 1365 } 1366 1367 addInSec(DT_SYMTAB, In.DynSymTab); 1368 addInt(DT_SYMENT, sizeof(Elf_Sym)); 1369 addInSec(DT_STRTAB, In.DynStrTab); 1370 addInt(DT_STRSZ, In.DynStrTab->getSize()); 1371 if (!Config->ZText) 1372 addInt(DT_TEXTREL, 0); 1373 if (In.GnuHashTab) 1374 addInSec(DT_GNU_HASH, In.GnuHashTab); 1375 if (In.HashTab) 1376 addInSec(DT_HASH, In.HashTab); 1377 1378 if (Out::PreinitArray) { 1379 addOutSec(DT_PREINIT_ARRAY, Out::PreinitArray); 1380 addSize(DT_PREINIT_ARRAYSZ, Out::PreinitArray); 1381 } 1382 if (Out::InitArray) { 1383 addOutSec(DT_INIT_ARRAY, Out::InitArray); 1384 addSize(DT_INIT_ARRAYSZ, Out::InitArray); 1385 } 1386 if (Out::FiniArray) { 1387 addOutSec(DT_FINI_ARRAY, Out::FiniArray); 1388 addSize(DT_FINI_ARRAYSZ, Out::FiniArray); 1389 } 1390 1391 if (Symbol *B = Symtab->find(Config->Init)) 1392 if (B->isDefined()) 1393 addSym(DT_INIT, B); 1394 if (Symbol *B = Symtab->find(Config->Fini)) 1395 if (B->isDefined()) 1396 addSym(DT_FINI, B); 1397 1398 bool HasVerNeed = InX<ELFT>::VerNeed->getNeedNum() != 0; 1399 if (HasVerNeed || In.VerDef) 1400 addInSec(DT_VERSYM, InX<ELFT>::VerSym); 1401 if (In.VerDef) { 1402 addInSec(DT_VERDEF, In.VerDef); 1403 addInt(DT_VERDEFNUM, getVerDefNum()); 1404 } 1405 if (HasVerNeed) { 1406 addInSec(DT_VERNEED, InX<ELFT>::VerNeed); 1407 addInt(DT_VERNEEDNUM, InX<ELFT>::VerNeed->getNeedNum()); 1408 } 1409 1410 if (Config->EMachine == EM_MIPS) { 1411 addInt(DT_MIPS_RLD_VERSION, 1); 1412 addInt(DT_MIPS_FLAGS, RHF_NOTPOT); 1413 addInt(DT_MIPS_BASE_ADDRESS, Target->getImageBase()); 1414 addInt(DT_MIPS_SYMTABNO, In.DynSymTab->getNumSymbols()); 1415 1416 add(DT_MIPS_LOCAL_GOTNO, [] { return In.MipsGot->getLocalEntriesNum(); }); 1417 1418 if (const Symbol *B = In.MipsGot->getFirstGlobalEntry()) 1419 addInt(DT_MIPS_GOTSYM, B->DynsymIndex); 1420 else 1421 addInt(DT_MIPS_GOTSYM, In.DynSymTab->getNumSymbols()); 1422 addInSec(DT_PLTGOT, In.MipsGot); 1423 if (In.MipsRldMap) { 1424 if (!Config->Pie) 1425 addInSec(DT_MIPS_RLD_MAP, In.MipsRldMap); 1426 // Store the offset to the .rld_map section 1427 // relative to the address of the tag. 1428 addInSecRelative(DT_MIPS_RLD_MAP_REL, In.MipsRldMap); 1429 } 1430 } 1431 1432 // Glink dynamic tag is required by the V2 abi if the plt section isn't empty. 1433 if (Config->EMachine == EM_PPC64 && !In.Plt->empty()) { 1434 // The Glink tag points to 32 bytes before the first lazy symbol resolution 1435 // stub, which starts directly after the header. 1436 Entries.push_back({DT_PPC64_GLINK, [=] { 1437 unsigned Offset = Target->PltHeaderSize - 32; 1438 return In.Plt->getVA(0) + Offset; 1439 }}); 1440 } 1441 1442 addInt(DT_NULL, 0); 1443 1444 getParent()->Link = this->Link; 1445 this->Size = Entries.size() * this->Entsize; 1446 } 1447 1448 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *Buf) { 1449 auto *P = reinterpret_cast<Elf_Dyn *>(Buf); 1450 1451 for (std::pair<int32_t, std::function<uint64_t()>> &KV : Entries) { 1452 P->d_tag = KV.first; 1453 P->d_un.d_val = KV.second(); 1454 ++P; 1455 } 1456 } 1457 1458 uint64_t DynamicReloc::getOffset() const { 1459 return InputSec->getVA(OffsetInSec); 1460 } 1461 1462 int64_t DynamicReloc::computeAddend() const { 1463 if (UseSymVA) 1464 return Sym->getVA(Addend); 1465 if (!OutputSec) 1466 return Addend; 1467 // See the comment in the DynamicReloc ctor. 1468 return getMipsPageAddr(OutputSec->Addr) + Addend; 1469 } 1470 1471 uint32_t DynamicReloc::getSymIndex() const { 1472 if (Sym && !UseSymVA) 1473 return Sym->DynsymIndex; 1474 return 0; 1475 } 1476 1477 RelocationBaseSection::RelocationBaseSection(StringRef Name, uint32_t Type, 1478 int32_t DynamicTag, 1479 int32_t SizeDynamicTag) 1480 : SyntheticSection(SHF_ALLOC, Type, Config->Wordsize, Name), 1481 DynamicTag(DynamicTag), SizeDynamicTag(SizeDynamicTag) {} 1482 1483 void RelocationBaseSection::addReloc(RelType DynType, InputSectionBase *IS, 1484 uint64_t OffsetInSec, Symbol *Sym) { 1485 addReloc({DynType, IS, OffsetInSec, false, Sym, 0}); 1486 } 1487 1488 void RelocationBaseSection::addReloc(RelType DynType, 1489 InputSectionBase *InputSec, 1490 uint64_t OffsetInSec, Symbol *Sym, 1491 int64_t Addend, RelExpr Expr, 1492 RelType Type) { 1493 // Write the addends to the relocated address if required. We skip 1494 // it if the written value would be zero. 1495 if (Config->WriteAddends && (Expr != R_ADDEND || Addend != 0)) 1496 InputSec->Relocations.push_back({Expr, Type, OffsetInSec, Addend, Sym}); 1497 addReloc({DynType, InputSec, OffsetInSec, Expr != R_ADDEND, Sym, Addend}); 1498 } 1499 1500 void RelocationBaseSection::addReloc(const DynamicReloc &Reloc) { 1501 if (Reloc.Type == Target->RelativeRel) 1502 ++NumRelativeRelocs; 1503 Relocs.push_back(Reloc); 1504 } 1505 1506 void RelocationBaseSection::finalizeContents() { 1507 // When linking glibc statically, .rel{,a}.plt contains R_*_IRELATIVE 1508 // relocations due to IFUNC (e.g. strcpy). sh_link will be set to 0 in that 1509 // case. 1510 InputSection *SymTab = Config->Relocatable ? In.SymTab : In.DynSymTab; 1511 if (SymTab && SymTab->getParent()) 1512 getParent()->Link = SymTab->getParent()->SectionIndex; 1513 else 1514 getParent()->Link = 0; 1515 1516 if (In.RelaIplt == this || In.RelaPlt == this) 1517 getParent()->Info = In.GotPlt->getParent()->SectionIndex; 1518 } 1519 1520 RelrBaseSection::RelrBaseSection() 1521 : SyntheticSection(SHF_ALLOC, 1522 Config->UseAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR, 1523 Config->Wordsize, ".relr.dyn") {} 1524 1525 template <class ELFT> 1526 static void encodeDynamicReloc(typename ELFT::Rela *P, 1527 const DynamicReloc &Rel) { 1528 if (Config->IsRela) 1529 P->r_addend = Rel.computeAddend(); 1530 P->r_offset = Rel.getOffset(); 1531 P->setSymbolAndType(Rel.getSymIndex(), Rel.Type, Config->IsMips64EL); 1532 } 1533 1534 template <class ELFT> 1535 RelocationSection<ELFT>::RelocationSection(StringRef Name, bool Sort) 1536 : RelocationBaseSection(Name, Config->IsRela ? SHT_RELA : SHT_REL, 1537 Config->IsRela ? DT_RELA : DT_REL, 1538 Config->IsRela ? DT_RELASZ : DT_RELSZ), 1539 Sort(Sort) { 1540 this->Entsize = Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1541 } 1542 1543 static bool compRelocations(const DynamicReloc &A, const DynamicReloc &B) { 1544 bool AIsRel = A.Type == Target->RelativeRel; 1545 bool BIsRel = B.Type == Target->RelativeRel; 1546 if (AIsRel != BIsRel) 1547 return AIsRel; 1548 return A.getSymIndex() < B.getSymIndex(); 1549 } 1550 1551 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *Buf) { 1552 if (Sort) 1553 std::stable_sort(Relocs.begin(), Relocs.end(), compRelocations); 1554 1555 for (const DynamicReloc &Rel : Relocs) { 1556 encodeDynamicReloc<ELFT>(reinterpret_cast<Elf_Rela *>(Buf), Rel); 1557 Buf += Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1558 } 1559 } 1560 1561 template <class ELFT> unsigned RelocationSection<ELFT>::getRelocOffset() { 1562 return this->Entsize * Relocs.size(); 1563 } 1564 1565 template <class ELFT> 1566 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection( 1567 StringRef Name) 1568 : RelocationBaseSection( 1569 Name, Config->IsRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL, 1570 Config->IsRela ? DT_ANDROID_RELA : DT_ANDROID_REL, 1571 Config->IsRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ) { 1572 this->Entsize = 1; 1573 } 1574 1575 template <class ELFT> 1576 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() { 1577 // This function computes the contents of an Android-format packed relocation 1578 // section. 1579 // 1580 // This format compresses relocations by using relocation groups to factor out 1581 // fields that are common between relocations and storing deltas from previous 1582 // relocations in SLEB128 format (which has a short representation for small 1583 // numbers). A good example of a relocation type with common fields is 1584 // R_*_RELATIVE, which is normally used to represent function pointers in 1585 // vtables. In the REL format, each relative relocation has the same r_info 1586 // field, and is only different from other relative relocations in terms of 1587 // the r_offset field. By sorting relocations by offset, grouping them by 1588 // r_info and representing each relocation with only the delta from the 1589 // previous offset, each 8-byte relocation can be compressed to as little as 1 1590 // byte (or less with run-length encoding). This relocation packer was able to 1591 // reduce the size of the relocation section in an Android Chromium DSO from 1592 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size. 1593 // 1594 // A relocation section consists of a header containing the literal bytes 1595 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two 1596 // elements are the total number of relocations in the section and an initial 1597 // r_offset value. The remaining elements define a sequence of relocation 1598 // groups. Each relocation group starts with a header consisting of the 1599 // following elements: 1600 // 1601 // - the number of relocations in the relocation group 1602 // - flags for the relocation group 1603 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta 1604 // for each relocation in the group. 1605 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info 1606 // field for each relocation in the group. 1607 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and 1608 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for 1609 // each relocation in the group. 1610 // 1611 // Following the relocation group header are descriptions of each of the 1612 // relocations in the group. They consist of the following elements: 1613 // 1614 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset 1615 // delta for this relocation. 1616 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info 1617 // field for this relocation. 1618 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and 1619 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for 1620 // this relocation. 1621 1622 size_t OldSize = RelocData.size(); 1623 1624 RelocData = {'A', 'P', 'S', '2'}; 1625 raw_svector_ostream OS(RelocData); 1626 auto Add = [&](int64_t V) { encodeSLEB128(V, OS); }; 1627 1628 // The format header includes the number of relocations and the initial 1629 // offset (we set this to zero because the first relocation group will 1630 // perform the initial adjustment). 1631 Add(Relocs.size()); 1632 Add(0); 1633 1634 std::vector<Elf_Rela> Relatives, NonRelatives; 1635 1636 for (const DynamicReloc &Rel : Relocs) { 1637 Elf_Rela R; 1638 encodeDynamicReloc<ELFT>(&R, Rel); 1639 1640 if (R.getType(Config->IsMips64EL) == Target->RelativeRel) 1641 Relatives.push_back(R); 1642 else 1643 NonRelatives.push_back(R); 1644 } 1645 1646 llvm::sort(Relatives, [](const Elf_Rel &A, const Elf_Rel &B) { 1647 return A.r_offset < B.r_offset; 1648 }); 1649 1650 // Try to find groups of relative relocations which are spaced one word 1651 // apart from one another. These generally correspond to vtable entries. The 1652 // format allows these groups to be encoded using a sort of run-length 1653 // encoding, but each group will cost 7 bytes in addition to the offset from 1654 // the previous group, so it is only profitable to do this for groups of 1655 // size 8 or larger. 1656 std::vector<Elf_Rela> UngroupedRelatives; 1657 std::vector<std::vector<Elf_Rela>> RelativeGroups; 1658 for (auto I = Relatives.begin(), E = Relatives.end(); I != E;) { 1659 std::vector<Elf_Rela> Group; 1660 do { 1661 Group.push_back(*I++); 1662 } while (I != E && (I - 1)->r_offset + Config->Wordsize == I->r_offset); 1663 1664 if (Group.size() < 8) 1665 UngroupedRelatives.insert(UngroupedRelatives.end(), Group.begin(), 1666 Group.end()); 1667 else 1668 RelativeGroups.emplace_back(std::move(Group)); 1669 } 1670 1671 unsigned HasAddendIfRela = 1672 Config->IsRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0; 1673 1674 uint64_t Offset = 0; 1675 uint64_t Addend = 0; 1676 1677 // Emit the run-length encoding for the groups of adjacent relative 1678 // relocations. Each group is represented using two groups in the packed 1679 // format. The first is used to set the current offset to the start of the 1680 // group (and also encodes the first relocation), and the second encodes the 1681 // remaining relocations. 1682 for (std::vector<Elf_Rela> &G : RelativeGroups) { 1683 // The first relocation in the group. 1684 Add(1); 1685 Add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG | 1686 RELOCATION_GROUPED_BY_INFO_FLAG | HasAddendIfRela); 1687 Add(G[0].r_offset - Offset); 1688 Add(Target->RelativeRel); 1689 if (Config->IsRela) { 1690 Add(G[0].r_addend - Addend); 1691 Addend = G[0].r_addend; 1692 } 1693 1694 // The remaining relocations. 1695 Add(G.size() - 1); 1696 Add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG | 1697 RELOCATION_GROUPED_BY_INFO_FLAG | HasAddendIfRela); 1698 Add(Config->Wordsize); 1699 Add(Target->RelativeRel); 1700 if (Config->IsRela) { 1701 for (auto I = G.begin() + 1, E = G.end(); I != E; ++I) { 1702 Add(I->r_addend - Addend); 1703 Addend = I->r_addend; 1704 } 1705 } 1706 1707 Offset = G.back().r_offset; 1708 } 1709 1710 // Now the ungrouped relatives. 1711 if (!UngroupedRelatives.empty()) { 1712 Add(UngroupedRelatives.size()); 1713 Add(RELOCATION_GROUPED_BY_INFO_FLAG | HasAddendIfRela); 1714 Add(Target->RelativeRel); 1715 for (Elf_Rela &R : UngroupedRelatives) { 1716 Add(R.r_offset - Offset); 1717 Offset = R.r_offset; 1718 if (Config->IsRela) { 1719 Add(R.r_addend - Addend); 1720 Addend = R.r_addend; 1721 } 1722 } 1723 } 1724 1725 // Finally the non-relative relocations. 1726 llvm::sort(NonRelatives, [](const Elf_Rela &A, const Elf_Rela &B) { 1727 return A.r_offset < B.r_offset; 1728 }); 1729 if (!NonRelatives.empty()) { 1730 Add(NonRelatives.size()); 1731 Add(HasAddendIfRela); 1732 for (Elf_Rela &R : NonRelatives) { 1733 Add(R.r_offset - Offset); 1734 Offset = R.r_offset; 1735 Add(R.r_info); 1736 if (Config->IsRela) { 1737 Add(R.r_addend - Addend); 1738 Addend = R.r_addend; 1739 } 1740 } 1741 } 1742 1743 // Don't allow the section to shrink; otherwise the size of the section can 1744 // oscillate infinitely. 1745 if (RelocData.size() < OldSize) 1746 RelocData.append(OldSize - RelocData.size(), 0); 1747 1748 // Returns whether the section size changed. We need to keep recomputing both 1749 // section layout and the contents of this section until the size converges 1750 // because changing this section's size can affect section layout, which in 1751 // turn can affect the sizes of the LEB-encoded integers stored in this 1752 // section. 1753 return RelocData.size() != OldSize; 1754 } 1755 1756 template <class ELFT> RelrSection<ELFT>::RelrSection() { 1757 this->Entsize = Config->Wordsize; 1758 } 1759 1760 template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() { 1761 // This function computes the contents of an SHT_RELR packed relocation 1762 // section. 1763 // 1764 // Proposal for adding SHT_RELR sections to generic-abi is here: 1765 // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg 1766 // 1767 // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks 1768 // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ] 1769 // 1770 // i.e. start with an address, followed by any number of bitmaps. The address 1771 // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63 1772 // relocations each, at subsequent offsets following the last address entry. 1773 // 1774 // The bitmap entries must have 1 in the least significant bit. The assumption 1775 // here is that an address cannot have 1 in lsb. Odd addresses are not 1776 // supported. 1777 // 1778 // Excluding the least significant bit in the bitmap, each non-zero bit in 1779 // the bitmap represents a relocation to be applied to a corresponding machine 1780 // word that follows the base address word. The second least significant bit 1781 // represents the machine word immediately following the initial address, and 1782 // each bit that follows represents the next word, in linear order. As such, 1783 // a single bitmap can encode up to 31 relocations in a 32-bit object, and 1784 // 63 relocations in a 64-bit object. 1785 // 1786 // This encoding has a couple of interesting properties: 1787 // 1. Looking at any entry, it is clear whether it's an address or a bitmap: 1788 // even means address, odd means bitmap. 1789 // 2. Just a simple list of addresses is a valid encoding. 1790 1791 size_t OldSize = RelrRelocs.size(); 1792 RelrRelocs.clear(); 1793 1794 // Same as Config->Wordsize but faster because this is a compile-time 1795 // constant. 1796 const size_t Wordsize = sizeof(typename ELFT::uint); 1797 1798 // Number of bits to use for the relocation offsets bitmap. 1799 // Must be either 63 or 31. 1800 const size_t NBits = Wordsize * 8 - 1; 1801 1802 // Get offsets for all relative relocations and sort them. 1803 std::vector<uint64_t> Offsets; 1804 for (const RelativeReloc &Rel : Relocs) 1805 Offsets.push_back(Rel.getOffset()); 1806 llvm::sort(Offsets.begin(), Offsets.end()); 1807 1808 // For each leading relocation, find following ones that can be folded 1809 // as a bitmap and fold them. 1810 for (size_t I = 0, E = Offsets.size(); I < E;) { 1811 // Add a leading relocation. 1812 RelrRelocs.push_back(Elf_Relr(Offsets[I])); 1813 uint64_t Base = Offsets[I] + Wordsize; 1814 ++I; 1815 1816 // Find foldable relocations to construct bitmaps. 1817 while (I < E) { 1818 uint64_t Bitmap = 0; 1819 1820 while (I < E) { 1821 uint64_t Delta = Offsets[I] - Base; 1822 1823 // If it is too far, it cannot be folded. 1824 if (Delta >= NBits * Wordsize) 1825 break; 1826 1827 // If it is not a multiple of wordsize away, it cannot be folded. 1828 if (Delta % Wordsize) 1829 break; 1830 1831 // Fold it. 1832 Bitmap |= 1ULL << (Delta / Wordsize); 1833 ++I; 1834 } 1835 1836 if (!Bitmap) 1837 break; 1838 1839 RelrRelocs.push_back(Elf_Relr((Bitmap << 1) | 1)); 1840 Base += NBits * Wordsize; 1841 } 1842 } 1843 1844 return RelrRelocs.size() != OldSize; 1845 } 1846 1847 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &StrTabSec) 1848 : SyntheticSection(StrTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0, 1849 StrTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB, 1850 Config->Wordsize, 1851 StrTabSec.isDynamic() ? ".dynsym" : ".symtab"), 1852 StrTabSec(StrTabSec) {} 1853 1854 // Orders symbols according to their positions in the GOT, 1855 // in compliance with MIPS ABI rules. 1856 // See "Global Offset Table" in Chapter 5 in the following document 1857 // for detailed description: 1858 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 1859 static bool sortMipsSymbols(const SymbolTableEntry &L, 1860 const SymbolTableEntry &R) { 1861 // Sort entries related to non-local preemptible symbols by GOT indexes. 1862 // All other entries go to the beginning of a dynsym in arbitrary order. 1863 if (L.Sym->isInGot() && R.Sym->isInGot()) 1864 return L.Sym->GotIndex < R.Sym->GotIndex; 1865 if (!L.Sym->isInGot() && !R.Sym->isInGot()) 1866 return false; 1867 return !L.Sym->isInGot(); 1868 } 1869 1870 void SymbolTableBaseSection::finalizeContents() { 1871 if (OutputSection *Sec = StrTabSec.getParent()) 1872 getParent()->Link = Sec->SectionIndex; 1873 1874 if (this->Type != SHT_DYNSYM) { 1875 sortSymTabSymbols(); 1876 return; 1877 } 1878 1879 // If it is a .dynsym, there should be no local symbols, but we need 1880 // to do a few things for the dynamic linker. 1881 1882 // Section's Info field has the index of the first non-local symbol. 1883 // Because the first symbol entry is a null entry, 1 is the first. 1884 getParent()->Info = 1; 1885 1886 if (In.GnuHashTab) { 1887 // NB: It also sorts Symbols to meet the GNU hash table requirements. 1888 In.GnuHashTab->addSymbols(Symbols); 1889 } else if (Config->EMachine == EM_MIPS) { 1890 std::stable_sort(Symbols.begin(), Symbols.end(), sortMipsSymbols); 1891 } 1892 1893 size_t I = 0; 1894 for (const SymbolTableEntry &S : Symbols) 1895 S.Sym->DynsymIndex = ++I; 1896 } 1897 1898 // The ELF spec requires that all local symbols precede global symbols, so we 1899 // sort symbol entries in this function. (For .dynsym, we don't do that because 1900 // symbols for dynamic linking are inherently all globals.) 1901 // 1902 // Aside from above, we put local symbols in groups starting with the STT_FILE 1903 // symbol. That is convenient for purpose of identifying where are local symbols 1904 // coming from. 1905 void SymbolTableBaseSection::sortSymTabSymbols() { 1906 // Move all local symbols before global symbols. 1907 auto E = std::stable_partition( 1908 Symbols.begin(), Symbols.end(), [](const SymbolTableEntry &S) { 1909 return S.Sym->isLocal() || S.Sym->computeBinding() == STB_LOCAL; 1910 }); 1911 size_t NumLocals = E - Symbols.begin(); 1912 getParent()->Info = NumLocals + 1; 1913 1914 // We want to group the local symbols by file. For that we rebuild the local 1915 // part of the symbols vector. We do not need to care about the STT_FILE 1916 // symbols, they are already naturally placed first in each group. That 1917 // happens because STT_FILE is always the first symbol in the object and hence 1918 // precede all other local symbols we add for a file. 1919 MapVector<InputFile *, std::vector<SymbolTableEntry>> Arr; 1920 for (const SymbolTableEntry &S : llvm::make_range(Symbols.begin(), E)) 1921 Arr[S.Sym->File].push_back(S); 1922 1923 auto I = Symbols.begin(); 1924 for (std::pair<InputFile *, std::vector<SymbolTableEntry>> &P : Arr) 1925 for (SymbolTableEntry &Entry : P.second) 1926 *I++ = Entry; 1927 } 1928 1929 void SymbolTableBaseSection::addSymbol(Symbol *B) { 1930 // Adding a local symbol to a .dynsym is a bug. 1931 assert(this->Type != SHT_DYNSYM || !B->isLocal()); 1932 1933 bool HashIt = B->isLocal(); 1934 Symbols.push_back({B, StrTabSec.addString(B->getName(), HashIt)}); 1935 } 1936 1937 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *Sym) { 1938 // Initializes symbol lookup tables lazily. This is used only 1939 // for -r or -emit-relocs. 1940 llvm::call_once(OnceFlag, [&] { 1941 SymbolIndexMap.reserve(Symbols.size()); 1942 size_t I = 0; 1943 for (const SymbolTableEntry &E : Symbols) { 1944 if (E.Sym->Type == STT_SECTION) 1945 SectionIndexMap[E.Sym->getOutputSection()] = ++I; 1946 else 1947 SymbolIndexMap[E.Sym] = ++I; 1948 } 1949 }); 1950 1951 // Section symbols are mapped based on their output sections 1952 // to maintain their semantics. 1953 if (Sym->Type == STT_SECTION) 1954 return SectionIndexMap.lookup(Sym->getOutputSection()); 1955 return SymbolIndexMap.lookup(Sym); 1956 } 1957 1958 template <class ELFT> 1959 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &StrTabSec) 1960 : SymbolTableBaseSection(StrTabSec) { 1961 this->Entsize = sizeof(Elf_Sym); 1962 } 1963 1964 static BssSection *getCommonSec(Symbol *Sym) { 1965 if (!Config->DefineCommon) 1966 if (auto *D = dyn_cast<Defined>(Sym)) 1967 return dyn_cast_or_null<BssSection>(D->Section); 1968 return nullptr; 1969 } 1970 1971 static uint32_t getSymSectionIndex(Symbol *Sym) { 1972 if (getCommonSec(Sym)) 1973 return SHN_COMMON; 1974 if (!isa<Defined>(Sym) || Sym->NeedsPltAddr) 1975 return SHN_UNDEF; 1976 if (const OutputSection *OS = Sym->getOutputSection()) 1977 return OS->SectionIndex >= SHN_LORESERVE ? (uint32_t)SHN_XINDEX 1978 : OS->SectionIndex; 1979 return SHN_ABS; 1980 } 1981 1982 // Write the internal symbol table contents to the output symbol table. 1983 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *Buf) { 1984 // The first entry is a null entry as per the ELF spec. 1985 memset(Buf, 0, sizeof(Elf_Sym)); 1986 Buf += sizeof(Elf_Sym); 1987 1988 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf); 1989 1990 for (SymbolTableEntry &Ent : Symbols) { 1991 Symbol *Sym = Ent.Sym; 1992 1993 // Set st_info and st_other. 1994 ESym->st_other = 0; 1995 if (Sym->isLocal()) { 1996 ESym->setBindingAndType(STB_LOCAL, Sym->Type); 1997 } else { 1998 ESym->setBindingAndType(Sym->computeBinding(), Sym->Type); 1999 ESym->setVisibility(Sym->Visibility); 2000 } 2001 2002 ESym->st_name = Ent.StrTabOffset; 2003 ESym->st_shndx = getSymSectionIndex(Ent.Sym); 2004 2005 // Copy symbol size if it is a defined symbol. st_size is not significant 2006 // for undefined symbols, so whether copying it or not is up to us if that's 2007 // the case. We'll leave it as zero because by not setting a value, we can 2008 // get the exact same outputs for two sets of input files that differ only 2009 // in undefined symbol size in DSOs. 2010 if (ESym->st_shndx == SHN_UNDEF) 2011 ESym->st_size = 0; 2012 else 2013 ESym->st_size = Sym->getSize(); 2014 2015 // st_value is usually an address of a symbol, but that has a 2016 // special meaining for uninstantiated common symbols (this can 2017 // occur if -r is given). 2018 if (BssSection *CommonSec = getCommonSec(Ent.Sym)) 2019 ESym->st_value = CommonSec->Alignment; 2020 else 2021 ESym->st_value = Sym->getVA(); 2022 2023 ++ESym; 2024 } 2025 2026 // On MIPS we need to mark symbol which has a PLT entry and requires 2027 // pointer equality by STO_MIPS_PLT flag. That is necessary to help 2028 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs. 2029 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt 2030 if (Config->EMachine == EM_MIPS) { 2031 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf); 2032 2033 for (SymbolTableEntry &Ent : Symbols) { 2034 Symbol *Sym = Ent.Sym; 2035 if (Sym->isInPlt() && Sym->NeedsPltAddr) 2036 ESym->st_other |= STO_MIPS_PLT; 2037 if (isMicroMips()) { 2038 // Set STO_MIPS_MICROMIPS flag and less-significant bit for 2039 // a defined microMIPS symbol and symbol should point to its 2040 // PLT entry (in case of microMIPS, PLT entries always contain 2041 // microMIPS code). 2042 if (Sym->isDefined() && 2043 ((Sym->StOther & STO_MIPS_MICROMIPS) || Sym->NeedsPltAddr)) { 2044 if (StrTabSec.isDynamic()) 2045 ESym->st_value |= 1; 2046 ESym->st_other |= STO_MIPS_MICROMIPS; 2047 } 2048 } 2049 if (Config->Relocatable) 2050 if (auto *D = dyn_cast<Defined>(Sym)) 2051 if (isMipsPIC<ELFT>(D)) 2052 ESym->st_other |= STO_MIPS_PIC; 2053 ++ESym; 2054 } 2055 } 2056 } 2057 2058 SymtabShndxSection::SymtabShndxSection() 2059 : SyntheticSection(0, SHT_SYMTAB_SHNDX, 4, ".symtab_shndxr") { 2060 this->Entsize = 4; 2061 } 2062 2063 void SymtabShndxSection::writeTo(uint8_t *Buf) { 2064 // We write an array of 32 bit values, where each value has 1:1 association 2065 // with an entry in .symtab. If the corresponding entry contains SHN_XINDEX, 2066 // we need to write actual index, otherwise, we must write SHN_UNDEF(0). 2067 Buf += 4; // Ignore .symtab[0] entry. 2068 for (const SymbolTableEntry &Entry : In.SymTab->getSymbols()) { 2069 if (getSymSectionIndex(Entry.Sym) == SHN_XINDEX) 2070 write32(Buf, Entry.Sym->getOutputSection()->SectionIndex); 2071 Buf += 4; 2072 } 2073 } 2074 2075 bool SymtabShndxSection::empty() const { 2076 // SHT_SYMTAB can hold symbols with section indices values up to 2077 // SHN_LORESERVE. If we need more, we want to use extension SHT_SYMTAB_SHNDX 2078 // section. Problem is that we reveal the final section indices a bit too 2079 // late, and we do not know them here. For simplicity, we just always create 2080 // a .symtab_shndxr section when the amount of output sections is huge. 2081 size_t Size = 0; 2082 for (BaseCommand *Base : Script->SectionCommands) 2083 if (isa<OutputSection>(Base)) 2084 ++Size; 2085 return Size < SHN_LORESERVE; 2086 } 2087 2088 void SymtabShndxSection::finalizeContents() { 2089 getParent()->Link = In.SymTab->getParent()->SectionIndex; 2090 } 2091 2092 size_t SymtabShndxSection::getSize() const { 2093 return In.SymTab->getNumSymbols() * 4; 2094 } 2095 2096 // .hash and .gnu.hash sections contain on-disk hash tables that map 2097 // symbol names to their dynamic symbol table indices. Their purpose 2098 // is to help the dynamic linker resolve symbols quickly. If ELF files 2099 // don't have them, the dynamic linker has to do linear search on all 2100 // dynamic symbols, which makes programs slower. Therefore, a .hash 2101 // section is added to a DSO by default. A .gnu.hash is added if you 2102 // give the -hash-style=gnu or -hash-style=both option. 2103 // 2104 // The Unix semantics of resolving dynamic symbols is somewhat expensive. 2105 // Each ELF file has a list of DSOs that the ELF file depends on and a 2106 // list of dynamic symbols that need to be resolved from any of the 2107 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n) 2108 // where m is the number of DSOs and n is the number of dynamic 2109 // symbols. For modern large programs, both m and n are large. So 2110 // making each step faster by using hash tables substiantially 2111 // improves time to load programs. 2112 // 2113 // (Note that this is not the only way to design the shared library. 2114 // For instance, the Windows DLL takes a different approach. On 2115 // Windows, each dynamic symbol has a name of DLL from which the symbol 2116 // has to be resolved. That makes the cost of symbol resolution O(n). 2117 // This disables some hacky techniques you can use on Unix such as 2118 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.) 2119 // 2120 // Due to historical reasons, we have two different hash tables, .hash 2121 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new 2122 // and better version of .hash. .hash is just an on-disk hash table, but 2123 // .gnu.hash has a bloom filter in addition to a hash table to skip 2124 // DSOs very quickly. If you are sure that your dynamic linker knows 2125 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a 2126 // safe bet is to specify -hash-style=both for backward compatibilty. 2127 GnuHashTableSection::GnuHashTableSection() 2128 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, Config->Wordsize, ".gnu.hash") { 2129 } 2130 2131 void GnuHashTableSection::finalizeContents() { 2132 if (OutputSection *Sec = In.DynSymTab->getParent()) 2133 getParent()->Link = Sec->SectionIndex; 2134 2135 // Computes bloom filter size in word size. We want to allocate 12 2136 // bits for each symbol. It must be a power of two. 2137 if (Symbols.empty()) { 2138 MaskWords = 1; 2139 } else { 2140 uint64_t NumBits = Symbols.size() * 12; 2141 MaskWords = NextPowerOf2(NumBits / (Config->Wordsize * 8)); 2142 } 2143 2144 Size = 16; // Header 2145 Size += Config->Wordsize * MaskWords; // Bloom filter 2146 Size += NBuckets * 4; // Hash buckets 2147 Size += Symbols.size() * 4; // Hash values 2148 } 2149 2150 void GnuHashTableSection::writeTo(uint8_t *Buf) { 2151 // The output buffer is not guaranteed to be zero-cleared because we pre- 2152 // fill executable sections with trap instructions. This is a precaution 2153 // for that case, which happens only when -no-rosegment is given. 2154 memset(Buf, 0, Size); 2155 2156 // Write a header. 2157 write32(Buf, NBuckets); 2158 write32(Buf + 4, In.DynSymTab->getNumSymbols() - Symbols.size()); 2159 write32(Buf + 8, MaskWords); 2160 write32(Buf + 12, Shift2); 2161 Buf += 16; 2162 2163 // Write a bloom filter and a hash table. 2164 writeBloomFilter(Buf); 2165 Buf += Config->Wordsize * MaskWords; 2166 writeHashTable(Buf); 2167 } 2168 2169 // This function writes a 2-bit bloom filter. This bloom filter alone 2170 // usually filters out 80% or more of all symbol lookups [1]. 2171 // The dynamic linker uses the hash table only when a symbol is not 2172 // filtered out by a bloom filter. 2173 // 2174 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2), 2175 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf 2176 void GnuHashTableSection::writeBloomFilter(uint8_t *Buf) { 2177 unsigned C = Config->Is64 ? 64 : 32; 2178 for (const Entry &Sym : Symbols) { 2179 // When C = 64, we choose a word with bits [6:...] and set 1 to two bits in 2180 // the word using bits [0:5] and [26:31]. 2181 size_t I = (Sym.Hash / C) & (MaskWords - 1); 2182 uint64_t Val = readUint(Buf + I * Config->Wordsize); 2183 Val |= uint64_t(1) << (Sym.Hash % C); 2184 Val |= uint64_t(1) << ((Sym.Hash >> Shift2) % C); 2185 writeUint(Buf + I * Config->Wordsize, Val); 2186 } 2187 } 2188 2189 void GnuHashTableSection::writeHashTable(uint8_t *Buf) { 2190 uint32_t *Buckets = reinterpret_cast<uint32_t *>(Buf); 2191 uint32_t OldBucket = -1; 2192 uint32_t *Values = Buckets + NBuckets; 2193 for (auto I = Symbols.begin(), E = Symbols.end(); I != E; ++I) { 2194 // Write a hash value. It represents a sequence of chains that share the 2195 // same hash modulo value. The last element of each chain is terminated by 2196 // LSB 1. 2197 uint32_t Hash = I->Hash; 2198 bool IsLastInChain = (I + 1) == E || I->BucketIdx != (I + 1)->BucketIdx; 2199 Hash = IsLastInChain ? Hash | 1 : Hash & ~1; 2200 write32(Values++, Hash); 2201 2202 if (I->BucketIdx == OldBucket) 2203 continue; 2204 // Write a hash bucket. Hash buckets contain indices in the following hash 2205 // value table. 2206 write32(Buckets + I->BucketIdx, I->Sym->DynsymIndex); 2207 OldBucket = I->BucketIdx; 2208 } 2209 } 2210 2211 static uint32_t hashGnu(StringRef Name) { 2212 uint32_t H = 5381; 2213 for (uint8_t C : Name) 2214 H = (H << 5) + H + C; 2215 return H; 2216 } 2217 2218 // Add symbols to this symbol hash table. Note that this function 2219 // destructively sort a given vector -- which is needed because 2220 // GNU-style hash table places some sorting requirements. 2221 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &V) { 2222 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce 2223 // its type correctly. 2224 std::vector<SymbolTableEntry>::iterator Mid = 2225 std::stable_partition(V.begin(), V.end(), [](const SymbolTableEntry &S) { 2226 return !S.Sym->isDefined(); 2227 }); 2228 2229 // We chose load factor 4 for the on-disk hash table. For each hash 2230 // collision, the dynamic linker will compare a uint32_t hash value. 2231 // Since the integer comparison is quite fast, we believe we can 2232 // make the load factor even larger. 4 is just a conservative choice. 2233 // 2234 // Note that we don't want to create a zero-sized hash table because 2235 // Android loader as of 2018 doesn't like a .gnu.hash containing such 2236 // table. If that's the case, we create a hash table with one unused 2237 // dummy slot. 2238 NBuckets = std::max<size_t>((V.end() - Mid) / 4, 1); 2239 2240 if (Mid == V.end()) 2241 return; 2242 2243 for (SymbolTableEntry &Ent : llvm::make_range(Mid, V.end())) { 2244 Symbol *B = Ent.Sym; 2245 uint32_t Hash = hashGnu(B->getName()); 2246 uint32_t BucketIdx = Hash % NBuckets; 2247 Symbols.push_back({B, Ent.StrTabOffset, Hash, BucketIdx}); 2248 } 2249 2250 std::stable_sort( 2251 Symbols.begin(), Symbols.end(), 2252 [](const Entry &L, const Entry &R) { return L.BucketIdx < R.BucketIdx; }); 2253 2254 V.erase(Mid, V.end()); 2255 for (const Entry &Ent : Symbols) 2256 V.push_back({Ent.Sym, Ent.StrTabOffset}); 2257 } 2258 2259 HashTableSection::HashTableSection() 2260 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") { 2261 this->Entsize = 4; 2262 } 2263 2264 void HashTableSection::finalizeContents() { 2265 if (OutputSection *Sec = In.DynSymTab->getParent()) 2266 getParent()->Link = Sec->SectionIndex; 2267 2268 unsigned NumEntries = 2; // nbucket and nchain. 2269 NumEntries += In.DynSymTab->getNumSymbols(); // The chain entries. 2270 2271 // Create as many buckets as there are symbols. 2272 NumEntries += In.DynSymTab->getNumSymbols(); 2273 this->Size = NumEntries * 4; 2274 } 2275 2276 void HashTableSection::writeTo(uint8_t *Buf) { 2277 // See comment in GnuHashTableSection::writeTo. 2278 memset(Buf, 0, Size); 2279 2280 unsigned NumSymbols = In.DynSymTab->getNumSymbols(); 2281 2282 uint32_t *P = reinterpret_cast<uint32_t *>(Buf); 2283 write32(P++, NumSymbols); // nbucket 2284 write32(P++, NumSymbols); // nchain 2285 2286 uint32_t *Buckets = P; 2287 uint32_t *Chains = P + NumSymbols; 2288 2289 for (const SymbolTableEntry &S : In.DynSymTab->getSymbols()) { 2290 Symbol *Sym = S.Sym; 2291 StringRef Name = Sym->getName(); 2292 unsigned I = Sym->DynsymIndex; 2293 uint32_t Hash = hashSysV(Name) % NumSymbols; 2294 Chains[I] = Buckets[Hash]; 2295 write32(Buckets + Hash, I); 2296 } 2297 } 2298 2299 // On PowerPC64 the lazy symbol resolvers go into the `global linkage table` 2300 // in the .glink section, rather then the typical .plt section. 2301 PltSection::PltSection(bool IsIplt) 2302 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, 2303 Config->EMachine == EM_PPC64 ? ".glink" : ".plt"), 2304 HeaderSize(!IsIplt || Config->ZRetpolineplt ? Target->PltHeaderSize : 0), 2305 IsIplt(IsIplt) { 2306 // The PLT needs to be writable on SPARC as the dynamic linker will 2307 // modify the instructions in the PLT entries. 2308 if (Config->EMachine == EM_SPARCV9) 2309 this->Flags |= SHF_WRITE; 2310 } 2311 2312 void PltSection::writeTo(uint8_t *Buf) { 2313 // At beginning of PLT or retpoline IPLT, we have code to call the dynamic 2314 // linker to resolve dynsyms at runtime. Write such code. 2315 if (HeaderSize > 0) 2316 Target->writePltHeader(Buf); 2317 size_t Off = HeaderSize; 2318 // The IPlt is immediately after the Plt, account for this in RelOff 2319 unsigned PltOff = getPltRelocOff(); 2320 2321 for (auto &I : Entries) { 2322 const Symbol *B = I.first; 2323 unsigned RelOff = I.second + PltOff; 2324 uint64_t Got = B->getGotPltVA(); 2325 uint64_t Plt = this->getVA() + Off; 2326 Target->writePlt(Buf + Off, Got, Plt, B->PltIndex, RelOff); 2327 Off += Target->PltEntrySize; 2328 } 2329 } 2330 2331 template <class ELFT> void PltSection::addEntry(Symbol &Sym) { 2332 Sym.PltIndex = Entries.size(); 2333 RelocationBaseSection *PltRelocSection = In.RelaPlt; 2334 if (IsIplt) { 2335 PltRelocSection = In.RelaIplt; 2336 Sym.IsInIplt = true; 2337 } 2338 unsigned RelOff = 2339 static_cast<RelocationSection<ELFT> *>(PltRelocSection)->getRelocOffset(); 2340 Entries.push_back(std::make_pair(&Sym, RelOff)); 2341 } 2342 2343 size_t PltSection::getSize() const { 2344 return HeaderSize + Entries.size() * Target->PltEntrySize; 2345 } 2346 2347 // Some architectures such as additional symbols in the PLT section. For 2348 // example ARM uses mapping symbols to aid disassembly 2349 void PltSection::addSymbols() { 2350 // The PLT may have symbols defined for the Header, the IPLT has no header 2351 if (!IsIplt) 2352 Target->addPltHeaderSymbols(*this); 2353 size_t Off = HeaderSize; 2354 for (size_t I = 0; I < Entries.size(); ++I) { 2355 Target->addPltSymbols(*this, Off); 2356 Off += Target->PltEntrySize; 2357 } 2358 } 2359 2360 unsigned PltSection::getPltRelocOff() const { 2361 return IsIplt ? In.Plt->getSize() : 0; 2362 } 2363 2364 // The string hash function for .gdb_index. 2365 static uint32_t computeGdbHash(StringRef S) { 2366 uint32_t H = 0; 2367 for (uint8_t C : S) 2368 H = H * 67 + toLower(C) - 113; 2369 return H; 2370 } 2371 2372 GdbIndexSection::GdbIndexSection() 2373 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index") {} 2374 2375 // Returns the desired size of an on-disk hash table for a .gdb_index section. 2376 // There's a tradeoff between size and collision rate. We aim 75% utilization. 2377 size_t GdbIndexSection::computeSymtabSize() const { 2378 return std::max<size_t>(NextPowerOf2(Symbols.size() * 4 / 3), 1024); 2379 } 2380 2381 // Compute the output section size. 2382 void GdbIndexSection::initOutputSize() { 2383 Size = sizeof(GdbIndexHeader) + computeSymtabSize() * 8; 2384 2385 for (GdbChunk &Chunk : Chunks) 2386 Size += Chunk.CompilationUnits.size() * 16 + Chunk.AddressAreas.size() * 20; 2387 2388 // Add the constant pool size if exists. 2389 if (!Symbols.empty()) { 2390 GdbSymbol &Sym = Symbols.back(); 2391 Size += Sym.NameOff + Sym.Name.size() + 1; 2392 } 2393 } 2394 2395 static std::vector<InputSection *> getDebugInfoSections() { 2396 std::vector<InputSection *> Ret; 2397 for (InputSectionBase *S : InputSections) 2398 if (InputSection *IS = dyn_cast<InputSection>(S)) 2399 if (IS->Name == ".debug_info") 2400 Ret.push_back(IS); 2401 return Ret; 2402 } 2403 2404 static std::vector<GdbIndexSection::CuEntry> readCuList(DWARFContext &Dwarf) { 2405 std::vector<GdbIndexSection::CuEntry> Ret; 2406 for (std::unique_ptr<DWARFUnit> &Cu : Dwarf.compile_units()) 2407 Ret.push_back({Cu->getOffset(), Cu->getLength() + 4}); 2408 return Ret; 2409 } 2410 2411 static std::vector<GdbIndexSection::AddressEntry> 2412 readAddressAreas(DWARFContext &Dwarf, InputSection *Sec) { 2413 std::vector<GdbIndexSection::AddressEntry> Ret; 2414 2415 uint32_t CuIdx = 0; 2416 for (std::unique_ptr<DWARFUnit> &Cu : Dwarf.compile_units()) { 2417 Expected<DWARFAddressRangesVector> Ranges = Cu->collectAddressRanges(); 2418 if (!Ranges) { 2419 error(toString(Sec) + ": " + toString(Ranges.takeError())); 2420 return {}; 2421 } 2422 2423 ArrayRef<InputSectionBase *> Sections = Sec->File->getSections(); 2424 for (DWARFAddressRange &R : *Ranges) { 2425 InputSectionBase *S = Sections[R.SectionIndex]; 2426 if (!S || S == &InputSection::Discarded || !S->Live) 2427 continue; 2428 // Range list with zero size has no effect. 2429 if (R.LowPC == R.HighPC) 2430 continue; 2431 auto *IS = cast<InputSection>(S); 2432 uint64_t Offset = IS->getOffsetInFile(); 2433 Ret.push_back({IS, R.LowPC - Offset, R.HighPC - Offset, CuIdx}); 2434 } 2435 ++CuIdx; 2436 } 2437 2438 return Ret; 2439 } 2440 2441 template <class ELFT> 2442 static std::vector<GdbIndexSection::NameAttrEntry> 2443 readPubNamesAndTypes(const LLDDwarfObj<ELFT> &Obj, 2444 const std::vector<GdbIndexSection::CuEntry> &CUs) { 2445 const DWARFSection &PubNames = Obj.getGnuPubNamesSection(); 2446 const DWARFSection &PubTypes = Obj.getGnuPubTypesSection(); 2447 2448 std::vector<GdbIndexSection::NameAttrEntry> Ret; 2449 for (const DWARFSection *Pub : {&PubNames, &PubTypes}) { 2450 DWARFDebugPubTable Table(Obj, *Pub, Config->IsLE, true); 2451 for (const DWARFDebugPubTable::Set &Set : Table.getData()) { 2452 // The value written into the constant pool is Kind << 24 | CuIndex. As we 2453 // don't know how many compilation units precede this object to compute 2454 // CuIndex, we compute (Kind << 24 | CuIndexInThisObject) instead, and add 2455 // the number of preceding compilation units later. 2456 uint32_t I = 2457 lower_bound(CUs, Set.Offset, 2458 [](GdbIndexSection::CuEntry CU, uint32_t Offset) { 2459 return CU.CuOffset < Offset; 2460 }) - 2461 CUs.begin(); 2462 for (const DWARFDebugPubTable::Entry &Ent : Set.Entries) 2463 Ret.push_back({{Ent.Name, computeGdbHash(Ent.Name)}, 2464 (Ent.Descriptor.toBits() << 24) | I}); 2465 } 2466 } 2467 return Ret; 2468 } 2469 2470 // Create a list of symbols from a given list of symbol names and types 2471 // by uniquifying them by name. 2472 static std::vector<GdbIndexSection::GdbSymbol> 2473 createSymbols(ArrayRef<std::vector<GdbIndexSection::NameAttrEntry>> NameAttrs, 2474 const std::vector<GdbIndexSection::GdbChunk> &Chunks) { 2475 typedef GdbIndexSection::GdbSymbol GdbSymbol; 2476 typedef GdbIndexSection::NameAttrEntry NameAttrEntry; 2477 2478 // For each chunk, compute the number of compilation units preceding it. 2479 uint32_t CuIdx = 0; 2480 std::vector<uint32_t> CuIdxs(Chunks.size()); 2481 for (uint32_t I = 0, E = Chunks.size(); I != E; ++I) { 2482 CuIdxs[I] = CuIdx; 2483 CuIdx += Chunks[I].CompilationUnits.size(); 2484 } 2485 2486 // The number of symbols we will handle in this function is of the order 2487 // of millions for very large executables, so we use multi-threading to 2488 // speed it up. 2489 size_t NumShards = 32; 2490 size_t Concurrency = 1; 2491 if (ThreadsEnabled) 2492 Concurrency = 2493 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), NumShards); 2494 2495 // A sharded map to uniquify symbols by name. 2496 std::vector<DenseMap<CachedHashStringRef, size_t>> Map(NumShards); 2497 size_t Shift = 32 - countTrailingZeros(NumShards); 2498 2499 // Instantiate GdbSymbols while uniqufying them by name. 2500 std::vector<std::vector<GdbSymbol>> Symbols(NumShards); 2501 parallelForEachN(0, Concurrency, [&](size_t ThreadId) { 2502 uint32_t I = 0; 2503 for (ArrayRef<NameAttrEntry> Entries : NameAttrs) { 2504 for (const NameAttrEntry &Ent : Entries) { 2505 size_t ShardId = Ent.Name.hash() >> Shift; 2506 if ((ShardId & (Concurrency - 1)) != ThreadId) 2507 continue; 2508 2509 uint32_t V = Ent.CuIndexAndAttrs + CuIdxs[I]; 2510 size_t &Idx = Map[ShardId][Ent.Name]; 2511 if (Idx) { 2512 Symbols[ShardId][Idx - 1].CuVector.push_back(V); 2513 continue; 2514 } 2515 2516 Idx = Symbols[ShardId].size() + 1; 2517 Symbols[ShardId].push_back({Ent.Name, {V}, 0, 0}); 2518 } 2519 ++I; 2520 } 2521 }); 2522 2523 size_t NumSymbols = 0; 2524 for (ArrayRef<GdbSymbol> V : Symbols) 2525 NumSymbols += V.size(); 2526 2527 // The return type is a flattened vector, so we'll copy each vector 2528 // contents to Ret. 2529 std::vector<GdbSymbol> Ret; 2530 Ret.reserve(NumSymbols); 2531 for (std::vector<GdbSymbol> &Vec : Symbols) 2532 for (GdbSymbol &Sym : Vec) 2533 Ret.push_back(std::move(Sym)); 2534 2535 // CU vectors and symbol names are adjacent in the output file. 2536 // We can compute their offsets in the output file now. 2537 size_t Off = 0; 2538 for (GdbSymbol &Sym : Ret) { 2539 Sym.CuVectorOff = Off; 2540 Off += (Sym.CuVector.size() + 1) * 4; 2541 } 2542 for (GdbSymbol &Sym : Ret) { 2543 Sym.NameOff = Off; 2544 Off += Sym.Name.size() + 1; 2545 } 2546 2547 return Ret; 2548 } 2549 2550 // Returns a newly-created .gdb_index section. 2551 template <class ELFT> GdbIndexSection *GdbIndexSection::create() { 2552 std::vector<InputSection *> Sections = getDebugInfoSections(); 2553 2554 // .debug_gnu_pub{names,types} are useless in executables. 2555 // They are present in input object files solely for creating 2556 // a .gdb_index. So we can remove them from the output. 2557 for (InputSectionBase *S : InputSections) 2558 if (S->Name == ".debug_gnu_pubnames" || S->Name == ".debug_gnu_pubtypes") 2559 S->Live = false; 2560 2561 std::vector<GdbChunk> Chunks(Sections.size()); 2562 std::vector<std::vector<NameAttrEntry>> NameAttrs(Sections.size()); 2563 2564 parallelForEachN(0, Sections.size(), [&](size_t I) { 2565 ObjFile<ELFT> *File = Sections[I]->getFile<ELFT>(); 2566 DWARFContext Dwarf(make_unique<LLDDwarfObj<ELFT>>(File)); 2567 2568 Chunks[I].Sec = Sections[I]; 2569 Chunks[I].CompilationUnits = readCuList(Dwarf); 2570 Chunks[I].AddressAreas = readAddressAreas(Dwarf, Sections[I]); 2571 NameAttrs[I] = readPubNamesAndTypes<ELFT>( 2572 static_cast<const LLDDwarfObj<ELFT> &>(Dwarf.getDWARFObj()), 2573 Chunks[I].CompilationUnits); 2574 }); 2575 2576 auto *Ret = make<GdbIndexSection>(); 2577 Ret->Chunks = std::move(Chunks); 2578 Ret->Symbols = createSymbols(NameAttrs, Ret->Chunks); 2579 Ret->initOutputSize(); 2580 return Ret; 2581 } 2582 2583 void GdbIndexSection::writeTo(uint8_t *Buf) { 2584 // Write the header. 2585 auto *Hdr = reinterpret_cast<GdbIndexHeader *>(Buf); 2586 uint8_t *Start = Buf; 2587 Hdr->Version = 7; 2588 Buf += sizeof(*Hdr); 2589 2590 // Write the CU list. 2591 Hdr->CuListOff = Buf - Start; 2592 for (GdbChunk &Chunk : Chunks) { 2593 for (CuEntry &Cu : Chunk.CompilationUnits) { 2594 write64le(Buf, Chunk.Sec->OutSecOff + Cu.CuOffset); 2595 write64le(Buf + 8, Cu.CuLength); 2596 Buf += 16; 2597 } 2598 } 2599 2600 // Write the address area. 2601 Hdr->CuTypesOff = Buf - Start; 2602 Hdr->AddressAreaOff = Buf - Start; 2603 uint32_t CuOff = 0; 2604 for (GdbChunk &Chunk : Chunks) { 2605 for (AddressEntry &E : Chunk.AddressAreas) { 2606 uint64_t BaseAddr = E.Section->getVA(0); 2607 write64le(Buf, BaseAddr + E.LowAddress); 2608 write64le(Buf + 8, BaseAddr + E.HighAddress); 2609 write32le(Buf + 16, E.CuIndex + CuOff); 2610 Buf += 20; 2611 } 2612 CuOff += Chunk.CompilationUnits.size(); 2613 } 2614 2615 // Write the on-disk open-addressing hash table containing symbols. 2616 Hdr->SymtabOff = Buf - Start; 2617 size_t SymtabSize = computeSymtabSize(); 2618 uint32_t Mask = SymtabSize - 1; 2619 2620 for (GdbSymbol &Sym : Symbols) { 2621 uint32_t H = Sym.Name.hash(); 2622 uint32_t I = H & Mask; 2623 uint32_t Step = ((H * 17) & Mask) | 1; 2624 2625 while (read32le(Buf + I * 8)) 2626 I = (I + Step) & Mask; 2627 2628 write32le(Buf + I * 8, Sym.NameOff); 2629 write32le(Buf + I * 8 + 4, Sym.CuVectorOff); 2630 } 2631 2632 Buf += SymtabSize * 8; 2633 2634 // Write the string pool. 2635 Hdr->ConstantPoolOff = Buf - Start; 2636 parallelForEach(Symbols, [&](GdbSymbol &Sym) { 2637 memcpy(Buf + Sym.NameOff, Sym.Name.data(), Sym.Name.size()); 2638 }); 2639 2640 // Write the CU vectors. 2641 for (GdbSymbol &Sym : Symbols) { 2642 write32le(Buf, Sym.CuVector.size()); 2643 Buf += 4; 2644 for (uint32_t Val : Sym.CuVector) { 2645 write32le(Buf, Val); 2646 Buf += 4; 2647 } 2648 } 2649 } 2650 2651 bool GdbIndexSection::empty() const { return Chunks.empty(); } 2652 2653 EhFrameHeader::EhFrameHeader() 2654 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr") {} 2655 2656 // .eh_frame_hdr contains a binary search table of pointers to FDEs. 2657 // Each entry of the search table consists of two values, 2658 // the starting PC from where FDEs covers, and the FDE's address. 2659 // It is sorted by PC. 2660 void EhFrameHeader::writeTo(uint8_t *Buf) { 2661 typedef EhFrameSection::FdeData FdeData; 2662 2663 std::vector<FdeData> Fdes = In.EhFrame->getFdeData(); 2664 2665 Buf[0] = 1; 2666 Buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4; 2667 Buf[2] = DW_EH_PE_udata4; 2668 Buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4; 2669 write32(Buf + 4, In.EhFrame->getParent()->Addr - this->getVA() - 4); 2670 write32(Buf + 8, Fdes.size()); 2671 Buf += 12; 2672 2673 for (FdeData &Fde : Fdes) { 2674 write32(Buf, Fde.PcRel); 2675 write32(Buf + 4, Fde.FdeVARel); 2676 Buf += 8; 2677 } 2678 } 2679 2680 size_t EhFrameHeader::getSize() const { 2681 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs. 2682 return 12 + In.EhFrame->NumFdes * 8; 2683 } 2684 2685 bool EhFrameHeader::empty() const { return In.EhFrame->empty(); } 2686 2687 VersionDefinitionSection::VersionDefinitionSection() 2688 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t), 2689 ".gnu.version_d") {} 2690 2691 static StringRef getFileDefName() { 2692 if (!Config->SoName.empty()) 2693 return Config->SoName; 2694 return Config->OutputFile; 2695 } 2696 2697 void VersionDefinitionSection::finalizeContents() { 2698 FileDefNameOff = In.DynStrTab->addString(getFileDefName()); 2699 for (VersionDefinition &V : Config->VersionDefinitions) 2700 V.NameOff = In.DynStrTab->addString(V.Name); 2701 2702 if (OutputSection *Sec = In.DynStrTab->getParent()) 2703 getParent()->Link = Sec->SectionIndex; 2704 2705 // sh_info should be set to the number of definitions. This fact is missed in 2706 // documentation, but confirmed by binutils community: 2707 // https://sourceware.org/ml/binutils/2014-11/msg00355.html 2708 getParent()->Info = getVerDefNum(); 2709 } 2710 2711 void VersionDefinitionSection::writeOne(uint8_t *Buf, uint32_t Index, 2712 StringRef Name, size_t NameOff) { 2713 uint16_t Flags = Index == 1 ? VER_FLG_BASE : 0; 2714 2715 // Write a verdef. 2716 write16(Buf, 1); // vd_version 2717 write16(Buf + 2, Flags); // vd_flags 2718 write16(Buf + 4, Index); // vd_ndx 2719 write16(Buf + 6, 1); // vd_cnt 2720 write32(Buf + 8, hashSysV(Name)); // vd_hash 2721 write32(Buf + 12, 20); // vd_aux 2722 write32(Buf + 16, 28); // vd_next 2723 2724 // Write a veraux. 2725 write32(Buf + 20, NameOff); // vda_name 2726 write32(Buf + 24, 0); // vda_next 2727 } 2728 2729 void VersionDefinitionSection::writeTo(uint8_t *Buf) { 2730 writeOne(Buf, 1, getFileDefName(), FileDefNameOff); 2731 2732 for (VersionDefinition &V : Config->VersionDefinitions) { 2733 Buf += EntrySize; 2734 writeOne(Buf, V.Id, V.Name, V.NameOff); 2735 } 2736 2737 // Need to terminate the last version definition. 2738 write32(Buf + 16, 0); // vd_next 2739 } 2740 2741 size_t VersionDefinitionSection::getSize() const { 2742 return EntrySize * getVerDefNum(); 2743 } 2744 2745 // .gnu.version is a table where each entry is 2 byte long. 2746 template <class ELFT> 2747 VersionTableSection<ELFT>::VersionTableSection() 2748 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t), 2749 ".gnu.version") { 2750 this->Entsize = 2; 2751 } 2752 2753 template <class ELFT> void VersionTableSection<ELFT>::finalizeContents() { 2754 // At the moment of june 2016 GNU docs does not mention that sh_link field 2755 // should be set, but Sun docs do. Also readelf relies on this field. 2756 getParent()->Link = In.DynSymTab->getParent()->SectionIndex; 2757 } 2758 2759 template <class ELFT> size_t VersionTableSection<ELFT>::getSize() const { 2760 return (In.DynSymTab->getSymbols().size() + 1) * 2; 2761 } 2762 2763 template <class ELFT> void VersionTableSection<ELFT>::writeTo(uint8_t *Buf) { 2764 Buf += 2; 2765 for (const SymbolTableEntry &S : In.DynSymTab->getSymbols()) { 2766 write16(Buf, S.Sym->VersionId); 2767 Buf += 2; 2768 } 2769 } 2770 2771 template <class ELFT> bool VersionTableSection<ELFT>::empty() const { 2772 return !In.VerDef && InX<ELFT>::VerNeed->empty(); 2773 } 2774 2775 template <class ELFT> 2776 VersionNeedSection<ELFT>::VersionNeedSection() 2777 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t), 2778 ".gnu.version_r") { 2779 // Identifiers in verneed section start at 2 because 0 and 1 are reserved 2780 // for VER_NDX_LOCAL and VER_NDX_GLOBAL. 2781 // First identifiers are reserved by verdef section if it exist. 2782 NextIndex = getVerDefNum() + 1; 2783 } 2784 2785 template <class ELFT> void VersionNeedSection<ELFT>::addSymbol(Symbol *SS) { 2786 auto &File = cast<SharedFile<ELFT>>(*SS->File); 2787 if (SS->VerdefIndex == VER_NDX_GLOBAL) { 2788 SS->VersionId = VER_NDX_GLOBAL; 2789 return; 2790 } 2791 2792 // If we don't already know that we need an Elf_Verneed for this DSO, prepare 2793 // to create one by adding it to our needed list and creating a dynstr entry 2794 // for the soname. 2795 if (File.VerdefMap.empty()) 2796 Needed.push_back({&File, In.DynStrTab->addString(File.SoName)}); 2797 const typename ELFT::Verdef *Ver = File.Verdefs[SS->VerdefIndex]; 2798 typename SharedFile<ELFT>::NeededVer &NV = File.VerdefMap[Ver]; 2799 2800 // If we don't already know that we need an Elf_Vernaux for this Elf_Verdef, 2801 // prepare to create one by allocating a version identifier and creating a 2802 // dynstr entry for the version name. 2803 if (NV.Index == 0) { 2804 NV.StrTab = In.DynStrTab->addString(File.getStringTable().data() + 2805 Ver->getAux()->vda_name); 2806 NV.Index = NextIndex++; 2807 } 2808 SS->VersionId = NV.Index; 2809 } 2810 2811 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *Buf) { 2812 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs. 2813 auto *Verneed = reinterpret_cast<Elf_Verneed *>(Buf); 2814 auto *Vernaux = reinterpret_cast<Elf_Vernaux *>(Verneed + Needed.size()); 2815 2816 for (std::pair<SharedFile<ELFT> *, size_t> &P : Needed) { 2817 // Create an Elf_Verneed for this DSO. 2818 Verneed->vn_version = 1; 2819 Verneed->vn_cnt = P.first->VerdefMap.size(); 2820 Verneed->vn_file = P.second; 2821 Verneed->vn_aux = 2822 reinterpret_cast<char *>(Vernaux) - reinterpret_cast<char *>(Verneed); 2823 Verneed->vn_next = sizeof(Elf_Verneed); 2824 ++Verneed; 2825 2826 // Create the Elf_Vernauxs for this Elf_Verneed. The loop iterates over 2827 // VerdefMap, which will only contain references to needed version 2828 // definitions. Each Elf_Vernaux is based on the information contained in 2829 // the Elf_Verdef in the source DSO. This loop iterates over a std::map of 2830 // pointers, but is deterministic because the pointers refer to Elf_Verdef 2831 // data structures within a single input file. 2832 for (auto &NV : P.first->VerdefMap) { 2833 Vernaux->vna_hash = NV.first->vd_hash; 2834 Vernaux->vna_flags = 0; 2835 Vernaux->vna_other = NV.second.Index; 2836 Vernaux->vna_name = NV.second.StrTab; 2837 Vernaux->vna_next = sizeof(Elf_Vernaux); 2838 ++Vernaux; 2839 } 2840 2841 Vernaux[-1].vna_next = 0; 2842 } 2843 Verneed[-1].vn_next = 0; 2844 } 2845 2846 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() { 2847 if (OutputSection *Sec = In.DynStrTab->getParent()) 2848 getParent()->Link = Sec->SectionIndex; 2849 getParent()->Info = Needed.size(); 2850 } 2851 2852 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const { 2853 unsigned Size = Needed.size() * sizeof(Elf_Verneed); 2854 for (const std::pair<SharedFile<ELFT> *, size_t> &P : Needed) 2855 Size += P.first->VerdefMap.size() * sizeof(Elf_Vernaux); 2856 return Size; 2857 } 2858 2859 template <class ELFT> bool VersionNeedSection<ELFT>::empty() const { 2860 return getNeedNum() == 0; 2861 } 2862 2863 void MergeSyntheticSection::addSection(MergeInputSection *MS) { 2864 MS->Parent = this; 2865 Sections.push_back(MS); 2866 } 2867 2868 MergeTailSection::MergeTailSection(StringRef Name, uint32_t Type, 2869 uint64_t Flags, uint32_t Alignment) 2870 : MergeSyntheticSection(Name, Type, Flags, Alignment), 2871 Builder(StringTableBuilder::RAW, Alignment) {} 2872 2873 size_t MergeTailSection::getSize() const { return Builder.getSize(); } 2874 2875 void MergeTailSection::writeTo(uint8_t *Buf) { Builder.write(Buf); } 2876 2877 void MergeTailSection::finalizeContents() { 2878 // Add all string pieces to the string table builder to create section 2879 // contents. 2880 for (MergeInputSection *Sec : Sections) 2881 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2882 if (Sec->Pieces[I].Live) 2883 Builder.add(Sec->getData(I)); 2884 2885 // Fix the string table content. After this, the contents will never change. 2886 Builder.finalize(); 2887 2888 // finalize() fixed tail-optimized strings, so we can now get 2889 // offsets of strings. Get an offset for each string and save it 2890 // to a corresponding StringPiece for easy access. 2891 for (MergeInputSection *Sec : Sections) 2892 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2893 if (Sec->Pieces[I].Live) 2894 Sec->Pieces[I].OutputOff = Builder.getOffset(Sec->getData(I)); 2895 } 2896 2897 void MergeNoTailSection::writeTo(uint8_t *Buf) { 2898 for (size_t I = 0; I < NumShards; ++I) 2899 Shards[I].write(Buf + ShardOffsets[I]); 2900 } 2901 2902 // This function is very hot (i.e. it can take several seconds to finish) 2903 // because sometimes the number of inputs is in an order of magnitude of 2904 // millions. So, we use multi-threading. 2905 // 2906 // For any strings S and T, we know S is not mergeable with T if S's hash 2907 // value is different from T's. If that's the case, we can safely put S and 2908 // T into different string builders without worrying about merge misses. 2909 // We do it in parallel. 2910 void MergeNoTailSection::finalizeContents() { 2911 // Initializes string table builders. 2912 for (size_t I = 0; I < NumShards; ++I) 2913 Shards.emplace_back(StringTableBuilder::RAW, Alignment); 2914 2915 // Concurrency level. Must be a power of 2 to avoid expensive modulo 2916 // operations in the following tight loop. 2917 size_t Concurrency = 1; 2918 if (ThreadsEnabled) 2919 Concurrency = 2920 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), NumShards); 2921 2922 // Add section pieces to the builders. 2923 parallelForEachN(0, Concurrency, [&](size_t ThreadId) { 2924 for (MergeInputSection *Sec : Sections) { 2925 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) { 2926 size_t ShardId = getShardId(Sec->Pieces[I].Hash); 2927 if ((ShardId & (Concurrency - 1)) == ThreadId && Sec->Pieces[I].Live) 2928 Sec->Pieces[I].OutputOff = Shards[ShardId].add(Sec->getData(I)); 2929 } 2930 } 2931 }); 2932 2933 // Compute an in-section offset for each shard. 2934 size_t Off = 0; 2935 for (size_t I = 0; I < NumShards; ++I) { 2936 Shards[I].finalizeInOrder(); 2937 if (Shards[I].getSize() > 0) 2938 Off = alignTo(Off, Alignment); 2939 ShardOffsets[I] = Off; 2940 Off += Shards[I].getSize(); 2941 } 2942 Size = Off; 2943 2944 // So far, section pieces have offsets from beginning of shards, but 2945 // we want offsets from beginning of the whole section. Fix them. 2946 parallelForEach(Sections, [&](MergeInputSection *Sec) { 2947 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) 2948 if (Sec->Pieces[I].Live) 2949 Sec->Pieces[I].OutputOff += 2950 ShardOffsets[getShardId(Sec->Pieces[I].Hash)]; 2951 }); 2952 } 2953 2954 static MergeSyntheticSection *createMergeSynthetic(StringRef Name, 2955 uint32_t Type, 2956 uint64_t Flags, 2957 uint32_t Alignment) { 2958 bool ShouldTailMerge = (Flags & SHF_STRINGS) && Config->Optimize >= 2; 2959 if (ShouldTailMerge) 2960 return make<MergeTailSection>(Name, Type, Flags, Alignment); 2961 return make<MergeNoTailSection>(Name, Type, Flags, Alignment); 2962 } 2963 2964 template <class ELFT> void elf::splitSections() { 2965 // splitIntoPieces needs to be called on each MergeInputSection 2966 // before calling finalizeContents(). 2967 parallelForEach(InputSections, [](InputSectionBase *Sec) { 2968 if (auto *S = dyn_cast<MergeInputSection>(Sec)) 2969 S->splitIntoPieces(); 2970 else if (auto *Eh = dyn_cast<EhInputSection>(Sec)) 2971 Eh->split<ELFT>(); 2972 }); 2973 } 2974 2975 // This function scans over the inputsections to create mergeable 2976 // synthetic sections. 2977 // 2978 // It removes MergeInputSections from the input section array and adds 2979 // new synthetic sections at the location of the first input section 2980 // that it replaces. It then finalizes each synthetic section in order 2981 // to compute an output offset for each piece of each input section. 2982 void elf::mergeSections() { 2983 std::vector<MergeSyntheticSection *> MergeSections; 2984 for (InputSectionBase *&S : InputSections) { 2985 MergeInputSection *MS = dyn_cast<MergeInputSection>(S); 2986 if (!MS) 2987 continue; 2988 2989 // We do not want to handle sections that are not alive, so just remove 2990 // them instead of trying to merge. 2991 if (!MS->Live) { 2992 S = nullptr; 2993 continue; 2994 } 2995 2996 StringRef OutsecName = getOutputSectionName(MS); 2997 uint32_t Alignment = std::max<uint32_t>(MS->Alignment, MS->Entsize); 2998 2999 auto I = llvm::find_if(MergeSections, [=](MergeSyntheticSection *Sec) { 3000 // While we could create a single synthetic section for two different 3001 // values of Entsize, it is better to take Entsize into consideration. 3002 // 3003 // With a single synthetic section no two pieces with different Entsize 3004 // could be equal, so we may as well have two sections. 3005 // 3006 // Using Entsize in here also allows us to propagate it to the synthetic 3007 // section. 3008 return Sec->Name == OutsecName && Sec->Flags == MS->Flags && 3009 Sec->Entsize == MS->Entsize && Sec->Alignment == Alignment; 3010 }); 3011 if (I == MergeSections.end()) { 3012 MergeSyntheticSection *Syn = 3013 createMergeSynthetic(OutsecName, MS->Type, MS->Flags, Alignment); 3014 MergeSections.push_back(Syn); 3015 I = std::prev(MergeSections.end()); 3016 S = Syn; 3017 Syn->Entsize = MS->Entsize; 3018 } else { 3019 S = nullptr; 3020 } 3021 (*I)->addSection(MS); 3022 } 3023 for (auto *MS : MergeSections) 3024 MS->finalizeContents(); 3025 3026 std::vector<InputSectionBase *> &V = InputSections; 3027 V.erase(std::remove(V.begin(), V.end(), nullptr), V.end()); 3028 } 3029 3030 MipsRldMapSection::MipsRldMapSection() 3031 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, Config->Wordsize, 3032 ".rld_map") {} 3033 3034 ARMExidxSentinelSection::ARMExidxSentinelSection() 3035 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX, 3036 Config->Wordsize, ".ARM.exidx") {} 3037 3038 // Write a terminating sentinel entry to the end of the .ARM.exidx table. 3039 // This section will have been sorted last in the .ARM.exidx table. 3040 // This table entry will have the form: 3041 // | PREL31 upper bound of code that has exception tables | EXIDX_CANTUNWIND | 3042 // The sentinel must have the PREL31 value of an address higher than any 3043 // address described by any other table entry. 3044 void ARMExidxSentinelSection::writeTo(uint8_t *Buf) { 3045 assert(Highest); 3046 uint64_t S = Highest->getVA(Highest->getSize()); 3047 uint64_t P = getVA(); 3048 Target->relocateOne(Buf, R_ARM_PREL31, S - P); 3049 write32le(Buf + 4, 1); 3050 } 3051 3052 // The sentinel has to be removed if there are no other .ARM.exidx entries. 3053 bool ARMExidxSentinelSection::empty() const { 3054 for (InputSection *IS : getInputSections(getParent())) 3055 if (!isa<ARMExidxSentinelSection>(IS)) 3056 return false; 3057 return true; 3058 } 3059 3060 bool ARMExidxSentinelSection::classof(const SectionBase *D) { 3061 return D->kind() == InputSectionBase::Synthetic && D->Type == SHT_ARM_EXIDX; 3062 } 3063 3064 ThunkSection::ThunkSection(OutputSection *OS, uint64_t Off) 3065 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 3066 Config->Wordsize, ".text.thunk") { 3067 this->Parent = OS; 3068 this->OutSecOff = Off; 3069 } 3070 3071 void ThunkSection::addThunk(Thunk *T) { 3072 Thunks.push_back(T); 3073 T->addSymbols(*this); 3074 } 3075 3076 void ThunkSection::writeTo(uint8_t *Buf) { 3077 for (Thunk *T : Thunks) 3078 T->writeTo(Buf + T->Offset); 3079 } 3080 3081 InputSection *ThunkSection::getTargetInputSection() const { 3082 if (Thunks.empty()) 3083 return nullptr; 3084 const Thunk *T = Thunks.front(); 3085 return T->getTargetInputSection(); 3086 } 3087 3088 bool ThunkSection::assignOffsets() { 3089 uint64_t Off = 0; 3090 for (Thunk *T : Thunks) { 3091 Off = alignTo(Off, T->Alignment); 3092 T->setOffset(Off); 3093 uint32_t Size = T->size(); 3094 T->getThunkTargetSym()->Size = Size; 3095 Off += Size; 3096 } 3097 bool Changed = Off != Size; 3098 Size = Off; 3099 return Changed; 3100 } 3101 3102 // If linking position-dependent code then the table will store the addresses 3103 // directly in the binary so the section has type SHT_PROGBITS. If linking 3104 // position-independent code the section has type SHT_NOBITS since it will be 3105 // allocated and filled in by the dynamic linker. 3106 PPC64LongBranchTargetSection::PPC64LongBranchTargetSection() 3107 : SyntheticSection(SHF_ALLOC | SHF_WRITE, 3108 Config->Pic ? SHT_NOBITS : SHT_PROGBITS, 8, 3109 ".branch_lt") {} 3110 3111 void PPC64LongBranchTargetSection::addEntry(Symbol &Sym) { 3112 assert(Sym.PPC64BranchltIndex == 0xffff); 3113 Sym.PPC64BranchltIndex = Entries.size(); 3114 Entries.push_back(&Sym); 3115 } 3116 3117 size_t PPC64LongBranchTargetSection::getSize() const { 3118 return Entries.size() * 8; 3119 } 3120 3121 void PPC64LongBranchTargetSection::writeTo(uint8_t *Buf) { 3122 assert(Target->GotPltEntrySize == 8); 3123 // If linking non-pic we have the final addresses of the targets and they get 3124 // written to the table directly. For pic the dynamic linker will allocate 3125 // the section and fill it it. 3126 if (Config->Pic) 3127 return; 3128 3129 for (const Symbol *Sym : Entries) { 3130 assert(Sym->getVA()); 3131 // Need calls to branch to the local entry-point since a long-branch 3132 // must be a local-call. 3133 write64(Buf, 3134 Sym->getVA() + getPPC64GlobalEntryToLocalEntryOffset(Sym->StOther)); 3135 Buf += Target->GotPltEntrySize; 3136 } 3137 } 3138 3139 bool PPC64LongBranchTargetSection::empty() const { 3140 // `removeUnusedSyntheticSections()` is called before thunk allocation which 3141 // is too early to determine if this section will be empty or not. We need 3142 // Finalized to keep the section alive until after thunk creation. Finalized 3143 // only gets set to true once `finalizeSections()` is called after thunk 3144 // creation. Becuase of this, if we don't create any long-branch thunks we end 3145 // up with an empty .branch_lt section in the binary. 3146 return Finalized && Entries.empty(); 3147 } 3148 3149 InStruct elf::In; 3150 3151 template GdbIndexSection *GdbIndexSection::create<ELF32LE>(); 3152 template GdbIndexSection *GdbIndexSection::create<ELF32BE>(); 3153 template GdbIndexSection *GdbIndexSection::create<ELF64LE>(); 3154 template GdbIndexSection *GdbIndexSection::create<ELF64BE>(); 3155 3156 template void elf::splitSections<ELF32LE>(); 3157 template void elf::splitSections<ELF32BE>(); 3158 template void elf::splitSections<ELF64LE>(); 3159 template void elf::splitSections<ELF64BE>(); 3160 3161 template void EhFrameSection::addSection<ELF32LE>(InputSectionBase *); 3162 template void EhFrameSection::addSection<ELF32BE>(InputSectionBase *); 3163 template void EhFrameSection::addSection<ELF64LE>(InputSectionBase *); 3164 template void EhFrameSection::addSection<ELF64BE>(InputSectionBase *); 3165 3166 template void PltSection::addEntry<ELF32LE>(Symbol &Sym); 3167 template void PltSection::addEntry<ELF32BE>(Symbol &Sym); 3168 template void PltSection::addEntry<ELF64LE>(Symbol &Sym); 3169 template void PltSection::addEntry<ELF64BE>(Symbol &Sym); 3170 3171 template void MipsGotSection::build<ELF32LE>(); 3172 template void MipsGotSection::build<ELF32BE>(); 3173 template void MipsGotSection::build<ELF64LE>(); 3174 template void MipsGotSection::build<ELF64BE>(); 3175 3176 template class elf::MipsAbiFlagsSection<ELF32LE>; 3177 template class elf::MipsAbiFlagsSection<ELF32BE>; 3178 template class elf::MipsAbiFlagsSection<ELF64LE>; 3179 template class elf::MipsAbiFlagsSection<ELF64BE>; 3180 3181 template class elf::MipsOptionsSection<ELF32LE>; 3182 template class elf::MipsOptionsSection<ELF32BE>; 3183 template class elf::MipsOptionsSection<ELF64LE>; 3184 template class elf::MipsOptionsSection<ELF64BE>; 3185 3186 template class elf::MipsReginfoSection<ELF32LE>; 3187 template class elf::MipsReginfoSection<ELF32BE>; 3188 template class elf::MipsReginfoSection<ELF64LE>; 3189 template class elf::MipsReginfoSection<ELF64BE>; 3190 3191 template class elf::DynamicSection<ELF32LE>; 3192 template class elf::DynamicSection<ELF32BE>; 3193 template class elf::DynamicSection<ELF64LE>; 3194 template class elf::DynamicSection<ELF64BE>; 3195 3196 template class elf::RelocationSection<ELF32LE>; 3197 template class elf::RelocationSection<ELF32BE>; 3198 template class elf::RelocationSection<ELF64LE>; 3199 template class elf::RelocationSection<ELF64BE>; 3200 3201 template class elf::AndroidPackedRelocationSection<ELF32LE>; 3202 template class elf::AndroidPackedRelocationSection<ELF32BE>; 3203 template class elf::AndroidPackedRelocationSection<ELF64LE>; 3204 template class elf::AndroidPackedRelocationSection<ELF64BE>; 3205 3206 template class elf::RelrSection<ELF32LE>; 3207 template class elf::RelrSection<ELF32BE>; 3208 template class elf::RelrSection<ELF64LE>; 3209 template class elf::RelrSection<ELF64BE>; 3210 3211 template class elf::SymbolTableSection<ELF32LE>; 3212 template class elf::SymbolTableSection<ELF32BE>; 3213 template class elf::SymbolTableSection<ELF64LE>; 3214 template class elf::SymbolTableSection<ELF64BE>; 3215 3216 template class elf::VersionTableSection<ELF32LE>; 3217 template class elf::VersionTableSection<ELF32BE>; 3218 template class elf::VersionTableSection<ELF64LE>; 3219 template class elf::VersionTableSection<ELF64BE>; 3220 3221 template class elf::VersionNeedSection<ELF32LE>; 3222 template class elf::VersionNeedSection<ELF32BE>; 3223 template class elf::VersionNeedSection<ELF64LE>; 3224 template class elf::VersionNeedSection<ELF64BE>; 3225