1 //===- SyntheticSections.cpp ----------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains linker-synthesized sections. Currently, 10 // synthetic sections are created either output sections or input sections, 11 // but we are rewriting code so that all synthetic sections are created as 12 // input sections. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "SyntheticSections.h" 17 #include "Config.h" 18 #include "DWARF.h" 19 #include "EhFrame.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 "Thunks.h" 27 #include "Writer.h" 28 #include "lld/Common/CommonLinkerContext.h" 29 #include "lld/Common/DWARF.h" 30 #include "lld/Common/Strings.h" 31 #include "lld/Common/Version.h" 32 #include "llvm/ADT/STLExtras.h" 33 #include "llvm/ADT/SetOperations.h" 34 #include "llvm/ADT/StringExtras.h" 35 #include "llvm/BinaryFormat/Dwarf.h" 36 #include "llvm/BinaryFormat/ELF.h" 37 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h" 38 #include "llvm/Support/Endian.h" 39 #include "llvm/Support/LEB128.h" 40 #include "llvm/Support/Parallel.h" 41 #include "llvm/Support/TimeProfiler.h" 42 #include <cstdlib> 43 44 using namespace llvm; 45 using namespace llvm::dwarf; 46 using namespace llvm::ELF; 47 using namespace llvm::object; 48 using namespace llvm::support; 49 using namespace lld; 50 using namespace lld::elf; 51 52 using llvm::support::endian::read32le; 53 using llvm::support::endian::write32le; 54 using llvm::support::endian::write64le; 55 56 constexpr size_t MergeNoTailSection::numShards; 57 58 static uint64_t readUint(uint8_t *buf) { 59 return config->is64 ? read64(buf) : read32(buf); 60 } 61 62 static void writeUint(uint8_t *buf, uint64_t val) { 63 if (config->is64) 64 write64(buf, val); 65 else 66 write32(buf, val); 67 } 68 69 // Returns an LLD version string. 70 static ArrayRef<uint8_t> getVersion() { 71 // Check LLD_VERSION first for ease of testing. 72 // You can get consistent output by using the environment variable. 73 // This is only for testing. 74 StringRef s = getenv("LLD_VERSION"); 75 if (s.empty()) 76 s = saver().save(Twine("Linker: ") + getLLDVersion()); 77 78 // +1 to include the terminating '\0'. 79 return {(const uint8_t *)s.data(), s.size() + 1}; 80 } 81 82 // Creates a .comment section containing LLD version info. 83 // With this feature, you can identify LLD-generated binaries easily 84 // by "readelf --string-dump .comment <file>". 85 // The returned object is a mergeable string section. 86 MergeInputSection *elf::createCommentSection() { 87 auto *sec = make<MergeInputSection>(SHF_MERGE | SHF_STRINGS, SHT_PROGBITS, 1, 88 getVersion(), ".comment"); 89 sec->splitIntoPieces(); 90 return sec; 91 } 92 93 // .MIPS.abiflags section. 94 template <class ELFT> 95 MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags flags) 96 : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags"), 97 flags(flags) { 98 this->entsize = sizeof(Elf_Mips_ABIFlags); 99 } 100 101 template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *buf) { 102 memcpy(buf, &flags, sizeof(flags)); 103 } 104 105 template <class ELFT> 106 std::unique_ptr<MipsAbiFlagsSection<ELFT>> MipsAbiFlagsSection<ELFT>::create() { 107 Elf_Mips_ABIFlags flags = {}; 108 bool create = false; 109 110 for (InputSectionBase *sec : ctx.inputSections) { 111 if (sec->type != SHT_MIPS_ABIFLAGS) 112 continue; 113 sec->markDead(); 114 create = true; 115 116 std::string filename = toString(sec->file); 117 const size_t size = sec->content().size(); 118 // Older version of BFD (such as the default FreeBSD linker) concatenate 119 // .MIPS.abiflags instead of merging. To allow for this case (or potential 120 // zero padding) we ignore everything after the first Elf_Mips_ABIFlags 121 if (size < sizeof(Elf_Mips_ABIFlags)) { 122 error(filename + ": invalid size of .MIPS.abiflags section: got " + 123 Twine(size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags))); 124 return nullptr; 125 } 126 auto *s = 127 reinterpret_cast<const Elf_Mips_ABIFlags *>(sec->content().data()); 128 if (s->version != 0) { 129 error(filename + ": unexpected .MIPS.abiflags version " + 130 Twine(s->version)); 131 return nullptr; 132 } 133 134 // LLD checks ISA compatibility in calcMipsEFlags(). Here we just 135 // select the highest number of ISA/Rev/Ext. 136 flags.isa_level = std::max(flags.isa_level, s->isa_level); 137 flags.isa_rev = std::max(flags.isa_rev, s->isa_rev); 138 flags.isa_ext = std::max(flags.isa_ext, s->isa_ext); 139 flags.gpr_size = std::max(flags.gpr_size, s->gpr_size); 140 flags.cpr1_size = std::max(flags.cpr1_size, s->cpr1_size); 141 flags.cpr2_size = std::max(flags.cpr2_size, s->cpr2_size); 142 flags.ases |= s->ases; 143 flags.flags1 |= s->flags1; 144 flags.flags2 |= s->flags2; 145 flags.fp_abi = elf::getMipsFpAbiFlag(flags.fp_abi, s->fp_abi, filename); 146 }; 147 148 if (create) 149 return std::make_unique<MipsAbiFlagsSection<ELFT>>(flags); 150 return nullptr; 151 } 152 153 // .MIPS.options section. 154 template <class ELFT> 155 MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo reginfo) 156 : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"), 157 reginfo(reginfo) { 158 this->entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo); 159 } 160 161 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *buf) { 162 auto *options = reinterpret_cast<Elf_Mips_Options *>(buf); 163 options->kind = ODK_REGINFO; 164 options->size = getSize(); 165 166 if (!config->relocatable) 167 reginfo.ri_gp_value = in.mipsGot->getGp(); 168 memcpy(buf + sizeof(Elf_Mips_Options), ®info, sizeof(reginfo)); 169 } 170 171 template <class ELFT> 172 std::unique_ptr<MipsOptionsSection<ELFT>> MipsOptionsSection<ELFT>::create() { 173 // N64 ABI only. 174 if (!ELFT::Is64Bits) 175 return nullptr; 176 177 SmallVector<InputSectionBase *, 0> sections; 178 for (InputSectionBase *sec : ctx.inputSections) 179 if (sec->type == SHT_MIPS_OPTIONS) 180 sections.push_back(sec); 181 182 if (sections.empty()) 183 return nullptr; 184 185 Elf_Mips_RegInfo reginfo = {}; 186 for (InputSectionBase *sec : sections) { 187 sec->markDead(); 188 189 std::string filename = toString(sec->file); 190 ArrayRef<uint8_t> d = sec->content(); 191 192 while (!d.empty()) { 193 if (d.size() < sizeof(Elf_Mips_Options)) { 194 error(filename + ": invalid size of .MIPS.options section"); 195 break; 196 } 197 198 auto *opt = reinterpret_cast<const Elf_Mips_Options *>(d.data()); 199 if (opt->kind == ODK_REGINFO) { 200 reginfo.ri_gprmask |= opt->getRegInfo().ri_gprmask; 201 sec->getFile<ELFT>()->mipsGp0 = opt->getRegInfo().ri_gp_value; 202 break; 203 } 204 205 if (!opt->size) 206 fatal(filename + ": zero option descriptor size"); 207 d = d.slice(opt->size); 208 } 209 }; 210 211 return std::make_unique<MipsOptionsSection<ELFT>>(reginfo); 212 } 213 214 // MIPS .reginfo section. 215 template <class ELFT> 216 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo reginfo) 217 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"), 218 reginfo(reginfo) { 219 this->entsize = sizeof(Elf_Mips_RegInfo); 220 } 221 222 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *buf) { 223 if (!config->relocatable) 224 reginfo.ri_gp_value = in.mipsGot->getGp(); 225 memcpy(buf, ®info, sizeof(reginfo)); 226 } 227 228 template <class ELFT> 229 std::unique_ptr<MipsReginfoSection<ELFT>> MipsReginfoSection<ELFT>::create() { 230 // Section should be alive for O32 and N32 ABIs only. 231 if (ELFT::Is64Bits) 232 return nullptr; 233 234 SmallVector<InputSectionBase *, 0> sections; 235 for (InputSectionBase *sec : ctx.inputSections) 236 if (sec->type == SHT_MIPS_REGINFO) 237 sections.push_back(sec); 238 239 if (sections.empty()) 240 return nullptr; 241 242 Elf_Mips_RegInfo reginfo = {}; 243 for (InputSectionBase *sec : sections) { 244 sec->markDead(); 245 246 if (sec->content().size() != sizeof(Elf_Mips_RegInfo)) { 247 error(toString(sec->file) + ": invalid size of .reginfo section"); 248 return nullptr; 249 } 250 251 auto *r = reinterpret_cast<const Elf_Mips_RegInfo *>(sec->content().data()); 252 reginfo.ri_gprmask |= r->ri_gprmask; 253 sec->getFile<ELFT>()->mipsGp0 = r->ri_gp_value; 254 }; 255 256 return std::make_unique<MipsReginfoSection<ELFT>>(reginfo); 257 } 258 259 InputSection *elf::createInterpSection() { 260 // StringSaver guarantees that the returned string ends with '\0'. 261 StringRef s = saver().save(config->dynamicLinker); 262 ArrayRef<uint8_t> contents = {(const uint8_t *)s.data(), s.size() + 1}; 263 264 return make<InputSection>(nullptr, SHF_ALLOC, SHT_PROGBITS, 1, contents, 265 ".interp"); 266 } 267 268 Defined *elf::addSyntheticLocal(StringRef name, uint8_t type, uint64_t value, 269 uint64_t size, InputSectionBase §ion) { 270 Defined *s = makeDefined(section.file, name, STB_LOCAL, STV_DEFAULT, type, 271 value, size, §ion); 272 if (in.symTab) 273 in.symTab->addSymbol(s); 274 275 if (config->emachine == EM_ARM && !config->isLE && config->armBe8 && 276 (section.flags & SHF_EXECINSTR)) 277 // Adding Linker generated mapping symbols to the arm specific mapping 278 // symbols list. 279 addArmSyntheticSectionMappingSymbol(s); 280 281 return s; 282 } 283 284 static size_t getHashSize() { 285 switch (config->buildId) { 286 case BuildIdKind::Fast: 287 return 8; 288 case BuildIdKind::Md5: 289 case BuildIdKind::Uuid: 290 return 16; 291 case BuildIdKind::Sha1: 292 return 20; 293 case BuildIdKind::Hexstring: 294 return config->buildIdVector.size(); 295 default: 296 llvm_unreachable("unknown BuildIdKind"); 297 } 298 } 299 300 // This class represents a linker-synthesized .note.gnu.property section. 301 // 302 // In x86 and AArch64, object files may contain feature flags indicating the 303 // features that they have used. The flags are stored in a .note.gnu.property 304 // section. 305 // 306 // lld reads the sections from input files and merges them by computing AND of 307 // the flags. The result is written as a new .note.gnu.property section. 308 // 309 // If the flag is zero (which indicates that the intersection of the feature 310 // sets is empty, or some input files didn't have .note.gnu.property sections), 311 // we don't create this section. 312 GnuPropertySection::GnuPropertySection() 313 : SyntheticSection(llvm::ELF::SHF_ALLOC, llvm::ELF::SHT_NOTE, 314 config->wordsize, ".note.gnu.property") {} 315 316 void GnuPropertySection::writeTo(uint8_t *buf) { 317 uint32_t featureAndType = config->emachine == EM_AARCH64 318 ? GNU_PROPERTY_AARCH64_FEATURE_1_AND 319 : GNU_PROPERTY_X86_FEATURE_1_AND; 320 321 write32(buf, 4); // Name size 322 write32(buf + 4, config->is64 ? 16 : 12); // Content size 323 write32(buf + 8, NT_GNU_PROPERTY_TYPE_0); // Type 324 memcpy(buf + 12, "GNU", 4); // Name string 325 write32(buf + 16, featureAndType); // Feature type 326 write32(buf + 20, 4); // Feature size 327 write32(buf + 24, config->andFeatures); // Feature flags 328 if (config->is64) 329 write32(buf + 28, 0); // Padding 330 } 331 332 size_t GnuPropertySection::getSize() const { return config->is64 ? 32 : 28; } 333 334 BuildIdSection::BuildIdSection() 335 : SyntheticSection(SHF_ALLOC, SHT_NOTE, 4, ".note.gnu.build-id"), 336 hashSize(getHashSize()) {} 337 338 void BuildIdSection::writeTo(uint8_t *buf) { 339 write32(buf, 4); // Name size 340 write32(buf + 4, hashSize); // Content size 341 write32(buf + 8, NT_GNU_BUILD_ID); // Type 342 memcpy(buf + 12, "GNU", 4); // Name string 343 hashBuf = buf + 16; 344 } 345 346 void BuildIdSection::writeBuildId(ArrayRef<uint8_t> buf) { 347 assert(buf.size() == hashSize); 348 memcpy(hashBuf, buf.data(), hashSize); 349 } 350 351 BssSection::BssSection(StringRef name, uint64_t size, uint32_t alignment) 352 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, alignment, name) { 353 this->bss = true; 354 this->size = size; 355 } 356 357 EhFrameSection::EhFrameSection() 358 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame") {} 359 360 // Search for an existing CIE record or create a new one. 361 // CIE records from input object files are uniquified by their contents 362 // and where their relocations point to. 363 template <class ELFT, class RelTy> 364 CieRecord *EhFrameSection::addCie(EhSectionPiece &cie, ArrayRef<RelTy> rels) { 365 Symbol *personality = nullptr; 366 unsigned firstRelI = cie.firstRelocation; 367 if (firstRelI != (unsigned)-1) 368 personality = 369 &cie.sec->template getFile<ELFT>()->getRelocTargetSym(rels[firstRelI]); 370 371 // Search for an existing CIE by CIE contents/relocation target pair. 372 CieRecord *&rec = cieMap[{cie.data(), personality}]; 373 374 // If not found, create a new one. 375 if (!rec) { 376 rec = make<CieRecord>(); 377 rec->cie = &cie; 378 cieRecords.push_back(rec); 379 } 380 return rec; 381 } 382 383 // There is one FDE per function. Returns a non-null pointer to the function 384 // symbol if the given FDE points to a live function. 385 template <class ELFT, class RelTy> 386 Defined *EhFrameSection::isFdeLive(EhSectionPiece &fde, ArrayRef<RelTy> rels) { 387 auto *sec = cast<EhInputSection>(fde.sec); 388 unsigned firstRelI = fde.firstRelocation; 389 390 // An FDE should point to some function because FDEs are to describe 391 // functions. That's however not always the case due to an issue of 392 // ld.gold with -r. ld.gold may discard only functions and leave their 393 // corresponding FDEs, which results in creating bad .eh_frame sections. 394 // To deal with that, we ignore such FDEs. 395 if (firstRelI == (unsigned)-1) 396 return nullptr; 397 398 const RelTy &rel = rels[firstRelI]; 399 Symbol &b = sec->template getFile<ELFT>()->getRelocTargetSym(rel); 400 401 // FDEs for garbage-collected or merged-by-ICF sections, or sections in 402 // another partition, are dead. 403 if (auto *d = dyn_cast<Defined>(&b)) 404 if (!d->folded && d->section && d->section->partition == partition) 405 return d; 406 return nullptr; 407 } 408 409 // .eh_frame is a sequence of CIE or FDE records. In general, there 410 // is one CIE record per input object file which is followed by 411 // a list of FDEs. This function searches an existing CIE or create a new 412 // one and associates FDEs to the CIE. 413 template <class ELFT, class RelTy> 414 void EhFrameSection::addRecords(EhInputSection *sec, ArrayRef<RelTy> rels) { 415 offsetToCie.clear(); 416 for (EhSectionPiece &cie : sec->cies) 417 offsetToCie[cie.inputOff] = addCie<ELFT>(cie, rels); 418 for (EhSectionPiece &fde : sec->fdes) { 419 uint32_t id = endian::read32<ELFT::TargetEndianness>(fde.data().data() + 4); 420 CieRecord *rec = offsetToCie[fde.inputOff + 4 - id]; 421 if (!rec) 422 fatal(toString(sec) + ": invalid CIE reference"); 423 424 if (!isFdeLive<ELFT>(fde, rels)) 425 continue; 426 rec->fdes.push_back(&fde); 427 numFdes++; 428 } 429 } 430 431 template <class ELFT> 432 void EhFrameSection::addSectionAux(EhInputSection *sec) { 433 if (!sec->isLive()) 434 return; 435 const RelsOrRelas<ELFT> rels = sec->template relsOrRelas<ELFT>(); 436 if (rels.areRelocsRel()) 437 addRecords<ELFT>(sec, rels.rels); 438 else 439 addRecords<ELFT>(sec, rels.relas); 440 } 441 442 // Used by ICF<ELFT>::handleLSDA(). This function is very similar to 443 // EhFrameSection::addRecords(). 444 template <class ELFT, class RelTy> 445 void EhFrameSection::iterateFDEWithLSDAAux( 446 EhInputSection &sec, ArrayRef<RelTy> rels, DenseSet<size_t> &ciesWithLSDA, 447 llvm::function_ref<void(InputSection &)> fn) { 448 for (EhSectionPiece &cie : sec.cies) 449 if (hasLSDA(cie)) 450 ciesWithLSDA.insert(cie.inputOff); 451 for (EhSectionPiece &fde : sec.fdes) { 452 uint32_t id = endian::read32<ELFT::TargetEndianness>(fde.data().data() + 4); 453 if (!ciesWithLSDA.contains(fde.inputOff + 4 - id)) 454 continue; 455 456 // The CIE has a LSDA argument. Call fn with d's section. 457 if (Defined *d = isFdeLive<ELFT>(fde, rels)) 458 if (auto *s = dyn_cast_or_null<InputSection>(d->section)) 459 fn(*s); 460 } 461 } 462 463 template <class ELFT> 464 void EhFrameSection::iterateFDEWithLSDA( 465 llvm::function_ref<void(InputSection &)> fn) { 466 DenseSet<size_t> ciesWithLSDA; 467 for (EhInputSection *sec : sections) { 468 ciesWithLSDA.clear(); 469 const RelsOrRelas<ELFT> rels = sec->template relsOrRelas<ELFT>(); 470 if (rels.areRelocsRel()) 471 iterateFDEWithLSDAAux<ELFT>(*sec, rels.rels, ciesWithLSDA, fn); 472 else 473 iterateFDEWithLSDAAux<ELFT>(*sec, rels.relas, ciesWithLSDA, fn); 474 } 475 } 476 477 static void writeCieFde(uint8_t *buf, ArrayRef<uint8_t> d) { 478 memcpy(buf, d.data(), d.size()); 479 // Fix the size field. -4 since size does not include the size field itself. 480 write32(buf, d.size() - 4); 481 } 482 483 void EhFrameSection::finalizeContents() { 484 assert(!this->size); // Not finalized. 485 486 switch (config->ekind) { 487 case ELFNoneKind: 488 llvm_unreachable("invalid ekind"); 489 case ELF32LEKind: 490 for (EhInputSection *sec : sections) 491 addSectionAux<ELF32LE>(sec); 492 break; 493 case ELF32BEKind: 494 for (EhInputSection *sec : sections) 495 addSectionAux<ELF32BE>(sec); 496 break; 497 case ELF64LEKind: 498 for (EhInputSection *sec : sections) 499 addSectionAux<ELF64LE>(sec); 500 break; 501 case ELF64BEKind: 502 for (EhInputSection *sec : sections) 503 addSectionAux<ELF64BE>(sec); 504 break; 505 } 506 507 size_t off = 0; 508 for (CieRecord *rec : cieRecords) { 509 rec->cie->outputOff = off; 510 off += rec->cie->size; 511 512 for (EhSectionPiece *fde : rec->fdes) { 513 fde->outputOff = off; 514 off += fde->size; 515 } 516 } 517 518 // The LSB standard does not allow a .eh_frame section with zero 519 // Call Frame Information records. glibc unwind-dw2-fde.c 520 // classify_object_over_fdes expects there is a CIE record length 0 as a 521 // terminator. Thus we add one unconditionally. 522 off += 4; 523 524 this->size = off; 525 } 526 527 // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table 528 // to get an FDE from an address to which FDE is applied. This function 529 // returns a list of such pairs. 530 SmallVector<EhFrameSection::FdeData, 0> EhFrameSection::getFdeData() const { 531 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff; 532 SmallVector<FdeData, 0> ret; 533 534 uint64_t va = getPartition().ehFrameHdr->getVA(); 535 for (CieRecord *rec : cieRecords) { 536 uint8_t enc = getFdeEncoding(rec->cie); 537 for (EhSectionPiece *fde : rec->fdes) { 538 uint64_t pc = getFdePc(buf, fde->outputOff, enc); 539 uint64_t fdeVA = getParent()->addr + fde->outputOff; 540 if (!isInt<32>(pc - va)) 541 fatal(toString(fde->sec) + ": PC offset is too large: 0x" + 542 Twine::utohexstr(pc - va)); 543 ret.push_back({uint32_t(pc - va), uint32_t(fdeVA - va)}); 544 } 545 } 546 547 // Sort the FDE list by their PC and uniqueify. Usually there is only 548 // one FDE for a PC (i.e. function), but if ICF merges two functions 549 // into one, there can be more than one FDEs pointing to the address. 550 auto less = [](const FdeData &a, const FdeData &b) { 551 return a.pcRel < b.pcRel; 552 }; 553 llvm::stable_sort(ret, less); 554 auto eq = [](const FdeData &a, const FdeData &b) { 555 return a.pcRel == b.pcRel; 556 }; 557 ret.erase(std::unique(ret.begin(), ret.end(), eq), ret.end()); 558 559 return ret; 560 } 561 562 static uint64_t readFdeAddr(uint8_t *buf, int size) { 563 switch (size) { 564 case DW_EH_PE_udata2: 565 return read16(buf); 566 case DW_EH_PE_sdata2: 567 return (int16_t)read16(buf); 568 case DW_EH_PE_udata4: 569 return read32(buf); 570 case DW_EH_PE_sdata4: 571 return (int32_t)read32(buf); 572 case DW_EH_PE_udata8: 573 case DW_EH_PE_sdata8: 574 return read64(buf); 575 case DW_EH_PE_absptr: 576 return readUint(buf); 577 } 578 fatal("unknown FDE size encoding"); 579 } 580 581 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to. 582 // We need it to create .eh_frame_hdr section. 583 uint64_t EhFrameSection::getFdePc(uint8_t *buf, size_t fdeOff, 584 uint8_t enc) const { 585 // The starting address to which this FDE applies is 586 // stored at FDE + 8 byte. And this offset is within 587 // the .eh_frame section. 588 size_t off = fdeOff + 8; 589 uint64_t addr = readFdeAddr(buf + off, enc & 0xf); 590 if ((enc & 0x70) == DW_EH_PE_absptr) 591 return addr; 592 if ((enc & 0x70) == DW_EH_PE_pcrel) 593 return addr + getParent()->addr + off + outSecOff; 594 fatal("unknown FDE size relative encoding"); 595 } 596 597 void EhFrameSection::writeTo(uint8_t *buf) { 598 // Write CIE and FDE records. 599 for (CieRecord *rec : cieRecords) { 600 size_t cieOffset = rec->cie->outputOff; 601 writeCieFde(buf + cieOffset, rec->cie->data()); 602 603 for (EhSectionPiece *fde : rec->fdes) { 604 size_t off = fde->outputOff; 605 writeCieFde(buf + off, fde->data()); 606 607 // FDE's second word should have the offset to an associated CIE. 608 // Write it. 609 write32(buf + off + 4, off + 4 - cieOffset); 610 } 611 } 612 613 // Apply relocations. .eh_frame section contents are not contiguous 614 // in the output buffer, but relocateAlloc() still works because 615 // getOffset() takes care of discontiguous section pieces. 616 for (EhInputSection *s : sections) 617 target->relocateAlloc(*s, buf); 618 619 if (getPartition().ehFrameHdr && getPartition().ehFrameHdr->getParent()) 620 getPartition().ehFrameHdr->write(); 621 } 622 623 GotSection::GotSection() 624 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 625 target->gotEntrySize, ".got") { 626 numEntries = target->gotHeaderEntriesNum; 627 } 628 629 void GotSection::addConstant(const Relocation &r) { relocations.push_back(r); } 630 void GotSection::addEntry(Symbol &sym) { 631 assert(sym.auxIdx == symAux.size() - 1); 632 symAux.back().gotIdx = numEntries++; 633 } 634 635 bool GotSection::addTlsDescEntry(Symbol &sym) { 636 assert(sym.auxIdx == symAux.size() - 1); 637 symAux.back().tlsDescIdx = numEntries; 638 numEntries += 2; 639 return true; 640 } 641 642 bool GotSection::addDynTlsEntry(Symbol &sym) { 643 assert(sym.auxIdx == symAux.size() - 1); 644 symAux.back().tlsGdIdx = numEntries; 645 // Global Dynamic TLS entries take two GOT slots. 646 numEntries += 2; 647 return true; 648 } 649 650 // Reserves TLS entries for a TLS module ID and a TLS block offset. 651 // In total it takes two GOT slots. 652 bool GotSection::addTlsIndex() { 653 if (tlsIndexOff != uint32_t(-1)) 654 return false; 655 tlsIndexOff = numEntries * config->wordsize; 656 numEntries += 2; 657 return true; 658 } 659 660 uint32_t GotSection::getTlsDescOffset(const Symbol &sym) const { 661 return sym.getTlsDescIdx() * config->wordsize; 662 } 663 664 uint64_t GotSection::getTlsDescAddr(const Symbol &sym) const { 665 return getVA() + getTlsDescOffset(sym); 666 } 667 668 uint64_t GotSection::getGlobalDynAddr(const Symbol &b) const { 669 return this->getVA() + b.getTlsGdIdx() * config->wordsize; 670 } 671 672 uint64_t GotSection::getGlobalDynOffset(const Symbol &b) const { 673 return b.getTlsGdIdx() * config->wordsize; 674 } 675 676 void GotSection::finalizeContents() { 677 if (config->emachine == EM_PPC64 && 678 numEntries <= target->gotHeaderEntriesNum && !ElfSym::globalOffsetTable) 679 size = 0; 680 else 681 size = numEntries * config->wordsize; 682 } 683 684 bool GotSection::isNeeded() const { 685 // Needed if the GOT symbol is used or the number of entries is more than just 686 // the header. A GOT with just the header may not be needed. 687 return hasGotOffRel || numEntries > target->gotHeaderEntriesNum; 688 } 689 690 void GotSection::writeTo(uint8_t *buf) { 691 // On PPC64 .got may be needed but empty. Skip the write. 692 if (size == 0) 693 return; 694 target->writeGotHeader(buf); 695 target->relocateAlloc(*this, buf); 696 } 697 698 static uint64_t getMipsPageAddr(uint64_t addr) { 699 return (addr + 0x8000) & ~0xffff; 700 } 701 702 static uint64_t getMipsPageCount(uint64_t size) { 703 return (size + 0xfffe) / 0xffff + 1; 704 } 705 706 MipsGotSection::MipsGotSection() 707 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16, 708 ".got") {} 709 710 void MipsGotSection::addEntry(InputFile &file, Symbol &sym, int64_t addend, 711 RelExpr expr) { 712 FileGot &g = getGot(file); 713 if (expr == R_MIPS_GOT_LOCAL_PAGE) { 714 if (const OutputSection *os = sym.getOutputSection()) 715 g.pagesMap.insert({os, {}}); 716 else 717 g.local16.insert({{nullptr, getMipsPageAddr(sym.getVA(addend))}, 0}); 718 } else if (sym.isTls()) 719 g.tls.insert({&sym, 0}); 720 else if (sym.isPreemptible && expr == R_ABS) 721 g.relocs.insert({&sym, 0}); 722 else if (sym.isPreemptible) 723 g.global.insert({&sym, 0}); 724 else if (expr == R_MIPS_GOT_OFF32) 725 g.local32.insert({{&sym, addend}, 0}); 726 else 727 g.local16.insert({{&sym, addend}, 0}); 728 } 729 730 void MipsGotSection::addDynTlsEntry(InputFile &file, Symbol &sym) { 731 getGot(file).dynTlsSymbols.insert({&sym, 0}); 732 } 733 734 void MipsGotSection::addTlsIndex(InputFile &file) { 735 getGot(file).dynTlsSymbols.insert({nullptr, 0}); 736 } 737 738 size_t MipsGotSection::FileGot::getEntriesNum() const { 739 return getPageEntriesNum() + local16.size() + global.size() + relocs.size() + 740 tls.size() + dynTlsSymbols.size() * 2; 741 } 742 743 size_t MipsGotSection::FileGot::getPageEntriesNum() const { 744 size_t num = 0; 745 for (const std::pair<const OutputSection *, FileGot::PageBlock> &p : pagesMap) 746 num += p.second.count; 747 return num; 748 } 749 750 size_t MipsGotSection::FileGot::getIndexedEntriesNum() const { 751 size_t count = getPageEntriesNum() + local16.size() + global.size(); 752 // If there are relocation-only entries in the GOT, TLS entries 753 // are allocated after them. TLS entries should be addressable 754 // by 16-bit index so count both reloc-only and TLS entries. 755 if (!tls.empty() || !dynTlsSymbols.empty()) 756 count += relocs.size() + tls.size() + dynTlsSymbols.size() * 2; 757 return count; 758 } 759 760 MipsGotSection::FileGot &MipsGotSection::getGot(InputFile &f) { 761 if (f.mipsGotIndex == uint32_t(-1)) { 762 gots.emplace_back(); 763 gots.back().file = &f; 764 f.mipsGotIndex = gots.size() - 1; 765 } 766 return gots[f.mipsGotIndex]; 767 } 768 769 uint64_t MipsGotSection::getPageEntryOffset(const InputFile *f, 770 const Symbol &sym, 771 int64_t addend) const { 772 const FileGot &g = gots[f->mipsGotIndex]; 773 uint64_t index = 0; 774 if (const OutputSection *outSec = sym.getOutputSection()) { 775 uint64_t secAddr = getMipsPageAddr(outSec->addr); 776 uint64_t symAddr = getMipsPageAddr(sym.getVA(addend)); 777 index = g.pagesMap.lookup(outSec).firstIndex + (symAddr - secAddr) / 0xffff; 778 } else { 779 index = g.local16.lookup({nullptr, getMipsPageAddr(sym.getVA(addend))}); 780 } 781 return index * config->wordsize; 782 } 783 784 uint64_t MipsGotSection::getSymEntryOffset(const InputFile *f, const Symbol &s, 785 int64_t addend) const { 786 const FileGot &g = gots[f->mipsGotIndex]; 787 Symbol *sym = const_cast<Symbol *>(&s); 788 if (sym->isTls()) 789 return g.tls.lookup(sym) * config->wordsize; 790 if (sym->isPreemptible) 791 return g.global.lookup(sym) * config->wordsize; 792 return g.local16.lookup({sym, addend}) * config->wordsize; 793 } 794 795 uint64_t MipsGotSection::getTlsIndexOffset(const InputFile *f) const { 796 const FileGot &g = gots[f->mipsGotIndex]; 797 return g.dynTlsSymbols.lookup(nullptr) * config->wordsize; 798 } 799 800 uint64_t MipsGotSection::getGlobalDynOffset(const InputFile *f, 801 const Symbol &s) const { 802 const FileGot &g = gots[f->mipsGotIndex]; 803 Symbol *sym = const_cast<Symbol *>(&s); 804 return g.dynTlsSymbols.lookup(sym) * config->wordsize; 805 } 806 807 const Symbol *MipsGotSection::getFirstGlobalEntry() const { 808 if (gots.empty()) 809 return nullptr; 810 const FileGot &primGot = gots.front(); 811 if (!primGot.global.empty()) 812 return primGot.global.front().first; 813 if (!primGot.relocs.empty()) 814 return primGot.relocs.front().first; 815 return nullptr; 816 } 817 818 unsigned MipsGotSection::getLocalEntriesNum() const { 819 if (gots.empty()) 820 return headerEntriesNum; 821 return headerEntriesNum + gots.front().getPageEntriesNum() + 822 gots.front().local16.size(); 823 } 824 825 bool MipsGotSection::tryMergeGots(FileGot &dst, FileGot &src, bool isPrimary) { 826 FileGot tmp = dst; 827 set_union(tmp.pagesMap, src.pagesMap); 828 set_union(tmp.local16, src.local16); 829 set_union(tmp.global, src.global); 830 set_union(tmp.relocs, src.relocs); 831 set_union(tmp.tls, src.tls); 832 set_union(tmp.dynTlsSymbols, src.dynTlsSymbols); 833 834 size_t count = isPrimary ? headerEntriesNum : 0; 835 count += tmp.getIndexedEntriesNum(); 836 837 if (count * config->wordsize > config->mipsGotSize) 838 return false; 839 840 std::swap(tmp, dst); 841 return true; 842 } 843 844 void MipsGotSection::finalizeContents() { updateAllocSize(); } 845 846 bool MipsGotSection::updateAllocSize() { 847 size = headerEntriesNum * config->wordsize; 848 for (const FileGot &g : gots) 849 size += g.getEntriesNum() * config->wordsize; 850 return false; 851 } 852 853 void MipsGotSection::build() { 854 if (gots.empty()) 855 return; 856 857 std::vector<FileGot> mergedGots(1); 858 859 // For each GOT move non-preemptible symbols from the `Global` 860 // to `Local16` list. Preemptible symbol might become non-preemptible 861 // one if, for example, it gets a related copy relocation. 862 for (FileGot &got : gots) { 863 for (auto &p: got.global) 864 if (!p.first->isPreemptible) 865 got.local16.insert({{p.first, 0}, 0}); 866 got.global.remove_if([&](const std::pair<Symbol *, size_t> &p) { 867 return !p.first->isPreemptible; 868 }); 869 } 870 871 // For each GOT remove "reloc-only" entry if there is "global" 872 // entry for the same symbol. And add local entries which indexed 873 // using 32-bit value at the end of 16-bit entries. 874 for (FileGot &got : gots) { 875 got.relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) { 876 return got.global.count(p.first); 877 }); 878 set_union(got.local16, got.local32); 879 got.local32.clear(); 880 } 881 882 // Evaluate number of "reloc-only" entries in the resulting GOT. 883 // To do that put all unique "reloc-only" and "global" entries 884 // from all GOTs to the future primary GOT. 885 FileGot *primGot = &mergedGots.front(); 886 for (FileGot &got : gots) { 887 set_union(primGot->relocs, got.global); 888 set_union(primGot->relocs, got.relocs); 889 got.relocs.clear(); 890 } 891 892 // Evaluate number of "page" entries in each GOT. 893 for (FileGot &got : gots) { 894 for (std::pair<const OutputSection *, FileGot::PageBlock> &p : 895 got.pagesMap) { 896 const OutputSection *os = p.first; 897 uint64_t secSize = 0; 898 for (SectionCommand *cmd : os->commands) { 899 if (auto *isd = dyn_cast<InputSectionDescription>(cmd)) 900 for (InputSection *isec : isd->sections) { 901 uint64_t off = alignToPowerOf2(secSize, isec->addralign); 902 secSize = off + isec->getSize(); 903 } 904 } 905 p.second.count = getMipsPageCount(secSize); 906 } 907 } 908 909 // Merge GOTs. Try to join as much as possible GOTs but do not exceed 910 // maximum GOT size. At first, try to fill the primary GOT because 911 // the primary GOT can be accessed in the most effective way. If it 912 // is not possible, try to fill the last GOT in the list, and finally 913 // create a new GOT if both attempts failed. 914 for (FileGot &srcGot : gots) { 915 InputFile *file = srcGot.file; 916 if (tryMergeGots(mergedGots.front(), srcGot, true)) { 917 file->mipsGotIndex = 0; 918 } else { 919 // If this is the first time we failed to merge with the primary GOT, 920 // MergedGots.back() will also be the primary GOT. We must make sure not 921 // to try to merge again with isPrimary=false, as otherwise, if the 922 // inputs are just right, we could allow the primary GOT to become 1 or 2 923 // words bigger due to ignoring the header size. 924 if (mergedGots.size() == 1 || 925 !tryMergeGots(mergedGots.back(), srcGot, false)) { 926 mergedGots.emplace_back(); 927 std::swap(mergedGots.back(), srcGot); 928 } 929 file->mipsGotIndex = mergedGots.size() - 1; 930 } 931 } 932 std::swap(gots, mergedGots); 933 934 // Reduce number of "reloc-only" entries in the primary GOT 935 // by subtracting "global" entries in the primary GOT. 936 primGot = &gots.front(); 937 primGot->relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) { 938 return primGot->global.count(p.first); 939 }); 940 941 // Calculate indexes for each GOT entry. 942 size_t index = headerEntriesNum; 943 for (FileGot &got : gots) { 944 got.startIndex = &got == primGot ? 0 : index; 945 for (std::pair<const OutputSection *, FileGot::PageBlock> &p : 946 got.pagesMap) { 947 // For each output section referenced by GOT page relocations calculate 948 // and save into pagesMap an upper bound of MIPS GOT entries required 949 // to store page addresses of local symbols. We assume the worst case - 950 // each 64kb page of the output section has at least one GOT relocation 951 // against it. And take in account the case when the section intersects 952 // page boundaries. 953 p.second.firstIndex = index; 954 index += p.second.count; 955 } 956 for (auto &p: got.local16) 957 p.second = index++; 958 for (auto &p: got.global) 959 p.second = index++; 960 for (auto &p: got.relocs) 961 p.second = index++; 962 for (auto &p: got.tls) 963 p.second = index++; 964 for (auto &p: got.dynTlsSymbols) { 965 p.second = index; 966 index += 2; 967 } 968 } 969 970 // Update SymbolAux::gotIdx field to use this 971 // value later in the `sortMipsSymbols` function. 972 for (auto &p : primGot->global) { 973 if (p.first->auxIdx == 0) 974 p.first->allocateAux(); 975 symAux.back().gotIdx = p.second; 976 } 977 for (auto &p : primGot->relocs) { 978 if (p.first->auxIdx == 0) 979 p.first->allocateAux(); 980 symAux.back().gotIdx = p.second; 981 } 982 983 // Create dynamic relocations. 984 for (FileGot &got : gots) { 985 // Create dynamic relocations for TLS entries. 986 for (std::pair<Symbol *, size_t> &p : got.tls) { 987 Symbol *s = p.first; 988 uint64_t offset = p.second * config->wordsize; 989 // When building a shared library we still need a dynamic relocation 990 // for the TP-relative offset as we don't know how much other data will 991 // be allocated before us in the static TLS block. 992 if (s->isPreemptible || config->shared) 993 mainPart->relaDyn->addReloc({target->tlsGotRel, this, offset, 994 DynamicReloc::AgainstSymbolWithTargetVA, 995 *s, 0, R_ABS}); 996 } 997 for (std::pair<Symbol *, size_t> &p : got.dynTlsSymbols) { 998 Symbol *s = p.first; 999 uint64_t offset = p.second * config->wordsize; 1000 if (s == nullptr) { 1001 if (!config->shared) 1002 continue; 1003 mainPart->relaDyn->addReloc({target->tlsModuleIndexRel, this, offset}); 1004 } else { 1005 // When building a shared library we still need a dynamic relocation 1006 // for the module index. Therefore only checking for 1007 // S->isPreemptible is not sufficient (this happens e.g. for 1008 // thread-locals that have been marked as local through a linker script) 1009 if (!s->isPreemptible && !config->shared) 1010 continue; 1011 mainPart->relaDyn->addSymbolReloc(target->tlsModuleIndexRel, *this, 1012 offset, *s); 1013 // However, we can skip writing the TLS offset reloc for non-preemptible 1014 // symbols since it is known even in shared libraries 1015 if (!s->isPreemptible) 1016 continue; 1017 offset += config->wordsize; 1018 mainPart->relaDyn->addSymbolReloc(target->tlsOffsetRel, *this, offset, 1019 *s); 1020 } 1021 } 1022 1023 // Do not create dynamic relocations for non-TLS 1024 // entries in the primary GOT. 1025 if (&got == primGot) 1026 continue; 1027 1028 // Dynamic relocations for "global" entries. 1029 for (const std::pair<Symbol *, size_t> &p : got.global) { 1030 uint64_t offset = p.second * config->wordsize; 1031 mainPart->relaDyn->addSymbolReloc(target->relativeRel, *this, offset, 1032 *p.first); 1033 } 1034 if (!config->isPic) 1035 continue; 1036 // Dynamic relocations for "local" entries in case of PIC. 1037 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l : 1038 got.pagesMap) { 1039 size_t pageCount = l.second.count; 1040 for (size_t pi = 0; pi < pageCount; ++pi) { 1041 uint64_t offset = (l.second.firstIndex + pi) * config->wordsize; 1042 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, l.first, 1043 int64_t(pi * 0x10000)}); 1044 } 1045 } 1046 for (const std::pair<GotEntry, size_t> &p : got.local16) { 1047 uint64_t offset = p.second * config->wordsize; 1048 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, 1049 DynamicReloc::AddendOnlyWithTargetVA, 1050 *p.first.first, p.first.second, R_ABS}); 1051 } 1052 } 1053 } 1054 1055 bool MipsGotSection::isNeeded() const { 1056 // We add the .got section to the result for dynamic MIPS target because 1057 // its address and properties are mentioned in the .dynamic section. 1058 return !config->relocatable; 1059 } 1060 1061 uint64_t MipsGotSection::getGp(const InputFile *f) const { 1062 // For files without related GOT or files refer a primary GOT 1063 // returns "common" _gp value. For secondary GOTs calculate 1064 // individual _gp values. 1065 if (!f || f->mipsGotIndex == uint32_t(-1) || f->mipsGotIndex == 0) 1066 return ElfSym::mipsGp->getVA(0); 1067 return getVA() + gots[f->mipsGotIndex].startIndex * config->wordsize + 0x7ff0; 1068 } 1069 1070 void MipsGotSection::writeTo(uint8_t *buf) { 1071 // Set the MSB of the second GOT slot. This is not required by any 1072 // MIPS ABI documentation, though. 1073 // 1074 // There is a comment in glibc saying that "The MSB of got[1] of a 1075 // gnu object is set to identify gnu objects," and in GNU gold it 1076 // says "the second entry will be used by some runtime loaders". 1077 // But how this field is being used is unclear. 1078 // 1079 // We are not really willing to mimic other linkers behaviors 1080 // without understanding why they do that, but because all files 1081 // generated by GNU tools have this special GOT value, and because 1082 // we've been doing this for years, it is probably a safe bet to 1083 // keep doing this for now. We really need to revisit this to see 1084 // if we had to do this. 1085 writeUint(buf + config->wordsize, (uint64_t)1 << (config->wordsize * 8 - 1)); 1086 for (const FileGot &g : gots) { 1087 auto write = [&](size_t i, const Symbol *s, int64_t a) { 1088 uint64_t va = a; 1089 if (s) 1090 va = s->getVA(a); 1091 writeUint(buf + i * config->wordsize, va); 1092 }; 1093 // Write 'page address' entries to the local part of the GOT. 1094 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l : 1095 g.pagesMap) { 1096 size_t pageCount = l.second.count; 1097 uint64_t firstPageAddr = getMipsPageAddr(l.first->addr); 1098 for (size_t pi = 0; pi < pageCount; ++pi) 1099 write(l.second.firstIndex + pi, nullptr, firstPageAddr + pi * 0x10000); 1100 } 1101 // Local, global, TLS, reloc-only entries. 1102 // If TLS entry has a corresponding dynamic relocations, leave it 1103 // initialized by zero. Write down adjusted TLS symbol's values otherwise. 1104 // To calculate the adjustments use offsets for thread-local storage. 1105 // http://web.archive.org/web/20190324223224/https://www.linux-mips.org/wiki/NPTL 1106 for (const std::pair<GotEntry, size_t> &p : g.local16) 1107 write(p.second, p.first.first, p.first.second); 1108 // Write VA to the primary GOT only. For secondary GOTs that 1109 // will be done by REL32 dynamic relocations. 1110 if (&g == &gots.front()) 1111 for (const std::pair<Symbol *, size_t> &p : g.global) 1112 write(p.second, p.first, 0); 1113 for (const std::pair<Symbol *, size_t> &p : g.relocs) 1114 write(p.second, p.first, 0); 1115 for (const std::pair<Symbol *, size_t> &p : g.tls) 1116 write(p.second, p.first, 1117 p.first->isPreemptible || config->shared ? 0 : -0x7000); 1118 for (const std::pair<Symbol *, size_t> &p : g.dynTlsSymbols) { 1119 if (p.first == nullptr && !config->shared) 1120 write(p.second, nullptr, 1); 1121 else if (p.first && !p.first->isPreemptible) { 1122 // If we are emitting a shared library with relocations we mustn't write 1123 // anything to the GOT here. When using Elf_Rel relocations the value 1124 // one will be treated as an addend and will cause crashes at runtime 1125 if (!config->shared) 1126 write(p.second, nullptr, 1); 1127 write(p.second + 1, p.first, -0x8000); 1128 } 1129 } 1130 } 1131 } 1132 1133 // On PowerPC the .plt section is used to hold the table of function addresses 1134 // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss 1135 // section. I don't know why we have a BSS style type for the section but it is 1136 // consistent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI. 1137 GotPltSection::GotPltSection() 1138 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize, 1139 ".got.plt") { 1140 if (config->emachine == EM_PPC) { 1141 name = ".plt"; 1142 } else if (config->emachine == EM_PPC64) { 1143 type = SHT_NOBITS; 1144 name = ".plt"; 1145 } 1146 } 1147 1148 void GotPltSection::addEntry(Symbol &sym) { 1149 assert(sym.auxIdx == symAux.size() - 1 && 1150 symAux.back().pltIdx == entries.size()); 1151 entries.push_back(&sym); 1152 } 1153 1154 size_t GotPltSection::getSize() const { 1155 return (target->gotPltHeaderEntriesNum + entries.size()) * 1156 target->gotEntrySize; 1157 } 1158 1159 void GotPltSection::writeTo(uint8_t *buf) { 1160 target->writeGotPltHeader(buf); 1161 buf += target->gotPltHeaderEntriesNum * target->gotEntrySize; 1162 for (const Symbol *b : entries) { 1163 target->writeGotPlt(buf, *b); 1164 buf += target->gotEntrySize; 1165 } 1166 } 1167 1168 bool GotPltSection::isNeeded() const { 1169 // We need to emit GOTPLT even if it's empty if there's a relocation relative 1170 // to it. 1171 return !entries.empty() || hasGotPltOffRel; 1172 } 1173 1174 static StringRef getIgotPltName() { 1175 // On ARM the IgotPltSection is part of the GotSection. 1176 if (config->emachine == EM_ARM) 1177 return ".got"; 1178 1179 // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection 1180 // needs to be named the same. 1181 if (config->emachine == EM_PPC64) 1182 return ".plt"; 1183 1184 return ".got.plt"; 1185 } 1186 1187 // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit 1188 // with the IgotPltSection. 1189 IgotPltSection::IgotPltSection() 1190 : SyntheticSection(SHF_ALLOC | SHF_WRITE, 1191 config->emachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS, 1192 target->gotEntrySize, getIgotPltName()) {} 1193 1194 void IgotPltSection::addEntry(Symbol &sym) { 1195 assert(symAux.back().pltIdx == entries.size()); 1196 entries.push_back(&sym); 1197 } 1198 1199 size_t IgotPltSection::getSize() const { 1200 return entries.size() * target->gotEntrySize; 1201 } 1202 1203 void IgotPltSection::writeTo(uint8_t *buf) { 1204 for (const Symbol *b : entries) { 1205 target->writeIgotPlt(buf, *b); 1206 buf += target->gotEntrySize; 1207 } 1208 } 1209 1210 StringTableSection::StringTableSection(StringRef name, bool dynamic) 1211 : SyntheticSection(dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, name), 1212 dynamic(dynamic) { 1213 // ELF string tables start with a NUL byte. 1214 strings.push_back(""); 1215 stringMap.try_emplace(CachedHashStringRef(""), 0); 1216 size = 1; 1217 } 1218 1219 // Adds a string to the string table. If `hashIt` is true we hash and check for 1220 // duplicates. It is optional because the name of global symbols are already 1221 // uniqued and hashing them again has a big cost for a small value: uniquing 1222 // them with some other string that happens to be the same. 1223 unsigned StringTableSection::addString(StringRef s, bool hashIt) { 1224 if (hashIt) { 1225 auto r = stringMap.try_emplace(CachedHashStringRef(s), size); 1226 if (!r.second) 1227 return r.first->second; 1228 } 1229 if (s.empty()) 1230 return 0; 1231 unsigned ret = this->size; 1232 this->size = this->size + s.size() + 1; 1233 strings.push_back(s); 1234 return ret; 1235 } 1236 1237 void StringTableSection::writeTo(uint8_t *buf) { 1238 for (StringRef s : strings) { 1239 memcpy(buf, s.data(), s.size()); 1240 buf[s.size()] = '\0'; 1241 buf += s.size() + 1; 1242 } 1243 } 1244 1245 // Returns the number of entries in .gnu.version_d: the number of 1246 // non-VER_NDX_LOCAL-non-VER_NDX_GLOBAL definitions, plus 1. 1247 // Note that we don't support vd_cnt > 1 yet. 1248 static unsigned getVerDefNum() { 1249 return namedVersionDefs().size() + 1; 1250 } 1251 1252 template <class ELFT> 1253 DynamicSection<ELFT>::DynamicSection() 1254 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, config->wordsize, 1255 ".dynamic") { 1256 this->entsize = ELFT::Is64Bits ? 16 : 8; 1257 1258 // .dynamic section is not writable on MIPS and on Fuchsia OS 1259 // which passes -z rodynamic. 1260 // See "Special Section" in Chapter 4 in the following document: 1261 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 1262 if (config->emachine == EM_MIPS || config->zRodynamic) 1263 this->flags = SHF_ALLOC; 1264 } 1265 1266 // The output section .rela.dyn may include these synthetic sections: 1267 // 1268 // - part.relaDyn 1269 // - in.relaIplt: this is included if in.relaIplt is named .rela.dyn 1270 // - in.relaPlt: this is included if a linker script places .rela.plt inside 1271 // .rela.dyn 1272 // 1273 // DT_RELASZ is the total size of the included sections. 1274 static uint64_t addRelaSz(const RelocationBaseSection &relaDyn) { 1275 size_t size = relaDyn.getSize(); 1276 if (in.relaIplt->getParent() == relaDyn.getParent()) 1277 size += in.relaIplt->getSize(); 1278 if (in.relaPlt->getParent() == relaDyn.getParent()) 1279 size += in.relaPlt->getSize(); 1280 return size; 1281 } 1282 1283 // A Linker script may assign the RELA relocation sections to the same 1284 // output section. When this occurs we cannot just use the OutputSection 1285 // Size. Moreover the [DT_JMPREL, DT_JMPREL + DT_PLTRELSZ) is permitted to 1286 // overlap with the [DT_RELA, DT_RELA + DT_RELASZ). 1287 static uint64_t addPltRelSz() { 1288 size_t size = in.relaPlt->getSize(); 1289 if (in.relaIplt->getParent() == in.relaPlt->getParent() && 1290 in.relaIplt->name == in.relaPlt->name) 1291 size += in.relaIplt->getSize(); 1292 return size; 1293 } 1294 1295 // Add remaining entries to complete .dynamic contents. 1296 template <class ELFT> 1297 std::vector<std::pair<int32_t, uint64_t>> 1298 DynamicSection<ELFT>::computeContents() { 1299 elf::Partition &part = getPartition(); 1300 bool isMain = part.name.empty(); 1301 std::vector<std::pair<int32_t, uint64_t>> entries; 1302 1303 auto addInt = [&](int32_t tag, uint64_t val) { 1304 entries.emplace_back(tag, val); 1305 }; 1306 auto addInSec = [&](int32_t tag, const InputSection &sec) { 1307 entries.emplace_back(tag, sec.getVA()); 1308 }; 1309 1310 for (StringRef s : config->filterList) 1311 addInt(DT_FILTER, part.dynStrTab->addString(s)); 1312 for (StringRef s : config->auxiliaryList) 1313 addInt(DT_AUXILIARY, part.dynStrTab->addString(s)); 1314 1315 if (!config->rpath.empty()) 1316 addInt(config->enableNewDtags ? DT_RUNPATH : DT_RPATH, 1317 part.dynStrTab->addString(config->rpath)); 1318 1319 for (SharedFile *file : ctx.sharedFiles) 1320 if (file->isNeeded) 1321 addInt(DT_NEEDED, part.dynStrTab->addString(file->soName)); 1322 1323 if (isMain) { 1324 if (!config->soName.empty()) 1325 addInt(DT_SONAME, part.dynStrTab->addString(config->soName)); 1326 } else { 1327 if (!config->soName.empty()) 1328 addInt(DT_NEEDED, part.dynStrTab->addString(config->soName)); 1329 addInt(DT_SONAME, part.dynStrTab->addString(part.name)); 1330 } 1331 1332 // Set DT_FLAGS and DT_FLAGS_1. 1333 uint32_t dtFlags = 0; 1334 uint32_t dtFlags1 = 0; 1335 if (config->bsymbolic == BsymbolicKind::All) 1336 dtFlags |= DF_SYMBOLIC; 1337 if (config->zGlobal) 1338 dtFlags1 |= DF_1_GLOBAL; 1339 if (config->zInitfirst) 1340 dtFlags1 |= DF_1_INITFIRST; 1341 if (config->zInterpose) 1342 dtFlags1 |= DF_1_INTERPOSE; 1343 if (config->zNodefaultlib) 1344 dtFlags1 |= DF_1_NODEFLIB; 1345 if (config->zNodelete) 1346 dtFlags1 |= DF_1_NODELETE; 1347 if (config->zNodlopen) 1348 dtFlags1 |= DF_1_NOOPEN; 1349 if (config->pie) 1350 dtFlags1 |= DF_1_PIE; 1351 if (config->zNow) { 1352 dtFlags |= DF_BIND_NOW; 1353 dtFlags1 |= DF_1_NOW; 1354 } 1355 if (config->zOrigin) { 1356 dtFlags |= DF_ORIGIN; 1357 dtFlags1 |= DF_1_ORIGIN; 1358 } 1359 if (!config->zText) 1360 dtFlags |= DF_TEXTREL; 1361 if (ctx.hasTlsIe && config->shared) 1362 dtFlags |= DF_STATIC_TLS; 1363 1364 if (dtFlags) 1365 addInt(DT_FLAGS, dtFlags); 1366 if (dtFlags1) 1367 addInt(DT_FLAGS_1, dtFlags1); 1368 1369 // DT_DEBUG is a pointer to debug information used by debuggers at runtime. We 1370 // need it for each process, so we don't write it for DSOs. The loader writes 1371 // the pointer into this entry. 1372 // 1373 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some 1374 // systems (currently only Fuchsia OS) provide other means to give the 1375 // debugger this information. Such systems may choose make .dynamic read-only. 1376 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG. 1377 if (!config->shared && !config->relocatable && !config->zRodynamic) 1378 addInt(DT_DEBUG, 0); 1379 1380 if (part.relaDyn->isNeeded() || 1381 (in.relaIplt->isNeeded() && 1382 part.relaDyn->getParent() == in.relaIplt->getParent())) { 1383 addInSec(part.relaDyn->dynamicTag, *part.relaDyn); 1384 entries.emplace_back(part.relaDyn->sizeDynamicTag, 1385 addRelaSz(*part.relaDyn)); 1386 1387 bool isRela = config->isRela; 1388 addInt(isRela ? DT_RELAENT : DT_RELENT, 1389 isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel)); 1390 1391 // MIPS dynamic loader does not support RELCOUNT tag. 1392 // The problem is in the tight relation between dynamic 1393 // relocations and GOT. So do not emit this tag on MIPS. 1394 if (config->emachine != EM_MIPS) { 1395 size_t numRelativeRels = part.relaDyn->getRelativeRelocCount(); 1396 if (config->zCombreloc && numRelativeRels) 1397 addInt(isRela ? DT_RELACOUNT : DT_RELCOUNT, numRelativeRels); 1398 } 1399 } 1400 if (part.relrDyn && part.relrDyn->getParent() && 1401 !part.relrDyn->relocs.empty()) { 1402 addInSec(config->useAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR, 1403 *part.relrDyn); 1404 addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ, 1405 part.relrDyn->getParent()->size); 1406 addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT, 1407 sizeof(Elf_Relr)); 1408 } 1409 // .rel[a].plt section usually consists of two parts, containing plt and 1410 // iplt relocations. It is possible to have only iplt relocations in the 1411 // output. In that case relaPlt is empty and have zero offset, the same offset 1412 // as relaIplt has. And we still want to emit proper dynamic tags for that 1413 // case, so here we always use relaPlt as marker for the beginning of 1414 // .rel[a].plt section. 1415 if (isMain && (in.relaPlt->isNeeded() || in.relaIplt->isNeeded())) { 1416 addInSec(DT_JMPREL, *in.relaPlt); 1417 entries.emplace_back(DT_PLTRELSZ, addPltRelSz()); 1418 switch (config->emachine) { 1419 case EM_MIPS: 1420 addInSec(DT_MIPS_PLTGOT, *in.gotPlt); 1421 break; 1422 case EM_SPARCV9: 1423 addInSec(DT_PLTGOT, *in.plt); 1424 break; 1425 case EM_AARCH64: 1426 if (llvm::find_if(in.relaPlt->relocs, [](const DynamicReloc &r) { 1427 return r.type == target->pltRel && 1428 r.sym->stOther & STO_AARCH64_VARIANT_PCS; 1429 }) != in.relaPlt->relocs.end()) 1430 addInt(DT_AARCH64_VARIANT_PCS, 0); 1431 addInSec(DT_PLTGOT, *in.gotPlt); 1432 break; 1433 case EM_RISCV: 1434 if (llvm::any_of(in.relaPlt->relocs, [](const DynamicReloc &r) { 1435 return r.type == target->pltRel && 1436 (r.sym->stOther & STO_RISCV_VARIANT_CC); 1437 })) 1438 addInt(DT_RISCV_VARIANT_CC, 0); 1439 [[fallthrough]]; 1440 default: 1441 addInSec(DT_PLTGOT, *in.gotPlt); 1442 break; 1443 } 1444 addInt(DT_PLTREL, config->isRela ? DT_RELA : DT_REL); 1445 } 1446 1447 if (config->emachine == EM_AARCH64) { 1448 if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI) 1449 addInt(DT_AARCH64_BTI_PLT, 0); 1450 if (config->zPacPlt) 1451 addInt(DT_AARCH64_PAC_PLT, 0); 1452 1453 if (config->androidMemtagMode != ELF::NT_MEMTAG_LEVEL_NONE) { 1454 addInt(DT_AARCH64_MEMTAG_MODE, config->androidMemtagMode == NT_MEMTAG_LEVEL_ASYNC); 1455 addInt(DT_AARCH64_MEMTAG_HEAP, config->androidMemtagHeap); 1456 addInt(DT_AARCH64_MEMTAG_STACK, config->androidMemtagStack); 1457 if (mainPart->memtagDescriptors->isNeeded()) { 1458 addInSec(DT_AARCH64_MEMTAG_GLOBALS, *mainPart->memtagDescriptors); 1459 addInt(DT_AARCH64_MEMTAG_GLOBALSSZ, mainPart->memtagDescriptors->getSize()); 1460 } 1461 } 1462 } 1463 1464 addInSec(DT_SYMTAB, *part.dynSymTab); 1465 addInt(DT_SYMENT, sizeof(Elf_Sym)); 1466 addInSec(DT_STRTAB, *part.dynStrTab); 1467 addInt(DT_STRSZ, part.dynStrTab->getSize()); 1468 if (!config->zText) 1469 addInt(DT_TEXTREL, 0); 1470 if (part.gnuHashTab && part.gnuHashTab->getParent()) 1471 addInSec(DT_GNU_HASH, *part.gnuHashTab); 1472 if (part.hashTab && part.hashTab->getParent()) 1473 addInSec(DT_HASH, *part.hashTab); 1474 1475 if (isMain) { 1476 if (Out::preinitArray) { 1477 addInt(DT_PREINIT_ARRAY, Out::preinitArray->addr); 1478 addInt(DT_PREINIT_ARRAYSZ, Out::preinitArray->size); 1479 } 1480 if (Out::initArray) { 1481 addInt(DT_INIT_ARRAY, Out::initArray->addr); 1482 addInt(DT_INIT_ARRAYSZ, Out::initArray->size); 1483 } 1484 if (Out::finiArray) { 1485 addInt(DT_FINI_ARRAY, Out::finiArray->addr); 1486 addInt(DT_FINI_ARRAYSZ, Out::finiArray->size); 1487 } 1488 1489 if (Symbol *b = symtab.find(config->init)) 1490 if (b->isDefined()) 1491 addInt(DT_INIT, b->getVA()); 1492 if (Symbol *b = symtab.find(config->fini)) 1493 if (b->isDefined()) 1494 addInt(DT_FINI, b->getVA()); 1495 } 1496 1497 if (part.verSym && part.verSym->isNeeded()) 1498 addInSec(DT_VERSYM, *part.verSym); 1499 if (part.verDef && part.verDef->isLive()) { 1500 addInSec(DT_VERDEF, *part.verDef); 1501 addInt(DT_VERDEFNUM, getVerDefNum()); 1502 } 1503 if (part.verNeed && part.verNeed->isNeeded()) { 1504 addInSec(DT_VERNEED, *part.verNeed); 1505 unsigned needNum = 0; 1506 for (SharedFile *f : ctx.sharedFiles) 1507 if (!f->vernauxs.empty()) 1508 ++needNum; 1509 addInt(DT_VERNEEDNUM, needNum); 1510 } 1511 1512 if (config->emachine == EM_MIPS) { 1513 addInt(DT_MIPS_RLD_VERSION, 1); 1514 addInt(DT_MIPS_FLAGS, RHF_NOTPOT); 1515 addInt(DT_MIPS_BASE_ADDRESS, target->getImageBase()); 1516 addInt(DT_MIPS_SYMTABNO, part.dynSymTab->getNumSymbols()); 1517 addInt(DT_MIPS_LOCAL_GOTNO, in.mipsGot->getLocalEntriesNum()); 1518 1519 if (const Symbol *b = in.mipsGot->getFirstGlobalEntry()) 1520 addInt(DT_MIPS_GOTSYM, b->dynsymIndex); 1521 else 1522 addInt(DT_MIPS_GOTSYM, part.dynSymTab->getNumSymbols()); 1523 addInSec(DT_PLTGOT, *in.mipsGot); 1524 if (in.mipsRldMap) { 1525 if (!config->pie) 1526 addInSec(DT_MIPS_RLD_MAP, *in.mipsRldMap); 1527 // Store the offset to the .rld_map section 1528 // relative to the address of the tag. 1529 addInt(DT_MIPS_RLD_MAP_REL, 1530 in.mipsRldMap->getVA() - (getVA() + entries.size() * entsize)); 1531 } 1532 } 1533 1534 // DT_PPC_GOT indicates to glibc Secure PLT is used. If DT_PPC_GOT is absent, 1535 // glibc assumes the old-style BSS PLT layout which we don't support. 1536 if (config->emachine == EM_PPC) 1537 addInSec(DT_PPC_GOT, *in.got); 1538 1539 // Glink dynamic tag is required by the V2 abi if the plt section isn't empty. 1540 if (config->emachine == EM_PPC64 && in.plt->isNeeded()) { 1541 // The Glink tag points to 32 bytes before the first lazy symbol resolution 1542 // stub, which starts directly after the header. 1543 addInt(DT_PPC64_GLINK, in.plt->getVA() + target->pltHeaderSize - 32); 1544 } 1545 1546 if (config->emachine == EM_PPC64) 1547 addInt(DT_PPC64_OPT, getPPC64TargetInfo()->ppc64DynamicSectionOpt); 1548 1549 addInt(DT_NULL, 0); 1550 return entries; 1551 } 1552 1553 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() { 1554 if (OutputSection *sec = getPartition().dynStrTab->getParent()) 1555 getParent()->link = sec->sectionIndex; 1556 this->size = computeContents().size() * this->entsize; 1557 } 1558 1559 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *buf) { 1560 auto *p = reinterpret_cast<Elf_Dyn *>(buf); 1561 1562 for (std::pair<int32_t, uint64_t> kv : computeContents()) { 1563 p->d_tag = kv.first; 1564 p->d_un.d_val = kv.second; 1565 ++p; 1566 } 1567 } 1568 1569 uint64_t DynamicReloc::getOffset() const { 1570 return inputSec->getVA(offsetInSec); 1571 } 1572 1573 int64_t DynamicReloc::computeAddend() const { 1574 switch (kind) { 1575 case AddendOnly: 1576 assert(sym == nullptr); 1577 return addend; 1578 case AgainstSymbol: 1579 assert(sym != nullptr); 1580 return addend; 1581 case AddendOnlyWithTargetVA: 1582 case AgainstSymbolWithTargetVA: { 1583 uint64_t ca = InputSection::getRelocTargetVA(inputSec->file, type, addend, 1584 getOffset(), *sym, expr); 1585 return config->is64 ? ca : SignExtend64<32>(ca); 1586 } 1587 case MipsMultiGotPage: 1588 assert(sym == nullptr); 1589 return getMipsPageAddr(outputSec->addr) + addend; 1590 } 1591 llvm_unreachable("Unknown DynamicReloc::Kind enum"); 1592 } 1593 1594 uint32_t DynamicReloc::getSymIndex(SymbolTableBaseSection *symTab) const { 1595 if (!needsDynSymIndex()) 1596 return 0; 1597 1598 size_t index = symTab->getSymbolIndex(sym); 1599 assert((index != 0 || (type != target->gotRel && type != target->pltRel) || 1600 !mainPart->dynSymTab->getParent()) && 1601 "GOT or PLT relocation must refer to symbol in dynamic symbol table"); 1602 return index; 1603 } 1604 1605 RelocationBaseSection::RelocationBaseSection(StringRef name, uint32_t type, 1606 int32_t dynamicTag, 1607 int32_t sizeDynamicTag, 1608 bool combreloc, 1609 unsigned concurrency) 1610 : SyntheticSection(SHF_ALLOC, type, config->wordsize, name), 1611 dynamicTag(dynamicTag), sizeDynamicTag(sizeDynamicTag), 1612 relocsVec(concurrency), combreloc(combreloc) {} 1613 1614 void RelocationBaseSection::addSymbolReloc( 1615 RelType dynType, InputSectionBase &isec, uint64_t offsetInSec, Symbol &sym, 1616 int64_t addend, std::optional<RelType> addendRelType) { 1617 addReloc(DynamicReloc::AgainstSymbol, dynType, isec, offsetInSec, sym, addend, 1618 R_ADDEND, addendRelType ? *addendRelType : target->noneRel); 1619 } 1620 1621 void RelocationBaseSection::addAddendOnlyRelocIfNonPreemptible( 1622 RelType dynType, GotSection &sec, uint64_t offsetInSec, Symbol &sym, 1623 RelType addendRelType) { 1624 // No need to write an addend to the section for preemptible symbols. 1625 if (sym.isPreemptible) 1626 addReloc({dynType, &sec, offsetInSec, DynamicReloc::AgainstSymbol, sym, 0, 1627 R_ABS}); 1628 else 1629 addReloc(DynamicReloc::AddendOnlyWithTargetVA, dynType, sec, offsetInSec, 1630 sym, 0, R_ABS, addendRelType); 1631 } 1632 1633 void RelocationBaseSection::mergeRels() { 1634 size_t newSize = relocs.size(); 1635 for (const auto &v : relocsVec) 1636 newSize += v.size(); 1637 relocs.reserve(newSize); 1638 for (const auto &v : relocsVec) 1639 llvm::append_range(relocs, v); 1640 relocsVec.clear(); 1641 } 1642 1643 void RelocationBaseSection::partitionRels() { 1644 if (!combreloc) 1645 return; 1646 const RelType relativeRel = target->relativeRel; 1647 numRelativeRelocs = 1648 llvm::partition(relocs, [=](auto &r) { return r.type == relativeRel; }) - 1649 relocs.begin(); 1650 } 1651 1652 void RelocationBaseSection::finalizeContents() { 1653 SymbolTableBaseSection *symTab = getPartition().dynSymTab.get(); 1654 1655 // When linking glibc statically, .rel{,a}.plt contains R_*_IRELATIVE 1656 // relocations due to IFUNC (e.g. strcpy). sh_link will be set to 0 in that 1657 // case. 1658 if (symTab && symTab->getParent()) 1659 getParent()->link = symTab->getParent()->sectionIndex; 1660 else 1661 getParent()->link = 0; 1662 1663 if (in.relaPlt.get() == this && in.gotPlt->getParent()) { 1664 getParent()->flags |= ELF::SHF_INFO_LINK; 1665 getParent()->info = in.gotPlt->getParent()->sectionIndex; 1666 } 1667 if (in.relaIplt.get() == this && in.igotPlt->getParent()) { 1668 getParent()->flags |= ELF::SHF_INFO_LINK; 1669 getParent()->info = in.igotPlt->getParent()->sectionIndex; 1670 } 1671 } 1672 1673 void DynamicReloc::computeRaw(SymbolTableBaseSection *symtab) { 1674 r_offset = getOffset(); 1675 r_sym = getSymIndex(symtab); 1676 addend = computeAddend(); 1677 kind = AddendOnly; // Catch errors 1678 } 1679 1680 void RelocationBaseSection::computeRels() { 1681 SymbolTableBaseSection *symTab = getPartition().dynSymTab.get(); 1682 parallelForEach(relocs, 1683 [symTab](DynamicReloc &rel) { rel.computeRaw(symTab); }); 1684 // Sort by (!IsRelative,SymIndex,r_offset). DT_REL[A]COUNT requires us to 1685 // place R_*_RELATIVE first. SymIndex is to improve locality, while r_offset 1686 // is to make results easier to read. 1687 if (combreloc) { 1688 auto nonRelative = relocs.begin() + numRelativeRelocs; 1689 parallelSort(relocs.begin(), nonRelative, 1690 [&](auto &a, auto &b) { return a.r_offset < b.r_offset; }); 1691 // Non-relative relocations are few, so don't bother with parallelSort. 1692 llvm::sort(nonRelative, relocs.end(), [&](auto &a, auto &b) { 1693 return std::tie(a.r_sym, a.r_offset) < std::tie(b.r_sym, b.r_offset); 1694 }); 1695 } 1696 } 1697 1698 template <class ELFT> 1699 RelocationSection<ELFT>::RelocationSection(StringRef name, bool combreloc, 1700 unsigned concurrency) 1701 : RelocationBaseSection(name, config->isRela ? SHT_RELA : SHT_REL, 1702 config->isRela ? DT_RELA : DT_REL, 1703 config->isRela ? DT_RELASZ : DT_RELSZ, combreloc, 1704 concurrency) { 1705 this->entsize = config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1706 } 1707 1708 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *buf) { 1709 computeRels(); 1710 for (const DynamicReloc &rel : relocs) { 1711 auto *p = reinterpret_cast<Elf_Rela *>(buf); 1712 p->r_offset = rel.r_offset; 1713 p->setSymbolAndType(rel.r_sym, rel.type, config->isMips64EL); 1714 if (config->isRela) 1715 p->r_addend = rel.addend; 1716 buf += config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); 1717 } 1718 } 1719 1720 RelrBaseSection::RelrBaseSection(unsigned concurrency) 1721 : SyntheticSection(SHF_ALLOC, 1722 config->useAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR, 1723 config->wordsize, ".relr.dyn"), 1724 relocsVec(concurrency) {} 1725 1726 void RelrBaseSection::mergeRels() { 1727 size_t newSize = relocs.size(); 1728 for (const auto &v : relocsVec) 1729 newSize += v.size(); 1730 relocs.reserve(newSize); 1731 for (const auto &v : relocsVec) 1732 llvm::append_range(relocs, v); 1733 relocsVec.clear(); 1734 } 1735 1736 template <class ELFT> 1737 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection( 1738 StringRef name, unsigned concurrency) 1739 : RelocationBaseSection( 1740 name, config->isRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL, 1741 config->isRela ? DT_ANDROID_RELA : DT_ANDROID_REL, 1742 config->isRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ, 1743 /*combreloc=*/false, concurrency) { 1744 this->entsize = 1; 1745 } 1746 1747 template <class ELFT> 1748 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() { 1749 // This function computes the contents of an Android-format packed relocation 1750 // section. 1751 // 1752 // This format compresses relocations by using relocation groups to factor out 1753 // fields that are common between relocations and storing deltas from previous 1754 // relocations in SLEB128 format (which has a short representation for small 1755 // numbers). A good example of a relocation type with common fields is 1756 // R_*_RELATIVE, which is normally used to represent function pointers in 1757 // vtables. In the REL format, each relative relocation has the same r_info 1758 // field, and is only different from other relative relocations in terms of 1759 // the r_offset field. By sorting relocations by offset, grouping them by 1760 // r_info and representing each relocation with only the delta from the 1761 // previous offset, each 8-byte relocation can be compressed to as little as 1 1762 // byte (or less with run-length encoding). This relocation packer was able to 1763 // reduce the size of the relocation section in an Android Chromium DSO from 1764 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size. 1765 // 1766 // A relocation section consists of a header containing the literal bytes 1767 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two 1768 // elements are the total number of relocations in the section and an initial 1769 // r_offset value. The remaining elements define a sequence of relocation 1770 // groups. Each relocation group starts with a header consisting of the 1771 // following elements: 1772 // 1773 // - the number of relocations in the relocation group 1774 // - flags for the relocation group 1775 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta 1776 // for each relocation in the group. 1777 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info 1778 // field for each relocation in the group. 1779 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and 1780 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for 1781 // each relocation in the group. 1782 // 1783 // Following the relocation group header are descriptions of each of the 1784 // relocations in the group. They consist of the following elements: 1785 // 1786 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset 1787 // delta for this relocation. 1788 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info 1789 // field for this relocation. 1790 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and 1791 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for 1792 // this relocation. 1793 1794 size_t oldSize = relocData.size(); 1795 1796 relocData = {'A', 'P', 'S', '2'}; 1797 raw_svector_ostream os(relocData); 1798 auto add = [&](int64_t v) { encodeSLEB128(v, os); }; 1799 1800 // The format header includes the number of relocations and the initial 1801 // offset (we set this to zero because the first relocation group will 1802 // perform the initial adjustment). 1803 add(relocs.size()); 1804 add(0); 1805 1806 std::vector<Elf_Rela> relatives, nonRelatives; 1807 1808 for (const DynamicReloc &rel : relocs) { 1809 Elf_Rela r; 1810 r.r_offset = rel.getOffset(); 1811 r.setSymbolAndType(rel.getSymIndex(getPartition().dynSymTab.get()), 1812 rel.type, false); 1813 r.r_addend = config->isRela ? rel.computeAddend() : 0; 1814 1815 if (r.getType(config->isMips64EL) == target->relativeRel) 1816 relatives.push_back(r); 1817 else 1818 nonRelatives.push_back(r); 1819 } 1820 1821 llvm::sort(relatives, [](const Elf_Rel &a, const Elf_Rel &b) { 1822 return a.r_offset < b.r_offset; 1823 }); 1824 1825 // Try to find groups of relative relocations which are spaced one word 1826 // apart from one another. These generally correspond to vtable entries. The 1827 // format allows these groups to be encoded using a sort of run-length 1828 // encoding, but each group will cost 7 bytes in addition to the offset from 1829 // the previous group, so it is only profitable to do this for groups of 1830 // size 8 or larger. 1831 std::vector<Elf_Rela> ungroupedRelatives; 1832 std::vector<std::vector<Elf_Rela>> relativeGroups; 1833 for (auto i = relatives.begin(), e = relatives.end(); i != e;) { 1834 std::vector<Elf_Rela> group; 1835 do { 1836 group.push_back(*i++); 1837 } while (i != e && (i - 1)->r_offset + config->wordsize == i->r_offset); 1838 1839 if (group.size() < 8) 1840 ungroupedRelatives.insert(ungroupedRelatives.end(), group.begin(), 1841 group.end()); 1842 else 1843 relativeGroups.emplace_back(std::move(group)); 1844 } 1845 1846 // For non-relative relocations, we would like to: 1847 // 1. Have relocations with the same symbol offset to be consecutive, so 1848 // that the runtime linker can speed-up symbol lookup by implementing an 1849 // 1-entry cache. 1850 // 2. Group relocations by r_info to reduce the size of the relocation 1851 // section. 1852 // Since the symbol offset is the high bits in r_info, sorting by r_info 1853 // allows us to do both. 1854 // 1855 // For Rela, we also want to sort by r_addend when r_info is the same. This 1856 // enables us to group by r_addend as well. 1857 llvm::sort(nonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) { 1858 if (a.r_info != b.r_info) 1859 return a.r_info < b.r_info; 1860 if (a.r_addend != b.r_addend) 1861 return a.r_addend < b.r_addend; 1862 return a.r_offset < b.r_offset; 1863 }); 1864 1865 // Group relocations with the same r_info. Note that each group emits a group 1866 // header and that may make the relocation section larger. It is hard to 1867 // estimate the size of a group header as the encoded size of that varies 1868 // based on r_info. However, we can approximate this trade-off by the number 1869 // of values encoded. Each group header contains 3 values, and each relocation 1870 // in a group encodes one less value, as compared to when it is not grouped. 1871 // Therefore, we only group relocations if there are 3 or more of them with 1872 // the same r_info. 1873 // 1874 // For Rela, the addend for most non-relative relocations is zero, and thus we 1875 // can usually get a smaller relocation section if we group relocations with 0 1876 // addend as well. 1877 std::vector<Elf_Rela> ungroupedNonRelatives; 1878 std::vector<std::vector<Elf_Rela>> nonRelativeGroups; 1879 for (auto i = nonRelatives.begin(), e = nonRelatives.end(); i != e;) { 1880 auto j = i + 1; 1881 while (j != e && i->r_info == j->r_info && 1882 (!config->isRela || i->r_addend == j->r_addend)) 1883 ++j; 1884 if (j - i < 3 || (config->isRela && i->r_addend != 0)) 1885 ungroupedNonRelatives.insert(ungroupedNonRelatives.end(), i, j); 1886 else 1887 nonRelativeGroups.emplace_back(i, j); 1888 i = j; 1889 } 1890 1891 // Sort ungrouped relocations by offset to minimize the encoded length. 1892 llvm::sort(ungroupedNonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) { 1893 return a.r_offset < b.r_offset; 1894 }); 1895 1896 unsigned hasAddendIfRela = 1897 config->isRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0; 1898 1899 uint64_t offset = 0; 1900 uint64_t addend = 0; 1901 1902 // Emit the run-length encoding for the groups of adjacent relative 1903 // relocations. Each group is represented using two groups in the packed 1904 // format. The first is used to set the current offset to the start of the 1905 // group (and also encodes the first relocation), and the second encodes the 1906 // remaining relocations. 1907 for (std::vector<Elf_Rela> &g : relativeGroups) { 1908 // The first relocation in the group. 1909 add(1); 1910 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG | 1911 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela); 1912 add(g[0].r_offset - offset); 1913 add(target->relativeRel); 1914 if (config->isRela) { 1915 add(g[0].r_addend - addend); 1916 addend = g[0].r_addend; 1917 } 1918 1919 // The remaining relocations. 1920 add(g.size() - 1); 1921 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG | 1922 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela); 1923 add(config->wordsize); 1924 add(target->relativeRel); 1925 if (config->isRela) { 1926 for (const auto &i : llvm::drop_begin(g)) { 1927 add(i.r_addend - addend); 1928 addend = i.r_addend; 1929 } 1930 } 1931 1932 offset = g.back().r_offset; 1933 } 1934 1935 // Now the ungrouped relatives. 1936 if (!ungroupedRelatives.empty()) { 1937 add(ungroupedRelatives.size()); 1938 add(RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela); 1939 add(target->relativeRel); 1940 for (Elf_Rela &r : ungroupedRelatives) { 1941 add(r.r_offset - offset); 1942 offset = r.r_offset; 1943 if (config->isRela) { 1944 add(r.r_addend - addend); 1945 addend = r.r_addend; 1946 } 1947 } 1948 } 1949 1950 // Grouped non-relatives. 1951 for (ArrayRef<Elf_Rela> g : nonRelativeGroups) { 1952 add(g.size()); 1953 add(RELOCATION_GROUPED_BY_INFO_FLAG); 1954 add(g[0].r_info); 1955 for (const Elf_Rela &r : g) { 1956 add(r.r_offset - offset); 1957 offset = r.r_offset; 1958 } 1959 addend = 0; 1960 } 1961 1962 // Finally the ungrouped non-relative relocations. 1963 if (!ungroupedNonRelatives.empty()) { 1964 add(ungroupedNonRelatives.size()); 1965 add(hasAddendIfRela); 1966 for (Elf_Rela &r : ungroupedNonRelatives) { 1967 add(r.r_offset - offset); 1968 offset = r.r_offset; 1969 add(r.r_info); 1970 if (config->isRela) { 1971 add(r.r_addend - addend); 1972 addend = r.r_addend; 1973 } 1974 } 1975 } 1976 1977 // Don't allow the section to shrink; otherwise the size of the section can 1978 // oscillate infinitely. 1979 if (relocData.size() < oldSize) 1980 relocData.append(oldSize - relocData.size(), 0); 1981 1982 // Returns whether the section size changed. We need to keep recomputing both 1983 // section layout and the contents of this section until the size converges 1984 // because changing this section's size can affect section layout, which in 1985 // turn can affect the sizes of the LEB-encoded integers stored in this 1986 // section. 1987 return relocData.size() != oldSize; 1988 } 1989 1990 template <class ELFT> 1991 RelrSection<ELFT>::RelrSection(unsigned concurrency) 1992 : RelrBaseSection(concurrency) { 1993 this->entsize = config->wordsize; 1994 } 1995 1996 template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() { 1997 // This function computes the contents of an SHT_RELR packed relocation 1998 // section. 1999 // 2000 // Proposal for adding SHT_RELR sections to generic-abi is here: 2001 // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg 2002 // 2003 // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks 2004 // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ] 2005 // 2006 // i.e. start with an address, followed by any number of bitmaps. The address 2007 // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63 2008 // relocations each, at subsequent offsets following the last address entry. 2009 // 2010 // The bitmap entries must have 1 in the least significant bit. The assumption 2011 // here is that an address cannot have 1 in lsb. Odd addresses are not 2012 // supported. 2013 // 2014 // Excluding the least significant bit in the bitmap, each non-zero bit in 2015 // the bitmap represents a relocation to be applied to a corresponding machine 2016 // word that follows the base address word. The second least significant bit 2017 // represents the machine word immediately following the initial address, and 2018 // each bit that follows represents the next word, in linear order. As such, 2019 // a single bitmap can encode up to 31 relocations in a 32-bit object, and 2020 // 63 relocations in a 64-bit object. 2021 // 2022 // This encoding has a couple of interesting properties: 2023 // 1. Looking at any entry, it is clear whether it's an address or a bitmap: 2024 // even means address, odd means bitmap. 2025 // 2. Just a simple list of addresses is a valid encoding. 2026 2027 size_t oldSize = relrRelocs.size(); 2028 relrRelocs.clear(); 2029 2030 // Same as Config->Wordsize but faster because this is a compile-time 2031 // constant. 2032 const size_t wordsize = sizeof(typename ELFT::uint); 2033 2034 // Number of bits to use for the relocation offsets bitmap. 2035 // Must be either 63 or 31. 2036 const size_t nBits = wordsize * 8 - 1; 2037 2038 // Get offsets for all relative relocations and sort them. 2039 std::unique_ptr<uint64_t[]> offsets(new uint64_t[relocs.size()]); 2040 for (auto [i, r] : llvm::enumerate(relocs)) 2041 offsets[i] = r.getOffset(); 2042 llvm::sort(offsets.get(), offsets.get() + relocs.size()); 2043 2044 // For each leading relocation, find following ones that can be folded 2045 // as a bitmap and fold them. 2046 for (size_t i = 0, e = relocs.size(); i != e;) { 2047 // Add a leading relocation. 2048 relrRelocs.push_back(Elf_Relr(offsets[i])); 2049 uint64_t base = offsets[i] + wordsize; 2050 ++i; 2051 2052 // Find foldable relocations to construct bitmaps. 2053 for (;;) { 2054 uint64_t bitmap = 0; 2055 for (; i != e; ++i) { 2056 uint64_t d = offsets[i] - base; 2057 if (d >= nBits * wordsize || d % wordsize) 2058 break; 2059 bitmap |= uint64_t(1) << (d / wordsize); 2060 } 2061 if (!bitmap) 2062 break; 2063 relrRelocs.push_back(Elf_Relr((bitmap << 1) | 1)); 2064 base += nBits * wordsize; 2065 } 2066 } 2067 2068 // Don't allow the section to shrink; otherwise the size of the section can 2069 // oscillate infinitely. Trailing 1s do not decode to more relocations. 2070 if (relrRelocs.size() < oldSize) { 2071 log(".relr.dyn needs " + Twine(oldSize - relrRelocs.size()) + 2072 " padding word(s)"); 2073 relrRelocs.resize(oldSize, Elf_Relr(1)); 2074 } 2075 2076 return relrRelocs.size() != oldSize; 2077 } 2078 2079 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &strTabSec) 2080 : SyntheticSection(strTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0, 2081 strTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB, 2082 config->wordsize, 2083 strTabSec.isDynamic() ? ".dynsym" : ".symtab"), 2084 strTabSec(strTabSec) {} 2085 2086 // Orders symbols according to their positions in the GOT, 2087 // in compliance with MIPS ABI rules. 2088 // See "Global Offset Table" in Chapter 5 in the following document 2089 // for detailed description: 2090 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf 2091 static bool sortMipsSymbols(const SymbolTableEntry &l, 2092 const SymbolTableEntry &r) { 2093 // Sort entries related to non-local preemptible symbols by GOT indexes. 2094 // All other entries go to the beginning of a dynsym in arbitrary order. 2095 if (l.sym->isInGot() && r.sym->isInGot()) 2096 return l.sym->getGotIdx() < r.sym->getGotIdx(); 2097 if (!l.sym->isInGot() && !r.sym->isInGot()) 2098 return false; 2099 return !l.sym->isInGot(); 2100 } 2101 2102 void SymbolTableBaseSection::finalizeContents() { 2103 if (OutputSection *sec = strTabSec.getParent()) 2104 getParent()->link = sec->sectionIndex; 2105 2106 if (this->type != SHT_DYNSYM) { 2107 sortSymTabSymbols(); 2108 return; 2109 } 2110 2111 // If it is a .dynsym, there should be no local symbols, but we need 2112 // to do a few things for the dynamic linker. 2113 2114 // Section's Info field has the index of the first non-local symbol. 2115 // Because the first symbol entry is a null entry, 1 is the first. 2116 getParent()->info = 1; 2117 2118 if (getPartition().gnuHashTab) { 2119 // NB: It also sorts Symbols to meet the GNU hash table requirements. 2120 getPartition().gnuHashTab->addSymbols(symbols); 2121 } else if (config->emachine == EM_MIPS) { 2122 llvm::stable_sort(symbols, sortMipsSymbols); 2123 } 2124 2125 // Only the main partition's dynsym indexes are stored in the symbols 2126 // themselves. All other partitions use a lookup table. 2127 if (this == mainPart->dynSymTab.get()) { 2128 size_t i = 0; 2129 for (const SymbolTableEntry &s : symbols) 2130 s.sym->dynsymIndex = ++i; 2131 } 2132 } 2133 2134 // The ELF spec requires that all local symbols precede global symbols, so we 2135 // sort symbol entries in this function. (For .dynsym, we don't do that because 2136 // symbols for dynamic linking are inherently all globals.) 2137 // 2138 // Aside from above, we put local symbols in groups starting with the STT_FILE 2139 // symbol. That is convenient for purpose of identifying where are local symbols 2140 // coming from. 2141 void SymbolTableBaseSection::sortSymTabSymbols() { 2142 // Move all local symbols before global symbols. 2143 auto e = std::stable_partition( 2144 symbols.begin(), symbols.end(), 2145 [](const SymbolTableEntry &s) { return s.sym->isLocal(); }); 2146 size_t numLocals = e - symbols.begin(); 2147 getParent()->info = numLocals + 1; 2148 2149 // We want to group the local symbols by file. For that we rebuild the local 2150 // part of the symbols vector. We do not need to care about the STT_FILE 2151 // symbols, they are already naturally placed first in each group. That 2152 // happens because STT_FILE is always the first symbol in the object and hence 2153 // precede all other local symbols we add for a file. 2154 MapVector<InputFile *, SmallVector<SymbolTableEntry, 0>> arr; 2155 for (const SymbolTableEntry &s : llvm::make_range(symbols.begin(), e)) 2156 arr[s.sym->file].push_back(s); 2157 2158 auto i = symbols.begin(); 2159 for (auto &p : arr) 2160 for (SymbolTableEntry &entry : p.second) 2161 *i++ = entry; 2162 } 2163 2164 void SymbolTableBaseSection::addSymbol(Symbol *b) { 2165 // Adding a local symbol to a .dynsym is a bug. 2166 assert(this->type != SHT_DYNSYM || !b->isLocal()); 2167 symbols.push_back({b, strTabSec.addString(b->getName(), false)}); 2168 } 2169 2170 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *sym) { 2171 if (this == mainPart->dynSymTab.get()) 2172 return sym->dynsymIndex; 2173 2174 // Initializes symbol lookup tables lazily. This is used only for -r, 2175 // --emit-relocs and dynsyms in partitions other than the main one. 2176 llvm::call_once(onceFlag, [&] { 2177 symbolIndexMap.reserve(symbols.size()); 2178 size_t i = 0; 2179 for (const SymbolTableEntry &e : symbols) { 2180 if (e.sym->type == STT_SECTION) 2181 sectionIndexMap[e.sym->getOutputSection()] = ++i; 2182 else 2183 symbolIndexMap[e.sym] = ++i; 2184 } 2185 }); 2186 2187 // Section symbols are mapped based on their output sections 2188 // to maintain their semantics. 2189 if (sym->type == STT_SECTION) 2190 return sectionIndexMap.lookup(sym->getOutputSection()); 2191 return symbolIndexMap.lookup(sym); 2192 } 2193 2194 template <class ELFT> 2195 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &strTabSec) 2196 : SymbolTableBaseSection(strTabSec) { 2197 this->entsize = sizeof(Elf_Sym); 2198 } 2199 2200 static BssSection *getCommonSec(Symbol *sym) { 2201 if (config->relocatable) 2202 if (auto *d = dyn_cast<Defined>(sym)) 2203 return dyn_cast_or_null<BssSection>(d->section); 2204 return nullptr; 2205 } 2206 2207 static uint32_t getSymSectionIndex(Symbol *sym) { 2208 assert(!(sym->hasFlag(NEEDS_COPY) && sym->isObject())); 2209 if (!isa<Defined>(sym) || sym->hasFlag(NEEDS_COPY)) 2210 return SHN_UNDEF; 2211 if (const OutputSection *os = sym->getOutputSection()) 2212 return os->sectionIndex >= SHN_LORESERVE ? (uint32_t)SHN_XINDEX 2213 : os->sectionIndex; 2214 return SHN_ABS; 2215 } 2216 2217 // Write the internal symbol table contents to the output symbol table. 2218 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *buf) { 2219 // The first entry is a null entry as per the ELF spec. 2220 buf += sizeof(Elf_Sym); 2221 2222 auto *eSym = reinterpret_cast<Elf_Sym *>(buf); 2223 2224 for (SymbolTableEntry &ent : symbols) { 2225 Symbol *sym = ent.sym; 2226 bool isDefinedHere = type == SHT_SYMTAB || sym->partition == partition; 2227 2228 // Set st_name, st_info and st_other. 2229 eSym->st_name = ent.strTabOffset; 2230 eSym->setBindingAndType(sym->binding, sym->type); 2231 eSym->st_other = sym->stOther; 2232 2233 if (BssSection *commonSec = getCommonSec(sym)) { 2234 // When -r is specified, a COMMON symbol is not allocated. Its st_shndx 2235 // holds SHN_COMMON and st_value holds the alignment. 2236 eSym->st_shndx = SHN_COMMON; 2237 eSym->st_value = commonSec->addralign; 2238 eSym->st_size = cast<Defined>(sym)->size; 2239 } else { 2240 const uint32_t shndx = getSymSectionIndex(sym); 2241 if (isDefinedHere) { 2242 eSym->st_shndx = shndx; 2243 eSym->st_value = sym->getVA(); 2244 // Copy symbol size if it is a defined symbol. st_size is not 2245 // significant for undefined symbols, so whether copying it or not is up 2246 // to us if that's the case. We'll leave it as zero because by not 2247 // setting a value, we can get the exact same outputs for two sets of 2248 // input files that differ only in undefined symbol size in DSOs. 2249 eSym->st_size = shndx != SHN_UNDEF ? cast<Defined>(sym)->size : 0; 2250 } else { 2251 eSym->st_shndx = 0; 2252 eSym->st_value = 0; 2253 eSym->st_size = 0; 2254 } 2255 } 2256 2257 ++eSym; 2258 } 2259 2260 // On MIPS we need to mark symbol which has a PLT entry and requires 2261 // pointer equality by STO_MIPS_PLT flag. That is necessary to help 2262 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs. 2263 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt 2264 if (config->emachine == EM_MIPS) { 2265 auto *eSym = reinterpret_cast<Elf_Sym *>(buf); 2266 2267 for (SymbolTableEntry &ent : symbols) { 2268 Symbol *sym = ent.sym; 2269 if (sym->isInPlt() && sym->hasFlag(NEEDS_COPY)) 2270 eSym->st_other |= STO_MIPS_PLT; 2271 if (isMicroMips()) { 2272 // We already set the less-significant bit for symbols 2273 // marked by the `STO_MIPS_MICROMIPS` flag and for microMIPS PLT 2274 // records. That allows us to distinguish such symbols in 2275 // the `MIPS<ELFT>::relocate()` routine. Now we should 2276 // clear that bit for non-dynamic symbol table, so tools 2277 // like `objdump` will be able to deal with a correct 2278 // symbol position. 2279 if (sym->isDefined() && 2280 ((sym->stOther & STO_MIPS_MICROMIPS) || sym->hasFlag(NEEDS_COPY))) { 2281 if (!strTabSec.isDynamic()) 2282 eSym->st_value &= ~1; 2283 eSym->st_other |= STO_MIPS_MICROMIPS; 2284 } 2285 } 2286 if (config->relocatable) 2287 if (auto *d = dyn_cast<Defined>(sym)) 2288 if (isMipsPIC<ELFT>(d)) 2289 eSym->st_other |= STO_MIPS_PIC; 2290 ++eSym; 2291 } 2292 } 2293 } 2294 2295 SymtabShndxSection::SymtabShndxSection() 2296 : SyntheticSection(0, SHT_SYMTAB_SHNDX, 4, ".symtab_shndx") { 2297 this->entsize = 4; 2298 } 2299 2300 void SymtabShndxSection::writeTo(uint8_t *buf) { 2301 // We write an array of 32 bit values, where each value has 1:1 association 2302 // with an entry in .symtab. If the corresponding entry contains SHN_XINDEX, 2303 // we need to write actual index, otherwise, we must write SHN_UNDEF(0). 2304 buf += 4; // Ignore .symtab[0] entry. 2305 for (const SymbolTableEntry &entry : in.symTab->getSymbols()) { 2306 if (!getCommonSec(entry.sym) && getSymSectionIndex(entry.sym) == SHN_XINDEX) 2307 write32(buf, entry.sym->getOutputSection()->sectionIndex); 2308 buf += 4; 2309 } 2310 } 2311 2312 bool SymtabShndxSection::isNeeded() const { 2313 // SHT_SYMTAB can hold symbols with section indices values up to 2314 // SHN_LORESERVE. If we need more, we want to use extension SHT_SYMTAB_SHNDX 2315 // section. Problem is that we reveal the final section indices a bit too 2316 // late, and we do not know them here. For simplicity, we just always create 2317 // a .symtab_shndx section when the amount of output sections is huge. 2318 size_t size = 0; 2319 for (SectionCommand *cmd : script->sectionCommands) 2320 if (isa<OutputDesc>(cmd)) 2321 ++size; 2322 return size >= SHN_LORESERVE; 2323 } 2324 2325 void SymtabShndxSection::finalizeContents() { 2326 getParent()->link = in.symTab->getParent()->sectionIndex; 2327 } 2328 2329 size_t SymtabShndxSection::getSize() const { 2330 return in.symTab->getNumSymbols() * 4; 2331 } 2332 2333 // .hash and .gnu.hash sections contain on-disk hash tables that map 2334 // symbol names to their dynamic symbol table indices. Their purpose 2335 // is to help the dynamic linker resolve symbols quickly. If ELF files 2336 // don't have them, the dynamic linker has to do linear search on all 2337 // dynamic symbols, which makes programs slower. Therefore, a .hash 2338 // section is added to a DSO by default. 2339 // 2340 // The Unix semantics of resolving dynamic symbols is somewhat expensive. 2341 // Each ELF file has a list of DSOs that the ELF file depends on and a 2342 // list of dynamic symbols that need to be resolved from any of the 2343 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n) 2344 // where m is the number of DSOs and n is the number of dynamic 2345 // symbols. For modern large programs, both m and n are large. So 2346 // making each step faster by using hash tables substantially 2347 // improves time to load programs. 2348 // 2349 // (Note that this is not the only way to design the shared library. 2350 // For instance, the Windows DLL takes a different approach. On 2351 // Windows, each dynamic symbol has a name of DLL from which the symbol 2352 // has to be resolved. That makes the cost of symbol resolution O(n). 2353 // This disables some hacky techniques you can use on Unix such as 2354 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.) 2355 // 2356 // Due to historical reasons, we have two different hash tables, .hash 2357 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new 2358 // and better version of .hash. .hash is just an on-disk hash table, but 2359 // .gnu.hash has a bloom filter in addition to a hash table to skip 2360 // DSOs very quickly. If you are sure that your dynamic linker knows 2361 // about .gnu.hash, you want to specify --hash-style=gnu. Otherwise, a 2362 // safe bet is to specify --hash-style=both for backward compatibility. 2363 GnuHashTableSection::GnuHashTableSection() 2364 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, config->wordsize, ".gnu.hash") { 2365 } 2366 2367 void GnuHashTableSection::finalizeContents() { 2368 if (OutputSection *sec = getPartition().dynSymTab->getParent()) 2369 getParent()->link = sec->sectionIndex; 2370 2371 // Computes bloom filter size in word size. We want to allocate 12 2372 // bits for each symbol. It must be a power of two. 2373 if (symbols.empty()) { 2374 maskWords = 1; 2375 } else { 2376 uint64_t numBits = symbols.size() * 12; 2377 maskWords = NextPowerOf2(numBits / (config->wordsize * 8)); 2378 } 2379 2380 size = 16; // Header 2381 size += config->wordsize * maskWords; // Bloom filter 2382 size += nBuckets * 4; // Hash buckets 2383 size += symbols.size() * 4; // Hash values 2384 } 2385 2386 void GnuHashTableSection::writeTo(uint8_t *buf) { 2387 // Write a header. 2388 write32(buf, nBuckets); 2389 write32(buf + 4, getPartition().dynSymTab->getNumSymbols() - symbols.size()); 2390 write32(buf + 8, maskWords); 2391 write32(buf + 12, Shift2); 2392 buf += 16; 2393 2394 // Write the 2-bit bloom filter. 2395 const unsigned c = config->is64 ? 64 : 32; 2396 for (const Entry &sym : symbols) { 2397 // When C = 64, we choose a word with bits [6:...] and set 1 to two bits in 2398 // the word using bits [0:5] and [26:31]. 2399 size_t i = (sym.hash / c) & (maskWords - 1); 2400 uint64_t val = readUint(buf + i * config->wordsize); 2401 val |= uint64_t(1) << (sym.hash % c); 2402 val |= uint64_t(1) << ((sym.hash >> Shift2) % c); 2403 writeUint(buf + i * config->wordsize, val); 2404 } 2405 buf += config->wordsize * maskWords; 2406 2407 // Write the hash table. 2408 uint32_t *buckets = reinterpret_cast<uint32_t *>(buf); 2409 uint32_t oldBucket = -1; 2410 uint32_t *values = buckets + nBuckets; 2411 for (auto i = symbols.begin(), e = symbols.end(); i != e; ++i) { 2412 // Write a hash value. It represents a sequence of chains that share the 2413 // same hash modulo value. The last element of each chain is terminated by 2414 // LSB 1. 2415 uint32_t hash = i->hash; 2416 bool isLastInChain = (i + 1) == e || i->bucketIdx != (i + 1)->bucketIdx; 2417 hash = isLastInChain ? hash | 1 : hash & ~1; 2418 write32(values++, hash); 2419 2420 if (i->bucketIdx == oldBucket) 2421 continue; 2422 // Write a hash bucket. Hash buckets contain indices in the following hash 2423 // value table. 2424 write32(buckets + i->bucketIdx, 2425 getPartition().dynSymTab->getSymbolIndex(i->sym)); 2426 oldBucket = i->bucketIdx; 2427 } 2428 } 2429 2430 // Add symbols to this symbol hash table. Note that this function 2431 // destructively sort a given vector -- which is needed because 2432 // GNU-style hash table places some sorting requirements. 2433 void GnuHashTableSection::addSymbols(SmallVectorImpl<SymbolTableEntry> &v) { 2434 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce 2435 // its type correctly. 2436 auto mid = 2437 std::stable_partition(v.begin(), v.end(), [&](const SymbolTableEntry &s) { 2438 return !s.sym->isDefined() || s.sym->partition != partition; 2439 }); 2440 2441 // We chose load factor 4 for the on-disk hash table. For each hash 2442 // collision, the dynamic linker will compare a uint32_t hash value. 2443 // Since the integer comparison is quite fast, we believe we can 2444 // make the load factor even larger. 4 is just a conservative choice. 2445 // 2446 // Note that we don't want to create a zero-sized hash table because 2447 // Android loader as of 2018 doesn't like a .gnu.hash containing such 2448 // table. If that's the case, we create a hash table with one unused 2449 // dummy slot. 2450 nBuckets = std::max<size_t>((v.end() - mid) / 4, 1); 2451 2452 if (mid == v.end()) 2453 return; 2454 2455 for (SymbolTableEntry &ent : llvm::make_range(mid, v.end())) { 2456 Symbol *b = ent.sym; 2457 uint32_t hash = hashGnu(b->getName()); 2458 uint32_t bucketIdx = hash % nBuckets; 2459 symbols.push_back({b, ent.strTabOffset, hash, bucketIdx}); 2460 } 2461 2462 llvm::sort(symbols, [](const Entry &l, const Entry &r) { 2463 return std::tie(l.bucketIdx, l.strTabOffset) < 2464 std::tie(r.bucketIdx, r.strTabOffset); 2465 }); 2466 2467 v.erase(mid, v.end()); 2468 for (const Entry &ent : symbols) 2469 v.push_back({ent.sym, ent.strTabOffset}); 2470 } 2471 2472 HashTableSection::HashTableSection() 2473 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") { 2474 this->entsize = 4; 2475 } 2476 2477 void HashTableSection::finalizeContents() { 2478 SymbolTableBaseSection *symTab = getPartition().dynSymTab.get(); 2479 2480 if (OutputSection *sec = symTab->getParent()) 2481 getParent()->link = sec->sectionIndex; 2482 2483 unsigned numEntries = 2; // nbucket and nchain. 2484 numEntries += symTab->getNumSymbols(); // The chain entries. 2485 2486 // Create as many buckets as there are symbols. 2487 numEntries += symTab->getNumSymbols(); 2488 this->size = numEntries * 4; 2489 } 2490 2491 void HashTableSection::writeTo(uint8_t *buf) { 2492 SymbolTableBaseSection *symTab = getPartition().dynSymTab.get(); 2493 unsigned numSymbols = symTab->getNumSymbols(); 2494 2495 uint32_t *p = reinterpret_cast<uint32_t *>(buf); 2496 write32(p++, numSymbols); // nbucket 2497 write32(p++, numSymbols); // nchain 2498 2499 uint32_t *buckets = p; 2500 uint32_t *chains = p + numSymbols; 2501 2502 for (const SymbolTableEntry &s : symTab->getSymbols()) { 2503 Symbol *sym = s.sym; 2504 StringRef name = sym->getName(); 2505 unsigned i = sym->dynsymIndex; 2506 uint32_t hash = hashSysV(name) % numSymbols; 2507 chains[i] = buckets[hash]; 2508 write32(buckets + hash, i); 2509 } 2510 } 2511 2512 PltSection::PltSection() 2513 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt"), 2514 headerSize(target->pltHeaderSize) { 2515 // On PowerPC, this section contains lazy symbol resolvers. 2516 if (config->emachine == EM_PPC64) { 2517 name = ".glink"; 2518 addralign = 4; 2519 } 2520 2521 // On x86 when IBT is enabled, this section contains the second PLT (lazy 2522 // symbol resolvers). 2523 if ((config->emachine == EM_386 || config->emachine == EM_X86_64) && 2524 (config->andFeatures & GNU_PROPERTY_X86_FEATURE_1_IBT)) 2525 name = ".plt.sec"; 2526 2527 // The PLT needs to be writable on SPARC as the dynamic linker will 2528 // modify the instructions in the PLT entries. 2529 if (config->emachine == EM_SPARCV9) 2530 this->flags |= SHF_WRITE; 2531 } 2532 2533 void PltSection::writeTo(uint8_t *buf) { 2534 // At beginning of PLT, we have code to call the dynamic 2535 // linker to resolve dynsyms at runtime. Write such code. 2536 target->writePltHeader(buf); 2537 size_t off = headerSize; 2538 2539 for (const Symbol *sym : entries) { 2540 target->writePlt(buf + off, *sym, getVA() + off); 2541 off += target->pltEntrySize; 2542 } 2543 } 2544 2545 void PltSection::addEntry(Symbol &sym) { 2546 assert(sym.auxIdx == symAux.size() - 1); 2547 symAux.back().pltIdx = entries.size(); 2548 entries.push_back(&sym); 2549 } 2550 2551 size_t PltSection::getSize() const { 2552 return headerSize + entries.size() * target->pltEntrySize; 2553 } 2554 2555 bool PltSection::isNeeded() const { 2556 // For -z retpolineplt, .iplt needs the .plt header. 2557 return !entries.empty() || (config->zRetpolineplt && in.iplt->isNeeded()); 2558 } 2559 2560 // Used by ARM to add mapping symbols in the PLT section, which aid 2561 // disassembly. 2562 void PltSection::addSymbols() { 2563 target->addPltHeaderSymbols(*this); 2564 2565 size_t off = headerSize; 2566 for (size_t i = 0; i < entries.size(); ++i) { 2567 target->addPltSymbols(*this, off); 2568 off += target->pltEntrySize; 2569 } 2570 } 2571 2572 IpltSection::IpltSection() 2573 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".iplt") { 2574 if (config->emachine == EM_PPC || config->emachine == EM_PPC64) { 2575 name = ".glink"; 2576 addralign = 4; 2577 } 2578 } 2579 2580 void IpltSection::writeTo(uint8_t *buf) { 2581 uint32_t off = 0; 2582 for (const Symbol *sym : entries) { 2583 target->writeIplt(buf + off, *sym, getVA() + off); 2584 off += target->ipltEntrySize; 2585 } 2586 } 2587 2588 size_t IpltSection::getSize() const { 2589 return entries.size() * target->ipltEntrySize; 2590 } 2591 2592 void IpltSection::addEntry(Symbol &sym) { 2593 assert(sym.auxIdx == symAux.size() - 1); 2594 symAux.back().pltIdx = entries.size(); 2595 entries.push_back(&sym); 2596 } 2597 2598 // ARM uses mapping symbols to aid disassembly. 2599 void IpltSection::addSymbols() { 2600 size_t off = 0; 2601 for (size_t i = 0, e = entries.size(); i != e; ++i) { 2602 target->addPltSymbols(*this, off); 2603 off += target->pltEntrySize; 2604 } 2605 } 2606 2607 PPC32GlinkSection::PPC32GlinkSection() { 2608 name = ".glink"; 2609 addralign = 4; 2610 } 2611 2612 void PPC32GlinkSection::writeTo(uint8_t *buf) { 2613 writePPC32GlinkSection(buf, entries.size()); 2614 } 2615 2616 size_t PPC32GlinkSection::getSize() const { 2617 return headerSize + entries.size() * target->pltEntrySize + footerSize; 2618 } 2619 2620 // This is an x86-only extra PLT section and used only when a security 2621 // enhancement feature called CET is enabled. In this comment, I'll explain what 2622 // the feature is and why we have two PLT sections if CET is enabled. 2623 // 2624 // So, what does CET do? CET introduces a new restriction to indirect jump 2625 // instructions. CET works this way. Assume that CET is enabled. Then, if you 2626 // execute an indirect jump instruction, the processor verifies that a special 2627 // "landing pad" instruction (which is actually a repurposed NOP instruction and 2628 // now called "endbr32" or "endbr64") is at the jump target. If the jump target 2629 // does not start with that instruction, the processor raises an exception 2630 // instead of continuing executing code. 2631 // 2632 // If CET is enabled, the compiler emits endbr to all locations where indirect 2633 // jumps may jump to. 2634 // 2635 // This mechanism makes it extremely hard to transfer the control to a middle of 2636 // a function that is not supporsed to be a indirect jump target, preventing 2637 // certain types of attacks such as ROP or JOP. 2638 // 2639 // Note that the processors in the market as of 2019 don't actually support the 2640 // feature. Only the spec is available at the moment. 2641 // 2642 // Now, I'll explain why we have this extra PLT section for CET. 2643 // 2644 // Since you can indirectly jump to a PLT entry, we have to make PLT entries 2645 // start with endbr. The problem is there's no extra space for endbr (which is 4 2646 // bytes long), as the PLT entry is only 16 bytes long and all bytes are already 2647 // used. 2648 // 2649 // In order to deal with the issue, we split a PLT entry into two PLT entries. 2650 // Remember that each PLT entry contains code to jump to an address read from 2651 // .got.plt AND code to resolve a dynamic symbol lazily. With the 2-PLT scheme, 2652 // the former code is written to .plt.sec, and the latter code is written to 2653 // .plt. 2654 // 2655 // Lazy symbol resolution in the 2-PLT scheme works in the usual way, except 2656 // that the regular .plt is now called .plt.sec and .plt is repurposed to 2657 // contain only code for lazy symbol resolution. 2658 // 2659 // In other words, this is how the 2-PLT scheme works. Application code is 2660 // supposed to jump to .plt.sec to call an external function. Each .plt.sec 2661 // entry contains code to read an address from a corresponding .got.plt entry 2662 // and jump to that address. Addresses in .got.plt initially point to .plt, so 2663 // when an application calls an external function for the first time, the 2664 // control is transferred to a function that resolves a symbol name from 2665 // external shared object files. That function then rewrites a .got.plt entry 2666 // with a resolved address, so that the subsequent function calls directly jump 2667 // to a desired location from .plt.sec. 2668 // 2669 // There is an open question as to whether the 2-PLT scheme was desirable or 2670 // not. We could have simply extended the PLT entry size to 32-bytes to 2671 // accommodate endbr, and that scheme would have been much simpler than the 2672 // 2-PLT scheme. One reason to split PLT was, by doing that, we could keep hot 2673 // code (.plt.sec) from cold code (.plt). But as far as I know no one proved 2674 // that the optimization actually makes a difference. 2675 // 2676 // That said, the 2-PLT scheme is a part of the ABI, debuggers and other tools 2677 // depend on it, so we implement the ABI. 2678 IBTPltSection::IBTPltSection() 2679 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt") {} 2680 2681 void IBTPltSection::writeTo(uint8_t *buf) { 2682 target->writeIBTPlt(buf, in.plt->getNumEntries()); 2683 } 2684 2685 size_t IBTPltSection::getSize() const { 2686 // 16 is the header size of .plt. 2687 return 16 + in.plt->getNumEntries() * target->pltEntrySize; 2688 } 2689 2690 bool IBTPltSection::isNeeded() const { return in.plt->getNumEntries() > 0; } 2691 2692 RelroPaddingSection::RelroPaddingSection() 2693 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, 1, ".relro_padding") { 2694 } 2695 2696 // The string hash function for .gdb_index. 2697 static uint32_t computeGdbHash(StringRef s) { 2698 uint32_t h = 0; 2699 for (uint8_t c : s) 2700 h = h * 67 + toLower(c) - 113; 2701 return h; 2702 } 2703 2704 GdbIndexSection::GdbIndexSection() 2705 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index") {} 2706 2707 // Returns the desired size of an on-disk hash table for a .gdb_index section. 2708 // There's a tradeoff between size and collision rate. We aim 75% utilization. 2709 size_t GdbIndexSection::computeSymtabSize() const { 2710 return std::max<size_t>(NextPowerOf2(symbols.size() * 4 / 3), 1024); 2711 } 2712 2713 static SmallVector<GdbIndexSection::CuEntry, 0> 2714 readCuList(DWARFContext &dwarf) { 2715 SmallVector<GdbIndexSection::CuEntry, 0> ret; 2716 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) 2717 ret.push_back({cu->getOffset(), cu->getLength() + 4}); 2718 return ret; 2719 } 2720 2721 static SmallVector<GdbIndexSection::AddressEntry, 0> 2722 readAddressAreas(DWARFContext &dwarf, InputSection *sec) { 2723 SmallVector<GdbIndexSection::AddressEntry, 0> ret; 2724 2725 uint32_t cuIdx = 0; 2726 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) { 2727 if (Error e = cu->tryExtractDIEsIfNeeded(false)) { 2728 warn(toString(sec) + ": " + toString(std::move(e))); 2729 return {}; 2730 } 2731 Expected<DWARFAddressRangesVector> ranges = cu->collectAddressRanges(); 2732 if (!ranges) { 2733 warn(toString(sec) + ": " + toString(ranges.takeError())); 2734 return {}; 2735 } 2736 2737 ArrayRef<InputSectionBase *> sections = sec->file->getSections(); 2738 for (DWARFAddressRange &r : *ranges) { 2739 if (r.SectionIndex == -1ULL) 2740 continue; 2741 // Range list with zero size has no effect. 2742 InputSectionBase *s = sections[r.SectionIndex]; 2743 if (s && s != &InputSection::discarded && s->isLive()) 2744 if (r.LowPC != r.HighPC) 2745 ret.push_back({cast<InputSection>(s), r.LowPC, r.HighPC, cuIdx}); 2746 } 2747 ++cuIdx; 2748 } 2749 2750 return ret; 2751 } 2752 2753 template <class ELFT> 2754 static SmallVector<GdbIndexSection::NameAttrEntry, 0> 2755 readPubNamesAndTypes(const LLDDwarfObj<ELFT> &obj, 2756 const SmallVectorImpl<GdbIndexSection::CuEntry> &cus) { 2757 const LLDDWARFSection &pubNames = obj.getGnuPubnamesSection(); 2758 const LLDDWARFSection &pubTypes = obj.getGnuPubtypesSection(); 2759 2760 SmallVector<GdbIndexSection::NameAttrEntry, 0> ret; 2761 for (const LLDDWARFSection *pub : {&pubNames, &pubTypes}) { 2762 DWARFDataExtractor data(obj, *pub, config->isLE, config->wordsize); 2763 DWARFDebugPubTable table; 2764 table.extract(data, /*GnuStyle=*/true, [&](Error e) { 2765 warn(toString(pub->sec) + ": " + toString(std::move(e))); 2766 }); 2767 for (const DWARFDebugPubTable::Set &set : table.getData()) { 2768 // The value written into the constant pool is kind << 24 | cuIndex. As we 2769 // don't know how many compilation units precede this object to compute 2770 // cuIndex, we compute (kind << 24 | cuIndexInThisObject) instead, and add 2771 // the number of preceding compilation units later. 2772 uint32_t i = llvm::partition_point(cus, 2773 [&](GdbIndexSection::CuEntry cu) { 2774 return cu.cuOffset < set.Offset; 2775 }) - 2776 cus.begin(); 2777 for (const DWARFDebugPubTable::Entry &ent : set.Entries) 2778 ret.push_back({{ent.Name, computeGdbHash(ent.Name)}, 2779 (ent.Descriptor.toBits() << 24) | i}); 2780 } 2781 } 2782 return ret; 2783 } 2784 2785 // Create a list of symbols from a given list of symbol names and types 2786 // by uniquifying them by name. 2787 static std::pair<SmallVector<GdbIndexSection::GdbSymbol, 0>, size_t> 2788 createSymbols( 2789 ArrayRef<SmallVector<GdbIndexSection::NameAttrEntry, 0>> nameAttrs, 2790 const SmallVector<GdbIndexSection::GdbChunk, 0> &chunks) { 2791 using GdbSymbol = GdbIndexSection::GdbSymbol; 2792 using NameAttrEntry = GdbIndexSection::NameAttrEntry; 2793 2794 // For each chunk, compute the number of compilation units preceding it. 2795 uint32_t cuIdx = 0; 2796 std::unique_ptr<uint32_t[]> cuIdxs(new uint32_t[chunks.size()]); 2797 for (uint32_t i = 0, e = chunks.size(); i != e; ++i) { 2798 cuIdxs[i] = cuIdx; 2799 cuIdx += chunks[i].compilationUnits.size(); 2800 } 2801 2802 // The number of symbols we will handle in this function is of the order 2803 // of millions for very large executables, so we use multi-threading to 2804 // speed it up. 2805 constexpr size_t numShards = 32; 2806 const size_t concurrency = 2807 llvm::bit_floor(std::min<size_t>(config->threadCount, numShards)); 2808 2809 // A sharded map to uniquify symbols by name. 2810 auto map = 2811 std::make_unique<DenseMap<CachedHashStringRef, size_t>[]>(numShards); 2812 size_t shift = 32 - llvm::countr_zero(numShards); 2813 2814 // Instantiate GdbSymbols while uniqufying them by name. 2815 auto symbols = std::make_unique<SmallVector<GdbSymbol, 0>[]>(numShards); 2816 2817 parallelFor(0, concurrency, [&](size_t threadId) { 2818 uint32_t i = 0; 2819 for (ArrayRef<NameAttrEntry> entries : nameAttrs) { 2820 for (const NameAttrEntry &ent : entries) { 2821 size_t shardId = ent.name.hash() >> shift; 2822 if ((shardId & (concurrency - 1)) != threadId) 2823 continue; 2824 2825 uint32_t v = ent.cuIndexAndAttrs + cuIdxs[i]; 2826 size_t &idx = map[shardId][ent.name]; 2827 if (idx) { 2828 symbols[shardId][idx - 1].cuVector.push_back(v); 2829 continue; 2830 } 2831 2832 idx = symbols[shardId].size() + 1; 2833 symbols[shardId].push_back({ent.name, {v}, 0, 0}); 2834 } 2835 ++i; 2836 } 2837 }); 2838 2839 size_t numSymbols = 0; 2840 for (ArrayRef<GdbSymbol> v : ArrayRef(symbols.get(), numShards)) 2841 numSymbols += v.size(); 2842 2843 // The return type is a flattened vector, so we'll copy each vector 2844 // contents to Ret. 2845 SmallVector<GdbSymbol, 0> ret; 2846 ret.reserve(numSymbols); 2847 for (SmallVector<GdbSymbol, 0> &vec : 2848 MutableArrayRef(symbols.get(), numShards)) 2849 for (GdbSymbol &sym : vec) 2850 ret.push_back(std::move(sym)); 2851 2852 // CU vectors and symbol names are adjacent in the output file. 2853 // We can compute their offsets in the output file now. 2854 size_t off = 0; 2855 for (GdbSymbol &sym : ret) { 2856 sym.cuVectorOff = off; 2857 off += (sym.cuVector.size() + 1) * 4; 2858 } 2859 for (GdbSymbol &sym : ret) { 2860 sym.nameOff = off; 2861 off += sym.name.size() + 1; 2862 } 2863 // If off overflows, the last symbol's nameOff likely overflows. 2864 if (!isUInt<32>(off)) 2865 errorOrWarn("--gdb-index: constant pool size (" + Twine(off) + 2866 ") exceeds UINT32_MAX"); 2867 2868 return {ret, off}; 2869 } 2870 2871 // Returns a newly-created .gdb_index section. 2872 template <class ELFT> GdbIndexSection *GdbIndexSection::create() { 2873 llvm::TimeTraceScope timeScope("Create gdb index"); 2874 2875 // Collect InputFiles with .debug_info. See the comment in 2876 // LLDDwarfObj<ELFT>::LLDDwarfObj. If we do lightweight parsing in the future, 2877 // note that isec->data() may uncompress the full content, which should be 2878 // parallelized. 2879 SetVector<InputFile *> files; 2880 for (InputSectionBase *s : ctx.inputSections) { 2881 InputSection *isec = dyn_cast<InputSection>(s); 2882 if (!isec) 2883 continue; 2884 // .debug_gnu_pub{names,types} are useless in executables. 2885 // They are present in input object files solely for creating 2886 // a .gdb_index. So we can remove them from the output. 2887 if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes") 2888 s->markDead(); 2889 else if (isec->name == ".debug_info") 2890 files.insert(isec->file); 2891 } 2892 // Drop .rel[a].debug_gnu_pub{names,types} for --emit-relocs. 2893 llvm::erase_if(ctx.inputSections, [](InputSectionBase *s) { 2894 if (auto *isec = dyn_cast<InputSection>(s)) 2895 if (InputSectionBase *rel = isec->getRelocatedSection()) 2896 return !rel->isLive(); 2897 return !s->isLive(); 2898 }); 2899 2900 SmallVector<GdbChunk, 0> chunks(files.size()); 2901 SmallVector<SmallVector<NameAttrEntry, 0>, 0> nameAttrs(files.size()); 2902 2903 parallelFor(0, files.size(), [&](size_t i) { 2904 // To keep memory usage low, we don't want to keep cached DWARFContext, so 2905 // avoid getDwarf() here. 2906 ObjFile<ELFT> *file = cast<ObjFile<ELFT>>(files[i]); 2907 DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file)); 2908 auto &dobj = static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj()); 2909 2910 // If the are multiple compile units .debug_info (very rare ld -r --unique), 2911 // this only picks the last one. Other address ranges are lost. 2912 chunks[i].sec = dobj.getInfoSection(); 2913 chunks[i].compilationUnits = readCuList(dwarf); 2914 chunks[i].addressAreas = readAddressAreas(dwarf, chunks[i].sec); 2915 nameAttrs[i] = readPubNamesAndTypes<ELFT>(dobj, chunks[i].compilationUnits); 2916 }); 2917 2918 auto *ret = make<GdbIndexSection>(); 2919 ret->chunks = std::move(chunks); 2920 std::tie(ret->symbols, ret->size) = createSymbols(nameAttrs, ret->chunks); 2921 2922 // Count the areas other than the constant pool. 2923 ret->size += sizeof(GdbIndexHeader) + ret->computeSymtabSize() * 8; 2924 for (GdbChunk &chunk : ret->chunks) 2925 ret->size += 2926 chunk.compilationUnits.size() * 16 + chunk.addressAreas.size() * 20; 2927 2928 return ret; 2929 } 2930 2931 void GdbIndexSection::writeTo(uint8_t *buf) { 2932 // Write the header. 2933 auto *hdr = reinterpret_cast<GdbIndexHeader *>(buf); 2934 uint8_t *start = buf; 2935 hdr->version = 7; 2936 buf += sizeof(*hdr); 2937 2938 // Write the CU list. 2939 hdr->cuListOff = buf - start; 2940 for (GdbChunk &chunk : chunks) { 2941 for (CuEntry &cu : chunk.compilationUnits) { 2942 write64le(buf, chunk.sec->outSecOff + cu.cuOffset); 2943 write64le(buf + 8, cu.cuLength); 2944 buf += 16; 2945 } 2946 } 2947 2948 // Write the address area. 2949 hdr->cuTypesOff = buf - start; 2950 hdr->addressAreaOff = buf - start; 2951 uint32_t cuOff = 0; 2952 for (GdbChunk &chunk : chunks) { 2953 for (AddressEntry &e : chunk.addressAreas) { 2954 // In the case of ICF there may be duplicate address range entries. 2955 const uint64_t baseAddr = e.section->repl->getVA(0); 2956 write64le(buf, baseAddr + e.lowAddress); 2957 write64le(buf + 8, baseAddr + e.highAddress); 2958 write32le(buf + 16, e.cuIndex + cuOff); 2959 buf += 20; 2960 } 2961 cuOff += chunk.compilationUnits.size(); 2962 } 2963 2964 // Write the on-disk open-addressing hash table containing symbols. 2965 hdr->symtabOff = buf - start; 2966 size_t symtabSize = computeSymtabSize(); 2967 uint32_t mask = symtabSize - 1; 2968 2969 for (GdbSymbol &sym : symbols) { 2970 uint32_t h = sym.name.hash(); 2971 uint32_t i = h & mask; 2972 uint32_t step = ((h * 17) & mask) | 1; 2973 2974 while (read32le(buf + i * 8)) 2975 i = (i + step) & mask; 2976 2977 write32le(buf + i * 8, sym.nameOff); 2978 write32le(buf + i * 8 + 4, sym.cuVectorOff); 2979 } 2980 2981 buf += symtabSize * 8; 2982 2983 // Write the string pool. 2984 hdr->constantPoolOff = buf - start; 2985 parallelForEach(symbols, [&](GdbSymbol &sym) { 2986 memcpy(buf + sym.nameOff, sym.name.data(), sym.name.size()); 2987 }); 2988 2989 // Write the CU vectors. 2990 for (GdbSymbol &sym : symbols) { 2991 write32le(buf, sym.cuVector.size()); 2992 buf += 4; 2993 for (uint32_t val : sym.cuVector) { 2994 write32le(buf, val); 2995 buf += 4; 2996 } 2997 } 2998 } 2999 3000 bool GdbIndexSection::isNeeded() const { return !chunks.empty(); } 3001 3002 EhFrameHeader::EhFrameHeader() 3003 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr") {} 3004 3005 void EhFrameHeader::writeTo(uint8_t *buf) { 3006 // Unlike most sections, the EhFrameHeader section is written while writing 3007 // another section, namely EhFrameSection, which calls the write() function 3008 // below from its writeTo() function. This is necessary because the contents 3009 // of EhFrameHeader depend on the relocated contents of EhFrameSection and we 3010 // don't know which order the sections will be written in. 3011 } 3012 3013 // .eh_frame_hdr contains a binary search table of pointers to FDEs. 3014 // Each entry of the search table consists of two values, 3015 // the starting PC from where FDEs covers, and the FDE's address. 3016 // It is sorted by PC. 3017 void EhFrameHeader::write() { 3018 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff; 3019 using FdeData = EhFrameSection::FdeData; 3020 SmallVector<FdeData, 0> fdes = getPartition().ehFrame->getFdeData(); 3021 3022 buf[0] = 1; 3023 buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4; 3024 buf[2] = DW_EH_PE_udata4; 3025 buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4; 3026 write32(buf + 4, 3027 getPartition().ehFrame->getParent()->addr - this->getVA() - 4); 3028 write32(buf + 8, fdes.size()); 3029 buf += 12; 3030 3031 for (FdeData &fde : fdes) { 3032 write32(buf, fde.pcRel); 3033 write32(buf + 4, fde.fdeVARel); 3034 buf += 8; 3035 } 3036 } 3037 3038 size_t EhFrameHeader::getSize() const { 3039 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs. 3040 return 12 + getPartition().ehFrame->numFdes * 8; 3041 } 3042 3043 bool EhFrameHeader::isNeeded() const { 3044 return isLive() && getPartition().ehFrame->isNeeded(); 3045 } 3046 3047 VersionDefinitionSection::VersionDefinitionSection() 3048 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t), 3049 ".gnu.version_d") {} 3050 3051 StringRef VersionDefinitionSection::getFileDefName() { 3052 if (!getPartition().name.empty()) 3053 return getPartition().name; 3054 if (!config->soName.empty()) 3055 return config->soName; 3056 return config->outputFile; 3057 } 3058 3059 void VersionDefinitionSection::finalizeContents() { 3060 fileDefNameOff = getPartition().dynStrTab->addString(getFileDefName()); 3061 for (const VersionDefinition &v : namedVersionDefs()) 3062 verDefNameOffs.push_back(getPartition().dynStrTab->addString(v.name)); 3063 3064 if (OutputSection *sec = getPartition().dynStrTab->getParent()) 3065 getParent()->link = sec->sectionIndex; 3066 3067 // sh_info should be set to the number of definitions. This fact is missed in 3068 // documentation, but confirmed by binutils community: 3069 // https://sourceware.org/ml/binutils/2014-11/msg00355.html 3070 getParent()->info = getVerDefNum(); 3071 } 3072 3073 void VersionDefinitionSection::writeOne(uint8_t *buf, uint32_t index, 3074 StringRef name, size_t nameOff) { 3075 uint16_t flags = index == 1 ? VER_FLG_BASE : 0; 3076 3077 // Write a verdef. 3078 write16(buf, 1); // vd_version 3079 write16(buf + 2, flags); // vd_flags 3080 write16(buf + 4, index); // vd_ndx 3081 write16(buf + 6, 1); // vd_cnt 3082 write32(buf + 8, hashSysV(name)); // vd_hash 3083 write32(buf + 12, 20); // vd_aux 3084 write32(buf + 16, 28); // vd_next 3085 3086 // Write a veraux. 3087 write32(buf + 20, nameOff); // vda_name 3088 write32(buf + 24, 0); // vda_next 3089 } 3090 3091 void VersionDefinitionSection::writeTo(uint8_t *buf) { 3092 writeOne(buf, 1, getFileDefName(), fileDefNameOff); 3093 3094 auto nameOffIt = verDefNameOffs.begin(); 3095 for (const VersionDefinition &v : namedVersionDefs()) { 3096 buf += EntrySize; 3097 writeOne(buf, v.id, v.name, *nameOffIt++); 3098 } 3099 3100 // Need to terminate the last version definition. 3101 write32(buf + 16, 0); // vd_next 3102 } 3103 3104 size_t VersionDefinitionSection::getSize() const { 3105 return EntrySize * getVerDefNum(); 3106 } 3107 3108 // .gnu.version is a table where each entry is 2 byte long. 3109 VersionTableSection::VersionTableSection() 3110 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t), 3111 ".gnu.version") { 3112 this->entsize = 2; 3113 } 3114 3115 void VersionTableSection::finalizeContents() { 3116 // At the moment of june 2016 GNU docs does not mention that sh_link field 3117 // should be set, but Sun docs do. Also readelf relies on this field. 3118 getParent()->link = getPartition().dynSymTab->getParent()->sectionIndex; 3119 } 3120 3121 size_t VersionTableSection::getSize() const { 3122 return (getPartition().dynSymTab->getSymbols().size() + 1) * 2; 3123 } 3124 3125 void VersionTableSection::writeTo(uint8_t *buf) { 3126 buf += 2; 3127 for (const SymbolTableEntry &s : getPartition().dynSymTab->getSymbols()) { 3128 // For an unextracted lazy symbol (undefined weak), it must have been 3129 // converted to Undefined and have VER_NDX_GLOBAL version here. 3130 assert(!s.sym->isLazy()); 3131 write16(buf, s.sym->versionId); 3132 buf += 2; 3133 } 3134 } 3135 3136 bool VersionTableSection::isNeeded() const { 3137 return isLive() && 3138 (getPartition().verDef || getPartition().verNeed->isNeeded()); 3139 } 3140 3141 void elf::addVerneed(Symbol *ss) { 3142 auto &file = cast<SharedFile>(*ss->file); 3143 if (ss->versionId == VER_NDX_GLOBAL) 3144 return; 3145 3146 if (file.vernauxs.empty()) 3147 file.vernauxs.resize(file.verdefs.size()); 3148 3149 // Select a version identifier for the vernaux data structure, if we haven't 3150 // already allocated one. The verdef identifiers cover the range 3151 // [1..getVerDefNum()]; this causes the vernaux identifiers to start from 3152 // getVerDefNum()+1. 3153 if (file.vernauxs[ss->versionId] == 0) 3154 file.vernauxs[ss->versionId] = ++SharedFile::vernauxNum + getVerDefNum(); 3155 3156 ss->versionId = file.vernauxs[ss->versionId]; 3157 } 3158 3159 template <class ELFT> 3160 VersionNeedSection<ELFT>::VersionNeedSection() 3161 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t), 3162 ".gnu.version_r") {} 3163 3164 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() { 3165 for (SharedFile *f : ctx.sharedFiles) { 3166 if (f->vernauxs.empty()) 3167 continue; 3168 verneeds.emplace_back(); 3169 Verneed &vn = verneeds.back(); 3170 vn.nameStrTab = getPartition().dynStrTab->addString(f->soName); 3171 bool isLibc = config->relrGlibc && f->soName.starts_with("libc.so."); 3172 bool isGlibc2 = false; 3173 for (unsigned i = 0; i != f->vernauxs.size(); ++i) { 3174 if (f->vernauxs[i] == 0) 3175 continue; 3176 auto *verdef = 3177 reinterpret_cast<const typename ELFT::Verdef *>(f->verdefs[i]); 3178 StringRef ver(f->getStringTable().data() + verdef->getAux()->vda_name); 3179 if (isLibc && ver.starts_with("GLIBC_2.")) 3180 isGlibc2 = true; 3181 vn.vernauxs.push_back({verdef->vd_hash, f->vernauxs[i], 3182 getPartition().dynStrTab->addString(ver)}); 3183 } 3184 if (isGlibc2) { 3185 const char *ver = "GLIBC_ABI_DT_RELR"; 3186 vn.vernauxs.push_back({hashSysV(ver), 3187 ++SharedFile::vernauxNum + getVerDefNum(), 3188 getPartition().dynStrTab->addString(ver)}); 3189 } 3190 } 3191 3192 if (OutputSection *sec = getPartition().dynStrTab->getParent()) 3193 getParent()->link = sec->sectionIndex; 3194 getParent()->info = verneeds.size(); 3195 } 3196 3197 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *buf) { 3198 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs. 3199 auto *verneed = reinterpret_cast<Elf_Verneed *>(buf); 3200 auto *vernaux = reinterpret_cast<Elf_Vernaux *>(verneed + verneeds.size()); 3201 3202 for (auto &vn : verneeds) { 3203 // Create an Elf_Verneed for this DSO. 3204 verneed->vn_version = 1; 3205 verneed->vn_cnt = vn.vernauxs.size(); 3206 verneed->vn_file = vn.nameStrTab; 3207 verneed->vn_aux = 3208 reinterpret_cast<char *>(vernaux) - reinterpret_cast<char *>(verneed); 3209 verneed->vn_next = sizeof(Elf_Verneed); 3210 ++verneed; 3211 3212 // Create the Elf_Vernauxs for this Elf_Verneed. 3213 for (auto &vna : vn.vernauxs) { 3214 vernaux->vna_hash = vna.hash; 3215 vernaux->vna_flags = 0; 3216 vernaux->vna_other = vna.verneedIndex; 3217 vernaux->vna_name = vna.nameStrTab; 3218 vernaux->vna_next = sizeof(Elf_Vernaux); 3219 ++vernaux; 3220 } 3221 3222 vernaux[-1].vna_next = 0; 3223 } 3224 verneed[-1].vn_next = 0; 3225 } 3226 3227 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const { 3228 return verneeds.size() * sizeof(Elf_Verneed) + 3229 SharedFile::vernauxNum * sizeof(Elf_Vernaux); 3230 } 3231 3232 template <class ELFT> bool VersionNeedSection<ELFT>::isNeeded() const { 3233 return isLive() && SharedFile::vernauxNum != 0; 3234 } 3235 3236 void MergeSyntheticSection::addSection(MergeInputSection *ms) { 3237 ms->parent = this; 3238 sections.push_back(ms); 3239 assert(addralign == ms->addralign || !(ms->flags & SHF_STRINGS)); 3240 addralign = std::max(addralign, ms->addralign); 3241 } 3242 3243 MergeTailSection::MergeTailSection(StringRef name, uint32_t type, 3244 uint64_t flags, uint32_t alignment) 3245 : MergeSyntheticSection(name, type, flags, alignment), 3246 builder(StringTableBuilder::RAW, llvm::Align(alignment)) {} 3247 3248 size_t MergeTailSection::getSize() const { return builder.getSize(); } 3249 3250 void MergeTailSection::writeTo(uint8_t *buf) { builder.write(buf); } 3251 3252 void MergeTailSection::finalizeContents() { 3253 // Add all string pieces to the string table builder to create section 3254 // contents. 3255 for (MergeInputSection *sec : sections) 3256 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) 3257 if (sec->pieces[i].live) 3258 builder.add(sec->getData(i)); 3259 3260 // Fix the string table content. After this, the contents will never change. 3261 builder.finalize(); 3262 3263 // finalize() fixed tail-optimized strings, so we can now get 3264 // offsets of strings. Get an offset for each string and save it 3265 // to a corresponding SectionPiece for easy access. 3266 for (MergeInputSection *sec : sections) 3267 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) 3268 if (sec->pieces[i].live) 3269 sec->pieces[i].outputOff = builder.getOffset(sec->getData(i)); 3270 } 3271 3272 void MergeNoTailSection::writeTo(uint8_t *buf) { 3273 parallelFor(0, numShards, 3274 [&](size_t i) { shards[i].write(buf + shardOffsets[i]); }); 3275 } 3276 3277 // This function is very hot (i.e. it can take several seconds to finish) 3278 // because sometimes the number of inputs is in an order of magnitude of 3279 // millions. So, we use multi-threading. 3280 // 3281 // For any strings S and T, we know S is not mergeable with T if S's hash 3282 // value is different from T's. If that's the case, we can safely put S and 3283 // T into different string builders without worrying about merge misses. 3284 // We do it in parallel. 3285 void MergeNoTailSection::finalizeContents() { 3286 // Initializes string table builders. 3287 for (size_t i = 0; i < numShards; ++i) 3288 shards.emplace_back(StringTableBuilder::RAW, llvm::Align(addralign)); 3289 3290 // Concurrency level. Must be a power of 2 to avoid expensive modulo 3291 // operations in the following tight loop. 3292 const size_t concurrency = 3293 llvm::bit_floor(std::min<size_t>(config->threadCount, numShards)); 3294 3295 // Add section pieces to the builders. 3296 parallelFor(0, concurrency, [&](size_t threadId) { 3297 for (MergeInputSection *sec : sections) { 3298 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) { 3299 if (!sec->pieces[i].live) 3300 continue; 3301 size_t shardId = getShardId(sec->pieces[i].hash); 3302 if ((shardId & (concurrency - 1)) == threadId) 3303 sec->pieces[i].outputOff = shards[shardId].add(sec->getData(i)); 3304 } 3305 } 3306 }); 3307 3308 // Compute an in-section offset for each shard. 3309 size_t off = 0; 3310 for (size_t i = 0; i < numShards; ++i) { 3311 shards[i].finalizeInOrder(); 3312 if (shards[i].getSize() > 0) 3313 off = alignToPowerOf2(off, addralign); 3314 shardOffsets[i] = off; 3315 off += shards[i].getSize(); 3316 } 3317 size = off; 3318 3319 // So far, section pieces have offsets from beginning of shards, but 3320 // we want offsets from beginning of the whole section. Fix them. 3321 parallelForEach(sections, [&](MergeInputSection *sec) { 3322 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) 3323 if (sec->pieces[i].live) 3324 sec->pieces[i].outputOff += 3325 shardOffsets[getShardId(sec->pieces[i].hash)]; 3326 }); 3327 } 3328 3329 template <class ELFT> void elf::splitSections() { 3330 llvm::TimeTraceScope timeScope("Split sections"); 3331 // splitIntoPieces needs to be called on each MergeInputSection 3332 // before calling finalizeContents(). 3333 parallelForEach(ctx.objectFiles, [](ELFFileBase *file) { 3334 for (InputSectionBase *sec : file->getSections()) { 3335 if (!sec) 3336 continue; 3337 if (auto *s = dyn_cast<MergeInputSection>(sec)) 3338 s->splitIntoPieces(); 3339 else if (auto *eh = dyn_cast<EhInputSection>(sec)) 3340 eh->split<ELFT>(); 3341 } 3342 }); 3343 } 3344 3345 void elf::combineEhSections() { 3346 llvm::TimeTraceScope timeScope("Combine EH sections"); 3347 for (EhInputSection *sec : ctx.ehInputSections) { 3348 EhFrameSection &eh = *sec->getPartition().ehFrame; 3349 sec->parent = &eh; 3350 eh.addralign = std::max(eh.addralign, sec->addralign); 3351 eh.sections.push_back(sec); 3352 llvm::append_range(eh.dependentSections, sec->dependentSections); 3353 } 3354 3355 if (!mainPart->armExidx) 3356 return; 3357 llvm::erase_if(ctx.inputSections, [](InputSectionBase *s) { 3358 // Ignore dead sections and the partition end marker (.part.end), 3359 // whose partition number is out of bounds. 3360 if (!s->isLive() || s->partition == 255) 3361 return false; 3362 Partition &part = s->getPartition(); 3363 return s->kind() == SectionBase::Regular && part.armExidx && 3364 part.armExidx->addSection(cast<InputSection>(s)); 3365 }); 3366 } 3367 3368 MipsRldMapSection::MipsRldMapSection() 3369 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize, 3370 ".rld_map") {} 3371 3372 ARMExidxSyntheticSection::ARMExidxSyntheticSection() 3373 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX, 3374 config->wordsize, ".ARM.exidx") {} 3375 3376 static InputSection *findExidxSection(InputSection *isec) { 3377 for (InputSection *d : isec->dependentSections) 3378 if (d->type == SHT_ARM_EXIDX && d->isLive()) 3379 return d; 3380 return nullptr; 3381 } 3382 3383 static bool isValidExidxSectionDep(InputSection *isec) { 3384 return (isec->flags & SHF_ALLOC) && (isec->flags & SHF_EXECINSTR) && 3385 isec->getSize() > 0; 3386 } 3387 3388 bool ARMExidxSyntheticSection::addSection(InputSection *isec) { 3389 if (isec->type == SHT_ARM_EXIDX) { 3390 if (InputSection *dep = isec->getLinkOrderDep()) 3391 if (isValidExidxSectionDep(dep)) { 3392 exidxSections.push_back(isec); 3393 // Every exidxSection is 8 bytes, we need an estimate of 3394 // size before assignAddresses can be called. Final size 3395 // will only be known after finalize is called. 3396 size += 8; 3397 } 3398 return true; 3399 } 3400 3401 if (isValidExidxSectionDep(isec)) { 3402 executableSections.push_back(isec); 3403 return false; 3404 } 3405 3406 // FIXME: we do not output a relocation section when --emit-relocs is used 3407 // as we do not have relocation sections for linker generated table entries 3408 // and we would have to erase at a late stage relocations from merged entries. 3409 // Given that exception tables are already position independent and a binary 3410 // analyzer could derive the relocations we choose to erase the relocations. 3411 if (config->emitRelocs && isec->type == SHT_REL) 3412 if (InputSectionBase *ex = isec->getRelocatedSection()) 3413 if (isa<InputSection>(ex) && ex->type == SHT_ARM_EXIDX) 3414 return true; 3415 3416 return false; 3417 } 3418 3419 // References to .ARM.Extab Sections have bit 31 clear and are not the 3420 // special EXIDX_CANTUNWIND bit-pattern. 3421 static bool isExtabRef(uint32_t unwind) { 3422 return (unwind & 0x80000000) == 0 && unwind != 0x1; 3423 } 3424 3425 // Return true if the .ARM.exidx section Cur can be merged into the .ARM.exidx 3426 // section Prev, where Cur follows Prev in the table. This can be done if the 3427 // unwinding instructions in Cur are identical to Prev. Linker generated 3428 // EXIDX_CANTUNWIND entries are represented by nullptr as they do not have an 3429 // InputSection. 3430 static bool isDuplicateArmExidxSec(InputSection *prev, InputSection *cur) { 3431 // Get the last table Entry from the previous .ARM.exidx section. If Prev is 3432 // nullptr then it will be a synthesized EXIDX_CANTUNWIND entry. 3433 uint32_t prevUnwind = 1; 3434 if (prev) 3435 prevUnwind = read32(prev->content().data() + prev->content().size() - 4); 3436 if (isExtabRef(prevUnwind)) 3437 return false; 3438 3439 // We consider the unwind instructions of an .ARM.exidx table entry 3440 // a duplicate if the previous unwind instructions if: 3441 // - Both are the special EXIDX_CANTUNWIND. 3442 // - Both are the same inline unwind instructions. 3443 // We do not attempt to follow and check links into .ARM.extab tables as 3444 // consecutive identical entries are rare and the effort to check that they 3445 // are identical is high. 3446 3447 // If Cur is nullptr then this is synthesized EXIDX_CANTUNWIND entry. 3448 if (cur == nullptr) 3449 return prevUnwind == 1; 3450 3451 for (uint32_t offset = 4; offset < (uint32_t)cur->content().size(); offset +=8) { 3452 uint32_t curUnwind = read32(cur->content().data() + offset); 3453 if (isExtabRef(curUnwind) || curUnwind != prevUnwind) 3454 return false; 3455 } 3456 // All table entries in this .ARM.exidx Section can be merged into the 3457 // previous Section. 3458 return true; 3459 } 3460 3461 // The .ARM.exidx table must be sorted in ascending order of the address of the 3462 // functions the table describes. std::optionally duplicate adjacent table 3463 // entries can be removed. At the end of the function the executableSections 3464 // must be sorted in ascending order of address, Sentinel is set to the 3465 // InputSection with the highest address and any InputSections that have 3466 // mergeable .ARM.exidx table entries are removed from it. 3467 void ARMExidxSyntheticSection::finalizeContents() { 3468 // The executableSections and exidxSections that we use to derive the final 3469 // contents of this SyntheticSection are populated before 3470 // processSectionCommands() and ICF. A /DISCARD/ entry in SECTIONS command or 3471 // ICF may remove executable InputSections and their dependent .ARM.exidx 3472 // section that we recorded earlier. 3473 auto isDiscarded = [](const InputSection *isec) { return !isec->isLive(); }; 3474 llvm::erase_if(exidxSections, isDiscarded); 3475 // We need to remove discarded InputSections and InputSections without 3476 // .ARM.exidx sections that if we generated the .ARM.exidx it would be out 3477 // of range. 3478 auto isDiscardedOrOutOfRange = [this](InputSection *isec) { 3479 if (!isec->isLive()) 3480 return true; 3481 if (findExidxSection(isec)) 3482 return false; 3483 int64_t off = static_cast<int64_t>(isec->getVA() - getVA()); 3484 return off != llvm::SignExtend64(off, 31); 3485 }; 3486 llvm::erase_if(executableSections, isDiscardedOrOutOfRange); 3487 3488 // Sort the executable sections that may or may not have associated 3489 // .ARM.exidx sections by order of ascending address. This requires the 3490 // relative positions of InputSections and OutputSections to be known. 3491 auto compareByFilePosition = [](const InputSection *a, 3492 const InputSection *b) { 3493 OutputSection *aOut = a->getParent(); 3494 OutputSection *bOut = b->getParent(); 3495 3496 if (aOut != bOut) 3497 return aOut->addr < bOut->addr; 3498 return a->outSecOff < b->outSecOff; 3499 }; 3500 llvm::stable_sort(executableSections, compareByFilePosition); 3501 sentinel = executableSections.back(); 3502 // std::optionally merge adjacent duplicate entries. 3503 if (config->mergeArmExidx) { 3504 SmallVector<InputSection *, 0> selectedSections; 3505 selectedSections.reserve(executableSections.size()); 3506 selectedSections.push_back(executableSections[0]); 3507 size_t prev = 0; 3508 for (size_t i = 1; i < executableSections.size(); ++i) { 3509 InputSection *ex1 = findExidxSection(executableSections[prev]); 3510 InputSection *ex2 = findExidxSection(executableSections[i]); 3511 if (!isDuplicateArmExidxSec(ex1, ex2)) { 3512 selectedSections.push_back(executableSections[i]); 3513 prev = i; 3514 } 3515 } 3516 executableSections = std::move(selectedSections); 3517 } 3518 // offset is within the SyntheticSection. 3519 size_t offset = 0; 3520 size = 0; 3521 for (InputSection *isec : executableSections) { 3522 if (InputSection *d = findExidxSection(isec)) { 3523 d->outSecOff = offset; 3524 d->parent = getParent(); 3525 offset += d->getSize(); 3526 } else { 3527 offset += 8; 3528 } 3529 } 3530 // Size includes Sentinel. 3531 size = offset + 8; 3532 } 3533 3534 InputSection *ARMExidxSyntheticSection::getLinkOrderDep() const { 3535 return executableSections.front(); 3536 } 3537 3538 // To write the .ARM.exidx table from the ExecutableSections we have three cases 3539 // 1.) The InputSection has a .ARM.exidx InputSection in its dependent sections. 3540 // We write the .ARM.exidx section contents and apply its relocations. 3541 // 2.) The InputSection does not have a dependent .ARM.exidx InputSection. We 3542 // must write the contents of an EXIDX_CANTUNWIND directly. We use the 3543 // start of the InputSection as the purpose of the linker generated 3544 // section is to terminate the address range of the previous entry. 3545 // 3.) A trailing EXIDX_CANTUNWIND sentinel section is required at the end of 3546 // the table to terminate the address range of the final entry. 3547 void ARMExidxSyntheticSection::writeTo(uint8_t *buf) { 3548 3549 // A linker generated CANTUNWIND entry is made up of two words: 3550 // 0x0 with R_ARM_PREL31 relocation to target. 3551 // 0x1 with EXIDX_CANTUNWIND. 3552 uint64_t offset = 0; 3553 for (InputSection *isec : executableSections) { 3554 assert(isec->getParent() != nullptr); 3555 if (InputSection *d = findExidxSection(isec)) { 3556 for (int dataOffset = 0; dataOffset != (int)d->content().size(); 3557 dataOffset += 4) 3558 write32(buf + offset + dataOffset, 3559 read32(d->content().data() + dataOffset)); 3560 // Recalculate outSecOff as finalizeAddressDependentContent() 3561 // may have altered syntheticSection outSecOff. 3562 d->outSecOff = offset + outSecOff; 3563 target->relocateAlloc(*d, buf + offset); 3564 offset += d->getSize(); 3565 } else { 3566 // A Linker generated CANTUNWIND section. 3567 write32(buf + offset + 0, 0x0); 3568 write32(buf + offset + 4, 0x1); 3569 uint64_t s = isec->getVA(); 3570 uint64_t p = getVA() + offset; 3571 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p); 3572 offset += 8; 3573 } 3574 } 3575 // Write Sentinel CANTUNWIND entry. 3576 write32(buf + offset + 0, 0x0); 3577 write32(buf + offset + 4, 0x1); 3578 uint64_t s = sentinel->getVA(sentinel->getSize()); 3579 uint64_t p = getVA() + offset; 3580 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p); 3581 assert(size == offset + 8); 3582 } 3583 3584 bool ARMExidxSyntheticSection::isNeeded() const { 3585 return llvm::any_of(exidxSections, 3586 [](InputSection *isec) { return isec->isLive(); }); 3587 } 3588 3589 ThunkSection::ThunkSection(OutputSection *os, uint64_t off) 3590 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 3591 config->emachine == EM_PPC64 ? 16 : 4, ".text.thunk") { 3592 this->parent = os; 3593 this->outSecOff = off; 3594 } 3595 3596 size_t ThunkSection::getSize() const { 3597 if (roundUpSizeForErrata) 3598 return alignTo(size, 4096); 3599 return size; 3600 } 3601 3602 void ThunkSection::addThunk(Thunk *t) { 3603 thunks.push_back(t); 3604 t->addSymbols(*this); 3605 } 3606 3607 void ThunkSection::writeTo(uint8_t *buf) { 3608 for (Thunk *t : thunks) 3609 t->writeTo(buf + t->offset); 3610 } 3611 3612 InputSection *ThunkSection::getTargetInputSection() const { 3613 if (thunks.empty()) 3614 return nullptr; 3615 const Thunk *t = thunks.front(); 3616 return t->getTargetInputSection(); 3617 } 3618 3619 bool ThunkSection::assignOffsets() { 3620 uint64_t off = 0; 3621 for (Thunk *t : thunks) { 3622 off = alignToPowerOf2(off, t->alignment); 3623 t->setOffset(off); 3624 uint32_t size = t->size(); 3625 t->getThunkTargetSym()->size = size; 3626 off += size; 3627 } 3628 bool changed = off != size; 3629 size = off; 3630 return changed; 3631 } 3632 3633 PPC32Got2Section::PPC32Got2Section() 3634 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 4, ".got2") {} 3635 3636 bool PPC32Got2Section::isNeeded() const { 3637 // See the comment below. This is not needed if there is no other 3638 // InputSection. 3639 for (SectionCommand *cmd : getParent()->commands) 3640 if (auto *isd = dyn_cast<InputSectionDescription>(cmd)) 3641 for (InputSection *isec : isd->sections) 3642 if (isec != this) 3643 return true; 3644 return false; 3645 } 3646 3647 void PPC32Got2Section::finalizeContents() { 3648 // PPC32 may create multiple GOT sections for -fPIC/-fPIE, one per file in 3649 // .got2 . This function computes outSecOff of each .got2 to be used in 3650 // PPC32PltCallStub::writeTo(). The purpose of this empty synthetic section is 3651 // to collect input sections named ".got2". 3652 for (SectionCommand *cmd : getParent()->commands) 3653 if (auto *isd = dyn_cast<InputSectionDescription>(cmd)) { 3654 for (InputSection *isec : isd->sections) { 3655 // isec->file may be nullptr for MergeSyntheticSection. 3656 if (isec != this && isec->file) 3657 isec->file->ppc32Got2 = isec; 3658 } 3659 } 3660 } 3661 3662 // If linking position-dependent code then the table will store the addresses 3663 // directly in the binary so the section has type SHT_PROGBITS. If linking 3664 // position-independent code the section has type SHT_NOBITS since it will be 3665 // allocated and filled in by the dynamic linker. 3666 PPC64LongBranchTargetSection::PPC64LongBranchTargetSection() 3667 : SyntheticSection(SHF_ALLOC | SHF_WRITE, 3668 config->isPic ? SHT_NOBITS : SHT_PROGBITS, 8, 3669 ".branch_lt") {} 3670 3671 uint64_t PPC64LongBranchTargetSection::getEntryVA(const Symbol *sym, 3672 int64_t addend) { 3673 return getVA() + entry_index.find({sym, addend})->second * 8; 3674 } 3675 3676 std::optional<uint32_t> 3677 PPC64LongBranchTargetSection::addEntry(const Symbol *sym, int64_t addend) { 3678 auto res = 3679 entry_index.try_emplace(std::make_pair(sym, addend), entries.size()); 3680 if (!res.second) 3681 return std::nullopt; 3682 entries.emplace_back(sym, addend); 3683 return res.first->second; 3684 } 3685 3686 size_t PPC64LongBranchTargetSection::getSize() const { 3687 return entries.size() * 8; 3688 } 3689 3690 void PPC64LongBranchTargetSection::writeTo(uint8_t *buf) { 3691 // If linking non-pic we have the final addresses of the targets and they get 3692 // written to the table directly. For pic the dynamic linker will allocate 3693 // the section and fill it. 3694 if (config->isPic) 3695 return; 3696 3697 for (auto entry : entries) { 3698 const Symbol *sym = entry.first; 3699 int64_t addend = entry.second; 3700 assert(sym->getVA()); 3701 // Need calls to branch to the local entry-point since a long-branch 3702 // must be a local-call. 3703 write64(buf, sym->getVA(addend) + 3704 getPPC64GlobalEntryToLocalEntryOffset(sym->stOther)); 3705 buf += 8; 3706 } 3707 } 3708 3709 bool PPC64LongBranchTargetSection::isNeeded() const { 3710 // `removeUnusedSyntheticSections()` is called before thunk allocation which 3711 // is too early to determine if this section will be empty or not. We need 3712 // Finalized to keep the section alive until after thunk creation. Finalized 3713 // only gets set to true once `finalizeSections()` is called after thunk 3714 // creation. Because of this, if we don't create any long-branch thunks we end 3715 // up with an empty .branch_lt section in the binary. 3716 return !finalized || !entries.empty(); 3717 } 3718 3719 static uint8_t getAbiVersion() { 3720 // MIPS non-PIC executable gets ABI version 1. 3721 if (config->emachine == EM_MIPS) { 3722 if (!config->isPic && !config->relocatable && 3723 (config->eflags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC) 3724 return 1; 3725 return 0; 3726 } 3727 3728 if (config->emachine == EM_AMDGPU && !ctx.objectFiles.empty()) { 3729 uint8_t ver = ctx.objectFiles[0]->abiVersion; 3730 for (InputFile *file : ArrayRef(ctx.objectFiles).slice(1)) 3731 if (file->abiVersion != ver) 3732 error("incompatible ABI version: " + toString(file)); 3733 return ver; 3734 } 3735 3736 return 0; 3737 } 3738 3739 template <typename ELFT> void elf::writeEhdr(uint8_t *buf, Partition &part) { 3740 memcpy(buf, "\177ELF", 4); 3741 3742 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf); 3743 eHdr->e_ident[EI_CLASS] = config->is64 ? ELFCLASS64 : ELFCLASS32; 3744 eHdr->e_ident[EI_DATA] = config->isLE ? ELFDATA2LSB : ELFDATA2MSB; 3745 eHdr->e_ident[EI_VERSION] = EV_CURRENT; 3746 eHdr->e_ident[EI_OSABI] = config->osabi; 3747 eHdr->e_ident[EI_ABIVERSION] = getAbiVersion(); 3748 eHdr->e_machine = config->emachine; 3749 eHdr->e_version = EV_CURRENT; 3750 eHdr->e_flags = config->eflags; 3751 eHdr->e_ehsize = sizeof(typename ELFT::Ehdr); 3752 eHdr->e_phnum = part.phdrs.size(); 3753 eHdr->e_shentsize = sizeof(typename ELFT::Shdr); 3754 3755 if (!config->relocatable) { 3756 eHdr->e_phoff = sizeof(typename ELFT::Ehdr); 3757 eHdr->e_phentsize = sizeof(typename ELFT::Phdr); 3758 } 3759 } 3760 3761 template <typename ELFT> void elf::writePhdrs(uint8_t *buf, Partition &part) { 3762 // Write the program header table. 3763 auto *hBuf = reinterpret_cast<typename ELFT::Phdr *>(buf); 3764 for (PhdrEntry *p : part.phdrs) { 3765 hBuf->p_type = p->p_type; 3766 hBuf->p_flags = p->p_flags; 3767 hBuf->p_offset = p->p_offset; 3768 hBuf->p_vaddr = p->p_vaddr; 3769 hBuf->p_paddr = p->p_paddr; 3770 hBuf->p_filesz = p->p_filesz; 3771 hBuf->p_memsz = p->p_memsz; 3772 hBuf->p_align = p->p_align; 3773 ++hBuf; 3774 } 3775 } 3776 3777 template <typename ELFT> 3778 PartitionElfHeaderSection<ELFT>::PartitionElfHeaderSection() 3779 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_EHDR, 1, "") {} 3780 3781 template <typename ELFT> 3782 size_t PartitionElfHeaderSection<ELFT>::getSize() const { 3783 return sizeof(typename ELFT::Ehdr); 3784 } 3785 3786 template <typename ELFT> 3787 void PartitionElfHeaderSection<ELFT>::writeTo(uint8_t *buf) { 3788 writeEhdr<ELFT>(buf, getPartition()); 3789 3790 // Loadable partitions are always ET_DYN. 3791 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf); 3792 eHdr->e_type = ET_DYN; 3793 } 3794 3795 template <typename ELFT> 3796 PartitionProgramHeadersSection<ELFT>::PartitionProgramHeadersSection() 3797 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_PHDR, 1, ".phdrs") {} 3798 3799 template <typename ELFT> 3800 size_t PartitionProgramHeadersSection<ELFT>::getSize() const { 3801 return sizeof(typename ELFT::Phdr) * getPartition().phdrs.size(); 3802 } 3803 3804 template <typename ELFT> 3805 void PartitionProgramHeadersSection<ELFT>::writeTo(uint8_t *buf) { 3806 writePhdrs<ELFT>(buf, getPartition()); 3807 } 3808 3809 PartitionIndexSection::PartitionIndexSection() 3810 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".rodata") {} 3811 3812 size_t PartitionIndexSection::getSize() const { 3813 return 12 * (partitions.size() - 1); 3814 } 3815 3816 void PartitionIndexSection::finalizeContents() { 3817 for (size_t i = 1; i != partitions.size(); ++i) 3818 partitions[i].nameStrTab = mainPart->dynStrTab->addString(partitions[i].name); 3819 } 3820 3821 void PartitionIndexSection::writeTo(uint8_t *buf) { 3822 uint64_t va = getVA(); 3823 for (size_t i = 1; i != partitions.size(); ++i) { 3824 write32(buf, mainPart->dynStrTab->getVA() + partitions[i].nameStrTab - va); 3825 write32(buf + 4, partitions[i].elfHeader->getVA() - (va + 4)); 3826 3827 SyntheticSection *next = i == partitions.size() - 1 3828 ? in.partEnd.get() 3829 : partitions[i + 1].elfHeader.get(); 3830 write32(buf + 8, next->getVA() - partitions[i].elfHeader->getVA()); 3831 3832 va += 12; 3833 buf += 12; 3834 } 3835 } 3836 3837 void InStruct::reset() { 3838 attributes.reset(); 3839 riscvAttributes.reset(); 3840 bss.reset(); 3841 bssRelRo.reset(); 3842 got.reset(); 3843 gotPlt.reset(); 3844 igotPlt.reset(); 3845 relroPadding.reset(); 3846 armCmseSGSection.reset(); 3847 ppc64LongBranchTarget.reset(); 3848 mipsAbiFlags.reset(); 3849 mipsGot.reset(); 3850 mipsOptions.reset(); 3851 mipsReginfo.reset(); 3852 mipsRldMap.reset(); 3853 partEnd.reset(); 3854 partIndex.reset(); 3855 plt.reset(); 3856 iplt.reset(); 3857 ppc32Got2.reset(); 3858 ibtPlt.reset(); 3859 relaPlt.reset(); 3860 relaIplt.reset(); 3861 shStrTab.reset(); 3862 strTab.reset(); 3863 symTab.reset(); 3864 symTabShndx.reset(); 3865 } 3866 3867 constexpr char kMemtagAndroidNoteName[] = "Android"; 3868 void MemtagAndroidNote::writeTo(uint8_t *buf) { 3869 static_assert( 3870 sizeof(kMemtagAndroidNoteName) == 8, 3871 "Android 11 & 12 have an ABI that the note name is 8 bytes long. Keep it " 3872 "that way for backwards compatibility."); 3873 3874 write32(buf, sizeof(kMemtagAndroidNoteName)); 3875 write32(buf + 4, sizeof(uint32_t)); 3876 write32(buf + 8, ELF::NT_ANDROID_TYPE_MEMTAG); 3877 memcpy(buf + 12, kMemtagAndroidNoteName, sizeof(kMemtagAndroidNoteName)); 3878 buf += 12 + alignTo(sizeof(kMemtagAndroidNoteName), 4); 3879 3880 uint32_t value = 0; 3881 value |= config->androidMemtagMode; 3882 if (config->androidMemtagHeap) 3883 value |= ELF::NT_MEMTAG_HEAP; 3884 // Note, MTE stack is an ABI break. Attempting to run an MTE stack-enabled 3885 // binary on Android 11 or 12 will result in a checkfail in the loader. 3886 if (config->androidMemtagStack) 3887 value |= ELF::NT_MEMTAG_STACK; 3888 write32(buf, value); // note value 3889 } 3890 3891 size_t MemtagAndroidNote::getSize() const { 3892 return sizeof(llvm::ELF::Elf64_Nhdr) + 3893 /*namesz=*/alignTo(sizeof(kMemtagAndroidNoteName), 4) + 3894 /*descsz=*/sizeof(uint32_t); 3895 } 3896 3897 void PackageMetadataNote::writeTo(uint8_t *buf) { 3898 write32(buf, 4); 3899 write32(buf + 4, config->packageMetadata.size() + 1); 3900 write32(buf + 8, FDO_PACKAGING_METADATA); 3901 memcpy(buf + 12, "FDO", 4); 3902 memcpy(buf + 16, config->packageMetadata.data(), 3903 config->packageMetadata.size()); 3904 } 3905 3906 size_t PackageMetadataNote::getSize() const { 3907 return sizeof(llvm::ELF::Elf64_Nhdr) + 4 + 3908 alignTo(config->packageMetadata.size() + 1, 4); 3909 } 3910 3911 // Helper function, return the size of the ULEB128 for 'v', optionally writing 3912 // it to `*(buf + offset)` if `buf` is non-null. 3913 static size_t computeOrWriteULEB128(uint64_t v, uint8_t *buf, size_t offset) { 3914 if (buf) 3915 return encodeULEB128(v, buf + offset); 3916 return getULEB128Size(v); 3917 } 3918 3919 // https://github.com/ARM-software/abi-aa/blob/main/memtagabielf64/memtagabielf64.rst#83encoding-of-sht_aarch64_memtag_globals_dynamic 3920 constexpr uint64_t kMemtagStepSizeBits = 3; 3921 constexpr uint64_t kMemtagGranuleSize = 16; 3922 static size_t createMemtagDescriptors(const SmallVector<const Symbol *, 0> &symbols, 3923 uint8_t *buf = nullptr) { 3924 size_t sectionSize = 0; 3925 uint64_t lastGlobalEnd = 0; 3926 3927 for (const Symbol *sym : symbols) { 3928 if (!includeInSymtab(*sym)) 3929 continue; 3930 const uint64_t addr = sym->getVA(); 3931 const uint64_t size = sym->getSize(); 3932 3933 if (addr <= kMemtagGranuleSize && buf != nullptr) 3934 errorOrWarn("address of the tagged symbol \"" + sym->getName() + 3935 "\" falls in the ELF header. This is indicative of a " 3936 "compiler/linker bug"); 3937 if (addr % kMemtagGranuleSize != 0) 3938 errorOrWarn("address of the tagged symbol \"" + sym->getName() + 3939 "\" at 0x" + Twine::utohexstr(addr) + 3940 "\" is not granule (16-byte) aligned"); 3941 if (size == 0) 3942 errorOrWarn("size of the tagged symbol \"" + sym->getName() + 3943 "\" is not allowed to be zero"); 3944 if (size % kMemtagGranuleSize != 0) 3945 errorOrWarn("size of the tagged symbol \"" + sym->getName() + 3946 "\" (size 0x" + Twine::utohexstr(size) + 3947 ") is not granule (16-byte) aligned"); 3948 3949 const uint64_t sizeToEncode = size / kMemtagGranuleSize; 3950 const uint64_t stepToEncode = ((addr - lastGlobalEnd) / kMemtagGranuleSize) 3951 << kMemtagStepSizeBits; 3952 if (sizeToEncode < (1 << kMemtagStepSizeBits)) { 3953 sectionSize += computeOrWriteULEB128(stepToEncode | sizeToEncode, buf, sectionSize); 3954 } else { 3955 sectionSize += computeOrWriteULEB128(stepToEncode, buf, sectionSize); 3956 sectionSize += computeOrWriteULEB128(sizeToEncode - 1, buf, sectionSize); 3957 } 3958 lastGlobalEnd = addr + size; 3959 } 3960 3961 return sectionSize; 3962 } 3963 3964 bool MemtagDescriptors::updateAllocSize() { 3965 size_t oldSize = getSize(); 3966 std::stable_sort(symbols.begin(), symbols.end(), 3967 [](const Symbol *s1, const Symbol *s2) { 3968 return s1->getVA() < s2->getVA(); 3969 }); 3970 return oldSize != getSize(); 3971 } 3972 3973 void MemtagDescriptors::writeTo(uint8_t *buf) { 3974 createMemtagDescriptors(symbols, buf); 3975 } 3976 3977 size_t MemtagDescriptors::getSize() const { 3978 return createMemtagDescriptors(symbols); 3979 } 3980 3981 InStruct elf::in; 3982 3983 std::vector<Partition> elf::partitions; 3984 Partition *elf::mainPart; 3985 3986 template GdbIndexSection *GdbIndexSection::create<ELF32LE>(); 3987 template GdbIndexSection *GdbIndexSection::create<ELF32BE>(); 3988 template GdbIndexSection *GdbIndexSection::create<ELF64LE>(); 3989 template GdbIndexSection *GdbIndexSection::create<ELF64BE>(); 3990 3991 template void elf::splitSections<ELF32LE>(); 3992 template void elf::splitSections<ELF32BE>(); 3993 template void elf::splitSections<ELF64LE>(); 3994 template void elf::splitSections<ELF64BE>(); 3995 3996 template class elf::MipsAbiFlagsSection<ELF32LE>; 3997 template class elf::MipsAbiFlagsSection<ELF32BE>; 3998 template class elf::MipsAbiFlagsSection<ELF64LE>; 3999 template class elf::MipsAbiFlagsSection<ELF64BE>; 4000 4001 template class elf::MipsOptionsSection<ELF32LE>; 4002 template class elf::MipsOptionsSection<ELF32BE>; 4003 template class elf::MipsOptionsSection<ELF64LE>; 4004 template class elf::MipsOptionsSection<ELF64BE>; 4005 4006 template void EhFrameSection::iterateFDEWithLSDA<ELF32LE>( 4007 function_ref<void(InputSection &)>); 4008 template void EhFrameSection::iterateFDEWithLSDA<ELF32BE>( 4009 function_ref<void(InputSection &)>); 4010 template void EhFrameSection::iterateFDEWithLSDA<ELF64LE>( 4011 function_ref<void(InputSection &)>); 4012 template void EhFrameSection::iterateFDEWithLSDA<ELF64BE>( 4013 function_ref<void(InputSection &)>); 4014 4015 template class elf::MipsReginfoSection<ELF32LE>; 4016 template class elf::MipsReginfoSection<ELF32BE>; 4017 template class elf::MipsReginfoSection<ELF64LE>; 4018 template class elf::MipsReginfoSection<ELF64BE>; 4019 4020 template class elf::DynamicSection<ELF32LE>; 4021 template class elf::DynamicSection<ELF32BE>; 4022 template class elf::DynamicSection<ELF64LE>; 4023 template class elf::DynamicSection<ELF64BE>; 4024 4025 template class elf::RelocationSection<ELF32LE>; 4026 template class elf::RelocationSection<ELF32BE>; 4027 template class elf::RelocationSection<ELF64LE>; 4028 template class elf::RelocationSection<ELF64BE>; 4029 4030 template class elf::AndroidPackedRelocationSection<ELF32LE>; 4031 template class elf::AndroidPackedRelocationSection<ELF32BE>; 4032 template class elf::AndroidPackedRelocationSection<ELF64LE>; 4033 template class elf::AndroidPackedRelocationSection<ELF64BE>; 4034 4035 template class elf::RelrSection<ELF32LE>; 4036 template class elf::RelrSection<ELF32BE>; 4037 template class elf::RelrSection<ELF64LE>; 4038 template class elf::RelrSection<ELF64BE>; 4039 4040 template class elf::SymbolTableSection<ELF32LE>; 4041 template class elf::SymbolTableSection<ELF32BE>; 4042 template class elf::SymbolTableSection<ELF64LE>; 4043 template class elf::SymbolTableSection<ELF64BE>; 4044 4045 template class elf::VersionNeedSection<ELF32LE>; 4046 template class elf::VersionNeedSection<ELF32BE>; 4047 template class elf::VersionNeedSection<ELF64LE>; 4048 template class elf::VersionNeedSection<ELF64BE>; 4049 4050 template void elf::writeEhdr<ELF32LE>(uint8_t *Buf, Partition &Part); 4051 template void elf::writeEhdr<ELF32BE>(uint8_t *Buf, Partition &Part); 4052 template void elf::writeEhdr<ELF64LE>(uint8_t *Buf, Partition &Part); 4053 template void elf::writeEhdr<ELF64BE>(uint8_t *Buf, Partition &Part); 4054 4055 template void elf::writePhdrs<ELF32LE>(uint8_t *Buf, Partition &Part); 4056 template void elf::writePhdrs<ELF32BE>(uint8_t *Buf, Partition &Part); 4057 template void elf::writePhdrs<ELF64LE>(uint8_t *Buf, Partition &Part); 4058 template void elf::writePhdrs<ELF64BE>(uint8_t *Buf, Partition &Part); 4059 4060 template class elf::PartitionElfHeaderSection<ELF32LE>; 4061 template class elf::PartitionElfHeaderSection<ELF32BE>; 4062 template class elf::PartitionElfHeaderSection<ELF64LE>; 4063 template class elf::PartitionElfHeaderSection<ELF64BE>; 4064 4065 template class elf::PartitionProgramHeadersSection<ELF32LE>; 4066 template class elf::PartitionProgramHeadersSection<ELF32BE>; 4067 template class elf::PartitionProgramHeadersSection<ELF64LE>; 4068 template class elf::PartitionProgramHeadersSection<ELF64BE>; 4069