1 //===- SyntheticSections.cpp ----------------------------------------------===//
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
7 //===----------------------------------------------------------------------===//
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
14 //===----------------------------------------------------------------------===//
16 #include "SyntheticSections.h"
18 #include "InputFiles.h"
19 #include "LinkerScript.h"
20 #include "OutputSections.h"
21 #include "SymbolTable.h"
25 #include "lld/Common/DWARF.h"
26 #include "lld/Common/ErrorHandler.h"
27 #include "lld/Common/Memory.h"
28 #include "lld/Common/Strings.h"
29 #include "lld/Common/Version.h"
30 #include "llvm/ADT/SetOperations.h"
31 #include "llvm/ADT/SetVector.h"
32 #include "llvm/ADT/StringExtras.h"
33 #include "llvm/BinaryFormat/Dwarf.h"
34 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h"
35 #include "llvm/Object/ELFObjectFile.h"
36 #include "llvm/Support/Compression.h"
37 #include "llvm/Support/Endian.h"
38 #include "llvm/Support/LEB128.h"
39 #include "llvm/Support/MD5.h"
40 #include "llvm/Support/Parallel.h"
41 #include "llvm/Support/TimeProfiler.h"
46 using namespace llvm::dwarf;
47 using namespace llvm::ELF;
48 using namespace llvm::object;
49 using namespace llvm::support;
51 using namespace lld::elf;
53 using llvm::support::endian::read32le;
54 using llvm::support::endian::write32le;
55 using llvm::support::endian::write64le;
57 constexpr size_t MergeNoTailSection::numShards;
59 static uint64_t readUint(uint8_t *buf) {
60 return config->is64 ? read64(buf) : read32(buf);
63 static void writeUint(uint8_t *buf, uint64_t val) {
70 // Returns an LLD version string.
71 static ArrayRef<uint8_t> getVersion() {
72 // Check LLD_VERSION first for ease of testing.
73 // You can get consistent output by using the environment variable.
74 // This is only for testing.
75 StringRef s = getenv("LLD_VERSION");
77 s = saver.save(Twine("Linker: ") + getLLDVersion());
79 // +1 to include the terminating '\0'.
80 return {(const uint8_t *)s.data(), s.size() + 1};
83 // Creates a .comment section containing LLD version info.
84 // With this feature, you can identify LLD-generated binaries easily
85 // by "readelf --string-dump .comment <file>".
86 // The returned object is a mergeable string section.
87 MergeInputSection *elf::createCommentSection() {
88 return make<MergeInputSection>(SHF_MERGE | SHF_STRINGS, SHT_PROGBITS, 1,
89 getVersion(), ".comment");
92 // .MIPS.abiflags section.
94 MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags flags)
95 : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags"),
97 this->entsize = sizeof(Elf_Mips_ABIFlags);
100 template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *buf) {
101 memcpy(buf, &flags, sizeof(flags));
104 template <class ELFT>
105 MipsAbiFlagsSection<ELFT> *MipsAbiFlagsSection<ELFT>::create() {
106 Elf_Mips_ABIFlags flags = {};
109 for (InputSectionBase *sec : inputSections) {
110 if (sec->type != SHT_MIPS_ABIFLAGS)
115 std::string filename = toString(sec->file);
116 const size_t size = sec->data().size();
117 // Older version of BFD (such as the default FreeBSD linker) concatenate
118 // .MIPS.abiflags instead of merging. To allow for this case (or potential
119 // zero padding) we ignore everything after the first Elf_Mips_ABIFlags
120 if (size < sizeof(Elf_Mips_ABIFlags)) {
121 error(filename + ": invalid size of .MIPS.abiflags section: got " +
122 Twine(size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags)));
125 auto *s = reinterpret_cast<const Elf_Mips_ABIFlags *>(sec->data().data());
126 if (s->version != 0) {
127 error(filename + ": unexpected .MIPS.abiflags version " +
132 // LLD checks ISA compatibility in calcMipsEFlags(). Here we just
133 // select the highest number of ISA/Rev/Ext.
134 flags.isa_level = std::max(flags.isa_level, s->isa_level);
135 flags.isa_rev = std::max(flags.isa_rev, s->isa_rev);
136 flags.isa_ext = std::max(flags.isa_ext, s->isa_ext);
137 flags.gpr_size = std::max(flags.gpr_size, s->gpr_size);
138 flags.cpr1_size = std::max(flags.cpr1_size, s->cpr1_size);
139 flags.cpr2_size = std::max(flags.cpr2_size, s->cpr2_size);
140 flags.ases |= s->ases;
141 flags.flags1 |= s->flags1;
142 flags.flags2 |= s->flags2;
143 flags.fp_abi = elf::getMipsFpAbiFlag(flags.fp_abi, s->fp_abi, filename);
147 return make<MipsAbiFlagsSection<ELFT>>(flags);
151 // .MIPS.options section.
152 template <class ELFT>
153 MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo reginfo)
154 : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"),
156 this->entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo);
159 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *buf) {
160 auto *options = reinterpret_cast<Elf_Mips_Options *>(buf);
161 options->kind = ODK_REGINFO;
162 options->size = getSize();
164 if (!config->relocatable)
165 reginfo.ri_gp_value = in.mipsGot->getGp();
166 memcpy(buf + sizeof(Elf_Mips_Options), ®info, sizeof(reginfo));
169 template <class ELFT>
170 MipsOptionsSection<ELFT> *MipsOptionsSection<ELFT>::create() {
175 std::vector<InputSectionBase *> sections;
176 for (InputSectionBase *sec : inputSections)
177 if (sec->type == SHT_MIPS_OPTIONS)
178 sections.push_back(sec);
180 if (sections.empty())
183 Elf_Mips_RegInfo reginfo = {};
184 for (InputSectionBase *sec : sections) {
187 std::string filename = toString(sec->file);
188 ArrayRef<uint8_t> d = sec->data();
191 if (d.size() < sizeof(Elf_Mips_Options)) {
192 error(filename + ": invalid size of .MIPS.options section");
196 auto *opt = reinterpret_cast<const Elf_Mips_Options *>(d.data());
197 if (opt->kind == ODK_REGINFO) {
198 reginfo.ri_gprmask |= opt->getRegInfo().ri_gprmask;
199 sec->getFile<ELFT>()->mipsGp0 = opt->getRegInfo().ri_gp_value;
204 fatal(filename + ": zero option descriptor size");
205 d = d.slice(opt->size);
209 return make<MipsOptionsSection<ELFT>>(reginfo);
212 // MIPS .reginfo section.
213 template <class ELFT>
214 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo reginfo)
215 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"),
217 this->entsize = sizeof(Elf_Mips_RegInfo);
220 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *buf) {
221 if (!config->relocatable)
222 reginfo.ri_gp_value = in.mipsGot->getGp();
223 memcpy(buf, ®info, sizeof(reginfo));
226 template <class ELFT>
227 MipsReginfoSection<ELFT> *MipsReginfoSection<ELFT>::create() {
228 // Section should be alive for O32 and N32 ABIs only.
232 std::vector<InputSectionBase *> sections;
233 for (InputSectionBase *sec : inputSections)
234 if (sec->type == SHT_MIPS_REGINFO)
235 sections.push_back(sec);
237 if (sections.empty())
240 Elf_Mips_RegInfo reginfo = {};
241 for (InputSectionBase *sec : sections) {
244 if (sec->data().size() != sizeof(Elf_Mips_RegInfo)) {
245 error(toString(sec->file) + ": invalid size of .reginfo section");
249 auto *r = reinterpret_cast<const Elf_Mips_RegInfo *>(sec->data().data());
250 reginfo.ri_gprmask |= r->ri_gprmask;
251 sec->getFile<ELFT>()->mipsGp0 = r->ri_gp_value;
254 return make<MipsReginfoSection<ELFT>>(reginfo);
257 InputSection *elf::createInterpSection() {
258 // StringSaver guarantees that the returned string ends with '\0'.
259 StringRef s = saver.save(config->dynamicLinker);
260 ArrayRef<uint8_t> contents = {(const uint8_t *)s.data(), s.size() + 1};
262 return make<InputSection>(nullptr, SHF_ALLOC, SHT_PROGBITS, 1, contents,
266 Defined *elf::addSyntheticLocal(StringRef name, uint8_t type, uint64_t value,
267 uint64_t size, InputSectionBase §ion) {
268 auto *s = make<Defined>(section.file, name, STB_LOCAL, STV_DEFAULT, type,
269 value, size, §ion);
271 in.symTab->addSymbol(s);
275 static size_t getHashSize() {
276 switch (config->buildId) {
277 case BuildIdKind::Fast:
279 case BuildIdKind::Md5:
280 case BuildIdKind::Uuid:
282 case BuildIdKind::Sha1:
284 case BuildIdKind::Hexstring:
285 return config->buildIdVector.size();
287 llvm_unreachable("unknown BuildIdKind");
291 // This class represents a linker-synthesized .note.gnu.property section.
293 // In x86 and AArch64, object files may contain feature flags indicating the
294 // features that they have used. The flags are stored in a .note.gnu.property
297 // lld reads the sections from input files and merges them by computing AND of
298 // the flags. The result is written as a new .note.gnu.property section.
300 // If the flag is zero (which indicates that the intersection of the feature
301 // sets is empty, or some input files didn't have .note.gnu.property sections),
302 // we don't create this section.
303 GnuPropertySection::GnuPropertySection()
304 : SyntheticSection(llvm::ELF::SHF_ALLOC, llvm::ELF::SHT_NOTE,
305 config->wordsize, ".note.gnu.property") {}
307 void GnuPropertySection::writeTo(uint8_t *buf) {
308 uint32_t featureAndType = config->emachine == EM_AARCH64
309 ? GNU_PROPERTY_AARCH64_FEATURE_1_AND
310 : GNU_PROPERTY_X86_FEATURE_1_AND;
312 write32(buf, 4); // Name size
313 write32(buf + 4, config->is64 ? 16 : 12); // Content size
314 write32(buf + 8, NT_GNU_PROPERTY_TYPE_0); // Type
315 memcpy(buf + 12, "GNU", 4); // Name string
316 write32(buf + 16, featureAndType); // Feature type
317 write32(buf + 20, 4); // Feature size
318 write32(buf + 24, config->andFeatures); // Feature flags
320 write32(buf + 28, 0); // Padding
323 size_t GnuPropertySection::getSize() const { return config->is64 ? 32 : 28; }
325 BuildIdSection::BuildIdSection()
326 : SyntheticSection(SHF_ALLOC, SHT_NOTE, 4, ".note.gnu.build-id"),
327 hashSize(getHashSize()) {}
329 void BuildIdSection::writeTo(uint8_t *buf) {
330 write32(buf, 4); // Name size
331 write32(buf + 4, hashSize); // Content size
332 write32(buf + 8, NT_GNU_BUILD_ID); // Type
333 memcpy(buf + 12, "GNU", 4); // Name string
337 void BuildIdSection::writeBuildId(ArrayRef<uint8_t> buf) {
338 assert(buf.size() == hashSize);
339 memcpy(hashBuf, buf.data(), hashSize);
342 BssSection::BssSection(StringRef name, uint64_t size, uint32_t alignment)
343 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, alignment, name) {
348 EhFrameSection::EhFrameSection()
349 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame") {}
351 // Search for an existing CIE record or create a new one.
352 // CIE records from input object files are uniquified by their contents
353 // and where their relocations point to.
354 template <class ELFT, class RelTy>
355 CieRecord *EhFrameSection::addCie(EhSectionPiece &cie, ArrayRef<RelTy> rels) {
356 Symbol *personality = nullptr;
357 unsigned firstRelI = cie.firstRelocation;
358 if (firstRelI != (unsigned)-1)
360 &cie.sec->template getFile<ELFT>()->getRelocTargetSym(rels[firstRelI]);
362 // Search for an existing CIE by CIE contents/relocation target pair.
363 CieRecord *&rec = cieMap[{cie.data(), personality}];
365 // If not found, create a new one.
367 rec = make<CieRecord>();
369 cieRecords.push_back(rec);
374 // There is one FDE per function. Returns true if a given FDE
375 // points to a live function.
376 template <class ELFT, class RelTy>
377 bool EhFrameSection::isFdeLive(EhSectionPiece &fde, ArrayRef<RelTy> rels) {
378 auto *sec = cast<EhInputSection>(fde.sec);
379 unsigned firstRelI = fde.firstRelocation;
381 // An FDE should point to some function because FDEs are to describe
382 // functions. That's however not always the case due to an issue of
383 // ld.gold with -r. ld.gold may discard only functions and leave their
384 // corresponding FDEs, which results in creating bad .eh_frame sections.
385 // To deal with that, we ignore such FDEs.
386 if (firstRelI == (unsigned)-1)
389 const RelTy &rel = rels[firstRelI];
390 Symbol &b = sec->template getFile<ELFT>()->getRelocTargetSym(rel);
392 // FDEs for garbage-collected or merged-by-ICF sections, or sections in
393 // another partition, are dead.
394 if (auto *d = dyn_cast<Defined>(&b))
395 if (SectionBase *sec = d->section)
396 return sec->partition == partition;
400 // .eh_frame is a sequence of CIE or FDE records. In general, there
401 // is one CIE record per input object file which is followed by
402 // a list of FDEs. This function searches an existing CIE or create a new
403 // one and associates FDEs to the CIE.
404 template <class ELFT, class RelTy>
405 void EhFrameSection::addRecords(EhInputSection *sec, ArrayRef<RelTy> rels) {
407 for (EhSectionPiece &piece : sec->pieces) {
408 // The empty record is the end marker.
412 size_t offset = piece.inputOff;
413 uint32_t id = read32(piece.data().data() + 4);
415 offsetToCie[offset] = addCie<ELFT>(piece, rels);
419 uint32_t cieOffset = offset + 4 - id;
420 CieRecord *rec = offsetToCie[cieOffset];
422 fatal(toString(sec) + ": invalid CIE reference");
424 if (!isFdeLive<ELFT>(piece, rels))
426 rec->fdes.push_back(&piece);
431 template <class ELFT>
432 void EhFrameSection::addSectionAux(EhInputSection *sec) {
435 if (sec->areRelocsRela)
436 addRecords<ELFT>(sec, sec->template relas<ELFT>());
438 addRecords<ELFT>(sec, sec->template rels<ELFT>());
441 void EhFrameSection::addSection(EhInputSection *sec) {
444 alignment = std::max(alignment, sec->alignment);
445 sections.push_back(sec);
447 for (auto *ds : sec->dependentSections)
448 dependentSections.push_back(ds);
451 static void writeCieFde(uint8_t *buf, ArrayRef<uint8_t> d) {
452 memcpy(buf, d.data(), d.size());
454 size_t aligned = alignTo(d.size(), config->wordsize);
456 // Zero-clear trailing padding if it exists.
457 memset(buf + d.size(), 0, aligned - d.size());
459 // Fix the size field. -4 since size does not include the size field itself.
460 write32(buf, aligned - 4);
463 void EhFrameSection::finalizeContents() {
464 assert(!this->size); // Not finalized.
466 switch (config->ekind) {
468 llvm_unreachable("invalid ekind");
470 for (EhInputSection *sec : sections)
471 addSectionAux<ELF32LE>(sec);
474 for (EhInputSection *sec : sections)
475 addSectionAux<ELF32BE>(sec);
478 for (EhInputSection *sec : sections)
479 addSectionAux<ELF64LE>(sec);
482 for (EhInputSection *sec : sections)
483 addSectionAux<ELF64BE>(sec);
488 for (CieRecord *rec : cieRecords) {
489 rec->cie->outputOff = off;
490 off += alignTo(rec->cie->size, config->wordsize);
492 for (EhSectionPiece *fde : rec->fdes) {
493 fde->outputOff = off;
494 off += alignTo(fde->size, config->wordsize);
498 // The LSB standard does not allow a .eh_frame section with zero
499 // Call Frame Information records. glibc unwind-dw2-fde.c
500 // classify_object_over_fdes expects there is a CIE record length 0 as a
501 // terminator. Thus we add one unconditionally.
507 // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table
508 // to get an FDE from an address to which FDE is applied. This function
509 // returns a list of such pairs.
510 std::vector<EhFrameSection::FdeData> EhFrameSection::getFdeData() const {
511 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
512 std::vector<FdeData> ret;
514 uint64_t va = getPartition().ehFrameHdr->getVA();
515 for (CieRecord *rec : cieRecords) {
516 uint8_t enc = getFdeEncoding(rec->cie);
517 for (EhSectionPiece *fde : rec->fdes) {
518 uint64_t pc = getFdePc(buf, fde->outputOff, enc);
519 uint64_t fdeVA = getParent()->addr + fde->outputOff;
520 if (!isInt<32>(pc - va))
521 fatal(toString(fde->sec) + ": PC offset is too large: 0x" +
522 Twine::utohexstr(pc - va));
523 ret.push_back({uint32_t(pc - va), uint32_t(fdeVA - va)});
527 // Sort the FDE list by their PC and uniqueify. Usually there is only
528 // one FDE for a PC (i.e. function), but if ICF merges two functions
529 // into one, there can be more than one FDEs pointing to the address.
530 auto less = [](const FdeData &a, const FdeData &b) {
531 return a.pcRel < b.pcRel;
533 llvm::stable_sort(ret, less);
534 auto eq = [](const FdeData &a, const FdeData &b) {
535 return a.pcRel == b.pcRel;
537 ret.erase(std::unique(ret.begin(), ret.end(), eq), ret.end());
542 static uint64_t readFdeAddr(uint8_t *buf, int size) {
544 case DW_EH_PE_udata2:
546 case DW_EH_PE_sdata2:
547 return (int16_t)read16(buf);
548 case DW_EH_PE_udata4:
550 case DW_EH_PE_sdata4:
551 return (int32_t)read32(buf);
552 case DW_EH_PE_udata8:
553 case DW_EH_PE_sdata8:
555 case DW_EH_PE_absptr:
556 return readUint(buf);
558 fatal("unknown FDE size encoding");
561 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to.
562 // We need it to create .eh_frame_hdr section.
563 uint64_t EhFrameSection::getFdePc(uint8_t *buf, size_t fdeOff,
565 // The starting address to which this FDE applies is
566 // stored at FDE + 8 byte.
567 size_t off = fdeOff + 8;
568 uint64_t addr = readFdeAddr(buf + off, enc & 0xf);
569 if ((enc & 0x70) == DW_EH_PE_absptr)
571 if ((enc & 0x70) == DW_EH_PE_pcrel)
572 return addr + getParent()->addr + off;
573 fatal("unknown FDE size relative encoding");
576 void EhFrameSection::writeTo(uint8_t *buf) {
577 // Write CIE and FDE records.
578 for (CieRecord *rec : cieRecords) {
579 size_t cieOffset = rec->cie->outputOff;
580 writeCieFde(buf + cieOffset, rec->cie->data());
582 for (EhSectionPiece *fde : rec->fdes) {
583 size_t off = fde->outputOff;
584 writeCieFde(buf + off, fde->data());
586 // FDE's second word should have the offset to an associated CIE.
588 write32(buf + off + 4, off + 4 - cieOffset);
592 // Apply relocations. .eh_frame section contents are not contiguous
593 // in the output buffer, but relocateAlloc() still works because
594 // getOffset() takes care of discontiguous section pieces.
595 for (EhInputSection *s : sections)
596 s->relocateAlloc(buf, nullptr);
598 if (getPartition().ehFrameHdr && getPartition().ehFrameHdr->getParent())
599 getPartition().ehFrameHdr->write();
602 GotSection::GotSection()
603 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
605 // If ElfSym::globalOffsetTable is relative to .got and is referenced,
606 // increase numEntries by the number of entries used to emit
607 // ElfSym::globalOffsetTable.
608 if (ElfSym::globalOffsetTable && !target->gotBaseSymInGotPlt)
609 numEntries += target->gotHeaderEntriesNum;
612 void GotSection::addEntry(Symbol &sym) {
613 sym.gotIndex = numEntries;
617 bool GotSection::addDynTlsEntry(Symbol &sym) {
618 if (sym.globalDynIndex != -1U)
620 sym.globalDynIndex = numEntries;
621 // Global Dynamic TLS entries take two GOT slots.
626 // Reserves TLS entries for a TLS module ID and a TLS block offset.
627 // In total it takes two GOT slots.
628 bool GotSection::addTlsIndex() {
629 if (tlsIndexOff != uint32_t(-1))
631 tlsIndexOff = numEntries * config->wordsize;
636 uint64_t GotSection::getGlobalDynAddr(const Symbol &b) const {
637 return this->getVA() + b.globalDynIndex * config->wordsize;
640 uint64_t GotSection::getGlobalDynOffset(const Symbol &b) const {
641 return b.globalDynIndex * config->wordsize;
644 void GotSection::finalizeContents() {
645 size = numEntries * config->wordsize;
648 bool GotSection::isNeeded() const {
649 // We need to emit a GOT even if it's empty if there's a relocation that is
650 // relative to GOT(such as GOTOFFREL).
651 return numEntries || hasGotOffRel;
654 void GotSection::writeTo(uint8_t *buf) {
655 // Buf points to the start of this section's buffer,
656 // whereas InputSectionBase::relocateAlloc() expects its argument
657 // to point to the start of the output section.
658 target->writeGotHeader(buf);
659 relocateAlloc(buf - outSecOff, buf - outSecOff + size);
662 static uint64_t getMipsPageAddr(uint64_t addr) {
663 return (addr + 0x8000) & ~0xffff;
666 static uint64_t getMipsPageCount(uint64_t size) {
667 return (size + 0xfffe) / 0xffff + 1;
670 MipsGotSection::MipsGotSection()
671 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16,
674 void MipsGotSection::addEntry(InputFile &file, Symbol &sym, int64_t addend,
676 FileGot &g = getGot(file);
677 if (expr == R_MIPS_GOT_LOCAL_PAGE) {
678 if (const OutputSection *os = sym.getOutputSection())
679 g.pagesMap.insert({os, {}});
681 g.local16.insert({{nullptr, getMipsPageAddr(sym.getVA(addend))}, 0});
682 } else if (sym.isTls())
683 g.tls.insert({&sym, 0});
684 else if (sym.isPreemptible && expr == R_ABS)
685 g.relocs.insert({&sym, 0});
686 else if (sym.isPreemptible)
687 g.global.insert({&sym, 0});
688 else if (expr == R_MIPS_GOT_OFF32)
689 g.local32.insert({{&sym, addend}, 0});
691 g.local16.insert({{&sym, addend}, 0});
694 void MipsGotSection::addDynTlsEntry(InputFile &file, Symbol &sym) {
695 getGot(file).dynTlsSymbols.insert({&sym, 0});
698 void MipsGotSection::addTlsIndex(InputFile &file) {
699 getGot(file).dynTlsSymbols.insert({nullptr, 0});
702 size_t MipsGotSection::FileGot::getEntriesNum() const {
703 return getPageEntriesNum() + local16.size() + global.size() + relocs.size() +
704 tls.size() + dynTlsSymbols.size() * 2;
707 size_t MipsGotSection::FileGot::getPageEntriesNum() const {
709 for (const std::pair<const OutputSection *, FileGot::PageBlock> &p : pagesMap)
710 num += p.second.count;
714 size_t MipsGotSection::FileGot::getIndexedEntriesNum() const {
715 size_t count = getPageEntriesNum() + local16.size() + global.size();
716 // If there are relocation-only entries in the GOT, TLS entries
717 // are allocated after them. TLS entries should be addressable
718 // by 16-bit index so count both reloc-only and TLS entries.
719 if (!tls.empty() || !dynTlsSymbols.empty())
720 count += relocs.size() + tls.size() + dynTlsSymbols.size() * 2;
724 MipsGotSection::FileGot &MipsGotSection::getGot(InputFile &f) {
725 if (!f.mipsGotIndex.hasValue()) {
727 gots.back().file = &f;
728 f.mipsGotIndex = gots.size() - 1;
730 return gots[*f.mipsGotIndex];
733 uint64_t MipsGotSection::getPageEntryOffset(const InputFile *f,
735 int64_t addend) const {
736 const FileGot &g = gots[*f->mipsGotIndex];
738 if (const OutputSection *outSec = sym.getOutputSection()) {
739 uint64_t secAddr = getMipsPageAddr(outSec->addr);
740 uint64_t symAddr = getMipsPageAddr(sym.getVA(addend));
741 index = g.pagesMap.lookup(outSec).firstIndex + (symAddr - secAddr) / 0xffff;
743 index = g.local16.lookup({nullptr, getMipsPageAddr(sym.getVA(addend))});
745 return index * config->wordsize;
748 uint64_t MipsGotSection::getSymEntryOffset(const InputFile *f, const Symbol &s,
749 int64_t addend) const {
750 const FileGot &g = gots[*f->mipsGotIndex];
751 Symbol *sym = const_cast<Symbol *>(&s);
753 return g.tls.lookup(sym) * config->wordsize;
754 if (sym->isPreemptible)
755 return g.global.lookup(sym) * config->wordsize;
756 return g.local16.lookup({sym, addend}) * config->wordsize;
759 uint64_t MipsGotSection::getTlsIndexOffset(const InputFile *f) const {
760 const FileGot &g = gots[*f->mipsGotIndex];
761 return g.dynTlsSymbols.lookup(nullptr) * config->wordsize;
764 uint64_t MipsGotSection::getGlobalDynOffset(const InputFile *f,
765 const Symbol &s) const {
766 const FileGot &g = gots[*f->mipsGotIndex];
767 Symbol *sym = const_cast<Symbol *>(&s);
768 return g.dynTlsSymbols.lookup(sym) * config->wordsize;
771 const Symbol *MipsGotSection::getFirstGlobalEntry() const {
774 const FileGot &primGot = gots.front();
775 if (!primGot.global.empty())
776 return primGot.global.front().first;
777 if (!primGot.relocs.empty())
778 return primGot.relocs.front().first;
782 unsigned MipsGotSection::getLocalEntriesNum() const {
784 return headerEntriesNum;
785 return headerEntriesNum + gots.front().getPageEntriesNum() +
786 gots.front().local16.size();
789 bool MipsGotSection::tryMergeGots(FileGot &dst, FileGot &src, bool isPrimary) {
791 set_union(tmp.pagesMap, src.pagesMap);
792 set_union(tmp.local16, src.local16);
793 set_union(tmp.global, src.global);
794 set_union(tmp.relocs, src.relocs);
795 set_union(tmp.tls, src.tls);
796 set_union(tmp.dynTlsSymbols, src.dynTlsSymbols);
798 size_t count = isPrimary ? headerEntriesNum : 0;
799 count += tmp.getIndexedEntriesNum();
801 if (count * config->wordsize > config->mipsGotSize)
808 void MipsGotSection::finalizeContents() { updateAllocSize(); }
810 bool MipsGotSection::updateAllocSize() {
811 size = headerEntriesNum * config->wordsize;
812 for (const FileGot &g : gots)
813 size += g.getEntriesNum() * config->wordsize;
817 void MipsGotSection::build() {
821 std::vector<FileGot> mergedGots(1);
823 // For each GOT move non-preemptible symbols from the `Global`
824 // to `Local16` list. Preemptible symbol might become non-preemptible
825 // one if, for example, it gets a related copy relocation.
826 for (FileGot &got : gots) {
827 for (auto &p: got.global)
828 if (!p.first->isPreemptible)
829 got.local16.insert({{p.first, 0}, 0});
830 got.global.remove_if([&](const std::pair<Symbol *, size_t> &p) {
831 return !p.first->isPreemptible;
835 // For each GOT remove "reloc-only" entry if there is "global"
836 // entry for the same symbol. And add local entries which indexed
837 // using 32-bit value at the end of 16-bit entries.
838 for (FileGot &got : gots) {
839 got.relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
840 return got.global.count(p.first);
842 set_union(got.local16, got.local32);
846 // Evaluate number of "reloc-only" entries in the resulting GOT.
847 // To do that put all unique "reloc-only" and "global" entries
848 // from all GOTs to the future primary GOT.
849 FileGot *primGot = &mergedGots.front();
850 for (FileGot &got : gots) {
851 set_union(primGot->relocs, got.global);
852 set_union(primGot->relocs, got.relocs);
856 // Evaluate number of "page" entries in each GOT.
857 for (FileGot &got : gots) {
858 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
860 const OutputSection *os = p.first;
861 uint64_t secSize = 0;
862 for (BaseCommand *cmd : os->sectionCommands) {
863 if (auto *isd = dyn_cast<InputSectionDescription>(cmd))
864 for (InputSection *isec : isd->sections) {
865 uint64_t off = alignTo(secSize, isec->alignment);
866 secSize = off + isec->getSize();
869 p.second.count = getMipsPageCount(secSize);
873 // Merge GOTs. Try to join as much as possible GOTs but do not exceed
874 // maximum GOT size. At first, try to fill the primary GOT because
875 // the primary GOT can be accessed in the most effective way. If it
876 // is not possible, try to fill the last GOT in the list, and finally
877 // create a new GOT if both attempts failed.
878 for (FileGot &srcGot : gots) {
879 InputFile *file = srcGot.file;
880 if (tryMergeGots(mergedGots.front(), srcGot, true)) {
881 file->mipsGotIndex = 0;
883 // If this is the first time we failed to merge with the primary GOT,
884 // MergedGots.back() will also be the primary GOT. We must make sure not
885 // to try to merge again with isPrimary=false, as otherwise, if the
886 // inputs are just right, we could allow the primary GOT to become 1 or 2
887 // words bigger due to ignoring the header size.
888 if (mergedGots.size() == 1 ||
889 !tryMergeGots(mergedGots.back(), srcGot, false)) {
890 mergedGots.emplace_back();
891 std::swap(mergedGots.back(), srcGot);
893 file->mipsGotIndex = mergedGots.size() - 1;
896 std::swap(gots, mergedGots);
898 // Reduce number of "reloc-only" entries in the primary GOT
899 // by subtracting "global" entries in the primary GOT.
900 primGot = &gots.front();
901 primGot->relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
902 return primGot->global.count(p.first);
905 // Calculate indexes for each GOT entry.
906 size_t index = headerEntriesNum;
907 for (FileGot &got : gots) {
908 got.startIndex = &got == primGot ? 0 : index;
909 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
911 // For each output section referenced by GOT page relocations calculate
912 // and save into pagesMap an upper bound of MIPS GOT entries required
913 // to store page addresses of local symbols. We assume the worst case -
914 // each 64kb page of the output section has at least one GOT relocation
915 // against it. And take in account the case when the section intersects
917 p.second.firstIndex = index;
918 index += p.second.count;
920 for (auto &p: got.local16)
922 for (auto &p: got.global)
924 for (auto &p: got.relocs)
926 for (auto &p: got.tls)
928 for (auto &p: got.dynTlsSymbols) {
934 // Update Symbol::gotIndex field to use this
935 // value later in the `sortMipsSymbols` function.
936 for (auto &p : primGot->global)
937 p.first->gotIndex = p.second;
938 for (auto &p : primGot->relocs)
939 p.first->gotIndex = p.second;
941 // Create dynamic relocations.
942 for (FileGot &got : gots) {
943 // Create dynamic relocations for TLS entries.
944 for (std::pair<Symbol *, size_t> &p : got.tls) {
946 uint64_t offset = p.second * config->wordsize;
947 if (s->isPreemptible)
948 mainPart->relaDyn->addReloc(target->tlsGotRel, this, offset, s);
950 for (std::pair<Symbol *, size_t> &p : got.dynTlsSymbols) {
952 uint64_t offset = p.second * config->wordsize;
956 mainPart->relaDyn->addReloc(target->tlsModuleIndexRel, this, offset, s);
958 // When building a shared library we still need a dynamic relocation
959 // for the module index. Therefore only checking for
960 // S->isPreemptible is not sufficient (this happens e.g. for
961 // thread-locals that have been marked as local through a linker script)
962 if (!s->isPreemptible && !config->isPic)
964 mainPart->relaDyn->addReloc(target->tlsModuleIndexRel, this, offset, s);
965 // However, we can skip writing the TLS offset reloc for non-preemptible
966 // symbols since it is known even in shared libraries
967 if (!s->isPreemptible)
969 offset += config->wordsize;
970 mainPart->relaDyn->addReloc(target->tlsOffsetRel, this, offset, s);
974 // Do not create dynamic relocations for non-TLS
975 // entries in the primary GOT.
979 // Dynamic relocations for "global" entries.
980 for (const std::pair<Symbol *, size_t> &p : got.global) {
981 uint64_t offset = p.second * config->wordsize;
982 mainPart->relaDyn->addReloc(target->relativeRel, this, offset, p.first);
986 // Dynamic relocations for "local" entries in case of PIC.
987 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
989 size_t pageCount = l.second.count;
990 for (size_t pi = 0; pi < pageCount; ++pi) {
991 uint64_t offset = (l.second.firstIndex + pi) * config->wordsize;
992 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, l.first,
993 int64_t(pi * 0x10000)});
996 for (const std::pair<GotEntry, size_t> &p : got.local16) {
997 uint64_t offset = p.second * config->wordsize;
998 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, true,
999 p.first.first, p.first.second});
1004 bool MipsGotSection::isNeeded() const {
1005 // We add the .got section to the result for dynamic MIPS target because
1006 // its address and properties are mentioned in the .dynamic section.
1007 return !config->relocatable;
1010 uint64_t MipsGotSection::getGp(const InputFile *f) const {
1011 // For files without related GOT or files refer a primary GOT
1012 // returns "common" _gp value. For secondary GOTs calculate
1013 // individual _gp values.
1014 if (!f || !f->mipsGotIndex.hasValue() || *f->mipsGotIndex == 0)
1015 return ElfSym::mipsGp->getVA(0);
1016 return getVA() + gots[*f->mipsGotIndex].startIndex * config->wordsize +
1020 void MipsGotSection::writeTo(uint8_t *buf) {
1021 // Set the MSB of the second GOT slot. This is not required by any
1022 // MIPS ABI documentation, though.
1024 // There is a comment in glibc saying that "The MSB of got[1] of a
1025 // gnu object is set to identify gnu objects," and in GNU gold it
1026 // says "the second entry will be used by some runtime loaders".
1027 // But how this field is being used is unclear.
1029 // We are not really willing to mimic other linkers behaviors
1030 // without understanding why they do that, but because all files
1031 // generated by GNU tools have this special GOT value, and because
1032 // we've been doing this for years, it is probably a safe bet to
1033 // keep doing this for now. We really need to revisit this to see
1034 // if we had to do this.
1035 writeUint(buf + config->wordsize, (uint64_t)1 << (config->wordsize * 8 - 1));
1036 for (const FileGot &g : gots) {
1037 auto write = [&](size_t i, const Symbol *s, int64_t a) {
1041 writeUint(buf + i * config->wordsize, va);
1043 // Write 'page address' entries to the local part of the GOT.
1044 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
1046 size_t pageCount = l.second.count;
1047 uint64_t firstPageAddr = getMipsPageAddr(l.first->addr);
1048 for (size_t pi = 0; pi < pageCount; ++pi)
1049 write(l.second.firstIndex + pi, nullptr, firstPageAddr + pi * 0x10000);
1051 // Local, global, TLS, reloc-only entries.
1052 // If TLS entry has a corresponding dynamic relocations, leave it
1053 // initialized by zero. Write down adjusted TLS symbol's values otherwise.
1054 // To calculate the adjustments use offsets for thread-local storage.
1055 // https://www.linux-mips.org/wiki/NPTL
1056 for (const std::pair<GotEntry, size_t> &p : g.local16)
1057 write(p.second, p.first.first, p.first.second);
1058 // Write VA to the primary GOT only. For secondary GOTs that
1059 // will be done by REL32 dynamic relocations.
1060 if (&g == &gots.front())
1061 for (const std::pair<Symbol *, size_t> &p : g.global)
1062 write(p.second, p.first, 0);
1063 for (const std::pair<Symbol *, size_t> &p : g.relocs)
1064 write(p.second, p.first, 0);
1065 for (const std::pair<Symbol *, size_t> &p : g.tls)
1066 write(p.second, p.first, p.first->isPreemptible ? 0 : -0x7000);
1067 for (const std::pair<Symbol *, size_t> &p : g.dynTlsSymbols) {
1068 if (p.first == nullptr && !config->isPic)
1069 write(p.second, nullptr, 1);
1070 else if (p.first && !p.first->isPreemptible) {
1071 // If we are emitting PIC code with relocations we mustn't write
1072 // anything to the GOT here. When using Elf_Rel relocations the value
1073 // one will be treated as an addend and will cause crashes at runtime
1075 write(p.second, nullptr, 1);
1076 write(p.second + 1, p.first, -0x8000);
1082 // On PowerPC the .plt section is used to hold the table of function addresses
1083 // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss
1084 // section. I don't know why we have a BSS style type for the section but it is
1085 // consistent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI.
1086 GotPltSection::GotPltSection()
1087 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
1089 if (config->emachine == EM_PPC) {
1091 } else if (config->emachine == EM_PPC64) {
1097 void GotPltSection::addEntry(Symbol &sym) {
1098 assert(sym.pltIndex == entries.size());
1099 entries.push_back(&sym);
1102 size_t GotPltSection::getSize() const {
1103 return (target->gotPltHeaderEntriesNum + entries.size()) * config->wordsize;
1106 void GotPltSection::writeTo(uint8_t *buf) {
1107 target->writeGotPltHeader(buf);
1108 buf += target->gotPltHeaderEntriesNum * config->wordsize;
1109 for (const Symbol *b : entries) {
1110 target->writeGotPlt(buf, *b);
1111 buf += config->wordsize;
1115 bool GotPltSection::isNeeded() const {
1116 // We need to emit GOTPLT even if it's empty if there's a relocation relative
1118 return !entries.empty() || hasGotPltOffRel;
1121 static StringRef getIgotPltName() {
1122 // On ARM the IgotPltSection is part of the GotSection.
1123 if (config->emachine == EM_ARM)
1126 // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection
1127 // needs to be named the same.
1128 if (config->emachine == EM_PPC64)
1134 // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit
1135 // with the IgotPltSection.
1136 IgotPltSection::IgotPltSection()
1137 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
1138 config->emachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS,
1139 config->wordsize, getIgotPltName()) {}
1141 void IgotPltSection::addEntry(Symbol &sym) {
1142 assert(sym.pltIndex == entries.size());
1143 entries.push_back(&sym);
1146 size_t IgotPltSection::getSize() const {
1147 return entries.size() * config->wordsize;
1150 void IgotPltSection::writeTo(uint8_t *buf) {
1151 for (const Symbol *b : entries) {
1152 target->writeIgotPlt(buf, *b);
1153 buf += config->wordsize;
1157 StringTableSection::StringTableSection(StringRef name, bool dynamic)
1158 : SyntheticSection(dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, name),
1160 // ELF string tables start with a NUL byte.
1164 // Adds a string to the string table. If `hashIt` is true we hash and check for
1165 // duplicates. It is optional because the name of global symbols are already
1166 // uniqued and hashing them again has a big cost for a small value: uniquing
1167 // them with some other string that happens to be the same.
1168 unsigned StringTableSection::addString(StringRef s, bool hashIt) {
1170 auto r = stringMap.insert(std::make_pair(s, this->size));
1172 return r.first->second;
1174 unsigned ret = this->size;
1175 this->size = this->size + s.size() + 1;
1176 strings.push_back(s);
1180 void StringTableSection::writeTo(uint8_t *buf) {
1181 for (StringRef s : strings) {
1182 memcpy(buf, s.data(), s.size());
1183 buf[s.size()] = '\0';
1184 buf += s.size() + 1;
1188 // Returns the number of entries in .gnu.version_d: the number of
1189 // non-VER_NDX_LOCAL-non-VER_NDX_GLOBAL definitions, plus 1.
1190 // Note that we don't support vd_cnt > 1 yet.
1191 static unsigned getVerDefNum() {
1192 return namedVersionDefs().size() + 1;
1195 template <class ELFT>
1196 DynamicSection<ELFT>::DynamicSection()
1197 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, config->wordsize,
1199 this->entsize = ELFT::Is64Bits ? 16 : 8;
1201 // .dynamic section is not writable on MIPS and on Fuchsia OS
1202 // which passes -z rodynamic.
1203 // See "Special Section" in Chapter 4 in the following document:
1204 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1205 if (config->emachine == EM_MIPS || config->zRodynamic)
1206 this->flags = SHF_ALLOC;
1209 template <class ELFT>
1210 void DynamicSection<ELFT>::add(int32_t tag, std::function<uint64_t()> fn) {
1211 entries.push_back({tag, fn});
1214 template <class ELFT>
1215 void DynamicSection<ELFT>::addInt(int32_t tag, uint64_t val) {
1216 entries.push_back({tag, [=] { return val; }});
1219 template <class ELFT>
1220 void DynamicSection<ELFT>::addInSec(int32_t tag, InputSection *sec) {
1221 entries.push_back({tag, [=] { return sec->getVA(0); }});
1224 template <class ELFT>
1225 void DynamicSection<ELFT>::addInSecRelative(int32_t tag, InputSection *sec) {
1226 size_t tagOffset = entries.size() * entsize;
1228 {tag, [=] { return sec->getVA(0) - (getVA() + tagOffset); }});
1231 template <class ELFT>
1232 void DynamicSection<ELFT>::addOutSec(int32_t tag, OutputSection *sec) {
1233 entries.push_back({tag, [=] { return sec->addr; }});
1236 template <class ELFT>
1237 void DynamicSection<ELFT>::addSize(int32_t tag, OutputSection *sec) {
1238 entries.push_back({tag, [=] { return sec->size; }});
1241 template <class ELFT>
1242 void DynamicSection<ELFT>::addSym(int32_t tag, Symbol *sym) {
1243 entries.push_back({tag, [=] { return sym->getVA(); }});
1246 // The output section .rela.dyn may include these synthetic sections:
1249 // - in.relaIplt: this is included if in.relaIplt is named .rela.dyn
1250 // - in.relaPlt: this is included if a linker script places .rela.plt inside
1253 // DT_RELASZ is the total size of the included sections.
1254 static std::function<uint64_t()> addRelaSz(RelocationBaseSection *relaDyn) {
1256 size_t size = relaDyn->getSize();
1257 if (in.relaIplt->getParent() == relaDyn->getParent())
1258 size += in.relaIplt->getSize();
1259 if (in.relaPlt->getParent() == relaDyn->getParent())
1260 size += in.relaPlt->getSize();
1265 // A Linker script may assign the RELA relocation sections to the same
1266 // output section. When this occurs we cannot just use the OutputSection
1267 // Size. Moreover the [DT_JMPREL, DT_JMPREL + DT_PLTRELSZ) is permitted to
1268 // overlap with the [DT_RELA, DT_RELA + DT_RELASZ).
1269 static uint64_t addPltRelSz() {
1270 size_t size = in.relaPlt->getSize();
1271 if (in.relaIplt->getParent() == in.relaPlt->getParent() &&
1272 in.relaIplt->name == in.relaPlt->name)
1273 size += in.relaIplt->getSize();
1277 // Add remaining entries to complete .dynamic contents.
1278 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() {
1279 elf::Partition &part = getPartition();
1280 bool isMain = part.name.empty();
1282 for (StringRef s : config->filterList)
1283 addInt(DT_FILTER, part.dynStrTab->addString(s));
1284 for (StringRef s : config->auxiliaryList)
1285 addInt(DT_AUXILIARY, part.dynStrTab->addString(s));
1287 if (!config->rpath.empty())
1288 addInt(config->enableNewDtags ? DT_RUNPATH : DT_RPATH,
1289 part.dynStrTab->addString(config->rpath));
1291 for (SharedFile *file : sharedFiles)
1293 addInt(DT_NEEDED, part.dynStrTab->addString(file->soName));
1296 if (!config->soName.empty())
1297 addInt(DT_SONAME, part.dynStrTab->addString(config->soName));
1299 if (!config->soName.empty())
1300 addInt(DT_NEEDED, part.dynStrTab->addString(config->soName));
1301 addInt(DT_SONAME, part.dynStrTab->addString(part.name));
1304 // Set DT_FLAGS and DT_FLAGS_1.
1305 uint32_t dtFlags = 0;
1306 uint32_t dtFlags1 = 0;
1307 if (config->bsymbolic)
1308 dtFlags |= DF_SYMBOLIC;
1309 if (config->zGlobal)
1310 dtFlags1 |= DF_1_GLOBAL;
1311 if (config->zInitfirst)
1312 dtFlags1 |= DF_1_INITFIRST;
1313 if (config->zInterpose)
1314 dtFlags1 |= DF_1_INTERPOSE;
1315 if (config->zNodefaultlib)
1316 dtFlags1 |= DF_1_NODEFLIB;
1317 if (config->zNodelete)
1318 dtFlags1 |= DF_1_NODELETE;
1319 if (config->zNodlopen)
1320 dtFlags1 |= DF_1_NOOPEN;
1322 dtFlags1 |= DF_1_PIE;
1324 dtFlags |= DF_BIND_NOW;
1325 dtFlags1 |= DF_1_NOW;
1327 if (config->zOrigin) {
1328 dtFlags |= DF_ORIGIN;
1329 dtFlags1 |= DF_1_ORIGIN;
1332 dtFlags |= DF_TEXTREL;
1333 if (config->hasStaticTlsModel)
1334 dtFlags |= DF_STATIC_TLS;
1337 addInt(DT_FLAGS, dtFlags);
1339 addInt(DT_FLAGS_1, dtFlags1);
1341 // DT_DEBUG is a pointer to debug information used by debuggers at runtime. We
1342 // need it for each process, so we don't write it for DSOs. The loader writes
1343 // the pointer into this entry.
1345 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some
1346 // systems (currently only Fuchsia OS) provide other means to give the
1347 // debugger this information. Such systems may choose make .dynamic read-only.
1348 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG.
1349 if (!config->shared && !config->relocatable && !config->zRodynamic)
1350 addInt(DT_DEBUG, 0);
1352 if (OutputSection *sec = part.dynStrTab->getParent())
1353 this->link = sec->sectionIndex;
1355 if (part.relaDyn->isNeeded() ||
1356 (in.relaIplt->isNeeded() &&
1357 part.relaDyn->getParent() == in.relaIplt->getParent())) {
1358 addInSec(part.relaDyn->dynamicTag, part.relaDyn);
1359 entries.push_back({part.relaDyn->sizeDynamicTag, addRelaSz(part.relaDyn)});
1361 bool isRela = config->isRela;
1362 addInt(isRela ? DT_RELAENT : DT_RELENT,
1363 isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel));
1365 // MIPS dynamic loader does not support RELCOUNT tag.
1366 // The problem is in the tight relation between dynamic
1367 // relocations and GOT. So do not emit this tag on MIPS.
1368 if (config->emachine != EM_MIPS) {
1369 size_t numRelativeRels = part.relaDyn->getRelativeRelocCount();
1370 if (config->zCombreloc && numRelativeRels)
1371 addInt(isRela ? DT_RELACOUNT : DT_RELCOUNT, numRelativeRels);
1374 if (part.relrDyn && !part.relrDyn->relocs.empty()) {
1375 addInSec(config->useAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR,
1377 addSize(config->useAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ,
1378 part.relrDyn->getParent());
1379 addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT,
1382 // .rel[a].plt section usually consists of two parts, containing plt and
1383 // iplt relocations. It is possible to have only iplt relocations in the
1384 // output. In that case relaPlt is empty and have zero offset, the same offset
1385 // as relaIplt has. And we still want to emit proper dynamic tags for that
1386 // case, so here we always use relaPlt as marker for the beginning of
1387 // .rel[a].plt section.
1388 if (isMain && (in.relaPlt->isNeeded() || in.relaIplt->isNeeded())) {
1389 addInSec(DT_JMPREL, in.relaPlt);
1390 entries.push_back({DT_PLTRELSZ, addPltRelSz});
1391 switch (config->emachine) {
1393 addInSec(DT_MIPS_PLTGOT, in.gotPlt);
1396 addInSec(DT_PLTGOT, in.plt);
1399 addInSec(DT_PLTGOT, in.gotPlt);
1402 addInt(DT_PLTREL, config->isRela ? DT_RELA : DT_REL);
1405 if (config->emachine == EM_AARCH64) {
1406 if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI)
1407 addInt(DT_AARCH64_BTI_PLT, 0);
1408 if (config->zPacPlt)
1409 addInt(DT_AARCH64_PAC_PLT, 0);
1412 addInSec(DT_SYMTAB, part.dynSymTab);
1413 addInt(DT_SYMENT, sizeof(Elf_Sym));
1414 addInSec(DT_STRTAB, part.dynStrTab);
1415 addInt(DT_STRSZ, part.dynStrTab->getSize());
1417 addInt(DT_TEXTREL, 0);
1418 if (part.gnuHashTab)
1419 addInSec(DT_GNU_HASH, part.gnuHashTab);
1421 addInSec(DT_HASH, part.hashTab);
1424 if (Out::preinitArray) {
1425 addOutSec(DT_PREINIT_ARRAY, Out::preinitArray);
1426 addSize(DT_PREINIT_ARRAYSZ, Out::preinitArray);
1428 if (Out::initArray) {
1429 addOutSec(DT_INIT_ARRAY, Out::initArray);
1430 addSize(DT_INIT_ARRAYSZ, Out::initArray);
1432 if (Out::finiArray) {
1433 addOutSec(DT_FINI_ARRAY, Out::finiArray);
1434 addSize(DT_FINI_ARRAYSZ, Out::finiArray);
1437 if (Symbol *b = symtab->find(config->init))
1440 if (Symbol *b = symtab->find(config->fini))
1445 if (part.verSym && part.verSym->isNeeded())
1446 addInSec(DT_VERSYM, part.verSym);
1447 if (part.verDef && part.verDef->isLive()) {
1448 addInSec(DT_VERDEF, part.verDef);
1449 addInt(DT_VERDEFNUM, getVerDefNum());
1451 if (part.verNeed && part.verNeed->isNeeded()) {
1452 addInSec(DT_VERNEED, part.verNeed);
1453 unsigned needNum = 0;
1454 for (SharedFile *f : sharedFiles)
1455 if (!f->vernauxs.empty())
1457 addInt(DT_VERNEEDNUM, needNum);
1460 if (config->emachine == EM_MIPS) {
1461 addInt(DT_MIPS_RLD_VERSION, 1);
1462 addInt(DT_MIPS_FLAGS, RHF_NOTPOT);
1463 addInt(DT_MIPS_BASE_ADDRESS, target->getImageBase());
1464 addInt(DT_MIPS_SYMTABNO, part.dynSymTab->getNumSymbols());
1466 add(DT_MIPS_LOCAL_GOTNO, [] { return in.mipsGot->getLocalEntriesNum(); });
1468 if (const Symbol *b = in.mipsGot->getFirstGlobalEntry())
1469 addInt(DT_MIPS_GOTSYM, b->dynsymIndex);
1471 addInt(DT_MIPS_GOTSYM, part.dynSymTab->getNumSymbols());
1472 addInSec(DT_PLTGOT, in.mipsGot);
1473 if (in.mipsRldMap) {
1475 addInSec(DT_MIPS_RLD_MAP, in.mipsRldMap);
1476 // Store the offset to the .rld_map section
1477 // relative to the address of the tag.
1478 addInSecRelative(DT_MIPS_RLD_MAP_REL, in.mipsRldMap);
1482 // DT_PPC_GOT indicates to glibc Secure PLT is used. If DT_PPC_GOT is absent,
1483 // glibc assumes the old-style BSS PLT layout which we don't support.
1484 if (config->emachine == EM_PPC)
1485 add(DT_PPC_GOT, [] { return in.got->getVA(); });
1487 // Glink dynamic tag is required by the V2 abi if the plt section isn't empty.
1488 if (config->emachine == EM_PPC64 && in.plt->isNeeded()) {
1489 // The Glink tag points to 32 bytes before the first lazy symbol resolution
1490 // stub, which starts directly after the header.
1491 entries.push_back({DT_PPC64_GLINK, [=] {
1492 unsigned offset = target->pltHeaderSize - 32;
1493 return in.plt->getVA(0) + offset;
1499 getParent()->link = this->link;
1500 this->size = entries.size() * this->entsize;
1503 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *buf) {
1504 auto *p = reinterpret_cast<Elf_Dyn *>(buf);
1506 for (std::pair<int32_t, std::function<uint64_t()>> &kv : entries) {
1507 p->d_tag = kv.first;
1508 p->d_un.d_val = kv.second();
1513 uint64_t DynamicReloc::getOffset() const {
1514 return inputSec->getVA(offsetInSec);
1517 int64_t DynamicReloc::computeAddend() const {
1519 return sym->getVA(addend);
1522 // See the comment in the DynamicReloc ctor.
1523 return getMipsPageAddr(outputSec->addr) + addend;
1526 uint32_t DynamicReloc::getSymIndex(SymbolTableBaseSection *symTab) const {
1527 if (sym && !useSymVA)
1528 return symTab->getSymbolIndex(sym);
1532 RelocationBaseSection::RelocationBaseSection(StringRef name, uint32_t type,
1534 int32_t sizeDynamicTag)
1535 : SyntheticSection(SHF_ALLOC, type, config->wordsize, name),
1536 dynamicTag(dynamicTag), sizeDynamicTag(sizeDynamicTag) {}
1538 void RelocationBaseSection::addReloc(RelType dynType, InputSectionBase *isec,
1539 uint64_t offsetInSec, Symbol *sym) {
1540 addReloc({dynType, isec, offsetInSec, false, sym, 0});
1543 void RelocationBaseSection::addReloc(RelType dynType,
1544 InputSectionBase *inputSec,
1545 uint64_t offsetInSec, Symbol *sym,
1546 int64_t addend, RelExpr expr,
1548 // Write the addends to the relocated address if required. We skip
1549 // it if the written value would be zero.
1550 if (config->writeAddends && (expr != R_ADDEND || addend != 0))
1551 inputSec->relocations.push_back({expr, type, offsetInSec, addend, sym});
1552 addReloc({dynType, inputSec, offsetInSec, expr != R_ADDEND, sym, addend});
1555 void RelocationBaseSection::addReloc(const DynamicReloc &reloc) {
1556 if (reloc.type == target->relativeRel)
1557 ++numRelativeRelocs;
1558 relocs.push_back(reloc);
1561 void RelocationBaseSection::finalizeContents() {
1562 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1564 // When linking glibc statically, .rel{,a}.plt contains R_*_IRELATIVE
1565 // relocations due to IFUNC (e.g. strcpy). sh_link will be set to 0 in that
1567 if (symTab && symTab->getParent())
1568 getParent()->link = symTab->getParent()->sectionIndex;
1570 getParent()->link = 0;
1572 if (in.relaPlt == this)
1573 getParent()->info = in.gotPlt->getParent()->sectionIndex;
1574 if (in.relaIplt == this)
1575 getParent()->info = in.igotPlt->getParent()->sectionIndex;
1578 RelrBaseSection::RelrBaseSection()
1579 : SyntheticSection(SHF_ALLOC,
1580 config->useAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR,
1581 config->wordsize, ".relr.dyn") {}
1583 template <class ELFT>
1584 static void encodeDynamicReloc(SymbolTableBaseSection *symTab,
1585 typename ELFT::Rela *p,
1586 const DynamicReloc &rel) {
1588 p->r_addend = rel.computeAddend();
1589 p->r_offset = rel.getOffset();
1590 p->setSymbolAndType(rel.getSymIndex(symTab), rel.type, config->isMips64EL);
1593 template <class ELFT>
1594 RelocationSection<ELFT>::RelocationSection(StringRef name, bool sort)
1595 : RelocationBaseSection(name, config->isRela ? SHT_RELA : SHT_REL,
1596 config->isRela ? DT_RELA : DT_REL,
1597 config->isRela ? DT_RELASZ : DT_RELSZ),
1599 this->entsize = config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1602 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *buf) {
1603 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1605 // Sort by (!IsRelative,SymIndex,r_offset). DT_REL[A]COUNT requires us to
1606 // place R_*_RELATIVE first. SymIndex is to improve locality, while r_offset
1607 // is to make results easier to read.
1610 relocs, [&](const DynamicReloc &a, const DynamicReloc &b) {
1611 return std::make_tuple(a.type != target->relativeRel,
1612 a.getSymIndex(symTab), a.getOffset()) <
1613 std::make_tuple(b.type != target->relativeRel,
1614 b.getSymIndex(symTab), b.getOffset());
1617 for (const DynamicReloc &rel : relocs) {
1618 encodeDynamicReloc<ELFT>(symTab, reinterpret_cast<Elf_Rela *>(buf), rel);
1619 buf += config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1623 template <class ELFT>
1624 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection(
1626 : RelocationBaseSection(
1627 name, config->isRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL,
1628 config->isRela ? DT_ANDROID_RELA : DT_ANDROID_REL,
1629 config->isRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ) {
1633 template <class ELFT>
1634 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() {
1635 // This function computes the contents of an Android-format packed relocation
1638 // This format compresses relocations by using relocation groups to factor out
1639 // fields that are common between relocations and storing deltas from previous
1640 // relocations in SLEB128 format (which has a short representation for small
1641 // numbers). A good example of a relocation type with common fields is
1642 // R_*_RELATIVE, which is normally used to represent function pointers in
1643 // vtables. In the REL format, each relative relocation has the same r_info
1644 // field, and is only different from other relative relocations in terms of
1645 // the r_offset field. By sorting relocations by offset, grouping them by
1646 // r_info and representing each relocation with only the delta from the
1647 // previous offset, each 8-byte relocation can be compressed to as little as 1
1648 // byte (or less with run-length encoding). This relocation packer was able to
1649 // reduce the size of the relocation section in an Android Chromium DSO from
1650 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size.
1652 // A relocation section consists of a header containing the literal bytes
1653 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two
1654 // elements are the total number of relocations in the section and an initial
1655 // r_offset value. The remaining elements define a sequence of relocation
1656 // groups. Each relocation group starts with a header consisting of the
1657 // following elements:
1659 // - the number of relocations in the relocation group
1660 // - flags for the relocation group
1661 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta
1662 // for each relocation in the group.
1663 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info
1664 // field for each relocation in the group.
1665 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and
1666 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for
1667 // each relocation in the group.
1669 // Following the relocation group header are descriptions of each of the
1670 // relocations in the group. They consist of the following elements:
1672 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset
1673 // delta for this relocation.
1674 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info
1675 // field for this relocation.
1676 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and
1677 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for
1680 size_t oldSize = relocData.size();
1682 relocData = {'A', 'P', 'S', '2'};
1683 raw_svector_ostream os(relocData);
1684 auto add = [&](int64_t v) { encodeSLEB128(v, os); };
1686 // The format header includes the number of relocations and the initial
1687 // offset (we set this to zero because the first relocation group will
1688 // perform the initial adjustment).
1692 std::vector<Elf_Rela> relatives, nonRelatives;
1694 for (const DynamicReloc &rel : relocs) {
1696 encodeDynamicReloc<ELFT>(getPartition().dynSymTab, &r, rel);
1698 if (r.getType(config->isMips64EL) == target->relativeRel)
1699 relatives.push_back(r);
1701 nonRelatives.push_back(r);
1704 llvm::sort(relatives, [](const Elf_Rel &a, const Elf_Rel &b) {
1705 return a.r_offset < b.r_offset;
1708 // Try to find groups of relative relocations which are spaced one word
1709 // apart from one another. These generally correspond to vtable entries. The
1710 // format allows these groups to be encoded using a sort of run-length
1711 // encoding, but each group will cost 7 bytes in addition to the offset from
1712 // the previous group, so it is only profitable to do this for groups of
1713 // size 8 or larger.
1714 std::vector<Elf_Rela> ungroupedRelatives;
1715 std::vector<std::vector<Elf_Rela>> relativeGroups;
1716 for (auto i = relatives.begin(), e = relatives.end(); i != e;) {
1717 std::vector<Elf_Rela> group;
1719 group.push_back(*i++);
1720 } while (i != e && (i - 1)->r_offset + config->wordsize == i->r_offset);
1722 if (group.size() < 8)
1723 ungroupedRelatives.insert(ungroupedRelatives.end(), group.begin(),
1726 relativeGroups.emplace_back(std::move(group));
1729 // For non-relative relocations, we would like to:
1730 // 1. Have relocations with the same symbol offset to be consecutive, so
1731 // that the runtime linker can speed-up symbol lookup by implementing an
1733 // 2. Group relocations by r_info to reduce the size of the relocation
1735 // Since the symbol offset is the high bits in r_info, sorting by r_info
1736 // allows us to do both.
1738 // For Rela, we also want to sort by r_addend when r_info is the same. This
1739 // enables us to group by r_addend as well.
1740 llvm::stable_sort(nonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1741 if (a.r_info != b.r_info)
1742 return a.r_info < b.r_info;
1744 return a.r_addend < b.r_addend;
1748 // Group relocations with the same r_info. Note that each group emits a group
1749 // header and that may make the relocation section larger. It is hard to
1750 // estimate the size of a group header as the encoded size of that varies
1751 // based on r_info. However, we can approximate this trade-off by the number
1752 // of values encoded. Each group header contains 3 values, and each relocation
1753 // in a group encodes one less value, as compared to when it is not grouped.
1754 // Therefore, we only group relocations if there are 3 or more of them with
1757 // For Rela, the addend for most non-relative relocations is zero, and thus we
1758 // can usually get a smaller relocation section if we group relocations with 0
1760 std::vector<Elf_Rela> ungroupedNonRelatives;
1761 std::vector<std::vector<Elf_Rela>> nonRelativeGroups;
1762 for (auto i = nonRelatives.begin(), e = nonRelatives.end(); i != e;) {
1764 while (j != e && i->r_info == j->r_info &&
1765 (!config->isRela || i->r_addend == j->r_addend))
1767 if (j - i < 3 || (config->isRela && i->r_addend != 0))
1768 ungroupedNonRelatives.insert(ungroupedNonRelatives.end(), i, j);
1770 nonRelativeGroups.emplace_back(i, j);
1774 // Sort ungrouped relocations by offset to minimize the encoded length.
1775 llvm::sort(ungroupedNonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1776 return a.r_offset < b.r_offset;
1779 unsigned hasAddendIfRela =
1780 config->isRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0;
1782 uint64_t offset = 0;
1783 uint64_t addend = 0;
1785 // Emit the run-length encoding for the groups of adjacent relative
1786 // relocations. Each group is represented using two groups in the packed
1787 // format. The first is used to set the current offset to the start of the
1788 // group (and also encodes the first relocation), and the second encodes the
1789 // remaining relocations.
1790 for (std::vector<Elf_Rela> &g : relativeGroups) {
1791 // The first relocation in the group.
1793 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1794 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1795 add(g[0].r_offset - offset);
1796 add(target->relativeRel);
1797 if (config->isRela) {
1798 add(g[0].r_addend - addend);
1799 addend = g[0].r_addend;
1802 // The remaining relocations.
1804 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1805 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1806 add(config->wordsize);
1807 add(target->relativeRel);
1808 if (config->isRela) {
1809 for (auto i = g.begin() + 1, e = g.end(); i != e; ++i) {
1810 add(i->r_addend - addend);
1811 addend = i->r_addend;
1815 offset = g.back().r_offset;
1818 // Now the ungrouped relatives.
1819 if (!ungroupedRelatives.empty()) {
1820 add(ungroupedRelatives.size());
1821 add(RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1822 add(target->relativeRel);
1823 for (Elf_Rela &r : ungroupedRelatives) {
1824 add(r.r_offset - offset);
1825 offset = r.r_offset;
1826 if (config->isRela) {
1827 add(r.r_addend - addend);
1828 addend = r.r_addend;
1833 // Grouped non-relatives.
1834 for (ArrayRef<Elf_Rela> g : nonRelativeGroups) {
1836 add(RELOCATION_GROUPED_BY_INFO_FLAG);
1838 for (const Elf_Rela &r : g) {
1839 add(r.r_offset - offset);
1840 offset = r.r_offset;
1845 // Finally the ungrouped non-relative relocations.
1846 if (!ungroupedNonRelatives.empty()) {
1847 add(ungroupedNonRelatives.size());
1848 add(hasAddendIfRela);
1849 for (Elf_Rela &r : ungroupedNonRelatives) {
1850 add(r.r_offset - offset);
1851 offset = r.r_offset;
1853 if (config->isRela) {
1854 add(r.r_addend - addend);
1855 addend = r.r_addend;
1860 // Don't allow the section to shrink; otherwise the size of the section can
1861 // oscillate infinitely.
1862 if (relocData.size() < oldSize)
1863 relocData.append(oldSize - relocData.size(), 0);
1865 // Returns whether the section size changed. We need to keep recomputing both
1866 // section layout and the contents of this section until the size converges
1867 // because changing this section's size can affect section layout, which in
1868 // turn can affect the sizes of the LEB-encoded integers stored in this
1870 return relocData.size() != oldSize;
1873 template <class ELFT> RelrSection<ELFT>::RelrSection() {
1874 this->entsize = config->wordsize;
1877 template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() {
1878 // This function computes the contents of an SHT_RELR packed relocation
1881 // Proposal for adding SHT_RELR sections to generic-abi is here:
1882 // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg
1884 // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks
1885 // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ]
1887 // i.e. start with an address, followed by any number of bitmaps. The address
1888 // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63
1889 // relocations each, at subsequent offsets following the last address entry.
1891 // The bitmap entries must have 1 in the least significant bit. The assumption
1892 // here is that an address cannot have 1 in lsb. Odd addresses are not
1895 // Excluding the least significant bit in the bitmap, each non-zero bit in
1896 // the bitmap represents a relocation to be applied to a corresponding machine
1897 // word that follows the base address word. The second least significant bit
1898 // represents the machine word immediately following the initial address, and
1899 // each bit that follows represents the next word, in linear order. As such,
1900 // a single bitmap can encode up to 31 relocations in a 32-bit object, and
1901 // 63 relocations in a 64-bit object.
1903 // This encoding has a couple of interesting properties:
1904 // 1. Looking at any entry, it is clear whether it's an address or a bitmap:
1905 // even means address, odd means bitmap.
1906 // 2. Just a simple list of addresses is a valid encoding.
1908 size_t oldSize = relrRelocs.size();
1911 // Same as Config->Wordsize but faster because this is a compile-time
1913 const size_t wordsize = sizeof(typename ELFT::uint);
1915 // Number of bits to use for the relocation offsets bitmap.
1916 // Must be either 63 or 31.
1917 const size_t nBits = wordsize * 8 - 1;
1919 // Get offsets for all relative relocations and sort them.
1920 std::vector<uint64_t> offsets;
1921 for (const RelativeReloc &rel : relocs)
1922 offsets.push_back(rel.getOffset());
1923 llvm::sort(offsets);
1925 // For each leading relocation, find following ones that can be folded
1926 // as a bitmap and fold them.
1927 for (size_t i = 0, e = offsets.size(); i < e;) {
1928 // Add a leading relocation.
1929 relrRelocs.push_back(Elf_Relr(offsets[i]));
1930 uint64_t base = offsets[i] + wordsize;
1933 // Find foldable relocations to construct bitmaps.
1935 uint64_t bitmap = 0;
1938 uint64_t delta = offsets[i] - base;
1940 // If it is too far, it cannot be folded.
1941 if (delta >= nBits * wordsize)
1944 // If it is not a multiple of wordsize away, it cannot be folded.
1945 if (delta % wordsize)
1949 bitmap |= 1ULL << (delta / wordsize);
1956 relrRelocs.push_back(Elf_Relr((bitmap << 1) | 1));
1957 base += nBits * wordsize;
1961 // Don't allow the section to shrink; otherwise the size of the section can
1962 // oscillate infinitely. Trailing 1s do not decode to more relocations.
1963 if (relrRelocs.size() < oldSize) {
1964 log(".relr.dyn needs " + Twine(oldSize - relrRelocs.size()) +
1965 " padding word(s)");
1966 relrRelocs.resize(oldSize, Elf_Relr(1));
1969 return relrRelocs.size() != oldSize;
1972 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &strTabSec)
1973 : SyntheticSection(strTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0,
1974 strTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB,
1976 strTabSec.isDynamic() ? ".dynsym" : ".symtab"),
1977 strTabSec(strTabSec) {}
1979 // Orders symbols according to their positions in the GOT,
1980 // in compliance with MIPS ABI rules.
1981 // See "Global Offset Table" in Chapter 5 in the following document
1982 // for detailed description:
1983 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1984 static bool sortMipsSymbols(const SymbolTableEntry &l,
1985 const SymbolTableEntry &r) {
1986 // Sort entries related to non-local preemptible symbols by GOT indexes.
1987 // All other entries go to the beginning of a dynsym in arbitrary order.
1988 if (l.sym->isInGot() && r.sym->isInGot())
1989 return l.sym->gotIndex < r.sym->gotIndex;
1990 if (!l.sym->isInGot() && !r.sym->isInGot())
1992 return !l.sym->isInGot();
1995 void SymbolTableBaseSection::finalizeContents() {
1996 if (OutputSection *sec = strTabSec.getParent())
1997 getParent()->link = sec->sectionIndex;
1999 if (this->type != SHT_DYNSYM) {
2000 sortSymTabSymbols();
2004 // If it is a .dynsym, there should be no local symbols, but we need
2005 // to do a few things for the dynamic linker.
2007 // Section's Info field has the index of the first non-local symbol.
2008 // Because the first symbol entry is a null entry, 1 is the first.
2009 getParent()->info = 1;
2011 if (getPartition().gnuHashTab) {
2012 // NB: It also sorts Symbols to meet the GNU hash table requirements.
2013 getPartition().gnuHashTab->addSymbols(symbols);
2014 } else if (config->emachine == EM_MIPS) {
2015 llvm::stable_sort(symbols, sortMipsSymbols);
2018 // Only the main partition's dynsym indexes are stored in the symbols
2019 // themselves. All other partitions use a lookup table.
2020 if (this == mainPart->dynSymTab) {
2022 for (const SymbolTableEntry &s : symbols)
2023 s.sym->dynsymIndex = ++i;
2027 // The ELF spec requires that all local symbols precede global symbols, so we
2028 // sort symbol entries in this function. (For .dynsym, we don't do that because
2029 // symbols for dynamic linking are inherently all globals.)
2031 // Aside from above, we put local symbols in groups starting with the STT_FILE
2032 // symbol. That is convenient for purpose of identifying where are local symbols
2034 void SymbolTableBaseSection::sortSymTabSymbols() {
2035 // Move all local symbols before global symbols.
2036 auto e = std::stable_partition(
2037 symbols.begin(), symbols.end(), [](const SymbolTableEntry &s) {
2038 return s.sym->isLocal() || s.sym->computeBinding() == STB_LOCAL;
2040 size_t numLocals = e - symbols.begin();
2041 getParent()->info = numLocals + 1;
2043 // We want to group the local symbols by file. For that we rebuild the local
2044 // part of the symbols vector. We do not need to care about the STT_FILE
2045 // symbols, they are already naturally placed first in each group. That
2046 // happens because STT_FILE is always the first symbol in the object and hence
2047 // precede all other local symbols we add for a file.
2048 MapVector<InputFile *, std::vector<SymbolTableEntry>> arr;
2049 for (const SymbolTableEntry &s : llvm::make_range(symbols.begin(), e))
2050 arr[s.sym->file].push_back(s);
2052 auto i = symbols.begin();
2053 for (std::pair<InputFile *, std::vector<SymbolTableEntry>> &p : arr)
2054 for (SymbolTableEntry &entry : p.second)
2058 void SymbolTableBaseSection::addSymbol(Symbol *b) {
2059 // Adding a local symbol to a .dynsym is a bug.
2060 assert(this->type != SHT_DYNSYM || !b->isLocal());
2062 bool hashIt = b->isLocal();
2063 symbols.push_back({b, strTabSec.addString(b->getName(), hashIt)});
2066 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *sym) {
2067 if (this == mainPart->dynSymTab)
2068 return sym->dynsymIndex;
2070 // Initializes symbol lookup tables lazily. This is used only for -r,
2071 // -emit-relocs and dynsyms in partitions other than the main one.
2072 llvm::call_once(onceFlag, [&] {
2073 symbolIndexMap.reserve(symbols.size());
2075 for (const SymbolTableEntry &e : symbols) {
2076 if (e.sym->type == STT_SECTION)
2077 sectionIndexMap[e.sym->getOutputSection()] = ++i;
2079 symbolIndexMap[e.sym] = ++i;
2083 // Section symbols are mapped based on their output sections
2084 // to maintain their semantics.
2085 if (sym->type == STT_SECTION)
2086 return sectionIndexMap.lookup(sym->getOutputSection());
2087 return symbolIndexMap.lookup(sym);
2090 template <class ELFT>
2091 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &strTabSec)
2092 : SymbolTableBaseSection(strTabSec) {
2093 this->entsize = sizeof(Elf_Sym);
2096 static BssSection *getCommonSec(Symbol *sym) {
2097 if (!config->defineCommon)
2098 if (auto *d = dyn_cast<Defined>(sym))
2099 return dyn_cast_or_null<BssSection>(d->section);
2103 static uint32_t getSymSectionIndex(Symbol *sym) {
2104 if (getCommonSec(sym))
2106 if (!isa<Defined>(sym) || sym->needsPltAddr)
2108 if (const OutputSection *os = sym->getOutputSection())
2109 return os->sectionIndex >= SHN_LORESERVE ? (uint32_t)SHN_XINDEX
2114 // Write the internal symbol table contents to the output symbol table.
2115 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *buf) {
2116 // The first entry is a null entry as per the ELF spec.
2117 memset(buf, 0, sizeof(Elf_Sym));
2118 buf += sizeof(Elf_Sym);
2120 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2122 for (SymbolTableEntry &ent : symbols) {
2123 Symbol *sym = ent.sym;
2124 bool isDefinedHere = type == SHT_SYMTAB || sym->partition == partition;
2126 // Set st_info and st_other.
2128 if (sym->isLocal()) {
2129 eSym->setBindingAndType(STB_LOCAL, sym->type);
2131 eSym->setBindingAndType(sym->computeBinding(), sym->type);
2132 eSym->setVisibility(sym->visibility);
2135 // The 3 most significant bits of st_other are used by OpenPOWER ABI.
2136 // See getPPC64GlobalEntryToLocalEntryOffset() for more details.
2137 if (config->emachine == EM_PPC64)
2138 eSym->st_other |= sym->stOther & 0xe0;
2140 eSym->st_name = ent.strTabOffset;
2142 eSym->st_shndx = getSymSectionIndex(ent.sym);
2146 // Copy symbol size if it is a defined symbol. st_size is not significant
2147 // for undefined symbols, so whether copying it or not is up to us if that's
2148 // the case. We'll leave it as zero because by not setting a value, we can
2149 // get the exact same outputs for two sets of input files that differ only
2150 // in undefined symbol size in DSOs.
2151 if (eSym->st_shndx == SHN_UNDEF || !isDefinedHere)
2154 eSym->st_size = sym->getSize();
2156 // st_value is usually an address of a symbol, but that has a
2157 // special meaning for uninstantiated common symbols (this can
2158 // occur if -r is given).
2159 if (BssSection *commonSec = getCommonSec(ent.sym))
2160 eSym->st_value = commonSec->alignment;
2161 else if (isDefinedHere)
2162 eSym->st_value = sym->getVA();
2169 // On MIPS we need to mark symbol which has a PLT entry and requires
2170 // pointer equality by STO_MIPS_PLT flag. That is necessary to help
2171 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs.
2172 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt
2173 if (config->emachine == EM_MIPS) {
2174 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2176 for (SymbolTableEntry &ent : symbols) {
2177 Symbol *sym = ent.sym;
2178 if (sym->isInPlt() && sym->needsPltAddr)
2179 eSym->st_other |= STO_MIPS_PLT;
2180 if (isMicroMips()) {
2181 // We already set the less-significant bit for symbols
2182 // marked by the `STO_MIPS_MICROMIPS` flag and for microMIPS PLT
2183 // records. That allows us to distinguish such symbols in
2184 // the `MIPS<ELFT>::relocate()` routine. Now we should
2185 // clear that bit for non-dynamic symbol table, so tools
2186 // like `objdump` will be able to deal with a correct
2188 if (sym->isDefined() &&
2189 ((sym->stOther & STO_MIPS_MICROMIPS) || sym->needsPltAddr)) {
2190 if (!strTabSec.isDynamic())
2191 eSym->st_value &= ~1;
2192 eSym->st_other |= STO_MIPS_MICROMIPS;
2195 if (config->relocatable)
2196 if (auto *d = dyn_cast<Defined>(sym))
2197 if (isMipsPIC<ELFT>(d))
2198 eSym->st_other |= STO_MIPS_PIC;
2204 SymtabShndxSection::SymtabShndxSection()
2205 : SyntheticSection(0, SHT_SYMTAB_SHNDX, 4, ".symtab_shndx") {
2209 void SymtabShndxSection::writeTo(uint8_t *buf) {
2210 // We write an array of 32 bit values, where each value has 1:1 association
2211 // with an entry in .symtab. If the corresponding entry contains SHN_XINDEX,
2212 // we need to write actual index, otherwise, we must write SHN_UNDEF(0).
2213 buf += 4; // Ignore .symtab[0] entry.
2214 for (const SymbolTableEntry &entry : in.symTab->getSymbols()) {
2215 if (getSymSectionIndex(entry.sym) == SHN_XINDEX)
2216 write32(buf, entry.sym->getOutputSection()->sectionIndex);
2221 bool SymtabShndxSection::isNeeded() const {
2222 // SHT_SYMTAB can hold symbols with section indices values up to
2223 // SHN_LORESERVE. If we need more, we want to use extension SHT_SYMTAB_SHNDX
2224 // section. Problem is that we reveal the final section indices a bit too
2225 // late, and we do not know them here. For simplicity, we just always create
2226 // a .symtab_shndx section when the amount of output sections is huge.
2228 for (BaseCommand *base : script->sectionCommands)
2229 if (isa<OutputSection>(base))
2231 return size >= SHN_LORESERVE;
2234 void SymtabShndxSection::finalizeContents() {
2235 getParent()->link = in.symTab->getParent()->sectionIndex;
2238 size_t SymtabShndxSection::getSize() const {
2239 return in.symTab->getNumSymbols() * 4;
2242 // .hash and .gnu.hash sections contain on-disk hash tables that map
2243 // symbol names to their dynamic symbol table indices. Their purpose
2244 // is to help the dynamic linker resolve symbols quickly. If ELF files
2245 // don't have them, the dynamic linker has to do linear search on all
2246 // dynamic symbols, which makes programs slower. Therefore, a .hash
2247 // section is added to a DSO by default. A .gnu.hash is added if you
2248 // give the -hash-style=gnu or -hash-style=both option.
2250 // The Unix semantics of resolving dynamic symbols is somewhat expensive.
2251 // Each ELF file has a list of DSOs that the ELF file depends on and a
2252 // list of dynamic symbols that need to be resolved from any of the
2253 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n)
2254 // where m is the number of DSOs and n is the number of dynamic
2255 // symbols. For modern large programs, both m and n are large. So
2256 // making each step faster by using hash tables substantially
2257 // improves time to load programs.
2259 // (Note that this is not the only way to design the shared library.
2260 // For instance, the Windows DLL takes a different approach. On
2261 // Windows, each dynamic symbol has a name of DLL from which the symbol
2262 // has to be resolved. That makes the cost of symbol resolution O(n).
2263 // This disables some hacky techniques you can use on Unix such as
2264 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.)
2266 // Due to historical reasons, we have two different hash tables, .hash
2267 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new
2268 // and better version of .hash. .hash is just an on-disk hash table, but
2269 // .gnu.hash has a bloom filter in addition to a hash table to skip
2270 // DSOs very quickly. If you are sure that your dynamic linker knows
2271 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a
2272 // safe bet is to specify -hash-style=both for backward compatibility.
2273 GnuHashTableSection::GnuHashTableSection()
2274 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, config->wordsize, ".gnu.hash") {
2277 void GnuHashTableSection::finalizeContents() {
2278 if (OutputSection *sec = getPartition().dynSymTab->getParent())
2279 getParent()->link = sec->sectionIndex;
2281 // Computes bloom filter size in word size. We want to allocate 12
2282 // bits for each symbol. It must be a power of two.
2283 if (symbols.empty()) {
2286 uint64_t numBits = symbols.size() * 12;
2287 maskWords = NextPowerOf2(numBits / (config->wordsize * 8));
2290 size = 16; // Header
2291 size += config->wordsize * maskWords; // Bloom filter
2292 size += nBuckets * 4; // Hash buckets
2293 size += symbols.size() * 4; // Hash values
2296 void GnuHashTableSection::writeTo(uint8_t *buf) {
2297 // The output buffer is not guaranteed to be zero-cleared because we pre-
2298 // fill executable sections with trap instructions. This is a precaution
2299 // for that case, which happens only when -no-rosegment is given.
2300 memset(buf, 0, size);
2303 write32(buf, nBuckets);
2304 write32(buf + 4, getPartition().dynSymTab->getNumSymbols() - symbols.size());
2305 write32(buf + 8, maskWords);
2306 write32(buf + 12, Shift2);
2309 // Write a bloom filter and a hash table.
2310 writeBloomFilter(buf);
2311 buf += config->wordsize * maskWords;
2312 writeHashTable(buf);
2315 // This function writes a 2-bit bloom filter. This bloom filter alone
2316 // usually filters out 80% or more of all symbol lookups [1].
2317 // The dynamic linker uses the hash table only when a symbol is not
2318 // filtered out by a bloom filter.
2320 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2),
2321 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf
2322 void GnuHashTableSection::writeBloomFilter(uint8_t *buf) {
2323 unsigned c = config->is64 ? 64 : 32;
2324 for (const Entry &sym : symbols) {
2325 // When C = 64, we choose a word with bits [6:...] and set 1 to two bits in
2326 // the word using bits [0:5] and [26:31].
2327 size_t i = (sym.hash / c) & (maskWords - 1);
2328 uint64_t val = readUint(buf + i * config->wordsize);
2329 val |= uint64_t(1) << (sym.hash % c);
2330 val |= uint64_t(1) << ((sym.hash >> Shift2) % c);
2331 writeUint(buf + i * config->wordsize, val);
2335 void GnuHashTableSection::writeHashTable(uint8_t *buf) {
2336 uint32_t *buckets = reinterpret_cast<uint32_t *>(buf);
2337 uint32_t oldBucket = -1;
2338 uint32_t *values = buckets + nBuckets;
2339 for (auto i = symbols.begin(), e = symbols.end(); i != e; ++i) {
2340 // Write a hash value. It represents a sequence of chains that share the
2341 // same hash modulo value. The last element of each chain is terminated by
2343 uint32_t hash = i->hash;
2344 bool isLastInChain = (i + 1) == e || i->bucketIdx != (i + 1)->bucketIdx;
2345 hash = isLastInChain ? hash | 1 : hash & ~1;
2346 write32(values++, hash);
2348 if (i->bucketIdx == oldBucket)
2350 // Write a hash bucket. Hash buckets contain indices in the following hash
2352 write32(buckets + i->bucketIdx,
2353 getPartition().dynSymTab->getSymbolIndex(i->sym));
2354 oldBucket = i->bucketIdx;
2358 static uint32_t hashGnu(StringRef name) {
2360 for (uint8_t c : name)
2361 h = (h << 5) + h + c;
2365 // Add symbols to this symbol hash table. Note that this function
2366 // destructively sort a given vector -- which is needed because
2367 // GNU-style hash table places some sorting requirements.
2368 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &v) {
2369 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce
2370 // its type correctly.
2371 std::vector<SymbolTableEntry>::iterator mid =
2372 std::stable_partition(v.begin(), v.end(), [&](const SymbolTableEntry &s) {
2373 return !s.sym->isDefined() || s.sym->partition != partition;
2376 // We chose load factor 4 for the on-disk hash table. For each hash
2377 // collision, the dynamic linker will compare a uint32_t hash value.
2378 // Since the integer comparison is quite fast, we believe we can
2379 // make the load factor even larger. 4 is just a conservative choice.
2381 // Note that we don't want to create a zero-sized hash table because
2382 // Android loader as of 2018 doesn't like a .gnu.hash containing such
2383 // table. If that's the case, we create a hash table with one unused
2385 nBuckets = std::max<size_t>((v.end() - mid) / 4, 1);
2390 for (SymbolTableEntry &ent : llvm::make_range(mid, v.end())) {
2391 Symbol *b = ent.sym;
2392 uint32_t hash = hashGnu(b->getName());
2393 uint32_t bucketIdx = hash % nBuckets;
2394 symbols.push_back({b, ent.strTabOffset, hash, bucketIdx});
2397 llvm::stable_sort(symbols, [](const Entry &l, const Entry &r) {
2398 return l.bucketIdx < r.bucketIdx;
2401 v.erase(mid, v.end());
2402 for (const Entry &ent : symbols)
2403 v.push_back({ent.sym, ent.strTabOffset});
2406 HashTableSection::HashTableSection()
2407 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") {
2411 void HashTableSection::finalizeContents() {
2412 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2414 if (OutputSection *sec = symTab->getParent())
2415 getParent()->link = sec->sectionIndex;
2417 unsigned numEntries = 2; // nbucket and nchain.
2418 numEntries += symTab->getNumSymbols(); // The chain entries.
2420 // Create as many buckets as there are symbols.
2421 numEntries += symTab->getNumSymbols();
2422 this->size = numEntries * 4;
2425 void HashTableSection::writeTo(uint8_t *buf) {
2426 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2428 // See comment in GnuHashTableSection::writeTo.
2429 memset(buf, 0, size);
2431 unsigned numSymbols = symTab->getNumSymbols();
2433 uint32_t *p = reinterpret_cast<uint32_t *>(buf);
2434 write32(p++, numSymbols); // nbucket
2435 write32(p++, numSymbols); // nchain
2437 uint32_t *buckets = p;
2438 uint32_t *chains = p + numSymbols;
2440 for (const SymbolTableEntry &s : symTab->getSymbols()) {
2441 Symbol *sym = s.sym;
2442 StringRef name = sym->getName();
2443 unsigned i = sym->dynsymIndex;
2444 uint32_t hash = hashSysV(name) % numSymbols;
2445 chains[i] = buckets[hash];
2446 write32(buckets + hash, i);
2450 PltSection::PltSection()
2451 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt"),
2452 headerSize(target->pltHeaderSize) {
2453 // On PowerPC, this section contains lazy symbol resolvers.
2454 if (config->emachine == EM_PPC64) {
2459 // On x86 when IBT is enabled, this section contains the second PLT (lazy
2460 // symbol resolvers).
2461 if ((config->emachine == EM_386 || config->emachine == EM_X86_64) &&
2462 (config->andFeatures & GNU_PROPERTY_X86_FEATURE_1_IBT))
2465 // The PLT needs to be writable on SPARC as the dynamic linker will
2466 // modify the instructions in the PLT entries.
2467 if (config->emachine == EM_SPARCV9)
2468 this->flags |= SHF_WRITE;
2471 void PltSection::writeTo(uint8_t *buf) {
2472 // At beginning of PLT, we have code to call the dynamic
2473 // linker to resolve dynsyms at runtime. Write such code.
2474 target->writePltHeader(buf);
2475 size_t off = headerSize;
2477 for (const Symbol *sym : entries) {
2478 target->writePlt(buf + off, *sym, getVA() + off);
2479 off += target->pltEntrySize;
2483 void PltSection::addEntry(Symbol &sym) {
2484 sym.pltIndex = entries.size();
2485 entries.push_back(&sym);
2488 size_t PltSection::getSize() const {
2489 return headerSize + entries.size() * target->pltEntrySize;
2492 bool PltSection::isNeeded() const {
2493 // For -z retpolineplt, .iplt needs the .plt header.
2494 return !entries.empty() || (config->zRetpolineplt && in.iplt->isNeeded());
2497 // Used by ARM to add mapping symbols in the PLT section, which aid
2499 void PltSection::addSymbols() {
2500 target->addPltHeaderSymbols(*this);
2502 size_t off = headerSize;
2503 for (size_t i = 0; i < entries.size(); ++i) {
2504 target->addPltSymbols(*this, off);
2505 off += target->pltEntrySize;
2509 IpltSection::IpltSection()
2510 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".iplt") {
2511 if (config->emachine == EM_PPC || config->emachine == EM_PPC64) {
2517 void IpltSection::writeTo(uint8_t *buf) {
2519 for (const Symbol *sym : entries) {
2520 target->writeIplt(buf + off, *sym, getVA() + off);
2521 off += target->ipltEntrySize;
2525 size_t IpltSection::getSize() const {
2526 return entries.size() * target->ipltEntrySize;
2529 void IpltSection::addEntry(Symbol &sym) {
2530 sym.pltIndex = entries.size();
2531 entries.push_back(&sym);
2534 // ARM uses mapping symbols to aid disassembly.
2535 void IpltSection::addSymbols() {
2537 for (size_t i = 0, e = entries.size(); i != e; ++i) {
2538 target->addPltSymbols(*this, off);
2539 off += target->pltEntrySize;
2543 PPC32GlinkSection::PPC32GlinkSection() {
2548 void PPC32GlinkSection::writeTo(uint8_t *buf) {
2549 writePPC32GlinkSection(buf, entries.size());
2552 size_t PPC32GlinkSection::getSize() const {
2553 return headerSize + entries.size() * target->pltEntrySize + footerSize;
2556 // This is an x86-only extra PLT section and used only when a security
2557 // enhancement feature called CET is enabled. In this comment, I'll explain what
2558 // the feature is and why we have two PLT sections if CET is enabled.
2560 // So, what does CET do? CET introduces a new restriction to indirect jump
2561 // instructions. CET works this way. Assume that CET is enabled. Then, if you
2562 // execute an indirect jump instruction, the processor verifies that a special
2563 // "landing pad" instruction (which is actually a repurposed NOP instruction and
2564 // now called "endbr32" or "endbr64") is at the jump target. If the jump target
2565 // does not start with that instruction, the processor raises an exception
2566 // instead of continuing executing code.
2568 // If CET is enabled, the compiler emits endbr to all locations where indirect
2569 // jumps may jump to.
2571 // This mechanism makes it extremely hard to transfer the control to a middle of
2572 // a function that is not supporsed to be a indirect jump target, preventing
2573 // certain types of attacks such as ROP or JOP.
2575 // Note that the processors in the market as of 2019 don't actually support the
2576 // feature. Only the spec is available at the moment.
2578 // Now, I'll explain why we have this extra PLT section for CET.
2580 // Since you can indirectly jump to a PLT entry, we have to make PLT entries
2581 // start with endbr. The problem is there's no extra space for endbr (which is 4
2582 // bytes long), as the PLT entry is only 16 bytes long and all bytes are already
2585 // In order to deal with the issue, we split a PLT entry into two PLT entries.
2586 // Remember that each PLT entry contains code to jump to an address read from
2587 // .got.plt AND code to resolve a dynamic symbol lazily. With the 2-PLT scheme,
2588 // the former code is written to .plt.sec, and the latter code is written to
2591 // Lazy symbol resolution in the 2-PLT scheme works in the usual way, except
2592 // that the regular .plt is now called .plt.sec and .plt is repurposed to
2593 // contain only code for lazy symbol resolution.
2595 // In other words, this is how the 2-PLT scheme works. Application code is
2596 // supposed to jump to .plt.sec to call an external function. Each .plt.sec
2597 // entry contains code to read an address from a corresponding .got.plt entry
2598 // and jump to that address. Addresses in .got.plt initially point to .plt, so
2599 // when an application calls an external function for the first time, the
2600 // control is transferred to a function that resolves a symbol name from
2601 // external shared object files. That function then rewrites a .got.plt entry
2602 // with a resolved address, so that the subsequent function calls directly jump
2603 // to a desired location from .plt.sec.
2605 // There is an open question as to whether the 2-PLT scheme was desirable or
2606 // not. We could have simply extended the PLT entry size to 32-bytes to
2607 // accommodate endbr, and that scheme would have been much simpler than the
2608 // 2-PLT scheme. One reason to split PLT was, by doing that, we could keep hot
2609 // code (.plt.sec) from cold code (.plt). But as far as I know no one proved
2610 // that the optimization actually makes a difference.
2612 // That said, the 2-PLT scheme is a part of the ABI, debuggers and other tools
2613 // depend on it, so we implement the ABI.
2614 IBTPltSection::IBTPltSection()
2615 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt") {}
2617 void IBTPltSection::writeTo(uint8_t *buf) {
2618 target->writeIBTPlt(buf, in.plt->getNumEntries());
2621 size_t IBTPltSection::getSize() const {
2622 // 16 is the header size of .plt.
2623 return 16 + in.plt->getNumEntries() * target->pltEntrySize;
2626 // The string hash function for .gdb_index.
2627 static uint32_t computeGdbHash(StringRef s) {
2630 h = h * 67 + toLower(c) - 113;
2634 GdbIndexSection::GdbIndexSection()
2635 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index") {}
2637 // Returns the desired size of an on-disk hash table for a .gdb_index section.
2638 // There's a tradeoff between size and collision rate. We aim 75% utilization.
2639 size_t GdbIndexSection::computeSymtabSize() const {
2640 return std::max<size_t>(NextPowerOf2(symbols.size() * 4 / 3), 1024);
2643 // Compute the output section size.
2644 void GdbIndexSection::initOutputSize() {
2645 size = sizeof(GdbIndexHeader) + computeSymtabSize() * 8;
2647 for (GdbChunk &chunk : chunks)
2648 size += chunk.compilationUnits.size() * 16 + chunk.addressAreas.size() * 20;
2650 // Add the constant pool size if exists.
2651 if (!symbols.empty()) {
2652 GdbSymbol &sym = symbols.back();
2653 size += sym.nameOff + sym.name.size() + 1;
2657 static std::vector<GdbIndexSection::CuEntry> readCuList(DWARFContext &dwarf) {
2658 std::vector<GdbIndexSection::CuEntry> ret;
2659 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units())
2660 ret.push_back({cu->getOffset(), cu->getLength() + 4});
2664 static std::vector<GdbIndexSection::AddressEntry>
2665 readAddressAreas(DWARFContext &dwarf, InputSection *sec) {
2666 std::vector<GdbIndexSection::AddressEntry> ret;
2669 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) {
2670 if (Error e = cu->tryExtractDIEsIfNeeded(false)) {
2671 warn(toString(sec) + ": " + toString(std::move(e)));
2674 Expected<DWARFAddressRangesVector> ranges = cu->collectAddressRanges();
2676 warn(toString(sec) + ": " + toString(ranges.takeError()));
2680 ArrayRef<InputSectionBase *> sections = sec->file->getSections();
2681 for (DWARFAddressRange &r : *ranges) {
2682 if (r.SectionIndex == -1ULL)
2684 // Range list with zero size has no effect.
2685 InputSectionBase *s = sections[r.SectionIndex];
2686 if (s && s != &InputSection::discarded && s->isLive())
2687 if (r.LowPC != r.HighPC)
2688 ret.push_back({cast<InputSection>(s), r.LowPC, r.HighPC, cuIdx});
2696 template <class ELFT>
2697 static std::vector<GdbIndexSection::NameAttrEntry>
2698 readPubNamesAndTypes(const LLDDwarfObj<ELFT> &obj,
2699 const std::vector<GdbIndexSection::CuEntry> &cus) {
2700 const LLDDWARFSection &pubNames = obj.getGnuPubnamesSection();
2701 const LLDDWARFSection &pubTypes = obj.getGnuPubtypesSection();
2703 std::vector<GdbIndexSection::NameAttrEntry> ret;
2704 for (const LLDDWARFSection *pub : {&pubNames, &pubTypes}) {
2705 DWARFDataExtractor data(obj, *pub, config->isLE, config->wordsize);
2706 DWARFDebugPubTable table;
2707 table.extract(data, /*GnuStyle=*/true, [&](Error e) {
2708 warn(toString(pub->sec) + ": " + toString(std::move(e)));
2710 for (const DWARFDebugPubTable::Set &set : table.getData()) {
2711 // The value written into the constant pool is kind << 24 | cuIndex. As we
2712 // don't know how many compilation units precede this object to compute
2713 // cuIndex, we compute (kind << 24 | cuIndexInThisObject) instead, and add
2714 // the number of preceding compilation units later.
2715 uint32_t i = llvm::partition_point(cus,
2716 [&](GdbIndexSection::CuEntry cu) {
2717 return cu.cuOffset < set.Offset;
2720 for (const DWARFDebugPubTable::Entry &ent : set.Entries)
2721 ret.push_back({{ent.Name, computeGdbHash(ent.Name)},
2722 (ent.Descriptor.toBits() << 24) | i});
2728 // Create a list of symbols from a given list of symbol names and types
2729 // by uniquifying them by name.
2730 static std::vector<GdbIndexSection::GdbSymbol>
2731 createSymbols(ArrayRef<std::vector<GdbIndexSection::NameAttrEntry>> nameAttrs,
2732 const std::vector<GdbIndexSection::GdbChunk> &chunks) {
2733 using GdbSymbol = GdbIndexSection::GdbSymbol;
2734 using NameAttrEntry = GdbIndexSection::NameAttrEntry;
2736 // For each chunk, compute the number of compilation units preceding it.
2738 std::vector<uint32_t> cuIdxs(chunks.size());
2739 for (uint32_t i = 0, e = chunks.size(); i != e; ++i) {
2741 cuIdx += chunks[i].compilationUnits.size();
2744 // The number of symbols we will handle in this function is of the order
2745 // of millions for very large executables, so we use multi-threading to
2747 constexpr size_t numShards = 32;
2748 size_t concurrency = PowerOf2Floor(
2749 std::min<size_t>(hardware_concurrency(parallel::strategy.ThreadsRequested)
2750 .compute_thread_count(),
2753 // A sharded map to uniquify symbols by name.
2754 std::vector<DenseMap<CachedHashStringRef, size_t>> map(numShards);
2755 size_t shift = 32 - countTrailingZeros(numShards);
2757 // Instantiate GdbSymbols while uniqufying them by name.
2758 std::vector<std::vector<GdbSymbol>> symbols(numShards);
2759 parallelForEachN(0, concurrency, [&](size_t threadId) {
2761 for (ArrayRef<NameAttrEntry> entries : nameAttrs) {
2762 for (const NameAttrEntry &ent : entries) {
2763 size_t shardId = ent.name.hash() >> shift;
2764 if ((shardId & (concurrency - 1)) != threadId)
2767 uint32_t v = ent.cuIndexAndAttrs + cuIdxs[i];
2768 size_t &idx = map[shardId][ent.name];
2770 symbols[shardId][idx - 1].cuVector.push_back(v);
2774 idx = symbols[shardId].size() + 1;
2775 symbols[shardId].push_back({ent.name, {v}, 0, 0});
2781 size_t numSymbols = 0;
2782 for (ArrayRef<GdbSymbol> v : symbols)
2783 numSymbols += v.size();
2785 // The return type is a flattened vector, so we'll copy each vector
2787 std::vector<GdbSymbol> ret;
2788 ret.reserve(numSymbols);
2789 for (std::vector<GdbSymbol> &vec : symbols)
2790 for (GdbSymbol &sym : vec)
2791 ret.push_back(std::move(sym));
2793 // CU vectors and symbol names are adjacent in the output file.
2794 // We can compute their offsets in the output file now.
2796 for (GdbSymbol &sym : ret) {
2797 sym.cuVectorOff = off;
2798 off += (sym.cuVector.size() + 1) * 4;
2800 for (GdbSymbol &sym : ret) {
2802 off += sym.name.size() + 1;
2808 // Returns a newly-created .gdb_index section.
2809 template <class ELFT> GdbIndexSection *GdbIndexSection::create() {
2810 // Collect InputFiles with .debug_info. See the comment in
2811 // LLDDwarfObj<ELFT>::LLDDwarfObj. If we do lightweight parsing in the future,
2812 // note that isec->data() may uncompress the full content, which should be
2814 SetVector<InputFile *> files;
2815 for (InputSectionBase *s : inputSections) {
2816 InputSection *isec = dyn_cast<InputSection>(s);
2819 // .debug_gnu_pub{names,types} are useless in executables.
2820 // They are present in input object files solely for creating
2821 // a .gdb_index. So we can remove them from the output.
2822 if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes")
2824 else if (isec->name == ".debug_info")
2825 files.insert(isec->file);
2828 std::vector<GdbChunk> chunks(files.size());
2829 std::vector<std::vector<NameAttrEntry>> nameAttrs(files.size());
2831 parallelForEachN(0, files.size(), [&](size_t i) {
2832 // To keep memory usage low, we don't want to keep cached DWARFContext, so
2833 // avoid getDwarf() here.
2834 ObjFile<ELFT> *file = cast<ObjFile<ELFT>>(files[i]);
2835 DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file));
2836 auto &dobj = static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj());
2838 // If the are multiple compile units .debug_info (very rare ld -r --unique),
2839 // this only picks the last one. Other address ranges are lost.
2840 chunks[i].sec = dobj.getInfoSection();
2841 chunks[i].compilationUnits = readCuList(dwarf);
2842 chunks[i].addressAreas = readAddressAreas(dwarf, chunks[i].sec);
2843 nameAttrs[i] = readPubNamesAndTypes<ELFT>(dobj, chunks[i].compilationUnits);
2846 auto *ret = make<GdbIndexSection>();
2847 ret->chunks = std::move(chunks);
2848 ret->symbols = createSymbols(nameAttrs, ret->chunks);
2849 ret->initOutputSize();
2853 void GdbIndexSection::writeTo(uint8_t *buf) {
2854 // Write the header.
2855 auto *hdr = reinterpret_cast<GdbIndexHeader *>(buf);
2856 uint8_t *start = buf;
2858 buf += sizeof(*hdr);
2860 // Write the CU list.
2861 hdr->cuListOff = buf - start;
2862 for (GdbChunk &chunk : chunks) {
2863 for (CuEntry &cu : chunk.compilationUnits) {
2864 write64le(buf, chunk.sec->outSecOff + cu.cuOffset);
2865 write64le(buf + 8, cu.cuLength);
2870 // Write the address area.
2871 hdr->cuTypesOff = buf - start;
2872 hdr->addressAreaOff = buf - start;
2874 for (GdbChunk &chunk : chunks) {
2875 for (AddressEntry &e : chunk.addressAreas) {
2876 uint64_t baseAddr = e.section->getVA(0);
2877 write64le(buf, baseAddr + e.lowAddress);
2878 write64le(buf + 8, baseAddr + e.highAddress);
2879 write32le(buf + 16, e.cuIndex + cuOff);
2882 cuOff += chunk.compilationUnits.size();
2885 // Write the on-disk open-addressing hash table containing symbols.
2886 hdr->symtabOff = buf - start;
2887 size_t symtabSize = computeSymtabSize();
2888 uint32_t mask = symtabSize - 1;
2890 for (GdbSymbol &sym : symbols) {
2891 uint32_t h = sym.name.hash();
2892 uint32_t i = h & mask;
2893 uint32_t step = ((h * 17) & mask) | 1;
2895 while (read32le(buf + i * 8))
2896 i = (i + step) & mask;
2898 write32le(buf + i * 8, sym.nameOff);
2899 write32le(buf + i * 8 + 4, sym.cuVectorOff);
2902 buf += symtabSize * 8;
2904 // Write the string pool.
2905 hdr->constantPoolOff = buf - start;
2906 parallelForEach(symbols, [&](GdbSymbol &sym) {
2907 memcpy(buf + sym.nameOff, sym.name.data(), sym.name.size());
2910 // Write the CU vectors.
2911 for (GdbSymbol &sym : symbols) {
2912 write32le(buf, sym.cuVector.size());
2914 for (uint32_t val : sym.cuVector) {
2915 write32le(buf, val);
2921 bool GdbIndexSection::isNeeded() const { return !chunks.empty(); }
2923 EhFrameHeader::EhFrameHeader()
2924 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr") {}
2926 void EhFrameHeader::writeTo(uint8_t *buf) {
2927 // Unlike most sections, the EhFrameHeader section is written while writing
2928 // another section, namely EhFrameSection, which calls the write() function
2929 // below from its writeTo() function. This is necessary because the contents
2930 // of EhFrameHeader depend on the relocated contents of EhFrameSection and we
2931 // don't know which order the sections will be written in.
2934 // .eh_frame_hdr contains a binary search table of pointers to FDEs.
2935 // Each entry of the search table consists of two values,
2936 // the starting PC from where FDEs covers, and the FDE's address.
2937 // It is sorted by PC.
2938 void EhFrameHeader::write() {
2939 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
2940 using FdeData = EhFrameSection::FdeData;
2942 std::vector<FdeData> fdes = getPartition().ehFrame->getFdeData();
2945 buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4;
2946 buf[2] = DW_EH_PE_udata4;
2947 buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4;
2949 getPartition().ehFrame->getParent()->addr - this->getVA() - 4);
2950 write32(buf + 8, fdes.size());
2953 for (FdeData &fde : fdes) {
2954 write32(buf, fde.pcRel);
2955 write32(buf + 4, fde.fdeVARel);
2960 size_t EhFrameHeader::getSize() const {
2961 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs.
2962 return 12 + getPartition().ehFrame->numFdes * 8;
2965 bool EhFrameHeader::isNeeded() const {
2966 return isLive() && getPartition().ehFrame->isNeeded();
2969 VersionDefinitionSection::VersionDefinitionSection()
2970 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t),
2971 ".gnu.version_d") {}
2973 StringRef VersionDefinitionSection::getFileDefName() {
2974 if (!getPartition().name.empty())
2975 return getPartition().name;
2976 if (!config->soName.empty())
2977 return config->soName;
2978 return config->outputFile;
2981 void VersionDefinitionSection::finalizeContents() {
2982 fileDefNameOff = getPartition().dynStrTab->addString(getFileDefName());
2983 for (const VersionDefinition &v : namedVersionDefs())
2984 verDefNameOffs.push_back(getPartition().dynStrTab->addString(v.name));
2986 if (OutputSection *sec = getPartition().dynStrTab->getParent())
2987 getParent()->link = sec->sectionIndex;
2989 // sh_info should be set to the number of definitions. This fact is missed in
2990 // documentation, but confirmed by binutils community:
2991 // https://sourceware.org/ml/binutils/2014-11/msg00355.html
2992 getParent()->info = getVerDefNum();
2995 void VersionDefinitionSection::writeOne(uint8_t *buf, uint32_t index,
2996 StringRef name, size_t nameOff) {
2997 uint16_t flags = index == 1 ? VER_FLG_BASE : 0;
3000 write16(buf, 1); // vd_version
3001 write16(buf + 2, flags); // vd_flags
3002 write16(buf + 4, index); // vd_ndx
3003 write16(buf + 6, 1); // vd_cnt
3004 write32(buf + 8, hashSysV(name)); // vd_hash
3005 write32(buf + 12, 20); // vd_aux
3006 write32(buf + 16, 28); // vd_next
3009 write32(buf + 20, nameOff); // vda_name
3010 write32(buf + 24, 0); // vda_next
3013 void VersionDefinitionSection::writeTo(uint8_t *buf) {
3014 writeOne(buf, 1, getFileDefName(), fileDefNameOff);
3016 auto nameOffIt = verDefNameOffs.begin();
3017 for (const VersionDefinition &v : namedVersionDefs()) {
3019 writeOne(buf, v.id, v.name, *nameOffIt++);
3022 // Need to terminate the last version definition.
3023 write32(buf + 16, 0); // vd_next
3026 size_t VersionDefinitionSection::getSize() const {
3027 return EntrySize * getVerDefNum();
3030 // .gnu.version is a table where each entry is 2 byte long.
3031 VersionTableSection::VersionTableSection()
3032 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t),
3037 void VersionTableSection::finalizeContents() {
3038 // At the moment of june 2016 GNU docs does not mention that sh_link field
3039 // should be set, but Sun docs do. Also readelf relies on this field.
3040 getParent()->link = getPartition().dynSymTab->getParent()->sectionIndex;
3043 size_t VersionTableSection::getSize() const {
3044 return (getPartition().dynSymTab->getSymbols().size() + 1) * 2;
3047 void VersionTableSection::writeTo(uint8_t *buf) {
3049 for (const SymbolTableEntry &s : getPartition().dynSymTab->getSymbols()) {
3050 write16(buf, s.sym->versionId);
3055 bool VersionTableSection::isNeeded() const {
3057 (getPartition().verDef || getPartition().verNeed->isNeeded());
3060 void elf::addVerneed(Symbol *ss) {
3061 auto &file = cast<SharedFile>(*ss->file);
3062 if (ss->verdefIndex == VER_NDX_GLOBAL) {
3063 ss->versionId = VER_NDX_GLOBAL;
3067 if (file.vernauxs.empty())
3068 file.vernauxs.resize(file.verdefs.size());
3070 // Select a version identifier for the vernaux data structure, if we haven't
3071 // already allocated one. The verdef identifiers cover the range
3072 // [1..getVerDefNum()]; this causes the vernaux identifiers to start from
3073 // getVerDefNum()+1.
3074 if (file.vernauxs[ss->verdefIndex] == 0)
3075 file.vernauxs[ss->verdefIndex] = ++SharedFile::vernauxNum + getVerDefNum();
3077 ss->versionId = file.vernauxs[ss->verdefIndex];
3080 template <class ELFT>
3081 VersionNeedSection<ELFT>::VersionNeedSection()
3082 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t),
3083 ".gnu.version_r") {}
3085 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() {
3086 for (SharedFile *f : sharedFiles) {
3087 if (f->vernauxs.empty())
3089 verneeds.emplace_back();
3090 Verneed &vn = verneeds.back();
3091 vn.nameStrTab = getPartition().dynStrTab->addString(f->soName);
3092 for (unsigned i = 0; i != f->vernauxs.size(); ++i) {
3093 if (f->vernauxs[i] == 0)
3096 reinterpret_cast<const typename ELFT::Verdef *>(f->verdefs[i]);
3097 vn.vernauxs.push_back(
3098 {verdef->vd_hash, f->vernauxs[i],
3099 getPartition().dynStrTab->addString(f->getStringTable().data() +
3100 verdef->getAux()->vda_name)});
3104 if (OutputSection *sec = getPartition().dynStrTab->getParent())
3105 getParent()->link = sec->sectionIndex;
3106 getParent()->info = verneeds.size();
3109 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *buf) {
3110 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs.
3111 auto *verneed = reinterpret_cast<Elf_Verneed *>(buf);
3112 auto *vernaux = reinterpret_cast<Elf_Vernaux *>(verneed + verneeds.size());
3114 for (auto &vn : verneeds) {
3115 // Create an Elf_Verneed for this DSO.
3116 verneed->vn_version = 1;
3117 verneed->vn_cnt = vn.vernauxs.size();
3118 verneed->vn_file = vn.nameStrTab;
3120 reinterpret_cast<char *>(vernaux) - reinterpret_cast<char *>(verneed);
3121 verneed->vn_next = sizeof(Elf_Verneed);
3124 // Create the Elf_Vernauxs for this Elf_Verneed.
3125 for (auto &vna : vn.vernauxs) {
3126 vernaux->vna_hash = vna.hash;
3127 vernaux->vna_flags = 0;
3128 vernaux->vna_other = vna.verneedIndex;
3129 vernaux->vna_name = vna.nameStrTab;
3130 vernaux->vna_next = sizeof(Elf_Vernaux);
3134 vernaux[-1].vna_next = 0;
3136 verneed[-1].vn_next = 0;
3139 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const {
3140 return verneeds.size() * sizeof(Elf_Verneed) +
3141 SharedFile::vernauxNum * sizeof(Elf_Vernaux);
3144 template <class ELFT> bool VersionNeedSection<ELFT>::isNeeded() const {
3145 return isLive() && SharedFile::vernauxNum != 0;
3148 void MergeSyntheticSection::addSection(MergeInputSection *ms) {
3150 sections.push_back(ms);
3151 assert(alignment == ms->alignment || !(ms->flags & SHF_STRINGS));
3152 alignment = std::max(alignment, ms->alignment);
3155 MergeTailSection::MergeTailSection(StringRef name, uint32_t type,
3156 uint64_t flags, uint32_t alignment)
3157 : MergeSyntheticSection(name, type, flags, alignment),
3158 builder(StringTableBuilder::RAW, alignment) {}
3160 size_t MergeTailSection::getSize() const { return builder.getSize(); }
3162 void MergeTailSection::writeTo(uint8_t *buf) { builder.write(buf); }
3164 void MergeTailSection::finalizeContents() {
3165 // Add all string pieces to the string table builder to create section
3167 for (MergeInputSection *sec : sections)
3168 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3169 if (sec->pieces[i].live)
3170 builder.add(sec->getData(i));
3172 // Fix the string table content. After this, the contents will never change.
3175 // finalize() fixed tail-optimized strings, so we can now get
3176 // offsets of strings. Get an offset for each string and save it
3177 // to a corresponding SectionPiece for easy access.
3178 for (MergeInputSection *sec : sections)
3179 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3180 if (sec->pieces[i].live)
3181 sec->pieces[i].outputOff = builder.getOffset(sec->getData(i));
3184 void MergeNoTailSection::writeTo(uint8_t *buf) {
3185 for (size_t i = 0; i < numShards; ++i)
3186 shards[i].write(buf + shardOffsets[i]);
3189 // This function is very hot (i.e. it can take several seconds to finish)
3190 // because sometimes the number of inputs is in an order of magnitude of
3191 // millions. So, we use multi-threading.
3193 // For any strings S and T, we know S is not mergeable with T if S's hash
3194 // value is different from T's. If that's the case, we can safely put S and
3195 // T into different string builders without worrying about merge misses.
3196 // We do it in parallel.
3197 void MergeNoTailSection::finalizeContents() {
3198 // Initializes string table builders.
3199 for (size_t i = 0; i < numShards; ++i)
3200 shards.emplace_back(StringTableBuilder::RAW, alignment);
3202 // Concurrency level. Must be a power of 2 to avoid expensive modulo
3203 // operations in the following tight loop.
3204 size_t concurrency = PowerOf2Floor(
3205 std::min<size_t>(hardware_concurrency(parallel::strategy.ThreadsRequested)
3206 .compute_thread_count(),
3209 // Add section pieces to the builders.
3210 parallelForEachN(0, concurrency, [&](size_t threadId) {
3211 for (MergeInputSection *sec : sections) {
3212 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) {
3213 if (!sec->pieces[i].live)
3215 size_t shardId = getShardId(sec->pieces[i].hash);
3216 if ((shardId & (concurrency - 1)) == threadId)
3217 sec->pieces[i].outputOff = shards[shardId].add(sec->getData(i));
3222 // Compute an in-section offset for each shard.
3224 for (size_t i = 0; i < numShards; ++i) {
3225 shards[i].finalizeInOrder();
3226 if (shards[i].getSize() > 0)
3227 off = alignTo(off, alignment);
3228 shardOffsets[i] = off;
3229 off += shards[i].getSize();
3233 // So far, section pieces have offsets from beginning of shards, but
3234 // we want offsets from beginning of the whole section. Fix them.
3235 parallelForEach(sections, [&](MergeInputSection *sec) {
3236 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3237 if (sec->pieces[i].live)
3238 sec->pieces[i].outputOff +=
3239 shardOffsets[getShardId(sec->pieces[i].hash)];
3243 MergeSyntheticSection *elf::createMergeSynthetic(StringRef name, uint32_t type,
3245 uint32_t alignment) {
3246 bool shouldTailMerge = (flags & SHF_STRINGS) && config->optimize >= 2;
3247 if (shouldTailMerge)
3248 return make<MergeTailSection>(name, type, flags, alignment);
3249 return make<MergeNoTailSection>(name, type, flags, alignment);
3252 template <class ELFT> void elf::splitSections() {
3253 llvm::TimeTraceScope timeScope("Split sections");
3254 // splitIntoPieces needs to be called on each MergeInputSection
3255 // before calling finalizeContents().
3256 parallelForEach(inputSections, [](InputSectionBase *sec) {
3257 if (auto *s = dyn_cast<MergeInputSection>(sec))
3258 s->splitIntoPieces();
3259 else if (auto *eh = dyn_cast<EhInputSection>(sec))
3264 MipsRldMapSection::MipsRldMapSection()
3265 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
3268 ARMExidxSyntheticSection::ARMExidxSyntheticSection()
3269 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX,
3270 config->wordsize, ".ARM.exidx") {}
3272 static InputSection *findExidxSection(InputSection *isec) {
3273 for (InputSection *d : isec->dependentSections)
3274 if (d->type == SHT_ARM_EXIDX && d->isLive())
3279 static bool isValidExidxSectionDep(InputSection *isec) {
3280 return (isec->flags & SHF_ALLOC) && (isec->flags & SHF_EXECINSTR) &&
3281 isec->getSize() > 0;
3284 bool ARMExidxSyntheticSection::addSection(InputSection *isec) {
3285 if (isec->type == SHT_ARM_EXIDX) {
3286 if (InputSection *dep = isec->getLinkOrderDep())
3287 if (isValidExidxSectionDep(dep)) {
3288 exidxSections.push_back(isec);
3289 // Every exidxSection is 8 bytes, we need an estimate of
3290 // size before assignAddresses can be called. Final size
3291 // will only be known after finalize is called.
3297 if (isValidExidxSectionDep(isec)) {
3298 executableSections.push_back(isec);
3302 // FIXME: we do not output a relocation section when --emit-relocs is used
3303 // as we do not have relocation sections for linker generated table entries
3304 // and we would have to erase at a late stage relocations from merged entries.
3305 // Given that exception tables are already position independent and a binary
3306 // analyzer could derive the relocations we choose to erase the relocations.
3307 if (config->emitRelocs && isec->type == SHT_REL)
3308 if (InputSectionBase *ex = isec->getRelocatedSection())
3309 if (isa<InputSection>(ex) && ex->type == SHT_ARM_EXIDX)
3315 // References to .ARM.Extab Sections have bit 31 clear and are not the
3316 // special EXIDX_CANTUNWIND bit-pattern.
3317 static bool isExtabRef(uint32_t unwind) {
3318 return (unwind & 0x80000000) == 0 && unwind != 0x1;
3321 // Return true if the .ARM.exidx section Cur can be merged into the .ARM.exidx
3322 // section Prev, where Cur follows Prev in the table. This can be done if the
3323 // unwinding instructions in Cur are identical to Prev. Linker generated
3324 // EXIDX_CANTUNWIND entries are represented by nullptr as they do not have an
3326 static bool isDuplicateArmExidxSec(InputSection *prev, InputSection *cur) {
3332 // Get the last table Entry from the previous .ARM.exidx section. If Prev is
3333 // nullptr then it will be a synthesized EXIDX_CANTUNWIND entry.
3334 ExidxEntry prevEntry = {ulittle32_t(0), ulittle32_t(1)};
3336 prevEntry = prev->getDataAs<ExidxEntry>().back();
3337 if (isExtabRef(prevEntry.unwind))
3340 // We consider the unwind instructions of an .ARM.exidx table entry
3341 // a duplicate if the previous unwind instructions if:
3342 // - Both are the special EXIDX_CANTUNWIND.
3343 // - Both are the same inline unwind instructions.
3344 // We do not attempt to follow and check links into .ARM.extab tables as
3345 // consecutive identical entries are rare and the effort to check that they
3346 // are identical is high.
3348 // If Cur is nullptr then this is synthesized EXIDX_CANTUNWIND entry.
3350 return prevEntry.unwind == 1;
3352 for (const ExidxEntry entry : cur->getDataAs<ExidxEntry>())
3353 if (isExtabRef(entry.unwind) || entry.unwind != prevEntry.unwind)
3356 // All table entries in this .ARM.exidx Section can be merged into the
3357 // previous Section.
3361 // The .ARM.exidx table must be sorted in ascending order of the address of the
3362 // functions the table describes. Optionally duplicate adjacent table entries
3363 // can be removed. At the end of the function the executableSections must be
3364 // sorted in ascending order of address, Sentinel is set to the InputSection
3365 // with the highest address and any InputSections that have mergeable
3366 // .ARM.exidx table entries are removed from it.
3367 void ARMExidxSyntheticSection::finalizeContents() {
3368 // The executableSections and exidxSections that we use to derive the final
3369 // contents of this SyntheticSection are populated before
3370 // processSectionCommands() and ICF. A /DISCARD/ entry in SECTIONS command or
3371 // ICF may remove executable InputSections and their dependent .ARM.exidx
3372 // section that we recorded earlier.
3373 auto isDiscarded = [](const InputSection *isec) { return !isec->isLive(); };
3374 llvm::erase_if(exidxSections, isDiscarded);
3375 // We need to remove discarded InputSections and InputSections without
3376 // .ARM.exidx sections that if we generated the .ARM.exidx it would be out
3378 auto isDiscardedOrOutOfRange = [this](InputSection *isec) {
3379 if (!isec->isLive())
3381 if (findExidxSection(isec))
3383 int64_t off = static_cast<int64_t>(isec->getVA() - getVA());
3384 return off != llvm::SignExtend64(off, 31);
3386 llvm::erase_if(executableSections, isDiscardedOrOutOfRange);
3388 // Sort the executable sections that may or may not have associated
3389 // .ARM.exidx sections by order of ascending address. This requires the
3390 // relative positions of InputSections and OutputSections to be known.
3391 auto compareByFilePosition = [](const InputSection *a,
3392 const InputSection *b) {
3393 OutputSection *aOut = a->getParent();
3394 OutputSection *bOut = b->getParent();
3397 return aOut->addr < bOut->addr;
3398 return a->outSecOff < b->outSecOff;
3400 llvm::stable_sort(executableSections, compareByFilePosition);
3401 sentinel = executableSections.back();
3402 // Optionally merge adjacent duplicate entries.
3403 if (config->mergeArmExidx) {
3404 std::vector<InputSection *> selectedSections;
3405 selectedSections.reserve(executableSections.size());
3406 selectedSections.push_back(executableSections[0]);
3408 for (size_t i = 1; i < executableSections.size(); ++i) {
3409 InputSection *ex1 = findExidxSection(executableSections[prev]);
3410 InputSection *ex2 = findExidxSection(executableSections[i]);
3411 if (!isDuplicateArmExidxSec(ex1, ex2)) {
3412 selectedSections.push_back(executableSections[i]);
3416 executableSections = std::move(selectedSections);
3421 for (InputSection *isec : executableSections) {
3422 if (InputSection *d = findExidxSection(isec)) {
3423 d->outSecOff = offset;
3424 d->parent = getParent();
3425 offset += d->getSize();
3430 // Size includes Sentinel.
3434 InputSection *ARMExidxSyntheticSection::getLinkOrderDep() const {
3435 return executableSections.front();
3438 // To write the .ARM.exidx table from the ExecutableSections we have three cases
3439 // 1.) The InputSection has a .ARM.exidx InputSection in its dependent sections.
3440 // We write the .ARM.exidx section contents and apply its relocations.
3441 // 2.) The InputSection does not have a dependent .ARM.exidx InputSection. We
3442 // must write the contents of an EXIDX_CANTUNWIND directly. We use the
3443 // start of the InputSection as the purpose of the linker generated
3444 // section is to terminate the address range of the previous entry.
3445 // 3.) A trailing EXIDX_CANTUNWIND sentinel section is required at the end of
3446 // the table to terminate the address range of the final entry.
3447 void ARMExidxSyntheticSection::writeTo(uint8_t *buf) {
3449 const uint8_t cantUnwindData[8] = {0, 0, 0, 0, // PREL31 to target
3450 1, 0, 0, 0}; // EXIDX_CANTUNWIND
3452 uint64_t offset = 0;
3453 for (InputSection *isec : executableSections) {
3454 assert(isec->getParent() != nullptr);
3455 if (InputSection *d = findExidxSection(isec)) {
3456 memcpy(buf + offset, d->data().data(), d->data().size());
3457 d->relocateAlloc(buf, buf + d->getSize());
3458 offset += d->getSize();
3460 // A Linker generated CANTUNWIND section.
3461 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3462 uint64_t s = isec->getVA();
3463 uint64_t p = getVA() + offset;
3464 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p);
3469 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3470 uint64_t s = sentinel->getVA(sentinel->getSize());
3471 uint64_t p = getVA() + offset;
3472 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p);
3473 assert(size == offset + 8);
3476 bool ARMExidxSyntheticSection::isNeeded() const {
3477 return llvm::find_if(exidxSections, [](InputSection *isec) {
3478 return isec->isLive();
3479 }) != exidxSections.end();
3482 bool ARMExidxSyntheticSection::classof(const SectionBase *d) {
3483 return d->kind() == InputSectionBase::Synthetic && d->type == SHT_ARM_EXIDX;
3486 ThunkSection::ThunkSection(OutputSection *os, uint64_t off)
3487 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 4,
3490 this->outSecOff = off;
3493 size_t ThunkSection::getSize() const {
3494 if (roundUpSizeForErrata)
3495 return alignTo(size, 4096);
3499 void ThunkSection::addThunk(Thunk *t) {
3500 thunks.push_back(t);
3501 t->addSymbols(*this);
3504 void ThunkSection::writeTo(uint8_t *buf) {
3505 for (Thunk *t : thunks)
3506 t->writeTo(buf + t->offset);
3509 InputSection *ThunkSection::getTargetInputSection() const {
3512 const Thunk *t = thunks.front();
3513 return t->getTargetInputSection();
3516 bool ThunkSection::assignOffsets() {
3518 for (Thunk *t : thunks) {
3519 off = alignTo(off, t->alignment);
3521 uint32_t size = t->size();
3522 t->getThunkTargetSym()->size = size;
3525 bool changed = off != size;
3530 PPC32Got2Section::PPC32Got2Section()
3531 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 4, ".got2") {}
3533 bool PPC32Got2Section::isNeeded() const {
3534 // See the comment below. This is not needed if there is no other
3536 for (BaseCommand *base : getParent()->sectionCommands)
3537 if (auto *isd = dyn_cast<InputSectionDescription>(base))
3538 for (InputSection *isec : isd->sections)
3544 void PPC32Got2Section::finalizeContents() {
3545 // PPC32 may create multiple GOT sections for -fPIC/-fPIE, one per file in
3546 // .got2 . This function computes outSecOff of each .got2 to be used in
3547 // PPC32PltCallStub::writeTo(). The purpose of this empty synthetic section is
3548 // to collect input sections named ".got2".
3549 uint32_t offset = 0;
3550 for (BaseCommand *base : getParent()->sectionCommands)
3551 if (auto *isd = dyn_cast<InputSectionDescription>(base)) {
3552 for (InputSection *isec : isd->sections) {
3555 isec->file->ppc32Got2OutSecOff = offset;
3556 offset += (uint32_t)isec->getSize();
3561 // If linking position-dependent code then the table will store the addresses
3562 // directly in the binary so the section has type SHT_PROGBITS. If linking
3563 // position-independent code the section has type SHT_NOBITS since it will be
3564 // allocated and filled in by the dynamic linker.
3565 PPC64LongBranchTargetSection::PPC64LongBranchTargetSection()
3566 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
3567 config->isPic ? SHT_NOBITS : SHT_PROGBITS, 8,
3570 uint64_t PPC64LongBranchTargetSection::getEntryVA(const Symbol *sym,
3572 return getVA() + entry_index.find({sym, addend})->second * 8;
3575 Optional<uint32_t> PPC64LongBranchTargetSection::addEntry(const Symbol *sym,
3578 entry_index.try_emplace(std::make_pair(sym, addend), entries.size());
3581 entries.emplace_back(sym, addend);
3582 return res.first->second;
3585 size_t PPC64LongBranchTargetSection::getSize() const {
3586 return entries.size() * 8;
3589 void PPC64LongBranchTargetSection::writeTo(uint8_t *buf) {
3590 // If linking non-pic we have the final addresses of the targets and they get
3591 // written to the table directly. For pic the dynamic linker will allocate
3592 // the section and fill it it.
3596 for (auto entry : entries) {
3597 const Symbol *sym = entry.first;
3598 int64_t addend = entry.second;
3599 assert(sym->getVA());
3600 // Need calls to branch to the local entry-point since a long-branch
3601 // must be a local-call.
3602 write64(buf, sym->getVA(addend) +
3603 getPPC64GlobalEntryToLocalEntryOffset(sym->stOther));
3608 bool PPC64LongBranchTargetSection::isNeeded() const {
3609 // `removeUnusedSyntheticSections()` is called before thunk allocation which
3610 // is too early to determine if this section will be empty or not. We need
3611 // Finalized to keep the section alive until after thunk creation. Finalized
3612 // only gets set to true once `finalizeSections()` is called after thunk
3613 // creation. Because of this, if we don't create any long-branch thunks we end
3614 // up with an empty .branch_lt section in the binary.
3615 return !finalized || !entries.empty();
3618 static uint8_t getAbiVersion() {
3619 // MIPS non-PIC executable gets ABI version 1.
3620 if (config->emachine == EM_MIPS) {
3621 if (!config->isPic && !config->relocatable &&
3622 (config->eflags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC)
3627 if (config->emachine == EM_AMDGPU) {
3628 uint8_t ver = objectFiles[0]->abiVersion;
3629 for (InputFile *file : makeArrayRef(objectFiles).slice(1))
3630 if (file->abiVersion != ver)
3631 error("incompatible ABI version: " + toString(file));
3638 template <typename ELFT> void elf::writeEhdr(uint8_t *buf, Partition &part) {
3639 // For executable segments, the trap instructions are written before writing
3640 // the header. Setting Elf header bytes to zero ensures that any unused bytes
3641 // in header are zero-cleared, instead of having trap instructions.
3642 memset(buf, 0, sizeof(typename ELFT::Ehdr));
3643 memcpy(buf, "\177ELF", 4);
3645 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3646 eHdr->e_ident[EI_CLASS] = config->is64 ? ELFCLASS64 : ELFCLASS32;
3647 eHdr->e_ident[EI_DATA] = config->isLE ? ELFDATA2LSB : ELFDATA2MSB;
3648 eHdr->e_ident[EI_VERSION] = EV_CURRENT;
3649 eHdr->e_ident[EI_OSABI] = config->osabi;
3650 eHdr->e_ident[EI_ABIVERSION] = getAbiVersion();
3651 eHdr->e_machine = config->emachine;
3652 eHdr->e_version = EV_CURRENT;
3653 eHdr->e_flags = config->eflags;
3654 eHdr->e_ehsize = sizeof(typename ELFT::Ehdr);
3655 eHdr->e_phnum = part.phdrs.size();
3656 eHdr->e_shentsize = sizeof(typename ELFT::Shdr);
3658 if (!config->relocatable) {
3659 eHdr->e_phoff = sizeof(typename ELFT::Ehdr);
3660 eHdr->e_phentsize = sizeof(typename ELFT::Phdr);
3664 template <typename ELFT> void elf::writePhdrs(uint8_t *buf, Partition &part) {
3665 // Write the program header table.
3666 auto *hBuf = reinterpret_cast<typename ELFT::Phdr *>(buf);
3667 for (PhdrEntry *p : part.phdrs) {
3668 hBuf->p_type = p->p_type;
3669 hBuf->p_flags = p->p_flags;
3670 hBuf->p_offset = p->p_offset;
3671 hBuf->p_vaddr = p->p_vaddr;
3672 hBuf->p_paddr = p->p_paddr;
3673 hBuf->p_filesz = p->p_filesz;
3674 hBuf->p_memsz = p->p_memsz;
3675 hBuf->p_align = p->p_align;
3680 template <typename ELFT>
3681 PartitionElfHeaderSection<ELFT>::PartitionElfHeaderSection()
3682 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_EHDR, 1, "") {}
3684 template <typename ELFT>
3685 size_t PartitionElfHeaderSection<ELFT>::getSize() const {
3686 return sizeof(typename ELFT::Ehdr);
3689 template <typename ELFT>
3690 void PartitionElfHeaderSection<ELFT>::writeTo(uint8_t *buf) {
3691 writeEhdr<ELFT>(buf, getPartition());
3693 // Loadable partitions are always ET_DYN.
3694 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3695 eHdr->e_type = ET_DYN;
3698 template <typename ELFT>
3699 PartitionProgramHeadersSection<ELFT>::PartitionProgramHeadersSection()
3700 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_PHDR, 1, ".phdrs") {}
3702 template <typename ELFT>
3703 size_t PartitionProgramHeadersSection<ELFT>::getSize() const {
3704 return sizeof(typename ELFT::Phdr) * getPartition().phdrs.size();
3707 template <typename ELFT>
3708 void PartitionProgramHeadersSection<ELFT>::writeTo(uint8_t *buf) {
3709 writePhdrs<ELFT>(buf, getPartition());
3712 PartitionIndexSection::PartitionIndexSection()
3713 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".rodata") {}
3715 size_t PartitionIndexSection::getSize() const {
3716 return 12 * (partitions.size() - 1);
3719 void PartitionIndexSection::finalizeContents() {
3720 for (size_t i = 1; i != partitions.size(); ++i)
3721 partitions[i].nameStrTab = mainPart->dynStrTab->addString(partitions[i].name);
3724 void PartitionIndexSection::writeTo(uint8_t *buf) {
3725 uint64_t va = getVA();
3726 for (size_t i = 1; i != partitions.size(); ++i) {
3727 write32(buf, mainPart->dynStrTab->getVA() + partitions[i].nameStrTab - va);
3728 write32(buf + 4, partitions[i].elfHeader->getVA() - (va + 4));
3730 SyntheticSection *next =
3731 i == partitions.size() - 1 ? in.partEnd : partitions[i + 1].elfHeader;
3732 write32(buf + 8, next->getVA() - partitions[i].elfHeader->getVA());
3741 std::vector<Partition> elf::partitions;
3742 Partition *elf::mainPart;
3744 template GdbIndexSection *GdbIndexSection::create<ELF32LE>();
3745 template GdbIndexSection *GdbIndexSection::create<ELF32BE>();
3746 template GdbIndexSection *GdbIndexSection::create<ELF64LE>();
3747 template GdbIndexSection *GdbIndexSection::create<ELF64BE>();
3749 template void elf::splitSections<ELF32LE>();
3750 template void elf::splitSections<ELF32BE>();
3751 template void elf::splitSections<ELF64LE>();
3752 template void elf::splitSections<ELF64BE>();
3754 template class elf::MipsAbiFlagsSection<ELF32LE>;
3755 template class elf::MipsAbiFlagsSection<ELF32BE>;
3756 template class elf::MipsAbiFlagsSection<ELF64LE>;
3757 template class elf::MipsAbiFlagsSection<ELF64BE>;
3759 template class elf::MipsOptionsSection<ELF32LE>;
3760 template class elf::MipsOptionsSection<ELF32BE>;
3761 template class elf::MipsOptionsSection<ELF64LE>;
3762 template class elf::MipsOptionsSection<ELF64BE>;
3764 template class elf::MipsReginfoSection<ELF32LE>;
3765 template class elf::MipsReginfoSection<ELF32BE>;
3766 template class elf::MipsReginfoSection<ELF64LE>;
3767 template class elf::MipsReginfoSection<ELF64BE>;
3769 template class elf::DynamicSection<ELF32LE>;
3770 template class elf::DynamicSection<ELF32BE>;
3771 template class elf::DynamicSection<ELF64LE>;
3772 template class elf::DynamicSection<ELF64BE>;
3774 template class elf::RelocationSection<ELF32LE>;
3775 template class elf::RelocationSection<ELF32BE>;
3776 template class elf::RelocationSection<ELF64LE>;
3777 template class elf::RelocationSection<ELF64BE>;
3779 template class elf::AndroidPackedRelocationSection<ELF32LE>;
3780 template class elf::AndroidPackedRelocationSection<ELF32BE>;
3781 template class elf::AndroidPackedRelocationSection<ELF64LE>;
3782 template class elf::AndroidPackedRelocationSection<ELF64BE>;
3784 template class elf::RelrSection<ELF32LE>;
3785 template class elf::RelrSection<ELF32BE>;
3786 template class elf::RelrSection<ELF64LE>;
3787 template class elf::RelrSection<ELF64BE>;
3789 template class elf::SymbolTableSection<ELF32LE>;
3790 template class elf::SymbolTableSection<ELF32BE>;
3791 template class elf::SymbolTableSection<ELF64LE>;
3792 template class elf::SymbolTableSection<ELF64BE>;
3794 template class elf::VersionNeedSection<ELF32LE>;
3795 template class elf::VersionNeedSection<ELF32BE>;
3796 template class elf::VersionNeedSection<ELF64LE>;
3797 template class elf::VersionNeedSection<ELF64BE>;
3799 template void elf::writeEhdr<ELF32LE>(uint8_t *Buf, Partition &Part);
3800 template void elf::writeEhdr<ELF32BE>(uint8_t *Buf, Partition &Part);
3801 template void elf::writeEhdr<ELF64LE>(uint8_t *Buf, Partition &Part);
3802 template void elf::writeEhdr<ELF64BE>(uint8_t *Buf, Partition &Part);
3804 template void elf::writePhdrs<ELF32LE>(uint8_t *Buf, Partition &Part);
3805 template void elf::writePhdrs<ELF32BE>(uint8_t *Buf, Partition &Part);
3806 template void elf::writePhdrs<ELF64LE>(uint8_t *Buf, Partition &Part);
3807 template void elf::writePhdrs<ELF64BE>(uint8_t *Buf, Partition &Part);
3809 template class elf::PartitionElfHeaderSection<ELF32LE>;
3810 template class elf::PartitionElfHeaderSection<ELF32BE>;
3811 template class elf::PartitionElfHeaderSection<ELF64LE>;
3812 template class elf::PartitionElfHeaderSection<ELF64BE>;
3814 template class elf::PartitionProgramHeadersSection<ELF32LE>;
3815 template class elf::PartitionProgramHeadersSection<ELF32BE>;
3816 template class elf::PartitionProgramHeadersSection<ELF64LE>;
3817 template class elf::PartitionProgramHeadersSection<ELF64BE>;