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/ErrorHandler.h"
26 #include "lld/Common/Memory.h"
27 #include "lld/Common/Strings.h"
28 #include "lld/Common/Threads.h"
29 #include "lld/Common/Version.h"
30 #include "llvm/ADT/SetOperations.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/BinaryFormat/Dwarf.h"
33 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h"
34 #include "llvm/Object/ELFObjectFile.h"
35 #include "llvm/Support/Compression.h"
36 #include "llvm/Support/Endian.h"
37 #include "llvm/Support/LEB128.h"
38 #include "llvm/Support/MD5.h"
43 using namespace llvm::dwarf;
44 using namespace llvm::ELF;
45 using namespace llvm::object;
46 using namespace llvm::support;
49 using namespace lld::elf;
51 using llvm::support::endian::read32le;
52 using llvm::support::endian::write32le;
53 using llvm::support::endian::write64le;
55 constexpr size_t MergeNoTailSection::numShards;
57 static uint64_t readUint(uint8_t *buf) {
58 return config->is64 ? read64(buf) : read32(buf);
61 static void writeUint(uint8_t *buf, uint64_t val) {
68 // Returns an LLD version string.
69 static ArrayRef<uint8_t> getVersion() {
70 // Check LLD_VERSION first for ease of testing.
71 // You can get consistent output by using the environment variable.
72 // This is only for testing.
73 StringRef s = getenv("LLD_VERSION");
75 s = saver.save(Twine("Linker: ") + getLLDVersion());
77 // +1 to include the terminating '\0'.
78 return {(const uint8_t *)s.data(), s.size() + 1};
81 // Creates a .comment section containing LLD version info.
82 // With this feature, you can identify LLD-generated binaries easily
83 // by "readelf --string-dump .comment <file>".
84 // The returned object is a mergeable string section.
85 MergeInputSection *elf::createCommentSection() {
86 return make<MergeInputSection>(SHF_MERGE | SHF_STRINGS, SHT_PROGBITS, 1,
87 getVersion(), ".comment");
90 // .MIPS.abiflags section.
92 MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags flags)
93 : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags"),
95 this->entsize = sizeof(Elf_Mips_ABIFlags);
98 template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *buf) {
99 memcpy(buf, &flags, sizeof(flags));
102 template <class ELFT>
103 MipsAbiFlagsSection<ELFT> *MipsAbiFlagsSection<ELFT>::create() {
104 Elf_Mips_ABIFlags flags = {};
107 for (InputSectionBase *sec : inputSections) {
108 if (sec->type != SHT_MIPS_ABIFLAGS)
113 std::string filename = toString(sec->file);
114 const size_t size = sec->data().size();
115 // Older version of BFD (such as the default FreeBSD linker) concatenate
116 // .MIPS.abiflags instead of merging. To allow for this case (or potential
117 // zero padding) we ignore everything after the first Elf_Mips_ABIFlags
118 if (size < sizeof(Elf_Mips_ABIFlags)) {
119 error(filename + ": invalid size of .MIPS.abiflags section: got " +
120 Twine(size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags)));
123 auto *s = reinterpret_cast<const Elf_Mips_ABIFlags *>(sec->data().data());
124 if (s->version != 0) {
125 error(filename + ": unexpected .MIPS.abiflags version " +
130 // LLD checks ISA compatibility in calcMipsEFlags(). Here we just
131 // select the highest number of ISA/Rev/Ext.
132 flags.isa_level = std::max(flags.isa_level, s->isa_level);
133 flags.isa_rev = std::max(flags.isa_rev, s->isa_rev);
134 flags.isa_ext = std::max(flags.isa_ext, s->isa_ext);
135 flags.gpr_size = std::max(flags.gpr_size, s->gpr_size);
136 flags.cpr1_size = std::max(flags.cpr1_size, s->cpr1_size);
137 flags.cpr2_size = std::max(flags.cpr2_size, s->cpr2_size);
138 flags.ases |= s->ases;
139 flags.flags1 |= s->flags1;
140 flags.flags2 |= s->flags2;
141 flags.fp_abi = elf::getMipsFpAbiFlag(flags.fp_abi, s->fp_abi, filename);
145 return make<MipsAbiFlagsSection<ELFT>>(flags);
149 // .MIPS.options section.
150 template <class ELFT>
151 MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo reginfo)
152 : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"),
154 this->entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo);
157 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *buf) {
158 auto *options = reinterpret_cast<Elf_Mips_Options *>(buf);
159 options->kind = ODK_REGINFO;
160 options->size = getSize();
162 if (!config->relocatable)
163 reginfo.ri_gp_value = in.mipsGot->getGp();
164 memcpy(buf + sizeof(Elf_Mips_Options), ®info, sizeof(reginfo));
167 template <class ELFT>
168 MipsOptionsSection<ELFT> *MipsOptionsSection<ELFT>::create() {
173 std::vector<InputSectionBase *> sections;
174 for (InputSectionBase *sec : inputSections)
175 if (sec->type == SHT_MIPS_OPTIONS)
176 sections.push_back(sec);
178 if (sections.empty())
181 Elf_Mips_RegInfo reginfo = {};
182 for (InputSectionBase *sec : sections) {
185 std::string filename = toString(sec->file);
186 ArrayRef<uint8_t> d = sec->data();
189 if (d.size() < sizeof(Elf_Mips_Options)) {
190 error(filename + ": invalid size of .MIPS.options section");
194 auto *opt = reinterpret_cast<const Elf_Mips_Options *>(d.data());
195 if (opt->kind == ODK_REGINFO) {
196 reginfo.ri_gprmask |= opt->getRegInfo().ri_gprmask;
197 sec->getFile<ELFT>()->mipsGp0 = opt->getRegInfo().ri_gp_value;
202 fatal(filename + ": zero option descriptor size");
203 d = d.slice(opt->size);
207 return make<MipsOptionsSection<ELFT>>(reginfo);
210 // MIPS .reginfo section.
211 template <class ELFT>
212 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo reginfo)
213 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"),
215 this->entsize = sizeof(Elf_Mips_RegInfo);
218 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *buf) {
219 if (!config->relocatable)
220 reginfo.ri_gp_value = in.mipsGot->getGp();
221 memcpy(buf, ®info, sizeof(reginfo));
224 template <class ELFT>
225 MipsReginfoSection<ELFT> *MipsReginfoSection<ELFT>::create() {
226 // Section should be alive for O32 and N32 ABIs only.
230 std::vector<InputSectionBase *> sections;
231 for (InputSectionBase *sec : inputSections)
232 if (sec->type == SHT_MIPS_REGINFO)
233 sections.push_back(sec);
235 if (sections.empty())
238 Elf_Mips_RegInfo reginfo = {};
239 for (InputSectionBase *sec : sections) {
242 if (sec->data().size() != sizeof(Elf_Mips_RegInfo)) {
243 error(toString(sec->file) + ": invalid size of .reginfo section");
247 auto *r = reinterpret_cast<const Elf_Mips_RegInfo *>(sec->data().data());
248 reginfo.ri_gprmask |= r->ri_gprmask;
249 sec->getFile<ELFT>()->mipsGp0 = r->ri_gp_value;
252 return make<MipsReginfoSection<ELFT>>(reginfo);
255 InputSection *elf::createInterpSection() {
256 // StringSaver guarantees that the returned string ends with '\0'.
257 StringRef s = saver.save(config->dynamicLinker);
258 ArrayRef<uint8_t> contents = {(const uint8_t *)s.data(), s.size() + 1};
260 auto *sec = 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, 4,
305 ".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::addSectionAux(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> void EhFrameSection::addSection(InputSectionBase *c) {
432 auto *sec = cast<EhInputSection>(c);
435 alignment = std::max(alignment, sec->alignment);
436 sections.push_back(sec);
438 for (auto *ds : sec->dependentSections)
439 dependentSections.push_back(ds);
441 if (sec->pieces.empty())
444 if (sec->areRelocsRela)
445 addSectionAux<ELFT>(sec, sec->template relas<ELFT>());
447 addSectionAux<ELFT>(sec, sec->template rels<ELFT>());
450 static void writeCieFde(uint8_t *buf, ArrayRef<uint8_t> d) {
451 memcpy(buf, d.data(), d.size());
453 size_t aligned = alignTo(d.size(), config->wordsize);
455 // Zero-clear trailing padding if it exists.
456 memset(buf + d.size(), 0, aligned - d.size());
458 // Fix the size field. -4 since size does not include the size field itself.
459 write32(buf, aligned - 4);
462 void EhFrameSection::finalizeContents() {
463 assert(!this->size); // Not finalized.
465 for (CieRecord *rec : cieRecords) {
466 rec->cie->outputOff = off;
467 off += alignTo(rec->cie->size, config->wordsize);
469 for (EhSectionPiece *fde : rec->fdes) {
470 fde->outputOff = off;
471 off += alignTo(fde->size, config->wordsize);
475 // The LSB standard does not allow a .eh_frame section with zero
476 // Call Frame Information records. glibc unwind-dw2-fde.c
477 // classify_object_over_fdes expects there is a CIE record length 0 as a
478 // terminator. Thus we add one unconditionally.
484 // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table
485 // to get an FDE from an address to which FDE is applied. This function
486 // returns a list of such pairs.
487 std::vector<EhFrameSection::FdeData> EhFrameSection::getFdeData() const {
488 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
489 std::vector<FdeData> ret;
491 uint64_t va = getPartition().ehFrameHdr->getVA();
492 for (CieRecord *rec : cieRecords) {
493 uint8_t enc = getFdeEncoding(rec->cie);
494 for (EhSectionPiece *fde : rec->fdes) {
495 uint64_t pc = getFdePc(buf, fde->outputOff, enc);
496 uint64_t fdeVA = getParent()->addr + fde->outputOff;
497 if (!isInt<32>(pc - va))
498 fatal(toString(fde->sec) + ": PC offset is too large: 0x" +
499 Twine::utohexstr(pc - va));
500 ret.push_back({uint32_t(pc - va), uint32_t(fdeVA - va)});
504 // Sort the FDE list by their PC and uniqueify. Usually there is only
505 // one FDE for a PC (i.e. function), but if ICF merges two functions
506 // into one, there can be more than one FDEs pointing to the address.
507 auto less = [](const FdeData &a, const FdeData &b) {
508 return a.pcRel < b.pcRel;
510 llvm::stable_sort(ret, less);
511 auto eq = [](const FdeData &a, const FdeData &b) {
512 return a.pcRel == b.pcRel;
514 ret.erase(std::unique(ret.begin(), ret.end(), eq), ret.end());
519 static uint64_t readFdeAddr(uint8_t *buf, int size) {
521 case DW_EH_PE_udata2:
523 case DW_EH_PE_sdata2:
524 return (int16_t)read16(buf);
525 case DW_EH_PE_udata4:
527 case DW_EH_PE_sdata4:
528 return (int32_t)read32(buf);
529 case DW_EH_PE_udata8:
530 case DW_EH_PE_sdata8:
532 case DW_EH_PE_absptr:
533 return readUint(buf);
535 fatal("unknown FDE size encoding");
538 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to.
539 // We need it to create .eh_frame_hdr section.
540 uint64_t EhFrameSection::getFdePc(uint8_t *buf, size_t fdeOff,
542 // The starting address to which this FDE applies is
543 // stored at FDE + 8 byte.
544 size_t off = fdeOff + 8;
545 uint64_t addr = readFdeAddr(buf + off, enc & 0xf);
546 if ((enc & 0x70) == DW_EH_PE_absptr)
548 if ((enc & 0x70) == DW_EH_PE_pcrel)
549 return addr + getParent()->addr + off;
550 fatal("unknown FDE size relative encoding");
553 void EhFrameSection::writeTo(uint8_t *buf) {
554 // Write CIE and FDE records.
555 for (CieRecord *rec : cieRecords) {
556 size_t cieOffset = rec->cie->outputOff;
557 writeCieFde(buf + cieOffset, rec->cie->data());
559 for (EhSectionPiece *fde : rec->fdes) {
560 size_t off = fde->outputOff;
561 writeCieFde(buf + off, fde->data());
563 // FDE's second word should have the offset to an associated CIE.
565 write32(buf + off + 4, off + 4 - cieOffset);
569 // Apply relocations. .eh_frame section contents are not contiguous
570 // in the output buffer, but relocateAlloc() still works because
571 // getOffset() takes care of discontiguous section pieces.
572 for (EhInputSection *s : sections)
573 s->relocateAlloc(buf, nullptr);
575 if (getPartition().ehFrameHdr && getPartition().ehFrameHdr->getParent())
576 getPartition().ehFrameHdr->write();
579 GotSection::GotSection()
580 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
582 // If ElfSym::globalOffsetTable is relative to .got and is referenced,
583 // increase numEntries by the number of entries used to emit
584 // ElfSym::globalOffsetTable.
585 if (ElfSym::globalOffsetTable && !target->gotBaseSymInGotPlt)
586 numEntries += target->gotHeaderEntriesNum;
589 void GotSection::addEntry(Symbol &sym) {
590 sym.gotIndex = numEntries;
594 bool GotSection::addDynTlsEntry(Symbol &sym) {
595 if (sym.globalDynIndex != -1U)
597 sym.globalDynIndex = numEntries;
598 // Global Dynamic TLS entries take two GOT slots.
603 // Reserves TLS entries for a TLS module ID and a TLS block offset.
604 // In total it takes two GOT slots.
605 bool GotSection::addTlsIndex() {
606 if (tlsIndexOff != uint32_t(-1))
608 tlsIndexOff = numEntries * config->wordsize;
613 uint64_t GotSection::getGlobalDynAddr(const Symbol &b) const {
614 return this->getVA() + b.globalDynIndex * config->wordsize;
617 uint64_t GotSection::getGlobalDynOffset(const Symbol &b) const {
618 return b.globalDynIndex * config->wordsize;
621 void GotSection::finalizeContents() {
622 size = numEntries * config->wordsize;
625 bool GotSection::isNeeded() const {
626 // We need to emit a GOT even if it's empty if there's a relocation that is
627 // relative to GOT(such as GOTOFFREL).
628 return numEntries || hasGotOffRel;
631 void GotSection::writeTo(uint8_t *buf) {
632 // Buf points to the start of this section's buffer,
633 // whereas InputSectionBase::relocateAlloc() expects its argument
634 // to point to the start of the output section.
635 target->writeGotHeader(buf);
636 relocateAlloc(buf - outSecOff, buf - outSecOff + size);
639 static uint64_t getMipsPageAddr(uint64_t addr) {
640 return (addr + 0x8000) & ~0xffff;
643 static uint64_t getMipsPageCount(uint64_t size) {
644 return (size + 0xfffe) / 0xffff + 1;
647 MipsGotSection::MipsGotSection()
648 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16,
651 void MipsGotSection::addEntry(InputFile &file, Symbol &sym, int64_t addend,
653 FileGot &g = getGot(file);
654 if (expr == R_MIPS_GOT_LOCAL_PAGE) {
655 if (const OutputSection *os = sym.getOutputSection())
656 g.pagesMap.insert({os, {}});
658 g.local16.insert({{nullptr, getMipsPageAddr(sym.getVA(addend))}, 0});
659 } else if (sym.isTls())
660 g.tls.insert({&sym, 0});
661 else if (sym.isPreemptible && expr == R_ABS)
662 g.relocs.insert({&sym, 0});
663 else if (sym.isPreemptible)
664 g.global.insert({&sym, 0});
665 else if (expr == R_MIPS_GOT_OFF32)
666 g.local32.insert({{&sym, addend}, 0});
668 g.local16.insert({{&sym, addend}, 0});
671 void MipsGotSection::addDynTlsEntry(InputFile &file, Symbol &sym) {
672 getGot(file).dynTlsSymbols.insert({&sym, 0});
675 void MipsGotSection::addTlsIndex(InputFile &file) {
676 getGot(file).dynTlsSymbols.insert({nullptr, 0});
679 size_t MipsGotSection::FileGot::getEntriesNum() const {
680 return getPageEntriesNum() + local16.size() + global.size() + relocs.size() +
681 tls.size() + dynTlsSymbols.size() * 2;
684 size_t MipsGotSection::FileGot::getPageEntriesNum() const {
686 for (const std::pair<const OutputSection *, FileGot::PageBlock> &p : pagesMap)
687 num += p.second.count;
691 size_t MipsGotSection::FileGot::getIndexedEntriesNum() const {
692 size_t count = getPageEntriesNum() + local16.size() + global.size();
693 // If there are relocation-only entries in the GOT, TLS entries
694 // are allocated after them. TLS entries should be addressable
695 // by 16-bit index so count both reloc-only and TLS entries.
696 if (!tls.empty() || !dynTlsSymbols.empty())
697 count += relocs.size() + tls.size() + dynTlsSymbols.size() * 2;
701 MipsGotSection::FileGot &MipsGotSection::getGot(InputFile &f) {
702 if (!f.mipsGotIndex.hasValue()) {
704 gots.back().file = &f;
705 f.mipsGotIndex = gots.size() - 1;
707 return gots[*f.mipsGotIndex];
710 uint64_t MipsGotSection::getPageEntryOffset(const InputFile *f,
712 int64_t addend) const {
713 const FileGot &g = gots[*f->mipsGotIndex];
715 if (const OutputSection *outSec = sym.getOutputSection()) {
716 uint64_t secAddr = getMipsPageAddr(outSec->addr);
717 uint64_t symAddr = getMipsPageAddr(sym.getVA(addend));
718 index = g.pagesMap.lookup(outSec).firstIndex + (symAddr - secAddr) / 0xffff;
720 index = g.local16.lookup({nullptr, getMipsPageAddr(sym.getVA(addend))});
722 return index * config->wordsize;
725 uint64_t MipsGotSection::getSymEntryOffset(const InputFile *f, const Symbol &s,
726 int64_t addend) const {
727 const FileGot &g = gots[*f->mipsGotIndex];
728 Symbol *sym = const_cast<Symbol *>(&s);
730 return g.tls.lookup(sym) * config->wordsize;
731 if (sym->isPreemptible)
732 return g.global.lookup(sym) * config->wordsize;
733 return g.local16.lookup({sym, addend}) * config->wordsize;
736 uint64_t MipsGotSection::getTlsIndexOffset(const InputFile *f) const {
737 const FileGot &g = gots[*f->mipsGotIndex];
738 return g.dynTlsSymbols.lookup(nullptr) * config->wordsize;
741 uint64_t MipsGotSection::getGlobalDynOffset(const InputFile *f,
742 const Symbol &s) const {
743 const FileGot &g = gots[*f->mipsGotIndex];
744 Symbol *sym = const_cast<Symbol *>(&s);
745 return g.dynTlsSymbols.lookup(sym) * config->wordsize;
748 const Symbol *MipsGotSection::getFirstGlobalEntry() const {
751 const FileGot &primGot = gots.front();
752 if (!primGot.global.empty())
753 return primGot.global.front().first;
754 if (!primGot.relocs.empty())
755 return primGot.relocs.front().first;
759 unsigned MipsGotSection::getLocalEntriesNum() const {
761 return headerEntriesNum;
762 return headerEntriesNum + gots.front().getPageEntriesNum() +
763 gots.front().local16.size();
766 bool MipsGotSection::tryMergeGots(FileGot &dst, FileGot &src, bool isPrimary) {
768 set_union(tmp.pagesMap, src.pagesMap);
769 set_union(tmp.local16, src.local16);
770 set_union(tmp.global, src.global);
771 set_union(tmp.relocs, src.relocs);
772 set_union(tmp.tls, src.tls);
773 set_union(tmp.dynTlsSymbols, src.dynTlsSymbols);
775 size_t count = isPrimary ? headerEntriesNum : 0;
776 count += tmp.getIndexedEntriesNum();
778 if (count * config->wordsize > config->mipsGotSize)
785 void MipsGotSection::finalizeContents() { updateAllocSize(); }
787 bool MipsGotSection::updateAllocSize() {
788 size = headerEntriesNum * config->wordsize;
789 for (const FileGot &g : gots)
790 size += g.getEntriesNum() * config->wordsize;
794 void MipsGotSection::build() {
798 std::vector<FileGot> mergedGots(1);
800 // For each GOT move non-preemptible symbols from the `Global`
801 // to `Local16` list. Preemptible symbol might become non-preemptible
802 // one if, for example, it gets a related copy relocation.
803 for (FileGot &got : gots) {
804 for (auto &p: got.global)
805 if (!p.first->isPreemptible)
806 got.local16.insert({{p.first, 0}, 0});
807 got.global.remove_if([&](const std::pair<Symbol *, size_t> &p) {
808 return !p.first->isPreemptible;
812 // For each GOT remove "reloc-only" entry if there is "global"
813 // entry for the same symbol. And add local entries which indexed
814 // using 32-bit value at the end of 16-bit entries.
815 for (FileGot &got : gots) {
816 got.relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
817 return got.global.count(p.first);
819 set_union(got.local16, got.local32);
823 // Evaluate number of "reloc-only" entries in the resulting GOT.
824 // To do that put all unique "reloc-only" and "global" entries
825 // from all GOTs to the future primary GOT.
826 FileGot *primGot = &mergedGots.front();
827 for (FileGot &got : gots) {
828 set_union(primGot->relocs, got.global);
829 set_union(primGot->relocs, got.relocs);
833 // Evaluate number of "page" entries in each GOT.
834 for (FileGot &got : gots) {
835 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
837 const OutputSection *os = p.first;
838 uint64_t secSize = 0;
839 for (BaseCommand *cmd : os->sectionCommands) {
840 if (auto *isd = dyn_cast<InputSectionDescription>(cmd))
841 for (InputSection *isec : isd->sections) {
842 uint64_t off = alignTo(secSize, isec->alignment);
843 secSize = off + isec->getSize();
846 p.second.count = getMipsPageCount(secSize);
850 // Merge GOTs. Try to join as much as possible GOTs but do not exceed
851 // maximum GOT size. At first, try to fill the primary GOT because
852 // the primary GOT can be accessed in the most effective way. If it
853 // is not possible, try to fill the last GOT in the list, and finally
854 // create a new GOT if both attempts failed.
855 for (FileGot &srcGot : gots) {
856 InputFile *file = srcGot.file;
857 if (tryMergeGots(mergedGots.front(), srcGot, true)) {
858 file->mipsGotIndex = 0;
860 // If this is the first time we failed to merge with the primary GOT,
861 // MergedGots.back() will also be the primary GOT. We must make sure not
862 // to try to merge again with isPrimary=false, as otherwise, if the
863 // inputs are just right, we could allow the primary GOT to become 1 or 2
864 // words bigger due to ignoring the header size.
865 if (mergedGots.size() == 1 ||
866 !tryMergeGots(mergedGots.back(), srcGot, false)) {
867 mergedGots.emplace_back();
868 std::swap(mergedGots.back(), srcGot);
870 file->mipsGotIndex = mergedGots.size() - 1;
873 std::swap(gots, mergedGots);
875 // Reduce number of "reloc-only" entries in the primary GOT
876 // by substracting "global" entries exist in the primary GOT.
877 primGot = &gots.front();
878 primGot->relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
879 return primGot->global.count(p.first);
882 // Calculate indexes for each GOT entry.
883 size_t index = headerEntriesNum;
884 for (FileGot &got : gots) {
885 got.startIndex = &got == primGot ? 0 : index;
886 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
888 // For each output section referenced by GOT page relocations calculate
889 // and save into pagesMap an upper bound of MIPS GOT entries required
890 // to store page addresses of local symbols. We assume the worst case -
891 // each 64kb page of the output section has at least one GOT relocation
892 // against it. And take in account the case when the section intersects
894 p.second.firstIndex = index;
895 index += p.second.count;
897 for (auto &p: got.local16)
899 for (auto &p: got.global)
901 for (auto &p: got.relocs)
903 for (auto &p: got.tls)
905 for (auto &p: got.dynTlsSymbols) {
911 // Update Symbol::gotIndex field to use this
912 // value later in the `sortMipsSymbols` function.
913 for (auto &p : primGot->global)
914 p.first->gotIndex = p.second;
915 for (auto &p : primGot->relocs)
916 p.first->gotIndex = p.second;
918 // Create dynamic relocations.
919 for (FileGot &got : gots) {
920 // Create dynamic relocations for TLS entries.
921 for (std::pair<Symbol *, size_t> &p : got.tls) {
923 uint64_t offset = p.second * config->wordsize;
924 if (s->isPreemptible)
925 mainPart->relaDyn->addReloc(target->tlsGotRel, this, offset, s);
927 for (std::pair<Symbol *, size_t> &p : got.dynTlsSymbols) {
929 uint64_t offset = p.second * config->wordsize;
933 mainPart->relaDyn->addReloc(target->tlsModuleIndexRel, this, offset, s);
935 // When building a shared library we still need a dynamic relocation
936 // for the module index. Therefore only checking for
937 // S->isPreemptible is not sufficient (this happens e.g. for
938 // thread-locals that have been marked as local through a linker script)
939 if (!s->isPreemptible && !config->isPic)
941 mainPart->relaDyn->addReloc(target->tlsModuleIndexRel, this, offset, s);
942 // However, we can skip writing the TLS offset reloc for non-preemptible
943 // symbols since it is known even in shared libraries
944 if (!s->isPreemptible)
946 offset += config->wordsize;
947 mainPart->relaDyn->addReloc(target->tlsOffsetRel, this, offset, s);
951 // Do not create dynamic relocations for non-TLS
952 // entries in the primary GOT.
956 // Dynamic relocations for "global" entries.
957 for (const std::pair<Symbol *, size_t> &p : got.global) {
958 uint64_t offset = p.second * config->wordsize;
959 mainPart->relaDyn->addReloc(target->relativeRel, this, offset, p.first);
963 // Dynamic relocations for "local" entries in case of PIC.
964 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
966 size_t pageCount = l.second.count;
967 for (size_t pi = 0; pi < pageCount; ++pi) {
968 uint64_t offset = (l.second.firstIndex + pi) * config->wordsize;
969 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, l.first,
970 int64_t(pi * 0x10000)});
973 for (const std::pair<GotEntry, size_t> &p : got.local16) {
974 uint64_t offset = p.second * config->wordsize;
975 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, true,
976 p.first.first, p.first.second});
981 bool MipsGotSection::isNeeded() const {
982 // We add the .got section to the result for dynamic MIPS target because
983 // its address and properties are mentioned in the .dynamic section.
984 return !config->relocatable;
987 uint64_t MipsGotSection::getGp(const InputFile *f) const {
988 // For files without related GOT or files refer a primary GOT
989 // returns "common" _gp value. For secondary GOTs calculate
990 // individual _gp values.
991 if (!f || !f->mipsGotIndex.hasValue() || *f->mipsGotIndex == 0)
992 return ElfSym::mipsGp->getVA(0);
993 return getVA() + gots[*f->mipsGotIndex].startIndex * config->wordsize +
997 void MipsGotSection::writeTo(uint8_t *buf) {
998 // Set the MSB of the second GOT slot. This is not required by any
999 // MIPS ABI documentation, though.
1001 // There is a comment in glibc saying that "The MSB of got[1] of a
1002 // gnu object is set to identify gnu objects," and in GNU gold it
1003 // says "the second entry will be used by some runtime loaders".
1004 // But how this field is being used is unclear.
1006 // We are not really willing to mimic other linkers behaviors
1007 // without understanding why they do that, but because all files
1008 // generated by GNU tools have this special GOT value, and because
1009 // we've been doing this for years, it is probably a safe bet to
1010 // keep doing this for now. We really need to revisit this to see
1011 // if we had to do this.
1012 writeUint(buf + config->wordsize, (uint64_t)1 << (config->wordsize * 8 - 1));
1013 for (const FileGot &g : gots) {
1014 auto write = [&](size_t i, const Symbol *s, int64_t a) {
1018 writeUint(buf + i * config->wordsize, va);
1020 // Write 'page address' entries to the local part of the GOT.
1021 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
1023 size_t pageCount = l.second.count;
1024 uint64_t firstPageAddr = getMipsPageAddr(l.first->addr);
1025 for (size_t pi = 0; pi < pageCount; ++pi)
1026 write(l.second.firstIndex + pi, nullptr, firstPageAddr + pi * 0x10000);
1028 // Local, global, TLS, reloc-only entries.
1029 // If TLS entry has a corresponding dynamic relocations, leave it
1030 // initialized by zero. Write down adjusted TLS symbol's values otherwise.
1031 // To calculate the adjustments use offsets for thread-local storage.
1032 // https://www.linux-mips.org/wiki/NPTL
1033 for (const std::pair<GotEntry, size_t> &p : g.local16)
1034 write(p.second, p.first.first, p.first.second);
1035 // Write VA to the primary GOT only. For secondary GOTs that
1036 // will be done by REL32 dynamic relocations.
1037 if (&g == &gots.front())
1038 for (const std::pair<const Symbol *, size_t> &p : g.global)
1039 write(p.second, p.first, 0);
1040 for (const std::pair<Symbol *, size_t> &p : g.relocs)
1041 write(p.second, p.first, 0);
1042 for (const std::pair<Symbol *, size_t> &p : g.tls)
1043 write(p.second, p.first, p.first->isPreemptible ? 0 : -0x7000);
1044 for (const std::pair<Symbol *, size_t> &p : g.dynTlsSymbols) {
1045 if (p.first == nullptr && !config->isPic)
1046 write(p.second, nullptr, 1);
1047 else if (p.first && !p.first->isPreemptible) {
1048 // If we are emitting PIC code with relocations we mustn't write
1049 // anything to the GOT here. When using Elf_Rel relocations the value
1050 // one will be treated as an addend and will cause crashes at runtime
1052 write(p.second, nullptr, 1);
1053 write(p.second + 1, p.first, -0x8000);
1059 // On PowerPC the .plt section is used to hold the table of function addresses
1060 // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss
1061 // section. I don't know why we have a BSS style type for the section but it is
1062 // consitent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI.
1063 GotPltSection::GotPltSection()
1064 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
1066 if (config->emachine == EM_PPC) {
1068 } else if (config->emachine == EM_PPC64) {
1074 void GotPltSection::addEntry(Symbol &sym) {
1075 assert(sym.pltIndex == entries.size());
1076 entries.push_back(&sym);
1079 size_t GotPltSection::getSize() const {
1080 return (target->gotPltHeaderEntriesNum + entries.size()) * config->wordsize;
1083 void GotPltSection::writeTo(uint8_t *buf) {
1084 target->writeGotPltHeader(buf);
1085 buf += target->gotPltHeaderEntriesNum * config->wordsize;
1086 for (const Symbol *b : entries) {
1087 target->writeGotPlt(buf, *b);
1088 buf += config->wordsize;
1092 bool GotPltSection::isNeeded() const {
1093 // We need to emit GOTPLT even if it's empty if there's a relocation relative
1095 return !entries.empty() || hasGotPltOffRel;
1098 static StringRef getIgotPltName() {
1099 // On ARM the IgotPltSection is part of the GotSection.
1100 if (config->emachine == EM_ARM)
1103 // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection
1104 // needs to be named the same.
1105 if (config->emachine == EM_PPC64)
1111 // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit
1112 // with the IgotPltSection.
1113 IgotPltSection::IgotPltSection()
1114 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
1115 config->emachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS,
1116 config->wordsize, getIgotPltName()) {}
1118 void IgotPltSection::addEntry(Symbol &sym) {
1119 assert(sym.pltIndex == entries.size());
1120 entries.push_back(&sym);
1123 size_t IgotPltSection::getSize() const {
1124 return entries.size() * config->wordsize;
1127 void IgotPltSection::writeTo(uint8_t *buf) {
1128 for (const Symbol *b : entries) {
1129 target->writeIgotPlt(buf, *b);
1130 buf += config->wordsize;
1134 StringTableSection::StringTableSection(StringRef name, bool dynamic)
1135 : SyntheticSection(dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, name),
1137 // ELF string tables start with a NUL byte.
1141 // Adds a string to the string table. If `hashIt` is true we hash and check for
1142 // duplicates. It is optional because the name of global symbols are already
1143 // uniqued and hashing them again has a big cost for a small value: uniquing
1144 // them with some other string that happens to be the same.
1145 unsigned StringTableSection::addString(StringRef s, bool hashIt) {
1147 auto r = stringMap.insert(std::make_pair(s, this->size));
1149 return r.first->second;
1151 unsigned ret = this->size;
1152 this->size = this->size + s.size() + 1;
1153 strings.push_back(s);
1157 void StringTableSection::writeTo(uint8_t *buf) {
1158 for (StringRef s : strings) {
1159 memcpy(buf, s.data(), s.size());
1160 buf[s.size()] = '\0';
1161 buf += s.size() + 1;
1165 // Returns the number of version definition entries. Because the first entry
1166 // is for the version definition itself, it is the number of versioned symbols
1167 // plus one. Note that we don't support multiple versions yet.
1168 static unsigned getVerDefNum() { return config->versionDefinitions.size() + 1; }
1170 template <class ELFT>
1171 DynamicSection<ELFT>::DynamicSection()
1172 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, config->wordsize,
1174 this->entsize = ELFT::Is64Bits ? 16 : 8;
1176 // .dynamic section is not writable on MIPS and on Fuchsia OS
1177 // which passes -z rodynamic.
1178 // See "Special Section" in Chapter 4 in the following document:
1179 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1180 if (config->emachine == EM_MIPS || config->zRodynamic)
1181 this->flags = SHF_ALLOC;
1184 template <class ELFT>
1185 void DynamicSection<ELFT>::add(int32_t tag, std::function<uint64_t()> fn) {
1186 entries.push_back({tag, fn});
1189 template <class ELFT>
1190 void DynamicSection<ELFT>::addInt(int32_t tag, uint64_t val) {
1191 entries.push_back({tag, [=] { return val; }});
1194 template <class ELFT>
1195 void DynamicSection<ELFT>::addInSec(int32_t tag, InputSection *sec) {
1196 entries.push_back({tag, [=] { return sec->getVA(0); }});
1199 template <class ELFT>
1200 void DynamicSection<ELFT>::addInSecRelative(int32_t tag, InputSection *sec) {
1201 size_t tagOffset = entries.size() * entsize;
1203 {tag, [=] { return sec->getVA(0) - (getVA() + tagOffset); }});
1206 template <class ELFT>
1207 void DynamicSection<ELFT>::addOutSec(int32_t tag, OutputSection *sec) {
1208 entries.push_back({tag, [=] { return sec->addr; }});
1211 template <class ELFT>
1212 void DynamicSection<ELFT>::addSize(int32_t tag, OutputSection *sec) {
1213 entries.push_back({tag, [=] { return sec->size; }});
1216 template <class ELFT>
1217 void DynamicSection<ELFT>::addSym(int32_t tag, Symbol *sym) {
1218 entries.push_back({tag, [=] { return sym->getVA(); }});
1221 // A Linker script may assign the RELA relocation sections to the same
1222 // output section. When this occurs we cannot just use the OutputSection
1223 // Size. Moreover the [DT_JMPREL, DT_JMPREL + DT_PLTRELSZ) is permitted to
1224 // overlap with the [DT_RELA, DT_RELA + DT_RELASZ).
1225 static uint64_t addPltRelSz() {
1226 size_t size = in.relaPlt->getSize();
1227 if (in.relaIplt->getParent() == in.relaPlt->getParent() &&
1228 in.relaIplt->name == in.relaPlt->name)
1229 size += in.relaIplt->getSize();
1233 // Add remaining entries to complete .dynamic contents.
1234 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() {
1235 elf::Partition &part = getPartition();
1236 bool isMain = part.name.empty();
1238 for (StringRef s : config->filterList)
1239 addInt(DT_FILTER, part.dynStrTab->addString(s));
1240 for (StringRef s : config->auxiliaryList)
1241 addInt(DT_AUXILIARY, part.dynStrTab->addString(s));
1243 if (!config->rpath.empty())
1244 addInt(config->enableNewDtags ? DT_RUNPATH : DT_RPATH,
1245 part.dynStrTab->addString(config->rpath));
1247 for (SharedFile *file : sharedFiles)
1249 addInt(DT_NEEDED, part.dynStrTab->addString(file->soName));
1252 if (!config->soName.empty())
1253 addInt(DT_SONAME, part.dynStrTab->addString(config->soName));
1255 if (!config->soName.empty())
1256 addInt(DT_NEEDED, part.dynStrTab->addString(config->soName));
1257 addInt(DT_SONAME, part.dynStrTab->addString(part.name));
1260 // Set DT_FLAGS and DT_FLAGS_1.
1261 uint32_t dtFlags = 0;
1262 uint32_t dtFlags1 = 0;
1263 if (config->bsymbolic)
1264 dtFlags |= DF_SYMBOLIC;
1265 if (config->zGlobal)
1266 dtFlags1 |= DF_1_GLOBAL;
1267 if (config->zInitfirst)
1268 dtFlags1 |= DF_1_INITFIRST;
1269 if (config->zInterpose)
1270 dtFlags1 |= DF_1_INTERPOSE;
1271 if (config->zNodefaultlib)
1272 dtFlags1 |= DF_1_NODEFLIB;
1273 if (config->zNodelete)
1274 dtFlags1 |= DF_1_NODELETE;
1275 if (config->zNodlopen)
1276 dtFlags1 |= DF_1_NOOPEN;
1278 dtFlags |= DF_BIND_NOW;
1279 dtFlags1 |= DF_1_NOW;
1281 if (config->zOrigin) {
1282 dtFlags |= DF_ORIGIN;
1283 dtFlags1 |= DF_1_ORIGIN;
1286 dtFlags |= DF_TEXTREL;
1287 if (config->hasStaticTlsModel)
1288 dtFlags |= DF_STATIC_TLS;
1291 addInt(DT_FLAGS, dtFlags);
1293 addInt(DT_FLAGS_1, dtFlags1);
1295 // DT_DEBUG is a pointer to debug informaion used by debuggers at runtime. We
1296 // need it for each process, so we don't write it for DSOs. The loader writes
1297 // the pointer into this entry.
1299 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some
1300 // systems (currently only Fuchsia OS) provide other means to give the
1301 // debugger this information. Such systems may choose make .dynamic read-only.
1302 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG.
1303 if (!config->shared && !config->relocatable && !config->zRodynamic)
1304 addInt(DT_DEBUG, 0);
1306 if (OutputSection *sec = part.dynStrTab->getParent())
1307 this->link = sec->sectionIndex;
1309 if (part.relaDyn->isNeeded()) {
1310 addInSec(part.relaDyn->dynamicTag, part.relaDyn);
1311 addSize(part.relaDyn->sizeDynamicTag, part.relaDyn->getParent());
1313 bool isRela = config->isRela;
1314 addInt(isRela ? DT_RELAENT : DT_RELENT,
1315 isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel));
1317 // MIPS dynamic loader does not support RELCOUNT tag.
1318 // The problem is in the tight relation between dynamic
1319 // relocations and GOT. So do not emit this tag on MIPS.
1320 if (config->emachine != EM_MIPS) {
1321 size_t numRelativeRels = part.relaDyn->getRelativeRelocCount();
1322 if (config->zCombreloc && numRelativeRels)
1323 addInt(isRela ? DT_RELACOUNT : DT_RELCOUNT, numRelativeRels);
1326 if (part.relrDyn && !part.relrDyn->relocs.empty()) {
1327 addInSec(config->useAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR,
1329 addSize(config->useAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ,
1330 part.relrDyn->getParent());
1331 addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT,
1334 // .rel[a].plt section usually consists of two parts, containing plt and
1335 // iplt relocations. It is possible to have only iplt relocations in the
1336 // output. In that case relaPlt is empty and have zero offset, the same offset
1337 // as relaIplt has. And we still want to emit proper dynamic tags for that
1338 // case, so here we always use relaPlt as marker for the begining of
1339 // .rel[a].plt section.
1340 if (isMain && (in.relaPlt->isNeeded() || in.relaIplt->isNeeded())) {
1341 addInSec(DT_JMPREL, in.relaPlt);
1342 entries.push_back({DT_PLTRELSZ, addPltRelSz});
1343 switch (config->emachine) {
1345 addInSec(DT_MIPS_PLTGOT, in.gotPlt);
1348 addInSec(DT_PLTGOT, in.plt);
1351 addInSec(DT_PLTGOT, in.gotPlt);
1354 addInt(DT_PLTREL, config->isRela ? DT_RELA : DT_REL);
1357 if (config->emachine == EM_AARCH64) {
1358 if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI)
1359 addInt(DT_AARCH64_BTI_PLT, 0);
1360 if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_PAC)
1361 addInt(DT_AARCH64_PAC_PLT, 0);
1364 addInSec(DT_SYMTAB, part.dynSymTab);
1365 addInt(DT_SYMENT, sizeof(Elf_Sym));
1366 addInSec(DT_STRTAB, part.dynStrTab);
1367 addInt(DT_STRSZ, part.dynStrTab->getSize());
1369 addInt(DT_TEXTREL, 0);
1370 if (part.gnuHashTab)
1371 addInSec(DT_GNU_HASH, part.gnuHashTab);
1373 addInSec(DT_HASH, part.hashTab);
1376 if (Out::preinitArray) {
1377 addOutSec(DT_PREINIT_ARRAY, Out::preinitArray);
1378 addSize(DT_PREINIT_ARRAYSZ, Out::preinitArray);
1380 if (Out::initArray) {
1381 addOutSec(DT_INIT_ARRAY, Out::initArray);
1382 addSize(DT_INIT_ARRAYSZ, Out::initArray);
1384 if (Out::finiArray) {
1385 addOutSec(DT_FINI_ARRAY, Out::finiArray);
1386 addSize(DT_FINI_ARRAYSZ, Out::finiArray);
1389 if (Symbol *b = symtab->find(config->init))
1392 if (Symbol *b = symtab->find(config->fini))
1397 bool hasVerNeed = SharedFile::vernauxNum != 0;
1398 if (hasVerNeed || part.verDef)
1399 addInSec(DT_VERSYM, part.verSym);
1401 addInSec(DT_VERDEF, part.verDef);
1402 addInt(DT_VERDEFNUM, getVerDefNum());
1405 addInSec(DT_VERNEED, part.verNeed);
1406 unsigned needNum = 0;
1407 for (SharedFile *f : sharedFiles)
1408 if (!f->vernauxs.empty())
1410 addInt(DT_VERNEEDNUM, needNum);
1413 if (config->emachine == EM_MIPS) {
1414 addInt(DT_MIPS_RLD_VERSION, 1);
1415 addInt(DT_MIPS_FLAGS, RHF_NOTPOT);
1416 addInt(DT_MIPS_BASE_ADDRESS, target->getImageBase());
1417 addInt(DT_MIPS_SYMTABNO, part.dynSymTab->getNumSymbols());
1419 add(DT_MIPS_LOCAL_GOTNO, [] { return in.mipsGot->getLocalEntriesNum(); });
1421 if (const Symbol *b = in.mipsGot->getFirstGlobalEntry())
1422 addInt(DT_MIPS_GOTSYM, b->dynsymIndex);
1424 addInt(DT_MIPS_GOTSYM, part.dynSymTab->getNumSymbols());
1425 addInSec(DT_PLTGOT, in.mipsGot);
1426 if (in.mipsRldMap) {
1428 addInSec(DT_MIPS_RLD_MAP, in.mipsRldMap);
1429 // Store the offset to the .rld_map section
1430 // relative to the address of the tag.
1431 addInSecRelative(DT_MIPS_RLD_MAP_REL, in.mipsRldMap);
1435 // DT_PPC_GOT indicates to glibc Secure PLT is used. If DT_PPC_GOT is absent,
1436 // glibc assumes the old-style BSS PLT layout which we don't support.
1437 if (config->emachine == EM_PPC)
1438 add(DT_PPC_GOT, [] { return in.got->getVA(); });
1440 // Glink dynamic tag is required by the V2 abi if the plt section isn't empty.
1441 if (config->emachine == EM_PPC64 && in.plt->isNeeded()) {
1442 // The Glink tag points to 32 bytes before the first lazy symbol resolution
1443 // stub, which starts directly after the header.
1444 entries.push_back({DT_PPC64_GLINK, [=] {
1445 unsigned offset = target->pltHeaderSize - 32;
1446 return in.plt->getVA(0) + offset;
1452 getParent()->link = this->link;
1453 this->size = entries.size() * this->entsize;
1456 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *buf) {
1457 auto *p = reinterpret_cast<Elf_Dyn *>(buf);
1459 for (std::pair<int32_t, std::function<uint64_t()>> &kv : entries) {
1460 p->d_tag = kv.first;
1461 p->d_un.d_val = kv.second();
1466 uint64_t DynamicReloc::getOffset() const {
1467 return inputSec->getVA(offsetInSec);
1470 int64_t DynamicReloc::computeAddend() const {
1472 return sym->getVA(addend);
1475 // See the comment in the DynamicReloc ctor.
1476 return getMipsPageAddr(outputSec->addr) + addend;
1479 uint32_t DynamicReloc::getSymIndex(SymbolTableBaseSection *symTab) const {
1480 if (sym && !useSymVA)
1481 return symTab->getSymbolIndex(sym);
1485 RelocationBaseSection::RelocationBaseSection(StringRef name, uint32_t type,
1487 int32_t sizeDynamicTag)
1488 : SyntheticSection(SHF_ALLOC, type, config->wordsize, name),
1489 dynamicTag(dynamicTag), sizeDynamicTag(sizeDynamicTag) {}
1491 void RelocationBaseSection::addReloc(RelType dynType, InputSectionBase *isec,
1492 uint64_t offsetInSec, Symbol *sym) {
1493 addReloc({dynType, isec, offsetInSec, false, sym, 0});
1496 void RelocationBaseSection::addReloc(RelType dynType,
1497 InputSectionBase *inputSec,
1498 uint64_t offsetInSec, Symbol *sym,
1499 int64_t addend, RelExpr expr,
1501 // Write the addends to the relocated address if required. We skip
1502 // it if the written value would be zero.
1503 if (config->writeAddends && (expr != R_ADDEND || addend != 0))
1504 inputSec->relocations.push_back({expr, type, offsetInSec, addend, sym});
1505 addReloc({dynType, inputSec, offsetInSec, expr != R_ADDEND, sym, addend});
1508 void RelocationBaseSection::addReloc(const DynamicReloc &reloc) {
1509 if (reloc.type == target->relativeRel)
1510 ++numRelativeRelocs;
1511 relocs.push_back(reloc);
1514 void RelocationBaseSection::finalizeContents() {
1515 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1517 // When linking glibc statically, .rel{,a}.plt contains R_*_IRELATIVE
1518 // relocations due to IFUNC (e.g. strcpy). sh_link will be set to 0 in that
1520 if (symTab && symTab->getParent())
1521 getParent()->link = symTab->getParent()->sectionIndex;
1523 getParent()->link = 0;
1525 if (in.relaPlt == this)
1526 getParent()->info = in.gotPlt->getParent()->sectionIndex;
1527 if (in.relaIplt == this)
1528 getParent()->info = in.igotPlt->getParent()->sectionIndex;
1531 RelrBaseSection::RelrBaseSection()
1532 : SyntheticSection(SHF_ALLOC,
1533 config->useAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR,
1534 config->wordsize, ".relr.dyn") {}
1536 template <class ELFT>
1537 static void encodeDynamicReloc(SymbolTableBaseSection *symTab,
1538 typename ELFT::Rela *p,
1539 const DynamicReloc &rel) {
1541 p->r_addend = rel.computeAddend();
1542 p->r_offset = rel.getOffset();
1543 p->setSymbolAndType(rel.getSymIndex(symTab), rel.type, config->isMips64EL);
1546 template <class ELFT>
1547 RelocationSection<ELFT>::RelocationSection(StringRef name, bool sort)
1548 : RelocationBaseSection(name, config->isRela ? SHT_RELA : SHT_REL,
1549 config->isRela ? DT_RELA : DT_REL,
1550 config->isRela ? DT_RELASZ : DT_RELSZ),
1552 this->entsize = config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1555 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *buf) {
1556 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1558 // Sort by (!IsRelative,SymIndex,r_offset). DT_REL[A]COUNT requires us to
1559 // place R_*_RELATIVE first. SymIndex is to improve locality, while r_offset
1560 // is to make results easier to read.
1563 relocs, [&](const DynamicReloc &a, const DynamicReloc &b) {
1564 return std::make_tuple(a.type != target->relativeRel,
1565 a.getSymIndex(symTab), a.getOffset()) <
1566 std::make_tuple(b.type != target->relativeRel,
1567 b.getSymIndex(symTab), b.getOffset());
1570 for (const DynamicReloc &rel : relocs) {
1571 encodeDynamicReloc<ELFT>(symTab, reinterpret_cast<Elf_Rela *>(buf), rel);
1572 buf += config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1576 template <class ELFT>
1577 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection(
1579 : RelocationBaseSection(
1580 name, config->isRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL,
1581 config->isRela ? DT_ANDROID_RELA : DT_ANDROID_REL,
1582 config->isRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ) {
1586 template <class ELFT>
1587 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() {
1588 // This function computes the contents of an Android-format packed relocation
1591 // This format compresses relocations by using relocation groups to factor out
1592 // fields that are common between relocations and storing deltas from previous
1593 // relocations in SLEB128 format (which has a short representation for small
1594 // numbers). A good example of a relocation type with common fields is
1595 // R_*_RELATIVE, which is normally used to represent function pointers in
1596 // vtables. In the REL format, each relative relocation has the same r_info
1597 // field, and is only different from other relative relocations in terms of
1598 // the r_offset field. By sorting relocations by offset, grouping them by
1599 // r_info and representing each relocation with only the delta from the
1600 // previous offset, each 8-byte relocation can be compressed to as little as 1
1601 // byte (or less with run-length encoding). This relocation packer was able to
1602 // reduce the size of the relocation section in an Android Chromium DSO from
1603 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size.
1605 // A relocation section consists of a header containing the literal bytes
1606 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two
1607 // elements are the total number of relocations in the section and an initial
1608 // r_offset value. The remaining elements define a sequence of relocation
1609 // groups. Each relocation group starts with a header consisting of the
1610 // following elements:
1612 // - the number of relocations in the relocation group
1613 // - flags for the relocation group
1614 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta
1615 // for each relocation in the group.
1616 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info
1617 // field for each relocation in the group.
1618 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and
1619 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for
1620 // each relocation in the group.
1622 // Following the relocation group header are descriptions of each of the
1623 // relocations in the group. They consist of the following elements:
1625 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset
1626 // delta for this relocation.
1627 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info
1628 // field for this relocation.
1629 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and
1630 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for
1633 size_t oldSize = relocData.size();
1635 relocData = {'A', 'P', 'S', '2'};
1636 raw_svector_ostream os(relocData);
1637 auto add = [&](int64_t v) { encodeSLEB128(v, os); };
1639 // The format header includes the number of relocations and the initial
1640 // offset (we set this to zero because the first relocation group will
1641 // perform the initial adjustment).
1645 std::vector<Elf_Rela> relatives, nonRelatives;
1647 for (const DynamicReloc &rel : relocs) {
1649 encodeDynamicReloc<ELFT>(getPartition().dynSymTab, &r, rel);
1651 if (r.getType(config->isMips64EL) == target->relativeRel)
1652 relatives.push_back(r);
1654 nonRelatives.push_back(r);
1657 llvm::sort(relatives, [](const Elf_Rel &a, const Elf_Rel &b) {
1658 return a.r_offset < b.r_offset;
1661 // Try to find groups of relative relocations which are spaced one word
1662 // apart from one another. These generally correspond to vtable entries. The
1663 // format allows these groups to be encoded using a sort of run-length
1664 // encoding, but each group will cost 7 bytes in addition to the offset from
1665 // the previous group, so it is only profitable to do this for groups of
1666 // size 8 or larger.
1667 std::vector<Elf_Rela> ungroupedRelatives;
1668 std::vector<std::vector<Elf_Rela>> relativeGroups;
1669 for (auto i = relatives.begin(), e = relatives.end(); i != e;) {
1670 std::vector<Elf_Rela> group;
1672 group.push_back(*i++);
1673 } while (i != e && (i - 1)->r_offset + config->wordsize == i->r_offset);
1675 if (group.size() < 8)
1676 ungroupedRelatives.insert(ungroupedRelatives.end(), group.begin(),
1679 relativeGroups.emplace_back(std::move(group));
1682 unsigned hasAddendIfRela =
1683 config->isRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0;
1685 uint64_t offset = 0;
1686 uint64_t addend = 0;
1688 // Emit the run-length encoding for the groups of adjacent relative
1689 // relocations. Each group is represented using two groups in the packed
1690 // format. The first is used to set the current offset to the start of the
1691 // group (and also encodes the first relocation), and the second encodes the
1692 // remaining relocations.
1693 for (std::vector<Elf_Rela> &g : relativeGroups) {
1694 // The first relocation in the group.
1696 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1697 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1698 add(g[0].r_offset - offset);
1699 add(target->relativeRel);
1700 if (config->isRela) {
1701 add(g[0].r_addend - addend);
1702 addend = g[0].r_addend;
1705 // The remaining relocations.
1707 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1708 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1709 add(config->wordsize);
1710 add(target->relativeRel);
1711 if (config->isRela) {
1712 for (auto i = g.begin() + 1, e = g.end(); i != e; ++i) {
1713 add(i->r_addend - addend);
1714 addend = i->r_addend;
1718 offset = g.back().r_offset;
1721 // Now the ungrouped relatives.
1722 if (!ungroupedRelatives.empty()) {
1723 add(ungroupedRelatives.size());
1724 add(RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1725 add(target->relativeRel);
1726 for (Elf_Rela &r : ungroupedRelatives) {
1727 add(r.r_offset - offset);
1728 offset = r.r_offset;
1729 if (config->isRela) {
1730 add(r.r_addend - addend);
1731 addend = r.r_addend;
1736 // Finally the non-relative relocations.
1737 llvm::sort(nonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1738 return a.r_offset < b.r_offset;
1740 if (!nonRelatives.empty()) {
1741 add(nonRelatives.size());
1742 add(hasAddendIfRela);
1743 for (Elf_Rela &r : nonRelatives) {
1744 add(r.r_offset - offset);
1745 offset = r.r_offset;
1747 if (config->isRela) {
1748 add(r.r_addend - addend);
1749 addend = r.r_addend;
1754 // Don't allow the section to shrink; otherwise the size of the section can
1755 // oscillate infinitely.
1756 if (relocData.size() < oldSize)
1757 relocData.append(oldSize - relocData.size(), 0);
1759 // Returns whether the section size changed. We need to keep recomputing both
1760 // section layout and the contents of this section until the size converges
1761 // because changing this section's size can affect section layout, which in
1762 // turn can affect the sizes of the LEB-encoded integers stored in this
1764 return relocData.size() != oldSize;
1767 template <class ELFT> RelrSection<ELFT>::RelrSection() {
1768 this->entsize = config->wordsize;
1771 template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() {
1772 // This function computes the contents of an SHT_RELR packed relocation
1775 // Proposal for adding SHT_RELR sections to generic-abi is here:
1776 // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg
1778 // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks
1779 // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ]
1781 // i.e. start with an address, followed by any number of bitmaps. The address
1782 // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63
1783 // relocations each, at subsequent offsets following the last address entry.
1785 // The bitmap entries must have 1 in the least significant bit. The assumption
1786 // here is that an address cannot have 1 in lsb. Odd addresses are not
1789 // Excluding the least significant bit in the bitmap, each non-zero bit in
1790 // the bitmap represents a relocation to be applied to a corresponding machine
1791 // word that follows the base address word. The second least significant bit
1792 // represents the machine word immediately following the initial address, and
1793 // each bit that follows represents the next word, in linear order. As such,
1794 // a single bitmap can encode up to 31 relocations in a 32-bit object, and
1795 // 63 relocations in a 64-bit object.
1797 // This encoding has a couple of interesting properties:
1798 // 1. Looking at any entry, it is clear whether it's an address or a bitmap:
1799 // even means address, odd means bitmap.
1800 // 2. Just a simple list of addresses is a valid encoding.
1802 size_t oldSize = relrRelocs.size();
1805 // Same as Config->Wordsize but faster because this is a compile-time
1807 const size_t wordsize = sizeof(typename ELFT::uint);
1809 // Number of bits to use for the relocation offsets bitmap.
1810 // Must be either 63 or 31.
1811 const size_t nBits = wordsize * 8 - 1;
1813 // Get offsets for all relative relocations and sort them.
1814 std::vector<uint64_t> offsets;
1815 for (const RelativeReloc &rel : relocs)
1816 offsets.push_back(rel.getOffset());
1817 llvm::sort(offsets);
1819 // For each leading relocation, find following ones that can be folded
1820 // as a bitmap and fold them.
1821 for (size_t i = 0, e = offsets.size(); i < e;) {
1822 // Add a leading relocation.
1823 relrRelocs.push_back(Elf_Relr(offsets[i]));
1824 uint64_t base = offsets[i] + wordsize;
1827 // Find foldable relocations to construct bitmaps.
1829 uint64_t bitmap = 0;
1832 uint64_t delta = offsets[i] - base;
1834 // If it is too far, it cannot be folded.
1835 if (delta >= nBits * wordsize)
1838 // If it is not a multiple of wordsize away, it cannot be folded.
1839 if (delta % wordsize)
1843 bitmap |= 1ULL << (delta / wordsize);
1850 relrRelocs.push_back(Elf_Relr((bitmap << 1) | 1));
1851 base += nBits * wordsize;
1855 return relrRelocs.size() != oldSize;
1858 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &strTabSec)
1859 : SyntheticSection(strTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0,
1860 strTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB,
1862 strTabSec.isDynamic() ? ".dynsym" : ".symtab"),
1863 strTabSec(strTabSec) {}
1865 // Orders symbols according to their positions in the GOT,
1866 // in compliance with MIPS ABI rules.
1867 // See "Global Offset Table" in Chapter 5 in the following document
1868 // for detailed description:
1869 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1870 static bool sortMipsSymbols(const SymbolTableEntry &l,
1871 const SymbolTableEntry &r) {
1872 // Sort entries related to non-local preemptible symbols by GOT indexes.
1873 // All other entries go to the beginning of a dynsym in arbitrary order.
1874 if (l.sym->isInGot() && r.sym->isInGot())
1875 return l.sym->gotIndex < r.sym->gotIndex;
1876 if (!l.sym->isInGot() && !r.sym->isInGot())
1878 return !l.sym->isInGot();
1881 void SymbolTableBaseSection::finalizeContents() {
1882 if (OutputSection *sec = strTabSec.getParent())
1883 getParent()->link = sec->sectionIndex;
1885 if (this->type != SHT_DYNSYM) {
1886 sortSymTabSymbols();
1890 // If it is a .dynsym, there should be no local symbols, but we need
1891 // to do a few things for the dynamic linker.
1893 // Section's Info field has the index of the first non-local symbol.
1894 // Because the first symbol entry is a null entry, 1 is the first.
1895 getParent()->info = 1;
1897 if (getPartition().gnuHashTab) {
1898 // NB: It also sorts Symbols to meet the GNU hash table requirements.
1899 getPartition().gnuHashTab->addSymbols(symbols);
1900 } else if (config->emachine == EM_MIPS) {
1901 llvm::stable_sort(symbols, sortMipsSymbols);
1904 // Only the main partition's dynsym indexes are stored in the symbols
1905 // themselves. All other partitions use a lookup table.
1906 if (this == mainPart->dynSymTab) {
1908 for (const SymbolTableEntry &s : symbols)
1909 s.sym->dynsymIndex = ++i;
1913 // The ELF spec requires that all local symbols precede global symbols, so we
1914 // sort symbol entries in this function. (For .dynsym, we don't do that because
1915 // symbols for dynamic linking are inherently all globals.)
1917 // Aside from above, we put local symbols in groups starting with the STT_FILE
1918 // symbol. That is convenient for purpose of identifying where are local symbols
1920 void SymbolTableBaseSection::sortSymTabSymbols() {
1921 // Move all local symbols before global symbols.
1922 auto e = std::stable_partition(
1923 symbols.begin(), symbols.end(), [](const SymbolTableEntry &s) {
1924 return s.sym->isLocal() || s.sym->computeBinding() == STB_LOCAL;
1926 size_t numLocals = e - symbols.begin();
1927 getParent()->info = numLocals + 1;
1929 // We want to group the local symbols by file. For that we rebuild the local
1930 // part of the symbols vector. We do not need to care about the STT_FILE
1931 // symbols, they are already naturally placed first in each group. That
1932 // happens because STT_FILE is always the first symbol in the object and hence
1933 // precede all other local symbols we add for a file.
1934 MapVector<InputFile *, std::vector<SymbolTableEntry>> arr;
1935 for (const SymbolTableEntry &s : llvm::make_range(symbols.begin(), e))
1936 arr[s.sym->file].push_back(s);
1938 auto i = symbols.begin();
1939 for (std::pair<InputFile *, std::vector<SymbolTableEntry>> &p : arr)
1940 for (SymbolTableEntry &entry : p.second)
1944 void SymbolTableBaseSection::addSymbol(Symbol *b) {
1945 // Adding a local symbol to a .dynsym is a bug.
1946 assert(this->type != SHT_DYNSYM || !b->isLocal());
1948 bool hashIt = b->isLocal();
1949 symbols.push_back({b, strTabSec.addString(b->getName(), hashIt)});
1952 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *sym) {
1953 if (this == mainPart->dynSymTab)
1954 return sym->dynsymIndex;
1956 // Initializes symbol lookup tables lazily. This is used only for -r,
1957 // -emit-relocs and dynsyms in partitions other than the main one.
1958 llvm::call_once(onceFlag, [&] {
1959 symbolIndexMap.reserve(symbols.size());
1961 for (const SymbolTableEntry &e : symbols) {
1962 if (e.sym->type == STT_SECTION)
1963 sectionIndexMap[e.sym->getOutputSection()] = ++i;
1965 symbolIndexMap[e.sym] = ++i;
1969 // Section symbols are mapped based on their output sections
1970 // to maintain their semantics.
1971 if (sym->type == STT_SECTION)
1972 return sectionIndexMap.lookup(sym->getOutputSection());
1973 return symbolIndexMap.lookup(sym);
1976 template <class ELFT>
1977 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &strTabSec)
1978 : SymbolTableBaseSection(strTabSec) {
1979 this->entsize = sizeof(Elf_Sym);
1982 static BssSection *getCommonSec(Symbol *sym) {
1983 if (!config->defineCommon)
1984 if (auto *d = dyn_cast<Defined>(sym))
1985 return dyn_cast_or_null<BssSection>(d->section);
1989 static uint32_t getSymSectionIndex(Symbol *sym) {
1990 if (getCommonSec(sym))
1992 if (!isa<Defined>(sym) || sym->needsPltAddr)
1994 if (const OutputSection *os = sym->getOutputSection())
1995 return os->sectionIndex >= SHN_LORESERVE ? (uint32_t)SHN_XINDEX
2000 // Write the internal symbol table contents to the output symbol table.
2001 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *buf) {
2002 // The first entry is a null entry as per the ELF spec.
2003 memset(buf, 0, sizeof(Elf_Sym));
2004 buf += sizeof(Elf_Sym);
2006 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2008 for (SymbolTableEntry &ent : symbols) {
2009 Symbol *sym = ent.sym;
2010 bool isDefinedHere = type == SHT_SYMTAB || sym->partition == partition;
2012 // Set st_info and st_other.
2014 if (sym->isLocal()) {
2015 eSym->setBindingAndType(STB_LOCAL, sym->type);
2017 eSym->setBindingAndType(sym->computeBinding(), sym->type);
2018 eSym->setVisibility(sym->visibility);
2021 // The 3 most significant bits of st_other are used by OpenPOWER ABI.
2022 // See getPPC64GlobalEntryToLocalEntryOffset() for more details.
2023 if (config->emachine == EM_PPC64)
2024 eSym->st_other |= sym->stOther & 0xe0;
2026 eSym->st_name = ent.strTabOffset;
2028 eSym->st_shndx = getSymSectionIndex(ent.sym);
2032 // Copy symbol size if it is a defined symbol. st_size is not significant
2033 // for undefined symbols, so whether copying it or not is up to us if that's
2034 // the case. We'll leave it as zero because by not setting a value, we can
2035 // get the exact same outputs for two sets of input files that differ only
2036 // in undefined symbol size in DSOs.
2037 if (eSym->st_shndx == SHN_UNDEF || !isDefinedHere)
2040 eSym->st_size = sym->getSize();
2042 // st_value is usually an address of a symbol, but that has a
2043 // special meaining for uninstantiated common symbols (this can
2044 // occur if -r is given).
2045 if (BssSection *commonSec = getCommonSec(ent.sym))
2046 eSym->st_value = commonSec->alignment;
2047 else if (isDefinedHere)
2048 eSym->st_value = sym->getVA();
2055 // On MIPS we need to mark symbol which has a PLT entry and requires
2056 // pointer equality by STO_MIPS_PLT flag. That is necessary to help
2057 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs.
2058 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt
2059 if (config->emachine == EM_MIPS) {
2060 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2062 for (SymbolTableEntry &ent : symbols) {
2063 Symbol *sym = ent.sym;
2064 if (sym->isInPlt() && sym->needsPltAddr)
2065 eSym->st_other |= STO_MIPS_PLT;
2066 if (isMicroMips()) {
2067 // We already set the less-significant bit for symbols
2068 // marked by the `STO_MIPS_MICROMIPS` flag and for microMIPS PLT
2069 // records. That allows us to distinguish such symbols in
2070 // the `MIPS<ELFT>::relocateOne()` routine. Now we should
2071 // clear that bit for non-dynamic symbol table, so tools
2072 // like `objdump` will be able to deal with a correct
2074 if (sym->isDefined() &&
2075 ((sym->stOther & STO_MIPS_MICROMIPS) || sym->needsPltAddr)) {
2076 if (!strTabSec.isDynamic())
2077 eSym->st_value &= ~1;
2078 eSym->st_other |= STO_MIPS_MICROMIPS;
2081 if (config->relocatable)
2082 if (auto *d = dyn_cast<Defined>(sym))
2083 if (isMipsPIC<ELFT>(d))
2084 eSym->st_other |= STO_MIPS_PIC;
2090 SymtabShndxSection::SymtabShndxSection()
2091 : SyntheticSection(0, SHT_SYMTAB_SHNDX, 4, ".symtab_shndx") {
2095 void SymtabShndxSection::writeTo(uint8_t *buf) {
2096 // We write an array of 32 bit values, where each value has 1:1 association
2097 // with an entry in .symtab. If the corresponding entry contains SHN_XINDEX,
2098 // we need to write actual index, otherwise, we must write SHN_UNDEF(0).
2099 buf += 4; // Ignore .symtab[0] entry.
2100 for (const SymbolTableEntry &entry : in.symTab->getSymbols()) {
2101 if (getSymSectionIndex(entry.sym) == SHN_XINDEX)
2102 write32(buf, entry.sym->getOutputSection()->sectionIndex);
2107 bool SymtabShndxSection::isNeeded() const {
2108 // SHT_SYMTAB can hold symbols with section indices values up to
2109 // SHN_LORESERVE. If we need more, we want to use extension SHT_SYMTAB_SHNDX
2110 // section. Problem is that we reveal the final section indices a bit too
2111 // late, and we do not know them here. For simplicity, we just always create
2112 // a .symtab_shndx section when the amount of output sections is huge.
2114 for (BaseCommand *base : script->sectionCommands)
2115 if (isa<OutputSection>(base))
2117 return size >= SHN_LORESERVE;
2120 void SymtabShndxSection::finalizeContents() {
2121 getParent()->link = in.symTab->getParent()->sectionIndex;
2124 size_t SymtabShndxSection::getSize() const {
2125 return in.symTab->getNumSymbols() * 4;
2128 // .hash and .gnu.hash sections contain on-disk hash tables that map
2129 // symbol names to their dynamic symbol table indices. Their purpose
2130 // is to help the dynamic linker resolve symbols quickly. If ELF files
2131 // don't have them, the dynamic linker has to do linear search on all
2132 // dynamic symbols, which makes programs slower. Therefore, a .hash
2133 // section is added to a DSO by default. A .gnu.hash is added if you
2134 // give the -hash-style=gnu or -hash-style=both option.
2136 // The Unix semantics of resolving dynamic symbols is somewhat expensive.
2137 // Each ELF file has a list of DSOs that the ELF file depends on and a
2138 // list of dynamic symbols that need to be resolved from any of the
2139 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n)
2140 // where m is the number of DSOs and n is the number of dynamic
2141 // symbols. For modern large programs, both m and n are large. So
2142 // making each step faster by using hash tables substiantially
2143 // improves time to load programs.
2145 // (Note that this is not the only way to design the shared library.
2146 // For instance, the Windows DLL takes a different approach. On
2147 // Windows, each dynamic symbol has a name of DLL from which the symbol
2148 // has to be resolved. That makes the cost of symbol resolution O(n).
2149 // This disables some hacky techniques you can use on Unix such as
2150 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.)
2152 // Due to historical reasons, we have two different hash tables, .hash
2153 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new
2154 // and better version of .hash. .hash is just an on-disk hash table, but
2155 // .gnu.hash has a bloom filter in addition to a hash table to skip
2156 // DSOs very quickly. If you are sure that your dynamic linker knows
2157 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a
2158 // safe bet is to specify -hash-style=both for backward compatibilty.
2159 GnuHashTableSection::GnuHashTableSection()
2160 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, config->wordsize, ".gnu.hash") {
2163 void GnuHashTableSection::finalizeContents() {
2164 if (OutputSection *sec = getPartition().dynSymTab->getParent())
2165 getParent()->link = sec->sectionIndex;
2167 // Computes bloom filter size in word size. We want to allocate 12
2168 // bits for each symbol. It must be a power of two.
2169 if (symbols.empty()) {
2172 uint64_t numBits = symbols.size() * 12;
2173 maskWords = NextPowerOf2(numBits / (config->wordsize * 8));
2176 size = 16; // Header
2177 size += config->wordsize * maskWords; // Bloom filter
2178 size += nBuckets * 4; // Hash buckets
2179 size += symbols.size() * 4; // Hash values
2182 void GnuHashTableSection::writeTo(uint8_t *buf) {
2183 // The output buffer is not guaranteed to be zero-cleared because we pre-
2184 // fill executable sections with trap instructions. This is a precaution
2185 // for that case, which happens only when -no-rosegment is given.
2186 memset(buf, 0, size);
2189 write32(buf, nBuckets);
2190 write32(buf + 4, getPartition().dynSymTab->getNumSymbols() - symbols.size());
2191 write32(buf + 8, maskWords);
2192 write32(buf + 12, Shift2);
2195 // Write a bloom filter and a hash table.
2196 writeBloomFilter(buf);
2197 buf += config->wordsize * maskWords;
2198 writeHashTable(buf);
2201 // This function writes a 2-bit bloom filter. This bloom filter alone
2202 // usually filters out 80% or more of all symbol lookups [1].
2203 // The dynamic linker uses the hash table only when a symbol is not
2204 // filtered out by a bloom filter.
2206 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2),
2207 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf
2208 void GnuHashTableSection::writeBloomFilter(uint8_t *buf) {
2209 unsigned c = config->is64 ? 64 : 32;
2210 for (const Entry &sym : symbols) {
2211 // When C = 64, we choose a word with bits [6:...] and set 1 to two bits in
2212 // the word using bits [0:5] and [26:31].
2213 size_t i = (sym.hash / c) & (maskWords - 1);
2214 uint64_t val = readUint(buf + i * config->wordsize);
2215 val |= uint64_t(1) << (sym.hash % c);
2216 val |= uint64_t(1) << ((sym.hash >> Shift2) % c);
2217 writeUint(buf + i * config->wordsize, val);
2221 void GnuHashTableSection::writeHashTable(uint8_t *buf) {
2222 uint32_t *buckets = reinterpret_cast<uint32_t *>(buf);
2223 uint32_t oldBucket = -1;
2224 uint32_t *values = buckets + nBuckets;
2225 for (auto i = symbols.begin(), e = symbols.end(); i != e; ++i) {
2226 // Write a hash value. It represents a sequence of chains that share the
2227 // same hash modulo value. The last element of each chain is terminated by
2229 uint32_t hash = i->hash;
2230 bool isLastInChain = (i + 1) == e || i->bucketIdx != (i + 1)->bucketIdx;
2231 hash = isLastInChain ? hash | 1 : hash & ~1;
2232 write32(values++, hash);
2234 if (i->bucketIdx == oldBucket)
2236 // Write a hash bucket. Hash buckets contain indices in the following hash
2238 write32(buckets + i->bucketIdx,
2239 getPartition().dynSymTab->getSymbolIndex(i->sym));
2240 oldBucket = i->bucketIdx;
2244 static uint32_t hashGnu(StringRef name) {
2246 for (uint8_t c : name)
2247 h = (h << 5) + h + c;
2251 // Add symbols to this symbol hash table. Note that this function
2252 // destructively sort a given vector -- which is needed because
2253 // GNU-style hash table places some sorting requirements.
2254 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &v) {
2255 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce
2256 // its type correctly.
2257 std::vector<SymbolTableEntry>::iterator mid =
2258 std::stable_partition(v.begin(), v.end(), [&](const SymbolTableEntry &s) {
2259 return !s.sym->isDefined() || s.sym->partition != partition;
2262 // We chose load factor 4 for the on-disk hash table. For each hash
2263 // collision, the dynamic linker will compare a uint32_t hash value.
2264 // Since the integer comparison is quite fast, we believe we can
2265 // make the load factor even larger. 4 is just a conservative choice.
2267 // Note that we don't want to create a zero-sized hash table because
2268 // Android loader as of 2018 doesn't like a .gnu.hash containing such
2269 // table. If that's the case, we create a hash table with one unused
2271 nBuckets = std::max<size_t>((v.end() - mid) / 4, 1);
2276 for (SymbolTableEntry &ent : llvm::make_range(mid, v.end())) {
2277 Symbol *b = ent.sym;
2278 uint32_t hash = hashGnu(b->getName());
2279 uint32_t bucketIdx = hash % nBuckets;
2280 symbols.push_back({b, ent.strTabOffset, hash, bucketIdx});
2283 llvm::stable_sort(symbols, [](const Entry &l, const Entry &r) {
2284 return l.bucketIdx < r.bucketIdx;
2287 v.erase(mid, v.end());
2288 for (const Entry &ent : symbols)
2289 v.push_back({ent.sym, ent.strTabOffset});
2292 HashTableSection::HashTableSection()
2293 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") {
2297 void HashTableSection::finalizeContents() {
2298 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2300 if (OutputSection *sec = symTab->getParent())
2301 getParent()->link = sec->sectionIndex;
2303 unsigned numEntries = 2; // nbucket and nchain.
2304 numEntries += symTab->getNumSymbols(); // The chain entries.
2306 // Create as many buckets as there are symbols.
2307 numEntries += symTab->getNumSymbols();
2308 this->size = numEntries * 4;
2311 void HashTableSection::writeTo(uint8_t *buf) {
2312 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2314 // See comment in GnuHashTableSection::writeTo.
2315 memset(buf, 0, size);
2317 unsigned numSymbols = symTab->getNumSymbols();
2319 uint32_t *p = reinterpret_cast<uint32_t *>(buf);
2320 write32(p++, numSymbols); // nbucket
2321 write32(p++, numSymbols); // nchain
2323 uint32_t *buckets = p;
2324 uint32_t *chains = p + numSymbols;
2326 for (const SymbolTableEntry &s : symTab->getSymbols()) {
2327 Symbol *sym = s.sym;
2328 StringRef name = sym->getName();
2329 unsigned i = sym->dynsymIndex;
2330 uint32_t hash = hashSysV(name) % numSymbols;
2331 chains[i] = buckets[hash];
2332 write32(buckets + hash, i);
2336 // On PowerPC64 the lazy symbol resolvers go into the `global linkage table`
2337 // in the .glink section, rather then the typical .plt section.
2338 PltSection::PltSection(bool isIplt)
2340 SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16,
2341 (config->emachine == EM_PPC || config->emachine == EM_PPC64)
2344 headerSize(!isIplt || config->zRetpolineplt ? target->pltHeaderSize : 0),
2346 // The PLT needs to be writable on SPARC as the dynamic linker will
2347 // modify the instructions in the PLT entries.
2348 if (config->emachine == EM_SPARCV9)
2349 this->flags |= SHF_WRITE;
2352 void PltSection::writeTo(uint8_t *buf) {
2353 if (config->emachine == EM_PPC) {
2354 writePPC32GlinkSection(buf, entries.size());
2358 // At beginning of PLT or retpoline IPLT, we have code to call the dynamic
2359 // linker to resolve dynsyms at runtime. Write such code.
2361 target->writePltHeader(buf);
2362 size_t off = headerSize;
2364 RelocationBaseSection *relSec = isIplt ? in.relaIplt : in.relaPlt;
2366 // The IPlt is immediately after the Plt, account for this in relOff
2367 size_t pltOff = isIplt ? in.plt->getSize() : 0;
2369 for (size_t i = 0, e = entries.size(); i != e; ++i) {
2370 const Symbol *b = entries[i];
2371 unsigned relOff = relSec->entsize * i + pltOff;
2372 uint64_t got = b->getGotPltVA();
2373 uint64_t plt = this->getVA() + off;
2374 target->writePlt(buf + off, got, plt, b->pltIndex, relOff);
2375 off += target->pltEntrySize;
2379 template <class ELFT> void PltSection::addEntry(Symbol &sym) {
2380 sym.pltIndex = entries.size();
2381 entries.push_back(&sym);
2384 size_t PltSection::getSize() const {
2385 return headerSize + entries.size() * target->pltEntrySize;
2388 // Some architectures such as additional symbols in the PLT section. For
2389 // example ARM uses mapping symbols to aid disassembly
2390 void PltSection::addSymbols() {
2391 // The PLT may have symbols defined for the Header, the IPLT has no header
2393 target->addPltHeaderSymbols(*this);
2395 size_t off = headerSize;
2396 for (size_t i = 0; i < entries.size(); ++i) {
2397 target->addPltSymbols(*this, off);
2398 off += target->pltEntrySize;
2402 // The string hash function for .gdb_index.
2403 static uint32_t computeGdbHash(StringRef s) {
2406 h = h * 67 + toLower(c) - 113;
2410 GdbIndexSection::GdbIndexSection()
2411 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index") {}
2413 // Returns the desired size of an on-disk hash table for a .gdb_index section.
2414 // There's a tradeoff between size and collision rate. We aim 75% utilization.
2415 size_t GdbIndexSection::computeSymtabSize() const {
2416 return std::max<size_t>(NextPowerOf2(symbols.size() * 4 / 3), 1024);
2419 // Compute the output section size.
2420 void GdbIndexSection::initOutputSize() {
2421 size = sizeof(GdbIndexHeader) + computeSymtabSize() * 8;
2423 for (GdbChunk &chunk : chunks)
2424 size += chunk.compilationUnits.size() * 16 + chunk.addressAreas.size() * 20;
2426 // Add the constant pool size if exists.
2427 if (!symbols.empty()) {
2428 GdbSymbol &sym = symbols.back();
2429 size += sym.nameOff + sym.name.size() + 1;
2433 static std::vector<InputSection *> getDebugInfoSections() {
2434 std::vector<InputSection *> ret;
2435 for (InputSectionBase *s : inputSections)
2436 if (InputSection *isec = dyn_cast<InputSection>(s))
2437 if (isec->name == ".debug_info")
2438 ret.push_back(isec);
2442 static std::vector<GdbIndexSection::CuEntry> readCuList(DWARFContext &dwarf) {
2443 std::vector<GdbIndexSection::CuEntry> ret;
2444 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units())
2445 ret.push_back({cu->getOffset(), cu->getLength() + 4});
2449 static std::vector<GdbIndexSection::AddressEntry>
2450 readAddressAreas(DWARFContext &dwarf, InputSection *sec) {
2451 std::vector<GdbIndexSection::AddressEntry> ret;
2454 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) {
2455 Expected<DWARFAddressRangesVector> ranges = cu->collectAddressRanges();
2457 error(toString(sec) + ": " + toString(ranges.takeError()));
2461 ArrayRef<InputSectionBase *> sections = sec->file->getSections();
2462 for (DWARFAddressRange &r : *ranges) {
2463 if (r.SectionIndex == -1ULL)
2465 InputSectionBase *s = sections[r.SectionIndex];
2466 if (!s || s == &InputSection::discarded || !s->isLive())
2468 // Range list with zero size has no effect.
2469 if (r.LowPC == r.HighPC)
2471 auto *isec = cast<InputSection>(s);
2472 uint64_t offset = isec->getOffsetInFile();
2473 ret.push_back({isec, r.LowPC - offset, r.HighPC - offset, cuIdx});
2481 template <class ELFT>
2482 static std::vector<GdbIndexSection::NameAttrEntry>
2483 readPubNamesAndTypes(const LLDDwarfObj<ELFT> &obj,
2484 const std::vector<GdbIndexSection::CuEntry> &cUs) {
2485 const DWARFSection &pubNames = obj.getGnuPubNamesSection();
2486 const DWARFSection &pubTypes = obj.getGnuPubTypesSection();
2488 std::vector<GdbIndexSection::NameAttrEntry> ret;
2489 for (const DWARFSection *pub : {&pubNames, &pubTypes}) {
2490 DWARFDebugPubTable table(obj, *pub, config->isLE, true);
2491 for (const DWARFDebugPubTable::Set &set : table.getData()) {
2492 // The value written into the constant pool is kind << 24 | cuIndex. As we
2493 // don't know how many compilation units precede this object to compute
2494 // cuIndex, we compute (kind << 24 | cuIndexInThisObject) instead, and add
2495 // the number of preceding compilation units later.
2497 lower_bound(cUs, set.Offset,
2498 [](GdbIndexSection::CuEntry cu, uint32_t offset) {
2499 return cu.cuOffset < offset;
2502 for (const DWARFDebugPubTable::Entry &ent : set.Entries)
2503 ret.push_back({{ent.Name, computeGdbHash(ent.Name)},
2504 (ent.Descriptor.toBits() << 24) | i});
2510 // Create a list of symbols from a given list of symbol names and types
2511 // by uniquifying them by name.
2512 static std::vector<GdbIndexSection::GdbSymbol>
2513 createSymbols(ArrayRef<std::vector<GdbIndexSection::NameAttrEntry>> nameAttrs,
2514 const std::vector<GdbIndexSection::GdbChunk> &chunks) {
2515 using GdbSymbol = GdbIndexSection::GdbSymbol;
2516 using NameAttrEntry = GdbIndexSection::NameAttrEntry;
2518 // For each chunk, compute the number of compilation units preceding it.
2520 std::vector<uint32_t> cuIdxs(chunks.size());
2521 for (uint32_t i = 0, e = chunks.size(); i != e; ++i) {
2523 cuIdx += chunks[i].compilationUnits.size();
2526 // The number of symbols we will handle in this function is of the order
2527 // of millions for very large executables, so we use multi-threading to
2529 size_t numShards = 32;
2530 size_t concurrency = 1;
2533 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), numShards);
2535 // A sharded map to uniquify symbols by name.
2536 std::vector<DenseMap<CachedHashStringRef, size_t>> map(numShards);
2537 size_t shift = 32 - countTrailingZeros(numShards);
2539 // Instantiate GdbSymbols while uniqufying them by name.
2540 std::vector<std::vector<GdbSymbol>> symbols(numShards);
2541 parallelForEachN(0, concurrency, [&](size_t threadId) {
2543 for (ArrayRef<NameAttrEntry> entries : nameAttrs) {
2544 for (const NameAttrEntry &ent : entries) {
2545 size_t shardId = ent.name.hash() >> shift;
2546 if ((shardId & (concurrency - 1)) != threadId)
2549 uint32_t v = ent.cuIndexAndAttrs + cuIdxs[i];
2550 size_t &idx = map[shardId][ent.name];
2552 symbols[shardId][idx - 1].cuVector.push_back(v);
2556 idx = symbols[shardId].size() + 1;
2557 symbols[shardId].push_back({ent.name, {v}, 0, 0});
2563 size_t numSymbols = 0;
2564 for (ArrayRef<GdbSymbol> v : symbols)
2565 numSymbols += v.size();
2567 // The return type is a flattened vector, so we'll copy each vector
2569 std::vector<GdbSymbol> ret;
2570 ret.reserve(numSymbols);
2571 for (std::vector<GdbSymbol> &vec : symbols)
2572 for (GdbSymbol &sym : vec)
2573 ret.push_back(std::move(sym));
2575 // CU vectors and symbol names are adjacent in the output file.
2576 // We can compute their offsets in the output file now.
2578 for (GdbSymbol &sym : ret) {
2579 sym.cuVectorOff = off;
2580 off += (sym.cuVector.size() + 1) * 4;
2582 for (GdbSymbol &sym : ret) {
2584 off += sym.name.size() + 1;
2590 // Returns a newly-created .gdb_index section.
2591 template <class ELFT> GdbIndexSection *GdbIndexSection::create() {
2592 std::vector<InputSection *> sections = getDebugInfoSections();
2594 // .debug_gnu_pub{names,types} are useless in executables.
2595 // They are present in input object files solely for creating
2596 // a .gdb_index. So we can remove them from the output.
2597 for (InputSectionBase *s : inputSections)
2598 if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes")
2601 std::vector<GdbChunk> chunks(sections.size());
2602 std::vector<std::vector<NameAttrEntry>> nameAttrs(sections.size());
2604 parallelForEachN(0, sections.size(), [&](size_t i) {
2605 ObjFile<ELFT> *file = sections[i]->getFile<ELFT>();
2606 DWARFContext dwarf(make_unique<LLDDwarfObj<ELFT>>(file));
2608 chunks[i].sec = sections[i];
2609 chunks[i].compilationUnits = readCuList(dwarf);
2610 chunks[i].addressAreas = readAddressAreas(dwarf, sections[i]);
2611 nameAttrs[i] = readPubNamesAndTypes<ELFT>(
2612 static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj()),
2613 chunks[i].compilationUnits);
2616 auto *ret = make<GdbIndexSection>();
2617 ret->chunks = std::move(chunks);
2618 ret->symbols = createSymbols(nameAttrs, ret->chunks);
2619 ret->initOutputSize();
2623 void GdbIndexSection::writeTo(uint8_t *buf) {
2624 // Write the header.
2625 auto *hdr = reinterpret_cast<GdbIndexHeader *>(buf);
2626 uint8_t *start = buf;
2628 buf += sizeof(*hdr);
2630 // Write the CU list.
2631 hdr->cuListOff = buf - start;
2632 for (GdbChunk &chunk : chunks) {
2633 for (CuEntry &cu : chunk.compilationUnits) {
2634 write64le(buf, chunk.sec->outSecOff + cu.cuOffset);
2635 write64le(buf + 8, cu.cuLength);
2640 // Write the address area.
2641 hdr->cuTypesOff = buf - start;
2642 hdr->addressAreaOff = buf - start;
2644 for (GdbChunk &chunk : chunks) {
2645 for (AddressEntry &e : chunk.addressAreas) {
2646 uint64_t baseAddr = e.section->getVA(0);
2647 write64le(buf, baseAddr + e.lowAddress);
2648 write64le(buf + 8, baseAddr + e.highAddress);
2649 write32le(buf + 16, e.cuIndex + cuOff);
2652 cuOff += chunk.compilationUnits.size();
2655 // Write the on-disk open-addressing hash table containing symbols.
2656 hdr->symtabOff = buf - start;
2657 size_t symtabSize = computeSymtabSize();
2658 uint32_t mask = symtabSize - 1;
2660 for (GdbSymbol &sym : symbols) {
2661 uint32_t h = sym.name.hash();
2662 uint32_t i = h & mask;
2663 uint32_t step = ((h * 17) & mask) | 1;
2665 while (read32le(buf + i * 8))
2666 i = (i + step) & mask;
2668 write32le(buf + i * 8, sym.nameOff);
2669 write32le(buf + i * 8 + 4, sym.cuVectorOff);
2672 buf += symtabSize * 8;
2674 // Write the string pool.
2675 hdr->constantPoolOff = buf - start;
2676 parallelForEach(symbols, [&](GdbSymbol &sym) {
2677 memcpy(buf + sym.nameOff, sym.name.data(), sym.name.size());
2680 // Write the CU vectors.
2681 for (GdbSymbol &sym : symbols) {
2682 write32le(buf, sym.cuVector.size());
2684 for (uint32_t val : sym.cuVector) {
2685 write32le(buf, val);
2691 bool GdbIndexSection::isNeeded() const { return !chunks.empty(); }
2693 EhFrameHeader::EhFrameHeader()
2694 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr") {}
2696 void EhFrameHeader::writeTo(uint8_t *buf) {
2697 // Unlike most sections, the EhFrameHeader section is written while writing
2698 // another section, namely EhFrameSection, which calls the write() function
2699 // below from its writeTo() function. This is necessary because the contents
2700 // of EhFrameHeader depend on the relocated contents of EhFrameSection and we
2701 // don't know which order the sections will be written in.
2704 // .eh_frame_hdr contains a binary search table of pointers to FDEs.
2705 // Each entry of the search table consists of two values,
2706 // the starting PC from where FDEs covers, and the FDE's address.
2707 // It is sorted by PC.
2708 void EhFrameHeader::write() {
2709 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
2710 using FdeData = EhFrameSection::FdeData;
2712 std::vector<FdeData> fdes = getPartition().ehFrame->getFdeData();
2715 buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4;
2716 buf[2] = DW_EH_PE_udata4;
2717 buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4;
2719 getPartition().ehFrame->getParent()->addr - this->getVA() - 4);
2720 write32(buf + 8, fdes.size());
2723 for (FdeData &fde : fdes) {
2724 write32(buf, fde.pcRel);
2725 write32(buf + 4, fde.fdeVARel);
2730 size_t EhFrameHeader::getSize() const {
2731 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs.
2732 return 12 + getPartition().ehFrame->numFdes * 8;
2735 bool EhFrameHeader::isNeeded() const {
2736 return isLive() && getPartition().ehFrame->isNeeded();
2739 VersionDefinitionSection::VersionDefinitionSection()
2740 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t),
2741 ".gnu.version_d") {}
2743 StringRef VersionDefinitionSection::getFileDefName() {
2744 if (!getPartition().name.empty())
2745 return getPartition().name;
2746 if (!config->soName.empty())
2747 return config->soName;
2748 return config->outputFile;
2751 void VersionDefinitionSection::finalizeContents() {
2752 fileDefNameOff = getPartition().dynStrTab->addString(getFileDefName());
2753 for (VersionDefinition &v : config->versionDefinitions)
2754 verDefNameOffs.push_back(getPartition().dynStrTab->addString(v.name));
2756 if (OutputSection *sec = getPartition().dynStrTab->getParent())
2757 getParent()->link = sec->sectionIndex;
2759 // sh_info should be set to the number of definitions. This fact is missed in
2760 // documentation, but confirmed by binutils community:
2761 // https://sourceware.org/ml/binutils/2014-11/msg00355.html
2762 getParent()->info = getVerDefNum();
2765 void VersionDefinitionSection::writeOne(uint8_t *buf, uint32_t index,
2766 StringRef name, size_t nameOff) {
2767 uint16_t flags = index == 1 ? VER_FLG_BASE : 0;
2770 write16(buf, 1); // vd_version
2771 write16(buf + 2, flags); // vd_flags
2772 write16(buf + 4, index); // vd_ndx
2773 write16(buf + 6, 1); // vd_cnt
2774 write32(buf + 8, hashSysV(name)); // vd_hash
2775 write32(buf + 12, 20); // vd_aux
2776 write32(buf + 16, 28); // vd_next
2779 write32(buf + 20, nameOff); // vda_name
2780 write32(buf + 24, 0); // vda_next
2783 void VersionDefinitionSection::writeTo(uint8_t *buf) {
2784 writeOne(buf, 1, getFileDefName(), fileDefNameOff);
2786 auto nameOffIt = verDefNameOffs.begin();
2787 for (VersionDefinition &v : config->versionDefinitions) {
2789 writeOne(buf, v.id, v.name, *nameOffIt++);
2792 // Need to terminate the last version definition.
2793 write32(buf + 16, 0); // vd_next
2796 size_t VersionDefinitionSection::getSize() const {
2797 return EntrySize * getVerDefNum();
2800 // .gnu.version is a table where each entry is 2 byte long.
2801 VersionTableSection::VersionTableSection()
2802 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t),
2807 void VersionTableSection::finalizeContents() {
2808 // At the moment of june 2016 GNU docs does not mention that sh_link field
2809 // should be set, but Sun docs do. Also readelf relies on this field.
2810 getParent()->link = getPartition().dynSymTab->getParent()->sectionIndex;
2813 size_t VersionTableSection::getSize() const {
2814 return (getPartition().dynSymTab->getSymbols().size() + 1) * 2;
2817 void VersionTableSection::writeTo(uint8_t *buf) {
2819 for (const SymbolTableEntry &s : getPartition().dynSymTab->getSymbols()) {
2820 write16(buf, s.sym->versionId);
2825 bool VersionTableSection::isNeeded() const {
2826 return getPartition().verDef || getPartition().verNeed->isNeeded();
2829 void elf::addVerneed(Symbol *ss) {
2830 auto &file = cast<SharedFile>(*ss->file);
2831 if (ss->verdefIndex == VER_NDX_GLOBAL) {
2832 ss->versionId = VER_NDX_GLOBAL;
2836 if (file.vernauxs.empty())
2837 file.vernauxs.resize(file.verdefs.size());
2839 // Select a version identifier for the vernaux data structure, if we haven't
2840 // already allocated one. The verdef identifiers cover the range
2841 // [1..getVerDefNum()]; this causes the vernaux identifiers to start from
2842 // getVerDefNum()+1.
2843 if (file.vernauxs[ss->verdefIndex] == 0)
2844 file.vernauxs[ss->verdefIndex] = ++SharedFile::vernauxNum + getVerDefNum();
2846 ss->versionId = file.vernauxs[ss->verdefIndex];
2849 template <class ELFT>
2850 VersionNeedSection<ELFT>::VersionNeedSection()
2851 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t),
2852 ".gnu.version_r") {}
2854 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() {
2855 for (SharedFile *f : sharedFiles) {
2856 if (f->vernauxs.empty())
2858 verneeds.emplace_back();
2859 Verneed &vn = verneeds.back();
2860 vn.nameStrTab = getPartition().dynStrTab->addString(f->soName);
2861 for (unsigned i = 0; i != f->vernauxs.size(); ++i) {
2862 if (f->vernauxs[i] == 0)
2865 reinterpret_cast<const typename ELFT::Verdef *>(f->verdefs[i]);
2866 vn.vernauxs.push_back(
2867 {verdef->vd_hash, f->vernauxs[i],
2868 getPartition().dynStrTab->addString(f->getStringTable().data() +
2869 verdef->getAux()->vda_name)});
2873 if (OutputSection *sec = getPartition().dynStrTab->getParent())
2874 getParent()->link = sec->sectionIndex;
2875 getParent()->info = verneeds.size();
2878 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *buf) {
2879 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs.
2880 auto *verneed = reinterpret_cast<Elf_Verneed *>(buf);
2881 auto *vernaux = reinterpret_cast<Elf_Vernaux *>(verneed + verneeds.size());
2883 for (auto &vn : verneeds) {
2884 // Create an Elf_Verneed for this DSO.
2885 verneed->vn_version = 1;
2886 verneed->vn_cnt = vn.vernauxs.size();
2887 verneed->vn_file = vn.nameStrTab;
2889 reinterpret_cast<char *>(vernaux) - reinterpret_cast<char *>(verneed);
2890 verneed->vn_next = sizeof(Elf_Verneed);
2893 // Create the Elf_Vernauxs for this Elf_Verneed.
2894 for (auto &vna : vn.vernauxs) {
2895 vernaux->vna_hash = vna.hash;
2896 vernaux->vna_flags = 0;
2897 vernaux->vna_other = vna.verneedIndex;
2898 vernaux->vna_name = vna.nameStrTab;
2899 vernaux->vna_next = sizeof(Elf_Vernaux);
2903 vernaux[-1].vna_next = 0;
2905 verneed[-1].vn_next = 0;
2908 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const {
2909 return verneeds.size() * sizeof(Elf_Verneed) +
2910 SharedFile::vernauxNum * sizeof(Elf_Vernaux);
2913 template <class ELFT> bool VersionNeedSection<ELFT>::isNeeded() const {
2914 return SharedFile::vernauxNum != 0;
2917 void MergeSyntheticSection::addSection(MergeInputSection *ms) {
2919 sections.push_back(ms);
2920 assert(alignment == ms->alignment || !(ms->flags & SHF_STRINGS));
2921 alignment = std::max(alignment, ms->alignment);
2924 MergeTailSection::MergeTailSection(StringRef name, uint32_t type,
2925 uint64_t flags, uint32_t alignment)
2926 : MergeSyntheticSection(name, type, flags, alignment),
2927 builder(StringTableBuilder::RAW, alignment) {}
2929 size_t MergeTailSection::getSize() const { return builder.getSize(); }
2931 void MergeTailSection::writeTo(uint8_t *buf) { builder.write(buf); }
2933 void MergeTailSection::finalizeContents() {
2934 // Add all string pieces to the string table builder to create section
2936 for (MergeInputSection *sec : sections)
2937 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
2938 if (sec->pieces[i].live)
2939 builder.add(sec->getData(i));
2941 // Fix the string table content. After this, the contents will never change.
2944 // finalize() fixed tail-optimized strings, so we can now get
2945 // offsets of strings. Get an offset for each string and save it
2946 // to a corresponding SectionPiece for easy access.
2947 for (MergeInputSection *sec : sections)
2948 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
2949 if (sec->pieces[i].live)
2950 sec->pieces[i].outputOff = builder.getOffset(sec->getData(i));
2953 void MergeNoTailSection::writeTo(uint8_t *buf) {
2954 for (size_t i = 0; i < numShards; ++i)
2955 shards[i].write(buf + shardOffsets[i]);
2958 // This function is very hot (i.e. it can take several seconds to finish)
2959 // because sometimes the number of inputs is in an order of magnitude of
2960 // millions. So, we use multi-threading.
2962 // For any strings S and T, we know S is not mergeable with T if S's hash
2963 // value is different from T's. If that's the case, we can safely put S and
2964 // T into different string builders without worrying about merge misses.
2965 // We do it in parallel.
2966 void MergeNoTailSection::finalizeContents() {
2967 // Initializes string table builders.
2968 for (size_t i = 0; i < numShards; ++i)
2969 shards.emplace_back(StringTableBuilder::RAW, alignment);
2971 // Concurrency level. Must be a power of 2 to avoid expensive modulo
2972 // operations in the following tight loop.
2973 size_t concurrency = 1;
2976 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), numShards);
2978 // Add section pieces to the builders.
2979 parallelForEachN(0, concurrency, [&](size_t threadId) {
2980 for (MergeInputSection *sec : sections) {
2981 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) {
2982 if (!sec->pieces[i].live)
2984 size_t shardId = getShardId(sec->pieces[i].hash);
2985 if ((shardId & (concurrency - 1)) == threadId)
2986 sec->pieces[i].outputOff = shards[shardId].add(sec->getData(i));
2991 // Compute an in-section offset for each shard.
2993 for (size_t i = 0; i < numShards; ++i) {
2994 shards[i].finalizeInOrder();
2995 if (shards[i].getSize() > 0)
2996 off = alignTo(off, alignment);
2997 shardOffsets[i] = off;
2998 off += shards[i].getSize();
3002 // So far, section pieces have offsets from beginning of shards, but
3003 // we want offsets from beginning of the whole section. Fix them.
3004 parallelForEach(sections, [&](MergeInputSection *sec) {
3005 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3006 if (sec->pieces[i].live)
3007 sec->pieces[i].outputOff +=
3008 shardOffsets[getShardId(sec->pieces[i].hash)];
3012 static MergeSyntheticSection *createMergeSynthetic(StringRef name,
3015 uint32_t alignment) {
3016 bool shouldTailMerge = (flags & SHF_STRINGS) && config->optimize >= 2;
3017 if (shouldTailMerge)
3018 return make<MergeTailSection>(name, type, flags, alignment);
3019 return make<MergeNoTailSection>(name, type, flags, alignment);
3022 template <class ELFT> void elf::splitSections() {
3023 // splitIntoPieces needs to be called on each MergeInputSection
3024 // before calling finalizeContents().
3025 parallelForEach(inputSections, [](InputSectionBase *sec) {
3026 if (auto *s = dyn_cast<MergeInputSection>(sec))
3027 s->splitIntoPieces();
3028 else if (auto *eh = dyn_cast<EhInputSection>(sec))
3033 // This function scans over the inputsections to create mergeable
3034 // synthetic sections.
3036 // It removes MergeInputSections from the input section array and adds
3037 // new synthetic sections at the location of the first input section
3038 // that it replaces. It then finalizes each synthetic section in order
3039 // to compute an output offset for each piece of each input section.
3040 void elf::mergeSections() {
3041 std::vector<MergeSyntheticSection *> mergeSections;
3042 for (InputSectionBase *&s : inputSections) {
3043 MergeInputSection *ms = dyn_cast<MergeInputSection>(s);
3047 // We do not want to handle sections that are not alive, so just remove
3048 // them instead of trying to merge.
3049 if (!ms->isLive()) {
3054 StringRef outsecName = getOutputSectionName(ms);
3056 auto i = llvm::find_if(mergeSections, [=](MergeSyntheticSection *sec) {
3057 // While we could create a single synthetic section for two different
3058 // values of Entsize, it is better to take Entsize into consideration.
3060 // With a single synthetic section no two pieces with different Entsize
3061 // could be equal, so we may as well have two sections.
3063 // Using Entsize in here also allows us to propagate it to the synthetic
3066 // SHF_STRINGS section with different alignments should not be merged.
3067 return sec->name == outsecName && sec->flags == ms->flags &&
3068 sec->entsize == ms->entsize &&
3069 (sec->alignment == ms->alignment || !(sec->flags & SHF_STRINGS));
3071 if (i == mergeSections.end()) {
3072 MergeSyntheticSection *syn =
3073 createMergeSynthetic(outsecName, ms->type, ms->flags, ms->alignment);
3074 mergeSections.push_back(syn);
3075 i = std::prev(mergeSections.end());
3077 syn->entsize = ms->entsize;
3081 (*i)->addSection(ms);
3083 for (auto *ms : mergeSections)
3084 ms->finalizeContents();
3086 std::vector<InputSectionBase *> &v = inputSections;
3087 v.erase(std::remove(v.begin(), v.end(), nullptr), v.end());
3090 MipsRldMapSection::MipsRldMapSection()
3091 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
3094 ARMExidxSyntheticSection::ARMExidxSyntheticSection()
3095 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX,
3096 config->wordsize, ".ARM.exidx") {}
3098 static InputSection *findExidxSection(InputSection *isec) {
3099 for (InputSection *d : isec->dependentSections)
3100 if (d->type == SHT_ARM_EXIDX)
3105 bool ARMExidxSyntheticSection::addSection(InputSection *isec) {
3106 if (isec->type == SHT_ARM_EXIDX) {
3107 exidxSections.push_back(isec);
3111 if ((isec->flags & SHF_ALLOC) && (isec->flags & SHF_EXECINSTR) &&
3112 isec->getSize() > 0) {
3113 executableSections.push_back(isec);
3114 if (empty && findExidxSection(isec))
3119 // FIXME: we do not output a relocation section when --emit-relocs is used
3120 // as we do not have relocation sections for linker generated table entries
3121 // and we would have to erase at a late stage relocations from merged entries.
3122 // Given that exception tables are already position independent and a binary
3123 // analyzer could derive the relocations we choose to erase the relocations.
3124 if (config->emitRelocs && isec->type == SHT_REL)
3125 if (InputSectionBase *ex = isec->getRelocatedSection())
3126 if (isa<InputSection>(ex) && ex->type == SHT_ARM_EXIDX)
3132 // References to .ARM.Extab Sections have bit 31 clear and are not the
3133 // special EXIDX_CANTUNWIND bit-pattern.
3134 static bool isExtabRef(uint32_t unwind) {
3135 return (unwind & 0x80000000) == 0 && unwind != 0x1;
3138 // Return true if the .ARM.exidx section Cur can be merged into the .ARM.exidx
3139 // section Prev, where Cur follows Prev in the table. This can be done if the
3140 // unwinding instructions in Cur are identical to Prev. Linker generated
3141 // EXIDX_CANTUNWIND entries are represented by nullptr as they do not have an
3143 static bool isDuplicateArmExidxSec(InputSection *prev, InputSection *cur) {
3149 // Get the last table Entry from the previous .ARM.exidx section. If Prev is
3150 // nullptr then it will be a synthesized EXIDX_CANTUNWIND entry.
3151 ExidxEntry prevEntry = {ulittle32_t(0), ulittle32_t(1)};
3153 prevEntry = prev->getDataAs<ExidxEntry>().back();
3154 if (isExtabRef(prevEntry.unwind))
3157 // We consider the unwind instructions of an .ARM.exidx table entry
3158 // a duplicate if the previous unwind instructions if:
3159 // - Both are the special EXIDX_CANTUNWIND.
3160 // - Both are the same inline unwind instructions.
3161 // We do not attempt to follow and check links into .ARM.extab tables as
3162 // consecutive identical entries are rare and the effort to check that they
3163 // are identical is high.
3165 // If Cur is nullptr then this is synthesized EXIDX_CANTUNWIND entry.
3167 return prevEntry.unwind == 1;
3169 for (const ExidxEntry entry : cur->getDataAs<ExidxEntry>())
3170 if (isExtabRef(entry.unwind) || entry.unwind != prevEntry.unwind)
3173 // All table entries in this .ARM.exidx Section can be merged into the
3174 // previous Section.
3178 // The .ARM.exidx table must be sorted in ascending order of the address of the
3179 // functions the table describes. Optionally duplicate adjacent table entries
3180 // can be removed. At the end of the function the executableSections must be
3181 // sorted in ascending order of address, Sentinel is set to the InputSection
3182 // with the highest address and any InputSections that have mergeable
3183 // .ARM.exidx table entries are removed from it.
3184 void ARMExidxSyntheticSection::finalizeContents() {
3185 if (script->hasSectionsCommand) {
3186 // The executableSections and exidxSections that we use to derive the
3187 // final contents of this SyntheticSection are populated before the
3188 // linker script assigns InputSections to OutputSections. The linker script
3189 // SECTIONS command may have a /DISCARD/ entry that removes executable
3190 // InputSections and their dependent .ARM.exidx section that we recorded
3192 auto isDiscarded = [](const InputSection *isec) { return !isec->isLive(); };
3193 llvm::erase_if(executableSections, isDiscarded);
3194 llvm::erase_if(exidxSections, isDiscarded);
3197 // Sort the executable sections that may or may not have associated
3198 // .ARM.exidx sections by order of ascending address. This requires the
3199 // relative positions of InputSections to be known.
3200 auto compareByFilePosition = [](const InputSection *a,
3201 const InputSection *b) {
3202 OutputSection *aOut = a->getParent();
3203 OutputSection *bOut = b->getParent();
3206 return aOut->sectionIndex < bOut->sectionIndex;
3207 return a->outSecOff < b->outSecOff;
3209 llvm::stable_sort(executableSections, compareByFilePosition);
3210 sentinel = executableSections.back();
3211 // Optionally merge adjacent duplicate entries.
3212 if (config->mergeArmExidx) {
3213 std::vector<InputSection *> selectedSections;
3214 selectedSections.reserve(executableSections.size());
3215 selectedSections.push_back(executableSections[0]);
3217 for (size_t i = 1; i < executableSections.size(); ++i) {
3218 InputSection *ex1 = findExidxSection(executableSections[prev]);
3219 InputSection *ex2 = findExidxSection(executableSections[i]);
3220 if (!isDuplicateArmExidxSec(ex1, ex2)) {
3221 selectedSections.push_back(executableSections[i]);
3225 executableSections = std::move(selectedSections);
3230 for (InputSection *isec : executableSections) {
3231 if (InputSection *d = findExidxSection(isec)) {
3232 d->outSecOff = offset;
3233 d->parent = getParent();
3234 offset += d->getSize();
3239 // Size includes Sentinel.
3243 InputSection *ARMExidxSyntheticSection::getLinkOrderDep() const {
3244 return executableSections.front();
3247 // To write the .ARM.exidx table from the ExecutableSections we have three cases
3248 // 1.) The InputSection has a .ARM.exidx InputSection in its dependent sections.
3249 // We write the .ARM.exidx section contents and apply its relocations.
3250 // 2.) The InputSection does not have a dependent .ARM.exidx InputSection. We
3251 // must write the contents of an EXIDX_CANTUNWIND directly. We use the
3252 // start of the InputSection as the purpose of the linker generated
3253 // section is to terminate the address range of the previous entry.
3254 // 3.) A trailing EXIDX_CANTUNWIND sentinel section is required at the end of
3255 // the table to terminate the address range of the final entry.
3256 void ARMExidxSyntheticSection::writeTo(uint8_t *buf) {
3258 const uint8_t cantUnwindData[8] = {0, 0, 0, 0, // PREL31 to target
3259 1, 0, 0, 0}; // EXIDX_CANTUNWIND
3261 uint64_t offset = 0;
3262 for (InputSection *isec : executableSections) {
3263 assert(isec->getParent() != nullptr);
3264 if (InputSection *d = findExidxSection(isec)) {
3265 memcpy(buf + offset, d->data().data(), d->data().size());
3266 d->relocateAlloc(buf, buf + d->getSize());
3267 offset += d->getSize();
3269 // A Linker generated CANTUNWIND section.
3270 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3271 uint64_t s = isec->getVA();
3272 uint64_t p = getVA() + offset;
3273 target->relocateOne(buf + offset, R_ARM_PREL31, s - p);
3278 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3279 uint64_t s = sentinel->getVA(sentinel->getSize());
3280 uint64_t p = getVA() + offset;
3281 target->relocateOne(buf + offset, R_ARM_PREL31, s - p);
3282 assert(size == offset + 8);
3285 bool ARMExidxSyntheticSection::classof(const SectionBase *d) {
3286 return d->kind() == InputSectionBase::Synthetic && d->type == SHT_ARM_EXIDX;
3289 ThunkSection::ThunkSection(OutputSection *os, uint64_t off)
3290 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS,
3291 config->wordsize, ".text.thunk") {
3293 this->outSecOff = off;
3296 void ThunkSection::addThunk(Thunk *t) {
3297 thunks.push_back(t);
3298 t->addSymbols(*this);
3301 void ThunkSection::writeTo(uint8_t *buf) {
3302 for (Thunk *t : thunks)
3303 t->writeTo(buf + t->offset);
3306 InputSection *ThunkSection::getTargetInputSection() const {
3309 const Thunk *t = thunks.front();
3310 return t->getTargetInputSection();
3313 bool ThunkSection::assignOffsets() {
3315 for (Thunk *t : thunks) {
3316 off = alignTo(off, t->alignment);
3318 uint32_t size = t->size();
3319 t->getThunkTargetSym()->size = size;
3322 bool changed = off != size;
3327 PPC32Got2Section::PPC32Got2Section()
3328 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 4, ".got2") {}
3330 bool PPC32Got2Section::isNeeded() const {
3331 // See the comment below. This is not needed if there is no other
3333 for (BaseCommand *base : getParent()->sectionCommands)
3334 if (auto *isd = dyn_cast<InputSectionDescription>(base))
3335 for (InputSection *isec : isd->sections)
3341 void PPC32Got2Section::finalizeContents() {
3342 // PPC32 may create multiple GOT sections for -fPIC/-fPIE, one per file in
3343 // .got2 . This function computes outSecOff of each .got2 to be used in
3344 // PPC32PltCallStub::writeTo(). The purpose of this empty synthetic section is
3345 // to collect input sections named ".got2".
3346 uint32_t offset = 0;
3347 for (BaseCommand *base : getParent()->sectionCommands)
3348 if (auto *isd = dyn_cast<InputSectionDescription>(base)) {
3349 for (InputSection *isec : isd->sections) {
3352 isec->file->ppc32Got2OutSecOff = offset;
3353 offset += (uint32_t)isec->getSize();
3358 // If linking position-dependent code then the table will store the addresses
3359 // directly in the binary so the section has type SHT_PROGBITS. If linking
3360 // position-independent code the section has type SHT_NOBITS since it will be
3361 // allocated and filled in by the dynamic linker.
3362 PPC64LongBranchTargetSection::PPC64LongBranchTargetSection()
3363 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
3364 config->isPic ? SHT_NOBITS : SHT_PROGBITS, 8,
3367 void PPC64LongBranchTargetSection::addEntry(Symbol &sym) {
3368 assert(sym.ppc64BranchltIndex == 0xffff);
3369 sym.ppc64BranchltIndex = entries.size();
3370 entries.push_back(&sym);
3373 size_t PPC64LongBranchTargetSection::getSize() const {
3374 return entries.size() * 8;
3377 void PPC64LongBranchTargetSection::writeTo(uint8_t *buf) {
3378 // If linking non-pic we have the final addresses of the targets and they get
3379 // written to the table directly. For pic the dynamic linker will allocate
3380 // the section and fill it it.
3384 for (const Symbol *sym : entries) {
3385 assert(sym->getVA());
3386 // Need calls to branch to the local entry-point since a long-branch
3387 // must be a local-call.
3389 sym->getVA() + getPPC64GlobalEntryToLocalEntryOffset(sym->stOther));
3394 bool PPC64LongBranchTargetSection::isNeeded() const {
3395 // `removeUnusedSyntheticSections()` is called before thunk allocation which
3396 // is too early to determine if this section will be empty or not. We need
3397 // Finalized to keep the section alive until after thunk creation. Finalized
3398 // only gets set to true once `finalizeSections()` is called after thunk
3399 // creation. Becuase of this, if we don't create any long-branch thunks we end
3400 // up with an empty .branch_lt section in the binary.
3401 return !finalized || !entries.empty();
3404 RISCVSdataSection::RISCVSdataSection()
3405 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 1, ".sdata") {}
3407 bool RISCVSdataSection::isNeeded() const {
3408 if (!ElfSym::riscvGlobalPointer)
3411 // __global_pointer$ is defined relative to .sdata . If the section does not
3412 // exist, create a dummy one.
3413 for (BaseCommand *base : getParent()->sectionCommands)
3414 if (auto *isd = dyn_cast<InputSectionDescription>(base))
3415 for (InputSection *isec : isd->sections)
3421 static uint8_t getAbiVersion() {
3422 // MIPS non-PIC executable gets ABI version 1.
3423 if (config->emachine == EM_MIPS) {
3424 if (!config->isPic && !config->relocatable &&
3425 (config->eflags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC)
3430 if (config->emachine == EM_AMDGPU) {
3431 uint8_t ver = objectFiles[0]->abiVersion;
3432 for (InputFile *file : makeArrayRef(objectFiles).slice(1))
3433 if (file->abiVersion != ver)
3434 error("incompatible ABI version: " + toString(file));
3441 template <typename ELFT> void elf::writeEhdr(uint8_t *buf, Partition &part) {
3442 // For executable segments, the trap instructions are written before writing
3443 // the header. Setting Elf header bytes to zero ensures that any unused bytes
3444 // in header are zero-cleared, instead of having trap instructions.
3445 memset(buf, 0, sizeof(typename ELFT::Ehdr));
3446 memcpy(buf, "\177ELF", 4);
3448 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3449 eHdr->e_ident[EI_CLASS] = config->is64 ? ELFCLASS64 : ELFCLASS32;
3450 eHdr->e_ident[EI_DATA] = config->isLE ? ELFDATA2LSB : ELFDATA2MSB;
3451 eHdr->e_ident[EI_VERSION] = EV_CURRENT;
3452 eHdr->e_ident[EI_OSABI] = config->osabi;
3453 eHdr->e_ident[EI_ABIVERSION] = getAbiVersion();
3454 eHdr->e_machine = config->emachine;
3455 eHdr->e_version = EV_CURRENT;
3456 eHdr->e_flags = config->eflags;
3457 eHdr->e_ehsize = sizeof(typename ELFT::Ehdr);
3458 eHdr->e_phnum = part.phdrs.size();
3459 eHdr->e_shentsize = sizeof(typename ELFT::Shdr);
3461 if (!config->relocatable) {
3462 eHdr->e_phoff = sizeof(typename ELFT::Ehdr);
3463 eHdr->e_phentsize = sizeof(typename ELFT::Phdr);
3467 template <typename ELFT> void elf::writePhdrs(uint8_t *buf, Partition &part) {
3468 // Write the program header table.
3469 auto *hBuf = reinterpret_cast<typename ELFT::Phdr *>(buf);
3470 for (PhdrEntry *p : part.phdrs) {
3471 hBuf->p_type = p->p_type;
3472 hBuf->p_flags = p->p_flags;
3473 hBuf->p_offset = p->p_offset;
3474 hBuf->p_vaddr = p->p_vaddr;
3475 hBuf->p_paddr = p->p_paddr;
3476 hBuf->p_filesz = p->p_filesz;
3477 hBuf->p_memsz = p->p_memsz;
3478 hBuf->p_align = p->p_align;
3483 template <typename ELFT>
3484 PartitionElfHeaderSection<ELFT>::PartitionElfHeaderSection()
3485 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_EHDR, 1, "") {}
3487 template <typename ELFT>
3488 size_t PartitionElfHeaderSection<ELFT>::getSize() const {
3489 return sizeof(typename ELFT::Ehdr);
3492 template <typename ELFT>
3493 void PartitionElfHeaderSection<ELFT>::writeTo(uint8_t *buf) {
3494 writeEhdr<ELFT>(buf, getPartition());
3496 // Loadable partitions are always ET_DYN.
3497 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3498 eHdr->e_type = ET_DYN;
3501 template <typename ELFT>
3502 PartitionProgramHeadersSection<ELFT>::PartitionProgramHeadersSection()
3503 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_PHDR, 1, ".phdrs") {}
3505 template <typename ELFT>
3506 size_t PartitionProgramHeadersSection<ELFT>::getSize() const {
3507 return sizeof(typename ELFT::Phdr) * getPartition().phdrs.size();
3510 template <typename ELFT>
3511 void PartitionProgramHeadersSection<ELFT>::writeTo(uint8_t *buf) {
3512 writePhdrs<ELFT>(buf, getPartition());
3515 PartitionIndexSection::PartitionIndexSection()
3516 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".rodata") {}
3518 size_t PartitionIndexSection::getSize() const {
3519 return 12 * (partitions.size() - 1);
3522 void PartitionIndexSection::finalizeContents() {
3523 for (size_t i = 1; i != partitions.size(); ++i)
3524 partitions[i].nameStrTab = mainPart->dynStrTab->addString(partitions[i].name);
3527 void PartitionIndexSection::writeTo(uint8_t *buf) {
3528 uint64_t va = getVA();
3529 for (size_t i = 1; i != partitions.size(); ++i) {
3530 write32(buf, mainPart->dynStrTab->getVA() + partitions[i].nameStrTab - va);
3531 write32(buf + 4, partitions[i].elfHeader->getVA() - (va + 4));
3533 SyntheticSection *next =
3534 i == partitions.size() - 1 ? in.partEnd : partitions[i + 1].elfHeader;
3535 write32(buf + 8, next->getVA() - partitions[i].elfHeader->getVA());
3544 std::vector<Partition> elf::partitions;
3545 Partition *elf::mainPart;
3547 template GdbIndexSection *GdbIndexSection::create<ELF32LE>();
3548 template GdbIndexSection *GdbIndexSection::create<ELF32BE>();
3549 template GdbIndexSection *GdbIndexSection::create<ELF64LE>();
3550 template GdbIndexSection *GdbIndexSection::create<ELF64BE>();
3552 template void elf::splitSections<ELF32LE>();
3553 template void elf::splitSections<ELF32BE>();
3554 template void elf::splitSections<ELF64LE>();
3555 template void elf::splitSections<ELF64BE>();
3557 template void EhFrameSection::addSection<ELF32LE>(InputSectionBase *);
3558 template void EhFrameSection::addSection<ELF32BE>(InputSectionBase *);
3559 template void EhFrameSection::addSection<ELF64LE>(InputSectionBase *);
3560 template void EhFrameSection::addSection<ELF64BE>(InputSectionBase *);
3562 template void PltSection::addEntry<ELF32LE>(Symbol &Sym);
3563 template void PltSection::addEntry<ELF32BE>(Symbol &Sym);
3564 template void PltSection::addEntry<ELF64LE>(Symbol &Sym);
3565 template void PltSection::addEntry<ELF64BE>(Symbol &Sym);
3567 template class elf::MipsAbiFlagsSection<ELF32LE>;
3568 template class elf::MipsAbiFlagsSection<ELF32BE>;
3569 template class elf::MipsAbiFlagsSection<ELF64LE>;
3570 template class elf::MipsAbiFlagsSection<ELF64BE>;
3572 template class elf::MipsOptionsSection<ELF32LE>;
3573 template class elf::MipsOptionsSection<ELF32BE>;
3574 template class elf::MipsOptionsSection<ELF64LE>;
3575 template class elf::MipsOptionsSection<ELF64BE>;
3577 template class elf::MipsReginfoSection<ELF32LE>;
3578 template class elf::MipsReginfoSection<ELF32BE>;
3579 template class elf::MipsReginfoSection<ELF64LE>;
3580 template class elf::MipsReginfoSection<ELF64BE>;
3582 template class elf::DynamicSection<ELF32LE>;
3583 template class elf::DynamicSection<ELF32BE>;
3584 template class elf::DynamicSection<ELF64LE>;
3585 template class elf::DynamicSection<ELF64BE>;
3587 template class elf::RelocationSection<ELF32LE>;
3588 template class elf::RelocationSection<ELF32BE>;
3589 template class elf::RelocationSection<ELF64LE>;
3590 template class elf::RelocationSection<ELF64BE>;
3592 template class elf::AndroidPackedRelocationSection<ELF32LE>;
3593 template class elf::AndroidPackedRelocationSection<ELF32BE>;
3594 template class elf::AndroidPackedRelocationSection<ELF64LE>;
3595 template class elf::AndroidPackedRelocationSection<ELF64BE>;
3597 template class elf::RelrSection<ELF32LE>;
3598 template class elf::RelrSection<ELF32BE>;
3599 template class elf::RelrSection<ELF64LE>;
3600 template class elf::RelrSection<ELF64BE>;
3602 template class elf::SymbolTableSection<ELF32LE>;
3603 template class elf::SymbolTableSection<ELF32BE>;
3604 template class elf::SymbolTableSection<ELF64LE>;
3605 template class elf::SymbolTableSection<ELF64BE>;
3607 template class elf::VersionNeedSection<ELF32LE>;
3608 template class elf::VersionNeedSection<ELF32BE>;
3609 template class elf::VersionNeedSection<ELF64LE>;
3610 template class elf::VersionNeedSection<ELF64BE>;
3612 template void elf::writeEhdr<ELF32LE>(uint8_t *Buf, Partition &Part);
3613 template void elf::writeEhdr<ELF32BE>(uint8_t *Buf, Partition &Part);
3614 template void elf::writeEhdr<ELF64LE>(uint8_t *Buf, Partition &Part);
3615 template void elf::writeEhdr<ELF64BE>(uint8_t *Buf, Partition &Part);
3617 template void elf::writePhdrs<ELF32LE>(uint8_t *Buf, Partition &Part);
3618 template void elf::writePhdrs<ELF32BE>(uint8_t *Buf, Partition &Part);
3619 template void elf::writePhdrs<ELF64LE>(uint8_t *Buf, Partition &Part);
3620 template void elf::writePhdrs<ELF64BE>(uint8_t *Buf, Partition &Part);
3622 template class elf::PartitionElfHeaderSection<ELF32LE>;
3623 template class elf::PartitionElfHeaderSection<ELF32BE>;
3624 template class elf::PartitionElfHeaderSection<ELF64LE>;
3625 template class elf::PartitionElfHeaderSection<ELF64BE>;
3627 template class elf::PartitionProgramHeadersSection<ELF32LE>;
3628 template class elf::PartitionProgramHeadersSection<ELF32BE>;
3629 template class elf::PartitionProgramHeadersSection<ELF64LE>;
3630 template class elf::PartitionProgramHeadersSection<ELF64BE>;